content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
\newcounter{conter1}
\setcounter{conter1}{0}
\newtheorem{theorem}[conter1]{Theorem}
\newtheorem{definitions}{Definition}
\newtheorem{theorems}{Theorem}
\newtheorem{lemmas}{Lemma}
\newtheorem{corollarys}{Corollary}
\newtheorem{examples}{Example}
\newtheorem{propositions}{Proposition}
In classical information theory, there are mainly two types of error models: the independent noise model proposed by Shannon and the adversarial noise model considered by Hamming. Errors in these two models are usually called random errors and burst errors. Correspondingly, there are random error-correcting codes (RECC) and burst error-correcting codes (BECC) to deal with these two different types of errors \cite{lin2004error}. In reality, channels tend to introduce errors which are localized in a short interval, i.e., the burst errors. These errors could be commonly found in communication systems and storage mediums, as a result of a stroke of lightning in wireless channels or scratch on a storage disc.
In the quantum regime, quantum errors can be independent or correlated in space and time. Hence there are counterparts of quantum random error-correcting codes {(see, e.g., \cite{steane1996error, calderbank1996good, calderbank1997quantum, Brun436, 6671483}) } and quantum burst error-correcting codes {\cite{jihao2017on, vatan1999spatially, kawabata2000quantum, tokiwa2005some}}. Analogous to the classical case, quantum channels commonly have memory \cite{PhysRevA.72.062323} or introduce errors which are localized \cite{caruso2014quantum}, i.e., quantum burst errors.
Vatan \textit{et al.}~\cite{vatan1999spatially} first considered spatially correlated qubit errors and constructed families of QBECCs using CSS construction \cite{calderbank1996good, calderbank1997quantum}. However, CSS construction yields QBECCs with inferior code rate. In \cite{kawabata2000quantum}, a quantum interleaver for QBECCs was proposed so that long QBECCs could be produced from short ones. However, this method highly relies on short efficient QBECCs, which are lacking at this moment.
In \cite{tokiwa2005some}, QBECCs of length up to $51$ were found using computer search.
The construction and investigation of QBECCs have received far less attention, compared to the development of standard QECCs or entanglement-assisted QECCs \cite{7501532, Brun436, 6461941, 6671483, 5714249, PhysRevA.79.032340, PhysRevA.76.062313}. Many important questions remain open. Currently, there is no general upper bound for correctable quantum burst errors, analygous to the classical Reiger bound: $n-k\geq 2\ell$, of an $[n,k]$ classical BECC, where $n$ is the code length, $k$ is the message size, and $\ell$ is the correctable length of burst errors.
In addition, there is an interesting class of quantum codes, called \emph{degenerate} codes, that have no classical correspondences. They can potentially store more quantum information or correct more quantum errors than nondegenerate codes. However, degenerate QBECCs have never been explored.
In this paper we generalize the theory of standard QECCs to QBECCs. We develop the stabilizer formalism for QBECCs and prove the corresponding quantum Reiger bound: $n-k\geq 4\ell$, for an $[[n,k]]$ QBECC that corrects a quantum burst error of length $\ell$ or less. The quantum Reiger bound further generalizes the quantum Singleton bound in QECCs. We obtain many new QBECCs in the stabilizer formalism via computer heuristic search, and these codes are better than existing QBECCs with the same code lengths. We show that the burst error-correcting abilities of most of these codes can achieve the quantum Reiger bound and thus are perfect codes.
In particular, several of our constructed QBECCs, that attain the quantum Reiger bound, are degenerate codes. Additionally, we propose a new concatenation construction of long QBECCs from two short component codes based on the quantum tensor product code structure \cite{jihao2017on} and the interleaving technique. Since only one of the component codes of QTPCs needs to satisfy the dual containing constraint, this construction method can largely facilitate the systematical construction of QBECCs. Finally, we perform numerical experiments on two of our constructed QBECCs by
measuring the entanglement fidelity over Markovian correlated depolarizing quantum memory channels \cite{caruso2014quantum}. It is known that the correlation errors in memory channels can lower the
performance of the entanglement fidelity of standard QECCs (see \cite{Cafaro2010quantum,caruso2014quantum}). But if we consider the extra burst error correction abilities of them, they can indeed outperform the best QECCs of the same lengths for random errors.
\section{Theory of Quantum Burst Error Correction Codes}
In this section, we introduce background of QECCs and develop the stabilizer formalism for QBECCs.
\subsection{Quantum Burst Error Correction Codes}
In a two-dimensional complex Hilbert space $\mathbb{C}^2$, a qubit $|v\rangle$ can be written as
$
|v\rangle=\alpha|0\rangle+\beta|1\rangle,
$
where $\alpha$ and $\beta$ are complex numbers satisfying $|\alpha|^2+|\beta|^2=1$. Two states $|v\rangle$ and $e^{i\theta}|v\rangle$, that are different up to a global phase $e^{i\theta}$, are considered to be the same in this paper. The Pauli matrices
\begin{equation*}
\label{Pauli_Matrices}
I_2= \left[
\begin{matrix}
1&0\\
0&1\\
\end{matrix}
\right],
X= \left[
\begin{matrix}
0&1\\
1&0\\
\end{matrix}
\right],
Z= \left[
\begin{matrix}
\setlength{\arraycolsep}{0.1pt}
1&0\\
0&-1\\
\end{matrix}
\right],
Y= \left[
\begin{matrix}
\setlength{\arraycolsep}{0.1pt}
0&-i\\
i&0\\
\end{matrix}
\right]
\end{equation*}
form a basis of the linear operators on $\mathbb{C}^2$, where $i=\sqrt{-1}$ and $Y=iXZ$. An $n$-qubit $|\psi\rangle$ is then a quantum state in the $n$-th tensor product of $\mathbb{C}^2$, i.e., $|\psi\rangle \in \mathbb{C}^{2^n}\equiv \mathbb{C}^2\otimes\mathbb{C}^2\otimes\cdots\otimes\mathbb{C}^2$.
Since it is possible to \emph{discretize} quantum errors \cite{nielsen2000quantum, 6778074}, we only need to consider a discrete set of quantum errors of $n$ qubits,
described by the following error group
\begin{equation}
\mathcal{G}_n=\{i^\lambda w_1\otimes\cdots\otimes w_n|0\leq\lambda\leq 3, w_i\in{I_2,X,Y,Z}\}.
\end{equation}
Furthermore, it is sufficient to consider the quotient group $\mathcal{\overline{G}}_n= \mathcal{G}_n/\{\pm1,\pm i\}$ of $\mathcal{G}_n$ since the global phase $i^\lambda$ in $\mathcal{G}_n$ is not important. Let $\overline{e}= w_1\otimes w_2\otimes\cdots\otimes w_n\in\mathcal{\overline{G}}_n$ and $e=i^\lambda\overline{e}\in\mathcal{G}_n$. We define the \emph{burst} length of $\bar{e}$ to be $\ell$, denoted by $\textrm{bl}_Q(e)=\textrm{bl}_Q(\overline{e})=\ell$, if the nonidentity matrices in $\overline{e}$ are confined to at most $\ell$ consecutive $w_i$'s.
The idea of a QECC is to encode quantum information into a subspace of some larger Hilbert space. An $((n,K))$ QECC $Q$ is defined to contain the subspace of dimension $K$ in $\mathbb{C}^{2^n}$. If $K=2^k$, then $Q$ is also written as $Q=[[n,k]]$. A QECC $Q$ can correct arbitrary errors from an error class $\varepsilon$ \cite{bennett1996mixed,knill1997theory} if
\begin{equation}\label{knill77}
\langle c_i|EE'|c_j\rangle=a_{(E,E')}\delta_{ij}
\end{equation}
for all $\langle c_i|c_j\rangle = \delta_{ij}$ and for all $E\neq E' \in\varepsilon$, where $|c_i\rangle$ and $|c_j\rangle\in Q$, and $a_{(E,E')}$ is a constant which depends only on $E$ and $E'$. If $\langle c_i|EE'|c_j\rangle=0$ for all $|c_i\rangle,|c_j\rangle\in Q$ and for all $E\neq E' \in\varepsilon$, then $Q$ is called a \emph{nondegenerate} quantum code.
The above error-correcting condition (\ref{knill77}) can be generalized to the burst error case.
\begin{propositions}
\label{proposition_burst}
The code $Q$ can correct any quantum burst errors of length $l$ or less if and only if
\begin{equation}
\label{burst-error-correction criterion}
\langle c_i|EE'|c_j\rangle=a_{(E,E')}\delta_{ij}
\end{equation}
for all $\langle c_i|c_j\rangle = \delta_{ij}$ and for all $\textrm{bl}(E),\textrm{bl}(E')\leq \ell$, where $|c_i\rangle$ and $|c_j\rangle\in Q$, $E$ and $E'\in\mathcal{G}_n$, and $a_{(E,E')}$ is a constant which depends only on $E$ and $E'$.
If $\langle c_i|EE'|c_j\rangle=0$ for all $|c_i\rangle,|c_j\rangle\in Q$ and for all $\textrm{bl}(E),\textrm{bl}(E')\leq \ell$, where $E\neq E' \in\mathcal{G}_n$, then $Q$ is a nondegenerate QBECC.
\end{propositions}
According to the group theoretic framework for QECCs in \cite{calderbank1997quantum,calderbank1998quantum}, we can also get the stabilizer formalism for QBECCs.
Let $(a|b)$ and $(a'|b')$ be two vectors in $\mathbb{F}_2^{2n}$, the \emph{symplectic inner product} of them is given by
\begin{equation}
((a|b),(a'|b'))_s=a\cdot b'+a'\cdot b.
\end{equation}
For a subspace $C$ of $\mathbb{F}_2^{2n}$, the \emph{symplectic dual} space $C^{\bot_s}$ of $C$ is given by
\begin{equation}
C^{\bot_s}=\{u\in \mathbb{F}_2^{2n}|\forall c\in C,(u,c)_s=0\}.
\end{equation}
We define the \emph{symplectic burst} length of a nonzero vector $(a|b)=(a_1\cdots a_n|b_1\cdots b_n)\in \mathbb{F}_2^{2n}$ to be the
largest integer $1\leq \ell\leq n$ such that $(a_i|b_i)\neq (0,0)$ and $(a_{i+\ell-1}|b_{i+\ell-1})\neq (0,0)$ for some $1\leq i\leq n$. We denote by $\textrm{bl}_s((a|b))=\ell$.
According to \cite{calderbank1998quantum}, each element $E\in\mathcal{G}_n$ can be written
uniquely as $ E=i^\lambda X(a)Z(b)$ where $1\leq\lambda\leq3$, $X(a)|\psi\rangle=|\psi+a\rangle$, $Z(b)|\psi\rangle=(-1)^{b\cdot \psi}|\psi\rangle$, and for $a,b\in\mathbb{F}_2^n$. It is easy to verify that
$\textrm{bl}_Q(E)=\textrm{bl}_Q(\overline{E})=\textrm{bl}_s((a|b))$.
Then we have the group framework for quantum burst error correction codes.
\begin{theorems}
\label{theorem_symplectic}
Suppose that there exists an $(n-k)$-dimensional linear subspace $C$ of $\mathbb{F}_2^{2n}$ which is contained in its symplectic dual $C^{\bot_s}$, i.e., $C\subseteq C^{\bot_s}$. Let $\ell$ be the largest integer such that for arbitrary two vectors $e_1\neq e_2\in\mathbb{F}_2^{2n}$ whose symplectic burst length $\leq \ell$ there is $e_1+e_2\notin C^{\bot_s} \backslash C$. Then there exists a quantum burst error correction code $Q=[[n,k]]$ which can correct arbitrary quantum burst errors of length $\ell$ or less. If all the $e_1+e_2\notin C^{\bot_s} \backslash \{0\}$, then $Q$ is a nondegenerate quantum burst error correction code.
\end{theorems}
\begin{IEEEproof}
The existence of the quantum code $Q$ has been shown in
\cite[Theorem 1]{calderbank1998quantum}. The burst error correction abilities of $Q$ can be obtained directly by combining
\cite[Lemma 1]{calderbank1998quantum} and Proposition \ref{proposition_burst}.
\end{IEEEproof}
As shown in \cite{calderbank1998quantum}, binary quantum codes can be constructed by using additive codes over $\mathbb{F}_4$. Define the
\emph{trace inner product} of two vectors $u,v\in \mathbb{F}_4^n$ by
\begin{equation}
(u,v)_{tr}= \sum_{i=1}^{n}(u_iv_j^2+u_i^2v_j).
\end{equation}
Let $C$ be an additive code over $\mathbb{F}_4$, then the \emph{trace dual} of $C$ with respect to the trace inner product is defined by
\begin{equation}
C^{\bot_{tr}}=\{v\in \mathbb{F}_4^{n}|\forall u\in C,(u,v)_{tr}=0\}.
\end{equation}
Then Theorem \ref{theorem_symplectic} can also be reformulated by using additive codes over $\mathbb{F}_4$ and by replacing the symplectic inner product with trace inner product.
\begin{theorems}
\label{Hermitian QBECCs}
Suppose that $C$ is an additive code over $\mathbb{F}_4$ which is contained in its trace dual $C^{\bot_{tr}}$, i.e., $C\subseteq C^{\bot_{tr}}$. Let $\ell$ be the largest integer such that for arbitrary two vectors $e_1\neq e_2\in\mathbb{F}_4^{n}$ whose burst length $\leq \ell$ there is $e_1+e_2\notin C^{\bot_{tr}} \backslash C$. Then there exists a binary quantum burst error correction code $Q=[[n,k]]$ which can correct arbitrary quantum burst errors of length $\ell$ or less. If all the $e_1+e_2\notin C^{\bot_{tr}} \backslash \{0\}$, then $Q$ is a nondegenerate quantum burst error correction code.
\end{theorems}
The CSS code construction \cite{calderbank1996good,steane1996error} provides a direct way to construct QECCs from classical linear codes. The CSS construction for QBECCs can be obtained from Theorem \ref{theorem_symplectic}.
\begin{corollarys}[CSS Construction]
\label{CSS QBECCs}
Let $C_1=[n,k_1]$ and $C_2=[n,k_2]$ be
two binary linear codes which have $\ell_1$ and $\ell_2$ burst error correction abilities, respectively,
and such that $C_2^\bot\subseteq C_1$. Let $\ell$ be the largest integer such that for arbitrary two vectors $e_1\neq e_2\in\mathbb{F}_2^{n}$ whose burst length $\leq \ell$ there is $e_1+e_2\notin (C_1 \backslash C_2^\bot)\cup (C_2 \backslash C_1^\bot)$.
Then there exists a binary quantum burst error correction code $Q=[[n,k_1+k_2-n]]$ which can correct arbitrary quantum burst errors of length $\ell$ or less and if $\ell=\min\{\ell_1,\ell_2\}$, then $Q$ is a nondegenerate code.
\end{corollarys}
\subsection{Quantum Reiger Bound}
For a classical code $C=[n,k]$ which can correct $\leq t$ random errors or can correct any burst errors of length $\leq \ell$, there exists two important upper bounds called the Singleton bound $n-k\geq 2t$ and the Reiger bound $n-k\geq 2\ell$ that constrain the random error correction and burst error correction abilities of $C$, respectively (see \cite{lin2004error}). In quantum codes, let $Q=[[n,k]]$ be a QECC which can correct $\leq t$ quantum random errors, there exists the quantum Singleton bound $n-k\geq 4t$ which is an upper bound for the quantum random error correction ability of code $Q$ (see \cite{nielsen2000quantum,calderbank1998quantum}).
In the following, we derive the quantum Reiger bound (QRB) which is an upper bound for the quantum burst error correction ability of code $Q$.
\begin{theorems} [Quantum Reiger Bound]
\label{quantumReigerbound}
If an $[[n,k]]$ QBECC $Q$ can correct quantum burst errors of length $\ell$, then it satisfies
\begin{equation}
n-k\geq 4\ell.
\end{equation}
\end{theorems}
\begin{IEEEproof}
The proof follows closely by that of the quantum Singleton bound given by Preskill (see \cite[p.32]{preskill1998physics} and \cite[p.568]{nielsen2000quantum}).
First of all, Lemma~\ref{no-cloning bound for QBECCs} in the Appendix says that if $Q$ can correct $\ell$ burst errors, then it must satisfy $n>4\ell$, a consequence following from the quantum no-cloning principle.
Then we introduce a $k$-qubit ancilla system $A$, and construct a pure state $|\Psi\rangle_{AQ}$ that is maximally entangled between the system $A$ and the $2^k$ codewords of the $[[n,k]]$ QBECC $Q$:
\begin{equation}
|\Psi\rangle_{AQ}=\frac{1}{\sqrt{2^k}}\sum|x\rangle_A|{x}\rangle_Q,
\end{equation}
where $\{|x\rangle_A\}$ denotes an orthonormal basis for the $2^k$-dimensional Hilbert space of the ancilla, and $\{|x\rangle_Q\}$ denotes an orthonormal basis for the $2^k$-dimensional code subspace.
It is obvious that
\begin{equation}
S(A)_\Psi=k=S(Q)_\Psi,
\end{equation}
where $S(A)_\rho= -\text{Tr} \rho_A \log \rho_A $ is the von Neumann entropy of a density operator $\rho_A$.
Next we divide the $n$-qubit QBECC $Q $ into three disjoint parts so that $Q^{(1)}$ and $Q^{(2)}$ consist of $2\ell$ qubits each and $Q^{(3)}$ consists of the remaining $n-4\ell$ qubits. If we trace out $Q^{(2)}$ and $Q^{(3)}$, the reduced density matrix that we obtained must contain no correlations between $Q^{(1)}$ and the ancilla $A$, a consequence following from Lemma \ref{Located errors for QBECCs} in the Appendix. This means that the entropy of
system $AQ^{(1)}$ is additive:
\begin{equation}
S( Q^{(2)}Q^{(3)})_\Psi=S(AQ^{(1)})_\Psi=S(A)_\Psi+S(Q^{(1)})_\Psi.
\end{equation}
Similarly,
\begin{equation}
S( Q^{(1)}Q^{(3)})_\Psi=S(AQ^{(2)})_\Psi=S(A)_\Psi+S(Q^{(2)})_\Psi.
\end{equation}
Furthermore, in general, the von Neumann entropy is subadditive, so that
\begin{eqnarray}
S( Q^{(1)}Q^{(3)})_\Psi\leq S(Q^{(1)})_\Psi +S(Q^{(3)})_\Psi\\
S( Q^{(2)}Q^{(3)})_\Psi\leq S(Q^{(2)})_\Psi +S(Q^{(3)})_\Psi.
\end{eqnarray}
Combining these inequalities with the equalities above, we find
\begin{eqnarray}
S(A)+S(Q^{(2)})_\Psi\leq S(Q^{(1)})_\Psi +S(Q^{(3)})_\Psi\\
S(A)+S(Q^{(1)})_\Psi\leq S(Q^{(2)})_\Psi +S(Q^{(3)})_\Psi.
\end{eqnarray}
Both inequalities can be simultaneously satisfied only if
\begin{equation}
S(A)_\Psi\leq S(Q^{(3)})_\Psi.
\end{equation}
Finally, we have
\begin{equation}
S(A)_\Psi=k\leq n-4\ell,
\end{equation}
since $S(Q^{(3)})$ is bounded above by its dimension $n - 4\ell$. We then conclude the quantum Reiger bound.
\end{IEEEproof}
\section{Construction of Quantum Burst Error Correction Codes}
In this section we provide two methods for constructing QBECCs: one by using computer search based on the stabilizer formalism, and the other by concatenating and interleaving quantum tensor product codes.
\subsection{Stabilizer QBECCs Constructed by Using Computer Search}
We create a program using Magma software (version V2.12-16) to search all possible cyclic codes to a reasonable length ($n\leq 41$) according to Proposition \ref{proposition_burst} and make the codes as
close as possible to the quantum Reiger bound. For simplicity, we only consider the construction of QBECCs from cyclic codes with odd length. We list new codes that are near or saturate the quantum Reiger bound found by this program in Table \ref{Computer Searching for QBECCs}.
The bold numbers ``$\textbf{1}-\textbf{3}$'' stand for the
coefficients and the superscript numbers stand for the exponents in the generator polynomials of the corresponding classical cyclic codes \cite{lin2004error}. Notice that the burst error-correcting abilities of most QBECCs in Table \ref{Computer Searching for QBECCs} can saturate the quantum Reiger bound. In particular, we get several \emph{degenerate} QBECCs which are the first class of QBECCs until now and they can saturate the quantum Reiger bound. Moreover, some of the constructed QBECCs are better than the QBECCs in Ref.~\cite{tokiwa2005some}, {e.g., the codes $[[15,3]]$, $[[35,17]]$ and $[[35,13]]$ in Table \ref{Computer Searching for QBECCs} have larger dimensions than the codes $[[15,2]]$, $[[35,14]]$ and $[[35,11]]$ in Ref.~\cite{tokiwa2005some}, respectively, but have the same burst error-correcting ability.}
We remark that some QBECCs of length up to 51 have been found by using computer search in Ref.~\cite{tokiwa2005some}. However, only CSS type QBECCs were considered and no degenerate codes were obtained in Ref.~\cite{tokiwa2005some}.
\newcommand{\tabincell}[2]{\begin{tabular}{@{}#1@{}}#2\end{tabular}}
\begin{table}
\renewcommand{\arraystretch}{0.1}
\setlength{\tabcolsep}{2pt}
\caption{Computer Searching for Quantum Burst Error Correction Codes}
\label{Computer Searching for QBECCs}
\centering
\footnotesize
\begin{tabular}[c]{|c|c|c|c|c|}
\hline
$[[n,k]]$& $l$ &Generator Polynomials&QRB&Degenerate? \\
\hline
$[[13,1]]$&3&$g=(\textbf{1}^6\textbf{2}^5\textbf{3}^3\textbf{2}^1\textbf{1}^0)$&3&False\\
\hline
$[[15,3]]$&3&$g=(\textbf{1}^6\textbf{2}^3\textbf{1}^0)$&3&False\\
\hline
$[[17,1]]$&4&$g=(\textbf{1}^8\textbf{3}^7\textbf{1}^6\textbf{1}^5\textbf{2}^4\textbf{1}^3\textbf{1}^2\textbf{3}^1\textbf{1}^0)$&4&False\\
\hline
$[[17,1]]$&4&$g=(\textbf{1}^8\textbf{3}^7\textbf{3}^5\textbf{3}^4\textbf{3}^3\textbf{3}^1\textbf{1}^
0)$&4&True\\
\hline
$[[21,9]]$&2&\tabincell{l}{$g_1=(\textbf{1}^{6}\textbf{1}^4\textbf{1}^1\textbf{1}^0)$,
\\$g_2=(\textbf{1}^{6}\textbf{1}^4\textbf{1}^2\textbf{1}^1\textbf{1}^0)$}&3&False\\
\hline
$[[23,1]]$&5&$(g=\textbf{1}^{11}\textbf{1}^9\textbf{1}^7\textbf{1}^6\textbf{1}^5\textbf{1}^1\textbf{1}^0)$&5&False\\
\hline
$[[25,1]]$&6&$g=(\textbf{1}^{12}\textbf{2}^{11}\textbf{1}^{10}\textbf{2}^7\textbf{3}^6\textbf{2}^5\textbf{1}^2\textbf{2}^1\textbf{1}^0)$&6&True\\
\hline
$[[25,5]]$&5&$g=(\textbf{1}^{10}\textbf{2}^{5}\textbf{1}^0)$&5&False\\
\hline
$[[29,1]]$&7&\tabincell{c}{$g=(\textbf{1}^{14}\textbf{2}^{13}\textbf{2}^{11}\textbf{3}^{10}\textbf{1}^{9}\textbf{3}^{8} $
\\\hspace{10mm}$\textbf{2}^7\textbf{3}^6\textbf{1}^5\textbf{3}^{4}\textbf{2}^3\textbf{2}^1\textbf{1}^0)$}&7&True\\
\hline
$[[35,25]]$&2&$g=(\textbf{1}^{5}\textbf{2}^{4}\textbf{3}^{2}\textbf{2}^{1}\textbf{1}^0)$&2&False\\
\hline
$[[35,19]]$&3&$g=(\textbf{1}^{8}\textbf{2}^{7}\textbf{3}^{6}\textbf{1}^{5}\textbf{3}^{3}\textbf{1}^{2}\textbf{1}^1\textbf{1}^0)$&4&False\\
\hline
$[[35,17]]$&4&$g=(\textbf{1}^{9}\textbf{3}^{7}\textbf{3}^{6}\textbf{3}^{5}\textbf{3}^{4}\textbf{2}^{3}\textbf{2}^{2}\textbf{2}^1\textbf{1}^0)$&4&False\\
\hline
$[[35,13]]$&5&$g=(\textbf{1}^{11}\textbf{3}^{10}\textbf{2}^{9}\textbf{1}^{8}\textbf{2}^{7}\textbf{2}^{6}\textbf{3}^{5}
\textbf{1}^{3}\textbf{2}^{2}\textbf{1}^1\textbf{1}^0)$&5&False\\
\hline
$[[35,7]]$&6&\tabincell{c}{$g=(\textbf{1}^{14}\textbf{2}^{13}\textbf{3}^{11}\textbf{2}^{10}\textbf{2}^{9}\textbf{3}^{8} $
\\\hspace{10mm}$\textbf{1}^7\textbf{1}^6\textbf{2}^5\textbf{3}^{3}\textbf{2}^2\textbf{3}^1\textbf{1}^0)$}&7&False\\
\hline
$[[41,1]]$&10&\tabincell{c}{$g_1=(\textbf{1}^{20}\textbf{1}^{18}\textbf{1}^{17}\textbf{1}^{16}\textbf{1}^{15}\textbf{1}^{14}\textbf{1}^{11}$
\\\hspace{10mm}$\textbf{1}^{10}\textbf{1}^{9}\textbf{1}^{6}\textbf{1}^{5}\textbf{1}^{4}\textbf{1}^{3}\textbf{1}^{2}\textbf{1}^{0})$,\\
$g_2=(\textbf{1}^{20}\textbf{1}^{19}\textbf{1}^{17}\textbf{1}^{16}\textbf{1}^{14}\textbf{1}^{11}$\\
\hspace{10mm}$\textbf{1}^{10}\textbf{1}^{9}\textbf{1}^{6}\textbf{1}^{4}\textbf{1}^{3}\textbf{1}^{1}\textbf{1}^{0})$}&10&True\\
\hline
\end{tabular}
\end{table}
\subsection{Concatenation Construction of QBECCs Based on Quantum Tensor Product Codes}
In this section we give a concatenation construction of long QBECCs from two short component codes based on the quantum tensor product codes structure \cite{jihao2017on,fan2017comments}.
Firstly we present a brief review of classical and quantum tensor product codes. Details could be found in, e.g., \cite{wolf1965codes,jihao2017on}.
Let $C_1=[n_1, k_1]_2$ be a linear code with a
parity check matrix $H_1$ and let $\rho_1=n_1-k_1$ be the number of check symbols. Let $C_2=[n_2, k_1]_{2^{\rho_1}} $ be a linear code over the extension filed $\mathbb{F}_{2^{\rho_1}}$ with
a parity check matrix $H_2$. Then the tensor product code (TPC) of $C_1$ and $C_2$ is denoted by $\mathcal{C}=C_2\otimes_HC_1$, and the parity check
matrix $\mathcal{C}$ is given by
\begin{equation}
H=H_2\otimes H_1.
\end{equation}
By selecting different types of component codes, TPCs can be designed to provide different error control abilities.
In \cite{jihao2017on}, a framework for the construction of quantum tensor product codes (QTPC), which can provide a wide variety of quantum error-correcting, error-detecting or error-locating
properties, was proposed. In particular, if one of the
component codes is selected as a BECC, then QTPCs can have multiple quantum burst error-correcting abilities, but provided these bursts fall in distinct subblocks.
\begin{theorems}[\cite{jihao2017on}
\label{CSS_TPC1_theorem}
Let $C_1=[n_1,k_1]_q$ be an $\ell_1$ burst error correction code, and let $C_2=[n_2,k_2]_{q^{\rho_1}}$ be an $\ell_2$ burst error correction code
over the extension field $\mathbb{F}_{q^{\rho_1}}$, and the numbers of check symbols are $\rho_1=n_1-k_1$ and $\rho_2=n_2-k_2$, respectively.
If $q=2$ and $C_1^\perp\subseteq C_1$ or
if $q=4$ and $C_1^{\perp_h}\subseteq C_1$, where $C_1^{\perp_h}$ is the Hermitian dual code of $C_1$, then there exists a QTPC $\mathcal{Q}=[[n_1n_2, n_1n_2-2\rho_1\rho_2]]$ which can correct $\ell_2$ or fewer bursts of burst errors each is a burst of length $\ell_1$ or less,
provided these bursts fall in distinct subblocks.
\end{theorems}
\iffalse
\begin{theorems}[\cite{jihao2017on}]
\label{CSS_TPC2}
Let $H_{1}$ be the parity check matrix of an $\ell_1$ burst error correction code $C_1=[n_1,k_1] $, and let $H_{2}$ be the parity check matrix of an
$\ell_2$ burst error correction code $C_2=[n_2,k_2]_{2^{\rho_1}}$ over the extension field $\mathbb{F}_{2^{\rho_1}}$,
and the numbers of check symbols are $\rho_1=n_1-k_1$ and $\rho_2=n_2-k_2$, respectively.
If the component code $C_2$ satisfies $C_2^\bot\subseteq C_2$, and $H_{1}H_{1}^T$ is of full rank, then
there exists a QTPC $\mathcal{Q}=[[n_1n_2, n_1n_2-2\rho_1\rho_2]]$ which can correct $\ell_2$ or fewer bursts of burst errors each is a burst of length $l_1$ or less,
provided these bursts fall in distinct subblocks.
\end{theorems}
\fi
Although we can use QTPCs to correct a single burst of errors since QTPCs have multiple burst error correction abilities, they are not efficient enough any more. To overcome this problem, we can interleave the encoded qubits before sending into the quantum channels, and deinterleave after receiving the qubits.
The whole interleaving/deinterleaving procedure is summarized as follows:
\begin{itemize}
\item[(1).] After the quantum encoding, we arrange the encoded $n_1n_2$ qubits into an $n_1\times n_2$ code array.
\item[(2).] Instead of transmitting the encoded qubits sequentially one by one, we do an interleaved transmission. Denote by $\ell_1$ the burst error correction ability of $C_1$.
If $\ell_1|n_1$ and
the component code $C_2$ can correct \emph{end-around} (see \cite{lin2004error}) burst errors, then we divide the $n_1\times n_2$ code array into $s=n_1/\ell_1$ subblocks by rows.
We do the transmission subblock by subblock, and in each subblock, we transmit the qubits column by column sequentially (each column contains $\ell_1$ qubits).
\item[(3).] After receiving all the $n_1n_2$ qubits, we deinterleave the qubits into an $n_1\times n_2$ code array so that the quantum decoding can be processed next,
and the deinterleaving is just the inverse of the interleaving. The deinterleaving/interleaving procedure can be accomplished by using quantum SWAP gates (see \cite{nielsen2000quantum}).
\end{itemize}
\begin{figure}[!t]
\centering
\includegraphics[width=3.2in]{ef1.png}
\caption{The entanglement fidelity (EF) of the two $[[13,1]]$ and $[[17,1]]$ codes with respect to the correlation degree $ 0\leq \mu\leq1 $, the error probability
is set to be $p=3\times 10^{-2}$.}
\label{ef1}
\end{figure}
Suppose that a single burst errors of length at most $\ell_1\ell_2$ happens among the $n_1n_2$ interleaved qubits. After the quantum transmition and deinterleaving, the $n_1n_2$ qubits are recovered to their original positions, but the single burst errors of length at most $\ell_1\ell_2$ has been dispersed into $\ell_2$ or fewer consecutive subblocks (end around) and each subblock contains a burst errors of length at most $\ell_1$. Thus the resultant QTPC $\mathcal{Q}$ can correct a single burst error of length at most
$\ell_1\ell_2$ according to Theorem \ref{CSS_TPC1_theorem} and Ref. ~\cite{jihao2017on}. Then we have the following result.
\begin{theorems}
\label{interleav burst error correction}
Let $C_1=[n_1,k_1]_q$ and $C_2=[n_2,k_2]_{q^{\rho_1}}$ be two component codes of a QTPC with parameters
$\mathcal{Q}=[[n_1n_2,n_1n_2-2\rho_1\rho_2]]$, where $C_1$ is an $\ell_1$ burst error correction code and
$C_2$ is an $\ell_2$ burst error correction code (end-around),
and the numbers of check symbols are $\rho_1=n_1-k_1$ and $\rho_2=n_2-k_2$, respectively. If $\ell_1|n_1$,
then there exists an $\ell_1\ell_2$ burst error correction quantum code $\mathcal{Q}=[[n_1n_2,n_1n_2-2\rho_1\rho_2]]$.
\end{theorems}
\begin{examples}
\label{interleav burst error correction corollarys}
We choose $C_1=[15,9]_4$ as a $3$ burst error correction cyclic code with the generator polynomial $g=(\textbf{1}^6\textbf{2}^3\textbf{1}^0)$ and it is Hermitian dual containing by Table \ref{Computer Searching for QBECCs}. Let $C_2=[n_2,n_2-2\ell_2,2\ell_2+1]_{4^{6}}$ be an MDS code over the extension field $\mathbb{F}_{4^{6}}$ with $2\leq n_2\leq 4^6+2$ and $1\leq \ell_2\leq\lfloor\frac{n_2-1}{2}\rfloor$. Then there exists a $3\ell_2$ burst error correction QTPC with parameters $\mathcal{Q}=[[15n_2,15n_2-24\ell_2]]$.
\end{examples}
\begin{figure}[!t]
\centering
\includegraphics[width=3.2in]{ef2.png}
\caption{ The entanglement fidelity (EF) of the two $[[13,1]]$ and $[[17,1]]$ codes with respect to the error probability
$ 1\times 10^{-5}\leq p\leq1\times10^{-1}$, the correlation degree is
set to be $\mu=0.5$.}
\label{ef2}
\end{figure}
\section{Performance of QBECCs over Markovian Correlated Quantum Memory Channels}
In this section we evaluate the performance of two specific QBECCs in the presence of correlated errors.
The channel model that we choose is a Markovian correlated depolarizing quantum channel \cite{caruso2014quantum,Cafaro2010quantum}:
\begin{eqnarray}
\Phi^{(n)}(\rho) =\sum_{i_1,\ldots,i_n=0}^{3}p_{i_n|i_{n-1}}^{(n)}\cdots p_{i_2|i_{1}}^{(2)}p_{i_1}^{(1)}\nonumber\\
\times E_{i_n}^{(n)}\cdots E_{i_1}^{(1)}\rho E_{i_1}^{(1)\dagger}\cdots E_{i_n}^{(n)\dagger},
\end{eqnarray}
where $\{E_{i_j}^{(j)}\}_{i,j}$ are the Pauli operators, and the conditional probabilities
satisfy the normalization condition
\begin{equation}
\sum\limits_{i_1,\ldots,i_n=0}^{3}p_{i_n|i_{n-1}}^{(n)}\cdots p_{i_2|i_{1}}^{(2)}p_{i_1}^{(1)}=1,
\end{equation}
where $p_{l|k}^{(j)}=(1-\mu)p_{l}+\mu\delta_{(k,l)}$ for $1\leq j\leq n$ and $ 0\leq k,l\leq 3$, and $p_{0}=1-p, p_{1,2,3}=p/3$
are the error probabilities in the depolarizing channel, $\mu\in[0,1]$ is the correlation degree.
Specifically, we show the performance of the two specific codes by measuring the entanglement fidelity $\mathcal{F}(p,\mu)$ as
a function of the error probability $p$ and the correlation degree $\mu$ \cite{nielsen2000quantum}.
The two specific codes considered here are two QBECCs $[[13,1]]$ and $[[17,1]]$ in Table \ref{Computer Searching for QBECCs} which can correct burst errors of length $\leq 3$ and of length $\leq 4$, respectively. Through the computation, we know that the minimum distances of the two codes are $5$ and $7$, then they can also correct $\leq2$ and $\leq3$ random errors, respectively, and they have achieved the upper bounds in Ref.~\cite{Grassl:codetables}. We plot the performance by means of entanglement fidelity of the two specific codes with respect to random errors or burst errors, versus the correlation degree $\mu$ or the error probability $p$ in Fig. \ref{ef1} or Fig. \ref{ef2}, respectively. For details about
the computation of the entanglement fidelity, see \cite{caruso2014quantum,Cafaro2010quantum,nielsen2000quantum}, and the computation results are put in the \href{https://www.dropbox.com/sh/on17nv0ca1b6j7j/AAD3fd0PORxyQH0TuhJlJR0ra?dl=0}{Cloud}.
\hspace{-2mm}
It is shown that the correlation errors do degrade the performance of the entanglement fidelity of the two codes in Fig.~\ref{ef1}. If the the correlation degree $\mu=0$ which means that errors are independent with each other, then the extra burst error correction abilities of the two codes do little help to improve the performance of entanglement fidelity of them, respectively, see Fig. \ref{ef1} and Fig. \ref{ef3}. However, if we consider the correlated errors when $0<\mu<1$, the performance of the entanglement fidelity can be improved largely, see Fig.~\ref{ef1} and Fig.~\ref{ef2}. In particular, in Fig.~\ref{ef2}, the $[[13,1]]$ code have better performance when considering its extra burst error correction ability compared to the $[[17,1]]$ code when only considering the
random error correction ability.
\begin{figure}[!t]
\centering
\includegraphics[width=3.2in]{ef3.png}
\caption{ The entanglement fidelity (EF) of the two $[[13,1]]$ and $[[17,1]]$ codes with respect to the error probability
$ 1\times 10^{-5}\leq p\leq1\times10^{-1}$, the correlation degree is
set to be $\mu=0$.}
\label{ef3}
\end{figure}
\section*{Acknowledgment}
J. Fan was supported by NJIT (Grant No. YKJ201719) and by NSFC (Grant No. 61403188). M.-H. Hsieh was supported in part by an ARC Future Fellowship under Grant FT140100574 and in part by U.S. the Army Research Office for Basic Scientific Research under Grant W911NF-17-1-0401.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}
|
\section{Introduction\label{sec:introduction}}
Indirect evidence for new physics in processes involving charmed mesons may be
obtained in two ways:
\begin{itemize}
\item Very rare processes could be measured with rates that exceed unambiguously
predictions within the Standard Model.
\item Very precise amplitude relations predicted within the Standard Model could be
violated experimentally.
\end{itemize}
Two properties of charmed mesons, $D^0$--$\bar D^0$ mixing and CP violation in singly Cabibbo-suppressed (SCS) $D$ meson decays, have been suggested as potential cases
for the first scenario~\cite{Bianco:2003vb,Golowich:2007ka,Grossman:2006jg}.
Recently we derived a very precise nonlinear relation among four ratios of amplitudes for
$D^0 \to \pi^+K^-, K^+\pi^-, K^+K^-, \pi^+\pi^-$~\cite{Gronau:2013xba}, valid up to fourth order U-spin breaking.
While precise amplitude relations have already been proposed for hadronic $B$ meson
decays (see for instance Ref.\,\cite{Gronau:1990ka}), this particular relation provides a first
case for the second scenario in hadronic $D$ decays.
Measurements of the difference between CP asymmetries in $D^0\to K^+K^-$ and
$D^0\to\pi^+\pi^-$ of order $10^{-3}$~\cite{Aaltonen:2011se,Aaij:2011in,HFAG} have been
shown to be consistent with Standard Model estimates~\cite{Brod:2011re}. This has been
used to obtain model-dependent constraints on new $|\Delta C|=1$ operators occurring in a number of models~\cite{Isidori:2011qw}. The amplitude relation derived in
Ref.~\cite{Gronau:2013xba}, involving a precision of $10^{-3}$, has been
shown to agree with experiment at this same high accuracy. The purpose of this
letter is to study the possibility of using this excellent agreement for obtaining
model-independent constraints on new $|\Delta C|=1$ operators.
The proof in~\cite{Gronau:2013xba} of the nonlinear amplitude relation is based largely on the fact that the charm-changing weak hamiltonian transforms as a U-spin triplet~\cite{Feldmann:2012js,Gronau:2012kq}.
Thus we will distinguish between two classes of models involving new $|\Delta C|=1$ operators behaving distinctly under U-spin. In the first case we will assume these operators to transform like
U-spin scalars. This rather broad class of models includes supersymmetric and extra-dimensional models involving new QCD penguin and chromomagnetic dipole operators. Constraints on such models from CP asymmetries in SCS D decays and from $D^0$-$\bar D^0$ mixing have been studied in Refs.~\cite{Grossman:2006jg} and~\cite{DaRold:2012sz}. A second class of models includes new $|\Delta C|=1$ operators transforming like $U=1, U_3=0$. Constraints on such operators from CP asymmetries in SCS decays have been discussed in Ref.~\cite{Hiller}.
Other probes for new physics have been suggested in Ref.\,\cite{Grossman:2012eb}
in terms of isospin sum rules for CP asymmetries in SCS $D$ decays.
Section~\ref{sec:SM} summarizes briefly arguments used in
Ref.~\cite{Gronau:2013xba} leading to a precise nonlinear relation among four ratios
of amplitudes for $D^0 \to \pi^+K^-, K^+\pi^-, K^+K^-, \pi^+\pi^-$. In Sections~\ref{sec:U=0}
and~\ref{sec:U=1} we study separately contributions of new $U=0$ and $U=1, U_3=0$
operators potentially modifying this relation. Section~\ref{sec:CP asymmetries} contains a
discussion of CP asymmetries in $D^0\to K^+K^-$ and $D^0 \to \pi^+\pi^-$ in these two
classes of models. Conclusions are given in Section~\ref{sec:conclusion}.
\section{Precise amplitude relation in the Standard Model\label{sec:SM}}
Hadronic weak decays of charmed mesons are conveniently studied using U-spin
symmetry under which the quark pair $(d,s)$ transforms like a doublet.
The effective Hamiltonian operators for Cabibbo-favored (CF), singly Cabibbo-suppressed (SCS) and doubly Cabibbo-suppressed charm decays transform like three components $U_3=-1, 0, +1$
of $U=1$, excluding corresponding CKM factors, $\cos^2\theta_C, -\cos\theta_C\sin\theta_C, \sin^2\theta_C$.
We neglect a term in the SCS Hamiltonian proportional to a tiny CKM factor $V^*_{cb}V_{ub}$, where $|V^*_{cb}V_{ub}|/\cos\theta_C\sin\theta_C \simeq 0.7\times 10^{-3}$~\cite{PDG}.
This term would be responsible for CP asymmetries of this order in $D^0\to K^+K^-$ and
$D^0\to\pi^+\pi^-$. Since the measured CP asymmetries are at most of this order~\cite{HFAG}, while our approximation for amplitudes involves uncertainties of this same order, we will neglect
CP asymmetries also in the next two sections discussing new physics, returning to discuss them
in Section \ref{sec:CP asymmetries}.
The $D^0$ is a U-spin singlet, the states $-|\pi^+K^-\rangle, \frac{1}{\sqrt{2}}|K^+K^--\pi^+\pi^-\rangle, |K^+\pi^-\rangle$ are three components $|U_3=-1, 0, +1\rangle$ of $|U=1\rangle$ while
$\frac{1}{\sqrt{2}}|K^+K^- + \pi^+\pi^-\rangle$ is a singlet.
The matrix element of the $U=1$ Hamiltonian vanishes for the latter state,
$\langle K^+K^-|{\cal H}_{\rm eff}|D^0\rangle = - \langle \pi^+\pi^-|{\cal H}_{\rm eff}|D^0\rangle$.
Thus in the U-spin symmetry limit the four amplitudes for $D^0 \to \pi^+K^-, K^+K^-, \pi^+\pi^-, K^+\pi^-$ are given by a common $U=1$ amplitude $A$. Consequently the multiple ratio of the four decay amplitudes is given
by ratios of CKM factors~\cite{Kingsley:1975fe}:
\begin{eqnarray}
& & A(D^0\to \pi^+K^-) : A(D^0\to K^+K^-) : A(D^0\to \pi^+\pi^-) : A(D^0 \to K^+\pi^-)
\nonumber\\
& &
= \cos^2\theta_C : \cos\theta_C\sin\theta_C : -\cos\theta_C\sin\theta_C : -\sin^2\theta_C~.
\end{eqnarray}
U-spin breaking in amplitudes is treated perturbatively in terms of two distinct parameters
proportional to $(m_s - m_d)/\Lambda_{\rm QCD}$, taken separately for $D^0\to \pi^+K^-,
K^+\pi^-$ and $D^0 \to K^+K^-, \pi^+\pi^-$. Symmetry breaking of order $k$ in
an amplitude $\langle f|{\cal H}_W|D^0\rangle$ is obtained by introducing in ${\cal H}_W$
or in $|f\rangle$ $k$ powers of a quark mass-difference operator $\bar s s- \bar d d$ behaving
like $U=1, U_3=0$. The decays $D^0 \to K^+K^-, \pi^+\pi^-$ obtain also a first order U-spin breaking correction from a $U=0$ penguin operator. This and a simple sign property for U-spin
addition, $(1, -1;n, 0|1, -1) = (-1)^n(1, 1;n, 0|1, 1)$, lead to the following two properties of U-spin breaking~\cite{Gronau:2013xba}:
\begin{itemize}
\item Symmetry breaking effects in $D^0\to \pi^+K^-, K^+\pi^-$ and $D^0 \to K^+K^-, \pi^+\pi^-$
are described by two different parameters, to be denoted $\epsilon_1$ and $\epsilon_2$,
respectively.
\item In each one of these two pairs of processes U-spin breaking corrections of even (odd)
order have equal magnitudes and same (opposite) signs.
\end{itemize}
Expanding up to third order U-spin breaking one has
\begin{eqnarray}\label{Amp}
A(D^0 \to \pi^+K^-) & = & \cos^2\theta_CA[1 - \epsilon_1 + a_1(\epsilon_1)^2
- a'_1(\epsilon_1)^3]~,
\nonumber\\
A(D^0\to K^+\pi^-) & = & -\sin^2\theta_CA[1 + \epsilon_1 + a_1(\epsilon_1)^2
+ a'_1(\epsilon_1)^3]~,
\nonumber\\
A(D^0 \to K^+K^-) & = & \cos\theta_C\sin\theta_CA[1 + \epsilon_2 + a_2(\epsilon_2)^2
+ a'_2(\epsilon_2)^3]~, \nonumber\\
A(D^0 \to \pi^+\pi^-) & = & -\cos\theta_C\sin\theta_CA[1 - \epsilon_2 + a_2(\epsilon_2)^2
- a'_2(\epsilon_2)^3]~,
\end{eqnarray}
where $a_{1,2} \sim a'_{1,2} \sim 1$.
Defining four independent ratios of amplitudes $R_i, (i=1,2,3,4)$ one finds:
\begin{eqnarray}\label{Ri}
R_1 & \equiv & \frac{|A(D^0 \to K^+\pi^-)|}{|A(D^0 \to \pi^+K^-)|\tan^2\theta_C}
= 1 + 2[{\rm Re}\,\epsilon_1 + ({\rm Re}\,\epsilon_1)^2] = 1.118 \pm 0.014~,
\nonumber\\
R_2 & \equiv & \frac{|A(D^0 \to K^+K^-)|}{|A(D^0\to \pi^+\pi^-)|} =
1 + 2[{\rm Re}\,\epsilon_2 + ({\rm Re}\,\epsilon_2)^2] = 1.814 \pm 0.018~,
\nonumber\\
R_3 & \equiv & \frac{|A(D^0 \to K^+K^-)| + |A(D^0\to \pi^+\pi^-)|}
{|A(D^0 \to \pi^+K^-)|\tan\theta_C + |A(D^0\to K^+\pi^-)|\tan^{-1}\theta_C}
\nonumber\\
& = & 1 + \frac{1}{2}[({\rm Im}\,\epsilon_2)^2 - ({\rm Im}\,\epsilon_1)^2] +
{\rm Re}\,[a_2(\epsilon_2)^2 - a_1(\epsilon_1)^2] = 1.056 \pm 0.008~,
\nonumber\\
R_4 & \equiv & \sqrt{\frac{|A(D^0 \to K^+K^-)||A(D^0\to \pi^+\pi^-)|}
{|A(D^0\to \pi^+K^-)||A(D^0\to K^+\pi^-)|}}
= 1 + \frac{1}{2}[({\rm Im}\,\epsilon_2)^2 - ({\rm Im}\,\epsilon_1)^2]
\nonumber\\
& + & {\rm Re}\,[a_2(\epsilon_2)^2 - a_1(\epsilon_1)^2]
- \frac{1}{2}[({\rm Re}\,\epsilon_2)^2 - ({\rm Re}\,\epsilon_1)^2]
= 1.012 \pm 0.007~.
\end{eqnarray}
$R_1$ and $R_2$ involve additional third order terms while corrections to $R_3$
and $R_4$ start at fourth order. [The last term of Eq,\,(21) in \cite{Gronau:2013xba}
should be of fourth order.]
In the above we used the following third order expansion:
\begin{eqnarray}
|1 \pm \epsilon + a\epsilon^2 \pm a'\epsilon^3| = 1 & \pm & {\rm Re}\,\epsilon +
\frac{1}{2}({\rm Im}\,\epsilon)^2 + {\rm Re}(a\epsilon^2) \pm {\rm Re}(a'\epsilon^3)
\nonumber\\
& \mp & \frac{1}{2}{\rm Re}\,\epsilon\,({\rm Im}\,\epsilon)^2
\mp {\rm Im}\,\epsilon\,{\rm Im}(a\epsilon^2)~.
\end{eqnarray}
\begin{table}[h]
\caption{Amplitudes in units of
$10^{-1}({\rm GeV}/c)^{-1/2}$ for $D^0$ decays to pairs involving a charged pion
and kaon.\label{tab:A}}
\begin{center}
\begin{tabular}{c c } \hline \hline
Decay mode & $|A|=\sqrt{{\cal B}/p^*}$ \\ \hline
$D^0\to \pi^+K^-$ & $2.1228$ \\
$D^0 \to K^+\pi^- $ & $0.1268 \pm 0.0012$ \\
$D^0 \to K^+K^-$ & $0.7076 \pm 0.0052$ \\
$D^0 \to \pi^+\pi^-$ & $0.3900 \pm 0.0027$ \\
\hline \hline
\end{tabular}
\end{center}
\end{table}
Numerical values on the right hand sides of (\ref{Ri}) have been obtained using
$\tan\theta_C=0.2312\pm 0.0009$ and branching
fractions for $D^0\to K^+\pi^-, K^+K^-, \pi^+\pi^-$ measured relative to
$D^0 \to \pi^+K^-$~\cite{PDG}. Table \ref{tab:A} quotes magnitudes of the four amplitudes
defined by $|A| \equiv \sqrt{{\cal B}/p^*}$. Note that the first amplitude involves no error as
the three others are measured relative to its magnitude.
The second order expressions in (\ref{Ri}) imply ${\rm Re}\epsilon_1 =0.056 \pm 0.006, {\rm Re}\epsilon_2 = 0.311 \pm 0.006$, and a nonlinear relation among these four ratios which holds up
to fourth order U-spin breaking:
\begin{equation}\label{NLrel}
\Delta R \equiv R_3 - R_4 + \frac{1}{8}\left[(\sqrt{2R_1-1} - 1)^2 - (\sqrt{2R_2 - 1} -1)^2\right]=0~.
\end{equation}
This relation is satisfied extremely well by current experiments for which one finds
\begin{equation}
\Delta R_{\rm exp} = -0.003 \pm 0.002~.
\end{equation}
We have neglected in Eq.\,(\ref{NLrel}) a fourth order U-spin breaking correction and an
isospin breaking term suppressed also by U-spin breaking. This leads to an uncertainty
of order $10^{-3}$~\cite{Gronau:2013xba}.
\section{Contributions of a new $U=0$ operator\label{sec:U=0}}
A new $U=0$ operator does not contribute to $D^0 \to \pi^\pm K^\mp$ where final states
have $U_3=\mp 1$. This is true in the U-spin symmetry limit and also when including U-spin
breaking of arbitrary order. We will assume that contributions of this operator in
$D^0\to K^+K^-$ and $D^0\to \pi^+\pi^-$ are subleading, namely of order $\epsilon_2$
or smaller, as are the U-spin breaking
contributions of a $U=0$ penguin operator occurring in the CKM
framework~\cite{Gronau:2013xba}. In these two processes the new $U=0$ operator has
equal contributions in the U-spin symmetry limit and first order U-spin breaking terms of
equal magnitudes and opposite signs. Normalizing these two contributions by the $U=1$ amplitude we denote them by $\cos\theta_C\sin\theta_CAn$ and
$\cos\theta_C\sin\theta_CAn\epsilon$, respectively:
\begin{eqnarray}\label{amp0}
A(D^0 \to K^+K^-) & = & \cos\theta_C\sin\theta_CA[1 + \epsilon_2 + a_2(\epsilon_2)^2
+ a'_2(\epsilon_2)^3 + n + n\epsilon]~,
\nonumber\\
A(D^0 \to \pi^+\pi^-) & = & -\cos\theta_C\sin\theta_CA[1 - \epsilon_2 + a_2(\epsilon_2)^2
- a'_2(\epsilon_2)^3 - n + n\epsilon]~.
\end{eqnarray}
Here we wish to study the effects of these new terms on (\ref{Ri}) and (\ref{NLrel}).
Expanding ratios of amplitudes up to and including terms of second order in $n$ and
in the U-spin breaking parameters $\epsilon_1, \epsilon_2, \epsilon$, one obtains
\begin{eqnarray}\label{RiU=0}
R_1 & = & 1 + 2[{\rm Re}\,\epsilon_1 + ({\rm Re}\,\epsilon_1)^2]~,
\nonumber\\
R_2 & = & 1 + 2\left[{\rm Re}\,(\epsilon_2 + n) + [{\rm Re}\,(\epsilon_2+n)]^2\right]~,
\nonumber\\
R_3 & = & 1 + \frac{1}{2}[({\rm Im}\,(\epsilon_2+n))^2 - ({\rm Im}\,\epsilon_1)^2] +
{\rm Re}\,[a_2(\epsilon_2)^2 - a_1(\epsilon_1)^2 + n\epsilon]
\nonumber\\
R_4 & = & 1
+ \frac{1}{2}[({\rm Im}\,(\epsilon_2+n))^2 - ({\rm Im}\,\epsilon_1)^2] +
{\rm Re}\,[a_2(\epsilon_2)^2 - a_1(\epsilon_1)^2 +n\epsilon]
\nonumber\\
& &
~~- \frac{1}{2}[({\rm Re}\,(\epsilon_2+n))^2 - ({\rm Re}\,\epsilon_1)^2]~.
\end{eqnarray}
These results correspond to a substituting $\epsilon_2 \to \epsilon_2 + n$
and $a_2(\epsilon_2)^2 \to a_2(\epsilon_2)^2 + n\epsilon$ in (\ref{Ri}).
Thus the ratio $R_2$ involves a term which is first order in the $U=0$ amplitude. It implies
${\rm Re}(\epsilon_2 + n) = 0.311 \pm 0.006$.
The ratios $R_3$ and $R_4$ include identical second order terms depending on the
$U=0$ amplitude that cancel in their difference occurring in $\Delta R$. Since
Eqs.\,(\ref{RiU=0}) have the same structure as Eqs.\,(\ref{Ri}), with a mere substitution,
$\epsilon_2 \to \epsilon_2 + n, a_2(\epsilon_2)^2 \to a_2(\epsilon_2)^2 + n\epsilon$,
the nonlinear relation (\ref{NLrel}) still holds. That is, {\em Eq.\,(\ref{NLrel}) is unaffected
by arbitrary new $U=0$ operators and cannot be used to constrain such operators}.
\section{Contributions of a new $U=1, U_3=0$ operator\label{sec:U=1}}
Consider now models with new $U=1, U_3=0$ operators. The effect of such
operators on (\ref{Amp}) is to modify the overall factor in the amplitudes for
$D^0\to K^+K^-, \pi^+\pi^-,$ and to replace the U-spin breaking parameter $\epsilon_2$
by a new parameter $\epsilon'_2$, corresponding to U-spin breaking in the total $U=1,
U_3=0$ amplitude. Using a parameter $n$ to normalize the new amplitude by the
U-spin invariant amplitude $A$, one has
\begin{eqnarray}\label{KKpipi}
A(D^0 \to K^+K^-) & = & \cos\theta_C\sin\theta_CA(1 + n)[1+ \epsilon'_2 +
a_2(\epsilon'_2)^2 + a'_2(\epsilon'_2)^3]~,
\nonumber\\
A(D^0 \to \pi^+\pi^-) & = & -\cos\theta_C\sin\theta_CA(1 + n)[1 - \epsilon'_2 +
a_2(\epsilon'_2)^2 - a'_2(\epsilon'_2)^3]~.
\end{eqnarray}
Expanding the four ratios of amplitudes up to second order, we note that
$R_1$ and $R_2$ are essentially unaffected relative to (\ref{Ri}) while
$R_3$ and $R_4$ obtain an overall factor $|1+n|$:
\begin{eqnarray}
R_1 & = & 1 + 2[{\rm Re}\,\epsilon_1 + ({\rm Re}\,\epsilon_1)^2]~,
\nonumber\\
R_2 & = & 1 + 2[{\rm Re}\,\epsilon'_2 + ({\rm Re}\,\epsilon'_2)^2]~~~~
{\rm implying}~{\rm Re}\,\epsilon'_2=0.311\pm 0.006~,
\nonumber\\
R_3 & = & |1 + n|\left[1 + \frac{1}{2}[({\rm Im}\,\epsilon'_2)^2 - ({\rm Im}\,\epsilon_1)^2] +
{\rm Re}\,[a'_2(\epsilon'_2)^2 - a_1(\epsilon_1)^2]\right]~,
\nonumber\\
R_4 & = & |1+n|\left[ 1 + \frac{1}{2}[({\rm Im}\,\epsilon'_2)^2 - ({\rm Im}\,\epsilon_1)^2]
+ {\rm Re}\,[a'_2(\epsilon'_2)^2 - a_1(\epsilon_1)^2] \right.
\nonumber\\
~~~~~~& & \left.- \frac{1}{2}[({\rm Re}\,\epsilon'_2)^2 - ({\rm Re}\,\epsilon_1)^2]\right]~.
\end{eqnarray}
Thus the relation (\ref{NLrel}) is now modified to
\begin{equation}
R_3 - R_4 + \frac{1}{8}|1+n|\left[(\sqrt{2R_1-1} - 1)^2 - (\sqrt{2R_2 - 1} -1)^2\right] = 0~,
\end{equation}
leading to the following constraint on the complex parameter $n$ representing a new $U=1,
U_3=0$ amplitude:
\begin{equation}\label{1+n}
|1+n| = \frac{R_3 - R_4}{\frac{1}{8}[(\sqrt{2R_2-1} - 1)^2 - (\sqrt{2R_1 - 1} -1)^2]}~.
\end{equation}
Using values of amplitudes given in Table \ref{tab:A} we calculate
\begin{equation}\label{num1+n}
|1 + n| \simeq 1 + {\rm Re}\, n= (0.95 \pm 0.02)
\left[1 + {\cal O}\left((\epsilon_1)^2, (\epsilon'_2)^2\right)\right]~.
\end{equation}
Fourth order U-spin breaking corrections have been neglected in the numerator and
denominator of (\ref{1+n}), which by themselves are both of second order. Therefore
Eq.\,(\ref{num1+n}) is valid up to second order U-spin breaking. Second order terms
in $R_3$ and $R_4$ in (\ref{Ri}) have been shown to be between one and five percent.
Adding in quadrature this uncertainty and the experimental error in (\ref{num1+n}) we obtain
\begin{equation}\label{Ren}
{\rm Re}\,n = -0.05\pm 0.05~.
\end{equation}.
\section{CP asymmetries in $D^0\to K^+K^-, \pi^+\pi^-$\label{sec:CP asymmetries}}
In the preceding sections we have neglected CP asymmetries in $D^0\to K^+K^-, \pi^+\pi^-$,
which are expected to be at most of order $10^{-3}$ in the CKM framework in agreement with
experiments~\cite{HFAG}. We have also neglected these asymmetries in the presence of new operators, consistent with neglecting
fourth order U-spin breaking which introduces uncertainties of this order.
In general, hadronic matrix elements of these new operators may involve a CP-violating
phase $\phi$ and a strong phase $\delta$, $n = |n|e^{i\delta}e^{i\phi}$.
Consequently the two processes, $D^0\to K^+K^-, \pi^+\pi^-$ acquire nonzero CP asymmetries,
\begin{equation}\label{ACP}
A_{\rm CP}(D^0 \to f) \equiv
\frac{|A(D^0 \to f)|^2 - |A(\bar D^0 \to \bar f)}{|A(D^0 \to f)|^2 - |A(\bar D^0 \to \bar f)}~,
\end{equation}
which are proportional to $|n|\sin\delta\sin\phi$. The ratios of amplitudes $R_i$ in Eqs.(\ref{Ri})
are now defined in terms of CP-averaged amplitudes
\begin{equation}
|A(D^0 \to f)|_{\rm CPav} \equiv \sqrt{\frac{1}{2}\left[|A(D^0 \to f)|^2 + |A(\bar D^0\to \bar f)|^2\right]}~.
\end{equation}
These amplitudes involve a term $|n|\cos\delta\cos\phi$ instead of
${\rm Re}\,n$ occuring in the above discussion neglecting CP asymmetries. We now summarize the situation of $\Delta R$ and $A_{\rm CP}$ defined in (\ref{NLrel}) and (\ref{ACP}) in the presence of a new CP violating phase $\phi$.
\begin{itemize}
\item For new $U=0$ operators the ratio $R_2$ and the difference $R_3- R_4$ obtain
expressions as in (\ref{Ri}) where one substitutes ${\rm Re}\,\epsilon_2 \to {\rm Re}\,\epsilon_2
+ |n|\cos\delta\cos\phi$. Thus Eq.~(\ref{NLrel}) holds also for $\phi\ne 0$.
The two CP asymmetries are equal in the U-spin symmetry approximation and have opposite
signs
(as are small CKM asymmetries \cite{Feldmann:2012js}),
\begin{equation}\label{ACPU0}
A_{\rm CP}(D^0\to \pi^+\pi^-)\simeq -A_{\rm CP}(D^0 \to K^+K^-) \simeq 2|n|\sin\delta\sin\phi~.
\end{equation}
\item For new $U=1$ operators Eq.\,(\ref{Ren}) becomes
\begin{equation}\label{1+nU1}
|n|\cos\delta\cos\phi = -0.05 \pm 0.05~.
\end{equation}
The two asymmetries are exactly equal and have the same sign,
\begin{equation}\label{ACPU1}
A_{\rm CP}(D^0\to \pi^+\pi^-) = A_{\rm CP}(D^0 \to K^+K^-) \simeq -2|n|\sin\delta\sin\phi~.
\end{equation}
\end{itemize}
Experimental constraints of CP asymmetries on $|n|$ implied by (\ref{ACPU0}) or (\ref{ACPU1}) depend on unknown values of $\delta$ and $\phi$. An uncertainty in the factor $\sin\delta\sin\phi$ is intrinsic in all earlier work studying constraints on new physics from CP asymmetries in singly
Cabibbo-suppressed $D$ decays~\cite{Grossman:2006jg,Isidori:2011qw}.
One often assumes $\sin\delta\sin\phi\sim 1$, thereby obtaining the strongest possible
constraint on $|n|$ of order $10^{-3}$. However, the constraint becomes much weaker for
small values of $\phi$ or $\delta$ and no constraint is obtained for $\phi=0$ or $\delta =0$.
Our other restriction (\ref{1+nU1}) involves the factor $\cos\delta\cos\phi$ which is complementary
to $\sin\delta\sin\phi$, becoming maximal for $\phi=0, \delta=0$. Thus combining (\ref{1+nU1}) and
(\ref{ACPU1})
leads to a more robust constraint on $|n|$ than obtained by using merely the two CP
asymmetries.
We note in passing that the new amplitude $n$ is absorbed into the definitions of
$A(D^0\to K^+K^-)$ and $A(D^0 \to \pi^+\pi^-)$ in (\ref{KKpipi}).
Therefore it does not affect the contributions of these amplitudes to the $D^0-\bar D^0$
mixing parameter $y\equiv \Delta\Gamma/2\Gamma$. These contributions are only a small
fraction of the measured mixing parameter~\cite{Gronau:2012kq}.
\section{Conclusion\label{sec:conclusion}}
We have studied the effects of new physics operators on a precise U-spin relation
for $D^0$ decays to pairs involving a charged pion or kaon. We have shown that
this relation is unaffected by new $U=0$ operators, while its sensitivity to new $U=1$
operators is at a level
of a few percent characteristic of second order U-spin breaking.
The two classes of models involving $U=0$ and $U=1$ operators may be distinguished
by the relative sign of CP asymmetries in $D^0 \to K^+K^-$ and $D^0 \to \pi^+\pi^-$.
|
\section{Introduction}
Software model-checking emerged as a natural evolution of applying model checking
to verify hardware systems.
Some factors, among several ones, that still make software model checking
challenging are: the inherently dynamic nature of software components, the
heterogeneous nature of software systems and the relatively limited amount of
modular tools (both theoretical and practical) for verifying generic software
systems.
Software systems definable as an arbitrary number of identical copies of some
process template, are called parameterized systems, and are an example of
infinite state
systems \cite{Emerson1998}. Sometimes the nature of a software system is
heterogeneous, meaning that it combines several ``characteristics'' (e.g. a clock
synchronization algorithm is supposed to work with an arbitrary number of processes
but also to terminate within a certain time). The
scarcity of modular tools is witnessed by the fact that almost everyone trying to
model check a software system, has to build his/her own toolchain that applies
several intermediate steps (usually translations and abstractions) before building
a model that can be actually model checked.
Despite such obstacles, several industries already apply model checking as part
of their software design and/or software testing stages. (e.g.,
Microsoft \cite{ball2011}, NASA \cite{Mansouri2008},
Bell Labs.\cite{godefroid2005}, IBM \cite{ben2003},
UP4ALL\footnote{http://www.uppaal.com/index.php?sida=203\&rubrik=92 URL visited on April '14}). In the aerospace industry, the DO178C international standard
\cite{RTCA11a}
even consider software model checking (or more generally, software
verification) an alternative to software testing, under suitable assumptions.
The core of our work is an extension of the Emerson and Kahlon's Cutoff
Theorem \cite{Emerson2000} to \emph{parameterized and timed systems}.
Assuming a parameterized system based on Timed Automata $U_1, \dots, U_m$ that
synchronize using conjunctive Boolean guards,
the cutoff theorem allows to compute a list of positive numbers
$(c_1, \dots, c_m)$ such that, let $\phi$ be a given specification,
then:
\[
\begin{array}{l}
\forall i \in [1,m] . (\forall n_i \in [0,\infty) ~ . ~
(U_1,\dots,U_m)^{(n_1, \dots, n_m)} \models \phi ~ \textit{iff} \\
\hspace{2cm} \forall n_i \in [0,c_i] ~ . ~
(U_1,\dots,U_m)^{(n_1, \dots, n_m)} \models \phi)
\end{array}
\]
Intuitively, the proof shows that the cutoff configuration is
\emph{trace equivalent} to each ``bigger'' system.
The contribution of this work is multifold, w.r.t. the aforementioned factors: it
reduces the problem of model checking an \emph{infinite state} real-time software
system to model checking a finite number of finite state systems; it shows a
concrete example of how to combine verification algorithms from distinct domains,
to verify what we call a \emph{heterogeneous} software systems; the cutoff
theorem for real-time systems is a theoretical tool that can be applied as a
first step when verifying a parameterized and real-time algorithm.
A second contribution is methodological: this paper describes how
to exploit the cutoff theorem to model variables that store process identifiers
(PIDs) of processes participating to the distributed algorithm.
This is non trivial, since the former relies on the fact that processes should be symmetrical, thus indistinguishable.
In order to show this, we will use a popular benchmark protocol, viz. the
Fischer's protocol for mutual exclusion. To the best of our knowledge, this is
the first time that the Fischer's protocol has been verified using model checking
techniques, for an \emph{apriori} unknown number of processes.
\section{Related Work}
\label{sec:rw}
\textbf{Infinite State System.}
Timed Automata and Parameterized Systems are two examples of infinite
state systems \cite{Emerson1998}. In general, the problem of
model checking infinite state systems is undecidable \cite{Krzysztof1986}.
A classic approach to overcome this limitation, is to find
suitable subsets of infinite state systems that can be reduced to model checking
of finitely many finite state systems, e.g. identifying a precise abstraction
(e.g. clock-zones for Timed Automata \cite{Bengtsson2004}).
Other approaches are based on the idea of finding a finite-state abstraction
that is correct but not complete, such that a property verified for the abstract
system holds for the original system as well
\cite{Clarke1986,Zuck04a,German92a,Spegni2014}.
Some other approaches are based on the idea of building an invariant representing
the common behaviors exhibited by the system \cite{Kurshan89a}.
When a given relation over the invariant is satisfied, then the desired property
is satisfied by the original system.
Its limitation is that building the abstraction or the invariant is usually not
automatic. \\
\textbf{Cutoffs for Parameterized Systems.}
Concerning the use of cutoff for model checking parameterized systems, there
exists two main approaches: computing the cutoff number of process replications
or the cutoff length of paths.
The former consists in finding a finite number of process instances such that
if they satisfy a property then the same property is satisfied by an arbitrary
number of such processes.
Emerson and Kahlon \cite{Emerson2000} established a cutoff value of about the
number of template states, for a clique of interconnected process
skeletons. In the case of rings, a constant between 2 and 5 is enough
\cite{Emerson03}. For shared resources management algorithms
\cite{Bouajjani08a}, the cutoff value is the number of resources plus the
quantified processes (in the decidable fragment of processes with equal priority).
Other works proved that one process per template is enough, for certain grids
\cite{Pagliarecci2011}.
Recently, in \cite{Rubin2014a} it has been showed that certain parameterized
systems may admit a cutoff which is not computable, while Hanna \emph{et al.}
\cite{Hanna10} proposed a procedure to compute a cutoff for Input-Output Automata
that is independent of the communication topology.
On the other hand, computing the cutoff length of paths
of a parameterized system consists in finding an upper bound on the number
of nodes in its longest computation path.
When a property is satisfied within the bounded path, then the property holds for
a system with unbound paths, i.e., with an arbitrary number of process instances.
The classic work from German and Sistla \cite{German92a}, Emerson and
Namjoshi \cite{Emerson96a} proved that such a cutoff exists for the verification of
parameterized systems composed of a control process and
an arbitrary number of user processes against indexed \textsc{ltl}
properties.
Yang and Li \cite{Yang10a} proposed a sound and complete method to compute such
a cutoff for parameterized systems with only rendezvous actions.
In that work, the property itself is represented as an automaton.
Lately it has been also showed that parameterized systems on pairwise rendezvous
do not admit, in general, a cutoff \cite{Spegni2014}.
To the best of our knowledge, cutoff theorems have not been stated previously for
timed systems. Surprisingly enough, extending Emerson and Kahlon cutoff theorems
\cite{Emerson2000} to timed systems does not increase the cutoff value.\\
\textbf{Parameterized Networks of Timed or Hybrid Automata.}
The realm of real-time systems (timed automata and, more in general, hybrid
automata) with a finite but unknown number of instances has been explored.
Abdulla and Jonsson \cite{Jonsson} proposed in their seminal work to
reduce safety properties to reachability properties.
They worked with a network composed by an arbitrary set of identical timed
automata controlled by a controller (i.e. a finite timed automaton as well).
Abdulla et al. show also that checking safety properties in networks of timed
automata
with multiple clocks is an undecidable problem \cite{Abdulla04}, as well as the
problem of determining if a state is visited infinitely often, in the continuous
time model (in the discrete time model, instead, it is decidable)
\cite{Abdulla2003}.
It should be remarked that in their undecidability proof, the network of timed
automata must rely on synchronous rendezvous in order to prove the undecidability
results. This motivated us to explore timed automata with different synchronization
mechanisms in this work.
Ghilardi et al. \cite{Carioni2010}, reduced model checking
safety properties to reachability problem.
Similarly to Abdulla and Jonsson, they applied their approach to networks
composed by an arbitrary set of timed automata interacting with a controller.
Their original contribution consisted in the usage of
Satisfiability Modulo Theories techniques.
G\"{o}thel and Glesner \cite{Gothel10a} proposed a semi-automatic verification
methodology based on finding network invariants and using both theorem proving
and model checking.
Along the same line, Johnson and Mitra \cite{johnson12} proposed a
semi-automatic verification of safety properties for parameterized networks of
hybrid automata with rectangular dynamics.
They based their approach on a combination of invariant synthesis and
inductive invariant proving. Their main limitation is that specifications
are often not inductive properties (e.g. the
mutual exclusion property it is not an inductive property). In this case one
must show that a set of inductive invariants can imply the desired property.
This last step is often not fully automatic.
We consider systems composed of a finite number of templates, each of which can
be instantiated an arbitrary number of times.
We limit Timed Automata to synchronize using Conjunctive Guards, instead of the
classic Pairwise Rendezvous \cite{Bengtsson2004},
because, as already mentioned, parameterized systems with pairwise rendezvous do
not admit, in general, a cutoff \cite{Spegni2014}.
Finally, the verification proposed in this paper is completely automatic.
\section{Parameterized Networks of Timed Automata}
\label{sec:theoretical-aspects}
\label{ssec:pnta}
This work introduces Parameterized Networks of Timed Automata (PNTA),
an extensions of Timed Automata that synchronize using conjunctive Boolean
guards.
We also introduce Indexed-Timed CTL$^\star$, a temporal logic that integrates
TCTL and MTL \cite{Bouyer2009}, for reasoning about timed processes,
together with Indexed-CTL$^\star\setminus$X \cite{Emerson2000}, for reasoning
about parametric networks of processes. In the following definition we will
make use of a set of \emph{temporal constraints} $TC(C_l)$, defined as:
\[
\begin{array}{rclclclclcl}
\textit{TC}(C) &::=& \top ~ | ~ \neg~ TC(C) ~ | ~ TC(C) ~\vee~ TC(C) ~ | \\
&& C ~ \sim ~ C ~ | ~ C ~\sim~ \mathbb{Q}^{\geq 0} \\
\end{array}
\]
where $\sim ~ \in ~ \{ <, \leq, >, \geq, = \}$, $C$ is a set of clock variables and $\mathbb{Q}$ denotes the set of
rational numbers.
\begin{definition}[\bf Timed Automaton Template]
\label{def:ta}
A Timed Automaton (TA) Template $U_l$ is a tuple
$\langle S_l, \hat{s}_l, C_l, \Gamma_l, \tau_l, I_l\rangle$ where:
\begin{itemize}
\item $S_l$ is a finite set of states, or locations;
\item $\hat{s}_l \in S_l$ is a distinguished initial state;
\item $C_l$ is a finite set of clock variables;
\item $\Gamma_l$ is a finite set of Boolean guards built upon $S_l$;
\item $\tau_l \subseteq S_l \times TC(C_l) \times 2^{C_l} \times \Gamma_l \times S_l$
is a finite set of transitions;
\item $I_l : S_l \rightarrow TC(C_l)$ maps a state to an invariant, such that $I_l(\hat{s}_l)=\top$;
\end{itemize}
\end{definition}
We will denote with $|U_l| = |S_l|$ the size of the timed automaton. A network of
timed automata can be defined as a set of $k$ TA templates, where each TA
template (say $U_l$) is instantiated an arbitrary number (say $n_l$) of times.
\begin{definition}[\bf PNTA]
\label{def:pnta}
Let $(U_1, \dots, U_k)$ be a set of Timed Automaton templates.
Let $(n_1, \dots, n_k)$ be a set of natural numbers.
Then
$$
(U_1, \dots, U_k)^{(n_1, \dots, n_k)}
$$
is a Parameterized Network of Timed Automata denoting the asynchronous parallel
composition of timed automata $U_1^1 || \dots || U_1^{n_1} || \dots || U_k^1 || \dots || U_k^{n_k}$, such that for each $l \in [1,k]$ and $i \in [1,n_l]$, then
$U_l^i$ is the \emph{i-th copy} of $U_l$.
\end{definition}
Let us remark that every component of $U_l^i$ is a disjoint copy of the
corresponding template component.
In the following will be described how every process $U_l^i$, also called
instance, can take a local step after having checked that the neighbors' states
satisfy the transition (conjunctive) Boolean guard. In such system a process can
check it is ``safe'' to take a local step, but it cannot induce a move on a
different instance.
A PNTA based on conjunctive guards is defined as follows.
\begin{definition}[\bf PNTA with Conjunctive Guards]
\label{def:cg-pnta}
Let $(U_1, \dots, U_k)^{(n_1, \dots, n_k)}$ be a PNTA.
Then, it is a PNTA with Conjunctive Guards iff every $\gamma \in \Gamma_l^i$ is a
Boolean expression with the following form:
\[
\bigwedge_{\substack{m \in [1,n_1] \\ m \neq i}} (\hat{s}_l(m) \vee s_l^1(m) \vee \dots \vee s_l^p(m)) ~ \wedge ~
\bigwedge_{\substack{h \in [1,k] \\ h \neq l}}(\bigwedge_{j \in [1,n_j]} (\hat{s}_h(j) \vee s_h^1(j) \vee \dots \vee s_h^q(j)))
\]
where, for all $l \in [1,k]$, $i \in [1,n_l]$ and $p > 0$,
$\{ s_l^1, \dots, s_l^p \} \subseteq S_l$, $s_l(i) \in S_l^i$ and
$\hat{s}_l$ is the initial states of $U_l$. The initial states $\hat{s}_l(m)$ and
$\hat{s}_h(j)$ must be present.
\end{definition}
We remark that our definitions of Timed Automaton template, PNTA and PNTA with
Conjunctive Guards are variants of the notion of \emph{timed automata} and
\emph{networks of timed automata} found in literature (e.g. \cite{Bengtsson2004}).
The \emph{operational semantics} of PNTA with conjunctive guards is expressed as
a transition system over \emph{PNTA configurations}.
\begin{definition}[\bf PNTA Configuration]
\label{def:abs-conf}
~~ \\
Let $(U_1, \dots, U_k)^{(n_1, \dots, n_k)}$ be a PNTA.
Then a \emph{configuration} is a tuple:
$$
\mathfrak{c} = (\langle \overline{s}_1, \overline{u}_1 \rangle, \dots, \langle \overline{s}_k, \overline{u}_k \rangle)
$$
where, for each $l \in [1,k]$:
\begin{itemize}
\item $\overline{s}_l : [1,n_l] \to S_l$
maps an instance to its current state, and
\item $\overline{u}_l : [1,n_l] \to (C_l \to \mathbb{R}^{\geq 0})$,
maps an instance to its clock function, s.t.
\begin{equation}
\label{eq:invariant}
\forall i ~ . ~ \overline{u}_l(i) \models I_l^i(\overline{s}_l(i))
\end{equation}
\end{itemize}
$\mathfrak{C}$ is the set of all the configurations.
\end{definition}
Intuitively, let $(\dots, \langle \overline{s}_l, \overline{u}_l \rangle, \dots)$
be a configuration, then $\overline{s}_l(i)\in S_l$ denotes the state where
instance $U_l^i$ is in that configuration.
$\overline{u}_l(i)$ is the clock assignment function (i.e.,
$\overline{u}_l(i): C_l \to \mathbb{R}^{\geq 0}$) of instance $U_l^i$ in that
configuration.
In other words, for each $c\in C_l$, $\overline{u}_l(i)(c)$ is the current value
that the clock variable $c$ assumes for instance $U_l^i$.
Any assignment to such clock variables must satisfy the invariant for the
corresponding state (see Eqn. (\ref{eq:invariant})).
The notion of transition requires some auxiliary notations.
Let $l\in[1,k]$, and let $i\in[1,n_l]$, then we call:
\begin{itemize}
\item \emph{initial configuration} \\
$\hat{\mathfrak{c}} \in \mathfrak{C}$ such that,
for each $l\in[1,k]$, for each $i\in[1,n_k]$: \\
\hspace*{1.3cm} $\overline{s}_l(i)=\hat{s}_l^i$, and \\
\hspace*{1.3cm} $\forall c \in C_l$, $\overline{u}_l(i)(c) = 0$.
\item \emph{projection} \\
$\forall \mathfrak{c} =
(\langle \overline{s}_1, \overline{u}_1 \rangle, \dots,
\langle \overline{s}_l, \overline{u}_l \rangle, \dots,
\langle \overline{s}_k, \overline{u}_k \rangle) \in \mathfrak{C}$, \\
\hspace*{1.3cm} $\mathfrak{c}(l) = \langle \overline{s}_l, \overline{u}_l \rangle$, and \\
\hspace*{1.3cm} $\mathfrak{c}(l,i) = \langle \overline{s}_l(i), \overline{u}_l(i) \rangle$.
\item \emph{state-component} \\
$\forall \mathfrak{c} =
(\langle \overline{s}_1, \overline{u}_1 \rangle, \dots,
\langle \overline{s}_l, \overline{u}_l \rangle, \dots,
\langle \overline{s}_k, \overline{u}_k \rangle) \in \mathfrak{C}$, \\
\hspace*{1.3cm} $\textit{state}(\mathfrak{c}) = (\overline{s}_1, \dots,\overline{s}_l, \dots,\overline{s}_k)$, \\
\hspace*{1.3cm} $\textit{state}(\mathfrak{c}(l)) = \overline{s}_l$, and \\
\hspace*{1.3cm} $\textit{state}(\mathfrak{c}(l,i)) = \overline{s}_l(i)$.
\item \emph{clock-component} \\
$\forall \mathfrak{c} =
(\langle \overline{s}_1, \overline{u}_1 \rangle, \dots,
\langle \overline{s}_l, \overline{u}_l \rangle, \dots,
\langle \overline{s}_k, \overline{u}_k \rangle) \in \mathfrak{C}$,
$\forall c \in C_l$,\\
\hspace*{1.3cm} $\textit{clock}(\mathfrak{c}) = (\overline{u}_1, \dots,\overline{u}_l, \dots,\overline{u}_k)$, \\
\hspace*{1.3cm} $\textit{clock}(\mathfrak{c}(l)) = \overline{u}_l$, \\
\hspace*{1.3cm} $\textit{clock}(\mathfrak{c}(l,i)) = \overline{u}_l(i)$, thus \\
\hspace*{1.3cm} $\textit{clock}(\mathfrak{c}(l,i))(c) = \overline{u}_l(i)(c)$.
\item \emph{time increase} \\
$\forall c \in C_l . \forall d \in \mathbb{R}^{\geq 0} .
(\overline{u}_l + d)(i)(c) = \overline{u}_l(i)(c) + d$ \\
\hspace*{1.3cm} $(\textit{clock}(\mathfrak{c})+d) = (\overline{u}_1+d, \dots,\overline{u}_l+d, \dots,\overline{u}_k+d)$, \\
\hspace*{1.3cm} $(\textit{clock}(\mathfrak{c}(l))+d) = (\overline{u}_l+d)$, and \\
\hspace*{1.3cm} $(\textit{clock}(\mathfrak{c}(l,i))+d) = (\overline{u}_l+d)(i)$.
\item \emph{clock reset} \\
$\forall c \in C_l ~ . ~ \forall r \subseteq C_l ~ . ~ \forall j ~ . $ \\
\hspace*{1.3cm}
$\overline{u}_l[(i,r) \mapsto 0](j)(c) = \left\{
\begin{array}{ll}
0 & \textrm{if } i = j ~ \textit{and} ~ c \in r \\
\overline{u}_l(j)(c) & \textrm{otherwise}
\end{array}
\right.$
\item \emph{clock constraint evaluation} \\
$\overline{u}_l(i) \models g$ \emph{iff}
the clock values of instance $U_l^i$ denoted by $\overline{u}_l(i)$
satisfy the clock constraint $g$; the semantics $\models$ is defined as
usual by induction on the structure of $g$;
\item \emph{guard evaluation} \\
$\textit{state}(\mathfrak{c}) \models \gamma$ \emph{iff}
the set of states of $(U_1, \dots, U_k)^{(n_1, \dots, n_k)}$
denoted by $\textit{state}(\mathfrak{c})$
satisfies the Boolean guard $\gamma$; this predicate as well can be
defined by induction on the structure of $\gamma$.
\end{itemize}
\begin{definition}[\bf PNTA Transitions]
\label{def:sem-op}
The transitions among PNTA configurations are governed by the following rules:
\vspace{0.25cm}
\\
$
\textit{(delay)} \\
\hspace*{1.0cm} \mathfrak{c} \xrightarrow{d} \mathfrak{c}^\prime \hspace{1cm}
\textrm{if} \hspace{0.3cm} d \in \mathbb{R}^{\geq 0} \\
\hspace*{3.5cm} \textit{state}(\mathfrak{c}^\prime) = \textit{state}(\mathfrak{c}) \\
\hspace*{3.5cm} \textit{clock}(\mathfrak{c}^\prime) = (\textit{clock}(\mathfrak{c})+d) \\
\hspace*{3.5cm} \forall l, i, d^\prime \in [0,d] . \textit{clock}(\mathfrak{c}(l,i))+d^\prime \models I_l^i(\textit{state}(\mathfrak{c}(l,i)))
\vspace{0.25cm}
\\
\textit{(synchronization)} \\
\hspace*{1.0cm} \mathfrak{c} \xrightarrow{\gamma} \mathfrak{c}^\prime \hspace{0.5cm} \textrm{if} \hspace{0.3cm} \exists l\in[1,k] ~ . \exists i\in[1,n_l] ~ :\\
\hspace*{3.5cm} s \xrightarrow{g,r,\gamma} t \in \tau_l^i. \\
\hspace*{3.5cm} \textit{state}(\mathfrak{c}(l,i)) = s, \\
\hspace*{3.5cm} \textit{clock}(\mathfrak{c}(l,i)) \models g, \\
\hspace*{3.5cm} \textit{state}(\mathfrak{c}) \models \gamma, \\
\hspace*{3.5cm} \mathfrak{c}^\prime(h) = \mathfrak{c}(h) ~ \textrm{for each} ~ h \neq l, \\
\hspace*{3.5cm} \mathfrak{c}^\prime(l,j) = \mathfrak{c}(l,j) ~ \textrm{for each} ~ j \neq i, \\
\hspace*{3.5cm} \mathfrak{c}^\prime(l,i) = \langle t , \textit{clock}(\mathfrak{c}(l,i))[ (r,i) \mapsto 0] \rangle \\
\hspace*{3.5cm} \textit{clock}(\mathfrak{c}^\prime(l,i)) \models I_l^i(\textit{state}(\mathfrak{c}^\prime(l,i)) \\
$
\end{definition}
Let us define what is a \emph{timed-computation} for PNTA.
\begin{definition}[\bf Timed Computation]
\label{def:timed-computation}
Let $\hat{\mathfrak{c}_0}$ be an \emph{initial configuration}, a
\emph{timed-computation} $x$ is a finite or infinite sequence of pairs:
$$
x = (\mathfrak{c}_0, t_0) \dots (\mathfrak{c}_v, t_v) \dots
$$
s.t. $t_0 = 0$ and
$\forall v \geq 0 ~ . ~ (\exists d > 0 ~ . ~ \mathfrak{c}_v \xrightarrow{d} \mathfrak{c}_{v+1} ~ \wedge ~ t_{v+1} = t_v + d) ~ \vee ~ (\exists \gamma ~ . ~ \mathfrak{c}_v \xrightarrow{\gamma} \mathfrak{c}_{v+1} ~ \wedge ~ t_{v+1} = t_v)$
\end{definition}
In other words, a timed computation can be seen as a sequence of \emph{snapshots}
of the transition system configurations taken at successive times.
It should be noticed that, according to Emerson and Kahlon \cite{Emerson2000},
in this work, it has been adopted the so-called \emph{interleaving semantics}.
This means that in a transition between two configurations, only one instance
can change its state (see the \emph{synchronization} rule in
Def. \ref{def:sem-op}).
For the sake of conciseness, let us extend the notion of
\emph{projection}, \emph{state-component}, and \emph{clock-component} to timed computations.
Let $x = (\mathfrak{c}_0, t_0) \dots (\mathfrak{c}_v, t_v) \dots$ be a timed computation, let $x_v = ( \mathfrak{c}_v , t_v )$ be the $v$-th element of $x$, then
\[
\begin{array}{lllclll}
x(l) & = & ( \mathfrak{c}_0(l) , t_0 ) \dots ( \mathfrak{c}_v(l) , t_v ) \dots &&
\textit{clock}(x_v) & = & \textit{clock}( \mathfrak{c}_v ) \\
x(l,i) & = & ( \mathfrak{c}_0(l,i) , t_0 ) \dots ( \mathfrak{c}_v(l,i) , t_v ) \dots &&
\textit{clock}(x_v(l)) & = & \textit{clock}( \mathfrak{c}_v(l) ) \\
x_v(l) & = & ( \mathfrak{c}_v(l) , t_v ) &&
\textit{clock}(x_v(l,i)) & = & \textit{clock}( \mathfrak{c}_v(l,i) ) \\
x_v(l,i) & = & ( \mathfrak{c}_v(l,i) , t_v ) && \\
\textit{state}(x_v) & = & \textit{state}( \mathfrak{c}_v ) &&
\textit{time}(x_v) & = & t_v \\
\textit{state}(x_v(l)) & = & \textit{state}( \mathfrak{c}_v(l) ) &&
\textit{time}(x_v(l)) & = & t_v \\
\textit{state}(x_v(l,i)) & = & \textit{state}( \mathfrak{c}_v(l,i) ) &&
\textit{time}(x_v(l,i)) & = & t_v \\
\\
\end{array}
\]
$x(l,i)$ is called the \emph{local computation} of the $i$-th instance of automaton template $l$.
$\textit{time}(x_v)$, $\textit{time}(x_v(l))$, and $\textit{time}(x_v(l,i))$
are the \emph{time-components} of $x_v$, $x_v(l)$, and $x_v(l,i)$ respectively.
\begin{definition}[\bf Idle Local Computation]
\label{def:idle-computation}
Let $U_l^i = \langle S_l^i, \hat{s}_l^i, C_l^i, \tau_l^i, I_l^i \rangle$ be the $i$-th instance of the \emph{timed automaton template} $U_l$.
An \emph{idle local computation} $\hat{\mathfrak{s}}(l,i)$ is a timed local computation such that, for all $v \geq 0$:
\begin{eqnarray*}
\hat{\mathfrak{s}}(l,i) & = &
( \langle\hat{s}_l^i,\overline{u}_l(i)\rangle , t_0) \dots ( \langle\hat{s}_l^i,\overline{u}_l(i)+t_v\rangle , t_v) \dots \\
\hat{\mathfrak{s}}_v(l,i) & = & ( \langle\hat{s}_l^i,\overline{u}_l(i)+t_v\rangle , t_v)
\end{eqnarray*}
where $t_0=0$ and for each $c \in C_l$, $\overline{u}_l(i)(c) = 0$.
\end{definition}
It should be noticed that for each $v$, it must be
$\overline{u}_l(i)+t_v\models I_l^i(\hat{s}_l^i)$,
since $I_l^i(\hat{s}_l^i)=\top$ according to Def. \ref{def:ta}.
Intuitively, an idle local computation is an instance of the automaton template $U_l$ that stutters in its initial state.
\begin{definition}[\bf Stuttering]
\label{def:stuttering}
Let $x$ and $y$ be two timed computations.
Let $x = x_0 \cdot \ldots \cdot x_v \cdot x_{v+1} \ldots$
The timed computation $y$ is a \emph{stuttering} of the timed computation $x$ iff
for all $v \geq 0$, there exists $r \geq 0$, such that
\begin{eqnarray*}
y & = & x_0 \cdot \ldots \cdot x_v \cdot x_{v,\delta_1} \cdot x_{v,\delta_2} \cdot \ldots \cdot x_{v,\delta_r} \cdot x_{v+1} \ldots
\end{eqnarray*}
where $\delta_1, \delta_2, \dots, \delta_r \in \mathbb{R}^{\geq 0}$,
$\delta_1 \leq \delta_2 \leq \dots \leq \delta_r$,
$t_v+\delta_r \leq t_{v+1}$, and \\
\hspace*{1cm}$x_{v,\delta_1} = (\langle\textit{state}(x_v) , \textit{clock}(x_v)+\delta_1\rangle , t_v+\delta_1)$ \\
\hspace*{1cm}$x_{v,\delta_2} = (\langle\textit{state}(x_v) , \textit{clock}(x_v)+\delta_2\rangle , t_v+\delta_2)$ \\
\hspace*{1cm}$\dots$ \\
\hspace*{1cm}$x_{v,\delta_r} = (\langle\textit{state}(x_v) , \textit{clock}(x_v)+\delta_r\rangle , t_v+\delta_r)$ \\
\end{definition}
Intuitively, the above definition means that a stuttering of a given timed
computation $x$ can be generated by inserting an arbitrary number of
\emph{delay transitions} (see Def. \ref{def:sem-op}) short enough to not alter
the validity of temporal conditions of the original computation $x$.
It only represents a more detailed view (i.e. a finer sampling) of the interval between
a configuration and the next one without changing the original sequence of states.
For the purpose of this work, timed computations conforming to Def. \ref{def:timed-computation}
(i.e. each configuration complies with Eqn. (\ref{eq:invariant})) can be classified in three different kinds
of computation:
\begin{itemize}
\item \emph{Infinite Timed Computation}:
$x$ is a timed computation of infinite length.
\item \emph{Deadlocked Timed Computation}:
$x$ is a maximal finite timed computation, i.e. in it reaches a final
configuration where all transitions are disabled.
\item \emph{Finite Timed Computation}:
$x$ is a (not necessarily maximal) final timed computation, i.e. it is
either a deadlocked computation or a finite prefix of an infinite one.
\end{itemize}
\section{A Temporal Logic for PNTA}
\label{ssec:iltl}
A dedicated logic is needed in order to specify behaviors of a PNTA.
This logic, named Indexed-Timed-CTL$^\star$, allows to reason about real-time intervals
and temporal relations (until, before, after, \dots) in systems of arbitrary size.
While its satisfiability problem is undecidable, the problem of model checking a
PNTA is proved to be decidable, under certain conditions.
\begin{definition}[\bf Indexed-Timed-CTL$^\star$]
\label{def:ITCTL}
Let $\{ P_l \}_{l \in [1,k]}$ be finite sets of atomic propositions.
Let $p(l,i)$ be any atomic proposition such that $l \in [1,k]$, $i\in\mathbb{N}^{>0}$,
and $p\in P_l$.
Then, the set of ITCTL$^\star$ formulae is inductively defined as follows:
$$
\begin{array}{rcl}
\phi &::=& \top ~ | ~ p(l,i) ~ | ~ \phi ~ \wedge ~ \phi ~ | ~ \neg \phi ~ | ~ \bigwedge_{i_l} \phi ~ | ~ A \Phi ~ | ~ A_\textit{fin} \Phi ~ | ~ A_\textit{inf} \Phi \\
\Phi &::=& \phi ~|~ \Phi ~ \wedge ~ \Phi ~|~ \neg \Phi ~|~ \Phi ~ \mathcal{U}_{\sim q} ~ \Phi \\
\end{array}
$$
where $\sim ~ \in \{ <, \leq, \geq, >, = \}$ and $q \in \mathbb{Q}^{\geq 0}$.
\end{definition}
As usual for branching-time temporal logics, the terms in $\phi$ denote
\emph{state} formulae, while terms in $\Phi$ denote \emph{path} formulae.
For the purpose of this work it is enough to assume the set of atomic propositions
coincides with the set of states of a given PNTA, i.e. $P_l = S_l$, for every $l$.
The path quantifier $A_\textit{fin}$ (resp. $A_\textit{inf}$) is a variant of the
usual universal path quantifier $A$, restricted to paths that are of finite
length (resp. infinite length). Such variants are inspired by \cite{Emerson2000}.
Missing Boolean ($\lor,\to, \dots$) operators,
temporal operators ($\mathcal{G}, \mathcal{F}, \mathcal{W}, \dots$),
as well as path quantifiers ($E, E_\textit{fin}, E_\textit{inf}$)
can be defined as usual.
The semantics of ITCTL$^\star$ is defined w.r.t. a Kripke Structure
integrating the notions of parametric system size and continuous time semantics
\cite{Bouyer2009}.
The continuous time model requires that between any two configurations it
always exists a third state. It is possible, though, introduce \emph{continuous
time computation trees}
\cite{Alur1990}. Let us call \emph{s-path} a function
$\rho : \mathbb{R}^{\geq 0} \to \mathfrak{C}$ that intuitively maps
a time $t$ with the current system configuration at that time. The mapping
$\rho_{\rfloor_{t^\prime}}:[0,t^\prime) \to \mathfrak{C}$ is a
\emph{prefix} of $\rho$ iff
$\forall t < t^\prime .
\rho_{\rfloor_{t^\prime}}(t)=\rho(t)$. The mapping
$\rho_{\lfloor_{t^\prime}}:[t^\prime,\infty) \to \mathfrak{C}$
is a \emph{suffix} of $\rho$ iff
$\forall t \geq t^\prime .
\rho_{\lfloor_{t^\prime}}(t)=\rho(t)$.
Let us take a prefix $\rho_{\rfloor_{t^\prime}}$ and an s-path
$\rho^\prime$, then their \emph{concatenation} is defined as:
\[
(\rho_{\rfloor_{t^\prime}} \cdot \rho^\prime)(t) = \left\{
\begin{array}{ll}
\rho_{\rfloor_{t^\prime}}(t) & ~ \textrm{if } t < t^\prime \\
\rho^\prime(t - t^\prime) & ~ \textrm{else}
\end{array}
\right. \\
\]
Let $\Pi$ be a set of s-paths, then
$\rho_{\rfloor_{t^\prime}} \cdot \Pi =
\{\rho_{\rfloor_{t^\prime}} \cdot \rho^\prime : \rho^\prime \in \Pi\}$.
A \emph{continuous time computation tree} is a mapping
$f : \mathfrak{C} \to 2^{[ \mathbb{R}^{\geq 0} \to \mathfrak{C} ]}$ such that:
\[
\begin{array}{l}
\forall \mathfrak{c} \in \mathfrak{C}.
\forall \rho \in f(\mathfrak{c}).
\forall t \in \mathbb{R}^{\geq 0} ~ . ~
\rho_{\rfloor_t} \cdot f(\rho(t))
\subseteq f(\mathfrak{c}).
\end{array}
\]
For the purpose of this work, here only \emph{s-paths} defined over \emph{timed
computations} will be considered.
\begin{definition}[\bf PNTA s-paths] \label{def:spath} \\
For each timed computation
$x = (\mathfrak{c}_0, t_0) \dots (\mathfrak{c}_v, t_v) \dots,$
let us call \emph{PNTA s-path} the s-path
$\rho : \mathbb{R}^{\geq 0} \to \mathfrak{C}$ satisfying:
\[
\forall v . \forall t \in [t_v, t_{v+1}) ~ . ~ \rho(t) =
\langle s,c \rangle
\]
where $s = \textit{state}(\mathfrak{c}_v)$ and $c = \textit{clock}(\mathfrak{c}_v) + t - t_v$.
\end{definition}
It should be noticed that, according to the above construction, an infinite set of
timed computations can generate the same s-path $\rho$; let us denote such set
by $\mathit{tcomp}(\rho)$.
As a consequence, for each $y \in \mathit{tcomp}(\rho)$, there exists $x \in \mathit{tcomp}(\rho)$
such that $y$ is a stuttering of $x$ (see Def. \ref{def:stuttering}).
The continuous semantics of ITCTL$^\star$ can be defined as follows.
\begin{definition}[\bf Satisfiability of ITCTL$^\star$]
~~ \\
Let $(U_1, \dots, U_k)^{(n_1, \dots, n_k)}$ be a PNTA and $\mathfrak{c}$ be the
current configuration.
Let $\phi$ denote an ITCTL$^\star$ state formula, then the
\emph{satisfiability relation} $\mathfrak{c} \models \phi$
is defined by structural induction as follows:
\[
\begin{array}{rlcl}
\mathfrak{c} \models &\top \\
\mathfrak{c} \models &p(l,i) & \textit{ iff } & p = \textit{state}(\mathfrak{c}(l,i)) \\
\mathfrak{c} \models &\phi_1 ~ \wedge ~ \phi_2 & \textit{ iff } & \mathfrak{c} \models \phi_1 ~ \textit{and} ~ \mathfrak{c} \models \phi_2 \\
\mathfrak{c} \models &\neg \phi_1 & \textit{ iff } & \mathfrak{c} \not\models \phi_1 \\
\mathfrak{c} \models &A \phi_1 & \textit{ iff } & \rho \models \phi_1, \hfill
\textit{for all} ~ \rho \in f(\mathfrak{c}) ~ \textit{and} ~ \\
& && \hfill (| \rho | = \omega ~ \textit{or} ~ \textit{deadlock}(\rho)) \\
\mathfrak{c} \models &A_\textit{inf} \phi_1 & \textit{ iff } & \rho \models \phi_1, \hfill
\textit{for all} ~ \rho \in f(\mathfrak{c}) ~ \textit{and} ~ |\rho| = \omega \\
\mathfrak{c} \models &A_\textit{fin} \phi_1 & \textit{ iff } & \rho \models \phi_1, \hfill
\textit{for all} ~ \rho \in f(\mathfrak{c}) ~ \textit{and} ~ |\rho| < \omega \\
\
\
\mathfrak{c} \models &\bigwedge_{i_l} \phi(i_l) & \textit{ iff } & \mathfrak{c} \models \phi_1(i_l), \hfill \textit{for each } i_l \in [1,n_l] \\
\\
\rho \models &\phi_1 & \textit{ iff } & \rho(0) \models \phi_1 \\
\rho \models &\phi_1~\wedge~\phi_2 & \textit{ iff } & \rho \models \phi_1 ~ \textit{and} ~ \rho \models \phi_2 \\
\rho \models &\neg \phi_1 & \textit{ iff } & \rho \not\models \phi_1 \\
\rho \models &\phi_1 ~ \mathcal{U}_{\sim q} ~ \phi_2
& \textit{ iff } & \textit{for some} ~ t^\prime \sim q, \textit{where } \sim ~ \in \{ <, \leq,\geq,>,= \} \\
&& & \hfill \rho_{\lfloor_{t^\prime}} \models \phi_2, \textit{ and } \rho_{\lfloor_t} \models \phi_1 ~ \textit{for all} ~ t \in [0,t^\prime) \\
\end{array}
\]
where $| \rho| = \omega$ (resp. $|\rho| < \omega$, resp.
$\textit{deadlock}(\rho)$) denotes that the s-path $\rho$
has infinite length (resp. has finite length, resp. is deadlocked).
\end{definition}
Note that a finite s-path is not necessarily deadlocked, since it can be a
finite prefix of some infinite s-path.
When a given PNTA $(U_1, \dots, U_k)^{(n_1, \dots, n_k)}$
satisfies an ITCTL$^\star$ state-formula $\phi$ at its initial configuration $\hat{\mathfrak{c}}$,
this is denoted by
\[
(U_1, \dots, U_k)^{(n_1, \dots, n_k)} \models \phi
\]
\begin{theorem}[\bf Undecidability of ITCTL$^\star$]
The satisfiability problem for ITCTL$^\star$ is undecidable.
\end{theorem}
\begin{proof}
The satisfiability problem for TCTL is undecidable \cite{Alur1990}.
TCTL is included in ITCTL$^\star$, therefore the latter is undecidable.
\end{proof}
In the next section we will call \textit{IMTL} the fragment of ITCTL$^\star$
having formulae with the following forms:
$\bigwedge_{i_l} Q h(i_l)$, where $Q \in \{ A, A_\textit{fin}, A_\textit{inf} \}$
and in $h$ only Boolean ($\wedge$ and $\neg$) and temporal
($\mathcal{U}_{\sim q}$) operators are allowed. We will call \textit{IMITL} the
subset of IMTL where equality constraints (i.e. $\mathcal{U}_{= q}$) are excluded.
\section{Cutoff Theorem for PNTA with Conjunctive Guards}
\label{ssec:cutoff}
\label{ssec:conj-cutoff}
In this section we prove that a cutoff can be computed to make the PMCP of PNTAs
with conjunctive guards decidable, for a suitable set of formulae.
The system in which every template is instantiated as many times as its cutoff,
will be called the \emph{cutoff system}.
Given two instantiations $I = (U_1, \dots, U_k)^{(c_1, \dots,c_k)}$ and
$I^\prime = (U_1, \dots, U_k)^{(c_1^\prime, \dots,c_k^\prime)}$, such that all
$c_i^\prime \geq c_i$ and at least one $c_j^\prime > c_j$, it can be said that
$I^\prime$ is \emph{bigger} than $I$, written $I^\prime > I$.
The cutoff theorem states that given a cutoff system $I$, for each $I^\prime > I$,
both $I^\prime$ and $I$ satisfy the same subset of ITCTL$^\star$ formulae.
\begin{theorem}[\bf Conjunctive Cutoff Theorem]
\label{the:conj-cutoff}
\ \ \\
Let $(U_1, \dots, U_k)$ be a set of TA templates with conjunctive guards.
Let $\phi = \bigwedge_{i_{l_1},\dots,i_{l_h}}Q \Phi(i_{l_1},\dots,i_{l_h})$
where $Q \in \{ A, A_\textit{inf}, A_\textit{fin},
E, E_\textit{inf}, E_\textit{fin} \}$ and $\Phi$ is an IMTL formula and
$\{ l_1, \dots, l_h \} \subseteq [1,k]$. Then
$$
\begin{array}{l}
\forall (n_1,\dots,n_k) . (U_1, \dots, U_k)^{(n_1, \dots,n_k)} \models \phi
~ ~ \textit{iff } \\
\hspace{1cm} \forall (d_1,\dots,d_k) \preceq (c_1, \dots,c_k) . (U_1, \dots, U_k)^{(d_1, \dots,d_k)} \models \phi
\end{array}
$$
where the cutoff $(c_1,\dots,c_k)$ can be computed as follows:
\begin{itemize}
\item In case $Q \in \{ A_\textit{inf}, E_\textit{inf} \}$
(i.e., deadlocked or finite timed computations are ignored). Then
$c_{l}=2$ if $l\in\{l_1,\dots,l_h\}$, and
$c_{l}=1$ otherwise (i.e. $l \in [1,k] \setminus \{ l_1,\dots,l_h\}$).
\item In case $Q \in \{ A_\textit{fin}, E_\textit{fin} \}$ (i.e. finite timed
computations, either deadlocked or finite prefixes of infinite
computations). Then
$c_{l}=1$ for each $l$.
\item In case $Q \in \{ A, E \}$ (i.e., infinite and deadlocked). Then
$c_{l}=2|U_{l}|+1$ if $l\in\{l_1,\dots,l_h\}$;
$c_{l}=2|U_{l}|$ otherwise (i.e. $l \in [1,k] \setminus \{l_1,\dots,l_h\}$).
\end{itemize}
\end{theorem}
The proof of the Cutoff Theorem consists of three steps.
The first step (\emph{Conjunctive Monotonicity Lemma}) shows that
adding instances to the system does not alter the truth of logic formulae.
The second step (\emph{Conjunctive Bounding Lemma}) proves that removing an
instance beyond the cutoff number, does not alter the truth of logic formulae
either.
The third step (\emph{Conjunctive Truncation Lemma}) generalizes the
Conjunctive Bounding Lemma to a system that has two automaton templates with an
arbitrary number of instances.
The given proofs can be generalized to systems with an arbitrary number of
templates.
\begin{theorem}[\bf Conjunctive Monotonicity Lemma]
\label{th:monotonicity}
Let $U_1$ and $U_2$ be two TA templates with conjunctive guards.
Let $\Phi(1_l)$ be an IMTL formula, with $l \in \{1, 2\}$.
Then for any $n \in \mathbb{N}$
such that $n \geq 1$ we have:
$$
\begin{array}{cl}
(i) ~ & ~ \!\!(U_1,U_2)^{(1,n)} \models Q \Phi(1_2) \Rightarrow
(U_1,U_2)^{(1,n+1)} \models Q\Phi(1_2) \\
(ii) ~ & ~ \!\!(U_1,U_2)^{(1,n)} \models Q \Phi(1_1) \Rightarrow
(U_1,U_2)^{(1,n+1)} \models Q\Phi(1_1) \\
\end{array}
$$
where $Q \in \{ E, E_\textit{inf}, E_\textit{fin} \}$.
\end{theorem}
\ifextended
\begin{proof}
\ \ \\
(i) The first part of the theorem states that there exists a
\emph{s-path} $\rho$ of $(U_1,U_2)^{(1,n)}$ such that $\rho \models \Phi(1_2)$.
For each \emph{timed-computation} $x = (\mathfrak{c}_0,t_0)(\mathfrak{c}_1,t_1),\dots$, $x \in \mathit{tcomp}(\rho)$
it is possible to build a timed-word $y = (\mathfrak{y}_0, t_0)(\mathfrak{y}_1, t_1), \dots$ of $(U_1,U_2)^{(1,n+1)}$
such that at every step $v$,
$$
y_v = ( x_v(1,1) , x_v(2,1) , \dots, x_v(2,n), \hat{\mathfrak{s}}_v(2,n+1) )
$$
that is, the $n+1$-th instance of $U_2$ always remains (i.e. stutters) in its initial state $\hat{s}_2$,
the rest of the automaton instances behave as in $x$.
Since, in this case, each conjunctive guard has the following form
\begin{equation}
\label{guard}
(\hat{s}_1^1 \vee p_1^1 \vee \dots \vee q_1^1) ~ \wedge ~
\bigwedge_{\substack{1 < i \leq n}} (\hat{s}_2^{i} \vee p_2^{i} \vee \dots \vee q_2^{i})
\end{equation}
(i.e., it includes the initial state for each instance)
adding the $n+1$-th instance stuttering in its initial state does not change the truth value of such guards, then the conclusion holds. \\
(ii) The second part of the theorem follows from a very similar argument. \\
\qed
\end{proof}
\else
A detailed proof of the theorem is in the extended version of this paper
\cite{Spalazzi14}.
\fi
Intuitively, from any time
computation $x$ one can build a new time computation $y$ where each instance
behaves as in $x$, except for a new instance of $U_2$ that halts in its initial
state (remember that by definition the initial states don't falsify any
conjunctive guard).
\begin{theorem}[\bf Conjunctive Bounding Lemma]
\label{th:bounding}
Let $U_1$ and $U_2$ be two TA templates with conjunctive guards.
Let $\Phi(1_l)$ be an IMTL formula, with $l \in \{1, 2\}$.
Then for any $n \in \mathbb{N}$
such that $n \geq 1$ we have:
$$
\begin{array}{cl}
(i) & \forall n \geq c_2 . (U_1,U_2)^{(1,n)} \models Q \Phi(1_2) ~ \Rightarrow ~ (U_1,U_2)^{(1,c_2)} \models Q \Phi(1_2)\\
(ii) & \forall n \geq c_1 .(U_1,U_2)^{(1,n)} \models Q \Phi(1_1) ~ \Rightarrow ~ (U_1,U_2)^{(1,c_1)} \models Q \Phi(1_1) \\
\end{array}
$$
where $Q \in \{ E, E_\textit{inf}, E_\text{fin} \}$ and:
\begin{itemize}
\item $c_1 = 1$ and $c_2 = 2$, when $Q = E_\textit{inf}$;
\item $c_1 = c_2 = 1$, when $Q = E_\textit{fin}$;
\item $c_1 = 2|U_2|$ and $c_2 = 2|U_2| + 1$, when $Q = E$.
\end{itemize}
\end{theorem}
\ifextended
\begin{proof}
\ \ \\
(i) ($\Rightarrow$)
The first part of the theorem states that there exists a
\emph{s-path} $\rho$ of $(U_1,U_2)^{(1,n)}$ such that $\rho \models \Phi(1_2)$.
Let $x = (\mathfrak{c}_0,t_0)(\mathfrak{c}_1,t_1),\dots$ be a \emph{timed-computation} such that
$x \in \mathit{tcomp}(\rho)$.
Then, it is possible to distinguish three distinct cases:
$x$ is an infinite computation, a deadlocked computation, or a finite computation.
Let us suppose that $x$ is an \emph{infinite computation}.
This means that $x(1,1)$ is an infinite local computation or
there exists $i \leq n$ such that $x(2,i)$ is an infinite local computation
(they are not mutually exclusive).
Then, it is possible to build a timed computation
$y = (\mathfrak{y}_0, t_0)(\mathfrak{y}_1, t_1) \dots$ of $(U_1,U_2)^{(1,c_2)}$
as follows:
\begin{eqnarray}
&& \hspace*{-.7cm}
y(1,1) = x(1,1) \label{y11} \\
&& \hspace*{-.7cm}
y(2,1) = x(2,1) \label{y21} \\
&& \hspace*{-.7cm}
y(2,2) = \left\{
\begin{array}{ll}
x(2,i) & \textrm{if $x(1,1)$ and $x(2,1)$ are finite,} \\
& \textrm{and $x(2,i)$ with $i>1$ is infinite,} \\
\hat{\mathfrak{s}}(2,2) & \textrm{otherwise}
\end{array}
\right. \label{y22} \\
&& \hspace*{-.7cm}
\forall j \in [ 3 , c_2 ] . y(2,j) = \hat{\mathfrak{s}}(2,j) \label{y2j}
\end{eqnarray}
Rule \ref{y2j} can be applied only when $c_2=2|U_2| + 1$.
It should be noticed that $y$ preserves the local timed computation of $U_2^1$ and
is a timed computation of $(U_1,U_2)^{(1,c_2)}$ (consider that either $c_2=2$ or $c_2=2|U_2| + 1$).
Indeed, for each $v$, five cases can occur:
\begin{description}
\item[I.1.]
$\textit{state}(x_{v+1})=\textit{state}(x_v)$. \\
Therefore, $\textit{state}(y_{v+1})=\textit{state}(y_v)$ by construction.
\item[I.2.]
$\textit{state}(x_v(1,1))\xrightarrow{g,r,\gamma}\textit{state}(x_{v+1}(1,1))$. \\
This means that $\textit{state}(\mathfrak{c}_v)\models\gamma$.
As $\gamma$ has the form reported in Equation (\ref{guard}), this implies $\textit{state}(\mathfrak{y}_v)\models\gamma$
(instances $U_2^i$ eventually stuttering in their initial states do not prevent the progress of instance $U_1^1$).
Furthermore, Case I.2 means that $\textit{clock}(\mathfrak{c}_v(1,1))\models g$ and, thus,
$\textit{clock}(\mathfrak{y}_v(1,1))\models g$ by construction \linebreak (Rule \ref{y11}). \\
Therefore, $\textit{state}(y_v(1,1))\xrightarrow{g,r,\gamma}\textit{state}(y_{v+1}(1,1))$.
\item[I.3.]
$\textit{state}(x_v(2,1))\xrightarrow{g,r,\gamma}\textit{state}(x_{v+1}(2,1))$. \\
Similarly to Case I.2, it is possible to prove that
$\textit{state}(y_v(2,1))\xrightarrow{g,r,\gamma}\textit{state}(y_{v+1}(2,1))$.
\item[I.4.]
$\textit{state}(x_v(2,i))\xrightarrow{g,r,\gamma}\textit{state}(x_{v+1}(2,i))$ and \\ $y(2,2)=x(2,i)$.
Similarly to Case I.2, it is possible to prove that
$\textit{state}(y_v(2,2))\xrightarrow{g,r,\gamma}\textit{state}(y_{v+1}(2,2))$.
\item[I.5.]
$\textit{state}(x_v(2,i))\xrightarrow{g,r,\gamma}\textit{state}(x_{v+1}(2,i))$ and \\ $y(2,2)\neq x(2,i)$. \\
Therefore,
$\textit{state}(y_v)=\textit{state}(y_{v+1})$ by construction.
\end{description}
Let us suppose that $x$ is an \emph{deadlocked computation}, i.e. there exists $v$ such that,
for each $l, i$, for each $x_v(l,i) \xrightarrow{g,r,\gamma} q \in \tau_l^i$, $\textit{state}(x_v) \not\models \gamma$,
where $\gamma$ has the form reported in Equation (\ref{guard}).
This means that there exists $h, j$ such that $x_v(h,j) \not\in \{\hat{s}_h^{j} \vee p_h^{j} \vee \dots \vee q_h^{j}\}$ of $\gamma$.
Then, it is possible to build a timed computation
$y = (\mathfrak{y}_0, t_0)(\mathfrak{y}_1, t_1) \dots$ of $(U_1,U_2)^{(1,2|U_2| + 1)}$
that preserves the deadlock as follows:
\begin{eqnarray}
&& \hspace*{-.7cm}
y(1,1) = x(1,1) \label{y11b} \\
&& \hspace*{-.7cm}
y(2,1) = x(2,1) \label{y21b} \\
&& \hspace*{-.7cm}
\forall k \in [ 2 , 2|U_2| + 1 ] . \nonumber \\
&& \hspace*{-.2cm}
y(2,k) = \left\{
\begin{array}{ll}
x(2,j) & \textrm{\!\!\!if there exists a distinct $y_v(l,i)$} \\
& \textrm{\!\!\!that is deadlocked by $x_v(2,j)$,} \\
\hat{\mathfrak{s}}(2,k) & \textrm{\!\!\!otherwise}
\end{array}
\right. \label{y2jb}
\end{eqnarray}
It should be noticed that, at most, there are $|U_2|$ distinct states that deadlock a local computation.
Nevertheless, when a given local state $x_v(2,i)=p$ is deadlocked by a given local state $x_v(2,j)=p$
(i.e. $x_v(2,i)=x_v(2,j)$),
then the reverse holds as well, i.e., $x_v(2,j)$ is deadlocked by $x_v(2,i)$.
This means that in the worst case, the construction of $y$ requires at most $2|U_2|$ copies of local computations of $x$
(as guards are not reflexive).
Finally, $y$ needs at least an idle local computation of $U_2$ stuttering in its initial state to assure that each guard
used to build $x$ can be fired in $y$ as well.
This justifies the choice of $c_2=2|U_2| + 1$.
At this point, it is possible to prove, similarly to the previous case, that
$y$ preserves the local timed computation of $U_2^1$ and
is a timed computation of $(U_1,U_2)^{(1,c_2)}$.
Indeed, for each $v$, five cases can occur:
\begin{description}
\item[D.1.]
$\textit{state}(x_{v+1})=\textit{state}(x_v)$. \\
Therefore, $\textit{state}(y_{v+1})=\textit{state}(y_v)$ by construction.
\item[D.2.]
$\textit{state}(x_v(1,1))\xrightarrow{g,r,\gamma}\textit{state}(x_{v+1}(1,1))$. \\
This means that $\textit{state}(\mathfrak{c}_v)\models\gamma$.
As $\gamma$ has the form reported in Equation (\ref{guard}), this implies $\textit{state}(\mathfrak{y}_v)\models\gamma$
(instances $U_2^i$ eventually stuttering in their initial states do not prevent the progress of instance $U_1^1$).
Furthermore, Case I.2 means that $\textit{clock}(\mathfrak{c}_v(1,1))\models g$ and, thus,
$\textit{clock}(\mathfrak{y}_v(1,1))\models g$ by construction \linebreak (Rule \ref{y11}). \\
Therefore, $\textit{state}(y_v(1,1))\xrightarrow{g,r,\gamma}\textit{state}(y_{v+1}(1,1))$.
\item[D.3.]
$\textit{state}(x_v(2,1))\xrightarrow{g,r,\gamma}\textit{state}(x_{v+1}(2,1))$. \\
Similarly to Case I.2, it is possible to prove that
$\textit{state}(y_v(2,1))\xrightarrow{g,r,\gamma}\textit{state}(y_{v+1}(2,1))$.
\item[D.4.]
$\textit{state}(x_v(2,j))\xrightarrow{g,r,\gamma}\textit{state}(x_{v+1}(2,j))$ and \\ $y(2,k)=x(2,j)$.
Similarly to Case I.2, it is possible to prove that
$\textit{state}(y_v(2,k))\xrightarrow{g,r,\gamma}\textit{state}(y_{v+1}(2,k))$.
\end{description}
Let us suppose that $x$ is a \emph{finite computation}.
This means that each local computation is finite.
Then, it is possible to build a timed computation
$y = (\mathfrak{y}_0, t_0)(\mathfrak{y}_1, t_1) \dots$ of $(U_1,U_2)^{(1,c_2)}$
as follows:
\begin{eqnarray}
&& \hspace*{-.7cm}
y(1,1) = x(1,1) \label{y11t} \\
&& \hspace*{-.7cm}
y(2,1) = x(2,1) \label{y21t} \\
&& \hspace*{-.7cm}
\forall j \in [ 2 , c_2 ] . y(2,j) = \hat{\mathfrak{s}}(2,j) \label{y2jt}
\end{eqnarray}
Rule \ref{y2jt} can be applied only when $c_2=2|U_2| + 1$.
It should be noticed that $y$ preserves the local timed computation of $U_2^1$ and
is a timed computation of $(U_1,U_2)^{(1,1)}$ (consider that either $c_2=1$ or $c_2=2|U_2| + 1$).
Indeed, for each $v$, four cases can occur:
\begin{description}
\item[F.1.]
$\textit{state}(x_{v+1})=\textit{state}(x_v)$. \\
Therefore, $\textit{state}(y_{v+1})=\textit{state}(y_v)$ by construction.
\item[F.2.]
$\textit{state}(x_v(1,1))\xrightarrow{g,r,\gamma}\textit{state}(x_{v+1}(1,1))$. \\
This means that $\textit{state}(\mathfrak{c}_v)\models\gamma$.
As $\gamma$ has the form reported in Equation (\ref{guard}), this implies $\textit{state}(\mathfrak{y}_v)\models\gamma$
(instances $U_2^i$ eventually stuttering in their initial states do not prevent the progress of instance $U_1^1$).
Furthermore, Case F.2 means that $\textit{clock}(\mathfrak{c}_v(1,1))\models g$ and, thus,
$\textit{clock}(\mathfrak{y}_v(1,1))\models g$ by construction \linebreak (Rule \ref{y11t}). \\
Therefore, $\textit{state}(y_v(1,1))\xrightarrow{g,r,\gamma}\textit{state}(y_{v+1}(1,1))$.
\item[F.3.]
$\textit{state}(x_v(2,1))\xrightarrow{g,r,\gamma}\textit{state}(x_{v+1}(2,1))$. \\
Similarly to Case F.2, it is possible to prove that
$\textit{state}(y_v(2,1))\xrightarrow{g,r,\gamma}\textit{state}(y_{v+1}(2,1))$.
\item[F.4.]
$\textit{state}(x_v(2,i))\xrightarrow{g,r,\gamma}\textit{state}(x_{v+1}(2,i))$ and $i\geq 2$. \\
Therefore,
$\textit{state}(y_v)=\textit{state}(y_{v+1})$ by construction.
\end{description}
In conclusion, it has been built a new Krypke Structure such that
for each $x \in \textit{tcomp}(\rho)$, with $\rho$ starting from $x_0$, there
exists $y \in \textit{tcomp}(\rho^\prime)$ such that $\rho^\prime$ starts in $y_0$
and it preserves the local timed computation of $U_2^1$.
\ \ \\
(i)($\Leftarrow$) \\
The opposite direction can be easily proved by means of repeated applications of the
Monotonicity Lemma \ref{th:monotonicity}.
\ \ \\
(ii) \\
This part can be proved by applying similar arguments. \\
\qed
\end{proof}
\fi
\begin{theorem}[\bf Truncation Lemma]
\label{th:truncation}
Let $U_1$ and $U_2$ be two TA templates with conjunctive guards.
Let $\Phi(1_l)$ be an IMTL formula, with $l \in \{1, 2\}$, then:
$$
\begin{array}{l}
\forall n_1, n_2 \geq 1 . (U_1,U_2)^{(n_1,n_2)} \models Q \Phi(1_2) \textit{ iff }
(U_1,U_2)^{(n_1^\prime,n_2^\prime)} \models Q \Phi(1_2)
\end{array}
$$
where $Q \in \{ E, E_\textit{inf}, E_\text{inf} \}$,
$n_1^\prime = \textit{min}(n_1 , c_1)$, $n_2^\prime = \textit{min}(n_2 , c_2)$,
and:
\begin{itemize}
\item $c_1 = 1$ and $c_2 = 2$, when $Q = E_\textit{inf}$;
\item $c_1 = c_2 = 1$, when $Q = E_\textit{fin}$;
\item $c_1 = 2|U_2|$ and $c_2 = 2|U_2| + 1$, when $Q = E$.
\end{itemize}
\end{theorem}
\ifextended
\begin{proof}
First of all, let us prove that \\
$(U_1,U_2)^{(n_1,n_2)} \models E\Phi(1_2)$ iff
$(U_1,U_2)^{(n_1,n_2^\prime)} \models E\Phi(1_2)$. \\
If $n_2 \leq c_2$ it is straightforward. If $n_2 > c_2$, let us set
$V_1 = U_1^{n_1}$ and $V_2 = U_2$. Then: \\
$(U_1,U_2)^{(n_1,n_2)} \models E\Phi(1_2)$ iff \\
$(V_1,V_2)^{(1,n_2)} \models E\Phi(1_2)$ iff \\
$(V_1,V_2)^{(1,n_2^\prime)} \models E\Phi(1_2)$ (by Bounding Lemma \ref{th:bounding}) iff \\
$(U_1,U_2)^{(n_1,n_2^\prime)} \models E\Phi(1_2)$. \\
Now, let us prove the lemma.
If $n_1 \leq c_1$ it is straightforward.
If $n_1 > c_1$, let us set $V_1 = U_2^{n_2^\prime}$ and $V_2 = U_1$.
Then: \\
$(U_1,U_2)^{(n_1,n_2^\prime)} \models E\Phi(1_2)$ iff \\
$(U_2,U_1)^{(n_2^\prime,n_1)} \models E\Phi(1_2)$ iff \\
$(V_1,V_2)^{(1,n_1)} \models E\Phi(1_2/1_1)$ (where index $1_2$ of $U_2$ has been substituted by index $1_1$ of $V_1$) iff \\
$(V_1,V_2)^{(1,n_1^\prime)} \models E\Phi(1_2/1_1)$ (by Bounding Lemma \ref{th:bounding}) iff \\
$(U_2,U_1)^{(n_2^\prime,n_1^\prime)} \models E\Phi(1_2)$ iff \\
$(U_1,U_2)^{(n_1^\prime,n_2^\prime)} \models E\Phi(1_2)$. \\
\qed
\end{proof}
\else
The detailed proofs of Thm. \ref{th:bounding} and Thm. \ref{th:truncation} are
given in the extended version \cite{Spalazzi14}.
\fi
Thanks to the Truncation Lemma and the duality between operators $A$ and $E$,
the Conjunctive Cutoff Theorem can be easily proved.
The Cutoff Theorem together with the known decidability and complexity results of
the model checking problems for various timed temporal logics \cite{Bouyer2009}
justify the following decidability theorem.
\begin{theorem}[\bf Decidability Theorem]
\label{th:decidability}
\ \ \\
Let $(U_1, \dots, U_k)$ be a set of TA templates with conjunctive guards
and let $\phi = \bigwedge_{i_{l_1},\dots,i_{l_h}}Q \Phi(i_{l_1},\dots,i_{l_h})$
where $Q \in \{ A, A_\textit{inf}, A_\textit{fin}, E, E_\textit{inf},
E_\textit{fin} \}$ and $\{ l_1, \dots, l_h \} \in
[1,k]$.
The parameterized model checking problem (under the continuous time semantics)
$$
\begin{array}{lc}
\forall (n_1,\dots,n_k) \succeq (1,\dots,1) . (U_1, \dots, U_k)^{(n_1, \dots,n_k)} \models \phi
\end{array}
$$
is:
\begin{itemize}
\item \textsc{undecidable}
when $\Phi$ is an IMTL formula;
\item \textsc{decidable} and \textsc{2-expspace}
when $\Phi$ is an IMITL formula;
\item \textsc{decidable} and \textsc{expspace}
when $\phi$ is a TCTL formula.
\end{itemize}
\end{theorem}
\begin{proof}
\ifextended
For the first two results, consider that the Cutoff Theorem reduces the
parameterized model checking problem to an ordinary model checking problem.
The latter is undecidable for MTL and is decidable and \textsc{expspace}-Complete
(i.e. $\textsc{DSPACE}(2^{O(n)})$) for MITL \cite{Bouyer2009}.
Since the model in the parameterized model checking problem has at most
an exponential number of states (i.e. $n = O(k \cdot |U|^{|U|}$), where $|U| =
max(|U_1|,\dots,|U_k|$)), and it is invoked at most
$\Theta(|U|^k)$ times, then the following is an upper
bound on the complexity of the parameterized model checking problem:
$\Theta(|U|^k) \cdot 2^{O(2^{|U| log(|U|)})}$, thus the problem is \textsc{2-expspace}.
Concerning the third statement, an ordinary model checking problem for TCTL
is decidable and \textsc{pspace}-Complete (i.e. $O(n^p)$, for some $p$)
\cite{Bouyer2009}. The parameterized model checking problem invokes
$\Theta(|U|^k)$ times a TCTL model checking problem, whose state space is at most
exponential, thus the complexity of the former is
$\Theta(|U|^k) \cdot O(k^p \cdot 2^{p |U| log(|U|)})$, i.e. at most
\textsc{expspace}.
\else
For the first two results, consider that the Cutoff Theorem reduces the
parameterized model checking problem to an ordinary model checking problem.
The latter is undecidable for MTL and is decidable and \textsc{expspace}-Complete
(i.e. $\textsc{DSPACE}(2^{O(n)})$, for MITL \cite{Bouyer2009}. Since the model
has an exponential number of states (i.e. $n = 2^{|U| log(|U|)}$, where $U$ is the
``biggest'' template), the problem is
at most $\textsc{2-EXPSPACE}$.
Concerning the third statement, the TCTL model checking problem is
\textsc{pspace}-Complete \cite{Bouyer2009}. Again, since the
model has an exponential number of states,
the parameterized model checking problem is at most $\textsc{expspace}$.
A more detailed proof can be found in the extended version \cite{Spalazzi14}.
\fi
\end{proof}
\section{Case Study}
\label{sec:case-study}
\begin{figure*}[t]
\center
\begin{tikzpicture}[->,>=stealth,shorten >=1pt,auto,node distance=3cm,font=\scriptsize]
\node[initial,state] (qi) {\emph{init}};
\node[state] (b1) [right of=qi] {$b_1$};
\node[state] (b2) [right of=b1] {$b_2$};
\node[state] (cs) [right of=b2] {$cs$};
\path (qi) edge node {$v = 0, c := 0$} (b1);
\path (b1) edge node {$v := PID, c:= 0$} (b2);
\path (b2) edge node {$v = PID, c > k$} (cs)
edge [bend left] node {$v \neq PID, c > k$} (qi);
\path (cs) edge [bend right] node {$v := 0$} (qi);
\end{tikzpicture}
\caption{\label{fig:proc_fischer} Process in Fischer's protocol as a Timed Automaton with integer variables}
\end{figure*}
We use the Fischer's protocol for mutual exclusion to show how to model-check a
parameterized and timed systems. The protocol uses a single timed automaton
template, instantiated an arbitrary number of time. Fig. \ref{fig:proc_fischer}
depicts such template, where $\textit{inv}(b_1) = (c \leq k)$ \cite{Carioni2010}.
In Fischer's protocol every process (a) reads and writes a PID from and into a
shared variable, and (b) waits a constant amount of time between when it asks to
enter the critical section, and when it actually does so.
The Fischer's protocol cannot be directly modeled in our framework because of the
shared variable. We will first abstract the variable into a finite state system
with conjunctive guards, and subsequently we will present the results of our
verification. \\
\textbf{Abstracting Process Identifier.}
A variable can be modeled naively as an automaton with the structure of
a completely connected graph, whose vertices denote possible assigned values
(let us call $V$ such model).
The state space can thus be infinite or finite, but even in the latter case it is
usually too big and makes the verification task unfeasible.
\begin{figure}[t]
\begin{minipage}[t][3cm][t]{0.45\textwidth}
\centering
\begin{tikzpicture}[->,node distance=2cm]
\node[initial,state] (s0) {$s_0$};
\node[state] (s1) [above of=s0] {$s_1$};
\node[state] (s2) [above right of=s0] {$s_2$};
\node[state] (s3) [right of=s0,draw=none] {$\cdots$};
\path (s0) edge (s1)
edge (s2)
edge (s3);
\path (s1) edge (s2)
edge [bend left] (s0)
edge [bend left=75] (s3);
\path (s2) edge [bend left] (s0)
edge [bend right] (s1)
edge (s3);
\path (s3) edge [bend left] (s0)
edge [bend right] (s2)
edge [bend right=100] (s1);
\end{tikzpicture}
\caption{V: a shared variable}
\end{minipage} ~
\begin{minipage}[t][3cm][t]{0.45\textwidth}
\centering
\begin{tikzpicture}[->,node distance=2cm]
\node[initial,state] (np) {\emph{diffpid}};
\node[state] (mp) [right of=np] {\emph{mypid}};
\path (np) edge [bend left] (mp)
edge [loop above] (np);
\path (mp) edge [loop above] (mp)
edge [bend left] (np);
\end{tikzpicture}
\caption{W: a process-centric view of a shared PID variable}
\end{minipage}
\end{figure}
An abstract shared variable for PIDs can be defined, under the assumptions:
\begin{itemize}
\item the variable only stores PID values;
\item the variable is shared among all processes;
\item every PID value overwrites the previous values of the variable itself;
\item every process can compare the variable value only with its own
PID value.
\end{itemize}
\vspace{-0.2cm}
As in a predicate abstraction, we replace the shared variable with its
\emph{process-centric view}. The latter has only two relevant states: it is
either the same PID as the process, or it stores a different one.
We use $W$ to denote such process.
Every process $P$ is in a one-to-one relation with its own view of the variable.
We introduce a process template $P^\prime = P \times W$ that results from the
synchronous product of the $P$ and $W$. We could then model check a system
${P^\prime}^{(n)}$. Doing this, we would probably obtain many spurious
counter-examples, since two processes could have their copy of $W$ in state
\texttt{*\_\_Mypid}. Since no variable can store multiple values, this is
impossible. Conjunctive guards, though, allow to constraint
the system in such a way that no two processes can be in a state of the
\texttt{*\_\_Mypid} group. This solution rules out the undesired spurious behaviors,
and is very convenient since it can be applied whenever an algorithm uses a shared
variable. We thus define $P^{\prime\prime}$ to be the refined version of $P^\prime$
represented in Fig. \ref{fig:fischer_uppaal} using the Uppaal notation.
It is possible to show that the abstract system simulates the concrete system,
namely $(P \times V)^{(1,n)} \preceq (P \times W)^{(1,n)}$, for any positive $n$.
\begin{figure}[t]
\begin{minipage}[t][2.5cm][t]{0.45\linewidth}
\centering\includegraphics[width=\linewidth]{figures/fischer.pdf}
\caption{\label{fig:fischer_uppaal}$P^{\prime\prime} = (P \times W) + CG$ template}
\end{minipage}
~
\begin{minipage}[t][2.5cm][t]{0.5\linewidth}
\centering\includegraphics[width=\linewidth]{figures/fischer_reduced.pdf}
\caption{\label{fig:fischer_uppaal_reduced}Reduced $P^{\prime\prime}$}
\end{minipage}
\end{figure}
Fig. \ref{fig:fischer_uppaal} depicts template $P^{\prime\prime}$. Some of the
eight states
resulting from the product are not reached by any transition, and can thus be
removed from the model, implying a smaller cutoff. The model
manipulation up to this point can be completely automatized. We notice that it is
safe to remove state \verb+b2__diff+ and connect directly
state \verb+b2__Mypid+ with \verb+Init__Diff+, obtaining the reduced
system in Fig. \ref{fig:fischer_uppaal_reduced}.
Finally, let us remark that variable \verb+mypid+ in Figg. \ref{fig:fischer_uppaal} and \ref{fig:fischer_uppaal_reduced} is added to overcome Uppaal
syntax limitations that cannot refer directly to process states in guards and
specifications. The reduced system has $4$ states, and thus the cutoff is $9$. \\
\\
\textbf{Verification Results.}
Below are the formulae that have been model checked, together with the required
time and memory. \footnote{The experiments were run on an Intel Core2 Duo CPU
T5870 @ 2.0 Ghz with 4GB RAM, OS Linux 3.13-1-amd64} \\
\begin{tabular}{llccc}
~ & \textit{Formula} & \textit{Outcome} & \textit{Time (s)} & \textit{Mem. (MB)} \\
(1) & $\bigwedge_i E \mathcal{F}_{\geq 0} (\textit{CS\_mypid}(i))$ & \textit{true} & 0.01 & 155.2 \\
(2) & $\bigwedge_{i \neq j} A \mathcal{G}_{\geq 0} ! (\textit{CS\_mypid}(i) \wedge \textit{CS\_mypid}(j))$ & \textit{true} & 30.1 & 155.2 \\
(3) & $\bigwedge_i A \mathcal{F}_{\geq 0} (\textit{CS\_mypid}(i))$ & \textit{false} & 0.59 & 155.2 \\
\end{tabular} \vspace{0.2cm} \\
Formula (1) checks that a process can enter its critical section, while (2)
checks the actual mutual exclusion property. Finally (3) states that a process will always
be able to enter its critical section.
It is well known that while the Fischer's protocol ensures the
mutual exclusion property (i.e. formulae (1) and (2)), it also suffers from the
problem of processes to possibly starve (i.e. formula (3)).
\section{Conclusions}
In this work we presented the combined study of timed and parameterized systems.
We proved that a cutoff exists for PNTA with conjunctive guards and a subset of
ITCTL$^\star$ formulae. Moreover, the cutoff value is equal to the value
computed in Emerson and Kahlon's work for untimed systems\cite{Emerson2000}.
This proves that the parameterized model checking problem
is decidable for networks of timed automata with disjunctive guards, for a
suitable logic. We remark that for timed systems, applying
Thm. \ref{the:conj-cutoff} one obtains a considerably smaller cutoff than
applying the (untimed) Emerson and Kahlon's cutoff theorem after reducing the
original timed system to a finite state system by means of the traditional region
or zone abstractions.
Finally, we used the Fischer's protocol for mutual exclusion as a benchmark for
showing how to apply the cutoff theorem. We claim that the use of conjunctive
guards is convenient
for verifying systems based of shared variables, since they naturally express the
constraint that a variable can store only one value at any time.
As a follow-up of this work, we aim at two main goals: (a) finding more algorithms
for real-time and distributed systems that can be model checked using our
framework, and (b) extending the Emerson and Kahlon cutoff theorem also to PNTA
with Disjunctive Guards.
\bibliographystyle{plain}
|
\chapter[List of publications]{List of publications}
\vspace{0.5cm}%
The chapters of this thesis are based on the following research papers:
\begin{itemize}
\item C.~Corian\`o, L.~Delle~Rose, E.~Mottola, and M.~Serino \\
\newblock{\em Graviton vertices and the mapping of anomalous correlators to momentum space for a general conformal field theory} \\
\newblock {\em JHEP} 1208 (2012) 147 [arXiv:1203.1339 [hep-th]]
\item C.~Corian\`o, L.~Delle Rose, A.~Quintavalle, M.~Serino \\
\newblock {\em Dilaton interactions and the anomalous breaking of scale invariance in the Standard Model} \\
\newblock {\em JHEP} 1306 (2013) 077, [arXiv:1206.0590 [hep-ph]]
\item C.~Corian\`o, L.~Delle Rose, C.~Marzo, M.~Serino \\
\newblock {\em Higher order dilaton interactions in the nearly conformal limit of the Standard Model} \\
\newblock {\em Phys.Lett.} B717 182-187 (2012), [ arXiv:1207.2930 [hep-ph]]
\item C.~Corian\`o, L.~Delle Rose, C.~Marzo, M.~Serino \\
\newblock {\em Conformal trace relations from the dilaton Wess-Zumino action} \\
\newblock {\em Phys.Lett. } B726 (2013) 896-905, [ arXiv:1306.4248 [hep-th]]
\item C.~Corian\`o, C.~Marzo, L.~Delle Rose, M.~Serino \\
\newblock {\em The dilaton Wess-Zumino action in 6 dimensions from Weyl-gauging: local anomalies and trace relations} \\
\newblock {\em Class. Quantum Grav.} 31 (2014) 105009, [arXiv:1311.1804 [hep-th]]
\end{itemize}
\vspace{0.5cm}
The following papers are related to the topics presented in this thesis but are not discussed in detail:
\begin{itemize}
\item C.~Corian\`{o}, L.~Delle~Rose, A.~Quintavalle and M.~Serino \\
\newblock {\em The conformal anomaly and the neutral currents sector of the Standard Model} \\
\newblock {\em Phys.Lett.}, B700 (2011) 29-38 [arXiv:1101.1624 [hep-ph]]
\item C.~Corian\`{o}, L.~Delle~Rose, and M.~Serino \\
\newblock {\em Gravity and the neutral currents: effective interactions from the trace anomaly} \\
\newblock {\em Phys.Rev.}, D83 (2011) 125028 [arXiv:1102.4558 [hep-ph]]
\item C.~Corian\`o, L.~Delle Rose, M.~Serino \\
\newblock {\em Three and four point functions of stress energy tensors in D=3 for the analysis of cosmological non-gaussianities } \\
\newblock {\em JHEP} 1212 (2012) 090 [ arXiv:1210.0136 [hep-th]]
\item C.~Corian\`o, L.~Delle Rose, E.~Mottola, M.~Serino \\
\newblock {\em Solving the conformal constraints for scalar operators in momentum space and the evaluation of Feynman's master integrals} \\
\newblock {\em JHEP} 1307 (2013) 011 [arXiv:1304.6944 [hep-th]]
\item C.~Corian\`o, A.~Costantini, L.~Delle Rose, M.~Serino \\
\newblock {\em Sum rules and the effective action of the composite Axion/Dilaton/Axino supermultiplet in N = 1 theories }
\newblock {\em JHEP} 1406 (2014) 136 [arXiv:1402.6369 [hep-ph]]
\end{itemize}
\vspace{1cm}%
{\Large \bfseries Proceedings}
\begin{itemize}
\item L.~Delle Rose, M.~Serino
\newblock {\em Dilaton interactions in QCD and in the electroweak sector of the Standard Model}\\
\newblock {\em AIP Conf.Proc.} 1492 (2012) 210-213 [arXiv:1208.6432]
\item L.~Delle Rose, M.~Serino
\newblock {\em Massless scalar degrees of freedom in QCD and in the electroweak sector from the trace anomaly } \\
\newblock {\em AIP Conf.Proc.} 1492 (2012) 205-209 [arXiv:1208.6425]
\end{itemize}
\clearpage{\pagestyle{empty}\cleardoublepage}
\chapter[Introduction]{Introduction}
The first keystone of our present understanding of the fundamental interactions is the gauge symmetry.
In fact, the structure of the Standard Model of electroweak interactions is based on the symmetry group $SU(2)_L \times U(1)_Y$,
where $L$ stands for the weak isospin, which characterizes left-handed fermions, and $Y$ is the hypercharge quantum number. Similarly, the theory of strong interactions, QCD, is entirely based on the non abelian group describing a gauged colour symmetry, $SU(3)_C$.
The symmetry of these groups dictates the structure of the fermion multiplets which couple to the corresponding gauge currents and predicts the existence of four massless gauge bosons.
By incorporating the spontaneous breaking of the gauge symmetry via the Higgs mechanism, the Standard Model accounts for a gauge-invariant generation of the masses both of the fermions and of the $W^{\pm}/ Z$ bosons mediating
weak interactions, while keeping the photon massless.
Beside the symmetry principles, the second fundamental pillar of a consistent quantum field theory is renormalizablity,
which is invoked in order to remove the infinities plaguing the perturbative computations in order to obtain predictive results.
In the case of quantum electrodynamics or QED, the problem of renormalizability, for instance, was solved in the late $40$'s by Feynman
\cite{Feynman:1949hz,Feynman:1948fi,Feynman:1949zx,Feynman:1950ir,Feynman:1948km,Feynman:1951gn},
Schwinger \cite{Schwinger:1948iu,Schwinger:1948yk,Schwinger:1948yj,Schwinger:1949ra,Schwinger:1949zz, Schwinger:1951xk,Schwinger:1953tb}, Tomonaga \cite{Tomonaga:1946zz} and Dyson \cite{Dyson:1949bp,Dyson:1949ha}.
A similar result in the case of non abelian gauge theories was presented only 20 years later, after that Yang-Mills theories
were recognized as a possible description of the fundamental interactions. In this case, the proof of renormalizability was given by 't Hooft in 1971 \cite{THooft:1971fh,THooft:1971rn}.
Together, gauge symmetry and renormalizability allow to successfully account for three of the four fundamental interactions observed in nature, i.e. electromagnetism, weak and strong interactions, at least up to the highest energy which can be experimentally tested.
Nonetheless, this picture is far from being complete. In fact, gravitation is still out of this framework, and has so far defied any attempt to a proper quantization, consistent with perturbative unitarity and renormalizability.
Its formalism is obtained by gauging the $10$-parameter Poincar\'{e} group of rigid translations, boosts and rotations
acting in Minkowski space, which turns into the group of general coordinate transformations, also called diffeomorphisms.
Despite the fact that General Relativity is a gauge theory which has successfully passed all the experimental tests performed so far
and that almost a century has passed since 1916, when it was formulated for the first time,
a consistent quantum description of gravitation is still lacking.
This happens because no renormalization program is viable for the quantized Einstein's theory.
In fact, due to well known power counting arguments, a necessary condition for a field theory to be renormalizable
is that the coupling constants in the Lagrangian must not have negative mass dimensions.
The Einstein-Hilbert action contains one such parameter, namely Newton's constant $G$, whose dimension is $\left[ M \right]^{-2}$ and therefore the attempt to quantize General Relativity inevitably results in a non renormalizable theory.
This implies that the description of the dynamics of the gravitational field needs to be modified at very high energies,
i.e. near the Planck energy $E_{P} = (\hbar\, c^5/G)^{\frac{1}{2}} \approx 1.22\,\times\, 10^{19}$ GeV, where it is expected
to play a decisive role in the dynamics of the early universe.
At present, nobody has ever formulated a consistent quantum theory of gravitation, but there are two reasons why one can draw
really significant insights by coupling a relativistic field theory to a purely classical background metric.
First, the value of $E_P$ is bigger by fifteen orders of magnitude than the highest energy at which all the known quantum field theories
can be experimentally tested so far, i.e. $14$ TeV's at the Large Hadron Collider. This means that the description of
any scattering process will not be affected at all by the quantum fluctuations in the gravitational field.
Second, and more importantly for the purposes of this work, beside the symmetry under general coordinate transformations,
which is enjoyed by any system embedded in curved space according to the formalism of General Relativity, there is another one,
which is typical of a subset of theories containing no dimensioful parameters, i.e. \emph{Weyl symmetry}.
A theory is said to be Weyl-symmetric if its action is invariant under the local rescaling of the metric tensor
$g_{\mu\nu}\rightarrow e^{2\sigma(x)}g_{\mu\nu}$, where $\sigma(x)$ is a well-behaved function of the coordinates.
It is trivial to see that the condition expressing Weyl invariance is the tracelessness of the energy-momentum tensor,
${T^\mu}_\mu = 0$. \\
On the other hand, it is known that the possibility to define a traceless energy-momentum tensor, for instance by introducing
proper terms of improvement, is the condition obeyed in Minkowski space by field theories which are invariant under the conformal group
\cite{Polchinski:1987dy}.We remark that the conformal constraints in Minkowski space are implemented through rather nontrivial operators,
whereas Weyl symmetry is much more straightforward to study. In particular, the tracelessness condition on the stress-energy tensor, derived in curved space as a consequence of
Weyl symmetry, obviously remains valid in the flat limit as well. \\
Under some mild assumptions, it is possible to provide a simple algebraic criterion in curved space which allows to establish whether
a classical field theory is conformally invariant \cite{Iorio:1996ad} or not.
There is a great advantage in studying field theories through their formal embedding in a curved space-time background. In fact many aspects of conformal invariance can be studied much more easily than in Minkowski space,
by analysing the constraints of the abelian Weyl group.
This issue is reviewed in chapter \ref{Weyl}, to set the stage for all the subsequent work presented in this thesis.
So far for classical conformal field theories.
\vspace{0.5cm}
{\large \bfseries Conformal anomalies and the low-energy effective actions for gravity}
\vspace{0.5cm}
It is known that, due to renormalization effects, the naive tracelessness condition of the energy-momentum tensor that conformal field theories
enjoy at the classical level is not inherited by the vacuum expectation value of the corresponding quantum operator
\cite{Capper:1974ic,Capper:1973mv,Capper:1974ed,Duff:1977ay,Duff:1993wm,Birrell:1982ix}.
Terms violating the classical Weyl symmetry due to quantum effects are known as \emph{conformal} or \emph{trace anomalies}
and are of two kinds.
Terms of the first kind are proportional to the beta functions of the theory and depend on the background gauge fields,
so that they vanish only at the renormalization group fixed points for theories containing interactions,
whereas they do not appear at all in free field theories.
Terms of the second kind are c-number contributions depending on the background metric tensor and are always present in the trace of the
energy-momentum tensor of any field theory.
It goes without saying that for theories which are not conformally invariant at the classical level, these terms appear in the
trace of vaccum expectation value of the energy-momentum tensor together with the vacuum expectation value (vev) of the operators which classically break conformal symmetry,
e.g. the mass terms.
An outstanding feature of the conformal anomalies is that, whether they are seen as a consequence of renormalization of UV singularities \cite{Duff:1977ay}
or as an infrared effect \cite{Giannotti:2008cv}, they affect the physics of the system at all energy scales.
This circumstance is generally true for chiral anomalies and has led 't Hooft to formulate his famous anomaly matching conditions
\cite{THooft:1979bh}, which are a powerful constraint for chiral theories describing the low-energy limit of QCD.
For conformal anomalies, as we have mentioned, there are contributions which depend on the beta-functions of the theory, and hence on the energy scale,
but there are also pure c-number contributions, built out of the metric tensor, just as in the chiral case. \\
The existence of such scale-independent terms allows to look at the connection between conformal symmetry and General Relativity also in other
ways. In fact, since anomalies are a quantum effect that does not depend on the energy scale and are built out of the metric background field,
they affect the description of gravitation in the infrared regime of the quantum theory, providing the first corrections to the
classical Einstein-Hilbert action \cite{Mottola:2006ew}.
The dynamics of a quantum system in the infrared can be effectively described by a proper low-energy action encoding
the symmetries of the theory and describing the interactions of the degrees of freedom surviving in the infrared regime \cite{Wilson:1973jj},
after integrating out the ultraviolet modes. If the theory is anomalous, the effective action is modified and can always be thought as a sum of the anomalous and of the ordinary (non anomalous) one, which is homogeneous under the action of the anomalous symmetry transformation. \\
Historically, the first example of this kind is the Wess-Zumino effective action for the $SU(3)_{L}\times SU(3)_{R}$
flavour symmetry of low-energy QCD, describing the pion dynamics \cite{Wess:1971yu} and incorporating the effects of the chiral anomalies.
In this case, pions appear as (pseudo-) Goldstone bosons introduced specifically to solve the variational problem defined by the anomalous
Ward identities. \\
It is important to notice that anomalous effective actions solving the chiral constraints can also be defined without introducing
additional scalar fields, but in this case they are non-local (see e.g. \cite{Coriano:2008pg} for an overview). \\
Of course, any variational solution of the anomaly equation cannot account for the non anomalous part part of the effective action
and is always defined modulo homogeneous terms. Determining such contributions requires a separate effort. For example,
in the case of the one-particle irreducible effective action, it is necessary to explicitly evaluate the Feynman diagrams in the perturbative series,
taking into account all the degrees of freedom in the fundamental Lagrangian.
On the perturbative side, a signature of scalar degrees of freedom is present in computations of Feynman diagrams as well.
The chiral anomaly was discovered by Schwinger \cite{Schwinger:1951nm} and, later, by Adler, Bell and Jackiw
\cite{Adler:1969gk,Bell:1969ts} in the $AVV$ diagram describing the decay of an axial-vector current ($A$) into two vector currents ($V$).
Soon after that, it was pointed out by Dolgov and Zakharov that a salient feature of this diagram is the presence of a one-particle massless pole
\cite{Dolgov:1971ri}. Finally, this massless pole was interpreted as a signature of a scalar degree of freedom, identified with the pion,
interpolating between the axial and the vector currents. This interpretation is encoded in the modified PCAC relation connecting the divergence
of the axial current to the pion field plus the chiral anomaly term \cite{Adler:1969gk}, which successfully accounts for the experimental
pion decay rate into two photons.
The situation for conformal anomalies has shown to be quite similar.
A non-local effective action for the trace anomaly in $4$ dimensions was proposed for the first time in a paper by Deser, Duff and Isham
\cite{Deser:1976yx}, but this action has a rather complicated form, as it contains logarithmic terms such as
$\log\left((\Box + R)/\mu^2\right)$, where $\mu$ is a renormalization scale and $R$ the Ricci scalar, which are hard to expand
around the flat limit $g_{\mu\nu}= \eta_{\mu\nu}$. Moreover, the possibility of the existence of such logarithmic terms in the anomalous
effective action was subsequently ruled out through cohomological arguments in
\cite{Mazur:2001aa,Boulanger:2007ab,Boulanger:2007st}.
An action which provides the minimal variational solution of the anomaly equation in a general curved background and can be
easily expanded around flat space was found by Riegert in 1987 \cite{Riegert:1987kt}. \\
A salient feature of Riegert's effective action is that, just like the non-local action for chiral anomalies, it predicts the existence of
massless scalar poles coupled to the anomaly, as shown in \cite{Giannotti:2008cv}, so that, in order to complete the correspondence
between chiral and conformal anomalies, this pole should be found in perturbative computations. \\
Given the structure of the trace anomaly (see chapter \ref{TTTVertex}), the easiest way to look for such an interpolating scalar
state in perturbation theory is to evaluate explicitly the $TVV$ correlator in flat space, where $T$ is the energy-momentum tensor and $V$
stands for a vector current.
This Green function, at $1$-loop, gives the next-to-leading order contribution to the interaction between a graviton and two gauge bosons
and is affected by the trace anomaly. In \cite{Giannotti:2008cv,Armillis:2009pq}
its computation was performed in QED and the pole predicted by Riegert's action was found.
As Riegert's action holds for non abelian gauge theories as well, the computation was subsequently performed in QCD and in the Standard Model
\cite{Armillis:2010qk,Coriano:2011zk}, confirming the presence of this contribution in all cases.
A similar analysis of anomaly poles in three point functions was performed in a supersymmetric context for $\mathcal{N}=1$ super
Yang-Mills theory in \cite{Coriano:2014gja}. In this work, it is shown that anomaly poles appear in the $\mathcal{JVV}$ correlator, with
$\mathcal{J}$ the Ferrara-Zumino hypercurrent and $\mathcal{V}$ the vector supercurrent.
The next thing to accomplish, in order to complete this program, is to look for anomaly poles in the contributions to the trace anomaly
depending only on the metric tensor. This requires the study of correlation functions of the energy-momentum tensor alone,
as this is the quantum
operator sourced by the metric tensor. Actually, a trace anomaly already affects the $2$-point function of the energy-momentum tensor,
but it depends on the renormalization scheme, so that the first correlator where scheme-independent contributions can be found
is the $TTT$ vertex.
Chapter \ref{TTTVertex} of this thesis presents a complete $1$-loop computation of the three graviton vertex in $4$ dimensions,
in three different free field theories which are conformally invariant, namely the scalar field with a proper term of improvement,
the Dirac fermion and the abelian gauge boson.
The computation is performed in the off-shell kinematic configuration, by evaluating all the diagrams in the perturbative expansion
with the Passarino-Veltman tensor-reduction technique, implemented in a symbolic manipulation program.
The results are tested by checking the general covariance and the trace Ward identities which descend from the well known master equations
for the conservation and the anomalous trace of the energy-momentum tensor. Renormalization is performed in the $\overline{MS}$ scheme.
The general result is given in terms of a set of $499$ scalar coefficients multiplying a corresponding basis of rank-$6$ tensors.
Due to the size of the general result, explicit coefficients are provided only in the limit in which two of the three gravitons are on the mass-shell, for which we expand all the correlator on a basis of only $13$ tensors.
In the end, Riegert's effective action for the conformal anomaly is explicitly introduced and we briefly review one of its two possible
local formulations, i.e. in terms of two auxiliary scalar fields, which is discussed, for instance, in \cite{Mottola:2006ew}.
In this paper the infrared effective action for gravity is studied, with special focus on the terms induced by the anomaly,
which are shown to be relevant in the infrared and, as such, provide the first quantum corrections to General Relativity.
The possible appearance of scalar poles interpolating between the gravitons and the anomalous contributions to the $TTT$ vertex
is briefly discussed as a suggestion for further investigation. \\
\vspace{0.5cm}
{\large \bfseries Conformal symmetry in position and in momentum space}
\vspace{0.5cm}
In chapter \ref{Mapping}, we temporarily turn away from the discussion of the conformal anomaly effective actions and exploit our computation
of the $TTT$ correlator to elaborate on the connection between conformal invariance in position and momentum space. \\
The implications of the constraints of conformal invariance have been worked out mostly in position space, as for instance in
\cite{Osborn:1993cr, Erdmenger:1996yc}, where the structure of various important $3$-point correlators was established,
modulo a small set of constants.
On the other hand, explicit evaluations of correlators in specific conformally invariant field theories are performed through the usual Feynman
expansion, which is commonly and most easily implemented in momentum space.
This discrepancy is likely to hinder the comparison of results found in the two ways,
especially when it comes to such complicated correlators as the $TTT$, whose first explicit perturbative computation was performed
for the first time in \cite{Coriano:2012wp} and is discussed here in chapter \ref{TTTVertex}.
Motivated by the search for a clear-cut way to map the results obtained with these different approaches into each other,
we develop two methods of comparison to which chapter \ref{Mapping} is devoted.
The first method is called the inverse mapping procedure. It starts from the integral expression in momentum space of the $1$-loop diagrams
defining the correlators and proceeds with their (inverse) Fourier transform. This allows to set a precise correspondence between
Feynman diagrams with specific different topologies and the non-local terms in the position space expressions for
$3$-point functions which are provided in \cite{Osborn:1993cr}.
Counterterms, which correspond to local terms in position space, are considered separately,
as they have to be added by hand in most cases.
Clearly, this method makes sense only with theories for which a specific Lagrangian formulation exists and that are clearly defined in momentum
space. Nevertheless, the implications of conformal constraints are much more general, as they do not rely on a specific Lagrangian.
In this sense, after using extensively the inverse-mapping procedure to compare perturbative results with the constructions in
\cite{Osborn:1993cr}, we turn to the development of a second method, which works in the opposite direction.
It is a general algorithm to Fourier-transform position space results, finding their expressions in terms of integrals in momentum space.
In order to deal with non Fourier-integrable expressions, the use of an intermediate regulator is required, in the spirit of differential regularization
\cite{Freedman:1991tk}. An interesting result of this analysis is a criterion to establish whether the conformal correlator which is studied
can be realized in the framework of a Lagrangian theory or not.
In fact, the mapping procedure can end on momentum integrals containing logarithmic terms,
which clearly cannot be generated from any Lagrangian theory.
So, if the expressions of the transformed correlators contain combinations of such terms which cannot be re-expressed as
ordinary Feynman integrals, then it is established that the underlying conformal field theory cannot be formulated in terms of a local Lagrangian. \\
\vspace{0.5cm}
{\large \bfseries Dilatons and effective actions for conformal anomalies}
\vspace{0.5cm}
Coming back to the discussion of the effective actions for conformal anomalies, we have already mentioned that Riegert's nonlocal action can be expressed in a local form
at the cost of introducing two auxiliary scalar fields \cite{Mottola:2006ew}. The presence of two such fields is the consequence
of the existence of $2$ independent cocycles for the Weyl group in $4$ dimensions
(see \cite{Mazur:2001aa} and the discussion in chapter \ref{Recursive} of this work for details). \\
However, as the Weyl group is abelian, application of the general method of Wess and Zumino to the trace anomaly
\cite{Schwimmer:2010za} implies that a local effective action can be built in terms of one single (pseudo-)Goldstone boson,
which is usually called \emph{dilaton}. The construction of this effective action is discussed extensively in chapter \ref{Recursive}.
From chapter \ref{Effective} onwards, this thesis deals with dilaton interactions. \\
Dilaton states may be either fundamental or composite scalars. In the first case, they result from the compactification
of extra dimensions (graviscalars) or, on the other hand, they may appear as effective degrees of freedom of a more fundamental field theory, similar Nambu-Goldstone (NG) modes of a broken symmetry. While massless NG modes are always present in the case of a spontaneously broken global symmetry, for radiative breakings their massless nature is not
necessarily guaranteed. In fact, non perturbative effects may contribute with a mass term and shift the position of the massless poles encountered in the 1-particle irreducible (1PI) anomaly action.
If the dilaton is not a fundamental field, then, in close analogy with the pion case, it can be thought of as an effective state mediating the coupling of matter to the trace anomaly, according to the interactions derived from the Wess-Zumino action. \\
On the perturbative side, the pole identified in the $TVV$ correlator in \cite{Giannotti:2008cv}
suggests that there might be such effective state interacting with matter via the trace anomaly.
Moreover, as the pion is a composite state of fermions, it is quite natural to elaborate on the idea that the dilaton
is a composite state of particles belonging to a strongly interacting sector which might be accessible in the near future at high energy colliders.
This possibility was suggested, for instance, in \cite{Goldberger:2007zk}.
In this sense, the Wess-Zumino action for the conformal anomaly could describe the low-energy limit of a theory
whose more fundamental components might be revealed at energies higher than those probed so far at the LHC.
Then the dilaton could prove to be an effective degree of freedom surviving at energies lower than the scale $\Lambda$
at which conformal symmetry is broken.
Chapter \ref{Effective}, which is based on \cite{Coriano:2012nm}, discusses this scenario and presents complete $1$-loop
computations of the interactions of a graviscalar particle, derived from the compactification of large extra dimensions, with the
neutral gauge currents of the Standard Model. \\
Then we turn, in the same chapter, to a discussion of scale invariant extensions of the Standard Model and to the possibility that the anomaly
poles found in the $TVV$ in these theories \cite{Giannotti:2008cv,Armillis:2009pq,Armillis:2010qk,Coriano:2011zk}
might be describing the emergence of an effective dilaton.
We must mention that a very similar scenario shows up in the $\mathcal{N}=1$ super Yang-Mills theory,
where anomaly poles appearing the triangle correlator of the Ferrara hypercurrent and two vector supercurrents can be interpreted
as a signal of the exchange of a composite dilaton/axion/dilatino multiplet in the effective Lagrangian \cite{Coriano:2014gja}.
Chapter \ref{Traced4T}, based on \cite{Coriano:2012dg}, extends the results of chapter \ref{Effective}
by working in the conformal limit of the Standard Model, in which all the masses are set to zero.
In this regime, we present the computation of $3$- and $4$-point traced correlators of the energy-momentum
tensor, which exactly match dilaton self-interactions in the on-shell limit.
Techniques presented in chapter \ref{Mapping}, relying on the connection between the trace anomaly and the
gravitational counterterms for conformal field theories, are used to secure the correctness of the result.
Finally, chapter \ref{Recursive} deals with the Wess-Zumino action for the geometric sector of the conformal anomaly
both in $4$ and $6$ dimensions, putting together the results presented in \cite{Coriano:2013xua,Coriano:2013nja}.
The anomalous effective action is explicitly built by using the most general renormalization scheme, exploiting a cohomological
method presented for the first time in \cite{Mazur:2001aa}. Possible kinetic terms for the dilaton, which obviously cannot be derived by
the analysis of the anomalous constraints, are systematically reviewed.
After the derivation of the most general anomalous effective action, an interesting result is proven.
First, the Wess-Zumino anomalous effective action is written in another way,
i.e. as a perturbative functional expansion with respect to the dilaton field,
which is also a power series in the inverse conformal breaking scale, $1/\Lambda$.
Each term of this power series, in turn, is a well defined and simple combination of traced Green functions of the energy-momentum tensor.
For consistency, then, one can require that each term in the perturbative expansion must match the term proportional to the same power
of $1/\Lambda$ in the explicit expression of the anomalous effective action which is derived in the first part with cohomological methods.
Imposing this consistency condition results in an infinite set of recurrence relations,
which allow to compute traced correlators of the energy-momentum tensor to an arbitrarily high order.
\mainmatter
\clearpage{\pagestyle{empty}\cleardoublepage}
\chapter{Conformal symmetry and Weyl symmetry}\label{Weyl}
\section{Introduction}
In this introductory chapter, we set the stage for all the results to be presented in the rest of the thesis.
The computations that we present are performed by embedding the quantum field theories that we are going to investigate
in a curved metric background $g_{\mu\nu}$.
Hereafter, the term \emph{matter fields} will be used by us to refer to any fundamental field except for the metric tensor.
In particular we are going to present a short introduction on the concept of Weyl symmetry and its various realization in a curved background.
This will be useful for the analysis presented in the later chapters.
Weyl invariance in a curved background allows to address the issue of conformal invariance in any free-falling frame.
We recall that conformal invariance implies the possibility to define a traceless energy-momentum tensor,
as discussed, for instance, in \cite{Polchinski:1987dy}. The search for theories which exhibit scale invariance but not conformal invariance has
been at the center of several recent studies, as reviewed in \cite{Nakayama:2013is}. The advantage of dealing with a Weyl invariant theory in a
curved background respect to a conformal symmetric theory in flat background is in the different character of the two symmetry groups, as
the first one is abelian.
Therefore, one can derive specific implications for a certain conformal invariant theory by starting from a simpler Weyl invariant
theory on a curved background and the specialising the result to a local free falling frame \cite{Iorio:1996ad}.
In the following sections, we first introduce the conformal group, then proceed to discuss Weyl invariance in curved spacetime background.
Finally, we review the argument presented in \cite{Iorio:1996ad}. Our attention will be limited to the case of any spacetime dimensions except for the case of $d=2$ since in this case the conformal group is infinite dimensional.
\section{The conformal group}
We present a brief review, in $d > 2$ dimensions and euclidean space, of the transformations which identify the conformal group $SO(2,d)$.
These may be defined as the transformations $x_\mu \rightarrow x'_\mu(x)$ that preserve the infinitesimal length up to a local factor
\begin{eqnarray}
d x_\mu d x^\mu \rightarrow d x'_\mu d x'^\mu = \Omega(x)^{-2} d x_\mu d x^\mu \, .
\label{ConformalMeasure}
\end{eqnarray}
In the infinitesimal form, the conformal transformations are given by
\begin{eqnarray}
\label{xtransf}
x'_\mu(x) = x_\mu + a_\mu + \omega_{\mu\nu}\, x^\nu + \sigma\, x_\mu + b_\mu\, x^2 - 2\, b \cdot x\, x_\mu\ , ,
\end{eqnarray}
with
\begin{eqnarray}
\Omega(x) = 1 -\lambda(x) \qquad \mbox{and} \qquad \lambda(x) = \sigma - 2 b \cdot x \, .
\label{Omega}
\end{eqnarray}
The transformation in eq. (\ref{xtransf}) is defined by translations ($a_\mu$),
boosts and rotations ($\omega_{\mu\nu} = - \omega_{\nu\mu}$),
dilatations ($\sigma$) and special conformal transformations ($b_\mu$).
The first two define the Poincar\'{e} subgroup which leaves invariant the infinitesimal length and for which $\Omega(x) = 1$.
If we also consider the inversion
\begin{eqnarray}
x_\mu \rightarrow x'_\mu = \frac{x_\mu}{x^2} \, , \qquad \qquad \Omega(x) = x^2 \, ,
\end{eqnarray}
we can enlarge the conformal group to $O(2,d)$.
Special conformal transformations can be realized by a translation preceded and followed by an inversion. \\
The Poincar\'{e} subgroup containts the basic set of symmetries for any relativistic system.
Its albegra is given by
\begin{eqnarray}
i\, \left[ J^{\mu\nu}, J^{\rho\sigma}\right]
&=&
\delta^{\nu\rho}\, J^{\mu\sigma} - \delta^{\mu\rho}\, J^{\nu\sigma}
- \delta^{\mu\sigma}\, J^{\rho\nu} + \delta^{\nu\sigma}\, J^{\rho\mu} \, ,\nonumber \\
i\, \left[P^\mu, J^{\rho\sigma} \right]
&=&
\delta^{\mu\rho}\,P^\sigma - \delta^{\mu\sigma}\, P^\rho \, ,\nonumber \\
\left[P^\mu,P^\nu\right]
&=& 0 \, ,
\end{eqnarray}
where the $J$'s are the generators of the Lorentz group and the four components of the momentum $P^\mu$ generate rigid translations.
For scale invariant theories, which contain no dimensionful paramters, the Poincar\'{e} group can be extended by including the dilatation
generator $D$, corresponding to the fourth terms in the coordinate transformations in (\ref{xtransf}),
whose commutation relations with the other generators are
\begin{eqnarray}
\left[ P^\mu, D \right] &=& i\, P^\mu\, , \nonumber \\
\left[ J^{\mu\nu}, D \right] &=& 0\, .
\label{Dilatation}
\end{eqnarray}
Finally, it is possible to further extend this group so as to include the four special conformal transformations,
whose generators we call $K^\mu$, extending the algebra as
\begin{eqnarray}
\left [ K^\mu,D \right] &=& -i\, K^\mu \, , \nonumber \\
\left [ P^\mu, K^\nu \right] &=& 2\,i\,\delta^{\mu\nu}\,D + 2\,i\,J^{\mu\nu}\, , \nonumber \\
\left [ K^\mu,K^\nu \right] &=& 0 \, , \nonumber \\
\left [ J^{\rho\sigma}, K^\mu \right] &=& i\,\delta^{\mu\rho}\,K^\sigma - i\,\delta^{\mu\sigma}\,K^ \rho \, .
\label{Conformality}
\end{eqnarray}
By the way, it is clear from (\ref{Dilatation}) and (\ref{Conformality}) that scale invariance does not require conformal invariance,
but conformal invariance necessarily implies scale invariance.
Having specified the elements of the conformal group, we can define a quasi primary field $\mathcal O^i(x)$, where the index $i$ runs over the
representation of the group to which the field belongs, through the transformation property under a conformal transformation $g$
\begin{eqnarray}
\mathcal O^i(x) \stackrel{g}{\rightarrow} \mathcal O'^i(x') = \Omega(x)^\eta D^i_j(g) \mathcal O^j(x) \,,
\end{eqnarray}
where $\eta$ is the scaling dimension of the field and $D^i_j(g)$ denotes the representation of $O(1,d-1)$.
In the infinitesimal form we have
\begin{eqnarray}
\delta_g \mathcal O^i(x) =
- (L_g \mathcal O)^i(x) \,, \qquad \mbox{with} \qquad L_g = v \cdot \partial + \eta \, \lambda + \frac{1}{2}
\partial_{[ \mu} v_{\nu ]} \Sigma^{\mu\nu} \,,
\end{eqnarray}
where the vector $v_\mu$ is the infinitesimal coordinate variation $v_\mu = \delta_g x_\mu = x'_\mu(x) - x_\mu$ and
$(\Sigma_{\mu\nu})^i_j$ are the generators of $O(1,d-1)$ in the representation of the field $\mathcal O^i$.
The explicit form of the operator $L_g$ can be obtained from eq. (\ref{xtransf}) and eq. (\ref{Omega}) and is given by
\begin{align}
& \mbox{translations:} && L_g = a^\mu \partial_\mu \,, \nonumber \\
& \mbox{rotations:} && L_g = \frac{\omega^{\mu\nu}}{2} \left[ x_\nu \partial_\mu - x_\mu \partial_\nu - \Sigma_{\mu\nu} \right] \,,
\nonumber \\
& \mbox{scale transformations :} && L_g = \sigma \left[ x \cdot \partial + \eta \right] \,, \nonumber \\
& \mbox{special conformal transformations. :} && L_g = b^\mu \left[ x^2 \partial_\mu - 2 x_\mu \, x \cdot \partial - 2 \eta \, x_\mu - 2 x_\nu
{\Sigma_{\mu}}^{\nu} \right] \,.
\end{align}
As already remarked, invariance of a matter system under the conformal group implies the possibility to define a
traceless energy-momentum tensor $T_I^{\mu\nu}$ \cite{Polchinski:1987dy},
\begin{equation}
{T_I^\mu}_\mu = 0\, ,
\end{equation}
where the subscript $I$ indicates that possible improvement terms have been added to the minimal energy-momentum tensor
which is obtained from the sole requirement of invariance under the Poincar\'{e} group.
\section{Weyl symmetry and its connection to conformal invariance}
Now that conformal symmetry in flat space has been introduced, we can move on to curved space and discuss Weyl symmetry.
We assume that the reader is familiar with basic General Relativity, including the Vielbein formalism which is necessary to embed
fermions in a gravitational field, for which we refer to \cite{Birrell:1982ix,Weinberg:1972gc}.
In the following, we follow the discussion in \cite{Iorio:1996ad}.
Let us suppose that our theory in flat space is described by an action functional
\begin{equation}
\mathcal S = \int d^d x \, \mathcal{L}(\Phi,\partial_\mu\Phi)\, ,
\end{equation}
depending on the matter fields $\Phi$ and their first derivatives.
This theory can be easily embedded in curved space, replacing ordinary derivatives with diffeomorphism-invariant ones
\begin{equation}
\mathcal S = \int d^dx\, \sqrt{g}\, \mathcal{L}(\Phi,\nabla_\mu\Phi)\, ,
\end{equation}
The energy-momentum tensor of the theory is defined as the source of the gravitational
field appearing in Einstein's equations and it is given by
\begin{equation}
T^{\mu\nu} = \frac{2}{\sqrt{g}}\, \frac{\delta \mathcal{S}}{\delta g_{\mu\nu}}\, .
\end{equation}
For a Lagrangian in flat space written in a diffeomorphic invariant form, scale invariance is equivalent to global Weyl invariance.
The equivalence can be shown quite straightforwardly by rewriting a scale transformation acting on the coordinates of flat space
and the matter fields $\Phi$,
\begin{eqnarray} \label{ChIntFlatScaling}
x^\mu &\to& {x'}^{\mu}=e^\sigma x^\mu \, , \nonumber\\
\Phi(x)&\to& \Phi'(x')=e^{-d_\Phi \sigma}\Phi(x) \, ,
\end{eqnarray}
in terms of a rescaling of the metric tensor, the Vielbein and the matter fields
\begin{eqnarray} \label{ChIntCurvedScaling}
g_{\mu\nu}(x) &\to& e^{2\sigma}\, g_{\mu\nu}(x)\, , \nonumber\\
{V}_{a\,\rho}(x) &\to& e^{\sigma}\, V_{a\,\rho}(x)\, , \nonumber\\
\Phi(x)&\to& e^{-d_\Phi \sigma}\Phi(x)\, ,
\end{eqnarray}
leaving the coordinates $x$ of the manifold invariant. We have denoted with $d_\Phi$ the field scaling dimension,
which is deduced by an ordinary dimensional analysis of the Lagrangian density.
The reason why (\ref{ChIntFlatScaling}) can be traded for (\ref{ChIntCurvedScaling}) is that metric tensors
and the Vielbein always appear in order to contract derivative terms to obtain diffeomorphic scalars.
Once we move to a curved metric background, it is natural to promote the global scaling parameter $\sigma$
to a local function, so that the transformation laws of the metric, Vielbein and matter fields are
\begin{eqnarray}
\label{ChIntWeylTransf}
{g'}_{\mu\nu}(x) &=& e^{2\sigma(x)}\, g_{\mu\nu}(x)\, , \nonumber \\
{V'}_{a\,\rho}(x) &=& e^{\sigma(x)}\, V_{a\,\rho}(x)\, , \nonumber \\
\Phi'(x) &=& e^{-d_{\Phi}\, \sigma(x)}\, \Phi(x)\, .
\end{eqnarray}
The transformations of metric, Vielbein and matter fields shown in (\ref{ChIntWeylTransf}) define the abelian Weyl group.
It is natural to ask whether is possible to modify the theory in such a way that (\ref{ChIntWeylTransf}) leave the action functional invariant.
Historically, the scale symmetry was the first whose gauging was systematically studied, in an attempt, made by Weyl, to connect
electromagnetism with geometry. For this reason, this procedure was named Weyl-gauging.
It can be implemented in the same way as for QED, introducing an appropriate new field which takes a role similar to the vector potential.
This allows to define a new Lagrangian which is diffeomorphic and Weyl invariant at the same time.
For instance, for a free scalar theory described by the action
\begin{equation}
\mathcal{S}_{\phi} = \frac{1}{2}\, \int d^d x\, \sqrt{g} \, g^{\mu\nu} \, \partial_{\mu}\phi\, \partial_\nu \phi ,
\label{ChIntkin}
\end{equation}
the derivative terms are modified according to
\begin{equation} \label{ChIntDerTransf}
\partial_\mu \rightarrow \partial^W_\mu = \partial_\mu - d_{\phi}\, W_{\mu}\, ,
\end{equation}
where $W_{\mu}$ is a vector gauge field that shifts under a Weyl trasnformation as
\begin{equation} \label{ChIntWeylGaugeField}
W_{\mu} \rightarrow W_{\mu} - \partial_\mu \sigma \, .
\end{equation}
In the case of a covariant derivative acting on higher spin fields, such as a spin-$1$ field $v_\mu$, the Weyl-gauging has to be supplemented
with a prescription to render the generally covariant derivative Weyl invariant, which is to add to (\ref{ChIntDerTransf}) the modified
Christoffel connection
\begin{equation} \label{ChIntModChristoffel}
\hat\Gamma^\lambda_{\mu\nu} =
\Gamma^\lambda_{\mu\nu} + {\delta_\mu}^\lambda\, W_\nu + {\delta_\nu}^\lambda\, W_\mu - g_{\mu\nu}\, W^\lambda \, .
\end{equation}
It is easy to check that this Christoffel symbol is Weyl invariant.
So, pursuing closely the analogy with the gauging of a typical abelian theory,
we can define the Weyl covariant derivatives acting on vector fields as
\begin{eqnarray} \label{ChIntWeylChristoffel}
\nabla^W_\mu v_\nu &=& \partial_\mu v_\nu - d_v\, W_\mu v_\nu - \hat\Gamma^\lambda_{\mu\nu} v_\lambda \, , \nonumber\\
\nabla^W_\mu v_\nu &\rightarrow& e^{- d_v\sigma(x)}\, \nabla^W_\mu v_\nu \, ,
\end{eqnarray}
with an obvious generalisation to tensors of arbitrary rank. \\
Of course, the extension of such a derivative to the fermion case requires the Vielbein formalism and is obtained by the relation
\begin{equation} \label{ChIntWeylSpinConnection}
\nabla_{\mu} \rightarrow \nabla^W_\mu = \nabla_{\mu} - d_\psi\, W_{\mu} + 2\, {\Sigma_\mu}^\nu\, W_\nu\, , \quad
\Sigma^{\mu\nu} \equiv V^{a\mu}\, V^{b\nu} \Sigma_{ab}\, ,
\end{equation}
where we have denoted with $d_\psi$ the scaling dimension of the spinor field ($\psi$) and with
$\Sigma_{ab}$ the spinor generators of the Lorentz group.
If we Weyl-gauge the scalar action (\ref{ChIntkin}) according to the prescriptions in (\ref{ChIntDerTransf})
and (\ref{ChIntWeylGaugeField}), we obtain
\begin{eqnarray} \label{ChIntWeylGaugeFieldScalar}
S_{\phi,W}
&=&
\frac{1}{2}\, \int d^dx\, \sqrt{g}\, g^{\mu\nu}\, \partial^W_\mu \phi\, \partial^W_\nu \phi
= \frac{1}{2}\, \int d^dx\, \sqrt{g}\, g^{\mu\nu}\,\bigg( \partial_\mu - \frac{d-2}{2}\, W_\mu \bigg)\,\phi\,
\bigg( \partial_\nu - \frac{d-2}{2}\, W_\nu \bigg)\, \phi \nonumber \\
&=&
\frac{1}{2}\, \int d^dx\, \sqrt{g}\, g^{\mu\nu} \bigg\{ \partial_\mu \phi\, \partial_\nu\phi
- \frac{d-2}{2}\, \bigg( \phi\,W_\mu\,\partial_\nu\phi + \phi\,W_\nu\,\partial_\mu \phi - \frac{d-2}{2}\, W_\mu\,W_\nu\,\phi^2 \bigg)
\bigg\}
\end{eqnarray}
which, using $\phi\, \partial_{\mu}\phi = 1/2\, \partial_\mu \phi^2$ and integrating by parts, can be written in the form
\begin{equation}
\mathcal S_{\phi,W} =
\frac{1}{2}\, \int d^dx\, \sqrt{g}\, g^{\mu\nu}\, \bigg(\partial_\mu \phi\, \partial_\nu\phi
+ \phi^2\, \frac{d-2}{2}\, \Omega_{\mu\nu}(W) \bigg)\, ,
\label{ChIntgauged}
\end{equation}
where we have introduced the quantity
\begin{equation} \label{ChIntOmega}
\Omega_{\mu\nu}(W) = \nabla_\mu W_\nu - W_\mu\,W_\nu + \frac{1}{2}\,g_{\mu\nu}\, W^2 \, .
\end{equation}
The result of this procedure is a Weyl invariant Lagrangian in which the Weyl variation of the
ordinary kinetic term of $\phi$ is balanced by the variation of the $\Omega$ term.
This term plays a prominent role in the subsequent discussion, as we are going to ask whether it is possible to
enhance scale invariance to Weyl invariance without introducing the new degree of freedom $W_\mu$.
To this aim, we notice that there are two second rank tensors which can be built out of $W_\mu$ and its first covariant derivative,
namely $W_\mu W_\nu$ and $\nabla_\mu W_\nu$. \\
From (\ref{ChIntWeylGaugeField}), we can infer how they transform under a finite Weyl scaling,
\begin{equation}
\Delta (W_\mu W_\nu) = \sigma_\mu\sigma_\nu -(W_\mu\sigma_\nu+W_\nu\sigma_\mu)\, ,
\label{DeltaW2}
\end{equation}
where we habve introduced $ \sigma_\mu \equiv \partial_\mu \sigma $ to keep the notation easy.
To compute the finite Weyl-variation of $\Delta (\nabla_\mu W_\nu)$ we recall the
transformation rule for the Christoffel connection, $\Gamma_{\mu\nu}^\lambda$
\begin{equation}
\Delta \Gamma_{\mu\nu}^\lambda = g^{\lambda\sigma}
\bigl(g_{\mu\sigma}\sigma_\nu+
g_{\nu\sigma}\sigma_\mu-
g_{\mu\nu}\sigma_\sigma\bigr), \label{5.2}
\end{equation}
which implies
\begin{equation}
\Delta (\nabla_\mu W_\nu)= - \nabla_\mu\sigma_\nu - g_{\mu\nu}\, \sigma \cdot \sigma
+ 2\, \sigma_\mu\sigma_\nu + g_{\mu\nu}\, W \cdot \sigma - \left(W_\mu\, \sigma_\nu + W_\nu\, \sigma_\mu \right)\, .
\label{DeltaNablaW}
\end{equation}
If we notice that contracting (\ref{DeltaW2}) we obtain
\begin{equation}
\Delta (g_{\mu\nu}W \cdot W)= g_{\mu\nu}\left(\sigma \cdot \sigma - 2\, W \cdot \sigma \right) \, ,
\label{5.4}
\end{equation}
then we immediately conclude that the variation under a finite Weyl shift of $\Omega_{\mu\nu}[W]$
is independent of $W_\mu$ and symmetric. More precisely, it is given by
\begin{equation}
\Delta\Omega_{\mu\nu}[W] =
- \left(\nabla_\mu\sigma_\nu- \sigma_\mu\sigma_\nu
+ \frac{1}{2}\,g_{\mu\nu}\;\sigma\!\cdot\! \sigma \right) = - \Omega_{\mu\nu}[\sigma] \, .
\label{5.6}
\end{equation}
As $\Omega_{\mu\nu}[\sigma]$ depends only on the scaling parameter $\sigma$,
one can argue that it may be related to some purely geometrical object.
In fact, the variation of the Ricci tensor under a local scale trasnformation is given by
\begin{equation}
\Delta R_{\mu\nu}= R_{\mu\nu}[e^{2\sigma} g_{\mu\nu}]-R_{\mu\nu}[g_{\mu\nu}]
= g_{\mu\nu}\nabla^2 \sigma
+(n-2)\Bigl(\nabla_\mu \sigma_\nu-\sigma_\mu \sigma_\nu+g_{\mu\nu}\;\sigma\!\cdot\! \sigma\Bigr)\, .
\label{RicciVarInt}
\end{equation}
From (\ref{RicciVarInt}) we see that the tensor
\begin{equation}
S_{\mu\nu}= R_{\mu\nu}- \frac{g_{\mu\nu}}{2(n-1)}\, R
\label{DefineSInt}
\end{equation}
transforms under Weyl-scalings in the same way as $\Omega_{\mu\nu}$, i.e.
\begin{equation}
\Delta S_{\mu\nu}=(n-2)\Omega_{\mu\nu}[\sigma] \, .
\label{DeltaSInt}
\end{equation}
Now it is clear when Weyl-gauging can be replaced by a non-minimal coupling to the curvature.
Since, according to (\ref{DefineSInt}) and (\ref{DeltaSInt}), the Weyl variation of $\Omega_{\mu\nu}[W]$
is proportional to the variation of $S_{\mu\nu}$, we see that whenever $W_\mu$ appears in the action
only in the combination $\Omega_{\mu\nu}[W]$, it can be replaced by $S_{\mu\nu}$.
Replacing terms depending on $W_\mu$ with a non minimal couplig to the Ricci tensor is a procedure referred to as
\emph{Ricci gauging}.
We still have to explore under what conditions, in an action which is Weyl-gauged, the terms depending on $W_\mu$
appear only in the combination $\Omega_{\mu\nu}$.
We are going to see that the sufficient condition for Ricci gauging to be possible is preciely conformal invariance in flat space.
We can start our discussion by representing a conformal transformation of the metric in the form of a diffeomorphism
\begin{equation}
\frac{\partial x^\mu}{\partial x'^\alpha} \frac{\partial x^\nu}{\partial x'^\beta}g_{\mu\nu}(x) =
\hat g_{\alpha\beta}(x') \quad \text{and} \quad \hat g_{\alpha\beta}(x)= e^{\hat \sigma(x)}g_{\alpha\beta}(x) \, .
\label{6a.1}
\end{equation}
The functions $\hat\sigma$ in (\ref{6a.1}) form a subgroup of the group of local Weyl transformations that is induced
by conformal transformations, that we call the conformal Weyl group.
Given (\ref{6a.1}), we can characterize the functions $\hat \sigma$ via the condition
\begin{equation}
\frac{\partial x^\mu}{\partial x'^\alpha}\frac{\partial x^\nu}{\partial x'^\beta}R_{\mu\nu}(x)=
\hat{R}_{\alpha\beta}(y)= R_{\alpha\beta}[e^{2 \hat{\sigma}}g_{\mu\nu}](x')\, ,
\label{6a.2}
\end{equation}
which leads, in virtue of the transformation law of the tensor $S_{\mu\nu}$ under Weyl shifts, given by
(\ref{RicciVarInt}) and (\ref{DeltaSInt}), to the differential equation
\begin{equation}
(n-2)\sigma_{\alpha\beta}[\hat\sigma] = \hat S_{\mu\nu}-S_{\mu\nu}\, .
\label{6a.3}
\end{equation}
The existence of global solutions of (\ref{6a.3}) is a non-trivial problem in general.
Now, as in the flat-space limit $S_{\alpha\beta}$ vanishes, the condition (\ref{6a.3}) reduces to
\begin{equation}
\partial_\nu \hat \sigma_\mu - \hat\sigma_\mu\hat\sigma_\nu + \frac{g_{\mu\nu}}{2} \;\hat\sigma\!\cdot\!\hat\sigma = 0 \, .
\label{ChapIntDiffCondition}
\end{equation}
We find that the general solution of (\ref{ChapIntDiffCondition}) identifying the subset of the $\hat\sigma$ functions
defining the conformal group of flat space is simply
\begin{equation}
\hat\sigma(x)= \log\left( \frac{1}{1- 2 \, b \cdot x + b^2 \, x^2} \right) \, ,
\label{ChIntHatSigma}
\end{equation}
where $b$ is any constant vector. We see that (\ref{ChIntHatSigma}), for $b$ infinitesimal,
corresponds exactly to eq. (\ref{Omega}). When exponentiated as in (\ref{6a.1}),
it gives a metric tensor producing the variation of the line interval defined in (\ref{ConformalMeasure}). \\
Now suppose that an action $\mathcal{S}$ admits Ricci gauging, so that the gauged action satisfies the condition
\begin{equation}
\mathcal{S}(\Phi, S_{\mu\nu})= \mathcal{S}(\Phi', S_{\mu\nu}+ (n-2)\, \Omega_{\mu\nu}[\sigma])\, ,
\label{ca}
\end{equation}
where $\Phi'$ denotes the Weyl-transformed fields.
It follows in particular that, if $\Omega_{\mu\nu}[\sigma]=0$, the action is invariant even without gauging.
But the condition $\Omega_{\mu\nu}[\sigma]=0$ defines the conformal Weyl group in flat space,
as shown in eqs. (\ref{6a.3})-(\ref{ChIntHatSigma}); Ricci gauging is equivalent to the identity transformation,
as the Riemann tensor always vanishes in flat space \cite{Weinberg:1972gc}.
This proves that, \emph{if an action admits Ricci gauging, it is conformally invariant in flat space}. \\
All that remains to be proved is that conformal invariance in flat space is also sufficient for the action to admit Ricci gauging.
This result will be proved \emph{for actions which which contain only first derivatives of the conformally variant fields}.
Suppose that an action $\mathcal{S}_0$ is conformally invariant in flat space.
For infinitesimal conformal transformations we then have the conservation law
\begin{equation}
\delta \mathcal{S}_0 = \int d^dx \hat\sigma_{\mu}j^{\mu} = c_{\mu}\int j^{\mu}=0 \, ,
\label{VirialInt}
\end{equation}
where $j^\mu$ is the virial current \cite{Polchinski:1987dy,Callan:1970ze}, defined as
\begin{equation}
j^\mu = \pi_\nu \left( d_{\Phi} \delta^{\mu\nu} + 2\, \Sigma^{\mu\nu} \right) \Phi \, ,
\quad
\pi^\mu = \frac{\delta \mathcal{S}}{\delta (\partial_\mu\Phi)} \, ,
\label{VirialDefInt}
\end{equation}
where we remind that the $\Sigma$'s are the generators of the Lorentz group in the representation of the field $\Phi$
and $\hat\sigma = b_\nu\,x^\nu$.
Eq. (\ref{VirialInt}) holds for an arbitrary vector $b_\mu$, so it implies that $j^\mu$ is the divergence of a second rank tensor,
\begin{equation}
j^{\mu}= \partial_\nu J^{\mu\nu}\, .
\label{TotDivInt}
\end{equation}
As the actions we are considering contain only first derivatives of the conformally variant fields, the same must be true for $j^\mu$.
Therefore (\ref{TotDivInt}) is telling us that the tensor $J^{\mu\nu}$ depends only on the conformally variant fields
and not on their derivatives, for otherwise $j^\mu$ would contain higher derivatives of the same fields.
This implies that $j^\mu$ is at most linear in the derivatives of the conformally variant fields.
So, given the definition of the virial current (\ref{VirialDefInt}),
the action $\mathcal{S}_{0}$ must be at most quadratic in the same derivatives. \\
We have got the intermediate result that \emph{as far as we consider actions containing only first derivatives of the conformally variant fields,
conformal invariance is allowed only for those which are at most quadratic in such derivatives}. \\
In the case where the action is linear in the first derivatives the same argument implies that the virial current is identically zero.
This happens, for instance, for the Dirac fermion, which does vary under coformal transformations.
On the other hand, the argument does not impose any conditons on the conformally invariant fields. \\
Next we consider finite conformal transformations.
As our actions are at most quadratic in the first derivatives of the conformally variant fields
and there are no higher derivatives of any field whatsoever, their variation under finite conformal shifts is simply
\begin{equation}
\Delta \mathcal{S}_0 = \int d^dx\, \left(\hat\sigma_\mu j^\mu + \hat\sigma_\mu \hat\sigma_\nu K^{\mu\nu}\right)\, ,
\label{6.6}
\end{equation}
where $K^{\mu\nu}$ does not contain derivatives of the fields.
Using the relation (\ref{TotDivInt}) , we can integrate by parts and write (\ref{6.6}) as
\begin{equation}
\Delta \mathcal{S}_0 = \int d^dx\, \left( - J^{\mu\nu}\partial_\mu \hat \sigma_\nu +\hat\sigma_\mu\hat\sigma_\nu K^{\mu\nu}\right)\, .
\label{6.7}
\end{equation}
We can use (\ref{ChapIntDiffCondition}) to recast (\ref{6.7}) in the form
\begin{equation}
\Delta \mathcal{S}_0 = \int d^dx\, \hat\sigma_\mu \hat\sigma_\nu\,
\left(K^{\mu\nu}- J^{\mu\nu}+\frac{\delta^{\mu\nu}}{2}{J^\lambda}_\lambda \right)\, .
\label{6.8}
\end{equation}
The function $\hat\sigma$ is given in (\ref{ChIntHatSigma}):
it has a specific form but otherwise depends on an arbitrary four-vector $b_\mu$.
This is sufficient
to conclude that the integrand in (\ref{6.8}) must vanish identically, so that
\begin{equation}
K^{\mu\nu}= J^{\mu\nu}- \frac{\delta^{\mu\nu}}{2}J
\qquad \text{or} \qquad
J^{\mu\nu}= K^{\mu\nu}-\frac{\delta^{\mu\nu}}{n-2} K\, ,
\label{6.9}
\end{equation}
where we have denoted with $K$ and $J$ the traces.
Hence, invariance under finite conformal transformations implies that the virial tensor $J^{\mu\nu}$
is a specific linear function of the tensor $K^{\mu\nu}$ which appears in the quadratic expansion. \\
This allows to construct the Ricci-gauged action, for which we return to curved space,
where the results just derived can be written as
\begin{equation}
j^\mu = \nabla_\nu J^{\mu\nu}\, ,
\quad \text{with} \quad
J^{\mu\nu}= K^{\mu\nu}- \frac{g^{\mu\nu}}{n-2} T \, .
\label{8.9}
\end{equation}
From the discussion in the preceding section, we learn that the Weyl gauge field is introduces only to compensate the variation of the
derivatives of the fields which change under Weyl transformations, according to eq. (\ref{ChIntWeylSpinConnection}).
As the action is at most quadratic in the derivatives of the conformally variant fields, we can Weyl-gauge it by adding terms that are at most
quadratic in the Weyl field $W_\mu$, namely
\begin{equation}
\mathcal{S}= \mathcal{S}_0 + \int d^dx\, \sqrt{g}\, \left(W_\mu j^\mu + W_\mu W_\nu K^{\mu\nu}\right) \, .
\label{6.10}
\end{equation}
The form of the first term follows from eqs. (\ref{ChIntWeylSpinConnection}) and (\ref{VirialDefInt}),
while the tensor in the quadratic term is necessarily the same as in (\ref{6.6}), because derivative terms are the only ones that have
non trivial properties under both Weyl and conformal transformations.
By the same integration by part as above, this time in curved space, we find that
\begin{equation}
\mathcal{S} = \mathcal{S}_0 +
\int d^dx \, \sqrt{g}\, \left(- J^{\mu\nu}\, \nabla_\mu W_\nu + W_\mu W_\nu\, K^{\mu\nu}\right) \, ,
\label{6.11}
\end{equation}
which, using (\ref{6.9}), can be written as
\begin{equation}
\mathcal{S} =\mathcal{S}_0 - \int d^dx\, \sqrt{g}\, J^{\mu\nu}\, \Omega_{\mu\nu}[W] \, .
\label{6.12}
\end{equation}
This shows that, for theories which are conformally invariant in the flat limit, the supplementary terms brought by the Weyl-gauging
appear only in the form $\Omega_{\mu\nu}[W]$, which is exactly the condition for Ricci gauging.
Therefore we have proved that \emph{a necessary and sufficient condition for a scale invariant action $\mathcal{S}$ to allow
for Ricci gauging is that the flat-space limit of the ungauged action $\mathcal{S}_0$ is conformally invariant}.
The Ricci-gauging is achieved by (\ref{6.12}).
All the discussion so far is purely classical. At the quantum level, Weyl symmetry is violated by the trace anomaly,
which will be introduced in chapter \ref{TTTVertex}. \\
\clearpage{\pagestyle{empty}\cleardoublepage}
\chapter{The Three Graviton Vertex}\label{TTTVertex}
\section{Introduction}
In several recent works \cite{Giannotti:2008cv,Armillis:2009pq,Armillis:2010qk} certain correlation functions describing the
interaction between a gauge theory and gravity with massless fields in the internal loop and related therefore to the trace
anomalies in these theories have been analysed. The interesting property that such anomalous amplitudes contain massless poles in
2-particle intermediate states has been exposed in these investigations. In particular, this has been demonstrated in the $TVV$ amplitude
in QED, characterized by the insertion of the energy-momentum tensor ($T$) on 2-point functions of vector gauge currents ($V$).
As long as the gravitational field is kept as a classical external source, which will always be the case in our treatment,
this amplitude gives the $1$-loop contribution to the interaction between a gauge theory and gravity,
a part of which is mediated by the trace anomaly.
The complete evaluation of this amplitude in QCD and in the Standard Model \cite{Armillis:2010qk,Coriano:2011zk} confirms the
conclusion of \cite{Giannotti:2008cv}, namely the presence of an effective massless scalar, ``dilaton-like" degree of freedom in
intermediate 2-particle states, that is intimately connected with the trace anomaly, in the sense that the non-zero residue of the pole is
necessarily proportional to the coefficient of the anomaly. The perturbative results of
\cite{Giannotti:2008cv,Armillis:2009pq,Armillis:2010qk}
are also in agreement with the anomaly-induced gravitational effective action in 4 dimensions whose non-local form was found in
\cite{Riegert:1984kt}. It had been argued in \cite{Mottola:2006ew,Mottola:2010gp} that
the local covariant form of this anomalous effective action necessarily implies effective massless scalar degrees of freedom .
This is the 4-dimensional analogue of the anomaly-induced action in 2-dimensional CFT's coupled to a background metric generated
by the 2-dimensional trace anomaly and related to the central term in the infinite dimensional Virasoro algebra \cite{Polyakov:1981}.
The anomaly-induced scalar in the 2-dimensional case is the Liouville mode of
non-critical string theory on the 2-dimensional world sheet of the string.
In even dimensions greater than 2 it is important to recognize that the anomaly-induced effective action discussed in
\cite{Mottola:2006ew,Mazur:2001aa,Riegert:1984kt,Mottola:2010gp} is determined only up to Weyl invariant terms. The full quantum
effective action is not determined by the trace anomaly alone, and hence only when certain anomalous contributions to the $TVV$ or
other amplitudes are isolated from their non-anomalous parts should any comparison with the anomaly-induced effective action be made.
The non-anomalous components are dependent upon additional Weyl invariant terms in the quantum effective action,
corresponding to traceless parts of the Green functions of the theory.
While \cite{Giannotti:2008cv,Armillis:2009pq,Armillis:2010qk,Coriano:2011zk} are focused on the search for signatures
of massless scalar degrees of freedom in correlators describing the interaction of gauge fields with the background gravitational field,
exploiting the connection with Riegert's anomalous effective action, no such study has been attempted for the gravitational
field self-interactions, until recently. The simplest Green function accounting for such interaction, in the limit in which
the gravitational field is kept classical, is the $3$-graviton correlator, whose explicit evaluation is technically quite demanding. \\
In this chapter we present the first explicit perturbative $1$-loop computation of the three graviton vertex in $4$ dimensions.
The computation was performed in momentum space in a completely off-shell configuration,
but the remarkable complexity of the general result allows us to present here, in a compact form, only the expression with two of three gravitons
on the external lines in an on-shell configuration. \\
We discuss the derivation of general covariance and anomalous trace Ward identities for the correlator,
all of which were explicitly checked in order to secure the correctness of the result.
The computation of the necessary one loop tensor integrals with three denominators and up to rank $6$ that are necessary
for the evaluation of the correlator was performed with the Passarino-Veltman technique.
Though the original motivation for the explicit evaluation of the $\langle TTT \rangle$ Green function was the search of anomaly poles
in the geometric sector of the trace anomaly, we did not not attempt in this work to address the issue of their presence in the $TTT$ correlator.
Although this is an important motivation for initiating this study, the actual demonstration of the existence of such poles
requires a considerable additional effort, due to the extreme complexity of the result.
We expect to address this final point in a related work making use of the technical framework
and building upon the results of the present study.
\section{Conventions and the trace anomaly equation}
Before beginning the discussion of the $TTT$ correlator investigated in our work, we introduce our definitions and conventions.
We recall that the ordinary definition of the energy-momentum tensor (which we will address as EMT from now on)
in a classical theory described by an action $\mathcal S$, which is embedded in curved space, is
\begin{equation} \label{Ch1EMT}
T^{\mu\nu}(z) = -\frac{2}{\sqrt{g_z}}\frac{\delta \mathcal{S}}{\delta g_{\mu\nu}(z)} =
g^{\mu\alpha}(z)\,g^{\nu\beta}(z)\,\frac{2}{\sqrt{g_z}}\,\frac{\delta\mathcal S}{\delta g^{\alpha\beta}(z)} \, ,
\end{equation}
with $\textrm{det}\, g_{\mu\nu}(z)\equiv g_z$.
Now we introduce the generating functional of the theory in euclidean conventions, which we call $\mathcal W$,
\begin{equation}\label{Ch1Generating}
\mathcal W = \frac{1}{\mathcal{N}} \, \int \, \mathcal D\Phi \,
e^{- \mathcal S - \int d^4x\, \sqrt{g}\, A^a_\mu \,V^{a\,\mu} }\, ,
\end{equation}
where $\mathcal{N}$ is a normalization constant, $\Phi$ stands generally for all the quantum fields
of the theory and we have explicitly added the coupling of vector currents to background gauge fields $A^a_\mu$.
Given (\ref{Ch1EMT}), the vacuum expectation value (vev) of the EMT is given, in terms of $\mathcal W$, by
\begin{equation} \label{Ch1VEVEMT}
\left\langle T^{\mu\nu}(z) \right\rangle_s = \frac{2}{\sqrt{g_z}}\frac{\delta\, \mathcal W}{\delta\, g_{\mu\nu}(z)}\,,
\end{equation}
with the subscript $s$ meaning that the background fields are kept turned on.
From now on the dependence on coordinates will be dropped when not strictly necessary.
As for conformally invariant field theories, which will be the subject of most of this work, the trace of the EMT is zero at the classical
level, ${T^\mu}_\mu = 0$, one would naively expect this to be true also for the vev of the EMT,
\begin{equation}
g_{\mu\nu} \left\langle T^{\mu\nu} \right\rangle_s = 0\, .
\end{equation}
But this is known not to be true, as the quantum theory shows non vanishing terms in this trace.
In particular, when the matter system which is classically conformal invariant at the classical level is embedded in a background of
gauge fields $A^a_\mu$ and a gravitational field described by the metric tensor $g_{\mu\nu}$, it is found, in $4$ dimensions,
that the traced vev is given by \cite{Duff:1977ay,Duff:1993wm,Birrell:1982ix}
\begin{equation} \label{Ch1TraceAnomaly}
g_{\mu\nu}\left\langle T^{\mu\nu} \right\rangle_s \equiv \mathcal{A}[g, A] =
\sum_{I=f,s,V}n_I \, \bigg[ \beta_a(I)\, F + \beta_b(I)\, G + \beta_c(I)\,\Box R + \beta_d(I)\, R^2 \bigg]
- \frac{\kappa}{4} \, n_V \, F^{a\,\mu\nu}\, F^a_{\mu\nu} \, ,
\end{equation}
where $g$ and $A^a$ are short-hand notations respectively for the background metric and gauge field and
the coefficients $\beta_a$, $\beta_b$, $\beta_c$ and $\beta_d$ depend on the field content of the Lagrangian theory
(e.g. fermions, scalars, vector bosons) and we have a multiplicity factor $n_I$ for each particle species\footnote{Equivalent and more popular
notations are $c\equiv 16 \pi^2 \beta_a$ and $a\equiv -16 \pi^2 \beta_b$}. Actually the coefficient of $R^2$ must vanish identically
\begin{equation}
\beta_d \equiv 0
\end{equation}
since a non-zero $R^2$ term does not satisfy the Wess-Zumino consistency condition for conformal anomalies
\cite{Bonora:1983ff,AntMazMott:1992}. In addition, the value of $\beta_c$ is regularization dependent, corresponding to
the fact that it can be changed by the addition of an arbitrary local term in the effective action proportional
to the integral of $R^2$.
In particular, the values for $\beta_c$ reported in table \ref{Ch1AnomalyCoeff} hold in dimensional regularization,
for which one finds the constraint \cite{Duff:1977ay,Birrell:1982ix}
\begin{equation}\label{Ch1constraints}
\beta_c = -\frac{2}{3}\,\beta_a \, .
\end{equation}
Thus only $\beta_a$, $\beta_b$ and $\kappa$ correspond to true anomalies in the trace of the stress tensor.
For the purpose of this study, the gauge field sector of the trace anomaly is not concerned, so that from now on
we will assume that there is no background gauge field, implying that the last term on the r.h.s in eq. (\ref{Ch1TraceAnomaly}),
proportional to the squared field-strengths $F^a_{\mu\nu}$, is absent, so that
the trace anomaly functional depends only on the metric, $\mathcal A \equiv \mathcal{A}[g]$.
In table \ref{Ch1AnomalyCoeff} we list the values of the coefficients for the three conformal free field theories with
spin $0, \frac{1}{2}, 1$, in which the computations described in this work were performed.
\begin{table}
$$
\begin{array}{|c|c|c|c|}\hline
I & \beta_a(I)\times 2880\,\pi^2 & \beta_b(I)\times 2880\,\pi^2 & \beta_c(I)\times 2880\,\pi^2
\\ \hline\hline
S & \frac{3}{2} & -\frac{1}{2} & -1
\\ \hline
F & 9 & -\frac{11}{2} & -6
\\ \hline
V & 18 & -31 & -12
\\
\hline
\end{array}
$$
\caption{Anomaly coefficients for a conformally coupled scalar, a Dirac Fermion and a vector boson}
\label{Ch1AnomalyCoeff}
\end{table}
$\mathcal{A}[g]$ contains the diffeomorphism-invariants built out of the Riemann tensor,
${R^{\alpha}}_{\beta\gamma\delta}$, as well as the Ricci tensor $R_{\alpha\beta}$ and the scalar curvature $R$.
$G$ and $F$ in eq. (\ref{Ch1TraceAnomaly}) are the Euler density and the square of the Weyl tensor respectively.
All our conventions are listed in appendix \ref{Geometrical}.
Eq. (\ref{Ch1TraceAnomaly}) plays the role of a generating functional for the anomalous Ward identities of any underlying field theory.
These conditions are not necessarily linked to any Lagrangian, since the solution of these and of the other (non anomalous)
Ward identities - which typically constrain a given correlator - are based on generic requirements of conformal invariance.
Nevertheless, for our purposes, all these identities can be extracted from an ordinary generating functional, defined in terms
of a generic Lagrangian $\mathcal{L}$, which offers a convenient device to identify such relations.
Inserting these definitions in (\ref{Ch1TraceAnomaly}) and multiplying both sides by $\sqrt{g}$ we obtain
\begin{equation}
\label{Ch1TraceAnomalySymm}
2 \, g_{\mu\nu}\frac{\delta\, \mathcal{W}}{\delta \, g_{\mu\nu}}=
\sqrt{g} \, \mathcal A[g] \, .
\end{equation}
From (\ref{Ch1TraceAnomaly}) and (\ref{Ch1TraceAnomalySymm}) we can extract identities for the anomaly of
correlators involving $n$ insertions of energy-momentum tensors, just by taking $n$ functional derivatives of both sides
with respect to the metric of (\ref{Ch1TraceAnomalySymm}) and setting $g_{\mu\nu}=\delta_{\mu\nu}$ at the end.
For vertices with multiple insertions of gravitons, such as the $TTT$ vertex, which are really involved,
a successful test of the anomalous Ward identities is crucial in order to secure the correctness of the result of the perturbative computation.
\section{Definitions for the $TTT$ Amplitude}
\label{Ch1DiagTTT}
For a multi-graviton vertex, it is convenient to define the corresponding correlation function as the n-th
functional variation with respect to the metric of the generating functional $\mathcal W$ evaluated in the flat-space limit
\begin{eqnarray} \label{Ch1NPF}
\left\langle T^{\mu_1\nu_1}(x_1)...T^{\mu_n\nu_n}(x_n) \right\rangle
&=&
\bigg[\frac{2}{\sqrt{g_{x_1}}}...\frac{2}{\sqrt{g_{x_n}}} \,
\frac{\delta^n \mathcal{W}}{\delta g_{\mu_1\nu_1}(x_1)...\delta g_{\mu_n\nu_n}(x_n)}\bigg]
\bigg|_{g_{\mu\nu} = \delta_{\mu\nu}} \nonumber \\
&=&
2^n\, \frac{\delta^n \mathcal{W}}{\delta g_{\mu_1\nu_1}(x_1)...\delta g_{\mu_n\nu_n}(x_n)}\bigg|_{g_{\mu\nu} =
\delta_{\mu\nu}} \, ,
\end{eqnarray}
so that it is explicitly symmetric with respect to the exchange of any couple of metric tensors.
As we are going to deal with correlation functions evaluated in the flat-space limit all through the work,
we will omit to specify it from now on, so as to keep our notation easy.
The 3-point function we are interested in studying is found by evaluating (\ref{Ch1NPF}) for $n=3$,
\begin{eqnarray}\label{Ch13PF}
\left\langle T^{\mu\nu}(x_1)T^{\rho\sigma}(x_2)T^{\alpha\beta}(x_3)\right\rangle
&=&
8 \, \bigg[ - \, \left\langle \frac{\delta \mathcal S}{\delta g_{\mu\nu}(x_1)}\frac{\delta \mathcal S}{\delta g_{\rho\sigma}(x_2)}
\frac{\delta \mathcal S}{\delta g_{\alpha\beta}(x_3)}\right\rangle \nonumber \\
&& \hspace{-15mm}
+ \, \left\langle \frac{\delta^2 \mathcal S }{\delta g_{\alpha\beta}(x_3)\delta g_{\mu\nu}(x_1)}
\frac{\delta \mathcal S }{\delta g_{\rho\sigma}(x_2)} \right\rangle
+ \left\langle \frac{\delta^2 \mathcal S }{\delta g_{\rho\sigma}(x_2)\delta g_{\mu\nu}(x_1)}
\frac{\delta \mathcal S }{\delta g_{\alpha\beta}(x_3)} \right\rangle
\nonumber \\
&&\hspace{-15mm}
+ \, \left\langle \frac{\delta^2 \mathcal S}{\delta g_{\rho\sigma}(x_2)\delta g_{\alpha\beta}(x_3)}
\frac{\delta \mathcal S }{\delta g_{\mu\nu}(x_1)}\right\rangle
- \,\left\langle \frac{\delta^3 \mathcal S}{\delta g_{\rho\sigma}(x_2)\delta g_{\alpha\beta}(x_3)
\delta g_{\mu\nu}(x_1)}\right\rangle \bigg] \, .
\nonumber \\
\end{eqnarray}
The last term is identically zero in dimensional regularization, being proportional to a massless tadpole.
The Green function
\begin{equation}
\left\langle \frac{\delta\mathcal S}{\delta g_{\mu\nu}(x_1)}
\frac{\delta\mathcal S}{\delta g_{\rho\sigma}(x_2)}
\frac{\delta\mathcal S}{\delta g_{\alpha\beta}(x_3)}
\right\rangle \,
\label{Ch1Triangle}
\end{equation}
has the diagrammatic representation of a triangle topology, while the contributions
\begin{eqnarray} \label{Ch1bubbles}
\left\langle \frac{\d^2 \mathcal S}{\delta g_{\rho\sigma}(x_2)\delta g_{\alpha\beta}(x_3)}
\frac{\delta\mathcal S}{\delta g_{\mu\nu}(x_1)} \right\rangle \, ,
\quad
\left\langle \frac{\d^2 \mathcal S}{\delta g_{\alpha\beta}(x_3)\delta g_{\mu\nu}(x_1)}
\frac{\delta\mathcal S}{\delta g_{\rho\sigma}(x_2)} \right\rangle \, ,
\quad
\left\langle \frac{\d^2 \mathcal S}{\delta g_{\rho\sigma}(x_2)\delta g_{\mu\nu}(x_1)}
\frac{\delta\mathcal S}{\delta g_{\alpha\beta}(x_3)} \right\rangle
\end{eqnarray}
have the topology of 2-point functions and are traditionally named, in perturbative analysis, as ``bubbles''.
We decide to call them "$k$", "$q$" and "$p$" bubbles respectively, naming each one after the momentum
flowing into or out of the single graviton vertex.
The diagrammatic representation of the four contributions is show in fig. \ref{Ch1Fig.diagramsTTT}.
We convey to choose a dependence on the momenta such that $k$ is incoming at the point $x_1$
and $q$ and $p$ are outgoing at $x_2$ and $x_3$ respectively.
These conventions are summarized by the Fourier transform
\begin{equation}
\int \, d^4x_1\,d^4x_2\,d^4x_3\,
\left\langle T^{\mu\nu}(x_1)T^{\rho\sigma}(x_2)T^{\alpha\beta}(x_3)\right\rangle \,
e^{-i(k\cdot x_1 - q\cdot x_2 - p\cdot x_3)} =
(2\pi)^4\,\delta^{(4)}(k-p-q)\, \left\langle T^{\mu\nu}T^{\rho\sigma}T^{\alpha\beta}\right\rangle(q,p)\, .
\label{Ch13PFMom}
\end{equation}
Of course, for a $2$-point function we have
\begin{equation}\label{Ch12PFMom}
\int \, d^4x_2 \, d^4x_3 \, \left\langle T^{\rho\sigma}(x_2)T^{\alpha\beta}(x_3) \right\rangle\, e^{-i(q\cdot x_2 - p\cdot x_3)} =
(2\pi)^4\,\delta^{(4)}(p-q)\, \left\langle T^{\rho\sigma}T^{\alpha\beta} \right\rangle(p) \, .
\end{equation}
\begin{figure}[t]
\centering
\includegraphics[scale=0.8]{figures/triangle.eps}
\hspace{5mm}
\includegraphics[scale=0.8]{figures/pbubble.eps}
\hspace{5mm}
\includegraphics[scale=0.8]{figures/qbubble.eps}
\hspace{5mm}
\includegraphics[scale=0.8]{figures/kbubble.eps}
\hspace{5mm}
\caption{One loop expansion of the 3-graviton vertex in terms of the triangle and the self-energy-type contributions.
Shown here are the diagrams for the scalar particle. The expansion for the other two CFT's investigated can be obtained by replacing
the scalar by a fermion or a photon in the loops. In the former case one has to consider, for the triangle case, two inequivalent contributions,
distinguished by the direction of flow of the momentum flow of the fermion; for the latter, ghost corrections follow the same topologies.}
\label{Ch1Fig.diagramsTTT}
\end{figure}
It proves particularly useful to introduce a specific notation for the flat limit of functional derivatives with respect to the metric,
\begin{equation} \label{Ch1Flat}
\left[\mathcal F\right]^{\mu_{1}\nu_{1}\mu_{2}\nu_{2}\dots\mu_{n}\nu_{n}}(x_{1},x_{2},\dots,x_n) \equiv
\frac{\delta^n\, \mathcal F}{\delta g_{\mu_{1}\nu_{1}}(x_{1})\,\delta g_{\mu_{2}\nu_{2}}(x_{2})\dots\delta g_{\mu_n\nu_n}(x_{n})}
\bigg|_{g_{\mu\nu}=\delta_{\mu\nu}} \, ,
\end{equation}
for any functional (or function) $\mathcal F$ which depends on the background field $g_{\mu\nu}(x)$.
\subsection{General covariance Ward identities for the $TTT$}
\label{Ch1DiagTTTWardCov}
The requirement of general covariance for the generating functional $\mathcal W$
immediately leads to the master Ward identity for the conservation of the energy-momentum tensor.
Assuming that the integration measure is invariant under diffeomorphisms $\mathcal D \Phi' = \mathcal D \Phi'$,
standard manipulations lead to the master equation
\begin{equation} \label{Ch1masterWI0}
\nabla_\nu \left\langle T^{\mu\nu}(x_1) \right\rangle =
\nabla_\nu \bigg(\frac{2}{\sqrt{g_{x_1}}}\frac{\delta\mathcal W}{\delta g_{\mu\nu}(x_1)}\bigg) = 0 \, ,
\end{equation}
which, expanding the covariant derivative, can be written as
\begin{equation}
\frac{2}{\sqrt{g_{x_1}}}\bigg(\partial_{\nu}\frac{\delta\mathcal W}{\delta g_{\mu\nu}(x_1)}
- \Gamma^\lambda_{\lambda\nu}(x_1)\frac{\delta\mathcal W}{\delta g_{\mu\nu}(x_1)}
+ \Gamma^\mu_{\kappa\nu}(x_1)\frac{\delta\mathcal W}{\delta g_{\kappa\nu}(x_1)}
+ \Gamma^\nu_{\kappa\nu}(x_1)\frac{\delta\mathcal W}{\delta g_{\mu\kappa}(x_1)}\bigg) = 0 \, ,
\end{equation}
where the first of the three Christoffel symbols (for their definition see appendix \ref{Sign}) is generated by differentiation
of ${1}/{\sqrt{g_{x_1}}}$ in the definition of $T_{\mu\nu}(x_{1})$ in (\ref{Ch1masterWI0}) together with
\begin{equation}
\Gamma^\alpha_{\alpha\beta}= \frac{1}{2} \, g^{\alpha\gamma} \, \partial_\beta \, g_{\alpha\gamma} \,
\end{equation}
or, equivalently, as
\begin{equation}
2 \, \bigg(\partial_{\nu}\frac{\delta\mathcal W}{\delta g_{\mu\nu}(x_1)}
+ \Gamma^\mu_{\kappa\nu}(x_1)\frac{\delta\mathcal W}{\delta g_{\kappa\nu}(x_1)}\bigg) = 0\, .
\label{Ch1masterWI}
\end{equation}
By taking one and two functional derivatives of (\ref{Ch1masterWI}) with respect to $g_{\rho\sigma}(x_2)$ and
$g_{\rho\sigma}(x_2)$ and $g_{\alpha\beta}(x_3)$ respectively, one gets, in curved spacetime,
\begin{eqnarray}
&&
4 \, \bigg[ \partial_\nu \frac{\delta^2\mathcal W}{\delta g_{\rho\sigma}(x_2)\delta g_{\mu\nu}(x_1)}
+ \frac{\delta \Gamma^\mu_{\kappa\nu}(x_1)}{\delta g_{\rho\sigma}(x_2)}
\frac{\delta\mathcal W}{\delta g_{\kappa\nu}(x_1)}
+ \Gamma^\mu_{\kappa\nu}(x_1) \frac{\delta^2 \mathcal W}{\delta g_{\mu\nu}(x_1)\delta g_{\rho\sigma}(x_2)}\bigg]
= 0 \, , \label{Ch1WI2PFCoordinate} \\
&&
8 \bigg[\partial_\nu\frac{\delta^3\mathcal W}{\delta g_{\alpha\beta}(x_3)\delta g_{\rho\sigma}(x_2)\delta g_{\mu\nu}(x_1)}
+ \frac{\delta \Gamma^\mu_{\kappa\nu}(x_1)}{\delta g_{\rho\sigma}(x_2)}
\frac{\delta^2 \mathcal W}{\delta g_{\alpha\beta}(x_3)\delta g_{\kappa\nu}(x_1)}
+ \frac{\delta \Gamma^\mu_{\kappa\nu}(x_1)}{\delta g_{\alpha\beta}(x_3)}
\frac{\delta^2 \mathcal W}{\delta g_{\rho\sigma}(x_2)\delta g_{\kappa\nu}(x_3)}
\nonumber \\
&&
+ \frac{\delta^2\Gamma^\mu_{\kappa\nu}(x_1)}{\delta g_{\rho\sigma}(x_2)\delta g_{\alpha\beta}(x_3)}
\frac{\delta\mathcal W}{\delta g_{\mu\nu}(x_1)}
+ \Gamma^\mu_{\kappa\nu}(x_1)
\frac{\delta^3\mathcal W}{\delta g_{\rho\sigma}(x_2)\delta g_{\alpha\beta}(x_2)\delta g_{\kappa\nu}(x_1)}\bigg] = 0 \, .
\label{Ch1WI3PF}
\end{eqnarray}
As we are interested in the flat spacetime limit, we must evaluate (\ref{Ch1WI2PFCoordinate})
and (\ref{Ch1WI3PF}) by letting the Christoffel symbols go to zero.
Another simplification is obtained by noticing that the Green's functions
\begin{equation}
\left\langle \frac{\delta \mathcal S}{\delta g_{\mu\nu}(x_1)}\right\rangle = - \frac{\delta \mathcal W}{\delta g_{\mu\nu}(x_1)}
\end{equation}
and
\begin{equation}\label{Ch1Tadpole2PF}
\left\langle \frac{\delta^2 \mathcal S}{\delta g_{\mu\nu}(x_1)\delta g_{\alpha\beta}(x_3)}\right\rangle
\end{equation}
are proportional to massless tadpoles, so that we can ignore them in the expression of the $2$-point function,
\begin{equation}
\frac{\d^2 \mathcal W}{\delta g_{\alpha\beta}(x_3) \delta g_{\mu\nu}(z)}
= \left\langle \frac{\delta \mathcal S}{\delta g_{\mu\nu}(x_1)}
\frac{\delta \mathcal S}{\delta g_{\alpha\beta}(x_3)}\right\rangle
- \left\langle \frac{\d^2 \mathcal S}{\d g_{\alpha\beta}(x_3)\delta g_{\mu\nu}(x_1)} \right\rangle
= \left\langle \frac{\delta \mathcal S}{\delta g_{\mu\nu}(x_1)}
\frac{\delta \mathcal S}{\delta g_{\alpha\beta}(x_3)}\right\rangle \, .
\end{equation}
Thus the Ward identity for the $2$-point function in flat coordinates is immediately seen to be
\begin{equation}
\partial_{\nu} \left\langle T^{\mu\nu}(x_1) T^{\rho\sigma}(x_2) \right\rangle = 0 \, ,
\end{equation}
where, due to the vanishing of (\ref{Ch1Tadpole2PF}), we have set
\begin{equation} \label{Ch12PF}
\left\langle T^{\mu\nu}(x_1) T^{\rho\sigma}(x_2) \right\rangle
\equiv
4\, \left\langle \frac{\delta\mathcal S}{\delta g_{\mu\nu}(x_1)}
\frac{\delta\mathcal S}{\delta g_{\rho\sigma}(x_2)} \right\rangle \, .
\end{equation}
Obviously, its form in momentum space, exploiting the Fourier-transform (\ref{Ch12PFMom}), is
\begin{equation} \label{Ch1WI2PFMom}
p_{\mu}\left\langle T^{\mu\nu}T^{\rho\sigma} \right\rangle(p) = 0 \, .
\end{equation}
The terms surviving in (\ref{Ch1WI3PF}) are those in the first line.
In order to make them explicit, we evaluate the functional derivative of the Christoffel symbols using
the rules in appendix \ref{Sign}, namely (\ref{Ch1Christoffel}), (\ref{Ch1Tricks}) and (\ref{Ch1Tricks2}), finding
\begin{eqnarray}\label{Ch1GammaDerivatives1}
\frac{\delta\Gamma^\mu_{\kappa\nu}(x_1)}{\delta g_{\rho\sigma}(x_2)}
&=&
\frac{1}{2}\delta^{\mu\alpha}\bigg[-s^{\rho\sigma}_{\,\,\,\,\,\,\kappa\nu}\partial_\alpha
+ s^{\rho\sigma}_{\,\,\,\,\,\,\alpha\nu}\partial_\k + s^{\rho\sigma}_{\,\,\,\,\,\,\alpha\kappa} \partial_\nu\bigg]\delta^{4}(x_{12})\, ,
\end{eqnarray}
where we establish the convention $\delta^{4}(x_{12})\equiv \delta(x_1-x_2)$ and so on for the other couples of points.
Plugging this into (\ref{Ch1WI3PF}) and using (\ref{Ch12PF}), the second term becomes
\begin{eqnarray}
8 \, \frac{\delta\Gamma^\mu_{\k\nu}(x_1)}{\delta g_{\r\s}(x_2)}
\frac{\delta^2 \mathcal W}{\delta g_{\alpha\beta}(x_3)\delta g_{\kappa\nu}(x_1)}
&=&
\bigg[\delta^{\mu\r}\left\langle T^{\nu\sigma}(x_1)T^{\alpha\beta}(x_3)\right\rangle \, \partial_\nu +
\delta^{\mu\s}\left\langle T^{\nu\rho}(x_1)T^{\alpha\beta}(x_3)\right\rangle \, \partial_\nu
\nonumber\\
&& \hspace{5mm}
- \, \left\langle T^{\rho\sigma}(x_1)T^{\alpha\beta}(x_3)\right\rangle \, \partial^\mu\bigg]\delta^{4}(x_{12}) \, .
\end{eqnarray}
A completely analogous relation holds for the exchanged term
$\big(g_{\alpha\beta}(x_3)\leftrightarrow g_{\rho\sigma}(x_2)\big)$. \\
Finally, we can recast the Ward identity (\ref{Ch1WI3PF}) in the form
\begin{eqnarray} \label{Ch1WI3PFcoordinate}
\partial_\nu\left\langle T^{\mu\nu}(x_1)T^{\rho\sigma}(x_2)T^{\alpha\beta}(x_3) \right\rangle
&=&
\bigg[\left\langle T^{\rho\sigma}(x_1)T^{\alpha\beta}(x_3)\right\rangle\partial^\mu\delta^{4}(x_{12}) +
\left\langle T^{\alpha\beta}(x_1)T^{\rho\sigma}(x_2)\right\rangle\partial^\mu\delta^{4}(x_{13}) \bigg]\nonumber \\
&-&
\bigg[\delta^{\mu\rho}\left\langle T^{\nu\sigma}(x_1)T^{\alpha\beta}(x_3)\right\rangle
+ \delta^{\mu\sigma}\left\langle T^{\nu\rho}(x_1)T^{\alpha\beta}(x_3)\right\rangle\bigg]\partial_\nu\delta^{4}(x_{12})\nonumber\\
&-&
\bigg[\delta^{\mu\alpha}\left\langle T^{\nu\beta}(x_1)T^{\rho\sigma}(x_2)\right\rangle
+ \delta^{\mu\beta}\left\langle T^{\nu\alpha}(x_1)T^{\rho\sigma}(x_2)\right\rangle\bigg]\partial_\nu\delta^{4}(x_{13})\, ,\nonumber\\
\end{eqnarray}
having used the definitions (\ref{Ch1NPF}) and (\ref{Ch13PF}).\\
Fourier-transforming according to (\ref{Ch13PFMom}) and (\ref{Ch12PFMom}),
we get the Ward identity in momentum space that we need, i.e.
\begin{eqnarray}\label{Ch1WI3PFmomenta2a}
&& k_\nu \left\langle T^{\mu\nu}T^{\rho\sigma}T^{\alpha\beta} \right\rangle(q,p) =
p^\mu \left\langle T^{\rho\sigma}T^{\alpha\beta}\right\rangle(q)
+ q^\mu \left\langle T^{\rho\sigma}T^{\alpha\beta}\right\rangle(p) \nonumber \\
&&
- p_\nu \bigg[\delta^{\mu\beta} \left\langle T^{\nu\alpha}T^{\rho\sigma} \right\rangle(q)
+ \delta^{\mu\alpha}\left\langle T^{\nu\beta}T^{\rho\sigma} \right\rangle(q)\bigg]
- q_\nu \bigg[\delta^{\mu\sigma}\left\langle T^{\nu\rho}T^{\alpha\beta} \right\rangle(p)
+ \delta^{\mu\rho} \left\langle T^{\nu\sigma}T^{\alpha\beta}\right\rangle(p)\bigg].
\end{eqnarray}
Similar Ward identities can be obtained when we contract with the momenta of the other lines. These are going to be essential in
order to test the correctness of the computation once we turn to perturbation theory.
\subsection{The anomalous Ward identities for the TTT}
\label{Ch1DiagTTTWardAnom}
The anomalous Ward identities for the 3-graviton vertex is obtained performing two functional variations
of (\ref{Ch1TraceAnomalySymm}) and taking the flat-space limit, thereby obtaining
\begin{eqnarray}
\delta_{\mu\nu}\left\langle T^{\mu\nu}T^{\rho\sigma}T^{\alpha\beta} \right\rangle(q,p)
&=&
4 \, \left[\mathcal A[g]\right]^{\rho\sigma\alpha\beta}(q,p)
- 2 \, \left\langle T^{\rho\sigma} T^{\alpha\beta} \right\rangle(q) - 2 \, \left\langle T^{\rho\sigma}T^{\alpha\beta} \right\rangle(p)\nonumber\\
&=&
4 \, \bigg[ \beta_a\,\big(\big[F\big]^{\rho\sigma\alpha\beta}(q,p)
- \frac{2}{3} \big[\sqrt{g}\Box\,R\big]^{\rho\sigma\alpha\beta}(q,p)\big)
+ \beta_b\, \big[G\big]^{\rho\sigma\alpha\beta}(q,p) \bigg]\nonumber\\
&-&
2 \, \left\langle T^{\rho\sigma}T^{\alpha\beta} \right\rangle(q)
- 2 \, \left\langle T^{\rho\sigma}T^{\alpha\beta} \right\rangle(p) \, ,\label{Ch1munu3PFanomaly}
\end{eqnarray}
where $\left[ \dots \right]^{\alpha\beta\rho\sigma}(p,q)$ are contributions generated by functional derivatives of the anomaly,
according to the notation introduced in (\ref{Ch1Flat}).
We remark, if not obvious, that all the contractions with the metric tensor in the flat spacetime limit ($\delta_{\mu\nu}$)
should be understood as being 4-dimensional. This is the case for all the anomaly equations.
The various contributions to the trace anomaly are given in terms of
the functional derivatives of quadratic invariants in appendix \ref{Ch1Functionals}. Analogous anomalous
Ward identities can be obtained by tracing the other two pairs of indices.
\section{Three free field theory realizations of conformal symmetry}
\label{Ch1lags}
At this point we illustrate our perturbative computation in momentum space.
It was performed within the context of three free field theories, namely a conformally coupled (improven) scalar,
a Dirac fermion and the free Maxwell field.
The actions for the scalar and the fermion fields are respectively
\begin{eqnarray}\label{Ch1}
\mathcal{S}^s
&=&
\frac{1}{2} \, \int d^4 x \, \sqrt{g}\,
\bigg[g^{\mu\nu}\,\nabla_\mu\phi\,\nabla_\nu\phi - \chi\,R\,\phi^2 \bigg]\, ,\label{Ch1scalarAction}\\
\mathcal{S}^f
&=&
\frac{1}{2} \, \int d^4 x \, V \, {V_\alpha}^\rho\,
\bigg[\bar{\psi}\,\gamma^\alpha\,(\mathcal{D}_\rho\,\psi) - (\mathcal{D}_\rho\,\bar{\psi}) \, \gamma^\alpha\,\psi \bigg] \, .
\label{Ch1DiracAction}
\end{eqnarray}
Here $\chi$ is the parameter corresponding to the ``improvement term" one must add to the action of the free scalar field
so as to obtain a Weyl-invariant action; in particular, its value in $4$ dimensions has to be $\chi = 1/6$;
the symbol ${V_\alpha}^\rho$ is instead the Vielbein and $V(= \sqrt{g})$ its determinant,
needed to embed fermions in the curved background, with its covariant derivative $\mathcal{D}_\mu$ as
\begin{equation}
\mathcal{D}_\mu = \partial_\mu + \Gamma_\mu =
\partial_\mu + \frac{1}{2} \, \Sigma^{\alpha\beta} \, {V_\alpha}^\sigma \, \nabla_\mu\,V_{\beta\sigma} \, .
\end{equation}
The $\Sigma^{\alpha\beta}$ are the generators of the Lorentz group in the case of a spin $1/2$-field.
For more details on embedding classical fields into a curved background, we refer to \cite{Birrell:1982ix,Weinberg:1972gc}. \\
The action $\mathcal{S}_V$ for the photon field is given by the sum of three terms
\begin{equation}
\mathcal{S}_V = \mathcal{S}_M + \mathcal{S}_{gf} + \mathcal{S}_{gh}\, ,
\end{equation}
where the superscript $V$ stands for vector boson and the three contributions are the Maxwell action,
the gauge fixing contribution and the ghost action, that must be taken into account as well,
as gravity couples to any field with the same strength,
\begin{eqnarray}
\mathcal{S}_M &=& \frac{1}{4} \, \int d^4 x \, \sqrt{g} \, F^{\a\b} F_{\a\b}\, ,\\
\mathcal{S}_{gf} &=& \frac{1}{2 \xi} \, \int d^4 x \, \sqrt{g} \, \left( \nabla_{\alpha}A^\alpha \right)^2\, ,\\
\mathcal{S}_{gh} &=& - \int d^4 x \, \sqrt{g}\, \partial^\a \bar{c} \, \partial_\a c\, .
\end{eqnarray}
When performing formal manipulations with the \emph{Vielbein}, one has to correspondingly modify the definition of
the functional derivative with respect to the background metric, so that the EMT defined in (\ref{Ch1EMT})becomes,
in the fermion case,
\begin{equation} \label{Ch1FerTEI}
T^{\mu\nu} = - \frac{1}{V} \, V^{\alpha\mu} \, \frac{\delta\mathcal{S}}{\delta {V^\alpha}_\nu}\, .
\end{equation}
This tensor is not symmetric in general, but its antisymmetric part does not contribute to our calculations,
so that, for our purposes, we can adopt the symmetric definition
\begin{equation}\label{Ch1SymmFerTEI}
T^{\mu\nu}
\stackrel{def}{\equiv} - \frac{1}{2\,V}\bigg(
V^{\alpha\mu} \, \frac{\delta}{\delta {V^\alpha}_\nu} +
V^{\alpha\nu}\, \frac{\delta}{\delta {V^\alpha}_\mu}\bigg) \, \mathcal S \, .
\end{equation}
The EMT's for the scalar and the fermion are then
\begin{eqnarray}
T^{s\,\mu\nu}
&=&
\nabla^\mu \phi \, \nabla^\nu\phi - \frac{1}{2} \, g^{\mu\nu}\,g^{\alpha\beta}\,\nabla_\alpha \phi \, \nabla_\beta \phi
+ \chi \bigg[g^{\mu\nu} \Box - \nabla^\mu\,\nabla^\nu - \frac{1}{2}\,g^{\mu\nu}\,R + R^{\mu\nu} \bigg]\, \phi^2 \\
T^{f\,\mu\nu}
&=&
\frac{1}{4}\,
\bigg[ g^{\mu\rho}\,{V_\alpha}^\nu + g^{\nu\rho}\,{V_\alpha}^\mu - 2\,g^{\mu\nu}\,{V_\alpha}^\rho \bigg]
\bigg[\bar{\psi} \, \gamma^{\alpha} \, \left(\mathcal{D}_\rho \,\psi\right) -
\left(\mathcal{D}_\rho \, \bar{\psi}\right) \, \gamma^{\alpha} \, \psi \bigg],
\end{eqnarray}
while the EMT for the photon field is given by the sum of three terms
\begin{equation}
T_V^{\mu\nu} = T^{\mu\nu}_M + T^{\mu\nu}_{gf} + T^{\mu\nu}_{gh}\, ,
\end{equation}
with
\begin{eqnarray}
T^{\mu\nu}_M
&=&
F^{\mu\a}{F^\nu}_{\a} - \frac{1}{4}g^{\mu\nu}F^{\a\b}F_{\a\b} \, ,
\label{Ch1TEI M}
\\
T^{\mu\nu}_{gf}
&=&
\frac{1}{\xi}\left\{ A^\mu\nabla^\nu(\nabla_\rho A^\rho) + A^\nu\nabla^\mu(\nabla_\rho A^\rho ) -g^{\mu\nu}\,
\left[ A^\rho \nabla_\rho(\nabla_\sigma A^\sigma) + \frac{1}{2}(\nabla_\rho A^\rho)^2 \right]\right\}\, ,
\label{Ch1TEI g.f.}
\\
T^{\mu\nu}_{gh}
&=&
- \left(\partial^\mu\bar{c}\, \partial^\nu c + \partial^\nu\bar{c}\, \partial^\mu c - g^{\mu\nu}\partial^{\r}\bar{c}\,\partial_{\r}c \right)
\label{Ch1TEI gh}\, .
\end{eqnarray}
Now, to perform the explicit one loop computation, one must write down the integrals corresponding to the Feynman diagrams drawn in
fig. \ref{Ch1Fig.diagramsTTT}. Since the vev's of the third order derivatives correspond to massless tadpoles,
which can be consistently set to zero in dimensional regularization,
in order to write down the perturbative expansion vertices with no more than to gravitons are needed;
these require in turn to perform two variations of the action with respect to the background metric.
These results, together with the euclidean propagators, are listed in appendix \ref{Ch1Vertices}.
Once one has found them, tensor integrals must be evaluated. As mentioned in the introduction, they have been computed
by implementing in a symbolic manipulation program the Passarino-Veltman reduction technique.
Given the complexity of the result and to avoid any error, we have checked that all the expressions obtained
satisfy the corresponding Ward identities derived above. \\
Finally, we must remark that, as the $3$-graviton correlator does not have any gauge field on the external lines, one would expect it
not to depend on the gauge-fixing procedure, which enters only in the virtualities running in the loop.
This is known to happen in the case of the gauge-field $2$-point function and we explicitly checked that it is the case also for our
Green function, as expected. In other words, the computation of the correlator with the vertices listed in appendix \ref{Ch1Vertices}
is completely equivalent to the same computation performed by sending $1/\xi \rightarrow 0$ and omitting the diagrams with ghost loops.
This is another non trivial check of our computation in the gauge boson sector. \\
We end this section by providing the expressions of the $2$-point functions, which are necessary to test our computations.
Their general structure before renormalization is
\begin{eqnarray}
\left\langle T^{\mu\nu} \, T^{\alpha\beta} \right\rangle(p)
&=&
C_1(p)\, \bigg[ \frac{1}{2} \, \bigg( \Theta^{\mu\alpha}(p) \, \Theta^{\nu\beta}(p) +
\Theta^{\mu\beta}(p) \, \Theta^{\nu\alpha}(p)\bigg)
- \frac{1}{3} \, \Theta^{\mu\nu}(p) \, \Theta^{\alpha\beta}(p)\bigg]
+ \frac{C_2}{3}\,\Theta^{\mu\nu}(p) \, \Theta^{\alpha\beta}(p) \, , \nonumber \\
\end{eqnarray}
where the transverse tensor $\Theta$ is
\begin{equation} \label{DefineTheta}
\Theta^{\alpha\beta}(p) = \delta^{\alpha\beta}(p)\, p^2 - p^\alpha p^\beta \, .
\end{equation}
The values of the form factors for the three free field theories at hand are respectively given by
\begin{eqnarray}
C^s_1(p)
&=&
\frac{16 + 15 \, \mathcal{B}_0(p^2)}{14400\,\pi^2}\, ,
\qquad \hspace{4mm}
C^s_2 = - \frac{1}{1440\,\pi^2} \, , \\
C^f_1(p)
&=&
\frac{2 + 5 \,\mathcal{B}_0(p^2)}{800\,\pi^2} \, ,
\qquad \hspace{6mm}
C^f_2 = - \frac{1}{240\,\pi^2}\, , \\
C^V_1(p)
&=&
\frac{-11 + 10\, \mathcal{B}_0(p^2)}{800\,\pi^2} \, ,
\qquad
C^V_2 = - \frac{1}{120\,\pi^2}\, .
\end{eqnarray}
Here, $\mathcal{B}_0(p^2)$ is the $2$-point scalar integral, the only one that can appear in a massless $2$-point correlator
in the flat limit, as no scale is present and tadpoles can be consistently set to zero.
Its expression is
\begin{equation}
\mathcal{B}_0(p^2) =
\frac {1}{\pi^2} \int d^dl\, \frac{1}{l^2\,(l + p)^2} =
\frac{2}{\epsilon} - \gamma + \log \pi + 2 + \ln\left(\frac{\mu^2}{p^2}\right) \, ,
\label{BareB0}
\end{equation}
with $\gamma$ being the Euler constant. \\
The Ward identities discussed so far were tested before performing renormalization, but of course they hold even after the subtraction
of the ultraviolet singularities is performed.
For diffeomorphism invariance Ward identities this is easily understood, as the master equation
from which these constraints descends, (\ref{Ch1masterWI0}), can be very easily derived requiring the bare
generating functional to be invariant under general coordinate transformations.
Every correlator in such a Ward identity, for instance (\ref{Ch1WI3PFmomenta2a}), can be split into a finite part and an infinite one,
with the latter featuring a $1/\epsilon$ pole, so that the coefficients of $1/\epsilon$ and the finite (i.e. renormalized)
contributions on both sides can be separately equated.
The situation is rather different for trace Ward identities, because the master equation from which they descend,
(\ref{Ch1TraceAnomaly}), does not hold for the bare generating functional, but only for the renormalized one.
A thorough, detailed discussion of the last point, using the method of $\zeta$-function regularization and holding also for the
more general case of non conformal field theories, can be found in \cite{Birrell:1982ix}.
A more direct approach for CFT's, due to Duff \cite{Duff:1977ay} and enlightening
the relation between anomalies and counterterms, will be reviewed in chapter \ref{Mapping}.
\section{Renormalization of the $TTT$}
\label{Ch1Renormalization}
In this section we address the problem of the renormalization of the $3$-graviton vertex.
In \cite{Freedman:1974gs,Freedman:1974ze} it was shown that for scalar and gauge field theories,
with and without spontaneous symmetry breaking, the counterterms for the theory in flat space
are sufficient to renormalize Green functions with one insertion of the EMT and an arbitrary number of matter fields.
But for correlators including more than one insertion of the EMT, this is known not to be true.
In fact, the theories we are dealing with are not renormalizable, i.e. there are not enough constants in the Lagrangians
which could be split into a finite and an infinite part so as to produce the contributions needed to subtract infinities from
correlators involving multiple insertions of the EMT.
Actually, power-counting arguments can immediately show that the number of divergent correlators is infinite
when gravitational interactions are present, as remarked in the Introduction.
It is then necessary to introduce counterterms by hand.
For $d=4$ with dimensional regularization and in the $\overline{MS}$ renormalization scheme, it is known that the contribution
that must be added to the generating functional in order to remove $1$-loop divergences is \cite{Duff:1977ay,Duff:1993wm}
\begin{equation}\label{Ch1CounterAction}
S_{counter} =
- \frac{\mu^{-\epsilon}}{\bar\epsilon}\sum_{I=f,s,V}n_I \int d^d x \sqrt{g} \bigg( \beta_a(I) F + \beta_b(I) G\bigg) \, ,
\qquad \frac{2}{\bar\epsilon} = \frac{2}{\epsilon} - \gamma - \log \pi \, .
\end{equation}
Notice that one of the two objects that appear in the counterterm integral, $G$, is a total divergence
in 4 but not in $d$ dimensions. In particular, $G$ generates a counterterm which is effectively a projector on the extra $(4-d)$-dimensional
space and, as such, gives a contribution which needs to be included in order to perform a correct renormalization of the vertex,
in contrast with assertions that can be found in much of the literature on the subject.
For instance, in \cite{Deser:1993yx} the famous distinction between $A$ and $B$-type anomalies
for general (even) dimension was established, where the $B$-type are defined as the anomalies to which conformal invariants built out of
the Weyl tensor contribute, whereas $A$-type ones are those proportional to the Euler density contribution.
In \cite{Deser:1993yx} it was also claimed that the $A$-type anomalies are not associated with poles in $1/\epsilon$.
Such a result cannot be tested at the level of the $2$-point function, as the Euler anomaly does not contribute to it.
To the extent of our knowledge, the fact tha this point is not true was first pointed out in \cite{Osborn:1993cr} and has been confirmed
by our explicit computation in dimensional regularization in our approach.
Another way to show the presence of $1/\epsilon$ poles associated to type-$A$ anomalies relies on cohomological arguments,
which are used to build the Wess-Zumino effective action for conformal anomalies in dimensional regularization \cite{Mazur:2001aa}.
This approach is a starting point for the results presented in chapter \ref{Recursive}, where it is briefly reviewed.
The intimate relation between anomalies and counterterms will be further explored in chapter \ref{Mapping}.
We have used the 4-dimensional realization of $F$
\begin{equation}\label{Ch1RecallF}
F =
R^{\alpha\beta\gamma\delta}R_{\alpha\beta\gamma\delta} - 2\,R^{\alpha\beta}R_{\alpha\beta} + \frac{1}{3} \, R^2 \, .
\end{equation}
As remarked, $G$ does not contribute to every correlator. For instance, in the case of the $TT$, the counterterm is obtained by
functional differentiation twice of $S_{counter}$, but one can easily check (see eq. (\ref{Ch1Magic})) that the second variation of
$G$ vanishes in the flat limit. Hence, the only counterterm is given by
\begin{equation}
\label{Ch12PFCounterterm}
D_F^{\alpha\beta\rho\sigma}(x_1,x_2) =
4 \, \frac{\delta^2}{\delta g_{\alpha\beta}(x_1)\delta g_{\rho\sigma}(x_2)} \, \int\,d^d w\,\sqrt{g} \, F
\bigg|_{g_{\mu\nu}= \delta_{\mu\nu}} \, ,
\end{equation}
whose form in momentum space is
\begin{eqnarray}
D_F^{\alpha\beta\rho\sigma}(p)
&=&
4 \, \bigg[ \frac{1}{2}\,
\left( \Theta^{\mu\alpha}(p)\, \Theta^{\nu\beta}(p) + \Theta^{\mu\beta}(p)\, \Theta^{\nu\alpha}(p) \right)
- \frac{1}{3}\, \Theta^{\mu\nu}(p)\, \Theta^{\alpha\beta}(p)\bigg] \, ,
\label{Ch12PFCountertermMom}
\end{eqnarray}
where $\Theta$ was defined in (\ref{DefineTheta}).
We obtain the renormalized $2$-point function by adding it according to (\ref{Ch1CounterAction}), i.e.
\begin{equation}\label{Ch1Ren2PF2}
\left\langle T^{\alpha\beta}\,T^{\rho\sigma} \right\rangle_{ren}(p) =
\left\langle T^{\alpha\beta}\,T^{\rho\sigma} \right\rangle(p) - \frac{\beta_a}{\,{\bar\epsilon}} \, D_F^{\alpha\beta\rho\sigma}(p)\, .
\end{equation}
In the case of the $3$-graviton vertex the counterterm action (\ref{Ch1CounterAction}) generates the vertices
\begin{equation}
-\frac{\mu^{-\epsilon}}{\bar\epsilon}\, \bigg(\beta_a D_F^{\mu\nu\rho\sigma\alpha\beta}(x_1,x_2,x_3)
+ \beta_b \, D_G^{\mu\nu\rho\sigma\alpha\beta}(x_1,x_2,x_3)\bigg)\, ,
\end{equation}
where
\begin{eqnarray}
D_F^{\mu\nu\rho\sigma\alpha\beta}(x_1,x_2,x_3)
&=&
8 \, \frac{\delta^3}{\delta g_{\mu\nu}(x_1) \delta g_{\rho\sigma}(x_2)\delta g_{\alpha\beta}(x_3)}
\int\,d^d w\,\sqrt{g}\, F\bigg|_{g_{\mu\nu}=\delta_{\mu\nu}} \, ,
\label{Ch1DF}\\
D_G^{\mu\nu\rho\sigma\alpha\beta}(z,x,y)
&=&
8 \, \frac{\delta^3}{\delta g_{\mu\nu}(x_1) \delta g_{\rho\sigma}(x_2)\delta g_{\alpha\beta}(x_3)}
\int\,d^d w\,\sqrt{g}\, G\bigg|_{g_{\mu\nu}=\delta_{\mu\nu}} \, .
\label{Ch1DG}
\end{eqnarray}
The explicit form of (\ref{Ch1DF}) and (\ref{Ch1DG}) is derived by functionally deriving three times the general functional
\begin{equation}
\mathcal{I}(a,b,c) \equiv \int\,d^4 x\,\sqrt{g}\,
\big(a\,R^{abcd}R_{abcd} + b\,R^{ab}R_{ab} + c\, R^2 \big)\, ,
\end{equation}
with respect to the metric for appropriate $a, b$ and $c$, i.e.
\begin{eqnarray}
&&
a = 1 \, ,\quad b = -2 \, ,\quad c = \frac{1}{3} \, , \nonumber \\
&&
a = 1 \, ,\quad b = -4 \, ,\quad c = 1 \, . \nonumber
\end{eqnarray}
For convenience, the computations leading to the general result are reproduced in appendix \ref{Ch1FunctionalIntegral}.
The renormalized correlator is represented in fig. \ref{Ch1Fig.diagrams3grav}.
It goes without saying that the counterterms for the $3$-point function
in momentum space are given by the transform (\ref{Ch13PFMom}). \\
\begin{figure}[t]
\centering
\includegraphics[scale=0.7]{figures/CT1.eps}
\hspace{5mm}
\caption{$TTT$ and its counterterms generated with the choice of the square of the Weyl ($F$)
tensor in 4 dimensions and the Euler density ($G$).}
\label{Ch1Fig.diagrams3grav}
\end{figure}
It is known that $D_G^{\mu\nu\rho\sigma\alpha\beta}(q,p)$ is found to vanish identically in four dimensions.
In fact, its explicit form is
\begin{equation}\label{Ch1ExplicitDG}
D_G^{\mu\nu\rho\sigma\alpha\beta}(q,p) =
-240 \big(E^{\mu\sigma\alpha\gamma\kappa,\nu\rho\beta\delta\lambda} +
E^{\mu\rho\alpha\gamma\kappa,\nu\sigma\beta\delta\lambda} + \alpha\leftrightarrow \beta\big)
\,q_\gamma\, q_\delta\, p_\kappa\, p_\lambda\, ,
\end{equation}
where $E^{\mu\sigma\alpha\gamma\kappa,\nu\rho\beta\delta\lambda}$ is a projector onto
completely antisymmetric tensors with five indices, so that it would yield zero in four dimensions,
reflecting the fact that the integral of the Euler density is a topological invariant in integer dimensions.
We have explicitly checked that, given the structure of the counterterm Lagrangian in (\ref{Ch1CounterAction}), one necessarily needs
to include the contribution from the $G$ part of the functional, in the form given by $D_G$, in order to remove all
the divergences.
The fully renormalized 3-point correlator in momentum space can be written down as
\begin{equation}\label{Ch1Ren3PF}
\left\langle T^{\mu\nu}T^{\rho\sigma}T^{\alpha\beta} \right\rangle_{ren}(q,p) =
\left\langle T^{\mu\nu}T^{\rho\sigma}T^{\alpha\beta} \right\rangle_{bare}(q,p) -
\frac{\mu^{-\epsilon}}{{\bar\epsilon}}\,\bigg(\beta_a\,D_F^{\mu\nu\rho\sigma\alpha\beta}(q,p)
+ \beta_b\,D_G^{\mu\nu\rho\sigma\alpha\beta}(q,p)\bigg)\,
\end{equation}
and the goal is to proceed with an identification both of $D_F$ and $D_G$ from the
diagrammatic expansion in momentum space.
The cancellation of all of the ultraviolet poles, for suitable expressions of $D_F$ and $D_G$, has been thoroughly checked
from our explicit results.
At this point, it is necessary to comment about the difference between our approach and that followed in \cite{Osborn:1993cr},
where the choice of $F$ is slightly different from ours, since the authors essentially define
a counterterm which is given by an integral of the form
\begin{equation}
\mathcal{\tilde{S}}_{counter} =
- \frac{\mu^{-\epsilon}}{\bar\epsilon} \int d^4x \sqrt{g} \, \big( \beta_a \, F^{d} + \beta_b \, G \big) \, ,
\label{Ch1newren}
\end{equation}
based on the $d$-dimensional expression of the squared of the Weyl tensor ($F_d$).
Such a choice does not generate the anomaly contribution proportional to $\square R$ (sometimes refferred to as ``local anomaly'').
In fact, the authors choose to work with $\beta_c=0$ from the very beginning, since the inclusion of the local anomaly contribution amounts just
to a finite renormalization with respect to (\ref{Ch1newren}).
We briefly comment on the connection between the choice of the counterterm in (\ref{Ch1newren}) and the quoted finite
renormalization, discussed for the first time in \cite{Capper:1974ic}.
Notice that in $d$ dimensions, if we take the trace of the functional derivative in (\ref{Ch1Magic})
for $a=1$, $b = -{4}/({d-2})$, $c = {2}/((d-1)(d-2))$, which are the $d$-dimensional coefficients appearing in $F_d$,
we can explicitly check that the contribution proportional to $\square R$ in the anomalous trace cancels.
For this purpose we can expand the integrand of (\ref{Ch1newren}) around $d=4$ (in $\epsilon = 4 - d$) up to $O(\epsilon)$,
obtaining that the counterterm action can be separated in a polar plus a finite part, i.e.
\begin{equation}
\mathcal{\tilde{S}}_{counter} =
\mathcal S_{counter} + \mathcal S_{fin.\,ren.} = \mathcal S_{counter}
+ \beta_a \, \int d^4x\, \sqrt{g} \, \bigg(R^{\alpha\beta}\,R_{\alpha\beta} - \frac{5}{18} \, R^2\bigg) + O(\epsilon)\, .
\end{equation}
Recalling the definition (\ref{Ch1VEVEMT}) and using (\ref{Ch1Magic}),
we see that the contribution of this finite part to the vev of the EMT is
\begin{equation}
g_{\mu\nu} \left\langle \, T^{\mu\nu} \, \right\rangle_{fin.ren.} = -\beta_c \, \square R \, .
\end{equation}
\begin{figure}
\begin{center}
\includegraphics[scale=0.7]{figures/CT2.eps}
\hspace{5mm}
\caption{The contributions to the renormalized $TTT$ vertex from the square of the Weyl tensor in $d$-dimensions ($F_dd$)
and the Euler density ($G$).}
\label{Ch1Fig.diagrams3gravd}
\includegraphics[scale=0.7]{figures/CT3.eps}
\hspace{5mm}
\caption{The relation between the counterterm generated by $F_d$ and the same obtained from $F$. The difference is a finite
renormalization ($F_{fin}$) term in the counterterm Lagrangian, which generates the local contribution to the trace anomaly.}
\end{center}
\end{figure}
Comparing this with (\ref{Ch1TraceAnomaly}), we see that this extra contribution will cancel the local anomaly. \\
It is immediately realized that this approach is equivalent, for what concerns the anomaly, to supplying the action of the theory with
the finite renormalization usually met in the literature, i.e.
\begin{equation}
\mathcal S^{(2)}_{fin.\,ren.} \equiv - \frac{\beta_c}{12} \, \int d^4x\, \sqrt{g} \, R^2\, ,
\end{equation}
which is known to cancel the local anomaly, due to the similar relation
\begin{equation}
g_{\mu\nu} \, \frac{2}{\sqrt{g}} \, \frac{\delta \mathcal S^{(2)}_{fin.\,ren.}}{\delta g_{\mu\nu}} =
- \beta_c \,\square R\, ,
\end{equation}
which can be checked using (\ref{Ch1Magic}) once again .
\section{The renormalized 3-graviton vertex with two lines on-shell}
\label{Ch1ExplicitTTT}
In general, a rank-$6$ tensor depending on $2$ momenta can be expanded on a basis made up of $499$ tensors
built out of the $2$ momenta and the metric tensor $\delta_{\mu\nu}$.
The problem is that the scalar coefficients multiplying such tensors are really complicated, in the most general kinematic
configuration and it is no point reporting them explicitly.
Nevertheless, we have found that, in all of the three cases explicitly examined, if we go on-shell on the two outgoing gravitons,
the $TTT$ vertex can be expanded on a basis made up of just thirteen tensors.
This amounts to contracting the amplitude with polarization tensors $e^s_{\lambda\kappa}(p)$ which are transverse and traceless
\begin{equation}\label{Ch1helicity}
p^\lambda\,e^s_{\lambda\kappa}(p) = 0 \, , \quad {(e^s)^\lambda}_\lambda(p) = 0 \, ,
\end{equation}
where the superscript denotes the helicity state of the graviton carrying the momentum $p$.
Given the assignment of the momenta established in (\ref{Ch13PFMom}), it is immediate
to see that the contraction of the amplitude with the polarization tensors with the properties (\ref{Ch1helicity})
for the two outgoing gravitons is equivalent to the replacements
\begin{equation}
p^2 \rightarrow 0 \, , \quad q^2 \rightarrow 0 \, , \quad
p^\alpha \rightarrow 0 \, , \quad p^\beta \rightarrow 0 \, , \quad
q^\rho \rightarrow 0 \, , \quad q^\sigma \rightarrow 0 \, ,
\label{Ch1OnShell}
\end{equation}
so that these conditions can be used to select the tensors that are non-vanishing after the on-shell limit is taken.
We are going to express our result for the amplitude in terms of such tensors.
The expansion of our Green's function for a theory with $n_S$ scalars, $n_F$ fermions and $n_V$ vector bosons can be generally written as
\begin{equation}
\left\langle T^{\mu\nu}T^{\rho\sigma}T^{\alpha\beta} \right\rangle(q,p)\bigg|_{On-Shell} \equiv
\sum_{n_I=n_S,n_F,n_V} n_I \, \sum_{i=1}^{13} \Omega^I_i(s) \, t_i^{\mu\nu\rho\sigma\alpha\beta}(q,p)\, ,
\quad s \equiv k^2 = 2 \, p \cdot q \, .
\end{equation}
The form factors for the three theories that we consider in this chapter are listed in table \ref{Ch1FormFactors}.
\begin{table}
$$
\begin{array}{|c|c|c|c|}
\hline
i
& \Omega^S_{i}(s)
& \Omega^F_{i}(s)
& \Omega^V_{i}(s)
\\
\hline
\hline
1
& - \frac{1}{720\,\pi^2}\times \frac{1}{2 \, s}
& - \frac{1}{240\,\pi^2}\times \frac{1}{s}
& \frac{1}{1152\,\pi^2}\times \frac{72}{5 \, s} \\
\hline
2
& - \frac{1}{720\,\pi^2} \times \frac{1}{s}
& - \frac{1}{240\,\pi^2} \times \frac{1}{3 \, s}
& \frac{1}{1152\,\pi^2} \times \frac{64}{5 \, s}
\\
\hline
3
& - \frac{1}{720\,\pi^2}\times \frac{7 + 30 \, \mathcal{B}_0(s)}{120}
& \frac{1}{240\,\pi^2}\times \frac{13 - 30 \, \mathcal{B}_0(s)}{60}
& \frac{1}{1152\,\pi^2}\times \frac{82 - 120 \, \mathcal{B}_0(s)}{25}
\\
\hline
4
& - \frac{1}{720\,\pi^2}\times \frac{2 + 5 \, \mathcal{B}_0(s)}{10}
& \frac{1}{240\,\pi^2}\times \frac{7 - 70 \, \mathcal{B}_0(s)}{120}
& \frac{1}{1152\,\pi^2}\times \frac{2 \, (482 + 130 \, \mathcal{B}_0(s))}{25}
\\
\hline
5
& \frac{1}{720\,\pi^2}\times \frac{1}{6}
& - \frac{1}{240\,\pi^2}\times \frac{-1 + 10 \, \mathcal{B}_0(s)}{48}
& - \frac{1}{1152\,\pi^2}\times \frac{79 + 50 \, \mathcal{B}_0(s)}{5}
\\
\hline
6
& \frac{1}{720\,\pi^2}\times \frac{23 + 20 \, \mathcal{B}_0(s)}{20}
& \frac{1}{240\,\pi^2}\times \frac{33 + 70 \, \mathcal{B}_0(s)}{60}
& - \frac{1}{1152\,\pi^2}\times \frac{104 \, (22 + 5 \, \mathcal{B}_0(s))}{25}
\\
\hline
7
& - \frac{1}{720\,\pi^2}\times \frac{s\,(16 + 15 \, \mathcal{B}_0(s))}{20}
& - \frac{1}{240\,\pi^2}\times \frac{3\,s\,(2 + 5 \, \mathcal{B}_0(s))}{10}
& - \frac{1}{1152\,\pi^2}\times \frac{s\,(-11 + 10 \, \mathcal{B}_0(s))}{80}
\\
\hline
8
& - \frac{1}{720\,\pi^2}\times \frac{s\,(47 + 30 \, \mathcal{B}_0(s))}{80}
& - \frac{1}{240\,\pi^2}\times \frac{3\, s\,(9 + 10 \, \mathcal{B}_0(s))}{40}
& \frac{1}{1152\,\pi^2}\times \frac{s\,(2 + 5 \, \mathcal{B}_0(s))}{40}
\\
\hline
9
& \frac{1}{720\,\pi^2}\times \frac{s\,(2 + 5 \, \mathcal{B}_0(s))}{40}
& - \frac{1}{240\,\pi^2}\times \frac{7 s\,(1 - 10 \, \mathcal{B}_0(s))}{480}
& - \frac{1}{1152\,\pi^2}\times \frac{s\,(487 + 130 \, \mathcal{B}_0(s))}{50}
\\
\hline
10
& \frac{1}{720\,\pi^2}\times \frac{s \, (9 + 10 \, \mathcal{B}_0(s))}{20}
& \frac{1}{240\,\pi^2}\times \frac{s \, (137 + 430\, \mathcal{B}_0(s))}{480}
& - \frac{1}{1152\,\pi^2}\times \frac{s \, (883 - 230 \, \mathcal{B}_0(s))}{50}
\\
\hline
11
& - \frac{1}{720\,\pi^2}\times \frac{s \, (7 + 5 \, \mathcal{B}_0(s))}{20}
& - \frac{1}{240\,\pi^2}\times \frac{7 \, s \, (9 + 10 \, \mathcal{B}_0(s))}{240}
& \frac{1}{1152\,\pi^2}\times \frac{s \, (467 + 130 \, \mathcal{B}_0(s))}{25}
\\
\hline
12
& - \frac{1}{720\,\pi^2}\times \frac{s \, (121 + 90 \, \mathcal{B}_0(s))}{240}
& - \frac{1}{240\,\pi^2}\times \frac{s \, (97 + 130 \, \mathcal{B}_0(s))}{240}
& \frac{1}{1152\,\pi^2}\times \frac{2 \, s \, (299 + 35 \, \mathcal{B}_0(s))}{25}
\\
\hline
13
& \frac{1}{720\,\pi^2}\times \frac{5 \, s^2 \, (3 + 2 \, \mathcal{B}_0(s))}{32}
& \frac{1}{240\,\pi^2}\times \frac{5 \, s^2 \, (9 + 10 \, \mathcal{B}_0(s))}{96}
& -\frac{s^2 \, (13 - \mathcal{B}_0(s))}{1152\,\pi^2}
\\
\hline
\end{array}
$$
\caption{Form factors for the vertex $\left\langle T^{\mu\nu}T^{\rho\sigma}T^{\alpha\beta} \right\rangle (q,p)$ with the
graviton lines $(\alpha,\beta,p)$ and $(\rho,\sigma,q)$ on the mass shell.}
\label{Ch1FormFactors}
\end{table}
Below we give the explicit form of the 13 tensors $t_i^{\mu\nu\alpha\beta\rho\sigma}(p,q)$,
\begin{eqnarray}
t_1^{\mu \nu \rho \sigma\alpha \beta}(q,p)
&=&
\big( p^{\mu} p^{\nu} + q^{\mu} q^{\nu}\big) \, p^{\rho} p^{\sigma} q^{\alpha} q^{\beta} \nonumber\\
t_2^{\mu \nu \rho \sigma\alpha \beta}(q,p)
&=&
\big( p^{\mu} q^{\nu} + p^{\nu} q^{\mu} \big) \, p^{\rho} p^{\sigma} q^{\alpha} q^{\beta}
\nonumber \\
t_3^{\mu\nu\rho\sigma\alpha\beta}(q,p)
&=&
\big(p^{\mu} p^{\nu} + q^{\mu} q^{\nu} \big) \,
\big( p^{\sigma} q^{\beta} \delta^{\alpha \rho} + p^{\sigma} q^{\alpha} \delta^{\beta \rho}
+ p^{\rho} q^{\beta} \delta^{\alpha \sigma} + p^{\rho} q^{\alpha}\delta^{\beta \sigma} \big) \nonumber \\
t_4^{\mu\nu\rho\sigma\alpha\beta}(q,p)
&=&
p^{\rho} p^{\sigma} \, \big(
q^{\beta} q^{\nu} \delta^{\alpha \mu} + q^{\beta} q^{\mu} \delta^{\alpha \nu}
+ q^{\alpha} q^{\nu} \delta^{\beta \mu} + q^{\alpha} q^{\mu} \delta^{\beta \nu}
\big) \nonumber \\
&+&
q^{\alpha} q^{\beta} \, \big(
p^{\nu} p^{\sigma} \delta^{\mu \rho} + p^{\nu} p^{\rho} \delta^{\mu \sigma}
+ p^{\mu} p^{\sigma} \delta^{\nu \rho} + p^{\mu} p^{\rho} \delta^{\nu \sigma}
\big)
\nonumber \\
t_5^{\mu\nu\rho\sigma\alpha\beta}(q,p)
&=&
\big(p^{\mu} q^{\nu} + q^{\mu} p^{\nu}\big) \, \bigg(
p^{\rho} \big(q^{\alpha} \delta^{\beta\sigma} + q^{\beta} \delta^{\alpha\sigma} \big)
+ p^{\sigma} \big(q^{\alpha}\delta^{\beta \rho} + q^{\beta} \delta^{\alpha\rho} \big) \bigg) \nonumber \\
t_6^{\mu\nu\rho\sigma\alpha\beta}(q,p)
&=&
\delta^{\mu \nu} p^{\rho} p^{\sigma} q^{\alpha} q^{\beta}
\nonumber \\
t_7^{\mu\nu\rho\sigma\alpha\beta}(q,p)
&=&
p^{\rho}p^{\sigma} \, \big(\delta^{\mu\alpha}\delta^{\nu\beta} + \delta^{\mu\beta}\delta^{\nu\alpha}\big)
+ q^{\alpha}q^{\beta} \, \big(\delta^{\mu\rho}\delta^{\nu\sigma} + \delta^{\mu\sigma}\delta^{\nu\rho}\big) \nonumber\\
&-&
\frac{1}{2} \, \bigg(
p^{\mu}p^{\rho} \big( \delta^{\alpha\sigma}\delta^{\nu\beta} + \delta^{\beta\sigma}\delta^{\nu\alpha} \big)
+ p^{\nu}p^{\rho} \big( \delta^{\alpha\sigma}\delta^{\mu\beta} + \delta^{\beta\sigma}\delta^{\mu\alpha} \big) \nonumber \\
&+&
p^{\mu}p^{\sigma} \big( \delta^{\alpha\rho}\delta^{\nu\beta} + \delta^{\beta\rho}\delta^{\nu\alpha} \big)
+ p^{\nu}p^{\sigma} \big( \delta^{\alpha\rho}\delta^{\mu\beta} + \delta^{\beta\rho}\delta^{\mu\alpha} \big) \nonumber \\
&+&
q^{\mu}q^{\alpha} \big( \delta^{\beta\sigma}\delta^{\nu\rho} + \delta^{\beta\rho}\delta^{\nu\sigma} \big)
+ q^{\nu}q^{\alpha} \big( \delta^{\beta\sigma}\delta^{\mu\rho} + \delta^{\beta\rho}\delta^{\mu\sigma} \big)\nonumber \\
&+&
q^{\mu}q^{\beta} \big( \delta^{\alpha\sigma}\delta^{\nu\rho} + \delta^{\alpha\rho}\delta^{\nu\sigma} \big)
+ q^{\nu}q^{\beta} \big( \delta^{\alpha\sigma}\delta^{\mu\rho} + \delta^{\alpha\rho}\delta^{\mu\sigma} \big)
\bigg) \nonumber \\
t_8^{\mu\nu\rho\sigma\alpha\beta}(q,p)
&=&
\big(p^{\mu} p^{\nu}+q^{\mu} q^{\nu}\big) \,
\big(\delta^{\alpha \sigma} \delta^{\beta \rho} + \delta^{\alpha \rho} \delta^{\beta \sigma})
\nonumber \\
t_9^{\mu\nu\rho\sigma\alpha\beta}(q,p)
&=&
p^{\rho} \, \bigg(
q^{\mu} (\delta^{\alpha \sigma} \delta^{\beta \nu}+\delta^{\alpha \nu} \delta^{\beta \sigma})
+ q^{\nu} (\delta^{\alpha \sigma} \delta^{\beta \mu}+\delta^{\alpha \mu} \delta^{\beta \sigma})
\bigg) \nonumber \\
&+&
p^{\sigma} \, \bigg(
q^{\mu} (\delta^{\alpha \rho} \delta^{\beta \nu}+\delta^{\alpha \nu} \delta^{\beta \rho})
+ q^{\nu} (\delta^{\alpha \rho} \delta^{\beta \mu} + \delta^{\alpha \mu} \delta^{\beta \rho})
\bigg) \nonumber \\
&+&
q^{\alpha} \, \bigg(
p^{\mu} (\delta^{\beta \sigma} \delta^{\nu \rho} + \delta^{\beta \rho} \delta^{\nu \sigma})
+ p^{\nu} (\delta^{\beta \sigma} \delta^{\mu\rho} + \delta^{\beta \rho} \delta^{\mu \sigma})
\bigg) \nonumber \\
&+&
q^{\beta} \, \bigg(
p^{\mu} ( \delta^{\alpha\sigma}\delta^{\nu\rho} + \delta^{\alpha\rho} \delta^{\nu\sigma})
+ p^{\nu} ( \delta^{\alpha\sigma}\delta^{\mu\rho} + \delta^{\alpha\rho}\delta^{\mu\sigma})
\bigg)
\nonumber
\end{eqnarray}
\begin{eqnarray}
t_{10}^{\mu\nu\rho\sigma\alpha\beta}(q,p)
&=&
p^{\rho} \, \bigg(
q^{\alpha} (\delta^{\beta \nu} \delta^{\mu \sigma} + \delta^{\beta \mu} \delta^{\nu \sigma})
+ q^{\beta} (\delta^{\alpha \nu} \delta^{\mu \sigma} + \delta^{\alpha \mu} \delta^{\nu\sigma})
\big) \nonumber \\
&+&
p^{\sigma} \, \bigg(
q^{\alpha} (\delta^{\beta \nu} \delta^{\mu\rho} + \delta^{\beta \mu} \delta^{\nu \rho})
+ q^{\beta} (\delta^{\alpha \nu} \delta^{\mu \rho} + \delta^{\alpha \mu} \delta^{\nu \rho}) \bigg) \nonumber \\
&-& p.q \, \bigg(
\delta^{\alpha \rho} (\delta^{\beta \nu} \delta^{\mu \sigma} + \delta^{\beta \mu} \delta^{\nu \sigma})
+ \delta^{\alpha \nu} (\delta^{\beta \sigma} \delta^{\mu \rho} + \delta^{\beta \rho} \delta^{\mu \sigma}) \nonumber \\
&+&
\delta^{\alpha \mu} (\delta^{\beta \sigma} \delta^{\nu \rho} + \delta^{\beta \rho} \delta^{\nu \sigma} )
+ \delta^{\alpha \sigma} (\delta^{\beta \nu} \delta^{\mu \rho}+\delta^{\beta \mu} \delta^{\nu\rho})
\bigg) \nonumber \\
t_{11}^{\mu\nu\rho\sigma\alpha\beta}(q,p)
&=&
\big(p^{\nu} q^{\mu} + p^{\mu} q^{\nu}\big) \,
\big(\delta^{\alpha \sigma} \delta^{\beta \rho} + \delta^{\alpha \rho} \delta^{\beta \sigma}\big)
\nonumber \\
t_{12}^{\mu\nu\rho\sigma\alpha\beta}(q,p)
&=&
\delta^{\mu \nu} \,
\bigg( p^{\rho} \big( q^{\beta} \delta^{\alpha \sigma} + q^{\alpha} \delta^{\beta\sigma} \big)
+ p^{\sigma} \big( q^{\beta} \delta^{\alpha \rho} + q^{\alpha} \delta^{\beta \rho} \big)
\bigg)
\nonumber\\
t_{13}^{\mu\nu\rho\sigma\alpha\beta}(q,p)
&=&
\delta^{\mu \nu} \, \big(\delta^{\alpha \sigma} \delta^{\beta \rho}+\delta^{\alpha \rho} \delta^{\beta \sigma}\big).
\end{eqnarray}
In this limit, the correlator is affected by ultraviolet divergences coming from the $2$-point integrals $\mathcal{B}_0(s)$ .
This is true in the off-shell case too, as all the other contributions to the scalar coefficients
of its tensor expansion are made up of the three invariants $p^2$, $q^2$ and $p \cdot q$ plus the scalar 3-point integral
\begin{equation}
\mathcal {C}_0(s,s_1,s_2) =
\frac {1}{\pi^2} \int d^dl\, \frac{1}{l^2\,(l + p_1)^2\, (l + p_2)^2}\, , \quad s = (p_1+p_2)^2 \, , s_i= p_i^2\, , \quad i=1,2 \, ,
\end{equation}
which is finite for $d=4$.
The bare $2$-point integral in $4$ dimensions is defined in eq. (\ref{BareB0}).
After using the renormalization procedure discussed in the previous section in the $\overline{MS}$ scheme, it becomes
\begin{equation}
B_0^{\overline{MS}}(p^2) = 2 + \log\left(\frac{\mu^2}{p^2}\right)\, ,
\label{RenB0}
\end{equation}
For the sake of completeness, we report that we have checked that by taking the trace of these 13 tensors one reproduces the Weyl,
Euler and local contributions to the trace anomaly satisfied by the vertex, which in this on-shell case are given by
\begin{eqnarray}
\delta_{\mu\nu}\left\langle T^{\mu\nu}T^{\rho\sigma}T^{\alpha\beta}\right\rangle(q,p)\bigg|_{On-Shell}
&=&
4 \, \bigg\{\beta_a\,\bigg( \big[F\big]^{\rho\sigma\alpha\beta}(q,p)
- \frac{2}{3} \big[ \sqrt{g}\Box\,R\big]^{\rho\sigma\alpha\beta}(q,p) \bigg)
\nonumber\\
&+& \beta_b\, \big[G\big]^{\rho\sigma\alpha\beta}(q,p)\bigg\}\bigg|_{On-Shell} \, ,
\label{Ch1munu3PFanomaly1}
\end{eqnarray}
\begin{eqnarray}
\delta_{\alpha\beta}\left\langle T^{\mu\nu}T^{\rho\sigma}T^{\alpha\beta}\right\rangle(q,p)\bigg|_{On-Shell}
&=&
4 \, \bigg\{\beta_a\,\bigg(\big[F\big]^{\mu\nu\rho\sigma}(-k,q)
- \frac{2}{3} \big[\sqrt{g}\Box\,R\big]^{\mu\nu\rho\sigma}(-k,q)\bigg)
\nonumber\\
&+&
\beta_b\, \big[G\big]^{\mu\nu\rho\sigma}(-k,q)
- \frac{1}{2}\,\left\langle T^{\mu\nu}T^{\rho\sigma}\right\rangle(k)\bigg\}\bigg|_{On-Shell} \, ,
\label{Ch1albe3PFanomaly1}
\end{eqnarray}
\begin{eqnarray}
\delta_{\rho\sigma}\left\langle T^{\mu\nu}T^{\rho\sigma}T^{\alpha\beta}\right\rangle(q,p)\bigg|_{On-Shell}
&=&
4 \, \bigg\{\beta_a\,\bigg(\big[F\big]^{\mu\nu\alpha\beta}(-k,p)
- \frac{2}{3} \big[\sqrt{g}\Box\,R\big]^{\mu\nu\alpha\beta}(-k,p)\bigg)
\nonumber\\
&+&
\beta_b\, \big[G\big]^{\mu\nu\alpha\beta}(-k,p)
- \frac{1}{2} \,\left\langle T^{\mu\nu}T^{\alpha\beta}\right\rangle(k)\bigg\}\bigg|_{On-Shell}\, ,
\nonumber\\
\label{Ch1rosi3PFanomaly1}
\end{eqnarray}
with
\begin{eqnarray}
\left[F\right]^{\rho\sigma\alpha\beta}(q,p)\bigg|_{On-Shell}
&=&
2\, p^{\rho} \, p^{\sigma} q^{\alpha} q^{\beta}
- p\cdot q\,\bigg( p^{\sigma} q^{\beta} \delta^{\alpha\rho} - p^{\rho} q^{\beta} \delta^{\alpha\sigma}
- p^{\sigma} q^{\alpha} \delta^{\beta\rho} - p^{\rho} q^{\alpha} \delta^{\beta\sigma}\bigg)
\nonumber \\
&+&
(p\cdot q)^2\, \bigg(\delta^{\alpha\sigma} \delta^{\beta\rho} + \delta^{\alpha\rho} \delta^{\beta\sigma}\bigg) \, ,
\\
\left[G\right]^{\rho\sigma\alpha\beta}(q,p)\bigg|_{On-Shell}
&=&
2\, p^{\rho} p^{\sigma} q^{\alpha} q^{\beta}
- p\cdot q\,\bigg( p^{\sigma} q^{\beta} \delta^{\alpha\rho} - p^{\rho} q^{\beta} \delta^{\alpha \sigma}
- p^{\sigma} q^{\alpha} \delta^{\beta\rho} - p^{\rho} q^{\alpha} \delta^{\beta \sigma}\bigg)
\nonumber\\
&+&
(p\cdot q)^2\, \bigg(\delta^{\alpha \sigma} \delta^{\beta \rho} + \delta^{\alpha \rho} \delta^{\beta \sigma}\bigg) \, ,
\\
\left[\sqrt{g}\,\square R\right]^{\rho\sigma\alpha\beta}(q,p)\bigg|_{On-Shell}
&=&
\frac{1}{2}\, p\cdot q\,\bigg(p^{\sigma} q^{\beta} \delta^{\alpha\rho} + p^{\rho} q^{\beta} \delta^{\alpha \sigma}
+ p^{\sigma} q^{\alpha} \delta^{\beta\rho} + p^{\rho} q^{\alpha} \delta^{\beta \sigma}\bigg)
\nonumber\\
&-&
\frac{3}{2}\, (p\cdot q)^2 \bigg(g^{\alpha \sigma} \delta^{\beta \rho} - \delta^{\alpha \rho} \delta^{\beta \sigma}\bigg)\, .
\end{eqnarray}
\section{Conclusions and perspectives: the integrated anomaly and the nonlocal action}
Before coming to our conclusions, we offer here a brief discussion of the possible extensions of our analysis in the context
of the emergence of massless degrees of freedom in the computation of correlators of the form $TVV$ and $TTT$,
as predicted by Riegert's non local solution \cite{Riegert:1984kt} of the anomaly equation.
We recall that an action that formally solves the anomaly equation (\ref{Ch1TraceAnomaly}) takes the form
\begin{eqnarray}
S_{anom}[g,A]
&=&
\frac {1}{8}\int d^4x\sqrt{g}\int d^4x'\sqrt{-g'} \left(G + \frac{2}{3} \square R\right)_x\,
G_4(x,x')\,\left[ 2\,\beta_a\, F + \beta_b\, \left(G + \frac{2}{3} \square R\right)
- 2\, \frac{\kappa}{4}\, F_{\mu\nu}F^{\mu\nu}\right]_{x'}
\nonumber \\
&&
-\, \frac{\beta_a+\beta_b}{18}\, \int d^4x\,\sqrt{g}\, R^2
\label{Ch1Riegertactions}
\end{eqnarray}
where we keep the sum on all the contributions coming from different particle species implicit. \\
The notation $G_4(x,x')$ denotes the Green's function of the differential operator defined by
\begin{equation}
\Delta_4 \equiv \nabla_\mu\left(\nabla^\mu\nabla^\nu - 2 R^{\mu\nu} + \frac{2}{3}\, g^{\mu\nu}\, R \right)
\nabla_\nu = \square^2 - 2\, R^{\mu\nu}\nabla_\mu \nabla_\nu -\frac{1}{3}\, (\nabla^\mu R)\, \nabla_\mu
+ \frac{2}{3}\, \square R \, ,
\label{Ch1operator4}
\end{equation}
which is the only order-$4$ differential operator enjoying the invariance property under Weyl transformations
\begin{equation}
\sqrt{g}\, G_4(x,x')\rightarrow \sqrt{g}\, G_4(x,x')\, , \quad \text{for} \quad
g_{\mu\nu} \rightarrow e^{2\,\sigma(x)}\, g_{\mu\nu}(x)\, ,
\end{equation}
whereas the combination $\left( G + 2/3 \square R\right)$ transforms as
\begin{equation}
\left( G + \frac{2}{3}\, \square R \right) \rightarrow
\left( G + \frac{2}{3}\, \square R \right) + 4\, \Delta_4 \sigma\, , \quad \text{for} \quad
g_{\mu\nu} \rightarrow e^{2\,\sigma(x)}\, g_{\mu\nu}(x)\, ,
\end{equation}
which immediately helps understanding how the trace of the EMT derived from (\ref{Ch1operator4})
reproduces the trace anomaly in (\ref{Ch1TraceAnomaly}).
As shown in \cite{Giannotti:2008cv,Mottola:2006ew}, performing repeated variations of the "anomaly-induced" action
(\ref{Ch1Riegertactions}) with respect to the background metric $g_{\mu\nu}$ and to the $A_{\alpha}$ gauge field,
one can reproduce the anomalous contribution of correlators with multiple insertions of the EMT or of gauge currents.
Of course, this action does not reproduce the homogeneous contributions to the anomalous trace Ward identity (no variational solution
could). Because these require an independent computation in order to be identified,
such as the perturbative $1$-loop analysis undertaken in this work.
The action can be reformulated in such a way that its interactions become local \cite{Mottola:2006ew},
by introducing two auxiliary scalar fields.
After some manipulations, one can show that the apparently double pole in $G_4(x,x')$ reduces to a single pole
and the anomaly-induced action near a flat background takes the simpler form
\begin{equation}
S_{anom}[g,A] \rightarrow
- \frac{\kappa}{24}\int d^4x\sqrt{g}\int d^4x'\sqrt{-g'}\, R_x\,
\square^{-1}_{x,x'}\, [F_{\alpha\beta}F^{\alpha\beta}]_{x'} \, .
\label{Ch1SSimple}
\end{equation}
Notice that this action is valid to first order in metric variations around flat space. Its local expression is given by
\begin{equation}
S_{anom} [g,A;\varphi,\psi'] = \int\,d^4x\,\sqrt{g}\,
\left[ -\psi'\square\,\varphi - \frac{R}{3}\, \psi' + \frac{c}{2} F_{\alpha\beta}F^{\alpha\beta} \varphi\right] \, .
\label{Ch1effact}
\end{equation}
with $\psi'$ and $\varphi$ defined as in \cite{Giannotti:2008cv}.
$R_x$, in the equations above, is the linearised version of the Ricci scalar
\begin{equation}
R_x \equiv \partial^x_\mu\, \partial^x_\nu \, h^{\mu\nu} - \square \, h, \qquad h=\eta_{\mu\nu} \, h^{\mu\nu} \, .
\end{equation}
Eq. (\ref{Ch1effact}) shows the appearance of coupled massless degrees of freedom, whose interpretation was been offered in
\cite{Giannotti:2008cv}, to which we refer for further details, using the approach of dispersion relations.
This analysis, so far, has been limited to the $TVV$ correlator and could be extended, with a lot of additional effort,
to the case of the $TTT$ vertex whose explicit computation has been discussed in this work.
In particular, this analysis could test directly if the pole structure present in the
expression of the $TTT$ vertex will match the prediction of the same vertex once this is computed using (\ref{Ch1Riegertactions})
by functional differentiation with respect to the metric.
This point is technically very involved, since it requires a comparison between the off-shell result of a direct computation of the $TTT$
in perturbative field theory, as done in this work, with the anomalous part of the same correlator computed from Riegert' s variational solution.
We hope to come back to discuss this point in a related work.
\clearpage{\pagestyle{empty}\cleardoublepage}
\chapter{Conformal correlators in position and momentum space}\label{Mapping}
\section{Introduction}
The analysis of correlation functions in $d$-dimensional quantum field theory possessing conformal invariance has found widespread
interest over the years (see \cite{Fradkin:1997df} for an overview). Given the infinite dimensional character of the conformal algebra in $2$
dimensions, conformal field theories (CFT's from now on) for $d=2$ have received much more attention than their
$4$-dimensional counterparts.
In $d$ dimensional CFT's the structure of generic conformal correlators is not entirely fixed just by conformal symmetry,
but for $2$- and $3$-point functions the situation is rather special and these can be significantly constrained, up to a small number of constants.
From the CFT side, some important information, mainly due to \cite{Osborn:1993cr,Erdmenger:1996yc}, is available.
These results concern the $TOO$ - with $O$ denoting a generic scalar operator - $TVV$ and $TTT$ vertices, which are determined by
implementing the conformal constraints in position space. In the analysis of \cite{Osborn:1993cr}, in particular, it was shown for the first time
that some of these vertices are expressible in terms of few linearly independent tensor structures: specifically, their numbers are
$1$ for the $TOO$ vertex, $2$ for the $TVV$ and $3$ for the $TTT$. \\
Imposing the conformal Ward identities and identifying these tensor structures directly in momentum space turns out the be technically quite
involved. The main goal of the present chapter is to present the result of a systematic study, initiated in \cite{Coriano:2012wp},
enabling comparison of general results of d-dimensional CFT's based on position space analyses, such as those in
\cite{Osborn:1993cr, Erdmenger:1996yc}, with explicit realizations of anomalous $3$-point vertices in free CFT's,
most commonly expressed in momentum space.
Recent results of studies of $3$- and $4$-point functions in $d=3$ in the context of the $ADS/CFT$ correspondence
are contained in \cite{Bzowski:2011ab,Maldacena:2011nz,Raju:2012zs}.
Another significant difference between the position-space approach and the computations in momentum space is that
conformal anomalies necessarily arise quite differently in the two contexts.
In the former case, in fact, anomalous terms show up as ultra-local divergences proportional to delta functions or derivatives thereof
at coincident spatial points. Thus a very careful regularization procedure is required to
determine these anomalous ultra-local contributions which are absent for any finite point separation.
The special strategy followed in determining these anomalous ultra-local contributions in position space, developed in \cite{Osborn:1993cr},
deserves some comments. In this approach, the diffeomorphism and trace Ward identities are solved
in each case by combining a completely homogeneous solution, which is built on the ground of the requirements of Lorentz and
conformal invariance and is non local - i.e. obtained keeping the three points separate - with inhomogeneous terms.
The inhomogeneous terms are of two kinds: terms of the first kind are semi-local (two coincident points out of three),
are identified via the Operator Product Expansion of correlators and account for ordinary contact terms in the Ward identities;
the terms of the second kind are ultra-local (all three points coincident). In particular, the latter contributions
arise from the need to subtract ultraviolet singularities appearing in the ultra-local limit and are responsible for the corresponding trace anomalies.
It is clear that such a separation, based on the distinction of terms according to the separation/coincidence of points
at which operators are evaluated, does not make sense in momentum space. Here anomalies are thought either
as a remnant of the renormalization of ultraviolet divergences, that breaks conformal invariance via the introduction of a mass scale $\mu$
(see for example \cite{Duff:1977ay,Duff:1993wm,Birrell:1982ix}) or as an infrared effect, as can be seen in the dispersive approach
of \cite{Giannotti:2008cv}, where the trace anomaly is shown originating from the imposition of all the non anomalous Ward identities
and the spectral representation of the amplitude.
At first glance this appears to be quite different than the ultra-local delta function terms obtained in the position
space approach of \cite{Osborn:1993cr, Erdmenger:1996yc}.
Thus, the relationship between this approaches requires some clarification, and this is a principal motivation for the present work.
A necessary comment is that the eventual agreement of the two approaches may seem less surprising if it is remembered that coincident point
singularities in euclidean position space become light cone singularities in Minkowski spacetime, and these light cone singularities are associated
with the propagation of massless fields, which generally have long range infrared effects. \\
This chapter is composed of two main parts.
In the first part, building on the results of \cite{Osborn:1993cr, Erdmenger:1996yc},
we analyse the structure of the $3$-point correlators in configuration and in momentum space for a general CFT.
In particular we generalize previous studies of the $TVV$ correlator, perturbatively evaluated in $4$ dimensions in
\cite{Armillis:2009pq, Armillis:2010qk,Coriano:2011zk} in QED, QCD and the Standard Model, to $d$ dimensions.
We also study the $TTT$ vertex, whose computation in $4$ dimensions was presented in the previous chapter,
and perform a complete investigation of this correlator by the same approach.
We give particular emphasis to the discussion of the connection between the general approach of \cite{Osborn:1993cr}
and the perturbative picture. In particular, we give a diagrammatic interpretation of the various contact terms introduced
by Osborn and Petkou in order to solve the Ward identities for generic positions of the points of the correlators.
This allows to close a gap between their bootstrap method, previous investigations of the $TVV$
\cite{Giannotti:2008cv,Armillis:2009pq, Armillis:2010qk}, and the recent computation of the $TTT$ vertex.
We show that the perturbative analysis in momentum space in dimensional regularization is in complete agreement with their results. \\
It should be remarked that, in general, the momentum space formulation of the correlators of a CFT has remained largely unexplored
until the publication of \cite{Coriano:2013jba,Bzowski:2013sza}, where conformal constraints in momentum space for $3$-point functions are
systematically explored, proving, as expected, much more difficult to implement with respect to the corresponding position space constraints.
The lack of this investigation for such a long time is mainly due to the fact that momentum space is ideally suited for perturbative computations,
which in turn always stem from a Lagrangian formulation. This Lagrangian is often missing for CFT's,
which can be defined on the sole ground of symmetry principles.
This brings us to the second part of the chapter, contained in section \ref{Ch1direct}, where we discuss a general procedure to map
to momentum space any massless correlator given in position space and not necessarily related to a Lagrangian description.
The investigation of these correlators in momentum space reveals, in general, some specific facts, such as the presence of single and
multi-logarithmic integrands which, in general, cannot be re-expressed in terms of ordinary master integrals, typical of the Feynman
expansion. In particular, we conclude that, whenever the mentioned logarithmic integrals do not cancel, the theory does not possess
a Lagrangian formulation, because otherwise no such integral would appear in the perturbative expansion. \\
To address these points, one has to formulate an alternative and general approach to perform the transforms,
not directly linked to the Lagrangian realization, since in this case such representation may not exist.
The method that is proposed relies on a $d$-dimensional version of differential regularization, similar to the approach suggested in
\cite{Erdmenger:1996yc}. We use the standard technique of \emph{pulling out derivatives via partial integration}
in singular correlators in such a way to make them Fourier-integrable, i.e. expressible as integrals in momentum space.
This is combined with {\em the method of uniqueness} \cite{Kazakov:1986mu}, here generalized to tensor structures,
in order to formulate a complete and self-consistent procedure.
As in \cite{Osborn:1993cr, Erdmenger:1996yc} we need an extra regulator ($\omega$),
unrelated to the dimensional regularization parameter $\epsilon$.
Our approach is defined as a generic algorithm which can handle rather straightforwardly any massless correlator
written in configuration space.
The algorithm has been implemented in a symbolic manipulation program and can handle, in principle, correlators of any rank. \\
The aim of the method is to test the Fourier-integrability of a given correlator, by checking the cancellation of the singularities
in the extra regulator $\omega$ directly in momentum space, and to provide us with the direct expression of the transform.
\section{The correlators and the corresponding Ward identities}
We provide the basic definitions of the correlators that we are going to investigate.
We suppose that the theory admits an euclidean generating functional $\mathcal W$, in analogy with (\ref{Ch1Generating}),
which depends on the background metric $g_{\mu\nu}$, acting as source of the EMT $T^{\mu\nu}$,
on gauge fields $A^a_\mu$ coupled to the gauge currents $V^a_\mu$, and a source $J$ for each scalar operator $\mathcal{O}$
of the spectrum (for the sake of simplicity, we do not distinguish them ).
Thus, if the classical theory is described by an action $\mathcal S$, we are embedding this into a curved space via the coupling
to the metric $g$ and supplying it with additional source terms,
\begin{equation}
\mathcal{W}[g,A,J]= \int \mathcal{D}\Phi\,
e^{- \mathcal S - \int d^dx\, \sqrt{g}\, \left( A^a_\mu \,V^{a\,\mu} + J \, \mathcal O \right)}\, .
\label{GenPlusSources}
\end{equation}
Then, the functional averages of $\mathcal O$, $V$ and $T$ are obtained by differentiating
the generating functional with respect to the corresponding sources, i.e.
\begin{equation}
\left\langle O(x) \right\rangle_s = - \frac{1}{\sqrt{g_x}} \frac{\delta \mathcal{W}}{\delta J(x)}\, ,
\quad
\left\langle V^{a \, \mu}(x) \right\rangle_s = -\frac{1}{\sqrt{g_x}} \frac{\delta \mathcal{W}}{\delta A^a_\mu(x)}\, ,
\quad
\left\langle T^{\mu\nu}(x) \right\rangle_s = \frac{2}{\sqrt{g_x}}\frac{\delta\, \mathcal W}{\delta\, g_{\mu\nu}(x)}\, .
\end{equation}
The construction of the correlators is straightforward.
If the scalar operator $O$ is coupled to the source $J$, the three point function is defined via a triple functional derivative
with respect to $g_{\mu\nu}$ once and to the scalar source $J$ twice, evaluated switching off the sources at the end, i.e.
\begin{eqnarray}
\left\langle T^{\mu\nu}(x_1) O (x_2)O(x_3) \right\rangle
&=&
\bigg\{ \frac{\delta^2}{\delta J(x_2) \delta J(x_3)} \bigg[
\frac{2}{\sqrt{g_{x_1}}} \frac{\delta \mathcal{W}}{\delta g_{\mu\nu}(x_1)} \bigg]_{g=\delta}\bigg\}_{J=0} \nonumber \\
&& \hspace{-20mm}
=\, \left\langle T^{\mu\nu}(x_1) O(x_2) O(x_3) \right\rangle_{J=0} + \left\langle \frac{\delta T^{\mu\nu}[J](x_1)}{
\delta J(x_2)} O(x_3)
\right\rangle_{J= 0} + \left\langle \frac{\delta T^{\mu\nu}[J](x_1)}{\delta J(x_3)} O(x_2) \right\rangle_{J= 0}\, ,
\label{DefineTOO}
\end{eqnarray}
The second correlator that we will analyse will be the $VVV$ vertex, which is defined by the third functional derivative of the
generating functional with respect to the source gauge field $A^a_{\mu}(x)$
\begin{eqnarray}
\left\langle V^{a\,\mu}(x_1) V^{b\,\nu}(x_2) V^{c\,\rho}(x_3) \right\rangle = -
\frac{\delta^3 \mathcal{W}|_{g=\delta}}{\delta A^a_\mu(x_1) \delta A^b_\nu(x_2) \delta A^c_\rho(x_3)}\bigg |_{A=0}\, .
\end{eqnarray}
We remark that, due to Furry's theorem, the gauge theory has to be non abelian
in order to define a non vanishing $VVV$ correlator.
To derive the $TVV$ correlator, we can first perform a functional derivative with respect to the metric
and then insert the vector currents by functionally differentiating with respect to the gauge field sources $A$, specifically
\begin{eqnarray}
\left\langle T^{\mu\nu}(x_1)\, V^{a \, \alpha} (x_2)\, V^{b \, \beta} (x_3) \right\rangle
&=&
\bigg\{\frac{\delta^2}{\delta A^a_\alpha(x_2) \delta A^b_\beta(x_3)}
\bigg[\frac{2}{\sqrt{g_{x_1}}} \frac{\delta \mathcal{W}}{\delta g_{\mu\nu}(x_1)} \bigg]_{g=\delta} \bigg\}_{A=0}
\nonumber \\
&=&
\left\langle T^{\mu\nu}(x_1)\, V^{a\,\alpha}(x_2)\, V^{b\,\beta}(x_3) \right\rangle_{A=0} \nonumber \\
&&
+\, \left\langle \frac{\delta T^{\mu\nu}(x_1)}{\delta A^a_\alpha(x_2)} V^{b\,\beta}(x_3) \right\rangle_{A= 0} +
\left\langle \frac{\delta T^{\mu\nu}(x_1)}{\delta A^b_\beta(x_3)} V^{a\,\alpha}(x_2) \right\rangle_{A= 0}
\label{DefineTVV}
\end{eqnarray}
where $T_{\mu\nu}$ is calculated in the presence of the background source $A_\mu^a$. The first term in
the previous expression represents the insertion of the three operators, while the last two are contact terms, with the topology of
$2$-point functions, exploiting the linear dependence of the EMT from the source field $A$.
Finally, for the definition of the $TTT$ Green function, which obviously does not change after including the additional sources
in the generating functional, we refer to section (\ref{Ch1DiagTTT}).
Now we turn to the derivation of non-anomalous Ward identities, which hold for general dimensions,
by which we mean away from the (even) values of the space dimension for which the trace Ward identities become anomalous
\cite{Birrell:1982ix}. \\
We assume that the generating functional $W[g,A,J]$ is invariant under diffeomorphisms,
\begin{eqnarray}
\mathcal{W}[g',A',J'] = \mathcal{W}[g,A,J] \, ,
\end{eqnarray}
where $g'$ and $A'$ and $J'$ are transformed metric, gauge field and scalar source under the general infinitesimal
coordinate transformation $x^\mu \rightarrow {x'}^\mu = x^\mu + \epsilon^\mu(x)$, under which they change according to
\begin{equation}
\delta g_{\mu\nu} = \nabla_{\mu} \epsilon_{\nu} + \nabla_{\nu} \epsilon_{\mu} \, ,
\qquad
\delta A^{a}_{\mu}=
- \left( \epsilon^{\lambda} \nabla_{\lambda} A^a_{\mu} + A^{a \, \lambda} \nabla_{\mu} \epsilon_{\lambda} \right) \, ,
\qquad
\delta J = - \epsilon^\lambda \partial_\lambda J \, .
\end{equation}
Diffeomorphism invariance and gauge invariance of the generating functional respectively imply the relations
\begin{eqnarray}
&& \qquad
\nabla_{\mu} \left\langle T^{\mu\nu} \right\rangle + \nabla^{\nu} A^a_\mu \left\langle V^{a \, \mu} \right\rangle +
\nabla_{\mu} \left( A^{a\,\nu} \left\langle V^{a \, \mu}\right\rangle \right)+ \partial_\nu J\, \left\langle\mathcal{O}\right\rangle = 0 \, , \nonumber \\
&& \qquad
\nabla_{\mu} \left\langle V^{a \, \mu} \right\rangle + f^{abc} A^b_{\mu} \left\langle V^{c \mu}\right\rangle = 0 \, ,
\label{Ch1BaiscWard}
\end{eqnarray}
where $f^{abc}$ are the structure constants of the gauge group. \\
Naive conformal invariance gives the tracelessness condition
\begin{equation} \label{Ch1NaiveScaleWI}
g_{\mu\nu} \left\langle T^{\mu\nu} \right\rangle_s + \left( d - \eta \right)\, J\, \left\langle \mathcal{O}\right\rangle_s = 0 \, .
\end{equation}
This last Ward identity is naive, due to the appearance of an anomaly at quantum level, after renormalization of the correlator,
for even dimensions. However, it is the correct identity away from $d=2\,k$, for integer $k$.
In this respect, the functional differentiation of (\ref{Ch1BaiscWard}) and (\ref{Ch1NaiveScaleWI})
allows to derive ordinary Ward identities for the various correlators. \\
If we want to include anomalies in $4$-dimensional space, then remembering eq. (\ref{Ch1TraceAnomaly})
and the conditions on the $\beta$ coefficients discussed below it, we can write, in dimensional regularization,
\begin{equation}
g_{\mu\nu}\left\langle T^{\mu\nu} \right\rangle_s =
\sum_{I=f,s,V}n_I\, \bigg[ \beta_a(I)\, \bigg( F - \frac{2}{3}\, \square R \bigg)+ \beta_b(I)\, G \bigg]
- \frac{\kappa}{4} \, n_V\, F^{a\,\mu\nu}\,F^a_{\mu\nu} + F[J] \, ,
\end{equation}
where by $F[J]$ we have denoted a possible functional of the background source for the scalar operators,
whose form is not unique, but depends on the dimensions of the corresponding operators.
For example, in $4$ dimensions and for the operator $\mathcal{O} = \phi^2 $, where $\phi$
is the standard elementary scalar field, then $F[J] = \frac{p}{2}\, J^2$, with $p$ a c-number.
Now let us list the Ward identities implied by (\ref{Ch1BaiscWard}) for the various correlators.
In the case of the $TOO$ vertex one has the equation
\begin{equation}
\partial^{x_1}_\mu \left\langle T^{\mu\nu}(x_1)\mathcal{O}(x_2)\mathcal{O}(x_3) \right\rangle =
\partial^{x_1}_\nu \delta^{(d)}(x_{12}) \left\langle \mathcal{O}(x_1)\mathcal{O}(x_3)\right\rangle +
\partial^{x_1}_\nu \delta^{(d)}(x_{13}) \left\langle \mathcal{O}(x_1)\mathcal{O}(x_2)\right\rangle \, .
\label{DiffWardTOO}
\end{equation}
For the $VVV$ the conservation equation is
\begin{equation}
\partial^{x_1}_\mu \left\langle V^{a\,\mu}(x_1)V^{b\,\nu}(x_2)V^{c\,\rho}(x_3) \right\rangle =
f^{abd}\, \delta^{(d)}(x_{12}) \left\langle V^{d\,\nu}(x_1)V^{c\,\rho}(x_3) \right\rangle -
f^{acd}\, \delta^{(d)}(x_{13}) \left\langle V^{d\,\rho}(x_1)V^{c\,\nu}(x_2) \right\rangle \, ,
\label{DiffWardVVV}
\end{equation}
Finally, for the case of the $TVV$ we obtain
\begin{eqnarray}
\partial_{\mu}^{x_1} \left\langle T^{\mu\nu}(x_1) V^{a\,\alpha}(x_2) V^{b\,\beta}(x_3) \right\rangle
&=&
\partial^{\nu}_{x_1} \delta^d(x_{12}) \left\langle V^{a\,\alpha}(x_1) V^{b\,\beta}(x_3) \right\rangle
+ \partial^{\nu}_{x_1} \delta^d(x_{31}) \left\langle V^{a\,\alpha}(x_2) V^{b\,\beta}(x_1) \right\rangle \nonumber \\
&-&
\delta^{\nu\alpha} \partial^{x_1\,}_{\mu} \left( \delta^d(x_{12}) \left\langle V^{a\,\mu}(x_1) V^{b\,\beta}(x_3)
\right\rangle \right) -
\delta^{\nu\beta} \partial^{x_1\,}_{\mu} \left( \delta^d(x_{31}) \left\langle V^{a\,\alpha}(x_2) V^{b\,\mu}(x_1)
\right\rangle \right) \, , \nonumber \\
\label{TVVWardCoord}
\end{eqnarray}
together with the vector current Ward identities, following from gauge invariance,
\begin{eqnarray}
\partial_{\alpha}^{x_2} \left\langle T^{\mu\nu}(x_1) V^{a\,\alpha}(x_2) V^{b\,\beta}(x_3) \right\rangle &=& 0 \,, \qquad
\partial_{\beta}^{x_3} \left\langle T^{\mu\nu}(x_1) V^{a\,\alpha}(x_2) V^{b\,\beta}(x_3) \right\rangle = 0 \, .
\label{GaugeInvTVV}
\end{eqnarray}
The general covariance Ward identity for the $TTT$ vertex was already given in section \ref{Ch1DiagTTTWardCov}.
Now let us move to trace Ward identities for both general $d$ and $4$ dimensions.
Discarding the $VVV$, the naive identity (\ref{Ch1NaiveScaleWI}) gives the non-anomalous conditions, for $d$ dimensions
\begin{eqnarray}
\delta_{\mu\nu} \, \left\langle T^{\mu\nu}(x_1) \mathcal{O}(x_2)\mathcal{O}(x_2) \right\rangle
&=&
\left( d-\eta \right)\, \left( \delta^{(d)}(x_{12})\,\left\langle \mathcal{O}(x_1)\mathcal{O}(x_3)\right\rangle +
\delta^{(d)}(x_{13})\,\left\langle \mathcal{O}(x_1)\mathcal{O}(x_2)\right\rangle \right) \, ,
\label{TraceWardTOO} \\
\delta_{\mu\nu} \, \left\langle T^{\mu\nu}(x_1) V^{a\,\alpha}(x_2) V^{b\,\beta}(x_3) \right\rangle
&=&
0 \, ,
\label{TraceWardTVV}
\end{eqnarray}
which become, when anomalies are properly included,
\begin{eqnarray}
\delta_{\mu\nu} \, \left\langle T^{\mu\nu}(x_1) \mathcal{O}(x_2)\mathcal{O}(x_2) \right\rangle
&=&
\left( 4-\eta \right)\, \left( \delta^{(d)}(x_{12})\,\left\langle \mathcal{O}(x_1)\mathcal{O}(x_3)\right\rangle +
\delta^{(d)}(x_{13})\,\left\langle \mathcal{O}(x_1)\mathcal{O}(x_2)\right\rangle \right) \nonumber
+ \frac{\delta^2 F[J](x_1)}{\delta J(x_2)\delta J(x_3)} \, ,
\label{AnTraceWardTOO}
\\
\delta_{\mu\nu} \, \left\langle T^{\mu\nu}(x_1) V^{a\,\alpha}(x_2) V^{b\,\beta}(x_3) \right\rangle
&=&
\delta^{ab}\, \kappa\, \left( \partial^\beta\delta^{(4)}(x_{12})\partial^\alpha\delta^{(4)}(x_{13}) -
\delta^{\alpha\beta}\, \partial^\lambda\delta^{(4)}(x_{12})\partial_\lambda\delta^{(4)}(x_{13}) \right) \, .
\label{AnTraceWardTVV}
\end{eqnarray}
The trace Ward identity for the $TTT$ was given in section \ref{Ch1DiagTTTWardAnom} in momentum space in 4 dimensions.
Here we report the coordinate-space versions of the non anomalous identity, holding for general $d$.
\begin{equation}
\delta_{\mu\nu}\left\langle T^{\mu\nu}(x_1)T^{\alpha\beta}(x_2)T^{\rho\sigma}(x_3) \right\rangle =
-2\, \bigg( \delta^{(4)}(x_{12}) + \delta^{(4)}(x_{13} ) \bigg)\,
\left\langle T^{\alpha\beta}(x_2)T^{\rho\sigma}(x_3)\right\rangle \, ,
\label{NoAnTraceTTT}
\end{equation}
and of the anomalous identity, valid only for $d=4$
\begin{eqnarray}
\delta_{\mu\nu}\left\langle T^{\mu\nu}(x_1)T^{\alpha\beta}(x_2)T^{\rho\sigma}(x_3) \right\rangle
&=&
4\, \bigg( \beta_a\, \left[ F(x_1) - \frac{2}{3}\,\Box R(x_1) \right]^{\rho\sigma\alpha\beta}(x_2,x_3)
+ \beta_b\, \left[ G(x_1)\right]^{\rho\sigma\alpha\beta}(x_2,x_3) \bigg) \nonumber \\
&&
-\, 2\, \bigg( \delta^{(4)}(x_{12}) + \delta^{(4)}(x_{13} ) \bigg)\,
\left\langle T^{\alpha\beta}(x_2)T^{\rho\sigma}(x_3)\right\rangle \, ,
\label{AnTraceTTT}
\end{eqnarray}
with the compact notation for functional derivatives in the flat space limit introduced in (\ref{Ch1Flat}).
\section{Inverse mappings: correlators in position space from the momentum space Feynman expansion}
Having by now defined all the fundamental (anomalous and regular) Ward identities which allow to test the correctness
of the correlators we are interested in studying, we turn to compare the expressions of these correlators in position space
with their perturbative realizations in free field theories in momentum space.
We remind that an important result of \cite{Osborn:1993cr} is the identification of the solution of the Ward identities in terms of a set of
constants and of certain linearly independent tensor structures in euclidean position space. Consistency requires that the Fourier transforms
of these tensor structures must occur in direct computations of the same vertex functions in free field theories in momentum space,
which are defined, in turn, via certain $1$-loop integrals that can be computed according to a well-defined set of Feynman rules,
once the Lagrangian theory has been specified.
This implies that, after establishing the combination of these integrals defining the Green function in momentum space,
we can use them to infer what those tensor structures must be, and find the exact correspondence between CFT amplitudes
in position space and momentum space {\it a posteriori}.
Obviously, this is only possible provided that we have enough linearly independent vertex functions for different free theories to determine
the linear combinations uniquely. We call this procedure an \emph{inverse mapping}, as it allows to re-express the correlators
of \cite{Osborn:1993cr} in such a form that their Fourier-integrability is explicit.
By \emph{integrable} we mean, in this context, a function of coordinates whose Fourier transform is not divergent and contains
no additional regulator, in a sense made precise in section \ref{Ch1pull}.
This result is obtained by pulling out derivatives of the corresponding diagrams, in the spirit of differential regularization \cite{Freedman:1991tk}
in such a way that integrability becomes trivial.
More technical details on the inverse mapping method are explicitly provided in appendix \ref{Ch1InverseTTT}, where the $TTT$
correlator is used as an illustrative example.
\subsection{The TOO case}
The first correlator that we are going to investigate is the $TOO$. In the perspective of comparing the coordinate space results of
\cite{Osborn:1993cr} with momentum space perturbative expansions, it is also the most ambiguous, as such a perturbative
expansion requires establishing the scaling dimensions of the scalar operators once for all.
For this reason, we are going to perform such a test only in one specific case, i.e. for $\mathcal O = \phi^2$.
The general structure of this Green function in coordinate space - for non coincident points - is
\begin{equation}
\label{Ch1TOOPO}
\left\langle T_{\mu\nu}(x_1) \, O(x_2) \, O(x_3) \right\rangle =
\frac{a}{(x^2_{12})^{d/2} \, (x^2_{23})^{\eta-d/2} \, (x^2_{31})^{d/2}} \, h^1_{\mu\nu}(\hat{X}_{23}) \, ,
\end{equation}
where $a$ is a constant, $\eta$ the scaling dimension of the scalar operator $O $ and where
\begin{eqnarray} \label{Ch1TOOstructures}
\hat{X}_\mu = \frac{X_\mu}{\sqrt{X^2}} \, , \nonumber \qquad
h^1_{\mu\nu}(\hat X) = \hat{X}_\mu \, \hat{X}_\nu - \frac{1}{d} \, \delta_{\mu\nu} \, ,
\end{eqnarray}
where
\begin{eqnarray}
x_{ij} \equiv x_i - x_j \, , \qquad X_{ij} = - X_{ji} \equiv
\frac{x_{ik}}{x^2_{ik}} - \frac{x_{jk}}{x^2_{jk}} \, , \quad i,j,k = 1,2,3 \, , \quad k\neq i \, , k \neq j \, .
\end{eqnarray}
In the short-distance limits of its external points, this vertex is singular for $\eta \rightarrow d/2$ and needs regularization.
In \cite{Osborn:1993cr} the Ward identities are solved through the analysis of the short distance limits of (\ref{Ch1TOOPO}),
by which we mean the limits $x_1 \rightarrow x_2$, $x_1 \rightarrow x_3$.
Some singular terms are found and thus the authors are forced to regularize them
with the method of differential regularization \cite{Freedman:1991tk}, which finally gives the modified expression
\begin{eqnarray}
\left\langle T_{\mu\nu}(x_1) \, O(x_2) \, O(x_3) \right\rangle
&=&
\frac{a}{(x^2_{12})^{d/2} \, (x^2_{23})^{\eta-d/2} \, (x^2_{31})^{d/2}} \, h^1_{\mu\nu}(\hat{X}_{23})
\nonumber \\
&&
+\, \left[ \hat A_{\mu\nu}(x_{12}) - A_{\mu\nu}(x_{12}) +
\hat A_{\mu\nu}(x_{31}) - A_{\mu\nu}(x_{31})\right] \, \frac{N}{(x_{23}^2)^{\eta}} \, ,
\label{CompleteTOO}
\end{eqnarray}
where we have introduced the structures
\begin{eqnarray}
A_{\mu\nu}(s) = \frac{a}{N}\, \frac{1}{s^d}\,
\left( \frac{s_\mu s_\nu}{s^2} - \frac{1}{d} \delta_{\mu\nu} \right) \, ,
\qquad
\hat A_{\mu\nu}(s) = \frac{a}{N\,d} \left( \frac{\partial_{\mu}\partial_{\nu}}{d-2} \, \frac{1}{s^{d-2}} +
\frac{\eta -d +1}{\eta}\, S_d \, \delta_{\mu\nu} \, \delta^d(s) \right) \, ,
\label{Ahat}
\end{eqnarray}
and $N$ is defined as the normalization constant of the $2$-point function of the scalar operator
\begin{equation}
\left\langle \mathcal{O}(x_1)\mathcal{O}(x_2) \right\rangle = \frac{N}{(x^2_{12})^\eta}\, .
\end{equation}
It is important to make some comments at this point, as the $TOO$ is the simplest function, between those addressed in this chapter,
to require such a regularization procedure in order to account for the inhomogeneous terms in the Ward identities.
Notice that, in general, one can define a differentially regularized tensor $\hat{A}$ as
\begin{equation}
\hat{A}_{\mu\nu}(s) = \frac{a}{N\,d}\,
\left( \frac{1}{d-2}\partial_\mu\partial_\nu \frac{2}{s^{d-2}} + C\,\delta_{\mu\nu}\,S_d \delta^{(d)}(s) \right) \, ,
\end{equation}
which \emph{exactly coincides with $A_{\mu\nu}(s)$ for $s \neq 0$} and where the only difference between them is in the
$\delta$-function term, whose coefficient is not fixed \emph{a priori}, reflecting the arbitrariness typical of any regularization.
It is precisely this kind of term that discerns (\ref{CompleteTOO}) from (\ref{Ch1TOOPO}).
The ambiguity in its coefficient is solved by requiring the Ward identities to be satisfied by (\ref{CompleteTOO}).
In this way, the terms $\hat{A}- A$ are the contact contributions (sometimes called \emph{semi-local} in the literature)
which consistently account for the r.h.s. of (\ref{DiffWardTOO}) and (\ref{TraceWardTOO}).
Essentially the same argument holds for the more complicated $3$-point functions we are going to discuss in the next sections:
only the formulas are more complicated, due to the increasing number of tensor structures.
In the expression above $S_d$ is the "volume" of the sphere in $d$-dimensions,
\begin{equation}
S_d = 2\,\pi^{\frac{d}{2}}/\Gamma(d/2) \, .
\end{equation}
Introducing (\ref{Ahat}) into (\ref{CompleteTOO}), we find the explicit expression
\begin{eqnarray}
\left\langle T_{\mu\nu}(x_1) \, O(x_2) \, O(x_3) \right\rangle
&=&
\frac{a}{(d-2)^2}\bigg\{
( \partial_{\mu}^{12} \partial_{\nu}^{31} + \partial_{\nu}^{12} \partial_{\mu}^{31} ) +
\frac{d-2}{d} ( \partial_{\mu \nu}^{12} + \partial_{\mu\nu}^{31} )
\bigg\} \frac{1}{(x^2_{12})^{d/2-1} (x^2_{23})^{\eta-d/2+1} (x^2_{31})^{d/2-1}} \nonumber \\
&+&
a\, \frac{ x^2_{12} x^2_{23} + x^2_{31} x^2_{23} -
(x^2_{23})^2}{(x^2_{12})^{d/2} (x^2_{23})^{\eta-d/2+1} (x^2_{31})^{d/2}} \frac{\delta_{\mu\nu}}{d}+
a\, \frac{\eta -d+1}{d\, \eta} \, S_d \, \delta_{\mu\nu} \, \frac{\delta^d(x_{12}) +
\delta^d(x_{31}) }{(x_{23}^2)^\eta}\, ,
\label{ExpCompleteTOO}
\end{eqnarray}
where, from now on, we set $\partial^{12}_\mu \equiv \frac{\partial}{\partial x_{12\,\mu}}$ and
$\partial^{12}_{\mu\nu} \equiv \frac{\partial}{\partial x_{12\,\mu}}\frac{\partial}{\partial x_{12\,\nu}}$.
Notice that the first term of the second line proportional to $\delta_{\mu\nu}$ is not manifestly integrable,
but one can use identities such as
$x_{12}^2+ x_{13}^2 - x_{23}^2=2 x_{12}\cdot x_{13}$ in order to rewrite it in the form
\begin{equation}
\frac{ x^2_{12} \, x^2_{23} + x^2_{31} \, x^2_{23} - (x_{23}^2)^2}{(x^2_{12})^{d/2} \, (x^2_{23})^{d/2}\,
(x^2_{31})^{d/2}}=
\frac{2 }{(d-2)^2} \partial^{12}_\mu \partial^{31 \,\mu}
\frac{1}{ (x_{12}^2)^{d/2-1} (x_{31}^2)^{d/2-1} (x_{23}^2)^{\eta-d/2+1}}
\end{equation}
which shows its integrability when $\eta < d-1$.
Now, in order to test the consistency of the result (\ref{Ch1TOOPO}) obtained from the application of the conformal Ward identities
for the $TOO$, we can consider a particular scalar free field theory.
We suppose for instance that the scalar operator $\mathcal O$ is given by $\mathcal O = \phi^2$ with dimensions $\eta = d-2$, whose EMT is
\begin{eqnarray}
T_{\mu\nu}
&=&
\partial_\mu \phi \, \partial_\nu\phi - \frac{1}{2} \, \delta_{\mu\nu}\,\partial_\alpha \phi \, \partial^\alpha \phi
+ \frac{1}{4}\,\frac{d-2}{d-1}\, \bigg[\delta_{\mu\nu} \partial^2 - \partial_\mu\,\partial_\nu\bigg]\, \phi^2
\end{eqnarray}
which is conserved and traceless in $d$ dimensions. \\
Using the Feynman rules in momentum space together with the expression of a scalar propagator,
after applying the inverse mapping procedure detailed in appendix \ref{Ch1InverseTTT},
we obtain the $T\phi^2\phi^2$ correlation function in $d$ dimensions
\begin{eqnarray}
\label{Ch1TphiphiDer}
\left\langle T_{\mu\nu}(x_1) \phi^2 (x_2) \phi^2 (x_3) \right\rangle
&=&
\frac{2 a (d-1)}{d (d-2)^2} \bigg[ \partial_{\mu}^{12}
\partial_{\nu}^{31} + \partial_{\nu}^{12} \partial_{\mu}^{31} - \delta_{\mu\nu} \partial^{12} \cdot \partial^{31} -
\frac{d-2}{2(d-1)} \bigg( - \partial_{\mu\nu}^{12} - \partial_{\mu\nu}^{31} + \partial_{\mu}^{12} \partial_{\nu}^{31}
\nonumber \\
&+&
\partial_{\nu}^{12} \partial_{\mu}^{31} + \delta_{\mu\nu} \left(\partial^2_{12} + \partial^2_{31} - 2 \partial^{12} \cdot
\partial^{31} \right) \bigg) \bigg] \frac{1}{(x_{12}^2)^{d/2-1} (x_{23}^2)^{d/2-1} (x_{31}^2)^{d/2-1}} \nonumber \\
&-&
a \frac{d-1}{d(d-2)} S_d \delta_{\mu\nu} \frac{\delta^d(x_{12}) + \delta^d(x_{31}) }{(x_{23}^2)^{d-2}} \,.
\end{eqnarray}
The equivalence of this expression with the solution given in (\ref{ExpCompleteTOO}) can be explicitly checked.
We remark that (\ref{Ch1TphiphiDer}) is clearly integrable and does not require any intermediate regularization.
The first term in the previous expression comes from the triangle topology diagram while the last two,
proportional to the delta functions, are contact terms with $2$-point topology (see eq. (\ref{DefineTOO})).
\subsection{The $VVV$ case}
The $VVV$ vertex function is pretty easy to handle with the inverse mapping procedure.
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.7]{figures/VVVferm.eps}\qquad\qquad
\includegraphics[scale=0.7]{figures/VVVscalar.eps}
\caption{The fermion and the scalar sectors contributing to the conformal VVV vertex in any dimension.}
\label{Ch1VVV}
\end{center}
\end{figure}
In \cite{Osborn:1993cr} the general CFT requirements fix the structure of the $VVV$ to be
\begin{eqnarray} \label{Ch1VVVcoord}
\left\langle V_{\mu}^a(x_1) V_{\nu}^b (x_2) V_{\rho}^c(x_3) \right\rangle
&=&
\frac{f^{abc}}{(x^2_{12})^{d/2-1}\,(x^2_{23})^{d/2-1}\,(x^2_{31})^{d/2-1}} \,
\bigg\{ (a - 2b) \, X_{23\,\mu}\,X_{31\,\nu}\,X_{12\,\rho} \nonumber \\
&-&
b \left[\frac{1}{x^2_{23}} \, X_{23\,\mu} \, I_{\nu\rho}(x_{23}) +
\frac{1}{x^2_{31}} \, X_{31\,\nu} \, I_{\mu\rho}(x_{31})+
\frac{1}{x^2_{12}} \, X_{12\,\rho}\,I_{\mu\nu}(x_{12} ) \right] \bigg\} \, ,
\end{eqnarray}
where $I_{\mu\nu}(x)$ is the inversion operator defined as
\begin{equation}
I^{\mu\nu}(x)=
\delta^{\mu\nu} - 2 \frac{x^\mu x^\nu}{x^2} \, .
\end{equation}
The correlator is Fourier-integrable, although this is not immediately evident from (\ref{Ch1VVVcoord}). The simplest way to prove this point
consists in showing that (\ref{Ch1VVVcoord}) can be reproduced in $d$-dimensions by the combination of the scalar and the fermion
sectors of a free field theory.
For this purpose we use two realizations of the vector current $V_{\mu}^a$, using scalar and fermion fields
\begin{eqnarray}
V_{\mu}^a = \phi^{*} t^a \left(\partial_{\mu} \phi \right) - \left( \partial_{\mu} \phi^{*}\right) t^a \phi \, ,
\qquad
V_{\mu}^a = \bar \psi \, t^a \gamma_{\mu} \psi \, .
\label{SAndF}
\end{eqnarray}
The diagrammatic expansion of this correlator consists of two triangle diagrams, the direct and the exchanged, both in the scalar and fermion
sectors. These two types of diagrams are shown in fig. \ref{Ch1VVV} and it is well known how to write down their expressions
in momentum space, using the corresponding Feynman rules. \\
Performing the inverse mapping procedure, we find the result
\begin{eqnarray}
\left\langle V_{\mu}^a(x_1) V_{\nu}^b (x_2) V_{\rho}^c(x_3) \right\rangle^{f} &=&
- \frac{ c_f \, f^{a b c} }{(d-2)^3} \Delta_{\mu \alpha \nu \beta \rho \gamma} \partial^{\alpha}_{12}\partial^{\beta}_{23}
\partial^{\gamma}_{31} \frac{1}{(x_{12}^2)^{d/2-1} (x_{23}^2)^{d/2-1} (x_{31}^2)^{d/2-1}} \, , \\
\left\langle V_{\mu}^a(x_1) V_{\nu}^b (x_2) V_{\rho}^c(x_3) \right\rangle^{s}
&=&
\frac{c_s \, f^{a b c}}{(d-2)^2} \left( \partial_{\mu}^{12} + \partial_{\mu}^{31} \right) \left( \partial_{\nu}^{23} +
\partial_{\nu}^{12} \right) \left( \partial_{\rho}^{31} +
\partial_{\rho}^{23} \right) \frac{1}{(x_{12}^2)^{d/2-1} (x_{23}^2)^{d/2-1} (x_{31}^2)^{d/2-1}} \nonumber \\
\end{eqnarray}
for the fermion and the scalar sector respectively, where we have introduced the operator
\begin{eqnarray}
\label{Ch1Delta6}
\Delta_{\mu \alpha \nu \beta \rho \gamma} = \frac{1}{4}\,
Tr \left[ \gamma_{\mu} \gamma_{\alpha} \gamma_{\nu} \gamma_{\beta} \gamma_{\rho} \gamma_{\gamma} \right] \, ,
\end{eqnarray}
and $c_f, c_s$ are normalization constants whose numerical values are irrelevant here.
Written in this form, with derivatives pulled out, the two expressions are manifestly integrable.
Tracing over the $\gamma$ matrices and applying the derivatives over all the denominators,
we generate the result of \cite{Osborn:1993cr} by taking a linear combination of these two sectors
\begin{eqnarray}
\left\langle V_{\mu}^a(x_1) V_{\nu}^b (x_2) V_{\rho}^c(x_3) \right\rangle =
\bigg( a \, t^a_{\mu\nu\rho} + b \, t^b_{\mu\nu\rho} \bigg)
\frac{ f^{abc}}{(x_{12}^2)^{d/2-1} (x_{23}^2)^{d/2-1} (x_{31}^2)^{d/2-1}} \, ,
\end{eqnarray}
where
\begin{eqnarray}
t^a_{\mu\nu\rho} &=& \frac{1}{d(d-2)^2} \left( \partial_{\mu}^{12} +
\partial_{\mu}^{31} \right) \left( \partial_{\nu}^{23} + \partial_{\nu}^{12} \right) \left( \partial_{\rho}^{31} +
\partial_{\rho}^{23} \right) - \frac{1}{d} t^b_{\mu\nu\rho} \, , \\
t^b_{\mu\nu\rho} &=& - \frac{1}{(d-2)^3}\,
\Delta_{\mu \alpha \nu \beta \rho \gamma}\partial^{\alpha}_{12}\partial^{\beta}_{23}\partial^{\gamma}_{31} \, .
\end{eqnarray}
We have explicitly checked the equivalence between this expression and eq. (\ref{Ch1VVVcoord}).
No additional term is required, in position space, to account for the general covariance Ward identity (\ref{DiffWardVVV}).
\subsection{The $TVV$ case}
The next correlator that we are going to discuss is the $TVV$, for which, together with the $TTT$ vertex,
the analysis required to confirm the correspondence between the position space solutions given in \cite{Osborn:1993cr}
and perturbative computations in momentum space is much more involved, due to the growth of the number of tensors structures.
We begin with the expression of the $TVV$ in position space at separate points, which is
\begin{equation}
\left\langle T_{\mu\nu}(x_1) V^a_\alpha(x_2) V^b_\beta(x_3) \right\rangle =
\frac{\delta^{ab}}{(x^2_{12})^{d/2} \, (x^2_{31})^{d/2} \,
(x^2_{23})^{d/2-1}} \, I_{\alpha\sigma}(x_{12}) \, I_{\beta\rho}(x_{31}) \, t_{\mu\nu\rho\sigma}(X_{23}) \, ,
\label{TVVSepCoord}
\end{equation}
where the structure $t_{\mu\nu\rho\sigma}(X)$ is given by the following combination of $4$-indices structures,
which are traceless with respect to $(\mu,\nu)$,
\begin{equation}
t_{\mu\nu\rho\sigma}(X) =
a\, h^1_{\mu\nu}(\hat X)\, \delta_{\rho\sigma} +
b\, h^{1}_{\mu\nu}(\hat X)\,h^{1}_{\rho\sigma}(X) +
c\, h^{2}_{\mu\nu\rho\sigma}(\hat X) +
e\, h^{3}_{\mu\nu\rho\sigma}\, ,
\end{equation}
with $h^1_{\mu\nu}$ already introduced in (\ref{Ch1TOOstructures}), whereas
\begin{eqnarray}
h^{2}_{\mu\nu\rho\sigma}(\hat X)
&=&
\hat{X}_\mu\, \hat{X}_\rho\, \delta_{\nu\sigma} +
\hat{X}_\nu\, \hat{X}_\sigma\, \delta_{\mu\sigma} +
\left( \rho \leftrightarrow \sigma \right) -
\frac{4}{d}\, \hat{X}_\mu\, \hat{X}_\nu\, \delta_{\rho\sigma} -
\frac{4}{d}\, \hat{X}_\rho\, \hat{X}_\sigma\, \delta_{\mu\nu} +
\frac{4}{d^2}\, \delta_{\mu\nu}\, \delta_{\rho\sigma} \, ,
\nonumber \\
h^{3}_{\mu\nu\rho\sigma}
&=&
\delta_{\mu\rho}\delta_{\nu\sigma} + \delta_{\mu\sigma}\delta_{\nu\rho} -
\frac{2}{d}\, \delta_{\mu\nu}\delta_{\rho\sigma} \, .
\end{eqnarray}
On the other hand, if one considers the Ward identity (\ref{TVVWardCoord}) at separate points, i.e. with vanishing r.h.s. ,
the four coefficients $a$,$b$,$c$ and $e$ are found to be constrained by
\begin{equation}
d\,a -2\,b +2\,(d-2)\,c = 0 \, , \quad b= d\,(d-2)\,e\, ,
\label{ConstraintsTVV}
\end{equation}
so that there are two independent contributions to the $TVV$ vertex for general dimensions.
As usual, the next step in the analysis of \cite{Osborn:1993cr} is to study the $2$-point coincidence limits
$x_1\rightarrow x_2$, $x_1\rightarrow x_3$ in order to identify the terms which are responsible
for the r.h.s. of the Ward identity (\ref{TVVWardCoord}).
Unlike the case of the $VVV$ and similarly to the case of the $TOO$, to which we refer for the details concerning the regularization procedure
of the short-distance singularities, some of the terms appearing in this limit on the r.h.s. of (\ref{TVVWardCoord})
are found to need regularization. Then differential regularization is used to pull out derivatives
and find some regularization-dependent terms proportional to $\delta$-functions.
The solution obtained is connected to the normalization constant of the vector current (unrenormalized) $2$-point function, which is
\begin{equation}
\left\langle V^{a}_\mu (x_1)V^{b}_\nu (x_2) \right\rangle =
C_V \, \frac{I_{\mu\nu}(x_{12})}{(x^2_{12})^{d-1}} \, .
\label{VVPoistion}
\end{equation}
so that the complete, unrenormalized $TVV$ correlator is given by
\begin{eqnarray}
\left\langle T_{\mu\nu}(x_1) V^a_\alpha(x_2) V^b_\beta(x_3) \right\rangle
&=&
\frac{\delta^{ab}}{(x^2_{12})^{d/2} \, (x^2_{31})^{d/2} \,
(x^2_{23})^{d/2-1}} \, I_{\alpha\sigma}(x_{12}) \, I_{\beta\rho}(x_{31}) \, t_{\mu\nu\rho\sigma}(X_{23}) \nonumber \\
&&
+ \delta^{ab}\,C_V\, \left[ \hat A_{\mu\nu\alpha\rho}(x_{12}) - A_{\mu\nu\alpha\rho}(x_{12}) \right]
\frac{I_{\rho\beta}(x_{23})}{(x_{23}^2)^{d-1}} \nonumber\\
&&
+ \delta^{ab}\,C_V \, \left[ \hat A_{\mu\nu\sigma\beta}(x_{31}) - A_{\mu\nu\sigma\beta}(x_{31}) \right]
\frac{I_{\sigma\alpha}(x_{23})}{(x_{23}^2)^{d-1}}\, .
\label{Ch1TVVcoord}
\end{eqnarray}
Here, the structures $A$ and $\hat{A}$ are respectively
\begin{eqnarray}
C_V\, A_{\mu\nu\rho\sigma}(s)
&=&
\frac{1}{s^d}\, I_{\rho\alpha}(s)\, t_{\mu\nu\alpha\sigma}(s) \, , \nonumber \\
C_V\, \hat{A}_{\mu\nu\rho\sigma}(s)
&=&
\bigg[
\frac{2e}{d(d-2)}\delta_{\mu\nu}\partial_\rho\partial_\sigma - \frac{2c}{d(d-2)}\,\delta_{\rho\sigma} \partial_\mu\partial_\nu
- \frac{c+d e}{d(d-2)}\left(\delta_{\nu\rho}\partial_\mu\partial_\sigma + \delta_{\mu\rho}\partial_\nu\partial_\sigma \right) \nonumber \\
&&
+\, \frac{c-(d-2)e}{d(d-2)}\left( \delta_{\nu\sigma}\partial_\mu\partial_\rho + \delta_{\mu\sigma}\partial_\nu\partial_\rho \right)
\bigg]\, \frac{1}{s^{d-2}}
-\, \frac{e}{d(d-4)}\,\partial_\mu\partial_\nu\partial_\rho\partial_\sigma \frac{1}{s^{d-4}} \nonumber \\
&&
+\, \frac{1}{d}\, \bigg[ 2e\, \delta_{\mu\nu}\delta_{\rho\sigma}- \left(c + d e\right)
\left(\delta_{\mu\sigma}\delta_{\nu\rho}+\delta_{\mu\rho}\delta_{\nu\sigma}\right) \bigg]\,S_d\, \delta^{(d)}(s)\, .
\label{hatATVV}
\end{eqnarray}
Again, $\hat{A}_{\mu\nu\rho\sigma}(s) = {A}_{\mu\nu\rho\sigma}(s) $ for $s\neq 0$ and the coefficients in front
of the $\delta$-function terms in the last line of (\ref{hatATVV}) are determined only after imposing the Ward identities.
Now we are ready to check the correspondence between the complete expression of the unrenormalized $TVV$ in position space,
(\ref{Ch1TVVcoord}), and the inverse-mapped momentum space $1$-loop computations.
We have drawn the diagrammatic structure of the $TVV$ in fig. \ref{Ch1TVVfigure}.
Using the information that the most general $TVV$ is parametrized by just two independent constants, we conclude that,
in any dimension, it can be fully constructed as a linear combination of two contributions coming from independent free theories,
the fermion and the scalar. \\
Therefore we can write
\begin{eqnarray}
\left\langle T_{\mu\nu}(x_1) V^a_{\alpha} (x_2) V^b_{\beta} (x_3) \right\rangle
&=&
\sum_{I=s,f}n_I\, \bigg( \left\langle T_{\mu\nu}(x_1) V^a_\alpha(x_2) V^b_\beta(x_3) \right\rangle^I_{A=0} \nonumber \\
&&
+\, \left\langle \frac{\delta T_{\mu\nu}(x_1)}{\delta A^{a\,\alpha}(x_2)} V^b_\beta(x_3) \right\rangle^I_{A= 0}
+ \left\langle \frac{\delta T_{\mu\nu}(x_1)}{\delta A^{b \,\beta}(x_3)} V^a_\alpha(x_2) \right\rangle^I_{A= 0} \bigg)
\label{Ch1contactTVV}
\end{eqnarray}
where the sum is over the same scalar (s) and fermion (f) sectors introduced for the $VVV$ and $n_I$ stands for the number
of corresponding fields.
For the diagrammatic interpretation of the various contributions to this correlator,
(except for the counterterm, which will be addressed in the next section),
among the terms above, the first one corresponds to the triangle topology, while the remaining two are the two bubbles
(see fig. \ref{Ch1TVVfigure}).
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.7]{figures/triangleTVV.eps}\qquad
\includegraphics[scale=0.7]{figures/Tbubble1TVV.eps}\qquad
\includegraphics[scale=0.7]{figures/Tbubble2TVV.eps}
\caption{The fermion/scalar sectors in the TVV vertex.}
\label{Ch1TVVfigure}
\end{center}
\end{figure}
Using the Feynman rules in momentum space and going through the inverse-mapping procedure,
one can obtain the following parametrization of the triangle contribution to the $TVV$ vertex for scalars within the loop,
\begin{eqnarray}
\left\langle T_{\mu\nu}(x_1) V^a_{\alpha} (x_2) V^b_{\beta} (x_3) \right\rangle_{A=0}^{s} =
\nonumber
c \, \delta^{ab} \frac{2(d-1)}{d (d-2)^3} \bigg[ \partial_{\mu}^{12} \partial_{\nu}^{31} +
\partial_{\nu}^{12} \partial_{\mu}^{31} - \delta_{\mu\nu} \partial^{12} \cdot \partial^{31}
\nonumber \\
- \frac{d-2}{2(d-1)} \bigg( - \partial_{\mu\nu}^{12} - \partial_{\mu\nu}^{31}
+ \partial_{\mu}^{12} \partial_{\nu}^{31} + \partial_{\nu}^{12} \partial_{\mu}^{31} +
\delta_{\mu\nu} \left(\partial^2_{12} + \partial^2_{31} - 2 \partial^{12} \cdot \partial^{31} \right) \bigg) \bigg] \times
\nonumber \\
\times \left( \partial_{\alpha}^{12} + \partial_{\alpha}^{23}\right)
\left( \partial_{\beta}^{31} + \partial_{\beta}^{23}\right)
\frac{1}{(x_{12}^2)^{d/2-1} (x_{23}^2)^{d/2-1} (x_{31}^2)^{d/2-1}}\, .
\label{Ch1scalarsector}
\end{eqnarray}
whereas for fermions we have
\begin{eqnarray}
\left\langle T_{\mu\nu}(x_1) V^a_{\alpha} (x_2) V^b_{\beta} (x_3) \right\rangle_{A=0}^{f} =
\frac{c \, \delta^{ab}}{d(d-2)^3} \, A_{\mu\nu\xi\eta}\, \Delta_{\xi\rho\alpha\sigma\beta\lambda}\,
(\partial_{\eta}^{12} + \partial_{\eta}^{31} ) \, \partial^{\rho}_{12} \partial^{\sigma}_{23} \partial^{\lambda}_{31}
\nonumber\\
\times \frac{1}{(x_{12}^2)^{d/2-1} (x_{23}^2)^{d/2-1} (x_{31}^2)^{d/2-1}} \, ,
\label{Ch1fermionsector}
\end{eqnarray}
where $ \Delta_{\mu\rho\alpha\sigma\beta\lambda}$ is defined in eq. (\ref{Ch1Delta6})
and $A_{\mu\nu\rho\sigma}$ in appendix \ref{Ch1Vertices}. \\
In (\ref{Ch1scalarsector})-(\ref{Ch1fermionsector}) $c$ is a normalization constant and
these terms can be seen to exactly correspond to the expression (\ref{TVVSepCoord}), holding for separate points, if one sets
\begin{eqnarray}
&&
e = \frac{c}{d-2} \, , \quad c = \frac{1}{S_d^3}\, \frac{d}{2(d-1)}\, ,
\nonumber \\
&&
e=0 \, , \quad \hspace{7mm} c = \frac{1}{S_d^3}\, \frac{d\,2^{\frac{d}{2}}}{2}\, ,
\end{eqnarray}
in the scalar and fermion sector respectively, the values of the other coefficients following from (\ref{ConstraintsTVV}).
The only subtle difference to notice is in the scalar sector, where the $\partial^2_{12}$ and $\partial^2_{31}$ terms,
which are proportional to $\delta_{\mu\nu}$, vanish in the non-coincident point limit and their sum is given by
\begin{eqnarray}
&&
- \frac{c \, \delta^{ab} }{d (d-2)^2} \delta_{\mu\nu} \left(\partial^2_{12} + \partial^2_{31} \right) \left(
\partial_{\alpha}^{12} + \partial_{\alpha}^{23}\right) \left( \partial_{\beta}^{31} + \partial_{\beta}^{23}\right)
\frac{1}{(x_{12}^2)^{d/2-1} (x_{23}^2)^{d/2-1} (x_{31}^2)^{d/2-1}} \nonumber \\
&&
=
\frac{2 c \, \delta^{ab} }{d (d-2)} S_d \delta_{\mu\nu} \bigg[\partial_{\alpha}^{23} \left( \partial_{\beta}^{31} +
\partial_{\beta}^{23}\right) \frac{\delta^d(x_{12}) }{ (x_{23}^2)^{d/2-1} (x_{31}^2)^{d/2-1}} +
\partial_{\beta}^{23} \left( \partial_{\alpha}^{12}
+ \partial_{\alpha}^{23}\right) \frac{ \delta^d(x_{31})}{(x_{12}^2)^{d/2-1} (x_{23}^2)^{d/2-1} } \bigg]\,.
\label{Ch1top2a}
\end{eqnarray}
They have the topology of $2$-point functions and must be carefully summed to those arising from the bubble diagrams
in order to reproduce exactly the contributions identified as $\hat A - A$ in (\ref{Ch1TVVcoord}).
The bubble contributions are
\begin{eqnarray}
\left\langle \frac{\delta T_{\mu\nu}(x_1)}{\delta A^{a\,\alpha}(x_2)} V^b_\beta(x_3) \right\rangle_{A= 0}^{s}
&=&
\frac{c \, \delta^{ab} (d-1)}{d (d-2)^2} S_d \delta^d(x_{12})
\left((\partial^{23}_{\mu} + \partial^{31}_{\mu} ) \delta_{\nu\alpha} +
(\partial^{23}_{\nu} + \partial^{31}_{\nu} ) \delta_{\mu \alpha}
- \delta_{\mu \nu} (\partial^{23}_{\alpha} + \partial^{31}_{\alpha}) \right)\times \nonumber \\
&&
\times (\partial^{23}_{\beta} + \partial^{31}_{\beta} )
\frac{1}{(x_{31}^2)^{d/2-1} (x_{23}^2)^{d/2-1}} \label{Ch1top2b} \, , \\
\left\langle \frac{\delta T_{\mu\nu}(x_1)}{\delta A^{b \,\beta}(x_3)} V^a_\alpha(x_2) \right\rangle_{A= 0}^{s}
&=&
\frac{c \, \delta^{ab} (d-1)}{d (d-2)^2} S_d \delta^d(x_{31})
\left((\partial^{23}_{\mu} + \partial^{12}_{\mu} ) \delta_{\nu \beta} + (\partial^{23}_{\nu} + \partial^{12}_{\nu} )
\delta_{\mu \beta} - \delta_{\mu \nu} (\partial^{23}_{\alpha} + \partial^{12}_{\beta}) \right)\times
\nonumber \\
&&
\times (\partial^{23}_{\alpha} + \partial^{12}_{\alpha} )
\frac{1}{(x_{12}^2)^{d/2-1} (x_{23}^2)^{d/2-1}} \label{Ch1top2c}\, ,
\end{eqnarray}
and we thoroughly checked that the sum of (\ref{Ch1top2a}), (\ref{Ch1top2b}) and (\ref{Ch1top2c}) corresponds
to the solution of the Ward identities (\ref{Ch1TVVcoord}) for $e=c/(d-2)$ and applying (\ref{ConstraintsTVV}). \\
Similarly, the contact terms for the fermion sector in the diagrammatic expansion are found to be, after the inverse mapping,
\begin{eqnarray}
\left\langle \frac{\delta T_{\mu\nu}(x_1)}{\delta A^{a\,\alpha}(x_2)} V^b_\beta(x_3) \right\rangle_{A= 0}^{f}
&= &
\frac{c \, \delta^{ab}}{d(d-2)^2} S_d \delta^d(x_{12})
\Delta^{(2)}_{\mu \nu \alpha \beta \rho \sigma} \partial^{\rho}_{31}
\frac{1}{(x_{31}^2)^{d/2-1}} \partial^{\sigma}_{31} \frac{1}{(x_{31}^2)^{d/2-1}} \, , \\
\left\langle \frac{\delta T_{\mu\nu}(x_1)}{\delta A^{b \,\beta}(x_3)} V^a_\alpha(x_2) \right\rangle_{A= 0}^{f}
&=&
\frac{c \, \delta^{ab}}{d(d-2)^2} S_d \delta^d(x_{31})
\Delta^{(2)}_{\mu \nu \beta \alpha \rho \sigma} \partial^{\rho}_{12}
\frac{1}{(x_{12}^2)^{d/2-1}} \partial^{\sigma}_{31} \frac{1}{(x_{12}^2)^{d/2-1}}\, ,
\end{eqnarray}
with $\Delta^{(2)}$ defined by
\begin{eqnarray}
\Delta^{(2)}_{\mu \nu \alpha \beta \rho \sigma}
&=&
\delta_{\alpha\nu } \delta_{\beta\sigma}\delta_{\mu\rho }
+ \delta_{\alpha\mu } \delta_{\beta\sigma}\delta_{\nu\rho }
+ \delta_{\alpha\nu } \delta_{\beta\rho} \delta_{\mu\sigma }
+ \delta_{\alpha\mu } \delta_{\beta\rho} \delta_{\nu\sigma }
- \delta_{\alpha\nu } \delta_{\beta\mu} \delta_{\rho\sigma }
- \delta_{\alpha\mu } \delta_{\beta\nu} \delta_{\rho\sigma } \nonumber \\
&-&
2 \, \delta_{\mu\nu} \, \left(\delta_{\alpha\rho}\delta_{\beta\sigma} + \delta_{\alpha\sigma}\delta_{\beta\rho}
- \delta_{\alpha\beta}\delta_{\rho\sigma} \right)\, .
\end{eqnarray}
These two contributions exactly match the $\hat A - A$ terms in (\ref{Ch1TVVcoord}) for $e=0$ and applying (\ref{ConstraintsTVV}),
so that we can conclude that our check for the agreement between the position space solution and the momentum space
perturbative computation is successful also for the $TVV$. \\
To conclude with the $TVV$, we notice that the Green functions discussed so far are unrenormalized.
The issue of renormalization will be addressed separately in section \ref{Counterterms}.
\subsection{The $TTT$ case}\label{Ch1InverseMappingTTT}
Now we are ready to turn to the analysis of the $3$-graviton vertex, whose perturbative computation was presented in chapter
\ref{TTTVertex}.
The general structure of the $TTT$ correlator in position space for separate points is \cite{Osborn:1993cr}
\begin{equation} \label{Ch1bareTTT}
\left\langle T^{\mu\nu}(x_1) \, T^{\rho\sigma}(x_2) \, T^{\alpha\beta}(x_3) \right\rangle =
\frac{1}{(x^2_{12})^{d/2} \, (x^2_{23})^{d/2} \, (x^2_{31})^{d/2}} \,
\mathcal{I}^{\mu\nu,\mu'\nu'}\, \mathcal{I}^{\rho\sigma,\rho'\sigma'} \, t^{\mu'\nu'\rho'\sigma'\alpha\beta}(X_{12}) \, ,
\end{equation}
\begin{equation} \label{Ch1Inversion}
\mathcal{I}^{\mu \nu,\alpha \beta} (s) =
I^{\mu\rho}(s)I^{\nu\sigma}(s) \epsilon_{T}^{\rho\sigma,\alpha\beta} \, ,
\quad s= x - y \, ,
\end{equation}
where the tensor
\begin{equation}\label{Ch1epislon}
\epsilon_{T}^{\mu\nu,\alpha\beta} =
\frac{1}{2} \, (\delta^{\mu\alpha} \delta^{\nu \beta} + \delta^{\mu \beta} \delta^{\nu \alpha} \bigl)
- \frac{1}{d} \, \delta^{\mu \nu} \delta^{\alpha \beta}
\end{equation}
is the projector onto the space of symmetric traceless tensors. \\
If we introduce
\begin{eqnarray}
h^{4\,\mu\nu\rho\sigma\alpha\beta}(X)
&=&
h^{3\,\mu\nu\alpha\rho} \hat{X}^{\sigma}\hat{X}^{\beta} + h^{3\,\mu\nu\alpha\sigma} \hat{X}^{\rho}\hat{X}^{\beta}
+ (\alpha \leftrightarrow \beta) \nonumber \\
&&
-\, \frac{2}{d}\, \delta^{\rho\sigma}\, h^{2\,\mu\nu\alpha\beta}(\hat{X})
- \frac{2}{d}\, \delta^{\alpha\beta}\, h^{2\,\mu\nu\rho\sigma}(\hat{X})
- \frac{8}{d^2}\, \delta^{\rho\sigma}\, \delta^{\alpha\beta}\, h^{1\,\mu\nu}(\hat{X})\, , \nonumber \\
h^{5\,\mu\nu\rho\sigma\alpha\beta}(\hat{X})
&=&
\bigg[ \left( \delta^{\mu\rho}\delta^{\nu\alpha}\delta^{\rho\beta} + \delta^{\nu\rho}\delta^{\mu\alpha}\delta^{\rho\beta} +
\left( \rho \leftrightarrow \sigma \right) \bigg] + \bigg[ \alpha \leftrightarrow \beta \right]
\nonumber \\
&&
-\, \frac{4}{d}\,\delta^{\mu\nu}\,h^{3\,\rho\sigma\alpha\beta} -
\frac{4}{d}\,\delta^{\rho\sigma}\,h^{3\,\mu\nu\alpha\beta} -
\frac{4}{d}\,\delta^{\alpha\beta}\,h^{3\,\mu\nu\rho\sigma} -
\frac{8}{d^2}\, \delta^{\mu\nu}\delta^{\rho\sigma}\delta^{\alpha\beta}\, ,
\end{eqnarray}
then the rank-$6$ tensor $t^{\mu\nu\rho\sigma\alpha\beta}(X)$ is written as
\begin{eqnarray}
t^{\mu\nu\rho\sigma\alpha\beta}(X)
&=& a\, h^{5\,\mu\nu\rho\sigma\alpha\beta} + b\, h^{4\,\alpha\beta\mu\nu\rho\sigma}(\hat X) +
b'\, \bigg(h^{4\, \mu\nu\rho\sigma\alpha\beta}(\hat X) + h^{4\,\rho\sigma\mu\nu\alpha\beta}(\hat X)\bigg)
\nonumber \\
&&
+\, c\, h^{3\,\mu\nu\rho\sigma}h^{1\,\alpha\beta}(\hat X) +
c' \, \bigg (h^{3\, \rho\sigma\alpha\beta}h^{1\,\mu\nu}(\hat X) +
h^{3\, \mu\nu\alpha\beta}h^{1\,\rho\sigma}(\hat X) \bigg) \nonumber \\
&&
+\, e\, h^{2\,\mu\nu\rho\sigma}(\hat X)h^{1\,\alpha\beta}(\hat X) + e'\, \bigg(
h^{2\, \rho\sigma\alpha\beta}(\hat X)h^{1\,\mu\nu}(\hat X) + h^{2\, \mu\nu\alpha\beta}(\hat X)h^{1\, \rho\sigma}(\hat X)\bigg)
\nonumber \\
&&
+\, f \, h^{1\,\mu\nu}(\hat X)\, h^{1\,\rho\sigma}(\hat X)\, h^{1\,\alpha\beta}(\hat X) \, ,
\end{eqnarray}
with the constraints
\begin{eqnarray}
&&
b+b' = -2\,a\, , \quad c' = c\, , \quad e + e' = -4\,b' - 2\,c \nonumber \\
&&
d^2\,a + 2\,(b+b') - (d-2)\,b' - d\,c + e' = 0 \, , \nonumber \\
&&
d\,(d+2)(2\,b'+c)+4\,(e+e') + f = 0\, ,
\end{eqnarray}
leaving only three unconstrained coefficients, say $a$, $b$ and $c$.
The $TTT$ needs regularization in the coincidence limits too, so we introduce the $2$-point function for the energy-momentum tensor,
\begin{equation}
\left\langle T^{\mu\nu}(x_1)T^{\alpha\beta}(x_2) \right\rangle =
C_T\, \frac{\mathcal{I}^{\mu\nu\alpha\beta}(x_{12})}{(x_{12}^2)^{d/2}}\, .
\label{TTPoistion}
\end{equation}
Following arguments similar to those for the $TOO$ and $TVV$ vertices, the authors in \cite{Osborn:1993cr} arrive at the
unrenormalized expression
\begin{eqnarray}
\left\langle T^{\mu\nu}(x_1) \, T^{\rho\sigma}(x_2) \, T^{\alpha\beta}(x_3) \right\rangle
&=&
\frac{1}{(x^2_{12})^{d/2} \, (x^2_{23})^{d/2} \, (x^2_{31})^{d/2}} \,
\mathcal{I}^{\mu\nu,\mu'\nu'}\, \mathcal{I}^{\rho\sigma,\rho'\sigma'} \, t^{\mu'\nu'\rho'\sigma'\alpha\beta}(X_{12}) \nonumber \\
&& \hspace{-5mm}
+\, \frac{C_T}{2}\, \bigg[
\hat{A}^{\mu\nu\rho\sigma\gamma\delta}(x_{12}) - A^{\mu\nu\rho\sigma\gamma\delta}(x_{12}) \bigg]\,
\bigg( \frac{\mathcal{I}^{\gamma\delta\alpha\beta}(x_{23})}{(x^2_{23})^{d/2}} +
\frac{\mathcal{I}^{\gamma\delta\alpha\beta}(x_{13})}{(x^2_{13})^{d/2}} \bigg) \nonumber \\
&& \hspace{-5mm}
+\, \frac{C_T}{2}\, \bigg[
\hat{A}^{\mu\nu\alpha\beta\gamma\delta}(x_{13}) - A^{\mu\nu\alpha\beta\gamma\delta}(x_{13}) \bigg]\,
\bigg( \frac{\mathcal{I}^{\gamma\delta\rho\sigma}(x_{32})}{(x^2_{32})^{d/2}} +
\frac{\mathcal{I}^{\gamma\delta\rho\sigma}(x_{12})}{(x^2_{12})^{d/2}} \bigg) \nonumber \\
&& \hspace{-5mm}
+\, \frac{C_T}{2}\, \bigg[
\hat{A}^{\rho\sigma\alpha\beta\gamma\delta}(x_{23}) - A^{\rho\sigma\alpha\beta\gamma\delta}(x_{23}) \bigg]\,
\bigg( \frac{\mathcal{I}^{\gamma\delta\mu\nu}(x_{31})}{(x^2_{31})^{d/2}} +
\frac{\mathcal{I}^{\gamma\delta\mu\nu}(x_{21})}{(x^2_{21})^{d/2}} \bigg)\, ,
\label{HatATTT}
\end{eqnarray}
where the $A$ tensor is found to be, by computing the short distance limit of (\ref{Ch1bareTTT}),
\begin{equation}
A^{\mu\nu\rho\sigma\alpha\beta}(s) = \frac{1}{(s^2)^{d/2}}\, t^{\mu\nu\rho\sigma\alpha\beta}(s)\, ,
\label{UnregATTT}
\end{equation}
whereas its regularized counterpart is given by
\begin{eqnarray}
\hat{A}^{\mu\nu\rho\sigma\alpha\beta}(s)
&=&
\mathcal{D.\,R.}\left[A^{\mu\nu\rho\sigma\alpha\beta}(s)\right] +
\bigg( C\, h^{5\, \mu\nu\rho\sigma\alpha\beta}(s) + D\, \left( \delta^{\mu\nu} \, h^{3\,\rho\sigma\alpha\beta} +
\delta^{\rho\sigma}\, h^{3\,\mu\nu\alpha\beta} \right) \bigg)\, S_d\, \delta^{(d)}(s) \, ,
\end{eqnarray}
where the short-hand notation on the r.h.s. stands for the differentially regularized version of $A$, obtained, just as in the
case of the $TOO$ and the $TVV$ vertices, by re-expressing it as a (lengthy) combination of derivatives of powers of $s^2$
which are lower than $d$ (compare with eq. (\ref{UnregATTT})) which is no point reporting here explicitly
and can be found in eqs. (6.37)-(6.38) of \cite{Osborn:1993cr}. \\
Next comes the imposition of the Ward identities: by requiring the general covariance (\ref{Ch1WI3PFcoordinate}) and the non anomalous
trace (\ref{NoAnTraceTTT}) constraints to be satisfied, it is found that the values of the so far arbitrary coefficients $C$ and $D$ are
\begin{equation}
C = \frac{(d-2)(2\,a+b)-d\,c}{d\,(d+2)} \, , \quad D = \frac{C_T}{d} \, .
\end{equation}
As the general solution of the Ward identities, for any CFT, is parametrized by $3$ independent constants,
we conclude that computing the correlator in the $3$ independent free field theories considered in chapter \ref{TTTVertex}
is enough to account for the complete unrenormalized result in $4$ dimensions, whereas for $d\neq 4$ the spin $1$ sector is not conformally
invariant and we cannot build the general expression just by superposing the scalar and the fermion sectors.
However, the combination of the scalar and fermion sectors corresponds to a special solution for $d=3$,
where the $TTT$ is parametrized just by $2$ independent constants, whereas in $d=2$ there is just $1$ such constant
\cite{Osborn:1993cr, Erdmenger:1996yc}.
As done before for the $TOO$, $VVV$ and $TVV$ correlators, here we check the result (\ref{Ch1bareTTT}) building explicitly
the correlator in position space anti-transforming the diagrammatic expansion in free field theory.
This allows to come up with an expression for this vertex which is manifestly integrable.
We will be using the Feynman rules obtained from the Lagrangian descriptions for scalars, fermions and spin $1$
in configuration space, given in section \ref{Ch1lags}.
We start by testing the non-coincident case, for which we can omit the contact terms.
This corresponds only to the diagrams with triangle topology.
We give the expressions in $d$ dimensions for the scalar and the fermion cases, while - as already remarked -
we have to limit our analysis to $d=4$ for the spin-$1$ vector boson. Moreover, in the vector case the gauge-fixing and ghost parts
of the amplitude have to cancel since the vertex is gauge invariant, as explained in section \ref{Ch1lags}, so that, performing our inverse
mapping, we include in the interactions vertices for the vector field only the Maxwell contributions.
We have
\begin{eqnarray}
\left\langle \frac{\delta\mathcal S}{\delta g_{\mu\nu}(x_1)}\,\frac{\delta\mathcal S}{\delta g_{\rho\sigma}(x_2)}
\frac{\delta\mathcal S}{\delta g_{\alpha\beta}(x_3)} \right\rangle^{s}
&=&
C^{s}_{TTT} V_{\mathcal{S}\phi\phi}^{\mu\nu}(i\,\partial^{12},-i\,\partial^{31})\,
V_{\mathcal{S}\phi\phi}^{\rho\sigma}(i\,\partial^{23},-i\,\partial^{12})\,
V_{\mathcal{S}\phi\phi}^{\alpha\beta}(i\,\partial^{31},-i\,\partial^{23})
\nonumber\\
&\times&
\frac{1}{(x_{12}^2)^{d/2-1}\,(x_{23}^2)^{d/2-1}\,(x_{31}^2)^{d/2-1}}
\, , \label{Ch1ScalarTriangle}\\
\left\langle \frac{\delta\mathcal S}{\delta g_{\mu\nu}(x_1)}\,\frac{\delta\mathcal S}{\delta g_{\rho\sigma}(x_2)}
\frac{\delta\mathcal S}{\delta g_{\alpha\beta}(x_3)} \right\rangle^{f}
&=&\nonumber
\end{eqnarray}
\begin{eqnarray}
C^{f}_{TTT}\,(-1)\,\,\bigg( \textrm{Tr}\big[V_{\mathcal{S}\bar\psi \psi}^{\mu\nu}(i\,\partial^{12},-i\,\partial^{31})\,
\,i\,\gamma\cdot\partial^{12}\,V_{\mathcal{S}\bar\psi \psi}^{\rho\sigma}(i\,\partial^{23},-i\,\partial^{12})\,
\,i\,\gamma\cdot\partial^{23}\,V_{\mathcal{S}\bar\psi \psi}^{\alpha\beta}(i\,\partial^{31},-i\,\partial^{23})\,
i\,\gamma\cdot\partial^{31}\big]
&&\,
\nonumber\\
\hspace{-10mm}
+\textrm{Tr}\big[ V_{\mathcal{S}\bar\psi \psi}^{\mu\nu}(i\,\partial^{31},-i\,\partial^{12})\,
i\,\gamma\cdot\,\partial^{31}\,V_{\mathcal{S}\bar\psi \psi}^{\alpha\beta}(i\,\partial^{23},-i\,\partial^{31})\,
i\,\gamma\cdot\,\partial^{12}\,V_{\mathcal{S}\bar\psi \psi}^{\rho\sigma}(i\,\partial^{12},-i\,\partial^{23})\,
i\,\gamma\cdot\partial^{12}\big]\bigg)\,
\nonumber\\
\times
\frac{1}{(x_{12}^2)^{d/2-1}\,(x_{23}^2)^{d/2-1}\,(x_{31}^2)^{d/2-1}} \, , &&
\nonumber \\
\label{Ch1FermionTriangle}
\end{eqnarray}
\begin{eqnarray}
\left\langle \frac{\delta\mathcal S}{\delta g_{\mu\nu}(x_1)}\,\frac{\delta\mathcal S}{\delta g_{\rho\sigma}(x_2)}
\frac{\delta\mathcal S}{\delta g_{\alpha\beta}(x_3)} \right\rangle^V
&=&
C^{V}_{TTT}\,(-1)^3\,
V_{\mathcal{S}AA}^{\mu\nu\gamma\delta}(i\,\partial^{12},-i\,\partial^{31})\,
V_{\mathcal{S}AA}^{\rho\sigma\zeta\xi}(i\,\partial^{23},-i\,\partial^{12})\,
V_{\mathcal{S}AA}^{\alpha\beta\chi\omega}(i\,\partial^{31},-i\,\partial^{23})\,
\nonumber\\
&\times&
\frac{\delta_{\gamma\xi}\,\delta_{\delta\chi}\delta_{\zeta\omega}}{x_{12}^2\,x_{23}^2\,x_{31}^2}
\bigg|_{\frac{1}{\xi}\rightarrow 0}\, .
\label{Ch1VectorTriangle}
\end{eqnarray}
Due to the complexity of the expressions, we have chosen an implicit notation in which the dependences of the vertices on the coordinates
are obtained by replacing the momenta of the vertices in appendix \ref{Ch1Vertices} with appropriate derivatives with respect to the external
position variables. For instance
\begin{equation}
V^{\mu\nu}_{\mathcal{S}\phi\phi}(p,q)\to V^{\mu\nu}_{\mathcal{S}\phi\phi}(\hat{p},\hat{q})=
V^{\mu\nu}_{\mathcal{S}\phi\phi}(i\, \partial^{12},- i\, \partial^{23}) \, ,
\end{equation}
with
\begin{equation}
\hat{p}\to i\, \partial^{12} \qquad\qquad \hat{q}\to - i\, \partial^{23} \, .
\end{equation}
Explicitly
\begin{equation}
V^{\mu\nu}_{\mathcal{S}\phi\phi}(i\, \partial^{12},- i\, \partial^{23}) =
\frac{1}{2}\,(i\,\partial_{12\,\alpha}) \, (- i\,\partial_{23\,\beta}) \, C^{\mu\nu\alpha\beta} +
\chi \bigg( \delta^{\mu\nu} \left( i\,\partial_{12} - i\,\partial_{23} \right)^2
- \left( i\,\partial_{12}^{\mu} - i\,\partial_{23}^{\mu}\right)\,
\left( i\,\partial_{12}^{\nu} - i\,\partial_{23}^{\nu} \right) \bigg) \, .
\end{equation}
The replacements of $p,q$ and $l$ by the operator expressions $\hat{p},\hat{q}$ and $\hat{l}$ are specific for each vertex.
In appendix \ref{Ch1InverseTTT} we provide some more details on this procedure. Notice that we have chosen the coupling parameter
for the scalar field in $d$ dimensions at the corresponding conformal value $\chi = (d-2)/4(d-1)$.
Expanding the derivatives contained in each vertex, the expression given in (\ref{Ch1bareTTT}) is recovered by setting
\begin{equation}
C^{s}_{TTT} = -\frac{8}{S_d^3\,(d-2)^3}\, ,
\quad C^{f}_{TTT} = \frac{2^{d/2+1}}{S_d^3\,(d-2)^3}\, ,
\quad C^{V}_{TTT} = \frac{1}{S_4^3} \, .
\label{OverAllTTT}
\end{equation}
It was explicitly checked that the results in (\ref{Ch1ScalarTriangle})-(\ref{Ch1VectorTriangle}) with
overall coefficients (\ref{OverAllTTT}) match the result for separate points presented in (\ref{Ch1bareTTT}),
with $a$, $b$ and $c$ assuming the values corresponding to the respective theories, as listed in \cite{Osborn:1993cr}:
\begin{eqnarray}
&&
a = \frac{1}{8\,S_d^3}\, \frac{d^3}{(d-1)^3}\, , \quad
b = - \frac{1}{8\,S_d^3}\, \frac{d^4}{(d-1)^3}\, , \quad
c = - \frac{1}{8\,S_d^3}\, \frac{d^2\,(d-2)^2}{(d-1)^3}\, , \quad \text{for the scalar}
\nonumber \\
&&
a = 0 \, , \quad
b = -\frac{1}{16\,S_d^3}\, d^2\,2^{\frac{d}{2}} \, , \quad
c = -\frac{1}{8\,S_d^3}\, d^2\,2^{\frac{d}{2}} \, , \quad \text{for the fermion}
\nonumber \\
&&
a = - \frac{16}{S_4^3} \, , \quad
b = 0 \, , \quad
c = - \frac{64}{S_4^3}\, , \quad \text{for the photon} \, .
\end{eqnarray}
Next we compute the contributions with the topology of $2$-point functions,
which are needed to account for the behaviour of the vertex in the short distance limits of two coincident points.
In coordinate space we can write them in a manifestly integrable form by pulling out derivatives in the same way as
for the triangle diagram. We replace the momenta with derivatives with respect to the corresponding
coordinates acting on propagators, obtaining very compact expressions for the vertex.
As already mentioned, more details on this computation can be found in appendix \ref{Ch1InverseTTT},
whereas here we just quote the results.
In the scalar case we have
\begin{eqnarray}
\left\langle \frac{\delta^2\mathcal{S}}{\delta g_{\mu\nu}(x_1)\delta g_{\alpha\beta}(x_3)}\,
\frac{\delta\mathcal{S}}{\delta g_{\rho\sigma}(x_2)} \right\rangle^{s}
&=&
\frac{C^{s}_{TTT}}{2}\,
V^{\rho\sigma}_{\mathcal{S}\phi\phi}(i\,\partial^{23},-i\,\partial^{12})\,
V^{\mu\nu\alpha\beta}_{\mathcal{S}\mathcal{S}\phi\phi}(i\,\partial^{12},-i\,\partial^{23},i\,\partial^{23}-i\,\partial^{31})\, \nonumber\\
&& \times \frac{\delta^{(d)}(x_{31})}{(x^2_{12})^{d/2-1}(x^2_{23})^{d/2-1}} \, ,
\nonumber
\end{eqnarray}
\begin{eqnarray}
\left\langle \frac{\delta^2\mathcal{S}}{\delta g_{\mu\nu}(x_1)\delta g_{\rho\sigma}(x_2)}\,
\frac{\delta\mathcal{S}}{\delta g_{\alpha\beta}(x_3)} \right\rangle^{s}
&=&
\frac{C^{s}_{TTT}}{2}\,
V^{\alpha\beta}_{\mathcal{S}\phi\phi}(i\,\partial^{31},-i\,\partial^{23})\,
V^{\mu\nu\alpha\beta}_{\mathcal{S}\mathcal{S}\phi\phi}(i\,\partial^{23},-i\,\partial^{31},-i\,\partial^{23}+i\,\partial^{12})
\nonumber\\
&&\times \frac{\delta^{(d)}(x_{12})}{(x^2_{23})^{d/2-1}(x^2_{31})^{d/2-1}} \, ,
\nonumber
\end{eqnarray}
\begin{eqnarray}
\left\langle \frac{\delta^2\mathcal{S}}{\delta g_{\alpha\beta}(x_3)\delta g_{\rho\sigma}(x_2)}\,
\frac{\delta\mathcal{S}}{\delta g_{\mu\nu}(x_1)} \right\rangle^{s}
&=&
\frac{C^{s}_{TTT}}{2}\,
V^{\mu\nu}_{\mathcal{S}\phi\phi}(i\,\partial^{12},-i\,\partial^{31})\,
V^{\alpha\beta\rho\sigma}_{\mathcal{S}\mathcal{S}\phi\phi}
(i\,\partial^{31},-i\,\partial^{12},i\,\partial^{12}-i\,\partial^{23})\, \nonumber \\
&& \times \frac{\delta^{(d)}(x_{23})}{(x^2_{12})^{d/2-1}(x^2_{31})^{d/2-1}}\, .
\label{Ch1ScalarKBubble}
\end{eqnarray}
Notice that in the three contributions above, the $p,q,$ and $l$ dependence of the vertices correspond to mappings
onto $\hat{p}, \hat{q}$ and $\hat{l}$ which are specific for each bubble.
Similarly, in the fermion sector we obtain
\begin{eqnarray}
\left\langle \frac{\delta^2\mathcal{S}}{\delta g_{\mu\nu}(x_1)\delta g_{\alpha\beta}(x_3)}\,
\frac{\delta\mathcal{S}}{\delta g_{\rho\sigma}(x_2)} \right\rangle^{f}
&=&
- C^{f}_{TTT}\,\delta^{(d)}(x_{31})\, \textrm{tr}\,
\big[
V^{\mu\nu\alpha\beta}_{\mathcal{S}\mathcal{S}\bar\psi \psi}(i\,\partial^{12},-i\,\partial^{23})\,
i\,\gamma\cdot\partial^{12}
V^{\rho\sigma}_{\mathcal{S}\bar{\psi }\psi }(i\,\partial^{23},-i\,\partial^{12})\, i\, \gamma\cdot\partial^{23}\big] \nonumber \\
&\times&
\frac{1}{(x^2_{23})^{d/2-1}(x^2_{12})^{d/2-1}} \, ,
\end{eqnarray}
and similar expressions for the $k-$ and $p$-bubbles.
Finally, for the spin-1 vector field we have
\begin{eqnarray}
\left\langle \frac{\delta^2\mathcal{S}}{\delta g_{\mu\nu}(x_1)\delta g_{\alpha\beta}(x_3)}\,
\frac{\delta\mathcal{S}}{\delta g_{\rho\sigma}(x_2)} \right\rangle^V
&=&
\frac{C^{V}_{TTT}}{2}\, \,\delta^{(d)}(x_{31})\,
V^{\mu\nu\rho\alpha\beta\chi}_{\mathcal{S}\mathcal{S}AA}(i\,\partial^{12},-i\,\partial^{23})\,
V^{\rho\sigma\tau\omega}_{\mathcal{S}AA}(i\,\partial^{23},-i\,\partial^{12})
\,\frac{\delta_{\zeta\tau}\,\delta_{\chi\omega}}{x^2_{12}\,x^2_{23}}\bigg|_{\frac{1}{\xi}\rightarrow 0} \, , \nonumber \\
\end{eqnarray}
and similarly for the other bubble-type contributions.
Again, we find that these results are in exact correspondence with the contact terms $\hat A - A$ given in (\ref{HatATTT}),
which completes our check successfully.
The complete structure of the $TTT$ vertex in $4$ dimensions and in position space is thus obtained by combining the
triangle and the ``k",``p" and ``q"-bubble topologies in the form
\begin{eqnarray}
\left\langle T^{\mu\nu}(x_1)\,T^{\rho\sigma}(x_2)\,T^{\alpha\beta}(x_3) \right\rangle
&=&
\sum_{I=s,f,V} 8 \, C^{I}_{TTT}\,
\bigg[- \left\langle \frac{\delta \mathcal{S}}{\delta g_{\mu\nu}(x_1)}\,
\frac{\delta \mathcal{S}}{\delta g_{\sigma\rho}(x_3)}\,
\frac{\delta \mathcal{S}}{\delta g_{\alpha\beta}(x_2)} \right\rangle^I
\nonumber\\
&&\hspace{-45mm}
+ \, \left\langle \frac{\delta^2\mathcal{S}}{\delta g_{\mu\nu}(x_1)\,\delta g_{\alpha\beta}(x_3)}\,
\frac{\delta\mathcal{S}}{\delta g_{\rho\sigma}(x_2)}\right\rangle^I +
\left\langle \frac{\delta^2\mathcal{S}}{\delta g_{\mu\nu}(x_1)\,\delta g_{\rho\sigma}(x_2)}
\frac{\delta\mathcal{S}}{\delta g_{\alpha\beta}(x_3)}\right\rangle^I
+ \left\langle \frac{\delta^2\mathcal{S}}{\delta g_{\alpha\beta}(x_3)\,\delta g_{\rho\sigma}(x_2)}\,
\frac{\delta\mathcal{S}}{\delta g_{\mu\nu}(x_1)}\right\rangle^I \bigg] \, .
\nonumber \\
\end{eqnarray}
This expression is in agreement with the form of the unrenormalized energy-momentum tensor
three point function given in \cite{Osborn:1993cr}. The integrability of this result is manifest, due to the $(d/2-1)$ exponent of
each propagator in position space, which corresponds, generically, to a $1/l^2$ behaviour in momentum space. \\
We are now ready to discuss the renormalization of the correlators discussed so far, elaborating on the meaning of the counterterms
and their relation to the trace anomaly.
\section{Counterterms and their relation to the trace anomaly}
\label{Counterterms}
So far, in comparing the position-space results of \cite{Osborn:1993cr} with perturbative computations in momentum space,
we discarded the issue of renormalization of the divergent correlators, particularly of the $TVV$ and $TTT$ vertices
(the $VVV$ is finite, whereas the ultraviolet behaviour of the $TOO$ is of no particular interest and
and will not be considered any longer).
The reason was simply that the inverse mapping procedure we used naturally allows to establish a direct correspondence
between $3$- and $2$-point function topologies and solutions of the conformal constraints at separate points or in the
coincidence limits in which $2$ out of $3$ points are pinched.
On the other hand, divergent contributions, corresponding to poles in $1/\epsilon$ in the dimensional regularization scheme,
are found, in position space, in the limit in which all the three points coincide $x_1 \approx x_2 \approx x_3$, which come
from the high-momentum region in the loop integrals defining our correlators in the perturbative picture
(see, in particular, eq. (8.13) of \cite{Osborn:1993cr}).
In this section we discuss the structure of counterterms for conformal $2$- and $3$-point functions
and show how one can derive them by imposing Ward identities on the renormalized vertices and telling the divergent
contributions apart from the finite ones we already treated in the previous section.
We also comment on the relation between the traces of the counterterms in the analytically continued dimension $d$
and the trace anomaly of the corresponding correlators.
The results of this discussion are complementary with the ones of the previous section and, together, they complete the
study of the correspondence between position and momentum space results for conformal correlators, which was
the first goal of this chapter.
\subsection{The counterterms for $2$-point functions}
As we are going to see, the interpretation of the anomaly and of its origin,
in the process of renormalization, can be different, depending on the way the correlator is represented.
In fact, the anomaly can be attributed either to the renormalized amplitude in $4$ dimensions or, alternatively,
to the specific structure of the counterterm in dimensional regularization, which violates conformal invariance in $d$ dimensions,
while being traceless for $d=4$.
In the first case the anomaly emerges as a feature of the $d=4$ renormalized amplitude and,
specifically, of its 4-dimensional trace (in different even dimensions there will be a similar mechanism at work).
We start by illustrating the case of the $TT$, which allows to discuss both the renormalization of a $2$-point function
and the connection of counterterms and trace anomalies.
Together with the discussion of the counterterms for the $VV$, which follows, this part is a warm-up exercise in view of the analysis
of the counterterms for $3$-point functions that we will discuss afterwards.
In the $TT$ case conformal symmetry fixes this correlator up to constant. \\
Recalling eq. (\ref{TTPoistion}), the conformal EMT $2$-point function is given by
\begin{equation}\label{Ch12PFOsborn}
\left\langle T^{\mu \nu}(x) \, T^{\alpha \beta} (0)\right\rangle =
\frac{C_T}{x^{2d}} \, \mathcal{I}^{\mu\nu ,\alpha\beta}(x) \, ,
\end{equation}
with $\mathcal{I}^{\mu\nu,\alpha\beta}(s)$ defined in (\ref{Ch1Inversion}) and (\ref{Ch1epislon}). \\
In order to move in the framework of differential regularization, we pull out some derivatives and rewrite our correlator as
\begin{equation}
\label{Ch12nd2PFOsborn}
\left\langle T^{\mu\nu}(x) \, T^{\alpha\beta}(0) \right\rangle
= \frac{C_T}{4\,d\,(d-2)^2\, (d+1)} \, \hat{\Delta}^{(d)\,\mu\nu\alpha\beta}
\frac{1}{(x^2)^{d - 2}} \, ,
\end{equation}
where the differential operator $\hat\Delta^{d\,\mu\nu\alpha\beta}$ is defined as
\begin{eqnarray}
\hat{\Delta}^{(d)\,\mu\nu\alpha\beta}
&=&
\frac{1}{2}\left( \hat{\Theta}^{\mu\alpha} \hat{\Theta}^{\nu\beta} +
\hat{ \Theta}^{\mu\beta}\hat{ \Theta}^{\nu\alpha} \right)
- \frac{1}{d-1}\hat{ \Theta^{\mu\nu}} \hat{\Theta}^{\alpha\beta}\, ,
\quad
\text{with}
\quad
\hat{\Theta}^{\mu\nu} = \partial^\mu \partial^\nu - \delta^{\mu\nu} \, \Box \, ,
\nonumber \\
\partial_\mu \, \hat{\Delta}^{(d)\,\mu\nu\alpha\beta}
&=&
0 \, , \quad
\delta^{(d)}_{\mu\nu} \, \hat{\Delta}^{(d)\,\mu\nu\alpha\beta} = 0 \, .
\label{Ch1TransverseDeltaCoord}
\end{eqnarray}
This form of the $TT$ correlator is Fourier-integrable (again, for the meaning of integrability see section \ref{Ch1direct},
in particular eq. (\ref{Ch1fund})). It is also characterized by a UV divergence in the limit $x\to 0$.
To move to momentum space we can split the ${1}/{(x^2)^{d-2}}$ term
into the product of two ${1}/{(x^2)^{d/2-1}}$ factors and apply straightforwardly the fundamental transform
in eq. (\ref{Ch1fund}), obtaining
\begin{eqnarray} \label{Ch1TransTT}
\left\langle T^{\mu\nu} \, T^{\alpha\beta} \right\rangle (p)
&\equiv&
\int \, d^d x \, \left\langle T^{\mu\nu}(x) T^{\alpha\beta}(0) \right\rangle
e^{- i\, p \cdot x} \nonumber\\
&=&\frac{C_T}{4\,d\,(d-2)^2\, (d+1)} \, \int \, d^d x \, e^{- i\, p \cdot x}
\,\hat{ \Delta}^{(d)\,\mu\nu\alpha\beta}\, \frac{1}{(x^2)^{d/2-1}} \, \frac{1}{(x^2)^{d/2-1}} \nonumber \\
&=&
\frac{ (2\pi)^d \, C(d/2 - 1)^2 \, C_T}{4\,d\,(d-2)^2\, (d+1)} \,
\Delta^{(d)\,\mu\nu\alpha\beta}(p) \, \int \, d^d l \, \frac{1}{l^2 (l+p)^2} \, ,
\end{eqnarray}
where we use the momentum space counterparts of the operators introduced in (\ref{Ch1TransverseDeltaCoord}),
\begin{eqnarray} \label{Ch1TransverseDeltaMom}
\Theta^{\mu\nu}(p)
&=&
\delta^{\mu\nu} \, p^2 - p^\mu \, p^\nu \, , \nonumber \\
\Delta^{(d)\,\mu\nu\alpha\beta}(p)
&=&
\frac{1}{2} \, \bigg(\Theta^{\mu\alpha}(p) \, \Theta^{\nu\beta}(p) + \Theta^{\mu\beta}(p) \, \Theta^{\nu\alpha}(p) \bigg)
- \frac{1}{d-1} \, \Theta^{\mu\nu}(p) \, \Theta^{\alpha\beta}(p) \, .
\end{eqnarray}
Of course, due to conformal invariance, in $d$ dimensions the $TT$ correlator is anomaly-free,
as apparent from (\ref{Ch1TransTT}),
\begin{equation}
\delta_{\mu\nu} \left\langle T^{\mu\nu} \, T^{\alpha\beta} \right\rangle \, (p) =
\delta_{\alpha\beta} \left\langle T^{\mu\nu} \, T^{\alpha\beta} \right\rangle \, (p) = 0 \, .
\end{equation}
Now, as we move to $d=4$ the correlator in momentum space has a UV singularity, coming from the $2$-point integral, shown in
eqs. (\ref{BareB0}) and (\ref{RenB0}).
Then one has to plug (\ref{BareB0}) into (\ref{Ch1TransTT}), \emph{expand all the result},
including the $1/(d-1)$ factor in $\Delta^{(d)\,\mu\nu\alpha\beta}(p)$, around $d=4$
and discard the terms that are $O(\epsilon)$, so as to end with the general expression of the bare $2$-point correlator,
already met in eq. (\ref{BareB0}),
\begin{eqnarray} \label{Ch12PFp}
\left\langle T^{\mu\nu} \, T^{\alpha\beta} \right\rangle_{bare}(p)
&=&
C_1(p)\, \bigg[ \frac{1}{2} \, \bigg( \Theta^{\mu\alpha}(p) \, \Theta^{\nu\beta}(p) +
\Theta^{\mu\beta}(p) \, \Theta^{\nu\alpha}(p)\bigg)
- \frac{1}{3} \, \Theta^{\mu\nu}(p) \, \Theta^{\alpha\beta}(p)\bigg]
+ \frac{C_2}{3}\,\Theta^{\mu\nu}(p) \, \Theta^{\alpha\beta}(p) \, \nonumber\\
&\equiv&
C_1(p)\, \Delta^{(4)\,\mu\nu\alpha\beta}(p) +
\frac{C_2}{3}\Theta^{\mu\nu}(p) \, \Theta^{\alpha\beta}(p) \, .
\end{eqnarray}
It must be pointed out that, whereas both the contributions to the unrenormalized correlator in the last line of (\ref{Ch12PFp})
separately respect the energy-momentum conservation Ward identity for the $2$-point function (\ref{Ch1WI2PFMom}),
only the first one, proportional to $\Delta^{(4)\,\alpha\beta\rho\sigma}(p)$ and carrying the divergence, is traceless in $d=4$,
while tracing the second, finite term we obtain the anomalous relation
\begin{equation}
\delta^{(4)}_{\mu\nu} \left\langle T^{\mu\nu}T^{\alpha\beta} \right\rangle(p) =
\frac{C_2}{3}\, \delta^{(4)}_{\mu\nu}\, \Theta^{\mu\nu}(p)\,\Theta^{\alpha\beta}(p) =
C_2\, p^2 \, \Theta^{\alpha\beta}(p) = 2 \, \beta_c\,\big[\Box R \big]^{\alpha\beta}(p) \, .
\end{equation}
The last equality can be checked directly from eq. (\ref{Ch1TraceAnomaly}), by computing the first functional derivative of its r.h.s.
around flat space, which leaves $ \Box R $ as the only contribution to the $TT$ anomaly.
The superscript $4$ on the Kronecker $\delta$ means that the metric is $4$-dimensional, as usual. \\
The singular contribution in eq. (\ref{Ch12PFp}) can be eliminated by the ordinary renormalization procedure
in the $\overline{MS}$ scheme, leaving a result that is finite and whose trace can be taken {\em directly in 4 dimensions}.
The last two equations allow to fix the final structure of the fully renormalized correlator in the form
\begin{equation}\label{Ch1Ren2PF1}
\left\langle T^{\mu\nu} \, T^{\alpha\beta} \right\rangle_{ren}(p) =
\left\langle T^{\mu\nu}\,T^{\alpha\beta}\right\rangle_{bare}(p)
+ 6 \, \beta_c\, \frac{\mu^{-\epsilon}}{\bar{\epsilon}} \, \Delta^{(4)\,\mu\nu\alpha\beta}(p)=
\left\langle T^{\mu\nu} \, T^{\alpha\beta}\right\rangle_{bare}(p)
- 4 \, \beta_a\, \frac{\mu^{-\epsilon}}{\bar{\epsilon}} \, \Delta^{(4)\,\mu\nu\alpha\beta}(p)\, ,
\end{equation}
where we have used eq. (\ref{Ch1constraints}) in the last step. \\
So far, the anomaly can be unambiguously attributed to the regularization procedure, not to the counterterm, which is traceless
in the physical dimension where traces are taken. \\
Now we want to explore a second approach to the problem, which is the one exploited in \cite{Osborn:1993cr} and that is particularly
suited to renormalization in position space. \\
To explain how to switch over to this point of view, let us write the renormalized correlator \emph{around the physical dimension} $d=4$,
but without doing any series expansion. Its form is
\begin{equation}
\left\langle T^{\mu\nu} \, T^{\alpha\beta} \right\rangle(p)
= \frac{C_T}{4\,d\,(d-2)^2\, (d+1)} \, \hat{\Delta}^{(d)\,\mu\nu\alpha\beta}(p)\, \mathcal{B}_0(p^2)
- 4\, \beta_a\, \frac{\mu^{-\epsilon}}{\bar{\epsilon}} \, \Delta^{(4)\,\mu\nu\alpha\beta}(p) \, ,
\label{Ch1super}
\end{equation}
where the counterterm is meant to remove the ultraviolet singularity as $d \rightarrow 4$ and it is implicitly meant
that the values of $C_T$ and $\beta_a$ depend on the field content of the theory. \\
In other words, one keeps everything $d$-dimensional and subtracts from it the $4$-dimensional ultraviolet divergence,
which does not exist away from $d=4$. \\
As noticed above, the counterterm is traceless for $d=4$ (i.e. contracting the indices with a 4-dimensional metric)
but not in general dimensions $d$.
The key observation is that, as the correlator is written in $d$ dimensions, it is natural to compute its trace by contracting it with the
$d$-dimensional metric. \\
We are free to split the $\delta^{(d)}_{\mu\nu}$ into a direct sum ($\oplus$)
of a 4-dimensional ($\delta_{\mu\nu}\equiv \delta^{(4)}_{\mu\nu}$) and of a $(d-4)$-dimensional metrics acting
on the subspaces $E_4$ and $E_{d-4}$ of $d$-dimensional euclidean space $E_d$
\begin{eqnarray}
E_d = E_4 + \oplus E_{d-4} \, , \nonumber \\
\delta^{(d)}_{\mu\nu} = \delta^{(4)}_{\mu\nu} + \delta^{(d-4)}_{\mu\nu} \, .
\end{eqnarray}
Then, by taking the trace we obtain
\begin{equation} \label{Ch1BrokenTrace}
\delta^{(d)}_{\mu\nu} \Delta^{(4)\mu\nu\alpha\beta}(p) =
\delta^{(4)}_{\mu\nu}\Delta^{(4)\mu\nu\alpha\beta}(p) +
\delta^{(d-4)}_{\mu\nu}\Delta^{(4)\mu\nu\alpha\beta}(p) =
\delta^{(d-4)}_{\mu\nu} \Delta^{(4)\mu\nu\alpha\beta}(p) \, .
\end{equation}
To arrive at (\ref{Ch1BrokenTrace}) we have used the tracelessness property
\begin{equation}
\delta^{(4)}_{\mu\nu} \Delta^{(4)\mu\nu\alpha\beta}(p) =0 \, .
\end{equation}
Thus, we find that the $d$-dimensional trace of $\Delta^{(4)}$ is $O(\epsilon)$
\begin{equation}
\delta^{(d)}_{\mu\nu}\, \hat{\Delta}^{(4)\,\mu\nu\alpha\beta} =
\frac{\epsilon}{3}\, p^2 \, \Theta^{\alpha\beta}(p) \, .
\end{equation}
It is then apparent that the trace of the renormalized $TT$ correlator around the physical dimension gives the correct anomaly.
In particular, the trace operation cancels the ${1}/{\epsilon}$ pole of the counterterm
\begin{equation}
\delta^{(d)}_{\mu\nu}\left\langle T^{\mu\nu} T^{\alpha\beta} \right\rangle(p) =
-4 \, \frac{\beta_a }{\bar{\epsilon}} \, \delta^{(d-4)}_{\mu\nu} \, \Delta^{(4)\,\mu\nu\alpha\beta}(p)
= 2\, \beta_c\, \, p^2\, \Theta^{\alpha\beta}(p) + O(\epsilon)\, ,
\label{Ch1count}
\end{equation}
which is finite as $\epsilon\to 0$ and reproduces the expected anomaly.
From this point of view, the anomaly can be attributed to the counterterm. \\
Here the $TT$ case is used only as an illustrative example. This procedure is very general and can be applied to any correlator.
In the following, we present simple relations that allow to extend this argument to arbitrary correlators involving
vector currents and/or EMT's on their external lines.
This will complete our discussion of the mapping between momentum and position space solutions of the conformal Ward identities
and open the way to the computations presented in chapters \ref{Traced4T} and \ref{Recursive},
for which the relation between counterterms of CFT's and conformal anomalies is of paramount importance. \\
Now we explain the reason why this approach is ideally suited for trace anomalies in position space CFT's. \\
As already pointed out, in position space ultraviolet singularities appear, in $2$ as in $3$-point functions, in the form of $1/\epsilon$ poles,
only for completely coincident points (see, in particular, eq. (8.13) of \cite{Osborn:1993cr}), i.e. \emph{1-loop divergences are local}.
This allows to write down the solution of the Ward identities as a sum of three pieces.
The first piece is built on the grounds of conformal invariance constraints for general $d$ dimensions, keeping
all the points separate. It respects naive Ward identities. \\
The second term contains $delta$-functions forcing no more than two points to coincide and is obtained by regularizing
the terms appearing in the short-distance limits $x_1 \rightarrow x_2$ and $x_1 \rightarrow x_3$.
These are needed to satisfy the (yet unrenormalized) Ward identities.
All this was reviewed in the previous section and, so far, neither ultraviolet singularities nor trace anomalies appear,
as everything is computed in general $d$ dimensions. \\
In the limit in which all the three points coincide, one has to add a counterterm to remove ultraviolet singularities.
This term is thus proportional to $1/\epsilon$ times $\delta$-functions and derivatives thereof enforcing all the three points to coincide.
As all the rest of the correlator is kept $d$-dimensional, if the Green function contains EMT's,
the trace anomaly cannot descend but \emph{from the trace of the counterterm, taken in $d$ dimensions}.
For example, let us write this down the renormalized $TT$ in position space,
\begin{equation}
\left\langle T^{\mu\nu}(x) \, T^{\alpha\beta}(0) \right\rangle
= \frac{C_T}{4\,d\,(d-2)^2 \,(d+1)} \, \hat{\Delta}^{(d)\,\mu\nu\alpha\beta}
\frac{1}{x^{2d - 4}} -4\, \beta_a\, \frac{\mu^{-\epsilon}}{\bar{\epsilon}} \,
\hat{\Delta}^{(4)\,\mu\nu\alpha\beta} \, \delta^d(x-y) \, ,
\label{Ch1SuperPoistion}
\end{equation}
from which the local structure of the counterterm is manifest. \\
To prepare for the discussion of the $TVV$ counterterm, we end by briefly recalling the structure of the famous vector $2$-point function,
given in (\ref{VVPoistion}) in position space for separate points and becoming, in momentum space
\begin{equation}
\left\langle V^a_\alpha \, V^b_\beta \right\rangle(p) =
\delta^{ab}\, C_V\, \Theta_{\alpha\beta}(p)\, B_0(p^2) \, ,
\label{VVd}
\end{equation}
the coefficient $C_V$ depends on the nature of the vector current $V$.
For the cases of the complex scalar and the fermion field of eq. (\ref{SAndF}), this coefficient is
\begin{equation}
C^s_V = - \frac{1}{2^{d+1}\, \pi^{d-2}\,(d-1)}\, , \qquad
C^f_V = - \frac{d-2}{2^{d-1}\, \pi^{d-2}\,(d-1)} \, .
\end{equation}
eq. (\ref{VVd}) can be expanded around $d=4$ to give
\begin{equation}
\left\langle V^a_\alpha \, V^b_\beta \right\rangle_{bare}(p) =
\delta^{ab}\, \Theta_{\alpha\beta} \bigg( c_1 + \frac{c_2}{\bar\epsilon} \bigg)\, ,
\label{ExpandVVd}
\end{equation}
where the values of $c_1$ and $c_2$ do not matter and which shows that both the bare part of the $VV$ and the counterterms
that has to be added, being both proportional to $\Theta_{\alpha\beta}(p)$, separately satisfy the gauge invariance constraint,
\begin{equation}
p_{\alpha}\left\langle V^a_\alpha \, V^b_\beta \right\rangle_{bare}(p) = 0 \, .
\end{equation}
\subsection{Connection between counterterms and trace anomalies}\label{CountAnom}
We review a method to derive the trace anomaly for CFT's.
The method is originally due to Duncan and Duff \cite{Duff:1977ay,Duncan:1976pv} and allows to generalize the discussion
of the previous section to arbitrary correlators of EMT's and vector currents (we discard scalar operators). \\
As we have seen above, an argument in position space shows that the trace of the counterterms in $d$ dimensions has to yield
the trace anomaly of the corresponding correlator.
Let us consider a general euclidean CFT described by a generating functional (\ref{GenPlusSources}),
depending on the background gauge fields $A^a_\mu$ and metric $g_{\mu\nu}$.
For some values of the space dimension, for instance $d=4$, the generating functional is affected by ultraviolet singularities.
In the framework of dimensional regularization, $1$-loop divergences are parametrized by a pole in $\epsilon = 4-d$.
For free field theories there are no higher order contributions and the generating functional is completely determined by
the sum of the bare part (\ref{GenPlusSources}) plus the $1$-loop counterterms.
These are strongly constrained by the requirement of Weyl invariance for $d=4$, which implies that the dimensionally continued
generating functional must consist of a combination of contributions, say $\mathcal{C}$, enjoying the following properties:
\begin{itemize}
\item they must depend only on the background gauge fields $A^a_\mu$ and metric $g_{\mu\nu}$; \\
\item they must be invariant under gauge and general coordinate transformations ; \\
\item as there are no dimensionful constants in the bare theory, their mass dimension must be $4$ ; \\
\item they must be Weyl invariant in 4 dimensions, i.e.
\begin{equation}
\lim_{d \rightarrow 4} \bigg\{ \frac{2}{\sqrt{g}}\, g_{\mu\nu} \frac{\delta\mathcal{C} }{\delta g_{\mu\nu}}\bigg\} = 0 \, .
\end{equation}
\end{itemize}
The set of these terms is well known in $4$ dimensions, where the only one depending on the gauge fields is the squared field-strength,
$F^{a\,\mu\nu}\,F^a_{\mu\nu}$, whereas the other two possible contributions were studied in chapter \ref{TTTVertex}
and are the Weyl tensor squared $F$ and the Euler density in $4$ dimensions, $G$. \\
We conclude that the renormalized generating functional for a CFT is written, in the $\overline{MS}$ scheme, in the form
\begin{equation}
\mathcal{W}_{ren}[g,A] \equiv \mathcal{W}[g,A] - \frac{\mu^{-\epsilon}}{\bar\epsilon}\mathcal{W}_{Ct}[g,A] =
\mathcal{W}[g,A] - \frac{\mu^{-\epsilon}}{\bar\epsilon}\,
\int d^dx\, \sqrt{g}\, \bigg( c_F \, F + c_g\, G + c_A\, F^{a\,\mu\nu}\,F^a_{\mu\nu} \bigg) \, .
\label{RenW}
\end{equation}
From this renormalized generating functional we can derive the vacuum expectation value of the EMT, using its definition
(\ref{Ch1VEVEMT}), and thus we can infer its trace. As (\ref{RenW}) is written in dimensional regularization,
it is clear that the trace must be taken in $d$ dimensions.
Moreover, the bare action always enjoys conformal invariance in the limit $d\rightarrow 4$
(for scalars and fermions this is true already in $d$ dimensions) as we have seen in the example of the EMT $2$-point function,
\begin{equation}
\lim_{d\rightarrow 4} \bigg[ \frac{2}{\sqrt{g}}\, g_{\mu\nu}\, \frac{\delta \mathcal W[g,A] }{\delta g_{\mu\nu}}\bigg] = 0 \, ,
\end{equation}
so it does not contribute to the trace anomaly. \\
We have to compute the contribution of the counterterm, taking the limit $d\rightarrow 4$ after tracing, i.e.
\begin{equation}
g_{\mu\nu} \left\langle T^{\mu\nu} \right\rangle_s \equiv
\lim_{d\rightarrow 4} \bigg\{ g_{\mu\nu}\, \frac{2}{\sqrt{g}}\, g_{\mu\nu}\, \frac{\delta}{\delta g_{\mu\nu}} \,
\bigg[ -\frac{\mu^{-\epsilon}}{\bar\epsilon}\, \int d^dx\, \sqrt{-g}\,
\bigg( c_F\, F + c_G\, G + c_A\, F^{a\,\mu\nu}\,F^a_{\mu\nu} \bigg) \bigg] \bigg\} \, .
\end{equation}
It is easy to show, following the procedure illustrated in appendix \ref{Ch1FunctionalIntegral}, that the following relations hold
\begin{eqnarray}
&&
\frac{2}{\sqrt{g}}\, g_{\mu\nu}\, \frac{\delta}{\delta g_{\mu\nu}}\, \int d^d x\,\sqrt{g}\, F =
- \epsilon \, \left(F - \frac{2}{3}\, \Box R\right)\, , \nonumber \\
&&
\frac{2}{\sqrt{g}}\, g_{\mu\nu}\, \frac{\delta}{\delta g_{\mu\nu}}\, \int d^d x\, \sqrt{g}\, G =
- \epsilon \, G \, , \nonumber \\
&&
\frac{2}{\sqrt{g}}\, g_{\mu\nu}\, \frac{\delta}{\delta g_{\mu\nu}}\, \int d^d x\, \sqrt{g}\, F^{a\,\mu\nu}\,F^a_{\mu\nu} =
- \epsilon \, F^{a\,\mu\nu}\, F^a_{\mu\nu} \, .
\label{CTVariations}
\end{eqnarray}
We see that, in this approach, the trace anomaly is intimately connected to the counterterms, as the $O(\epsilon)$ contributions
in (\ref{CTVariations}) cancel the $\epsilon$-pole in the $1$-loop counterterm, yielding the finite result (\ref{Ch1TraceAnomaly})
if one sets the values of the coefficients
\begin{equation}
c_F = \beta_a \, , \qquad c_G = \beta_b \, , \qquad c_A = -\frac{\kappa}{4} \, .
\end{equation}
Thus, we conclude that an equivalent form of (\ref{Ch1TraceAnomaly}) is
\begin{equation}
\frac{2}{\sqrt{g}}\, g_{\mu\nu}\, \frac{\delta \mathcal{W}_{Ct}[g,A]}{\delta g_{\mu\nu}} =
- \epsilon\, \mathcal{A}[g,A] \, ,
\label{dTraceCt}
\end{equation}
where the trace is meant to be $d$-dimensional, due to the presence of the counterterm $\mathcal{W}_{Ct}$,
computed in dimensional regularization. Of course we are neglecting $O(\epsilon^2)$ terms due to the difference between
$\epsilon$ and $\bar\epsilon$. This is the master equation generating the trace Ward identities satisfied by the counterterms
in $d$ dimensions and extensively used in \cite{Osborn:1993cr}. We have checked that the counterterms for the $TVV$
and $TTT$ correlators studied in this chapter satisfy such identities, as we are going do discuss below. \\
Finally, we still have to remark that, for interacting CFT's, as Yang-Mills gauge bosons in $d=4$, divergences that are higher order
than $1/\epsilon$ may exist, in general, so that the renormalized generating functional might be of the form
\begin{eqnarray}
\mathcal{W}_{ren}[g,A] =
\mathcal{W}[g,A] - \frac{\mu^{-\epsilon}}{\epsilon}\, \mathcal{W}^{(1)}_{Ct}[g,A]
- \frac{\mu^{-\epsilon}}{\epsilon^2}\, \mathcal{W}^{(2)}_{Ct}[g,A] + \dots \, ,
\end{eqnarray}
where the superscripts stand for the order of the divergence in $1/\epsilon$.
Now, as the vev of the renormalized EMT is finite, so has to be its trace, then the condition
\begin{equation}
g_{\mu\nu}\, \frac{\delta \mathcal{W}^{(n)}_{Ct}[g,A]}{\delta g_{\mu\nu}} = O(\epsilon^n)\, , \quad n=1,2,3,\dots
\end{equation}
should hold.
To our knowledge, there are no conformal invariants depending only on the metric whose Weyl variation vanishes faster than linearly for $d
\rightarrow 4$ \cite{Mazur:2001aa}. Moreover, concerning the gauge sector of the trace anomaly, the same form as in
(\ref{Ch1TraceAnomaly}) of the gauge field contribution to the trace anomaly was derived in \cite{Adler:1976zt} for QED,
whereas in \cite{Collins:1976yq} it was shown that it also holds for non abelian gauge theories after proper resummations
are performed before taking the limit $d\rightarrow 4$. These results imply that the trace anomaly is completely determined
by the $1$-loop contributions to the counterterm and is thus given by (\ref{Ch1TraceAnomaly}).
\subsection{The counterterm for the $TVV$}
We now turn to the issue of the renormalization of the $TVV$ in $4$ dimensions, which will complete the test of the correspondence
between the position space solution of \cite{Osborn:1993cr} and diagrammatic momentum space computations in dimensional
regularization. \\
The renormalized $3$-point function has to satisfy the requirement of general covariance (\ref{TVVWardCoord})
as well as the anomalous Ward identity (\ref{AnTraceWardTVV}).
As explained at the end of section \ref{Ch1lags}, the implications of general covariance for the counterterms immediately descend
from the Ward identities for the corresponding Green functions.
Specifically, we write the renormalized $VV$ and $TVV$ correlator as
\begin{eqnarray}
\langle V^a_{\alpha}V^b_{\beta} \rangle_{ren}(p)
&=&
\langle V^a_{\alpha}V^b_{\beta} \rangle_{bare}(p)
+ \frac{\mu^{-\epsilon}}{\bar\epsilon}\,\frac{\kappa}{4}\, D_{\alpha\beta}(p) \, ,
\nonumber \\
\langle T_{\mu\nu}V^a_{\alpha}V^b_{\beta} \rangle_{ren}(p,q)
&=&
\langle T_{\mu\nu}V^a_{\alpha}V^b_{\beta} \rangle_{bare}(p,q) +
\frac{\mu^{-\epsilon}}{\bar\epsilon}\,\frac{\kappa}{4}\, D_{\mu\nu\alpha\beta}(p,q)\, ,
\label{RenTVVAndVV}
\end{eqnarray}
where the counterterms are derived according to the results of the previous section,
applying the definitions of the $TVV$ and the $VV$ vertices to the renormalized generating functional (\ref{RenW}).
Their expressions in position space are found to be
\begin{eqnarray}
D_{\alpha\beta}(x_1,x_2)
&=&
\frac{\delta^2 F^{c\,\gamma\delta}F^c_{\gamma\delta}}{\delta A^{a\,\alpha}(x_2)\delta A^{b\,\beta}(x_3)}\bigg|_{A=0} \, ,
\nonumber \\
D_{\mu\nu\alpha\beta}(x_1,x_2,x_3)
&=&
\frac{\delta^2}{\delta A^{a\,\alpha}(x_2) \delta A^{b\,\beta}(x_3)} \bigg[ \frac{2}{\sqrt{g_{x_1}}}
\frac{\delta F^{c\,\gamma\delta}F^c_{\gamma\delta}}{\delta g_{\mu\nu}(x_1)}\bigg]_{g=\delta} \bigg|_{A=0} \, ,
\label{PoistionTVVCt}
\end{eqnarray}
and we denote with $D_{\alpha\beta}(p)$, $D_{\mu\nu\alpha\beta}(p,q)$ their Fourier transforms. \\
We find that (\ref{PoistionTVVCt}) match the expressions presented in \cite{Osborn:1993cr} and satisfy the
constraints that are obtained by plugging (\ref{RenTVVAndVV}) into the general covariance and gauge invariance Ward identities,
(\ref{TVVWardCoord}) and (\ref{GaugeInvTVV}), then passing to momentum space and isolating the divergent parts, i.e.
\begin{eqnarray} \label{Ch1DivWardTVV}
(p + q)^\mu \, D_{\mu\nu\alpha\beta} (p,q)
&=&
q_\nu \, D_{\alpha\beta}(p) - \delta_{\nu\beta} \, q^\mu \, D_{\mu\alpha}(p) +
p_\nu \, D_{\alpha\beta}(q) - \delta_{\nu\alpha} \, p^\mu \, D_{\mu\beta}(q) \, ,
\nonumber \\
p^\alpha \, D_{\mu\nu\alpha\beta}(p,q)
&=&
q^\beta \, D_{\mu\nu\alpha\beta}(p,q) = 0 \, .
\end{eqnarray}
We provide the explicit form of the counterterms in momentum space,
\begin{eqnarray}
D_{\alpha\beta}(p)
&=&
\Theta_{\alpha\beta}(p)\, , \nonumber \\
D_{\mu\nu\alpha\beta}(p,q)
&=&
\delta_{\alpha\beta} \, ( p_\mu \, q_\nu + q_\mu \, p_\nu )
- (\delta_{\beta\nu} \, p_\mu + \delta_{\beta\mu} \, p_\nu ) \, q_\alpha -
( \delta_{\mu\alpha} \, q_\nu + \delta_{\alpha\nu} \, q_\mu ) \, p_\beta
\nonumber \\
&&
+\, p \cdot q \, (\delta_{\mu\beta}\, \delta_{\nu\alpha} + \delta_{\mu\alpha} \, \delta_{\nu\beta} )
- \delta_{\mu\nu} \, ( p \cdot q \, \delta_{\alpha\beta} - q_\alpha \, p_\beta ) \, .
\end{eqnarray}
All that is left to check and is easily done is that the counterterm for the $TVV$ satisfies the $d$-dimensional trace relation discussed
in \ref{CountAnom} and encoded in eq. (\ref{dTraceCt}), which is
\begin{equation}
\delta^{(d)\, \mu\nu} \, D_{\mu\nu\alpha\beta}(p,q) =
- \epsilon \, ( q_\alpha \, p_\beta - p \cdot q \, \delta_{\alpha\beta} ) \, ,
\end{equation}
reproducing the correct anomaly. \\
We know that the identification of the divergent parts of the $TVV$ correlator can be performed diagrammatically.
We just mention that the general form of the $TVV$ amplitude can be expanded in a basis of $13$ tensor structures
defined in \cite{Giannotti:2008cv}.
A complete perturbative analysis shows that there is only $1$ tensor structure which is affected by the renormalization
procedure, which coincides with the $D_{\mu\nu\alpha\beta}(p,q)$ counterterm introduced above.
As found by direct computations in \cite{Giannotti:2008cv,Armillis:2009pq} for QED, in \cite{, Armillis:2010qk} for QCD
and in \cite{Coriano:2011zk} for the electroweak sector, renormalization of the $TVV$ vertex always affects only this tensor structure.
Given the complexity of the computations and the wide difference between the general CFT
approach and the ordinary diagrammatic one, this agreement is obviously non trivial.
\subsection{The counterterms for the $TTT$}
For the case of the $TTT$, the discussion of the derivation of the counterterms was already done in chapter \ref{TTTVertex},
so here we do not report it. The definitions of the counterterms for the EMT $2$- and $3$-point functions (\ref{Ch12PFCounterterm}),
(\ref{Ch1DF}) and (\ref{Ch1DG}) are identical, so that it is no point repeating the discussion.
Nevertheless, it is instructive to see how one can derive the analogue of the $TTT$ counterterm (\ref{Ch1Ren3PF})
by using the Ward identities to constrain the scalar coefficients and knowing the counterterm just for the $2$-point function.
In this case we are bound to introduce the generic counterterms to the $TTT$ vertex
\begin{equation} \label{Ch1Ren3PFansatz}
\left\langle T^{\mu\nu}T^{\rho\sigma}T^{\alpha\beta} \right\rangle_{ren}(p,q) =
\left\langle T^{\mu\nu}T^{\rho\sigma}T^{\alpha\beta} \right\rangle_{bare}(p,q)
+ \frac{1}{\bar\epsilon} \,
\bigg(C_F \, D_F^{\mu\nu\alpha\beta\rho\sigma}(p,q) + C_G \, D_G^{\mu\nu\alpha\beta\rho\sigma}(p,q)\bigg), \,
\end{equation}
written in terms of arbitrary coefficients $C_F$ and $C_G$ that one cannot know \emph{a priori}.
With the addition of the counterterms, it is clear that the renormalized vertex must satisfy (\ref{Ch1WI3PFcoordinate})
and two similar identities which follow by exchanging indices and momenta properly. In fact,
renormalization has to preserve general covariance.
One can check that $D_G^{\mu\nu\alpha\beta\rho\sigma}(p,q)$ is transverse, as (\ref{Ch1ExplicitDG}) shows clearly,
\begin{equation} \label{Ch1DGConstraints1}
k_{\nu}D_G^{\mu\nu\alpha\beta\rho\sigma}(p,q) = 0 \, , \quad
p_{\alpha}D_G^{\mu\nu\alpha\beta\rho\sigma}(p,q) = 0 \, \quad
q_\sigma D_G^{\mu\nu\alpha\beta\rho\sigma}(p,q) =0\, ,
\end{equation}
so that, by inserting the expressions (\ref{Ch1Ren2PF2}) and (\ref{Ch1Ren3PFansatz}) into these Ward identities and taking
(\ref{Ch1DGConstraints1}) into account, one obtains three conditions on the F-contribution to the counterterm, the first being
\begin{eqnarray} \label{Ch1DFConstraints1}
&&
C_F \, k_\nu D_F^{\mu\nu\alpha\beta\rho\sigma}(p,q) =
- \beta_a \bigg\{q^\mu D_F^{\rho\sigma\alpha\beta}(p) + p^\mu D_F^{\alpha\beta\rho\sigma}(q) \nonumber\\
&&
- q_\nu \bigg[\delta^{\mu\rho}D_F^{\nu\sigma\alpha\beta}(p)
+ \delta^{\mu\sigma}D_F^{\nu\rho\alpha\beta}(p)\bigg]
- p_\nu \bigg[\delta^{\mu\alpha}D_F^{\nu\beta\rho\sigma}(q)
+ \delta^{\mu\beta}D_F^{\nu\alpha\rho\sigma}(q)\bigg]\bigg\}\, .
\end{eqnarray}
and the other two coming from a permutation of the indices and of the momenta.
In (\ref{Ch1DFConstraints1}) we have used the expression (\ref{Ch1Ren2PF2}) for the renormalized $2$-point function.
These three identities are easily seen to be satisfied if $C_F = - \beta_a$, as one can check with a symbolic calculus program. \\
Once the first coefficient is fixed, the same argument can be applied to the three anomalous trace identities in $d = 4 - \epsilon$
dimensions, which can be exploited in order to fix $C_G$.
These identities descend from the double functional derivative of (\ref{dTraceCt}) with respect to other $2$ metric tensors
and are
\begin{eqnarray}\label{Ch1CTTraces}
\delta_{\mu\nu}D_F^{\mu\nu\alpha\beta\rho\sigma}(p,q)
&=&
-4 \,\epsilon \,\bigg( \big[F\big]^{\alpha\beta\rho\sigma}(p,q)
- \frac{2}{3} \big[\sqrt{g}\Box\,R\big]^{\alpha\beta\rho\sigma}(p,q) \bigg)
- 2 \, \beta_a\, \bigg( D_F^{\alpha\beta\rho\sigma}(p) + D_F^{\alpha\beta\rho\sigma}(q)\bigg) \, , \nonumber\\
\label{Ch1DFConstraints2}\,
\delta_{\mu\nu}D_G^{\mu\nu\alpha\beta\rho\sigma}(p,q)
&=&
- 4 \, \epsilon \,\big[G\big]^{\alpha\beta\rho\sigma}(p,q) \, .
\label{Ch1DGConstraints2}
\end{eqnarray}
According to the previously established notation, $\big[F\big]^{\alpha\beta\rho\sigma}(p,q)$ and
$\big[G\big]^{\alpha\beta\rho\sigma}(p,q)$ are the Fourier-transformed second functional derivatives
of the squared Weyl tensor and the Euler density respectively.
Requiring (\ref{Ch1munu3PFanomaly}) to be satisfied by the renormalized 2 and 3-point correlators we get
\begin{eqnarray} \label{Ch1Trace}
\delta_{\mu\nu} \, \bigg( - \beta_a \,\ D_F^{\mu\nu\alpha\beta\rho\sigma}(p,q)
+ C_G \, D_G^{\mu\nu\alpha\beta\rho\sigma}(p,q) \bigg)
&=&
4 \, \epsilon \,
\bigg[ \beta_a \, \bigg(\left[ F \right]^{\alpha\beta\rho\sigma}(p,q)
- \frac{2}{3}\left[ \sqrt{g}\Box\,R \right]^{\alpha\beta\rho\sigma}(p,q) \bigg)
\nonumber \\
&+&
\beta_b \, \left[ G \right]^{\alpha\beta\rho\sigma}(p,q) \bigg]
- 2 \, \beta_a\, \bigg(D_F^{\alpha\beta\rho\sigma}(p) + D_F^{\alpha\beta\rho\sigma}(q)\bigg)\, ,\nonumber\\
\end{eqnarray}
and other two similar equations, obtained by shuffling indices and momenta as for the general covariance Ward identities.
Solving these conditions allows one to obtain the relation $C_G = -{\beta_b}$, as expected.
\section{Handling massless correlators: a direct approach for general dimensions} \label{Ch1direct}
In the previous chapters we have tried to compare perturbative results in free field theory with general ones coming from imposing
conformal symmetry requirements on certain correlators. We have also seen that in this case one can work backward from
the explicit free field theory representation of these correlators in momentum space and match them with the general solutions
given by the conformal constraints in position space.
This is the case of the $TOO$, $VVV$ and $TVV$ correlators in general dimensions, while for the $TTT$
the $4$-dimensional solution of the Ward identities is completely matched by a combination of scalar, vector and fermion sectors.
As we consider the same $3$-graviton vertex in $d$ dimensions, the vector contribution is not conformally invariant,
and therefore the combination of the scalar and the fermion sectors does not match the most general $d$-dimensional solution.
This raises the issue if there is, in general, a free field theory that can reproduce a given CFT correlator, and there is no simple answer.
The goal of CFT, in fact, is to bootstrap certain correlation functions bypassing, if necessary, a Lagrangian formulation.
In fact, one of the main features of the standard CFT approach in the identification of the correlators is to work in position space
with no reference to a Lagrangian. The finiteness of the Fourier transform is a necessary requirement in order to proceed with the identification,
if this turns out to exist, of the corresponding Lagrangian field theory, since this could always be defined in momentum space.
Checking the finiteness (in momentum space) of a general solution given in position space is not an obvious step,
since a correlator in position space such as the $TTT$ contains several hundreds of terms,
most of them characterized by a divergent Fourier expression in momentum space.
For this reason here we are going to illustrate a very general algorithm that allows to transform correlators of such a complexity using
a direct approach. Our analysis will be formulated in general but illustrated with few examples up to correlators of rank-$4$.
For obvious reasons, we will be choosing, as a test of our approach, some of the Green functions defined in the previous sections.
These, in fact, as we have seen, can be deduced from a Lagrangian formulation and therefore their expressions in momentum space
are well defined. \\
Obviously, we need some intermediate regularization of the integrands (in position space) of these correlators in order
to proceed with the definition of the Fourier transform of each individual term.
This is obtained by introducing a power-like regulator $(\omega)$ which is the analogous of the
$\epsilon$ regulator of ordinary dimensional regularization but otherwise completely unrelated to it.
The algorithm implements a sequence of integration by parts before proceeding with the identification of the
$\omega$-regulated transforms. As a consistency condition, the correlators that we investigate have finite Fourier expressions,
as expected, and we check the direct cancellation of all their Fourier singularities, which appear as poles (double and single) in $1/\omega$.
The finite products of the procedure, which correspond to the Fourier space integrands, manifest specific logarithmic terms.
These, in general, are a new feature of the momentum space form of a given CFT correlator.
They are expected to appear once we rewrite any CFT correlation function from position to momentum space.
In some cases, these log terms can be rewritten as ordinary (non-logarithmic) integrals,
while in other cases this may not be possible, and we can think of the log-integrals, in all these second cases,
as of new irreducible contributions.
In the correlators that we explicitly investigate, obviously, we know beforehand that they have to be matched by free field theories.
In this case, a brute force application of the algorithm would produce log-integrals which are, therefore, reducible to ordinary
(non logarithmic) ones. When the $\omega$ singularities cancel, which indicates that it is possible to recollect the terms in position space
(and using integration by parts) in such a way that the Fourier expression is manifestly finite, the logarithmic terms are absent.
The use of the previous (Fourier-integrable) vertices allows to test this approach and show its consistency. \\
Before proceeding with an explicit discussion of the method, we list here, for definiteness, the steps that have to be followed in
order to transform the expression of any given CFT correlator in position space to momentum space:
\begin{enumerate}
\item expansion of the correlator into its single tensor components; \\
\item rewriting of each component in terms of some ``R-substitutions", that we will define below; \\
\item application of the dimensional shift $d\to d- 2 \omega$ which can be performed generically in the expression resulting from point $2$ ; \\
\item implementation of the transform. The transform is implemented by eq. (\ref{Ch1fund}) for each single difference $x_{ij}= x_i - x_j$.
\end{enumerate}
As we are going to describe below, this method and the regularization imposed by the dimensional shift allows to test quite
straightforwardly the integrability of any correlator, a point already emphasized in \cite{Erdmenger:1996yc}
where this regularization has been first introduced.
The transform can be applied in several independent ways. These features share some similarities with the so called
``method of uniqueness" (see for instance \cite{Kazakov:1986mu}) used for massless integrals in momentum or in configuration space.
\subsection{Pulling out derivatives}\label{Ch1pull}
One of the main steps that we will follow in the computation of the transform of the position space expression of the correlators
consists in rewriting a given position space tensor in terms of derivatives of other terms. We call this rule a ``derivative
relation." It allows one to reduce the degree of singularity of a given tensor structure, when the variables are coincident, in the
spirit of differential regularization. Differently from the standard approach of differential regularization, which is $4$-dimensional,
we will be working in $d$ dimensions. We will be using the term ``integrable" to refer to expressions
for which the Fourier transform exists and that are well defined in $d$ dimensions, although they may be singular for $d=4$.
Derivative relations, combined with the basic transform
\begin{eqnarray} \label{Ch1fund}
\frac{1}{(x^2)^\alpha}
&=&
\frac{1}{4^\alpha \pi^{d/2}}\, \frac{\Gamma(d/2 - \alpha)}{\Gamma(\alpha)}\,
\int d^d l \, \frac{e^{il\cdot x}}{(l^2)^{d/2 - \alpha}}
\equiv C(\alpha) \, \int d^d l \, \frac{e^{i l\cdot x}}{(l^2)^{d/2 - \alpha}} \nonumber \\
C(\alpha)
&=&
\frac{1}{4^{\alpha}\,\pi^{d/2}} \frac{\Gamma(d/2 - \alpha)}{\Gamma(\alpha)}
\end{eqnarray}
allow one to perform a direct mapping of these correlators to momentum space.
We proceed with a few examples to show how the lowering of the singularity takes place.
We start from tensors of rank $1$.
We use the relation
\begin{equation}
\frac{x^\mu}{(x^2)^\alpha} =
-\frac{1}{2 (\alpha -1)} \partial_\mu \frac{1}{(x^2)^{\alpha -1}} =
- \frac{i}{2^{2\alpha-1} \pi^{d/2}} \, \frac{\Gamma(d/2 +1 - \alpha) }{\Gamma(\alpha) } \,
\int d^d l \, e^{i l\cdot x}\frac{l_\mu}{(l^2)^{d/2 - \alpha +1}}
\end{equation}
to extract the derivative, where in the last step we have used (\ref{Ch1fund}).
Notice that by using (\ref{Ch1fund}) with $\alpha=d/2-1$ one can immediately obtain the equation
\begin{equation}
\Box \frac{1}{(x^2)^{d/2-1}}= - \frac{4\,\pi^{d/2 }}{\Gamma(d/2-1)} \, \delta^{(d)}(x)\, ,
\label{Ch1deltaeq}
\end{equation}
which otherwise needs Gauss' theorem to be derived.
Scalar $2$-point functions describing loops in position space are next in difficulty.
As an illustration, consider the generalized $2$-point function
\begin{equation}
\frac{1}{[(x-y)^2]^\alpha [(x-y)^2]^\beta}\, .
\end{equation}
Using (\ref{Ch1fund}) separately for the $1/[(x-y)^2]^\alpha$ and the $1/[(x-y)^2]^\beta$ factors,
the Fourier transform ($\mathcal{F T}$) of this expression is found to be
\begin{eqnarray}\label{Ch1FourierB0General}
\mathcal{FT} \left[ \frac{1}{[(x-y)^2]^\alpha [(x-y)^2]^\beta} \right]
&\equiv&
\int\, d^d x \, d^d y \, \frac{e^{- i ( p\cdot x + q \cdot y )}}{[(x-y)^2]^\alpha [(x-y)^2]^\beta}
\nonumber \\
&=&
(2\pi)^{2d} \, C(\alpha) \, C(\beta)\, \int d^d l \, \frac{1}{[l^2]^\alpha [( l+p)^2 ]^\beta}\, .
\end{eqnarray}
The requirement of uniqueness for the transform allows to reformulate it by combining
the powers of the propagators into a single factor,
\begin{equation}
\mathcal{FT}\left[\frac{1}{[(x-y)^2]^{\alpha + \beta}}\right] =
(2\pi)^{2d} \, \frac{C(\alpha + \beta)}{(p^2)^{d/2 - \alpha - \beta}}\, ,
\label{Ch1unsplit}
\end{equation}
giving, for consistency, a functional relation for the integral in (\ref{Ch1FourierB0General})
\begin{eqnarray} \label{Ch1B0general}
\int\,d^dl \, \frac{1}{[l^2]^\alpha[(l+p)^2]^\beta} =
\frac{C(\alpha + \beta)}{C(\alpha) \, C(\beta)} \frac{1}{(p^2)^{d/2 - \alpha - \beta}} =
\pi^{d/2}\, \frac{\Gamma(d/2-\alpha)\Gamma(d/2-\beta)\Gamma(\alpha+\beta-d/2)}{\Gamma(\alpha)\Gamma(\beta)
\Gamma(d-\alpha-\beta)}\, \frac{1}{(p^2)^{\alpha+\beta-d/2}}\, .
\nonumber \\
\end{eqnarray}
In the $TT$ and $TVV$ cases, position space expressions such as $x^{\mu_1}...x^{\mu_n}/ (x^2)^\alpha$ up to rank $4$ are common,
and the use of derivative relations - before proceeding with their final transform to momentum space - can be done in several ways.
Also in this case, as for the scalar functions, uniqueness shows that the result does not depend on the way we combine the factors
at the denominators with the corresponding numerators.
In order to deal with tensor expressions in position space, we introduce some notation.
We denote by
\begin{equation}
{R^n}_{\mu_1\dots \mu_n}(x,\alpha)
\equiv
\frac{x_{\mu_1} \dots x_{\mu_n}}{(x^2)^\alpha} \, ,
\end{equation}
the ratio between a generic tensor monomial in the vector $x$ and a power of $x^2$.
This notation is meant to help us denote in a compact way the tensor structures appearing in the expansion of any tensor correlator.
We call these expressions ``R-terms". \\
After some differential and algebraic manipulation we can easily derive the first four R-terms,
\begin{eqnarray}
{R^1}_\mu(x,\alpha)
&=&
-\frac{1}{2 \, (\alpha-1)} \, \partial_\mu \, \frac{1}{(x^2)^{\alpha-1}} \, , \nonumber \\
{R^2}_{\mu\nu}(x,\alpha)
&=&
\frac{1}{4 \, (\alpha-2) \, (\alpha-1)} \, \partial_\mu\,\partial_\nu \, \frac{1}{(x^2)^{\alpha-2}}
+ \frac{\delta_{\mu\nu}}{2\,(\alpha-1)} \,
\frac{1}{(x^2)^{\alpha-1}} \, , \nonumber \\
{R^3}_{\mu\nu\rho}(x,\alpha)
&=&
- \frac{1}{8 (\alpha-3)(\alpha-2)(\alpha-1)} \,\partial_\mu\,\partial_\nu \, \partial_\rho \frac{1}{(x^2)^{\alpha-3}} +
\frac{1}{2 (\alpha-1)}\, \big[ \delta_{\mu\nu}{R^1}_\rho
+ \delta_{\mu\rho}{R^1}_\nu + \delta_{\nu\rho}{R^1}_\mu \big](x,\alpha-1)\, , \nonumber \\
{R^4}_{\mu\nu\rho\sigma}(x,\alpha)
&=&
\frac{1}{16(\alpha-4)(\alpha-3)(\alpha-2)(\alpha-1)} \,
\partial_\mu \, \partial_\nu \, \partial_\rho \, \partial_\sigma \, \frac{1}{(x^2)^{\alpha-4}} \nonumber \\
&+&
\frac{1}{2(\alpha-1)} \, \big[
\delta_{\mu\nu} {R^2}_{\rho\sigma} + \delta_{\rho\sigma} {R^2}_{\mu\nu} +
\delta_{\mu\rho} {R^2}_{\nu\sigma} + \delta_{\nu\sigma} {R^2}_{\mu\rho} +
\delta_{\mu\sigma} {R^2}_{\nu\rho} + \delta_{\nu\rho} {R^2}_{\mu\sigma}
\big](x,\alpha-1)
\nonumber \\
&-&
\frac{1}{4(\alpha-2)(\alpha-1)} \, (\delta_{\mu\nu}\delta_{\rho\sigma} + \delta_{\mu\rho}\delta_{\nu\sigma}
+ \delta_{\mu\sigma}\delta_{\nu\rho})\, \frac{1}{(x^2)^{\alpha-2}} \label{Ch1Rterms} \, .
\end{eqnarray}
The use of R-terms allows to extract immediately the leading singularities of the correlators, as we show below.
One can use several different forms of R-substitutions for a given tensor component.
For example, a rank $2$ tensor can be rewritten in R-form in several ways
\begin{eqnarray}
\frac{(x -y)_\mu(x-y)_\nu}{[(x-y)^2]^{d+1}}
&=&
{R^2}_{\mu\nu}(x-y,d+1)
\nonumber \\
&=&
{R^1}_\mu(x-y,d/2+1)\, {R^1}_\nu(x-y,d/2)
\nonumber \\
&=&
\frac{1}{(x-y)^2} \, {R^1}_\mu(x-y,d/2) \, {R^1}_\nu(x-y,d/2) \, .
\end{eqnarray}
The derivative relations in the three cases shown above are obviously different, but the transform is unique.
One can also artificially rewrite the numerators at will by introducing trivial identities in position space,
without affecting the final expression of the mapping.
We will be using this method in order to extract some of the logarithmic integrals generated by this procedure.
Obviously, this is possible only if we guarantee an intermediate regularization.
We implement it by a dimensional shift of the exponents of the propagators.
The regulator will allow to smooth out the singularity of the correlators around
the value $\alpha=d/2$, which is the critical value beyond which a function such as $1/[x^2]^{\alpha}$ is not integrable,
according to (\ref{Ch1fund}).
The structure of the singularities in position space of the corresponding scalars and tensor correlators can be
identified using the basic transform. For instance, using (\ref{Ch1fund}) for $\alpha=d/2$ one encounters a pole
in the expression of the transform. For this reason we regulate dimensionally in position space
such a singularity by shifting $d\rightarrow d- 2 \omega$. At the same time we compensate with a regularization scale $\mu$ to
preserve the dimension of the redefined correlator. A similar approach has been discussed in \cite{Dunne:1992ws}, in an attempt to relate
differential and dimensional regularization. However, in our case as in \cite{Erdmenger:1996yc} $\omega$ is an independent regulator which
serves to test integrability in momentum space, and for this reason is combined with a fundamental transform which is given by
\begin{equation}
\frac{\mu^{2 \omega}}{[x^2]^{d/2- \omega}} =
\frac{\mu^{2 \omega}}{4^{d/2- \omega} \pi^{d/2}} \, \frac{\Gamma(\omega)}{\Gamma(d/2-\omega)} \, \int d^d l \,
\frac{e^{i l\cdot x}}{[l^2]^{\omega}} \, ,
\end{equation}
that we can expand around $\omega \sim 0$ to obtain
\begin{equation} \label{Ch1second}
\frac{\mu^{2\omega}}{[x^2]^{d/2- \omega}} =
\frac{\pi^{d/2}}{\Gamma(d/2)} \, \delta^{(d)}(x) \, \left[\frac{1}{\omega} - \gamma + \log 4 + \psi (d/2) \right] -
\frac{1}{(4\pi )^{d/2 } \, \Gamma(d/2)} \, \int d^d l \,{e^{i l\cdot x}} \log\left(\frac{l^2}{\mu^2}\right) + O(\omega) \, .
\end{equation}
The subtraction of this pole in $d$ dimensions is obviously related to the need of redefining correlators which are not integrable,
in analogy with the approach followed in differential regularization.
The most popular example is $1/[x^2]^2$, which has no transform for $d=4$, but is rewritten in the derivative form as \cite{Freedman:1991tk}
\begin{equation} \label{Ch1defin}
\frac{1}{x^4}=\Box \, G(x^2)\, ,
\end{equation}
where $G(x^2)$ is defined by
\begin{equation}
G(x^2)= \frac{\log x^2 M^2}{x^2} + c \, ,
\end{equation}
with $c$ being a constant. This second approach can be easily generalized to $d$ dimensions.
One can use derivative relations such as
\begin{equation}
\frac{1}{[x^2]^\alpha}= \frac{1}{2(\alpha -1)(2 \alpha - d)} \, \Box \, \frac{1}{[x^2]^{\alpha -1}}
\label{Ch1squareeq}
\end{equation}
which is correct as far as $\alpha\neq d/2$.
For $\alpha=d/2 $ this relation misses the singularity at $x=0$, which is apparent from (\ref{Ch1deltaeq}).
For this reason, as far as $\alpha=d/2 - \omega$ eq. (\ref{Ch1squareeq}) remains valid and it can be used
together with (\ref{Ch1deltaeq}) and an expansion in $\omega$ to give
\begin{eqnarray} \label{Ch1third}
\frac{\mu^{2 \omega}}{[x^2]^{d/2- \omega}}
&=&
- \frac{1}{2\,\omega}\frac{\mu^{2 \omega}}{d - 2 - 2 \omega} \, \Box
\frac{1}{[x^2]^{d/2-1-\omega}}
\nonumber \\
&=&
\frac{1}{4 - 2\,d}\left(\frac{1}{\omega} + \frac{2}{d-2} \right)\Box \frac{1}{[x^2]^{d/2-1}}-
\frac{1}{2 (d-2)}\Box\frac{ \log (\mu^2 x^2)}{[x^2]^{d/2-1}} \nonumber \\
&=&
\frac{\pi^{d/2}}{\Gamma(d/2)} \, \left(\frac{1}{\omega}
+ \frac{2}{d-2}\right) \, \delta^{(d)}(x) - \frac{1}{2(d-2)} \, \Box\frac{\log (\mu^2 x^2)}{(x^2)^{d/2-1}}.
\end{eqnarray}
The $d$-dimensional version of differential regularization can be obtained by requiring the subtraction of all
the terms in (\ref{Ch1third}) which are proportional to $\delta^d(x)$, giving
\begin{equation}
\frac{1}{[x^2]^{d/2}} \equiv -\frac{1}{2 (d-2)}\Box\frac{\log (\mu^2 x^2)}{(x^2)^{d/2-1}}\, .
\end{equation}
This procedure clearly agrees with the traditional version of differential regularization in $d=4$ \cite{Freedman:1991tk},
\begin{equation}
\frac{1}{x^4}\equiv-\frac{1}{4}\Box\frac{\log(x^2 \mu^2)}{x^2} \, .
\end{equation}
Notice that this analysis shows that, according to (\ref{Ch1third}),
the logarithmic integral in (\ref{Ch1second}) is given by
\begin{eqnarray} \label{Ch1LogInt}
\int d^d l e^{i l \cdot x}\log \left(\frac{l^2}{\mu^2}\right)
&=&
(2\pi)^d \, \left[ - \gamma + \log 4 + \psi (d/2) - \frac{2}{d-2} \right] \, \delta^{(d)}(x)
+ \frac{(4\pi)^{d/2}}{2(d-2)}\Gamma(d/2) \, \Box \, \frac{ \log(\mu^2 x^2)}{[x^2]^{d/2-1}}
\nonumber \\
&=&
\frac{(4\pi)^{d/2}}{2(d-2)}\Gamma(d/2) \, \Box \, \frac{ \log(\bar{\mu}^2 x^2)}{[x^2]^{d/2-1}} \, ,
\end{eqnarray}
having redefined the regularization scale properly
\begin{equation} \label{Ch1massscale}
\log \bar{\mu}^2 = \log \mu^2 + \gamma - \log 4 - \psi (d/2) + \frac{2}{d-2} \, .
\end{equation}
Notice that also regulated (but singular) correlators can be mapped in several ways to momentum space, with identical results,
exactly as for no singular correlators. For instance, we can take $1/[x^2]^{d/2}$ and use on it eq. (\ref{Ch1fund}) once
\begin{eqnarray} \label{Ch1transf}
\int d^d x \, e^{i k \cdot x}\frac{1}{[x^2]^{d/2}}
\to
\int d^d x \, e^{i k \cdot x}\frac{\mu^{2\omega}}{[x^2]^{d/2-\omega}}
&=&
\frac{1}{4^{d/2-\omega}\, \pi^{d/2}}\, \frac{\Gamma(\omega)}{\Gamma(d/2- \omega)}\,
\int d^d x\, d^d l\, e^{i (k + l) \cdot x}\frac{\mu^{2\omega}}{[l^2]^{\omega}}
\nonumber \\
&=&
4^{\omega}\, \pi^{d/2}\, \frac{\Gamma(\omega)}{\Gamma(d/2- \omega)} \, \frac{\mu^{2\omega}}{[k^2]^{\omega}}\, ,
\end{eqnarray}
twice
\begin{eqnarray}
\int d^dx \, \frac{\mu^{2\omega}}{x^2 [x^2]^{d/2-1-\omega}}
&=&
\frac{1}{4^{d/2-\omega} \pi^d}\, \frac{\Gamma(d/2 - 1)\, \Gamma(1+\omega)}{\Gamma(d/2-1-\omega)}
\int d^dx\, d^d l_1\, d^d l_2\, e^{i (k+ l_1 + l_2) \cdot x} \, \frac{\mu^{2\omega}}{[l_1^2]^{d/2-1} [l_2^2]^{1+\omega}}
\nonumber \\
&=&
4^{\omega}\, \pi^{d/2}\, \frac{\Gamma(\omega)}{\Gamma(d/2- \omega)}\,
\frac{\mu^{2\omega}}{[k^2]^{\omega}}\, ,
\end{eqnarray}
(where in the last step (\ref{Ch1B0general}) was used) or any number of times, obtaining the same transform.
As one can easily work out, the use of the dimensional regulator generates, after a Laurent expansion in $\omega$, some
logarithmic integrals in momentum space. As we shall show, if the $1/\omega$ poles cancel, then these integrals can be avoided, in the sense
that it will be possible to rewrite the correlator in such a way that they are absent.
This means that in this case one has to go back and try to rewrite the correlator in such a way that it takes an explicitly finite form already
in position space. In this case the mapping of the correlators onto momentum space is similar to the usual Feynman expansion
typical of perturbation theory. The condition of Fourier transformability is obviously necessary in order to have, eventually,
a Lagrangian description of the correlator.
On the other hand, if the same poles do not cancel, then the logarithms are a significant aspect of the correlator which,
for sure, cannot be reproduced by a local field theory Lagrangian anyhow, in particular not by a free field theory.
We have left to appendix \ref{Ch1Distributional} a few more examples on the correct handling of these distributional identities.
\subsection{Regularization of tensors}
The regularization of other tensor contributions using this extension of differential regularization can be handled
in a similar and straightforward way.
The use of the derivative relations on the R-terms, that map the tensor structures into derivative of less singular terms,
combined at the last stage with the basic transform, allows to get full control of any correlator and guarantees
its consistent mapping onto momentum space.
We provide a few examples to illustrate the procedure.
Consider, for instance, the tensor structure
\begin{equation}
t_{\mu} = \frac{(x-y)_\mu}{[(x-y)^2]^{d/2 +1}} \, ,
\end{equation}
whose R-form is, trivially,
\begin{equation}
t_{\mu}
=
{R^1}_\mu\left( x-y,\frac{d}{2}+1\right)
=
- \frac{1}{d} \, \partial_\mu \, \frac{1}{[(x-y)^2]^{d/2}} \, ,
\end{equation}
where the derivative is intended with respect to $x-y$.
Now we send $d \to d - 2\omega$ in the exponent of the denominator, introducing the proper mass scale,
since $d/2$ is a critical value for integrability.
This allows us to use the basic transform (\ref{Ch1fund}), getting
\begin{equation}
t_\mu(\omega) =
- \frac{i \, \mu^{2\omega}}{(d- 2\omega) \, 4^{d/2-\omega} \, \pi^{d/2}} \,
\frac{\Gamma(\omega)}{\Gamma(d/2-\omega)} \, \int \, d^d l \, \frac{l_\mu}{[l^2]^\omega} \, e^{i l\cdot (x-y)} \, .
\end{equation}
We can expand in $\omega$ obtaining
\begin{eqnarray}
t_\mu(\omega)
&=&
\frac{i}{d \, 2^d \pi^{d/2} \, \Gamma(d/2)}
\bigg[
- \bigg(\frac{1}{\omega} + \frac{2}{d} - \gamma + \log 4 + \psi (d/2)\bigg) \,
\int d^d l\, e^{i l\cdot (x-y)}\, l_{\mu}
\nonumber\\
&&\hspace{25mm}
+ \int d^d l\, e^{i l\cdot (x-y)} \, l_{\mu}\, \log\left(\frac{l^2}{\mu^2}\right)
\bigg]
+ O(\omega) \nonumber \\
&=& \frac{\pi^{d/2}}{d\,\Gamma(d/2)} \, \partial_\mu
\bigg[
- \bigg( \frac{1}{\omega} + \frac{4(d-1)}{d(d-2)} \bigg) \,
\delta^{(d)}(x-y)
+ \frac{\Gamma(d/2)}{2(d-2)\pi^{d/2}} \, \Box \frac{\log(\bar{\mu}^2(x-y)^2)}{[(x-y)^2]^{d/2-1}}
\bigg] \, ,
\end{eqnarray}
where in the last step we have used (\ref{Ch1LogInt}).
Notice that the strength of the singularity has increased from $\delta(x)/\omega$ to $\partial_\mu\delta(x)/\omega$,
due to the higher power $(d/2)$ of the denominator in position space. It is clear that for finite correlators
these singular contributions must cancel.
In general, the introduction of the regulator $\omega$ allows to perform algorithmically the transform of any lengthy
expression, leaving its implementation to a symbolic calculus program. Obviously, for finite correlators this approach might
look redundant, but it can be extremely useful in order to check the cancellation of all the multiple and single
pole singularities in a very efficient way. We will present more examples of this approach in the next sections.
A more involved example is given by
\begin{equation}
t_{\mu\nu} = \frac{(x-y)_\mu (x-y)_\nu}{[(x-y)^2]^{d/2+1}} \, ,
\end{equation}
to which corresponds the regulated expression
\begin{equation}
t_{\mu\nu}(\omega)= \frac{\mu^{2\omega}(x-y)_\mu (x-y)_\nu}{[(x-y)^2]^{d/2+1 - \omega}}
\end{equation}
and a minimal R-form given by
\begin{equation}
t_{\mu\nu}(\omega) = \mu^{2\omega} \, {R^2}_{\mu\nu}\left(x - y,\frac{d}{2} + 1 - \omega\right) \, .
\end{equation}
Using the list of replacements given in (\ref{Ch1Rterms}), the derivative form of $t_{\mu\nu}$ is given by
\begin{equation}
t_{\mu\nu }(\omega)= \frac{\mu^{2\omega}}{(d - 2 - 2 \omega) \, (d - 2\omega)} \, \partial_\mu \, \partial_\nu \,
\frac{1}{[(x-y)^2]^{d/2 - \omega -1}}
+ \frac{\delta_{\mu\nu}}{d + 2 - 2\omega} \, \frac{\mu^{2\omega}}{[(x-y)^2]^{d/2 - \omega}} \, ,
\end{equation}
whose singularities are all contained in the second term, whose Fourier transform is given by
\begin{equation}
\mathcal{FT}\bigg[\frac{\delta_{\mu\nu}}{d + 2 - 2\omega} \, \frac{\mu^{2\omega}}{[(x-y)^2]^{d/2 - \omega}}\bigg]
= \frac{1}{\omega} \, \frac{\delta_{\mu\nu}}{2^d \, \pi^{d/2} \, (d+2) \, \Gamma(d/2)} + O(\omega^0) \, ,
\end{equation}
where we have omitted the regular terms. The procedure therefore allows to identify quite straightforwardly
the leading singularities of any tensor in position space, giving, in this specific case,
\begin{equation}
\frac{(x-y)_\mu (x-y)_\nu}{[(x-y)^2]^{d/2+1 - \omega}} \sim
\frac{1}{\omega} \, \frac{\delta_{\mu\nu}}{2^d \, \pi^{d/2} \, (d+2) \, \Gamma(d/2)} \, .
\end{equation}
We can repeat the procedure for correlators of higher rank. The singularities, after performing all the substitutions,
are proportional to the non-derivative terms isolated by the repeated replacement of eq. (\ref{Ch1Rterms}).
\subsection{Regularization of $3$-point functions}
In the case of $3$-point functions, the analysis of the corresponding singularities can be extracted quite simply. \\
Let us consider, for instance, the identity
\begin{eqnarray} \label{Ch1fundFor3}
&&
\mathcal{FT}\bigg[\frac{1}{[(x-y)^2]^{\alpha_1}[(z-x)^2]^{\alpha_2}[(y-z)^2]^{\alpha_3 }}\bigg]
\equiv
\int \, d^d x \, d^d y \, d^d z \,
\frac{e^{- i(k\cdot z+ p\cdot x + q\cdot y)}}{[(x-y)^2]^{\alpha_1}[(z-x)^2]^{\alpha_2}[(y-z)^2]^{\alpha_3 }}
\nonumber \\
&=&
(2\pi)^{3d} \, \prod_{i=1}^{3}\left( \frac{\Gamma(d/2 - \alpha_i)}{4^{\alpha_i} \pi^{d /2}\Gamma(\alpha_i)}\right) \,
\delta^{(d)}(k+p+q)\,
\int \frac{d^dl}{[l^2]^{d/2-\alpha_1} [(l+p)^2]^{d/2-\alpha_2} [(l-q)^2]^{d/2-\alpha_3}}\, ,\nonumber \\
\end{eqnarray}
obtained using the fundamental transform (\ref{Ch1fund}), where all the physical momenta $(k,p,q)$ are treated as incoming.
The convention for matching the momenta in (\ref{Ch1fund}) with the couples of coordinate is
\begin{equation}
l_1 \leftrightarrow x-y\, , \qquad l_2 \leftrightarrow z-x\, , \qquad l_3 \leftrightarrow y-z \, ,
\end{equation}
and the shift $ l \rightarrow l-q $ (which is always possible in a regularized expression) has been performed at the end. \\
It is clear that the pre-factor on the r.h.s. of this relation has poles for $\alpha_i = d/2 + n$, with $n \geq 0$.
At the same time the loop integral is asymptotically divergent if $d = \sum_i \alpha_i $, where it develops a logarithmic singularity.
In dimensional regularization such a singularity corresponds to a single pole in $\epsilon= d -\sum_i \alpha_i$.
One can be more specific by discussing further examples of typical $3$-point functions. \\
For instance, consider the tensor structure
\begin{equation}
{\mathcal{Q}^1}_{\alpha\beta\mu\nu} =
\frac{(y-z)_\alpha \, (y-z)_\beta \, (y-z)_\mu \, (y-z)_\nu}
{[(x-y)^2]^{d/2+1} \, [(z-x)^2]^{d/2-1} \, [(y-z)^2]^{d/2+1}} \, ,
\end{equation}
which appears in the $TVV$ correlator and can be reduced to its R-form in several ways.
We use a minimal substitution and have
\begin{equation}
{\mathcal{Q}^1}_{\alpha\beta\mu\nu} = \frac{1}{[(x-y)^2]^{d/2+1}} \, \frac{1}{[(z-x)^2]^{d/2-1}} \,
{R^4}_{\alpha\beta\mu\nu}\left(y-z,\frac{d}{2}+1\right) \, ,
\end{equation}
after which an application of the derivative reductions in (\ref{Ch1Rterms}) gives
\begin{eqnarray}
{\mathcal{Q}^1}_{\alpha\beta\mu\nu}
&=&
\frac{1}{(d-6)\,(d-4)\,(d-2)\,d}\, \frac{1}{[(x-y)^2]^{d/2+1}} \, \frac{1}{[(x-z)^2]^{d/2 - 1}}
\nonumber \\
&\times&
\bigg\{
\partial_\alpha \, \partial_\beta \, \partial_\mu \, \partial_\nu \, \frac{1}{[(y-z)^2]^{d/2 - 3}}
+ (d-6)\,(d-4) \,
\frac{\delta_{\mu\nu}\,\delta_{\alpha\beta} + \delta_{\mu\alpha}\,\delta_{\nu\beta} + \delta_{\mu\beta}\,\delta_{\nu\alpha}}
{[(y - z)^2]^{d/2 - 1}}
\nonumber \\
&+&
(d-6) \, \left(
\delta_{\mu\nu} \, \partial_\alpha \, \partial_\beta + \delta_{\alpha\beta} \, \partial_\mu \, \partial_\nu
+ \delta_{\mu\alpha} \, \partial_\nu \, \partial_\beta + \delta_{\nu\beta} \, \partial_\mu \, \partial_\alpha
+ \delta_{\nu\alpha} \, \partial_\mu \, \partial_\beta + \delta_{\mu\beta} \, \partial_\nu \, \partial_\alpha
\right) \, \frac{1}{[(y-z)^2]^{d/2-2}}
\bigg\} \, .
\nonumber\\
\end{eqnarray}
Before moving to momentum space, a quick glance at this equation shows that its transform does not exist.
This appears obvious from the presence of the overall factor $1/[(x-y)^2]^{d/2+1}$ which needs regularization.
The mapping can be performed using the rules defined above, which give, for instance, for the coefficient of
$\delta_{\mu\nu}\, \delta_{\alpha\beta} + \delta_{\mu\alpha}\, \delta_{\nu\beta} + \delta_{\mu\beta}\, \delta_{\nu\alpha}$,
\begin{eqnarray}
&&
\mathcal{F T}
\bigg[
\frac{1}{d\,(d-2)} \,
\frac{\mu^{2\omega}}{[(x-y)^2]^{d/2 + 1 - \omega} \, [(z-x)^2]^{d/2 - 1} \, [(y-z)^2]^{d/2 - 1}}
\bigg]
\nonumber \\
&=&
\frac{(2\pi)^{3d} \, \delta^{(d)}(k+p+q)}{d\,(d-2)} \, \frac{4^{1+\omega}}{(4\pi)^{3d/2}}\,
\frac{\Gamma(\omega-1)}{\Gamma(d/2-1)^2 \,\Gamma(d/2-1-\omega)} \,
\int \, d^d l \, \frac{\mu^{2\omega}}{(l^2)^{\omega-1}\,(l+p)^2\,(l-q)^2}\nonumber
\end{eqnarray}
\begin{eqnarray}
&=&
\frac{\delta^{(d)}(k+p+q)}{d(d-2)}\, \frac{4\,\pi^{3d/2}}{\Gamma(d/2-1)^3}\,
\bigg[
-\frac{1}{\omega} \int d^dl\, \frac{l^2}{(l+p)^2(l-q)^2}
+\, \int d^dl\, \frac{l^2\, \log\left(l^2/\bar{\mu}^2\right)}{(l+p)^2(l-q)^2}
\bigg]
+ O(\omega). \nonumber \\
\end{eqnarray}
In a similar way, the Fourier transform of the first term is
\begin{eqnarray}
&&
\mathcal{F T}\bigg[\frac{1}{(d-6)\,(d-4)\,(d-2)\,d}\, \frac{\mu^{2\omega}}{[(x-y)^2]^{d/2+1-\omega}} \,
\frac{1}{[(z-x)^2]^{d/2-1}} \, \partial_\mu\,\partial_\nu\,\partial_\alpha\,\partial_\beta \, \frac{1}{[(y-z)^2]^{d/2-3}}\bigg] \nonumber \\
&=&
\frac{(2\pi)^{3d}\,\delta^{(d)(k+p+q)}}{(d-6)\,(d-4)\,(d-2)\,d}\, \frac{4^{3+\omega}}{(4\pi)^{3d/2}}\,
\frac{2\,\Gamma(\omega-1)}{\Gamma(d/2-3)\,\Gamma(d/2-1)\,\Gamma(d/2+1-\omega)}
\nonumber \\
&\times&
\int d^dl\, \frac{(l-q)_\alpha\,(l-q)_\beta\,(l-q)_\mu\,(l-q)_\nu}{(l^2)^{\omega-1}\,(l+p)^2\,[(l-q)]^3}
\nonumber \\
&=&
\frac{\delta^{(d)(k+p+q)}}{d\,(d-2)}\,\frac{32\,\pi^{3d/2}}{\Gamma(d/2-1)^3}\,
\bigg[
- \frac{1}{\omega}\, \int d^dl\, \frac{l^2\,(l-q)_\alpha\,(l-q)_\beta\,(l-q)_\mu\,(l-q)_\nu}{(l+p)^2[(l-q)^2]^3}
\bigg]
\nonumber\\
&& \hspace{40mm}
+\, \int d^dl\, \frac{\log\left(l^2/\bar{\mu}^2\right)\, (l-q)_\alpha\,(l-q)_\beta\,(l-q)_\mu\,(l-q)_\nu}
{(l+p)^2[(l-q)^2]^3} + O(\omega)\, ,
\end{eqnarray}
illustrating quite clearly how the general procedure can be implemented.
Of course, the regularization can be performed by sending $d\to d - 2 \omega$ - with no distinction among the
various terms - or, alternatively, one can regulate only the non integrable terms. The two approaches, in a generic
computation, will differ only at $O(\omega)$ and as such they are equivalent.
Another important comment concerns the possibility of performing an explicit computation of the logarithmic integrals.
They are indeed calculable in terms of generalized hypergeometric functions (for general $\omega$), but the small $\omega$
expansion of these functions is rather difficult to re-express as a combination of ordinary functions and polylogs.
This is due to the need of performing a double expansion (in $\epsilon$ and in $\omega$) if we move to $d=4$ and insist, as we
should, on the use of dimensional regularization in the computation of the momentum integrals. This difficulty is attributed to
the absence of simple expansions of hypergeometric functions (ordinary and generalized) about non integer (real) values of their indices.
However, if the $1/\omega$ terms for a combination of terms similar to those shown above cancel,
there are some steps which can be taken in order to simplify this final part of the computation.
\subsection{Application to the $VVV$ case}
To illustrate the general procedure through a specific example, we reconsider the $VVV$ case, that we know to be integrable.
We expand the position space correlator and perform the R-substitutions (\ref{Ch1Rterms}). The direct algorithm gives
an expression which is not immediately recognized as being integrable and is
\begin{eqnarray} f^{abc} \, \bigg\{\frac{(a - 2 \, b)}{(d-2)^3} \,
&\times&
\bigg[ \partial^{31}_\mu \, \frac{1}{(x^2_{31})^{d/2-1}}
\, \partial^{12}_\nu \, \frac{1}{(x^2_{12})^{d/2-1}} \, \partial^{23}_\rho \, \frac{1}{(x^2_{23})^{d/2-1}} \nonumber \\
&+&
\partial^{12}_\mu \, \frac{1}{(x^2_{12})^{d/2-1}} \, \partial^{23}_\nu \, \frac{1}{(x^2_{23})^{d/2-1}} \, \partial^{31}_\rho \,
\frac{1}{(x^2_{31})^{d/2-1}} \bigg] \nonumber \\
+ \frac{a}{d \, (d-2)^2} \,
&\times&
\bigg[ \frac{1}{(x^2_{12})^{d/2-1}} \left(
\partial^{31}_\mu \, \frac{1}{(x^2_{31})^{d/2-1}} \, \partial^{23}_\nu \, \partial^{23}_\rho \, \frac{1}{(x^2_{23})^{d/2-1}} +
\partial^{23}_\nu \, \frac{1}{(x^2_{23})^{d/2-1}} \, \partial^{31}_\mu \, \partial^{31}_\rho \, \frac{1}{(x^2_{31})^{d/2-1}} \right)
\nonumber \\
&+&
\frac{1}{(x^2_{23})^{d/2-1}} \left(
\partial^{31}_\rho \, \frac{1}{(x^2_{31})^{d/2-1}} \, \partial^{12}_\mu \, \partial^{12}_\nu \, \frac{1}{(x^2_{12})^{d/2-1}} +
\partial^{12}_\nu \, \frac{1}{(x^2_{12})^{d/2-1}} \, \partial^{31}_\mu \, \partial^{31}_\rho \, \frac{1}{(x^2_{31})^{d/2-1}}
\right)
\nonumber \\ &+&
\frac{1}{(x^2_{31})^{d/2-1}} \left(
\partial^{23}_\rho \, \frac{1}{(x^2_{23})^{d/2-1}} \, \partial^{12}_\mu \, \partial^{12}_\nu \, \frac{1}{(x^2_{12})^{d/2-1}} +
\partial^{12}_\mu \, \frac{1}{(x^2_{12})^{d/2-1}} \, \partial^{23}_\nu \, \partial^{23}_\rho \, \frac{1}{(x^2_{23})^{d/2-1}}
\right) \bigg]
\nonumber \\
- \frac{1}{d-2} \, \left( b - \frac{a}{d+2} \right) \,
&\times&
\bigg[ \frac{1}{(x^2_{31})^{d/2-1}} \, \left(
\frac{\delta_{\mu\nu}}{(x^2_{12})^{d/2}} \, \partial^{23}_\rho \, \frac{1}{(x^2_{23})^{d/2-1}} +
\frac{ \delta_{\nu\rho}}{(x^2_{23})^{d/2}} \, \partial^{12}_\mu\, \frac{1}{(x^2_{12})^{d/2-1}}\right) \nonumber \\
&+&
\frac{1}{(x^2_{23})^{d/2-1}} \, \left(
\frac{\delta_{\mu\nu}}{(x^2_{12})^{d/2}} \, \partial^{31}_\rho \, \frac{1}{(x^2_{31})^{d/2-1}} +
\frac{\delta_{\mu\rho}}{(x^2_{31})^{d/2}} \, \partial^{12}_\nu\, \frac{1}{(x^2_{12})^{d/2-1}} \right) \nonumber \\ &+&
\frac{1}{(x^2_{12})^{d/2-1}} \, \left(
\frac{\delta_{\mu\rho}}{(x^2_{31})^{d/2}} \, \partial^{23}_\nu \, \frac{1}{(x^2_{23})^{d/2-1}} +
\frac{\delta_{\nu\rho}}{(x^2_{23})^{d/2}} \, \partial^{31}_\mu\, \frac{1}{(x^2_{31})^{d/2-1}}\right) \bigg] \bigg\} \, .
\end{eqnarray}
The apparent non-integrability is due to terms of the form ${1}/{(x^2_{ij})^{d/2}}$ in the last addend.
For this reason, ignoring any further information, to test the approach we proceed with a regularization of the non-integrable terms.
The expression in momentum space is obtained by sending $ d \to d - 2\omega$ in all the terms of the form
${1}/{(x^2_{ij})^{d/2}}$. Expanding in $\omega$ the result, one can show that, as expected, the $1/\omega$ terms cancel,
proving its integrability. We fill in few more details to clarify this point.
A typical not manifestly integrable term in $VVV$ is
\begin{equation}
\frac{1}{(x^2_{31})^{d/2-1}} \,
\frac{1}{(x^2_{12})^{d/2}} \, \partial^{23}_\rho \, \frac{1}{(x^2_{23})^{d/2-1}} +
\frac{1}{(x^2_{23})^{d/2-1}} \,
\frac{1}{(x^2_{12})^{d/2}} \, \partial^{31}_\rho \, \frac{1}{(x^2_{31})^{d/2-1}} \, ,
\end{equation}
which in momentum space after $\omega$ regularization gives (omitting an irrelevant constant)
\begin{equation}
\mu^{2 \omega}\, \Gamma(\omega)\, \int d^d l\, \frac{ 2 l^\rho - q^\rho}{(l^2) (l-q)^2 [(l+p)^2]^{\omega}}.
\end{equation}
Expanding in $\omega$, the residue of the pole is given by the integral
\begin{equation}
\int d^d l\frac{2 l^\rho - q^\rho}{l^2 (l-q)^2}
\end{equation}
which vanishes in dimensional regularization.
The finite term is logarithmic,
\begin{equation}
\int d^d l\frac{\log \left((l+p)^2/\mu^2\right)\, (2 l^\rho - q^\rho)}{l^2 (l-q)^2} \, .
\end{equation}
The scale dependence also disappears, since the $\log\mu^2$ term is also multiplied by the same vanishing integral. Obviously, the
non trivial part of the computation is in the appearance of a finite logarithmic integral which, due to the finiteness of the
correlator, has to be re-expressed in terms of other non-logarithmic contributions, i.e. of ordinary Feynman integrals.
There is no simple way to relate one single integral to an ordinary non-logarithmic contribution unless one performs the entire
computation and expresses the result in terms of special polylogarithmic functions, using consistency. For correlators which are
integrable, however, it is possible to relate two log integrals to regular Feynman integrals.
Single log integrals, at least in this case, can also be evaluated explicitly, as we illustrate in appendix \ref{Ch1Distributional}.
By applying the algorithm we get
\begin{eqnarray}
&&
\left\langle {V^a}_{\mu} \, {V^b}_{\nu} \, {V^c}_{\rho} \right\rangle \, (p,q) =
(2\pi)^{3d} \, \delta^{(d)}(k+p+q) \, i \, f^{abc}
\nonumber \\
&\times&
\bigg\{
C(d/2-1)^3 \,
\bigg[
\frac{a\,(6-4d)+2\,b\,d}{d(d-2)^3}
\bigg(2 \, J_{\mu\nu\rho}(p,-q) + (p+q)_{\mu} \, J_{\nu\rho}(p,-q) + p_{\nu} \, J_{\mu\rho}(p,-q)
\nonumber \\
&& \hspace{50mm}
- \, q_{\rho} \, J_{\mu\nu}(p,-q) - p_{\nu}\,q_{\mu}\,J_{\rho}(p,-q) - p_{\mu}\,q_{\rho}\,J_{\nu}(p,-q)\bigg)
\bigg]
\nonumber \\
&+&
\frac{a}{d(d - 2)^2} \,
\bigg(- 2 \, \left(p_\mu + q_\mu\right) \, \big( p_\nu\,J_\rho(p,-q) + q_\rho\,J_\nu(p,-q)\big)
+ q_\rho p_\nu \big(2 \, J_\mu(p,-q) + (p-q)_\mu \, J(p,-q) \big) \bigg)
\nonumber \\
&-&
\frac{C(d/2-1)^2}{(4\pi)^{d/2}\,\Gamma(d/2)\,(d-2)} \,
\bigg(\frac{a}{d+2} - b \bigg) \,
\bigg[
\delta_{\mu\nu} \, \bigg( 2 \, IL_\rho(p,0,-q) - q_\rho \, IL(p,0,-q)\bigg) \nonumber \\
&&
\hspace{54mm}
+ \, \delta_{\mu\rho} \, \bigg( 2 \, IL_\nu(-q,0,p) + p_\nu \, IL(-q,0,p)\bigg)
\nonumber \\
&&
\hspace{54mm}
+ \, \delta_{\nu\rho} \, \bigg( 2 \, IL_\mu(q,0,k) + k_\mu \, IL(q,0,k)\bigg)
\bigg]
\bigg\} \, .
\end{eqnarray}
The notations introduced for the momentum space integrals here and in the following point are explained in appendix \ref{Ch1Distributional}.
One can easily show the scale independence of the result, which is related to the finiteness of the expressions
and to the fact that the logarithmic contributions, in this case, are an artefact of the approach.
For this reason, when the scale independence of the regulated expressions has been proven, then one can go back and try to
rewrite the correlator in such a way that it is manifestly integrable. Obviously this may not be straightforward,
especially if the expression in position space is given by hundreds of terms. If, even after proving the finiteness of the expression,
one is unable to rewrite it in an integrable form, one can always apply the algorithm that we have presented, generating the logarithmic integrals.
Pairs of log integrals can be related to ordinary Feynman integrals by applying appropriate tricks.
We have illustrated in an appendix an example where we discuss the computation of the single log-integral appearing in $VVV$.
In the case of the $TOO$ one encounters both single and double-log integrals.
\subsection{Application to the $TOO$ case and double logs}
A similar analysis can be pursued in the $TOO$ case. Also for this correlator we can apply a direct approach in order to show
the way to proceed in the test of its regularity. Using our basic transform (\ref{Ch1fundFor3}) and introducing the regulator
$\omega$ to regulate the intermediate singularities, we can easily transform it to momentum space
\begin{eqnarray}
&&
\mathcal{FT}\bigg[ \left\langle T_{\mu\nu}(x_1) \, O(x_2) \, O(x_3) \right\rangle \bigg]
\equiv
\left\langle T_{\mu\nu} \, O \, O \right\rangle(p,q) =
(2\pi)^{3d} \, \delta^{(d)}(k+p+q) \, a
\nonumber\\
&\times&
\bigg\{
\frac{C(d/2-1)^3}{d\,(d-2)^2}\,
\bigg[
- 4\, (d-1)\, J_{\mu\nu}(p,-q) - 2 \, (d-1) \, \bigg( (q_\nu - p_\nu)\, J_\mu(p,-q) + (q_\mu - p_\mu)\, J_\nu(p,-q) \bigg)
\nonumber \\
&+&
\bigg(d\,(p_\mu q_\nu + p_\nu q_\mu ) - (d-2)\, (p_\mu p_\nu + q_\mu q_\nu) \bigg) \, J(p,-q)
\bigg]
\nonumber \\
&+&
\frac{C(d/2-1)^2\, C(d/2-\omega)}{d}\, \delta_{\mu\nu}\,
\bigg(\int\,d^dl\,\frac{\mu^{2\omega}}{l^2[(l+p)^2]^{\omega}(l-q)^2}
+ \int\,d^dl\,\frac{\mu^{2\omega}}{l^2(l+p)^2[(l-q)^2]^{\omega}} \bigg)
\nonumber \\
&-&
\frac{C(d/2-1)\, C(d/2-\omega)^2}{d}\, \delta_{\mu\nu}\,
\int\,d^dl\,\frac{(\mu^{2\omega})^2}{[l^2]^2[(l+p)^2]^{\omega}[(l-q)^2]^{\omega}}
\bigg\} \, .
\end{eqnarray}
Once we perform an expansion in $\omega$, the expression above is affected by double and single poles,
which are expected to vanish so as to guarantee a finite result.
The coefficient of the double pole is easily seen to take the form
\begin{equation}
- \delta_{\mu\nu}\, \frac{a\, (2\,\pi^2)^d\, d\, C(d/2-1)}{\Gamma(d/2)^2} \, I(0) \, ,
\end{equation}
where the integral vanishes in dimensional regularization, being a massless tadpole. \\
The coefficient of the simple pole is instead given by
\begin{eqnarray} \label{Ch1simple}
&&
\delta_{\mu\nu}\, \frac{4^d\, \pi^{5d/2}\, C(d/2-1)^2}{d\, \Gamma(d/2)}\,
\bigg\{
\frac{1}{\Gamma(d/2-2) \, \Gamma(d/2)^2} \,
\bigg[
2 \, \bigg( \gamma - \log 4 - \psi (d/2) \bigg) \, I(0)
\nonumber \\
&+&
\bigg( IL(p,0,0) + IL(-q,0,0) \bigg) \bigg] +
\frac{1}{\Gamma(d/2-1)^2 \, \Gamma(d/2)} \,
\bigg[
I(p) + I(q)
\bigg]
\bigg\} \, .
\end{eqnarray}
The first term of (\ref{Ch1simple}) vanishes as in the case of the double pole, while for the remaining contributions we use the relation
\begin{eqnarray}\label{Ch1IntegralTrick}
IL(p,0,0) = \int d^dl\,\frac{\log\left(\frac{(l+p)^2}{\mu^2}\right)}{[l^2]^2} =
- \frac{\partial}{\partial\omega} \, \int d^d l \, \frac{\mu^{2\omega}}{[l^2]^2 \, [(l+p)^2]^\omega} \bigg|_{\omega=0} \, .
\end{eqnarray}
It is easy to see that the contributions in the last line in (\ref{Ch1simple}) cancel
after inserting the explicit value for the $2$-point function in (\ref{Ch1B0general}).
The finite part of the expression is found to be, after removing some additional tadpoles,
\begin{eqnarray}
&&
\left\langle T_{\mu\nu} \, O \, O \right\rangle(p,q) =
(2\pi)^{3d} \, \delta^{(d)}(k+p+q)\, a\,
\nonumber \\
&\times&
\bigg\{
\frac{C(d/2-1)^3}{d\,(d-2)^2}\,
\bigg[
- 4\, (d-1)\, J_{\mu\nu}(p,-q) - 2 \, (d-1) \, \bigg( (q_\nu - p_\nu)\, J_\mu(p,-q) + (q_\mu - p_\mu)\, J_\nu(p,-q) \bigg)
\nonumber \\
&&\hspace{25mm}
+\, \bigg(d\,(p_\mu q_\nu + p_\nu q_\mu ) - (d-2)\, (p_\mu p_\nu + q_\mu q_\nu) \bigg) \, J(p,-q) \bigg]
\nonumber \\
&-&
\delta_{\mu\nu}\,
\bigg[
\frac{C(d/2-1)^2}{d\,\pi^{d/2}\,2^{d}\,\Gamma(d/2)} \,
\bigg( \left(\gamma -\log 4 -\psi (d/2)\right)\, \big(I(p) + I(-q) \big) + \big(IL(p,0,-q) + IL(-q,0,p) \big) \bigg)
\nonumber \\
&& \hspace{-2mm}
+ \, \frac{C(d/2-1)}{3\,d\,2^{2d+1}\, \pi^d\,\Gamma(d/2)^2}\,
\bigg(
12\, \left(\gamma - \log 4 - \psi (d/2)\right)\, \big(IL(p,0,0) + IL(-q,0,0)\big)
\nonumber \\
&&\hspace{33mm}
+\, 3 \, \big(ILL(p,p,0,0) + 2\, ILL(p,-q,0) + ILL(-q,-q,0,0) \big)
\bigg)
\bigg]
\bigg\} \, ,
\end{eqnarray}
where now also double logarithmic integrals have appeared. Using the relations (\ref{Ch1B0general}) and (\ref{Ch1IntegralTrick}),
the terms proportional to $(\gamma - \log 4 - \psi (d/2))$, which are just a remain of the regularization procedure, cancel out,
leaving us with the simplified result
\begin{eqnarray}
\label{Ch1toologs}
&&
\left\langle T_{\mu\nu} \, O \, O \right\rangle(p,q) =
(2\pi)^{3d} \, \delta^{(d)}(k+p+q)\, a\,
\nonumber \\
&\times&
\bigg\{
\frac{C(d/2-1)^3}{d\,(d-2)^2}\,
\bigg[
- 4\, (d-1)\, J_{\mu\nu}(p,-q) - 2 \, (d-1) \, \bigg( (q_\nu - p_\nu)\, J_\mu(p,-q) + (q_\mu - p_\mu)\, J_\nu(p,-q) \bigg)
\nonumber \\
&&\hspace{25mm}
+\, \bigg(d\,(p_\mu q_\nu + p_\nu q_\mu ) - (d-2)\, (p_\mu p_\nu + q_\mu q_\nu) \bigg) \, J(p,-q) \bigg]
\nonumber \\
&-&
\delta_{\mu\nu}\, \frac{C(d/2-1)}{d\,(4\pi)^d\,\Gamma(d/2)^2}
\bigg[
(4\pi)^{d/2}\,\Gamma(d/2) \,C(d/2-1)\, \bigg(\big(IL(p,0,-q) + IL(-q,0,p) \big) \bigg)
\nonumber \\
&& \hspace{30mm}
+ \, (d-4)\, \big(ILL(p,p,0) + 2\, ILL(p,-q,0,0) + ILL(-q,-q,0,0) \big)
\bigg]
\bigg\}.
\end{eqnarray}
It is slightly lengthy but quite straightforward to show that (\ref{Ch1toologs}) can be re-expressed in terms of ordinary Feynman integrals.
This can be obtained by reducing all the tensor integrals (logarithmic and non-logarithmic) to scalar form.
After the reduction, one can check directly that specific combinations of logarithmic integrals can be expressed in terms of ordinary master
integrals. In this case these relations hold since the integrands of the logarithmic expansion (linear combinations thereof) are equivalent to
non-logarithmic ones, given the finiteness of the correlators.
Obviously for a correlator which is not integrable such a correspondence does not exist and the logarithmic integrals cannot be avoided.
This would be another signal, obviously, that the theory does not have a realization in terms of a local Lagrangian,
since a Lagrangian field theory has a diagrammatic description only in terms of ordinary Feynman integrals.
We conclude this section with few more remarks concerning the treatment of correlators with more general scaling dimensions.
For instance one could consider correlators of the generic form
\begin{equation}
\left\langle O_i(x_i) O_j(x_j)O_k(x_k)\right\rangle =
\frac{\lambda_{ijk}}{((x_i-x_j)^2)^{\Delta_i + \Delta_j - \Delta_k} ((x_j - x_k)^2)^{\Delta_j +
\Delta_k - \Delta_i}((x_k-x_i)^2)^{\Delta_k + \Delta_i - \Delta_j}}.
\end{equation}
In this case their expression in momentum space can be found by applying Mellin-Barnes methods. They can be reconducted to
integrals in momentum space of the form
\begin{equation}
J(\nu_1,\nu_2,\nu_3)=\int \frac{d^d l}{(l^2)^{\nu_1}((l-k)^2)^{\nu_2}((l+p)^2)^{\nu_3}} \, ,
\end{equation}
\begin{equation}
\nu_1=d/2-\Delta_i - \Delta_j + \Delta_k \, ,
\qquad
\nu_2=d/2-\Delta_j - \Delta_k + \Delta_i \, ,
\qquad
\nu_3=d/2-\Delta_k - \Delta_i + \Delta_j \, ,
\end{equation}
which can be expressed \cite{Davydychev:1992xr} in terms of generalized hypergeometric functions $F_4[a,b,c,d;x,y]$
of two variables $(x,y)$, the two ratios of the 3 external momenta. The computation of these integrals with arbitrary exponents at the
denominators is by now standard lore in perturbation theory, with recursion relations which allow to relate shifts in the exponents
in a systematic way. The problem is more involved for correlators which require an intermediate regularization in order to be transformed
to momentum space. In this case one can show, in general, that the pole structure (in $1/\omega$) of these can be worked out closely,
but the finite $O(1)$ contributions involve derivatives of generalized hypergeometric functions respect to their indices $a,b,c,d$.
Only in some cases the latter can be re-expressed in terms of polylogarithmic functions, which are typical and common
in ordinary perturbation theory. The possibility to achieve this is essentially related to finding simple expansions of the hypergeometric functions
around non integer (and not just rational) indicial points.
For integrable correlators, the analysis of Mellin-Barnes methods remains, however, a significant option, which will probably deserve a closer look.
\section{Conclusions}
The work presented in this chapter has the main goal to close the gap between position space and momentum space
analysis of $3$-point CFT's correlators characterized by the presence of one and three EMT's,
and to provide a general method to establish whether correlators built in position space CFT's admit or not a Lagrangian formulation.
We have tried to map position space and momentum space approaches, showing their interrelation, using free field theory
realizations of the general solutions of these correlators in order to establish their expression in momentum space.
A parallel has been drawn between the approach to renormalization typical of standard perturbation theory and the same
approach based on the solution of the anomalous Ward identities, as discussed in \cite{Osborn:1993cr,Erdmenger:1996yc}.
As a non trivial test of the equivalence of both methods in $4$ dimensions, we have verified that the counterterms predicted
by the general analysis in position space coincide with those obtained from momentum space in the Lagrangian predictions derived from
$1$-loop free field theory calculations.
We have also discussed a, in the second part, a general algorithm that should prove useful to regulate and
map correlators from position space to momentum space, and we have illustrated how to perform such a mapping
in a systematic way with a number of examples.
The method can be applied to the analysis of more complex correlators.
The power of the approach has been shown by re-analyisng conformal correlators investigated in the first part,
offering a complete test of its consistency.
\clearpage{\pagestyle{empty}\cleardoublepage}
\chapter{Dilaton interactions and the anomalous breaking of scale invariance in the Standard Model}\label{Effective}
\section{Introduction}
In this chapter we discuss the main features of dilaton interactions for fundamental and effective dilaton fields.
In particular, we elaborate on the various ways in which dilatons can couple to the Standard Model and on the role played by the conformal
anomaly as a way to characterize their interactions.
In the case of a dilaton derived from a metric compactification (graviscalar), we present the structure of the radiative corrections
to its decay into two photons, a photon and a $Z$, two $Z$ gauge bosons and two gluons, together with their renormalization
properties. We prove that, in the electroweak sector, the renormalization of the theory is guaranteed only if the Higgs is
conformally coupled. For such a dilaton, its coupling to the trace anomaly is quite general, and determines, for instance,
an enhancement of its decay rates into two photons and two gluons.
We then turn our attention to theories containing a non-gravitational (effective) dilaton, which, in our perturbative analysis,
manifests as a pseudo-Nambu Goldstone mode of the dilatation current ($J_D$). The infrared coupling of such a state to the
$2$-photons and to the $2$-gluons sector, together with the corresponding anomaly enhancements of its decay rates in these channels,
is critically analysed. \\
Dilatons are part of the low energy effective action of several different types of theories, from string theory to theories with compactified extra
dimensions, but they may appear also in appropriate bottom-up constructions. For instance, in scale invariant extensions of the Standard Model,
the introduction of a dilaton field allows to recover scale invariance, which is violated by the Higgs potential, by introducing a new, enlarged,
Lagrangian. This is characterized both by a spontaneous breaking of the conformal and of the electroweak symmetries.
In this case, one can formulate simple scale invariant extensions of the potential which can accommodate, via spontaneous breaking,
two separate scales: the electroweak scale ($v$), related to the vev of the Higgs field, and the conformal symmetry breaking scale
($\Lambda$), related to the vev of a new field $\Sigma=\Lambda + \rho$, with $\rho$ being the dilaton. The second scale can be
fine-tuned in order to proceed with a direct phenomenological analysis and is, therefore, of utmost relevance in the search for new
physics at the LHC.
In a bottom-up approach, and this will be one of the main points that we will address in our analysis, the dilaton of the effective
scale invariant Lagrangian can also be interpreted as a composite scalar, with the dilatation current taking the role of an operator
which interpolates between this state and the vacuum. We will relate this interpretation to the appearance of an anomaly pole in the
correlation function involving the dilatation current ($J_D$)
and two neutral currents ($V, V'$) of the Standard Model, providing evidence, in the ordinary perturbative picture, in favour of such
a statement.
One of the main issues which sets a difference between the various types of dilatons is, indeed, the contribution coming from the
anomaly, which is expected to be quite large. Dilatons obtained from compactifications with large extra dimensions and a low gravity
scale, for instance, carry this coupling, which is phenomenologically relevant.
The same coupling is present in the case of an effective dilaton, appearing as a Goldstone mode of the dilatation current, with some
differences that we will specify in a second part of the chapter. The analysis will be carried out in analogy to the pion case, which in
a perturbative picture is associated with the appearance of an anomaly pole in the $AVV$ diagram (with $A$ being the axial current).
This chapter is organized as follows. In a first part we will characterize the leading $1$-loop interactions of a dilaton derived from a
Kaluza-Klein compactification of the gravitational metric. The set-up is analogous to that presented in
\cite{Han:1998sg,Giudice:2000av} for a compactified theory with large extra dimensions and it involves all the neutral currents of
the Standard Model. We present also a discussion of the same interaction in the QCD case for off-shell gluons.
These interactions are obtained by tracing the $TVV$ vertex, with $T$ denoting the (symmetric and conserved)
energy-momentum tensor (EMT) of the Standard Model. This study is accompanied by an explicit proof of the renormalizability of these
interactions in the case of a conformally coupled Higgs scalar.
In a second part then we turn our discussion towards models in which dilatons are introduced from the ground up, starting with simple
examples which should clarify - at least up to operators of dimension 4 - how one can proceed with the formulation of scale invariant
extensions of the Standard Model. Some of the more technical material concerning this point has been left to the appendices, where
we illustrate the nature of the coupling of the dilaton to the mass dependent terms of the corresponding Lagrangian. The goal of
these technical additions is to clarify that a fundamental (i.e. not a composite) dilaton, in a {\em classical} scale invariant
extension of a given Lagrangian, does not necessarily couple to the anomaly, but only to massive states, exactly as in the Higgs
case. For an effective dilaton, instead, the Lagrangian is derived at tree level on the basis of classical scale invariance, as for
a fundamental dilaton, but needs to be modified with the addition of an anomalous contribution, due to the composite nature of the
scalar, in close analogy to the pion case.
As we are going to show, if the dilaton is a composite state, identified with the anomaly pole of
the $J_D VV$ correlator, an infrared coupling of this pole (i.e. a non-zero residue) is necessary in order to claim the presence of an
anomaly enhancement in the $VV$ decay channel, with the $VV$ denoting on-shell physical asymptotic states, in a typical $S$-matrix
approach. Here our reasoning follows quite closely the chiral anomaly case, where the anomaly pole of the $AVV$ diagram, which
describes the pion exchange between the axial vector ($A$) and the vector currents, is infrared coupled only if $V$ denote physical
asymptotic states.
Clearly, our argument relies on a perturbative picture and is, in this respect, admittedly limited, forcing this issue to be resolved
at experimental level, as in the pion case. We recall that in the pion the enhancement is present in the di-photon channel and not in
the 2-gluon decay channel.
Perturbation theory, in any case, allows to link the enhancement of a certain dilaton production/decay channel, to the virtuality of
the gauge currents in the initial or the final state.
We conclude with a discussion of the possible phenomenological implications of our results at the level of anomaly-enhanced dilaton
decays, after pointing out the difference between the various ways in which the requirement of scale invariance (classical or
quantum) can be realized in a typical scale invariant extension of the Standard Model Lagrangian.
\subsection{The energy-momentum tensor}
We start with a brief summary of the structure of the Standard Model interactions with a $4D$ gravitational background, which is
convenient in order to describe both the coupling of the graviscalar dilaton, emerging from the Kaluza-Klein compactification, and of
a graviton at tree level and at higher orders.
In the background metric $g_{\mu\nu}$ the action takes the form
\begin{equation} S = S_G + S_{SM} + S_{I}= -\frac{1}{\kappa^2}\int d^4 x \sqrt{-g}\, R + \int d^4 x
\sqrt{-g}\mathcal{L}_{SM} + \chi \int d^4 x \sqrt{-g}\, R \, H^\dag H \, ,
\end{equation}
where $\kappa^2=16 \pi G_N$, with $G_N$ being the four dimensional Newton's constant and $\mathcal H$ is the Higgs doublet.
We recall that the EMT in our conventions is defined as
\begin{equation} T_{\mu\nu}(x) = \frac{2}{\sqrt{-g(x)}}\frac{\d [S_{SM} + S_I ]}{\d g^{\mu\nu}(x)},
\end{equation}
or, in terms of the SM Lagrangian, as
\begin{equation} \label{TEI spaziocurvo}
\frac{1}{2} \sqrt{-g} T_{\mu\nu}{\equiv} \frac{\partial(\sqrt{-g}\mathcal{L})}
{\partial g^{\mu\nu}} - \frac{\partial}{\partial x^\s}\frac{\partial(\sqrt{-g}\mathcal{L})}{\partial(\partial_\s g^{\mu\nu})}\, ,
\end{equation}
which is classically covariantly conserved ($g^{\mu\rho}T_{\mu\nu; \rho}=0$). In flat spacetime, the covariant derivative is
replaced by the ordinary derivative, giving the ordinary conservation equation ($ \partial_\mu T^{\mu\nu} = 0$).
We use the convention $\eta_{\mu\nu}=(1,-1,-1,-1)$ for the metric in flat spacetime, parametrizing its deviations
from the flat case as
\begin{equation}\label{QMM} g_{\mu\nu}(x) \equiv \h_{\mu\nu} + \kappa \, h_{\mu\nu}(x)\,,\end{equation}
with the symmetric rank-2 tensor $h_{\mu\nu}(x)$ accounting for its fluctuations.
In this limit, the coupling of the Lagrangian to gravity is given by the term
\begin{equation}\label{Lgrav} \mathcal{L}_{grav}(x) = -\frac{\kappa}{2}T^{\mu\nu}(x)h_{\mu\nu}(x). \end{equation}
In the case of theories with extra spacetime dimensions the structure of the corresponding Lagrangian can be found in
\cite{Han:1998sg,Giudice:2000av}. For instance, in the case of a compactification over a $S_1$ circle of a 5-dimensional theory to
4D, equation (\ref{Lgrav}) is modified in the form
\begin{equation}
\label{Lgrav1} \mathcal{L}_{grav}(x) = -\frac{\kappa}{2} T^{\mu\nu}(x) \left(h_{\mu\nu}(x) + \rho(x) \, \eta_{\mu\nu} \right)
\end{equation}
which is sufficient in order to describe dilaton $(\rho)$ interactions with the fields of the Standard Model at leading order in
$\kappa$, as in our case. In this case the graviscalar field $\rho$ is related to the $g_{55}$ component of the 5D metric and
describes its massless Kaluza-Klein mode. The compactification generates an off-shell coupling of $\rho$ to the trace of the
symmetric EMT. Notice that in this construction the fermions are
assumed to live on the 4D brane and their interactions can be described by the ordinary embedding of the fermion Lagrangian of the
Standard Model to a curved 4D gravitational background. We use the spin connection $\Omega$ induced by the
curved metric $g_{\mu\nu}$. This allows to define a spinor derivative $\mathcal{D}$ which transforms covariantly under local Lorentz
transformations. If we denote with $\underline{a},\underline{b}$ the Lorentz indices of a local free-falling frame, and denote with
$\s^{\underline{a}\underline{b}}$ the generators of the Lorentz group in the spinor representation, the spin connection takes
the form
\begin{equation}
\Omega_\mu(x) = \frac{1}{2}\s^{\underline{a}\underline{b}}V_{\underline{a}}^{\,\nu}(x)V_{\underline{b}\nu;\mu}(x)\, ,
\end{equation}
where we have introduced the Vielbein $V_{\underline{a}}^\mu(x)$. The covariant derivative of a spinor in a given representation
$(R)$ of the gauge symmetry group, expressed in curved $(\mathcal{D}_{\mu})$ coordinates is then given by
\begin{equation} \mathcal{D}_{\mu} = \frac{\partial}{\partial x^\mu} + \Omega_\mu + A_\mu,\end{equation}
where $A_\mu\equiv A_\mu^a\, T^{a (R)}$ are the gauge fields and $T^{a (R)}$ the group generators,
giving a Lagrangian of the form
\begin{eqnarray} \mathcal{L}& = & \sqrt{-g} \bigg\{\frac{i}{2}\bigg[\bar\psi\g^\mu(\mathcal{D}_\mu\psi)
- (\mathcal{D}_\mu\bar\psi)\g^\mu\psi \bigg] - m\bar\psi\psi\bigg\}\, .
\end{eqnarray}
The derivation of the complete dilaton/gauge/gauge vertex in the Standard Model requires the computation of the trace of the EMT
${T^\mu}_\mu$ (for the tree-level contributions), and of a large set of 1-loop 3-point functions.
These are diagrams characterized by the insertion of the trace into 2-point functions of gauge currents.
The full EMT is given by a minimal tensor $T_{Min}^{\mu\nu}$ (without improvement) and by a term of improvement,
$T_I^{\mu\nu}$, originating from $S_I$, as
\begin{eqnarray}
T^{\mu\nu} = T_{Min}^{\mu\nu} + T_I^{\mu\nu} \,,
\end{eqnarray}
where the minimal tensor is decomposed into gauge, ghost, Higgs, Yukawa and gauge fixing (g.fix.) contributions which can be found in
\cite{Coriano:2011zk}
\begin{eqnarray}
T_{Min}^{\mu\nu} =
T_{gauge}^{\mu\nu} + T_{ferm.}^{\mu\nu} + T_{Higgs}^{\mu\nu} + T_{Yukawa}^{\mu\nu} + T_{g.fix.}^{\mu\nu} + T_{ghost}^{\mu\nu}.
\end{eqnarray}
Concerning the structure of the EMT of improvement, we introduce the ordinary parametrizations of the Higgs field
\begin{equation}
\label{Higgsparam}
H = \left(\begin{array}{c} -i \phi^{+} \\ \frac{1}{\sqrt{2}}(v + h + i \phi) \end{array}\right)
\end{equation}
and of its conjugate $H^\dagger$, expressed in terms of $h$, $\phi$ and $\phi^{\pm}$, corresponding to the physical Higgs and the
Goldstone bosons of the $Z$ and $W^{\pm}$ respectively. As usual, $v$ denotes the Higgs vacuum expectation value. This expansion
generates a non-vanishing EMT, induced by $S_I$, given by
\begin{eqnarray}
\label{Timpr}
T^{\mu\nu}_I = - 2 \chi (\partial^\mu \partial^\nu - \eta^{\mu \nu} \Box) H^\dag H =
- 2 \chi (\partial^\mu \partial^\nu - \eta^{\mu\nu} \Box) \left( \frac{h^2}{2} + \frac{\phi^2}{2} + \phi^+ \phi^- + v \, h\right)\, .
\end{eqnarray}
Notice that this term is generated by a Lagrangian which does not survive the flat spacetime limit. We are going to show by an
explicit computation that $T_I$, if properly included with $\chi=1/6$, guarantees the renormalizability of the model.
\section{One loop electroweak corrections to dilaton-gauge-gauge vertices}
In this section we will present results for the structure of the radiative corrections to the dilaton/gauge/gauge vertices in the
case of two photons, photon/$Z$ and $Z Z$ gauge currents. We have included in appendix \ref{rules} the list of the relevant tree
level interactions extracted from the SM Lagrangian
introduced above and which have been used in the computation of these corrections. We identify three classes of contributions,
denoted as $\mathcal{A}$, $\Sigma$ and $\Delta$, with the $\mathcal A$-term coming from the conformal anomaly while the $\Sigma$ and
$\Delta$ terms are related to the exchange of fermions, gauge bosons and scalars (Higgs/Goldstones). The separation between the
anomaly part and the remaining terms is typical of the $TVV$ interaction. In particular one can check that in a
mass-independent renormalization scheme, such as dimensional regularization with minimal subtraction, this separation can be verified
at least at one loop level and provides a realization of the (anomalous) conformal Ward identity
\begin{equation}
\Gamma^{\alpha\beta}(z,x,y)
\equiv \eta_{\mu\nu} \left\langle T^{\mu\nu}(z) V^{\alpha}(x) V^{\beta}(y) \right\rangle
= \frac{\delta^2 \mathcal A(z)}{\delta A_{\alpha}(x) \delta A_{\beta}(y)} + \left\langle {T^\mu}_\mu(z) V^{\alpha}(x) V^{\beta}(y)
\right\rangle,
\label{traceid1}
\end{equation}
where we have denoted by $\mathcal A(z)$ the anomaly and $A_{\alpha}$ the gauge sources coupled to the current $V^{\alpha}$. Notice
that in the expression above $\Gamma^{\alpha\beta}$ denotes a generic
dilaton/gauge/gauge vertex, which is obtained form the $TVV$ vertex by tracing the spacetime indices $\mu\nu$. A simple way to test
the validity of (\ref{traceid1}) is to compute the renormalized vertex $\langle T^{\mu\nu} V^\alpha V'^\beta\rangle$ (i.e. the
graviton/gauge/gauge vertex) and perform afterwards its 4-dimensional trace. This allows to identify the left-hand-side of this
equation. On the other hand, the insertion of the trace of $T^{\mu\nu}$ (i.e. $T^\mu_\mu$ )into a two point function $VV'$, allows to
identify the second term on the right-hand-side of (\ref{traceid1}), $\langle T^{\mu}_\mu(z) V^{\alpha}(x) V^{\beta}(y)\rangle$.
The difference between the two terms so computed can be checked to correspond to the $\mathcal A$-term, obtained by two
differentiations of the anomaly functional $\mathcal A$.
We recall that, in general, when scalars are conformally coupled, this takes the form
\begin{eqnarray} \label{TraceAnomaly}
\mathcal A(z)
&=& \sum_{i} \frac{\beta_i}{2 g_i} \, F^{\alpha\beta}_i(z) F^i_{\alpha\beta}(z) +... \,,
\end{eqnarray}
where $\beta_i$ are clearly the mass-independent $\beta$ functions of the gauge fields
and $g_i$ the corresponding coupling constants, while the ellipsis refer to curvature-dependent terms.
We present explicit results starting for the $\rho VV'$ vertices ($V,V' = \gamma, Z$), denoted as $\Gamma_{VV'}^{\alpha \beta}$,
which are decomposed in momentum space in the form
\begin{equation}
\Gamma_{VV'}^{\alpha \beta}(k,p,q) = (2\, \pi)^4\, \delta(k-p-q) \frac{i}{\Lambda}
\left( \mathcal A^{\alpha \beta}(p,q) + \Sigma^{\alpha \beta}(p,q) + \Delta^{\alpha \beta}(p,q)\right),
\end{equation}
where
\begin{equation}
\mathcal A^{\alpha \beta}(p,q) = \int d^4 x\, d^4 y \, e^{i p \cdot x + i q\cdot y}\,
\frac{\delta^2 \mathcal A(0)}{\delta A^\alpha(x)\delta A^\beta(y)}
\end{equation}
and
\begin{equation}
\Sigma^{\alpha \beta}(p,q) + \Delta^{\alpha \beta}(p,q) = \int d^4 x\, d^4 y\, e^{ i p \cdot x + i q\cdot y}\,
\left\langle {T^\mu}_\mu(0) V^\alpha(x) V^\beta(y) \right\rangle \,.
\end{equation}
We have denoted with $\Sigma^{\alpha \beta}$ the cut vertex contribution to $\Gamma^{\alpha\beta}_{\rho VV'}$,
while $\Delta^{\alpha \beta}$ includes the dilaton-Higgs mixing on the dilaton line, as shown in fig. \ref{figuremix}.
Notice that $\Sigma^{\alpha \beta}$ and $\Delta^{\alpha \beta}$ take contributions in two cases, specifically if the theory has an
explicit (mass dependent) breaking and/or if the scalar - which in this case is the Higgs field - is not conformally coupled.
The $\mathcal A^{\alpha\beta}(p,q)$ represents the conformal anomaly while $\Lambda$ is dilaton interaction scale.
\subsection{The $\rho\gamma\gamma$ vertex}
The interaction between a dilaton and two photons is identified by the diagrams in
figs. \ref{figuretriangle}, \ref{figuretadpole}, \ref{figuremix} and is summarized by the expression
\begin{eqnarray}
\Gamma_{\gamma \gamma}^{\alpha\beta}(p,q) = \frac{i}{\Lambda} \bigg[ \mathcal A^{\alpha\beta}(p,q) +
\Sigma^{\alpha\beta}(p,q) + \Delta^{\alpha\beta}(p,q) \bigg] \,,\end{eqnarray}
with the anomaly contribution given by
\begin{eqnarray}
\label{Agammagamma}
\mathcal A^{\alpha\beta} = \frac{\alpha}{\pi}\, \bigg[ -\frac{2}{3}\sum_{f} Q_{f}^2 + \frac{5}{2} + 6\,\chi\bigg]\,
u^{\alpha\beta}(p,q) \stackrel{\chi\rightarrow\frac{1}{6}}{=} - 2\, \frac{\beta_e}{e}\, u^{\alpha\beta}(p,q) \, ,
\end{eqnarray}
where
\begin{eqnarray}
\label{utensor}
u^{\alpha\beta}(p,q) = (p\cdot q) \eta^{\alpha\beta} - q^{\alpha}p^{\beta} \,,
\end{eqnarray}
and the explicit scale-breaking term $\Sigma^{\alpha \beta}$ which splits into
\begin{eqnarray}
\label{Sigmagammagamma}
\Sigma^{\alpha\beta}(p,q) = \Sigma_F^{\alpha\beta}(p,q) + \Sigma_B^{\alpha\beta}(p,q) +\Sigma_I^{\alpha\beta}(p,q) \,.
\end{eqnarray}
We obtain for the on-shell photon case ($p^2 = q^2 = 0$)
\begin{eqnarray}
\Sigma_F^{\alpha\beta}(p,q)
&=&
\frac{\alpha}{\pi}\, \sum_f Q_f^2 m_f^2 \left[ \frac{4}{s} + 2 \left(\frac{4
m_f^2}{s}-1\right) \mathcal C_0\left(s,0,0,m_f^2,m_f^2,m_f^2 \right)\right]\, u^{\alpha\beta}(p,q) \, , \nonumber \\
\Sigma_B^{\alpha\beta}(p,q)
&=&
\frac{\alpha}{\pi}\, \left[ 6 M_W^2 \left(1-2\frac{M_W^2}{s}\right)
\mathcal C_0(s,0,0,M_W^2,M_W^2,M_W^2) - 6 \frac{M_W^2}{s} - 1 \right]\, u^{\alpha\beta}(p,q) \, , \nonumber \\
\Sigma_I^{\alpha\beta}(p,q)
&=&
\frac{\alpha}{\pi}\, 6 \chi \bigg[ 2 M_W^2 \mathcal C_0\left(s,0,0,M_W^2,M_W^2,M_W^2\right) \,
u^{\alpha\beta}(p,q) \nonumber \\
&-&
M_W^2\, \frac{s}{2}\, \mathcal C_0(s,0,0,M_W^2,M_W^2,M_W^2) \, \eta^{\alpha \beta} \bigg]\, ,
\end{eqnarray}
while the mixing contributions are given by
\begin{eqnarray}
\label{DeltaHgammagamma}
\Delta^{\alpha\beta}(p,q)
&=&
\frac{\alpha}{\pi (s - M_H^2)} 6 \chi \bigg\{ 2 \sum_f Q_f^2 m_f^2 \bigg[ 2 + (4 m_f^2 -s ) \mathcal C_0(s,0,0,m_f^2,m_f^2,m_f^2)
\bigg] \nonumber \\
&+&
M_H^2 + 6 M_W^2 + 2 M_W^2 (M_H^2 + 6 M_W^2 -4 s) \mathcal C_0(s,0,0,M_W^2,M_W^2,M_W^2)\bigg\} u^{\alpha \beta}(p,q) \nonumber \\
&+&
\frac{\alpha}{\pi} 3 \chi s \, M_W^2 \, \mathcal C_0(s,0,0,M_W^2,M_W^2,M_W^2) \eta^{\alpha \beta} \, ,
\end{eqnarray}
with $\alpha$ the fine structure constant. The scalar integrals are defined in appendix \ref{scalars}.
The $\Sigma$'s and $\Delta$ terms are the contributions obtained from the insertion on the photon 2-point function of the trace of
the EMT, ${T^\mu}_\mu$. Notice that $\Sigma_I$ includes all the trace insertions which originate from the terms of improvement $T_I$
except for those which are bilinear in the Higgs-dilaton fields and
which have been collected in $\Delta$. The analysis of the Ward and Slavnov-Taylor identities for the graviton-vector-vector correlators shows that these can be consistently solved only if we include the graviton-Higgs mixing on the graviton line.
We have included contributions proportional both to fermions ($F$) and boson ($B$) loops, beside the $\Sigma_I$.
A conformal limit on these contributions can be performed by sending to zero all the mass terms, which is equivalent to sending
the vev $v$ to zero and requiring a conformal coupling of the Higgs $(\chi=1/6)$.
In the $v\to 0$ limit, but for a generic parameter $\chi$, we obtain
\begin{equation}
\lim_{v\to 0} \left( \Sigma^{\a\b} + \Delta^{\a\b} \right)
= \lim_{v\to 0} \left(\Sigma_{B}^{\a\b} + \Sigma_{I}^{\a\b}\right)
= \frac{\alpha}{\pi} (6 \chi -1) u^{\alpha \beta}(p,q),
\end{equation}
which, in general, is non-vanishing.
Notice that, among the various contributions, only the exchange of a boson or the term of improvement contribute in this limit and
their sum vanishes only if the Higgs is conformally coupled $(\chi = \frac{1}{6})$. \\
Finally, we give the decay rate of the dilaton into two on-shell photons in the simplified case in which we remove the term of
improvement by sending $\chi \to 0$
\begin{eqnarray}
\Gamma(\rho \rightarrow \gamma\gamma)
&=&
\frac{\alpha^2\,m_{\rho}^3}{256\,\Lambda^2\,\pi^3} \, \bigg| \beta_{2} + \beta_{Y}
-\left[ 2 + 3\, x_W +3\,x_W\,(2-x_W)\,f(x_W) \right]
+ \frac{8}{3} \, x_t\left[1 + (1-x_t)\,f(x_t) \right] \bigg|^2 \,,
\label{PhiGammaGamma}
\end{eqnarray}
where the contributions to the decay, beside the anomaly term, come from the $W$ and the fermion (top) loops
and $\beta_2 (= 19/6)$ and $\beta_Y (= -41/6)$ are the $SU(2)$ and $U(1)$ $\beta$ functions respectively.
Here, as well as in the other decay rates evaluated all through the paper, the $x_i$ are defined as
\begin{equation} \label{x}
x_i = \frac{4\, m_i^2}{m^2_\rho} \, ,
\end{equation}
with the index "$i$" labelling the corresponding massive particle, and $x_t$ denoting the contribution from the top quark,
which is the only massive fermion running in the loop.
The function $f(x)$ is given by
\begin{eqnarray}
\label{fx}
f(x) =
\begin{cases}
\arcsin^2(\frac{1}{\sqrt{x}})\, , \quad \mbox{if} \quad \, x \geq 1 \\
-\frac{1}{4}\,\left[ \ln\frac{1+\sqrt{1-x}}{1-\sqrt{1-x}} - i\,\pi \right]^2\, , \quad \mbox{if} \quad \, x < 1.
\end{cases}
\end{eqnarray}
which originates from the scalar $3$-point master integral through the relation
\begin{equation} \label{C03m}
C_0(s,0,0,m^2,m^2,m^2) = - \frac{2}{s} \, f(\frac{4\,m^2}{s}) \, .
\end{equation}
\subsection{The $\rho\gamma Z$ vertex}
The interaction between a dilaton, a photon and a $Z$ boson is described by the $\Gamma^{\alpha \beta}_{\gamma Z}$ correlation
function (figs.. \ref{figuretriangle}, \ref{figuretadpole}, \ref{figuremix}). In the on-shell case, with the kinematic defined by
\begin{equation}
p^2 = 0 \, \quad q^2 = M_Z^2 \, \quad k^2 = (p+q)^2 = s \,,
\end{equation}
the vertex $\Gamma^{\alpha \beta}_{\gamma Z}$ is expanded as
\begin{eqnarray}
\Gamma^{\alpha \beta}_{\gamma Z} &=& \frac{i}{\Lambda} \bigg[ \mathcal A^{\alpha\beta}(p,q)
+ \Sigma^{\alpha\beta}(p,q) + \Delta^{\alpha\beta}(p,q) \bigg] \nonumber \\
&=&
\frac{i}{\Lambda}\, \Bigg\{\,
\left[ \frac{1}{2}\,\left(s - M^2_Z\right)\,\eta^{\alpha\beta} - q^\alpha\,p^\beta\right] \,
\left( \mathcal A_{\gamma Z} + \Phi_{\gamma Z}(p,q)\right)
+\eta^{\alpha\beta}\, \Xi_{\gamma Z}(p,q) \Bigg\} \, .
\end{eqnarray}
The anomaly contribution is
\begin{equation}
\mathcal A_{\gamma Z} = \frac{\alpha}{\pi\,s_w c_w}\,\left[-\frac{1}{3}\sum_{f}C^f_v\,Q_f + \frac{1}{12}\,(37-30s_w^2)
+ 3\, \chi\, (c_w^2 - s_w^2) \right]\, ,
\end{equation}
where $s_w$ and $c_w$ to denote the sine and cosine of the $\theta$-Weinberg angle.
Here $\Delta^{\alpha \beta}$ is the external leg correction on the dilaton line and the form factors $\Phi(p,q)$ and $\Xi(p,q)$
are introduced to simplify the computation of the decay rate and decomposed as
\begin{eqnarray}
\Phi_{\gamma Z}(p,q) &=& \Phi^{\Sigma}_{\gamma Z}(p,q) + \Phi^{\Delta}_{\gamma Z}(p,q) \,, \nonumber \\
\Xi_{\gamma Z}(p,q) &=& \Xi^{\Sigma}_{\gamma Z}(p,q) + \Xi^{\Delta}_{\gamma Z}(p,q) \,,
\end{eqnarray}
in order to distinguish the contributions to the external leg corrections ($\Delta$) from those to the cut vertex ($\Sigma$).
They are given by
\begin{eqnarray}
\Phi^{\Sigma}_{\gamma Z}(p,q)
&=&
\frac{\alpha}{\pi\, s_w\,c_w}\, \Bigg\{
\sum_f C_v^f \, Q_f \, \Bigg[\frac{2\, m_f^2}{s - M_Z^2} + \frac{2 m_f^2 \, M_Z^2}{(s - M_Z^2)^2}\, \mathcal D_0(s,M_Z^2,m_f^2,m_f^2)
\nonumber \\
&& \hspace{-20mm}
- m_f^2\, \left(1 - \frac{4\, m_f^2}{s - M_Z^2}\right)\, \mathcal C_0(s,0,M_Z^2,m_f^2,m_f^2,m_f^2) \Bigg]
- \Bigg[ \frac{M_Z^2}{2\,\left(s - M_Z^2\right)}\, (12\, s_w^4 - 24\, s_w^2 + 11)
\nonumber\\
&& \hspace{-20mm}
\frac{M_Z^2}{2\,\left(s - M_Z^2 \right)^2}
\left[2\, M_Z^2\,\Big(6\, s_w^4 - 11\, s_w^2 + 5 \Big)- 2\, s_w^2\, s + s \right]\, \mathcal D_0(s,M_Z^2,M_W^2,M_W^2)
\nonumber \\
&& \hspace{-20mm}
+ \frac{M_Z^2\, c_w^2}{s - M_Z^2}\,
\Big[2\, M_Z^2\, \left(6\, s_w^4 - 15\, s_w^2 + 8 \right) + s\, \left(6\, s_w^2 - 5\right)\Big]\,
\mathcal C_0(s,0,M_Z^2,M_W^2,M_W^2,M_W^2) \Bigg]
\nonumber \\
&& \hspace{-20mm}
+ \frac{3 \chi \,(c_w^2 - s_w^2)\,}{s - M_Z^2}\,
\bigg[M_Z^2 + s\, \bigg( 2\, M_W^2 \, \mathcal C_0(s,0,M_Z^2,M_W^2,M_W^2,M_W^2)
+ \frac{M_Z^2}{s-M_Z^2}\, \mathcal D_0(s,M_Z^2,M_W^2,M_W^2) \bigg) \bigg]
\Bigg\} \, ,
\nonumber
\end{eqnarray}
\begin{eqnarray}
\Xi^{\Sigma}_{\gamma Z}(p,q)
&=&
\frac{\alpha}{\pi}
\Bigg\{ - \frac{c_w\, M_Z^2}{ s_w} \mathcal B_0(0,M_W^2,M_W^2) + 3\, s\, \chi \, s_w^2\, M_Z^2 \,
\mathcal C_0(s,0,M_Z^2,M_W^2,M_W^2,M_W^2) \Bigg\} \, , \nonumber \\
\Phi^{\Delta}_{\gamma Z}(p,q)
&=&
\frac{3\,\alpha\,s\,\chi}{\pi s_w c_w (s-M_H^2)(s-M_Z^2)}
\bigg\{ 2 \sum_f m_f^2 C_v^f Q_f \bigg[ 2 + 2 \frac{M_Z^2}{s-M_Z^2} \mathcal D_0(s,M_Z^2,m_f^2,m_f^2) \nonumber \\
&& \hspace{-20mm}
+ (4 m_f^2 + M_Z^2 - s) \mathcal C_0(s,0,M_Z^2,m_f^2,m_f^2,m_f^2) \bigg] + M_H^2(1-2 s_w^2) + 2 M_Z^2 (6 s_w^4 - 11 s_w^2 +5) \nonumber \\
&& \hspace{-20mm}
+ \frac{M_Z^2}{s-M_Z^2} ( M_H^2 (1-2 s_w^2) + 2 M_Z^2 (6 s_w^4 -11 s_w^2 + 5) ) \mathcal D_0(s,M_Z^2,M_W^2,M_W^2) \nonumber \\
&& \hspace{-20mm}
+ 2 M_W^2 \mathcal C_0(s,0,M_Z^2,M_W^2,M_W^2,M_W^2) ( M_H^2 (1-2 s_w^2) + 2 M_Z^2 (6 s_w^4 - 15 s_w^2 + 8) + 2 s (4 s_w^2-3)) \bigg\}
\nonumber \\
\Xi^{\Delta}_{\gamma Z}(p,q)
&=&
\frac{3\,\alpha\,s\, \chi\, c_w}{\pi\, s_w}\, M_Z^2 \bigg\{ \frac{2}{s-M_H^2} \mathcal B_0(0, M_W^2,M_W^2)
- s_w^2 \mathcal C_0(s,0,M_Z^2,M_W^2,M_W^2,M_W^2) \bigg\}\, .
\end{eqnarray}
As for the previous case, we give the decay rate in the simplified limit $\chi \to 0$ which is easily found to be
\begin{eqnarray}
\Gamma(\rho\rightarrow \gamma Z)
&=&
\frac{9\,m_{\rho}^3}{1024\,\Lambda^2\,\pi} \, \sqrt{1-x_Z} \, \bigg( |\Phi^{\Sigma}_{\gamma Z}|^2(p,q)\,m_{\rho}^4\,(x_Z-4)^2 +
48 \, Re\, \left\{\Phi^{\Sigma}_{\gamma Z}(p,q)\,\Xi^{\Sigma\,*}_{\gamma Z}(p,q) \, m_{\rho}^2\,(x_Z-4)\right\}
\nonumber \\
&& \hspace{35mm}
- \, 192 \, |\Xi^{\Sigma}_{\gamma Z}|^2(p,q) \bigg) \, ,
\label{RateRhoGammaZ}
\end{eqnarray}
where $Re$ is the symbol for the real part.
\subsection{The $\rho Z Z$ vertex}
The expression for the $\Gamma^{\alpha\beta}_{Z Z}$ vertex (figs.. \ref{figuretriangle},\ref{figuretadpole},\ref {figuremix}) defining
the $\rho ZZ$ interaction is presented here in the kinematic limit given by $k^2 = (p+q)^2 = s$, $p^2 = q^2 = M_Z^2$ with two
on-shell $Z$ bosons. The completely cut correlator takes contributions from a fermion sector, a $W$ gauge boson sector, a $Z-H$
sector together with a term of improvement.
There is also an external leg correction $\Delta^{\alpha \beta}$ on the dilaton line which is much more involved than in the previous
cases because there are contributions coming from the minimal EMT and from the improven EMT . \\
At one loop order we have
\begin{eqnarray} \label{1loopZZ}
&&
\Gamma_{Z Z}^{\alpha\beta}(p,q)
\equiv
\frac{i}{\Lambda}\, \left[ \mathcal A^{\alpha\beta}(p,q) + \Sigma^{\alpha\beta}(p,q)+ \Delta^{\alpha\beta}(p,q) \right] \nonumber \\
&& =
\frac{i}{\Lambda}\,
\Bigg\{
\left[\left(\frac{s}{2} - M^2_Z \right)\, \eta^{\alpha\beta} - q^\alpha\,p^\beta \right]\,
\left( \mathcal A_{ZZ} + \Phi^{\Sigma}_{ZZ}(p,q) + \Phi^{\Delta}_{ZZ}(p,q) \right)
+ \eta^{\alpha\beta}\, \left(\Xi^{\Sigma}_{ZZ}(p,q) + \Xi^{\Delta}_{ZZ}(p,q) \right) \Bigg\} ,
\end{eqnarray}
where again $\Sigma$ stands for the completely cut vertex and $\Delta$ for the external leg corrections and
we have introduced for convenience the separation
\begin{eqnarray}
\Phi^{\Sigma}_{ZZ}(p,q) &=& \Phi^{F}_{ZZ}(p,q) + \Phi^{W}_{ZZ}(p,q) + \Phi^{ZH}_{ZZ}(p,q) + \Phi^{I}_{ZZ}(p,q) \,, \nonumber \\
\Xi^{\Sigma}_{ZZ}(p,q) &=& \Xi^{F}_{ZZ}(p,q) + \Xi^{W}_{ZZ}(p,q) + \Xi^{ZH}_{ZZ}(p,q) + \Xi^{I}_{ZZ}(p,q)\, .
\end{eqnarray}
The form factors are given in appendix \ref{VZZ}, while here we report only the purely anomalous contribution
\begin{equation}
\mathcal A_{ZZ} = \frac{\alpha}{6 \pi c_w^2 s_w^2}\,
\left\{ -\sum_{f} \left({C_a^f}^2+{C_v^f}^2\right) + \frac{60\,s_w^6 - 148\,s_w^2 + 81}{4} - \frac{7}{4}
+ 18\,\chi\,\left[ 1 - 2\,s_w^2\,c_w^2 \right]\right\}\, .
\end{equation}
Finally, we give the decay rate expression for the $\rho \to Z Z$ process.
At leading order it can be computed from the tree level amplitude
\begin{eqnarray}
\mathcal M^{\alpha\beta}(\rho \rightarrow ZZ)
&=&
\frac{2}{\Lambda}\,M_Z^2\, \eta^{\alpha\beta} \, ,
\end{eqnarray}
and it is given by
\begin{eqnarray}
\Gamma(\rho \rightarrow ZZ)
&=&
\frac{ m_{\rho}^3}{32\,\pi \Lambda^2} \, (1-x_Z)^{1/2}\, \left[ 1 - x_Z + \frac{3}{4}\,x_Z^2 \right].
\label{PhiZZ}
\end{eqnarray}
Including the $1$-loop corrections defined in eq. (\ref{1loopZZ}), one gets the decay rate at next-to-leading order
\begin{eqnarray}
\Gamma(\rho\rightarrow ZZ)
&=&
\frac{m_{\rho}^3}{32\,\pi\,\Lambda^2} \,\sqrt{1-x_Z}\, \bigg\{ 1 - x_Z + \frac{3}{4}\,x_Z^2
+ \frac{3}{x_Z} \, \bigg[4\, Re\, \{\Phi^{\Sigma}_{ZZ}(p,q)\}(1-x_Z + \frac{3}{4}\,x_Z^2)
\nonumber \\
&-&
Re\, \{\Xi^{\Sigma}_{ZZ}(p,q)\}\,m_{\rho}^2\, \left( \frac{3}{4}\,x_Z^3 - \frac{3}{2}\,x_Z^2 \right) \bigg] \bigg\}\, .
\end{eqnarray}
\begin{figure}[t]
\centering
\subfigure[]{\includegraphics[scale=.7]{figures/dil_fermtriangle}}
\hspace{.2cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_Wtriangle}}
\hspace{.2cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_phitriangle}}
\hspace{.2cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_WWphitriangle}}
\hspace{.2cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_Wphiphitriangle}} \\
\hspace{.2cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_etaptriangle}}
\hspace{.2cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_etamtriangle}}
\hspace{.2cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_ZZHtriangle}}
\hspace{.2cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_phiphiHtriangle}}
\hspace{.2cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_HHZtriangle}}
\hspace{.2cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_HHphitriangle}}
\caption{Amplitudes of triangle topology contributing to the $\rho \gamma\gamma$, $\rho \gamma Z$ and $\rho ZZ$ interactions. They
include fermion $(F)$, gauge bosons $(B)$ and contributions from the term of improvement (I). Diagrams (a)-(g) contribute to all the
three channels while (h)-(k) only in the $\rho ZZ$ case.}
\label{figuretriangle}
\end{figure}
\begin{figure}[t]
\centering
\subfigure[]{\includegraphics[scale=.7]{figures/dil_fermtbubble}}
\hspace{.2cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_Wtbubble}}
\hspace{.2cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_phitbubble}}
\hspace{.2cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_phiWtbubble}}
\hspace{.2cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_Wphitbubble}}
\hspace{.2cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_etaptbubble}} \\
\subfigure[]{\includegraphics[scale=.7]{figures/dil_etamtbubble}}
\hspace{.2cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_Wsbubble}}
\hspace{.2cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_phipmsbubble}}
\hspace{.2cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_Wtadpole}}
\hspace{.2cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_phitadpole}}
\hspace{.2cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_ZHtbubble}} \\
\hspace{.2cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_phiHtbubble}}
\hspace{.2cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_Hsbubble}}
\hspace{.2cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_phisbubble}}
\hspace{.2cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_Htadpole}}
\hspace{.2cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_phi0tadpole}}
\caption{Bubble and tadpole-like diagrams for $\rho \gamma\gamma$ $\rho \gamma Z $ and $\rho Z Z$.
Amplitudes (l)-(q) contribute only in the $\rho ZZ$ channel.}
\label{figuretadpole}
\end{figure}
\subsection{Renormalization of dilaton interactions in the broken electroweak phase}
\label{renorm}
In this section we address the renormalization properties of the correlation functions given above. Although the
proof is quite cumbersome, one can check, from our previous results, that the 1-loop renormalization of the Standard Model
Lagrangian is sufficient to cancel all the singularities in the cut
vertices independently of whether the Higgs is conformally coupled or not. Concerning the uncut vertices, instead, the term of
improvement plays a significant role in the determination of Green functions which are ultraviolet finite. In particular such a term
has to appear with $\chi=1/6$ in order to guarantee the cancellation of a singularity present in the 1-loop 2-point function
describing the Higgs dilaton mixing ($\Sigma_{\rho H}$).
The problem arises only in the $\Gamma^{\alpha\beta}_{ZZ}$ correlator, where the $\Sigma_{\rho H}$ $2$-point function is present as
an external leg correction on the dilaton line.
The finite parts of the counterterms are determined in the on-shell renormalization scheme, which is widely used in the electroweak
theory. In this scheme the renormalization conditions are fixed in terms of the physical parameters to all orders in perturbation
theory and the wave-function normalizations of the fields are obtained by requiring a unit residue of the full 2-point functions on
the physical particle poles.
From the counterterm Lagrangian we compute the corresponding counterterm to the trace of the EMT. As we have already mentioned,
one can also verify from the explicit computation that the terms of improvement, in the conformally coupled case, are necessary
to renormalize the vertices containing an intermediate scalar with an external bilinear mixing (dilaton/Higgs).
The counterterm vertices for the correlators with a dilaton insertion are
\begin{eqnarray}
\delta [\rho \gamma \gamma]^{\alpha\beta} &=& 0
\\
\delta [\rho \gamma Z]^{\alpha\beta} &=& - \frac{i}{\Lambda} \delta Z_{Z\gamma} \, M_Z^2 \, \eta^{\alpha\beta} \, ,
\\
\delta [\rho Z Z]^{\alpha\beta} &=& - 2 \frac{i}{\Lambda} (M_Z^2 \, \delta Z_{ZZ} + \delta M^2_Z) \, \eta^{\alpha\beta} \, ,
\end{eqnarray}
where the counterterm coefficients are defined in terms of the 2-point functions of the fundamental fields as
\begin{eqnarray}
\delta Z_{Z \gamma} = 2 \frac{\Sigma_T^{\gamma Z}(0)}{M_Z^2} \,, \quad \delta Z_{ZZ} = - Re \frac{\partial
\Sigma_T^{ZZ}(k^2)}{\partial k^2} \bigg |_{k^2 = M_Z^2} \,, \quad
\delta M_Z^2 = Re \, \Sigma_T^{ZZ}(M_Z^2) \,,
\end{eqnarray}
and are defined in appendix \ref{SigmaSM}.
It follows then that the $\rho \gamma \gamma$ interaction must be finite, as one can find by a direct inspection of the
$\Gamma^{\alpha\beta}_{\gamma \gamma}$ vertex, while the others require the subtraction of their divergences.
\begin{figure}[t]
\centering
\subfigure[]{\includegraphics[scale=.7]{figures/dil_mixHblob}}
\hspace{.2cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_blobH}}
\hspace{.2cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_mixHblobH}}
\caption{External leg corrections. Diagrams (b) and (c) appear only in the $\rho Z Z$ sector.}
\label{figuremix}
\end{figure}
These counterterms are sufficient to remove the divergences of the completely cut graphs which do
not contain a bilinear mixing, once we set on-shell the external gauge lines. This occurs both for those diagrams which do not
involve the terms of improvement and for those involving $T_I$.
Regarding those contributions which involve the bilinear mixing on the external dilaton line, we encounter two different situations. \\
In the $\rho \gamma Z$ vertex the insertion of the bilinear mixing $\rho H$ generates a reducible diagram of the form Higgs/photon/Z
whose renormalization is guaranteed, within the Standard Model, by the use of the Higgs/photon/Z counterterm
\begin{eqnarray}
\delta [H\gamma Z]^{\alpha\beta} = \frac{e \, M_Z}{2 s_w c_w} \delta Z_{Z\gamma}\, \eta^{\alpha\beta} \,.
\end{eqnarray}
As a last case, we discuss the contribution to $\rho ZZ$ coming from the bilinear mixing, already mentioned above. The corrections on
the dilaton line involve the dilaton/Higgs mixing $\Sigma_{\rho H}$, the Higgs self-energy $\Sigma_{HH}$ and the term of improvement
$\Delta^{\alpha\beta}_{I\,,HZZ}$, which introduces the Higgs/Z/Z vertex (or $HZZ$) of the Standard Model.
The Higgs self-energy and the $HZZ$ vertex, in the Standard Model, are renormalized with the usual counterterms
\begin{eqnarray}
\delta [HH](k^2)
&=&
(\delta Z_H \, k^2 - M_H^2 \delta Z_H - \delta M_H^2) \, ,
\\
\delta [HZZ]^{\alpha\beta}
&=&
\frac{e \, M_Z}{s_w \, c_w} \bigg[ 1 + \delta Z_e + \frac{2 s_w^2
- c_w^2}{c_w^2} \frac{\delta s_w}{s_w} + \frac{1}{2} \frac{\delta M_W^2}{M_W^2} + \frac{1}{2} \delta Z_H + \delta Z_{ZZ} \bigg]
\, \eta^{\alpha\beta} \,,
\end{eqnarray}
where
\begin{eqnarray}
&&
\delta Z_H =
- Re \frac{\partial \Sigma_{HH}(k^2)}{\partial k^2} \bigg|_{k^2=M_H^2}\, ,\quad \delta M_H^2 = Re \Sigma_{HH}(M_H^2) \, , \quad
\delta Z_e = - \frac{1}{2} \delta Z_{\gamma \gamma} + \frac{s_w}{2 c_w} \delta Z_{Z \gamma} \,, \nonumber \\
&& \delta s_w = - \frac{c_w^2}{2 s_w} \left( \frac{\delta M_W^2}{M_W^2} - \frac{\delta M_Z^2}{M_Z^2} \right) \,, \quad \delta M_W^2 =
Re \Sigma^{WW}_T(M_W^2) \,, \quad
\delta Z_{\gamma \gamma} = - \frac{\partial \Sigma_T^{\gamma \gamma}(k^2)}{\partial k^2} \bigg|_{k^2 = 0} \,.
\end{eqnarray}
The self-energy $\Sigma_{\rho H}$ is defined by the minimal contribution generated by ${{T_{Min}}^\mu}_\mu$
and by a second term derived from ${{T_I}^\mu}_\mu$.
This second term, with the conformal coupling $\chi = \frac{1}{6}$, is necessary in order to ensure the renormalizability
of the dilaton/Higgs mixing.
In fact, the use of the minimal EMT in the computation of this self-energy involves a divergence of the form
\begin{eqnarray}
\delta [\rho H]_{Min} = - 4 \frac{i}{\Lambda}\, \delta t \, , \label{CThH}
\end{eqnarray}
with $\delta t$ fixed by the condition of cancellation of the Higgs tadpole $T_{ad}$ ($\delta t + T_{ad} = 0$).
A simple analysis of the divergences in $\Sigma_{Min, \, \rho H}$ shows that the counterterm given in eq. (\ref{CThH}) is not
sufficient to remove all the singularities of this correlator unless we also include the renormalization constant provided by the term of improvement
which is given by
\begin{eqnarray}
\delta [\rho H]_{I}(k) = - \frac{i}{\Lambda} 6 \, \chi \, v \bigg[ \delta v + \frac{1}{2} \delta Z_H \bigg] k^2\,, \qquad
\qquad \text{with} \quad \chi = \frac{1}{6} \,,
\end{eqnarray}
and
\begin{eqnarray}
\delta v = v \bigg( \frac{1}{2} \frac{\delta M_W^2}{M_W^2} + \frac{\delta s_w}{s_w} - \delta Z_e \bigg) \,.
\end{eqnarray}
One can show explicitly that this counterterm indeed ensures the finiteness of $\Sigma_{\rho H}$.
\begin{figure}[t]
\centering
\subfigure[]{\includegraphics[scale=.7]{figures/dil_triangle_gg}\label{QCD_NLOa}}
\hspace{.5cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_fermiontbubble_gg}\label{QCD_NLOb}}
\hspace{.5cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_fermionububble_gg}\label{QCD_NLOc}}
\hspace{.5cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_gluonloop_gg}\label{QCD_NLOd}}
\hspace{.5cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_gluontbubble_gg}\label{QCD_NLOe}}
\hspace{.5cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_gluonububble_gg}\label{QCD_NLOf}}\\
\hspace{.5cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_gluonsbubble_gg}\label{QCD_NLOg}}
\hspace{.5cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_gluontadpole}\label{QCD_NLOh}}
\hspace{.5cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_ghostloop_gg}\label{QCD_NLOi}}
\hspace{.5cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_ghosttbubble_gg}\label{QCD_NLOl}}
\hspace{.5cm}
\subfigure[]{\includegraphics[scale=.7]{figures/dil_ghostububble_gg}\label{QCD_NLOm}}
\caption{QCD vertices at next-to-leading order. In the on-shell gluon case only diagram (a)contributes.}
\label{QCD_NLO}
\end{figure}
\section{The off-shell dilaton-gluon-gluon vertex in QCD}
After a discussion of the leading corrections to the vertices involving one dilaton and two electroweak currents we investigate the
interaction of a dilaton and two gluons beyond leading order, giving the expression of the full off-shell vertex. The corresponding
interaction with two on-shell gluons has been computed in \cite{Giudice:2000av} and is simply given by the contributions of the
anomaly and of the quark loop. We will come back to rediscuss the on-shell case in the second part of this work, where we will
stress some specific perturbative features of this interaction.
We show in fig. \ref{QCD_NLO} a list of the NLO QCD contributions to dilaton interactions. As we have just mentioned, in the
$2$-gluon on-shell case one can show by an explicit computation that each of these contributions vanishes, except for diagram $(a)$,
which is non-zero when a massive fermion runs in the loop. For this specific reason, in the parton model, the production of the
dilaton in $pp$ collisions at the LHC is mediated by the diagram of gluon fusion, which involves a top quark in a loop.
We find convenient to express the result of the off-shell $\Gamma^{\alpha\beta}_{g g}$ vertex in the form
\begin{eqnarray} \label{OffShellQCD}
\Gamma_{g g}^{\alpha \beta}(p,q) =
\frac{i}{\Lambda}\, \bigg\{ A^{00}(p,q) \eta^{\alpha \beta} + A^{11}(p,q) p^{\alpha} p^{\beta} + A^{22}(p,q) q^{\alpha} q^{\beta}
+ A^{12}(p,q) p^{\alpha}q^{\beta} + A^{21}(p,q) q^{\alpha}p^{\beta}\bigg\} \, ,
\end{eqnarray}
where $A^{ij}(p,q) = A^{ij}_g(p,q) + A^{ij}_q(p,q)$ which are diagonal ($\propto \delta^{a b}$) in colour space.\\
After an explicit computation, we find
\begin{eqnarray}
A^{00}_g(p,q)
&=&
-\delta_{ab}\,\frac{g^2 \, N_C}{16 \pi^2} \bigg\{
2 \left( p^2 + q^2 + \frac{11}{3} p \cdot q \right)
+ (p^2-q^2) \bigg[ \mathcal B_0(p^2,0,0) - \mathcal B_0(q^2,0,0)\bigg]
\nonumber \\
&&
+ \left(p^4 + q^4 - 2(p^2 + q^2) p \cdot q -6 p^2 q^2 \right) \mathcal C_0((p+q)^2, p^2, q^2,0,0,0)
\bigg\} \, ,
\nonumber
\end{eqnarray}
\begin{eqnarray}
A^{11}_g(p,q) = A^{22}_g(q,p)
&=&
\delta_{ab} \, \frac{g^2 \, N_C \,}{16\,\pi^2 } \,
\bigg\{ 2 + \frac{1}{p \cdot q^2 - p^2 \, q^2} \,
\bigg[ (p+q)^2 \, p \cdot q \, \mathcal B_0((p+q)^2,0,0)
\nonumber \\
&&
- p^2 \, (q^2 + p \cdot q) \, \mathcal B_0(p^2,0,0) - (2 p \cdot q^2 - p^2 \, q^2 + p \cdot q \, q^2) \mathcal B_0(q^2,0,0)
\nonumber \\
&&
+ \left( p^2 \, q^2 ( 5 q^2 - p^2) + 2 p \cdot q^2 (p^2 + p \cdot q - 2 \, q^2 ) \right) \,
\mathcal C_0((p+q)^2, p^2, q^2,0,0,0) \bigg] \bigg\} \, ,
\nonumber
\end{eqnarray}
\begin{eqnarray}
A^{12}_g(p,q)
&=&
\delta_{ab}\, \frac{g^2 \, N_C}{4 \pi^2} \, p \cdot q \, \mathcal C_0((p+q)^2, p^2, q^2,0,0,0)\, ,
\nonumber
\end{eqnarray}
\begin{eqnarray}
A^{21}_g(p,q)
&=&
\delta_{ab}\, \frac{g^2 \, N_C}{24\,\pi^2 } \,
\bigg\{11 + \frac{3}{2}\,\frac{1}{p \cdot q^2 - p^2 \, q^2} (p^2 + q^2) \, \bigg[(p^2 + p \cdot q)\mathcal B_0(p^2,0,0)
\nonumber \\
&&
+ (q^2 + p \cdot q) \mathcal B_0(q^2,0,0) - (p+q)^2 \mathcal B_0((p+q)^2,0,0)
\nonumber \\
&&
- \left( p \cdot q (p^2 + 4 p \cdot q + q^2) - 2 \, p^2 \, q^2 \right) \mathcal C_0((p+q)^2, p^2, q^2,0,0,0) \bigg]\bigg\} \, ,
\nonumber
\end{eqnarray}
\begin{eqnarray}
A^{00}_q(p,q) &=&
\delta_{ab} \frac{g^2}{8 \pi^2} \sum_{i=1}^{n_f} \bigg\{ \frac{2}{3} p\cdot q - 2\, m_i^2
%
+ \frac{m_i^2}{p \cdot q^2 - p^2 q^2} \bigg[
p^2 \left(p \cdot q +q^2\right) \mathcal B_0(p^2,m_i^2,m_i^2 ) \nonumber \\
&+&
q^2 \left(p^2+p \cdot q \right) \mathcal B_0(q^2,m_i^2,m_i^2) -
\left(p^2 \left(p \cdot q+2 q^2\right) + p \cdot q \, q^2\right) \mathcal B_0( (p+q)^2,m_i^2,m_i^2) \nonumber \\
&-&
\left(p^2 q^2 \left(p^2+q^2 -4 m_i^2 \right)+4 m_i^2 \, p \cdot q^2 +4 p^2 \, q^2 \, p \cdot q -2 \, p \cdot q^3\right)
\mathcal C_0 ((p+q)^2,p^2,q^2,m_i^2,m_i^2,m_i^2)
\bigg] \bigg\} , \nonumber
\end{eqnarray}
\begin{eqnarray}
A^{11}_q(p,q) &=& A^{22}_q(q,p) =
\delta_{ab} \frac{g^2}{ 8\,\pi^2} \sum_{i=1}^{n_f} \frac{2 \, m_i^2 \, q^2}{ p \cdot q^2 - p^2 q^2} \bigg\{ -2 + \frac{1}{p \cdot q^2 - p^2 q^2} \bigg[
\left(q^2 \left(p^2+3 \, p \cdot q \right) \right. \nonumber \\
&+& \left. 2 \, p \cdot q^2\right) \mathcal B_0(q^2,m_i^2,m_i^2)
+ \left(p^2 \left(3 \, p \cdot q+q^2\right)+2 p \cdot q^2\right) \mathcal B_0(p^2,m_i^2,m_i^2) \nonumber \\
&-& \left(p^2 \left(3 \, p \cdot q+2 q^2\right) + p \cdot q \left(4 \, p \cdot q+3 q^2\right)\right) \mathcal B_0((p+q)^2,m_i^2,m_i^2) \nonumber \\
&-& \left(2 \, p \cdot q^2 \left(2 m_i^2+p^2+q^2\right) + p^2 q^2 \left(p^2+q^2-4 m_i^2 \right) \right.) \nonumber \\
&+& \left. 4 p^2 q^2 \, p \cdot q +2 \, p \cdot q ^3\right)
\mathcal C_0((p+q)^2,p^2,q^2,m_i^2,m_i^2,m_i^2)
\bigg] \bigg\} , \nonumber
\end{eqnarray}
\begin{eqnarray}
A^{12}_q(p,q) &=&
\delta_{ab} \frac{g^2}{ 8\,\pi^2} \sum_{i=1}^{n_f} \frac{2 \, m_i^2 \, p \cdot q^2}{ p \cdot q^2 - p^2 q^2} \bigg\{ 2 + \frac{1}{p \cdot q^2 - p^2 q^2} \bigg[
-\left(q^2 \left(p^2+3 \, p \cdot q \right)+ 2 \, p \cdot q^2\right) \mathcal B_0(q^2,m_i^2,m_i^2) \nonumber \\
&-&
\left(p^2 \left(3 \, p \cdot q +q^2\right)+ 2 \, p \cdot q^2\right) \mathcal B_0(p^2,m_i^2,m_i^2) +
\left(p^2 \left(3 \, p \cdot q +2 q^2\right)+p \cdot q \left(4 \, p \cdot q +3 q^2 \right)\right) \nonumber \\
&\times& \mathcal B_0( (p+q)^2,m_i^2,m_i^2)+
\left(2 \, p \cdot q^2 \left(2 m_i^2+p^2+q^2\right)+p^2 q^2 \left(p^2+q^2 -4 m_i^2 \right) \right. \nonumber \\
&+& \left. 4 \, p^2 \, q^2 \, p \cdot q +2\, p \cdot q^3\right)
\mathcal C_0((p+q)^2,p^2,q^2,m_i^2,m_i^2,m_i^2)
\bigg] \bigg\} , \nonumber
\end{eqnarray}
\begin{eqnarray}
A^{21}_q(p,q) &=& \delta_{ab} \frac{g^2}{ 8\,\pi^2} \sum_{i=1}^{n_f} \bigg\{ - \frac{2}{3} + \frac{2 \, m_i^2 \, p \cdot q}{p \cdot q - p^2 q^2} + \frac{m_i^2}{(p \cdot q -p^2 q^2)^2} \bigg[
- p^2 \left(q^2 \left(2 p^2+3 \, p \cdot q \right)+p \cdot q ^2\right) \nonumber \\
&\times& \mathcal B_0(p^2, m_i^2,m_i^2)
- q^2 \left(p^2 \left(3 \, p \cdot q+2 q^2\right)+p \cdot q^2\right) \mathcal B_0( q^2, m_i^2,m_i^2)
+ \left(2 p^4 q^2+p^2 \left(6 \, p \cdot q \, q^2 \right. \right. \nonumber \\
&+& \left. \left. p \cdot q^2+2 q^4\right)+p \cdot q^2 q^2\right) \mathcal B_0( (p+q)^2,m_i^2,m_i^2)
+ p \cdot q \left(p^2 q^2 \left(3 p^2+3 q^2 - 4 m_i^2\right) \right. \nonumber \\
&+& \left. 4 \, m_i^2 \, p \cdot q^2
+8 p^2 \, q^2 \, p \cdot q -2 \, p \cdot q ^3\right)
\mathcal C_0( (p+q)^2,p^2,q^2,m_i^2,m_i^2,m_i^2)
\bigg] \bigg\},
\end{eqnarray}
where $N_C$ is the number of colours, $n_f$ is the number of flavour and $m_i$ the mass of the quark.
In the on-shell gluon case, eq. (\ref{OffShellQCD}) reproduces the same interaction responsible for Higgs production at LHC augmented by
an anomaly term.
This is given by
\begin{eqnarray} \label{OnShellQCD}
\Gamma^{\alpha\beta}_{gg}(p,q) =
\frac{i}{\Lambda} \Phi(s) \, u^{\alpha\beta}(p,q) \, ,
\end{eqnarray}
with $u^{\alpha\beta}(p,q)$ defined in eq. (\ref{utensor}), and with the gluon/quark contributions included in the $\Phi(s)$ form factor ($s = k^2 = (p+q)^2$)
\begin{eqnarray} \label{OnShellPhi}
\Phi(s)
=
- \delta^{ab} \frac{g^2}{24\,\pi^2} \, \bigg\{
\, \left(11\, N_C - 2\, n_f \right) + 12 \, \sum_{i=1}^{n_f} m_i^2 \,
\bigg[ \frac{1}{s} \, - \, \frac{1} {2 }\mathcal C_0 (s, 0, 0, m_i^2, m_i^2, m_i^2) \bigg(1-\frac{4 m_i^2}{ s}\bigg) \bigg]
\bigg\} \, ,
\end{eqnarray}
where the first mass independent terms represent the contribution of the anomaly, while the others are the explicit mass corrections. \\
The decay rate of a dilaton in two gluons can be evaluated from the on-shell limit in eq. (\ref{OnShellQCD}) and it is given by
\begin{eqnarray}
\Gamma(\rho \rightarrow gg)
&=&
\frac{\alpha_s^2\,m_\rho^3}{32\,\pi^3 \Lambda^2} \, \bigg| \beta_{QCD} + x_t\left[1 + (1-x_t)\,f(x_t) \right] \bigg|^2 \,,
\label{Phigg}
\end{eqnarray}
where we have taken the top quark as the only massive fermion and $x_i$ and $f(x_i)$ are defined in eq. (\ref{x}) and eq. (\ref{fx})
respectively. Moreover we have set
$\beta_{QCD} = 11 N_C/3 - 2 \, n_f/3$ for the QCD $\beta$ function.
\section{Non-gravitational dilatons from scale invariant extensions of the Standard Model}
\label{NonGrav}
As we have pointed out in the introduction, a dilaton may appear in the spectrum of different extensions of the Standard Model not
only as a result of the compactification of extra spacetime dimensions, but also as an effective state, related to the breaking of
dilatation symmetry. In this respect, notice that in its actual formulation
the Standard Model is not scale invariant, but can be such, at classical level, if we slightly modify the scalar potential with the
introduction of a dynamical field $\Sigma$ that
allows to restore this symmetry and acquires a vacuum expectation value. This task is accomplished by the replacement of every
dimensionful parameter $m$ according to $m \rightarrow m \frac{\Sigma}{\Lambda}$, where $\Lambda$ is the classical conformal
breaking scale.
In the case of the Standard Model, classical scale invariance can be easily accommodated with a simple change of the scalar potential.
This is defined, obviously, modulo a constant, therefore we may consider, for instance, two equivalent choices
\begin{eqnarray}
V_1(H, H^\dagger)&=& - \mu^2 H^\dagger H +\lambda(H^\dagger H)^2 =
\lambda \left( H^\dagger H - \frac{\mu^2}{2\lambda}\right)^2 - \frac{\mu^4}{4 \lambda}\nonumber \\
V_2(H,H^\dagger)&=&\lambda \left( H^\dagger H - \frac{\mu^2}{2\lambda}\right)^2
\end{eqnarray}
which give two {\em different} scale invariant extensions
\begin{eqnarray}
V_1(H,H^\dagger, \Sigma)&=&- \frac{\mu^2\Sigma^2}{\Lambda^2} H^\dagger H +\lambda(H^\dagger H)^2 \nonumber \\
V_2(H,H^\dagger, \Sigma)&=& \lambda \left( H^\dagger H - \frac{\mu^2\Sigma^2}{2\lambda \Lambda^2}\right)^2 \,,
\end{eqnarray}
where $H$ is the Higgs doublet, $\lambda$ is its dimensionless coupling constant, while $\mu$ has the dimension of a mass and,
therefore, is the only term involved in the scale invariant extension. More details of this analysis can be found in
the next section.\\
The invariance of the potential under the addition of constant terms, typical of any Lagrangian, is lifted once we
require the presence of a dilatation symmetry. Only the second choice $(V_2)$ guarantees the existence of a stable ground state
characterized by a spontaneously
broken phase. In $V_2$ we parameterize the Higgs, as usual, around the electroweak vev $v$ as in eq. (\ref{Higgsparam}),
and indicate with $\Lambda$ the vev of the dilaton field $\Sigma = \Lambda + \rho$,
and we have set $\phi^+ = \phi = 0$ in the unitary gauge. \\
The potential $V_2$ has a massless mode due to the existence of a flat direction.
Performing a diagonalization of the mass matrix we define the two mass eigenstates $\rho_0$ and $h_0$, which are given by
\begin{equation}
\left( \begin{array}{c}
{\rho_0}\\
h_0 \\
\end{array} \right)
=\left( \begin{array}{cc}
\cos\alpha & \sin\alpha \\
-\sin\alpha & \cos\alpha \\
\end{array} \right)
\left( \begin{array}{c}
\rho\\
{h} \\
\end{array} \right)
\end{equation}
with
\begin{equation}
\cos\alpha=\frac{1}{\sqrt{1 + v^2/\Lambda^2}}\qquad \qquad \sin\alpha=\frac{1}{\sqrt{1 + \Lambda^2/v^2}}.
\end{equation}
We denote with ${\rho_0}$ the massless dilaton generated by this potential, while
$h_0$ will describe a massive scalar, interpreted as a new Higgs field, whose mass is given by
\begin{equation}
m_{h_0}^2= 2\lambda v^2 \left( 1 +\frac{v^2}{\Lambda^2}\right) \qquad \textrm{with} \qquad v^2=\frac{\mu^2}{\lambda},
\end{equation}
and with $m_h^2=2 \lambda v^2$ being the mass of the Standard Model Higgs.
The Higgs mass, in this case, is corrected by the new scale of the spontaneous breaking of the dilatation symmetry ($\Lambda$),
which remains a free parameter.
The vacuum degeneracy of the scale invariant model can be lifted by the introduction of
extra (explicit breaking) terms which give a small mass to the dilaton field.
To remove such degeneracy, one can introduce, for instance, the term
\begin{equation}
\mathcal{L}_{break}
= \frac{1}{2} m_{\rho}^2 {\rho}^2 + \frac{1}{3!}\, {m_{\rho}^2} \frac{{\rho}^3}{\Lambda} + \dots \, ,
\end{equation}
where $m_{\rho}$ represents the dilaton mass.
It is clear that in this approach the coupling of the dilaton to the anomaly has to be added by hand.
The obvious question to address, at this point, is if one can identify in the effective action of the Standard Model
an effective state which may interpolate between the dilatation current of the same model and the final state with two
neutral currents, for example with two photons. The role of the following sections will be to show
rigorously that such a state can be identified in ordinary perturbation theory in the form of an anomaly pole.
We will interpret this scalar exchange as a composite state whose interactions with the rest of the Standard Model are
defined by the conditions of scale and gauge invariance. In this respect, the Standard Model Lagrangian, enlarged by the introduction
of a potential of the form $V_2(H,H^\dagger,\Sigma)$, which is expected to capture the dynamics of this pseudo-Goldstone mode, could
take the role of a workable model useful for a phenomenological analysis.
We will show rigorously that this state couples to the conformal anomaly by a direct analysis of the $J_DVV$ correlator,
in the form of an anomaly pole, with $J_D$ and $V$ being the dilatation and a vector current respectively.
Usual polology arguments support the fact that a pole in a correlation function is there to indicate that a specific state can be created by
a field operator in the Lagrangian of the theory, or, alternatively, as a composite particle of the same elementary fields.
Obviously, a perturbative hint of the existence of such intermediate state does not correspond to a complete
description of the state, in the same way as the discovery of an anomaly pole in the $AVV$ correlator of QCD (with $A$ being the
axial current) is not equivalent to a proof of the existence of the pion.
\subsection{A classical scale invariant Lagrangian with a dilaton field}
\label{classical}
In this section we briefly describe the construction of a scale invariant theory to clarify some of the issues
concerning the coupling of a dilaton. In particular, the example has the goal to
illustrate that in a classical scale invariant extension of a given theory, the dilaton couples only to operators which are mass dependent,
and thus scale breaking, before the extension. We take the case of a fundamental dilaton field (not a composite) introduced
in this type of extensions.
A scale invariant extension of a given Lagrangian can be obtained if we promote all the dimensionful constants to dynamical
fields.
We illustrate this point in the case of a simple interacting scalar field theory incorporating the Higgs mechanism.
At a second stage we will derive the structure of the dilaton interaction at order $1/\Lambda$, where
$\Lambda$ is the scale characterizing the spontaneous breaking of the dilatation symmetry.
Our toy model consists in a real singlet scalar with a potential of the kind of $V_2(\phi)$ introduced in section \ref{NonGrav},
\begin{equation}
\label{original}
\mathcal L = \frac{1}{2} (\partial \phi)^2 - V_2(\phi) =
\frac{1}{2} (\partial \phi)^2 + \frac{\mu^2}{2}\, \phi^2 - \lambda\, \frac{\phi^4}{4} - \frac{\mu^4}{4\,\lambda}\, ,
\end{equation}
obeying the classical equation of motion
\begin{equation} \label{scalarEOM}
\square \phi = \mu^2\,\phi - \lambda\, \phi^3\, .
\end{equation}
Obviously this theory is not scale invariant due to the appearance of the mass term $\mu$. This feature is reflected in the trace of the EMT.
Indeed the canonical EMT of such a theory and its trace are
\begin{eqnarray}
T^{\mu\nu}_{c}(\phi)
&=&
\partial^\mu \phi\, \partial^\nu \phi
- \frac{1}{2}\,\eta^{\mu\nu} \bigg[ (\partial \phi)^2 + \mu^2 \,\phi^2
- \lambda\,\frac{\phi^4}{2} - \frac{\mu^4}{2\,\lambda} \bigg] \, ,
\nonumber \\
{T_{c}^\mu}_\mu(\phi) &=&
- (\partial\phi)^2 - 2\, \mu^2 \,\phi^2 +\lambda\, \phi^4 + \frac{\mu^4}{\lambda} \, .
\end{eqnarray}
Improving the EMT of the scalar field in such a way as to make its trace proportional only to the scale breaking parameter,
i.e. the mass $\mu$, which is done by adding an extra contribution $T_I^{\mu\nu}(\phi, \chi)$,
\begin{equation}
T_I^{\mu\nu}(\phi,\chi)= \chi\, \left(\eta^{\mu\nu} \square \phi^2 - \partial^\mu \partial^\nu \phi^2 \right) \, ,
\end{equation}
where the parameter $\chi$ is left generic.
The combination of the canonical plus the improvement EMT,
$T^{\mu\nu} \equiv T_c^{\mu\nu} + T_I^{\mu\nu}$ has the off-shell trace
\begin{equation}
{T^\mu}_\mu(\phi,\chi)=
(\partial\phi)^2\, \left( 6 \chi - 1 \right) - 2\, \mu^2\, \phi^2
+ \lambda\, \phi^4 + \frac{\mu^4}{\lambda} + 6 \chi \phi\, \square \phi\, .
\end{equation}
Using the equation of motion (\ref{scalarEOM}) and choosing $\chi=1/6$ the trace relation given above
becomes proportional uniquely to the scale breaking term $\mu$
\begin{equation} \label{ImprovenTrace}
{T^\mu}_\mu(\phi,1/6) = - \mu^2 \phi^2 + \frac{\mu^4}{\lambda} \, .
\end{equation}
The scale invariant extension of the Lagrangian given in eq. (\ref{original}) is achieved by promoting the mass terms to dynamical fields by the replacement
\begin{equation}
\label{rep}
\mu \to \frac{\mu}{\Lambda} \, \Sigma,
\end{equation}
obtaining
\begin{equation}
\label{sigmaphi}
\mathcal L =
\frac{1}{2}\, (\partial \phi)^2 +\frac{1}{2} (\partial \Sigma)^2
+ \frac{ \mu^2}{2\,\Lambda^2}\, \Sigma^2\, \phi^2 - \lambda \frac{\phi^4}{4}
- \frac{\mu^4}{4\,\lambda \, \Lambda^4}\, \Sigma^4
\end{equation}
where we have used eq. (\ref{rep}) and introduced a kinetic term for the dilaton $\Sigma$.
Obviously, the new Lagrangian is dilatation invariant, as one can see from the trace of the improven EMT
\begin{equation}
{T^{\mu}}_{\mu}(\phi,\Sigma,\chi,\chi') = \left( 6\, \chi - 1 \right)\, (\partial\phi)^2
+ \left( 6 \chi^\prime -1\right)\, (\partial\Sigma)^2
+ 6 \chi\, \phi\, \square \phi + 6 \chi^\prime\, \Sigma\, \square \Sigma
- 2\, \frac{\mu^2}{\Lambda^2}\, \Sigma^2\, \phi^2 + \lambda\, \phi^4
+ \frac{1}{\lambda}\,\frac{\mu^4}{\Lambda^4}\,\Sigma^4 \,,
\end{equation}
which vanishes upon using the equations of motion for the $\Sigma$ and $\phi$ fields,
\begin{eqnarray}
\square \phi &=& \frac{\mu^2}{\Lambda^2}\, \Sigma^2\, \phi - \lambda\, \phi^3\, ,
\nonumber \\
\square \Sigma &=& \frac{\mu^2}{\Lambda^2}\, \Sigma\, \phi^2 - \frac{1}{\lambda}\, \frac{\mu^4}{\Lambda^4}\,\Sigma^3 \, ,
\end{eqnarray}
and setting the $\chi, \chi'$ parameters at the special value $\chi=\chi^\prime=1/6$.
As we have already discussed in section \ref{NonGrav}, the scalar potential $V_2$ allows to perform
the spontaneous breaking of the scale symmetry around a stable minimum point,
giving the dilaton and the scalar field the vacuum expectation values $\Lambda$ and $v$ respectively
\begin{equation}
\Sigma = \Lambda + \rho \, , \quad \phi = v + h\, .
\end{equation}
For our present purposes, it is enough to expand the Lagrangian (\ref{sigmaphi}) around the vev for the dilaton field,
as we are interested in the structure of the couplings of its fluctuation $\rho$
\begin{equation} \label{Manifest}
\mathcal L = \frac{1}{2}\, (\partial\phi)^2 + \frac{1}{2}\, (\partial\rho)^2 + \frac{\mu^2}{2}\, \phi^2
- \lambda \, \frac{\phi^4}{4} - \frac{\mu^4}{4\,\lambda}
- \frac{\rho}{\Lambda}\,\left(- \mu^2\, \phi^2 + \frac{\mu^4}{\lambda}\right) + \dots\, ,
\end{equation}
where the ellipsis refer to terms that are higher order in $1/\Lambda$.
It is clear, from (\ref{ImprovenTrace}) and (\ref{Manifest}), that one can write an dilaton Lagrangian
at order $1/\Lambda$, as
\begin{equation} \label{RhoInteraction}
\mathcal L_{\rho} = (\partial\rho)^2 - \frac{\rho}{\Lambda}\, {T^{\mu}}_{\mu}(\phi,1/6) + \dots\, ,
\end{equation}
where the equations of motion have been used in the trace of the energy-momentum tensor.
Expanding the scalar field around $v$ would render the previous equation more complicated and we omit it for definiteness.
We only have to mention that a mixing term $\sim \rho\, h$ shows up and it has to be removed diagonalizing
the mass matrix, switching from interaction to mass eigenstates exactly in the way we discussed in section \ref{NonGrav},
to which we refer for the details.
It is clear, from this simple analysis, that a dilaton, in general, does not couple to the anomaly,
but only to the sources of explicit breaking of scale invariance, i.e. to the mass terms.
The coupling of a dilaton to an anomaly is, on the other hand, necessary,
if the state is interpreted as a composite pseudo Nambu-Goldstone mode of the dilatation symmetry.
Thus, this coupling has to be introduced by hand, in strict analogy with the chiral case.
\subsection{The $J_DVV$ and $TVV$ vertices}
This effective degree of freedom emerges both from the spectral analysis of the $TVV$ \cite{Giannotti:2008cv, Armillis:2009pq} and,
as we are now going to show, of the $J_D VV$
vertices, being the two vertices closely related.
We recall that the dilatation current can be defined as
\begin{equation}
J_D^\mu(z)= z_\nu T^{\mu \nu}(z) \qquad \textrm{with} \qquad \partial\cdot J_D = {T^\mu}_\mu.
\label{def}
\end{equation}
The $T^{\mu\nu}$ has to be symmetric and on-shell traceless for a classical scale invariant theory, and includes, at
quantum level, the contribution from the trace anomaly together with the additional terms describing the explicit breaking of the
dilatation symmetry.
The separation between the anomalous and the explicit contributions to the breaking of dilatation symmetry is present
in all the analysis that we have performed on the $TVV$ vertex in dimensional regularization.
In this respect, the analogy between these types of correlators and the $AVV$ diagram of the chiral anomaly goes
quite far, since in the $AVV$ case such a separation has been shown to hold in the Longitudinal/Transverse
(L/T) solution of the anomalous Ward identities \cite{Knecht:2003xy, Jegerlehner:2005fs, Armillis:2009sm}.
This has been verified in perturbation theory in the same scheme.
We recall that the $U(1)_A$ current is characterized by an anomaly pole which describes the interaction between the
Nambu-Goldstone mode, generated by the breaking of the chiral symmetry, and the gauge currents.
In momentum space this corresponds to the nonlocal vertex
\begin{equation}
\label{AVVpole}
V_{\textrm{anom}}^{\lambda \mu\nu}(k,p,q)= \frac{k^\lambda}{k^2}\epsilon^{\mu \nu \alpha \beta}p_\alpha q_\beta +...
\end{equation}
with $k$ being the momentum of the axial-vector current and $p$ and $q$ the momenta of the two photons.
In the equation above, the ellipsis refer to terms which are suppressed at large energy.
In this regime, this allows to distinguish the operator accounting for the chiral anomaly (i.e. $\square^{-1}$ in coordinate space)
from the contributions due to mass corrections.
Polology arguments can be used to relate the appearance of such a pole to the pion state
around the scale of chiral symmetry breaking.
To identify the corresponding pole in the dilatation current of the $J_D VV$ correlator at zero momentum transfer, one can follow the
analysis of \cite{Horejsi:1997yn}, where it is shown that the appearance of the trace anomaly is related to the presence of a
superconvergent sum rule in the spectral density of this correlator. At non-zero momentum transfer the derivation of a similar
behaviour can be obtained by an explicit computation of the spectral density of the
$TVV$ vertex \cite{Giannotti:2008cv} or of the entire correlator, as done for QED and QCD
\cite{Armillis:2009pq, Armillis:2010qk} and as we will show next.
Using the relation between $J_D^\mu$ and the EMT $T^{\mu\nu}$ we introduce the $J_DVV$ correlator
\begin{eqnarray}
\Gamma_D^{\mu\alpha\beta}(k,p)
&\equiv&
\int d^4 z\, d^4 x\, e^{-i k \cdot z + i p \cdot x}\,
\left\langle J^\mu_D(z) V^\alpha(x)V^\beta(0)\right\rangle
\label{gammagg}
\end{eqnarray}
which can be related to the $TVV$ correlator
\begin{eqnarray}
\Gamma^{\mu\nu\alpha\beta}(k,p)&\equiv& \int d^4 z\, d^4 x\, e^{-i k \cdot z + i p \cdot x}\,
\left\langle T^{\mu \nu}(z) V^\alpha(x) V^\beta(0)\right\rangle
\end{eqnarray}
according to
\begin{eqnarray}
\Gamma_D^{\mu\alpha\beta}(k,p)&=&
i \frac{\partial}{\partial k^\nu}\Gamma^{\mu\nu\alpha\beta}(k,p) \,.
\end{eqnarray}
As we have already mentioned, this equation allows us to identify a pole term in the $J_DVV$ diagram from the corresponding pole
structure in the $TVV$ vertex.
In the following we will show the emergence of the anomaly poles in the QED and QCD cases.
\subsection{The dilaton anomaly pole in the QED case}
\label{SecPoleQED}
For definiteness, it is convenient to briefly review the characterization of the $TVV$ vertex in the QED case with a massive fermion
(see \cite{Armillis:2009pq} for more details).
The full amplitude can be expanded in a specific basis of $13$ tensors first identified in \cite{Giannotti:2008cv}
\begin{eqnarray}
\Gamma^{\mu\nu\alpha\beta}(p,q) = \, \sum_{i=1}^{13} F_i (s; s_1, s_2,m^2)\phi_i^{\mu\nu\alpha\beta}(p,q)\,,
\label{Gamt}
\end{eqnarray}
where the $13$ invariant amplitudes $F_i$ are functions of the kinematic invariants $s=k^2=(p+q)^2$, \mbox{$s_1=p^2$}, $s_2=q^2$,
with $p$ and $q$ the momenta of the external photons,
and of the internal fermion mass $m$. The list of the tensor structures $\phi_i$ can be found in \cite{Giannotti:2008cv}. The number
of these form factors reduces from $13$ to $3$ in the case of on-shell photons. For our purposes, being interested in the appearance
of the anomaly poles, we only need the contributions that generate a non zero trace. These come from the tensors
$\phi_1^{\mu\nu\alpha\beta}$ and $\phi_2^{\mu\nu\alpha\beta}$ which are
\begin{eqnarray}
\phi_1^{\mu\nu\alpha\beta}&=&\left(k^2 \eta^{\mu\nu} - k^{\mu } k^{\nu}\right) u^{\alpha\beta}(p,q), \nonumber \\
\phi_2^{\mu\nu\alpha\beta}&=&\left(k^2 \eta^{\mu\nu} - k^{\mu } k^{\nu}\right) w^{\alpha\beta}(p,q),
\end{eqnarray}
where
\begin{eqnarray}
&&u^{\alpha\beta}(p,q) \equiv (p\cdot q) \eta^{\alpha\beta} - q^{\alpha}p^{\beta}\,,\nonumber \\
&&w^{\alpha\beta}(p,q) \equiv p^2 q^2 \eta^{\alpha\beta} + (p\cdot q) p^{\alpha}q^{\beta}
- q^2 p^{\alpha}p^{\beta} - p^2 q^{\alpha}q^{\beta}.\,
\label{uwdef}
\end{eqnarray}
For two on-shell final state photons ($s_1=s_2=0$) and a massive fermion we obtain
\begin{equation}
\label{oom}
{F_1 (s;\,0,\,0,\,m^2)} =
F_{1\, pole} \, + \, \frac{e^2 \, m^2}{3 \, \pi ^2 \, s^2} \,
- \frac{e^2 \, m^2}{3 \, \pi^2 \, s} \, \mathcal C_0 (s, 0, 0, m^2,m^2,m^2)
\bigg[\frac{1}{2}-\frac{2 \,m^2}{ s}\bigg], \\
\end{equation}
where
\begin{equation} F_{1\, pole}=- \frac{e^2 }{18 \, \pi^2 s}
\end{equation}
and the scalar $3$-point function $ \mathcal C_0 (s, 0,0,m^2,m^2,m^2) $ is given by
\begin{equation}
\mathcal C_0 (s, 0,0,m^2,m^2,m^2) = \frac{1}{2 s} \, \log^2 \frac{a_3+1}{a_3-1}, \qquad \mbox{with} \quad a_3 = \sqrt {1-4m^2/s} \,.
\end{equation}
In the massless fermion case two properties of this expansion are noteworthy: 1) the trace anomaly
takes contribution only from a single tensor structure $(\phi_1)$ and invariant amplitude $(F_1)$ which coincides with the pole term;
2) the residue of this pole as $s\to 0$ is non-zero, showing that the pole is coupled in the infrared. Notice that the form factor
$F_2$, which in general gives a non-zero contribution to the trace in the presence of mass terms, is multiplied by a tensor
structure ($\phi_2$) which {\em vanishes} when the two photons are on-shell.
Therefore, similarly to the case of the chiral anomaly, also in this case the anomaly is {\em entirely} given by the appearance
of an anomaly pole.
We stress that this result is found to be exact in dimensional regularization, which is a mass independent scheme:
at perturbative level, the anomalous breaking of the dilatation symmetry, related to an anomaly pole in the spectrum of all
the gauge-invariant correlators studied in this work, is separated from the sources of {\em explicit} breaking. The latter are related to the mass
parameters and/or to the gauge bosons virtualities $p^2$ and $q^2$.
To analyse the implications of the pole behaviour discussed so far for the $TVV$ vertex and its connection with the $J_DVV$
correlator, we limit our attention on the anomalous contribution ($F_1\, \phi_1^{\mu\nu\alpha\beta}$),
which we rewrite in the form
\begin{equation}
\Gamma_{pole}^{\mu\nu\alpha \beta}(k,p)\equiv
- \frac{e^2}{18\pi^2}\frac{1}{k^2}\left( \eta^{\mu \nu} k^2 - k^\mu k^\nu\right) u^{\alpha \beta}(p,q) \, ,
\qquad q = k - p \, .
\label{uref}
\end{equation}
This implies that the $J_D VV$ correlator acquires a pole as well
\begin{eqnarray}
\Gamma_{D\, pole}^{\mu\alpha\beta}
&=&
- i \frac{e^2}{18 \pi^2}\frac{\partial}{\partial k^\nu}
\left[ \frac{1}{k^2}\, \left( \eta^{\mu\nu } k^2 - k^{\mu} k^{\nu}\right) u^{\alpha \beta}(p,k-p) \right]
\end{eqnarray}
and acting with the derivative on the right hand side we finally obtain
\begin{equation}
\Gamma_{D \, pole}^{\mu\alpha\beta}(k,p)=
i\, \frac{e^2}{6 \pi^2}\frac{k^\mu}{k^2}u^{\alpha \beta}(p,k-p) - i \frac{e^2}{18 \pi^2}\frac{1}{k^2}
\left( \eta^{\mu\nu } k^2 - k^{\mu} k^{\nu}\right)\frac{\partial}{\partial k_\nu}u^{\alpha \beta}(p,k-p).
\end{equation}
Notice that the first contribution on the right hand side of the previous equation corresponds to an anomaly pole,
shown pictorially in fig. \ref{dilatonpole}. In fact, by taking a derivative of
the dilatation current only this term will contribute to the corresponding Ward identity
\begin{equation}
k_\mu \, \Gamma_D^{\mu\alpha\beta}(k,p) = i \frac{e^2}{6 \pi^2}u^{\alpha \beta}(p,k-p),
\label{res}
\end{equation}
which is the expression in momentum space of the usual relation $\partial J_D\sim FF$, while the second term trivially vanishes.
Notice that the pole in (\ref{res}) has disappeared, and we are left just with its residue on the r.h.s., or,
equivalently, the pole is removed in eq. (\ref{uref}) if we trace the two indices $(\mu,\nu)$.
\begin{figure}[t]
\begin{center}
\includegraphics[scale=1.2]{figures/JDAA.eps}
\end{center}
\caption{Exchange of a dilaton pole mediated by the $J_D V V$ correlator. }
\label{dilatonpole}
\end{figure}
\subsection{The dilaton anomaly pole in the QCD case}
The analysis presented for the dilatation current of QED can be immediately generalized to the case of QCD.
Following a similar reasoning, we expand the on-shell $TVV$ vertex, with $V$ denoting now the gluon, as
\begin{eqnarray}
\Gamma^{\mu\nu\alpha\beta}(p,q) = \delta^{a b} \sum_{i = 1}^{3} \Phi_i(s; 0,0) \, t_i^{\, \mu\nu\alpha\beta}(p,q) \qquad \mbox{with}
\quad p^2 = q^2 = 0 \, ,
\end{eqnarray}
with the tensor basis given by
\begin{eqnarray}
t_1^{\, \mu \nu \a \b} (p,q) &=&
(s \, \eta^{\mu\nu} - k^{\mu}k^{\nu}) \, u^{\a \b} (p,q),
\label{widetilde1} \nonumber \\
t_2^{\, \mu \nu \a \b} (p,q) &=& - 2 \, u^{\a \b} (p,q) \left[ s \, \eta^{\mu \nu} + 2 (p^\mu \, p^\nu + q^\mu \, q^\nu )
- 4 \, (p^\mu \, q^\nu + q^\mu \, p^\nu) \right],
\label{widetilde2} \nonumber \\
t^{\, \mu \nu \alpha \beta}_{3} (p,q) &=&
\big(p^{\mu} q^{\nu} + p^{\nu} q^{\mu}\big)g^{\alpha\beta}
+ \frac{s}{2} \left(\eta^{\alpha\nu} \eta^{\beta\mu} + \eta^{\alpha\mu} \eta^{\beta\nu}\right) \nonumber \\
&&
- \eta^{\mu\nu} \left(\frac{s}{2} \eta^{\alpha \beta}- q^{\alpha} p^{\beta}\right)
-\left(\eta^{\beta\nu} p^{\mu} + \eta^{\beta\mu} p^{\nu}\right)q^{\alpha}
- \big (\eta^{\alpha\nu} q^{\mu} + \eta^{\alpha\mu} q^{\nu }\big)p^{\beta},
\label{widetilde3}
\end{eqnarray}
where $\delta^{a b}$ is the diagonal matrix in colour space.
Again we have $s=k^2 = (p+q)^2$, with the virtualities of the two gluons being $p^2 = q^2 = 0$.
Notice that in the massless fermion limit only the first ($t_1$) of these 3 form factors contributes to the anomaly.
The corresponding on-shell form factors with massive quarks are
\begin{eqnarray}
\Phi_{1}(s;0,0) &=& - \frac{g^2}{72 \pi^2 \,s}\left(2 n_f - 11 N_C \right) + \frac{g^2}{6 \pi^2}\sum_{i=1}^{n_f} m_i^2 \, \bigg\{
\frac{1}{s^2} \, - \, \frac{1} {2 s}\mathcal C_0 (s, 0, 0, m_i^2,m_i^2,m_i^2)
\bigg[1-\frac{4 m_i^2}{ s}\bigg] \bigg\}, \,
\label{Phi1} \nonumber \\
\Phi_{2}(s;0,0) &=& - \frac{g^2}{288 \pi^2 \,s}\left(n_f - N_C\right) \nonumber \\
&-& \frac{g^2}{24 \pi^2} \sum_{i=1}^{n_f} m_i^2 \, \bigg\{ \frac{1}{s^2}
+ \frac{ 3}{ s^2} \mathcal D_0 (s, 0, 0, m_i^2, m_i^2)
+ \frac{ 1}{s } \mathcal C_0(s, 0, 0, m_i^2,m_i^2,m_i^2 )\, \left[ 1 + \frac{2 m_i^2}{s}\right]\bigg\},
\label{Phi2} \nonumber \\
\Phi_{3}(s;0,0) &=& \frac{g^2}{288 \pi^2}\left(11 n_f - 65 N_C\right) - \frac{g^2 \, N_C}{8 \pi^2}
\bigg[ \frac{11}{6} \mathcal B_0(s,0,0) - \mathcal B_0(0,0,0) + s \,\mathcal C_0(s,0,0,0,0,0) \bigg] \nonumber \\
&+& \frac{g^2}{8 \pi^2} \sum_{i=1}^{n_f}\bigg\{ \frac{1}{3}\mathcal B_0(s, m_i^2,m_i^2) + m_i^2 \, \bigg[
\frac{1}{s}
+ \frac{5}{3 s} \mathcal D_0 (s, 0, 0, m_i^2) + \mathcal C_0 (s, 0,0,m_i^2,m_i^2,m_i^2) \,\left[1 + \frac{2 m_i^2}{s}\right]
\bigg]\bigg\} , \nonumber \\
\label{Phi3}
\end{eqnarray}
where $m_i$ denotes the quark mass, and we have summed over the fermion flavours $(n_f)$, while $N_C$ denotes the number of colours.
Notice the appearance of the $1/s$ pole in $\Phi_1$, which saturates the contribution to the trace
anomaly in the massless limit which becomes
\begin{equation}
\Phi_{1}(s;0,0) = - \frac{g^2}{72 \pi^2 \,s}\left(2 n_f - 11 N_C\right).
\label{polepole}
\end{equation}
As for the QED case, this is the only invariant amplitude which contributes to the anomalous trace part of the correlator.
The pole completely accounts for the trace anomaly and is clearly inherited by the QCD dilatation current,
for the same reasonings discussed above.
\subsection{Mass corrections to the dilaton pole}
The discussion of the mass corrections to the massless dilaton can follow quite closely the strategy adopted in
the pion case using partially conserved axial currents (PCAC) techniques. Also in this case, as for PCAC in the past,
one can assume a partially conserved dilaton current (PCDC) in order to relate the decay amplitude of the dilaton $f_\rho$
to its mass $m_\rho$ and to the vacuum energy. \\
For this goal we define the 1-particle transition amplitudes for the dilatation current and the EMT between the vacuum and a dilaton
state with momentum $p_\mu$
\begin{eqnarray}
\langle 0| J^\mu_D(x) |\rho, p \rangle
&=&
- i \, f_\rho \, p^\mu \, e^{-i p \cdot x}
\nonumber \\
\langle 0| T^{\mu\nu}(x) |\rho, p \rangle
&=&
\frac{f_\rho}{3} \, \left( p^\mu p^\nu - \eta^{\mu\nu} \, p^2\right)\, e^{-i p \cdot x},
\end{eqnarray}
both of them giving
\begin{equation}
\label{rel1}
\partial_\mu \langle 0| J^\mu_D(x) |\rho, p\rangle = \eta_{\mu\nu} \langle 0| T^{\mu\nu}(x) |\rho, p \rangle = - f_\rho \, m_\rho^2 \, e^{-i p \cdot x}.
\end{equation}
We introduce the dilaton interpolating field $\rho(x)$ via a PCDC relation
\begin{equation}
\partial_\mu J^\mu_D(x) = - f_\rho \, m_\rho^2 \, \rho(x)
\label{pcdcrel}
\end{equation}
with
\begin{equation}
\langle 0|\rho(x)|\rho, p\rangle = e^{-i p \cdot x}
\end{equation}
and the matrix element
\begin{equation}
\mathcal A^\mu(q)= \int d^4 x \, e^{i q \cdot x} \,\langle 0| T \{ J^\mu_D(x) {T^\alpha}_\alpha(0) \} |0 \rangle,
\label{www}
\end{equation}
where $T\{\ldots\}$ denotes the time ordered product. \\
Using dilaton pole dominance we can rewrite the contraction of $q_\mu$ with this correlator as
\begin{eqnarray}
\label{inter}
\lim_{q_\mu \to 0} q_\mu \, \mathcal A^\mu(q) = f_\rho \,\left\langle \rho, q=0| {T^\alpha}_\alpha(0) |0\right\rangle \,,
\end{eqnarray}
where the soft limit $q_\mu \to 0$ with $q^2 \gg m_\rho^2 \sim 0$ has been taken. \\
At the same time the dilatation Ward identity on the amplitude $\mathcal A^{\mu}(q)$ in eq. (\ref{www}) gives
\begin{eqnarray}
\label{WIJD}
q_\mu \mathcal A^\mu(q)
&=&
i \int d^4 x\,e^{i q \cdot x}\, \frac{\partial}{\partial x^\mu}\,
\left\langle 0 \left| T\{ J^\mu_D(x) {T^\alpha}_\alpha(0) \} \right|0\right\rangle
\nonumber \\
&=&
i \int d^4 x \, e^{i q \cdot x} \,
\left\langle 0 \left| T\{ \partial_\mu J^\mu_D(x) {T^\alpha}_\alpha(0)\} \right|0\right\rangle
+ i \int d^4 x \, e^{i q \cdot x} \, \delta(x_0) \,
\left\langle 0\left| \left[ J_D^0(x), {T^\alpha}_\alpha(0)\right] \right|0\right\rangle.
\end{eqnarray}
The commutator of the time component of the dilatation charge density and the trace of the EMT can be rewritten as
\begin{equation}
\left[ J_D^0(0, {\bf x}), {T^\alpha}_\alpha(0)\right] = -i \delta^3 ({\bf x})\left(
d_T + x\cdot \partial\right){T^\alpha}_\alpha(0)
\label{derr}
\end{equation}
where $d_T$ is the canonical dimension of the EMT $(d_T=4)$.
Inserting eq. (\ref{derr}) in the Ward identity (\ref{WIJD}) and neglecting the first term due to the nearly
conserved dilatation current ($m_\rho \sim 0 $), we are left with
\begin{equation}
q_\mu \mathcal A^\mu(q)= d_T \, \left\langle 0|{T^\alpha}_\alpha(0) |0\right\rangle.
\label{interdue}
\end{equation}
In the soft limit, with $q^2 \gg m_\rho^2$, comparing eq. (\ref{inter}) and eq. (\ref{interdue}) we obtain
\begin{eqnarray}
\lim_{q^\mu\to 0} q^\mu \mathcal{A}_\mu =
f_\rho \, \left\langle \rho, q=0| {T^\alpha}_\alpha(0) |0\right\rangle = d_T \, \left\langle 0|{T^\alpha}_\alpha(0) |0\right\rangle.
\end{eqnarray}
Introducing the vacuum energy density $\epsilon_{vac}=\left\langle 0|T^0_0 |0\right\rangle = \frac{1}{4} \left\langle
0|T^\alpha_\alpha(0) |0\right\rangle$ and using the relation in eq. (\ref{rel1}) we have
\begin{equation}
\left\langle\rho, p=0| {T^\mu}_\mu |0\right\rangle =- f_\rho m_\rho^2 = \frac{d_T}{f_\rho} \epsilon_{vac}
\end{equation}
from which we finally obtain ($d_T = 4$)
\begin{equation}
f_\rho^2 m_\rho^2 = -16 \, \epsilon_{vac}.
\end{equation}
This equation fixes the decay amplitude of the dilaton in terms of its mass and the vacuum energy.
Notice that $\epsilon_{vac}$ can be related both to the anomaly and possibly to explicit contributions of the
breaking of the dilatation symmetry since
\begin{equation}
\label{beta}
\epsilon_{vac}= \frac{1}{4} \, \left\langle 0 \left| \frac{\beta(g)}{2 g} F_{\mu\nu}F^{\mu\nu} \right|0\right\rangle + ...
\end{equation}
where the ellipsis saturate the anomaly equation with extra mass-dependent contributions, which may be far larger in size then
the anomaly term.
In (\ref{beta}) we have assumed, for simplicity, the coupling of the pole to a single gauge field, with
a beta function $\beta(g)$, but obviously, it can be generalized to several gauge fields.
In the case of higher dimensional operators we would get
\begin{equation}
\epsilon_{vac} = \frac{1}{4} \, \left\langle 0 \left| \frac{\beta(g)}{2 g} \, F_{\mu\nu}\,F^{\mu\nu} \right|0\right\rangle
+ \sum_i g_i \, (d_i-4) \, \left\langle 0| O_i | 0 \right\rangle,
\end{equation}
valid around the scale at which the PCDC approximation holds. Therefore a massless pole can be corrected nonperturbatively according
to some completion theory, causing its mass to shift. Perturbation theory
gives indications about the interpolating fields which can couple to it, as we have seen by exploiting the chiral analogy, but not
more than that. The corrections are model-dependent and can be the subject of additional
phenomenological searches, but the dilatation current takes the role, with no doubt, of an interpolating field for the propagation of
such a scalar intermediate state.
\section{The infrared coupling of an anomaly pole and the anomaly enhancement}
It is easy to figure out from the results of the previous sections that the coupling of a (graviscalar) dilaton to the anomaly
causes a large enhancement of its $2$-photons and $2$-gluons decays.
One of the features of the graviscalar interaction is that its coupling includes
anomalous contributions which are part both of the $2$-photons and of the $2$-gluons cross sections. For this reason,
if an enhancement with respect to the Standard Model rates is found only in one of these two channels and it is associated
to the exchange of a spin zero intermediate state, this result could be used to rule out the exchange of a graviscalar.
On the other hand, for an effective dilaton, identified by an anomaly pole in the $J_D VV$ correlator of the Standard Model, the case is more
subtle, since the coupling of this effective state to the anomaly has to be introduced by hand. This state should be identified, in the perturbative
picture, with the corresponding anomaly pole. The situation, here, is closely similar to the pion case:
in fact, in perturbative QCD, the anomalous $AVV$ diagram is characterized by the presence of an anomaly pole in the variable
$k^2$, with $k$ denoting the momentum of the axial-vector current, which is explicitly shown in eq. (\ref{AVVpole}).
It is interesting to note that this structure has a non-vanishing residue for on-shell photons and for massless quarks running
in the loop. In this case the pole is said to be \emph{infrared coupled}.
This feature, supplemented by usual polology arguments, leads to a $\pi\to \gamma \gamma$ decay rate which
is enhanced with respect to the non-anomalous case.
On the other hand, if the photons are virtual or the quarks are massive the anomaly pole decouples, namely, its residue is zero.
We refer to \cite{Armillis:2009sm} for more details.
The same behaviour is shared by the conformally anomalous $TVV$ diagram \cite{Armillis:2009pq}, so
let us illustrate this important point in the QED case by considering the off-shell correlator.
We denote with $s_1$ and $s_2$ the virtualities of the two final state photons and with $m$ the mass
of the fermion running in the loops.
The case with on-shell photons and a massive fermion has already been discussed in section \ref{SecPoleQED}. There we have shown that
the anomaly pole has a non-vanishing residue only in the conformal limit, when all masses are set to zero.
Indeed, in the case of a massive fermion, besides the fact that the anomaly pole anyway appears in the corresponding invariant
amplitude $F_1 (s;\,0,\,0,\,m^2)$, as one can see from eq. (\ref{oom}), it will decouple, showing a zero residue
\begin{eqnarray}
\lim_{s\rightarrow0} \, s \, \Gamma^{\mu\nu\a\b}(s;\,0,\,0,\,m^2) =0 \,.
\end{eqnarray}
As for the chiral anomaly case, the absence of the internal fermion masses is not sufficient to guarantee the infrared coupling of
the anomaly pole. Indeed, if $m=0$ but the photons are taken off-shell, being characterized by non zero virtualities $s_1$ and $s_2$,
one can check that the entire correlator is completely free from anomaly poles as
\begin{eqnarray}
\lim_{s\rightarrow0} \, s \, \Gamma^{\mu\nu\a\b}(s;\,s_1,\,s_2,\,0) =0 \,.
\end{eqnarray}
The computation of this limit needs the explicit results for all the invariant amplitudes $F_i$, which are not given here
due to their lengthy expressions but can be found in \cite{Armillis:2009pq}. \\
One should be aware of the fact that the same pole is present in the $AVV$ diagram when $VV$ are now the gluons. If the two gluons
are on-shell, as in the 2-photon case, the perturbative anomaly pole is again infrared coupled. Obviously, such an an enhancement is
not observable, since the gluons cannot be on-shell, because of confinement. In the perturbative picture, a
non-zero virtuality of the two gluons is then sufficient to exclude an infrared coupling of the anomaly pole.
We feel, however, that a simple perturbative analysis may not be completely sufficient to decide whether or not the coupling of such a state to
the gluon anomaly takes place. On the other hand, there is no doubt, by the same reason, that such a coupling should occur in the 2-photon
case, being the photons massless asymptotic states. In this case the corresponding anomaly pole of the $J_D \gamma\gamma$ vertex is
infrared coupled.
In general, in the case of an effective dilaton, one is allowed to write down a Lagrangian which is assumed to be scale invariant and, at a second
stage, introduce the direct coupling of this state to the trace anomaly. The possibility of coupling such a state to the photon and to the gluons or
just to the photons, for instance, is a delicate issue for which a simple perturbative approach is unable to offer a definite answer.
\section{Quantum conformal invariance and dilaton couplings at low energy}
Similar enhancements are present in the case of quantum scale invariant extensions of the Standard Model \cite{Goldberger:2007zk}, where
one assumes that the spectrum of the theory is extended with new massive states in order to set the $\beta$ functions of the gauge couplings to
vanish. In a quantum scale invariant theory such as the one discussed in \cite{Goldberger:2007zk}, the dilaton couples only to massive states,
but the heavy mass limit and the condition of the vanishing of the complete $\beta$ functions, leave at low energy a dilaton interaction
proportional only to the $\beta$ functions of the low energy states.
The "remnant" low energy interaction is mass-independent and coincides with that due to a typical anomalous coupling,
although its origin is of different nature, since anomalous contributions are genuinely mass-independent.
For this reason, the decays of a dilaton produced by such extensions carries anomaly-like enhancements as in the graviscalar case.
Obviously, such enhancements to the low energy states of the Standard Model would also be typical of the decay of a Higgs field,
which couples proportionally to the mass of an intermediate state, if quantum scale invariance is combined with the decoupling of a heavy sector.
This, in general, causes an enhancement of the Higgs decay rates into photons and gluons. A partial enhancement only of the di-photon
channel could be accomplished, in this approach, by limiting the above quantum scale invariant arguments only to the electroweak sector.
As a second example, we consider the situation in which all the SM fields are embedded in a (quantum) Conformal Field Theory (CFT)
extension \cite{Goldberger:2007zk} and we discuss the (loop-induced) couplings of the dilaton to the massless gauge bosons.
At tree level the dilaton of \cite{Goldberger:2007zk} couples to the SM fields only through their masses, as the fundamental dilaton which we
have discussed previously, and, in this respect, it behaves like the SM Higgs, without scale anomaly contributions.
For this reason the dilaton interaction with the massless gauge bosons is induced by quantum effects mediated by heavy particles
running in the loops (in this context heavier or lighter is referred to the dilaton mass), and not by anomalous terms.
When the mass $m_i$ of the particle running in the loop is much greater than the dilaton mass, the coupling to the massless
gauge bosons becomes
\begin{eqnarray}
\mathcal L_{\rho} = \frac{\alpha_s}{8 \pi} \sum_i b_g^i \, \frac{\rho}{\Lambda} (F_{g \, \mu\nu}^a)^2 Ê
+ Ê \frac{\alpha_{em}}{8 \pi} \sum_i b_{em}^i \, \frac{\rho}{\Lambda} (F_{\gamma \, \mu\nu})^2 Ê \,,
\end{eqnarray}
where $b_{em}^i$ and $b_{g}^i$ are the contributions of the heavy field $i$ to the $1$-loop $\beta$ function (computed in the
$\overline{MS}$ scheme) for the electromagnetic and strong coupling constants respectively. The $\beta$ functions are normalized as
\begin{eqnarray}
\beta_i = \frac{g^3}{16 \pi^2} b^i \,.
\end{eqnarray}
Note that this result is independent from the heavy mass $m_i$ as one can prove by analysing the structure of the mass corrections
of the dilaton coupling, which reads as
\begin{equation}
\Gamma_{\rho V V}\sim \frac{g^2}{\pi^2 \Lambda} \, m_i^2 \,
\bigg[ \frac{1}{s} - \frac{1} {2 }\mathcal C_0 (s, 0, 0, m_i^2, m_i^2, m_i^2) \bigg(1-\frac{4 m_i^2}{ s}\bigg)\bigg]
\sim
\frac{g^2}{\pi^2 \Lambda} \, \frac{1}{6} + O\left( \frac{s}{m_i^2} \right)
\end{equation}
where $s=m_{\rho}^2$ is fixed at the dilaton mass and we have performed the large mass limit of the amplitude using
\begin{eqnarray}
\mathcal C_0(s, 0, 0, m_i^2, m_i^2, m_i^2) \sim - \frac{1}{2 m_i^2} \left( 1 + \frac{1}{12} \frac{s}{m_i^2}
+ O(\frac{s^2}{m_i^4} ) \right)
\end{eqnarray}
valid for $m_i^2 \gg s = m_{\rho}^2$. This shows that in the case of heavy fermions,
the dependence on the fermion mass cancels.
Obviously, this limit generates an effective coupling which is proportional to the $\beta$ function related to the heavy flavours.
The same reasonings can be employed to the Higgs case as well. It clear that this coupling to the massless gauge bosons is dependent
from new heavy states and, therefore, from the UV completion of the SM. This is certainly the case for the Standard Model Higgs whose
double photon decay is one of the most important decay channel for new physics discoveries. \\
For the dilaton case the situation is slightly different. Surely we do not understand the details of the CFT extension, nor its
particle spectrum, but nevertheless we know that the conformal symmetry is realized at the quantum level. Therefore the complete
$\beta$ functions, including the contribution from all states, must vanish
\begin{eqnarray}
\beta = \frac{g^3}{16 \pi^2} Ê\bigg[ \sum_{i} b^i Ê+ \sum_{j} b^j \bigg] = 0 \,,
\end{eqnarray}
where $i$ and $j$ run over the heavy and light states respectively. Exploiting the consequence of the quantum conformal symmetry, the
dilaton couplings to the massless gauge bosons become
\begin{eqnarray}
\mathcal L_{\rho} =
- \frac{\alpha_s}{8 \pi} \sum_j b_g^j \, \frac{\rho}{\Lambda} (F_{g \, \mu\nu}^a)^2 Ê - Ê \frac{\alpha_{em}}{8 \pi}
\sum_j b_{em}^j \, \frac{\rho}{\Lambda} (F_{\gamma \, \mu\nu})^2 Ê \,,
\end{eqnarray}
in which the dependence from the $\beta$ functions of the light states is now explicit. We emphasize that the appearance of the light
states contributions to the $\beta$ functions is a consequence of the vanishing of the complete $\beta$, and, therefore, of the CFT
extension and not the result of a direct coupling of the dilaton to the anomaly.
\section{Conclusions}
We have presented a general discussion of dilaton interactions with the neutral currents sector of the Standard Model. In the case of a
fundamental graviscalar as a dilaton, we have presented the complete electroweak corrections to the corresponding interactions and we have
discussed the renormalization properties of the same vertices. In particular, we have shown that the renormalizability of the dilaton vertices is
inherited directly from that of the Standard Model only if the Higgs sector is characterized by a conformal coupling ($\chi$)
fixed at the value $1/6$.
Then we have moved to an analysis of the analytic structure of the $J_D VV$ correlator, showing that it supports an anomaly pole as an
interpolating state, which indicates that such a state can be interpreted as the Nambu-Goldstone (effective dilaton) mode of the anomalous
breaking of the dilatation symmetry.
In fact, the trace anomaly seems to bring in some important information concerning the dynamics of the Standard Model, aspects that we have
tried to elucidate. For this reason, we have extended a previous analysis of ours of the $TVV$ vertex, performed in the broken electroweak
phase and in QCD, in order to characterize the dynamical behaviour of the analogous $J_D VV$ correlator. The latter carries relevant
information on the anomalous breaking of the dilatation symmetry in the Standard Model. In fact, as we move to high energy, far above the
electroweak scale, the Lagrangian of the Standard Model becomes approximately scale invariant. This approximate dilatation
symmetry is broken by a quantum anomaly and its signature, as we have shown in our analysis, is in the appearance of
an anomaly pole in the $J_DVV$ correlator. The same pole might appear in correlators with multiple insertions of $J_D$, but the proof
of their existence is far more involved and requires further investigations.
This pole is clearly massless in the perturbative picture, and accounts for the anomalous breaking
of this approximate scale invariance.
\clearpage{\pagestyle{empty}\cleardoublepage}
\chapter{Higher order dilaton interactions in the nearly conformal limit of the Standard Model}\label{Traced4T}
\section{Introduction}
In the previous chater we have elaborated on dilaton interactions with neutral gauge currents in QCD and in the Standard Model
and we have seen that the two main features distinguishing dilaton vertices from those of the Higgs field are respectively the appearance of the
conformal symmetry breaking scale $\Lambda$ and the anomalous enhancements, which are present both for the fundamental graviscalar and
the effective, composite dilaton field.
The whole analysis was carried out including all the mass terms in the $1$-loop radiative corrections.
In this chapter we turn to the investigation of dilaton self-interactions in the nearly conformal limit of the Standard Model,
in which conformal symmetry is broken only by the dilatation anomaly, through a hierarchy of anomalous Ward identities for the divergence
of its dilatation current. In this approximation, the identities allow to extract the coupling of the dilaton to the trace anomaly,
which we compute up to the quartic order in the conformal breaking scale. Our approach can be easily extended to discuss the
anomaly contributions to the dilaton effective action to an arbitrarily high order. They allow to make a distinction between
the Higgs and the dilaton at a phenomenological level.
The possibility that the Standard Model is characterized at high energy by a nearly conformal dynamics has motivated several investigations
spanning considerable time \cite{Goldberger:2007zk,Buchmuller:1987uc, Buchmuller:1988cj,Shaposhnikov:2008xi, Meissner:2006zh}.
If not for a quadratic term present in the Higgs potential, the model would in fact enjoy a dilatation symmetry which
is broken by the vev of the Higgs field in the process of spontaneous symmetry breaking.
A dilaton couples to the trace of the energy-momentum tensor (EMT) ${T^\mu}_\mu$, and the coupling is affected by a trace anomaly. The
trace anomaly equation plays a key role in characterizing the dynamics of the dilaton interactions, with a breaking of the dilatation symmetry which
is enforced by two different contributions.
They can be easily identified from the structure of the corresponding Ward identity satisfied by the
dilaton $(\rho)VV$ vertex \cite{Coriano:2012nm}, with $V$ denoting a neutral (or charged) vector current, but also of higher vertices, such
as the cubic ($TTT$) and quartic ($TTTT$) dilaton interactions, which are part of the dilaton effective action as well.
One specific contribution is the coupling of the dilaton to the anomaly, the second one being related to explicit mass terms generated
at the electroweak scale.
In fact, the basic trace anomaly equation which takes the role of the generating functional of all the Ward identities
satisfied by the dilaton vertices is given by
\begin{equation}
g_{\mu\nu}\left\langle T^{\mu\nu} \right\rangle_s = \mathcal A[g] + \left\langle {T^\mu}_{\mu} \right\rangle_s \, ,
\label{Ch2TraceAnomaly1}
\end{equation}
where the anomalous $(\mathcal A[g])$ and the explicit $(\left\langle T^{\mu\nu} \right\rangle)$ contributions are clearly separated.
As usual, gravity plays simply an auxiliary role, since one takes the flat limit in all the hierarchical
Ward identities which are obtained from (\ref{Ch2TraceAnomaly1}) after the functional differentiations.
The goal of this chapter is to stress on some specific features of the dilaton effective action which follow up from (\ref{Ch2TraceAnomaly1})
and which are related to the structure of the anomalous contributions. In particular, in a nearly conformal phase of the Standard Model,
which can be approximated by an exact $SU(3)_C \times SU(2)_L \times U(1)_Y$ gauge theory,
cubic and quartic contributions to the dilaton dynamics are essentially fixed by the anomaly and can be extracted, with some effort,
from a diagrammatic analysis of (\ref{Ch2TraceAnomaly1}) expanded up to the fourth order in the metric.
This is the approach that we will be following in our case and on which we are going to elaborate.
In particular, we will present the expressions of such contributions.
These interactions set a key distinction between a Higgs and a dilaton at every order, being the Higgs not affected by the scale anomaly,
and can provide the basis for a direct phenomenological analysis of possible dilaton interactions at the LHC.
Their derivation will bypass the direct diagrammatic computation, relying instead on the connection between conformal anomalies
and counterterms in dimensional regularization, which was thoroughly discussed in section \ref{CountAnom}.
\section{Anomalous interactions from the Ward identities}
To illustrate the role of the anomaly in a more direct way and its possible significance in setting a distinction between the
Higgs and the dilaton, we recall that the interaction of the dilaton $\rho$ with the Standard Model fields is given to first order by
\begin{eqnarray}
\label{Ch2Lint}
\mathcal L_{int} = - \frac{1}{\Lambda} \rho \, {T^\mu}_\mu \, ,
\end{eqnarray}
where $\Lambda$ is the conformal breaking scale, which remains a free parameter of the effective action.
For convenience, let us recall the ordinary definition of the EMT of the Standard Model
\begin{equation} \label{Ch2EMTCh2}
T^{\mu\nu} = -\frac{2}{\sqrt{g}}\frac{\delta\,\mathcal{S}}{\delta g_{\mu\nu}} \, ,
\end{equation}
in terms of the action $\mathcal{ S}$, so that its quantum average in terms of the euclidean generating functional of the theory,
$\mathcal {W}$, depending from the background metric $g_{\mu\nu}(x)$,
\begin{equation}\label{Ch2GeneratingCh2}
\mathcal W = \frac{1}{\mathcal{N}} \, \int \, \mathcal D\Phi \, e^{-\, \mathcal{S}} \, .
\end{equation}
is given by
\begin{equation} \label{Ch2VEVEMTCh2}
\left\langle T^{\mu\nu} \right\rangle_s = \frac{2}{\sqrt{g}}\frac{\delta\, \mathcal W}{\delta\, g_{\mu\nu}}\, ,
\end{equation}
where the background fields are kept switched on, as the subscript $s$ indicates.
\begin{figure}[t]
\centering
\includegraphics[scale=0.8]{figures/3dil.eps}
\hspace{1cm}\includegraphics[scale=0.8]{figures/4dil.eps}
\caption{Cubic and quartic dilaton interactions. In the nearly conformal limit the computation of the interactions involves virtual
scalars, spin 1 and fermion exchanges. }
\label{Ch2rgg1}
\end{figure}
The identification of the anomaly contributions to dilaton interactions, on general grounds, requires an analysis of the anomalous Ward identities
satisfied by the respective correlators. In this chapter we will concentrate on the extraction of the anomalous contribution to the quartic dilaton
interactions, using as a fundamental scenario the Standard Model in the unbroken phase (i.e. with $v\to 0$).
In this approximation, explicit trace insertions vanish for on-shell massless final states (i.e. for gauge fields)
and we neglect all the contributions related both to the Higgs and to virtual corrections with a
massive dilaton in the loops. In the same limit all the mixing contributions related to a possible term of improvement are not present
\cite{Coriano:2012nm} and the computation of the anomalous terms amounts to the extraction of some finite parts.
Explicit (i.e. non anomalous) corrections, also present in the fundamental Ward identity, are calculable, but they are model-dependent.
In fact, they require the introduction of some extra potential for the dilaton/Higgs system.
Its form necessarily has to rely on extra assumptions, such as the specific choice of breaking of the conformal symmetry,
as illustrated in chapter \ref{Effective}.
They are obtained by inserting the trace (${T^\mu}_\mu$) operator on generic correlators involving all the fields of the
Standard Model (plus other dilaton lines). In the conformal limit such contributions, which have been discussed in \cite{Coriano:2012nm}
in the case of a single dilaton, drop out and the computations simplify considerably.
In this approximation the breaking of the dilatation symmetry does not contain any explicit scale-dependent term, and it is only due to
the anomaly, which is induced by renormalization. We call this approximation "nearly conformal".
As we have pointed out before, the breaking of the dilatation symmetry shows up, at a perturbative level,
with the appearance of a massless pole in the $J_D VV$ correlator in the neutral and charged current sectors of the theory, with a residue which is
proportional to a specific beta-function of the theory, related to the final state. This takes the role of a Nambu-Goldstone mode of the broken
dilatation symmetry and it has been shown to affect each gauge invariant sector of the dilaton-to-two gauge bosons matrix elements.
To extract the anomalous contributions of the higher order interactions shown in fig. \ref{Ch2rgg1}
we start from the explicit expression of the anomaly, which in $d=4$ is given by \cite{Duff:1977ay,Duff:1993wm}
\begin{eqnarray} \label{Ch2TraceAnomalyCh2}
\mathcal A[g]
&=&
\sum_{I=f,s,V}n_I \bigg[\beta_a(I)\, F + \beta_b(I)\, G + \beta_c(I) \, \square R \bigg] \, .
\end{eqnarray}
$\mathcal{A}[g]$ contains the diffeomorphism-invariants built out of the Riemann tensor, ${R^{\alpha}}_{\beta\gamma\delta}$,
the Ricci tensor $R_{\alpha\beta}$ and the scalar curvature $R$. In particular, $G$ and $F$ in eq. (\ref{Ch2TraceAnomalyCh2})
are the Euler density and the square of the Weyl tensor respectively, given in appendix \ref{Geometrical}, with coefficients $\beta_a$
and $\beta_b$ depending on the field content of the Lagrangian (scalar, fermion,vector) and we have a multiplicity factor $n_I$
for each particle species. The form of the trace anomaly in (\ref{Ch2TraceAnomalyCh2}) already contains the
constraint $\beta_d = 0 $ discussed in chapter \ref{TTTVertex} and we are discarding the gauge sector,
which is not concerned for the purposes of this chapter.
In addition, the value of $\beta_c$ is regularization-dependent, corresponding to
the fact that it can be changed by the addition of an arbitrary local $R^2$ term in the effective action.
The values given in table \ref{Ch1AnomalyCoeff} are those obtained in dimensional regularization,
in which the relation $\beta_c = -{2}/{3}\,\beta_a$ is found to hold.
The values of the coefficients for three theories of spin $0, \frac{1}{2}, 1$, that we are going to consider
are listed in chapter \ref{TTTVertex}, table \ref{Ch1AnomalyCoeff}.
Taking (\ref{Ch2VEVEMTCh2}) into account, in terms of $\mathcal{W}$ the fundamental trace anomaly
equation can be rewritten in the form
\begin{equation}
\label{Ch2TraceAnomalySymmCh2}
2 \, g_{\mu\nu}\frac{\delta \mathcal{W}}{\delta g_{\mu\nu}}= \sqrt{g} \, \mathcal A[g]
\end{equation}
and plays the role of a generating functional for the anomalous Ward identities of any underlying
Lagrangian field theory being, therefore, model independent.
From (\ref{Ch2TraceAnomalySymmCh2}) we can extract several identities satisfied by the anomaly term, for correlators involving $n$
insertions of energy-momentum tensors, by performing $n-1$ functional derivatives with respect to the metric of both sides of
(\ref{Ch2TraceAnomalySymmCh2}) and taking the trace of the result at the very end.
\section{EMT's and Correlators}
In perturbation theory, imposing the conservation Ward identity for the EMT and the Ward identities for the vector currents - whenever these are
present - is sufficient to obtain the corresponding anomalous term from the complete diagrammatic expansion.
In particular, in dimensional regularization, the anomaly comes for free at the end of the computations, but this is a demanding job.
In the case of the $TVV$,
for instance, it is a common practice to perform a direct computation, since only one term ($\sim F^{a\,\mu\nu}(z)\,F^a_{\mu\nu} (z)$)
can appear in the anomaly. We have omitted it in (\ref{Ch2TraceAnomalyCh2}), since our analysis is focused on the anomaly-induced
radiative corrections to correlators involving only dilaton self-interactions. A general discussion of contributions containing neutral currents
(the $TVV$ vertex) has been given in \cite{Coriano:2011zk, Coriano:2012nm} in the Standard Model and in
\cite{Giannotti:2008cv, Armillis:2009pq} for QED. However, things are far more involved for vertices containing multiple insertions of EMT's,
such as the $TTT$ and $TTTT$, and it is convenient to infer the structure of the anomalous corrections without having to perform a complete
diagrammatic analysis. In any case, a successful test of the anomalous Ward identities is crucial in order to secure the correctness of the result of
the computation.
As mentioned above, in the nearly conformal limit of the Standard Model we will need to consider a scalar, a fermion and an abelian vector theory
coupled to a background gravitational field. In fact the non abelian character of the theory is not essential in the study of the higher order terms
to the dilaton effective action. In this case we can reconstruct the entire contribution to the anomaly from the abelian case
by correcting the result just by one extra multiplicity factor.
We recall here the EMT's for the theories that we consider, i.e.
\begin{eqnarray}
T^{\mu\nu}_{\phi}
&=&
\nabla^\mu \phi \, \nabla^\nu\phi - \frac{1}{2} \, g^{\mu\nu}\,g^{\alpha\beta}\,\nabla_\alpha \phi \, \nabla_\beta \phi
+ \chi \bigg[g^{\mu\nu} \Box - \nabla^\mu\,\nabla^\nu - \frac{1}{2}\,g^{\mu\nu}\,R + R^{\mu\nu} \bigg]\, \phi^2
\label{Ch2ScalarEMT}\\
T^{\mu\nu}_{\psi}
&=&
\frac{1}{4} \,
\bigg[ g^{\mu\lambda}\,{V_\alpha}^\nu + g^{\nu\lambda}\,{V_\alpha}^\mu - 2\,g^{\mu\nu}\,{V_\alpha}^\lambda \bigg]
\bigg[\bar{\psi} \, \gamma^a \, \left(\mathcal{D}_\lambda \,\psi\right) -
\left(\mathcal{D}_\lambda \, \bar{\psi}\right) \, \gamma^a \, \psi \bigg] \, ,
\label{Ch2FermionEMT} \\
T^{\mu\nu}_V
&=&
F^{\mu\a}{F^\nu}_{\a} - \frac{1}{4}g^{\mu\nu}F^{\a\b}F_{\a\b} \, ,
\label{Ch2MaxwellEMT}
\end{eqnarray}
where ${V_\alpha}^\nu$ is the \emph{vierbein} needed to embed the fermion in the gravitational background
and the corresponding covariant derivative is $ \mathcal{D}_\mu = \partial_\mu + \Gamma_\mu =
\partial_\mu + \frac{1}{2} \, \Sigma^{\alpha\beta} \, {V_\alpha}^\sigma \, \nabla_\mu\,V_{\beta\sigma} \, $
where the $\Sigma^{\alpha\beta}$ are the generators of the Lorentz group in the case of a spin $1/2$-field. \\
It is convenient to define the correlation functions with $n$ external insertions of EMT's, which can be effectively thought as
gravitons, as functional derivatives of order $n$ of $\mathcal W$, evaluated in the flat limit
\begin{eqnarray} \label{Ch2NPFCh2}
\left\langle T^{\mu_1\nu_1}(x_1)...T^{\mu_n\nu_n}(x_n) \right\rangle
&=&
\bigg[\frac{2}{\sqrt{g_{x_1}}}...\frac{2}{\sqrt{g_{x_n}}} \,
\frac{\delta^n \mathcal{W}}{\delta g_{\mu_1\nu_1}(x_1)...\delta g_{\mu_n\nu_n}(x_n)}\bigg]
\bigg|_{g_{\mu\nu} = \delta_{\mu\nu}} \nonumber \\
&=&
2^n\, \frac{\delta^n \mathcal{W}}{\delta g_{\mu_1\nu_1}(x_1)...\delta g_{\mu_n\nu_n}(x_n)}\bigg|_{g_{\mu\nu} =
\delta_{\mu\nu}} \, .
\end{eqnarray}
for any functional (or function) $\mathcal F$ which depends on the background field $g_{\mu\nu}(x)$.
Denoting with
\begin{equation}
\left\langle \mathcal{O} \right\rangle = \int\, \mathcal{D}\Phi\, \mathcal{O}\, e^{-S} \,
\end{equation}
the vacuum expectation values of each operator, with $\mathcal{S}$ the generic action, we obtain
\begin{eqnarray}
\left\langle T^{\mu\nu}(x)\right\rangle
&=&
- 2 \, \left\langle \left[\mathcal S\right]^{\mu\nu}(x) \right\rangle \label{Ch21PF} \\
\left\langle T^{\mu_{1}\nu_{1}}(x_1)T^{\mu_{2}\nu_{2}}(x_2)\right\rangle
&=&
4 \, \bigg[
\left\langle \left[\mathcal S\right]^{\mu_{1}\nu_{1}}(x_{1})\, \left[\mathcal S\right]^{\mu_{2}\nu_{2}}(x_{2}) \right\rangle
- \, \left\langle \left[\mathcal S\right]^{\mu_{1}\nu_{1}\mu_{2}\nu_{2}}(x_{1},x_{2})\right\rangle \bigg] \, , \label{Ch22PFCh2}\\
\left\langle T^{\mu_{1}\nu_{1}}(x_1)T^{\mu_{2}\nu_{2}}(x_2)T^{\mu_{3}\nu_{3}}(x_3)\right\rangle
&=&
8 \, \bigg[ - \left\langle \left[\mathcal S \right]^{\mu_{1}\nu_{1}}(x_{1})\left[\mathcal S \right]^{\mu_{2}\nu_{2}}(x_{2})
\left[\mathcal S \right]^{\mu_{3}\nu_{3}}(x_{3}) \right\rangle
\nonumber \\
&& \hspace{-50mm}
+ \,\bigg( \left\langle \left[\mathcal S\right]^{\mu_{1}\nu_{1}\mu_{2}\nu_{2}}(x_{1},x_{2})\,\left[\mathcal S\right]^{\mu_{3}\nu_{3}}(x_{3}) \right\rangle
+ 2\, \text{perm.} \bigg)
- \,\left\langle \left[\mathcal S\right]^{\mu_{1}\nu_{1}\mu_{2}\nu_{2}\mu_{3}\nu_{3}}(x_{1},x_{2},x_{3})\right\rangle \bigg] \label{Ch23PFCh2} \, ,
\end{eqnarray}
\begin{eqnarray}
\left\langle T^{\mu_{1}\nu_{1}}(x_1)T^{\mu_{2}\nu_{2}}(x_2)T^{\mu_{3}\nu_{3}}(x_3)T^{\mu_{4}\nu_{4}}(x_4)\right\rangle
&=&
16 \, \bigg[
\left\langle \left[\mathcal S \right]^{\mu_{1}\nu_{1}}(x_{1})\left[\mathcal S \right]^{\mu_{2}\nu_{2}}(x_{2})
\left[\mathcal S \right]^{\mu_{3}\nu_{3}}(x_{3})\left[\mathcal S \right]^{\mu_{4}\nu_{4}}(x_{4}) \right\rangle
\nonumber \\
&& \hspace{-80 mm}
- \,\bigg( \left\langle \left[\mathcal S\right]^{\mu_{1}\nu_{1}\mu_{2}\nu_{2}}(x_{1},x_{2})\left[\mathcal S\right]^{\mu_{3}\nu_{3}}(x_{3})
\left[\mathcal S\right]^{\mu_{4}\nu_{4}}(x_{4}) \right\rangle + 5 \, \text{perm.} \bigg)
+ \bigg( \left\langle \left[\mathcal S\right]^{\mu_{1}\nu_{1}\mu_{2}\nu_{2}}(x_{1},x_{2})\left[\mathcal
S\right]^{\mu_{3}\nu_{3}\mu_{4}\nu_{4}}(x_{3},x_{4})\right\rangle + 2 \, \text{perm.} \bigg)
\nonumber \\
&& \hspace{-80mm}
+ \,\bigg( \left\langle \left[\mathcal S\right]^{\mu_{1}\nu_{1}\mu_{2}\nu_{2}\mu_{3}\nu_{3}}(x_{1},x_{2},x_{3})\left[\mathcal
S\right]^{\mu_{4}\nu_{4}}(x_{4})\right\rangle + 3 \, \text{perm.} \bigg)
- \, \left\langle \left[\mathcal S\right]^{\mu_{1}\nu_{1}\mu_{2}\nu_{2}\mu_{3}\nu_{3}\mu_{4}\nu_{4}}(x_{1},x_{2},x_{3},x_{4}) \right\rangle \bigg] \, .
\label{Ch24PF}
\end{eqnarray}
Notice that in dimensional regularization
\begin{equation}
\left\langle \left[\mathcal S\right]^{\mu\nu}(x) \right\rangle= \left\langle \left[\mathcal
S\right]^{\mu_{1}\nu_{1}\mu_{2}\nu_{2}}(x_{1},x_{2})\right\rangle
=\left\langle \left[\mathcal S\right]^{\mu_{1}\nu_{1}\mu_{2}\nu_{2}\mu_{3}\nu_{3}}(x_{1},x_{2},x_{3})\right\rangle
= \left\langle \left[\mathcal S\right]^{\mu_{1}\nu_{1}\mu_{2}\nu_{2}\mu_{3}\nu_{3}\mu_{4}\nu_{4}}(x_{1},x_{2},x_{3},x_{4}) \right\rangle=0
\end{equation}
being proportional to massless tadpoles.
In particular, this implies that, to perform a perturbative computation of a correlator of order $n$,
one would be needing interaction vertices with at most $n-1$ gravitons.
Concerning the diagrammatic structure of each contribution, the correlator
\begin{equation}
\left\langle \left[\mathcal S \right]^{\mu_{1}\nu_{1}}(x_{1})\left[\mathcal S \right]^{\mu_{2}\nu_{2}}(x_{2})
\left[\mathcal S \right]^{\mu_{3}\nu_{3}}(x_{3})\left[\mathcal S \right]^{\mu_{4}\nu_{4}}(x_{4}) \right\rangle
\end{equation}
has a box topology;
\begin{equation}
\left\langle \left[\mathcal S \right]^{\mu_{1}\nu_{1}}(x_{1})\left[\mathcal S \right]^{\mu_{2}\nu_{2}}(x_{2})
\left[\mathcal S \right]^{\mu_{3}\nu_{3}}(x_{3}) \right\rangle
\end{equation}
which is the first contribution to the graviton $3$-point function, and
\begin{equation}
\left\langle \left[\mathcal S\right]^{\mu_{1}\nu_{1}\mu_{2}\nu_{2}}(x_{1},x_{2})
\left[\mathcal S\right]^{\mu_{3}\nu_{3}}(x_{3})\left[\mathcal S\right]^{\mu_{4}\nu_{4}}(x_{4}) \right\rangle \, ,
\end{equation}
which corresponds to a contact term in $\left\langle TTTT \right\rangle$, are represented by triangles. \\
The remaining contributions,
\begin{equation}
\left\langle \left[\mathcal S\right]^{\mu_{1}\nu_{1}}(x_{1})\, \left[\mathcal S\right]^{\mu_{2}\nu_{2}}(x_{2}) \right\rangle \, ,
\end{equation}
the contact terms $\left\langle TTT \right\rangle$, which are
\begin{equation}
\left\langle \left[\mathcal S\right]^{\mu_{1}\nu_{1}\mu_{2}\nu_{2}}(x_{1},x_{2})\,\left[\mathcal S\right]^{\mu_{3}\nu_{3}}(x_{3}) \right\rangle
\end{equation}
and the two remaining types of diagrams which enter into $\left\langle TTTT \right\rangle$,
\begin{equation}
\left\langle \left[\mathcal S\right]^{\mu_{1}\nu_{1}\mu_{2}\nu_{2}}(x_{1},x_{2})\left[\mathcal S\right]^{\mu_{3}\nu_{3}\mu_{4}\nu_{4}}(x_{3},x_{4})\right\rangle
\end{equation}
and
\begin{equation}
\left\langle \left[\mathcal S\right]^{\mu_{1}\nu_{1}\mu_{2}\nu_{2}\mu_{3}\nu_{3}}(x_{1},x_{2},x_{3})\left[\mathcal S\right]^{\mu_{4}\nu_{4}}(x_{4})\right\rangle \, ,
\end{equation}
have the topologies of 2-point functions.
Our conventions for the choice of the external momenta, which are taken to be all incoming, are defined via the Fourier transform
\begin{eqnarray}
&&
\int \, d^4x_{1}\,d^4x_{2}\,d^4x_{3}\,d^4x_{4}\,
\left\langle T^{\mu_{1}\nu_{1}}(x_{1})T^{\mu_{2}\nu_{2}}(x_{2})T^{\mu_{3}\nu_{3}}(x_{3})T^{\mu_{4}\nu_{4}}(x_{4})\right\rangle \,
e^{-i(k_{1}\cdot x_{1} + k_{2}\cdot x_{2} + k_{3}\cdot x_{3} + k_{4} \cdot x_{4})} =
\nonumber \\
&& \hspace{40mm}
(2\pi)^4\,\delta^{(4)}(k_{1} + k_{2} + k_{3} + k_{4})\,
\left\langle T^{\mu_{1}\nu_{1}}T^{\mu_{2}\nu_{2}}T^{\mu_{3}\nu_{3}}T^{\mu_{4}\nu_{4}}\right\rangle(k_{2},k_{3},k_{4}) \, ,
\label{Ch24PFMom}
\end{eqnarray}
and similar for the $3$- and $2$-point functions.
\section{Ward identities}
We start from the analysis of the general covariance Ward identities, which partially overlaps with the
discussion in chapter \ref{TTTVertex}, which is limited to the $TTT$ vertex. \\
The identities we look for are obtained from the functional relation
\begin{equation}\label{Ch2masterWI0Ch2}
\nabla_{\nu_{1}} \left\langle T^{\mu_{1}\nu_{1}}(x_1) \right\rangle
= \nabla_{\nu_{1}} \bigg(\frac{2}{\sqrt{g_{x_1}}}\frac{\delta\mathcal W}{\delta g_{\mu_{1}\nu_{1}}(x_1)}\bigg)= 0 \,
\end{equation}
which, after an expansion, becomes
\begin{eqnarray} \label{Ch2InterWard}
&&
\frac{2}{\sqrt{g_{x_1}}}\bigg(\partial_{\nu_{1}}\frac{\delta\mathcal W}{\delta g_{\mu_{1}\nu_{1}}(x_1)}
- \Gamma^\lambda_{\lambda\nu_{1}}(x_1)\frac{\delta\mathcal W}{\delta g_{\mu_{1}\nu_{1}}(x_1)}
+ \Gamma^{\mu_{1}}_{\kappa\nu_{1}}(x_1)\frac{\delta\mathcal W}{\delta g_{\kappa\nu_{1}}(x_1)}
+ \Gamma^{\nu_{1}}_{\kappa\nu_{1}}(x_1)\frac{\delta\mathcal W}{\delta g_{\mu_{1}\kappa}(x_1)}\bigg) = 0. \nonumber\\
\end{eqnarray}
Cancelling the second and fourth terms in parentheses, (\ref{Ch2InterWard}) takes the simpler form
\begin{eqnarray} \label{Ch2masterWICh2}
&&
2 \, \bigg(\partial_{\nu_{1}}\frac{\delta\mathcal W}{\delta g_{\mu_{1}\nu_{1}}(x_1)}
+ \Gamma^{\mu_{1}}_{\kappa\nu_{1}}(x_1)\frac{\delta\mathcal W}{\delta g_{\kappa\nu_{1}}(x_1)}\bigg)
= 0\, .
\end{eqnarray}
The Ward identities we are interested in are obtained by functional differentiation of (\ref{Ch2masterWICh2}) and give
\begin{eqnarray}
&&
4\, \bigg[ \partial_{\nu_{1}} \frac{\delta^2\mathcal W}{\delta g_{\mu_{2}\nu_{2}}(x_2)\delta g_{\mu_{1}\nu_{1}}(x_1)}
+ \frac{\delta \Gamma^{\mu_{1}}_{\kappa\nu_{1}}(x_1)}{\delta g_{\mu_{2}\nu_{2}}(x_2)}
\frac{\delta\mathcal W}{\delta g_{\kappa\nu_{1}}(x_1)}
+ \Gamma^{\mu_{1}}_{\kappa\nu_{1}}(x_1)
\frac{\delta^2 \mathcal W}{\delta g_{\mu_{2}\nu_{2}}(x_2)\delta g_{\kappa\nu_{1}}(x_1)}\bigg] = 0 \, ,
\label{Ch2WI2PFCoordinateCh2}
\end{eqnarray}
for single
\begin{eqnarray}
&&
8\, \bigg[\partial_{\nu_{1}}
\frac{\delta^3\mathcal W}{\delta g_{\mu_{3}\nu_{3}}(x_3)\delta g_{\mu_{2}\nu_{2}}(x_2)\delta g_{\mu_{1}\nu_{1}}(x_1)}
+ \frac{\delta \Gamma^{\mu_{1}}_{\kappa\nu_{1}}(x_1)}{\delta g_{\mu_{2}\nu_{2}}(x_2)}
\frac{\delta^2 \mathcal W}{\delta g_{\mu_{3}\nu_{3}}(x_3)\delta g_{\kappa\nu_{1}}(x_1)}
+ \frac{\delta \Gamma^{\mu_{1}}_{\kappa\nu_{1}}(x_1)}{\delta g_{\mu_{3}\nu_{3}}(x_3)}
\frac{\delta^2 \mathcal W}{\delta g_{\mu_{2}\nu_{2}}(x_2)\delta g_{\kappa\nu_{1}}(x_3)} \nonumber\\
&&
+ \frac{\delta^2\Gamma^{\mu_{1}}_{\kappa\nu_{1}}(x_1)}{\delta g_{\mu_{2}\nu_{2}}(x_2)\delta g_{\mu_{3}\nu_{3}}(x_3)}
\frac{\delta\mathcal W}{\delta g_{\kappa\nu_{1}}(x_1)}
+ \Gamma^{\mu_{1}}_{\kappa\nu_{1}}(x_1)
\frac{\delta^3\mathcal W}{\delta g_{\mu_{2}\nu_{2}}(x_2)\delta g_{\mu_{3}\nu_{3}}(x_2)\delta g_{\kappa\nu_{1}}(x_1)}\bigg] = 0 \, ,
\label{Ch2WI3PFCoordinate}
\end{eqnarray}
double, and
\begin{eqnarray}
&&
16\, \bigg[\partial_{\nu_{1}}\frac{\delta^4\mathcal W}
{\delta g_{\mu_{4}\nu_{4}}(x_4)\delta g_{\mu_{3}\nu_{3}}(x_3)\delta g_{\mu_{2}\nu_{2}}(x_2)\delta g_{\mu_{1}\nu_{1}}(x_1)}
+ \bigg(\frac{\delta \Gamma^{\mu_{1}}_{\kappa\nu_{1}}(x_1)}{\delta g_{\mu_{2}\nu_{2}}(x_2)}
\frac{\delta^3 \mathcal W}{\delta g_{\mu_{4}\nu_{4}}(x_4) \delta g_{\mu_{3}\nu_{3}}(x_3)\delta g_{\kappa\nu_{1}}(x_1)}
\nonumber \\
&&
+ \big( 2 \leftrightarrow 4, 2 \leftrightarrow 3\big) \bigg)
+ \bigg( \frac{\delta^2 \Gamma^{\mu_{1}}_{\kappa\nu_{1}}(x_1)}{\delta g_{\mu_{3}\nu_{3}}(x_3) \delta g_{\mu_{2}\nu_{2}}(x_2)}
\frac{\delta^2 \mathcal W}{\delta g_{\mu_{4}\nu_{4}}(x_4)\delta g_{\kappa\nu_{1}}(x_1)} +
\big( 2 \leftrightarrow 4, 3 \leftrightarrow 4 \big) \bigg)
\nonumber \\
&&
+ \frac{\delta^3 \Gamma^{\mu_{1}}_{\kappa\nu_{1}}(x_1)}
{\delta g_{\mu_{4}\nu_{4}}(x_4) \delta g_{\mu_{3}\nu_{3}}(x_3) \delta g_{\mu_{2}\nu_{2}}(x_2)}
\frac{\delta \mathcal W}{\delta g_{\kappa\nu_{1}}(x_1)}
+ \Gamma^{\mu_{1}}_{\kappa\nu_{1}}(x_1) \frac{\delta^4 \mathcal W}
{\delta g_{\mu_{4}\nu_{4}}(x_4)\delta g_{\mu_{3}\nu_{3}}(x_3)\delta g_{\mu_{2}\nu_{2}}(x_2)\delta g_{\kappa\nu_{1}}(x_1)}=0
\label{Ch2WI4PFCoordinate}
\end{eqnarray}
triple differentiations of the master equation (\ref{Ch2masterWI0Ch2}).
In the Ward identity satisfied by the $4$-point function we have left implicit the contributions obtained by permuting the action of
the functional derivatives.
To move to the flat spacetime limit of (\ref{Ch2WI2PFCoordinateCh2}) and (\ref{Ch2WI3PFCoordinate}),
we use the notations in (\ref{Ch1Flat}) and set to zero the contributions from the massless tadpoles, obtaining
\begin{eqnarray}
\partial_{\nu_{1}} \left\langle T^{\mu_{1}\nu_{1}}(x_1) T^{\mu_{2}\nu_{2}}(x_2) \right\rangle &=& 0 \, , \label{Ch2WI2PFCoordinateFlat} \\
\partial_{\nu_{1}} \left\langle T^{\mu_{1}\nu_{1}}(x_1)T^{\mu_{2}\nu_{2}}(x_2)T^{\mu_{3}\nu_{3}}(x_3) \right\rangle
&=&
- 2\, \left[\Gamma^{\mu_{1}}_{\kappa\nu_{1}}(x_1)\right]^{\mu_{2}\nu_{2}}(x_2)
\left\langle T^{\kappa\nu_{1}}(x_1)T^{\mu_{3}\nu_{3}}(x_3)\right\rangle
\nonumber \\
&&
- 2\, \left[\Gamma^{\mu_{1}}_{\kappa\nu_{1}}(x_1)\right]^{\mu_{3}\nu_{3}}(x_3)
\left\langle T^{\kappa\nu_{1}}(x_1)T^{\mu_{3}\nu_{3}}(x_3)\right\rangle \, ,
\label{Ch2WI3PFCoordinateFlat} \\
\partial_{\nu_{1}} \left\langle T^{\kappa\nu_{1}}(x_1)T^{\mu_{2}\nu_{2}}(x_2)T^{\mu_{3}\nu_{3}}(x_3)T^{\mu_{4}\nu_{4}}(x_4) \right\rangle
&=&
- 2\, \bigg(\left[\Gamma^{\mu_{1}}_{\kappa\nu_{1}}(x_1)\right]^{\mu_{2}\nu_{2}}(x_2)
\left\langle T^{\kappa\nu_{1}}(x_1) T^{\mu_{3}\nu_{3}}(x_3)T^{\mu_{4}\nu_{4}}(x_4) \right\rangle
\nonumber \\
&& \hspace{-40mm}
+ \big( 2 \leftrightarrow 3, 2 \leftrightarrow 4 \big) \bigg)
- 4\, \bigg( \left[\Gamma^{\mu_{1}}_{\kappa\nu_{1}}(x_1)\right]^{\mu_{2}\nu_{2}\mu_{3}\nu_{3}}(x_2,x_3)
\left\langle T^{\kappa\nu_{1}}(x_1)T^{\mu_{4}\nu_{4}}(x_4) \right\rangle
+ \big( 2 \leftrightarrow 4, 3 \leftrightarrow 4 \big) \bigg) \, , \label{Ch2WI4PFCoordinateFlat}
\nonumber \\
\end{eqnarray}
which after some manipulations give the transversality constraint for the $2$-point functions and
\begin{eqnarray}
k_{1\,\nu_{1}} \left\langle T^{\mu_{1}\nu_{1}}T^{\mu_{3}\nu_{3}}T^{\mu_{2}\nu_{2}} \right\rangle(k_{2},k_{3})
&=&
- k_{3}^{\mu_{1}} \left\langle T^{\mu_{3}\nu_{3}}T^{\mu_{2}\nu_{2}}\right\rangle(k_{2})
- k_{2}^{\mu_{1}} \left\langle T^{\mu_{2}\nu_{2}}T^{\mu_{3}\nu_{3}}\right\rangle(k_{3}) \nonumber \\
&+&
k_{3\, \nu_{1}} \bigg[\delta^{\mu_{1}\nu_{3}} \left\langle T^{\nu_{1}\mu_{3}}T^{\mu_{2}\nu_{2}} \right\rangle(k_{2})
+ \delta^{\mu_{1}\mu_{3}}\left\langle T^{\nu_{1}\nu_{3}}T^{\mu_{2}\nu_{2}} \right\rangle(k_{2})\bigg]
\nonumber \\
&+&
k_{2\, \nu_{1}} \bigg[\delta^{\mu_{1}\nu_{2}}\left\langle T^{\nu_{1}\mu_{2}}T^{\mu_{3}\nu_{3}} \right\rangle(k_{3})
+ \delta^{\mu_{1}\mu_{2}} \left\langle T^{\nu_{1}\nu_{2}}T^{\mu_{3}\nu_{3}}\right\rangle(k_{3})\bigg] \, .
\nonumber \\
\label{Ch2WI3PFCh2}
\end{eqnarray}
Similarly, in the case of the $4$-point function $TTTT$, using (\ref{Ch24PFMom}) we obtain
\begin{eqnarray}
&&
k_{1\,\nu_{1}} \, \left\langle T^{\mu_{1}\nu_{1}}T^{\mu_{3}\nu_{3}}T^{\mu_{2}\nu_{2}}T^{\mu_{4}\nu_{4}} \right\rangle(k_{2},k_{3},k_{4}) =
\bigg[ - k_{2}^{\mu_{1}} \, \left\langle T^{\mu_{2}\nu_{2}}T^{\mu_{3}\nu_{3}}T^{\mu_{4}\nu_{4}}\right\rangle(k_{3},k_{4})
\nonumber \\
&+&
k_{2\,\nu_{1}}\, \bigg( \delta^{\mu_{1}\nu_{2}}\, \left\langle T^{\nu_{1}\mu_{2}}T^{\mu_{3}\nu_{3}}T^{\mu_{4}\nu_{4}}\right\rangle(k_{3},k_{4})
+ \delta^{\mu_{1}\mu_{2}}\, \left\langle T^{\nu_{1}\nu_{2}}T^{\mu_{3}\nu_{3}}T^{\mu_{4}\nu_{4}}\right\rangle(k_{3},k_{4}) \bigg)
+ \big( 2 \leftrightarrow 3, 2 \leftrightarrow 4\big) \bigg]
\nonumber \\
&+&
\bigg[ 2\, k_{2\,\nu_{1}}\, \bigg(
\left[g^{\mu_{1}\mu_{2}}\right]^{\mu_{3}\nu_{3}}\, \left\langle T^{\nu_{1}\nu_{2}} T^{\mu_{4}\nu_{4}}\right\rangle(k_{4}) +
\left[g^{\mu_{1}\nu_{2}}\right]^{\mu_{3}\nu_{3}}\, \left\langle T^{\nu_{1}\mu_{2}}T^{\mu_{4}\nu_{4}} \right\rangle(k_{4}) \bigg)
\nonumber \\
&+& \hspace{2mm}
2\, k_{3\,\nu_{1}}\, \bigg(
\left[g^{\mu_{1}\mu_{3}}\right]^{\mu_{2}\nu_{2}}\, \left\langle T^{\nu_{1}\nu_{3}} T^{\mu_{4}\nu_{4}} \right\rangle(k_{4}) +
\left[g^{\mu_{1}\nu_{3}}\right]^{\mu_{2}\nu_{2}}\, \left\langle T^{\nu_{1}\mu_{3}} T^{\mu_{4}\nu_{4}} \right\rangle(k_{4}) \bigg)
\nonumber \\
&+&
\bigg( k_2^{\nu_{3}} \delta^{\mu_{1}\mu_{3}} + k_2^{\mu_{3}} \delta^{\mu_{1}\nu_{3}}\bigg)\,
\left\langle T^{\mu_{2}\nu_{2}}T^{\mu_{3}\nu_{3}}\right\rangle(k_{4})
+ \bigg( k_3^{\nu_{2}} \delta^{\mu_{1}\mu_{2}} + k_3^{\mu_{2}} \delta^{\mu_{1}\nu_{2}}\bigg)\,
\left\langle T^{\mu_{3}\nu_{3}}T^{\mu_{2}\nu_{2}}\right\rangle(k_{4}) + \big( 2 \leftrightarrow 4, 3 \leftrightarrow 4 \big) \bigg] \, .
\nonumber \\
\label{Ch2WI4PF}
\end{eqnarray}
Similar identities are obtained for the momenta of the other external gravitons.
\section{Counterterms}
Coming to a discussion of the counterterms to the $4$-dilaton amplitude, these are obtained from the $1$-loop Lagrangian
which accounts for the gravitational counterterms to pure graviton amplitudes in the $\overline{MS}$ scheme,
\begin{equation}\label{Ch2CounterActionCh2}
S_{counter} = - \frac{\mu^{-\epsilon}}{\bar\epsilon}
\sum_{I=s,f,V}n_I \int d^d x \sqrt{g}\, \bigg( \beta_a(I) \, F + \beta_b(I) \, G \bigg)\, .
\end{equation}
Again, the dimensional parameter is $\epsilon= 4 - d$.
In the case of the $4$-graviton vertex the counterterm action (\ref{Ch2CounterActionCh2}) generates the vertex
\begin{equation}
-\frac{\mu^{-\epsilon}}{\bar\epsilon}\, \bigg(
\beta_a\, D_F^{\mu_{1}\nu_{1}\mu_{2}\nu_{2}\mu_{3}\nu_{3}\mu_{4}\nu_{4}}(x_1,x_2,x_3,x_4) +
\beta_b\, D_G^{\mu_{1}\nu_{1}\mu_{2}\nu_{2}\mu_{3}\nu_{3}\mu_{4}\nu_{4}}(x_1,x_2,x_3,x_4)\bigg)\, ,
\end{equation}
where
\begin{eqnarray}
D_F^{\mu_{1}\nu_{1}\mu_{2}\nu_{2}\mu_{3}\nu_{3}\mu_{4}\nu_{4}}(x_1,x_2,x_3,x_4)
&=&
2^4 \, \frac{\delta^4}
{\delta g_{\mu_{1}\nu_{1}}(x_1)\delta g_{\mu_{2}\nu_{2}}(x_2)\delta g_{\mu_{3}\nu_{3}}(x_3)\delta g_{\mu_{4}\nu_{4}}(x_4)}
\int\,d^d w\,\sqrt{g}\, F\, ,
\label{Ch2DFCh2}\\
D_G^{\mu_{1}\nu_{1}\mu_{2}\nu_{2}\mu_{3}\nu_{3}\mu_{4}\nu_{4}}(x_1,x_2,_3,x_4)
&=&
2^4 \, \frac{\delta^4}
{\delta g_{\mu_{1}\nu_{1}}(x_1)\delta g_{\mu_{2}\nu_{2}}(x_2)\delta g_{\mu_{3}\nu_{3}}(x_3)\delta g_{\mu_{4}\nu_{4}}(x_4)}
\int\,d^d w\,\sqrt{g}\, G
\label{Ch2DGCh2}\,,
\end{eqnarray}
and similarly for the $2$- and $3$-point correlators.
Using these expressions, the fully renormalized $2$-, $3$- and $4$-point correlators in momentum space can be written down as
\begin{eqnarray}
\left\langle T^{\mu_{1}\nu_{1}}T^{\mu_{2}\nu_{2}} \right\rangle_{ren}(k_{2})
&=&
\left\langle T^{\mu_{1}\nu_{1}}T^{\mu_{2}\nu_{2}} \right\rangle_{bare}(k_{2}) -
\frac{\mu^{-\epsilon}}{\bar\epsilon}\, \beta_a\, D_F^{\mu_{1}\nu_{1}\mu_{2}\nu_{2}}(k_{2}) \, ,
\label{Ch2Ren2PF}
\nonumber \\
\left\langle T^{\mu_{1}\nu_{1}}T^{\mu_{2}\nu_{2}}T^{\mu_{3}\nu_{3}} \right\rangle_{ren}(k_{2},k_{3})
&=&
\left\langle T^{\mu_{1}\nu_{1}}T^{\mu_{2}\nu_{2}}T^{\mu_{3}\nu_{3}} \right\rangle_{bare}(k_{2},k_{3})
\nonumber \\
&-&
\frac{\mu^{-\epsilon}}{\bar\epsilon} \, \bigg(
\beta_a\, D_F^{\mu_{1}\nu_{1}\mu_{2}\nu_{2}\mu_{3}\nu_{3}}(k_{2},k_{3}) + \beta_b\, D_G^{\mu_{1}\nu_{1}\mu_{2}\nu_{2}\mu_{3}\nu_{3}}(k_{2},k_{3}) \bigg)\, ,
\label{Ch2Ren3PFCh2}
\nonumber \\
\left\langle T^{\mu_{1}\nu_{1}}T^{\mu_{2}\nu_{2}}T^{\mu_{3}\nu_{3}}T^{\mu_{4}\nu_{4}} \right\rangle_{ren}(k_{2},k_{3},k_{4})
&=&
\left\langle T^{\mu_{1}\nu_{1}}T^{\mu_{2}\nu_{2}}T^{\mu_{3}\nu_{3}}T^{\mu_{4}\nu_{4}} \right\rangle_{bare}(k_{2},k_{3},k_{4})
\nonumber \\
&-&
\frac{\mu^{-\epsilon}}{\bar\epsilon}\,\bigg(
\beta_a\, D_F^{\mu_{1}\nu_{1}\mu_{2}\nu_{2}\mu_{3}\nu_{3}\mu_{4}\nu_{4}}(k_{2},k_{3},k_{4}) +
\beta_b\, D_G^{\mu_{1}\nu_{1}\mu_{2}\nu_{2}\mu_{3}\nu_{3}\mu_{4}\nu_{4}}(k_{2},k_{3},k_{4}) \bigg)\, .
\nonumber \\
\label{Ch2Ren4PF}
\end{eqnarray}
From these relations and from (\ref{Ch2WI3PFCh2}), (\ref{Ch2WI4PF}) it is clear that counterterms must be related by the same
general covariance Ward identities which relate the bare correlators. One can also separately check these identites for F- and G- counterterms just
by writing them down and equating the coefficients of $\beta_a$ and $\beta_b$.
We omit the explicit forms of the counterterms, which are necessary in order to test all these constraints.
We have checked all of them with a symbolic calculus program.
A second, powerful constraint on the counterterms comes from the anomalous Ward identities for the Green functions at hand,
which are obtained through functional derivation of (\ref{Ch2TraceAnomalySymmCh2}),
passing to the flat space limit and using the definition (\ref{Ch2NPFCh2}).
A direct computation gives the equations
\begin{eqnarray}
\delta_{\mu_{1}\nu_{1}}\, \left\langle T^{\mu_{1}\nu_{1}}T^{\mu_{2}\nu_{2}}\right\rangle(k_{2})
&=&
2 \, \left[\sqrt{g}\, \mathcal{A}\right]^{\mu_{2}\nu_{2}}(k_{2}) = 2 \, \left[ \Box R\right]^{\mu_{2}\nu_{2}}(k_{2})\, , \label{Ch2AWard2PF} \\
\delta_{\mu_{1}\nu_{1}}\, \left\langle T^{\mu_{1}\nu_{1}}T^{\mu_{2}\nu_{2}}T^{\mu_{3}\nu_{3}} \right\rangle(k_{2},k_{3})
&=&
4 \, \left[\sqrt{g}\, \mathcal{A}\right]^{\mu_{2}\nu_{2}\mu_{3}\nu_{3}}(k_{2},k_{3})
- 2 \, \left\langle T^{\mu_{2}\nu_{2}}T^{\mu_{3}\nu_{3}} \right\rangle(k_{2})
- 2 \, \left\langle T^{\mu_{2}\nu_{2}}T^{\mu_{3}\nu_{3}} \right\rangle(k_{3})\nonumber\\
&& \hspace{-40mm}
=\, 4 \, \bigg[ \beta_a\,\bigg(\left[F\right]^{\mu_{3}\nu_{3}\mu_{2}\nu_{2}}(k_{2},k_{3})
- \frac{2}{3} \left[\sqrt{g}\, \Box R\right]^{\mu_{3}\nu_{3}\mu_{2}\nu_{2}}(k_{2},k_{3})\bigg)
+ \beta_b\, \left[G\right]^{\mu_{3}\nu_{3}\mu_{2}\nu_{2}}(k_{2},k_{3}) \bigg]\nonumber\\
&-&
2 \, \left\langle T^{\mu_{2}\nu_{2}}T^{\mu_{3}\nu_{3}} \right\rangle(k_{2})
- 2 \, \left\langle T^{\mu_{2}\nu_{2}}T^{\mu_{3}\nu_{3}} \right\rangle(k_{3}) \, , \label{Ch2AWard3PF} \\
\delta_{\mu_{1}\nu_{1}}\, \left\langle T^{\mu_{1}\nu_{1}}T^{\mu_{2}\nu_{2}}T^{\mu_{3}\nu_{3}}T^{\mu_{4}\nu_{4}} \right\rangle(k_{2},k_{3},k_{4})
&=&
8 \, \left[ \sqrt{g}\, \mathcal{A}\right]^{\mu_{2}\nu_{2}\mu_{3}\nu_{3}\mu_{4}\nu_{4}}(k_{2},k_{3},k_{4})
\nonumber \\
&& \hspace{-60mm}
-\, 2 \, \left\langle T^{\mu_{2}\nu_{2}}T^{\mu_{3}\nu_{3}}T^{\mu_{4}\nu_{4}} \right\rangle(k_{2},k_{3})
- 2 \, \left\langle T^{\mu_{2}\nu_{2}}T^{\mu_{3}\nu_{3}}T^{\mu_{4}\nu_{4}} \right\rangle(k_{2},k_{4})
- 2 \, \left\langle T^{\mu_{2}\nu_{2}}T^{\mu_{3}\nu_{3}}T^{\mu_{4}\nu_{4}} \right\rangle(k_{3},k_{4})
\nonumber\\
&& \hspace{-70mm}
=\, 8 \, \bigg[ \beta_a\, \bigg(\left[\sqrt{g}\,F\right]^{\mu_{2}\nu_{2}\mu_{3}\nu_{3}\mu_{4}\nu_{4}}(k_{2},k_{3},k_{4})
- \frac{2}{3} \left[\sqrt{g}\, \Box R\right]^{\mu_{2}\nu_{2}\mu_{3}\nu_{3}\mu_{4}\nu_{4}}(k_{2},k_{3},k_{4})\bigg)
+ \beta_b\, \left[\sqrt{g}\, G\right]^{\mu_{2}\nu_{2}\mu_{3}\nu_{3}\mu_{4}\nu_{4}}(k_{2},k_{3},k_{4}) \bigg]
\nonumber\\
&& \hspace{-70mm}
- 2 \, \left\langle T^{\mu_{2}\nu_{2}}T^{\mu_{3}\nu_{3}}T^{\mu_{4}\nu_{4}} \right\rangle(k_{2},k_{3})
- 2 \, \left\langle T^{\mu_{2}\nu_{2}}T^{\mu_{3}\nu_{3}}T^{\mu_{4}\nu_{4}} \right\rangle(k_{2},k_{4})
- 2 \, \left\langle T^{\mu_{2}\nu_{2}}T^{\mu_{3}\nu_{3}}T^{\mu_{4}\nu_{4}} \right\rangle(k_{3},k_{4}) \, . \label{Ch2AWard4PF}
\end{eqnarray}
The explicit expressions of the multiple functional derivatives of the various operators in square bracket $([\,\,])$
are very lengthy and we omit them.
At this stage, we can extract from (\ref{Ch2AWard4PF}) four trace identities
(one for each graviton) for the counterterms of the $4$-point functions (\ref{Ch2DFCh2}) and (\ref{Ch2DGCh2}),
relating them to the corresponding $2$- and $3$-point ones.
As these counterterms have been independently tested through general covariance Ward identities, this provides a
useful test of the anomaly contributions to the $4$-point function as well, which is used to deduce the form of the quartic dilaton interactions.
The identity, involving traces in $d$ dimensions, is a direct consequence of eq. (\ref{dTraceCt}), which was extensively discussed,
and its form is
\begin{eqnarray}
\delta^{d}_{\mu_{1}\nu_{1}}\, D_{F}^{\mu_{1}\nu_{1}\mu_{2}\nu_{2}\mu_{3}\nu_{3}\mu_{4}\nu_{4}}(k_{2},k_{3},k_{4})
&=&
- \frac{\epsilon}{2}\, \left(\left[\sqrt{g}\, F\right]^{\mu_{2}\nu_{2}\mu_{3}\nu_{3}\mu_{4}\nu_{4}}(k_{2},k_{3},k_{4})
- \frac{2}{3}\, \left[\sqrt{g}\,\Box R \right]^{\mu_{2}\nu_{2}\mu_{3}\nu_{3}\mu_{4}\nu_{4}}(k_{2},k_{3},k_{4})\right)
\nonumber \\
&-&
2\, D_{F}^{\mu_{2}\nu_{2}\mu_{3}\nu_{3}\mu_{4}\nu_{4}}(k_{3},k_{4}) - 2\, D_{F}^{\mu_{2}\nu_{2}\mu_{3}\nu_{3}\mu_{4}\nu_{4}}(k_{2},k_{4})
- 2\, D_{F}^{\mu_{2}\nu_{2}\mu_{3}\nu_{3}\mu_{4}\nu_{4}}(k_{2},k_{3}) \, ,
\nonumber \\
\delta^{d}_{\mu_{1}\nu_{1}}\, D_{G}^{\mu_{1}\nu_{1}\mu_{2}\nu_{2}\mu_{3}\nu_{3}\mu_{4}\nu_{4}}(k_{2},k_{3},k_{4})
&=&
- \frac{\epsilon}{2}\, \left[\sqrt{g}\, G \right]^{\mu_{2}\nu_{2}\mu_{3}\nu_{3}\mu_{4}\nu_{4}}(k_{2},k_{3},k_{4})
\nonumber\\
&-&
2\, D_{G}^{\mu_{2}\nu_{2}\mu_{3}\nu_{3}\mu_{4}\nu_{4}}(k_{3},k_{4}) - 2\, D_{G}^{\mu_{2}\nu_{2}\mu_{3}\nu_{3}\mu_{4}\nu_{4}}(k_{2},k_{4})
- 2\, D_{G}^{\mu_{2}\nu_{2}\mu_{3}\nu_{3}\mu_{4}\nu_{4}}(k_{2},k_{3}) \, .
\nonumber \\
\label{TraceConstraint}
\end{eqnarray}
The superscript $d$ in (\ref{TraceConstraint}) indicates that the trace has to be taken in $d = 4 - \epsilon$ dimensions
and the $3$-point function counterterms were computed and tested in \cite{Coriano:2012wp},
as discussed in chapter \ref{TTTVertex}.
The correctness of the counterterms computed for the $4$-point function was already put to a test by the general covariance
Ward identities. Now we can understand why it was worth computing them. In fact, the identities (\ref{TraceConstraint})
relate these counterterms to the functional derivatives of the trace anomaly, whose computation is not less difficult.
Nevertheless, with well checked expressions for the counterterms, we can use eq. (\ref{TraceConstraint}) to test
our functional derivatives of the trace anomaly. The test is highly non trivial, given the complexity of the expressions involved.
We succesfully performed it.
From the discussion in section \ref{CountAnom}, it follows that the traced $3$- and $4$- point correlators of the EMT exactly coincide
with the traces of their counterterms in dimensional regularization. These correlators are expected to be the constituents of the
dilaton perturbative effective action, which we indicate as $\Gamma[\rho]$ and can be tentatively written as the functional series expansion
\begin{eqnarray}
\Gamma[\rho]
&=&
\dots + \int d^4x_{1} d^4x_{2} d^4x_{3} \,
\left\langle {T^{\mu_{1}}}_{\mu_{1}}(x_{1}){T^{\mu_{2}}}_{\mu_{2}}(x_{2}){T^{\mu_{3}}}_{\mu_{3}} (x_{3}) \right\rangle\,
\rho(x_{1})\rho(x_{2})\rho(x_{3}) \nonumber \\
&&
+\, \int d^4x_{1} d^4x_{2} d^4x_{3} d^4x_{4} \,
\left\langle {T^{\mu_{1}}}_{\mu_{1}}(x_{1}){T^{\mu_{2}}}_{\mu_{2}}(x_{2})
{T^{\mu_{3}}}_{\mu_{3}} (x_{3}){T^{\mu_{4}}}_{\mu_{4}} (x_{4}) \right\rangle\, \rho(x_{1})\rho(x_{2})\rho(x_{3})\rho(x_{4})
+ \dots
\label{EffectiveDilaton}
\end{eqnarray}
We will see in the next chapter that the solution of the anomalous constraints generates an expression which is
more involved than (\ref{EffectiveDilaton}).
Nevertheless, the traced correlators of the EMT are the basic constituents of both actions and, in the on-shell limit, they exactly coincide.
\section{Three and four dilaton interactions from the trace anomaly}
From (\ref{Ch2AWard2PF})-(\ref{Ch2AWard4PF}) and from the knowledge of the trace anomalies therein, that we have explicitly
computed, one can get the form of the off-shell 3- and $4$-dilaton ($\rho$) interactions, which are found to be
\begin{eqnarray} \label{Ch23DVertex}
\mathcal V_{\rho\rho\rho}^{\phi}
&=&
- \frac{1}{\Lambda^3}\, \frac{3\,\left( k_{2}^4+ k_{3}^4\right) + 6 \, k_{2}\cdot k_{3} \, \left( k_{2}^2 + k_{3}^2 \right)
+ 4 \,(k_{2}\cdot k_{3})^2 + 5 \, k_{2}^2\,k_{3}^2}{360\,\pi^2}
\, , \nonumber \\
\mathcal V_{\rho\rho\rho}^{\psi}
&=&
- \frac{1}{\Lambda^3}\, \frac{18\,\left( k_{2}^4+ k_{3}^4\right) + 36 \, k_{2}\cdot k_{3} \, \left( k_{2}^2 + k_{3}^2 \right)
+ 29 \,(k_{2}\cdot k_{3})^2 + 25 \, k_{2}^2\,k_{3}^2}{360\, \pi^2} \, ,
\nonumber \\
\mathcal V_{\rho\rho\rho}^V
&=&
- \frac{1}{\Lambda^3}\, \frac{18\,\left( k_{2}^4+ k_{3}^4\right) + 36 \, k_{2}\cdot k_{3} \, \left( k_{2}^2 + k_{3}^2 \right)
+ 49 \,(k_{2}\cdot k_{3})^2 + 5 \, k_{2}^2\,k_{3}^2}{180\,\pi^2} \, ,
\end{eqnarray}
together with the new quartic dilaton interactions
\begin{eqnarray} \label{Ch24DVertex}
\mathcal V_{\rho\rho\rho\rho}^{\phi}
&=&
- \frac{1}{\Lambda^4}\,\frac{1}{60\,\pi^2}\, \bigg[
3\, \left((k_{2}^2)^2 + (k_{3}^2)^2 + (k_{4}^2)^2 \right) +
6\, \left( k_{4}^2\, k_{4} \cdot (k_{2} +k_{3}) + k_{3}^2\, k_{3} \cdot(k_{2} + k_{4}) + k_{2}^2\, k_{2}\cdot(k_{3}+k_{4}) \right)
\nonumber \\
&& \hspace{-10mm}
+ \, 4\, \left( ( k_{2} \cdot k_{4})^2 + ( k_{2} \cdot k_{3})^2 + ( k_{3} \cdot k_{4})^2 \right) +
6\, \left( k_{2} \cdot k_{3}\, k_{2} \cdot k_{4} + k_{3} \cdot k_{2}\, k_{3} \cdot k_{4} + k_{4} \cdot k_{2}\, k_{4} \cdot k_{3} \right)
\nonumber \\
&& \hspace{-10mm}
+\, 5\, \left( k_{2}^2\, k_{3}^2 + k_{3}^2\, k_{4}^2 + k_{2}^2\, k_{4}^2 \right)
+ 5\, \left( k_{2}^2\, k_{3} \cdot k_{4} + k_{3}^2\, k_{2} \cdot k_{4} + k_{4}^2\, k_{2} \cdot k_{3} \right) \bigg] \, ,
\nonumber \\
\mathcal V_{\rho\rho\rho\rho}^{\psi}
&=&
- \frac{1}{\Lambda^4}\, \frac{1}{120\,\pi^2}\,
\, \bigg[
36\, \left((k_{2}^2)^2 + (k_{3}^2)^2 + (k_{4}^2)^2 \right) +
72\, \left( k_{4}^2\, k_{4} \cdot (k_{2} +k_{3}) + k_{3}^2\, k_{3} \cdot(k_{2} + k_{4}) + k_{2}^2\, k_{2}\cdot(k_{3}+k_{4}) \right)
\nonumber \\
&& \hspace{-10mm}
+ \, 58\, \left( ( k_{2} \cdot k_{4})^2 + ( k_{2} \cdot k_{3})^2 + ( k_{3} \cdot k_{4})^2 \right) +
82\, \left( k_{2} \cdot k_{3}\, k_{2} \cdot k_{4} + k_{3} \cdot k_{2}\, k_{3} \cdot k_{4} + k_{4} \cdot k_{2}\, k_{4} \cdot k_{3} \right)
\nonumber \\
&& \hspace{-10mm}
+\, 50\, \left( k_{2}^2\, k_{3}^2 + k_{3}^2\, k_{4}^2 + k_{2}^2\, k_{4}^2 \right)
+ 55\, \left( k_{2}^2\, k_{3} \cdot k_{4} + k_{3}^2\, k_{2} \cdot k_{4} + k_{4}^2\, k_{2} \cdot k_{3} \right) \bigg] \, ,
\nonumber \\
\mathcal V_{\rho\rho\rho\rho}^V
&=&
- \frac{1}{\Lambda^4}\, \frac{1}{60\,\pi^2}\,
\, \bigg[
36\, \left((k_{2}^2)^2 + (k_{3}^2)^2 + (k_{4}^2)^2 \right) +
72\, \left( k_{4}^2\, k_{4} \cdot (k_{2} +k_{3}) + k_{3}^2\, k_{3} \cdot(k_{2} + k_{4}) + k_{2}^2\, k_{2}\cdot(k_{3}+k_{4}) \right)
\nonumber \\
&& \hspace{-10mm}
+ \,98 \, \left( ( k_{2} \cdot k_{4})^2 + ( k_{2} \cdot k_{3})^2 + ( k_{3} \cdot k_{4})^2 \right) +
122\, \left( k_{2} \cdot k_{3}\, k_{2} \cdot k_{4} + k_{3} \cdot k_{2}\, k_{3} \cdot k_{4} + k_{4} \cdot k_{2}\, k_{4} \cdot k_{3} \right)
\nonumber \\
&& \hspace{-10mm}
+\,10 \, \left( k_{2}^2\, k_{3}^2 + k_{3}^2\, k_{4}^2 + k_{2}^2\, k_{4}^2 \right)
+ 35 \, \left( k_{2}^2\, k_{3} \cdot k_{4} + k_{3}^2\, k_{2} \cdot k_{4} + k_{4}^2\, k_{2} \cdot k_{3} \right) \bigg] \, .
\end{eqnarray}
Both the cubic and the quartic terms can be easily modified to account for all the contributions
generated in the nearly conformal limit of the Standard Model, in the form
\begin{eqnarray}
\mathcal V_{\rho\rho\rho / \rho \rho\rho\rho} = \sum_{i=s,f,V} N_i \, V^i_{\rho\rho\rho / \rho \rho\rho\rho}
\end{eqnarray}
where $N_V = 8+3+1 = 12$ is the number of gauge fields, $N_f = 3 \times 6 + 3 + 3/2 = 45/2$ is the number of Dirac fermions, where
the factor 1/2 is due to the fermion chirality, and $N_s = 4$ counts the real scalars of the Higgs doublet.
These corrections, as we have already remarked, are typical of the dilaton interactions and can be derived
without any explicit diagrammatic computation. They provide the starting ground for an analysis of the dilaton effective action,
and characterize the terms which allows to differentiate between the Higgs and the dilaton at the radiative level.
\section{Conclusions}
The analysis of dilaton interactions and of their role in the context of the electroweak symmetry
breaking is particularly interesting at phenomenological level.
In fact, the Standard Model, in the limit in which we drop the Higgs vev, is conformally invariant at high energy,
with a breaking of scale invariance, in this limit, which is related only to the trace anomaly.
The anomalous coupling of the dilaton is responsible for setting a remarkable difference between this state and the Higgs, a property
which remains valid - with no distinction - even if the dilaton is assumed to be a fundamental or a composite scalar or a graviscalar.
We have shown that the anomalous corrections at any order to the dilaton effective action, in the conformal limit, can be extracted
from a general (and model independent) analysis of the Ward identities, with no further input.
We have illustrated the approach up to the quartic order.
\clearpage{\pagestyle{empty}\cleardoublepage}
\chapter{Conformal Trace Relations from the Dilaton Wess-Zumino Action}\label{Recursive}
\section{Introduction}
Anomaly-induced actions play a considerable role among effective field theories.
Simple instances of these types of actions are theories with chiral fermions in the presence of anomalous abelian symmetries
\cite{Wess:1971yu,Treiman:1986ep,Faddeev:1984jp,Babelon:1986sv}, other examples involve conformal
\cite{Duff:1977ay,Deser:1993yx,Deser:1999zv,Polyakov:1981re} and superconformal anomalies \cite{Chaichian:2000wr}.
Direct computations of these actions can be performed in ordinary perturbation theory by the usual Feynman expansion at 1 loop,
but alternative approaches are also possible. In fact, an action which reproduces the same anomaly at low energy can be constructed
quite directly, just as a variational solution of the anomaly condition, without any reference to the diagrammatic expansion.
In gravity, typical examples are anomaly actions such as the Riegert action \cite{Riegert:1987kt}, or the
Wess-Zumino (WZ from now on) dilaton action \cite{Antoniadis:1992xu},
which reproduce the anomaly either with a non-local (Riegert) or with a local
(WZ) effective operator, using a dilaton field in the latter case \cite{Codello:2012sn}. These types of actions are not unique,
since possible contributions which are conformally invariant are not identified by the variational procedure.
It should also be mentioned that a prolonged interest in these actions has been and is linked to the study of the irreversibility of the
Renormalization Group (RG) flow in various dimensions
\cite{Jack:1990eb,Cappelli:1990yc, Anselmi:1997ys, Komargodski:2011vj,Yonekura:2012kb}
and of the trace anomaly matching \cite{Schwimmer:2010za},
since Zamolodchikov's proof of his $c$-theorem for $d=2$ \cite{Zamolodchikov:1986gt}.
A salient feature of some of these anomaly actions, if formulated in a local form, as in the WZ case, is the inclusion
of extra degrees of freedom compared to the original tree-level action. In the case of the chiral anomaly this additional degree of
freedom is the axion $(\theta(x))$, which is linearly coupled to the anomaly functional in the form of a $(\theta/M) F\tilde{F}$ term
- the anomaly coupling - with $F$ and $\tilde{F}$ denoting the field strength of the gauge field and its dual respectively.
The anomaly interaction is accompanied by a new scale ($M$). This is the scale at which the anomalous symmetry starts to play a role
in the effective theory. A large value of $M$, for instance, is then associated with a decoupling of the anomaly in the low energy theory.
In the $1$-particle irreducible (1PI) effective action this is obtained - in the chiral case - by allowing the mass of the fermions ($\sim M$)
that run in the anomaly loops to grow large. The underlying idea of keeping the anomaly interaction in the form of a local operator at low energy -
such as the $(\theta/M) F\tilde{F}$ term - while removing part of the physical spectrum, is important in the study of
the renormalization group (RG) flows of large classes of theories, both for chiral and for conformal anomalies.
For conformal anomalies \cite{Duff:1977ay}, which is the case of interest in this chapter, the pattern is similar to the
chiral case, with the introduction of a dilaton field in place of the axion in order to identify the structure of the corresponding WZ action,
and the inclusion of a conformal scale ($\Lambda$). As in the chiral case, one of the significant features of the WZ conformal anomaly
action is the presence of a linear coupling of the Goldstone mode of the broken symmetry (the dilaton) to the anomaly functional,
but with a significant variant. In this case, in fact, this linear term has to be corrected by additional contributions, due to the non invariance
of the anomaly functional under a Weyl transformation.
This procedure, which allows to identify the structure of WZ action, goes under the name of the {\em Noether method} (see for instance
\cite{Coriano:2013xua,Elvang:2012st}) and has to be iterated several times, due to the structure of the anomaly functional,
before reaching an end. Given the fact that anomaly functional takes a different form in each spacetime dimension, the anomaly action
will involve interactions of the dilaton field of different orders in each dimension.
In \cite{Coriano:2013xua} we have investigated an alternative approach, useful for the computation of this action,
which exploits the structure of the counterterms in dimensional regularization and their {\em Weyl-gauging}, bypassing altogether the
Noether procedure. This approach has been discussed in $d=4$ by several authors
\cite{Antoniadis:1992xu,Tomboulis:1988gw}, and in a cohomological context in \cite{Mazur:2001aa}.
This construction in dimensions higher than $2$ or $4$ is interesting for several reasons.
The WZ action was been used in \cite{Elvang:2012st} in the attempt to generalize the proof of the weak a-theorem provided in
\cite{Komargodski:2011vj}, although some additional work is required, due to the complexity of dispersive analysis of scattering
amplitudes of more than $4$ particles.
At the same time it plays an equally important role in the study of the AdS/CFT correspondence. An example is the investigation of the
anomaly matching between conformal tensor multiplets on the six dimensional boundary and a stack of M5 branes of $AdS_7$ supergravity in
the bulk \cite{Bastianelli:1999ab, Bastianelli:2000hi}. We will present, as an application of our formalism, the expression of the WZ action for
this specific CFT realization in $d=6$. \\
In the two previous chapters, we have focused our attention on interactions concerning a dilaton field, featuring a coupling to the trace anomaly,
which we have interpreted as an effective, low-energy signature of a possible hidden conformal sector that could be investigated in future
data analysis at hadron colliders, maybe at the LHC. \\
In this final chapter, we turn to a different kind of analysis, switching from the phenomenological aspects to a more formal
application of dilaton interactions. We look at our dilaton as at a WZ Goldstone boson for CFT's, which
means that its self-interactions stem only from the trace anomaly.
These self-interactions are related, in turn, to definite combinations of traced EMT correlation functions, which coincide with
those analysed in chapter \ref{Traced4T} in the on-shell limit.
We will start our investigation by introducing our conventions for the anomalous equations and
the structure of correlation functions of traces of the EMT for a generic CFT.
Then we give a brief account of the method of Weyl-gauging, which was already extensively discussed, for scale invariant theories,
in chapter \ref{Weyl}. This gives us the chance to introduce the quadratic kinetic term for the dilaton field.
We also comment on the Weyl-gauging for theories containing dimensionful constants, as the dilaton might actually be a massive state.
This way, we can account for a mass term preserving Weyl symmetry.
In the past, the gauging has been discussed in various ways both in the context of extensions of the Standard Model
\cite{Buchmuller:1987uc,Buchmuller:1988cj} and in cosmology, where it has been shown that the introduction of an extra scalar brings
to a dynamical adjustment of the cosmological constant \cite{Tomboulis:1988gw}.
Recent discussions of the role of the dilaton in quantum gravity can be found in \cite{Henz:2013oxa, Percacci:2011uf}.
This review gives us the chance to classify the possible kinetic terms for the dilaton, on the ground of conformal symmetry requirements
and limiting ourselves to terms that are at most marginal in a wilsonian sense.
Then we turn to the original part of this chapter and apply the method of Weyl-gauging to the counterterms of a CFT in order
to determine the WZ conformal anomaly action and show that in any even ($d = 2 k$) dimensions all the hierarchy of correlation
functions involving traces of the EMT of a CFT is determined in terms of those of lower orders, up to $2 k$,
which are, in turn, completely fixed by the conformal anomaly.
For every even dimension $d$, it is the highest order of dilaton anomalous self-interactions which corresponds to the maximum
order of the independent traced correlators that are necessary to fix the entire hierarchy.
This order, in turn, equals the space dimension, as discussed in \cite{Coriano:2013xua,Coriano:2013nja}.
It turns out that all the correlators which are order $2k+1$ or higher are recursively generated by the first $2k$,
through a simple algorithm that is discussed in detail.
The method also allows to compute the first $2k$ traced correlation functions of the EMT just by knowing the structure of the WZ action,
thus providing an alternative way to the counterterm approach used in chapter \ref{Traced4T}.
We work out explicitly the cases $d=4$ and $d=6$, while the case in $2$ dimensions is left to appendix \ref{Ch42D}.
Both in the $4$- and in the $6$-dimensional case, we derive the dilaton effective action by taking into account the counterterms
required within a minimal subtraction scheme, in a sense that will be made clearer below.
We mention that general results on the structure of the WZ action in any even dimensions have been presented in
\cite{Baume:2013ika}, using the general form of the Euler density, which is sufficient to identify the
nonlocal structure of the anomaly in a specific scheme. However, the identification of the contributions related to the
so called \emph{local anomalies} requires a separate effort, that we undertake here.
This more general approach allows us to set a distinction between the nonlocal and local contributions to the anomalous effective action.
As an application in $6$ dimensions, we present the WZ action for the (2,0) tensor multiplet, which has been investigated in the past in the
context of the $AdS_7/CFT_6$ holographic anomaly matching.
Of the first $6$ independent correlators, we give the explicit expressions in the most general scheme of the first $4$.
We have computed the order $5$ and $6$ Green functions, but they are lengthy and add nothing essential to our discussion. \\
The gauging procedure, especially in $6$ dimensions, in a general scheme, is quite demanding from the technical side.
We set a distinction between the operators in the anomaly that are responsible for the universal (scheme-independent)
contributions, i.e. the Euler density and the Weyl invariants, and the operators which are responsible for the scheme-dependence,
such as the $\square R $ term in $4$ dimensions, establishing clearly the relation between their contribution to the WZ
effective action and the counterterms in the minimal scheme. This clarifies the difference with the prescription of \cite{Baume:2013ika}.
We then move to the analysis of the structure of the traced correlators and of their hierarchy, showing how to solve it in terms of the
first $d$ correlation functions. We have left to appendix \ref{Geometrical} a discussion of some of the more technical steps.
appendix \ref{Ch42D} includes the consistency checks of the recursion relations satisfied by the traced correlators
in $2$ dimensions, presenting the expressions of the first traced Green functions up to rank $6$ in this case.
\section{Conventions}\label{RecConv}
For practical reasons, we recollect here all the basic definitions necessary for the purpose of this chapter.
In a generic euclidean field theory, defining the generating functional of the theory $\mathcal{W}$ as
\begin{equation}
\mathcal{W}[g] = \int \mathcal D \Phi\, e^{- \mathcal S}\, ,
\end{equation}
where $\mathcal{S}$ is the generic euclidean action depending on the set of all the quantum fields ($\Phi$) and on the
background metric ($g$), the vev of the EMT is given by
\begin{equation} \label{Ch3EMTvev}
\left\langle T^{\mu\nu}(x) \right\rangle_s
= \frac{2}{\sqrt{g_x}}\, \frac{\delta \mathcal{W}[g]}{\delta g_{\mu\nu}(x)}
= \frac{2}{\sqrt{g_x}}\, \frac{\delta}{\delta g_{\mu\nu}(x)} \int\, \mathcal{D}\Phi\, e^{-S}\, ,
\end{equation}
and contains the response to the metric fluctuations keeping the background sources turned on, as the superscript $s$ indicates;
for the purpose of this chapter, the only background field is the metric tensor $g_{\mu\nu}$,
whereas $g_x\equiv \left|g_{\mu\nu}(x)\right|$ is its determinant.
For CFT's in a even dimensional space which are coupled to a background metric and neither scalar fields nor vector currents,
the trace anomaly condition takes the general symbolic form
\begin{eqnarray} \label{Ch3TraceAnomaly}
g_{\mu\nu} \langle T^{\mu\nu} \rangle_s = \mathcal{A}[g] \, ,
\end{eqnarray}
An anomalous relation of the form (\ref{Ch3TraceAnomaly}) holds in any even dimension, where $\mathcal{A}[g]$ is a
scalar functional depending only on the metric tensor, which can be written, in complete generality,
as \cite{Duff:1977ay,Duff:1993wm,Deser:1993yx}
\begin{equation}
\label{Ch4anom}
\mathcal{A}[g] =
\sum_{i} c_i\, \left( I_i + \nabla_\mu J^\mu_i \right) - (-1)^{d/2}\, a\, E_d \, ,
\end{equation}
where $\sqrt{g}\, I_i$ are the conformal invariants available in $d$ dimensions, whose number rapidly increases with $d$,
whereas $E_d$ is the Euler density in $d$ dimensions. They are both defined in appendix \ref{Geometrical}. \\
The contribution coming from the Euler density is usually denoted as the $A$ part of the anomaly, while the rest is called the $B$ part.
For every even value of the space dimension, say $2k$, the $I_i$'s are $2k$-dimensional combinations of the Weyl tensor for $2k$
dimensions and covariant derivatives thereof.
The total derivative terms $\nabla_\mu J_i^\mu$ are known under the name of \emph{local anomaly contributions} and are sometimes omitted,
as they are scheme-dependent. This scheme dependence has already been discussed in section \ref{Ch1Renormalization}
for the case $d=4$ and will be further explored in the following in $d=4$ and $d=6$.
The total derivatives can be removed by adding proper local counterterms to the action, as we thoroughly show in sections
\ref{WZ4} and \ref{WZ6}.
Let us provide the explicit expressions of the trace anomaly in $4$ and $6$ dimensions. In the first case, in order to stick to the notations
of the previous chapters, we set $- a = \beta_b$ and $ c = \beta_a$ and have the anomaly functional
\begin{equation}
\mathcal{A}[g] = \beta_a\, \left( F - \frac{2}{3}\, \Box R \right) + \beta_b\, G \, ,
\label{4DANOMALY}
\end{equation}
where $F$ is the squared $4$-dimensional Weyl tensor and $G$ is the Euler density, both defined appendix \ref{Geometrical}
and we are omitting the distinction between the various kinds of contributions that can go into the $\beta$ coefficients.
The specific expression of (\ref{Ch4anom}) for $d=6$ is instead
\begin{equation} \label{Ch4anom6D}
\mathcal{A}[g] =
\sum_{i=1}^3 c_i\, \left( I_i + \nabla_\mu J^\mu_i \right) + a\, E_6\, ,
\end{equation}
where $\sqrt{g}\, I_i\, , (i=1,2,3)$, are the three Weyl invariants available in $6$ dimensions.
Our goal will be to determine the structure of the dilaton WZ action in the most general case for both $d=4$ and $d=6$,
with the inclusion of the contributions related to the total derivative terms and to completely clarify the relation of the latter to the
various possible choices of the counterterms.
As already emphasized, multiple stress-energy tensor correlators can be defined in various ways,
differing by contact terms. These depend on the positions of the $g^{-\frac{1}{2}}$ factors entering in the definition
of the EMT respect to the functional derivatives.
We choose to define the Green function of $n$ EMT's in flat space in the completely symmetric fashion as
\begin{equation} \label{Ch3NPF}
\langle T^{\mu_1\nu_1}(x_1)\ldots T^{\mu_n\nu_n}(x_n)\rangle
\equiv
\frac{2^n}{\sqrt{g_{x_{1}}}\ldots \sqrt{g_{x_n}}}
\frac{\delta^n \mathcal{W}[g]}{\delta g_{\mu_1\nu_1}(x_{1})\ldots \ldots\delta g_{\mu_n\nu_n}(x_n)}
\bigg|_{g_{\mu\nu}=\delta_{\mu\nu}} \, .
\end{equation}
It is also useful to recall some notation to denote the functional derivatives
with respect to the metric of generic functionals in the limit of a flat background
\begin{eqnarray} \label{Ch3funcder}
\left[f(x)\right]^{\mu_{1}\nu_{1}\dots\mu_{n}\nu_{n}}(x_{1},\dots,x_n)
\equiv
\frac{\delta^n\, f(x)}{\delta g_{\mu_n\nu_n}(x_{n}) \, \ldots\, \delta g_{\mu_{1}\nu_{1}}(x_{1})}
\bigg|_{g_{\mu\nu}=\delta_{\mu\nu}}
\end{eqnarray}
and the corresponding expression with traced indices
\begin{equation}
\left[f(x)\right]^{\mu_{1}\dots\mu_n}_{\,\,\,\mu_{1}\dots\nu_n}\left(x_{1},x_{2},\dots,x_n\right)
\equiv \delta_{\mu_{1}\nu_{1}}\dots\delta_{\mu_{n}\nu_{n}}\,
\left[f(x)\right]^{\mu_{1}\nu_{1}\dots\mu_{n}\nu_{n}}\left(x_{1},\dots,x_n\right)\, ,
\end{equation}
where the curved euclidean metric $g_{\mu\nu}$ is replaced by $\delta_{\mu\nu}$.
It is clear that, in any CFT in even dimensions, the only object which plays a role in the
determination of the traces of these correlators is the anomaly functional, as one can realize by a direct computation.
Specifically, from (\ref{Ch3TraceAnomaly}) one can derive trace identities for the $n$-point correlation functions.
In fact, in momentum space the entire hierarchy, which is generated by functional differentiation of (\ref{Ch3TraceAnomaly}),
takes the form
\begin{eqnarray}
\label{Ch3hier}
\left\langle T(k_{1}) \, \dots \, T(k_{n+1})\right\rangle
&=&
2^n\, \left[\sqrt{g}\, \mathcal A \right]^{\mu_{1}\dots\mu_n}_{\,\,\,\mu_{1}\dots\nu_n}\left(k_{1},\dots,k_{n+1}\right)
\nonumber \\
&&
-\, 2 \sum_{i=1}^{n} \left\langle T(k_{1})\dots T(k_{i-1})T(k_{i+1})\dots T(k_{n+1}+k_i) \right\rangle \, .
\end{eqnarray}
In the expression above we have introduced the notation $T \equiv {T^\mu}_\mu$ to denote the trace of the EMT. All
the momenta characterizing the vertex are taken as incoming. \\
The identity (\ref{Ch3hier}) relates a $n$-point correlator to correlators of order $n-1$, together with the completely
traced derivatives of the anomaly functionals. In $4$ dimensions, these are
$\sqrt{g}\,F, \sqrt{g}\,G$ and $\sqrt{g}\,\square R$.
For $\sqrt{g}\,F$, which is a conformal invariant, the completely traced functional derivatives are identically zero.
For $\sqrt{g}\,G$ these are non vanishing at any arbitrary order $n \geq 2$, i.e. for the $TTT$ and higher order functions,
whereas $\sqrt{g}\,\Box R$ contributes also to the trace of the $2$-point function. \\
In order to characterize the expansion of the scalars appearing in the trace anomaly equation for $d=6$ ,
we introduce the basis of dimension $6$ scalars obtained from the Riemann tensor, its contractions and derivatives, which is given
by the $15$ terms in table \ref{Ktable}, according to the conventions of \cite{Henningson:1998gx}
\begin{table}[h]
$$
\begin{array}{lll}
K_1 = R^3 &
K_2 = R\,R^{\mu\nu}\,R_{\mu\nu} &
K_3 = R\,R^{\mu\nu\lambda\kappa}\,R_{\mu\nu\lambda\kappa} \\
K_4 = {R_\mu}^\nu\, {R_\nu}^\alpha\, {R_\alpha}^\mu &
K_5 = R^{\mu\nu}\, R^{\lambda\kappa}\, R_{\mu\lambda\kappa\nu} &
K_6 = R_{\mu\nu}\, R^{\mu\alpha\lambda\kappa}\, {R^\nu}_{\alpha\lambda\kappa} \\
K_7 = R_{\mu\nu\lambda\kappa}\, R^{\mu\nu\alpha\beta}\, {R^{\lambda\kappa}}_{\alpha\beta} &
K_8 = R_{\mu\nu\lambda\kappa}\, R^{\mu\alpha\beta\kappa}\, {{R^\nu}_{\alpha\beta}}^\lambda &
K_9 = R\square R \\
K_{10} = R_{\mu\nu}\square R^{\mu\nu} &
K_{11} = R_{\mu\nu\lambda\kappa}\square R^{\mu\nu\lambda\kappa} &
K_{12} = \partial_{\mu}R\, \partial^{\mu}R \\
K_{13} = \nabla_{\rho}R_{\mu\nu}\, \nabla^{\rho}R^{\mu\nu} &
K_{14} = \nabla_{\rho}R_{\mu\nu\alpha\beta}\, \nabla^{\rho}R^{\mu\nu\alpha\beta} &
K_{15} = \nabla_{\lambda}R_{\mu\kappa}\, \nabla^{\kappa}R^{\mu\lambda} \\
\end{array}
$$
\caption{Basis of dimension-$6$ scalars on which the Euler density and the conformal invariants in $6$ dimensions are expanded.}
\label{Ktable}
\end{table}
\\
In terms of such basis, the Euler density takes the form
\begin{equation} \label{Ch4Euler}
E_6 = K_1 - 12\, K_2 + 3\, K_3 + 16\, K_4 - 24\, K_5 - 24\, K_6 + 4\, K_7 + 8\, K_8 \, .
\end{equation}
Defining a Weyl transformation of the metric as
\begin{equation}
g_{\mu\nu}(x) \to e^{2\, \sigma(x)}\, g_{\mu\nu}(x),
\label{Ch4WeylT}
\end{equation}
the three Weyl invariants (modulo a $\sqrt{g}$ factor) in $d=6$, restricted to operators of dimension $6$, {$I_i\, , i=1,2,3$},
are given by the expressions (see appendix \ref{Geometrical} for their definitions in terms of the Weyl and Riemann tensors)
\begin{eqnarray} \label{Ch4WeylInv6}
I_1
&=&
\frac{19}{800}\, K_1 - \frac{57}{160}\, K_2 + \frac{3}{40}\, K_3 + \frac{7}{16}\, K_4
- \frac{9}{8}\, K_5 - \frac{3}{4}\, K_6 + K_8
\, ,\nonumber \\
I_2
&=&
\frac{9}{200}\, K_1 - \frac{27}{40}\, K_2 + \frac{3}{10}\, K_3 + \frac{5}{4}\, K_4 - \frac{3}{2}\, K_5 - 3\, K_6 + K_7
\, ,
\nonumber \\
I_3 &=&
\frac{1}{25}\, K_1 - \frac{2}{5}\, K_2 + \frac{2}{5}\, K_3 + \frac{1}{5}\, K_9 - 2\, K_{10} + 2\, K_{11}
+ K_{13} + K_{14} - 2\,K_{15} \, .
\end{eqnarray}
It is easy to prove that for the three scalars defined above the products $\sqrt{g}\, I_i$ are Weyl invariant in $6$ dimensions, i.e.,
denoting with $\delta_W$ the operator implementing an infinitesimal Weyl transformation,
\begin{equation}
\delta_{W} I_i = -6 \, \sigma I_{i}\, .
\end{equation}
The origin of the derivative terms $\nabla_\mu J^\mu_i$ in eq. (\ref{Ch4anom6D}) is discussed in section \ref{WZ6}.
Their explicit expressions are
\begin{eqnarray}
\nabla_\mu J_1^\mu
&=&
- \frac{3}{800}\, \nabla_\mu \bigg[
- 5\, \bigg( 44\, R^{\lambda\kappa}\, \nabla^\mu R_{\lambda\kappa} - 50\, R_{\lambda\kappa}\, \nabla^\sigma R^{\mu\rho}
- 3\, R^{\mu\nu}\, \partial_\nu R - 4\, R_{\nu\lambda\kappa\alpha }\, \nabla^\mu R^{\nu\lambda\kappa\alpha}
\nonumber \\
&& \hspace{20mm}
+\, 40\, R^{\mu\lambda\nu\kappa}\, \nabla_\nu R_{\lambda\kappa} \bigg) + 19\, R\, \partial^\mu R \bigg] \, ,
\nonumber \\
\nabla_\mu J_2^\mu
&=&
- \frac{3}{200}\, \nabla_\mu \bigg[
- 5\, \bigg( 4\, R^{\lambda\kappa}\, \nabla^\mu R_{\lambda\kappa}
+ 10\, R_{\lambda\kappa}\, \nabla^\kappa R^{\mu\lambda} + 7\, R^{\mu\nu}\, \partial_\nu R
- 4\, R_{\nu\lambda\kappa\alpha }\, \nabla^\mu R^{\nu\lambda\kappa\alpha}
\nonumber \\
&& \hspace{20mm}
-\, 40\, R^{\mu\lambda\nu\kappa}\, \nabla_\nu R_{\lambda\kappa}\, \bigg) + 9 R\, \partial^\mu R \bigg] \, ,
\nonumber \\
\nabla_\mu J_3^\mu
&=&
\frac{1}{25}\, \nabla_\mu \bigg[
10\, \bigg( 2\,\partial^\mu \square R - 5\, \nabla_\nu \square R^{\mu\nu} + R_{\nu\rho}\, \nabla^\mu R^{\nu\rho}
-2\, R^{\mu\nu}\, \partial_\nu R - R_{\nu\lambda\kappa\alpha }\, \nabla^\mu R^{\nu\lambda\kappa\alpha}
\nonumber \\
&& \hspace{15mm}
-\, 10\, R^{\mu\lambda\nu\kappa}\, \nabla_\nu R_{\lambda\kappa}\, \bigg) - 3\, R\, \partial^\mu R \bigg] \, .
\end{eqnarray}
For about the completely traced derivatives of the anomaly functionals in $d=6$,
$\sqrt{g}\,I_{i}$, $\sqrt{g}\,E_6$ and $\sqrt{g}\,\nabla_\mu J^\mu_i$,
those of $\sqrt{g}\,I_i$, which are conformal invariants, are identically zero.
Concerning $\sqrt{g}\,E_6$, which is cubic in the Riemann tensor, these contributions are non vanishing at any arbitrary order $n \geq 3$.
Finally, $\sqrt{g}\,\nabla_\mu J^\mu_i$ contribute to lower order functions as well.
In particular, being $\nabla_\mu J^\mu_{1}$ and $\nabla_\mu J^\mu_{2}$
at least quadratic in the Riemann tensor, they give non-vanishing contributions
from order $3$ onwards, whereas $\nabla_\mu J^\mu_3$ contains a term which is linear in $R$
and thus contributes a non-vanishing trace to the $2$-point function as well.
\section{Overview of Weyl-gauging }
The goal of this section is to extend the method of Weyl-gauging, already discussed in chapter \ref{Weyl},
to non scale invariant theories.
We first give a quick overview of the method for classical theories and later discuss its application to the renormalized quantum
effective action of a CFT, so as to provide the basics for our derivation of the WZ action for conformal anomalies.
We also comment on possible kinetic terms for the dilaton, which obviously cannot be determined by the anomaly equation.
Their possible form is suggested on the sole requirement of conformal invariance.
\subsection{Weyl-gauging for scale invariant theories}
Scale invariance in flat space is equivalent, once the Lagrangian has been merged in a gravitational background,
to global Weyl invariance.
The equivalence is shown by rewriting a scale transformation acting on the coordinates of flat space and the fields $\Phi$,
\begin{eqnarray} \label{Ch4FlatScaling}
x^\mu &\to& {x'}^{\mu}=e^\sigma x^\mu \, , \nonumber\\
\Phi(x)&\to& \Phi'(x')=e^{-d_\Phi \sigma}\Phi(x) \, ,
\end{eqnarray}
in terms of a rescaling of the metric tensor, the Vielbein and the matter fields
\begin{eqnarray} \label{Ch4CurvedScaling}
g_{\mu\nu}(x) &\to& e^{2\sigma}\, g_{\mu\nu}(x)\, , \nonumber\\
{V}_{a\,\rho}(x) &\to& e^{\sigma}\, V_{a\,\rho}(x)\, , \nonumber\\
\Phi(x)&\to& e^{-d_\Phi \sigma}\Phi(x)\, ,
\end{eqnarray}
leaving the coordinates $x$ unchanged, where $d_\Phi$ is the canonical scaling dimension of the field $\Phi$.
In a curved metric background, we can promote $\sigma$ to a local function, and thus turn the global scale transformations
(\ref{Ch4CurvedScaling}) into
\begin{eqnarray}
\label{Ch4WeylTransf}
{g'}_{\mu\nu}(x) &=& e^{2\sigma(x)}\, g_{\mu\nu}(x)\, , \nonumber \\
{V'}_{a\,\rho}(x) &=& e^{\sigma(x)}\, V_{a\,\rho}(x)\, , \nonumber \\
\Phi'(x) &=& e^{-d_{\Phi}\, \sigma(x)}\, \Phi(x)\, ,
\end{eqnarray}
leave the fundamental Lagrangian invariant. \\
Following closely the analogy with quantum electrodynamics, derivative terms are modified according to the prescription
\begin{equation} \label{Ch4DerTransf}
\partial_\mu \rightarrow \partial^W_\mu = \partial_\mu - d_{\phi}\, W_{\mu}\, ,
\end{equation}
where $W_{\mu}$ is the Weyl vector gauge field, that shifts under Weyl scaling as
\begin{equation} \label{Ch4WeylGaugeField}
W_{\mu} \rightarrow W_{\mu} - \partial_\mu \sigma\, ,
\end{equation}
just as for a gauge transformation of the vector potential.
In the case of higher spin fields, e.g. a vector field $v_\mu$, the Weyl-gauging has to be supplemented
with a prescription to render the general covariant derivative Weyl invariant, which is to add to (\ref{Ch4DerTransf}) the modified
Christoffel connection
\begin{equation} \label{Ch4ModChristoffel}
\hat\Gamma^\lambda_{\mu\nu} =
\Gamma^\lambda_{\mu\nu} + {\delta_\mu}^\lambda\, W_\nu + {\delta_\nu}^\lambda\, W_\mu - g_{\mu\nu}\, W^\lambda\, .
\end{equation}
This hatted Christoffel symbol is Weyl invariant, so that we can define the Weyl covariant derivatives acting on vector fields as
\begin{eqnarray} \label{Ch4WeylChristoffel}
\nabla^W_\mu v_\nu
&=&
\partial_\mu v_\nu - d_v\, W_\mu v_\nu - \hat\Gamma^\lambda_{\mu\nu} v_\lambda \, ,
\nonumber\\
\nabla^W_\mu v_\nu
&\rightarrow&
e^{- d_v\sigma(x)}\, \nabla^W_\mu v_\nu \, ,
\end{eqnarray}
which is obviously generalized to tensors of arbitrary rank. \\
In order to include fermions, we can define the covariant derivative
\begin{equation} \label{Ch4WeylSpinConnection}
\nabla_{\mu} \rightarrow \nabla^W_\mu = \nabla_{\mu} - d_\psi\, W_{\mu} + 2\, {\Sigma_\mu}^\nu\, W_\nu\, , \quad
\Sigma^{\mu\nu} \equiv {V_a}^\mu\, {V_b}^\nu \Sigma_{ab}\, ,
\end{equation}
where $\Sigma_{ab}$ are the spinor generators of the Lorentz group.
More details on the Weyl-gauging for scale invariant theories are given in chapter \ref{Weyl}.
Here we want to discuss kinetic terms for the additional degree of freedom that is introduced when a theory is Weyl-gauged.
For instance, the Weyl vector field $W_{\mu}$ can be rendered dynamical by the inclusion
of a kinetic term built out of an appropriate field strength
\begin{equation}
F^W_{\mu\nu}\equiv\partial_{\mu} W_{\nu} - \partial_{\nu} W_{\mu} \, ,
\label{Ch4FieldStrength}
\end{equation}
which is manifestly Weyl invariant.
A second possibility is to maintain the expression of $W_\mu$ identifying it with the gradient of a dilaton field,
\begin{equation}
\label{Ch4dil}
W_\mu(x) = \frac{\partial_{\mu} \rho(x)}{\Lambda}\, .
\end{equation}
As we will shortly point out below, this second choice offers an interesting physical interpretation, in connection with the breaking
of the conformal symmetry, which is related to the conformal scale $\Lambda$. Notice that in this second case the
$\Omega$ term generates non trivial cubic and quartic interactions between
the original scalar and the dilaton
\begin{equation} \label{Ch4Omegatau}
\Omega_{\mu\nu}\left(\frac{\partial\rho}{\Lambda}\right) =
\frac{\nabla_\mu \partial_\nu\rho}{\Lambda} - \frac{\partial_\mu\rho\,\partial_\nu\rho}{\Lambda^2}
+ \frac{1}{2}\,g_{\mu\nu}\, \frac{\left(\partial\rho \right)^2}{\Lambda^2},
\end{equation}
which bring the Weyl-gauged action for the scalar field (\ref{ChIntgauged}) to the form
\begin{equation}
\mathcal S_{\phi,\partial\rho} =
\frac{1}{2}\, \int d^dx\, \sqrt{g}\, g^{\mu\nu}\, \left(
\partial_\mu \phi\, \partial_\nu\phi + \frac{d-2}{2}\, \phi^2\, \frac{\Box\rho}{\Lambda}
+ \left(\frac{d-2}{2}\right)^2 \, \phi^2\, \frac{\left(\partial\rho\right)^2}{\Lambda^2} \right) \, .
\label{Ch4gaugedtau}
\end{equation}
As the field strength $F^W_{\mu\nu}$ in (\ref{Ch4FieldStrength}), on account of (\ref{Ch4dil}) is obviously zero, the standard way
to give a kinetic term to the dilaton is by introducing a conformally coupled scalar field $\chi$ and imposing the field redefinition
\begin{equation}
\chi(\rho) \equiv \Lambda^{\frac{d-2}{2}}\, e^{-\frac{(d-2)\,\rho}{2\Lambda}}.
\label{Ch4chitau}
\end{equation}
%
It is clear that, with the choice (\ref{Ch4dil}), we are no longer gauging the Weyl group, but we are just using a compensation
procedure to introduce an additional scalar field which we later render dynamical via the kinetic term for $\chi$.
As we shall see in the next section, this is mandatory if one wants to achieve Weyl invariance for a theory containing dimensional
parameters.
At this point, the dynamics of the combined scalar/dilaton/graviton system is described by the Weyl invariant action
\begin{equation}
\mathcal{S}= S_{\chi(\rho), imp} + S_{\phi,\partial \rho} \, ,
\end{equation}
having combined (\ref{ChIntgauged}), where $\phi$ is replaced by $\chi$, and (\ref{Ch4gaugedtau}).
The kinetic action for $\chi$, $S_{\chi(\rho), imp}$, takes the form
\begin{equation} \label{Ch4DilatonKinetic}
\mathcal S_{\chi(\rho), imp} = \frac{\Lambda^{d-2}}{2}\, \int d^dx\, \sqrt{g}\, e^{-\frac{(d-2)\,\rho}{\Lambda}}\, \bigg(
\frac{(d-2)^2}{4\,\Lambda^2}\, g^{\mu\nu}\,\partial_\mu \rho\, \partial_\nu\rho - \frac{1}{4}\, \frac{d-2}{d-1}\, R \bigg) \, .
\end{equation}
The Weyl-gauging procedure, as we have described it so far, is possible only when we take as a starting point a scale invariant Lagrangian,
with dimensionless constants.
Things are different when an action is not scale invariant in flat space, and in that case the same gauging requires some extra steps.
We illustrate this point below and discuss the modification of the procedure outlined above, by considering again a scalar theory as an example.
This approach exemplifies a situation which is typical in theories with spontaneous breaking of the ordinary gauge symmetry,
such as the Standard Model.
\subsection{Weyl-gauging for non scale invariant theories }
We consider a free scalar theory with a mass term
\begin{equation}
\label{Ch4tre}
\mathcal{S}_2 = \frac{1}{2}\, \int d^d x \sqrt{g}\, \left(g^{\mu\nu}\,\partial_{\mu}\phi\, \partial_\nu \phi + m^2\, \phi^2 \right) \, .
\end{equation}
Scale invariance is lost, but it can be recovered.
There are two ways to promote this action to a scale invariant one. The first is simply to render the mass term dynamical
\begin{equation}
\label{Ch4redef}
m\to m\, \frac{\Sigma}{\Lambda}\, ,
\end{equation}
using a second scalar field, $\Sigma$. The action (\ref{Ch4tre}),
with the replacement (\ref{Ch4redef}), can be extended with the inclusion of the
kinetic term for $\Sigma$. The inclusion of $\Sigma$ and the addition of two conformal couplings (i.e. of two Ricci gaugings) both for
$\phi$ and $\Sigma$ brings to the new action
\begin{equation}
\label{Ch4tre1}
\mathcal{S}^{\Sigma}_2 = \int d^d x \sqrt{g}\, \left[ \frac{1}{2}\,g^{\mu\nu}\,
\bigg(\partial_{\mu}\phi\, \partial_\nu \phi + \partial_{\mu}\Sigma \, \partial_\nu \Sigma \bigg)
+ \frac{1}{2}\, m^2\, \frac{\Sigma^2}{\Lambda^2}\, \phi^2
+ \frac{1}{4}\, \frac{d-2}{d-1}\,R\, \bigg(\phi^2 + \Sigma^2\bigg) \right]\, ,
\end{equation}
which is Weyl invariant in curved space.
These types of actions play a role in the context of Higgs-dilaton mixing in conformal invariant extension of the Standard
Model, where $\phi$ is replaced by the Higgs doublet and $\Sigma$ is assumed to acquire a vacuum expectation value (vev)
which coincides with the conformal breaking scale $\Lambda$ ($\langle \Sigma \rangle=\Lambda$)
(see for instance \cite{Coriano:2012nm}).
The mixing is induced by a simple extension of (\ref{Ch4tre1}), where the mass term is generated via the scale invariant potential
\begin{equation}
\mathcal{S}_{pot}=
\lambda\, \int d^4 x\, \sqrt{g}\, \left( \phi^2 - \frac{\mu^2}{2\,\lambda}\frac{\Sigma^2}{\Lambda^2} \right)^2 \,
\label{Ch4example}
\end{equation}
(with $m=\mu$). This choice provides a clear example of a Weyl invariant Lagrangian that allows a spontanous breaking of the
$Z_2$ symmetry of the scalar sector $\phi$, following the breaking of the conformal symmetry
($\langle\Sigma \rangle=\Lambda, \textrm{with} \langle \rho\rangle=0$).
The theory is obviously Weyl invariant, but the contributions proportional to the Ricci scalar $R$ do not survive in the flat limit. \\
We have to mention that the approach to Weyl-gauging discussed so far for scale invariant and non scale invariant theories is not unique.
In fact, a second alternative in the construction of a Weyl invariant Lagrangian in curved space,
starting from (\ref{Ch4tre}), using the compensation procedure, which amounts to the replacements
\begin{eqnarray} \label{Ch4Compensate}
m &\to& m \, e^{- \rho/\Lambda}\, ,\nonumber \\
g_{\mu\nu}&\to& \hat{g}_{\mu\nu}\equiv g_{\mu\nu}\, e^{-2 {\rho/\Lambda }} \, \nonumber \\
\phi&\to & \hat{\phi}\equiv \phi \, e^{\rho/\Lambda} \, , \nonumber\\
\partial_\mu \phi &\to& \partial_\mu \hat{\phi}= e^{\rho/\Lambda}\, \partial^W_\mu\, \phi \, .
\end{eqnarray}
It is immediately seen that, for instance, that the application of the replacements (\ref{Ch4Compensate})
to the scalar field action (\ref{Ch4tre}) give
\begin{equation}
\hat{\mathcal{S}}_2
\equiv
\mathcal{S}_2(\hat{g},\hat{\phi})=
\frac{1}{2}\, \int d^4 x\, \sqrt{g}\, \bigg[ g^{\mu\nu}\,\partial_{\mu}\phi\, \partial_\nu \phi
+ g^{\mu\nu}\,\Omega_{\mu\nu}\left(\frac{\partial \rho}{\Lambda}\right)\phi^2 + m^2\, e^{-2\,\rho/\Lambda}\, \phi^2
\bigg] \, ,
\label{Ch4newflat}
\end{equation}
where $\Omega(\partial\rho/\Lambda)$ was defined in (\ref{Ch4Omegatau}).
Also in this case, the compensator $\rho$ ca be promoted to a dynamical field by adding to
$\hat{\mathcal{S}}_2$ the kinetic contribution of a conformally coupled scalar (\ref{Ch4DilatonKinetic}),
thereby obtaining the total action
\begin{equation}
\mathcal{S}_T \equiv \hat{\mathcal{S}}_2 + \mathcal{S}_{\chi(\rho), imp}\, .
\end{equation}
Notice that in this case we choose not to require the Ricci gauging of the $\Omega\left(\partial\rho/\Lambda \right)$
term in $\hat{\mathcal{S}}_2$, but we leave it as it is, thereby generating additional interactions between the dilaton and the scalar $\phi$
in flat space. Obviously, also following this second route, we can incorporate spontaneous breaking of the $Z_2$ symmetry of the
$\phi$ field after the breaking of conformal invariance (with $\langle\Sigma\rangle=\Lambda)$. This is obtained, as before, by the inclusion of
the potential (\ref{Ch4example}).
In this second approach the $\Omega(\partial \rho/\Lambda)$ terms are essential in order to differentiate between the two residual dilaton
interactions in flat space. In the context of Weyl invariant extensions of the Standard Model, such terms are naturally present in the analysis
of \cite{Buchmuller:1988cj}.
\subsection{The dynamical dilaton}\label{DynDil}
So far, we have reviewed how to build standard kinetic terms for a dilaton field, essentially by requiring Weyl invariance to be preserved.
Nevertheless, a general Lagrangian describing the dilaton dynamics could contain, in principle, higher derivative contributions.
It is trivial that any diffeomorphism-invariant functional of the field-enlarged metric $\hat{g}_{\mu\nu}$ in (\ref{Ch4Compensate})
is Weyl invariant. Thus, by applying the Weyl-gauging procedure to all the infinite set of diffeomorphism invariant functionals which can be built
out of the metric tensor and of increasing mass dimension, one can identify the homogeneous terms of the anomaly action.
Beside, there will be the anomalous contributions, which are accounted for by the WZ action. The latter can be added in
order to identify a consistent anomaly action.
As we will see, these terms play no role in the determination of the hierarchy (\ref{Ch3hier}), so that the ambiguity intrinsic to their choice
is no harm for our results.
The Weyl invariant terms may take the form of any scalar contraction of $\hat R_{\mu\nu\rho\sigma}$, $\hat R_{\mu\nu}$, $\hat R$ù
and Weyl covariant derivatives thereof and can be classified by their mass dimension. Typical examples are
\begin{equation}
\mathcal{J}_n\sim \frac{1}{\Lambda^{2(n-2)}}\int d^4 x \sqrt{\hat{g}}\hat{R}^n \, .
\end{equation}
In principle, all these terms can be included into $\Gamma_0[\hat{g}]\equiv \Gamma_0[g,\rho]$ which describes
the non anomalous part of the renormalized action
\begin{equation}
\Gamma_0[\hat{g}]\sim \sum_n \mathcal J_n[\hat{g}]\, .
\end{equation}
Here we recall the structure of the operators that are at most marginal from the Renormalization Group viewpoint.
The first term that can be included is trivial, corresponding to a cosmological constant contribution
\begin{equation}
\mathcal S^{(0)}_\rho = \Lambda^d\, \int d^dx\, \sqrt{\hat{g}}=
\Lambda^d\, \int d^dx\, \sqrt{g}\,e^{-\frac{d\,\rho}{\Lambda}} \, .
\end{equation}
Here the superscript number in round brackets in $\mathcal{S}^{(2 n)}$ denotes the order of the contribution in the derivative expansion,
so to distinguish the scaling behaviour of the various terms under the variation of the length scale.
These terms can be included into the effective action of the theory, that we name $\Gamma_0[\hat{g}]\equiv \Gamma_0[g,\rho]$,
which describes the non anomalous part of the interactions,
\begin{equation}
\Gamma_0[\hat{g}]\sim \sum_n \mathcal J_n[\hat{g}] \, .
\end{equation}
The next contribution to $\Gamma_0$ is the kinetic term for the dilaton, which can be obtained in two ways.
The first method is to consider the Weyl-gauged Einstein-Hilbert term
\begin{eqnarray}
\int d^dx\, \sqrt{\hat g}\, \hat R
&=&
\int d^dx\, \sqrt{g}\, e^{\frac{(2-d)\,\rho}{\Lambda}}\, \bigg[
R - 2\, \left(d-1\right)\, \frac{\Box\rho}{\Lambda} +
\left(d-1\right)\,\left(d-2\right)\, \frac{\left(\partial\rho\right)^2}{\Lambda^2} \bigg] \nonumber \\
&=&
\int d^dx\, \sqrt{g}\, e^{\frac{(2-d)\,\rho}{\Lambda}}\, \bigg[R -
\left(d-1\right)\,\left(d-2\right)\, \frac{\left(\partial\rho\right)^2}{\Lambda^2} \bigg]\, ,
\end{eqnarray}
with the inclusion of an appropriate normalization
\begin{eqnarray} \label{Ch3EH}
\mathcal S^{(2)}_{\rho} = - \frac{\Lambda^{d-2}\, \left(d-2\right)}{8\,\left(d-1\right)}\, \int d^dx\, \sqrt{\hat g}\, \hat R \, ,
\end{eqnarray}
which reverts the sign in front of the Einstein term. We recall that the extraction of a conformal factor ($\tilde{\sigma}$)
from the Einstein-Hilbert term from a fiducial metric $\bar{g}_{\mu\nu}$ ($g_{\mu\nu}=\bar{g}_{\mu\nu}e^{\tilde{\sigma}}$)
generates a kinetic term for ($\tilde{\sigma}$) which is ghost-like. In this case the non-local anomaly action, which in perturbation theory
takes Riegert's form \cite{Riegert:1984kt}, can be rewritten in the WZ form but at the cost of sacrificing covariance, due
to the specific choice of the fiducial metric.
An alternative method consists in writing down the usual conformal invariant action for a scalar field $\chi$
in a curved background
\begin{eqnarray} \label{Ch3ScalarImprovenChi}
\mathcal S^{(2)}_{\chi} = \frac{1}{2}\, \int d^dx\, \sqrt{g}\, \bigg(
g^{\mu\nu}\,\partial_\mu \chi\, \partial_\nu\chi - \frac{1}{4}\,\frac{d-2}{d-1}\, R\,\chi^2 \bigg) \, .
\end{eqnarray}
By the field redefinition $\chi\equiv\Lambda^{\frac{d-2}{2}}\, e^{-\frac{(d-2)\,\rho}{2\Lambda}}$
eq. (\ref{Ch3ScalarImprovenChi}) becomes
\begin{equation} \label{Ch3DilatonKinetic}
\mathcal S^{(2)}_{\rho} = \frac{\Lambda^{d-2}}{2}\, \int d^dx\, \sqrt{g}\, e^{-\frac{(d-2)\,\rho}{\Lambda}}\, \bigg(
\frac{(d-2)^2}{4\,\Lambda^2}\, g^{\mu\nu}\,\partial_\mu \rho\, \partial_\nu\rho - \frac{1}{4}\, \frac{d-2}{d-1}\, R \bigg) \, ,
\end{equation}
which, for $d=4$, reduces to the familiar form
\begin{equation} \label{Ch3KinTau}
\mathcal S^{(2)}_{\rho} = \frac{1}{2}\, \int d^4 x\, \sqrt{g}\, e^{-\frac{2\,\rho}{\Lambda}}\, \bigg(
g^{\mu\nu}\,\partial_\mu \rho\, \partial_\nu\rho - \frac{\Lambda^2}{6}\, R \bigg)\,
\end{equation}
and coincides with the previous expression (\ref{Ch3EH}), obtained from the formal Weyl invariant construction. \\
In four dimensions we can build the following possible subleading contributions (in $1/\Lambda$) to the effective action which, when gauged,
can contribute to the fourth order dilaton action
\begin{equation}
S^{(4)}_\rho = \int d^4x\, \sqrt{g}\, \bigg( \alpha\, R^{\mu\nu\rho\sigma}\, R_{\mu\nu\rho\sigma}
+ \beta\, R^{\mu\nu}\, R_{\mu\nu} + \gamma\, R^2 + \delta\, \Box R\bigg)\, .
\end{equation}
The fourth term ($\sim \Box R$) is just a total divergence, whereas two of the remaining three terms can be
traded for the squared Weyl tensor $F$ and the Euler density $G$.
As $\sqrt{g}\, F$ is Weyl invariant and $G$ is a topological term, neither of them contributes,
when gauged according to (\ref{Ch4Compensate}), so that the only non vanishing $4$-derivative term
in the dilaton effective action in four dimensions is
\begin{equation} \label{Ch3UpToMarginal}
S^{(4)}_{\rho} = \gamma\, \int d^4x\, \sqrt{\hat{g}}\, \hat{R}^2 =
\gamma\, \int d^4x\, \sqrt{g}\,
\bigg[ R - 6\, \bigg( \frac{\Box\rho}{\Lambda} - \frac{\left(\partial\rho\right)^2}{\Lambda^2} \bigg) \bigg]^2 \, .
\end{equation}
with $\gamma$ a dimensionless constant.
Thus, we have got the final form of the dilaton effective action in $d=4$ up to order four in the derivatives of the metric tensor
\begin{eqnarray}
\label{Ch3tot}
S_{\rho} &=&
S^{(0)}_{\rho} + S^{(2)}_{\rho} + S^{(4)}_{\rho} + \dots =
\int d^4x\, \sqrt{\hat{g}}\, \bigg\{ \alpha
- \frac{\Lambda^{d-2}\, \left(d-2\right)}{8\,\left(d-1\right)}\, \hat{R} + \gamma\, \hat{R}^2\,
\bigg\} + \dots\, ,
\end{eqnarray}
where the ellipsis refer to additional operators which are suppressed in $1/\Lambda$.
In flat space ($g_{\mu\nu}\rightarrow \delta_{\mu\nu}$), (\ref{Ch3UpToMarginal}) becomes
\begin{equation}
\mathcal S_{\rho} = \int d^4x\, \bigg[ e^{-\frac{4\,\rho}{\Lambda}}\, \alpha +
\frac{1}{2}\, e^{-\frac{2\,\rho}{\Lambda}} \, \left(\partial\rho\right)^2 +
36\,\gamma\, \bigg( \frac{\Box\rho}{\Lambda} - \frac{\left(\partial\rho\right)^2}{\Lambda^2} \bigg) \bigg] + \dots
\end{equation}
where the ellipsis refer to higher dimensional contributions.
In general we can identify $\Gamma_0[\hat{g}]$ with $S_{\rho}$ as given in (\ref{Ch3tot}), thereby fixing the Weyl invariant
contribution to $\Gamma_{\textrm{ren}}$. \\
Now we consider the case of $6$ dimensions, where operators are marginal up to dimension $6$.
From (\ref{Ch3DilatonKinetic}), the kinetic term in $6$ dimensions is just
\begin{equation} \label{KinRho6}
\mathcal S_{\chi(\tau), imp} =
\int d^6 x\, \sqrt{g}\, e^{-\frac{4\,\tau}{\Lambda}}\,
\bigg( 2\, \Lambda^2\, g^{\mu\nu}\, \partial_\mu \tau\, \partial_\nu\tau - \frac{\Lambda^4}{10}\, R \bigg)\, .
\end{equation}
The possible $4$-derivative terms ($n=2$) are
\begin{equation}
\label{Ch44der}
\int d^6x\, \sqrt{\hat g}\, \bigg( \alpha\, \hat R^{\mu\nu\lambda\kappa}\, \hat R_{\mu\nu\lambda\kappa}
+ \beta\, \hat R^{\mu\nu}\, \hat R_{\mu\nu}+ \gamma\, \hat R^2 + \delta\, \hat \Box \hat R \bigg)\, .
\end{equation}
The $\square R$ contribution in this expression can be obviously omitted, being a total derivative.
We can also replace the Riemann tensor with the Weyl tensor squared and remain with only two
(as $\sqrt{\hat g}\,\hat C^{\mu\nu\lambda\kappa}\, \hat C_{\mu\nu\lambda\kappa}=
\sqrt{g}\,C^{\mu\nu\lambda\kappa}\, C_{\mu\nu\lambda\kappa}\, e^{\frac{2\rho}{\Lambda}} $)
non trivial contributions, $\hat R^{\mu\nu}\, \hat R_{\mu\nu}$ and $\hat R^2$.
We present here the expression of (\ref{Ch44der}) for a conformally flat metric, while the result for a general gravitational background
can be computed exploiting the Weyl-gauged tensors given in appendix \ref{Geometrical},
\begin{eqnarray}
S^{(4)}_\rho
&=&
\int d^6x\, \sqrt{\hat g}\, \bigg( \alpha\, \hat R^{\mu\nu}\, \hat R_{\mu\nu} + \beta\, \hat R^2\bigg)
\nonumber \\
&=&
\int d^6x\, e^{-\frac{2\,\rho}{\Lambda}}\, \bigg[
100\, \alpha\, \bigg( \frac{\Box\rho}{\Lambda} - 2\,\frac{\left(\partial\rho\right)^2}{\Lambda^2} \bigg)^2
+ 2\, \beta\, \bigg( 15\,\frac{\left(\Box\rho\right)^2}{\Lambda^2}
- \frac{68}{\Lambda^3}\, \Box\rho\, \left(\partial\rho\right)^2 + 72\, \frac{\left(\partial\rho\right)^4}{\Lambda^4} \bigg)
\bigg]\, .
\end{eqnarray}
The last contributions that are significant down to the infrared regime are the marginal ones, i.e. the 6-derivative operators.
To derive them we follow the analysis in \cite{Elvang:2012st}. We use the basis of diffeomorphic invariants which are order 6 in the derivatives,
on which the $I_i's$ are expanded (see eq. (\ref{Ch4WeylInv6})). It is made of 11 elements, 6 of which contain the Riemann tensor, that
can be traded for a combination of the Weyl tensor and the Ricci tensor and scalar, so that we are left with only the 5 terms in
$(K_1 - K_{11})$ (see table \ref{Ktable}) that do not contain the Riemann tensor.
As we are going to write down the result only in the flat limit, we can exploit two additional constraints.
Indeed in \cite{Anselmi:1999uk} it was shown that, in this case, the integral of
\begin{equation}
R^3 - 11\, R\,R^{\mu\nu}\, R_{\mu\nu} + 30\, {R_\mu}^\nu\,{R_\nu}^\alpha\,{R_\alpha}^\mu
- 6\, R\Box R + 20\, R^{\mu\nu}\Box R_{\mu\nu}
\end{equation}
vanishes, so that we can use this result to eliminate $R^{\mu\nu}\Box R_{\mu\nu}$. \\
Then, as the Euler density can be written in the form
\begin{eqnarray}
E_6
&=&
\frac{21}{100}\, R^3 - \frac{27}{20}\, R\, R^{\mu\nu}\, R_{\mu\nu}
+ \frac{3}{2}\, {R_\mu}^\nu\,{R_\nu}^\alpha\, {R_\alpha}^\mu
+ 4\, C_{\mu\nu\lambda\kappa}\,{C^{\mu\nu}} _{\alpha\beta}\, C^{\lambda\kappa\alpha\beta}
\nonumber \\
&&
-\, 8\, C_{\mu\nu\lambda\kappa}\,C^{\mu\alpha\lambda\beta}\, {{{C^\nu}_\alpha}^\kappa}_\beta
-6\, R_{\mu\nu}\, C^{\mu\alpha\lambda\kappa}\, {C^\nu}_{\alpha\lambda\kappa}
+\frac{6}{5}\, R\, C^{\mu\nu\lambda\kappa}\, C_{\mu\nu\lambda\kappa}
- 3\, R^{\mu\nu}\, R^{\lambda\kappa}\, C_{\mu\lambda\kappa\nu} \, ,
\end{eqnarray}
it is apparent that only the first three terms are non vanishing on a conformally flat metric.
Now, as in the effective action these contributions are integrated and the Euler density is a total derivative,
one can thereby replace ${R_\mu}^\nu\,{R_\nu}^\alpha\,{R_\alpha}^\mu$ for $R^3$ and $R\,R^{\mu\nu}\,R_{\mu\nu}$.
In the end, Weyl-gauging $R^3$, $R^{\mu\nu}\,R_{\mu\nu}$ and $R\Box R$ is sufficient to account for all the possible
6-derivative terms of the dilaton effective action which do not vanish in the flat space limit.
After some integrations by parts, one can write the overall contribution as
\begin{eqnarray}
S^{(6)}_\rho
&=&
\int d^6x\, \sqrt{\hat{g}}\, \bigg[ \gamma\, \hat{R}^3
+ \delta \hat{R}\,\hat R^{\mu\nu}\,\hat R_{\mu\nu}\, + \zeta\, \hat R \hat \Box \hat R \bigg] \nonumber \\
&=&
\int d^6x\, 20\, \bigg[ \frac{1}{\Lambda^3}\,
\bigg( 5\,\zeta\, \Box^2\rho\, \Box\rho -(50\, \gamma + 7\, \delta - 30\, \zeta)\, (\Box\rho)^3
- 8 \,(\delta + 5\,\zeta)\, \Box\rho\, (\partial\pd\rho)^2 \bigg)\nonumber \\
&& \hspace{14mm}
+\, \frac{1}{\Lambda^4}\, \bigg( 50\, (6\,\gamma + \delta - 2\,\zeta)\, (\Box\rho)^2\, (\partial\rho)^2
- 16\, (\delta + 5\, \zeta)\, \Box\rho\, \partial_\mu \partial_\nu\rho\, \partial^\mu\rho\, \partial^\nu\rho
\nonumber \\
&& \hspace{14mm}
+\, 8\,(2\,\delta + 5\,\zeta)\, (\partial\rho)^2 (\partial\pd\rho)^2 \bigg)
- \frac{120}{\Lambda^5}\,(5\,\gamma + \delta - \zeta)\, \Box\rho\,(\partial\rho)^4
+ \frac{80}{\Lambda^6}\,(5\,\gamma + \delta - \zeta)\,(\partial\rho)^6
\bigg].
\end{eqnarray}
We have introduced the compact notation $\left(\partial\rho\right)^n \equiv
\left( \partial_\lambda\rho\,\partial^\lambda\rho \right)^{n/2}\, , \left(\partial\pd\rho\right)^2 \equiv
\partial_\mu\partial_\nu\rho\, \partial^\mu\partial^\nu\rho $ to denote multiple derivatives of the dilaton field.
The Weyl invariant part of the dilaton effective action is then given by
\begin{equation}
\Gamma_0[g,\rho] =
\mathcal{S}^{(0)}_\rho + \mathcal{S}^{(2)}_\rho + \mathcal{S}^{(4)}_\rho + \mathcal{S}^{(6)}_\rho
+ \dots \, ,
\end{equation}
where the ellipsis denote all the possible higher-order, irrelevant terms.
\subsection{ Weyl-gauging of the renormalized effective action}
The starting relation of our argument is the cocycle condition satisfied by the WZ anomaly-induced action.
We recall that a WZ action is constructed by solving the constraints coming from the conformal anomaly and
differs from the effective action computed using perturbation theory and integrating out the matter fields.
For instance, direct computations of several correlators \cite{Giannotti:2008cv,Armillis:2009pq} have shown that these
are in agreement with the expression predicted by the non-local anomaly action proposed by Riegert \cite{Riegert:1984kt}.
In this respect the WZ and the Riegert's action show significantly different features.
A WZ form of the non-local anomaly action is regained from Riegert's expression only at the cost of
sacrificing covariance, by the choice of a fiducial metric \cite{Antoniadis:1991fa}. However, the WZ action becomes generally
covariant at the price of introducing one extra field, the dilaton, which plays a key role in order to extract information on some significant
implications of the anomaly, as we are going to show.
We will derive this action from the Weyl-gauging of the counterterms for CFT's in dimensional regularization \cite{Mazur:2001aa}.
We will be using the term {\em renormalized action} to denote the anomaly-induced action which is given by the sum
of the Weyl-invariant (non anomalous) terms, denoted by $\Gamma_0$, and of the counterterms
$\Gamma_{\textrm{Ct}}$ which one extracts in ordinary perturbation theory \cite{Duff:1977ay}. Explicitly
\begin{equation}
\label{oneloop}
\Gamma_{\textrm{ren}}[g, \rho]=\Gamma_0[g,\rho] + \Gamma_{\textrm{Ct}}[g],
\end{equation}
where the dependence on the dilaton $\rho$ in $\Gamma_0$ is generated by the Weyl-gauging of diffeomorphism invariant functionals of
the metric, discussed in the previous section.
This action correctly reproduces the anomaly, which is generated by the Weyl variation of $\Gamma_{\textrm{Ct}}$.
The change in the notation with respect to (\ref{RenW}) requires some comment.
In (\ref{RenW}) the generating functional $\mathcal{W}$ was used, whereas here we are replacing it by the effective
action $\Gamma$. This notation is more general and better suits the conventions that are met in the literature.
In fact, even though we are dealing only with exactly conformal field theories, we know that CFT's are believed to be
both the ultraviolet and infrared limits of any RG flow; but despite the fact that in the two phases the mathematics is the same,
yet the physics is quite different, because down in the infrared regime a lot of the microspcopic degrees of freedom
showing up at high energies have been integrated out.
This is why we choose to switch from the rather generic generating functional $\mathcal W$ to the
more common $\Gamma$, which typically indicates effective actions.
Of course, the different numbers of degrees of freedom in the two CFT's show up in the different values of the scalar coefficients
$c$ and $a$ in the trace anomaly equation (\ref{Ch4anom}).
We have also hidden the $1/\epsilon$ pole of the counterterm inside $\Gamma_{Ct}$, just for convenience, so that the following relation
holds (see eq. (\ref{RenW}))
\begin{equation}
\Gamma_{Ct}[g] = -\frac{\mu^{-\epsilon}}{\bar\epsilon}\, \mathcal{W}_{Ct}[g]\, .
\end{equation}
The cocycle condition summarizes the response of the functional $\Gamma$ under a Weyl-gauging of the metric, which is
given by (see also eq. (\ref{Ch4Compensate}))
\begin{equation}
\label{fet}
g_{\mu\nu} \to \hat{g}_{\mu\nu} = g_{\mu\nu}\, e^{- 2\, \rho/\Lambda} \, .
\end{equation}
In particular, we can define the Weyl-gauged renormalized effective action,
\begin{equation}
\hat\Gamma_{\textrm{ren}}[g,\rho] \equiv \Gamma_{0}[g,\rho] + \Gamma_{\textrm{Ct}}[\hat{g}]\, ,
\label{HattedGamma}
\end{equation}
and the WZ action, $\Gamma_{WZ}$, is identified from the relation
\begin{equation}
\label{coc}
\hat\Gamma_{\textrm{ren}}[g,\rho] = \Gamma_{\textrm{ren}}[g,\rho] - \Gamma_{\textrm{WZ}}[g,\rho] \, .
\end{equation}
We recall that the Weyl transformation of the field-enlarged system is realised by
\begin{eqnarray}
g'_{\mu\nu} &=& g_{\mu\nu}\, e^{2\sigma}\, , \nonumber \\
\rho' &=& \rho + \sigma\, .
\label{WeylVar}
\end{eqnarray}
The defining condition of the WZ action is that its variation equals the trace anomaly, i.e.
\begin{equation}
\delta_W \Gamma_{WZ}[g,\rho] = \int d^dx\, \sqrt{g}\, \sigma\, \mathcal{A}[g]\, .
\label{WZVar}
\end{equation}
We will exploit explicitly the relation (\ref{coc}) in $4$ and $6$ dimensions in the main text and for $d=2$ in appendix \ref{Ch42D}.
It is possible to prove it in complete generality for arbitrary even dimensions, under very general hypotheses.
The proof combines cohomological arguments and dimensional regularization, explicitly deriving the algorithm to be used below,
showing that it automatically generates effective actions satifying the WZ consistency condition for the Weyl group.
The proof lies beyond the goal of this chapter, so that we refer to the original paper by Mazur and Mottola for the details \cite{Mazur:2001aa}.
Here we limit to a sketchy derivation of the result. First, we observe that the Weyl-gauged the effective action (\ref{HattedGamma})
is naturally expected to consist of the non gauged action plus a function of the dilaton and the metric, say $\Gamma_{1}[g,\rho]$
\begin{equation}
\Gamma_{ren}[\hat{g}] = \Gamma_{ren}[g] + \Gamma_{1}[g,\rho] \, .
\end{equation}
Recalling the discussion of section \ref{CountAnom}, where it was shown that the Weyl variation of the renormalized generating functional
in dimensional regularization is the trace anomaly, we can write this condition,
with the generating functional replaced by the effective action $\Gamma$, in the equivalent integral form
\begin{equation}
\delta_W \Gamma_{ren}[g] = \int d^dx\, \sqrt{g}\, \mathcal{A}[g]\, ,
\label{DeltaGammaRen}
\end{equation}
which is the same constraint satisfied by the WZ action, (\ref{WZVar}).
Observing that any functional of the hatted metric $\hat{g}_{\mu\nu}$ is Weyl invariant by construction\footnote{This observation
is definitely true for functionals that are defined in the physical dimensions. For the counterterms in dimensional regularization
some care is needed. For details, see \cite{Mazur:2001aa}}, it follows that the Weyl variation of the hatted renormalized effective
action (\ref{HattedGamma}) is expected to vanish, which, taking (\ref{DeltaGammaRen}) into account, is
explicitly written as
\begin{equation}
\int d^dx\, \sqrt{g}\, \mathcal{A}[g] + \delta_W \Gamma_{1}[g,\rho] = 0\, .
\label{WanishWeylGamma}
\end{equation}
Considering (\ref{WanishWeylGamma}) together with (\ref{WZVar}), it is natural to identify $\Gamma_{1}$
with the WZ effective action,
\begin{equation}
\Gamma_{1}[g,\rho] \equiv - \Gamma_{WZ}[g,\rho]\, ,
\label{IdentifyWZ}
\end{equation}
thus obtaining (\ref{coc}).
Of course, our argument is of variational nature, so we are allowed to write (\ref{IdentifyWZ}) modulo Weyl invariant terms.
Nevertheless, these can always be absorbed into the part of the effective action giving the kinetic terms for the dilaton,
which were reviewed above.
\section{The WZ effective action for $d=4$}\label{WZ4}
Having reviewed the structure of the Weyl invariant operators in the dilaton effective action $\Gamma_{\textrm{ren}}$ and the cocycle
condition, we move on to the hard core of this chapter and construct the WZ effective actions, which are the mean we will exploit
to derive the recursive algorithm for the computation of traced EMT correlation functions of arbitrary order.
\subsection{The counterterms in $4$ dimensions} \label{Ch3Counterterms}
One standard approach followed in the derivation of the WZ anomaly action is the Noether method, in which $\rho$ is linearly coupled
to the anomaly. Further terms are then introduced in order to correct for the non invariance under Weyl transformations of the
anomaly functional itself. This approach, reviewed in appendix \ref{Ch3WessZumino}, does not make transparent the functional
dependence of the WZ action on the Weyl invariant metric $\hat{g}_{\mu\nu}$, which motivates our analysis.
Here, instead, we proceed with a construction of the same effective action by applying the Weyl-gauging procedure to the renormalized
effective action. For definiteness, in the following we briefly review the discussion of the connection of the counterterms in to the trace anomaly
and the scheme dependence of the latter.
Following the discussion in \cite{Duff:1977ay}, we start by introducing the counterterm action
\begin{equation}\label{Ch3CounterAction4}
{\Gamma}_{\textrm{Ct}}[g] =
- \frac{\mu^{-\epsilon}}{\epsilon}\int d^d x\, \sqrt{g}\, \bigg( \beta_a F + \beta_b G\bigg) \, , \quad \epsilon = 4 - d \, ,
\end{equation}
where $\mu$ is a regularization scale. It is this form of ${\Gamma}_{\textrm{Ct}}$, which is part of
$\Gamma_{\textrm{ren}}$ to induce the anomaly condition
\begin{equation}
\label{Ch3qt}
\frac{2}{\sqrt{g}}g_{\mu\nu}\frac{\delta{\Gamma_{\textrm{ren}}}[g]}{\delta g_{\mu\nu}}\bigg|_{d\rightarrow 4} =
\frac{2}{\sqrt{g}}g_{\mu\nu}\frac{\delta{\Gamma_{\textrm{Ct}}}[g]}{\delta g_{\mu\nu}}\bigg|_{d\rightarrow 4} =
\mathcal{A}[g].
\end{equation}
In (\ref{Ch3qt}) we have exploited the Weyl invariance of the non anomalous action $\Gamma_0[g]$
\begin{equation} \label{Ch3BareActionInvariance}
g_{\mu\nu}\frac{\delta{\Gamma}_0[g]}{\delta g_{\mu\nu}}\bigg|_{d\rightarrow 4} = 0 \, ,
\end{equation}
with the anomaly generated entirely by the counterterm action ${\Gamma}_{\textrm{Ct}}[g]$. This follows from the well known relations
\begin{eqnarray} \label{Ch3NTracesCTF}
\frac{2}{\sqrt{g}}\, g_{\mu\nu}\, \frac{\delta}{\delta g_{\mu\nu}}\, \int d^d x\,\sqrt{g}\, F
&=&
-\epsilon \, \left(F - \frac{2}{3}\, \Box R\right)\, ,
\\
\label{Ch3NTracesCTG}
\frac{2}{\sqrt{g}}\, g_{\mu\nu}\, \frac{\delta}{\delta g_{\mu\nu}}\, \int d^d x\, \sqrt{g}\, G
&=&
-\epsilon \, G \,,
\end{eqnarray}
which give
\begin{equation} \label{Ch3TraceAnomaly4d}
\left\langle T \right\rangle =
\frac{2}{\sqrt{g}}\, g_{\mu\nu}\, \frac{\delta{\Gamma}_{\textrm{Ct}}[g]}{\delta g_{\mu\nu}}\bigg|_{d\rightarrow 4} =
\beta_a\, \left( F - \frac{2}{3}\, \Box R\right) + \beta_b\, G \, .
\end{equation}
The $\Box R$ term in eq. ~(\ref{Ch3NTracesCTF}) is prescription dependent and can be avoided if the
$F$-counterterm is chosen to be conformal invariant in $d$ dimensions, i.e. using the square $F_d$ of the Weyl tensor in $d$
dimensions (see appendix (\ref{Geometrical})),
\begin{equation} \label{Ch3FdCounterterm}
{\Gamma}^{d}_{\textrm{Ct}}[g] =
- \frac{\mu^{-\epsilon}}{\epsilon}\, \int d^d x\, \sqrt{g}\, \bigg( \beta_a F_d + \beta_b G\bigg)\, .
\end{equation}
In fact, expanding (\ref{Ch3FdCounterterm}) around $d=4$ and computing the $O(\epsilon)$
contribution to the vev of the traced EMT we find
\begin{eqnarray}
\int d^d x\,\sqrt{g}\, F_d
&=& \int d^d x\,\sqrt{g}\, \bigg[ F - \epsilon\, \bigg( R^{\alpha\beta}R_{\alpha\beta} - \frac{5}{18}\, R^2 \bigg)
+ O\big(\epsilon^2\big) \bigg] \, , \\
\frac{2}{3}\, \Box R
&=&
\frac{2}{\sqrt{g}}\, g_{\mu\nu}\, \frac{\delta}{\delta g_{\mu\nu}}\,
\int d^4 x\, \sqrt{g}\, \left( R^{\alpha\beta} R_{\alpha\beta} - \frac{5}{18} R^2 \right) \, .
\end{eqnarray}
These formulae, combined with (\ref{Ch3NTracesCTF}), give
\begin{equation} \label{Ch3NTracesCT2}
\frac{2}{\sqrt{g}}\, g_{\mu\nu}\, \frac{\delta}{\delta g_{\mu\nu}}\, \int d^d x\,\sqrt{g}\, F_d
= - \epsilon\, F + O\big(\epsilon^2\big) \, ,
\end{equation}
in which the $\square R$ term is now absent.
In general, one may want to vary arbitrarily the coefficient in front of the $\Box R$ anomaly in (\ref{Ch3TraceAnomaly}).
This can be obtained by the inclusion of the counterterm
\begin{equation} \label{Ch3ll}
\beta_{\textrm{fin}}\, \int d^4x\, \sqrt{g}\, R^2\, ,
\end{equation}
where $\beta_{\textrm{fin}}$ is an arbitrary parameter, and the subscript $\textrm{fin}$ stands for "finite",
given that (\ref{Ch3ll}) is just a finite, prescription-dependent contribution.
In fact, the relation
\begin{equation} \label{Ch3LocalAnomaly}
\frac{2}{\sqrt{g}}\, g_{\mu\nu}\, \frac{\delta}{\delta g_{\mu\nu}}\, \int d^4 x\, \sqrt{g}\, R^2 = 12\, \Box R
\end{equation}
allows to modify at will the coefficient in front of $\square R$ in the anomaly functional. This is obtained by adding the finite
contribution (\ref{Ch3ll}) to the action of the theory and by tuning appropriately the coefficient $\beta_{\textrm{fin}}$.
When (\ref{Ch3ll}) is present, the overall counterterm is
\begin{equation}
{\Gamma}_{\textrm{Ct}}[g] + \beta_{\textrm{fin}}\, \int d^4x\, \sqrt{g}\, R^2
\end{equation}
and the modified trace anomaly equation becomes
\begin{equation}
\left\langle T \right\rangle =
\beta_a\, F + \beta_b\, G -\frac{2}{3}\, \bigg( \beta_a -18\, \beta_{\textrm{fin}} \bigg)\, \Box R \, .
\label{Ch3ModifiedTraceAnomalies}
\end{equation}
Nevertheless, the contribution (\ref{Ch3ll}) breaks the conformal symmetry of the theory.
This implies that the only modification of the effective action which modifies the coefficient of $\square R$ in the trace anomaly
and is, at the same time, consistent with conformal symmetry is the replacement of $F$ with $F_d$,
removing the $\Box R$ anomaly altogether.
\subsection{Weyl-gauging of the counterterms in $4$ dimensions}\label{Ch3GaugeCount}
At this point we illustrate the practical implementation of the gauging procedure on the renormalized effective action.
It is natural to expand the gauged counterterms in a double power series with respect to $\epsilon = 4-d$
and $\kappa_{\Lambda}\equiv1/\Lambda$ around $(\epsilon , \kappa_\Lambda )=( 0,0)$.
Their formal expansions are
\begin{eqnarray}
- \frac{1}{\epsilon}\,\int d^dx \, \sqrt{\hat{g}} \, \hat{F}\,
&=&
- \frac{1}{\epsilon}\, \int d^dx \sum_{i,j=0}^{\infty}\frac{1}{i!j!}\, \epsilon^i\,\left( \kappa_\Lambda\right)^j\,
\frac{\partial^{i+j}\left(\sqrt{\hat g}\, \hat F\, \right)}{\partial \epsilon^i\,\partial \kappa_\Lambda^{j}}\, ,
\nonumber \\
- \frac{1}{\epsilon}\,\int d^dx \, \sqrt{\hat{g}} \, \hat{G}
&=&
- \frac{1}{\epsilon}\, \int d^dx \sum_{i,j=0}^{\infty}\frac{1}{i!j!}\, \epsilon^i\,\left( \kappa_\Lambda\right)^j\,
\frac{\partial^{i+j}\left(\sqrt{\hat g}\, \hat G \right)}{\partial \epsilon^i\,\partial \kappa_\Lambda^{j}}\, .
\label{Ch3PreGauging}
\end{eqnarray}
It is clear that only the $O(\epsilon)$ contributions are significant, as every higher order term would still be $O(\epsilon)$ at least
after the division by $\epsilon$, therefore vanishing for $d \rightarrow 4$. \\
On the other hand, we observe that in the gauged Riemann tensor there are no more than two dilaton fields (see appendix \ref{Geometrical}).
But $\kappa_\Lambda$ can appear in the gauged counterterms only through the dilaton, so that the conditions
\begin{equation}
\frac{\partial^{n} \left(\sqrt{\hat{g}} \, \hat{F}\right)}{\partial\kappa_\Lambda^{n}} = O(\epsilon^2)\, ,
\qquad
\frac{\partial^{n} \left(\sqrt{\hat{g}} \, \hat{G}\right)}{\partial\kappa_\Lambda^{n}} = O(\epsilon^2) \, ,
\quad n \geq 5
\end{equation}
are found to hold.
Finally, in (\ref{Ch3PreGauging}) there are terms which are $O(1/\epsilon)$ and deserve special attention.
Taking, for example, the first of (\ref{Ch3PreGauging}), they are $F/\epsilon$ plus something more.
But everything differing from $F/\epsilon$ is found to vanish after proper integrations by parts and the same holds for the other
gauged counterterm.
Thus we have obtained the intermediate results
\begin{eqnarray}
\label{Ch3effe}
- \frac{\mu^{-\epsilon}}{\epsilon}\,\int d^dx \, \sqrt{\hat{g}} \, \hat{F} \,
&=&
- \frac{\mu^{-\epsilon}}{\epsilon}\,\int d^dx \, \sqrt{g}\, F
+\int d^4x \, \sqrt{g}\, \bigg\{\frac{1}{\Lambda}\,
\bigg( - \rho\, F - \frac{4}{3}\, R\,\Box\rho + 4\, R^{\alpha\beta}\, \nabla_\alpha\partial_\beta\rho \bigg)
\nonumber \\
&&
+\, \frac{2}{\Lambda^2}\, \bigg( 2\, R^{\alpha\beta}\,\partial_\alpha\rho\,\partial_\beta\rho - \frac{R}{3}\, \left(\partial\rho\right)^2
+ \left(\Box\rho\right)^2 - 2\, \nabla_\beta\partial_\alpha\rho\, \nabla^\beta\partial^\alpha\rho
\bigg)
\nonumber \\
&&
-\, \frac{8}{\Lambda^3}\, \partial^\alpha\rho\,\partial^\beta\rho\, \nabla_\beta\partial_\alpha\rho
- \frac{2}{\Lambda^4}\, \left(\left(\partial\rho\right)^2\right)^2\, \bigg\}\, ,
\end{eqnarray}
\begin{eqnarray}
\label{Ch3gi}
- \frac{\mu^{-\epsilon}}{\epsilon}\,\int d^dx \, \sqrt{\hat{g}} \, \hat{G} \,
&=&
- \frac{\mu^{-\epsilon}}{\epsilon}\,\int d^dx \, \sqrt{g}\, G
+ \int d^4x \, \sqrt{g}\, \bigg\{ \frac{1}{\Lambda}\,
\bigg( - \rho\, G - 4\, R\, \Box \rho + 8\, R^{\alpha\beta}\, \nabla_\beta \partial_\alpha \rho \bigg)
\nonumber \\
&& \hspace{-5mm}
+\, \frac{2}{\Lambda^2}\, \bigg[ 2\, R\, \rho\,\Box\rho + R\, \left(\partial\rho\right)^2
+ 4\, R^{\alpha\beta}\, \bigg(\partial_\alpha\rho\,\partial_\beta\rho - \rho\,\nabla_\beta \partial_\alpha\rho \bigg)
+ 6\, \left( \Box\rho \right)^2 - 6\, \nabla^\beta\partial^\alpha\rho\,\nabla_\beta\partial_\alpha\rho \bigg]
\nonumber \\
&& \hspace{-5mm}
-\, \frac{4}{\Lambda^3}\, \bigg( 2\, R^{\alpha\beta}\,\rho\, \partial_\alpha\rho\, \partial_\beta\rho + 2\, \rho\, \left(\Box\rho\right)^2 +
5\, \left(\partial\rho\right)^2\, \Box\rho + 6\,\partial^\alpha\rho\,\partial^\beta\rho\,\nabla_\beta\partial_\alpha\rho -
2\, \rho\, \nabla^\beta\partial^\alpha\rho\, \nabla_\beta\partial_\alpha\rho \bigg)
\nonumber \\
&& \hspace{-5mm}
+\, \frac{2}{\Lambda^4}\, \bigg(
4\, \rho\, \left(\partial\rho\right)^2\, \Box\rho +
3\, \left( \left(\partial\rho\right)^2 \right)^2 +
8\, \rho\, \partial^\alpha\rho\,\partial^\beta\rho\, \nabla_\beta\partial_\alpha \rho \bigg)\bigg\} \, .
\end{eqnarray}
The expressions above can be simplified using integrations by parts and the identity
for the commutator of covariant derivatives of a vector,
\begin{equation}
[\nabla_\mu,\nabla_\nu]\, v_\rho = {R^\lambda}_{\rho\mu\nu}\,v_\lambda \, .
\end{equation}
After these manipulations we find that the Weyl-gauging of the counterterms gives
\begin{eqnarray}
- \frac{\mu^{-\epsilon}}{\epsilon}\,\int d^dx \, \sqrt{\hat{g}} \, \hat{F} \,
&=&
- \frac{\mu^{-\epsilon}}{\epsilon}\, \int d^dx \, \sqrt{g} \, F +
\int d^4 x \, \sqrt{g}\, \bigg[ - \frac{\rho}{\Lambda}\, \bigg( F - \frac{2}{3} \Box R \bigg)
- \frac{2}{\Lambda^2}\, \bigg( \frac{R}{3}\, \left(\partial\rho\right)^2 + \left(\Box \rho \right)^2 \bigg)
\nonumber \\
&& \hspace{50mm}
+ \frac{4}{\Lambda^3}\, \left(\partial\rho\right)^2\,\Box\rho \,
- \frac{2}{\Lambda^4}\, \left(\left(\partial\rho\right)^2\right)^2 \bigg] \, ,
\label{Ch3GaugingF} \\
- \frac{\mu^{-\epsilon}}{\epsilon}\,\int d^dx \, \sqrt{\hat{g}} \, \hat{G}
&=&
- \frac{\mu^{-\epsilon}}{\epsilon}\, \int d^dx \, \sqrt{g} \, G +
\int d^4 x \, \sqrt{g}\, \bigg[ - \frac{\rho}{\Lambda}\, G
+ \frac{4}{\Lambda^2}\, \left( R^{\alpha\beta} - \frac{R}{2}\,g^{\alpha\beta} \right)\, \partial_\alpha\rho\,\partial_\beta\rho
\nonumber \\
&& \hspace{50mm}
+ \, \frac{4}{\Lambda^3}\, \left(\partial\rho\right)^2\, \Box \rho
- \frac{2}{\Lambda^4}\, \left(\left(\partial\rho\right)^2 \right)^2 \bigg]\, .
\label{Ch3GaugingG}
\end{eqnarray}
As a consistency check, one can apply a Weyl transformation to (\ref{Ch3GaugingF}) and (\ref{Ch3GaugingG}),
and see that it gives zero, which is the basic requirement for a WZ action.
Recalling the defining relation for the WZ action, (\ref{coc}), we can finally write
\begin{eqnarray}
\Gamma_{WZ}[g,\rho]
&=&
\int d^4x\, \sqrt{g}\, \bigg\{ \beta_a\, \bigg[ \frac{\rho}{\Lambda}\, \bigg( F - \frac{2}{3} \Box R \bigg) +
\frac{2}{\Lambda^2}\, \bigg( \frac{R}{3}\, \left(\partial\rho\right)^2 + \left( \Box \rho \right)^2 \bigg) -
\frac{4}{\Lambda^3}\, \left(\partial\rho\right)^2\,\Box \rho +
\frac{2}{\Lambda^4}\, \left(\partial\rho\right)^4 \bigg] \nonumber \\
&&
\hspace{15mm}
+\, \beta_{b}\, \bigg[ \frac{\rho}{\Lambda}\,G -
\frac{4}{\Lambda^2}\, \bigg( R^{\alpha\beta} - \frac{R}{2}\,g^{\alpha\beta} \bigg)\, \partial_\alpha\rho\, \partial_\beta\rho -
\frac{4}{\Lambda^3} \, \left(\partial\rho\right)^2\,\Box \rho +
\frac{2}{\Lambda^4}\, \left( \left(\partial\rho\right)^2 \right)^2 \bigg]\bigg\} \, .
\label{Ch3Effective4d}
\end{eqnarray}
Notice that the ambiguity in the choice of the Weyl tensor discussed above - i.e. between F and $F_d$ - implies that no dilaton
vertex is expected to emerge from the gauging of the $F_d$-counterterm, as it is conformal invariant.
This is indeed the case and we find the relation
\begin{eqnarray} \label{Ch3GaugingFd}
&& \hspace{-8 mm}
-\frac{\mu^{-\epsilon}}{\epsilon}\,\int d^d x \, \sqrt{\hat{g}} \, \hat{F}_d \, =
-\frac{\mu^{-\epsilon}}{\epsilon}\,\int d^d x \, \sqrt{g} \, F_{d}
- \int d^4 x \,\sqrt{g}\, \frac{\rho}{\Lambda}\, F\, ,
\end{eqnarray}
that modifies the structure of the WZ action eliminating from (\ref{Ch3Effective4d}) all the terms multiplying $\beta_a$
but $\left(\rho/\Lambda\right)\, F$.
Finally we remark that, in the case in which a finite counterterm of the kind (\ref{Ch3ll}) is present,
the formulae of this section are modified according to the simple prescription (see eq. (\ref{Ch3ModifiedTraceAnomalies})),
\begin{equation} \label{Ch3LocalToBeta}
\beta_a \rightarrow \beta_a - 18\, \beta_{\textrm{fin}}\, ,
\end{equation}
as it is possibile to render all the quantum effective action Weyl invariant. This is obtained, as discussed above,
by the Weyl-gauging of the complete counterterm
\begin{equation}
\Gamma_{\textrm{Ct}}[g] =
-\frac{\mu^{-\epsilon}}{\epsilon}\, \int d^dx\, \sqrt{g}\, \bigg( \beta_a\,F + \beta_b\,G \bigg) +
\beta_{\textrm{fin}}\, \int d^4x\, \sqrt{g}\, R^2\, .
\end{equation}
In this case the compensating WZ action for $\int d^4x\, \sqrt{g}\, R^2$ can be generated by the relation
\begin{eqnarray}
\int d^4x\, \sqrt{\hat g}\, \hat{R}^2
&=&
\int d^4x\, \sqrt{g}\, R^2 +
18\, \int d^4 x \, \sqrt{g}\, \bigg[ - \frac{2}{3}\, \frac{\rho}{\Lambda}\, \Box R
+ \frac{2}{\Lambda^2}\, \bigg( \frac{R}{3}\, \left(\partial\rho\right)^2 + \left(\Box \rho \right)^2 \bigg)
\nonumber \\
&& \hspace{47mm}
- \frac{4}{\Lambda^3}\, \left(\partial\rho\right)^2\,\Box\rho \,
+ \frac{2}{\Lambda^4}\, \left(\left(\partial\rho\right)^2\right)^2 \bigg] \, .
\end{eqnarray}
Comparing the result given above with (\ref{Ch3GaugingF}), eq. (\ref{Ch3LocalToBeta}) follows immediately.
\section{The WZ effective action for $d=6$}\label{WZ6}
We repeat step by step the procedure outlined for the $4$-dimensional case, discussing the possible choices of the counterterms
in $6$ dimensions in full generality, highlighting the details of the difference between the choice of the Weyl invariants
$\sqrt{g}\, I_i$ and their $d$-dimensional counterparts, $\sqrt{g}\, I^d_i$, in particular deriving the finite renormalization
distinguishing them. These results are then used in the study of the various possible WZ actions.
\subsection{The counterterms in $6$ dimensions} \label{Ch4Counterterms}
As we have discussed above, we construct the effective action by applying the Weyl-gauging procedure to the renormalized effective action,
which breaks scale invariance via the anomaly. First we must introduce the $1$-loop counterterm action,
which is given by the integrals of all the possible Weyl invariants and of the Euler density, each continued to $d$ dimensions,
\begin{equation}\label{Ch4CounterAction6}
{\Gamma}_{\textrm{Ct}}[g] =
- \frac{\mu^{-\epsilon}}{\epsilon}\int d^d x\, \sqrt{g}\, \bigg( \sum_{i=1}^{3} c_i\, I_i + a E_6 \bigg) \, ,
\quad \epsilon = 6 - d \, ,
\end{equation}
where $\mu$ is a regularization scale.
This form of ${\Gamma}_{\textrm{Ct}}$ induces the anomaly relation
\begin{equation} \label{Ch4AnomalyCondition}
\frac{2}{\sqrt{g}}g_{\mu\nu}\frac{\delta{\Gamma_{\textrm{ren}}}[g]}{\delta g_{\mu\nu}}\bigg|_{d\rightarrow 6} =
\frac{2}{\sqrt{g}}g_{\mu\nu}\frac{\delta{\Gamma_{\textrm{Ct}}}[g]}{\delta g_{\mu\nu}}\bigg|_{d\rightarrow 6}=
\mathcal{A}[g] \, .
\end{equation}
where we have exploited once again the Weyl invariance of the non anomalous action $\Gamma_0[g]$ in $6$ dimensions
\begin{equation} \label{Ch4BareActionInvariance}
g_{\mu\nu}\frac{\delta{\Gamma}_0[g]}{\delta g_{\mu\nu}}\bigg|_{d\rightarrow 6} = 0, \,
\end{equation}
with the anomaly generated entirely by the counterterm action ${\Gamma}_{\textrm{Ct}}[g]$, due to the relations
\begin{eqnarray} \label{Ch4NTracesCTI}
\frac{2}{\sqrt{g}}\, g_{\mu\nu}\, \frac{\delta}{\delta g_{\mu\nu}}\, \int d^d x\,\sqrt{g}\, I_i
&=&
-\epsilon \, \bigg( I_i + \nabla_\mu J^\mu_i \bigg) \, ,
\\
\label{Ch4NTracesCTE}
\frac{2}{\sqrt{g}}\, g_{\mu\nu}\, \frac{\delta}{\delta g_{\mu\nu}}\, \int d^d x\, \sqrt{g}\, E_6
&=&
-\epsilon \, E_6 \, ,
\end{eqnarray}
so that from (\ref{Ch4AnomalyCondition}) we find
\begin{equation} \label{Ch4TraceAnomaly6d}
\left\langle T \right\rangle =
\frac{2}{\sqrt{g}}\, g_{\mu\nu}\, \frac{\delta{\Gamma}_{\textrm{Ct}}[g]}{\delta g_{\mu\nu}}\bigg|_{d\rightarrow 4} =
\sum_{i=1}^3 c_i\, \left( I_i + \nabla_\mu J_i^\mu \right) + a E_6 \, .
\end{equation}
The explicit expressions of the derivative terms was given in section \ref{RecConv}
and can be obtained through the functional variations listed in appendix \ref{Ch4Geometrical3}.
These terms above are renormalization prescription dependent and are not present if, instead of the counterterms
$\sqrt{g}\, I_{i}$, one chooses scalars that are conformal invariant in $d$ dimensions, i.e. the $I^d_i$'s defined in
appendix \ref{Geometrical}. Notice that the inclusion of $d$-dimensional counterterms simplifies considerably the computation of the
dilaton WZ action, as shown in \cite{Baume:2013ika}. In fact, in this scheme, the contribution of the $I^d_i$'s to the same action
is just linear in the dilaton field and can be derived from the counterterm
\begin{equation} \label{Ch4IdCounterterm}
{\Gamma}^{d}_{\textrm{Ct}}[g] = - \frac{\mu^{-\epsilon}}{\epsilon}\, \int d^d x\, \sqrt{g}\,
\bigg( \sum_{i=1}^{3} c_i\, I^d_i + a E_6 \bigg).
\end{equation}
It can be explicitly checked that by expanding (\ref{Ch4IdCounterterm}) around $d=6$ and computing the order $O(\epsilon)$ contribution
to the vev of the traced EMT one obtains the relation
\begin{equation} \label{Ch4NTracesCT1}
\frac{2}{\sqrt{g}}\, g_{\mu\nu}\, \frac{\delta}{\delta g_{\mu\nu}}\, \int d^d x\,\sqrt{g}\, I^d_i =
- \epsilon\, I_i + O\big( \epsilon^2 \big). \,
\end{equation}
In this simplified scheme, it is possible to give the structure of the WZ action in any even dimension \cite{Baume:2013ika},
just by adding to the contribution of such invariants the one coming from the Euler density $E_d$, being the
total derivative terms $\nabla_\mu J^\mu_i$ absent.
\subsection{General scheme-dependence of the trace anomaly in $6$ dimensions}\label{Ch4SchemDep}
In this section we establish a connection between the two renormalization schemes used to derive the dilaton WZ action, with the inclusion of
invariant counterterms of $B$ type which are either $d$ or $6$-dimensional, in close analogy with the $4$-dimensional case.
For $d=6$ we proceed in a similar way. We expand the $d$-dimensional counterterms around $d=6$ to identify the finite contributions as
\begin{equation} \label{Ch4SeriesEps}
I_i^d = I_i + (d-6)\, \frac{\partial I^d_i}{\partial d}\bigg|_{d=6} = I_i - \epsilon\, \frac{\partial I^d_i}{\partial d}\bigg|_{d=6}\, .
\end{equation}
Using (\ref{Ch4SeriesEps}) in the $d$-dimensional counterterms, we have
\begin{equation}
-\frac{1}{\epsilon}\, \int d^d x\, \sqrt{g}\, I^d_i =
-\frac{1}{\epsilon} \int d^d x\, \sqrt{g}\, I_i + \int d^dx\, \sqrt{g}\, \frac{\partial I^d_i}{\partial d}\bigg|_{d=6}\, .
\end{equation}
This implies, due to (\ref{Ch4NTracesCTI}) and (\ref{Ch4NTracesCT1}), that
\begin{eqnarray}
-\frac{1}{\epsilon}\, \frac{2}{\sqrt{g}}\, g_{\mu\nu}\, \frac{\delta}{\delta g_{\mu\nu}}\, \int d^d x\,\sqrt{g}\, I_i
&=&
I_i - \frac{2}{\sqrt{g}}\, g_{\mu\nu}\,
\frac{\delta}{\delta g_{\mu\nu}}\, \int d^dx\, \sqrt{g}\, \frac{\partial I^d_i}{\partial d}\bigg|_{d=6}
\end{eqnarray}
and hence we conclude that a finite counterterms which can account for the $i$-th total derivative term in the trace anomaly
\begin{eqnarray}
\frac{2}{\sqrt{g}}\, g_{\mu\nu}\, \frac{\delta}{\delta g_{\mu\nu}}\,
\int d^dx\, \sqrt{g}\, \frac{\partial I^d_i}{\partial d}\bigg|_{d=6}
&=&
- \nabla_\mu J^\mu_i\, .
\end{eqnarray}
This clearly identifies the terms that we can add to (\ref{Ch4CounterAction6}) in order to arbitrarily vary the coefficients $c_i$
in (\ref{Ch4TraceAnomaly6d}). They are given by the derivatives of the $d$-dimensional terms $I^d_i$ evaluated at $d=6$,
linearly combined with arbitrary coefficients $c'_i$
\begin{equation}
\Gamma'_{\textrm{Ct}}[g] =
- \frac{\mu^{-\epsilon}}{\epsilon}\int d^d x\, \sqrt{g}\, \bigg( \sum_{i=1}^{3} c_i\, I_i + a E_6 \bigg)
+ \int d^6x\, \sqrt{g}\, \sum_{i=1}^3 c'_i\, \frac{\partial I_i^d}{\partial d}\bigg|_{d=6}
\label{Ch4CounterAction6Mod}
\end{equation}
which gives the modified trace anomaly relation
\begin{eqnarray}
\left\langle T' \right\rangle \equiv
\frac{2}{\sqrt{g}}\, g_{\mu\nu}\, \frac{\delta \Gamma'_{\textrm{Ct}}[g]}{\delta g_{\mu\nu}}\bigg|_{d\rightarrow 4}=
\sum_{i=1}^3 c_i\, I_i + a E_6 + \sum_{i=1}^3 \left( c_i - c'_i \right)\, \nabla_\mu J_i^\mu \, .
\label{Ch4CounterAction6Mod1}
\end{eqnarray}
Then, the choice $c'_i=c_i$ in (\ref{Ch4CounterAction6Mod}) allows to move back to the scheme in which the local anomaly
contribution is not present.
We list the three local counterterms of (\ref{Ch4CounterAction6Mod}),
\begin{eqnarray}
\frac{\partial I^d_1}{\partial d}\bigg|_{d=6}
&=&
\frac{1}{16000}\, \bigg(-307\, K_1 + 3465\, K_2 - 540\, K_3 - 3750\, K_4 + 6000\, K_5 + 3000\, K_6\bigg) \, , \nonumber \\
\frac{\partial I^d_2}{\partial d}\bigg|_{d=6}
&=&
\frac{1}{4000}\, \bigg( -167\, K_1 + 1965\, K_2 - 540\, K_3 - 2750\, K_4 + 3000\, K_5 + 3000\, K_6\bigg) \, , \nonumber \\
\frac{\partial I^d_3}{\partial d}\bigg|_{d=6}
&=&
\frac{1}{500}\, \bigg(-18\,K_1 + 140\,K_2 - 90\, K_3 - 70\,K_9 + 500\,K_{10} - 250\,K_{11} + 25\,K_{12}
\nonumber \\
&& \hspace{11mm}
-\, 625\,K_{13} + 750\,K_{15}\bigg) \, .\nonumber\\
\end{eqnarray}
Finally, in general one might also be interested to generate an anomaly functional in which
the derivative terms appear in combinations that are different from those in eq. (\ref{Ch4anom6D}).
For this goal, one should use proper linear combinations of the $K_i$ according to the relations listed in (\ref{Ch4Geometrical3}).
\subsection{Weyl-gauging of the counterterms in $6$ dimensions}
At this point we have to Weyl-gauge the renormalized effective action.
Again, we expand the gauged counterterms in a double power series with respect to $\epsilon = 6-d$
and $\kappa_{\Lambda}\equiv1/\Lambda$ around $(\epsilon , \kappa_\Lambda )=( 0,0)$.
Denoting generically with $A$ either the Euler density $E_6$ or the three invariants $I_i$' s, the expansion takes the form
\begin{equation}
- \frac{1}{\epsilon}\,\int d^dx \, \sqrt{\hat{g}} \, \hat{A} =
- \frac{1}{\epsilon}\, \int d^dx \sum_{i,j=0}^{\infty}\frac{1}{i!j!}\, \epsilon^i\,\left( \kappa_\Lambda\right)^j\,
\frac{\partial^{i+j}\left(\sqrt{\hat g}\, \hat A\, \right)}{\partial \epsilon^i\,\partial \kappa_\Lambda^{j}}\, ,
\label{Ch4PreGauging}
\end{equation}
only the $O(\epsilon)$ contributions are significant, due to the $1/\epsilon$ factor in front of the counterterms.
On the other hand, similarly to the case in $4$ dimensions, the condition
\begin{equation}
\frac{\partial^{n} \left(\sqrt{\hat{g}} \, \hat A\right)}{\partial\kappa_\Lambda^{n}} = O(\epsilon^2) \, ,
\quad n \geq 7 \,
\end{equation}
holds, as the Euler density and the three conformal invariants are at most cubic in the Riemann tensor
and in its double covariant derivatives and, besides, there are no terms with more than two dilatons in the gauged Riemann tensor.
All the terms which are of $O(1/\epsilon)$ in (\ref{Ch4PreGauging}) and are different from $I_i$' s or, respectively, $E_6$
are found to vanish after some integrations by parts. Therefore, after gauging the counterterms we end up with the general result
\begin{equation}
- \frac{\mu^{-\epsilon}}{\epsilon}\,\int d^dx \, \sqrt{\hat{g}} \, \hat{A} \, =
- \frac{\mu^{-\epsilon}}{\epsilon}\, \int d^dx \, \sqrt{g} \, A + \Sigma_{A} + O(\epsilon) \, .
\label{Ch4ExpandCT}
\end{equation}
where each $\Sigma_{A}$ term is related to the corresponding specific invariant $A$.
For instance, if $A=I_1$, then the corresponding $\Sigma$ term is $\Sigma_1$, and so on for each of the $I_i$'s,
whereas for the contribution of the Euler density we have $A=E_6$ and $\Sigma_A=\Sigma_a$. \\
Their explicit expressions are
\begin{eqnarray}
&&
\Sigma_1 = \int d^6 x \, \sqrt{g}\, \bigg\{- \frac{\rho}{\Lambda}\, \bigg( I_1 + \nabla_\mu J_1^\mu \bigg) \nonumber \\
&&
+\, \frac{1}{\Lambda^2}\, \bigg[ \frac{3}{4}\, R^{\mu\lambda\kappa\alpha}\,
{R^\nu}_{\lambda\kappa\alpha}\, \partial_\mu \rho\, \partial_\nu \rho
- \frac{3}{40}\, R^{\mu\nu\lambda\kappa}\, R_{\mu\nu\lambda\kappa}\, \left(\partial \rho\right)^2
- \frac{3}{10}\, R\, \left(\nabla \partial \rho\right)^2 \nonumber \\
&&
+\, \frac{9}{4}\, R^{\mu\lambda\kappa\nu}\, R_{\lambda\kappa}\, \partial_\mu \rho\, \partial_\nu \rho
- 3\, R^{\mu\nu\lambda\kappa}\, \nabla_\nu \partial_\lambda \rho\, \nabla_\mu \partial_\kappa \rho
-\frac{57}{800}\, R^2\, \left(\partial \rho\right)^2 \nonumber \\
&&
-\, \frac{21}{16}\, R^{\mu\lambda}\, {R_\lambda}^\nu\, \partial_\mu \rho\, \partial_\nu \rho
- \frac{9}{4}\, R^{\mu\nu}\, \Box \rho\,\nabla_\mu \partial_\nu \rho
+ \frac{57}{160}\, R^{\mu\nu}\,R_{\mu\nu}\, \left(\partial \rho\right)^2 \nonumber \\
&&
+\, \frac{3}{2}\, R^{\mu\nu} \nabla^\lambda \partial_\mu \rho \, \nabla_\lambda \partial_\nu \rho
+ \frac{57}{80}\, R\, R^{\mu\nu} \partial_\mu \rho \, \partial_\nu \rho
+ \frac{57}{160}\, R\, \left(\Box\rho\right)^2 \bigg] \nonumber \\
&&
+\, \frac{1}{\Lambda^3}\, \bigg[
- \frac{7}{16}\, \left(\Box \rho \right)^3
+ \frac{3}{2}\, \left(\nabla \partial \rho \right)^2 \, \Box \rho
- 6\, R^{\mu\nu\lambda\kappa}\, \partial_\rho \rho\, \partial_\nu \rho\, \nabla_\mu \partial_\sigma \rho \nonumber \\
&&
+ 3\, R^{\mu\nu}\, \nabla_\lambda \partial_\nu \rho \, \partial_\mu \rho\, \partial^\lambda \rho
-\frac{9}{4}\, R^{\mu\nu}\, \partial_\mu \rho \,\partial_\nu \rho \,\Box \rho
- \frac{3}{5}\, R\, \partial^\mu \rho\, \partial^\nu \rho\, \nabla_\mu \partial_\nu \rho \bigg] \nonumber \\
&&
+\, \frac{1}{\Lambda^4}\, \bigg[
- \frac{3}{2}\, \left(\partial \rho \right)^2\, \left(\nabla \partial \rho \right)^2
-\, \frac{3}{8}\, \left(\partial \rho \right)^2\, \left(\Box \rho\right)^2
+ \frac{3}{4}\, \partial^\mu \left( \partial\rho \right)^2\, \partial_\mu \left( \partial\rho \right)^2
-\frac{3}{20}\, R\,\left(\partial \rho\right)^4
\bigg] \nonumber \\
&&
+\, \frac{1}{\Lambda^5}\, \frac{3}{2}\, \left(\partial \rho\right)^4\, \Box \rho
- \frac{\left(\partial \rho\right)^6}{\Lambda^6}\, \bigg\},
\label{Ch4WZI1}
\end{eqnarray}
for $I_1$,
\begin{eqnarray}
&&
\Sigma_2 = \int d^6 x \, \sqrt{g}\, \bigg\{- \frac{\rho}{\Lambda}\, \bigg( I_2 + \nabla_\mu J_2^\mu \bigg) \nonumber \\
&&
+ \frac{1}{\Lambda^2}\, \bigg[
3\, R^{\mu\lambda\kappa\alpha}\, {R^\nu}_{\lambda\kappa\alpha}\, \partial_\mu \rho\, \partial_\nu \rho
+ \frac{27}{40}\, R\, \left(\Box\rho\right)^2
- \frac{6}{5}\, R\, \left(\nabla \partial \rho\right)^2 - \frac{27}{200}\, R^2\, \left(\partial \rho \right)^2 \nonumber \\
&&
-\, \frac{3}{10}\, R^{\mu\nu\lambda\kappa}\, R_{\mu\nu\lambda\kappa}\, \left(\partial \rho\right)^2
+ 3\, R^{\mu\lambda\kappa\nu}\, R_{\lambda\kappa}\, \partial_\mu \rho\, \partial_\nu \rho
- \frac{15}{4}\, R^{\mu\lambda}\, {R_\lambda}^\nu\, \partial_\mu \rho\, \partial_\nu\rho \nonumber \\
&&
-\, 3\, R^{\mu\nu}\, \nabla_\nu \partial_\mu \rho\, \Box\rho + \frac{27}{40}\, R^{\mu\nu}\, R_{\mu\nu}\, \left(\partial\rho\right)^2
+ 6\, R^{\mu\nu}\, \nabla^\lambda \partial_\mu \rho\, \nabla_\lambda \partial_\nu \rho
+\frac{27}{20}\, R\, R^{\mu\nu}\, \partial_\mu\rho\, \partial_\nu \rho \bigg] \nonumber
\end{eqnarray}
\begin{eqnarray}
&&
+\, \frac{1}{\Lambda^3}\, \bigg[
\frac{11}{4}\, \left(\Box \rho \right)^3 - 6\, \left(\nabla \partial \rho \right)^2\, \Box \rho
- 8\, R^{\mu\nu}\, \nabla_\lambda\partial_\nu \rho \, \partial_\mu\rho \, \partial^\lambda \rho
- 6\, R^{\mu\nu}\, \nabla_\mu \partial_\nu \rho \left(\partial\rho\right)^2
\nonumber \\
&&
+\, 8\, R^{\mu\nu\lambda\kappa}\, \partial_\nu\rho\, \partial_\lambda\rho\, \nabla_\mu\partial_\kappa\rho
+ 5\, R^{\mu\nu}\, \partial_\mu \rho \, \partial_\nu \rho\, \Box \rho
+ \frac{18}{5}\, R\, \partial^\mu \rho\, \partial^\nu \rho\, \nabla_\mu \partial_\nu \rho
+ 3\, R\, \left(\partial \rho\right)^2\, \Box \rho \bigg] \nonumber \\
&&
+\, \frac{1}{\Lambda^4}\, \bigg[
6\, \left(\partial \rho\right)^2\, \left(\nabla \partial \rho\right)^2 - \frac{9}{2}\, \left(\partial \rho\right)^2\, \left(\Box \rho\right)^2
- 3\, \partial^\mu \left( \partial\rho \right)^2\, \partial_\mu \left( \partial\rho \right)^2 - \frac{3}{5}\, R\, \left(\partial \rho\right)^4
\bigg] \nonumber \\
&&
+\, \frac{6}{\Lambda^5}\, \left(\partial \rho\right)^4\, \Box \rho - \frac{4}{\Lambda^6}\, \left(\partial \rho\right)^6
\bigg\},
\label{Ch4WZI2}
\end{eqnarray}
for the second invariant $I_2$ and
\begin{eqnarray}
&&
\Sigma_3 =
\int d^6 x \, \sqrt{g}\, \bigg\{ - \frac{\rho}{\Lambda}\, \bigg( I_3 + \nabla_\mu J_3^\mu \bigg)
\nonumber \\
&&
+\, \frac{1}{\Lambda^2}\, \bigg[ -\frac{3}{25} R^2 \left(\partial \rho \right)^2
+ \frac{13}{10}R^{\mu\nu}R\, \partial_\mu \rho\, \partial_\nu \rho
-\frac{2}{5}\, R^{\mu\nu\lambda\kappa}\, R_{\mu\nu\lambda\kappa}\, \left(\partial \rho \right)^2 \nonumber \\
&&
+\frac{9}{10}\, R\, \left(\Box \rho\right)^2 - \frac{3}{10}\, R\, \partial^\mu \rho \,\Box \partial_\mu \rho
- \frac{12}{5}\, R\, \left(\nabla \partial \rho \right)^2 - 5\, R^{\mu\lambda}\, {R_\lambda}^\nu\, \partial_\mu \rho\, \partial_\nu \rho \nonumber \\
&&
+ 7\, R^{\mu\nu}\, \nabla_\mu \partial_\nu \rho\, \Box \rho - 9\, R^{\mu\nu}\, \partial_\mu \rho \,\Box \partial_\nu \rho\,
- \Box \rho \, \Box^2 \rho + \frac{2}{5}\, R^{\mu\nu}\, R_{\mu\nu}\, \left(\partial \rho \right)^2 \nonumber \\
&&
+ 8\, R^{\mu\nu} \,\nabla_\lambda \nabla_\nu \partial_\mu \rho\, \partial^\lambda \rho
+ 16\, R^{\mu\nu\lambda\kappa}\,\nabla_\nu \partial_\lambda \rho \, \nabla_\mu \partial_\kappa \rho \bigg]
\nonumber \\
&&
+\, \frac{1}{\Lambda^3}\, \bigg[
2\, \left(\Box \rho \right)^3 - 8\, \left(\nabla \partial \rho \right)^2\, \Box\rho
- \frac{16}{5}\, R\,\partial^\mu \rho\, \partial^\nu \rho \, \nabla_\mu \partial_\nu \rho
+ 8\, R^{\mu\nu}\, \partial_\mu \rho\, \partial_\nu \rho\, \Box \rho
\nonumber \\
&&
+\, 32\, R^{\mu\nu\lambda\kappa}\,\partial_\nu \rho \, \partial_\rho \rho \, \nabla_\mu \partial_\sigma \rho
\bigg] \nonumber \\
&&
+\, \frac{1}{\Lambda^4}\, \bigg[
- 4\, \left(\partial \rho\right)^2\, \left(\Box \rho \right)^2 - 4\, \partial^\mu \left( \partial\rho \right)^2\, \partial_\mu \left( \partial\rho \right)^2
+ 16\, \left(\partial \rho\right)^2\, \left(\nabla \partial \rho\right)^2 - \frac{4}{5}\, R\, \left(\partial \rho\right)^4 \bigg]
\bigg\}
\label{Ch4WZI3}
\end{eqnarray}
for the third invariant $I_3$, while the contribution from the integrated Euler density is
\begin{eqnarray}
&&
\Sigma_a =
\int d^6 x \, \sqrt{g}\, \bigg\{- \frac{\rho}{\Lambda}\, E_6
+\, \frac{1}{\Lambda^2}\, \bigg[
12\, R^{\mu\lambda\kappa\alpha}\, {R^\nu}_{\lambda\kappa\alpha}\, \partial_\mu \rho\, \partial_\nu \rho
- 3\,R^{\mu\nu\lambda\kappa}\, R_{\mu\nu\lambda\kappa}\,\left(\partial \rho \right)^2 \nonumber \\
&&
+\, 24\, R^{\mu\lambda\kappa\nu}\, R_{\lambda\kappa}\, \partial_\mu \rho\, \partial_\nu \rho
+ 12\, R^{\mu\nu}\, R_{\mu\nu}\,\left(\partial \rho \right)^2
- 24\,R^{\mu\lambda}\,{R^\nu}_\lambda \, \partial_\mu \rho\, \partial_\nu \rho
+ 12\, R\, R^{\mu\nu}\, \partial_\mu \rho\, \partial_\nu \rho - 3\, R^2\,\left(\partial\rho \right)^2 \bigg] \nonumber \\
&&
+\, \frac{1}{\Lambda^3}\, \bigg[ 16\, R^{\mu\nu\lambda\kappa}\, \partial_\nu \rho \, \partial_\lambda \rho\, \nabla_\mu \partial_\kappa \rho
- 16\, R^{\mu\nu}\,\nabla_\mu \partial_\nu \rho\, \left(\partial \rho\right)^2
+ 32\, R^{\mu\nu}\,\nabla_\mu \partial_\lambda \rho\, \nabla^\lambda \partial_\nu \rho \nonumber \\
&&
-\, 8\, R\, \partial^\mu \rho\, \partial^\nu \rho\, \nabla_\mu \partial_\nu \rho
+ 8\, R\, \left( \partial \rho \right)^2\, \Box \rho - 16\, R^{\mu\nu}\, \partial_\mu \rho\, \partial_\nu \rho \bigg] \nonumber \\
&&
+\, \frac{1}{\Lambda^4}\, \bigg[ 24\, \left( \partial \rho \right)^2\, \left(\nabla \partial \rho\right)^2
- 24\, \left(\partial \rho \right)^2\, \left(\Box \rho \right)^2 - 6\, R\, \left(\partial \rho \right)^4 \bigg]
+ \frac{36}{\Lambda^5}\, \Box \rho\, \left(\partial \rho \right)^4 - \frac{24}{\Lambda^6}\, \left(\partial \rho\right)^6
\bigg\}. \,
\label{Ch4WZE6}
\end{eqnarray}
The derivation of (\ref{Ch4WZI1})-(\ref{Ch4WZE6}) is very involved
and we have used several integration by parts to get to the previous expressions.
The WZ effective action is then obtained from (\ref{coc}) and, in a general gravitational background,
it is just given by the combination of (\ref{Ch4WZI1})-(\ref{Ch4WZE6}) with the proper coefficients, up to a minus sign, i.e.
\begin{equation}
\label{Ch4WZfinal}
\Gamma_{WZ}[g,\rho] = - \bigg( \sum_{i=1}^3 c_i\, \Sigma_i + a\, \Sigma_a \bigg)\, .
\end{equation}
Before presenting the expression of the dilaton WZ action, we pause for a comment. It is clear from our analysis that the form of this action is not
unique, due to the renormalization scheme dependence of the counterterms which are chosen before performing the Weyl-gauging. As we have
seen in section \ref{Ch4SchemDep}, this ambiguity manifests in the coefficients $c'_i$ which parametrizes the local terms of the anomaly,
proportional to the derivatives of the currents $J^{\mu}_i$.
Obviously, it is preferable to be able to characterize this ambiguity in a more complete way,
and the analysis of the relation between counterterms in $6$ and in $d$ dimensions serves this purpose.
In fact, this allows to identify the functionals whose variation generates the local anomaly terms.
By proceeding in this way one is able to identify a 3-parameter class of renormalization schemes, related to the coefficients $c'_i$, which
become free parameters in the anomaly action.
For this purpose we exploit the relations
\begin{eqnarray}
- \frac{\mu^{-\epsilon}}{\epsilon}\, \int d^dx\,\sqrt{g}\, \hat I^d_i
&=&
- \frac{\mu^{-\epsilon}}{\epsilon}\, \int d^dx\,\sqrt{g}\, \hat I_i + \int d^6x\,\sqrt{g}\, \frac{\partial \hat I^d_i}{\partial d}\bigg|_{d=6}
\nonumber \\
&=&
- \frac{\mu^{-\epsilon}}{\epsilon}\, \int d^dx\,\sqrt{g}\, \hat I^d_i + \int d^6 x\, \sqrt{g}\, \frac{\rho}{\Lambda}\, I_i\, ,
\label{Ch4Additional}
\end{eqnarray}
where the last line follows from the transformations properties of $I^d_i$ in $d$ dimensions under Weyl scaling.
From (\ref{Ch4ExpandCT}) and (\ref{Ch4Additional}) we infer that
\begin{eqnarray}
\int d^6x\,\sqrt{g}\, \frac{\partial \hat I_i}{\partial d}\bigg|_{d=6} =
-\Sigma_{i} + \int d^6x\, \sqrt{g}\, \frac{\rho}{\Lambda}\, I_i \, .
\label{Ch4ModifyWZ}
\end{eqnarray}
Then can immediately use (\ref{Ch4ModifyWZ}) to infer that the WZ action corresponding to the modified effective action
(\ref{Ch4CounterAction6Mod}) is given by
\begin{equation}
\label{Ch4WZfinalMod}
\Gamma_{WZ}[g,\rho] = - \bigg( \sum_{i=1}^3 (c_i-c'_i)\, \Sigma_i + a\, \Sigma_a
+ \sum_{i=1}^3 c'_i\, \int d^6x\, \sqrt{g}\, \frac{\rho}{\Lambda}\, I_i \bigg)\, .
\end{equation}
In particular, it is clear that, choosing the counterterms with $I_i \rightarrow I^d_i$, as in eq. (\ref{Ch4IdCounterterm}),
we get just the so-called universal terms, as done in \cite{Baume:2013ika} for general even dimensions.
\begin{equation}
\label{Ch4WZfinalUniv}
\Gamma_{WZ}[g,\rho] = - \bigg( \sum_{i=1}^3 c_i\, \int d^6x\, \sqrt{g}\, \frac{\rho}{\Lambda}\, I_i + a\, \Sigma_a \bigg) \, .
\end{equation}
In the flat space limit $(g_{\mu\nu}\to \delta_{\mu\nu})$
there are obvious simplifications and (\ref{Ch4WZfinal})takes the form
\begin{eqnarray}
\Gamma_{WZ}[\delta,\rho]
&=&
- \int d^6x\, \sqrt{g}\, \bigg\{
- \frac{c_3}{\Lambda^2}\, \Box \rho\, \Box^2\rho
+ \frac{1}{\Lambda^3}\, \bigg[
\bigg( - \frac{7}{16}\, c_1 + \frac{11}{4}\, c_2 + 2\, c_3 \bigg)\, \left(\Box\rho\right)^3 \nonumber \\
&&
+ \bigg( \frac{3}{2}\, c_1 - 6\, c_2 - 8\, c_3 \bigg)\, \left(\partial \partial \rho \right)^2\, \Box\rho \bigg]
+ \frac{1}{\Lambda^4}\, \bigg[
\bigg( -\frac{3}{2}\, c_1 + 6\, c_2 + 16\, c_3 + 24\, a\bigg)\, \left(\partial\rho\right)^2\, \left(\partial\pd\rho\right)^2
\nonumber \\
&&
-\,
\bigg( \frac{3}{8}\, c_1 + \frac{9}{2}\, c_2 + 4\, c_3 + 24\, a \bigg)\, \left(\partial\rho\right)^2\, \left(\Box\rho\right)^2
+ \bigg( \frac{3}{4}\, c_1 - 3\, c_2 - 4\, c_3 \bigg)\, \partial^\mu\left(\partial\rho\right)^2\, \partial_\mu\left(\partial\rho\right)^2
\bigg] \nonumber \\
&&
\frac{1}{\Lambda^5}\, \bigg( \frac{3}{2}\, c_1 + 6\, c_2 + 36\, a \bigg)\, \left(\partial\rho\right)^4\, \Box\rho
- \frac{1}{\Lambda^6}\, \bigg( c_1 +4\, c_2 + 24\, a \bigg)\, \left(\partial\rho\right)^6
\bigg\}\, .
\label{Ch4Effective6dFlat}
\end{eqnarray}
The structures of the flat space limits of (\ref{Ch4WZfinalMod}) and (\ref{Ch4WZfinalUniv}) follow trivially.
Having obtained the most general form for the WZ action for conformal anomalies in 6 dimensions,
we now turn to discuss one specific example in $d=6$, previously studied within the AdS/CFT correspondence.
This provides an application of the results of the previous sections.
\subsection{The WZ action action for a free CFT: the $(2,0)$ tensor multiplet}
In this section we are going to determine the coefficients of the WZ action for the (2,0) tensor multiplet in $d=6$, which has been investigated
in the past in the context of the $AdS_7/CFT_6$ holographic anomaly matching.
Free field realizations of CFT's are particularly useful in the analysis of the anomalies and their matching between theories
in regimes of strong and weak coupling, allowing to relate free and interacting theories of these types.
In this respect, the analysis of correlation functions which can be uniquely fixed by the symmetry is crucial in order to compute the anomaly for
theories characterized by different field contents in general spacetime dimensions. This is the preliminary step in order to investigate the
matching with other realizations which share the same anomaly content. These are correlation functions which contain up to 3 EMT's and
that can be determined uniquely, in any dimensions, modulo a set of
coefficients, such as the number of fermions, scalars and/or spin 1,
which can be fixed within a specific field theory realization \cite{Osborn:1993cr, Erdmenger:1996yc}
While in $d=4$ these correlation functions can be completely identified by considering a generic theory which combines free scalar, fermions and
gauge fields \cite{Osborn:1993cr, Erdmenger:1996yc,Coriano:2012wp}, in $d$ dimensions scalars and fermions need to be
accompanied not by a spin 1 (a $1$-form) but by a $\kappa$ -form ($d=2 \kappa +2$).
In $d=6$ this is a 2-form, $B_{\mu\nu}$ \cite{Bastianelli:2000hi}.
Coming to specific realizations and use of CFT's in $d=6$, we mention that, for instance, the dynamics of a single M5 brane is described by a
free $\mathcal N=(2,0)$ tensor multiplet which contains 5 scalars, 2 Weyl fermions and a 2-form whose strength is anti-selfdual. For $N$
coincident $M5$ branes, at large $N$ values, the anomaly matching between the free field theories realizations and the interacting $(2,0)$
CFT's, investigated in the $AdS_7\times S^4$ supergravity description, has served as an interesting test of the correspondence between the $A$
and $B$ parts of the anomalies in both theories \cite{Bastianelli:1999ab,Bastianelli:2000hi}.
We have summarized in table \ref{Ch4AnomalyCoeff} the coefficients
of the WZ anomaly action in the case of a scalar, a fermion and a non-chiral $B_{\mu\nu}$ form, which are the fields appearing
in the (2,2)CFT. Anomalies in the (2,0) and the (2,2) theories are related just by a factor 1/2, after neglecting the gravitational
anomalies related to the imaginary parts of the (2,0) multiplet \cite{Bastianelli:2000hi}.
We have extracted the anomaly coefficients in table \ref{Ch4AnomalyCoeff}
from \cite{Bastianelli:2000hi}, having performed a redefinition of the third invariant $I_3$
in the structure of the anomaly functional (\ref{Ch4anom6D}). We choose to denote with $\tilde I_i, \tilde J_i$ and $\tilde c_i$ the anomaly
operators and coefficients in \cite{Bastianelli:2000hi}
\begin{equation} \label{Ch4Correspondence}
\tilde I_1 = I_1\, , \quad \tilde I_2 = I_2\, , \quad \tilde I_3 = 3\, I_3 + 8\, I_1 -2\, I_2 \, .
\end{equation}
Actually in \cite{Bastianelli:2000hi} the third conformal invariant, whose implicit expression can be found in \cite{Arakelian:1995ye},
is given by
\begin{equation} \label{Ch4I3tilde}
\tilde I_3 \equiv
C^{\alpha\gamma\lambda\kappa} \bigg(
{\delta_\alpha}^\beta\, \Box - 4\, {R_\alpha}^\beta
+ \frac{6}{5}\, {\delta_\alpha}^\beta \,R \bigg)\, C_{\beta\gamma\lambda\kappa}
+ \bigg( 8\, {\delta_\alpha}^\kappa\, {\delta_\beta}^\lambda - \frac{1}{2}\, g_{\alpha\beta}\, g^{\kappa\lambda} \bigg)\,
\nabla_\kappa\nabla_\lambda C^{\alpha\gamma\lambda\kappa}\, {C^\beta}_{\gamma\lambda\kappa}\, ,
\end{equation}
which differs from our choice, reported in appendix \ref{Geometrical}.
The relation in (\ref{Ch4Correspondence}) between the third invariant $\tilde{I}_3$ and $I_3$ can be derived
expanding (\ref{Ch4I3tilde}) on the basis of the $K$-scalars given in table \ref{Ktable}
and comparing it to the third of (\ref{Ch4WeylInv6}).
In light of (\ref{Ch4Correspondence}), as the conformal anomalies depend only on the field content of the theory, i.e.
\begin{equation} \label{Ch4Matching}
\mathcal{A}[g]=\sum_{i=1}^3 c_i\, \left( I_i + \nabla_\mu J^\mu_i \right)=
\sum_{i=1}^3 \tilde c_i\, \left( \tilde I_i + \nabla_\mu \tilde J^\mu_i \right) \, ,
\end{equation}
by replacing (\ref{Ch4Correspondence}) on the r.h.s. of (\ref{Ch4Matching}),
we conclude that the relations between the anomaly coefficients $\tilde c_i$ and $c_i$ are
\begin{equation} \label{Ch4MatchCoefficients}
c_1 = \tilde c_1 + 8\, \tilde c_3\, , \quad
c_2 = \tilde c_2 - 2\, \tilde c_3\, , \quad
c_3 = 3\, \tilde c_3 \, .
\end{equation}
The WZ action can be derived from eq. (\ref{Ch4WZfinal}) by inserting the expressions of the $c_i$'s and $a$ extracted from
table \ref{Ch4AnomalyCoeff}.
These can be specialized to the scalar (S), fermion (F) and to the 2-form (B) cases, thereby generating via (\ref{Ch4Matching}) the
corresponding anomaly functionals. For the $(2,0)$ tensor multiplet this is obtained from the relation
\begin{equation}
\mathcal A^{T}[g] = \frac{1}{2}\, \bigg( 10\, \mathcal A^{S}[g] + 2\, \mathcal A^{F}[g] + \mathcal A^{B}[g] \bigg).
\end{equation}
\begin{table}
$$
\begin{array}{|c|c|c|c|c|}\hline
I & c_1\times 7!\,(4\,\pi)^3 & c_2 \times 7!\,(4\,\pi)^3 & c_3 \times 7!\,(4\,\pi)^3 & a \times 7!\,(4\,\pi)^3
\\ \hline\hline
S & \frac{20}{3} & -\frac{7}{3} & 6 & -\frac{5}{72}
\\ \hline
F & \frac{64}{3} & -112 & 120 & -\frac{191}{72}
\\ \hline
B & - \frac{3688}{3} & -\frac{3458}{3} & 540 & -\frac{221}{4}
\\ \hline
T & -560 & -700 & 420 & -\frac{245}{8}
\\
\hline
\end{array}
$$
\caption{Anomaly coefficients for a conformally coupled scalar (S), a Dirac Fermion (F), a 2-form field (B)
and the chiral $(2,0)$ tensor multiplet (T), to be normalized by an overall $1/ (7!\,(4\pi)^3)$}
\label{Ch4AnomalyCoeff}
\end{table}
\section{Dilaton interactions and constraints from $\Gamma_{WZ}$}
So far, we have extracted the structure of the WZ action and thus of the anomaly-related dilaton interactions via
the Weyl-gauging of the effective action, adding to the results which can be found in the literature all the scheme-dependent terms.
We now turn to analysing the Weyl-gauged effective action of the CTF's with a perturbative approach.
This analysis will generate a new expression of the WZ effective action, in terms of traced correlators of the EMT.
We wiil then require this expression to match the form obtained in the previous sections, for consistency.
From this the recursive algorithm will follow. \\
We proceed with a Taylor expansion in $\kappa_{\Lambda}$ of the gauged metric which is given by
\begin{equation} \label{Ch3SeriesInG}
\hat{g}_{\mu\nu} = g_{\mu\nu}\, e^{-2\,\kappa_{\Lambda}\rho} =
\bigg(\d_{\mu\nu} + \kappa\, h_{\mu\nu} \bigg)\, e^{-2\,\kappa_{\Lambda}\rho} =
\bigg(\delta_{\mu\nu} + \kappa\, h_{\mu\nu} \bigg)\,
\sum_{n=0}^{\infty} \frac{(-2)^n}{n!}\,(\kappa_{\Lambda}\,\rho)^n \, ,
\end{equation}
where $\kappa = \sqrt{16\,\pi\,G_{N}}$, with $G_{N}$ the Newton constant in $4$ ($6$) dimensions.
As we are considering only the dilaton contributions, we focus on the functional expansion of the renormalized and Weyl-gauged effective
action $\hat\Gamma_{\textrm{ren}}[g,\rho]$ with respect to $\kappa_{\Lambda}$.
This is easily done using the rule for the derivation of composite functionals,
\begin{equation} \label{Ch3CompositeDiff}
\frac{\partial\hat\Gamma_{\textrm{ren}}[g,\rho]}{\partial\kappa_{\Lambda}} =
\int d^dy\, \frac{\delta\hat\Gamma_{\textrm{ren}}[g,\rho]}{\delta\hat{g}_{\mu\nu}(x)}
\frac{\partial\hat{g}_{\mu\nu}(x)}{\partial\kappa_{\Lambda}}\, .
\end{equation}
Applying (\ref{Ch3CompositeDiff}) repeatedly and taking (\ref{Ch3SeriesInG}) into account, the perturbative series takes the form
\begin{eqnarray} \label{Ch3Expansion}
\hat\Gamma_{\textrm{ren}}[g,\rho]
&=&
\Gamma_{\textrm{ren}}[g,\rho]
+\, \frac{1}{2!\,\Lambda^2}\, \int d^d x_{1} d^d x_{2}\,
\frac{\delta^2\hat\Gamma_{\textrm{ren}}[g,\rho]}{\delta\hat{g}_{\mu_{1}\nu_{1}}(x_{1})\delta\hat{g}_{\mu_{2}\nu_{2}}(x_{2})}
\frac{\partial \hat{g}_{\mu_{1} \nu_{1}}(x_{1})}{\partial \kappa_\Lambda}\frac{\partial \hat{g}_{\mu_{2}\nu_{2}}(x_{2})}{\partial\kappa_\Lambda}
\nonumber \\
&& \hspace{10mm}
+\, \frac{1}{3!\, \Lambda^3}\, \bigg(\int d^d x_1 d^d x_{2} d^d x_{3}\,
\frac{\delta^3\hat\Gamma_{\textrm{ren}}[g,\rho]}
{\delta\hat{g}_{\mu_{1}\nu_{1}}(x_{1})\delta\hat{g}_{\mu_{2}\nu_{2}}(x_{2})\delta\hat{g}_{\mu_{3}\nu_{3}}(x_{3})}
\frac{\partial \hat{g}_{\mu_{1} \nu_{1}}(x_{1})}{\partial \kappa_\Lambda}\frac{\partial \hat{g}_{\mu_{2} \nu_{2}}(x_{2})}{\partial \kappa_\Lambda}
\frac{\partial \hat{g}_{\mu_{3} \nu_{3}}(x_{3})}{\partial \kappa_\Lambda}
\nonumber \\
&& \hspace{23mm}
+\, 3\, \int d^d x_{1} d^d x_{2}\,
\frac{\delta^2\hat\Gamma_{\textrm{ren}}[g,\rho]}{\delta\hat{g}_{\mu_{1}\nu_{1}}(x_{1})\delta\hat{g}_{\mu_{2}\nu_{2}}(x_{2})}
\frac{\partial^2 \hat{g}_{\mu_{1}\nu_{1}}(x_{1})}{\partial \kappa_\Lambda^2}\frac{\partial\hat{g}_{\mu_{2}\nu_{2}}(x_{2})}{\partial \kappa_\Lambda}
\bigg) +\ldots
\end{eqnarray}
Until now we have always defined the Green functions of the EMT in terms of the ordinary generating functional $\mathcal W$,
accorging to (\ref{Ch3NPF}). It goes without saying that, either at UV or IR fixed points of the RG flow, this definition is
generalized through the simple replacement $\mathcal{W} \rightarrow \Gamma$.
As we are interested in the flat space limit of the dilaton action, we can write (\ref{Ch3Expansion})
by taking the limit of a conformally flat background metric
$(\hat{g}_{\mu\nu}\rightarrow \hat \delta_{\mu\nu} \equiv\delta_{\mu\nu}\, e^{- 2\,\kappa_{\Lambda}\rho})$
obtaining
\begin{eqnarray} \label{Ch3FinalExp}
\hat\Gamma_{\textrm{ren}}[\delta,\rho]
&=&
\Gamma_{\textrm{ren}}[\delta,\rho] +
\frac{1}{2!\,\Lambda^2}\,
\int d^d x_{1} d^d x_{2}\, \langle T(x_{1}) T(x_{2})\rangle\, \rho(x_{1})\rho(x_{2})
\nonumber \\
&&
-\, \frac{1}{3!\,\Lambda^3}\, \bigg[ \int d^d x_{1} d^d x_{2} d^d x_{3}\,\langle T(x_{1}) T(x_{2}) T(x_{3})\rangle\, \rho(x_{1})\rho(x_{2})\rho(x_{3})
\nonumber \\
&& \hspace{20mm}
+\, 6\,\int d^d x_{1} d^d x_{2}\, \langle T(x_{1}) T(x_{2})\rangle\, (\rho(x_{1}))^2\rho(x_{2}) \bigg] + \ldots \, ,
\end{eqnarray}
where we have used eq. (\ref{Ch3NPF}) in the definition of the EMT' s correlators and the obvious relation
\begin{equation} \label{Ch3MetricDil}
\frac{\partial^n \hat{g}_{\mu\nu}(x)}{\partial \kappa_{\Lambda}^n}
\bigg|_{g_{\mu\nu}=\delta_{\mu\nu},\kappa_{\Lambda}= 0} =
\left(-2\right)^n \, \left(\rho(x)\right)^n\, \delta_{\mu\nu}\, .
\end{equation}
From (\ref{Ch3FinalExp}) one may identify the expression of
$\Gamma_{WZ}= \Gamma_{\textrm{ren}}[\delta,\rho] - \hat\Gamma_{\textrm{ren}}[\delta,\rho]$
written in terms of the traced n-point correlators of stress-energy tensors.
This has to coincide with eq. (\ref{Ch3Effective4d}) evaluated in the conformally flat limit and given by
\begin{equation} \label{Ch3FlatWZ}
- \Gamma_{WZ}[\delta,\rho] =
- \int d^4x\, \bigg[
\frac{2\,\beta_a}{\Lambda^2}\, \left( \Box \rho \right)^2
+ \left(\beta_a + \beta_b\right)\, \bigg( - \frac{4}{\Lambda^3}\, \left(\partial\rho\right)^2\,\Box \rho
+ \frac{2}{\Lambda^4}\, \left(\partial\rho\right)^4 \bigg) \bigg ] \, .
\end{equation}
At this point, a comparison between the dilaton vertices extracted from (\ref{Ch3FinalExp}) and (\ref{Ch3FlatWZ}) allows to establish a
consistency condition between the first four of such vertices and a relation among the entire hierarchy of the traced correlators,
which is the key result of this chapter.
For this purpose we denote by $\mathcal I_n(x_{1},\dots,x_n)$ the dilaton vertices obtained by functional differentiation of
$\Gamma_{\textrm{ren}}[\hat{\delta}]$,
\begin{equation} \label{Ch3FuncDiffWZ}
\mathcal I_{n}(x_{1},\dots,x_n) =
\frac{\delta^n \left(\hat\Gamma_{\textrm{ren}}[\delta,\rho]-\Gamma_{\textrm{ren}}[\delta,\rho]\right)}
{\delta\rho(x_{1})\dots\delta\rho(x_n)}
= - \frac{\delta^n \Gamma_{WZ}[\delta,\rho]}{\delta\rho(x_{1})\dots\delta\rho(x_n)}
\end{equation}
in coordinate space, which we can promptly transformed to momentum space.
The expressions of such vertices for the first six orders in $\kappa_{\Lambda}$ are given by
\begin{eqnarray}
{\mathcal I}_2(k_{1},-k_{1})
&=&
\frac{1}{\Lambda^2}\, \langle T(k_{1}) T(-k_{1})\rangle \, ,
\nonumber \\
{\mathcal I}_3(k_{1},k_{2},k_{3})
&=&
-\frac{1}{\Lambda^3}\, \bigg[
\langle T(k_{1}) T(k_{2}) T(k_{3}) \rangle
+\, 2\, \bigg( \langle T(k_{1}) T(-k_{1})\rangle + \langle T(k_{2}) T(-k_{2})\rangle + \langle T(k_{3}) T(-k_{3}) \rangle \bigg) \bigg]\, ,
\nonumber \\
{\mathcal I}_4(k_{1},k_{2},k_{3},k_{4})
&=&
\frac{1}{\Lambda^4}\, \Biggl[
\langle T(k_{1}) T(k_{2}) T(k_{3}) T(k_{4}) \rangle
+\, 2\,\sum_{\mathcal T\left\{4,(k_{i_1},k_{i_2})\right\}}
\langle T(k_{i_1}) T(k_{i_2}) T(-k_{i_1}-k_{i_2})\rangle
\nonumber \\
&& \hspace{6mm}
+\, 2\, \sum_{\mathcal T\left\{4,(k_{i_1},k_{i_2})\right\}}
\langle T(k_{i_1}+k_{i_2}) T(-k_{i_1}-k_{i_2})\rangle
+ 4\, \sum_{i=1}^{4} \langle T(k_i) T(-k_i)\rangle \Biggr]\, ,
\nonumber \\
{\mathcal I}_5(k_{1},k_{2},k_{3},k_{4},k_{5})
&=&
- \frac{1}{\Lambda^5}\, \Biggl[
\langle T(k_{1}) T(k_{2}) T(k_{3}) T(k_{4}) T(k_{5}) \rangle
\nonumber \\
&&
+\, 2\, \sum_{\mathcal T\left\{5,(k_{i_1},k_{i_2},k_{i_3})\right\}}
\langle T(k_{i_1}) T(k_{i_2}) T(k_{i_3}) T(-k_{i_1}-k_{i_2}-k_{i_3}) \rangle
\nonumber \\
&&
+ \,4\, \Biggl(
\sum_{\mathcal T\left\{5,(k_{i_1},k_{i_2})\right\}}
\langle T(k_{i_1}) T(k_{i_2}) T(-k_{i_1}-k_{i_2}) \rangle
\nonumber \\
&&
+\, \sum_{\mathcal T\left\{5,[(k_{i_1},k_{i_2}),(k_{i_3},k_{i_4})]\right\}}
\langle T(k_{i_1}+k_{i_2}) T(k_{i_3}+k_{i_4}) T(-k_{i_1}-k_{i_2}-k_{i_3}-k_{i_4}) \rangle \Biggr)
\nonumber \\
&&
+\, 8\, \Biggl(
\sum_{\mathcal T\left\{5,(k_{i_1},k_{i_2})\right\}} \langle T(k_{i_1}+k_{i_2}) T(-k_{i_1}-k_{i_2}) \rangle +
\sum_{i=1}^{5} \langle T(k_i) T(-k_i)\rangle \Biggr)
\Biggr]\, ,
\label{Ch3DilIntStructure} \\
{\mathcal I}_6(k_{1},k_{2},k_{3},k_{4},k_{5},k_{6})
&=&
\kappa_{\Lambda}^6\, \Biggl[
\langle T(k_{1}) T(k_{2}) T(k_{3}) T(k_{4}) T(k_{5}) T(k_{6}) \rangle
\nonumber \\
&&
+\, 2\, \sum_{\mathcal T\left\{6,(k_{i_1},k_{i_2})\right\}}
\langle T(k_{i_1}+k_{i_2}) T(k_{i_3}) T(k_{i_4}) T(k_{i_5}) T(k_{i_6}) \rangle
\nonumber \\
&&
+ \,4\, \Biggl( \sum_{\mathcal T\left\{6,(k_{i_1},k_{i_2},k_{i_3})\right\}}
\langle T(k_{i_1}+k_{i_2}+k_{i_3}) T(k_{i_4}) T(k_{i_5}) T(k_{i_6}) \rangle
\nonumber \\
&&
+ \, \sum_{\mathcal T\left\{6,[(k_{i_1},k_{i_2}),(k_{i_3},k_{i_4})]\right\}}
\langle T(k_{i_1}+k_{i_2}) T(k_{i_3}+k_{i_4}) T(k_{i_5}) T(k_{i_6}) \rangle
\Biggr) \nonumber \\
&&
+\, 8\, \Biggl(
\sum_{\mathcal T\left\{6,(k_{i_1},k_{i_2},k_{i_3},k_{i_4})\right\}}
\langle T(k_{i_1}+k_{i_2}+k_{i_3}+k_{i_4}) T(k_{i_5}) T(k_{i_6}) \rangle
\nonumber
\end{eqnarray}
\begin{eqnarray}
&&
+\, \sum_{\mathcal T\left\{6,[(k_{i_1},k_{i_2},k_{i_3}),(k_{i_4},k_{i_5})]\right\}}
\langle T(k_{i_1}+k_{i_2}+k_{i_3}) T(k_{i_4}+k_{i_5}) T(k_{i_6}) \rangle
\nonumber \\
&&
+\, \sum_{\mathcal T\left\{6,[(k_{i_1},k_{i_2}),(k_{i_3},k_{i_4})]\right\}}
\langle T(k_{i_1}+k_{i_2}) T(k_{i_3}+k_{i_4}) T((k_{i_5}+k_{i_6})) \rangle\Biggr)
\nonumber \\
&&
+\, 16\, \Biggl(
\frac{1}{2}\, \sum_{\mathcal T\left\{6,(k_{i_1},k_{i_2},k_{i_3})\right\}}
\langle T(k_{i_1}+k_{i_2}+k_{i_3}) T(-k_{i_1}-k_{i_2}-k_{i_3}) \rangle
\nonumber \\
&&
+\, \, \sum_{\mathcal T\left\{6,(k_{i_1},k_{i_2})\right\}} \langle T(k_{i_1}+k_{i_2}) T(-k_{i_1}-k_{i_2}) \rangle
+ \sum_{i=1}^{6} \langle T(k_i) T(-k_i)\rangle
\Biggr)
\Biggr].
\label{Ch4DilIntStructure6}
\end{eqnarray}
These results can be easily extended to any higher order.
The recipe, in this respect, is really simple: \\
in order to construct the vertex at order $n$ one has to sum to the $n$-point function
all the lower order functions in the hierarchy, down to $n=2$, partitioning the momenta in all the possible ways
and symmetrising each single contribution.
The normalization factor in front of the correlator of order-$k$ is always $2^{n-k}$, while the factor in front of the
vertex of order $n$ is $(-\kappa_\Lambda)^n$. Notice that, for $n$ even, we have an additional $1/2$ factor in front of the contributions from
the $2$-point functions in which each EMT carries $n/2$ momenta, to avoid double counting.
Notice that the expressions of the dilaton interactions in (\ref{Ch4DilIntStructure6}) have been derived withour any reference to the
dimensions of space, so that they hold for any even dimension. As such, they can be thoroughly tested in the simplest case, i.e. $d=2$.
We did check them in $2$ dimensions, as illustrated in appendix \ref{Ch42D}.
We pause for a moment to clarify the notation used in
(\ref{Ch3DilIntStructure}) for the organization of the momenta and the meaning of the symbol $\mathcal T$.
For example $\mathcal T\left\{4,(k_{i_1},k_{i_2})\right\} $ denotes the six pairs of distinct momenta in the case of the four point
functions
\begin{equation}
\mathcal T\left\{4,(k_{i_1},k_{i_2})\right\} =
\left\{(k_{1},k_{2}),(k_{1},k_{3}),(k_{1},k_{4}),(k_{2},k_{3}),(k_{2},k_{4}),(k_{3},k_{4}) \right\} \, ,
\end{equation}
where we are combining the 4 momenta $k_1,...k_4$ into all the possible pairs, for a total of $\binom{4}{2}$ terms.
With five momenta $\left(k_{1},k_{2},k_{3},k_{4},k_{5}\right)$ the available pairs are
\begin{eqnarray}
\mathcal T\left\{5,(k_{i_1},k_{i_2})\right\}
&=&
\left\{(k_{1},k_{2}),(k_{1},k_{3}),(k_{1},k_{4}),(k_{1},k_{5}),(k_{2},k_{3}),(k_{2},k_{4}),(k_{2},k_{5}),(k_{3},k_{4}),(k_{3},k_{5}),(k_{4},k_{5}) \right\} \, \nonumber\\
\end{eqnarray}
while the possible triples are
\begin{eqnarray}
\mathcal T\left\{5,(k_{i_1},k_{i_2},k_{i_3})\right\}
&=&
\left\{(k_{1},k_{2},k_{3}),(k_{1},k_{2},k_{4}),(k_{1},k_{2},k_{5}),(k_{1},k_{3},k_{4}),(k_{1},k_{3},k_{5}),
\right.
\nonumber \\
&&
\left.
\hspace{1mm}
(k_{1},k_{4},k_{5}),(k_{2},k_{3},k_{4}),(k_{2},k_{3},k_{5}),(k_{2},k_{4},k_{5}),(k_{3},k_{4},k_{5}) \right\}\, .
\end{eqnarray}
As we move to higher orders, the description of the momentum dependence gets slightly more involved and we need to distribute the external
momenta into two pairs. The notation $\mathcal T\left\{5, [(k_{i_1},k_{i_2}),(k_{i_3},k_{i_4})]\right\}$ denotes the set of
independent paired couples which can be generated out of $5$ momenta.
Their number is $15$ and they are given by
\begin{eqnarray}
\mathcal T\left\{5,[(k_{i_1},k_{i_2}),(k_{i_3},k_{i_4})]\right\}
&=&
\left\{[(k_{1},k_{2}),(k_{3},k_{4})],[(k_{1},k_{2}),(k_{3},k_{5})],[(k_{1},k_{2}),(k_{4},k_{5})]
\right.
\nonumber \\
&& \hspace{-55mm}
\left.
[(k_{1},k_{3}),(k_{2},k_{4})],[(k_{1},k_{3}),(k_{2},k_{5})],[(k_{1},k_{3}),(k_{4},k_{5})],[(k_{1},k_{4}),(k_{2},k_{3})],
[(k_{1},k_{4}),(k_{2},k_{5})],[(k_{1},k_{4}),(k_{3},k_{5})],
\right.
\nonumber \\
&& \hspace{-55mm}
\left.
[(k_{1},k_{5}),(k_{2},k_{3})],[(k_{1},k_{5}),(k_{2},k_{4})],[(k_{1},k_{5}),(k_{3},k_{4})],[(k_{2},k_{3}),(k_{4},k_{5})],
[(k_{2},k_{4}),(k_{3},k_{5})],[(k_{2},k_{5}),(k_{3},k_{4})] \right\}\, .
\end{eqnarray}
\subsection{The recursive relation in $4$ dimensions}
It is obvious that a direct computation of $\mathcal{I}_2, \mathcal{I}_3$ and $\mathcal{I}_4$
from the anomaly action (\ref{Ch3FlatWZ}) allows to extract the explicit structure of these vertices in momentum space
\begin{eqnarray} \label{Ch3DilatonInt}
{{\mathcal I}}_2(k_{1},-k_{1})
&=&
-\frac{4}{\Lambda^2}\, \beta_a\, {k_{1}}^4\, , \nonumber \\
{{\mathcal I}}_3(k_{1},k_{2},k_{3})
&=&
\frac{8}{\Lambda^3} \, \bigg(\beta_a + \beta_b \bigg)\, \bigg(
k_{1}^2\, k_{2}\cdotk_{3} + k_{2}^2\, k_{1}\cdotk_{3} + k_{3}^2\, k_{1}\cdotk_{2} \bigg)
\nonumber \\
{{\mathcal I}}_4(k_{1},k_{2},k_{3},k_{4})
&=&
- \frac{16}{\Lambda^4}\,\bigg(\beta_a + \beta_b \bigg)\, \bigg(
k_{1}\cdotk_{2}\, k_{3}\cdotk_{4} + k_{1}\cdotk_{3}\, k_{2}\cdotk_{4} + k_{1}\cdotk_{4}\, k_{2}\cdotk_{3} \bigg)\, ,
\end{eqnarray}
(with $k_{i}^n \equiv (k_i^2)^{n/2}$). These relations can be used together with (\ref{Ch3FinalExp}) in order to extract the
structure of the $2$- $3$- and $4$-point functions of the traced correlators, solving an elementary linear system.
Their expressions are very easily found to be
\begin{eqnarray} \label{Ch3BuildingBlocks}
\langle T(k_{1}) T(-k_{1}) \rangle
&=&
- 4\, \beta_a\, {k_{1}}^4 \, ,
\nonumber \\
\langle T(k_{1}) T(k_{2}) T(k_{3}) \rangle
&=&
8 \bigg[
- \bigg( \beta_a+\beta_b \bigg)\,\bigg( f_{3}(k_{1},k_{2},k_{3})+f_{3}(k_{2},k_{1},k_{3})+f_{3}(k_{3},k_{1},k_{2})\bigg)
+ \beta_a\, \sum_{i=1}^{3} k_i^4 \bigg]\, ,
\nonumber \\
\langle T(k_{1}) T(k_{2}) T(k_{3}) T(k_{4})\rangle
&=&
8\, \bigg\{
6\, \bigg( \beta_a + \beta_b \bigg)\, \bigg[
\sum_{\mathcal T\left\{4,[(k_{i_1},k_{i_2}),(k_{i_3},k_{i_4})]\right\}}
k_{i_i}\cdot k_{i_2}\, k_{i_3}\cdot k_{i_4}
\nonumber \\
&&
+\, f_{4}(k_{1}\,k_{2},k_{3},k_{4}) + f_{4}(k_{2}\,k_{1},k_{3},k_{4}) + f_{4}(k_{3}\,k_{1},k_{2},k_{4}) + f_{4}(k_{4}\,k_{1},k_{2},k_{3}) \bigg]
\nonumber \\
&&
-\, \beta_a\,
\bigg( \sum_{\mathcal T\left\{4,(k_{i_1},k_{i_2})\right\}}(k_{i_1} + k_{i_2})^4
+ 4\, \sum_{i=1}^{4} k_{i}^4 \bigg)
\bigg\}\, ,
\end{eqnarray}
where we have introduced the compact notations
\begin{eqnarray}
f_{3}(k_a,k_b,k_c)
&=&
k_a^2\, k_b \cdot k_c \, ,
\nonumber \\
f_{4}(k_a,k_b,k_c,k_d)
&=&
k_a^2\, \left( k_b \cdot k_c + k_b \cdot k_d + k_c \cdot k_d \right)\, .
\end{eqnarray}
The third and fourth order results, in particular, were established in \cite{Coriano:2012dg}
via the explicit computation of the first three functional derivatives of the anomaly $\mathcal A[g]$
and exploiting recursively the hierarchical relations (\ref{Ch3hier}).
here comes the second significant result: it is quite immediate to realize that the hierarchy in eq. (\ref{Ch3hier})
can be entirely re-expressed in terms of the first four traced correlators.
For this purpose, one has just to notice that $\Gamma_{WZ}[\hat{\delta}]$ is quartic in $\rho$, with $\mathcal{I}_n=0$, for $n\ge 5$.
Therefore, for instance, the absence of vertices with $5$ dilaton external lines, which sets $\mathcal{I}_5=0$,
combined with the 4 fundamental traces in (\ref{Ch3BuildingBlocks}), are sufficient to completely fix the structure of the 5-point
function, which takes the form
\begin{eqnarray}
&&
\langle T(k_{1}) T(k_{2}) T(k_{3}) T(k_{4}) T(k_{5}) \rangle =
16\, \Biggl\{
-24\,\bigg(\beta_a + \beta_b\bigg)\, \Biggl[
\sum_{\mathcal T\left\{5,[(k_{i_1},k_{i_2}),(k_{i_3},k_{i_4})]\right\}}
k_{i_1}\cdot k_{i_2}\,k_{i_3}\cdot k_{i_4}
\nonumber \\
&& \hspace{-10mm}
+\, f_{5}(k_{1},k_{2},k_{3},k_{4},k_{5}) + f_{5}(k_{2},k_{1},k_{3},k_{4},k_{5}) + f_{5}(k_{3},k_{1},k_{2},k_{4},k_{5})
+ f_{5}(k_{4},k_{1},k_{2},k_{3},k_{5}) + f_{5}(k_{5},k_{1},k_{2},k_{3},k_{4}) \Biggr]
\nonumber \\
&& \hspace{-10mm}
+\, \beta_a\, \Biggl[
\sum_{\mathcal T\left\{5,(k_{i_1},k_{i_2},k_{i_3})\right\}}
\left(k_{i_1} + k_{i_2} + k_{i_3} \right)^4
+ 3\, \sum_{\mathcal T\left\{5,(k_{i_1},k_{i_2})\right\}} \left(k_{i_1} + k_{i_2} \right)^4
+ 12\, \sum_{i=1}^{5} k_{i}^4
\Biggr]
\Biggr\}\, ,
\label{Ch35T}
\end{eqnarray}
where $f_5$ is defined as
\begin{equation}
f_{5}(k_a,k_b,k_c,k_d,k_e) =
k_a^2\, \left( k_b \cdot k_c + k_b \cdot k_d + k_b \cdot k_e + k_c \cdot k_d + k_c \cdot k_e + k_d \cdot k_e \right)\, .
\end{equation}
The construction that we have outlined can be extended to any arbitrary traced $n$-point function of the EMT:
it just takes one to apply the simple recipe to express the $n$-dilaton interactions, which is given under eq. (\ref{Ch4DilIntStructure6}).
These relations can be compared, for consistency, with their equivalent expression obtained directly from the hierarchy (\ref{Ch3hier}).
In general, this requires the computation of functional derivatives of the anomaly functional $\mathcal{A}$ up to the relevant order.
One can check by a direct computation using (\ref{Ch3hier}) the agreement with (\ref{Ch35T}) up to the $5$-th order,
as we explicitly did to test the correctness of our result.
All the results given in this section can be easily generalized with the inclusion of a counterterm (\ref{Ch3ll}),
using the prescription (\ref{Ch3LocalToBeta}), as discussed above.
\subsection{The recursive relation in $6$ dimensions}
At this point we can move on to the evaluation of dilaton interactions in $6$ dimensions and, consequently,
of the first $6$ traced correlators, being clear from (\ref{Ch4DilIntStructure6}) that a direct computation of
$\mathcal{I}_2 - \mathcal{I}_6$ from the anomaly action (\ref{Ch4Effective6dFlat}) allows to extract the structure of these
Green functions.
The dilaton interactions are straightforwarly computed,
\begin{eqnarray} \label{Ch4DilatonInt}
{\mathcal I}_2(k_{1},-k_{1})
&=&
\frac{2}{\Lambda^2}\, c_3\, k_1^6\, , \nonumber \\
{\mathcal I}_3(k_{1},k_{2},k_{3})
&=&
\frac{1}{\Lambda^3}\, \bigg[
\bigg( \frac{21}{8}\, c_1 - \frac{33}{2}\,c_2 - 12\,c_3 \bigg)\, k_1^2\,k_2^2\,k_3^2 \nonumber \\
&& \hspace{4mm}
+\, \bigg( - 3\, c_1 + 12\, c_2 + 16\,c_3 \bigg)\, \bigg( k_1^2\, \left(k_2\cdot k_3\right)^2
+ k_2^2\, \left(k_1\cdot k_3\right)^2 + k_3^2\, \left(k_1\cdot k_2\right)^2 \bigg) \bigg] \, ,
\nonumber \\
{\mathcal I}_4(k_{1},k_{2},k_{3},k_{4})
&=&
\frac{1}{\Lambda^4}\, \bigg[
\bigg( 6\, c_1 - 24\,c_2 - 64\,c_3 - 96 \,a\bigg)\,
\sum_{\mathcal T\left\{4,(k_{i_1},k_{i_2})\right\}} k_{i_1}\cdot k_{i_2}\, \left(k_{i_3}\cdot k_{i_4}\right)^2 \nonumber \\
&& \hspace{-10mm}
+\, \bigg(\frac{3}{2}\, c_1 + 18\, c_2 + 16 \,c_3 + 96\,a \bigg)\,
\sum_{\mathcal T\left\{4,(k_{i_1},k_{i_2})\right\}} k_{i_1}\cdot k_{i_2}\, k_{i_3}^2\, k_{i_4}^2
\nonumber \\
&& \hspace{-10mm}
+\, \bigg( - 6\, c_1 + 24\,c_2 + 32\,c_3 \bigg)\,
\sum_{\mathcal T\left\{4,\left[(k_{i_1},k_{i_2}),(k_{i_3},k_{i_4})\right]\right\}}
(k_{i_1}+k_{i_2})\cdot(k_{i_3}+k_{i_4})\, k_{i_1}\cdot k_{i_2}\, k_{i_3}\cdot k_{i_4} \bigg] \, ,
\nonumber
\end{eqnarray}
\begin{eqnarray}
{\mathcal I}_5(k_{1},k_{2},k_{3},k_{4},k_{5})
&=&
- \frac{12}{\Lambda^5}\, \bigg(c_1 + 4\,c_2 + 24\,a \bigg) \nonumber \\
&&
\times \sum_{\mathcal T\left\{5,(k_{i_1},k_{i_2},k_{i_3},k_{i_4})\right\}}
k_{i_5}^2\, \left( k_{i_1} \cdot k_{i_2} \, k_{i_3} \cdot k_{i_4} + k_{i_1} \cdot k_{i_3} \, k_{i_2} \cdot k_{i_4}
+ k_{i_1} \cdot k_{i_4} \, k_{i_2} \cdot k_{i_3} \right)\, , \nonumber \\
{\mathcal I}_6(k_{1},k_{2},k_{3},k_{4},k_{5},k_{6})
&=&
\frac{48}{\Lambda^6}\, \bigg(c_1 + 4\, c_2 + 24\,a\bigg)\,
\sum_{\mathcal T\left\{6,\left[(k_{i_1},k_{i_2}),(k_{i_3},k_{i_4}),(k_{i_5},k_{i_6})\right]\right\}}
k_{i_1}\cdot k_{i_2}\, k_{i_3}\cdot k_{i_4}\, k_{i_5}\cdot k_{i_6} \, , \nonumber \\
\end{eqnarray}
(with $k_{i}^n \equiv (k_i^2)^{n/2}$).
These vertices can be used together with the relations (\ref{Ch4DilIntStructure6}) in order to extract the structure of the traced correlators.
Solving the linear system, we find that the first two of them are given by
\begin{eqnarray} \label{Ch4BuildingBlocks23}
\langle T(k_{1}) T(-k_{1}) \rangle
&=&
2\, c_3\, k_{1}^6 \, ,
\nonumber \\
\langle T(k_{1}) T(k_{2}) T(k_{3}) \rangle
&=&
\bigg( 3\, c_1 - 12\,c_2 - 16\,c_3 \bigg)\,
\bigg( k_{1}^2\,\left( k_{2}\cdotk_{3} \right)^2 + k_{2}^2\,\left( k_{1}\cdotk_{3} \right)^2 + k_{3}^2\,\left( k_{1}\cdotk_{2} \right)^2 \bigg)
\nonumber \\
&&
- \bigg( \frac{21}{8}\, c_1 - \frac{33}{2}\, c_2 - 12\, c_3 \bigg)\, k_{1}^2\,k_{2}^2\,k_{3}^2
-\, 4\,c_3\,\bigg( k_{1}^6 + k_{2}^6 + k_{3}^6 \bigg) \, .
\end{eqnarray}
The structure of the $4$-point Green function is much more complicated and can be expressed in a compact notation through the expression
\begin{eqnarray}
\langle T(k_{1}) T(k_{2}) T(k_{3}) T(k_{4})\rangle
&=&
\bigg[ 4\, c_3\, \bigg(7\, f^{2i,2i,2i} + 6\, f^{2i,2i,ij} + 3\, f^{2i,2i,2j}
+ 12\, f^{2i,ij,ij} + 12\, f^{2i,2j,ij} + 8\, f^{ij,ij,ij} \bigg)
\nonumber \\
&&
+\, \bigg(-18\, c_1 + 72\, c_2 + 96\, c_3 \bigg)\, f^{2i,jk,jk}
+ 4\, \bigg(24\, a + 3\, c_1 - 12\, c_2 - 8\, c_3\bigg)\, f^{2i,2j,kl}
\nonumber \\
&&
-\, 6\, \bigg(16\, a + c_1 - 4\,c_2\bigg) f^{ij,kl,kl}
+ \bigg(\frac{63}{4}\, c_1 - 99\, c_2 - 72\, c_3 \bigg)\, f^{2i,2j,2k}
\nonumber \\
&&
+\, \bigg(-6\,c_1 + 24\,c_2 + 32\,c_3 \bigg)\, \bigg( 2\, \, f^{2i,jk,jl} + f^{ij,ik,jl} \bigg)
\bigg]\, .
\label{Ch4BuildingBlocks4}
\end{eqnarray}
Here we have introduced a compact notation for the basis of the $12$ scalar functions $f^{\dots}(k_{1},k_{2},k_{3},k_{4})$
on which the correlator is expanded, leaving their dependence on the momenta implicit not to make the formula clumsy.
As dimensional analysis forces every term in the Green function to be the product of three scalar products of momenta,
the role of the tree superscripts on each of the $f$'s is to specify the way in which the momenta are distributed.
We present below the expressions of the first four scalar $f$'s, from which it should be clear how to derive the explicit forms of all the
others. We obtain
\begin{eqnarray}
f^{2i,2i,2i}(k_{1},k_{2},k_{3},k_{4})
&=&
\sum_{i=1}^{4} (k_i)^6 \, , \nonumber \\
f^{2i,2i,ij}(k_{1},k_{2},k_{3},k_{4})
&=&
\sum_{i=1}^{4} (k_i)^4\, \sum_{j\neq i} k_i \cdot k_j\, , \nonumber \\
f^{2i,2i,2j}(k_{1},k_{2},k_{3},k_{4})
&=&
\sum_{i=1}^{4} (k_i)^4\, \sum_{j\neq i} k_j^2\, , \nonumber \\
f^{2i,ij,ij}(k_{1},k_{2},k_{3},k_{4})
&=&
\sum_{i=1}^{4} (k_i)^2\, \sum_{j\neq i} \left(k_i \cdot k_j\right)^2\, .
\end{eqnarray}
Notice that each one of the $f$'s is completely symmetric with respect to any permutation of the momenta, as for the whole correlator.
The structure of the $5$- and $6$-point functions is essentially similar to (\ref{Ch4BuildingBlocks23}),
although they require much broader bases of scalar functions to account for all their terms and we do not report them explicitly.
Again, it is clear that the hierarchy in eq. (\ref{Ch3hier}) can be entirely re-expressed in terms of the first $6$
traced correlators. In fact, one notices that $\Gamma_{WZ}[\hat{\delta}]$ is at most of order $6$ in $\rho$, which implies
\begin{equation} \label{Ch4NoInt}
\mathcal I_n(x_{1},\dots,x_n) = 0\, , \quad n \geq 7 \, .
\end{equation}
Therefore, for instance, the absence of vertices with $7$ dilaton external lines, which sets $\mathcal{I}_7=0$
and the knowledge of the first $6$ fundamental Green functions are sufficient to completely fix the structure of the $7$-point function,
and so for the vertices of higher orders. In this way one can determine all the others recursively, up to the desired order.
The consistency of these relations could be checked, in principle, by a direct comparison with their expression obtained directly
from the hierarchy (\ref{Ch3hier}). This requires the explicit computation of functional derivatives of the anomaly functional $\mathcal{A}$
up to the relevant order, which is a much more time-consuming task.
\section{Conclusions}
Our analysis has had the goal of showing that the infinite hierarchy of fully traced correlation functions generated by the
anomaly constraint in a generic CFT in even dimensions has as fundamental building blocks, in $d$ dimensions, only the first $d$
correlators. For instance, in $d=4$ only correlators with $2, 3$ and $4$ traces are necessary to identify the entire hierarchy.
This result can be simply derived from the structure of the WZ action, which only contains dilaton
interactions up to the quartic order. Non anomalous terms, which are homogeneous under Weyl transformations and can be
of arbitrarily higher orders in $\rho$, do not play any role in this construction.
The WZ action can also be determined, in general, by the Noether method, where the dilaton is coupled directly to the anomaly
and corrections are included in order to take care of the Weyl non-invariance of the functional.
Alternatively, the same action is fixed by the cocycle condition, which shows that its functional dependence
on the dilaton field takes place via the Weyl-gauging of the metric tensor. In our analysis we have introduced an expression of the
anomaly-induced action in which the anomaly contribution is generated directly by the counterterms, evaluated in dimensional
regularization.
The WZ conformal anomaly actions that we have derived include all the contributions related to the local part of the anomaly.
This result adds full generality to the analysis of dilaton effective actions, which carry an intrinsic regularization scheme dependence,
due to the appearance in the anomaly functional of terms that are different from the Euler density and the $d$-dimensional Weyl invariants
in the specific space dimension.
In general, the extraction of these extra contributions, as one can figure out from our study, is very involved, with a level of difficulty that grows
with the dimensionality of the space in which the underlying CFT is formulated. Our main results for WZ dilaton actions, especially in $d=6$,
are remarkably simplified in flat space.
Comparing our results with those of the previous literature \cite{Bastianelli:1999ab}, we have given the form of the WZ action
in the case of the CFT of the $(2,0)$ tensor multiplet, which, in the past, has found application in the $AdS_7/CFT_6$ correspondence.
\clearpage{\pagestyle{empty}\cleardoublepage}
\begin{appendix}
\chapter{Appendix}\label{Ch1ComputeTTT}
\section{Sign conventions}\label{Sign}
The definition of the Fourier transform of a $n$-EMT's correlation functions, which holds for any other $n$-point function as well,
is given by
\begin{equation}
\int \, d^dx_{1}\, \dots d^d x_n\, \left\langle T^{\mu_{1}\nu_{1}}(x_{1})\dots T^{\mu_n\nu_n}(x_n)\right\rangle \,
e^{-i(k_{1}\cdot x_{1} + \dots + k_n \cdot x_n)} = (2\pi)^d\,
\delta^{(d)}\left( \sum_{i=1}^n k_i \right)\,\left\langle T^{\mu_{1}\nu_{1}}(k_{1})\dots T^{\mu_n\nu_n}(k_n)\right\rangle \, ,
\label{Ch3NPFMom}
\end{equation}
where all the momenta are conventionally taken to be incoming. \\
The covariant derivatives of a contravariant vector $A^\mu$ and of a covariant one $B_\mu$ are respectively
\begin{eqnarray}
\nabla_{\nu} A^\mu \equiv \partial_\nu A^\mu + \Gamma^\mu_{\nu\r}A^\r\, ,\\
\nabla_{\nu} B_\mu \equiv \partial_\nu B_\mu - \Gamma^\r_{\nu\mu}B_\r\, ,
\end{eqnarray}
with the Christoffel symbols defined as
\begin{equation}\label{Ch1Christoffel}
\Gamma^{\a}_{\b\g} = \frac{1}{2}g^{\a\k}\left[-\partial_\k g_{\b\g} + \partial_\b g_{\k\g} + \partial_\g g_{\k\b} \right]\, .
\end{equation}
Our definition of the Riemann tensor is
\begin{eqnarray} \label{Ch1Tensors}
{R^\lambda}_{\mu\kappa\nu}
&=&
\partial_\nu \Gamma^\lambda_{\mu\kappa} - \partial_\kappa \Gamma^\lambda_{\mu\nu}
+ \Gamma^\lambda_{\nu\eta}\Gamma^\eta_{\mu\kappa} - \Gamma^\lambda_{\kappa\eta}\Gamma^\eta_{\mu\nu}.
\end{eqnarray}
The Ricci tensor is defined by the contraction $R_{\mu\nu} = {R^{\lambda}}_{\mu\lambda\nu}$
and the scalar curvature by $R = g^{\mu\nu}R_{\mu\nu}$.\\
The functional variations with respect to the metric tensor are computed using the relations
\begin{eqnarray}\label{Ch1Tricks}
\delta \sqrt{g} = -\frac{1}{2} \sqrt{g}\, g_{\a\b}\,\delta g^{\a \b}\quad &&
\delta \sqrt{g} = \frac{1}{2} \sqrt{g}\, g^{\a\b}\,\delta g_{\a \b} \nonumber \\
\delta g_{\mu\nu} = - g_{\mu\a} g_{\nu\b}\, \delta g^{\a\b} \quad&&
\delta g^{\mu\nu} = - g^{\mu\a} g^{\nu\b}\, \delta g_{\a\b}\,
\end{eqnarray}
The structure $s^{\a\b\g\delta} $ has been repeatedly used throughout the calculations: it comes from
\begin{eqnarray}\label{Ch1Tricks2}
- \frac{\d g^{\a\b}(z)}{\d g_{\g\d}(x)}\bigg|_{g_{\mu\nu}=\delta_{\mu\nu}} =
\frac{1}{2}\left[\delta^{\a\g}\delta^{\b\delta} + \delta^{\a\delta}\delta^{\b\g}\right]\, \delta^{(4)}(z-x)
= s^{\mu\nu\alpha\beta}\, \delta^{(4)}(z-x)\, .
\end{eqnarray}
The variations of the Christoffel symbols are tensors themselves and their expression is
\begin{eqnarray} \label{Ch3deltaChristoffel}
\delta \Gamma^\alpha_{\beta\gamma}
&=&
\frac{1}{2}\,g^{\alpha\lambda}\big[
- \nabla_{\lambda}(\delta g_{\beta\gamma}) + \nabla_{\gamma}(\delta g_{\beta\lambda}) + \nabla_{\beta}(\delta g_{
\gamma\lambda})
\big]\, ,
\nonumber\\
\nabla_\rho \delta\Gamma^\alpha_{\beta\gamma}
&=&
\frac{1}{2}\,g^{\alpha\lambda}\big[ -
\nabla_{\rho}\nabla_{\lambda}(\delta g_{\beta\gamma}) +
\nabla_{\rho}\nabla_{\gamma} (\delta g_{\beta\lambda})
+ \nabla_{\rho}\nabla_{\beta}(\delta g_{\gamma\lambda}) \big]\, .
\end{eqnarray}
\section{Results for Weyl-gauging}
In the Weyl-gauging of the counterterms, we use the following relations
\begin{eqnarray}
{\hat{\Gamma}}^\alpha_{\beta\gamma}
&=&
\Gamma^\alpha_{\beta\gamma} + \frac{1}{\Lambda}\, \bigg( {\delta_\beta}^\alpha\, \nabla_\gamma\rho
+ {\delta_\gamma}^\alpha\, \nabla_\beta\rho - g_{\beta\gamma}\, \nabla^\alpha\rho \bigg) \, ,
\nonumber \\
\hat {R^\mu}_{\nu\lambda\sigma}
&=&
{R^\mu}_{\nu\lambda\sigma}
+ g_{\nu\lambda}\, \bigg( \frac{\nabla_{\sigma}\partial^\mu\rho}{\Lambda} + \frac{\partial^\mu\rho\, \partial_\sigma\rho}{\Lambda^2} \bigg)
- g_{\nu\sigma}\, \bigg( \frac{\nabla_{\lambda}\partial^\mu\rho}{\Lambda} + \frac{\partial^\mu\rho\, \partial_\lambda\rho}{\Lambda^2} \bigg)
\nonumber \\
&& +\,
{\delta^\mu}_\sigma\,\bigg( \frac{\nabla_{\lambda}\partial_\nu\rho}{\Lambda} + \frac{\partial_\nu\rho\,\partial_\lambda\rho}{\Lambda^2} \bigg) -
{\delta^\mu}_\lambda\, \bigg( \frac{\nabla_{\sigma}\partial_\nu\rho}{\Lambda} + \frac{\partial_\nu\rho\,\partial_\sigma\rho}{\Lambda^2} \bigg) +
\bigg( {\delta^\mu}_\lambda\, g_{\nu\sigma} - {\delta^\mu}_\sigma\, g_{\nu\lambda} \bigg)\,
\frac{(\partial\rho)^2}{\Lambda^2} \, ,
\nonumber \\
\hat R_{\mu\nu}
&=&
R_{\mu\nu} - g_{\mu\nu}\, \bigg( \frac{\Box\rho}{\Lambda}
- (d-2)\,\frac{(\partial\rho)^2}{\Lambda^2}\bigg)
- (d-2)\, \bigg( \frac{\nabla_\mu \partial_\nu\rho}{\Lambda} + \frac{\partial_\mu\rho\,\partial_\nu\rho}{\Lambda^2} \bigg)\, ,
\nonumber \\
\hat R
&\equiv&
\hat g^{\mu\nu}\, \hat R_{\mu\nu} =
e^{\frac{2\,\rho}{\Lambda}}\bigg[ R - 2\, (d-1)\, \frac{\Box \rho}{\Lambda}
+ (d-1)\,(d-2)\, \frac{(\partial\rho)^2}{\Lambda^2} \bigg]\, .
\label{Ch3GaugeRiemann}
\end{eqnarray}
When specialized to the case of Weyl transformations, for which $ \delta_{W} g_{\mu\nu} = 2 \sigma g_{\mu\nu}$,
the variations of the Christoffel symbols and their covariant derivatives (\ref{Ch3deltaChristoffel})
are
\begin{eqnarray} \label{Ch3deltaWeylChristoffel}
\delta_{W} \Gamma^\alpha_{\beta\gamma}
&=&
- g_{\beta\gamma}\, \partial^\alpha \sigma + {\delta_\beta}^\alpha\, \partial_\gamma \sigma + {\delta_\gamma}^\alpha\, \partial_\beta\sigma
\quad \Rightarrow \quad \delta_{W} \Gamma^\alpha_{\alpha\gamma} = d\, \partial_\gamma \sigma \, , \nonumber \\
\nabla_\lambda \delta_{W} \Gamma^\alpha_{\beta\gamma}
&=&
- g_{\beta\gamma}\, \nabla_\lambda\partial^\alpha\sigma + {\delta_\beta}^\alpha\, \nabla_\lambda\partial_\gamma\sigma
+ {\delta_\gamma}^\alpha\, \nabla_\lambda\partial_\beta\sigma
\quad \Rightarrow \quad \delta_{W} \nabla_{\lambda}\Gamma^\alpha_{\alpha\gamma} = d\, \nabla_\lambda \partial_\gamma\sigma \, .
\end{eqnarray}
Using the Palatini identity
\begin{equation} \label{Ch3Palatini}
\delta {R^\alpha}_{\beta\gamma\lambda} =
\nabla_{\lambda}(\delta\Gamma^\alpha_{\beta\gamma}) - \nabla_{\gamma}(\delta\Gamma^\alpha_{\beta\lambda})
\quad \Rightarrow \quad
\delta R_{\alpha\beta} =
\nabla_{\alpha} (\delta\Gamma^\lambda_{\beta\lambda}) - \nabla_{\lambda}(\delta\Gamma^\lambda_{\alpha\beta})
\end{equation}
we obtain the expressions for the Weyl variations of the Riemann and Ricci tensors
\begin{eqnarray} \label{Ch3deltaWeylRiemann}
\delta_{W} {R^\alpha}_{\beta\gamma\delta}
&=&
g_{\beta\delta}\, \nabla_{\gamma}\partial^\alpha\sigma
- g_{\beta\gamma}\, \nabla_{\delta}\partial^\alpha \sigma
+ {\delta_{\gamma}}^{\alpha}\, \nabla_{delta}\partial_\beta\sigma
- {\delta_{delta}}^{\alpha}\, \nabla_{\gamma}\partial_\beta\sigma \, , \nonumber \\
\delta_{W} R_{\alpha\beta}
&=&
g_{\alpha\beta}\, \Box \sigma + (d-2)\, \nabla_{\alpha}\partial_\beta\sigma \, .
\end{eqnarray}
It is also customary to introduce the Cotton tensor,
\begin{equation} \label{Ch4Cotton}
\tilde{C}_{\alpha\beta\gamma} = \nabla_{\gamma} K_{\alpha\beta} - \nabla_{\beta} K_{\alpha\gamma}\, ,
\quad \text{where}\, \quad
K_{\alpha\beta} = \frac {1}{d-2}\, \bigg( R_{\alpha\beta} - \frac{g_{\alpha\beta}}{2\,(d-1)}\, R \bigg)\, .
\end{equation}
Using (\ref{Ch3deltaWeylChristoffel})-(\ref{Ch3deltaWeylRiemann})
one can easily show that the variation of the Cotton tensor is simply given by
\begin{equation} \label{Ch4deltaWeylCotton}
\delta_{W} \tilde{C}_{\alpha\beta\gamma} = - \partial_\lambda\sigma\, {C^\lambda}_{\alpha\beta\gamma}\, ,
\end{equation}
which is expressed in terms of the Weyl tensor.
\section{Weyl invariants and Euler densities in $2$, $4$ and $6$ dimensions}\label{Geometrical}
It is well known that the object one has to deal with in order to construct Weyl-invariant objects for general dimensions $d$
is the traceless part of the Riemann tensor, called the Weyl tensor, defined by
\begin{equation} \label{WeyldDef}
C_{\alpha\beta\gamma\delta} = R_{\alpha\beta\gamma\delta} -
\frac{1}{d-2}( g_{\alpha\gamma} \, R_{\delta\beta} + g_{\alpha\delta} \, R_{\gamma\beta}
- g_{\beta\gamma} \, R_{\delta\alpha} - g_{\beta\delta} \, R_{\gamma\alpha} ) +
\frac{1}{(d-1)(d-2)} \, ( g_{\alpha\gamma} \, g_{\delta\beta} + g_{\alpha\delta} \, g_{\gamma\beta}) R\, .
\end{equation}
This object enjoys the same symmetry properties of the Riemann tensor, i.e.
\begin{equation}
C_{\alpha\beta\gamma\delta} = - C_{\b\a\g\d}= C_{\b\a\d\g} = C_{\d\g\b\a} \, ,
\end{equation}
and, moreover, is traceless with respect to any couple of its indices.
It is invariant under Weyl scalings of the metric
\begin{equation}
\delta_W {C^{\a}}_{\b\g\d} = 0 \, .
\end{equation}
It is apparent from (\ref{WeyldDef}), it is not defined for $d=2$,so that no Weyl-invariant object depending only on the metric
can be built out of it.
In $4$ dimensions, the only quantity which is Weyl invariant, when multiplied by $\sqrt{-g}$,
is the Weyl tensor squared, which is given by
\begin{equation} \label{Ch1Weyl}
F \equiv
C^{\alpha\beta\gamma\delta}C_{\alpha\beta\gamma\delta} =
R^{\alpha\beta\gamma\delta}R_{\alpha\beta\gamma\delta} - 2\, R^{\alpha\beta}R_{\alpha\beta} + \frac{1}{3}R^2 \, .
\end{equation}
Its d-dimensional version, which we call $F_d$, is instead
\begin{equation} \label{Ch1Weyld}
F_d \equiv
C^{\alpha\beta\gamma\delta}C_{\alpha\beta\gamma\delta} =
R^{\alpha\beta\gamma\delta}R_{\alpha\beta\gamma\delta} - \frac{4}{d-2}R^{\alpha\beta}R_{\alpha\beta}
+ \frac{2}{(d-2)(d-1)}R^2
\end{equation}
There are three dimension-6 scalars that are Weyl invariant when multiplied by $\sqrt{g}$.
Their choice is not unique at all, as one can always take linear combinations of them.
In particular, the literature is full of different choices for the third one, which involves differential operators.
In this work we adopt the definition of $I_3$ given, for general dimensions, in \cite{Parker:1987}
\begin{eqnarray} \label{Ch4WeylInvd}
I^d_1 &\equiv&
C_{\mu\nu\alpha\beta}\, C^{\mu\rho\sigma\beta}\, {{C^\nu}_{\rho\sigma}}^\alpha =
\frac{d^2+d-4}{(d-1)^2\,(d-2)^3}\, K_1 - \frac{3\,(d^2+d-4)}{(d-1)\,(d-2)^3}\, K_2 + \frac{3}{2\,d^2-6\,d+4}\, K_3
\nonumber \\
&& \hspace{35mm}
+\, \frac{6\,d-8}{(d-2)^3}\, K_4 - \frac{3\,d}{(d-2)^2}\, K_5 - \frac{3}{d-2}\, K_6 + K_8
\nonumber \\
I^d_2 &\equiv&
C_{\mu\nu\alpha\beta}\, C^{\alpha\beta\rho\sigma}\, {C^{\mu\nu}}_{\rho\sigma} =
\frac{8\,(2\,d-3)}{(d-1)^2\,(d-2)^3}\, K_1 + \frac{72-48\,d}{(d-1)\,(d-2)^3}\, K_2 + \frac{6}{d^2-3\,d+2}\, K3
\nonumber \\
&& \hspace{35mm}
+\, \frac{16\,(d-1)}{(d-2)^2}\, K_4 - \frac{24}{(d-2)^2}\, K_5 - \frac{12}{d-2}\, K_6 + K_7
\nonumber \\
I^d_3
&\equiv&
\frac{d-10}{d-2}\, \bigg( \nabla^{\alpha}C^{\beta\gamma\rho\sigma}\, \nabla_{\alpha}C_{\beta\gamma\rho\sigma}
- 4\, (d-2)\, \tilde{C}^{\gamma\rho\sigma}\, \tilde{C}_{\gamma\rho\sigma}\bigg) +
\frac{4}{d-2}\, \bigg(\square + \frac{2}{(d-1)}\, R\bigg)\,C^{\alpha\beta\rho\sigma}\,C_{\alpha\beta\rho\sigma}
\nonumber \\
&=&
\frac{16}{(d^2-3d+2)^2}\, K_{1} - \frac{32}{(d-1)\,(d-2)^2}\, K_{2} +
\frac{8}{d^2-3\,d+2} K_{3} + \frac{16}{(d-1)\,(d-2)^2}\, K_{9}
\nonumber \\
&& \hspace{-5mm}
-\, \frac{32}{(d-2)^2}\, K_{10} + \frac{8}{d-2}\, K_{11} + \frac{4\,(d-6)}{(d-1)\,(d-2)^2}\, K_{12}
+ \frac{88-12\,d}{(d-2)^2}\, K_{13} + K_{14} + \frac{8\,(d-10)}{(d-2)^2}\, K_{15}\, . \nonumber \\
\end{eqnarray}
We mention that the general problem of constructing all the possible Weyl-invariant coordinate scalars
depending on the metric tensor was solved in full generality in \cite{Fefferman:2007rka}.
The other quantity with which one can construct an integral which is Weyl invariant (in fact, a constant) for general dimensions
is the Euler density, defined, for a general even number of dimensions $d=2k$, as
\begin{equation} \label{EulerdDef}
E_{2k} = \frac{1}{2^k}\, \delta_{\mu_1 a_1\nu_1b_1\dots\mu_k a_k\nu_k b_k}\,
R^{\mu_1\nu_1\lambda_1 \kappa_1}\dots R^{\mu_k \nu_k a_k b_k}\, .
\end{equation}
The antisymmetric Kronecker symbol is defined by
\begin{equation}
\delta_{\nu_1 a_1 \nu_2 a_2\dots \nu_n a_n } =
n!\, \sum_{\mathcal P(a_1,\dots,a_n)}(-1)^{T_{\mathcal P}}\,
g_{\nu_1 \mathcal P(a_1)}\dots g_{\nu_n \mathcal P(a_n)}\, ,
\end{equation}
where $T_{\mathcal P}$ \`{e} is the number of inversions
the permutation $\mathcal P$ of the $n$ numbers $a_1, \dots a_n$ is made of.
By applying the general definition \ref{EulerdDef}, we find that in $2$, $4$ and $6$ dimensions respectively, it is given by the expressions
\begin{eqnarray}
E_2 &=& R \, , \nonumber \\
E_4 &\equiv&G =
R^{\alpha\beta\gamma\delta}R_{\alpha\beta\gamma\delta} - 4\,R^{\alpha\beta}R_{\alpha\beta} + R^2\, , \nonumber \\
E_6
&=&
\frac{21}{100}\, R^3 - \frac{27}{20}\, R\, R^{\mu\nu}\, R_{\mu\nu}
+ \frac{3}{2}\, {R_\mu}^\nu\,{R_\nu}^\alpha\, {R_\alpha}^\mu
+ 4\, C_{\mu\nu\rho\sigma}\,{C^{\mu\nu}} _{\alpha\beta}\, C^{\rho\sigma\alpha\beta}
\nonumber \\
&&
-\, 8\, C_{\mu\nu\rho\sigma}\,C^{\mu\alpha\rho\beta}\, {{{C^\nu}_\alpha}^\sigma}_\beta
-6\, R_{\mu\nu}\, C^{\mu\alpha\rho\sigma}\, {C^\nu}_{\alpha\rho\sigma}
+\frac{6}{5}\, R\, C^{\mu\nu\rho\sigma}\, C_{\mu\nu\rho\sigma}
- 3\, R^{\mu\nu}\, R^{\rho\sigma}\, C_{\mu\rho\sigma\nu} \, .
\label{VariousEuler}
\end{eqnarray}
Of course, in the last expression $C$ is the Weyl tensor in $6$ dimensions.
\section{Functional derivation of Riemann-quadratic integrals} \label{Ch1FunctionalIntegral}
In this appendix we explicitly show how to evaluate the functional variation of the most general integral
which is quadratic in the Riemann tensor and its contractions, which we call $\mathcal{I}(a,b,c)$,
\begin{equation}
\mathcal{I}(a,b,c)
\equiv
\int\,d^d x\,\sqrt{g}\, K\,
\equiv
\int\,d^d x\,\sqrt{g}\,
\big(a\,R^{\alpha\beta\gamma\delta}R_{\alpha\beta\gamma\delta} + b\,R^{\alpha\beta}R_{\alpha\beta} + c\, R^2 \big)\, ,
\end{equation}
needed to compute the counterterms found in section \ref{Ch1Renormalization}. \\
The same techniques apply to integrals which are higher order in the Riemann tensor,
as the one listed below for the case of $6$ dimensions.
Our index conventions for the Riemann and Ricci tensors are those in (\ref{Ch1Tensors}).
We have
\begin{eqnarray}
\delta (R^{\alpha\beta\gamma\delta}R_{\alpha\beta\gamma\delta})
&=&
\delta(g_{\alpha\sigma}g^{\beta\eta}g^{\gamma\zeta}g^{\delta\rho}
{R^\alpha}_{\beta\gamma\delta}{R^\sigma}_{\eta\zeta\rho}) \nonumber\\
&=&
\delta (g_{\alpha\sigma}g^{\beta\eta}g^{\gamma\zeta}g^{\delta\rho})
{R^\alpha}_{\beta\gamma\delta}{R^\sigma}_{\eta\zeta\rho}
+ g_{\alpha\sigma}g^{\beta\eta}g^{\gamma\zeta}g^{\delta\rho}\delta
({R^\alpha}_{\beta\gamma\delta}{R^\sigma}_{\eta\zeta\rho}) \nonumber\\
&=&
\delta (g_{\alpha\sigma}g^{\beta\eta}g^{\gamma\zeta}g^{\delta\rho})
{R^\alpha}_{\beta\gamma\delta}{R^\sigma}_{\eta\zeta\rho}
+ 2\,\delta ({R^\alpha}_{\beta\gamma\delta}){R_\alpha}^{\beta\gamma\delta}\, ,
\end{eqnarray}
Using (\ref{Ch1Tricks}) and (\ref{Ch1Tricks2}) and the product rule for derivatives one easily finds out that
the variation can be written at first as
\begin{eqnarray}
\delta \mathcal I(a,b,c)
&=& \int\,d^dx\,\sqrt{g}\,\bigg\{ \bigg[ \frac{1}{2}g^{\mu\nu}K - 2a\, R^{\mu\alpha\beta\gamma}{R^\nu}_{\alpha\beta\gamma}
- 2b\,R^{\mu\alpha}{R^\nu}_\alpha - 2c\,R R^{\mu\nu}\bigg]\delta g_{\mu\nu}\nonumber\\
&&\hspace{20mm}
+ \, 2a\, {R_\alpha}^{\beta\gamma\delta}\delta {R^\alpha}_{\beta\gamma\delta}
+ 2b\, R^{\alpha\beta}\delta R_{\alpha\beta} + 2c\, R\, g^{\alpha\beta}\, \delta R_{\alpha\beta}\bigg\}\, .
\end{eqnarray}
Now we have to exploit the Palatini identities
\begin{equation} \label{Ch1Palatini}
\delta {R^\alpha}_{\beta\gamma\delta}
=
(\delta\Gamma^\alpha_{\beta\gamma})_{;\delta} - (\delta\Gamma^a_{\beta\delta})_{;\gamma} \quad \Rightarrow \quad
\delta R_{\beta\delta}
=
(\delta\Gamma^\lambda_{\beta\lambda})_{;\delta} - (\delta\Gamma^\lambda_{\beta\delta})_{;\lambda}\, ,
\end{equation}
and the Bianchi identities,
\begin{eqnarray}\label{Ch1Bianchi}
R_{\alpha\beta\gamma\delta;\eta} + R_{\alpha\beta\eta\gamma;\delta} + R_{\alpha\beta\delta\eta;\gamma}
&=& 0
\quad \Rightarrow \quad
R_{\beta\delta;\eta} - R_{\beta\eta;\delta} + {R^\gamma}_{\beta\delta\eta;\gamma} = 0 \nonumber \\
\Rightarrow \quad
R_{;\delta}
&=&
2\,{R^\alpha}_{\delta;\alpha}
\quad \Leftrightarrow \quad
\big( R^{\alpha\beta} - \frac{1}{2}g^{\alpha\beta} R \big)_{;\beta} = 0 \, .
\end{eqnarray}
After an integration by parts and a reshuffling of indices we get
\begin{eqnarray}\label{Ch1deltaISecond}
\delta \mathcal I(a,b,c)
&=&
\int\,d^dx\,\sqrt{g}\,\bigg\{\bigg[ \frac{1}{2}g^{\mu\nu}K - 2\big( a\, R^{\mu\alpha\beta\gamma}{R^\nu}_{\alpha\beta\gamma}
+ b\,R^{\mu\alpha}{R^\nu}_\alpha + c\,R R^{\mu\nu}\big)\bigg]\delta g_{\mu\nu}\nonumber\\
&&
+ \,\left[ 4a\,g_{\beta\delta}\,g^{\gamma\eta}\,(\delta\Gamma^\delta_{\alpha\gamma})_{;\eta}
- (4a+2b)\,(\delta \Gamma^\gamma_{\alpha\beta})_{;\gamma}
+ (4c+2b)\, (\delta \Gamma^\lambda_{\alpha\lambda})_{;\beta}
- 4c\, g^{\eta\delta}\,g_{\gamma\alpha}\,(\delta \Gamma^\gamma_{\eta\delta})_{;\beta}\right]\,R^{\alpha\beta} \bigg\}.\nonumber\\
\end{eqnarray}
The variations of the Christoffel symbols and of their covariant derivatives
in terms of covariant derivatives of the metric tensors variations are
\begin{eqnarray}\label{Ch1deltaChristoffel}
\delta \Gamma^\alpha_{\beta\gamma}
&=&
\frac{1}{2}\,g^{\alpha\delta}\big[-(\delta g_{\beta\gamma})_{;\delta}
+ (\delta g_{\beta\delta})_{;\gamma} +(\delta g_{\gamma\delta})_{;\beta} \big]\, ,\nonumber\\
(\delta\Gamma^\alpha_{\beta\gamma})_{;\delta}
&=&
\frac{1}{2}\,g^{\alpha\eta}\big[-(\delta g_{\beta\gamma})_{;\eta;\delta} + (\delta g_{\beta\eta})_{;\gamma;\delta}
+ (\delta g_{\gamma\eta})_{;\beta;\delta} \big]\, .
\end{eqnarray}
Now we use them to rewrite (\ref{Ch1deltaISecond}) as
\begin{eqnarray}\label{Ch1Bivio}
\delta \mathcal I(a,b,c)
&=&
\int\,d^dx\,\sqrt{g}\bigg\{\bigg[ \frac{1}{2}g^{\mu\nu}K - 2\big( a\, R^{\mu\alpha\beta\gamma}{R^\nu}_{\alpha\beta\gamma}
+ b\,R^{\mu\alpha}{R^\nu}_\alpha + c\,R R^{\mu\nu}\big)\bigg]\delta g_{\mu\nu}\nonumber\\
&&\hspace{18,5mm}
+ \, \bigg[2a\,\big[-(\delta g_{\alpha\delta})_ {;\beta;\gamma} +(\delta g_{\alpha\beta})_ {;\gamma;\delta}
+ (\delta g_{\beta\delta})_ {;\alpha;\gamma} \big]\nonumber\\
&& \hspace{19mm}
- \, (2a + b)\,\big[- (\delta g_{\alpha\beta})_ {;\delta;\gamma} +(\delta g_{\alpha\delta})_ {;\beta;\gamma}
+ (\delta g_{\beta\delta})_ {;\alpha;\gamma} \big] + (2c + b)\, (\delta g_{\gamma\delta})_ {;\alpha;\beta}\nonumber\\
&& \hspace{19mm}
- \, 2c\, \big[- (\delta g_{\gamma\delta})_ {;\alpha;\beta} +(\delta g_{\alpha\delta})_ {;\gamma;\beta}
+ (\delta g_{\alpha\gamma})_ {;\delta;\beta} \big]\bigg]g^{\gamma\delta}\,R^{\alpha\beta} \bigg\}\, .
\end{eqnarray}
The presence of the factor $g^{cd}R^{ab}$ imposes two symmetry constraints on the terms in the last contribution in square brackets. By adding and subtracting $-(4a+2b)\,(\delta g_{ac})_{;d;b}$ we obtain the expression
\begin{eqnarray}
\delta \mathcal I(a,b,c)
&=& \int\,d^dx\,\sqrt{g}\,\bigg\{\bigg[\frac{1}{2}g^{\mu\nu}K
- 2\big( a\, R^{\mu\alpha\beta\gamma}{R^\nu}_{\alpha\beta\gamma}
+ b \, R^{\mu\alpha}{R^\nu}_\alpha + c\,R R^{\mu\nu}\big)\bigg]\delta g_{\mu\nu}\nonumber\\
&&\hspace{18,5mm}
+ \, \bigg[(4a+2b)\,\big[(\delta g_{\alpha\gamma})_ {;\beta;\delta}
- (\delta g_{\alpha\gamma})_ {;\delta;\beta}\big]+
(4a + b)(\delta g_{\alpha\beta})_ {;\gamma;\delta} + (4c + b)\, (\delta g_{\gamma\delta})_ {;\alpha;\beta}\nonumber\\
&& \hspace{18,5mm}
- \, (4a+2b+4c)\, (\delta g_{\alpha\gamma})_ {;\delta;\beta}\bigg]g^{\gamma\delta}\,R^{\alpha\beta} \bigg\}\, .
\label{Ch1HalfFirstFunctional}
\end{eqnarray}
The commutation of covariant derivatives allows us to write
\begin{eqnarray}
g^{\gamma\delta} \big[ (\delta g_{\alpha\gamma})_{;\beta;\delta}
- (\delta g_{\alpha\gamma})_{;\delta;\beta} \big]R^{\alpha\beta}
&=&
g^{\gamma\delta} \big[ -\delta g_{\alpha\sigma} {R^\sigma}_{\gamma\delta\beta}
- \delta g_{\gamma\sigma} {R^\sigma}_{\alpha\beta\delta} \big]R^{\alpha\beta}\nonumber\\
&=&
g^{\gamma\delta} \big[-s^{\mu\nu}_{\alpha\sigma}{R^\sigma}_{\gamma\beta\delta}
- s^{\mu\nu}_{c\sigma}{R^\sigma}_{\alpha\beta\delta}\big]R^{\alpha\beta} \,
\delta g_{\mu\nu}\nonumber\\
&=&
(- R^{\mu\alpha}{R^\nu}_\alpha + R^{\mu\alpha\nu\beta}R_{\alpha\beta}) \delta g_{\mu\nu}\, .
\end{eqnarray}
Inserting this back into (\ref{Ch1HalfFirstFunctional}) we get
\begin{eqnarray}
\delta \mathcal I(a,b,c) =
\nonumber\\
&& \hspace{-20mm}
\int\,d^dx\,\sqrt{g}\,\bigg\{\bigg[ \frac{1}{2}g^{\mu\nu}K
- 2a\, R^{\mu\alpha\beta\gamma}{R^\nu}_{\alpha\beta\gamma} + 4a\,R^{\mu\alpha}{R^\nu}_\alpha
-(4a+2b)\, R^{\mu\alpha\nu\beta}R_{\alpha\beta} - 2c \, R R^{\mu\nu}\bigg]\delta g_{\mu\nu}\nonumber\\
&&
+ \, \bigg[(4a + b)(\delta g_{\alpha\beta})_ {;\gamma;\delta}
+ (4c + b)\, (\delta g_{\gamma\delta})_ {;\alpha;\beta}
- (4a+2b+4c)\, (\delta g_{\alpha\gamma})_ {;\delta;\beta}\bigg]g^{\gamma\delta}\,R^{\alpha\beta}\bigg\}\, .
\nonumber\\
\end{eqnarray}
If the coefficients are $a = c = 1$ and $b=-4$, i.e. if the integrand is the Euler density,
the last three terms are zero. \\
All that is left to do is a double integration by parts for each one of the last three terms,
to factor out $\delta g_{\mu\nu}$.
This is easily performed and the final result can be written as
\begin{eqnarray} \label{Ch1Magic}
\frac{\delta}{\delta g_{\mu\nu}} \mathcal I(a,b,c)
&=&
\frac{\delta}{\delta g_{\mu\nu}} \int\,d^d x\,\sqrt{g\,}
\big( a\,R^{\alpha\beta\gamma\delta}R_{\alpha\beta\gamma\delta} + b\,R^{\alpha\beta}R_{\alpha\beta}
+ c\,R^2 \big)\nonumber\\
&=&
\sqrt{g}\, \bigg\{\frac{1}{2}g^{\mu\nu}K
- 2a\, R^{\mu\alpha\beta\gamma}{R^\nu}_{\alpha\beta\gamma}
+ 4a\,R^{\mu\alpha}{R^\nu}_\alpha -(4a+2b)\, R^{\mu\alpha\nu\beta}R_{\alpha\beta} - 2c \, R R^{\mu\nu}\nonumber\\
&& \hspace{8mm}
+ \, (4a + b)\,\Box{R^{\mu\nu}} + (4c + b)\,g^{\mu\nu}{R^{\alpha\beta}}_{;\alpha;\beta}
- (4a+2b+4c){{R^{\nu\beta}}_{;\beta}}^{;\mu}\bigg\}\, .
\end{eqnarray}
\section{Functional variations in $6$ dimensions}\label{Ch4Geometrical3}
The results for the trace anomaly in $6$ dimensions, presented in section \ref{Ch4Counterterms},
are obtained by computing the functional variations of the integrals of the $K_i$ in dimensional regularization,
which can be obtained with the same techniques employed in appendix \ref{Ch1FunctionalIntegral}.
A simple counting of the metric tensors needed to contract all the indices for any $K_i$ shows that
\begin{equation}
\delta_{W} \int d^dx\, \sqrt{g}\, K_i = \int d^dx\, \sqrt{g}\, \left[- \epsilon\, K_i + D(K_i) \right]\, \sigma
\, , \quad \epsilon = 6-d\, ,
\end{equation}
where the second term on the right hand side, $D(K_i)$, is a total derivative contribution.
We give the complete list of these terms below. We obtain
\begin{eqnarray} \label{Ch4FuncVar}
D(K_1) &=& 12\, (d-1)\,\nabla_\mu \left(R\,\partial^\mu R \right) \nonumber \\
D(K_2) &=& \nabla_\mu \bigg[ 4\,(d-1)\, R_{\nu\lambda}\, \nabla^\mu R^{\nu\lambda}
+ 2\,\left(d-2\right)\, R^{\mu\nu}\, \partial_\nu R + \left(d+2\right)\, R\,\partial^\mu R \bigg]\nonumber \\
D(K_3) &=& 4\, \nabla_\mu \bigg[ R\, \partial^\mu R + 2\, R^{\mu\nu}\, \partial_\nu R
+\left(d-1\right)\, R_{\nu\lambda\kappa\alpha}\, \nabla^\mu R^{\nu\lambda\kappa\alpha} \bigg] \nonumber \\
D(K_4) &=& 3\, \nabla_\mu \bigg[ \frac{d-2}{2}\,R^{\mu\nu}\, \partial_\nu R
+\left(d-2\right) R_{\nu\lambda}\, \nabla^\lambda R^{\mu\nu}
+2\, R_{\nu\rho}\, \nabla^\mu R^{\nu\rho} \bigg] \nonumber \\
D(K_5) &=& \nabla_\mu \bigg[ -R\, \partial^\mu R - R^{\mu\nu}\, \partial_\nu R
+2\, \left(d-1\right)\, R_{\nu\lambda}\, \nabla^\lambda R^{\mu\nu}
-2\,d\, R_{\nu\lambda}\,\nabla^\mu R^{\nu\lambda}
+ 2\,\left(d-2\right)\, R^{\mu\lambda\nu\kappa}\, \nabla_\kappa R_{\nu\lambda} \bigg] \nonumber \\
D(K_6) &=& \nabla_\mu \bigg[ 2\, R^{\mu\nu}\, \partial_\nu R + 4\, R_{\nu\lambda}\, \nabla^\mu R^{\nu\lambda}
+\frac{d+2}{2}\, R_{\nu\lambda\kappa\alpha}\, \nabla^\mu R^{\nu\lambda\kappa\alpha}
- 2\,d\, R^{\mu\lambda\nu\kappa}\, \nabla_\kappa R_{\nu\lambda} \bigg] \nonumber \\
D(K_7) &=& 6\, \nabla_\mu \bigg[ R_{\nu\lambda\kappa\alpha}\, \nabla^\mu R^{\nu\lambda\kappa\alpha}
+4\, R^{\nu\lambda\kappa\mu}\, \nabla_\lambda R_{\nu\kappa} \bigg] \nonumber \\
D(K_8) &=& 3\, \nabla_\mu \bigg[ \frac{1}{2}\,R_{\nu\lambda\kappa\alpha}\, \nabla^\mu R^{\nu\lambda\kappa\alpha}
+ 2\, R^{\nu\lambda}\, \left( \nabla_\lambda R_{\mu\nu} - \nabla_\mu R_{\nu\lambda} \right) \bigg] \nonumber \\
D(K_9) &=& \nabla_\mu \bigg[ 4\,\left(d-1\right)\, \partial^\mu \square R - \left(d-2\right)\, R\,\partial^\mu R \bigg] \nonumber \\
D(K_{10}) &=& \nabla_\mu \bigg[ 2\,\partial^\mu \square R + 2\,\left(d-2\right)\,\nabla_\nu \square R^{\mu\nu}
+2\, R^{\mu\nu}\, \partial_\nu R - 4\, R_{\nu\lambda}\, \nabla^\lambda R^{\mu\nu}
-\left(d-2\right)\, R_{\nu\lambda}\,\nabla^\mu R^{\nu\lambda} \bigg] \nonumber \\
D(K_{11}) &=& \nabla_\mu \bigg[ 8\, \nabla_\nu \square R^{\mu\nu}
- \left(d+2\right)\, R_{\nu\lambda\kappa\alpha}\, \nabla^{\mu}R^{\nu\lambda\kappa\alpha}
- 16\, R^{\mu\lambda\nu\kappa}\,\nabla_\kappa R_{\nu\lambda} \bigg] \nonumber \\
D(K_{12}) &=& 4\, \nabla_\mu \bigg[ R\,\partial^\mu R - \left(d-1\right)\, \partial^\mu \square R \bigg] \nonumber \\
D(K_{13}) &=& 2\,\ \nabla_\mu \bigg[ - \partial^\mu\square R - \left(d-2\right)\, \nabla_\nu \square R^{\mu\nu}
- R^{\mu\nu}\, \partial_\nu R + 2\, R_{\nu\lambda}\, \nabla^{\lambda} R^{\mu\nu}
+ 2\, R_{\nu\lambda}\, \nabla^\mu R^{\nu\lambda} \bigg] \nonumber \\
D(K_{14}) &=& 8\, \nabla_\mu \bigg[ - \nabla_\nu \square R^{\mu\nu}
+R_{\nu\lambda\kappa\alpha}\, \nabla^{\mu} R^{\nu\lambda\kappa\alpha}
+ 2\, R^{\mu\lambda\nu\kappa}\,\nabla_\kappa R_{\nu\lambda} \bigg] \nonumber \\
D(K_{15}) &=& \nabla_\mu \bigg[ - \partial^\mu\square R - 2\,\left(d-2\right)\, \nabla_\nu \square R^{\mu\nu}
-3\, R^{\mu\nu}\, \partial_\nu R + 6\, R_{\nu\lambda}\, \nabla^\lambda R^{\mu\nu} \nonumber \\
&& \hspace{10mm}
+ 2 \, R_{\nu\lambda}\, \nabla^\mu R^{\nu\lambda}
- 2 \, \left(d-2\right)\, R^{\mu\lambda\nu\kappa}\,\nabla_\kappa R_{\nu\lambda} \, \bigg]\, .
\end{eqnarray}
\section{ List of functional derivatives}
\label{Ch1Functionals}
We list here the contributions to the trace anomalies for three point function
coming from the elementary quadratic objects. They are given by
\begin{eqnarray}
\big[R_{\lambda\mu\kappa\nu}R^{\lambda\mu\kappa\nu}\big]^{\alpha\beta\rho\sigma}(p,q)
&=& p \cdot q\, \big[p \cdot q \big(\delta^{\alpha\rho}\delta^{\beta\sigma} + \delta^{\alpha\sigma}\delta^{\beta\rho}\big)
- \big(\delta^{\alpha\rho}p^{\sigma}q^{\beta} + \delta^{\alpha\sigma}p^{\rho}q^{\beta}\nonumber\\
&+& \delta^{\beta\rho}p^{\sigma}q^{\alpha} + \delta^{\beta\sigma}p^{\rho}q^{\alpha}\big)\big]
+ 2 \, p^{\rho}p^{\sigma}q^{\alpha}q^{\beta}\, , \nonumber
\end{eqnarray}
\begin{eqnarray}
\big[R_{\mu\nu}R^{\mu\nu}\big]^{\alpha\beta\rho\sigma}(p,q)
&=& \frac{1}{4} p \cdot q \big(\delta^{\alpha\rho}p^\beta q^\sigma + \delta^{\alpha\sigma}p^\beta q^\rho
+ \delta^{\beta\rho}p^\alpha q^\sigma + \delta^{\beta\sigma}p^\alpha q^\rho \big)\nonumber\\
&+&
\frac{1}{2}(p \cdot q)^2 \delta^{\alpha\beta}\delta^{\rho\sigma}
+ \frac{1}{4}p^2 q^2\big(\delta^{\alpha\rho}\delta^{\beta\sigma} + \delta^{\alpha\sigma}\delta^{\beta\rho}\big)\nonumber\\
&-&
\bigg[\frac{1}{4} p^2\big(q^\alpha q^\rho \delta^{\beta\sigma}+ q^\alpha q^\sigma \delta^{\beta\rho}
+ q^\beta q^\rho \delta^{\alpha\sigma}+ q^\beta q^\sigma \delta^{\alpha\rho}\big)\nonumber\\
&+&
\frac{1}{2}\delta^{\alpha\beta}\big( p\cdot q\,(p^\rho q^\sigma + p^\sigma q^\rho) - q^2 p^\rho p^\sigma \big)
+ (\alpha,\beta,p)\leftrightarrow(\rho,\sigma,q)\bigg]\, ,\nonumber
\end{eqnarray}
\begin{eqnarray}
\big[R^2\big]^{\alpha\beta\rho\sigma}(p,q)
&=& 2\big(p^\alpha p^\beta q^\rho q^\sigma - p^2 q^\rho q^\sigma \delta^{\alpha\beta}
- q^2 p^\alpha p^\beta \delta^{\rho\sigma} + p^2 \, q^2 \delta^{\alpha\beta}\delta^{\rho\sigma} \big)\, , \nonumber
\end{eqnarray}
\begin{eqnarray}\label{Ch1QuadraticFunctionals}
\big[\Box\,R\big]^{\alpha\beta\rho\sigma}(p,q)
&=&(p+q)^2\bigg\{-\frac{1}{2}\delta^{\alpha\beta}\big(p^\rho q^\sigma + p^\sigma q^\rho + 2\,p^\rho p^\sigma\big)
- \frac{1}{2}\delta^{\rho\sigma}\big(q^\alpha p^\beta + q^\beta q^\alpha + 2\,q^\alpha q^\beta \big)\nonumber\\
&+&
\frac{1}{2} p \cdot q \, \delta^{\alpha\beta}\delta^{\rho\sigma}
+\frac{1}{4}\big(p^\rho q^\beta \delta^{\alpha\sigma} + p^\rho q^\alpha \delta^{\beta\sigma}
+ p^\sigma q^\beta \delta^{\alpha\rho} + p^\sigma q^\alpha \delta^{\beta\rho}\big)\nonumber\\
&+&
\frac{1}{2}\bigg[\big(q^\rho p^\beta \delta^{\alpha\sigma} + q^\rho p^\alpha \delta^{\beta\sigma}
+ q^\sigma p^\beta \delta^{\alpha\rho} + q^\sigma p^\alpha \delta^{\beta\rho}\big)\nonumber\\
&+&
\delta^{\alpha\rho}\big(p^\beta p^\sigma + q^\beta q^\sigma \big)
+ \delta^{\alpha\sigma}\big(p^\beta p^\rho + q^\beta q^\rho \big)
+\delta^{\beta\rho}\big(p^\alpha p^\sigma + q^\alpha q^\sigma \big)\nonumber\\
&+&
\delta^{\beta\sigma}\big(p^\alpha p^\rho + q^\alpha q^\rho \big)
-\big(\delta^{\alpha\sigma}\delta^{\beta\rho} + \delta^{\alpha\rho}\delta^{\beta\sigma}\big)
\big(p^2 +q^2 + \frac{3}{2}p \cdot q\big)\bigg]\bigg\}\nonumber\\
&+&
\frac{1}{2}\big(p^2 \delta^{\alpha\beta} - p^\alpha p^\beta\big)\big(p \cdot q \, \delta^{\rho\sigma}
- (p^\rho q^\sigma + p^\sigma q^\rho) - 2\,p^\rho p^\sigma \big)\nonumber\\
&+& \frac{1}{2}\big(q^2 \delta^{\rho\sigma} - q^\sigma q^\rho\big)\big(p \cdot q \, \delta^{\alpha\beta}
- (p^\alpha q^\beta + p^\beta q^\alpha) - 2\,q^\alpha q^\beta \big)\, ,
\end{eqnarray}
\section{ Graviton interaction vertices}\label{Ch1Vertices}
Here we list the vertices which are needed for the momentum space computation of the $TTT$ correlator.
Notice that they are computed differentiating the first and second functional derivatives of the action,
because this allows to keep multi-graviton correlators symmetric (see \ref{Ch13PF}).
\begin{itemize}
\item{graviton - scalar - scalar vertex}
\\ \\
\begin{minipage}{95pt}
\includegraphics[scale=0.7]{figures/Tphiphi.eps}
\end{minipage}
\begin{minipage}{70pt}
\begin{eqnarray}
&\equiv& V_{\mathcal{S}\phi\phi}^{\mu\nu}(p,q)=
\frac{1}{2}\,p_\alpha \, q_\beta \, C^{\mu\nu\alpha\beta}
+\chi \bigg( \delta^{\mu\nu} \left( p + q \right)^2 - \left( p^\mu + q^\mu \right)\,\left( p^\nu + q^\nu \right) \bigg)\, ,
\nonumber
\end{eqnarray}
\end{minipage}
\item{graviton - fermion - fermion vertex}
\\ \\
\begin{minipage}{95pt}
\includegraphics[scale=0.7]{figures/Tpsipsi.eps}
\end{minipage}
\begin{minipage}{70pt}
\begin{eqnarray}
&\equiv& V_{\mathcal{S} \bar{\psi}\psi}^{\mu\nu}(p,q) =
\frac{1}{8} \, A^{\mu\nu\alpha\lambda}\, \gamma_\alpha \,\left(p_\lambda - q_\lambda \right)\, , \nonumber
\end{eqnarray}
\end{minipage}
\item{graviton - photon - photon vertex}
\\ \\
\begin{minipage}{95pt}
\includegraphics[scale=0.7]{figures/TAA.eps}
\end{minipage}
\begin{minipage}{70pt}
\begin{eqnarray}
&\equiv& V_{\mathcal{S}AA}^{\mu\nu\tau\omega}(p,q)
= \frac{1}{2}\,\bigg[ p \cdot q\, C^{\mu\nu\tau\omega} + D^{\mu\nu\tau\omega}(p,q)
+ \frac{1}{\xi}E^{\mu\nu\tau\omega}(p,q) \bigg] \, , \nonumber
\end{eqnarray}
\end{minipage}
\item{graviton - ghost - ghost vertex}
\\ \\
\begin{minipage}{95pt}
\includegraphics[scale=0.7]{figures/Tcc.eps}
\end{minipage}
\begin{minipage}{70pt}
\begin{eqnarray}
&\equiv& V_{\mathcal{S}\bar{c}c}^{\mu\nu}(p,q) = - V_{\mathcal{S}\phi\phi}^{\mu\nu}(p,q)\bigg|_{\chi=0} \, , \nonumber
\end{eqnarray}
\end{minipage}
\item{graviton - graviton - scalar - scalar vertex}
\\ \\
\begin{minipage}{95pt}
\includegraphics[scale=0.7]{figures/TTphiphi.eps}
\end{minipage}
\begin{minipage}{70pt}
\begin{eqnarray}
\equiv V_{\mathcal{S}\mathcal{S}\phi\phi}^{\mu\nu\rho\sigma}(p,q,l) &=&
\frac{1}{2}\, p\cdot q\, s^{\mu\nu\rho\sigma} - \frac{1}{4}\, G^{\mu\nu\rho\sigma}(p,q)
+ \frac{1}{4}\, \delta^{\rho\sigma}\, p_\alpha\, q_\beta\, C^{\mu\nu\alpha\beta} \nonumber \\
&& \hspace{-10mm}
+\, \chi \, \bigg\{\bigg[ \bigg(\delta^{\mu\lambda}\, \delta^{\alpha\kappa}\, \delta^{\nu\beta}
+ \delta^{\mu\alpha}\,\delta^{\nu\kappa}\,\delta^{\beta\lambda} - \delta^{\mu\kappa}\,\delta^{\nu\lambda}\,\delta^{\alpha\beta}
- \delta^{\mu\nu}\,\delta^{\alpha\lambda}\,\delta^{\beta\kappa}\bigg)\, s^{\rho\sigma}_{\lambda\kappa} \nonumber \\
&& \hspace{4mm}
+\, \frac{1}{2} \, \delta^{\rho\sigma} \, \bigg(\delta^{\mu\alpha}\,\delta^{\nu\beta} -
\delta^{\mu\nu}\, \delta^{\alpha\beta}\bigg)\bigg]
\left(p_\alpha \, q_\beta + p_\beta \, q_\alpha + p_\alpha \, p_\beta + q_\alpha \, q_\beta \right)\nonumber \\
&&
+\, \bigg[\bigg(\delta^{\mu\nu}\,\delta^{\alpha\beta} - \delta^{\mu\alpha}\,\delta^{\nu\beta} \bigg)
\big[\Gamma^\lambda_{\alpha\beta}\big]^{\rho\sigma}(l)\, i\, \left( \, p_\lambda + q_\lambda\right) \nonumber \\
&&
+\, \bigg(\delta^{\mu\alpha}\,\delta^{\nu\beta} - \frac{1}{2}\,\delta^{\mu\nu}\,\delta^{\alpha\beta}\bigg)\,
\big[R_{\alpha\beta}\big]^{\rho\sigma}(l) \, \bigg]\bigg\}\, , \nonumber
\end{eqnarray}
\end{minipage}
\item{graviton - graviton - fermion - fermion vertex}
\\ \\
\begin{minipage}{95pt}
\includegraphics[scale=0.7]{figures/TTpsipsi.eps}
\end{minipage}
\begin{minipage}{70pt}
\begin{eqnarray}
\equiv
V_{\mathcal{S}\mathcal{S} \bar{\psi}\psi}^{\mu\nu\rho\sigma}(p,q)
&=&
\frac{1}{16} \bigg[- 4 \, s^{\mu\nu\rho\sigma} - 2 \, \delta^{\mu\nu}\,s^{\alpha\lambda\rho\sigma}
+ 2 \, \delta^{\alpha\mu}\,s^{\nu\lambda\rho\sigma} + 2\, \delta^{\alpha\nu}\,s^{\mu\lambda\rho\sigma} \nonumber \\
&& \hspace{10mm}
+\, \delta^{\mu\lambda}\,s^{\alpha\nu\rho\sigma} + \delta^{\nu\lambda}\,s^{\alpha\mu\rho\sigma}
+ \delta^{\rho\sigma}\, A^{\mu\nu\alpha\lambda} \bigg]\,\gamma_{\alpha}\,(p_\lambda - q_\lambda) \, , \nonumber
\end{eqnarray}
\end{minipage}
\item{graviton - graviton - photon - photon vertex}
\\ \\
\begin{minipage}{95pt}
\includegraphics[scale=0.7]{figures/TTAA.eps}
\end{minipage}
\begin{minipage}{70pt}
\begin{eqnarray}
\equiv V_{\mathcal{S}\mathcal{S} AA}^{\mu\nu\rho\sigma\tau\omega}(p,q,l)
&=&
\frac{1}{2} \, \bigg\{
\bigg[ B^{\alpha\mu\rho\sigma\beta\lambda\gamma\nu} + \frac{1}{4}\, B^{\mu\nu\rho\sigma\alpha\lambda\gamma\beta} \bigg]\,
{F_{\alpha\beta\gamma\lambda}}^{\tau\omega} (p,q) \nonumber \\
&&
+\, \frac{1}{\xi}\bigg( H^{\mu\nu\rho\sigma\tau\omega}(p,q,l) + I^{\mu\nu\rho\sigma\tau\omega}(p,q,l) \bigg) \bigg\} \nonumber \\
&&
+\, \frac{1}{4}\, \delta^{\rho\sigma}\, \bigg[ p \cdot q \, C^{\mu\nu\tau\omega}
+ D^{\mu\nu\tau\omega}(p,q) + \frac{1}{\xi}E^{\mu\nu\tau\omega}(p,q) \bigg] \, , \nonumber
\end{eqnarray}
\end{minipage}
\item{graviton - graviton - ghost - ghost vertex}
\\ \\
\begin{minipage}{95pt}
\includegraphics[scale=0.7]{figures/TTcc.eps}
\end{minipage}
\begin{minipage}{70pt}
\begin{eqnarray}
\equiv V_{\mathcal{S}\mathcal{S}\bar{c}c}^{\mu\nu\rho\sigma}(p,q,l) =
- V_{\mathcal{S}\mathcal{S}\phi\phi}^{\mu\nu\rho\sigma}(p,q,l)\bigg|_{\chi=0}\, .
\nonumber
\end{eqnarray}
\end{minipage}
\end{itemize}
\vspace{0.5cm}%
We have simplified the notation by introducing, for convenience, the tensor structures
\begin{eqnarray}
A^{\mu\nu\alpha\lambda} = 2\,\delta^{\mu\nu}\,\delta^{\alpha\lambda}
- \delta^{\alpha\mu}\,\delta^{\lambda\nu} - \delta^{\alpha\nu}\,\delta^{\lambda\mu}
\nonumber
\end{eqnarray}
\begin{eqnarray}
B^{\alpha\mu\rho\sigma\beta\lambda\gamma\nu}
&=&
s^{\alpha\mu\rho\sigma} \, \delta^{\beta\lambda} \, \delta^{\gamma\nu}
+ s^{\beta\lambda\rho\sigma} \, \delta^{\alpha\mu} \, \delta^{\gamma\nu}
+ s^{\gamma\nu\rho\sigma} \, \delta^{\alpha\mu} \, \delta^{\beta\lambda}
\nonumber
\end{eqnarray}
\begin{eqnarray}
C^{\mu\nu\alpha\beta}
&=&
\delta^{\mu\alpha} \delta^{\nu\beta} + \delta^{\mu\beta} \delta^{\nu\alpha} - \delta^{\mu\nu} \delta^{\alpha\beta}\, ,
\nonumber
\end{eqnarray}
\begin{eqnarray}
D^{\mu\nu\rho\sigma} (p,q)
&=&
\delta^{\mu\nu} p^\sigma q^\rho + \delta^{\rho\sigma} \big(p^\mu q^\nu + p^\nu q^\mu\big)
- \delta^{\mu\sigma} p^\nu q^\rho - \delta^{\mu\rho} p^\sigma q^\nu
- \delta^{\nu\sigma}p^\mu q^\rho - \delta^{\nu\rho} p^\sigma q^\mu
\nonumber
\end{eqnarray}
\begin{eqnarray}
E^{\mu\nu\rho\sigma} (p, q)
&=&
\delta^{\mu\nu}\, \big[p^\rho p^\sigma + q^\rho q^\sigma + p^\rho q^\sigma \big]
- \big[\delta^{\nu\sigma} p^\mu p^\r + \delta^{\nu\rho}
q^\mu q^\sigma + \delta^{\mu\sigma} p^\nu p^\rho + \delta^{\mu\rho} q^\nu q^\sigma \big]\, ,
\nonumber
\end{eqnarray}
\begin{eqnarray}
F^{\mu\nu\rho\sigma\tau\omega} (p, q)
&=&
- \delta^{\tau\rho} \delta^{\omega\mu} p^{\sigma} q^{\nu} + \delta^{\tau\rho} \delta^{\omega\nu} p^{\sigma} q^{\mu}
+ \delta^{\tau\sigma} \delta^{\omega\mu} p^{\rho} q^{\nu} - \delta^{\tau\sigma} \delta^{\omega\nu} p^{\rho} q^{\mu}
+ (\tau,p) \leftrightarrow (\omega,q)
\nonumber
\end{eqnarray}
\begin{eqnarray}
G^{\mu\nu\rho\sigma}(p,q)
&=&
\delta^{\mu\sigma} \big[p^{\rho} q^{\nu} + q^{\rho} p^{\nu}\big]
+ \delta^{\nu\sigma}\big[p^{\rho} q^{\mu} + q^{\rho} p^{\mu}\big]+
\delta^{\mu\rho}\big[p^{\sigma} q^{\nu} + q^{\sigma} p^{\nu}\big]
+ \delta^{\nu\rho}\big[p^{\sigma} q^{\mu} + q^{\sigma} p^{\mu}\big]
\nonumber\\
&-&
\delta^{\mu\nu}\big[p^{\rho} q^{\sigma} + q^{\rho} p^{\sigma} \big]
\nonumber
\end{eqnarray}
\begin{eqnarray}
H^{\mu\nu\rho\sigma\tau\omega}(p,q,l)
&=&
\bigg[
\bigg(s^{\mu\omega\rho\sigma}\, \delta^{\nu\lambda} + s^{\nu\lambda\rho\sigma}\, \delta^{\mu\omega} \bigg)\, p_\lambda\, p^\tau
+ \delta^{\mu\omega}\,\bigg(s^{\lambda\tau\rho\sigma}\, l^\nu + s^{\lambda\tau\rho\sigma}\, p^\nu\bigg)\, p_\lambda
\nonumber \\
&+&
\frac{1}{2} \delta^{\mu\omega}\,
\left(p + l\right)^\nu\, \bigg(- l^{\tau}\, \delta^{\rho\sigma} + 2\, l_\lambda\, s^{\tau\lambda\rho\sigma}\bigg)
+ (\mu \leftrightarrow \nu)\bigg] + (\tau,p) \leftrightarrow (\omega,q)
\nonumber
\end{eqnarray}
\begin{eqnarray}
I^{\mu\nu\rho\sigma\tau\omega}(p,q,l)
&=&
\delta^{\mu\nu}\,\bigg\{
\frac{1}{2}\, \delta^{\rho\sigma}\, l^\tau\, \left(p + q + l\right)^\omega
- s^{\lambda\tau\rho\sigma} \, \bigg[q^\omega\, p_\lambda + l_\lambda\, \left( p + q + l \right)^\omega \bigg]
- s^{\lambda\omega\rho\sigma} \, \bigg[p^\tau \, p_\lambda + q_\lambda\, \left( q + l \right)^\tau \bigg]
\bigg\}
\nonumber \\
&-&
s^{\mu \nu \rho \sigma}\, \bigg(p^\omega\, p^\tau + q^\omega\, p^\tau\bigg) + (\tau,p) \leftrightarrow (\omega,q).
\nonumber
\end{eqnarray}
We have performed all our computations in the Feynman gauge ($\xi=1$)
The euclidean propagators of the fields in this case are
\begin{eqnarray}
\left\langle\phi \, \phi \right\rangle (p)
&=&
\frac{1}{p^2}
\nonumber\\
\left\langle \bar{\psi}\, \psi \right\rangle (p)
&=&
\frac{p\cdot\gamma}{p^2} \, .
\nonumber \\
\left\langle A^\mu \, A^\nu \right\rangle (p)
&=&
\frac{\delta^{\mu\nu}}{p^2} \, ,
\nonumber\\
\left\langle \bar{c}\, c \right\rangle (p)
&=&
- \frac{1}{p^2} \, .
\end{eqnarray}
\section{ Comments on the inverse mapping} \label{Ch1InverseTTT}
In this appendix we offer some calculational details in the derivation of the expression of the $TTT$
correlator in position space via the inverse mapping procedure. The remarks apply as well to any other correlator.\\
For example, eq. (\ref{Ch1ScalarTriangle}) refers to the contribution coming from the triangle diagram shown in fig.
\ref{Ch1Fig.diagramsTTT}. We assign the loop momentum $l$ to flow from the upper external point
($x_3$) to the lower one ($x_2$) on the right, the other two flows being determined by momentum conservation.
We denote the third external point as $x_1$.
For the scalar case, for instance, the complete $1$-loop triangle diagram is
\begin{equation}\label{Ch1TriangleLoop}
8\, \int\, \frac{d^dl}{(2\pi)^d} \,
\frac{V^{\mu\nu}_{\mathcal{S}\phi\phi}(l-q,-l-p)V^{\rho\sigma}_{\mathcal{S}\phi\phi}(l,-l+q)
V^{\alpha\beta}_{\mathcal{S}\phi\phi}(l+p,-l)}{l^2\,(l-q)^2\,(l+p)^2}
\end{equation}
The vertices are defined in appendix \ref{Ch1Vertices}.
The first argument in each vertex denotes the momentum of the incoming particle, the second argument
is the momentum of the outgoing one.
A typical term appearing in the loop integral is then
\begin{equation}\label{Ch1TypicalInverse3}
I\equiv\int\, \frac{d^dl}{(2\pi)^d} \,
\frac{l^{\mu}\,l^{\nu}\,(l+p)^{\rho}\,(l+p)^{\sigma}(l-q)^{\alpha}\,(l-q)^{\beta}}{l^2\,(l-q)^2\,(l+p)^2}\, .
\end{equation}
From (\ref{Ch1fund}) the propagators in configuration space are
\begin{equation}
\frac{1}{l^2\,(l-q)^2\,(l+p)^2} =
C(1)^3 \, \int\,d^d x_{12}\,d^d x_{23}\,d^d x_{31}\,
\frac{e^{i\,[l\cdot x_{23}+(l-q)\cdot x_{12}+(l+p)\cdot x_{31}]}}{(x^2_{12})^{d/2-1}\,(x^2_{23})^{d/2-1}\,(x^2_{31})^{d/2-1}}
\, ,
\end{equation}
where $C\alpha)$ has been defined in (\ref{Ch1fund}).
It is straightforward to see that (\ref{Ch1TypicalInverse3}) is given by
\begin{eqnarray}\label{Ch1TypicalInverse3Coord}
\int\, \frac{d^dl}{(2\pi)^d} \,
\frac{l^{\mu}\,l^{\nu}\,(l+p)^{\rho}\,(l+p)^{\sigma}(l-q)^{\alpha}\,(l-q)^{\beta}}{l^2\,(l-q)^2\,(l+p)^2}
&=&
\nonumber\\
&& \hspace {-70mm}
C(1)^3\,\int\, \frac{d^dl}{(2\pi)^d}\,d^d x_{12}\,d^d x_{23}\,d^d x_{31}\,
\frac{(-i)^6\,\partial^\mu_{23}\,\partial^\nu_{23}\,\partial^\rho_{31}\,\partial^\sigma_{31}\,
\partial^\alpha_{12}\,\partial^\beta_{12} \,e^{i\,[l\cdot x_{23}+(l-q)\cdot x_{12}+(l+p)\cdot x_{31}]}}
{(x^2_{12})^{d/2-1}\,(x^2_{23})^{d/2-1}\,(x^2_{31})^{d/2-1}} \nonumber \\
\, .
\end{eqnarray}
We can now integrate by parts moving the derivatives onto the propagators, getting
\begin{eqnarray}
I&=&
C(1)^3\,\int\, \frac{d^dl}{(2\pi)^d}\,d^d x_{12}\,d^d x_{23}\,d^d x_{31}\,
e^{i\,[l\cdot x_{23}+(l-q)\cdot x_{12}+(l+p)\cdot x_{31}]}
\nonumber\\
&\times&
i^6\,\partial^\mu_{23}\,\partial^\nu_{23}\,\partial^\rho_{31}\,\partial^\sigma_{31}\,
\partial^\alpha_{12}\,\partial^\beta_{12}\,\frac{1}{(x^2_{12})^{d/2-1}\,(x^2_{23})^{d/2-1}\,(x^2_{31})^{d/2-1}} \, .
\end{eqnarray}
The second line is immediately identified with the coordinate space Green's function.\\
This can be done for each term of (\ref{Ch1TriangleLoop}), justifying the rule quoted
in section \ref{Ch1InverseMappingTTT}, that we have used for all the inverse mappings of the work.
According to this the correlators in coordinate space can be obtained replacing the momenta in the vertices with ``$i$" times
the respective derivative which then act directly on the propagators after a partial integration.
The same arguments could be applied to the bubbles. Nevertheless, we have seen in \ref{Ch1InverseMappingTTT}
that derivatives of delta functions appear in the scalar case.
These are generated by the dependence of the $V^{\mu\nu\rho\sigma}_{\mathcal{S}\mathcal{S}\phi\phi}(p,q,l)$
from the momentum $l$ of the graviton bringing the pair of indices $\rho\sigma$ (see appendix \ref{Ch1Vertices}).
They are due to coupling of the scalar with derivatives of the metric through the Ricci scalar $R$ in
the improvement term (see eq. (\ref{Ch1scalarAction})) and state that the graviton feels the metric gradient.
We discuss this below, showing how to inverse-map the third bubble in fig. (\ref{Ch1Fig.diagramsTTT}), getting
(\ref{Ch1ScalarKBubble}). \\
This bubble can be seen as the ($x_2\rightarrow x_3$) limit of the triangle
and its diagrammatic momentum-space expression at $1$-loop is
\begin{equation}\label{Ch1KBubbleLoop}
\int\, \frac{d^dl}{(2\pi)^d} \,
\frac{V^{\mu\nu}_{\mathcal{S}\phi\phi}(l-q,-l-p)
V^{\alpha\beta\rho\sigma}_{\mathcal{S}\mathcal{S}\phi\phi}(l+p,-l+q,-q)}{(l-q)^2\,(l+p)^2}\, .
\end{equation}
As the two propagators are expressed by
\begin{equation}
\frac{1}{(l+q)^2\,(l+p)^2} =
C(1)^2 \, \int\,d^d x_{12}\,d^d x_{31}\,
\frac{e^{i\,[(l-q)\cdot x_{12}+(l+p)\cdot x_{31}]}}{(x^2_{12})^{d/2-1}\,(x^2_{31})^{d/2-1}} \, ,
\end{equation}
the dependence of the second vertex on $p$ cannot be ascribed to neither of them.\\
Two typical terms encountered in (\ref{Ch1KBubbleLoop}) are
\begin{eqnarray}\label{Ch1TypicalInverse2}
\int\, \frac{d^dl}{(2\pi)^d} \,
\frac{(l+p)^{\rho}\,(l+p)^{\sigma}(l-q)^{\alpha}\,(l-q)^{\beta}}{(l-q)^2\,(l+p)^2} \, ,
\nonumber\\
\int\, \frac{d^dl}{(2\pi)^d} \,
\frac{(l+p)^{\rho}\,(l+p)^{\sigma}(l-q)^{\alpha}\,p^{\beta}}{(l-q)^2\,(l+p)^2} \, .
\end{eqnarray}
The first one is treated at once restricting the procedure used for the three point function to the case of two
propagators.\\
For the second one, the following relation is immediately checked:
\begin{eqnarray}\label{Ch1TypicalInverse2Coord}
\int\, \frac{d^dl}{(2\pi)^d} \,
\frac{(l+p)^{\rho}\,(l+p)^{\sigma}(l-q)^{\alpha}\,p^{\beta}}{(l-q)^2\,(l+p)^2}
&=&
\nonumber\\
&& \hspace {-70mm}
C(1)^2\,\int\, \frac{d^dl}{(2\pi)^d}\,d^d x_{12}\,d^d x_{23}\,d^d x_{31}\,
\delta^{(d)}(x_{23})\,\frac{(-i)^4\,\partial^\rho_{31}\,\partial^\sigma_{31}\,\partial^\alpha_{12}
\,(\partial_{31}-\partial_{23})^\beta\,
e^{i\,[l\cdot x_{23}+(l-q)\cdot x_{12}+(l+p)\cdot x_{31}]}}
{(x^2_{12})^{d/2-1}\,(x^2_{31})^{d/2-1}} \, . \nonumber \\
\end{eqnarray}
Notice that an integration by parts brings in a derivative on the delta functions giving
\begin{eqnarray}
C(1)^2\,\int\, \frac{d^dl}{(2\pi)^d}\,d^d x_{12}\,d^d x_{23}\,d^d x_{31}\,
e^{i\,[l\cdot x_{23}+(l-q)\cdot x_{12}+(l+p)\cdot x_{31}]}\,
(i)^4\,\partial^\rho_{31}\,\partial^\sigma_{31}\,\partial^\alpha_{12}\,(\partial_{31}-\partial_{23})^\beta\,
\frac{\delta^{d}(x_{23})}{(x^2_{12})^{d/2-1}\,(x^2_{31})^{d/2-1}} \, . \nonumber \\
\end{eqnarray}
This approach has been followed in all the derivations of the expressions given in (\ref{Ch1InverseMappingTTT}).\\
The integration on $l$ brings about a $\delta^{(d)}(x_{12}+x_{23}+x_{31})$, so that it is natural to chose
the parameterization
\begin{equation}
x_{12} = x_1 - x_2\, , \quad x_{23} = x_2 - x_3\, , \quad x_{31} = x_3 - x_1\, .
\end{equation}
A more inviolved example is the 4-particle vertex. For instance the
$V_{\mathcal{S}\mathcal{S}\phi\phi}(i\, \partial_{31},- i\, \partial_{12},i \, (\partial_{12}-\partial_{23}))$
is obtained from $V_{\mathcal{S}\mathcal{S}\phi\phi}(p,q,l)$ with the functional replacements
\begin{equation}
p\to \hat{p}=i\, \partial_{31},\qquad q \to \hat{q}=- i\, \partial_{12}\qquad l\to \hat{l}=i \, (\partial_{12}-\partial_{23})
\end{equation}
giving
\begin{eqnarray}
V_{\mathcal{S}\mathcal{S}\phi\phi}^{\mu\nu\rho\sigma}
(i\, \partial_{31},- i\, \partial_{12},i \, (\partial_{12}-\partial_{23})) =
&&
\nonumber\\
&& \hspace{-60mm} \frac{1}{2}\, i\, \partial_{31}\cdot (-i)\,\partial_{12} s^{\mu\nu\rho\sigma}
- \frac{1}{4}\, G^{\mu\nu\rho\sigma}(i\,\partial_{31},-i\,\partial_{12})
+ \frac{1}{4}\, \delta^{\rho\sigma}\, i\,\partial_{31\,\alpha} \, (-i)\,\partial_{12\,\beta}\, C^{\mu\nu\alpha\beta}
\nonumber\\
&&\hspace{-60mm}
+ \, \chi \, \bigg\{\bigg[ \bigg(\delta^{\mu\lambda}\,\delta^{\alpha\kappa}\,\delta^{\nu\beta}
+ \delta^{\mu\alpha}\,\delta^{\nu\kappa}\,\delta^{\beta\lambda}
- \delta^{\mu\kappa}\,\delta^{\nu\lambda}\,\delta^{\alpha\beta}
- \delta^{\mu\nu}\,\delta^{\alpha\lambda}\,\delta^{\beta\kappa}\bigg)
s^{\rho\sigma}_{\lambda\kappa}
\nonumber \\
&& \hspace{-60mm}
+\, \frac{1}{2} \, \delta^{\rho\sigma} \, \bigg(\delta^{\mu\alpha}\,\delta^{\nu\beta} -
\delta^{\mu\nu}\,\delta^{\alpha\beta}\bigg)\bigg]
\left(i\,\partial_{31\,\alpha} \, (-i)\partial_{12\,\beta} + i\,\partial_{31\,\beta} \,(-i)\partial_{12\,\alpha}
+ i\,\partial_{31\,\alpha} \, i\,\partial_{31\,\beta} + (-i)\partial_{12\,\alpha} \,(-i)\partial_{12\,\beta} \right)
\nonumber \\
&&\hspace{-60mm}
- \, \bigg[\bigg(\delta^{\mu\alpha}\,\delta^{\nu\beta} - \delta^{\mu\nu}\,\delta^{\alpha\beta}\bigg)
\bigg(\big[\Gamma^\lambda_{\alpha\beta}\big]^{\rho\sigma}\left(i\,(\partial_{12}-\partial_{23})\right)\bigg)\,
(-i) \, \left( i \, \partial_{31\,\lambda} + \, (-i)\, \partial_{12\,\lambda}\right)
\nonumber \\
&& \hspace{-60mm}
+ \, \frac{1}{2} \, \bigg(\delta^{\mu\alpha}\,\delta^{\nu\beta} - \frac{1}{2}\,\delta^{\mu\nu}\,\delta^{\alpha\beta}\bigg)\,
\bigg( \big[R_{\alpha\beta}\big]^{\rho\sigma}\left(i\,(\partial_{12}-\partial_{23})\right)\bigg) \, \bigg]\bigg\} \, .
\end{eqnarray}
\section{ Regularizations and distributional identities} \label{Ch1Distributional}
We add few more comments and examples which illustrate the regularization that we have applied in the
computation of the various correlators.
The computation of the logarithmic integrals requires some care due to the distributional nature of some of these formulas.
As an example we consider the integrals
\begin{eqnarray}
H_1 = \int d^d l\, e^{i l\cdot x}\, \frac{\mu^{2 \omega}}{[l^2]^{1 + \omega}} \qquad
H_2 = \int d^d l\, e^{i l\cdot x}\, \frac{\mu^{2 \omega}}{[l^2]^{ \omega}} \qquad
H_3 = \int d^d l\, e^{i l\cdot x}\, \log\left(\frac{l^2}{\mu^2}\right)
\end{eqnarray}
We can relate them in the form
\begin{equation}
H_3 = -\frac{\partial}{\partial \omega}H_2 \bigg|_{\omega=0}
= \Box \left(\frac{\partial}{\partial \omega}H_1 \bigg|_{\omega=0}\right)
\end{equation}
In the two cases we get, using (\ref{Ch1fund})
\begin{equation}
-\frac{\partial}{\partial \omega}H_2 \bigg|_{\omega=0} = - \frac{(4\,\pi)^{d/2} \, \Gamma(d/2)}{(x^2)^{d/2}}
\end{equation}
and
\begin{equation}
\frac{\partial}{\partial \omega}H_1 \bigg|_{\omega=0}
= \frac{2^{d-2} \pi^{d/2} \Gamma(d/2-1)}{[x^2]^{d/2-1}}
\bigg( \log(x^2 \mu^2) + \gamma - \log 4 - \psi \left( \frac{d-2}{2}\right) \bigg)
\label{Ch1intermed}
\end{equation}
By redefining the regularization scale $\mu$ with eq. (\ref{Ch1massscale}) we clearly obtain from (\ref{Ch1intermed})
\begin{equation}
\int d^d l \frac{\log( l^2/\mu^2) e^{i l\cdot x}}{l^2} =
2^{d-2}\pi^{d/2} \Gamma(d/2-1) \frac{\log x^2 \bar{\mu}^2}{[x^2]^{d/2-1}}
\end{equation}
and
\begin{equation} \label{Ch1First}
H_3 = \Box \left(\frac{\partial}{\partial\omega} H_1\bigg|_{\omega=0} \right) =
2^{d - 2} \pi^{d/2} \Gamma(d/2-1) \Box \left( \frac{\log x^2 \bar{\mu}^2}{[x^2]^{d/2-1}}\right)
\end{equation}
The use of $H_2$ instead gives
\begin{equation}
H_3 = -\frac{\partial}{\partial \omega}H_2 \bigg|_{\omega=0}
= - \frac{2^d \pi^{d/2} \Gamma(d/2)}{[x^2]^{d/2}}
\end{equation}
Notice that this second relation coincides with (\ref{Ch1First}) away from the point $x=0$,
but differs from it right on the singularity, since
\begin{equation}
\Box \frac{\log x^2\mu^2}{[x^2]^{d/2-1}}
= - 2 \, (d-2) \bigg( \frac{\pi^{d/2}}{\Gamma(d/2)} \, \log( x^2\mu^2) \, \delta^d (x)
+ \frac{1}{[x^2]^{d/2}} \bigg)
\end{equation}
For this reason we take (\ref{Ch1First}) as the regularized expression of $H_3$, in agreement
with the standard approach of differential regularization. \\
The direct method discussed in the second part of the work, though very general and applicable to any correlator, introduces in
momentum space some logarithmic integrals which are more difficult to handle. They take the role of the ordinary master integrals of
perturbation theory. The scalar integrals needed for the tensor reduction of the logarithmic contributions in the text are defined
in (\ref{Ch1StandInt}).
After a shift of the momentum in the argument of the logarithm, a standard tensor reduction gives
\begin{eqnarray}
IL_\mu(0,p_1,p_2)
&=&
{CL}_1(p_1,p_2)\, p_{1\,\mu} + {CL}_2(p_1,p_2)\, p_{2\,\mu} \, ,
\nonumber\\
{CL}_1(p_1,p_2)
&=&
\frac{({p_1}^2 - p_1\cdot p_2){p_2}^2\, IL(0,p_1,p_2) + ({p_2}^2 - p_1\cdot p_2)\, {IL^\mu}_\mu(0,p_1,p_2)}
{2\, (p_1\cdot p_2)^2 - {p_1}^2\, {p_2}^2}\, ,
\nonumber \\
{CL}_2(p_1,p_2)
&=&
\frac{({p_2}^2 - p_1\cdot p_2){p_1}^2\, IL(0,p_1,p_2) + ({p_1}^2 - p_1\cdot p_2)\, {IL^\mu}_\mu(0,p_1,p_2)}
{2\, (p_1\cdot p_2)^2 - {p_1}^2\, {p_2}^2}\, .
\end{eqnarray}
To complete the computation of the $VVV$ correlator we need the explicit form of the logarithmic integrals
in terms of ordinary logarithmic and polylogarithmic functions.
We define
\begin{equation}
\mathcal I \equiv \int d^d l \frac{\log \left(l^2 / \mu^2\right)}{(l+p_1)^2 (l-p_2)^2} =
- \frac{\partial}{\partial \lambda}
\int d^dl\, \frac{\mu^{2 \lambda}}{(l^2)^\lambda\,(l+p_1)^2\,(l-p_2)^2}_{\lambda = 0} \, .
\end{equation}
The logarithmic integral is identified from the term of O($\lambda$) in the series expansion of the previous expression.
Because the coefficient in front of the parametric integral starts at this order, we just need to know the zeroth
order expansion of the integrand, which we separate into two terms. The first one is integrable
\begin{eqnarray}
I_1 = \int_0^1 d t \frac{t^{-\epsilon} (y t)^{1-\epsilon-\lambda}}{A(t)^{1-\epsilon}} =
\int_0^1 d t \frac{t^{-\epsilon} (y t)^{1-\epsilon}}{A(t)^{1-\epsilon}} + O(\lambda) \equiv I_1^{(0)} + O(\lambda)\, ,
\end{eqnarray}
while the last term has a singularity in $t=0$ which must be factored out and re-expressed in terms of a pole in $\lambda$
\begin{eqnarray}
I_2 &=& - \int_0^1 d t \frac{t^{-\epsilon} (x / t)^{1-\epsilon-\lambda}}{A(t)^{1-\epsilon}} =
- \frac{ x^{1-\epsilon-\lambda}}{\lambda} \int_0^1 d t \frac{1}{A(t)^{1-\epsilon}} \frac{d }{dt} t^\lambda \nonumber \\
&=&
- \frac{ x^{1-\epsilon-\lambda}}{\lambda} \bigg[ 1 - (\epsilon -1) \int_0^1 d t \frac{t^\lambda}{A(t)^{1-\epsilon}}
\left(\frac{1}{t-t_1} + \frac{1}{t-t_2}\right)\bigg] \nonumber \\
&=&
\frac{x^{1-\epsilon} }{\lambda} \bigg\{ -1 + (\epsilon -1) \int_0^1 d t \frac{1}{A(t)^{1-\epsilon}}
\left(\frac{1}{t-t_1} + \frac{1}{t-t_2}\right)\bigg] \bigg\} \nonumber \\
&+&
x^{1-\epsilon} \bigg[ \log x + (\epsilon -1) \int_0^1 d t \frac{\log \left(t/x\right)}{A(t)^{1-\epsilon}}
\left(\frac{1}{t-t_1} + \frac{1}{t-t_2}\right)\bigg] + O(\lambda) \equiv \frac{1}{\lambda} I_2^{(-1)} + I_2^{(0)}
+ O(\lambda)\, , \nonumber \\
\end{eqnarray}
where $t_1$ and $t_2$ are the two roots of $A(t) = y t^2 + (1-x-y)t + x$.
We are now able to write down the full $\lambda$-expansion of $J(1,1,\lambda)$ and to extract the logarithmic
integral $\mathcal I$
\begin{eqnarray}
\mathcal I = - \frac{\pi^{2-\epsilon} i^{1+2\epsilon}}{(p_3^2)^{\epsilon}} \frac{ \Gamma(1-\epsilon) \Gamma(2-\epsilon)
\Gamma(\epsilon )}{ \Gamma(2-2\epsilon)} \frac{1}{\epsilon-1} \bigg\{ I_1^{(0)} + I_2^{(0)} \bigg\} \,.
\end{eqnarray}
The previous expression can be expanded in $d=4-2\epsilon$ dimensions in which it manifests a $1/\epsilon$
pole of ultraviolet origin %
\begin{eqnarray}
\mathcal I =
\frac{\pi^{2-\epsilon} i^{1+2\epsilon}}{(p_3^2)^{\epsilon}} \left( -\frac{1}{\epsilon}
+ \gamma \right)\bigg[ A(x,y) + \epsilon \, B(x,y) \bigg] + O(\epsilon) \,,
\end{eqnarray}
where $A(x,y)$ and $B(x,y)$ are defined from the $\epsilon$-expansion of the two integrals $I_1^{(0)}$ and $I_2^{(0)}$ as
\begin{eqnarray}
A(x,y) &=& x \log x + \int_0^1 \frac{dt}{A(t)} \bigg[ y t - x \log \left( t/x \right)
\left(\frac{1}{t-t_1} + \frac{1}{t-t_2}\right)\bigg] \,, \\
B(x,y) &=& - x \log^2 x + \int_0^1 \frac{dt}{A(t)} \bigg[ y t \left( \log\left(t-t_1 \right)
+ \log\left(t-t_2 \right) - 2 \log t \right) \nonumber \\
&-& x \log\left(t/x \right) \left(\frac{1}{t-t_1} + \frac{1}{t-t_2}\right)
\left( \log\left(t-t_1 \right) + \log\left(t-t_2 \right) - \log \left( x/y\right) -1\right) \bigg].
\end{eqnarray}
We introduce here a systematic short-hand notation to denote the momentum-space integrals.
We define
\begin{eqnarray} \label{Ch1StandInt}
I_{\mu_1,\dots,\mu_n}(p)
&=&
\int d^d l \, \frac{l_{\mu_1}\, \dots \, l_{\mu_n}}{l^2\,(l+p)^2}\, ,
\nonumber\\
J_{\mu_1,\dots,\mu_n}(p_1,p_2)
&=&
\int d^d l \, \frac{l_{\mu_1}\, \dots \, l_{\mu_n}}{l^2\,(l+p_1)^2(l+p_2)^2}\, ,
\nonumber \\
IL_{\mu_{1}\dots\mu_{n}}(p_1,p_2,p_3)
&=&
\int d^dl \, \frac { l_{\mu_1}\, \dots l_{\mu_n} \, \log \left((l+p_1)^2/\mu^2\right)}{(l+p_2)^2(l+p_3)^2} \, ,
\nonumber\\
ILL_{\mu_{1}\dots\mu_{n}}(p_1,p_2,p_3,p_4)
&=&
\int d^d l \frac{l_{\mu_1}\, \dots l_{\mu_n} \,
\log \left((l+p_1)^2/\mu^2\right)\log \, \left((l+p_2)^2/\mu^2\right)}{(l+p_3)^2(l+p_4)^2}\, .
\end{eqnarray}
For correlators which are finite, the double logarithmic contributions will appear in combinations
that can be re-expressed in terms of ordinary Feynman integrals.
\section{ The Wess-Zumino action in $4$ dimensions by the Noether method}\label{Ch3WessZumino}
The dilaton effective action that we have obtained by the Weyl-gauging of the counterterms in chapter \ref{Recursive}
can be recovered also through an iterative technique, that we briefly review.
In this second approach one begins by requiring that the variation of the dilaton effective action under the Weyl transformations
be equal to the anomaly
\begin{eqnarray}
\delta_W \Gamma_{WZ}[g,\tau]
&=&
\int d^4x\, \sqrt{g}\, \sigma\, \bigg[ \beta_{a}\, \left( F - \frac{2}{3}\, \Box R \right) + \beta_b\, G \bigg]\, .
\label{Ch3VarWZ}
\end{eqnarray}
It is natural to start with the ansatz
\begin{equation} \label{Ch3ansatz1}
\Gamma^{(1)}_{WZ}[g,\tau] =
\int d^4 x\, \sqrt{g}\, \frac{\tau}{\Lambda} \, \bigg[ \beta_{a}\, \left( F - \frac{2}{3}\, \Box R \right) + \beta_b\, G \bigg] \, .
\end{equation}
As $\sqrt{g}\, F$ is Weyl invariant, the variation of $\tau$ saturates the $F$-contribution in (\ref{Ch3VarWZ}).
But $\sqrt{g}\, G$ and $\sqrt{g}\, \Box R$ are not conformally invariant. Their variations under Weyl scalings
introduce additional terms that must be taken into account. The general strategy is to compute the infinitesimal variation
of these terms and to add terms which are quadratic in the derivatives of the dilaton and that cancel this extra contributions.
But these, when Weyl-transformed, will generate additional terms which must be compensated in turn.
The iteration stops at the fourth order in the dilaton field.
We go through all the computation of the WZ action in some detail.
The piece of (\ref{Ch3ansatz1}) whose first Weyl variation is most easily worked out is the contribution $\sqrt{g}\,\tau\, \Box R$.
We integrate it by parts twice and find that
\begin{eqnarray}
\delta_W \int d^4x\, \sqrt{g}\, \frac{\tau}{\Lambda}\, \Box R
&=&
\delta_W \int d^4x\, \sqrt{g}\,g^{\mu\nu}\,g^{\rho\sigma}\, R_{\mu\nu}\, \frac{1}{\Lambda}\, \nabla_\sigma\partial_\rho\tau\, ,
\nonumber \\
&=&
\int d^4x\, \sqrt{g} \bigg( \Box\sigma\, R - \frac{1}{\Lambda}\, g^{\rho\sigma}\,\delta_W \Gamma^\lambda_{\rho\sigma}\,
\partial_\lambda\tau\, R + \frac{1}{\Lambda}\,\Box \tau\, g^{\mu\nu}\, \delta_W R_{\mu\nu} \bigg)\, .
\end{eqnarray}
Using (\ref{Ch3deltaWeylChristoffel}) and (\ref{Ch3deltaWeylRiemann}) this turns into
\begin{equation} \label{Ch3FirstWeylBoxR}
\delta_W \int d^4x\, \sqrt{g}\, \frac{\tau}{\Lambda}\, \Box R =
\int d^4 x\, \sqrt{g}\, \bigg( \sigma\, \Box R + \frac{2}{\Lambda}\, R\, \partial_\lambda\tau\, \partial^\lambda\sigma
+ \frac{6}{\Lambda}\, \Box\tau\Box\sigma \bigg) \, .
\end{equation}
Now we perform an infinitesimal Weyl variation of the contribution $\sqrt{g}\,\tau\,G$ and we find
\begin{eqnarray}
\delta_{W} \int d^4x\, \sqrt{g}\, \frac{\tau}{\Lambda}\, G
&=&
\int d^4 x\, \sqrt{g}\, \bigg\{ \sigma\, G +
\frac{2}{\Lambda}\,\tau \bigg[ {R_{\alpha}}^{\beta\gamma\delta}\delta_W {R^{\alpha}}_{\beta\gamma\delta}
- \left(4\,R^{\alpha\beta} - g^{\alpha\beta}R \right) \delta_W R_{\alpha\beta} \bigg] \bigg\}\, .
\end{eqnarray}
After using the algebraic symmetries of the Riemann and Ricci tensors and relabeling the indices we obtain
\begin{eqnarray}
\delta_{W} \int d^4x\, \sqrt{g}\, \frac{\tau}{\Lambda} \, G
&=&
\int d^4 x\, \sqrt{g}\, \bigg\{
\sigma\, G + \frac{\tau}{\Lambda}\, \bigg[ 4\,R^{\lambda\gamma\beta\alpha}
- 4\,\bigg( g^{\alpha\gamma}R^{\beta\lambda} + g^{\beta\lambda} R^{\alpha\gamma}
- 2\,g^{\lambda\gamma}\,R^{\alpha\beta} \bigg) \nonumber \\
&& \hspace{35mm}
+\, 2\, \bigg( g^{\alpha\gamma}g^{\beta\lambda} - g^{\alpha\beta}g^{\gamma\lambda} \bigg)\, R \bigg]\,
\nabla_\lambda\nabla_\beta(\delta_W g_{\alpha\gamma})
\bigg\}\, . \nonumber \\
\end{eqnarray}
Now we set $\delta_W g_{\alpha\gamma} = 2\, \sigma g_{\alpha\gamma}$ and after a double integration by parts we obtain
\begin{equation} \label{Ch3FirstWeylG}
\delta_{W} \int d^4x\, \sqrt{g}\, \frac{\tau}{\Lambda} \, G =
\int d^4 x\, \sqrt{g}\, \bigg\{ \sigma\, G
+ \frac{8}{\Lambda}\,\bigg[ \bigg( R^{\alpha\beta}- \frac{1}{2}\, g^{\alpha\beta}\,R \bigg)\, \partial_\alpha\sigma\, \partial_\beta\tau
\bigg] \bigg\} \, .
\end{equation}
Combined together, eqs. (\ref{Ch3FirstWeylBoxR}) and (\ref{Ch3FirstWeylG}) give the Weyl variation of the first ansatz (\ref{Ch3ansatz1})
\begin{eqnarray} \label{Ch3Var1}
\delta_{W} \Gamma^{(1)}_{WZ}[g,\tau]
&=&
\int d^4 x\, \sqrt{g}\, \bigg\{
\sigma\, \bigg[\beta_a\, \left(F - \frac{2}{3}\, \Box R \right) + \beta_b\, G \bigg]
\nonumber \\
&& \hspace{14mm}
+ \frac{1}{\Lambda}\, \bigg[ \beta_a\,
\left( \frac{4}{3}\,R\, \partial^\lambda\tau\, \partial_\lambda\sigma + 4\, \Box \tau\, \Box \sigma \right)
+\, 8\, \beta_b\, \partial_\alpha\sigma\, \partial_\beta\tau\, \left( R^{\alpha\beta} - \frac{g^{\alpha\beta}}{2}\, R \right) \bigg]
\bigg\}\, .
\end{eqnarray}
In order to cancel second line in the integrand in (\ref{Ch3Var1}) we correct with the second ansatz
\begin{eqnarray} \label{Ch3ansatz2}
\Gamma^{(2)}_{WZ}[g,\tau] = \Gamma^{(1)}_{WZ}[g,\tau]
+ \frac{1}{\Lambda^2}\,\int d^4 x\, \sqrt{g}\, \bigg\{
\beta_a\, \bigg(\frac{2}{3}\,R\, \left(\partial\tau\right)^2 + 2\, \left(\Box\tau\right)^2 \bigg)
- 4\, \beta_b\, \bigg( R^{\alpha\beta} - \frac{g^{\alpha\beta}}{2}\, R \bigg)\, \partial_\alpha\tau\, \partial_\beta\tau \bigg\}\, . \nonumber \\
\end{eqnarray}
We then find that the variation of this second ansatz is given by
\begin{eqnarray} \label{Ch3Var2}
\delta_{W} \Gamma^{(2)}_{WZ}[g,\tau]
&=&
\int d^4 x\, \sqrt{g}\, \bigg\{
\sigma\, \bigg[\beta_a\, \left(F - \frac{2}{3}\, \Box R \right) + \beta_b\, G \bigg]
+ \frac{1}{\Lambda^2}\, \bigg[ \beta_a\, \left( 8\, \Box \tau\, \partial^\lambda\tau\, \partial_\lambda\sigma
+ 4\, \left(\partial\tau\right)^2\, \Box \sigma \right)
\nonumber \\
&& \hspace{16mm}
+\, 8\, \beta_b\, \left( \left(\partial\tau\right)^2 \,\Box \sigma
- \partial_\alpha\tau\,\partial_\beta\tau\, \nabla^\beta\partial^\alpha\sigma \right)
\bigg]
\bigg\}\, .
\end{eqnarray}
It is then necessary to compensate for terms which are cubic in the dilaton.
The structure of the spurious contributions in (\ref{Ch3Var2}) suggests the third ansatz
\begin{eqnarray} \label{Ch3ansatz3}
\Gamma^{(3)}_{WZ}[g,\tau] = \Gamma^{(2)}_{WZ}[g,\tau]
- \frac{4}{\Lambda^3}\,\int d^4 x\, \sqrt{g}\,
\left( \beta_a + \beta_b \right)\, \left(\partial\tau\right)^2\,\Box \tau \, .
\end{eqnarray}
The variation of (\ref{Ch3ansatz3}) is
\begin{eqnarray} \label{Ch3Var3}
\delta_{W} \Gamma^{(3)}_{WZ}[g,\tau]
&=&
\int d^4 x\, \sqrt{g}\, \bigg\{
\sigma\, \bigg[\beta_a\, \left(F - \frac{2}{3}\, \Box R \right) + \beta_b\, G \bigg]
+ \frac{4}{\Lambda^2}\, \beta_b\, \left( 2\, \nabla_\beta\partial_\alpha\tau\, \partial^\beta\tau\, \partial^\alpha\sigma
+ \left(\partial\tau\right)^2\, \Box \sigma \right) \nonumber \\
&& \hspace{22mm}
-\, \frac{8}{\Lambda^3}\, \left(\beta_a + \beta_b\right)\, \partial^\alpha\tau\,\partial_\alpha\tau\, \partial^\beta\tau\, \partial_\beta\sigma
\bigg\} \nonumber \\
&=&
\int d^4 x\, \sqrt{g}\, \bigg\{
\sigma\, \bigg[\beta_a\, \left(F - \frac{2}{3}\, \Box R \right) + \beta_b\, G \bigg]
-\, \frac{8}{\Lambda^3}\, \left(\beta_a + \beta_b\right)\,
\partial^\alpha\tau\,\partial_\alpha\tau\, \partial^\beta\tau\,\partial_\beta\sigma \bigg\}\, ,
\end{eqnarray}
which finally allows to infer the structure of the complete WZ action
\begin{eqnarray} \label{Ch3ansatz4}
\Gamma_{WZ}[g,\tau] = \Gamma^{(3)}_{WZ}[g,\tau]
+ \frac{2}{\Lambda^4}\, \int d^4 x\, \sqrt{g}\, \left( \beta_a + \beta_b \right)\,
\left(\left(\partial\tau\right)^2\right)^2 \, .
\end{eqnarray}
eq. (\ref{Ch3ansatz4}) coincides with (\ref{Ch3Effective4d}) and it is easy to check that no more terms are needed to ensure
that (\ref{Ch3VarWZ}) holds.
\section{ The case $d=2$ as a direct check of the recursive formulae}\label{Ch42D}
Here we discuss how to cross-check the relations (\ref{Ch4DilIntStructure6}),
elaborating on the $2$-dimensional case, which is the easiest one of course. \\
It is clear that the expressions of the dilaton vertices $\mathcal{I}_n$ given in (\ref{Ch4DilIntStructure6}) do not depend on the working
dimensions. Therefore we take $d=2$ and check the agreement between the correlators that result from (\ref{Ch4DilIntStructure6})
and those found by a direct functional differentiation of the anomaly via the hierarchy (\ref{Ch3hier}). This provides a strong
check of the correctness of (\ref{Ch4DilIntStructure6}). In fact, the equation of the trace anomaly in 2 dimensions takes the form
\begin{equation} \label{Ch4TraceAnomaly2D}
\langle T \rangle = - \frac{c}{24\,\pi}\, R\, ,
\end{equation}
where $c = n_s + n_f$, with $n_s$ and $n_f$ being the numbers of free scalar and fermion fields respectively.
It is derived from the counterterm
\begin{equation}\label{Ch4Counterterm2D}
\Gamma_{\textrm{Ct}}[g] = - \frac{\mu^{\epsilon}}{\epsilon}\, \frac{c}{24\,\pi} \,\int d^d x\, \sqrt{g}\, R\, ,
\quad \epsilon = d - 2\, .
\end{equation}
The Weyl-gauging procedure for the integral of the scalar curvature gives
\begin{equation} \label{Ch4GaugeCT2D}
-\frac{\mu^{\epsilon}}{\epsilon}\, \int d^dx \, \sqrt{\hat g}\, \hat R =
- \frac{\mu^{\epsilon}}{\epsilon}\, \int d^dx \, \sqrt{g} \, R +
\int d^2 x \, \sqrt{g}\, \left[ \frac{\tau}{\Lambda}\, R + \frac{1}{\Lambda^2}\, \left(\partial\tau\right)^2 \right]\, .
\end{equation}
The second term in (\ref{Ch4GaugeCT2D}) is, modulo a constant, the Wess-Zumino action in $2$ dimensions,
\begin{equation}
\Gamma_{WZ}[g,\tau] = - \frac{c}{24\,\pi}\, \int d^2x\, \sqrt{g}\, \left[
\frac{\tau}{\Lambda}\, R + \frac{1}{\Lambda^2}\, \left(\partial\tau\right)^2 \right] \, ,
\end{equation}
from which we can extract the 2-dilaton amplitude according to (\ref{Ch3FuncDiffWZ})
\begin{equation}
\mathcal I_{2}(k_{1}\,-k_{1}) = \frac{1}{\Lambda^2}\, \langle T(k_{1}) T(-k_{1}) \rangle = \frac{c}{12\,\pi}\, k_{1}^2 \, .
\end{equation}
Starting from the 2-dilaton vertex, which is the only non-vanishing one,
exploiting (\ref{Ch4NoInt}) and inverting the remaining relations, we get the Green functions
\begin{eqnarray}
\langle T(k_{1}) T(k_{2}) T(k_{3}) \rangle
&=&
- \frac{c}{6\,\pi}\, \left( k_{1}^2 + k_{2}^2 + k_{3}^2 \right)\, ,
\nonumber \\
\langle T(k_{1}) T(k_{2}) T(k_{3}) T(k_{4}) \rangle
&=&
\frac{c}{\pi}\, \left( k_{1}^2 + k_{2}^2 + k_{3}^2 + k_{4}^2 \right)\, , \nonumber \\
\langle T(k_{1}) T(k_{2}) T(k_{3}) T(k_{4}) T(k_{5}) \rangle
&=&
-\frac{8\,c}{\pi}\, \left( k_{1}^2 + k_{2}^2 + k_{3}^2 + k_{4}^2 + k_{5}^2 \right)\, , \nonumber \\
\langle T(k_{1}) T(k_{2}) T(k_{3}) T(k_{4}) T(k_{5}) T(k_{6}) \rangle
&=&
\frac{80\,c}{\pi}\, \left( k_{1}^2 + k_{2}^2 + k_{3}^2 + k_{4}^2 + k_{5}^2 + k_{6}^2 \right)\, .
\end{eqnarray}
These results exactly agree with the combinations of completely traced multiple functional derivatives of the anomaly
(\ref{Ch4TraceAnomaly2D}) that one derives from (\ref{Ch3hier}), providing a consistency check of our recursive formulas
(\ref{Ch4DilIntStructure6}). \\
We also report the expression we obtain of the order-six correlator in $d=4$.
\begin{eqnarray}
&&
\langle T(k_{1}) T(k_{2}) T(k_{3}) T(k_{4}) T(k_{5}) T(k_{6}) \rangle =
32\, \Biggl\{
120\,\bigg(\beta_a + \beta_b\bigg)\, \Biggl[
\sum_{\mathcal T\left\{6,[(k_{i_1},k_{i_2}),(k_{i_3},k_{i_4})]\right\}}
k_{i_1}\cdot k_{i_2}\,k_{i_3}\cdot k_{i_4} \nonumber \\
&&
+\, f_{6}(k_{1},k_{2},k_{3},k_{4},k_{5},k_{6}) + f_{6}(k_{2},k_{1},k_{3},k_{4},k_{5},k_{6}) + f_{6}(k_{3},k_{1},k_{2},k_{4},k_{5},k_{6}) \nonumber \\
&&
+\, f_{6}(k_{4},k_{1},k_{2},k_{3},k_{5},k_{6}) + f_{6}(k_{5},k_{1},k_{2},k_{3},k_{4},k_{6}) + f_{6}(k_{6},k_{1},k_{2},k_{3},k_{4},k_{5}) \Biggr] \nonumber \\
&& \hspace{-10mm}
-\, \beta_a\, \Biggl[
\sum_{\mathcal T\left\{6,(k_{i_1},k_{i_2},k_{i_3},k_{i_4})\right\}}
\left(k_{i_1} + k_{i_2} + k_{i_3}+ k_{i_4} \right)^4
+ 4\, \sum_{\mathcal T\left\{6,(k_{i_1},k_{i_2},k_{i_3})\right\}} \left(k_{i_1} + k_{i_2} + k_{i_3} \right)^4
\nonumber \\
&&
+\, 11\, \sum_{\mathcal T\left\{6,(k_{i_1},k_{i_2})\right\}} \left(k_{i_1} + k_{i_2} \right)^4
+ 48\, \sum_{i=1}^{5} k_{i}^4
\Biggr]
\Biggr\}\, ,
\label{Ch4BuildingBlocks}
\end{eqnarray}
where we have introduced the compact notation
\begin{eqnarray}
f_{6}(k_a,k_b,k_c,k_d,k_e,k_f)
&=&
k_a^2\, \left( k_b \cdot k_c + k_b \cdot k_d + k_b \cdot k_e + k_b \cdot k_f + k_c \cdot k_d + k_c \cdot k_e
\right.
\nonumber \\
&& \hspace{5mm}
\left.
+\, k_c \cdot k_f + k_d \cdot k_e + k_d \cdot k_f + k_e \cdot k_f \right)\, .
\end{eqnarray}
\chapter{Appendix}\label{Effective dilaton}
\section{Dilaton interaction vertices}
\label{rules}
The Feynman rules used throughout the paper are collected here. We have
\begin{itemize}
\item{ dilaton - gauge boson - gauge boson vertex}
\\ \\
\begin{minipage}{95pt}
\includegraphics[scale=1.0]{figures/FR_dilVV.eps}
\end{minipage}
\begin{minipage}{70pt}
\begin{eqnarray}
= V^{\alpha\beta}_{\rho VV}(k_1,k_2) =
- \frac{2\, i}{\Lambda} \bigg\{M_V^2 \, \eta^{\alpha\beta}
- \frac{1}{\xi}\, \left( k_1^\alpha\, k_1^\beta + k_2^\alpha\, k_2^\beta + 2\, k_1^\alpha\, k_2^\beta \right) \bigg\}
\nonumber
\end{eqnarray}
\end{minipage}
\begin{eqnarray}
\label{FRdilVV}
\end{eqnarray}
where $V$ stands for the gluons or for the vector gauge bosons $A, Z$ and $W^{\pm}$ and,
if the gauge bosons are gluons, a color-conserving $\delta_{ab}$ matrix must be included.
\\ \\
\item{dilaton - fermion - fermion vertex}
\\ \\
\begin{minipage}{95pt}
\includegraphics[scale=1.0]{figures/FR_dilFF.eps}
\end{minipage}
\begin{minipage}{70pt}
\begin{eqnarray}
&=&
V_{\rho \bar\psi\psi}(k_1,k_2) =
- \frac{i}{2\,\Lambda} \, \bigg\{ 3 \, \left( k \! \! \! /_1 - k \! \! \! /_2 \right) + 8 \, m_f \bigg\}
\nonumber
\end{eqnarray}
\end{minipage}
\begin{eqnarray}
\label{FRdilFF}
\end{eqnarray}
If the fermions are quarks, the vertex must be multiplied by the identity color matrix $\delta_{a b}$. \\ \\
\item{dilaton - ghost - ghost vertex }
\\ \\
\begin{minipage}{95pt}
\includegraphics[scale=1.0]{figures/FR_dilghgh.eps}
\end{minipage}
\begin{minipage}{70pt}
\begin{eqnarray}
&=&
V_{\rho \bar c c}(k_1,k_2)
= - \frac{2 \, i}{\Lambda} \bigg\{ k_1 \cdot k_2 + 2 \, M_{\eta}^2\bigg\}
\nonumber
\end{eqnarray}
\end{minipage}
\begin{eqnarray}
\label{FRdilUU}
\end{eqnarray}
where $\eta$ denotes both the QCD ghost fields $c^{a}$ or the electroweak ghost fields $\eta^{+}$, $\eta^{-}$ ed $\eta^Z$. In the QCD
case one must include a color-conserving $\delta_{ab}$ matrix.
\\ \\
\item{dilaton - scalar - scalar vertex}
\\ \\
\begin{minipage}{95pt}
\includegraphics[scale=1.0]{figures/FR_dilSS.eps}
\end{minipage}
\begin{minipage}{70pt}
\begin{eqnarray}
=
V_{\rho SS}(k_1,k_2)
&=& - \frac{2 \, i}{\Lambda} \bigg\{ k_1 \cdot k_2 + 2 \, M_S^2 \bigg\}
\nonumber \\
&=&
\frac{i}{\Lambda} \,6\,\chi\, (k_1+k_2)^2 \nonumber
\end{eqnarray}
\end{minipage}
\begin{eqnarray}
\label{FRdilSS}
\end{eqnarray}
where $S$ stands for the Higgs $H$ and the Goldstones $\phi$ and $\phi^{\pm}$.
The first expression is the contribution coming from
the minimal energy-momentum tensor while the second is due to the term of improvement.
\\ \\
\item{dilaton - Higgs vertex}
\\ \\
\begin{minipage}{95pt}
\includegraphics[scale=1.0]{figures/FR_dilH.eps}
\end{minipage}
\begin{minipage}{70pt}
\begin{eqnarray}
\qquad
&=&
V_{I,\, \rho H }(k)
= - \frac{i}{\Lambda} \frac{12\,\chi\, s_w M_W}{e} \, k^2 \nonumber
\end{eqnarray}
\end{minipage}
\begin{eqnarray}
\label{FRdilH}
\end{eqnarray}
This vertex is derived from the term of improvement of the energy-momentum tensor and it is a feature of the electroweak symmetry
breaking because it is proportional to the Higgs vev.
\\ \\
\item{ dilaton - three gauge boson vertex}
\\ \\
\begin{minipage}{95pt}
\includegraphics[scale=1.0]{figures/FR_dilVVV.eps}
\end{minipage}
\begin{minipage}{70pt}
\begin{eqnarray}
&=&
V^{\alpha\beta\gamma}_{\rho V V V}
= 0
\nonumber
\end{eqnarray}
\end{minipage}
\begin{eqnarray}
\label{FRdilVVV}
\end{eqnarray}
where $V_1$, $V_2$, $V_3$ stand for gluon, photon, $Z$ and $W^{\pm}$.
\\ \\
\item{dilaton - gauge boson - scalar - scalar vertex }
\\ \\
\begin{minipage}{95pt}
\includegraphics[scale=1.0]{figures/FR_dilVSS.eps}
\end{minipage}
\begin{minipage}{70pt}
\begin{eqnarray}
&=&
V^\alpha_{\rho VSS}(k_2,k_3)
= - \frac{2\,i}{\Lambda}\, e \, \mathcal C_{V S_1 S_2} \, \left( k_2^\alpha - k_3^\alpha \right)
\nonumber
\end{eqnarray}
\end{minipage}
\begin{eqnarray}
\label{FRdilVSS}
\end{eqnarray}
with $\mathcal C_{V S_1 S_2}$ given by
\begin{eqnarray}
\mathcal C_{A\phi^{+}\phi^{-}} = 1 \qquad
\mathcal C_{Z\phi^{+}\phi^{-}} = \frac{c_w^2 - s_w^2}{2 s_w \, c_w} \qquad
\mathcal C_{Z H \phi} = \frac{i}{2 s_w \, c_w}. \nonumber
\end{eqnarray}
\\ \\
\item{dilaton - gauge boson - ghost - ghost vertex}
\\ \\
\begin{minipage}{95pt}
\includegraphics[scale=1.0]{figures/FR_dilVghgh.eps}
\end{minipage}
\begin{minipage}{70pt}
\begin{eqnarray}
&=&
V^\alpha_{\rho V \bar\eta\eta}(k_2)
= - \frac{2\,i}{\Lambda} \, \mathcal C_{V \eta} \, k_2^\alpha
\nonumber
\end{eqnarray}
\end{minipage}
\begin{eqnarray}
\label{FRdilVUU}
\end{eqnarray}
where $V$ denotes the $g^a$, $A$, $Z$ gauge bosons and $\eta$ the ghosts $c^b$, $\eta^{+}$, $\eta^{-}$.
The coefficients $\mathcal C$ are defined as
\begin{eqnarray}
\mathcal C_{g^a c^b} = f^{abc}\, g \qquad
\mathcal C_{A \eta^{+}} = e \qquad
\mathcal C_{A \eta^{-}} = -e \qquad
\mathcal C_{Z \eta^{+}} = e \, \frac{c_w}{s_w} \qquad
\mathcal C_{Z \eta^{-}} = - e \, \frac{c_w}{s_w}. \nonumber
\end{eqnarray}
\item{dilaton - gauge boson - gauge boson - scalar vertex}
\\ \\
\begin{minipage}{95pt}
\includegraphics[scale=1.0]{figures/FR_dilVVS.eps}
\end{minipage}
\begin{minipage}{70pt}
\begin{eqnarray}
&=&
V^{\alpha\beta}_{\rho VVS}
= - \frac{2}{\Lambda} \, e \, \mathcal C_{V_1 V_2 S} \, M_W \, \eta^{\alpha\beta}
\nonumber
\end{eqnarray}
\end{minipage}
\begin{eqnarray}
\label{FRdilVVS}
\end{eqnarray}
where $V$ stands for $A$, $Z$ and $W^{\pm}$ and $S$ for $\phi^{\pm}$
and $H$. The coefficients are defined as
\begin{eqnarray}
\mathcal C_{A W^{+} \phi^{-}} = 1 \qquad
\mathcal C_{A W^{-} \phi^{+}} = -1 \qquad
\mathcal C_{Z W^{+} \phi^{-}} = - \frac{s_w}{c_w} \qquad \nonumber \\
\mathcal C_{Z W^{-} \phi^{+}} = \frac{s_w}{c_w} \qquad
\mathcal C_{Z Z H} = - \frac{i}{s_w \, c_w^2} \qquad
\mathcal C_{W^{+} W^{-} H} = - \frac{i}{c_w}. \nonumber
\end{eqnarray}
\\ \\
\item{dilaton - scalar - ghost - ghost vertex}
\\ \\
\begin{minipage}{95pt}
\includegraphics[scale=1.0]{figures/FR_dilSghgh.eps}
\end{minipage}
\begin{minipage}{70pt}
\begin{eqnarray}
&=&
V{\rho S\bar\eta\eta}
= - \frac{4 \, i}{\Lambda}\, e \, \mathcal C_{S \eta} \, M_W
\nonumber
\end{eqnarray}
\end{minipage}
\begin{eqnarray}
\label{FRdilSUU}
\end{eqnarray}
where $S = H$ and $\eta$ denotes $\eta^{+}$, $\eta^{-}$ and $\eta^{z}$. The vertex is defined with the coefficients
\begin{eqnarray}
\mathcal C_{H \eta^{+}} = \mathcal C_{H \eta^{-}} = \frac{1}{2 s_w} \qquad \mathcal C_{H \eta^{z}} = \frac{1}{2 s_w \, c_w}.
\nonumber
\end{eqnarray}
\item{dilaton - three scalar vertex}
\\ \\
\begin{minipage}{95pt}
\includegraphics[scale=1.0]{figures/FR_dilSSS.eps}
\end{minipage}
\begin{minipage}{70pt}
\begin{eqnarray}
&=&
V_{\rho SSS}
= - \frac{4 \, i}{\Lambda} \, e \, \mathcal C_{S_1 S_2 S_3}
\nonumber
\end{eqnarray}
\end{minipage}
\begin{eqnarray}
\label{FRdilSSS}
\end{eqnarray}
with $S$ denoting $H$, $\phi$ and $\phi^{\pm}$. We have defined the coefficients
\begin{eqnarray}
\mathcal C_{H \phi \phi} = \mathcal C_{H \phi^{+} \phi^{-}} = \frac{1}{2 s_w \, c_w}
\frac{M_H^2}{M_Z} \qquad \mathcal C_{H H H} = \frac{3}{2 s_w \, c_w} \frac{M_H^2}{M_Z}.
\nonumber
\end{eqnarray}
\\ \\
\item{dilaton - scalar - fermion - fermion vertex}
\\ \\
\begin{minipage}{95pt}
\includegraphics[scale=1.0]{figures/FR_dilSFF.eps}
\end{minipage}
\begin{minipage}{70pt}
\begin{eqnarray}
&=&
V_{\rho S\bar\psi\psi}
= - \frac{2\,i}{\Lambda} \, \frac{e}{s_w \, c_w} \frac{m_f}{M_Z}
\nonumber
\end{eqnarray}
\end{minipage}
\begin{eqnarray}
\label{FRdilSFF}
\end{eqnarray}
where $S$ is only the Higgs scalar $H$.
\\ \\
\item{dilaton - gluon - fermion - fermion vertex}
\\ \\
\begin{minipage}{95pt}
\includegraphics[scale=1.0]{figures/FR_dilgFF.eps}
\end{minipage}
\begin{minipage}{70pt}
\begin{eqnarray}
&=&
V^{a\,\alpha}_{\rho g\bar\psi\psi}
= \frac{3\, i}{\Lambda}\, g \, T^a \, \gamma^\alpha\, .
\nonumber
\end{eqnarray}
\end{minipage}
\begin{eqnarray}
\label{FRdilgFF}
\end{eqnarray}
\\ \\
\item{dilaton - photon - fermion - fermion vertex}
\\ \\
\begin{minipage}{95pt}
\includegraphics[scale=1.0]{figures/FR_dilAFF.eps}
\end{minipage}
\begin{minipage}{70pt}
\begin{eqnarray}
&=&
V^{\alpha}_{\rho \gamma\bar\psi\psi}
= Q_f \, e \frac{3\,i}{\Lambda}\, \gamma^\alpha
\nonumber
\end{eqnarray}
\end{minipage}
\begin{eqnarray}
\label{FRdilAFF}
\end{eqnarray}
where $Q_f$ is the fermion charge expressed in units of $e$.
\\ \\
\item{dilaton - Z - fermion - fermion vertex}
\\ \\
\begin{minipage}{95pt}
\includegraphics[scale=1.0]{figures/FR_dilZFF.eps}
\end{minipage}
\begin{minipage}{70pt}
\begin{eqnarray}
&=&
V^{\alpha}_{\rho Z\bar\psi\psi}
= \frac{3 \,i }{2\, \Lambda s_w \, c_w} \, e \, (C_v^f - C_a^f \gamma^5) \, \gamma^\alpha
\nonumber
\end{eqnarray}
\end{minipage}
\begin{eqnarray}
\label{FRdilZFF}
\end{eqnarray}
where $C_v^f$ and $C_a^f$ are the vector and axial-vector couplings of the $Z$ gauge boson
to the fermion ($f$). Their expressions are
\begin{eqnarray}
C_v^f = I^f_3 - 2 s_w^2 \, Q^f \qquad \qquad C_a^f = I^f_3. \nonumber
\end{eqnarray}
$I^f_3$ denotes the 3rd component of the isospin.
\\ \\
\item{dilaton - four gauge bosons vertex}
\\ \\
\begin{minipage}{95pt}
\includegraphics[scale=1.0]{figures/FR_dilVVVV.eps}
\end{minipage}
\begin{minipage}{70pt}
\begin{eqnarray}
&=&
V^{\alpha\beta\gamma\delta}_{\rho VVVV}
\qquad \qquad = 0 \nonumber
\end{eqnarray}
\end{minipage}
\begin{eqnarray}
\label{FRdilVVVV}
\end{eqnarray}
where $V_1$, $V_2$, $V_3$ and $V_4$ denote $g$, $A$, $Z$ or $W^{\pm}$.
\\ \\
\item{dilaton - gauge boson - gauge boson - scalar - scalar vertex}
\\ \\
\begin{minipage}{95pt}
\includegraphics[scale=1.0]{figures/FR_dilVVSS.eps}
\end{minipage}
\begin{minipage}{70pt}
\begin{eqnarray}
&=&
V^{\alpha\beta}_{\rho VVSS}
\qquad = \frac{2 \, i}{\Lambda} \, e^2 \, \mathcal C_{V_1 V_2 S_1 S_2} \, \eta^{\alpha\beta}
\nonumber
\end{eqnarray}
\end{minipage}
\begin{eqnarray}
\label{FRdilVVSS}
\end{eqnarray}
where $V_1$ and $V_2$ denote the neutral gauge bosons $A$ and $Z$, while the possible scalars are
$\phi$, $\phi^{\pm}$ and $H$. The coefficients are
\begin{eqnarray}
\mathcal C_{A A \phi^+ \phi^-} = 2
\qquad
\mathcal C_{A Z \phi^+ \phi^-} = \frac{c_w^2 - s_w^2}{s_w \, c_w}\qquad \mathcal C_{Z Z \phi^+ \phi^-} =
\frac{\left( c_w^2 - s_w^2 \right)^2}{2 s_w^2 \,c_w^2}
\qquad
\mathcal C_{Z Z \phi \phi} = \mathcal C_{Z Z H H} = \frac{1}{2 s_w^2 \, c_w^2} . \nonumber
\end{eqnarray}
\end{itemize}
\section{The scalar integrals}
\label{scalars}
We collect in this appendix the definition of the scalar integrals appearing in the computation of the correlators.
$1$-, $2$- and $3$-point functions are denoted, respectively as $\mathcal A_0$, $\mathcal B_0$ and $\mathcal C_0$, with
\begin{eqnarray}
\mathcal A_0 (m_0^2) &=& \frac{1}{i \pi^2}\int d^n l \, \frac{1}{l^2 - m_0^2} \,, \nonumber\\
\mathcal B_0 (k^2, m_0^2,m_1^2) &=& \frac{1}{i \pi^2} \int d^n l \, \frac{1}{(l^2 - m_0^2) \, ((l + k )^2 - m_1^2 )} \,, \nonumber \\
%
\mathcal C_0 ((p+q)^2, p^2, q^2, m_0^2,m_1^2,m_2^2) &=& \frac{1}{i \pi^2} \int d^n l \, \frac{1}{(l^2 - m_0^2) \, ((l + p )^2 - m_1^2 ) \, ((l -q )^2 - m_2^2 ) } \,.
\end{eqnarray}
We have also used the finite combination of $2$-point scalar integrals
\begin{eqnarray}
\mathcal D_0 (p^2, q^2, m_0^2,m_1^2) = \mathcal B_0 (p^2, m_0^2,m_1^2) - \mathcal B_0 (q^2, m_0^2,m_1^2) \,.
\end{eqnarray}
The explicit expressions of $\mathcal A_0$, $\mathcal B_0$ and $\mathcal C_0$ can be found in \cite{Denner:1991kt}.
\section{Contributions to $\mathcal V_{\rho Z Z }$}
\label{VZZ}
\begin{itemize}
\item {\bf The fermion sector}
\begin{eqnarray}
&&
\Phi^{F}_{ZZ}(p,q)
=\sum_f \bigg\{
\frac{ \alpha \, m_f^2}{\pi\, s c_w^2 s_w^2 \left(s-4 M_Z^2\right)} \left(s-2 M_Z^2\right)
\left(C_a^{f \, 2} + C_v^{f\, 2}\right)
\nonumber \\
&&
+ \frac{2\, \alpha \, m_f^2}{ \pi\, s c_w^2 \left(s-4 M_Z^2\right)^2 s_w^2}
\left[\left(2 M_Z^4-3 s M_Z^2+s^2\right) C_a^{f \, 2}+C_v^{f \, 2} M_Z^2 \left(2 M_Z^2 + s \right)\right]
\mathcal D_0(s, M_Z^2,m_f^2,m_f^2)
\nonumber \\
&&
+ \frac{\alpha \, m_f^2}{2 \,\pi \, s c_w^2 \left(s - 4 M_Z^2\right){}^2 s_w^2}
\left(s-2 M_Z^2\right) \big\{\left[4 M_Z^4-2 \left(8 m_f^2+s \right) M_Z^2+s \left(4 m_f^2+s \right)\right] C_a^{f\,2}
\nonumber \\
&&
+ C_v^{f \, 2} \left[4 M_Z^4+2 \left(3 s - 8 m_f^2\right) M_Z^2-s \left(s - 4 m_f^2\right) \right] \big\}
\vphantom{\frac{m_f^2}{\left(s-4 M_Z^2\right)}}\, \mathcal C_0(s, M_Z^2,M_Z^2 ,m_f^2,m_f^2,m_f^2)\bigg\}
\nonumber \\
&& \Xi^{F}_{ZZ}(p,q) =
\sum_f \bigg\{ - \frac{\alpha\, m_f^2}{\pi \, s c_w^2 \left(s-4 M_Z^2\right) s_w^2}
\left[\left(2 M_Z^4-4 s M_Z^2+s^2\right) C_a^{f \, 2}+2 C_v^{f \, 2} M_Z^4 \right]
\nonumber \\
&&
- \frac{\alpha\,m_f^2}{ \pi\, c_w^2 s_w^2}C_a^{f\,2}\mathcal B_0(s,m_f^2,m_f^2)
- \frac{2\,\alpha\, m_f^2 \,M_Z^2}{ \pi \, s c_w^2 \left(s-4 M_Z^2\right){}^2 s_w^2}
\big[s^2 C_a^{f \, 2} - 2 \left( C_a^{f \, 2} + C_v^{f \, 2} \right) M_Z^4
\nonumber \\
&&
+ 2 s \left(C_v^{f \, 2}-C_a^{f \, 2}\right) M_Z^2 \big]\, \mathcal D_0(s, M_Z^2, m_f^2,m_f^2)
- \frac{\alpha \, m_f^2}{ \pi\, s c_w^2 \left(s - 4 M_Z^2\right){}^2 s_w^2}
\bigg[ \left[ 4 M_Z^8
\right.
\nonumber \\
&&
\left.
-2 \left(8 m_f^2 + 5 s \right) M_Z^6 + 3 s \left(12 m_f^2+s \right) M_Z^4 - 16 s^2 m_f^2 M_Z^2
+ 2 s^3 m_f^2 \right] C_a^{f \,2} \nonumber \\
&&
+ C_v^{f \, 2} M_Z^4 \left[4 M_Z^4-2 \left(8 m_f^2+s \right) M_Z^2 + s \left(4 m_f^2+s \right)\right] \bigg]\,
\mathcal C_0(s, M_Z^2,M_Z^2 ,m_f^2,m_f^2,m_f^2) \vphantom{\frac{m_f^2}{\left(s-4 M_Z^2\right)}} \bigg\} \nonumber
\end{eqnarray}
\item{\bf The $W$ boson sector}
\begin{eqnarray}
&&
\Phi^{W}_{ZZ}(p,q)=
\frac{\alpha}{ s_w^2 \, c_w^2 \, \pi }
\bigg[
\frac{M_Z^2}{2\,s\left(s-4 M_Z^2\right)} \big[ 2 M_Z^2 \left(-12 s_w^6+32 s_w^4-29 s_w^2+9\right)
+ s \left(12 s_w^6-36 s_w^4+33 s_w^2-10\right)\big] \bigg]
\nonumber
\end{eqnarray}
\begin{eqnarray}
&&
+ \frac{\alpha \, M_Z^2}{ 2\, s_w^2 \, c_w^2 \,\pi\,s\,(s-4 M_Z)^2}
[ 4 M_Z^4 (12 s_w^6 -32 s_w^4 +29 s_w^2 - 9)
\nonumber\\
&&
+ 2 M_Z^2 s (s_w^2 - 2)(12 s_w^4 - 12 s_w^2 +1) +
s^2 (-4 s_w^4+8s_w^2-5) ] \mathcal D_0(s, M_Z^2, M_W^2,M_W^2)
\nonumber\\
&&
+ \frac{\alpha \, M_Z^2}{2\,\pi \, s_w^2 \, c_w^2 \, s \, (s-4 M_Z)^2}
\bigg[-4 M_Z^6(s_w^2-1)(4 s_w^2-3)(12 s_w^4 - 20s_w^2 + 9)
\nonumber \\
&&
+ 2 M_Z^4 s (18 s_w^4-34s_w^2 + 15)(4(s_w^2-3)s_w^2+7) - 2M_Z^2 s^2 (12 s_w^8-96s_w^6 +201s_w^4-157s_w^2+41)+
\nonumber \\
&&
s^3(-12 s_w^6+32s_w^4-27s_w^2+7) \bigg]\, \mathcal C_0(s, M_Z^2,M_Z^2 ,M_W^2,M_W^2,M_W^2)
\nonumber \\
&&
\Xi^{W}_{ZZ}(p,q) =
\frac{\alpha\,M_Z^2}{2 \,\pi\, s_w^2\,c_w^2\,s (s-4M_Z^2)}
\bigg[2 M_Z^4 \left(-12 s_w^6+32 s_w^4-29 s_w^2+9\right)
\nonumber\\
&&
+ s M_Z^2 \left[4\left(s_w^4+ s_w^2\right)-7\right]- 2 s^2 \left(s_w^2-1\right) \bigg]
+ \frac{\alpha M_Z^2}{\pi \,s_w^2c_w^2 }(-2s_w^4 + 3s_w^2 - 1)\, \mathcal B_0(s,M_W^2,M_W^2)
\nonumber \\
&&
+ \frac{\alpha \, M_Z^2}{\pi\, s_w^2 \, c_w^2\, s (s-M_Z^2)^2}
\big[\vphantom{\left(s_w^2-1\right)^2}2 M_Z^6 \left(12 s_w^6-32 s_w^4+29 s_w^2-9\right)
+ s M_Z^4 \left(-24 s_w^6 + 92 s_w^4 - 110 s_w^2 + 41 \right)
\nonumber\\
&&
+ s^2 M_Z^2 \left(-12 s_w^4+26 s_w^2-13\right)+ 2 s^3 \left(s_w^2-1\right)^2\big]\mathcal D_0(s, M_Z^2, M_W^2,M_W^2)
\nonumber \\
&&
+ \frac{\alpha\,M_Z^2}{4 \,\pi \,s_w^2\,c_w^2\,s(s-M_Z^2)^2}
\big[ -8 M_Z^8 \left(s_w^2 - 1\right) \left(4 s_w^2-3\right) \left(12 s_w^4-20 s_w^2+9\right)
\nonumber \\
&&
+ 4 s M_Z^6 \left(24 s_w^8-60 s_w^6+30 s_w^4+25 s_w^2 - 18 \right)
+ 2 s^2 M_Z^4 (-20 s_w^6+76 s_w^4 - 103 s_w^2+46)
\nonumber \\
&&
+ s^3 M_Z^2 \left(-4 s_w^4+24 s_w^2-19\right) - 2 s^4 \left(s_w^2-1\right)\big]\,
\mathcal C_0(s, M_Z^2,M_Z^2 ,M_W^2,M_W^2,M_W^2)\vphantom{\frac{M_Z^2}{(s-4M_Z^2)}} \nonumber
\end{eqnarray}
\item{\bf The $(Z,H)$ sector}
\begin{eqnarray}
&&
\Phi^{ZH}_{ZZ}(p,q)
=
\frac{- \alpha}{4\,\pi\,s\,c_w^2 s_w^2\, \left( s - 4 M_Z^2 \right)}
\left\{\vphantom{\frac{M_H^2}{\left(s-4 M_Z^2\right)}} \left[M_H^2 \left(s-2 M_Z^2\right)+3 s M_Z^2- 2 M_Z^4\right]
\right.
\nonumber\\
&&
\left.
+ 2 \left(M_Z^2-M_H^2\right)\left(\mathcal A_0(M_Z^2)-\mathcal A_0(M_H^2)\right)
\right.
\nonumber\\
&&
\left.
+ \frac{1}{\left(s-4 M_Z^2\right)} \left[2 M_H^2 \left(s M_Z^2-2 M_Z^4+s^2\right)+3 s^2 M_Z^2-14 s M_Z^4+8 M_Z^6\right]
\mathcal B_0(s, M_Z^2,M_Z^2)
\right.
\nonumber\\
&&\left.
- \frac{1}{\left(s-4 M_Z^2\right)}
\left(2 M_H^2+s\right) \left[2 M_H^2 \left(s-M_Z^2\right)-3 s M_Z^2\right]
\mathcal B_0(s, M_H^2,M_H^2)
\right.
\nonumber\\
&&\left.
+ \frac{2}{\left(s-4 M_Z^2\right)} \left[s M_H^4+6 \left(s-M_H^2\right) M_Z^4+\left(2 M_H^4-3 s M_H^2-3 s^2\right)
M_Z^2\right]\mathcal B_0(M_Z^2, M_Z^2,M_H^2)
\right.
\nonumber\\
&&
\left.
+ \frac{\left(2 M_H^2+s\right)}{\left(s-4 M_Z^2\right)} \bigg[M_Z^2 \left(-8 s M_H^2-2 M_H^4+s^2\right)
+ 2 M_Z^4 \left(4 M_H^2 + s\right) + 2 s M_H^4\bigg]\mathcal C_0(s, M_Z^2,M_Z^2 ,M_Z^2,M_H^2,M_H^2)
\right.
\nonumber\\
&&
\left.
+ \frac{M_H^2}{\left(s-4 M_Z^2\right)}
\left[2 M_H^2 \left(s M_Z^2-2 M_Z^4+s^2\right) - 20 s M_Z^4+16 M_Z^6+s^3\right]
\mathcal C_0(s, M_Z^2,M_Z^2 ,M_H^2,M_Z^2,M_Z^2) \right\}
\nonumber \\
&& \Xi^{ZH}_{ZZ}(p,q)=
- \frac{\alpha}{8\, \pi \, s\,c_w^2s_w^2 \left(s-4 M_Z^2\right)}
\bigg\{- 4 M_Z^2 \left(M_Z^4+M_H^2 M_Z^2-3 s M_Z^2+s^2\right)
\nonumber\\
&&
+ 2 \left(M_H^2-M_Z^2 \right)\left(s-2 M_Z^2\right)\left(\mathcal A_0(M_Z^2)-\mathcal A_0(M_H^2)\right)
\nonumber\\
&&
- \frac{1}{\left(s-4 M_Z^2\right)} \left[\left(8 M_Z^6+s^3\right) M_H^2+M_Z^2 \left(s-4 M_Z^2\right) \left(s-2
M_Z^2\right)\left(3 s-2 M_Z^2\right)\right]\mathcal B_0(s, M_Z^2,M_Z^2) \nonumber\\
&&
+ \frac{1}{\left(s-4 M_Z^2\right)} \left[2 \left(4 M_H^4-s^2\right) M_Z^4
- s \left(2 M_H^2+s \right){}^2 M_Z^2+s^2 M_H^2\left(2 M_H^2+s \right)\right]\, \mathcal B_0(s, M_H^2,M_H^2)
\nonumber\\
&&
+ \frac{8 M_Z^2}{\left(s-4 M_Z^2\right)} \left[-s M_H^4-\left(3 M_H^2+5 s \right) M_Z^4
+ \left(M_H^2+s \right)\left(M_H^2+2 s \right) M_Z^2\right]\mathcal B_0(M_Z^2, M_Z^2,M_H^2)
\nonumber\\
&&
- \frac{\left(2 M_H^2+s \right)}{\left(s-4 M_Z^2\right)}
\left[4 \left(7 s-4 M_H^2\right) M_Z^6+4 \left(M_H^2-s \right) \left(M_H^2+3 s \right) M_Z^4
\right.
\nonumber\\
&&
\left.
+ 2 s \left(-M_H^4-2 s M_H^2+s^2\right) M_Z^2+s^2 M_H^4\right]\mathcal C_0(s, M_Z^2,M_Z^2 ,M_Z^2,M_H^2,M_H^2)
\nonumber
\end{eqnarray}
\begin{eqnarray}
&&
- \frac{1}{\left(s-4 M_Z^2\right)} \left[\left(8 M_Z^6+s^3\right) M_H^4
+ 4 M_Z^2 \left(s-4 M_Z^2\right) \left(2 M_Z^4-s M_Z^2+s^2\right) M_H^2
\right.
\nonumber \\
&&
\left.
+ 4 s M_Z^4 \left(s-4 M_Z^2\right){}^2\right]\mathcal C_0(s, M_Z^2,M_Z^2 ,M_H^2,M_Z^2,M_Z^2)
\bigg\} \nonumber
\end{eqnarray}
\item{\bf Term of improvement}
\begin{eqnarray}
&&
\Phi^{I}_{ZZ}(p,q)
= \frac{3\,\chi\,\alpha}{2\, \pi\, s_w^2\, c_w^2\,(s-4M_Z^2)^2}
\bigg\{ (c_w^2 - s_w^2)^2 \bigg[ 2 M_Z^2 s - 8 M_Z^4
\nonumber \\
&&
+ 2 M_Z^2 (s + 2 M_Z^2) \, \mathcal D_0 \left( s, M_Z^2 , M_W^2, M_W^2 \right)
+ 2 \left( c_w^2 M_Z^2 (8 M_Z^4 - 6 M_Z^2 s + s^2) - 2 M_Z^6 + 2 M_Z^4 s \right)
\nonumber \\
&&
\times \, \mathcal C_0 \left( s, M_Z^2, M_Z^2, M_W^2,M_W^2,M_W^2 \right) \bigg]
+ 2 M_Z^2 s - 8 M_Z^4 + 2 M_Z^2 (s + 2\, M_Z^2) \big[
\mathcal B_0 \left(s, M_Z^2, M_Z^2 \right) \nonumber \\
&&
- \mathcal B_0\left(M_Z^2, M_Z^2, M_H^2 \right) \big]
+ \left( 3 M_Z^2 s - 2 M_H^2 (s - M_Z^2) \right) \big[ \mathcal B_0
\left(s, M_H^2, M_H^2 \right) - \mathcal B_0 \left(s, M_Z^2, M_Z^2 \right)\big]
\nonumber \\
&&
+ M_H^2 \left( 2 M_H^2 (s-M_Z^2) + 8 M_Z^4 - 6 M_Z^2 s +
s^2 \right) \mathcal C_0\left( s, M_Z^2,M_Z^2, M_H^2,M_Z^2,M_Z^2\right)
\nonumber \\
&&
+ \left( 2 M_H^2 (M_H^2 - 4 M_Z^2)(s-M_Z^2) + s M_Z^2 (s+ 2 M_Z^2)\right) \mathcal C_0\left( s, M_Z^2,M_Z^2,
M_Z^2,M_H^2,M_H^2\right) \bigg\} \nonumber
\end{eqnarray}
\begin{eqnarray}
&& \Xi^{I}_{ZZ}(p,q) =
- \frac{3\,\chi\,\alpha\,s}{8\,\pi\, s_w^2 \, c_w^2 (s-4 M_Z^2)^2}
\, \bigg\{ (c_w^2 -s_w^2)^2 \bigg[ 4 M_Z^4 (s - 4 M_Z^2)
+ 8 M_Z^4 (s-M_Z^2) \, \mathcal D_0 \left( s, M_Z^2, M_W^2, M_W^2\right)
\nonumber \\
&&
+ 2 M_Z^4 \left[ s^2 - 2 M_Z^2 s + 4 M_Z^4 + 4 c_w^2 M_Z^2 (s - 4 M_Z^2) \right]\,\mathcal C_0 \left( s, M_Z^2, M_Z^2,
M_W^2,M_W^2,M_W^2 \right) \bigg]
\nonumber \\
&&
+ 4 M_Z^2 s_w^4 c_w^2 s (s - 4 M_Z^2)^2 \mathcal C_0 \left( s, M_Z^2, M_Z^2, M_W^2,M_W^2,M_W^2 \right)
+ 4 M_Z^2 (s-4 M_Z^2)
\nonumber \\
&&
+ \left[M_Z^2 s (s + 2 M_Z^2) - M_H^2 (s^2 - 2 M_Z^2 s + 4 M_Z^4)\right]
\big[ \mathcal B_0 \left(s, M_H^2, M_H^2 \right) - \mathcal B_0 \left(s, M_Z^2, M_Z^2 \right) \big]
\nonumber \\
&&
+ 8 M_Z^4 (s - M_Z^2) \big[ \mathcal B_0 \left(s, M_Z^2, M_Z^2 \right) - \mathcal B_0 \left(M_Z^2, M_Z^2, M_H^2 \right) \big]
+ M_H^2 \left[ 4 M_Z^4 (s-4 M_Z^2)
\right.
\nonumber \\
&&
\left.
+ M_H^2 (s^2 - 2 MZ^2 s + 4 M_Z^4)\right] \mathcal C_0\left( s, M_Z^2,M_Z^2, M_H^2,M_Z^2,M_Z^2\right)
+ \left[ M_H^2(M_H^2 - 4 M_Z^2) (s^2 - 2 M_Z^2 s + 4 M_Z^4)
\right.
\nonumber \\
&&
\left.
+ 2 M_Z^2 s (s^2 - 6 M_Z^2 s + 14 M_Z^4)\right] \mathcal C_0\left( s, M_Z^2,M_Z^2, M_Z^2,M_H^2,M_H^2 \right)
\bigg\} \, . \nonumber
\end{eqnarray}
\item{\bf External leg corrections}
\end{itemize}
The $\Delta^{\alpha\beta}(p,q)$ correlator is decomposed as
\begin{eqnarray}
\Delta^{\alpha\beta}(p,q)
&=&
\bigg[ \Sigma_{Min, \, \rho H}(k^2) + \Sigma_{I,\,\rho H}(k^2) \bigg]
\frac{1}{s - M_H^2} V_{HZZ}^{\alpha\beta} + \left( \frac{\Lambda}{i} \right) V_{I, \, \rho H}(k) \frac{1}{s - M_H^2}
\Sigma_{HH}(k^2) \frac{1}{s - M_H^2}
V_{HZZ}^{\alpha\beta}
\nonumber \\
&+&
\Delta^{\alpha\beta}_{I, \, HZZ}(p,q) \nonumber
\end{eqnarray}
where $\Sigma_{HH}(k^2)$ is the Higgs self-energy, $V_{HZZ}^{\alpha\beta}$ and
$ V_{I, \, \rho H}$ are tree level vertices defined in appendix (\ref{rules}) and $\Delta^{\alpha\beta}_{I, \, HZZ}(p,q)$
is expanded into the three contributions of improvement as
\begin{eqnarray}
\Delta^{\alpha\beta}_{I, \, HZZ}
&=&
\Delta^{\alpha\beta}_{(F), \, HZZ}(p,q) + \Delta^{\alpha\beta}_{(W), \, HZZ}(p,q) +
\Delta^{\alpha\beta}_{(Z,H), \, HZZ}(p,q) \nonumber \\
&=&
\left[ \left( \frac{s}{2} - M_Z^2 \right) \eta^{\alpha\beta} - q^\alpha p^\beta \right]\, \Phi^{\Delta}_{ZZ}(p,q) +
\eta^{\alpha\beta}\, \Xi^{\Delta}_{ZZ}(p,q)
\nonumber \\
&=&
\left[ \left( \frac{s}{2} - M_Z^2 \right) \eta^{\alpha\beta} - q^\alpha p^\beta \right]
\left(\Phi^{\Delta\,F}_{ZZ}(p,q) + \Phi^{\Delta\,W}_{ZZ}(p,q) + \Phi^{\Delta\,W}_{ZZ}(p,q)\right)
\nonumber \\
&& \hspace{35mm}
+\, \eta^{\alpha\beta}\, \left(\Xi^{\Delta\,F}_{ZZ}(p,q) + \Xi^{\Delta\,W}_{ZZ}(p,q) + \Xi^{\Delta\,W}_{ZZ}(p,q) \right)
\, . \nonumber
\end{eqnarray}
These are given by
\begin{eqnarray}
&&
\Phi^{\Delta\,F}_{ZZ}(p,q) =
- \sum_f \frac{6 \, \alpha \,\chi \, m_f^2}{\pi\, s_w^2 \, c_w^2 \, (s-M_H^2)(s - 4 M_Z^2)}
\bigg\{
( C_v^{f\, 2} + C_a^{f\, 2}) (s-2 M_Z^2) \nonumber \\
&&
+ \frac{2}{s-4 M_Z^2} [ M_Z^2 ( C_v^{f\, 2} + C_a^{f\, 2}) (s +2 M_Z^2) + C_a^{f \, 2} (s-4M_Z^2) s ] \mathcal
D_0(s,M_Z^2,m_f^2,m_f^2) \nonumber \\
&&
+ \frac{s - 2 M_Z^2}{2(s-4 M_Z^2)}
[ ( C_v^{f\, 2} + C_a^{f\, 2}) (4 m_f^2 (s-4 M_Z^2) + 4 M_Z^4 + 6 M_Z^2 s - s^2 ) + 2 C_a^{f \, 2} s (s-4 M_Z^2) ]
\times \nonumber \\
&&
\mathcal C_0(s, M_Z^2,M_Z^2,m_f^2,m_f^2,m_f^2)
\bigg\} \, , \nonumber \\
&& \Xi^{\Delta\,F}_{ZZ}(p,q)
= \sum_f \frac{18\,\alpha\,s\,\chi \, m_f^2}{\pi \, s_w^2 \, c_w^2 \, (s - M_H^2)(s -4 M_Z^2) } \,
\eta^{\alpha \beta}
\bigg\{
2 M_Z^2 ( C_v^{f\, 2} + C_a^{f\, 2}) \nonumber \\
&&
+ 2 s (s- 4 M_Z^2) C_a^{f \, 2} \mathcal B_0(s, m_f^2, m_f^2)
+ \frac{2}{s- 4M_Z^2} [ 2 M_Z^4 ( C_v^{f\, 2} + C_a^{f\, 2}) (s - M_Z^2)
+ C_a^{f \, 2} M_Z^2 (s-4M_Z^2) s ] \times \nonumber \\
&&
\mathcal D_0(s,M_Z^2,m_f^2,m_f^2)
+ \frac{1}{s -4 M_Z^2} [ ( C_v^{f\, 2} + C_a^{f\, 2}) M_Z^4 ( 4 m_f^2 (s -4 M_Z^2) + 4 M_Z^4 - 2 M_Z^2 s + s^2) \nonumber \\
&&
+ 2 C_a^{f \, 2} s (s - 4 M_Z^2) (m_f^2 (s- 4 M_Z^2) + M_Z^4)] \mathcal C_0(s, M_Z^2,M_Z^2,m_f^2,m_f^2,m_f^2)
\bigg\} \nonumber \\
&&
\Phi^{\Delta\,W}_{ZZ}(p,q) =
\frac{3\,\alpha \, \chi }{\pi\, s_w^2 \, c_w^2 \, (s-M_H^2)(s-4 M_Z^2)}
\bigg\{
\frac{s - 2 M_Z^2}{2} [ M_H^2 (1- 2 s_w^2)^2 \nonumber \\
&&
+ 2 M_Z^2 (-12 s_w^6 + 32 s_w^4 - 29 s_w^2 + 9)]
+ \frac{M_Z^2}{s - 4 M_Z^2} [ M_H^2 (1 - 2 s_w^2)^2 (s + 2 M_Z^2) \nonumber \\
&&
- 2 (s_w^2 -1) ( 2 M_Z^4 (12 s_w^4 - 20 s_w^2 + 9) + s M_Z^2 (12 s_w^4 - 20 s_w^2 +1 ) + 2 s^2) ]
\mathcal D_0(s, M_Z^2, M_W^2,M_W^2) \nonumber \\
&&
+ \frac{M_Z^2}{s - 4 M_Z^2} [ 2 (s_w^2 -1) ( 2 M_Z^6 (4 s_w^2-3)(12 s_w^4 - 20 s_w^2 + 9)
+ 2 M_Z^4 s (- 36 s_w^6 + 148 s_w^4 - 163 s_w^2 + 54) \nonumber \\
&&
+ M_Z^2 s^2 (12 s_w^6 - 96 s_w^4 + 125 s_w^2 - 43) + 4 s^3 ( 2 s_w^4 - 3 s_w^2 + 1) )
- M_H^2 (1- 2 s_w^2)^2 (M_Z^4 (8 s_w^2 -6) \nonumber \\
&&
+ 2 M_Z^2 s (2 - 3 s_w^2) + s^2 (s_w^2-1)) ] \mathcal C_0(s,M_Z^2,M_Z^2,M_W^2,M_W^2,M_W^2) \bigg\}\, , \nonumber \\
&& \Xi^{\Delta\,W}_{ZZ}(p,q)
=\frac{3\,\alpha\, \chi \, M_Z^2}{\pi\,s_w^2 \, c_w^2 \, (s-M_H^2)(s-4 M_Z^2)}
\bigg\{
- M_Z^2 [ M_H^2 (1- 2 s_w^2)^2 + 2 M_Z^2 (-12 s_w^6 + 32 s_w^4 - 29 s_w^2 + 9)] \nonumber \\
&&
+ \frac{s (s- 4 M_Z^2)}{2} (8 s_w^4 - 13 s_w^2 + 5) \mathcal B_0(s, M_W^2,M_W^2)
+ \frac{2}{s - 4 M_Z^2} [ M_H^2 M_Z^2 (1-2s_w^2)^2 (M_Z^2- s ) \nonumber \\
&&
- 2 (s_w^2 -1)(M_Z^6 (12 s_w^4 - 20 s_w^2 + 9)
- 3 M_Z^4 s ( 4 (s_w^2 -3)s_w^2 + 7) + M_Z^2 s^2 (7 - 8 s_w^2) + s^3 (s_w^2 -1) )]
\times \nonumber \\
&&
\mathcal D_0(s, M_Z^2, M_W^2,M_W^2)
+ \frac{1}{2 (s-4 M_Z^2)} [ M_H^2 (- 4 M_Z^6 (1 -2 s_w^2 )^2 (4 s_w^2-3)
+ 2 M_Z^4 s (24 s_w^6 - 28 s_w^4 + 6 s_w^2 - 1) \nonumber \\
&&
+ M_Z^2 s^2 (- 16 s^6 + 12 s_w^4 + 4 s_w^2 -1) + 2 s^3 s_w^4 (s_w^2 -1) )
+ 2 (s_w^2 -1) ( 4 M_Z^8 ( 4 s_w^2 -3)(12 s_w^4 - 20 s_w^2 + 9) \nonumber \\
&&
- 2 M_Z^6 s (24 s_w^6 - 52 s_w^4 + 6 s_w^2 + 15) + M_Z^4 s^2 (45 - 4 s_w^2 (s_w^2+ 13))
+ 2 M_Z^2 s^3 (4 s_w^4 + 2 s_w^2 -5) \nonumber \\
&&
- s^4(s_w^4 -1) ) ] \mathcal C_0(s,M_Z^2,M_Z^2,M_W^2,M_W^2,M_W^2)
\bigg\}\nonumber \\
&& \Phi^{\Delta\,ZH}_{ZZ}(p,q) =
\frac{\alpha \, \chi}{\pi\,s_w^2\,c_w^2\,(s - M_H^2)(s- 4 M_Z^2)}\,
\bigg\{
(2 M_H^2 + M_Z^2) (s - 2 M_Z^2) \nonumber \\
&&
- 2 (M_H^2 - M_Z^2) (\mathcal A_0(M_Z^2) - \mathcal A_0(M_H^2))
+ \frac{1}{s - 4 M_Z^2} [2 M_H^4 (s - M_Z^2) + 3 M_H^2 M_Z^2 s \nonumber \\
&&
+ 2 M_Z^2 (4 M_Z^4 - 9 M_Z^2 s + 2 s^2)] \mathcal B_0(s, M_Z^2, M_Z^2)
- \frac{3}{s - 4 M_Z^2} [2 M_H^4(s - M_Z^2)- 3 M_H^2 M_Z^2 s]
\mathcal B_0(s, M_H^2,M_H^2) \nonumber \\
&&
- \frac{2}{s-4 M_Z^2} [ M_H^2 (s + 2 M_Z^2)(4 M_Z^2 - M_H^2)
+ 2 M_Z^2 s (s - 4 M_Z^2)] \mathcal B_0(M_Z^2, M_Z^2, M_H^2) \nonumber
\end{eqnarray}
\begin{eqnarray}
&&
- \frac{3 M_H^2}{s- 4 M_Z^2} [2 M_H^2 (s-M_Z^2)(4 M_Z^2 - M_H^2)
- M_Z^2 s (s + 2 M_Z^2)] \mathcal C_0(s,M_Z^2,M_Z^2,M_Z^2,M_H^2,M_H^2)
+ \frac{M_H^2}{s - 4 M_Z^2} \times \nonumber \\
&&
[ 2 M_H^4 (s-M_Z^2) + M_H^2 (4 M_Z^4 - 2 M_Z^2 s + s^2) + 2 M_Z^2 (8 M_Z^4 - 14 M_Z^2 s + 3 s^2)] \mathcal
C_0(s,M_Z^2,M_Z^2,M_H^2,M_Z^2,M_Z^2)
\bigg\} \nonumber \\
&& \Xi^{\Delta\,ZH}_{ZZ}(p,q)
= - \frac{3\,\alpha\,\chi}{\pi\,s_w^2\,c_w^2\,(s- M_H^2)(s -4 M_Z^2)}\, \bigg\{ M_Z^4 (2 M_H^2 + M_Z^2)
\nonumber \\
&&
+ \frac{1}{2} (M_Z^2 - M_H^2) (s - 2 M_Z^2) (\mathcal A_0(M_Z^2) - \mathcal A_0(M_H^2))
+ \frac{1}{4 (s - 4 M_Z^2)}\, [ M_H^4 (4 M_Z^4 - 2 M_Z^2 s + s^2)
\nonumber \\
&&
+ M_H^2 M_Z^2 s (s + 2 M_Z^2) - M_Z^2 (16 M_Z^6 - 28 M_Z^4 s + 18
M_Z^2 s^2 - 3 s^3)] \, \mathcal B_0(s, M_Z^2, M_Z^2)
\nonumber \\
&&
- \frac{3 M_H^2}{4(s-4 M_Z^2)} [M_H^2 (4 M_Z^4 - 2 M_Z^2 s + s^2) - M_Z^2 s (s + 2 M_Z^2)] \mathcal B_0
(s, M_H^2, M_H^2)
\nonumber \\
&&
+ \frac{2}{s- 4 M_Z^2} [ M_Z^2 s (M_H^2 - 2 M_Z^2)^2 - M_H^2 M_Z^4 (M_H^2 - 4 M_Z^2)
- M_Z^4 s^2] \mathcal B_0(M_Z^2, M_Z^2, M_H^2)
\nonumber \\
&&
+ \frac{3 M_H^2}{4 (s- 4 M_Z^2)} \times
[ M_H^2 (4 M_Z^4 - 2 M_Z^2 s + s^2) (M_H^2 - 4 M_Z^2) + 2 M_Z^2 s (16 M_Z^4 - 6 M_Z^2 s + s^2)]\,\times
\nonumber \\
&&
\mathcal C_0(s, M_Z^2, M_Z^2, M_Z^2, M_H^2, M_H^2)
+ \frac{1}{4(s - 4 M_Z^2)} [ M_H^6 (4 M_Z^4 - 2 M_Z^2 s + s^2) + 2 M_H^4 M_Z^2 (s^2 - 4 M_Z^4) \nonumber \\
&&
- 4 M_H^2 M_Z^2 (8 M_Z^6 - 10 M_Z^4 s + 6 M_Z^2 s^2 - s^3)
+ 4 M_Z^4 s (s - 4 M_Z^2)^2 ] \mathcal C_0(s, M_Z^2, M_Z^2, M_H^2, M_Z^2, M_Z^2)
\bigg\} \nonumber
\end{eqnarray}
\begin{eqnarray}
\Sigma_{Min, \, \rho H}(k^2)
&=&
\frac{e }{48 \pi^2 \, s_w \, c_w \, M_Z} \bigg\{ \sum_f m_f^2 \bigg[ 3 (4 m_f^2 - k^2) \mathcal
B_0(k^2,m_f^2,m_f^2) + 12 \mathcal A_0(m_f^2) -2 k^2 + 12 m_f^2 \bigg] \nonumber \\
&-&
\frac{1}{2} \bigg[ 3 \left( k^2 (M_H^2 -6 M_W^2) + 2 M_W^2 (M_H^2 + 6 M_W^2) \right)
\mathcal B_0(k^2, M_W^2,M_W^2) + 6 (M_H^2 + 6 M_W^2) \mathcal A_0(M_W^2)
\nonumber \\
&&
- k^2 (M_H^2 + 18 M_W^2) + 6 M_W^2 (M_H^2 - 2 M_W^2) \bigg]
- \frac{1}{4} \bigg[ 3 \left(M_H^2 (2 M_Z^2 + s) + 12 M_Z^4 - 6 M_Z^2 s \right) \nonumber \\
&\times&
\mathcal B_0(s, M_Z^2,M_Z^2) + 9 M_H^2 (2 M_H^2 + s) \mathcal B_0(s, M_H^2,M_H^2)
+ 6 (M_H^2 + 6 M_Z^2) \mathcal A_0(M_Z^2) + 18 M_H^2 \mathcal A_0(M_H^2) \nonumber \\
&&
+ 2 (9 M_H^4 + M_H^2 (3 M_Z^2 -2 s) - 6 M_Z^4 - 9 M_Z^2 s ) \bigg]
\bigg\} \nonumber
\end{eqnarray}
\begin{eqnarray}
\Sigma_{I, \, \rho H}(k^2)
&=&
\frac{3 e}{16 \pi^2 \, s_w \, c_w} \frac{k^2 \, M_H^2}{M_Z^2}\chi \, \bigg[ \mathcal B_0(k^2, M_W^2, M_W^2)+
\frac{3}{2} \mathcal B_0(k^2,M_H^2,M_H^2) + \frac{1}{2} \mathcal B_0(k^2, M_Z^2, M_Z^2) \bigg] \nonumber
\end{eqnarray}
\section{Standard Model self-energies}
\label{SigmaSM}
We report here the expressions of the self-energies appearing in Section \ref{renorm} which define the renormalization conditions.
They are given by
\begin{eqnarray}
\Sigma^{\gamma\gamma}_T(p^2)
&=&
- \frac{\alpha}{4 \pi} \bigg\{ \frac{2}{3} \sum_f \, N_C^f 2 Q_f^2 \bigg[
-(p^2 + 2 m_f^2)B_0(p^2, m_f^2, m_f^2) Ê+ Ê2 m_f^2 B_0(0, m_f^2, m_f^2) + \frac{1}{3}p^2 \bigg] \nonumber \\
&+& Ê\bigg[ (3 p^2 + 4 M_W^2 ) B_0(p^2, M_W^2, M_W^2) - 4 M_W^2 B_0(0, M_W^2, M_W^2)\bigg]\bigg\}
\nonumber
\end{eqnarray}
\begin{eqnarray}
\Sigma^{ZZ}_T(p^2)
&=&
- \frac{\alpha}{4 \pi} \bigg\{ \frac{2}{3} \sum_f \,
N_C^f \bigg[ \frac{C_V^{f \, 2} + C_A^{f \, 2}}{2 s_w^2 c_w^2}
\bigg( -(p^2 + 2m_f^2) B_0(p^2, m_f^2, m_f^2) Ê
+ Ê2 m_f^2 B_0(0, m_f^2, m_f^2) + \frac{1}{3}p^2 \bigg) \nonumber\\
&&
+ \frac{3}{4 s_w^2 c_w^2} m_f^2 B_0(p^2,m_f^2, m_f^2) \bigg]
+ Ê\frac{1}{6 s_w^2 c_w^2}\bigg[ \bigg( (18 c_w^4 + 2 c_w^2 -\frac{1}{2})p^2
+ (24 c_w^4 + 16 c_w^2 -10)M_W^2 \bigg) \nonumber \\
&&
\times B_0(p^2, M_W^2, M_W^2) Ê- Ê(24 c_w^4 - 8 c_w^2 + 2)M_W^2 B_0(0, M_W^2, M_W^2)
+ (4 c_w^2-1)\frac{p^2}{3} Ê\bigg] \nonumber \\
&&
+ Ê \frac{1}{12 s_w^2 c_w^2} \bigg[ (2 M_H^2 -10 M_Z^2 - p^2) B_0(p^2, M_Z^2, M_H^2) Ê
- Ê 2 M_Z^2 B_0(0, M_Z^2, M_Z^2) - 2 M_H^2 B_0(0, M_H^2, M_H^2) Ê
\nonumber \\
&&
- Ê\frac{(M_Z^2 - M_H^2)^2}{p^2}\left( B_0(p^2, M_Z^2, M_H^2) Ê- B_0(0, M_Z^2, M_H^2) \right) -
\frac{2}{3} p^2 Ê \bigg] \bigg\} \nonumber
\end{eqnarray}
\begin{eqnarray}
\Sigma^{\gamma Z}_T(p^2)
&=&
\frac{\alpha}{4 \pi \, s_w \, c_w} \bigg\{ \frac{2}{3} \sum_f \, N_C^f \,
Q_f \, C_V^f \bigg[ (p^2 + 2m_f^2) B_0(p^2, m_f^2, m_f^2) - 2 m_f^2 B_0(0, m_f^2, m_f^2) -\frac{1}{3}p^2 \bigg] \nonumber \\
&&
- \frac{1}{3} \bigg[ \left( (9 c_w^2 + \frac{1}{2})p^2 + (12 c_w^2 + 4)M_W^2 \right) B_0(p^2, M_W^2, M_W^2)
- Ê(12 c_w^2 -2)M_W^2 B_0(0, M_W^2, M_W^2) + \frac{1}{3}p^2 \bigg]\bigg\} \nonumber
\end{eqnarray}
\begin{eqnarray}
\Sigma_{HH}(p^2)
&=&
- \frac{\alpha}{4 \pi} \bigg\{ \sum_{f} N_C^f \frac{m_f^2}{2 s_w^2 M_W^2}
\bigg[ 2 \mathcal A_0\left( m_f^2 \right) + (4 m_f^2 - p^2) \mathcal B_0 \left( p^2, m_f^2,m_f^2\right) \bigg] \nonumber \\
&&
- \frac{1}{2 s_w^2} \bigg[ \left(6 M_W^2 - 2p^2
+ \frac{M_H^4}{2 M_W^2} \right) \mathcal B_0 \left( p^2, M_W^2, M_W^2 \right)
+ \left( 3 + \frac{M_H^2}{2 M_W^2} \right) \mathcal A_0 \left( M_W^2 \right) - 6 M_W^2 \bigg] \nonumber \\
&&
- \frac{1}{4 s_w^2 \, c_w^2} \bigg[ \left(6 M_Z^2 - 2p^2
+ \frac{M_H^4}{2 M_Z^2} \right) \mathcal B_0 \left( p^2, M_Z^2, M_Z^2 \right)
+ \left( 3 + \frac{M_H^2}{2 M_Z^2} \right) \mathcal A_0 \left( M_Z^2 \right) - 6 M_Z^2 \bigg] \nonumber \\
&-& \frac{3}{8 s_w^2} \bigg[ 3 \frac{M_H^4}{M_W^2} \mathcal B_0 \left( p^2, M_H^2,M_H^2 \right)
+ \frac{M_H^2}{M_W^2} \mathcal A_0 \left( M_H^2 \right) \bigg] \bigg\} \nonumber
\end{eqnarray}
\begin{eqnarray}
\Sigma^{WW}_T(p^2)
&=&
-\frac{\alpha}{4 \pi} \bigg\{ \frac{1}{3 s_w^2}
\sum_i \bigg[ \left( \frac{m_{l,i}^2}{2} - p^2 \right)\mathcal B_0\left(p^2, 0, m_{l,i}^2 \right)
+ \frac{p^2}{3} + m_{l,i}^2 \mathcal B_0 \left( 0, m_{l,i}^2,m_{l,i}^2 \right) Ê \nonumber \\
&&
+ Ê\frac{m_{l,i}^4}{2 p^2} \left( \mathcal B_0\left(p^2, 0, m_{l,i}^2 \right)
- \mathcal B_0\left(0, 0, m_{l,i}^2 \right)\right) \bigg] + \frac{1}{s_w^2} \sum_{i,j}|V_{ij}|^2 \bigg[
\left( \frac{m_{u,i}^2 + m_{d,j}^2}{2} - p^2\right) \mathcal B_0 \left( p^2, m_{u,i}^2, m_{d,j}^2\right) \nonumber \\
&&
+ \frac{p^2}{3} + m_{u,i}^2 \mathcal B_0 \left( 0, m_{u,i}^2, m_{u,i}^2\right)
+ m_{d,j}^2 \mathcal B_0 \left( 0, m_{d,j}^2, m_{d,j}^2\right)
+ \frac{(m_{u,i}^2 - m_{d,j}^2)^2}{2 p^2} \big( \mathcal B_0 \left( p^2, m_{u,i}^2, m_{d,j}^2\right) \nonumber \\
&&
- \mathcal B_0 \left( 0, m_{u,i}^2, m_{d,j}^2\right)\big) \bigg] Ê
+ \frac{2}{3} \bigg[ (2 M_W^2 + 5 p^2) \mathcal B_0 \left( p^2,
M_W^2, \lambda^2 \right) - 2 M_W^2 \mathcal B_0 \left( 0, M_W^2, M_W^2\right) \nonumber \\
&&
- \frac{M_W^4}{p^2} \big( \mathcal B_0\left( p^2, M_W^2, \lambda^2 \right) - \mathcal B_0 \left( 0,M_W^2, \lambda^2 \right) \big)
+ \frac{p^2}{3} Ê\bigg] + \frac{1}{12 s_w^2} \bigg[ \big( (40 c_w^2 -1)p^2 \nonumber \\
&&
+(16 c_w^2 + 54 - 10 c_w^{-2}) M_W^2 \big) \mathcal B_0 \left(p^2, M_W^2, M_Z^2 \right)
- (16 c_w^2 + 2) \big( M_W^2 \mathcal B_0 \left( 0,M_W^2,M_W^2\right) \nonumber \\
&&
+ M_Z^2 \mathcal B_0 \left( 0, M_Z^2, M_Z^2\right) \big) + (4 c_w^2 -1) \frac{2 p^2}{3}
- Ê(8 c_w^2 +1) \frac{(M_W^2 - M_Z^2)^2}{p^2} \big( \mathcal B_0 \left( p^2, M_W^2,M_Z^2\right) \nonumber \\
&&
- \mathcal B_0 \left(0, M_W^2,M_Z^2 \right) \big) \bigg] + \frac{1}{12 s_w^2} \bigg[ (2 M_H^2 - 10 M_W^2 - p^2) \mathcal B_0
\left(p^2, M_W^2,M_H^2 \right) - 2 M_W^2 \mathcal B_0 \left(0,M_W^2,M_W^2 \right) \nonumber \\
&&
- 2 M_H^2 \mathcal B_0 \left( 0, M_H^2,M_H^2\right) - \frac{(M_W^2 -M_H^2)^2}{p^2} \big( \mathcal B_0 \left( p^2, M_W^2,
M_H^2\right) - \mathcal B_0 \left( 0,M_W^2,M_H^2\right) \big) - \frac{2 p^2}{3}\bigg] \bigg\}\, , \nonumber
\end{eqnarray}
where the subscripts $l$, $u$ and $d$ stand for "leptons", "up" and "down" (quarks) respectively.
The sum runs over the three generations and $\lambda$ is the photon mass introduced to regularize the infrared
divergences.
\end{appendix}
|
\section{#1}\setcounter{equation}{0}}
\newcommand{\subsectiono}[1]{\subsection{#1}}
\renewcommand{\theequation}{\thesection.\arabic{equation}}
\begin{document}
{}~
{}~
\hfill\vbox{\hbox{IMSc/2014/07/4}}
\vskip 2cm
\centerline{\Large \bf A cut-off tubular geometry of loop space}
\medskip
\vspace*{4.0ex}
\centerline{\large \rm Partha Mukhopadhyay }
\vspace*{4.0ex}
\centerline{\large \it The Institute of Mathematical Sciences}
\centerline{\large \it C.I.T. Campus, Taramani}
\centerline{\large \it Chennai 600113, India}
\medskip
\centerline{E-mail: <EMAIL>}
\vspace*{5.0ex}
\centerline{\bf Abstract}
\bigskip
Motivated by the computation of loop space quantum mechanics as indicated in \cite{semi-classical}, here we seek a better understanding of
the tubular geometry of loop space ${\cal L}{\cal M}$ corresponding to a Riemannian manifold ${\cal M}$ around the submanifold of vanishing loops. Our
approach is to first compute the tubular metric of $({\cal M}^{2N+1})_{C}$ around the diagonal submanifold, where $({\cal M}^N)_{C}$ is the
Cartesian product of $N$ copies of ${\cal M}$ with a cyclic ordering. This gives an infinite sequence of tubular metrics such that the one relevant to
${\cal L}{\cal M}$ can be obtained by taking the limit $N\to \infty$. Such metrics are computed by adopting an indirect method where the general
tubular expansion theorem of \cite{tubular} is crucially used. We discuss how the complete reparametrization isometry of loop space arises in
the large-$N$ limit and verify that the corresponding Killing equation is satisfied to all orders in tubular expansion. These tubular metrics can
alternatively be interpreted as some natural Riemannian metrics on certain bundles of tangent spaces of ${\cal M}$ which, for ${\cal M} \times {\cal M}$,
is the tangent bundle $T{\cal M}$.
\newpage
\tableofcontents
\baselineskip=18pt
\section{Introduction and summary}
\label{intro}
Configuration space of a non-linear sigma model (NLSM) \cite{review} with target space ${\cal M}$ is the corresponding free loop space ${\cal L} {\cal M}$ \cite{loop-space}.
\begin{eqnarray}
{\cal L} {\cal M} &=& C^{\infty}(S^1, {\cal M})~.
\end{eqnarray}
One can rewrite NLSM as a single particle model in ${\cal L} {\cal M}$ such that the complete geometry of the
configuration space is manifest \cite{witten82, frenkel, dwv, talks, semi-classical}. For {\it small loops} which are entirely contained in a single
normal neighbourhood \cite{normal} in ${\cal M}$, the configuration space can be described as a tubular neighbourhood \cite{spivak, FS, tubes} around $\Delta \hookrightarrow {\cal L} {\cal M}$ (see \cite{semi-classical}). Here $\Delta (\cong {\cal M})$ is the submanifold of vanishing loops, i.e. constant maps from $S^1$ to ${\cal M}$. It was demonstrated in \cite{semi-classical} that in a semi-classical limit ($\alpha' \to 0$) of the loop space quantum mechanics (LSQM) the string wavefunction localises on $\Delta$. This in turn implies that the semi-classical expansion is to be derived from tubular expansion of geometric quantities of ${\cal L} {\cal M}$. It was also shown that the effective quadratic action for the tachyon state was correctly reproduced at leading order.
There are two obstacles in performing the computation described in \cite{semi-classical},
\begin{enumerate}
\item
Lack of a detailed understanding of the tubular geometry in ${\cal L}{\cal M}$.
\item
${\cal L}{\cal M}$ is infinite dimensional. This causes divergences to appear in the computation.\footnote{These are analogue of the usual UV divergences of NLSM in the present approach.}
\end{enumerate}
To regularise the divergences, it is natural to think of considering a finite dimensional model which, in the limit of large dimensionality, will approach the loop space model. The goal of this work is to construct suitable finite dimensional geometries, perform the required limit and explicitly construct the tubular geometry in ${\cal L}{\cal M}$ in this way.
As will be discussed in \S \ref{tubnbh}, the finite dimensional {\it cut-off space} we consider is the tubular neighbourhood of the diagonal submanifold $\Delta$ embedded in $({\cal M}^{2N+1})_C$, where,
\begin{eqnarray}
{\cal M}^N &=& {\cal M} \times {\cal M} \times \cdots \times {\cal M} \quad [\hbox{$N$ factors}]~,
\end{eqnarray}
and the subscript $C$ indicates that the factors are cyclically ordered. The limit of large dimensionality is simply the limit $N\to \infty$. Notice that we use the same notation $\Delta$ for both the submanifold of vanishing loops in ${\cal L}{\cal M}$ and the diagonal submanifold of ${\cal M}^N$, as both of them are isomorphic to ${\cal M}$ and the details of their embeddings merge together at large $N$.
The plan of the paper is as follows. In \S \ref{tubnbh} we describe the construction of the relevant tubular neighbourhoods. We evaluate the tubular geometry around $\Delta \hookrightarrow {\cal M}^N$ and the subsequent large-$N$ limit in detail in \S \ref{tub:M^N} and
\S \ref{tub:LM} respectively with many technical details summarised in several appendices. Finally, in \S \ref{comments} we make comments regarding our choice of the cut-off space and some possible applications of our results besides its use in the study of LSQM.
Before we end this introduction, below we briefly discuss a couple of key points relevant to our analysis. The first question is: What do we actually mean by ``evaluating tubular geometry" in this case? To answer this we recall that given an arbitrary submanifold admitting a tubular neighbourhood, it is possible to perform covariant Taylor expansion of tensors around the submanifold \cite{tubes, tubular}. The expansion coefficients are tensors of the ambient space evaluated at the submanifold which describe various extrinsic properties of the embedding. In our case both the submanifold and the ambient space are constructed out of ${\cal M}$ only. Therefore, it should be possible to evaluate all the tubular expansion coefficients of, say vielbein, in terms of the intrinsic geometric data of ${\cal M}$ (hereafter called {\it ${\cal M}$-data}). A direct method was used in \cite{semi-classical} by explicitly constructing the Fermi normal coordinates (FNC) \cite{FS, tubes, tubular} and implementing the required coordinate transformation order-by-order. It is difficult to carry out such a method as the computation soon becomes involved enough. In this work we use an indirect method (to be discussed in \S \ref{s:indirect}) which utilizes the general results of \cite{tubular} very crucially and we are able to derive all-order-results this way.
The second question is: How do we know that the large-$N$-geometry we obtain this way is indeed the geometry of loop space? We show evidence for the case through arguments involving isometries. In particular, we show that the large-$N$-geometry admits the reparametrization isometry of loop space by verifying the relevant Killing equation to all orders in tubular expansion. We hope to present further evidences with the analysis of geodesics in \cite{cut-offII}.
\section{Construction of tubular neighbourhood}
\label{tubnbh}
Here we describe the construction of tubular neighbourhoods in ${\cal L} {\cal M}$ and ${\cal M}^N$. For ${\cal L} {\cal M}$ we review the intuitive manner in which it was introduced in \cite{semi-classical}. All our discussion about finite dimensional spaces will be rigorous and we would like to think
of ${\cal L}{\cal M}$ only as a limit.
We will use the notation ${\cal U}_S$ to refer to certain open normal neighbourhood in $S$, where $S$ will stand for ${\cal M}$, ${\cal M}^N$ or ${\cal L}{\cal M}$. We will see how such neighbourhoods are inter-related. A {\it small loop} in ${\cal M}$ is a loop which is entirely contained in a single normal neighbourhood in ${\cal M}$, say ${\cal U}_{{\cal M}}$. All such loop configurations within ${\cal U}_{{\cal M}}$ define a neighbourhood
${\cal U}_{{\cal L}{\cal M}}$ in ${\cal L}{\cal M}$ such that $\Delta \cap {\cal U}_{{\cal L}{\cal M}}$ is non-vanishing. In fact, given that $\Delta$ is the submanifold of vanishing loops, we must have: $\Delta \cap {\cal U}_{{\cal L}{\cal M}} \cong {\cal U}_{{\cal M}}$. This simply implies that the points in ${\cal U}_{{\cal M}}$, i.e. zero loops, can be identified with the points in $\Delta \cap {\cal U}_{{\cal L}{\cal M}}$. Therefore a non-zero loop in ${\cal U}_{{\cal M}}$ corresponds to a point in ${\cal U}_{{\cal L}{\cal M}}$ which is away from the submanifold. Given any point $P$ in a tubular neighbourhood, there exists a unique geodesic passing through $P$ that is orthogonal to the submanifold. The intersection of this geodesic and the submanifold is also a unique point $Q$ \cite{FS}. Since in our case $P\in {\cal U}_{{\cal L}{\cal M}}$ is a loop in ${\cal U}_{{\cal M}}$, one may view the unique point $Q\in {\cal U}_{{\cal M}}$ \footnote{Note that $Q$ is a point in $\Delta \cap {\cal U}_{{\cal L}{\cal M}}$, but the latter has been identified with ${\cal U}_{{\cal M}}$.} as an average, or in the language of \cite{semi-classical}, centre of mass (CM) of the loop. To define the CM we proceed as follows \cite{semi-classical}. Let $l: S^1 \to {\cal U}_{{\cal M}}$ be the loop corresponding to the point $P$ in ${\cal U}_{{\cal L} {\cal M}}$. Given an arbitrary point $q\in {\cal U}_{{\cal M}}$, we construct the pre-image of $l$ in $T_q{\cal M}$ under the exponential map based at $q$,
\begin{eqnarray}
t (q, \sigma) := {\hbox{Exp}}^{-1}_q \circ l(\sigma)~.
\label{t(q, s)}
\end{eqnarray}
Since exponential map is a diffeomorphism within a normal neighbourhood, the above map is one-to-one.
CM of $l$ is the unique point $Q$ for which,
\begin{eqnarray}
\oint d\sigma t(Q, \sigma) = 0~.
\label{oint-t}
\end{eqnarray}
The above prescription explicitly spells out how to identify, in a one-to-one manner, all possible small loop configurations in ${\cal U}_{{\cal M}}$ and
points in ${\cal U}_{{\cal L} {\cal M}}$ in the following way. First of all, the configurations of all possible small loops in ${\cal U}_{{\cal M}}$ are in one-to-one
correspondence with certain loop configurations in $T{\cal M}$ such that each loop in $T{\cal M}$ resides entirely in a single fiber with its average
position fixed at the corresponding base point. Zero section of $T{\cal M}$ is then identified with $\Delta \hookrightarrow {\cal L}{\cal M}$ and the loops in $T_Q{\cal M}$ are
identified with points on the geodesics in ${\cal L}{\cal M}$ that intersect $\Delta \hookrightarrow {\cal L}{\cal M}$ orthogonally at $Q\in \Delta$.\footnote{The
tubular neighbourhood theorem would then demand that the one-to-one map described above be a diffeomorphism and the identification of
loops in $T{\cal M}$ with points on geodesics in ${\cal L}{\cal M}$ be consistent. In this work we will construct this diffeomorphism explicitly for the
cut-off space while postponing the analysis of geodesics to \cite{cut-offII}. \label{f:diffeo} }
The above discussion however does not directly tell us how to write down the metric of ${\cal L}{\cal M}$ in the form of a tubular expansion in the sense of \cite{tubular}. Explicitly writing down this metric is the purpose of this work. To this end we define an infinite sequence of finite dimensional spaces ${\cal L}^{(N)}{\cal M}$ ($N$ being a positive integer) such that the tubular geometry of ${\cal L}{\cal M}$ can be understood as a large-$N$ limit.
Such a cut-off loop space should satisfy the following properties,
\begin{enumerate}
\item
${\cal L}^{(N)}{\cal M}$ admits $\Delta \cong {\cal M}$ as a submanifold.
\item
Tubular geometry around $\Delta \hookrightarrow {\cal L}^{(N)}{\cal M}$ approaches the tubular geometry around $\Delta \hookrightarrow {\cal L}{\cal M}$ in the limit $N\to \infty$.
\end{enumerate}
As mentioned in \S \ref{intro}, in this work we explore the following possibility,
\begin{eqnarray}
{\cal L}^{(N)}{\cal M} &=& ({\cal M}^{2N+1})_C ~,
\label{cut-off}
\end{eqnarray}
with the diagonal submanifold of ${\cal M}^{2N+1}$ playing the role of $\Delta$.
The construction of tubular neighbourhood around $\Delta \hookrightarrow {\cal M}^N$ is simply a discretised version of the above discussion regarding loop space. Since ${\cal U}_{{\cal M}^N} = ({\cal U}_{{\cal M}})^N$, a point in ${\cal U}_{{\cal M}^N}$, say $(l_1, l_2, \cdots ,l_N)$ is an $N$-point configuration in ${\cal U}_{{\cal M}}$. Following the loop-space-discussion we define the average position/CM by the unique point $Q \in {\cal U}_{{\cal M}} (\hbox{or } \Delta \cap {\cal U}_{{\cal M}^N})$ such that,
\begin{eqnarray}
\sum_{p=1}^N t_p(Q) &=& 0~,
\label{sum-tp}
\end{eqnarray}
where given $q\in {\cal U}_{{\cal M}}$,
\begin{eqnarray}
t_p(q) &\equiv& {\hbox{Exp}}_q^{-1} \circ l_p ~.
\label{tp-lp}
\end{eqnarray}
Notice that it is the condition in (\ref{sum-tp}) which singles out the diagonal submanifold as the space of all possible locations of the CM.
Therefore all possible $N$-point configurations contained entirely in a single normal neighbourhood in ${\cal M}$ are in one-to-one
correspondence with certain $N$-point configurations in $T{\cal M}$ such that in each such configuration all the $N$-points are in a single fiber
with the average fixed at the corresponding base point. Just like in the case of loop space, we then identify the zero section with $\Delta
\hookrightarrow {\cal M}^N$ such that the $N$-point configuration in $T_Q{\cal M}$ given by $\{t_p\}$ in (\ref{sum-tp}) is mapped to $(l_1, l_2,
\cdots , l_N)$ in the neighbourhood of $\Delta$ in ${\cal M}^N$. This is the basic construction that will be used in \S \ref{tub:M^N} to compute the
tubular metric in ${\cal M}^N$.
\section{Tubular geometry in ${\cal M}^N$}
\label{tub:M^N}
The purpose of this section is to explicitly work out the tubular geometry around $\Delta \hookrightarrow {\cal M}^N$ in terms of
${\cal M}$-data. Here we first pose the problem in more technical terms. It is a well-known fact in Riemannian geometry \cite{petersen} that an open neighbourhood around the diagonal of ${\cal M} \times {\cal M}$ is diffeomorphic to an open neighbourhood around the zero section of $T{\cal M}$. The construction described in the previous section is a generalisation of the same statement where ${\cal M} \times {\cal M}$ is replaced by ${\cal M}^N$ and $T{\cal M}$ by a bundle $T^{(N-1)}{\cal M}$ whose base is ${\cal M}$ and the fiber at $Q\in {\cal M}$ is given by\footnote{The reason that there are
$(N-1)$ additive factors in eq.(\ref{TQ}) is simply because in the $T{\cal M}$ description, as discussed below eq.(\ref{tp-lp}), the $N$-point configuration is given by $(N-1)$ independent tangent vectors of ${\cal M}$. },
\begin{eqnarray}
T_Q^{(N-1)}{\cal M} &=& T_Q {\cal M} \oplus \cdots \oplus T_Q{\cal M}~, \quad [(N-1)\hbox{ additive factors}] ~.
\label{TQ}
\end{eqnarray}
The relevant diffeomorphism may be considered to be a transformation between the coordinate systems which are natural in ${\cal M}^N$ and $T^{(N-1)}{\cal M}$. In the
natural coordinate system in ${\cal M}^N$, hereafter to be called {\it direct product coordinates} (DPC), the point $(l_1, l_2, \cdots , l_N) \in {\cal U}_{{\cal M}}$ is given by,
\begin{eqnarray}
\bar z^{\bar a} &=& (x_1^{\alpha_1}, x_2^{\alpha_2}, \cdots , x_N^{\alpha_N}) ~.
\label{z-bar-gen}
\end{eqnarray}
On the other hand the natural coordinate system in $T^{(N-1)}{\cal M}$ is taken to be the FNC relevant to the present case. The notation for FNC in a generic case is already set up in Appendix \ref{a:tubular}. See, for example, eq.(\ref{FNC}). In our special case, because of the particular structure of eq.(\ref{TQ}), this takes the following form,
\begin{eqnarray}
\hat z^a &=& (x^{\alpha}, \hat y^A) = (x^{\alpha}, \hat y_1^{\hat \alpha_1}, \cdots , \hat y_{N-1}^{\hat \alpha_{N-1}} ) ~,
\label{zhat}
\end{eqnarray}
where $x^{\alpha}$ is a general coordinate\footnote{We will eventually take $x^{\alpha}$ and $x_p^{\alpha_p}$ ($p=1, 2, \cdots , N$) to be the same local coordinate system in ${\cal M}$.} for $Q\in {\cal M}$ and $\hat y_{\hbox{a}}$ ($\hbox{a} = 1, 2, \cdots , (N-1)$) is the fiber coordinate along the $\hbox{a}$-th factor on the RHS of (\ref{TQ}).
The metric is of course known in DPC in terms of ${\cal M}$-data. A direct method of computing the tubular metric (i.e. the metric in $\hat z$-system given in tubular expansion form) will be to start with the metric in DPC and then perform the coordinate transformation $\bar z \to \hat z$. We will explicitly construct the full coordinate transformation in Appendix \ref{a:coord-construction} in terms of exponential maps. However, as mentioned earlier, this computation is very cumbersome. We will instead adopt an indirect method using which we are able to compute all-order results for vielbein.
As an interesting aside, notice that the above construction naturally gives a Riemannian metric on $T^{(N-1)}{\cal M}$. This Riemannian metric is nothing but the tubular metric that we have set out to compute. For $N=2$ the bundle under question is the tangent bundle $T{\cal M}$ and the explicit form of its metric will be written down to quartic order in eqs.(\ref{g-al-beta-2}-\ref{g-A-B-2}). There are other approaches of constructing natural Riemannian metrics on $T{\cal M}$ in the literature \cite{TM} and it may be interesting to explore if there is any relation between these two types of constructions.
The indirect method will be described in \S \ref{s:indirect}. The results for vielbein-expansion will be obtained using this method in \S \ref{s:exp-vielbein}. We show explicit form of the metric expansion up to quartic order for $N=2, 3$ in \S \ref{s:metric-M^N}. We test our results using the direct method up to second order in Appendix \ref{a:verification}.
\subsection{The indirect method}
\label{s:indirect}
The indirect method is given as follows. The work of \cite{tubular} proves a general theorem which describes the tubular expansion of vielbein around ${\cal M}$ sitting as a submanifold in an arbitrary ambient space ${\cal L}$ to all orders. This result, as well as certain notations to be used below, are summarised in Appendix \ref{a:tubular}. We take this general result and specialise to our case where ${\cal L} = {\cal M}^N$ and the submanfiold under question is the diagonal one. At this stage the expansion coefficients, which are tensors of ${\cal M}^N$ evaluated at the diagonal, are expressed in terms of FNC. Therefore the problem is to re-express all of them in terms of ${\cal M}$-data. This can be done, thanks to their tensorial nature, by transforming them under $\hat z \to \bar z$ to express them in terms of tensors in DPC, which are directly known in terms of ${\cal M}$-data. Notice that this requires a very limited amount of information about the coordinate transformation as the Jacobian matrix needs to be evaluated only at the submanifold. This is where the usefulness of the result in \cite{tubular} and the indirect method lies.
Below we will first compute this Jacobian matrix (restricted to the submanifold) following the general construction of \cite{FS} and then in \S \ref{s:exp-vielbein} we express all the quantities relevant to the expansion of vielbein in terms of ${\cal M}$-data. Although this suffices for our practical goal, we perform certain further analysis for completeness of our overall understanding. This also facilitates the verification done in Appendix \ref{a:verification}. There are various steps of this analysis and the entire discussion has been kept in Appendix \ref{a:coord-construction}.
We now proceed to compute the relevant Jacobian matrix. Going back to eqs.(\ref{z-bar-gen}) and (\ref{zhat}), we note that without loss of generality we may take each one of $x$ and $x_p$ ($\forall p$) (labelling the points $Q$ and $l_p$ respectively) to be the general coordinate system $U$ as described in Appendix \ref{sa:RNC}. This implies that the metric in DPC has the following block-diagonal form\footnote{Following notations similar to that of Appendix \ref{a:tubular}, we use lower case symbols with a bar to denote tensors of ${\cal M}^N$ in DPC. },
\begin{eqnarray}
\bar g_{\bar a \bar b}(\bar z) \to diag( {1\over N} G_{\alpha_1 \beta_1}(x_1), {1\over N} G_{\alpha_2 \beta_2}(x_2), \cdots , {1\over N} G_{\alpha_N \beta_N}(x_N)) ~,
\label{gbar}
\end{eqnarray}
where $G$ is the metric in ${\cal M}$. The factor of $1/N$ on the RHS is due to the following reason. In an $N$-dimensional Cartesian system there exists an $N$-dependent scaling between the length scales along the axes and the diagonal. The above definition and the coordinate transformation to be discussed below will ensure that the induced metric on the diagonal submanifold be given by $G$.
In order to define $\hat y^A$ in (\ref{zhat}), let us first denote the components of the tangent vector,
\begin{eqnarray}
(t_1, t_2, \cdots , t_N) \in T_{(x,x,\cdots , x)} {\cal M}^N = T_x{\cal M} \oplus \cdots \oplus T_x{\cal M} \quad (N\hbox{ additive factors})~,
\label{t-vector}
\end{eqnarray}
in DPC by $\bar \xi^{\bar a} = (\xi_1^{\alpha_1}, \xi_2^{\alpha_2}, \cdots , \xi_N^{\alpha_N})$. Then according to the relation between $l_p$ and $t_p$ as given by eq.(\ref{tp-lp}), we must have,
\begin{eqnarray}
x_p^{\alpha_p} &=& \delta^{\alpha_p}{}_{\alpha} x^{\alpha} + {\hbox{Exp}}_x^{\alpha_p}(\xi_p) ~, \cr
&=& \delta^{\alpha_p}{}_{\alpha} x^{\alpha} + \xi_p^{\alpha_p} + {\cal O}(\xi_p^2)~,
\label{xp-xip}
\end{eqnarray}
where ${\hbox{Exp}}_x: T_x{\cal M} \to {\cal M}$ \footnote{Since now onwards we will mostly work with coordinate description, by abuse of language, a point will usually be referred to by its coordinates in a given system. \label{pt-coord}} is the exponential map in ${\cal M}$ as given in eq.(\ref{Exp}). According to the construction of \cite{FS}, $\hat y$ and $\bar \xi$ are linearly related,
\begin{eqnarray}
\bar \xi^{\bar a} &=& \underline{K^{\bar a }{}_B} \hat y^B~,
\label{xibar-yhat}
\end{eqnarray}
where,
\begin{eqnarray}
K^{\bar a}{}_b = \left(\partial \bar z^{\bar a}\over \partial \hat z^b \right) ~,
\end{eqnarray}
is the Jacobian matrix for the transformation $\hat z \to \bar z$. Notice that $\bar \xi$ satisfies the following constraint,
\begin{eqnarray}
\sum_p \delta^{\alpha}{}_{\alpha_p} \xi^{\alpha_p}_p = 0~,
\label{xi-transverse}
\end{eqnarray}
(which is the same equation in (\ref{sum-tp}) in component form) and therefore eq.(\ref{xibar-yhat}) is invertible.
Below we will first construct $J=K^{-1}$ by demanding that the metric in FNC at the location of the submanifold, as dictated by the results mentioned in Appendix \ref{a:tubular}, be given by,
\begin{eqnarray}
\underline{\hat g_{a b}} &\to& diag(G_{\alpha \beta}(x), \eta_{AB})~, \quad \eta_{AB} = \eta_{\hat \alpha \hat \beta} \delta_{\hbox{a} \hbox{b}} ~,
\label{ghat-submanifold}
\end{eqnarray}
where $\eta_{\alpha \beta}$, given the fact that ${\cal M}$ is considered to be Riemannian, is simply given by the Kronecker delta. However, we will continue to use the symbol $\eta_{\alpha \beta}$ so that the expressions are generalizable to arbitrary signature. Before we go on to construct $J$, we explain the index notation adopted in the above equation. According to the index notation of eq.(\ref{zhat}), the transverse coordinates are denoted as
$\hat y^{\hat \alpha_1}_1, \hat y^{\hat \alpha_2}_2 , \cdots $. We use the notation $\hbox{a}, \hbox{b}, \cdots$ to denote the subscript when it is arbitrary. But in that case we can make our notation less clumsy by removing the subscript from the tensor index (i.e. by writing $\hat y_{\hbox{a}}^{\hat \alpha}, \hat y_{\hbox{b}}^{\hat \beta}, \cdots$) but keeping in mind the association between Roman and Greek alphabets as $\hbox{a} \leftrightarrow \hat \alpha$, $\hbox{b} \leftrightarrow \hat \beta$ etc. At the same time the upper case Roman indices can be thought of being associated to pairs in the following way: $A \leftrightarrow (\hbox{a}, \hat \alpha)$, $B \leftrightarrow (\hbox{b}, \hat \beta)$.\footnote{
\label{indices}
Notice that the types of indices considered so far, namely $\alpha_p$, $\alpha$ and $\hat \alpha_{\hbox{a}}$ (or $\hat \alpha$) are all tangent space indices of the same manifold, i.e. ${\cal M}$. Such indices are indistinguishable when they appear in a quantity that is intrinsic to ${\cal M}$. However, they play different roles from the point of view of ${\cal M}^N$.} Such an association is implied in the second equation of (\ref{ghat-submanifold}).
In order to construct $J$, we first look at the general way of decomposing a vector in ${\cal M}^N$ into components that are tangential and normal to the diagonal submanifold. Given an arbitrary element in $T_{(x,x,\cdots , x)}{\cal M}^N$ with components in DPC given by $\bar \eta^{\bar a} = (\eta_1^{\alpha_1}, \eta_2^{\alpha_2}, \cdots , \eta_N^{\alpha_N})$, one can define the tangential and normal parts as follows,
\begin{eqnarray}
\bar \eta_{\parallel}^{\alpha} = R^{\alpha}{}_{\bar b} \bar \eta^{\bar b} ~, \quad \bar \eta_{\parallel}^A = 0~,
\hbox{ and } \bar \eta_{\perp}^{\alpha} = 0 ~, \quad \bar \eta_{\perp}^A = R^A{}_{\bar b} \bar \eta^{\bar b} ~,
\end{eqnarray}
respectively, where $R = O \otimes {\hbox{ 1\kern-1.2mm l}}_{\dim {\cal M}}$ (${\hbox{ 1\kern-1.2mm l}}_{d}$ being the identity matrix of dimension $d$), i.e.
\begin{eqnarray}
R^{\alpha}{}_{\beta_p} = O_{0p} \delta^{\alpha}{}_{\beta_p} ~, \quad R^A{}_{\beta_p} = O_{\hbox{a} p} \delta^{\hat \alpha}{}_{\beta_p}~, \quad \hbox{a} = 1, 2, \cdots (N-1)~,
\label{R-def}
\end{eqnarray}
and $O$ is an $N\times N$ orthogonal matrix such that,
\begin{eqnarray}
O_{0p} &=& {1\over \sqrt{N}}~.
\label{O0p}
\end{eqnarray}
In fact, to preserve handedness of the coordinate system, we will always consider $O \in SO(N)$ in this work.
The above definition of transversality and the existing direct product structure imply that,
\begin{eqnarray}
\hat y^A &=& {1\over \sqrt{N}} {\cal E}^A{}_B (x) R^B{}_{\bar b} \bar \xi^{\bar b} ~,
\label{yh-xibar}
\end{eqnarray}
where $\bar \xi^{\bar a}$ satisfy the condition (\ref{xi-transverse}) and the matrix ${\cal E}(x)$ has a block-diagonal form,
\begin{eqnarray}
{\cal E}(x) = diag( {\cal E}_1(x), {\cal E}_2(x), \cdots , {\cal E}_{N-1}(x) )~,
\label{calE-def}
\end{eqnarray}
where the sub-matrix ${\cal E}_{\hbox{a}}(x)$ is unknown, to be determined below. Equations (\ref{yh-xibar}, \ref{xp-xip}) and (\ref{xibar-yhat}) imply,
\begin{eqnarray}
\underline{J^{\alpha}{}_{\beta_p}} &=& {1\over \sqrt{N}} R^{\alpha}{}_{\beta_p} \quad
\underline{J^A{}_{\beta_p}} = {1\over \sqrt{N}} {\cal E}^A{}_B(x) R^B{}_{\beta_p} ~,
\label{underline-J}
\\
\underline{K^{\alpha_p}{}_{\beta} } &=& \sqrt{N} (R^T)^{\alpha_p}{}_{\beta} ~, \quad
\underline{K^{\alpha_p}{}_B} = \sqrt{N} (R^T)^{\alpha_p}{}_C {\cal F}^C{}_B(x) ~,
\end{eqnarray}
where,
\begin{eqnarray}
{\cal F}(x) = diag ({\cal F}_1(x), {\cal F}_2(x), \cdots , {\cal F}_{N-1}(x) )~,
\end{eqnarray}
such that,
\begin{eqnarray}
{\cal F}_{\hbox{a}} (x) {\cal E}_{\hbox{a}}(x) &=& {\hbox{ 1\kern-1.2mm l}}_{\dim {\cal M}} ~.
\end{eqnarray}
Therefore to reach our final goal all we have to do is to find ${\cal E}_{\hbox{a}}(x)$ and ${\cal F}_{\hbox{a}}(x)$. This can be done by demanding that the coordinate transformation under consideration relate $\underline{\bar g_{\bar a \bar b}}$ as given in eq.(\ref{gbar}) and
$\underline{\hat g_{ab}}$ as given in eq.(\ref{ghat-submanifold}). This gives, upon using $O^TO={\hbox{ 1\kern-1.2mm l}}_N$,
\begin{eqnarray}
{\cal E}_{\hbox{a}}^{\alpha}{}_{\beta} (x) &=& E^{(\alpha)}{}_{\beta}(x)~, \quad {\cal F}_{\hbox{a}}^{\alpha}{}_{\beta}(x) = E_{(\beta)}{}^{\alpha}(x) ~, \quad \forall \hbox{a} = 1, 2, \cdots , (N-1)~,
\label{calE}
\end{eqnarray}
where $E^{(\alpha)}{}_{\beta}(x)$ is the vielbein of ${\cal M}$ (see Appendix \ref{sa:RNC}). Notice that the use of indices in the above equations does not seem to be compatible with the rules mentioned below eq.(\ref{ghat-submanifold}). This is because those rules do not apply to ${\cal E}_{\hbox{a}}$ and ${\cal F}_{\hbox{a}}$ as these are quantities intrinsic to ${\cal M}$ (see footnote \ref{indices}).
Our discussion so far enables one to relate any tensor in the two systems (FNC and DPC) at the submanifold. However, for a quantity which also caries an internal frame index, one has to find suitable basis for the frames as well in order to compare with the results summarized in Appendix \ref{a:tubular}. This is done simply by using the rotation matrix $R$. For example, given the vielbein components\footnote{The $N$-dependence of (\ref{ebar}) is obtained by requiring compatibility with (\ref{gbar}).},
\begin{eqnarray}
\bar e^{(\bar a)}{}_{\bar b} (\bar z) & \to & diag( {1\over \sqrt{N}} E^{(\alpha_1)}{}_{\beta_1}(x_1), {1\over \sqrt{N}} E^{(\alpha_2)}{}_{\beta_2}(x_2), \cdots ) ~,
\label{ebar}
\end{eqnarray}
in DPC, we define the tangential and transverse components as,
\begin{eqnarray}
\bar e_{\parallel}^{(\alpha)}{}_{\bar b} = R^{\alpha}{}_{\bar a} \bar e^{(\bar a)}{}_{\bar b}~, && \quad
\bar e_{\perp}^{(A)}{}_{\bar b} = R^{A}{}_{\bar a} \bar e^{(\bar a)}{}_{\bar b}~.
\label{ebar-prl-perp}
\end{eqnarray}
Then the vielbein components in FNC are defined as follows,
\begin{eqnarray}
\hat e^{(\alpha)}{}_b = \bar e_{\parallel}^{(\alpha)}{}_{\bar a} K^{\bar a}{}_b ~, && \quad
\hat e^{(A)}{}_b = \bar e_{\perp}^{(A)}{}_{\bar a} K^{\bar a}{}_b ~,
\label{ehat-ebar-prl-perp}
\end{eqnarray}
It is the vielbein components of (\ref{ehat-ebar-prl-perp}) that we need to identify with the ones whose tubular expansion has been discussed in Appendix \ref{a:tubular}. The same prescription for defining tangential and transverse internal indices as given in eqs.(\ref{ebar-prl-perp}) is to be used for arbitrary tensors. For example, each term in the tubular expansion of $\hat e^{(a)}{}_b$ is of the form $\underline{\hat t^{(a)}{}_b}$, which should be written as (in matrix notation),
\begin{eqnarray}
\underline{\hat{\hbox{t}}} &=& R \underline{\bar{\hbox{t}} K}~,
\label{that-tbar}
\end{eqnarray}
where $\bar{ \underline{\hbox{t}}}$ is the same tensor in DPC evaluated on the submanifold. Notice that the right hand side is entirely written in terms of ${\cal M}$-data.
\subsection{Expansion coefficients for vielbein}
\label{s:exp-vielbein}
Given the above discussion, we can now compute all the quantities appearing in eqs.(\ref{FNC-ehat}) in terms of ${\cal M}$-data using eq.(\ref{that-tbar}). However, there are a few points to be considered here.
The first one is to find the right $N$-dependence. To count the $N$-dependence systematically we introduce Weyl transformed tensors in DPC in the following manner.
Just like for the metric (see eq.(\ref{gbar})), given any tensor $T$ in ${\cal M}$, we construct a corresponding tensor $\bar t'$ in ${\cal M}^N$ which is block-diagonal, such that the $p$-th block is given by,
\begin{eqnarray}
t'_p(\bar z) = t'_p(x_p) = T(x_p) ~.
\label{t'-T}
\end{eqnarray}
For example, for a tensor of rank, say $(2,1)$,
\begin{eqnarray}
\bar t'^{\bar a \bar b}{}_{\bar c}(\bar z) &=& \left\{ \begin{array}{ll}
t'_p{}^{\alpha_p \beta_p}{}_{\xi_p} (x_p) = T^{\alpha_p \beta_p}{}_{\xi_p} (x_p)
& \hbox{for } \bar a = \alpha_p, \bar b = \beta_p, \bar c = \xi_p \cr & \cr
0 & \hbox{otherwise }
\end{array} \right\} ~, \forall p = 1, 2, \cdots, N ~. \cr &&
\label{example-tbar}
\end{eqnarray}
The above statements (\ref{t'-T}, \ref{example-tbar}) are, in fact, true not only for tensors, but for any quantity constructed out of vielbel and its derivatives. The primed tensors are related to the corresponding unprimed ones by a Weyl transformation.
\begin{eqnarray}
\bar{\hbox{t}} = N^{w\over 2} \bar {\hbox{t}}'~,
\label{weyl-weight}
\end{eqnarray}
where, $w$ is the Weyl-weight of the tensor. The above equation determines $N$-dependence of all the tensors. For example, $w=-2, 2, -1, 0$ for $\bar g_{\bar a\bar b}, \bar g^{\bar a \bar b}, \bar e^{(\bar a)}{}_{\bar b}$ and $\bar r^{\bar a}{}_{\bar b \bar c \bar d}$ respectively.
Our next concern is the following. The tubular expansion under consideration can be viewed as an expansion in powers of the vector $(t_1, t_2, \cdots , t_N)$ (see eq.(\ref{t-vector})). The expression for the expansion coefficients depends on the coordinate system chosen to describe this vector. For example, we could choose to use DPC, in which case the expansion parameter will be $\xi_p^{\alpha_p}$. Alternatively, we could also use FNC ($\hat y_{\hbox{a}}^{\hat \alpha}$) or any other coordinate system. The choice depends on the application. For example, if the (tubular) geometric structure of ${\cal M}^N$ is appearing in a dynamical model in ${\cal M}$, then it will be most suitable to expand in terms of $\xi_p$ as one is ultimately interested in a physical answer to be given completely in terms of ${\cal M}$-data. On the other hand, recalling our discussion at the beginning of \S \ref{tub:M^N}, we may also view the tubular geometry under consideration as a natural Riemannian geometry on the bundle $T^{(N-1)}{\cal M}$ where $\{ \hat y_{\hbox{a}}^{\hat \alpha} \}$ play the role of coordinates along fiber. From this point of view it will be natural to describe the geometry as an expansion in terms of $\{ \hat y_{\hbox{a}}^{\hat \alpha} \}$.
It turns out that the expressions look simpler if we use $\xi_p$ instead of $\hat y_{\hbox{a}}^{\hat \alpha}$. It will also turn out that this difference will not matter much when we extend the result to loop space in next section. Therefore below we choose to write the tubular expansion of vielbein in terms of $\xi_p$.
A typical term in this expansion is given by,
\begin{eqnarray}
\underline{\hat t^{(a)}{}_{b D^1 \cdots D^n} } \hat y^{D^1} \cdots \hat y^{D^n} &=& R^a{}_{\bar a}
\underline{\bar t^{(\bar a)}{}_{\bar b \bar d^1 \cdots \bar d^n} K^{\bar b}{}_b (K^{\bar d^1}{}_{D^1} J^{D^1}{}_{\bar e^1} ) \cdots (K^{\bar d^n}{}_{D^n} J^{D^n}{}_{\bar e^n} }) \bar \xi^{\bar e^1} \cdots \bar \xi^{\bar e^n}~, \cr
&=& N^{(w+1)\over 2} T^{(\hat \alpha)}{}_{(\hat \beta) \delta^1 \cdots \delta^n}(x) \sum_p O_{\hbox{a} p} O^T_{p \hbox{b}} \xi_p^{\delta^1} \cdots \xi_p^{\delta^n} ~, \hbox{a}, \hbox{b} = 0, 1,2, \cdots , \cr &&
\label{gen-coeff}
\end{eqnarray}
where we have used eq.(\ref{that-tbar}) and,
\begin{eqnarray}
\underline{K^{\bar d}{}_A J^A{}_{\bar e}} \bar \xi^{\bar e} = \xi_p^{\delta_p} ~.
\label{KJxi}
\end{eqnarray}
Notice that, to reduce clutter, in the second line we have specified the result for all values of the indices $a=(\alpha, A)$ and $b=(\beta, B)$ by allowing $\hbox{a}$ and $\hbox{b}$ to have the value $0$. According to our notation for indices, $\hat \alpha$ ($\hat \beta$) in the same equation should be replaced by $\alpha$ ($\beta$) whenever $\hbox{a}$ ($\hbox{b}$) possesses the value $0$.
Using the above results one can finally compute the expansion coefficients of vielbein. In addition to the last equation in (\ref{ehat0}), which remains the same, the final results are given by,
\begin{eqnarray}
\hat e_0^{(\alpha)}{}_{\beta} &=& E^{(\alpha)}{}_{\beta}(x) ~,
\label{ehat0-alpha} \\
\hat e_0^{(A)}{}_{\beta} &=& {1\over \sqrt{N}} \Omega_{\beta}{}^{(\hat \alpha)}{}_{\gamma}(x)
\sum_p O_{\hbox{a} p} \xi^{\gamma}_p ~,
\label{ehat0-A} \\
\underline{\hat \pi^{(a)}{}_{(b)}}(\{s\}_n, \hat y) &=& \sum_p O_{\hbox{a} p} O^T_{p \hbox{b}} \Pi_x^{(\hat \alpha)}{}_{(\hat \beta)}(\{s\}_n, \xi_p) ~, \quad \hbox{a}, \hbox{b} = 0, 1, \cdots , (N-1)~. \cr &&
\label{pihat-lowerbar}
\end{eqnarray}
The $\Pi$-matrix in the above equation is given by,
\begin{eqnarray}
\Pi_x(\{s\}_n, \xi) &=& (\xi. \nabla)^{s_1} {\cal R}(\xi;x) \cdots (\xi . \nabla)^{s_n} {\cal R}(\xi;x)~, \cr
(\xi. \nabla)^s [{\cal R}(\xi ;x)]^{(\alpha)}{}_{(\beta)} &=& \xi^{\alpha^1} \cdots \xi^{\alpha^s} \xi^{\gamma} \xi^{\delta} \nabla_{\alpha^1} \cdots \nabla_{\alpha^s} R^{(\alpha)}{}_{\gamma \delta (\beta)}(x) ~.
\label{Pihat}
\end{eqnarray}
Furthermore,
\begin{eqnarray}
\Omega_{\alpha}{}^{(\beta)}{}_{(\gamma)} &=& E^{(\beta)}{}_{\delta} \nabla_{\alpha} E_{(\gamma)}{}^{\delta} ~,
\label{spin-M}
\end{eqnarray}
$R^{\alpha}{}_{\beta \gamma \delta}$ and $\nabla$ are spin connection, Riemann tensor and covariant derivative of ${\cal M}$ respectively in the general coordinate system $U$ as described in Appendix \ref{sa:RNC}. The coordinate and non-coordinate indices are interchanged by the use of vielbein $E^{(\alpha)}{}_{\beta}$. For example, $\Omega_{\alpha}{}^{(\beta)}{}_{\gamma} = \Omega_{\alpha}{}^{(\beta)}{}_{(\delta)} E^{(\delta)}{}_{\gamma}$. Finally, notice that to reduce clutter we have packaged all the values of the indices $a$ and $b$ in eq.(\ref{pihat-lowerbar}) as was done in eq.(\ref{gen-coeff}).
\subsection{Some explicit results for metric-expansion}
\label{s:metric-M^N}
As we saw in the previous subsection, our method of computing tubular expansion of any tensor around $\Delta \hookrightarrow {\cal M}^N$ boils down to first writing down the expansion in the generic case of ${\cal M} \hookrightarrow {\cal L}$ and then specialize to ${\cal L} = {\cal M}^N$ and use the method as described in \S \ref{s:indirect} to express the results in terms of ${\cal M}$-data. We follow the same procedure to arrive at explicit results for metric-expansion up to quartic order for $N=2, 3$. The necessary details of the computation are given in Appendix \ref{a:metric} where we also argue that $\Delta \hookrightarrow {\cal M}^N$ is totally geodesic for any $N$.
For $N=2$, we express the results in terms of $\hat y_1^{\hat \alpha} = \hat y^{\hat \alpha}$ (the index $\hbox{a}$ possesses only one value, i.e. $1$) and it can be interpreted as a natural Riemannian metric on tangent bundle $T{\cal M}$. For $N=3$, to avoid complications we express the results in terms of $\bar \xi^{\bar a}$ which satisfies the constraint (\ref{xi-transverse}).
\noindent
\underline{\bf $N=2$}
The $SO(2)$ matrix is uniquely fixed to be as given in eq.(\ref{SO2}). The final results are,
\begin{eqnarray}
\hat g_{\alpha \beta}
&=& G_{\alpha \beta} + (\Omega_{\alpha}{}^{\eta}{}_{(\hat \xi^1)} \Omega_{\beta \eta (\hat \xi^2) }
+ R_{\alpha (\hat \xi^1 \hat \xi^2) \beta}) \hat y^{\hat \xi^1 \hat \xi^2}
+ \left\{\Omega_{\alpha}{}^{\eta}{}_{(\hat \xi^1)} \Omega_{\beta}{}^{\delta}{}_{(\hat \xi^2)} R_{\eta (\hat \xi^3 \hat \xi^4)
\delta } \right. \cr
&& + {1\over 3} (\Omega_{\alpha}{}^{\eta}{}_{(\hat \xi^1)} \nabla_{(\hat \xi^2)} R_{\eta (\hat \xi^3 \hat \xi^4) \beta}
+ \alpha \leftrightarrow \beta ) + {1\over 12} \nabla_{(\hat \xi^1)} \nabla_{(\hat \xi^2) }
R_{\alpha (\hat \xi^3 \hat \xi^4) \beta} \cr
&& \left. + {1\over 3} R_{\alpha (\hat \xi^1 \hat \xi^2) \eta } R^{\eta}{}_{(\hat \xi^3 \hat \xi^4) \beta} ) \right\} \hat y^{\hat \xi^1 \cdots \hat \xi^4} + O(\hat y^5) ~,
\label{g-al-beta-2} \\
\hat g_{\alpha \hat \beta } &=& \Omega_{\alpha}{}_{(\hat \beta \hat \xi)} \hat y^{\hat \xi}
+ ( {2\over 3} \Omega_{\alpha}{}^{\eta}{}_{(\hat \xi^1)} R_{\eta (\hat \xi^2 \hat \xi^3 \hat \beta) }
+ {1\over 4} \nabla_{(\hat \xi^1)} R_{\alpha (\hat \xi^2 \hat \xi^3 \hat \beta) } ) \hat y^{\hat \xi^1 \cdots \hat \xi^3} + O(\hat y^5) ~,
\label{g-al-B-2} \\
\hat g_{\hat \alpha \hat \beta}
&=& \eta_{\hat \alpha \hat \beta } + {1\over 3} R_{(\hat \alpha \hat \xi^1 \hat \xi^2 \hat \beta) } \hat y^{\hat \xi^1}
\hat y^{\hat \xi^2} + ( {1\over 20} \nabla_{(\hat \xi^1)} \nabla_{(\hat \xi^2)} R_{( \hat \alpha \hat \xi^3 \hat \xi^4 \hat \beta) } + {2\over 45} R_{(\hat \alpha \hat \xi^1 \hat \xi^2) \eta } R^{\eta}{}_{(\hat \xi^3 \hat \xi^4 \hat \beta) } ) \hat y^{\hat \xi^1 \cdots \hat \xi^4} \cr
&& + O(\hat y^5) ~,
\label{g-A-B-2}
\end{eqnarray}
where we have used the notation: $\hat y^{\hat \xi^1 \cdots \hat \xi^n} = \hat y^{\hat \xi^1} \cdots \hat y^{\hat \xi^n}$.
The geometric quantities appearing on the RHS, namely $G$, $\Omega$, $R$ and its covariant derivatives are all evaluated at
$x\in \Delta \cong {\cal M}$. Also note that according to our notations as explained below eq.(\ref{spin-M}),
\begin{eqnarray}
\nabla_{(\xi^1)} R_{\alpha (\xi^2 \xi^3) \beta} &=& E_{(\xi^1)}{}^{\eta^1} E_{(\xi^2)}{}^{\eta^2} E_{(\xi^3)}{}^{\eta^3} \nabla_{\eta^1} R_{\alpha \eta^2 \eta^3 \beta} ~, \cr
&=& E_{(\xi^1)}{}^{\eta^1} \nabla^{tot}_{\eta^1} (E_{(\xi^2)}{}^{\eta^2} E_{(\xi^3)}{}^{\eta^3} R_{\alpha \eta^2 \eta^3 \beta} ) ~,
\end{eqnarray}
where $\nabla^{tot}$ is the total covariant derivative which annihilates vielbein.
\noindent
\underline{\bf $N=3$}
The $SO(3)$ matrix is taken to be as given in eq.(\ref{SO3}). This is of course not a unique choice. The final results are given by,
\begin{eqnarray}
\hat g_{\alpha \beta}
&=& G_{\alpha \beta} + {1\over 3} \left[ (\Omega_{\alpha}{}^{\eta}{}_{\gamma^1} \Omega_{\beta \eta \gamma^2} + R_{\alpha \gamma^1 \gamma^2 \beta}) \sum_p \xi_p^{\gamma^1 \gamma^2} \right. \cr
&& + (R_{\alpha \gamma^1 \gamma^2 \eta} \Omega_{\beta}{}^{\eta}{}_{\gamma^3}
+ R_{\beta \gamma^1 \gamma^2 \eta} \Omega_{\alpha}{}^{\eta}{}_{\gamma^3}
+ {1\over 3} \nabla_{\gamma^1} R_{\alpha \gamma^2 \gamma^3 \beta} ) \sum_p \xi_p^{\gamma^1 \cdots \gamma^3} \cr
&& + \left\{ \Omega_{\alpha}{}^{\eta}{}_{\gamma^1} \Omega_{\beta}{}^{\delta}{}_{\gamma^2} R_{\eta \gamma^3 \gamma^4 \delta} + {1\over 3} ( \Omega_{\alpha}{}^{\eta}{}_{\gamma^1} \nabla_{\gamma^2} R_{\eta \gamma^3 \gamma^4 \beta} + \alpha \leftrightarrow \beta ) + {1\over 12} \nabla_{\gamma^1} \nabla_{\gamma^2} R_{\alpha \gamma^3 \gamma^4 \beta} \right. \cr
&& \left. \left. + {1\over 3} R_{\alpha \gamma^1 \gamma^2 \eta} R^{\eta}{}_{\gamma^3 \gamma^4 \beta} \right\}
\sum_p \xi_p^{\gamma^1 \cdots \gamma^4} + O(\xi^5) \right] ~.
\label{g-al-beta-3}
\\
&& \cr
\hat g_{\alpha \hat \beta_1} &=& {1\over \sqrt{6}} \left[ \Omega_{\alpha}{}_{(\hat \beta_1) \gamma } (-\xi_1^{\gamma} + \xi_2^{\gamma}) + {2\over 3} R_{\alpha \gamma^1 \gamma^2 (\hat \beta_1)} (- \xi_1^{\gamma^1 \gamma^2} +\xi_2^{\gamma^1 \gamma^2}) + ( {2\over 3} \Omega_{\alpha}{}^{\eta}{}_{\gamma^1} R_{\eta \gamma^2 \gamma^3 (\hat \beta_1) } \right. \cr
&& + {1\over 4} \nabla_{\gamma^1} R_{\alpha \gamma^2 \gamma^3 (\hat \beta_1) } )
(-\xi_1^{\gamma^1 \cdots \gamma^3} + \xi_2^{\gamma^1 \cdots \gamma^3} )
+ ({1\over 4} \Omega_{\alpha}{}^{\eta}{}_{\gamma^1} \nabla_{\gamma^2} R_{\eta \gamma^3 \gamma^4 (\hat \beta_1) } \cr
&& \left. + {1\over 15} \nabla_{\gamma^1} \nabla_{\gamma^2} R_{\alpha \gamma^3 \gamma^4 (\hat \beta_1) }
+ {2\over 15} R_{\alpha \gamma^1 \gamma^2 \eta} R^{\eta}{}_{\gamma^3 \gamma^4 (\hat \beta_1) } )
(- \xi_1^{\gamma^1 \cdots \gamma^4} + \xi_2^{\gamma^1 \cdots \gamma^4}) + O(\hat y^5) \right] ~, \cr &&
\label{g-al-betahat1-3} \\
\hat g_{\alpha \hat \beta_2} &=& {1\over 3\sqrt{2}} \left[ \Omega_{\alpha}{}_{(\hat \beta_2) \gamma }
(-\xi_1^{\gamma} - \xi_2^{\gamma} + 2 \xi_3^{\gamma})
+ {2\over 3} R_{\alpha \gamma^1 \gamma^2 (\hat \beta_2 )} (- \xi_1^{\gamma^1 \gamma^2}
- \xi_2^{\gamma^1 \gamma^2} + 2 \xi_3^{\gamma^1 \gamma^2} ) \right. \cr
&& + ( {2\over 3} \Omega_{\alpha}{}^{\eta}{}_{\gamma^1} R_{\eta \gamma^2 \gamma^3 (\hat \beta_2) }
+ {1\over 4} \nabla_{\gamma^1} R_{\alpha \gamma^2 \gamma^3 (\hat \beta_2 ) } )
(-\xi_1^{\gamma^1 \cdots \gamma^3} - \xi_2^{\gamma^1 \cdots \gamma^3} + 2 \xi_3^{\gamma^1 \cdots \gamma^3}) \cr
&& + ({1\over 4} \Omega_{\alpha}{}^{\eta}{}_{\gamma^1} \nabla_{\gamma^2} R_{\eta \gamma^3 \gamma^4 (\hat \beta_2 ) } + {1\over 15} \nabla_{\gamma^1} \nabla_{\gamma^2} R_{\alpha \gamma^3 \gamma^4 (\hat \beta_2 ) } \cr
&& \left. + {2\over 15} R_{\alpha \gamma^1 \gamma^2 \eta} R^{\eta}{}_{\gamma^3 \gamma^4 (\hat \beta_2 ) } )
(- \xi_1^{\gamma^1 \cdots \gamma^4} - \xi_2^{\gamma^1 \cdots \gamma^4} + 2 \xi_3^{\gamma^1 \cdots \gamma^4}) + O(\hat y^5) \right]
\label{g-al-betahat2-3}
\\
&& \cr
\hat g_{\hat \alpha_1 \hat \beta_1}
&=& \eta_{\hat \alpha_1 \hat \beta_1} + {1\over 2} \left[ {1\over 3} R_{(\hat \alpha_1) \gamma^1 \gamma^2 (\hat \beta_1) }
(\xi_1^{\gamma^1 \gamma^2} + \xi_2^{\gamma^1 \gamma^2})
+ {1\over 6} \nabla_{\gamma^1} R_{(\hat \alpha_1) \gamma^2 \gamma^3 (\hat \beta_1)}
(\xi_1^{\gamma^1 \cdots \gamma^3} + \xi_2^{\gamma^1 \cdots \gamma^3}) \right. \cr
&& \left. + ( {1\over 20} \nabla_{\gamma^1} \nabla_{\gamma^2} R_{(\hat \alpha_1) \gamma^3 \gamma^4 (\hat \beta_1) } + {2\over 45} R_{(\hat \alpha_1) \gamma^1 \gamma^2 \eta } R^{\eta}{}_{\gamma^3 \gamma^4 (\hat \beta_1)} )
(\xi_1^{\gamma^1 \cdots \gamma^4} + \xi_2^{\gamma^1 \cdots \gamma^4}) + O(\hat y^5) \right] ~. \cr &&
\label{g-alhat1-betahat1-3}
\\
\hat g_{\hat \alpha_1 \hat \beta_2}
&=& {1\over 2 \sqrt{3} } \left[ {1\over 3} R_{(\hat \alpha_1) \gamma^1 \gamma^2 (\hat \beta_2) }
(\xi_1^{\gamma^1 \gamma^2} - \xi_2^{\gamma^1 \gamma^2})
+ {1\over 6} \nabla_{\gamma^1} R_{(\hat \alpha_1) \gamma^2 \gamma^3 (\hat \beta_2)}
(\xi_1^{\gamma^1 \cdots \gamma^3} - \xi_2^{\gamma^1 \cdots \gamma^3}) \right. \cr
&& \left. + ( {1\over 20} \nabla_{\gamma^1} \nabla_{\gamma^2} R_{(\hat \alpha_1) \gamma^3 \gamma^4 (\hat \beta_2) } + {2\over 45} R_{(\hat \alpha_1) \gamma^1 \gamma^2 \eta } R^{\eta}{}_{\gamma^3 \gamma^4 (\hat \beta_2 )} ) (\xi_1^{\gamma^1 \cdots \gamma^4} - \xi_2^{\gamma^1 \cdots \gamma^4}) + O(\hat y^5) \right] ~. \cr &&
\label{g-alhat1-betahat2-3}
\\
\hat g_{\hat \alpha_2 \hat \beta_2}
&=& \eta_{\hat \alpha_2 \hat \beta_2}
+ {1\over 6 } \left[ {1\over 3} R_{(\hat \alpha_2) \gamma^1 \gamma^2 (\hat \beta_2) }
(\xi_1^{\gamma^1 \gamma^2} + \xi_2^{\gamma^1 \gamma^2} + 4 \xi_3^{\gamma^1 \gamma^2} ) \right. \cr
&& + {1\over 6} \nabla_{\gamma^1} R_{(\hat \alpha_1) \gamma^2 \gamma^3 (\hat \beta_1)}
(\xi_1^{\gamma^1 \cdots \gamma^3} + \xi_2^{\gamma^1 \cdots \gamma^3} + 4 \xi_3^{\gamma^1 \cdots \gamma^3} ) \cr
&& + ( {1\over 20} \nabla_{\gamma^1} \nabla_{\gamma^2} R_{(\hat \alpha_1) \gamma^3 \gamma^4 (\hat \beta_1) } + {2\over 45} R_{(\hat \alpha_1) \gamma^1 \gamma^2 \eta } R^{\eta}{}_{\gamma^3 \gamma^4 (\hat \beta_1)} ) (\xi_1^{\gamma^1 \cdots \gamma^4} + \xi_2^{\gamma^1 \cdots \gamma^4} + 4 \xi_3^{\gamma^1 \cdots \gamma^4} ) \cr
&& \left. + O(\hat y^5) \right] ~.
\label{g-alhat2-betahat2-3}
\end{eqnarray}
\section{Tubular geometry in ${\cal L}{\cal M}$}
\label{tub:LM}
We would now like to construct the tubular geometry of ${\cal L}{\cal M}$ near the submanifold of vanishing loops by performing a suitable large-$N$ limit of the construction described in the previous section. Note, however, that metric in ${\cal L} {\cal M}$ is well-known. We first show in \S \ref{s:LM-DPC} that this well-known form is nothing but the large-$N$ limit of the metric of ${\cal L}^{(N)}{\cal M}$ in DPC, i.e. the analogue of eq.(\ref{gbar}). Our goal here is to find analogues of eqs.(\ref{ehat0-alpha}-\ref{pihat-lowerbar}) and (\ref{g-al-beta-3}-\ref{g-alhat2-betahat2-3}) which will be done in \S \ref{s:FNC-LM} and \S \ref{s:metric-LM} respectively.
\subsection{Geometry of ${\cal L}{\cal M}$ and DPC}
\label{s:LM-DPC}
We discussed in \S \ref{s:exp-vielbein} how given a tensor in ${\cal M}$, one can construct a corresponding tensor in ${\cal M}^N$ in DPC. All the geometric quantities of ${\cal M}^N$, which are constructed out of vielbein and its derivatives, are of this type, as required by the discrete isometries (to be discussed in \S \ref{s:isometry}). All computations involving such tensors are expressible in terms of ${\cal M}$-data. Let $\bar v^{\bar a}(\bar z)$ and $V^{\alpha}(x)$ be the components of the corresponding tangent vector fields in ${\cal M}^N$ and ${\cal M}$ respectively. We relate them in the following way (see \S \ref{s:exp-vielbein} for notation),
\begin{eqnarray}
v_p^{\alpha_p}(x_p) &=& V^{\alpha_p}(x_p)~,
\end{eqnarray}
such that the lengths match on the diagonal ($x_p=x~, \forall p$),
\begin{eqnarray}
|\bar v (\bar z)|^2|_{x_p=x} &\equiv& \bar v^{\bar a}(\bar z) \bar v^{\bar b}(\bar z) \bar g_{\bar a \bar b}(\bar z) |_{x_p=x} = |V(x)|^2 ~.
\label{v-mod-squared}
\end{eqnarray}
The above discussion is also valid for ${\cal L}^{(N)}{\cal M}$ as given in (\ref{cut-off}). Restricting ourselves to this space, we now proceed to consider the large-$N$ limit. To this end, we introduce,
\begin{eqnarray}
\sigma &=& {2\pi \over m} (p-1)~, \quad m = 2N+1~,
\label{sigma}
\end{eqnarray}
which becomes, at large $N$, a continuous parameter with range from $0$ to $2\pi$ as $p$ goes from $1$ to $m$. We will identify this as the internal parameter of the loop. Moreover, at large $N$, we will restrict the values of $\{x_p\}$ to be such that the loop is smooth. This implies that the DPC given in eq.(\ref{z-bar-gen}) takes the following form at large $N$,
\begin{eqnarray}
\bar z^{\bar a} &\to & x^{\alpha}(\sigma)~,
\label{zbar-xsigma}
\end{eqnarray}
where $x(\sigma) \in C^{\infty}(S^1, {\cal M})$. Notice that the coordinate index $\alpha$ in eq.(\ref{z-bar-gen}) carried a discrete subindex corresponding to different copies of ${\cal M}$. This has now become a continuous parameter and we have removed it in the above equation for simplicity, with the (usual) understanding that the value of $\alpha$ is chosen independently for different values of $\sigma$. It is now clear how our notations for DPC must be transformed into the usual {\it loopy} notations in the large-$N$ limit. For example, for the tensor in (\ref{example-tbar}) we must have,
\begin{eqnarray}
t'_p{}^{\alpha_p \beta_p}{}_{\xi_p} (x_p) &\to& T^{\alpha \beta}{}_{\xi}(x(\sigma))~.
\end{eqnarray}
Therefore the norm in eq.(\ref{v-mod-squared}) takes the following form,
\begin{eqnarray}
|\bar v(\bar z)|^2 &\to& \oint {d\sigma \over 2\pi} V^{\alpha}(x(\sigma)) V^{\beta}(x(\sigma)) G_{\alpha \beta}(x(\sigma)) ~,
\label{LM-metric-usual}
\end{eqnarray}
where we have used the following large-$N$ property,
\begin{eqnarray}
{1\over m} \sum_{p=1}^m \cdots &\to& \oint {d\sigma \over 2\pi} \cdots ~.
\label{sum-int}
\end{eqnarray}
Equation (\ref{LM-metric-usual}) is the standard way of specifying the metric on ${\cal L}{\cal M}$ and we have shown here how this description is related to a large-$N$ limit in DPC.
\subsection{FNC and tubular geometry}
\label{s:FNC-LM}
Here we will implement the large-$N$ limit in the tubular construction as discussed in \S \ref{tub:M^N} applied to ${\cal L}^{(N)}{\cal M}$. While the general approach of \S \ref{tub:M^N} remains the same, we will incorporate certain important modifications.
Recall that to identify FNC, it was very crucial to first separate out directions which are tangential and orthogonal to the submanifold and then to scale the orthogonal coordinates in such a way that the transverse part of the metric is flat at the leading order everywhere on the submanifold - see eq.(\ref{ghat-submanifold}). This separation was done by using an orthogonal matrix which made the transverse coordinates real. However, if we allow the transverse coordinates to be complex, then the same can also be achieved through a specific unitary matrix corresponding to a discrete Fourier transform. In the large-$N$ limit such coordinates correspond to non-zero left and right moving momentum modes on the loop in $T{\cal M}$-description as explained below eq.(\ref{oint-t}). It is this system that we are going to use to describe the tubular geometry in ${\cal L}{\cal M}$.
Below we list the steps to be followed in order to translate tubular expressions in ${\cal L}^{(N)}{\cal M}$ to the corresponding ones in ${\cal L} {\cal M}$.
\begin{enumerate}
\item {\bf Complex FNC:} Replace orthogonal matrix $O$ by a unitary matrix $U$ (to be given below) and $O^T$ by $U^{\dagger}$.\footnote{The matrix $O^T$ appeared in various expressions because of the involvement of $\underline{K} = \underline{J}^{-1}$, which now contains $U^{\dagger}$. } The resultant FNC is complex and we rename the transverse indices in the following way,
\begin{eqnarray}
A \to (\hat \alpha, \hbox{a})~, \quad \hbox{a} = - N, - N + 1, \cdots , (\neq 0), \cdots N-1, N ~.
\end{eqnarray}
\item {\bf Discrete Fourier Transform (DFT) on $({\cal M}^{2N+1})_C$:} Use the following expressions for the unitary matrix elements,
\begin{eqnarray}
U_{0p} &=& {1\over \sqrt{m}} ~, \quad U_{\hbox{a} p} = {1\over \sqrt{m}} e^{-{2\pi i\over m} (p-1) \hbox{a}}~.
\label{U-def}
\end{eqnarray}
\item
{\bf Large $N$/continuum limit:} After introducing the parameter in eq.(\ref{sigma}), we take the large-$N$ limit.
\begin{itemize}
\item
In this limit the unitarity conditions are preserved in the following manner,
\begin{eqnarray}
\sum_p U_{\hbox{a} p} (U^{\dagger})_{p\hbox{a}'} &\to & \oint {d\sigma \over 2\pi} e^{-i(\hbox{a} -\hbox{a}')\sigma} = \delta_{\hbox{a}, \hbox{a}'}~, \label{sum-p} \\
\sum_{\hbox{a} =- N}^{N} (U^{\dagger})_{p\hbox{a}} U_{\hbox{a} p'} = {1\over m} \sum_{\hbox{a} = -N}^N e^{{2\pi i\over m}\hbox{a} (p-p')} &\to & {1\over 2\pi} \sum_{\hbox{a} \in \ZZ} e^{i\hbox{a} (\sigma-\sigma')} = \delta(\sigma-\sigma')~. \cr &&
\label{sum-a}
\end{eqnarray}
While (\ref{sum-p}) directly follows from (\ref{sum-int}), the limit in (\ref{sum-a}) is true in the following sense,
\begin{eqnarray}
\sum_p \left( {1\over m} \sum_{\hbox{a} = -N}^N e^{{2\pi i\over m}\hbox{a} (p-p')} \right) = 1 = \sum_{\hbox{a} \in \ZZ} \oint {d\sigma \over 2\pi} e^{i\hbox{a} (\sigma-\sigma')} ~.
\end{eqnarray}
\item
Finally, we follow the general prescription of transiting from the discrete DPC-notations to the usual loopy-notations as discussed in \S \ref{s:LM-DPC}. For example, for $\xi_p$ defined near eq.(\ref{xi-transverse}) we have,
\begin{eqnarray}
\xi_p^{\alpha_p} \to \xi^{\alpha}(\sigma) ~, \quad \hbox{such that} \quad \oint d\sigma \xi^{\alpha}(\sigma) = 0~.
\end{eqnarray}
Therefore following eq.(\ref{yh-xibar}), the FNC should read,
\begin{eqnarray}
\hat y^A &=& {1\over \sqrt{m}} E^{(\hat \alpha)}{}_{\beta}(x) \sum_p U_{\hbox{a} p} \xi_p^{\beta} \to \oint {d\sigma \over 2\pi} e^{-i\hbox{a} \sigma} \hat Y^{\hat \alpha}(\sigma)~,
\label{yhat-Yhat}
\end{eqnarray}
where $\hat Y^{\hat \alpha}(\sigma) = E^{(\hat \alpha)}{}_{\beta}(x) \xi^{\beta}(\sigma)$. Therefore the latter is the description of the loop in RNC (see eq.(\ref{Yh-Y'}, \ref{Y'-xi})) centred at $x \in {\cal M}$, the latter being the CM of the loop.
\end{itemize}
\end{enumerate}
Following the above steps one can re-work-out the expressions analogous to those in eqs.(\ref{ehat0-A}, \ref{pihat-lowerbar}). The final results are as follows,
\begin{eqnarray}
\hat e_0^{(A)}{}_{\beta}
&=& \Omega_{\beta}{}^{(\hat \alpha)}{}_{\delta}(x) \oint {d\sigma \over 2\pi} e^{-i\hbox{a} \sigma} \xi^{\delta}(\sigma)~, \cr
\underline{\hat \pi^{(a)}{}_{(b)}}(\{s\}_n, \hat y) &=& \oint {d\sigma \over 2\pi}
e^{-i(\hbox{a} -\hbox{b})\sigma} \Pi_x^{(\hat \alpha)}{}_{(\hat \beta)} (\{s\}_n,\xi (\sigma))~, \quad \hbox{a}, \hbox{b} \in \ZZ ~,
\end{eqnarray}
where just like in (\ref{pihat-lowerbar}), to reduce clutter we have combined four equations into one by allowing the indices $\hbox{a}$ and $\hbox{b}$ to take the value $0$.
Similar rules were suggested relating geometric quantities in ${\cal L}{\cal M}$ and the corresponding ones in ${\cal M}$ in general coordinates in earlier work \cite{dwv}. What we suggest here is that such rules better be defined for tubular expressions. This implies that in order to express a geometric quantity of ${\cal L} {\cal M}$ written in general coordinates in terms of ${\cal M}$-data, one may first write it in the form of tubular expansion and then evaluate each term in the expansion in terms of ${\cal M}$-data following the above procedure.
\subsection{Metric-expansion up to quartic order}
\label{s:metric-LM}
In order to perform tubular expansion of any tensor in ${\cal L}{\cal M}$, one uses a similar method as described at the beginning of \S \ref{s:metric-M^N}. The difference is that now one specialises to ${\cal L}={\cal L}^{(N)}{\cal M}$, replaces $O$ by $U$ etc and finally performs the continuum limit. The relevant details for the metric-expansion are given in Appendix \ref{a:metric}. The final results are given below,
\begin{eqnarray}
\hat g_{\alpha \beta}
&=& G_{\alpha \beta} + (\Omega_{\alpha}{}^{\eta}{}_{(\hat \xi^1)} \Omega_{\beta \eta (\hat \xi^2)} + R_{\alpha (\hat \xi^1 \hat \xi^2) \beta} ) \oint {d\sigma \over 2\pi} \hat Y^{\hat \xi^1 \hat \xi^2 }(\sigma) \cr
&& +( R_{\alpha (\hat \xi^1 \hat \xi^2) \eta } \Omega_{\beta}{}^{\eta}{}_{(\hat \xi^3) }
+ R_{\beta (\hat \xi^1 \hat \xi^2) \eta} \Omega_{\alpha}{}^{\eta}{}_{(\hat \xi^3)}
+ {1\over 3} \nabla_{(\hat \xi^1)} R_{\alpha (\hat \xi^2 \hat \xi^3) \beta} )
\oint {d\sigma \over 2\pi} \hat Y^{\hat \xi^1 \cdots \hat \xi^3}(\sigma) \cr
&& + \left\{ \Omega_{\alpha}{}^{\eta}{}_{(\hat \xi^1)} \Omega_{\beta}{}^{\zeta}{}_{(\hat \xi^2)} R_{\eta (\hat \xi^3 \hat \xi^4) \zeta} + {1\over 3} ( \nabla_{(\hat \xi^1)} R_{\alpha (\hat \xi^2 \hat \xi^3) \eta} \Omega_{\beta}{}^{\eta}{}_{(\hat \xi^4)} + \alpha \leftrightarrow \beta ) \right. \cr
&& \left. + {1\over 12} \nabla_{(\hat \xi^1)} \nabla_{(\hat \xi^2)} R_{\alpha (\hat \xi^3 \hat \xi^4) \beta} + {1\over 3} R_{\alpha (\hat \xi^1 \hat \xi^2) \eta} R^{\eta}{}_{(\hat \xi^3 \hat \xi^4) \beta} \right\}
\oint {d\sigma \over 2\pi} \hat Y^{\hat \xi^1 \cdots \hat \xi^4} (\sigma) + O(\hat Y^5) ~, \cr &&
\label{g-al-bet-LM}
\\
g_{\alpha B}
&=& \Omega_{\alpha}{}_{(\hat \beta \hat \xi)} \oint {d\sigma \over 2\pi} e^{i \hbox{b} \sigma} \hat Y^{\hat \xi}(\sigma)
+ {2\over 3} R_{\alpha (\hat \xi^1 \hat \xi^2 \hat \beta)} \oint {d\sigma \over 2\pi} e^{i\hbox{b} \sigma} \hat Y^{\hat \xi^1 \hat \xi^2}(\sigma) \cr
&& + ( {2\over 3} \Omega_{\alpha}{}^{\eta}{}_{(\hat \xi^1)} R_{\eta (\hat \xi^2 \hat \xi^3 \hat \beta) } + {1\over 4} \nabla_{(\hat \xi^1)} R_{\alpha (\hat \xi^2 \hat \xi^3 \hat \beta)} ) \oint {d\sigma \over 2\pi} e^{i\hbox{b} \sigma} \hat Y^{\hat \xi^1 \cdots \hat \xi^3}(\sigma) \cr
&& + ( {1\over 15} \nabla_{(\hat \xi^1)} \nabla_{(\hat \xi^2)} R_{\alpha (\hat \xi^3 \hat \xi^4 \hat \beta)} + {2\over 15} R_{\alpha (\hat \xi^1 \hat \xi^2) \eta} R^{\eta}{}_{(\hat \xi^3 \hat \xi^4 \hat \beta) } \cr
&& + {1\over 4} \Omega_{\alpha}{}^{\eta}{}_{(\hat \xi^1)} \nabla_{(\hat \xi^2)} R_{\eta (\hat \xi^3 \hat \xi^4 \hat \beta) } ) \oint {d\sigma \over 2\pi} e^{i\hbox{b} \sigma} \hat Y^{\hat \xi^1 \cdots \hat \xi^4}(\sigma) + O(\hat Y^5) ~,
\label{g-al-B-LM} \\
g_{AB}
&=& \eta_{AB} + {1\over 3} R_{(\hat \alpha \hat \xi^1 \hat \xi^2 \hat \beta)} \oint {d\sigma \over 2\pi} e^{i(\hbox{a} + \hbox{b})\sigma} \hat Y^{\hat \xi^1 \hat \xi^2}(\sigma) \cr
&& + {1\over 6} \nabla_{(\hat \xi^1)} R_{(\hat \alpha \hat \xi^2 \hat \xi^3 \hat \beta)} \oint {d\sigma \over 2\pi} e^{i(\hbox{a} + \hbox{b}) \sigma} \hat Y^{\hat \xi^1 \cdots \hat \xi^3}(\sigma) \cr
&& + ( {1\over 20} \nabla_{(\hat \xi^1)} \nabla_{(\hat \xi^2)} R_{(\hat \alpha \hat \xi^3 \hat \xi^4 \hat \beta)} + {2\over 45} R_{(\hat \alpha \hat \xi^1 \hat \xi^2) \eta} R^{\eta}{}_{(\hat \xi^3 \hat \xi^4 \hat \beta)} ) \oint {d\sigma\over 2\pi} e^{i(\hbox{a} + \hbox{b})\sigma} \hat Y^{\hat \xi^1 \cdots \hat \xi^4}(\sigma) \cr
&& + O(\hat Y^5) ~.
\label{g-A-B-LM}
\end{eqnarray}
where we have used the notation: $\hat Y^{\hat \alpha^1 \hat \alpha^2 \cdots}(\sigma) \equiv \hat Y^{\hat \alpha^1}(\sigma) \hat Y^{\hat \alpha^2}(\sigma) \cdots$ and
\begin{eqnarray}
\eta_{AB} &=& \eta_{\hat \alpha \hat \beta} \delta_{\hbox{a} + \hbox{b}, 0}~.
\end{eqnarray}
Notice that the above expression is different from the one in (\ref{ghat-submanifold}). This is simply because of the complex coordinates chosen here.
\subsection{Isometry}
\label{s:isometry}
Our discussion in this section so far shows how, given the tubular geometry around $\Delta \hookrightarrow {\cal L}^{(N)}{\cal M}$, a specific large-$N$ limit can be defined. As mentioned earlier, in order for it be the right tubular geometry of ${\cal L} {\cal M}$, it must satisfy the requirement of isometry. Here we will first discuss in \S \ref{ss:discrete} the discrete isometry of ${\cal L}^{(N)}{\cal M}$ and show how the reparametrization isometry arises in the large-$N$ limit. Then in \S \ref{ss:reparametrization} we will show that our large-$N$ geometry indeed satisfies the required Killing equation (in vielbein form) to all orders in tubular expansion.
\subsubsection{Discrete isometry and continuum limit}
\label{ss:discrete}
The discrete isometries of ${\cal M}^N$ that are independent of ${\cal M}$ are the ones that
permute factors of ${\cal M}$ in ${\cal M}^N$.
The transformation is given by,
\begin{eqnarray}
\bar z^{\bar a} \to \bar z'^{\bar a'} = ({\cal S}_N^{-1})^{\bar a'}{}_{\bar b} \bar z^{\bar b}
\label{perm-transf}
\end{eqnarray}
where ${\cal S}_N$ is the following matrix,
\begin{eqnarray}
{\cal S}_N = S_N \otimes {\hbox{ 1\kern-1.2mm l}}_d~,
\label{cS}
\end{eqnarray}
$S_N$ being an $N \times N$ permutation matrix, i.e.,
\begin{eqnarray}
(S_N)_{pq} &=& \delta_{\omega(p), q} = \delta_{p, \omega^{-1}(q)} ~,
\end{eqnarray}
where $\omega : \{ 1, 2, \cdots , N\} \to \{ 1, 2, \cdots , N\} $ is a bijection. Using eq.(\ref{ebar}), it is straightforward to show,
\begin{eqnarray}
\bar e'(z') &=& \bar e(z')~,
\end{eqnarray}
as required.
For an ordered product $({\cal M}^m)_C$, only cyclic permutations remain as isometries.
This is given by replacing $S_N$ above by the following,
\begin{eqnarray}
(C_N)_{pq} &=& \delta_{\psi(p), q}~,
\label{C-N}
\end{eqnarray}
where,
\begin{eqnarray}
\psi(p) &=& p + n ~ \hbox{mod } (m+1)~, \quad \forall n \in \{1, 2, \cdots , m \}~.
\label{psi-p}
\end{eqnarray}
We now look at the large-$N$ limit. The cyclic permutation becomes constant translation in terms of the parameter $\sigma$,
\begin{eqnarray}
\sigma \to \sigma' = \sigma + a~, \quad (a= {2\pi n \over m}) ~,
\end{eqnarray}
where $a$ remains finite when both $n$ and $m=2N+1$ become large. However, notice that (\ref{sigma}) is not the only way to introduce the continuous parameter $\sigma$. In the continuum limit one may introduce a {\it local density of points} in the following way. Consider a suitable range $\delta p$ centring around $p$ such that $\delta p/m$ is infinitesimally small. Then the most general way of introducing the continuous parameter is
\begin{eqnarray}
{\delta p \over m} = \sqrt{\gamma(\sigma)} \delta \sigma ~,
\end{eqnarray}
where $\delta \sigma$ and $\sigma$ correspond to $\delta p$ and $p$ respectively and $\sqrt{\gamma}$ is positive definite. For any other valid parametrization $\sigma'$ we must have,
\begin{eqnarray}
{d\sigma' \over d\sigma} &=& \sqrt{\gamma(\sigma) \over \gamma'(\sigma')} > 0~.
\end{eqnarray}
This is an orientation preserving diffeomorphism of the loop, i.e. an element of $\hbox{Diff}(S^1)$. In the discussion of \S \ref{s:FNC-LM} we fixed this ambiguity by choosing $\sqrt{\gamma(\sigma)} = 1/2\pi$.
\subsubsection{Reparametrization isometry}
\label{ss:reparametrization}
The reparametrization invariance that arises in the continuum limit manifests itself as an isometry of the loop space. The corresponding Killing vector is given by,
\begin{eqnarray}
\kappa^{\alpha} (\sigma) = \partial x^{\alpha}(\sigma)~, \quad [\partial \equiv {\partial \over \partial \sigma}]~.
\label{kappa-sigma}
\end{eqnarray}
In order to show that this is admitted by our large-$N$ geometry, we will prove that the following Killing equation in vielbein form \cite{tetrad} is satisfied,
\begin{eqnarray}
\hat \kappa^c \hat \partial_c \hat e^{(a)}{}_b + \hat \partial_b \hat \kappa^c \hat e^{(a)}{}_c = \chi^{(a)}{}_{(d)} \hat e^{(d)}{}_b ~, \quad [\hat \partial_a \equiv {\partial \over \partial \hat z^a}] ~,
\label{killing}
\end{eqnarray}
where $\hat \kappa^a$ are the components of the Killing vector in (\ref{kappa-sigma}) in FNC (as constructed in \S \ref{s:FNC-LM}) and
the matrix $\chi$ satisfies,
\begin{eqnarray}
\chi_{(ab)} + \chi_{(ba)} &=& 0~.
\label{chi}
\end{eqnarray}
Below we will first show that,
\begin{eqnarray}
\hat \kappa^{\alpha} = 0~, && \hat \kappa^A = i \hbox{a} \hat y^A~,
\label{kappah}
\end{eqnarray}
and then in Appendix \ref{a:killing-verification} we will verify that eq.(\ref{killing}) is indeed satisfied for $\hat \kappa$ given above to all orders in tubular expansion.
A heuristic argument that justifies eqs.(\ref{kappah}), which was used in \cite{semi-classical}, is as follows. Recall our $T{\cal M}$-description of loop configurations below eq.(\ref{oint-t}). According to this description and the subsequent construction of various coordinate systems,
loops can be described using the $T{\cal M}$ coordinates as $(x^{\alpha}, \hat Y^{\hat \alpha}(\sigma))$. Therefore, the reparametrization Killing vector is given in this description by $\hat \kappa^{\alpha}(\sigma) = \partial \hat Y^{\alpha}(\sigma)$.
Now notice that according to our construction in \S \ref{s:FNC-LM} the FNC in loop space, namely $\hat z^a = (x^{\alpha}, \hat y^A)$ is directly related to the above description through (\ref{yhat-Yhat}). This suggests,
\begin{eqnarray}
\hat \kappa^{\alpha} &=& \oint {d\sigma \over 2\pi} \partial \hat Y^{\alpha}(\sigma) = 0~, \quad \hat \kappa^A = \oint {d\sigma \over 2\pi} e^{-i\hbox{a} \sigma} \partial \hat Y^{\hat \alpha}(\sigma) = i\hbox{a} \hat y^A~.
\label{kappa-cal}
\end{eqnarray}
Our approach in this paper, on the other hand, has been to understand loop space as a large-$N$ limit. Therefore a more rigorous method to obtain the above result will be to first construct a suitable vector field in the cut-off space and then take the limit. This will be discussed in Appendix \ref{a:killing-finite}.
\section{Comments and outlook}
\label{comments}
Here we make some general comments regarding the choice of the cut-off space and certain more general physical applications (besides LSQM) that the results of this paper might end up finding.
\subsection{Choice of cut-off space}
\label{s:cut-off}
The cut-off space in eq.(\ref{cut-off}) is nothing but the total configuration space of a set of cyclically ordered
$(2N+1)$ number of particles which may be viewed as string bits. This is similar to the usual momentum cut-off on the worldsheet where the left and right moving Fourier modes are cut-off at the value $N$. As we have seen in \S \ref{tub:LM}, these modes, along with the zero/CM mode, are related to the string bits by a discrete Fourier transformation. There is another way to see why the number of string bits is taken to be $(2N+1)$, i.e. odd. The Killing vector field corresponding to reparametrization isometry of loop space happens to vanish identically on the submanifold of vanishing loops. In \cite{kobayashi} Kobayashi proved in finite dimensional case that the space of fixed points of a continuous isometry is a submanifold which (1) has even co-dimension and (2) is totally geodesic \cite{kobayashi-nomizu}. Although Kobayashi's theorem does not strictly apply in our infinite dimensional case, but $\Delta \hookrightarrow {\cal L}{\cal M}$ is a submanifold of even co-dimension in the following sense. The transverse directions are constituted by the non-zero left and right moving modes of the string and for each left moving transverse mode there exists a right moving one. Moreover the arguments of \cite{semi-classical} showed that the second fundamental form vanishes for $\Delta \hookrightarrow {\cal L}{\cal M}$, making it a totally geodesic submanifold - a feature that was crucially needed to get the right form of the tachyon effective equation. At finite $N$, again Kobayashi's theorem is invalid, but this time because of a different reason: there is no continuous isometry anymore, as the entire reparametrization isometry is replaced by the discrete isometry of cyclic permutations of the string bits. However, from the point of view of regularizing loop space quantum mechanics, it may be useful to preserve the above two features for $\Delta \hookrightarrow {\cal L}^{(N)}{\cal M}$. The property of even co-dimension dictates that we choose to work with odd number of string bits and, as we have seen explicitly in \S \ref{tub:M^N}, the diagonal submanifold of ${\cal M}^N$ is indeed a totally geodesic submanifold.
\subsection{Higher derivative corrections from finite models}
\label{s:HD}
A proper regularisation of LSQM would require a sensible finite-$N$ model to exist so that one can define cut-off versions of all possible worldsheet computations\footnote{Similar finite-$N$ string-bit models have appeared in various contexts in string theory \cite{thorn, discrete, pp-wave, ads}.}. If it is indeed possible to develop such a finite-$N$ theory, then that would be divergence-free and would describe a set of $(2N+1)$ number of interacting particles forming quantum bound states. Below we describe a line of thought with more general interests where these string inspired finite models may find applications as toy models.
Higher derivative (HD)/ $\alpha^{\prime}$ corrections come from the consideration of a string because it has an extended structure. This feature is independent of the nature of interactions that hold the extended body together. Therefore similar HD corrections are expected to appear in the effective theories of any other composite objects, in particular the naturally occurring ones\footnote{One may expect that this should also be true for gravitationally bound configurations, though the treatment for such classical bound configurations is expected to be very different from the quantum bound states, which is what we have in mind right now. }. Bound configurations are marked by the distinct feature that there exists an adiabatic decoupling between two different sets of degrees of freedom, namely the CM and {\it internal} degrees of freedom. These are {\it slow} and {\it fast} respectively in the Born-Oppenheimer sense. The question of interest is how to compute the HD corrections to the effective theory of the slow ones which are expected to encode the details of the {\it fast interactions}.
The approach of \cite{semi-classical} makes the above features explicit and tries to emphasise formulating a general mathematical framework (see also \cite{tubular}) of computing such corrections. One may imagine that such a framework should start from a covariant dynamical model written in positions space, i.e. the analogue of NLSM/LSQM. Then a semi-classical expansion is formulated by covariantly expanding the model around the space of all locations of the CM, which sits as a submanifold within the total configuration space.
It is of interest to investigate whether it is indeed possible to develop such a framework for naturally occurring bound configurations. While this question may not have a straightforward answer, the aforementioned string inspired finite models may be a suitable play ground for testing/developing this mathematical framework. In this case the tubular geometry around $\Delta \hookrightarrow {\cal L}^{(N)}{\cal M}$, which we have computed in this work, will be of direct use.
\begin{center}
{\bf Acknowledgement}
\end{center}
I would like to thank Indranil Biswas, T. R. Ramadas and S. Ramanan for helpful discussions. I am thankful to A. P. Balachandran and Sumit R. Das for interest and encouragement.
|
\section{Introduction}
The skew-symmetric non-Jacobi Courant bracket \cite{Co} on sections of ${\cal T}M:= TM\oplus T^\ast M$ was originally introduced by Courant to formulate the integrability condition defining a Dirac structure. However its nature became clear only due to the observation by Liu, Weinstein and Xu \cite{LWX} that $\cal T M$ endowed with the Courant bracket plays the role of a `double' object, in the sense of Drinfeld \cite{Dr}, for a pair of Lie algebroids over $M$. Whereas any Lie bialgebra has a double which is a Lie algebra, the double of a Lie bialgebroid is not a Lie algebroid, but a Courant algebroid -- a generalization of ${\cal T} M$ equipped with the Courant bracket. There is another way of viewing {\it Courant algebroids} as a generalization of Lie algebroids. This requires a change in the definition of the Courant bracket and the use of an analog of the non-antisymmetric Dorfman bracket \cite{Do}. The traditional Courant bracket then becomes the skew-symmetrization of the new one \cite{Roy}. This change replaces one defect with another: a version of the Jacobi identity is satisfied, while the bracket is no longer skew-symmetric. Such algebraic structures have been introduced by Loday \cite{Lo} under the name of {Leibniz algebras}. Canonical examples of Leibniz algebras arise often as {\em derived brackets} introduced by Kosmann-Schwarzbach \cite{K-S,YKS}. Since Leibniz brackets appear naturally in Geometry and Physics in the form of `algebroid brackets', i.e. brackets on sections of vector bundles, there were a number of attempts to formalize
the concept of {\em Leibniz algebroid} \cite{Ba,BV,G1,GKP,GM,ILMP,Ha,HM,KS,MM,SX,Wa}. Note also that a Leibniz algebroid is the horizontal categorification of a Leibniz algebra; vertical categorification leads to Leibniz $n$-algebras and Leibniz $n$-algebroids \cite{AP10,KMP11,KPQ,DP,BP12}.\medskip
In this article, we introduce the category of generalized Courant algebroids and show that it admits a free object on any anchored vector bundle. Our construction is based on the new concept of symmetric Leibniz algebroid. We compare this subclass of Leibniz algebroids with the subclass of Loday algebroids that was introduced and studied in~\cite{GKP}.
\medskip
The present paper is organized as follows. In Section 2, we recall the definitions of the categories of Leibniz algebroids \cite{ILMP}, of Leibniz pseudoalgebras -- their algebraic counterpart --, and of modules over them, as well as the classical notion of Courant algebroid. We then describe, in Section 3, two intersecting subclasses of Leibniz algebroids, namely the class of Loday algebroids (Definition \ref{DLodAld}) \cite{GKP}, which are Leibniz algebroids that admit a generalized right anchor and are thus geometric objects, and the class of symmetric Leibniz algebroids (Definition \ref{DSymLeiAld}), a new concept, made of Leibniz algebroids that satisfy two weak differentiability conditions on the left argument and contain Courant algebroids as a particular example. Examples of symmetric and nonsymmetric Leibniz and Loday algebroids are given. In Section 4, we motivate the definition of generalized Courant algebroids (Definitions \ref{DPrCrtAld} and \ref{DCrtAld}), which are specific symmetric Leibniz algebroids. The prototypical example of a generalized Courant algebroid is the one naturally associated to a symmetric Leibniz algebroid (Theorem \ref{Main}). This associated Courant algebroid is one of the two ingredients of the free Courant algebroid. Moreover, Theorem \ref{Main} allows to understand the origin of the definition of symmetric Leibniz algebroids. The second ingredient is the free symmetric Leibniz algebroid, which we construct in Section 5 (Theorem \ref{Main2} and Proposition \ref{Main3}). In Section 6, we combine the results of Section 4 and Section 5 to build the free Courant algebroid (Theorem \ref{Main4}).\medskip
\newcommand{\inser}[1]{\textcolor{blue}{#1}
\newcommand{\suppr}[1]{\textcolor{red}{\sout{#1}}
\newcommand{\define}[1]{\emph{#1}
\newcommand{\supprim}[1]{\textcolor{red}{\sout{#1}}
\newcommand{\frakX}{\mathfrak{X}}
\newcommand{\cat}[1]{\texttt{#1}
\section{Preliminaries}
\subsection{Notation and Conventions}
Unless otherwise specified, manifolds are made of a finite-dimensional smooth structure on a second-countable Hausdorff space.
If $[-,-]$ is a Leibniz bracket, we denote by $-\circ-$ its symmetrization, that is,
\begin{equation}
X \circ Y := [X,Y] + [Y,X]\;
\end{equation}
for any elements $X, Y$ of the Leibniz algebra.
\subsection{Anchored Vector Bundles and Anchored Modules}\label{subsec:anch}
\begin{defi}
If $M$ is a manifold, an \define{anchored vector bundle} over $M$ is a vector bundle $E \to M$ with a vector bundle morphism $a \colon E \to TM$, called its \define{anchor}.
If $R$ is a commutative unital ring and $\cA$ is a commutative unital $R$-algebra, an \define{anchored module} over $\cA$ is an $\cA$-module $\cE$ with an $\cA$-module morphism $a \colon \cE \to \Der \cA$, called its \define{derivative}.
\end{defi}
Of course, here $TM$ is the tangent bundle of $M$ and $\Der\cA$ is the module of derivations of $\cA.$ If $a \colon E \to TM$ is an anchor, we still denote by $a \colon \Gamma E \to \frakX(M) := \Der(\Ci(M))$ the corresponding operation on sections defined pointwise.
Obviously, if $E$ is an anchored vector bundle over $M$ with anchor $a$, then its space $\Gamma E$ of sections is an anchored module over $(\R,\Ci(M))$ with derivative $a$.\medskip
Morphisms of anchored vector bundles (resp., anchored modules) {over a fixed base} (resp., {over a fixed algebra}) are defined in an obvious way, and we obtain categories $\cat{AncVec}(M)$ and $\cat{AncMod}(\cA)$, respectively.
The algebroids (resp., pseudoalgebras) we are going to define in this article will be anchored vector bundles (resp., anchored modules) with extra structure.
They will form, together with their morphisms, categories that are concrete over ${\tt AncVec}(M)$ and ${\tt AncMod}(\cA)$, that is, admit a (faithful) forgetful functor to the latter.
One of our goals is to define left adjoints to these functors, or in other words, to define the free algebroid (resp., pseudoalgebra) of a given type on a given anchored vector bundle (resp., anchored module).
These constructions can be seen as the horizontal categorification of the free Lie, Leibniz$\ldots$ algebra on a given vector space.\medskip
For the different types of pseudoalgebras we are going to define, we will also define \define{modules} over them, using the following general principle: if $V$ is an $R$-module with extra structure, then a \define{$V$-module} is an $R$-module $W$ such that $V \oplus W$ is of the same type as $V$ and contains $V$ as a subobject and $W$ as an abelian ideal in an appropriate sense.
Similarly, a morphism of modules from the $V$-module $W$ to the $V'$-module $W'$ will be a morphism $V \oplus W \to V' \oplus W'$ sending $V$ to $V'$ and $W$ to $W'$.
It is possible to make these statements precise, but we prefer to keep them heuristic here and to work out the details below in the case of Leibniz modules.
\subsection{Leibniz Algebroids}
In this paper, we consider \emph{left Leibniz brackets}, i.e. bilinear brackets that satisfy the Jacobi identity
\begin{equation}\label{eq:jacobi}
[X,[Y,Z]]=[[X,Y],Z]+[Y,[X,Z]]\;.
\end{equation}
We first recall the definition of a Leibniz algebroid given in \cite{ILMP}. Note that this notion of Leibniz algebroid does not impose any differentiability requirement on the first argument of the bracket and is thus not a geometric concept.
\begin{defi}\label{LeiOid} A {\em Leibniz algebroid} is an anchored vector bundle $E\to M$ together with a Leibniz bracket $[-,-]$ on its space $\zG E$ of sections, which satisfy
\begin{equation}\label{axiom1}
[X,fY]=f[X,Y]+a(X)(f)Y,
\end{equation}
for any $f\in C^{\infty}(M)$ and $X,Y\in\Gamma E$.
\end{defi}
It is easily checked that the Leibniz rule (\ref{axiom1}) and the Jacobi identity imply that $a$ is a Leibniz algebra morphism: \begin{equation}\label{axtiom2}
a[X,Y]=[a(X),a(Y)],
\end{equation}
where the {\small RHS} bracket is the Lie bracket on $\Gamma TM$.
Let us also mention that here and in the following we consider {\it left} Leibniz brackets, i.e. bilinear brackets that satisfy the Jacobi identity $$[X,[Y,Z]]=[[X,Y],Z]+[Y,[X,Z]]\;.$$
We will essentially deal with the algebraic counterpart of Leibniz algebroids:
\begin{defi}\label{LeiPsAlg} Let $R$ be commutative unital ring and let $\cA$ be a commutative unital $R$-algebra. A {\em Leibniz pseudoalgebra} (or {\em Leibniz-Rinehart algebra}) over $(R,\cA)$ is an anchored module $(\cE,a)$ over $\cA$ endowed with a Leibniz $R$-algebra structure $[-,-]$, such that, for all $f\in\cA$ and $X,Y\in\cE,$ \begin{itemize}\item $[X,fY]=f[X,Y]+a(X)(f)\,Y$ and \item $a[X,Y]=[a(X),a(Y)]$, where the {\small RHS} is the commutator.\end{itemize} \end{defi}
If the $\cA$-module $\cE$ is faithful, the last requirement is again a consequence of the Leibniz rule and the Jacobi identity.\medskip
The space of sections of a Leibniz algebroid over $M$ is obviously a Leibniz pseudoalgebra over the $\R$-algebra $\Ci(M)$.\medskip
Of course, if, in Definitions \ref{LeiOid} and \ref{LeiPsAlg}, the Leibniz bracket is antisymmetric, we get a Lie algebroid and a Lie pseudoalgebra, respectively. \medskip
Leibniz algebroids over $M$ and Leibniz pseudoalgebras over $(R,\cA)$ are the objects of categories ${\tt LeiOid}\,M$ and ${\tt LeiPsAlg}\,(R,\cA)$, respectively. The morphisms of these categories are defined as follows.
\begin{defi}
Let $(E_{1},[-,-]_1,a_1)$ and $(E_{2},[-,-]_2,a_2)$ be two Leibniz algebroids over a same manifold $M$. A {\em Leibniz algebroid morphism} between them is a bundle map $\phi:E_{1}\to E_{2}$ such that $a_{2}\,\phi=a_{1}$ and $\phi[X,Y]_1=[\phi X,\phi Y]_2$, for any $X,Y\in\Gamma E_{1}$.
\end{defi}
\begin{defi}
Let $(\cE_{1},[-,-]_1,a_1)$ and $(\cE_{2},[-,-]_2,a_2)$ be two Leibniz pseudoalgebras over the same pair $(R,\cA)$. A {\em Leibniz pseudoalgebra morphism} between them is an $\cA$-module morphism $\phi:\cE_{1}\to \cE_{2}$ such that $a_{2}\,\phi=a_{1}$ and $\phi[X,Y]_1=[\phi X,\phi Y]_2$, for any $X,Y\in\cE_{1}$.
\end{defi}
We now define (bi)modules over Leibniz algeboids and pseudoalgebras.\medskip
Recall first the definition of a module over a Leibniz $R$-algebra $(V,[-,-])$.
By the general heuristic described above, this is an $R$-module $W$ with a Leibniz $R$-algebra structure $[-,-]$ on $V \oplus W$ containing $V$ as a subalgebra and $W$ as an abelian ideal.
Therefore, this bracket has to be the original bracket on $V \times V$, and 0 on $W \times W$, so it is determined by the values of $[x,w]$ and $[w,x]$, where $x \in V$ and $w \in W$.
Setting $\mu^l(x)(w) = [x,w]$ and $\mu^r(x)(w) = [w,x]$, we recover the usual notion: A {\it (bi)module} over a Leibniz $R$-algebra $(V,[-,-])$ is an $R$-module $W$ together with a left and a right {\it action} $\zm^l\in\op{Hom}_R(V\otimes_R W,W)$ and $\zm^r\in\op{Hom}_R(W\otimes_R V,W)$, which satisfy the following requirements \be\label{VVW}
\zm^r[x,y]=\zm^r(y)\zm^r(x)+\zm^l(x)\zm^r(y)\;,\ee \be\label{WVV}
\zm^r[x,y]=\zm^l(x)\zm^r(y)-\zm^r(y)\zm^l(x)\;,\ee \be\label{VWV}
\zm^l[x,y]=\zm^l(x)\zm^l(y)-\zm^l(y)\zm^l(x)\;,\ee for all $x,y\in V.$
In particular, let $\nabla$ be a representation of $(V,[-,-])$ on $W$, i.e. a Leibniz $R$-algebra morphism $V\to \op{End}_R(W)$. Then $\zm^\ell=\nabla$ and $\zm^r=-\nabla$ is a module structure over $V$ on $W$. \medskip
\begin{defi}\label{algebroidmodule}
Let $(E_{1},[-,-],a)$ be a Leibniz algebroid over $M$. A {\em module over the Leibniz algebroid} $E_1$ is a $C^{\infty}(M)$-module $\cE_2$ (not necessarily a locally free sheaf of $\Ci$-modules, i.e. not necessarily a module of sections of a vector bundle), which is a module $(\zm^\ell,\zm^r)$ over the Leibniz $\R$-algebra $\Gamma E_{1}$ whose left action satisfies the Leibniz rule
\be\label{LeibRule}
\zm^\ell(X)(fY)=f\zm^\ell(X)(Y)+a(X)(f)Y\;,
\ee
for any $f\in\Ci(M), X\in\Gamma E_{1}$, and $Y\in \cE_{2}$ (for convenience, we denote the left $\zG E_1$-action on $\cE_2$ and the Leibniz bracket on $\zG E_1$ by the same symbol).\end{defi}
\begin{rem}
In this definition, the $\Ci(M)$-module $\cE_2$ is not required to define a locally free sheaf of $\Ci(M)$-modules, i.e. it is not required to be a module of sections of a vector bundle.
\end{rem}
Similarly,
\begin{defi}\label{pseudoalgebramodule}
Let $(\cE_{1},[-,-],a)$ be a Leibniz pseudoalgebra over $(R,\cA)$. A {\em module over the Leibniz pseudoalgebra} $\cE_1$ is an $\cA$-module $\cE_2$ (hence an $R$-module), which is a module over the Leibniz $R$-algebra $\cE_{1}$ whose left action $\zm^\ell$ satisfies the Leibniz rule
$$
\zm^\ell(X)(fY)=f\zm^\ell(X)(Y)+a(X)(f)Y\;,
$$
for any $f\in\cA, X\in\cE_{1}$, and $Y\in \cE_{2}$.\end{defi}
In the case $\cE_1=\zG E_1$ and $\cE_2=\zG E_2$, where $E_1$ and $E_2$ are vector bundles over $M$, and $(\zm^\ell,\zm^r)=(\nabla,-\nabla)$, where $\nabla$ is a representation of $\zG E_1$ on $\zG E_2$, we deal with an $\R$-bilinear map $\nabla:\zG E_1\times \zG E_2\to \zG E_2$ such that, for any $f\in\Ci(M)$, $X,X_1,X_2\in\zG E_1$, and $Y\in\zG E_2$, $$\nabla_{[X_1,X_2]}=[\nabla_{X_1},\nabla_{X_2}]$$ and $$\nabla_X(fY)=a(X)(f)Y+f\nabla_XY\;.$$ If $\nabla$ is in addition $\Ci(M)$-linear in its first argument, the module structure is nothing but a flat $E_1$-connection on $E_2$. In the case of a Lie algebroid $E_1$, we thus recover the classical concept of $E_1$-module.
Note also that for any Leibniz algebroid $(E_1,[-,-],a)$, the $\Ci(M)$-module $\cE_2=\zG(M\times\R)=\Ci(M)$ and the actions $(\zm^\ell,\zm^r)=(a,-a)$ define on $\Ci(M)$ a module structure over the Leibniz algebroid $E_1$.
Let us emphasize that, just as for the Leibniz bracket on $\zG E_1$, we do not impose any differentiability condition on $\zm^r$.\medskip
Finally we define morphisms of modules over Leibniz pseudoalgebras.
From the above heuristic, a morphism from the $\cE_1$-module $\cE_2$ to the $\cE_1'$-module $\cE_2'$ should be a Leibniz pseudoalgebra morphism from $\cE_1 \oplus \cE_2$ to $\cE_1' \oplus \cE_2'$ sending $\cE_1$ to $\cE_1'$ and $\cE_2$ to $\cE_2'$. Unpacking this principle gives the following
\begin{defi} Let $(\cE_1,[-,-],a)$ and $(\cE'_1,[-,-]',a')$ be two Leibniz pseudoalgebras, and let $(\cE_2,\zm^\ell,\zm^r)$ and $(\cE'_2,\zm'^\ell,\zm'^r)$ be an $\cE_1$-module and an $\cE'_1$-module, respectively. A {\em morphism} between these two modules, is a pair $(\zf_1,\zf_2)$ made of a morphism $\zf_1:\cE_1\to\cE'_1$ of Leibniz pseudoalgebras and an $\cA$-linear map $\zf_2:\cE_2\to\cE'_2$, such that \be\label{RespActs}\zm'^\ell(\zf_1\times\zf_2)=\zf_2\,\zm^\ell\;\;\text{and}\;\;\zm'^r(\zf_2\times\zf_1)=\zf_2\,\zm^r\;.\ee\end{defi}
\subsection{Courant Algebroids}
As for the definition of Courant algebroids, we refer the reader to \cite{LWX}, \cite{Roy0}, \cite{GM2}, \cite{Uch}, and \cite{Kos3}.
\begin{defi}\label{CrtAld} A {\em Courant algebroid} is an anchored vector bundle $E\to M$, with anchor $a$, together with a Leibniz bracket $[-,-]$ on $\zG E$ and a bundle map $(-|-) \colon E\otimes E\to M\times\R$ that is in each fiber nondegenerate symmetric, called \define{scalar product}, which satisfy \be\label{4} a(X)(Y|Y)=2(X|[Y,Y])\;,\ee \be a(X)(Y|Y)=2([X,Y]|Y)\;,\label{5} \ee for any $X,Y\in\zG E$.
\end{defi}
\medskip\noindent The nondegeneracy of the scalar product allows us to identify $E$ with its dual $E^*$, and we will use this identification implicitly in the following. Note that (\ref{4}) is equivalent to \begin{equation}\label{4a} a(X)(Y|Z)=(X|Y\circ Z)\;,\end{equation} where $Y\circ Z$ denotes the symmetrized bracket. Similarly, (\ref{5}) easily implies the invariance of the scalar product, \begin{equation}\label{6}a(X)(Y|Z)=([X,Y]|Z)+(Y|[X,Z])\;,\end{equation} which in turn shows that $a$ is the anchor of the left adjoint map: \be\label{zr} [X,fY]=f[X,Y]+a(X)(f)Y\;. \ee Hence, a Courant algebroid is a particular Leibniz algebroid. When defining a derivation $D:\Ci(M)\to\zG E$ by \be\label{D}(Df|X)=a(X)(f)\;,\ee we get out of (\ref{4a}) that \be\label{4c}D(Y|Z)=Y\circ Z=[Y,Z]+[Z,Y]\;.\ee The fact that (\ref{4c}) is a consequence of the `invariance' condition (\ref{4a}) and the nondegeneracy of the scalar product, will be of importance later on. Let us moreover stress that (\ref{4c}) implies a differentiability condition for the first argument of the Leibniz bracket: \be\label{DiffCondFirst} [fX,Y]=f[X,Y]-a(Y)(f)X+(X|Y)Df\;.\ee
It is now clear that
\begin{prop} Courant algebroids $(E,[-,-],(-|-),a)$ are exactly the Leibniz algebroids $(E,[-,-],a)$ endowed with a scalar product $(-|-)$, such that, for any $X,Y,Z\in\zG E$,
\be\label{Inv1}a(X)(Y|Z)=([X,Y]|Z)+(Y|[X,Z])\;,\ee
\be\label{Inv2}a(X)(Y|Z)=(X|[Y,Z]+[Z,Y])\;.\ee
\end{prop}
As already indicated above, we view the conditions (\ref{Inv1}) and (\ref{Inv2}), as well as their consequence \be\label{Inv3}([X,Y]|Z)+(Y|[X,Z])=(X|[Y,Z]+[Z,Y])\;,\ee as the invariance properties of the scalar product. We will come back to this idea in Subsection \ref{Back}.\medskip
Finally, we define the algebraic version of Courant algebroids \cite{jub}.
\begin{defi}
A \define{Courant pseudoalgebra} over an $R$-algebra $\cA$ is an anchored module $(\cE,a)$ endowed with a Leibniz bracket $[-,-]$ and with a nondegenerate bilinear symmetric form $(-|-) \colon \cE \times \cE \to \cA$, such that
\be
([X,Y]|Y) = (X|[Y,Y])\;,\ee
\be a(X) (Y|Z) = ([X,Y]|Z) + (Y|[X,Z])\;,\ee
for any $X, Y, Z \in \cE$.
\end{defi}
As in the geometric case, these conditions imply, under the identification of $\cE$ with $\cE^*$ given by the scalar product, that
\begin{equation}
a(X) (Y|Z) = (X | Y \circ Z)\;,
\end{equation}
for any $X, Y, Z \in \cE$.\medskip
Obviously, if $E \to M$ is a Courant algebroid, then $\Gamma E$ is a Courant pseudoalgebra over $\Ci(M)$.
\section{Subclasses of Leibniz Algebroids}
\subsection{Loday Algebroids}
In \cite{GKP}, the authors consider specific Leibniz algebroids, called {\it Loday algebroids}, which have a right anchor satisfying a condition analogous to (\ref{DiffCondFirst}), and show that almost all Leibniz algebroids met in literature are Loday algebroids in their sense.
\begin{defi}\label{DLodAld} A {\em Loday algebroid} is a Leibniz algebroid $(E,[-,-],a)$ equipped with a derivation $$\textsf{D}:C^{\infty}(M)\to\Hom_{C^{\infty}(M)}(\zG (E^{\ot 2}),\zG E)\;,$$ such that
\begin{equation}\label{RightDiffProp}
[fX,Y]=f[X,Y]-a(Y)(f)X+(\textsf{D}f)(X,Y)\;,
\end{equation} for any $X, Y \in \zG E$ and $f \in \Ci(M)$.
\end{defi}
Let us mention that the right anchor $\textsf{\em D}$ can be viewed as a bundle map $\textsf{\em D}:E\to TM\otimes \op{End}E$ (whereas the left anchor is a bundle map $a:E\to TM$). Its local form is $$(\textsf{\em D}f)(X^ie_i,Y^je_j)=X^i \textsf{\em D}\,_{ij}^{\ell k} \p_k f\; Y^j e_\ell$$ (whereas the local form of $a$ is $$a(X^ie_i)(f)Y=X^i a_i^k \p_k f\; Y^\ell e_\ell\;)\;.$$
\begin{ex}[\cite{GKP}, Section 5] Leibniz algebra brackets, Courant-Dorfman brackets, twisted Courant-Dorfman brackets, Courant algebroid brackets, brackets associated to contact structures, Grassmann-Dorfman brackets, Grassmann-Dorfman brackets for Lie algebroids, Lie derivative brackets for Lie algebroids, Leibniz algebroid brackets associated to Nambu-Poisson structures... are Loday algebroid brackets.\end{ex}
For instance, it is clear from what was said above that, in the case of Courant algebroids, the derivation $\textsf{\em D}$ is given by $$\textsf{\em D}:\Ci(M)\times\zG E\times \zG E\ni (f,X,Y)\mapsto (X|Y)Df\in \zG E\;.$$
The algebraic version of Loday algebroids is defined as follows:
\begin{defi} A {\em Loday pseudoalgebra} is a Leibniz pseudoalgebra $(\cE,[-,-],a)$ over $(R,\cA)$ equipped with a derivation $$\textsf{D}:\cA\to\Hom_{\cA}(\cE\otimes_{\cA}\cE,\cE)\;,$$ such that
\begin{equation}\label{RightDiffProp}
[fX,Y]=f[X,Y]-a(Y)(f)X+(\textsf{D}f)(X,Y)\;.
\end{equation}
\end{defi}
\subsection{Symmetric Leibniz Algebroids}
We now introduce another subclass of Leibniz algebroids, {\it symmetric Leibniz algebroids}, which satisfy two weak differentiability conditions on the left argument, and contain Courant algebroids as a particular example.
\begin{defi}\label{DSymLeiAld} A {\em symmetric Leibniz algebroid} is a Leibniz algebroid $(E,[-,-],a)$ over $M$, such that, for any $f\in C^{\infty}(M)$, $X,Y\in\Gamma E$,
\be\label{S1}X\c fY-(fX)\c Y=0\;\;\text{and}\;,\ee
\be\label{S2}[fX,Y\c Z]-[X,Y]\c fZ-(fY)\c[X,Z]=0\;.\ee
\end{defi}
The definition can be equivalently formulated as follows:
\begin{prop} A {\em symmetric Leibniz algebroid} is a Leibniz algebroid $(E,[-,-],a)$ over $M$, such that, for any $f\in C^{\infty}(M)$, $X,Y\in\Gamma E$,
\be\label{S1b}X\c fY-(fX)\c Y=0\;\;\text{and}\;,\ee
\be\label{S2b}([fX,Y]-f[X,Y])\c Z + Y\c ([fX,Z]-f[X,Z])=0\;.\ee\end{prop}
\begin{proof} It suffices to show that, for a Leibniz algebroid that satisfies the first condition, the conditions (\ref{S2}) and (\ref{S2b}) are equivalent. Note first that the Jacobi identity implies that $[X,Y\circ Z]=[X,Y]\circ Z+Y\c [X,Z]$. It now follows that (\ref{S2}) is equivalent to $$[fX,Y]\c Z+Y\c [fX,Z]=[fX,Y\c Z]$$ $$=[X,Y]\c fZ+(fY)\c[X,Z]=(f[X,Y])\c Z+Y\c f[X,Z]\;.$$ \end{proof}
The first condition (\ref{S1b}) means that the symmetrized bracket is $\Ci(M)$-linear between the two variables $X,Y$. The second condition (\ref{S2b}) is a $\Ci(M)$-linearity condition in a combination of symmetrized products. \medskip
\begin{prop} A Loday algebroid $(E,[-,-],a,\textsf{D})$ over a manifold $M$ is a symmetric Leibniz algebroid if and only if, for all $f\in\Ci(M)$ and all $X,Y,Z\in\zG E,$ \be\label{LodS1}(\textsf{D}f)(X,Y)=(\textsf{D}f)(Y,X)\;\;\text{and}\;\ee \be\label{LodS2}(\textsf{D}f)(X,[Y,Z]+[Z,Y])=(\textsf{D}f)([X,Y],Z)+(\textsf{D}f)(Y,[X,Z])\;.\ee \end{prop}
\begin{proof} It follows from the differentiability properties (\ref{axiom1}) and (\ref{RightDiffProp}), and from the antisymmetry $a[X,Y]=-a[Y,X]$ of the bracket of vector fields, that the $\Ci(M)$-linearity conditions (\ref{S1}) and (\ref{S2}) are equivalent to (\ref{LodS1}) and (\ref{LodS2}).\end{proof}
\begin{ex}\label{ExSym1} Leibniz algebra brackets, Courant-Dorfman brackets, twisted Courant-Dorfman brackets, Courant algebroid brackets, brackets associated to contact structures, Grassmann-Dorfman brackets, Grassmann-Dorfman brackets for Lie algebroids... are Loday algebroid and symmetric Leibniz algebroid brackets. \end{ex}
Indeed, since for a Courant algebroid $(E,[-,-],(-|-),a)$ over $M$, with derivation $D:\Ci(M)\to \zG E$, we have $\textsf{\em D}=(-|-)D$, the $\Ci(M)$-linearity conditions (\ref{LodS1}) and (\ref{LodS2}) are direct consequences of the symmetry and the invariance properties of the scalar product. For definitions concerning the other examples, we refer the reader to \cite{GKP}, Section 5.
\begin{ex} Lie derivative brackets for Lie algebroids, Leibniz algebroid brackets associated to Nambu-Poisson structures... are Loday algebroid but $(\,$usually$\,)$ nonsymmetric Leibniz algebroid brackets.\end{ex}
We examine the first example. Let $(E,[-,-]_E,a_E)$ be a Lie algebroid, let $d^E:\zG(\w^\bullet E^*)\to \zG(\w^{\bullet +1}E^*)$ be the Lie algebroid differential, and denote by $${\cal L}^E:\zG E\times \zG(\w^\bullet E^*)\ni (X,\zw)\mapsto i_Xd^E\zw+d^Ei_X\zw\in \zG(\w^\bullet E^*)$$ the Lie algebroid Lie derivative, where $i_X$ is the interior product. There is a Leibniz bracket on sections of the vector bundle $E\oplus\w E^*$. Indeed, set, for any $X,Y\in\zG E$ and any $\zw,\zh\in\zG(\w E^*)$, \be\label{CDA1}[X+\zw,Y+\zh]=[X,Y]_{E}+{\cal L}_X^E\zh\;.\ee This is a Loday algebroid bracket with left anchor $a(X+\zw)=a_E(X)$ and right anchor $$(\textsf{\em D}h)(X+\zw,Y+\zh)=d^E h\w i_Y\zw+a_E(X)(h)\zh\;,$$ see \cite{GKP}, Section 5.\medskip
If this Loday algebroid $E\oplus \w E^*$ is symmetric, Condition (\ref{LodS1}) is satisfied in particular for 0-forms, i.e. we have $$a_E(X)(h)g=a_E(Y)(h)f\;,$$ for any $f,g,h\in\Ci(M)$ and any $X,Y\in\zG E$. If we choose $f=0$ and $g=1$, we find that $a_E=0$, so that the considered Lie algebroid $E$ is a Lie algebra bundle ({\small LAB}). Conversely, if $E$ is a {\small LAB}, we get $(\textsf{\em D}h)(X+\zw,Y+\zh)=0$, since $(d^E h)(-)=a_E(-)(h)=0$; hence, the conditions (\ref{LodS1}) and (\ref{LodS2}) are satisfied and $E\oplus \w E^*$ is a symmetric Leibniz algebroid.\medskip
Let us close this section with the definition of symmetric Leibniz pseudoalgebras:
\begin{defi} A {\em symmetric Leibniz pseudoalgebra} is a Leibniz pseudoalgebra $(\cE,[-,$ $-],a)$ over some $(R,\cA)$, such that (\ref{S1}) and (\ref{S2}), or, equivalently, (\ref{S1b}) and (\ref{S2b}), are satisfied for all $f\in\cA$ and all $X,Y,Z\in\cE$. We denote by ${\tt SymLeiPsAlg}$ $\,(R,\cA)$ $\subset$ ${\tt LeiPsAlg}\,(R,\cA)$ the full subcategory of symmetric Leibniz pseudoalgebras. \end{defi}
\begin{ex}\label{ExSym2} An example of a symmetric Leibniz pseudoalgebra is the free symmetric Leibniz pseudoalgebra over an anchored module that we describe in Theorem \ref{Main2} and in Proposition \ref{Main3}. Observe that this symmetric Leibniz bracket is not Loday.\end{ex}
\section{Generalized Courant Algebroids}
\subsection{Generalized Courant Algebroids}
As mentioned before, one of the goals of this paper is to extend the concept of Courant algebroid and to construct the free object in the category of generalized Courant algebroids over any anchored vector bundle.
\begin{defi}\label{DPrCrtAld} Let $(\cE_1,[-,-],a)$ be a Leibniz pseudoalgebra over $(R,\cA)$ and let $(\cE_2,\zm^\ell,\zm^r)$ be an $\cE_1$-module. Assume further that $(-|-):\cE_1\times\cE_1\to\cE_2$ is a symmetric $\cA$-bilinear $\cE_2$-valued map, such that, for any $X,Y,Z\in\cE_1$, the `invariance relations' \be\label{LeftAct} \zm^\ell(X)(Y|Z)=([X,Y]|Z)+(Y|[X,Z])\;,\ee \be\label{RightAct} -\zm^r(X)(Y|Z)=([Y,Z]+[Z,Y]|X)\;\;\text{and}\ee \be\label{Equal}([X,Y]|Z)+(Y|[X,Z])=([Y,Z]+[Z,Y]|X)\;\ee hold true. We refer to such a tuple $(\cE_1,\cE_2,[-,-],(-|-), a,\zm^\ell,\zm^r)$ as a {\em generalized pre-Courant pseudoalgebra}.\end{defi}
\begin{rem}\label{Natural} In the geometric situation $\cE_1=\zG E_1$, where $E_1\to M$ is a vector bundle over a manifold, $R=\R$ and $\cA=\Ci(M)$, we can take $(\cE_2,\zm^\ell,\zm^r)=(\Ci(M),a,-a)$, which is actually an $E_1$-module. If we now assume in addition that $(-|-)$ is nondegenerate, the generalized pre-Courant pseudoalgebra is a classical Courant algebroid $(E_1,[-,-],$ $(-|-),a)$.\end{rem}
\begin{prop}\label{Motiv} Any generalized pre-Courant pseudoalgebra with nondegenerate scalar product $(-|-)$ is a Loday pseudoalgebra and a symmetric Leibniz pseudoalgebra.\end{prop}
It is clear that, for any $X\in\cE_1$, we have $(X|-)\in\op{Hom}_\cA(\cE_1,\cE_2)$. By nondegenerate scalar product, we mean here that any $\zD\in\op{Hom}_\cA(\cE_1,\cE_2)$ reads $\zD=(X|-)$, and that the map $Y\mapsto (Y|-)$ is injective, so that $X$ is unique. Indeed, in the aforementioned geometric case, nondegeneracy in each fiber, see Definition \ref{CrtAld}, implies these requirements.
\begin{proof} We use the above notation; in particular $f\in\cA$ and $X,Y,Z\in\cE_1$. In view of the invariance relations and Leibniz rule for the left action, we get \be\label{R1}-\zm^r(X)(f(Y|Z))=f([Y,Z]+[Z,Y]|X)+a(X)(f)(Y|Z)\;.\ee On the other hand, the Leibniz rule for the bracket $[-,-]$ gives \be\label{R2}-\zm^r(X)(fY|Z)=([fY,Z]+[Z,fY]|X)=([fY,Z]|X)+(f[Z,Y]+a(Z)(f)Y|X)\;.\ee Note now that $a(-)(f)(Y|Z)\in\op{Hom}_\cA(\cE_1,\cE_2)$, so that there is a unique $(\textsf{\em D}f)(Y,Z)$ $\in\cE_1$ such that \be\label{R3}((\textsf{\em D}f)(Y,Z)|-)=a(-)(f)(Y|Z)\;.\ee The properties of the anchor and the scalar product imply that $\textsf{\em D}$ is a derivation $$\textsf{\em D}:\cA\to \op{Hom}_\cA(\cE_1\otimes_\cA\cE_1,\cE_1)\;.$$ It now follows from (\ref{R1}), (\ref{R2}), (\ref{R3}), and nondegeneracy that $$[fY,Z]=f[Y,Z]-a(Z)(f)Y+(\textsf{\em D}f)(Y,Z)\;,$$ i.e. that $(\cE_1,[-,-],a)$ is a Loday pseudoalgebra.\medskip
Furthermore, the latter is a symmetric Leibniz pseudoalgebra if and only if the conditions (\ref{LodS1}) and (\ref{LodS2}) are satisfied -- the proof in the geometric situation remains valid in the present algebraic case --. It is easily seen that these requirements are fulfilled due to the invariance relation (\ref{Equal}). \end{proof}
In view of Remark \ref{Natural}, it would now be natural to define generalized Courant pseudoalgebras as generalized pre-Courant pseudoalgebras with nondegenerate scalar product. Actually we choose a more general definition, see Proposition \ref{Motiv}:
\begin{defi}\label{DCrtAld} A {\em generalized Courant pseudoalgebra} is a symmetric generalized pre-Courant pseudoalgebra.\end{defi}
Generalized Courant pseudoalgebras are a full subcategory ${\tt CrtAld}$ of the category ${\tt PrCrtAld}$ of generalized pre-Courant pseudoalgebras.
\begin{defi} A {\em morphism} between two generalized (pre-)Courant pseudoalgebras $$(\cE_1,\cE_2,[-,-],(-|-),a,\zm^\ell,\zm^r)\;\;\text{and}\;\;(\cE'_1,\cE'_2,[-,-]',(-|-)',a',\zm'^\ell,\zm'^r)$$ over the same pair $(R,\cA)$, is a morphism $(\zf_1,\zf_2)$ from the $\cE_1$-module $\cE_2$ to the $\cE'_1$-module $\cE'_2$, such that \be\label{RespScProd}(-|-)'\,(\zf_1\times\zf_1)=\zf_2\,(-|-)\;.\ee\end{defi}
\subsection{Generalized Courant Algebroid Associated to a Symmetric Leibniz Algebroid}\label{Back}
The next theorem describes this Courant algebroid. It is also the motivation for the introduction of symmetric Leibniz algebroids. Moreover, it will turn out that the generalized Courant algebroid associated to a symmetric Leibniz algebroid is one of the two components of the free Courant algebroid.
\begin{theo}\label{Main} Let $(\cE,[-,-],a)$ be a symmetric Leibniz pseudoalgebra over $(R,\cA)$. Denote by $\odot$ the symmetric tensor product over $\cA$, take the subset \begin{equation}\label{Inv} \text{\em Inv}=\{[X,Y]\odot Z+Y\odot [X,Z]-X\odot([Y,Z]+[Z,Y]): X,Y,Z\in\cE\}\;\end{equation} of the $\cA$-module $\cE^{\odot 2}$, and let $\la\op{Inv}\ra$ be the $\cA$-submodule of $\cE^{\odot 2}$ generated by $\op{Inv}$. The quotient $${\cal R}(\cE)=\cE^{\odot 2}/\la\op{Inv}\ra$$ is an $\cE$-module with actions $\tilde{\zm}^\ell$ and $\tilde{\zm}^r$ induced by \be\label{LeftAct} \zm^\ell(X)(Y_1\odot Y_2)=[X,Y_1]\odot Y_2+Y_1\odot[X,Y_2]\ee and \be\label{RightAct} -\zm^r(X)(Y_1\odot Y_2)=(Y_1\c Y_2)\odot X\;.\ee These data, together with the {\em universal scalar product} \be\label{UnivScProd} (-|-): \cE\times\cE\ni (X,Y)\mapsto (X\odot Y)^{\;\widetilde{}}\in{\cal R}(\cE)\;,\ee define a generalized Courant pseudoalgebra \be\label{GenCrtAld}\cC(\cE):=(\cE,{\cal R}(\cE),[-,-],(-|-),a,\tilde{\zm}^\ell,\tilde{\zm}^r)\;.\ee\end{theo}
\begin{ex} All the examples of symmetric Leibniz brackets described in Example \ref{ExSym1} and Example \ref{ExSym2} thus give rise to generalized Courant algebroids.\end{ex}
\begin{rem} The associated generalized Courant algebroid is a very natural construction over a symmetric Leibniz pseudoalgebra, whose scalar product is the universal scalar product given by the symmetric tensor product and whose actions are the `invariant' Courant actions.\end{rem}
We first prove the following
\begin{lem} The $\cA$-module $\cE^{\odot 2}$ is an $\cE$-module for the actions $\zm^\ell$ and $\zm^r$.\end{lem}
\begin{proof} (i) We first show that $\zm^\ell(X)$ and $\zm^r(X)$ are well-defined on $\cE^{\odot 2}$ (note that we do of course not intend to show that they are $\cA$-linear on $\cE^{\odot 2}$; indeed, they are visibly only $R$-linear). Since the {\small RHS}s of (\ref{LeftAct}) and (\ref{RightAct}) are symmetric in $Y_1,Y_2$, it suffices to prove that they respect the `defining relations' of the tensor product over $\cA$. The only nonobvious condition is that $Y_{1}\odot fY_{2}=(fY_{1})\odot Y_{2}$ be preserved. And indeed, we have
\begin{eqnarray*}
\zm^\ell(X)(Y_{1}\odot fY_{2})&=&[X,Y_{1}]\odot fY_{2}+Y_{1}\odot[X,fY_{2}]\\
&=&[X,Y_{1}]\odot fY_{2}+Y_{1}\odot f[X,Y_{2}]+Y_{1}\odot a(X)(f)Y_{2}\\
&=&(f[X,Y_{1}])\odot Y_{2}+(fY_{1})\odot[X,Y_{2}]+(a(X)(f)Y_{1})\odot Y_{2}\\
&=&[X,fY_{1}]\odot Y_{2}+(fY_{1})\odot[X,Y_{2}]\\
&=&\zm^\ell(X)((fY_{1})\odot Y_{2})\;
\end{eqnarray*}
and
\begin{eqnarray*}
\zm^r(X)(Y_{1}\odot fY_{2})&=&-(Y_1\c fY_2)\odot X\\
&=&-((fY_1)\c Y_2)\odot X\\
&=&\zm^r(X)((fY_{1})\odot Y_{2})\;.
\end{eqnarray*}
(ii) It remains to check the `Leibniz morphism conditions' (\ref{VVW}), (\ref{WVV}), and (\ref{VWV}), as well as the Leibniz rule (\ref{LeibRule}). The Leibniz rule $$\zm^\ell(X)(Y_1\odot fY_2)=f\zm^\ell(X)(Y_1\odot Y_2)+a(X)(f)(Y_1\odot Y_2)$$ is clear from (i). The morphism conditions are also straightforwardly checked. To verify for instance $$\zm^r[X,Z]=\zm^r(Z)\zm^r(X)+\zm^\ell(X)\zm^r(Z)\;,$$ note first that the right adjoint action $[-,X]$ on a symmetrized product vanishes: $$[[Y_1,Y_2]+[Y_2,Y_1],X]=[Y_1,[Y_2,X]]-[Y_2,[Y_1,X]]+[Y_2,[Y_1,X]]-[Y_1,[Y_2,X]]=0\;.$$ We now get
$$\zm^r[X,Z](Y_1\odot Y_2)=-(Y_1\c Y_2)\odot [X,Z]\;,$$
\begin{eqnarray*}
\zm^r(Z)\zm^r(X)(Y_1\odot Y_2)&=& -\zm^r(Z)((Y_1\c Y_2)\odot X)\\
&=& ((Y_1\c Y_2)\c X)\odot Z\\
&=& ([Y_1\c Y_2,X]+[X,Y_1\c Y_2])\odot Z\\
&=& [X,Y_1\c Y_2]\odot Z\;,
\end{eqnarray*}
and
\begin{eqnarray*} \zm^\ell(X)\zm^r(Z)(Y_1\odot Y_2)&=&-\zm^\ell(X)((Y_1\c Y_2)\odot Z)\\
&=&-[X,Y_1\c Y_2]\odot Z-(Y_1\c Y_2)\odot [X,Z]\;.
\end{eqnarray*}
Hence, the result.
\end{proof}
The symmetric $\cA$-bilinear map
$$
\la -|-\ra:
\cE^{\times 2}\ni (X,Y) \mapsto X\odot Y\in \cE^{\odot 2}\;
$$ satisfies
\be \label{U1} \zm^\ell(X)\la Y|Z\ra=\la[X,Y]|Z\ra+\la Y|[X,Z]\ra\ee and \be\label{U2}-\zm^r(X)\la Y|Z\ra=\la X|[Y,Z]+[Z,Y]\ra\;,\ee
which are similar to (\ref{Inv1}) and (\ref{Inv2}). Since (\ref{Inv3}) does however not hold in general, we consider the quotient $\cA$-module $${\cal R}(\cE)=\cE^{\odot 2}/\la\op{Inv}\ra\;.$$
\begin{lem} The $\cA$-module ${\cal R}(\cE)$ is an $\cE$-module for the actions $\tilde{\zm}^\ell$ and $\tilde{\zm}^r$ induced by $\zm^\ell$ and $\zm^r$.\end{lem}
\begin{proof} It suffices to show that the actions descend to the quotient; indeed, the induced maps then inherit the required properties.\medskip
(i) Left action. Let $I(X,Y,Z)$, or just $I$, be any element in $\op{Inv}\subset\cE^{\odot 2}$, and let $f\in\cA$ and $W\in\cE$. Since $$\zm^\ell(W)(f I)=f\zm^\ell(W)(I)+a(W)(f)I\;,$$ we have $\zm^\ell(W)\la\op{Inv}\ra\subset\la\op{Inv}\ra,$ if $\zm^\ell(W)\op{Inv}\subset\la\op{Inv}\ra.$ The latter actually holds true:
$$ \zm^\ell(W)I(X,Y,Z) $$ $$= [W,[X,Y]]\odot Z+[X,Y]\odot[W,Z]+[W,Y]\odot[X,Z]
+Y\odot [W,[X,Z]] $$ $$
-[W,X]\odot(Y\c Z)-X\odot([W,Y]\c Z)-X\odot(Y\c[W,Z])$$
$$ =[[W,X],Y]\odot Z+[X,[W,Y]]\odot Z
+[X,Y]\odot[W,Z]+[W,Y]\odot[X,Z] $$ $$ +
Y\odot[[W,X],Z]+Y\odot[X,[W,Z]]
-[W,X]\odot(Y\c Z)-X\odot([W,Y]\c Z)-X\odot(Y\c[W,Z])$$
$$=I([W,X],Y,Z)+I(X,[W,Y],Z)+I(X,Y,[W,Z])\;.
$$
(ii) Right action. In view of the annihilation of symmetrized products by right adjoint actions and due to the symmetry condition (\ref{S2}), $$[fX,Y\c Z]=[X,Y]\c fZ+(fY)\c [X,Z]\;,$$ we get
$$\zm^r(W)(f I(X,Y,Z))$$ $$=\zm^r(W)\left([X,Y]\odot fZ+(fY)\odot [X,Z]-(fX)\odot(Y\c Z)\right)$$ $$= -([X,Y]\c fZ+(fY)\c [X,Z]-[fX,Y\c Z]-[Y\c Z,fX])\odot W$$ $$=0\;.$$ \end{proof}
It follows from (\ref{U1}) and (\ref{U2}) that $(\cE,[-,-],a)$, $({\cal R}(\cE),\tilde{\zm}^\ell,\tilde{\zm}^r)$, and the symmetric $\cA$-bilinear `universal scalar product' $$(-|-):\cE^{\times 2}\ni (X,Y)\mapsto \la X|Y\ra^{\;\widetilde{}}\in {\cal R}(\cE)\;$$ define a generalized Courant pseudoalgebra. This completes the proof of Theorem \ref{Main}.
\section{Free Symmetric Leibniz Algebroid}
The free symmetric Leibniz algebroid is the second ingredient needed for the construction of the free Courant algebroid.
\subsection{Leibniz Pseudoalgebra Ideals}
\newcommand{\cF}{{\cal F}}
\begin{defi} Let $(\cE,[-,-],a)$ be a Leibniz pseudoalgebra over $(R,\cA)$. A {\em Leibniz pseudoalgebra ideal} is an $\cA$-submodule $\cI\subset\cE$, which is a two-sided Leibniz $R$-algebra ideal, i.e. $[\cI,\cE]\subset\cI$ and $[\cE,\cI]\subset\cI$, and which is contained in the kernel of the anchor, i.e. $\cI\subset\ker a$.\end{defi}
\begin{prop} The quotient of a Leibniz pseudoalgebra by a Leibniz pseudoalgebra ideal is a Leibniz pseudoalgebra for the induced bracket and anchor.\end{prop}
\begin{proof} Obvious.\end{proof}
\begin{lem} Let $(\cE,[-,-],a)$ be a Leibniz pseudoalgebra over $(R,\cA)$, let \be\label{Ideal1} J_1=\{X\c fY-(fX)\c Y: f\in\cA,\, X,Y\in\cE\}\;\;\text{and}\ee \be\label{Ideal2} J_2=\{[fX,Y\c Z]-[X,Y]\c fZ-(fY)\c[X,Z]: f\in\cA,\, X,Y,Z\in\cE\}\;,\ee and denote by $\la J_1\ra$ $(\,$resp., $\la J_2\ra$$\,)$ the $\cA$-module generated by $J_1$ $(\,$resp., $J_2$$\,)$. The $\cA$-module $(J_1+J_2):=\la J_1\ra+\la J_2\ra$ is an ideal of the Leibniz pseudoalgebra $\cE$, so that the quotient $\cE/(J_1+J_2)$ inherits a symmetric Leibniz pseudoalgebra structure.\end{lem}
\begin{proof} The left adjoint action $[W,-]$, $W\in\cE$, satisfies the Leibniz rule with respect to the Leibniz bracket $(X,Y)\mapsto [X,Y]$, the symmetrized bracket $(X,Y)\mapsto X\c Y$, and the $\cA$-module structure $(f,X)\mapsto fX$. It follows that $$[W,X\c fY-(fX)\c Y]=$$ $$[W,X]\c fY+X\c f[W,Y]+X\c\, a(W)(f) Y-\;$$ $$(f[W,X])\c Y-\,(a(W)(f)X)\c Y-(fX)\c [W,Y]\in \la J_1\ra\;,$$ and similarly for $J_2$.
As for the right action $[-,W]$, $W\in\cE$, recall that it vanishes on every symmetrized bracket. Since the first term $[fX,Y\c Z]=[fX,Y]\c Z+Y\c[fX,Z]$ of an element of $J_2$ is symmetric as well, the sets $J_1$ and $J_2$ vanish under the right action.\medskip
For any $f\in\cA,$ $W\in\cE,$ and $Q\in J_{1},$ we have now $$Q\c fW-(fQ)\c W=[Q,fW]+[fW,Q]-[fQ,W]-[W,fQ]\in J_1\;,$$ with $[Q,fW]=0$, $[fW,Q]\in \la J_{1}\ra$, and \be\label{Ida}[W,fQ]=f[W,Q]+a(W)(f)Q\in \la J_{1}\ra\;.\ee Hence, \be\label{Idb}[fQ,W]\in\la J_1\ra\;.\ee
Similarly, if $Q\in J_{2}$, $$ W\c fQ-(fW)\c Q=[W,fQ]+[fQ,W]-[fW,Q]-[Q,fW]\in J_1\;,$$ with \be\label{Idc}[W,fQ]=f[W,Q]+a(W)(f)Q\in \la J_{2}\ra\;,\ee $[fW,Q]\in \la J_2\ra$, and $[Q,fW]=0$. Therefore, \be\label{Idd}[fQ,W]\in J_{1}+\la J_{2}\ra\;.\ee\medskip Equations (\ref{Ida}), (\ref{Idb}), (\ref{Idc}), and (\ref{Idd}) imply that the $\cA$-submodule $\la J_1\ra +\la J_2\ra\subset\cE$ is a Leibniz $R$-algebra ideal.\medskip
To see that $\la J_{1}\ra +\la J_{2}\ra$ is a Leibniz pseudoalgebra ideal, it now suffices to recall that $a$ is $\cA$-linear and that any symmetrized bracket belongs to $\ker a$.\end{proof}
\subsection{Free Leibniz Pseudoalgebra}
There exists a forgetful functor $$\op{For}:{\tt (Sym)LeiPsAlg}\,(R,\cA)\to {\tt AncMod}(\cA)\;.$$ We write for short $\cat{C} := \cat{(Sym)LeiPsAlg}\,(R,\cA)$ and $\cat{D} := \cat{AncMod}(\cA)$. Therefore, for any $\cM\in{\tt D}$, we can define the free (symmetric) Leibniz pseudoalgebra over $\cM$. It is made of an object $\op{F(S)}\cM\in {\tt C}$ and a ${\tt D}$-morphism $i:\cM\to \op{F(S)}\cM$, such that, for any object $\cE\in{\tt C}$ and any ${\tt D}$-morphism $\zf:\cM\to \cE$, there is a unique ${\tt C}$-morphism $\Phi:\op{F(S)}\cM\to \cE$, such that $\Phi\, i=\zf$ :
\begin{equation} \begin{tikzpicture}
\matrix (m) [matrix of math nodes, row sep=3em, column sep=3em]
{ \cM & \op{F(S)}\cM \\
& \cE \\ };
\path[->]
(m-1-1) edge node[left] {$\zf$} (m-2-2)
(m-1-1) edge node[auto] {$i$}(m-1-2)
(m-1-2) edge node[right] {$\Phi$} (m-2-2);
\end{tikzpicture}
\end{equation}
We first recall the construction of the free Leibniz algebra over an $R$-module \cite{LP}. Let $V$ be an $R$-module and let $\overline{T} V=\bigoplus_{k\ge 1}V^{\otimes_{\!R}\,k}$ be the reduced tensor $R$-module over $V.$ The {\em universal Leibniz bracket} $[-,-]$ is defined by the requirement
$$
v_{1}\ot v_{2}\ot\cdots\ot v_{k}
=[v_{1},[v_{2},...,[v_{k-2},[v_{k-1},v_{k}]]...]]\;,
$$
$v_i\in V.$\medskip
For instance, $$[v_{1},v_{2}]=v_{1}\ot v_{2}\in V^{\otimes_{\!R}\,2}\;,$$ $$[v_1,v_2\ot v_3]=[v_1,[v_2,v_3]]=v_1\ot v_2\ot v_3\in V^{\ot_{\! R}\,3}\;,$$ $$[v_1\ot v_2,v_{3}]=[[v_1,v_2],v_3]=[v_1,[v_2,v_3]]-[v_2,[v_1,v_3]]=v_{1}\ot v_{2}\ot v_{3}-v_{2}\ot v_{1}\ot v_{3}\in V^{\ot_{\!R}\,3}\;.$$
\vspace{0.1mm}\begin{theo}\label{Main2} Let $(\cM,a)$ be an anchored $\cA$-module. The free Leibniz $(R,\cA)$-pseudo-algebra $\op{F}\cM$ over $\cM$ is the triple $(\overline{T}\cM,[-,-]_{\op{Lei}},\op{F}a)$, where $\overline{T}\cM$ is endowed with the $\cA$-module structure defined inductively by \be\label{ModStr}
f(m_{1}\ot m_{2}\ot \ldots \ot m_{n}):=
m_{1}\ot f(m_{2}\ot\ldots \ot m_{n})
-a(m_{1})(f)(m_{2}\ot\ldots \ot m_{n})\;
\ee $(f\in\cA, m_i\in\cM)$, where $[-,-]_{\op{Lei}}$ is the universal Leibniz bracket on $\overline{T}\cM$, and where $$\op{F}a:\overline{T}\cM\to \op{Der}\cA$$ is induced by $a:\cM\to\op{Der}\cA$ .
\end{theo}
\begin{proof} We denote by $\op{F}^n\cM=\cM^{\ot_{\! R}\,n}$ (resp., $\op{F}_n\cM=\bigoplus_{1\le k\le n}\cM^{\ot_{\!R}\,k}$) the grading (resp., the filtration) of $\op{F}\cM$.\medskip
\newcommand{\mf}{\mathbf{m}}
(i) Module structure. Equation (\ref{ModStr}) provides a well-defined $\cA$-module structure on $\op{F}_{n}\cM$, if we are given a well-defined $\cA$-module structure on $\op{F}_{n-1}\cM$, $n\ge 2$. Since the {\small RHS} of (\ref{ModStr}) is $R$-multilinear, the `action' is well-defined from $\op{F}^n\cM$ into $\op{F}_n\cM$. We extend it by linearity to $\op{F}_n\cM=\op{F}^n\cM\,\oplus\, \op{F}_{n-1}\cM$. It is now straightforwardly checked that this extension satisfies all the $\cA$-module requirements, except, maybe, the condition $f(g\zm)=(fg)\zm$, where $f,g\in\cA$ and $\zm\in\op{F}_n\cM$. As for the latter, note first that, if $f,g\in\cA$, $m,m_i\in\cM$, and $\mf=m_2\ot\ldots\ot m_n$, we have $$f(m\ot g\mf)=m\ot f(g\mf)-a(m)(f)(g\mf)\;,$$ since $g\mf$ is a finite sum $g\mf=\sum_{i\le n-1}\mf_i$, where $\mf_i\in\op{F}^{i}\cM$ is a decomposed tensor. Thus,
\begin{eqnarray*}
f(g(m\ot\mf))&=&f(m\ot g\mf)-f(a(m)(g)\mf)\\
&=&m\ot (fg)\mf-(a(m)(f)g)\mf-(fa(m)(g))\mf\\
&=&m\ot (fg)\mf-a(m)(fg)\mf\\
&=&(fg)(m\ot\mf)\;.
\end{eqnarray*}
\noindent The $\cA$-module structures on the filters $\op{F}_n\cM$, $n\ge 1$, naturally induce an $\cA$-module structure on $\op{F}\cM$.\medskip
(ii) Universal anchor map. Since $\op{F}\cM$ is the free Leibniz algebra over $\cM$, the map $a:\cM\to\op{Der}\cA$ factors through the inclusion $\cM\to \op{F}\cM$:
$$
\begin{CD}
\cM @>{}>> \op{F}\cM \\
@| @V{\op{F}a}VV \\
\cM@>{a}>>\op{Der}\cA
\end{CD}
$$
The Leibniz algebra morphism $\op{F}a$ is actually $\cA$-linear. Indeed, in view of the decomposition $f\mf=\sum_{i\le n-1}\mf_i$, where $\mf_i\in\op{F}^{i}\cM$ is a decomposed tensor, we obtain first $m\ot f\mf=[m,f\mf]$, where the notation is the same as above. Since, by induction, $\op{F}a$ is $\cA$-linear on $\op{F}^{n-1}\cM$, we then have
\begin{eqnarray*}
\op{F}a(f(m\ot \mf))&=&\op{F}a(m\ot f\mf)-\op{F}a(a(m)(f)\mf)\\
&=&\op{F}a[m,f\mf]-a(m)(f)\op{F}a(\mf)\\
&=&[a(m),f \op{F}a(\mf)]-a(m)(f)\op{F}a(\mf)\\
&=&f[a(m),\op{F}a(\mf)]\\
&=&f\op{F}a(m\ot \mf)\;.
\end{eqnarray*}
(iii) Leibniz pseudoalgebra conditions. To see that $(\op{F}\cM,[-,-]_{\op{Lei}},\op{F}a)$ (in the following we omit subscript $\op{Lei}$) is a Leibniz pseudoalgebra, it now suffices to check that (\ref{axiom1}) is satisfied. We proceed by induction and assume that, for any $f\in\cA$, $\mf\in \op{F}_{n-1}\cM$ and $\mf'\in \op{F}\cM$, $n\ge 2$, the bracket $[\mf,f\mf']$ satisfies Condition (\ref{axiom1}). Indeed, for $n=2$, we have $$[m,f\mf']=m\ot f\mf'=f(m\ot \mf')+a(m)(f)\mf'=f[m,\mf']+\op{F}a(m)(f)\mf'\;.$$ It is easily seen that (\ref{axiom1}) is then also satisfied in $\op{F}_n\cM$:
\bea
[m\ot\mf, f\mf']
&=&[[m,\mf],f\mf']\\
&=&[m,[\mf,f\mf']]-[\mf,[m,f\mf']]\\
&=&[m,f[\mf,\mf']]+[m,\op{F}a(\mf)(f)\mf']
-[\mf,f[m,\mf']]-[\mf,a(m)(f)\mf']\\
&=&\cdots\\
&=&f[m,[\mf,\mf']]-f[\mf,[m,\mf']]
+[\op{F}a(m),\op{F}a(\mf)](f)\mf'\\
&=&f[m\ot \mf,\mf']+\op{F}a(m\ot\mf)(f)\mf'.
\eea
(iv) Freeness. It remains to show that the Leibniz $(R,\cA)$-pseudoalgebra $(\op{F}\cM,[-,$ $-],\op{F}a)$ (we omit $\op{Lei}$), together with the anchored $\cA$-module morphism $i:\cM\hookrightarrow \op{F}\cM$, is free. Let thus $(\cE,[-,-]',a')$ be any Leibniz $(R,\cA)$-pseudoalgebra and let $\zf:\cM\to \cE$ be any anchored $\cA$-module morphism. Since $(\op{F}\cM,[-,-],i)$ is the free Leibniz $R$-algebra over the $R$-module $\cM$, the $R$-linear map $\zf$ extends uniquely to a Leibniz $R$-algebra map $\op{F}\zf:\op{F}\cM\to\cE$. When assuming that $a'\op{F}\zf=\op{F}a$ (resp., that $\op{F}\zf$ is $\cA$-linear) on $\op{F}_{n-1}\cM$, the usual proof based on the observation that $m\ot f\mf=[m,f\mf]$ (resp., on this observation combined with (\ref{ModStr}) and (\ref{axiom1})) allows to see that the same property holds on $F_n\cM$.
\end{proof}
\subsection{Free Symmetric Leibniz Pseudoalgebra}
\begin{proposition}\label{Main3} Let $J_1$, $J_2$ be the ideals (\ref{Ideal1}) and (\ref{Ideal2}) associated to the free Leibniz $(R,\cA)$-pseudoalgebra $(\op{F}\cM,[-,-]_{\op{Lei}},\op{F}a)$ over an anchored $\cA$-module $(\cM,a)$. The quotient $\op{FS}\cM:=\op{F}\cM/(J_{1}+J_{2})$, with induced bracket, anchor and `inclusion', is the free symmetric Leibniz pseudoalgebra over the anchored module $\cM$.\end{proposition}
\begin{rem} The free symmetric Leibniz pseudoalgebra over an anchored module $(\cM,a)$ is the natural quotient of the free Leibniz pseudoalgebra over $(\cM,a)$. The latter is the reduced tensor $R$-module $\op{F}\cM=\overline{T}\cM$ over $\cM$, endowed with an $\cA$-module structure that encodes the anchor $a$, the universal Leibniz bracket $[-,-]_{\op{Lei}}$, and the induced anchor $\op{F}a$. In the geometric case, when $\cM=\zG(M)$, with $M\to B$ an anchored vector bundle over a manifold, the module $\overline{T}\cM$ is not a space of sections, since the tensor product in $\overline{T}\cM$ is over $R$.\end{rem}
\begin{proof} We characterize the classes and the mentioned induced data by the symbol `tilde'. It has already been said that $(\op{FS}\cM,[-,-]^{\,\widetilde{}}_{\op{Lei}},(\op{F}a)^{\;\widetilde{}}\;)$ is a symmetric Leibniz pseudoalgebra (as usual we will omit $\op{Lei}$). On the other hand, it is clear from the definition of $(\op{F}a)^{\;\widetilde{}}$ that $\tilde \imath:\cM\ni m\to \tilde m\in\op{FS}\cM$ is an anchored $\cA$-module map.
As for freeness, let $(\cE,[-,-]',a')$ be a symmetric Leibniz pseudoalgebra. Any anchored module map $\zf:\cM\to \cE$ uniquely extends to a Leibniz pseudoalgebra map $\op{F}\zf:\op{F}\cM\to\cE$. To see that $\op{F}\zf$ descends to $\op{FS}\cM$, it suffices to show that it vanishes on $J_1$ and $J_2$. Observe first that, for any $\zm,\zn\in\op{F}\cM$, we have $$\op{F}\zf(\zm\circ\zn)=\op{F}\zf(\zm)\circ'\op{F}\zf(\zn)\;.$$ It follows now from the $\cA$-linearity of $\op{F}\zf$ and the symmetry of $\cE$ that $\op{F}\zf$ annihilates $J_1$ and $J_2$. It is also straightforwardly checked that the induced map $(\op{F}\zf)^{\;\widetilde{}}:\op{FS}\cM\to\cE$ is a map of Leibniz pseudoalgebras such that $(\op{F}\zf)^{\;\widetilde{}}\;\tilde\imath=\zf$. As for uniqueness of this extension, note that any Leibniz pseudoalgebra morphism $\op{FS}\zf:\op{FS}\cM\to\cE$ that extends $\zf$, implements a Leibniz pseudoalgebra morphism $$(\op{FS}\zf)^{\;\bar{}}:\op{F}\cM\ni \zm\mapsto \op{FS}\zf(\tilde{\zm})\in\cE$$ that extends $\zf$; hence, $(\op{FS}\zf)^{\;\bar{}}=\op{F}\zf$ and $\op{FS}\zf=(\op{F}\zf)^{\;\widetilde{}}$.\end{proof}
\section{Free Courant Algebroid}\label{Repr}
There is a forgetful functor $\op{For}:{\tt CrtAld}\to {\tt AncMod}$ between the categories of generalized Courant $(R,\cA)$-pseudoalgebras and anchored $\cA$-modules.
\begin{theo}\label{Main4} The free Courant algebroid over an anchored module $(\cM,a)$ is the generalized Courant pseudoalgebra $${\cal C}(\op{FS}\cM)=(\op{FS}\cM,{\cal R}(\op{FS}\cM),[-,-]^{\;\widetilde{}}_{\op{Lei}},(-|-),(\op{F}a)^{\;\widetilde{}},\tilde{\zm}^\ell,\tilde{\zm}^r)\;,$$ associated to the free symmetric Leibniz pseudoalgebra over $\cM$, together with the anchored module map $\tilde{\imath}:\cM\to \op{FS}\cM$. More precisely, for any generalized
Courant pseudoalgebra $${\cal C}=(\cE_1,\cE_2,[-,-]',(-|-)',a',\zm'^\ell,\zm'^r)$$ and any anchored module map $\zf:\cM\to\cE_1$, there exists a unique morphism of generalized Courant pseudoalgebras $(\zf_1,\zf_2)$ from ${\cal C}(\op{FS}\cM)$ to $\cal C$, such that $\zf_1\,\tilde{\imath}=\zf$.\end{theo}
\newcommand{\tm}{\tilde{\zm}}
\newcommand{\tn}{\tilde{\zn}}
\newcommand{\tta}{\tilde{\zt}}
\begin{proof} Let $\cal C$ and $\zf$ be as in the statement of the theorem. Due to freeness of $\op{FS}\cM$, the anchored module map $\zf$ uniquely factors through $\op{FS}\cM$, thus leading to a unique Leibniz pseudoalgebra morphism $\zf_1:\op{FS}\cM\to \cE_1$ such that $\zf_1\,\tilde{\imath}=\zf$. As for $\zf_2$, note that the $\cA$-linear map $$\zf_2=(-|-)'(\zf_1\odot \zf_1):\op{FS}\cM\odot \op{FS}\cM\to \cE_1\odot \cE_1\to \cE_2$$ descends to the quotient ${\cal R}(\op{FS}\cM)$. Indeed, if $I(\tilde\zm,\tilde\zn,\tilde\zt)\in\op{Inv}$ (in the following we omit the symbol `tilde'), we have $$\zf_2(I(\zm,\zn,\zt))=$$ $$([\zf_1\zm,\zf_1\zn]'|\zf_1\zt)'+(\zf_1\zn|[\zf_1\zm,\zf_1\zt]')'-(\zf_1\zm|[\zf_1\zn,\zf_1\zt]'+[\zf_1\zt,\zf_1\zn]')'=0\;,$$ since $\cal C$ is a generalized Courant pseudoalgebra. The resulting $\cA$-linear map ${\cal R}(\op{FS}\cM)\to \cE_2$ will still be denoted by $\zf_2$. Since \be\label{UnivScProdS}(\zm|\zn)=(\zm\odot\zn)^{\;\widetilde{}}\;,\ee it is clear that the requirement (\ref{RespScProd}), i.e. $$(-|-)'(\zf_1\times\zf_1)=\zf_2(-|-)\;,$$ is satisfied. It now suffices to check that the conditions (\ref{RespActs}), i.e. $$\zm'^\ell(\zf_1\times\zf_2)=\zf_2\,\tilde\zm^\ell\;\;\text{and}\;\;\zm'^r(\zf_2\times\zf_1)=\zf_2\,\tilde\zm^r\;,$$ hold as well. Let us detail the last case. Let $\tm,\tn,\tta\in \op{FS}\cM$ (and omit again the `tilde'). Since $$\tilde\zm^r(\zm)(\zn\odot\zt)^{\;\widetilde{}}=(\zm^r(\zm)(\zn\odot\zt))^{\;\widetilde{}}=(-(\zn\c \zt)\odot\zm)^{\;\widetilde{}}\;,$$ the application of $\zf_2$ leads to $$-(\zf_1\zn\c' \zf_1\zt|\zf_1\zm)'\;.$$ The latter coincides with the value of $\zm'^r(\zf_2\times\zf_1)$ on the arguments $(\zn\odot\zt)^{\;\widetilde{}}$ and $\zm$. Finally, uniqueness of $\zf_1$ was already mentioned, and uniqueness of $\zf_2$ is a consequence of (\ref{RespScProd}) and (\ref{UnivScProdS}).
\end{proof}
\section{Acknowledgements}
Beno\^it Jubin thanks the Luxembourgian National Research Fund for support via AFR grant PDR 2012-1, 3963765. The research of N. Poncin was supported by Grant GeoAlgPhys 2011-2014 awarded by the University of Luxembourg. Kyousuke Uchino is grateful for an invitation to the University of Luxembourg which was the starting point of the present joint work. The authors would also like to thank Yannick Voglaire for useful discussions.
|
\section{Introduction}
Topological solitons and instantons \cite{Manton:2004tk} play a
significant role in diverse areas of physics such as quantum field
theories, string theory, cosmology \cite{Vilenkin:2000} and condensed
matter systems \cite{Volovik:2003}.
For instance, Yang-Mills instantons play an especially important
role in non-perturbative dynamics of quantum gauge-field theories.
Relations between solitons and instantons with different
dimensionalities are important for a unified understanding
of these objects, which may explain or even reveal unexpected
relations between field theories defined in different dimensions.
Yang-Mills instantons in pure Yang-Mills theory in four-dimensional
Euclidean space are particle-like topological solitons in $d=4+1$
dimensional spacetime.
When coupled to Higgs fields in the Higgs phase, they are unstable and
shrink in the bulk.
They can, however, stably live inside host solitons, in which they
transform themselves to other kinds of solitons:
They transform themselves into Skyrmions \cite{Eto:2005cc} when
trapped inside a non-Abelian domain wall \cite{Shifman:2003uh}, or
lumps \cite{Eto:2004rz} when inside a non-Abelian vortex
\cite{Hanany:2003hp,Eto:2005yh} and sine-Gordon (SG) kinks
\cite{Nitta:2013cn} when inside a monopole string, as summarized in
(d), (e), and (f) in Tab.~\ref{table:instantons}.
One such composites, namely a lump inside a non-Abelian vortex,
elegantly explains the coincidence of mass spectra between
field theories in different dimensions
\cite{Dorey:1998yh,Shifman:2004dr}.
If one compactifies the world volume of host solitons, instantons also
can exist, not only as such trapped solitons; they become twisted
closed domain walls, vortex sheets or monopole strings,
when the moduli $S^3$, $S^2$, or $S^1$ of these host solitons are
wound around their compact world volumes, also of the shape
$S^3$, $S^2$, or $S^1$, respectively \cite{Nitta:2013vaa}
[see (d), (e), and (f) in Tab.~\ref{table:instantons}].
\begin{table}[!htb]
\begin{tabular}{|c|c||c|c|c|c|c|c|c|c|} \hline
soliton & $\pi_n$ in & & host solitons & $\pi_n$ of & codim & moduli & w.v. & w.v. & $\pi_n$ on \\
/dim & bulk& & & host & & &shape & soliton & w.v.\\ \hline\hline
lump & $\pi_2(S^2)$ & (a) & $ {\mathbb C}P^1$ domain wall& $\pi_0$ & $1$ & $S^1$ & ${\mathbb R}^1$ or $S^1$ & SG kink & $\pi_1(S^1)$ \\
2+1 dim & & & & & & & & & \\
\hline
Skyrmion & $\pi_3(S^3)$ &(b) & NA domain wall & $\pi_0$ & $1$ & $S^2$ & ${\mathbb R}^2$ or $S^2$ & lump &$\pi_2(S^2)$\\
3+1 dim & & (c) & vortex string & $\pi_1$ & $2$ & $S^1$ & ${\mathbb R}^1$ or $S^1$ & SG kink & $\pi_1(S^1)$\\ \hline
Instanton & $\pi_3(G)$ &(d) &NA domain wall& $\pi_0$ & $1$ & $S^3$ & ${\mathbb R}^3$ or
$S^3$ & Skyrmion & $\pi_3(S^3)$ \\
4+1 dim & & (e) &NA vortex sheet & $\pi_1$ & $2$ & $S^2$ & ${\mathbb R}^2$ or $S^2$ & lump &$\pi_2(S^2)$\\
& & (f) &monopole string & $\pi_2$ & $3$ & $S^1$ & ${\mathbb R}^1$ or $S^1$ & SG kink & $\pi_1(S^1)$\\ \hline
\end{tabular}
\caption{Host solitons of
trapped instantons (Skyrmions).
{\small (a), (d), (e) and (f) were already pointed out in
Ref.~\cite{Nitta:2013vaa}.
The shape of the world-volume can be noncompact, ${\mathbb R}^n$,
or compact, $S^n$, corresponding to trapped and untrapped instantons
(Skyrmions), respectively.
w.v.~stands for ``world volume,'' NA denotes ``non-Abelian'' and $G$
denotes the gauge group. }
\label{table:instantons}}
\end{table}
A similar relation is known in $d=2+1$ dimensions, in which lumps
\cite{Polyakov:1975yp} or baby-Skyrmions
\cite{Piette:1994ug,Weidig:1998ii} become sine-Gordon kinks
\cite{Nitta:2012xq,Kobayashi:2013ju,Jennings:2013aea} when trapped
inside a ${\mathbb C}P^1$ domain wall
\cite{Abraham:1992vb,Arai:2002xa}, corresponding to (a) in
Tab.~\ref{table:instantons}.
This relation was already known earlier in condensed-matter systems
such as Josephson junctions of two superconductors
\cite{Ustinov:1998}, ferromagnets \cite{Chen:1977}, and $^3$He
superfluids \cite{Volovik:2003}.
When one compactifies the world volume of the domain wall to $S^1$, it
becomes an isolated lump or baby-Skyrmion as a closed domain line with
a twisted $U(1)$ modulus \cite{Kobayashi:2013ju}.
Similar twisted closed wall lines as vortices also exist in
condensed-matter systems such as p-wave superconductors
\cite{Garaud:2012}.
In this paper, we further pursue these relations between solitons with
different dimensionalities, by focusing on Skyrmions
\cite{Skyrme:1962vh} in $d=3+1$ dimensions, which are characterized by
the topological charge $\pi_3(S^3) \simeq {\mathbb Z}$, i.e.~the
baryon number.
It was already found that Skyrmions become baby-Skyrmions or lumps
when trapped inside a non-Abelian domain wall
\cite{Nitta:2012wi,Nitta:2012rq,Gudnason:2014nba}, and a spherical
domain wall with twisted $S^2$ moduli is a Skyrmion
\cite{Gudnason:2013qba}, both corresponding to (b) in
Tab.~\ref{table:instantons}
(this relation was generalized to $N$-dimensional Skyrmions becoming
$N-1$ dimensional Skyrmions inside non-Abelian domain walls in
\cite{Nitta:2012rq}).
The last piece which was missing in Tab.~\ref{table:instantons}
corresponds to (c), which we will work out explicitly in this paper.
We consider a potential term motivated by two-component Bose-Einstein
condensates (BECs) of ultracold atoms with repulsive interactions
\cite{Kasamatsu:2005} (see Appendix \ref{app:BEC}),
which are known to admit a variety of topological
solitons such as domain walls, vortices,
Skyrmions (as vortons)
\cite{3D-skyrmions,Nitta:2012hy,Metlitski:2003gj}
and D-brane solitons (vortices ending on a domain wall)
\cite{Kasamatsu:2010aq,Nitta:2012hy}.
We denote this model the BEC Skyrme model.
The model admits a global vortex solution with a $U(1)$ modulus.
We then deform the potential by a perturbation which introduces a
potential for the $U(1)$ modulus, and construct a stable Skyrmion as a
sine-Gordon kink trapped inside the straight vortex.
We achieve this by two approaches: the effective field theory on
solitons \cite{Gudnason:2014gla} which is based on the moduli
approximation \cite{Manton:1981mp} and by direct numerical
computations.
Together with the previous result, we have two kinds of solitons:
a domain wall and a vortex, able to host Skyrmions,
as illustrated in Fig.~\ref{fig:incarnation} (a) and (b).
\begin{figure}[!tp]
\begin{center}
\mbox{
\subfigure[]{\includegraphics[width=0.25\linewidth]{incarnation-wall-skyrmion}}\qquad\qquad
\subfigure[]{\includegraphics[width=0.2\linewidth]{incarnation-vortex-skyrmion}}}
\mbox{
\subfigure[]{\includegraphics[width=0.3\linewidth]{incarnation-compact-wall-skyrmion}}\quad
\subfigure[]{\includegraphics[width=0.3\linewidth]{incarnation-compact-vortex-skyrmion}}
}
\caption{Incarnation of Skyrmions.
(a) A lump or a baby-Skyrmion on a flat domain wall,
(b) A sine-Gordon kink on a straight vortex,
(c) A spherical domain wall with twisted $S^2$ moduli,
(d) A vortex ring with twisted $S^1$ modulus.
\label{fig:incarnation}}
\end{center}
\end{figure}
We then compactify the world-volume of the vortex-line to $S^1$ for
which we consider the BEC potential without any perturbations.
The configuration becomes a vortex ring with a twisted $U(1)$
modulus, that is, a vorton \cite{Davis:1988jq} (except for the time
dependence which is usually required for vortons).
In fact, it is known in the context of BECs that a Skyrmion is nothing
but a vorton \cite{3D-skyrmions,Nitta:2012hy,Metlitski:2003gj}.
Together with the previous results of \cite{Gudnason:2014nba}, we find
two possible incarnations of Skyrmions; one corresponds to a Skyrmion
as a spherical domain wall with the $S^2$ moduli twisted along the
$S^2$ world-volume \cite{Gudnason:2013qba}, while the other
corresponds to a closed vortex string with the $U(1)$ modulus twisted
along the $S^1$ world-volume, as illustrated in
Fig.~\ref{fig:incarnation} (c) and (d).
For a certain choice of perturbed potential which is described above,
a Skyrmion is attached by the same kind of vortices from both of its
sides.
With a different choice of perturbed potential term, we also construct
a (half-)Skyrmion attached by different kinds of vortices from both of
its sides.
In this case, we cannot compactify the world-volume since the vortices
attached from both of its sides are different.
The latter is similar to a confined monopole in the Higgs phase:
In the Higgs phase, magnetic fluxes of a 't Hooft-Polyakov monopole
\cite{'tHooft:1974qc} are squeezed to form vortices, and the monopole
becomes a kink inside a vortex
\cite{Tong:2003pz,Shifman:2004dr,Nitta:2010nd}.
This was the first prime example of composite
Bogomol'nyi-Prasad-Sommerfield (BPS) solitons,
see Refs.~\cite{Tong:2005un,Eto:2006pg,Shifman:2007ce} for a review.
We call our configurations confined Skyrmions.
This paper is organized as follows.
In Sec.~\ref{sec:model}, we present the Skyrme models we consider in
this paper. In particular, we use two different kinds of potential
terms for model 1 and 2, respectively.
Model 1 was already studied in our previous works, except for the
lump-type of Skyrmion residing in the domain wall, while model 2 is
introduced in this paper and is motivated by two-component BECs.
In Sec.~\ref{sec:wall-vortex}, we construct domain wall solutions for
model 1 and global vortex solutions for model 2, which will serve as
host solitons for our baby solitons: baby Skyrmions (or lumps) and
sine-Gordon kinks, respectively.
In Sec.~\ref{sec:effective}, we construct Skyrmions being
baby-Skyrmions (or lumps) and sine-Gordon kinks in the effective
theories of the domain wall and the vortex as host solitons in model 1
and model 2, respectively.
In Sec.~\ref{sec:numerical}, we provide full numerical solutions for
such composite Skyrmions.
We also construct a vortex ring with the twisted $U(1)$ modulus as a
Skyrmion in model 2.
Sec.~\ref{sec:summary} is devoted to a summary and discussion.
In Appendix \ref{app:BEC}, we explain a potential term of
two-component BECs and its relation to our model.
In Appendix \ref{app:half}, we show full numerical solutions of a
half-Skyrmion trapped inside a vortex in model 2.
\newpage
\section{Skyrme-like models}
\label{sec:model}
We consider the $SU(2)$ principal chiral model or the Skyrme model
with higher-derivative terms, in $d=3+1$ dimensions.
With the $SU(2)$ valued field $U(x)\in SU(2)$, the Lagrangian which we
consider is given by
\beq
\mathcal{L} = \frac{f_{\pi}^2}{16} \tr (\del_{\mu}U^{\dagger}
\del^{\mu} U) + \mathcal{L}_4 + \mathcal{L}_6 - V(U),
\eeq
with the Skyrme \cite{Skyrme:1962vh} and sixth-order derivative term,
respectively
\begin{align}
\mathcal{L}_4 &= \frac{\kappa}{32e^2} \tr
\left([U^\dagger \del_{\mu} U, U^\dagger \del_{\nu} U]^2\right),
\label{eq:Skyrme-term}\\
\mathcal{L}_6 &=
\frac{c_6}{36 e^4 f_\pi^2}\left(\epsilon^{\mu\nu\rho\sigma}
\tr\left[U^\dagger\del_\nu U U^\dagger\del_\rho U
U^\dagger\del_\sigma U\right]\right)^2.
\label{eq:6th_order-term}
\end{align}
The symmetry of the Lagrangian with $V=0$ is
$\tilde G = SU(2)_{L} \times SU(2)_{R}$ acting on $U$ as
$U \to U'= g_L U g_R^\dagger$.
This is spontaneously broken to $\tilde H \simeq SU(2)_{L+R}$
acting as
$U \to U'= g_{L+R} U g_{L+R}^\dagger$ so that the target space is
$\tilde G/\tilde H \simeq SU(2)_{\rm L-R}$.
The conventional potential term is
$V = m_{\pi}^2\tr(2{\bf 1}_2 - U - U^\dagger)$,
which breaks the symmetry $\tilde{G}$ to $SU(2)_{\rm L+R}$
\emph{explicitly}.
In this paper, we consider both cases where the higher-derivative
terms are turned off ($\kappa=c_6=0$) and where either the Skyrme term
or the sixth-order term is turned on.
A BPS model was discovered some years back \cite{Adam:2010fg},
which consists of only the sixth-order term as well as appropriate
potentials. This type of model is not in our parameter space here as
it corresponds to $f_\pi\to 0$ (and $\kappa=0$) which is not possible
as we will rescale away $f_\pi$, see below.
It is interesting, however, for phenomenological reasons due to the
possibility of parametrically small (classically) binding energies and
a large extended symmetry (volume-preserving diffeomorphisms). In this
paper, we consider systems which perhaps are closer to
condensed-matter systems and do have a kinetic term and large binding
energies here are thus not an immediate concern.
If we rewrite the Lagrangian in terms of two complex scalar fields
$\phi^{\rm T}=\{\phi_1(x),\phi_2(x)\}$, defined by
\beq
U =
\begin{pmatrix}
\phi_1 & -\phi_2^*\\
\phi_2 & \phi_1^*
\end{pmatrix},
\eeq
subject to the constraint
\beq
\det U = |\phi_1|^2 + |\phi_2|^2 = 1,
\eeq
then the Lagrangian can be rewritten as
\begin{align}
\mathcal{L}
&= \1{2} \partial_{\mu} \phi^\dagger \partial^{\mu}\phi
- \frac{\kappa}{4}
\left[ (\partial_{\mu} \phi^\dagger \partial^{\mu}\phi )^2
-\1{4}(\partial_{\mu} \phi^\dagger \partial_{\nu}\phi +
\partial_{\nu} \phi^\dagger \partial_{\mu}\phi )^2
\right] \non
&\phantom{=\ }
+ \frac{c_6}{144}
\left(\epsilon^{\mu\nu\rho\sigma} \left[
\phi^\dagger \left(\del_\nu\phi\del_j\phi^\dagger +
\sigma^2\del_i\phi^*\del_\rho\phi^{\rm T}\sigma^2\right)\del_\sigma\phi +
{\rm c.c.}\right]\right)^2
- V(\phi,\phi^*) \non
&= \1{2} \partial_{\mu} {\bf n} \cdot \partial^{\mu} {\bf n}
- \frac{\kappa}{4}
\left[ (\partial_{\mu} {\bf n} \cdot \partial^{\mu} {\bf n} )^2
- (\partial_{\mu} {\bf n} \cdot \partial_{\nu} {\bf n} )^2
\right] \non
&\phantom{=\ } +
\frac{c_6}{36}\left(\epsilon^{\mu\nu\rho\sigma}\epsilon^{ABCD}\del_\nu
n_A \del_\rho n_B \del_\sigma n_C n_D\right)^2
- V( {\bf n} ),
\label{eq:masterL_dimless}
\end{align}
where we have rescaled the Lagrangian density such that energy is
measured in units of $f_\pi/(2e)$ and length is measured in units of
$2/(ef_\pi)$ and we have introduced the real four-vector scalar fields
${\bf n}(x)=\{n_A(x)\}=\{n_1(x),n_2(x),n_3(x),n_4(x)\}$ satisfying
${\bf n}^2 =\sum_A n_A^2 = 1$,
defined by
$\phi_1 = n_1 + i n_2$ and $\phi_2 = n_3 + i n_4$
($A,B,C,D=1,2,3,4$).
The target space (the vacuum manifold with $V=0$)
$\mathcal{M}\simeq SU(2)\simeq S^3$ has a nontrivial homotopy group
\beq
\pi_3(\mathcal{M}) = {\mathbb Z},
\eeq
which admits Skyrmions as usual.
The baryon number (Skyrme charge), $B \in \pi_3(S^3)$, is defined as
\begin{align}
B &= -\1{24\pi^2} \int d^3x \;
\epsilon^{ijk} \, \tr \left( U^\dagger \partial_i U
U^\dagger \partial_j U U^\dagger \partial_k U\right) \non
&= \1{24\pi^2} \int d^3x \;
\epsilon^{ijk} \, \tr \left( U^\dagger \partial_i U
\partial_j U^\dagger \partial_k U\right) \non
&= \frac{1}{24 \pi^2} \int d^3x \left[\epsilon^{ijk}
\phi^\dagger \left(\del_i\phi\del_j\phi^\dagger +
\sigma^2\del_i\phi^*\del_j\phi^{\rm T}\sigma^2\right)\del_k\phi +
{\rm c.c.}\right] \non
&= -\frac{1}{12 \pi^2} \int d^3x\; \epsilon^{ABCD} \epsilon^{ijk}
\partial_i n_A \partial_j n_B \partial_k n_C n_D \non
&= -\frac{1}{2\pi^2} \int d^3x\; \epsilon^{ABCD}
\partial_1 n_A \partial_2 n_B \partial_3 n_C n_D.
\end{align}
Instead of the conventional potential term, we consider here potential
terms of the form
\beq
V = V_1 + V_2,
\eeq
where $V_1$ is the dominant potential and it admits a host soliton
such as a domain wall or a vortex while $V_2$ is a subdominant
potential admitting a soliton inside of the host soliton -- a baby
soliton.
We consider two theories: model 1 admitting Skyrmions as
baby-Skyrmions inside a domain wall; and model 2 admitting Skyrmions
as sine-Gordon solitons inside a vortex.
For model 1, we take the potential to be
\beq
V_1 = \frac{1}{2}M^2 (1-n_4^2), \qquad
V_2 = -\frac{1}{2}m_3^2 n_3^{a_3}
\eeq
with $a_3=1$ or $2$.
When $m_3$ is zero, $V_2$ vanishes and the potential admits two
discrete vacua: $n_4 = \pm 1$.
This allows for a domain wall interpolating between the two vacua
\cite{Kudryavtsev:1999zm}
as we show in the next section.
With a nonzero $m_3>0$, there still remain two vacua and a domain wall
interpolating between them as long as $m_3<M$
\cite{Nitta:2012wi,Nitta:2012rq,Gudnason:2014nba}.
For model 2, we take the potential to be
\begin{align}
V_1 &=\frac{1}{8}M^2\left[1-\big(\phi^\dag\sigma^3\phi\big)^2\right]
= \frac{1}{2} M^2 |\phi_1|^2|\phi_2|^2
= \frac{1}{2} M^2 \left(n_1^2 + n_2^2\right)
\left(n_3^2 + n_4^2\right), \non
V_2 &= - \frac{1}{2}m_3^2 n_3^{a_3},
\end{align}
with $a_3=1$ or $2$.
The potential $V_1$ is motivated by two-component BECs
(see Appendix \ref{app:BEC}), and admits
global vortices \cite{Kasamatsu:2005}
and Skyrmions as vortons
\cite{3D-skyrmions,Nitta:2012hy,Metlitski:2003gj}.
\section{Domain walls and vortices as host solitons}
\label{sec:wall-vortex}
In this section, we construct a domain wall and a vortex for models 1
and 2, respectively, with their respective $V_1$ potentials in the
limit $V_2=0$.
We will take into account the effect of $V_2$ in the next
sections.
\subsection{Model 1: the domain wall}
\label{sec:model1}
Model 1 has two discrete vacua and admits a domain wall solution
interpolating between them.
First, we consider only $V_1$ with $m_3=0$ ($V_2=0$).
With the Ansatz $\mathbf{n}=\{0,0,\sin f(x),\cos f(x)\}$
we have
\beq
\mathcal{L} = -\frac{1}{2}(\partial_x f)^2 - \frac{1}{2}M^2\sin^2 f.
\eeq
This Lagrangian density is the sine-Gordon model admitting a
domain-wall solution
\beq
f = 2\tan^{-1}\exp(\pm M x). \label{eq:SG}
\eeq
The domain wall in this type of model was first studied in Ref.~\cite{Kudryavtsev:1999zm}.
The most general solution is given by
\beq
\mathbf{n} = \{b_1\sin f(x),b_2\sin f(x),b_3\sin f(x),\cos f(x)\},
\eeq
which has moduli in the form of a constant three-vector $\mathbf{b}$
with unit length $\mathbf{b}^2=1$.
These are Nambu-Goldstone (NG) modes due to the spontaneously broken
$O(3)$ symmetry, which is broken down to $O(2)$ in the presence of the
domain wall (\ref{eq:SG}).
These $S^2$ moduli of the domain wall were discussed in
Refs.~\cite{Losev:2000mm,Nitta:2012wi,Nitta:2012rq}.
\subsection{Model 2: the vortex}
\label{sec:model2}
Model 2 allows for global vortices.
The vortex of $\phi_1$ traps $\phi_2$ in its core and carries a $U(1)$
modulus being the phase of $\phi_2$.
For constructing the vortex, we use the following Ansatz
\beq
\phi_1 = \sin f(r) e^{i\phi}, \qquad
\phi_2 = \cos f(r),
\eeq
where $r,\phi$ are polar coordinates in the plane. This simplifies the
Lagrangian density to
\beq
-\mathcal{L} =
\frac{1}{2}f_r^2
+\frac{1}{2r^2}\sin^2 f
+\frac{\kappa}{2r^2}\sin^2(f) f_r^2
+\frac{1}{8} M^2 \sin^2(2f),
\eeq
and the equation of motion reads
\beq
f_{rr} + \frac{1}{r} f_r -\frac{1}{2r^2}\sin 2f
+\frac{\kappa}{r^2}\sin^2 f\left(f_{rr} - \frac{1}{r} f_r\right)
+\frac{\kappa}{2r^2}\sin(2f) f_r^2
-\frac{1}{4} M^2 \sin 4f = 0.
\eeq
The boundary conditions for the vortex system are
\beq
f(0) = 0, \qquad
f(\infty) = \frac{\pi}{2}.
\eeq
Numerical solutions are shown in Fig.~\ref{fig:vortex} for
$\kappa=0,1,\ldots,4$.
\begin{figure}[!tp]
\begin{center}
\mbox{\subfigure[]{\includegraphics[width=0.48\linewidth]{bec_vortex1}}\quad
\subfigure[]{\includegraphics[width=0.48\linewidth]{bec_vortex1_condensates}}}\\
\mbox{\subfigure[]{\includegraphics[width=0.48\linewidth]{bec_vortex1_energy}}}
\caption{(a) Vortex profile function $f$ without the Skyrme term
$\kappa=0$ (blue solid curve) and with the Skyrme term
$\kappa=1,\ldots,4$ (dotted curves) for mass $M=3$.
(b) Condensate fields $|\phi_1|=\sin f$ and $|\phi_2|=\cos f$
($\phi_1$ vanishes at the origin).
(c) Corresponding energy densities.
\label{fig:vortex}}
\end{center}
\end{figure}
Notice that when $\kappa$ is turned on, the vortex widens up and the
energy density at the origin drops significantly.
This vortex is a particular solution, while the most general solution
is given by
\beq
\phi_1 = \sin f(r) e^{i\phi}, \qquad
\phi_2 = \cos f(r) e^{i\zeta},
\eeq
where $\mathbf{b}=\{\cos\zeta,\sin\zeta\}$ is a constant $U(1)$
modulus.
\section{Effective theory approach}\label{sec:effective}
In this section we will review the results of
Ref.~\cite{Gudnason:2014gla}. The main idea is to take a soliton solution
and integrate out the soliton to obtain the effective theory for the
moduli inhabiting the soliton in question. Here we will consider only
the leading-order effective Lagrangian. For details and a discussion
of the expansion, see Ref.~\cite{Gudnason:2014gla}. We will next consider
each model in turn.
\subsection{Model 1: Skyrmions as baby Skyrmions inside
a domain wall}\label{sec:eff-wall-skyrmion}
We start with the Lagrangian density \eqref{eq:masterL_dimless} and
integrate over the codimension of the domain wall.
Here we will take just the leading-order effective Lagrangian and
neglect the backreaction of the baby soliton to the domain wall. This
is a good approximation when there is a separation of scales between
the domain wall mass and the baby-soliton's typical scales.
At leading-order the domain wall is flat which leads to a drastic
simplification and the integration over its codimension yields
\begin{align}
-\mathcal{L}^{\rm eff} &=
\left(\frac{a_{2,0}}{M} + \kappa a_{2,2} M\right) \;
(\partial_\alpha\mathbf{b})^2
+\left(\frac{\kappa a_{4,0}}{2M} + c_6 a_{4,2} M\right)
\big(\partial_\alpha\mathbf{b}\times\partial_\beta\mathbf{b}\big)^2
- \frac{a_{2,0} m_3^2}{M} b_3^{a_3},
\label{eq:Lcodim1}
\end{align}
where $M$ is the mass scale of the domain wall, $\alpha,\beta=t,y,z$,
$a_3=1,2$ and we have defined dimensionless constants as follows
\begin{align}
a_{k,\ell} &\equiv
\frac{M}{2} \int dx \; \sin^{k} f \left(\frac{\partial_x f}{M}\right)^{\ell}
= \frac{1}{2}\int d\xi \mathop{\mathrm{sech}}\nolimits^{k+\ell}\xi
= \frac{\sqrt{\pi}\,\Gamma\!\left(\frac{k+\ell}{2}\right)}
{\Gamma\!\left(\frac{1+k+\ell}{2}\right)},
\end{align}
where $\xi=M x$ and the last two equalities have been evaluated using
the flat domain wall \eqref{eq:SG} and the result is given in terms of
the gamma function. Notice that the coefficients depend only on the
sum of $k$ and $\ell$ and so we can evaluate them as
$a_{k,\ell} = a_{k+\ell}$:
\beq
a_{2} = 1, \quad
a_{4} = \frac{2}{3}, \quad
a_{6} = \frac{8}{15}, \quad
\cdots
\eeq
Inserting the coefficients we get
\begin{align}
-\mathcal{L}^{\rm eff} &=
\left(\frac{1}{M} + \frac{2}{3}\kappa M\right) \;
(\partial_\alpha\mathbf{b})^2
+\left(\frac{\kappa}{3M} + \frac{8}{15} c_6 M\right)
\big(\partial_\alpha\mathbf{b}\times\partial_\beta\mathbf{b}\big)^2
- \frac{m_3^2}{M} b_3^{a_3}.
\label{eq:Lcodim1_coeff}
\end{align}
The existence of a baby-Skyrmion living on the domain wall requires a
non-vanishing $\kappa>0$ or $c_6>0$ as well as a non-zero $m_3$. The
size of the baby-Skyrmion can be estimated by a scaling argument
\cite{Derrick:1964ww} to be
\beq
\frac{1}{L} \sim
\sqrt[4]{\frac{15m_3^2}{5\kappa + 8c_6 M^2}}.
\eeq
A second kind of soliton which can inhabit the domain wall is the
lump, which exists when $\kappa=c_6=m_3=0$ and it will possess a size
modulus \cite{Nitta:2012wi}.
The (full 3D) baryon charge is composed by the domain wall charge and
the baby-Skyrmion charge and is given by
\begin{align}
B = \frac{1}{\pi} \int d^3x \; \mathcal{Q} f_x = Q,
\end{align}
where we have used that there is only a single domain wall in this
setup while
\beq
\mathcal{Q} = \frac{1}{8\pi} \epsilon^{ij}
\mathbf{n}\cdot\del_i\mathbf{n}\times\del_j\mathbf{n},
\eeq
is the baby-Skyrmion charge density and $Q$ the baby-Skyrmion number
(charge).
\subsection{Model 2: Skyrmions as kinks on a vortex}
\label{sec:eff-vortex-skyrmion}
The next and final type of soliton we will consider is the vortex
which has codimension two and a single world-volume direction.
We again take the Lagrangian density \eqref{eq:masterL_dimless} and
integrate over the two codimensions of the vortex to obtain
\begin{align}
-\mathcal{L}^{\rm eff} = \left[\frac{a_{2,0,0}}{M^2}
+ \kappa(a_{2,2,0} + a_{2,0,2})
+ 2c_6 a_{2,2,2} M^2\right] (\partial_\alpha \mathbf{b})^2
- \frac{a_{2,0,0} m_3^2}{M^2} b_1^2,
\end{align}
where $M$ is the mass scale of the vortex, $\alpha=t,z$ and the
dimensionless coefficients in the effective Lagrangian density read
\begin{align}
a_{k,\ell,m}
\equiv \pi M^{2-\ell-m} \int dr \; r^{1-\ell} \cos^k f\sin^\ell f (f_r)^m.
\label{eq:vortex_coefficients}
\end{align}
This effective theory possesses sine-Gordon kinks. Hence, the vortex
can bear sine-Gordon kinks in terms of its twisted $S^1$ modulus and
each of these kinks correspond to Skyrmions in the full 3-dimensional
theory.
Unfortunately, the vortex is not analytically integrable and hence we
need to evaluate the coefficients numerically. As we have defined the
coefficients in a dimensionless manner, they do not depend on the
value of the vortex mass scale, $M$, but they do depend on the value
of the fourth-order derivative term, $\kappa$ (or rather the
combination $\kappa M^2$). We give a set of numerically evaluated
coefficients for the effective Lagrangian density in
Tab.~\ref{tab:a_vortex_numerical}.
\begin{table}[!btp]
\begin{center}
\caption{Coefficients for the effective Lagrangian density for
sine-Gordon kinks living on a straight vortex for various values of
$\kappa M^2$. }
\label{tab:a_vortex_numerical}
\begin{tabular}{c||ccccc}
$\kappa M^2$ & $0$ & $1$ & $2$ & $3$ & $4$\\
\hline\hline
$a_{2,0,0}$ & 0.5106 & 0.7224 & 0.8678 & 0.9866 & 1.090\\
$a_{2,2,0}$ & 0.1616 & 0.1550 & 0.1519 & 0.1499 & 0.1484\\
$a_{2,0,2}$ & 0.1745 & 0.1816 & 0.1852 & 0.1877 & 0.1896\\
$a_{2,2,2}$ & 0.06072 & 0.04172 & 0.03438 & 0.03007 & 0.02712
\end{tabular}
\end{center}
\end{table}
Using again a scaling argument \cite{Derrick:1964ww}, we can estimate
the size of the sine-Gordon kink
\beq
\frac{1}{L} \sim \sqrt{\frac{a_{2,0,0}m_3^2}
{a_{2,0,0} + \kappa(a_{2,2,0} + a_{2,0,2}) M^2 + 2c_6 a_{2,2,2} M^4}},
\label{eq:kinksizeestimate}
\eeq
where the coefficients $a$ are functions of $\kappa M^2$, as shown in
Tab.~\ref{tab:a_vortex_numerical}. As examples, we can calculate the
kink sizes, see Tab.~\ref{tab:kink_sizes}.
\begin{table}[!tbp]
\begin{center}
\caption{Rough size estimate of sine-Gordon kinks using
Eq.~\eqref{eq:kinksizeestimate} as function of the parameters of the
effective theory. }
\label{tab:kink_sizes}
\begin{tabular}{c||rrr}
\raisebox{-2pt}{$\kappa M^2$}${\bf \backslash}$\raisebox{2pt}{$c_6 M^4$}
& $0$ & $1$ & $81$\\
\hline\hline
$0$ & $m_3$ & $0.90m_3$ & $0.22m_3$ \\
$1$ & $0.83m_3$ & $0.80m_3$ & $0.30m_3$\\
$9$ & $0.57m_3$ & $0.57m_3$ & $0.44m_3$
\end{tabular}
\end{center}
\end{table}
Thus from these rough estimates, we learn that the sixth-order
derivative term induces coefficients in the effective Lagrangian for
the kink which increases its size (for fixed masses). The fourth-order
derivative term also leads to an increase in the kink size, but to a
lesser extent.
The (full 3D) baryon charge is composed by the vortex charge and the
kink charges and is given by
\begin{align}
B = \frac{1}{16\pi^2} \int d^3x \; \frac{1}{r}\sin(f) f_r \zeta_z
= Q [\zeta]^{z=z_2}_{z=z_1} = Q P,
\end{align}
where $Q$ is the winding number of the vortex and $P$ is the number
kinks on the string.
\section{Numerical solutions}\label{sec:numerical}
In this section, we provide explicit numerical solutions.
Solutions of the baby-Skyrmion type in model 1 were already obtained
in Ref.~\cite{Gudnason:2014nba}.
Here we will add a new lump solution living on the domain wall, which
also carries baryon charge.
Our other new findings are confined Skyrmions residing on the vortex
string in model 2.
For both cases, we need the deformation $V_2$ of the potentials
for flat host solitons, while we do not need it for solutions
possessing an $S^n$ world-volume.
On the contrary, we need no higher-derivative terms for flat host
solitons, while we do need them for the solutions possessing an $S^n$
world-volume for the stability. An exception to the rule is the
baby-Skyrmion living on the domain wall, which needs both a
higher-derivative term as well as the potential $V_2$ (but the lump on
the domain wall needs neither of these).
For both models we use the relaxation method on a cubic square-lattice
of size $81^3$ (lattice points). We fix the boundary conditions
corresponding to the host solitons as described in
Sec.~\ref{sec:wall-vortex} and choose appropriate initial conditions
for the baby-solitons in question. Then we relax the initial guess
until the solution to the equations of motion is obtained with the
required precision. A cross check of the solutions is the calculation
of the topological baryon charge. We will now take the two models in
turn.
\subsection{Model 1: Skyrmions trapped inside a domain wall}
\label{sec:wall-skyrmion}
We begin with the (exact) domain wall solution of
Sec.~\ref{sec:model1} and add two types of baby-solitons. The first
example is the baby-Skyrmion, which was obtained in
\cite{Gudnason:2014nba} and we will only review it here for
completeness. This solution needs both the potential $V_2$ as well as
a higher-order derivative term. In Fig.~\ref{fig:DW_SL} we show the
baby-Skyrmion with $V_2$ setting $a_3=1$ and $\kappa=1,c_6=0$ (thus
only the Skyrme term is active), which is taken from
\cite{Gudnason:2014nba}.
\begin{figure}[!pt]
\begin{center}
\mbox{\subfigure[\ isosurfaces]{
\includegraphics[width=0.3\linewidth]{129skyrme_dw_sklump}}
\subfigure[\ energy density]{
\includegraphics[width=0.3\linewidth]{129skyrme_dw_sklump_energyslice}}
\subfigure[\ baryon charge density]{
\includegraphics[width=0.3\linewidth]{129skyrme_dw_sklump_chargeslice}
}}
\caption{The baby-Skyrmion living on the domain wall:
(a) 3D view of isosurfaces for the domain wall on which a
baby-Skyrmion resides; the magenta surfaces represent the energy
isosurfaces at a third of the maximum of the energy and
the blue surface at the center shows the baryon charge isosurface,
at half its maximum value. (b) and (c) show respectively the energy
density and baryon charge density at a $yz$-slice in the middle of
the domain wall (at $x=0$). The calculation is done on a $129^3$
cubic lattice, $B^{\rm numerical}=0.998$ and the potential used is
$V_2$ with $a_3=1$ and $M=4,m_3=2$.
This figure is taken from Ref.~\cite{Gudnason:2014nba}. }
\label{fig:DW_SL}
\end{center}
\end{figure}
The next solution, which is new, is the domain wall with a lump
solution inside. This solution also carries a full unit of
3-dimensional baryon (Skyrme) charge. It is obtained for $V_2=0$ and
no higher-derivative terms, i.e.~$\kappa=0,c_6=0$. The numerical
solution is shown in Fig.~\ref{fig:DW_LUMP}. Notice that the lump has
a size modulus and can thus take on any size. We also do not capture
the full baryon charge because we resolve only the center of the lump
with the finite lattice points. This is not a problem of the solution
but of the lattice size. A larger lattice will capture more of the
baryon charge (or alternatively a small lump on the same lattice with
a worse resolution).
\begin{figure}[!hpt]
\begin{center}
\mbox{\subfigure[\ isosurfaces]{
\includegraphics[width=0.3\linewidth]{walllump}}
\subfigure[\ energy density]{
\includegraphics[width=0.3\linewidth]{walllump_energyslice}}
\subfigure[\ baryon charge density]{
\includegraphics[width=0.3\linewidth]{walllump_baryonslice}}}
\caption{The domain wall with a trapped Skyrmion
in the theory without higher-derivative and potential terms.
(a) 3D view of isosurfaces for the energy density and baryon charge
density as in Fig.~\ref{fig:DW_SL}. (b) and (c) show respectively
the energy density and baryon charge density at a $yz$-slice in the
middle of the domain wall (at $x=0$). The calculation is done on an
$81^3$ cubic lattice, with $M=4$ and $B^{\rm numerical}=0.877$. }
\label{fig:DW_LUMP}
\end{center}
\end{figure}
\subsection{Model 2: Skyrmions confined by vortices}
\label{sec:vortex-skyrmion}
In this section we take the vortex solution of Sec.~\ref{sec:model2}
and add a sine-Gordon kink on its world-volume.
As already mentioned, for the straight vortex we need a finite
potential $V_2$ for the baby-soliton. If we choose a linear potential,
i.e.~$a_3=1$, the kink on the vortex corresponds to a full unit of
baryon charge, while for the quadratic potential, i.e.~$a_3=2$, each
kink corresponds to half a unit of baryon charge.
The straight vortex possesses sine-Gordon kinks if the potential $V_2$
is turned on even without higher-derivative terms.
The presence of the Skyrme term (the fourth-order derivative term)
widens both the vortex itself and the kink living on the vortex (see
Sec.~\ref{sec:eff-vortex-skyrmion}, whereas the sixth-order derivative
term does not alter the vortex solution, but it does widen the kink --
even more than the Skyrme term, see Tab.~\ref{tab:kink_sizes}.
First we will present the vortex with a full sine-Gordon kink living
on its world-volume in the case of no higher-derivative terms, see
Fig.~\ref{fig:vortexkink_ko}. This kink was made with kink mass
$m_3=0.22$, vortex mass $M=3$ and the kink length was measured as
\beq
L_{\rm kink} = \sqrt{\frac{\int d^3x \; z^2
\mathcal{E}_{\rm kink}}{\int d^3x \; \mathcal{E}_{\rm kink}}}
\sim 3.35,
\label{eq:kinklengthmeasure}
\eeq
which one can compare with the analytic formula
$\frac{\pi}{2\sqrt{3}m_3}\simeq 4.14$. The reason for the smaller
value in the numerical result is due to the binding energy present on
both sides of the center of the kink, see
Fig.~\ref{fig:vortexkink_ko}d.
\begin{figure}[!tbh]
\begin{center}
\mbox{\subfigure[\ isosurfaces]{
\includegraphics[width=0.4\linewidth]{vortexkink81_ko}}
\subfigure[\ energy density]{
\includegraphics[width=0.4\linewidth]{vortexkink81_ko_energyslice2}}}
\mbox{\subfigure[\ baryon charge density]{
\includegraphics[width=0.4\linewidth]{vortexkink81_ko_baryonslice}}
\subfigure[\ kink energy density]{
\includegraphics[width=0.4\linewidth]{vortexkink81_ko_kinkenergyslice}}}
\caption{The vortex with a trapped Skyrmion which is manifested as a
sine-Gordon kink on its world-volume, in the theory with no
higher-derivative terms.
(a) 3D view of isosurfaces for the energy density and baryon charge
density. (b) and (c) show respectively
the energy density and baryon charge density at a $yz$-slice through
the vortex (at $x=0$). (d) shows the energy of the kink (which is
the total energy with the vortex energy subtracted off). Notice the
negative dips on each side of the peak in the kink energy; we
interpret those as binding energy.
The calculation is done on an $81^3$ cubic lattice, with $M=3$,
$m_3=0.22$ and the baryon charge is $B^{\rm numerical}=0.991$. }
\label{fig:vortexkink_ko}
\end{center}
\end{figure}
Now we will consider a different example, namely the full sine-Gordon
kink in the vortex theory with the sixth-order derivative term turned
on ($c_6=1$) while the Skyrme term is still off ($\kappa=0$).
As already mentioned, it does not alter the vortex solution far away
from the kink, but it does increase the length of the kink living on
its world-volume. In Fig.~\ref{fig:vortexkink} we present this
solution and we have taken the kink mass to be $m_3=1$ and the vortex
mass $M=3$. The kink length was measured with
Eq.~\eqref{eq:kinklengthmeasure} to be $L_{\rm kink}\sim 2.44$.
The effective theory predicted this kink to be almost five times
longer than that without the sixth-order derivative term; whereas
numerically it is only about $2.7$ times longer (according to this
measure); but recall that the scaling argument is just a rough
estimate neglecting the actual integrals (or rather assuming them to
be of order one).
We chose the kink mass in the latter vortex solution (in
Fig.~\ref{fig:vortexkink}) to be $m_3=0.22$ such that the kink should
have approximately the same length as that with a sixth-order
derivative term with $m_3=1$. Measuring this ratio numerically we get
$\sim 0.73$ which on the one hand determines the accuracy of the
estimate; but also confirms that the effective theory is qualitatively
correct.
\begin{figure}[ht]
\begin{center}
\mbox{\subfigure[\ isosurfaces]{
\includegraphics[width=0.4\linewidth]{vortexkink81}}
\subfigure[\ energy density]{
\includegraphics[width=0.4\linewidth]{vortexkink81_energyslice2}}}
\mbox{\subfigure[\ baryon charge density]{
\includegraphics[width=0.4\linewidth]{vortexkink81_baryonslice}}
\subfigure[\ kink energy density]{
\includegraphics[width=0.4\linewidth]{vortexkink81_kinkenergyslice}}}
\caption{The vortex with a trapped Skyrmion which is manifested as a
sine-Gordon kink on its world-volume, in the theory with a
sixth-order derivative term.
(a) 3D view of isosurfaces for the energy density and baryon charge
density. (b) and (c) show respectively
the energy density and baryon charge density at a $yz$-slice through
the vortex (at $x=0$). (d) shows the energy of the kink (which is
the total energy with the vortex energy subtracted off). Notice that
instead of binding energy, the higher-derivative term induces
sub-peaks on the side of the kink.
The calculation is done on an $81^3$ cubic lattice, with $M=3$,
$m_3=1$ and the baryon charge is $B^{\rm numerical}=0.990$.}
\label{fig:vortexkink}
\end{center}
\end{figure}
The last example we will consider, is the vortex compactified on a
circle, $S^1$, without a potential $V_2$, which has a free theory
living on its world-volume.
Existence of this solution without angular momentum, requires
a higher-derivative term; here we will use only the sixth-order
derivative term.
In Fig.~\ref{fig:torus} is shown a numerical solution with vortex mass
$M=4$ and $c_6=1$.
Notice that the energy density is torus-like to some extent (it has a
valley almost halfway down in energy density) but the baryon charge
density remains as a ball-like object with a dip in the energy density
at the origin.
\begin{figure}[!hpt]
\begin{center}
\mbox{\subfigure[\ isosurfaces]{
\includegraphics[width=0.3\linewidth]{toruse6}}
\subfigure[\ energy density]{
\includegraphics[width=0.3\linewidth]{toruse6_energyslice}}
\subfigure[\ baryon charge density]{
\includegraphics[width=0.3\linewidth]{toruse6_baryonslice}}}
\caption{The vortex compactified on a circle in the theory with a
sixth-order derivative term and without the potential $V_2$.
(a) 3D view of isosurfaces for the energy density and baryon charge
density as in Fig.~\ref{fig:DW_SL}. (b) and (c) show respectively
the energy density and baryon charge density at a $yz$-slice (at
$x=0$). The calculation is done on an $81^3$ cubic lattice, with
$M=4$ and $B^{\rm numerical}=0.99990$. }
\label{fig:torus}
\end{center}
\end{figure}
\section{Summary and Discussion}
\label{sec:summary}
In this paper we have exhausted the possibilities (known so far) of
Skyrmions in different disguises. By trapping a Skyrmion on a domain
wall, it hosts a baby-Skyrmion while the full system has a
3-dimensional Skyrme (baryon) charge. If the domain wall is
compactified it is again a normal Skyrmion, but having its energy
distributed as in a ball-like object. Using the parameter space of the
model, it is possible to obtain a spherical shell-like object -- by
for instance having very large sixth-order derivative term, see
\cite{Gudnason:2013qba}.
In this paper, we find the new and last piece of the puzzle, i.e.~the
Skyrmion trapped on a vortex string, which looks like a sine-Gordon
kink on the vortex world-sheet. We find the existence of this object
by an effective theory approach \cite{Gudnason:2014gla} and by
explicit numerical calculations. The last object we find here is the
vortex compactified on a circle, which thus carries a Skyrme charge
by having a twist on its modulus. Kinks can furthermore live on
this torus-like object, but we leave such studies for future
developments.
Let us comment on the accuracy of the comparison between the lengths
predicted from the effective theory and the actual numerical
calculation we carried out. As emphasized both here and in
\cite{Gudnason:2014gla}, the effective theory relies heavily on the
separation of scales when taking only the leading-order contribution
into account. Higher-order corrections have not yet been calculated
explicitly, although it is straightforward. However, the numerical
solutions are all done on a finite square-lattice which makes a too
large separation of scale inconvenient, i.e.~memory and run-time
consuming, which is why we have only an order 3--4 between the mass
scales in the systems studied.
We have constructed a single sine-Gordon kink residing in a vortex in
this paper, however, it is also possible to make a sine-Gordon kink
crystal, which is described by the elliptic function ${\rm
sn}(x)$ \footnote{See for instance Ref.~\cite{Canfora:2014aia} for
sine-Gordon kink crystals in the context of the Skyrme model on
$S^2\times S^1$. }.
In Tab.~\ref{table:instantons}, we have summarized the
topological incarnations of
lumps (baby-Skyrmions, or sigma-model instantons),
Skyrmions, and Yang-Mills instantons.
There seems to be certain relations among
the homotopy in the bulk (resultant solitons),
the homotopy of host solitons,
and the homotopy of world-volume solitons,
but an exact mathematical correspondence is yet to be clarified.
We can do the same for Hopfions (knot solitons)
\cite{Faddeev:1996zj};
Hopfions can be realized as sine-Gordon kinks
on a toroidal domain wall \cite{Kobayashi:2013bqa}.
The BEC Skyrme model which we consider in this paper
also admits D-brane solitons \cite{Gauntlett:2000de}, that is,
vortices ending on a domain wall,
since the corresponding BECs admit them
\cite{Kasamatsu:2010aq,Nitta:2012hy}.
Therefore, this model admits various solitons with
various codimensions: domain walls, vortices and Skyrmions,
and their composites.
The dynamics of these solitons remain as an interesting problem to
explore.
For instance, Skyrmions were proposed to be created after the
annihilation of a brane and anti-brane
\cite{Nitta:2012hy,Nitta:2012kj}.
\section*{Acknowledgements}
The work of M.N. is supported in part by
a Grant-in-Aid for Scientific Research (No. 25400268)
and by the ``Topological Quantum Phenomena''
Grant-in-Aid for Scientific Research
on Innovative Areas (No. 25103720)
from the Ministry of Education, Culture, Sports, Science and
Technology (MEXT) of Japan.
SBG thanks Keio University (Yokohama, Japan) and Institute of Modern
Physics (Lanzhou, China) for hospitality.
|
\section{Introduction}
We begin with an apologia, since there is nothing still not known about the FP model [1]. However, there seems to be no published derivation\footnote{This was actually done [2] for the more general case of the system embedded in dS, rather than flat, space; however that derivation involved (an even number of) steps using inverse powers of $\Lambda$, hence singular and not applicable here; its final result of course does limit to ours. It was also shown in detail here that the helicity $0$ mode can be removed by suitably tuning $m^2/\Lambda$ in dS.} from its covariant and highly constrained form to the final unconstrained
canonical action in terms of its $2s+1=5$ helicity $(\pm 2,\pm 1,0)$ components. That form will displays that each mode propagates correctly and has manifestly positive energy, in the usual $L=p \, \dot q - H$, $H= \frac{1}{2}\, [p^2+ q\, (-\nabla^2+m^2)\, q]$ form. While [1] realized that positive energy was essential, it was displayed rather opaquely; a subsequent formulation [3] was likewise less than transparently presented (and contains distracting typos). lndeed, proper use of the constraints is not altogether trivial, making the correct process instructive as a (minor) exercise in (free) field theory.
Historically, it was not until FP's 1939 work that the problem of representing massive spins $>1$ by tensor fields -- involving many more than $2s+1$ components -- was raised, let alone solved. Given the current interest in massive gravity (mGR) with Einstein kinetic terms plus non-derivative mass terms involving a fixed, say flat, background, our summary may be useful. lndeed, this is a good place to note that-contrary to statements in the mGR literature -- the mass terms destroy the whole (ADM) asymptotic energy formulation of GR as a $2D$ surface integral at spatial infinity, just as the Coulomb asymptotic integral that counts total charge is lost in massive (Proca) vector theory: the Newtonian/Coulomb fields decay much too fast there for these integrals to contribute at all.
\section{The Derivation}
The action and field equations of the theory are the sum of linearized GR and the -unique-mass term that eliminates the 6th, ghost helicity $0$, degree of freedom (DoF). We work throughout in 1st order, $3+1$, canonical form, which simplifies the procedure and indeed starts directly in terms of the 6 conjugate pairs $(\pi^{ij}, h_{ij})$ rather than the 10 covariant $h_{\mu\nu}$. The action is (see e.g., [2])
\begin{equation}
\begin{aligned}
I &= \int d^4x\, \left\{ \pi^{ij} \dot h_{ij} - H(\pi,h)\right\}, \\
H &= \ ^3R_Q + \of{\pi_{ij}^2 - \frac{1}{2}\, \pi^2} + 4 \, n\, R_0 + 2\, N_i \, \dd_j \pi^{ij} + \frac{1}{4} \, m^2 \, (h_{ij}^2 - h_{ii}^2 - 4 \, n\, h_{ii} - 2 N_i^2),
\end{aligned}
\end{equation}
where (under the integral)
\begin{equation}
\ ^3 R_Q = \frac{1}{2} \, h_{ij} \, G^L_{ij} = -\frac{1}{4}\, [h_{\TT} \, \nabla^2 h_{\TT} - h_\T \, \nabla^2 h_\T], \, \, \, \, \, \, \, \, \, \, R_0 = (m^2 - \nabla^2) \, h_\T + m^2\, h_\LL;
\end{equation}
$n \sim \frac{1}{2} \, h_{00}$ is a Lagrange multiplier enforcing the linear constraint $R_0 =0$ , while $N_i=h_{0i}$ becomes an auxiliary field to be eliminated by completing squares, leaving only the six $(\pi,h)$ pairs -- indeed our whole process consists of juggling quadratic forms. Finally, we recall that the linearized 3D Einstein tensor $G^L_{ij} (h_{lm})$ is both identically conserved and independent of the longitudinal, gauge, parts of $h_{lm}$; The second ingredient, essential to the separation of the various helicity DoF in (1), is the usual orthogonal decomposition of any symmetric 3-tensor,
\begin{equation}
S_{ij} = S_{ij}^{\TT} + \frac{1}{2}\, (\delta_{ij} - \hat \dd_i \hat \dd_j) S^\T + [ \hat \dd_j S^\T_i + \hat \dd_j S^\T_{i}]
+ \hat \dd_i \hat \dd_j S^\LL, \, \, \, \, \, \dd_i \, S^\T_i \equiv 0, \, \, \, \, \, \hat \dd_i \equiv \dd_i/\sqrt{\nabla^2}.
\end{equation}
Completing squares in (1) removes the $N_i$ dependence of $H$ in favor of adding the term $2m^{-2} (\dd_j \pi^{ij})^2$ to $H$. There remains the elimination of the $R_0$ constraint, hence of one linear combination of the two helicity $0$ ($T$, $L$) modes. lt will be equally essential to use $\dot R_0 =0$ to further
eliminate one combination of their conjugate momenta $(\pi^\T,\pi^\LL)$ using $ \dot h \sim \pi$; constraints ``strike twice" in our 1st order form, since they are valid for all times\footnote{
That the original number, $6$, of $\pi^{ij} \dot h_{ij}$ kinetic terms decreases by one for every constraint is just Darboux's theorem on quadratic forms; in massless theory there are $4$ constraints, leaving just the two $(\pi^{\TT}, q^{\TT})$ pairs. We will see this more explicitly below.}.
The task before us then is to decompose (1) into a sum of three -- non- interacting -- orthogonal, two DoF sectors: Helicity $\pm 2$ (TT), helicity $\pm 1$ ``$T_i$", and the $(T,L)$ helicity-$0$. The latter's Hamiltonian is the source of difficulty, being a priori non-positive before using the $R_0(T,L)=0$ constraint. To keep the discussion compact, we first dispose of the helicity $>0$ sectors: that of TT is trivial to obtain, being unconstrained; we simply add up the TT terms in (1); dropping ``TT", we have
\begin{equation}
L = \pi^{ij} \, \dot h_{ij} - H \, \, \, \, \, \, \, \, \, \, H = \pi_{ij}^2 + \frac{1}{4} \, \of{ h_{ij,k}^2 + m^2 \, h_{ij}^2} \ge 0;
\end{equation}
Note that $H$ vanishes only for TT vacuum, $\pi=0=h$. The same is true of the transverse vector ($T_i$) part, though it still requires some field redefinitions to achieve the same final form; here (1) also easily yields (omitting ``$T_i$")
\begin{equation}
H = \pi^2 + 2 \, m^{-2} \, \of{\pi^{ij}_{\, \, \, ,j}}^2 + \frac{1}{2} \, m^2 \, h^2 \ge 0;
\end{equation}
while not (yet) very pretty, this H is also positive and vanishes at ($T_i$) vacuum, a result unaffected by the further field redefinitions required to reach the final $p\, \dot q - H$ form; we outline the process in the Appendix. [Recall, however, that correct energy functional form is only reached when the ``$(p,q)$" variables are redefined to ensure that the associated kinetic, ``$p\, \dot q$", term is itself free of unwanted numerical coefficients.]
We now face the final, $H(T,L)$, sector, where $R_0$ must be used -- twice. There,
\begin{equation}
\begin{aligned}
I[T,L] &= \int d^4 x \left[\frac{1}{2} \, \pi^\T \, \dot h_\T + \pi^\LL \, \dot h_\LL - V(h_\T, h_\LL) - K(\pi^\T, \pi^\LL)\right], \\
4 \, V(h) &\equiv h_\T \, \of{\nabla^2 - m^2} \, h_\T - 2 \, h_\T \, h_\LL , \, \, \, \, \, \, \, \, \, \, K(\pi) \equiv \frac{1}{2}\, \of{\pi^{\LL}}^2 - 2 \, \pi^\T \, \pi^\LL + 2 \, \pi^\LL \, \of{-m^{-2}\, \nabla^2} \, \pi^\LL.
\end{aligned}
\end{equation}
We now show that both potential and kinetic parts of $H$ are positive, using the $R_0$ constraint and its time derivative respectively.
Eliminating $h_\LL$ yields
\begin{equation}
V(h_\LL,h_\T) = \frac{3}{8} \, \left[ \of{h_{\T , i}}^2 + m^2 \, h_\T^2\right] \ge 0;
\end{equation}
again, $V$ only vanishes at vacuum, $h_\T=0$.
Next we find the $\dot R_0 = 0$ constraint between $\pi^\T$ and $\pi^\LL$: The two field equations for $\pi \sim \dot h$ obtained by varying (6) w.r.t. the $\pi$ are
\begin{equation}
\dot h_\LL - \pi^\LL + \pi^\T + 4 \, m^{-2} \, \nabla^2 \pi^\LL = 0, \, \, \, \, \, \, \, \, \, \, \dot h_\T + 2 \, \pi^\LL =0.
\end{equation}
Taking their appropriate vanishing linear combination, we learn that
\begin{equation}
m^2 \, \pi^T = \of{-2 \, \nabla^2 - m^2}\, \pi^\LL;
\end{equation}
hence finally
\begin{equation}
K(\pi^\LL,\pi^\T) \rightarrow K(\pi^\LL) =\frac{3}{2} \, \of{\pi^L}^2 \ge 0.
\end{equation}
We have now established $E\ge 0$ for the full theory, but one task is still to be completed: putting the helicity action into exact
$p\, \dot q- H(p,q)$ form. Even before this is done, one can already see that the (second order) field equations are uniformly $(\Box-m^2)\, h=0$, but it is an amusing exercise -- as well as a check on the result -- to do so. Using (9), it is easy to translate the $(T,L)$ sector's $\pi^\LL\, \dot h_\LL +\frac{1}{2} \, \pi^\T\, \dot h_\T$ into $\pi^\LL\, \dot h_\T$ form. At this penultimate point,
\begin{equation}
L (T,L) = -\frac{3}{2} \, \pi^\LL \, \dot h_\T - \frac{3}{2} \, [ \of{\pi^\LL}^2 + \frac{1}{4}\, \of{h_{T,i,i}}^2 + \frac{1}{4} \, m^2 \, h_\T^2 ];
\end{equation}
the obvious rescaling $(\pi,h) \rightarrow \sqrt{2/3} \, (-\pi,h)$ achieves the desired final canonical form of the helicity $0$ sector,
\begin{equation}
L(0) = \pi \, \dot h - H(\pi,h), \, \, \, \, \, \, \, \, \, \, H \equiv \pi^2 + \frac{1}{4} \, h \, \of{-\nabla^2 + m^2} \, h.
\end{equation}
Together with the vector mode in the Appendix, then, the total action is
\begin{equation}
L = \sum_{A=1}^5 p^A \, \dot q_A - \frac{1}{2}\, [ (p^A)^2 + q_A (-\nabla^2 + m^2) \, q_A ],
\end{equation}
after the (cosmetic) rescaling $\pi \rightarrow p^A/\sqrt{2}$, $h\rightarrow q_A \, \sqrt{2}$.
\section{Summary}
The physical correctness of the massive $s=2$ FP model has been displayed: each of its $2s+1=5$ helicity excitations obey $(\Box-m^2)\, h=0$, $E \ge 0$.
\section{Appendix: The helicity $\pm 1$ sector}
We consider here the remaining, helicity $1$, subspace involving only the ``$T_i$" parts of $(\pi,h)$ in (1). Clearly, neither $h\, G_\LL (h)$ nor the $R_0$ constraint involve $h_{\T\, i}$; only the mass term does: it contains $\frac{1}{2} \, m^2 \, (h_{\T \, i})^2$. Its $\pi$ sector involves $(\pi^\T_{i,j})^2$ as well as the quadratic term $2\, N^\T_i \dd_j \, \pi^{ij}$. Using the $\frac{1}{2}\, m^2 (N^\T_i)^2$ from the mass term, we complete the square to leave a net contribution $m^{-2}(\pi^{ij}_{\, \, \, ,j})^2 \sim 2 \, m^{-2} \, (\pi^\T_i \, \nabla^2 \pi^\T_i)$ there. At this point, then, dropping the $T_i$ indices, we find
\begin{equation}
L(T_i)= -2 \, \pi \, \dot h - \frac{1}{2} \, [m^2 h^2 + 4 \, m^{-2} \, \pi \, \of{m^2 - \nabla^2} \pi] ;
\end{equation}
the necessary redefinition is obvious:
\begin{equation}
\pi \rightarrow -\frac{1}{2} \, m \of{m^2 -\nabla^2}^{-1/2} \, \pi \, \, \, \, \, \, \, \, \, \, h \rightarrow m^{-1} \, \sqrt{m^2 - \nabla^2} \, h
\end{equation}
leads to the desired helicity $1$ canonical Lagrangian,
\begin{equation}
L(T_i) \rightarrow \, \pi \, \dot h - \frac{1}{2} \, [ \pi^2 + h \, \of{m^2 - \nabla^2} \, h ].
\end{equation}
\subsection*{Acknowledgements}
This work was supported in part by Grants NSF PHY- 1266107 and DOE \# DE-SC0011632. Collaboration with A. Waldron on [2] facilitated the present effort, as did composing help from J. Franklin and CPU help from G. Conrad.
|
\section{Introduction}
Electric dipole polarizability $\alpha$ of atoms, and ions is an important
property to quantify the response to an external electromagnetic field
\cite{bonin-97}. It is essential to have accurate values of $\alpha$ for atoms,
and ions in numerous state of the art experiments to probe
fundamental physics, and develop new technologies. An important example is
the accurate predictions of black-body radiation shift \cite{safronova-12a}
in optical atomic clocks \cite{diddams-04}, which has been realized with
optical lattice \cite{takamoto-05}, trapped ions \cite{diddams-01} and
ultracold atoms \cite{wilpers-02}. In theoretical atomic structure, and
properties calculations $\alpha$ serves as an excellent proxy to assess the
accuracy of theoretical many-body calculations. In the present work, the
studies on the $\alpha$ of Hg serves as an appraisal of the many-body
effects important for accurate structure, and properties calculations. This
is a prerequisite to study the permanent electric dipole moment of Hg
\cite{griffith-09} as a signature of parity- and time-reversal violations, and
probe physics beyond the standard model of particle physics. Given the
importance of $\alpha$, it has been studied using a variety of many-body
methods, and are discussed in a recent review by Mitroy and collaborators
\cite{mitroy-10}. Another reference we have found extremely valuable for our
studies on the $\alpha$ of neutral atoms is the Schwerdtfeger's updated
Table of $\alpha$ \cite{note1}, which originally appeared in the chapter
by the same author in the collected volume by
Maroulis \cite{schwerdtfeger-06}. The table provides an exhaustive list of
references on experimental, and theoretical results of $\alpha$ for the
electronic ground states of neutral elements.
In the present work we study the $\alpha$ of Zn, Cd and Hg using the
perturbed relativistic coupled-cluster (PRCC) theory. It is built
upon the coupled-cluster theory (CCT), first developed to address the nuclear
many-body problem \cite{coester-58,coester-60}, and later applied to studies
on atom and molecules \cite{cizek-69} . The CCT, and relativistic
version, relativistic coupled-cluster (RCC), are now extensively
used in atomic \cite{pal-07, mani-09,nataraj-11}, molecular \cite{isaev-04},
nuclear \cite{hagen-14}, and condensed matter physics \cite{li-14} many-body
calculations. In the PRCC theory, we add a second set of coupled-cluster
amplitudes to account for an additional interaction Hamiltonian. The method
is general, and can be adapted with ease to incorporate different forms
of interaction Hamiltonians. The detailed descriptions of the theory is
provided in a series of our previous works \cite{chattopadhyay-12a,
chattopadhyay-12b, chattopadhyay-13a, chattopadhyay-13b,chattopadhyay-14a}.
Besides the description of PRCC theory, through these works we had
explored the impact of Breit interaction \cite{chattopadhyay-12b},
improved diagrammatic evaluations \cite{chattopadhyay-13a},
vacuum polarization \cite{chattopadhyay-13b}, and triple excitation cluster
operators \cite{chattopadhyay-14a} in the unperturbed cluster operators.
A related method used for calculating electric dipole polarizabilities is to
consider the $z$-component of the dipole operator and define a set of
perturbed cluster operators \cite{sahoo-08,yashpal-13}. In the present work,
we report the inclusion of the dominant perturbative triples in
the PRCC theory, and improved validation of including Breit interaction in
the generation of orbital basis set and PRCC theory. Our earlier works, related
to Breit interaction, reported matching the Dirac-Coulomb-Breit SCF energies
with previous results. This, however, provides an assessment of the
implementation at a coarse grained level. A better comparison would be the
orbital energy corrections from the Breit interaction. This is what we
demonstrate for Hg, as we could get the data from a previous
work \cite{lindroth-89a}. This, we feel, is an important validation of our
implementation of Breit interaction.
The important feature of the present work is, it extends, and verify the
applicability of PRCC theory in the computation of $\alpha$ to the transition
elements. As expected, we get very good results, and we have gained
significant insight on the nature of the correlation effects with $d$
sub-shell as the immediate shell below the valence.
The remaining part of the paper consists of five sections. In next section,
Section II, we provide a brief discussion on the RCC theory. The description
of the linearized RCC and PRCC equations, along with the angular factors,
form the principal parts of subsections in this section. It must be
emphasized that the linearized RCC equations include the triple excitation
cluster amplitudes, with the representation we introduced in our previous
work \cite{chattopadhyay-14a}. The Section III provides a brief description
of how to compute $\alpha$ with PRCC, and is followed with an exposition on
the computational details in Section IV. The results and discussions are
given in Section V. We provide detailed analysis of our theoretical results,
and discuss, vis-a-vis previous results, relevant trends and prospects for
possible future improvements. We, then, end the main part of the paper with
conclusions. In the appendix, we have listed the angular factors of the
linearized RCCSDT and PRCC. With these, we feel, interested readers would
be able to implement these theories at the linear level without difficulty.
For the details on the nonlinear terms, the readers may refer our previous
work \cite{chattopadhyay-12b}. The results and equations presented in
this work are in atomic units ( $\hbar=m_e=e=1/4\pi\epsilon_0=1$). In
this system of units the velocity of light is $\alpha ^{-1}$, the inverse
of fine structure constant. For which we use the value
of $\alpha ^{-1} = 137.035\;999\;074$ \cite{mohr-12}.
\section{Relativistic Coupled-cluster theory}
The Dirac-Coulomb-Breit Hamiltonian $H^{\rm DCB}$ provides a good description
of neutral atom, and well suited for structure and properties calculations.
For an $N$-electron atom
\begin{eqnarray}
H^{\rm DCB} & = & \Lambda_{++}\sum_{i=1}^N \left [c\bm{\alpha}_i \cdot
\mathbf{p}_i + (\beta_i -1)c^2 - V_{N}(r_i) \right ]
\nonumber \\
& & + \sum_{i<j}\left [ \frac{1}{r_{ij}} + g^{\rm B}(r_{ij}) \right ]
\Lambda_{++},
\label{ham_dcb}
\end{eqnarray}
where $\bm{\alpha}$ and $\beta$ are the Dirac matrices, $\Lambda_{++}$ is an
operator which projects to the positive energy solutions and $V_{N}(r_{i})$ is
the nuclear potential. Sandwiching the Hamiltonian with $\Lambda_{++}$ ensures
that the effects of the negative energy continuum states are neglected in the
calculations. Another approach, which is better suited for numerical
computations, is to use the kinetically balanced finite basis sets
\cite{stanton-84,mohanty-90,grant-06,grant-10}. We use this method in the
present work to generate the orbital basis sets.
The last two terms, $1/r_{ij} $ and $g^{\rm B}(r_{ij})$ are the
Coulomb and Breit interactions, respectively. The later, Breit interaction,
represents the inter-electron magnetic interactions and is given by
\begin{equation}
g^{\rm B}(r_{12})= -\frac{1}{2r_{12}} \left [ \bm{\alpha}_1\cdot\bm{\alpha}_2
+ \frac{(\bm{\alpha_1}\cdot \mathbf{r}_{12})
(\bm{\alpha_2}\cdot\mathbf{r}_{12})}{r_{12}^2}\right].
\end{equation}
The Hamiltonian satisfies the eigen-value equation
\begin{equation}
H^{\rm DCB}|\Psi_{i}\rangle = E_{i}|\Psi_{i}\rangle ,
\end{equation}
where, $|\Psi_{i}\rangle$ is the exact atomic state and $E_i$ is the energy
of the atomic state. In the presence of external electromagnetic fields, the
Hamiltonian is modified with the addition of interaction terms. For external
static electric field, the interaction is
$H_{\rm int}=-\mathbf{d}\cdot\mathbf{E}_{\rm ext} $,
where $\mathbf{d}$ and $\mathbf{E}_{\rm ext}$ are the induced electric dipole
moment of the atom and external electric field, respectively. In the remaining
part of this section we give a brief description of RCC theory, which we use
to compute atomic state $|\Psi\rangle $ and PRCC to account for the
effects of $H_{\rm int}$ in the atomic state.
\subsection{Overview of RCC and PRCC theories}
In RCC theory we define the ground state atomic wavefunction of a
closed-shell atom as
\begin{equation}
|\Psi_0\rangle = e^{ T^{(0)}}|\Phi_0\rangle ,
\end{equation}
where $|\Phi_0\rangle$ is the reference state wave-function and $T^{(0)}$ is
the unperturbed cluster operator. The wave-function is modified when the atom
is subjected to an external static electric field $\mathbf{E}$, and the
interaction Hamiltonian is $ H_{\rm int} = -\mathbf{D}\cdot\mathbf{E}$,
where $\mathbf{D}$ is the induced electric dipole moment of the atom.
In the present work we define the perturbed ground state as
\begin{equation}
|\tilde{\Psi}_0\rangle = e^{T^{(0)} + \lambda \mathbf{T}^{(1)}\cdot\mathbf{E}}
|\Phi_0\rangle = e^{T^{(0)}}\left [ 1 + \lambda \mathbf{T^{(1)}\cdot
\mathbf{E}} \right ] |\Phi_0\rangle , \;\;\;\;
\label{psi_tilde}
\end{equation}
where $\mathbf{T}^{(1)}$ are the PRCC operators
\cite{chattopadhyay-12a,chattopadhyay-12b}. For an $N$-electron
closed-shell atom $T^{(0)} = \sum_{i=1}^N T_{i}^{(0)}$ and
${\mathbf T}^{(1)} = \sum_{i=1}^N {\mathbf T}_{i}^{(1)}$, where $i$ is the
order of excitation. In the coupled-cluster single and double (CCSD)
excitation approximation \cite{purvis-82},
\begin{subequations}
\begin{eqnarray}
T^{(0)} &=& T_{1}^{(0)} + T_{2}^{(0)}, \\
{\mathbf T}^{(1)} &=& {\mathbf T}_{1}^{(1)} + {\mathbf T}_{2}^{(1)}.
\end{eqnarray}
\end{subequations}
The CCSD is a good starting point for structure and properties calculations
of closed-shell atoms and ions. In the second quantized representation
\begin{subequations}
\begin{eqnarray}
T_1^{(0)} &= &\sum_{a,p} t_a^p {{a}_p^\dagger} a_a ,
\label{t1_def} \\
T_2^{(0)} &= &\frac{1}{4}\sum_{a,b,p,q}
t_{ab}^{pq} {{a}_p^\dagger}{{a}_q^\dagger}a_b a_a ,
\label{t2_def} \\
\mathbf{T}_1^{(1)} & = & \sum_{a,p} \tau_a^p \mathbf{C}_1 (\hat{r})
a_{p}^{\dagger}a_{a},
\label{pt1_def} \\
\mathbf{T}_2^{(1)} & = & \frac{1}{4}\sum_{a,b,p,q} \sum_{l,k}
\tau_{ab}^{pq}(l,k) \{ \mathbf{C}_l(\hat{r}_1)
\mathbf{C}_k(\hat{r}_2)\}^{1}
a_{p}^{\dagger}a_{q}^{\dagger}a_{b}a_{a}, \;\;\;\;\;\;\;\;
\label{pt2_def}
\end{eqnarray}
\end{subequations}
where $t_{\ldots}^{\ldots}$ and $\tau_{\ldots}^{\ldots}$ are the cluster
amplitudes, $a_i^{\dagger}$ ($a_i$) are single particle creation (annihilation)
operators and $abc\ldots$ ($pqr\ldots$) represent core (virtual) single
particle states or orbitals. To represent $\mathbf{T}_1^{(1)}$, a rank one
operator, we have used the $\mathbf{C}$-tensor of similar rank
$\mathbf{C}_1(\hat r)$. Coming to $\mathbf{T}_2^{(1)}$, to represent it two
$\mathbf{C}$-tensor operators of rank $l$ and $k$ are coupled to a rank one
tensor operator. In addition, the PRCC clusters are constrained by
other selection rules arising from parity and triangular conditions,
these are described in our previous work \cite{chattopadhyay-12b}.
With the inclusion of $T_3^{(0)}$ the RCC theory incorporates all the
correlation effects up to second order in the residual Coulomb interaction.
That is, the theory encapsulates all the many-body perturbation theory (MBPT)
diagrams \cite{lindgren-86} which are first and second order in the
residual Coulomb interaction. In addition, as it is coupled cluster theory,
it incorporates the connected single, double and triple excitations
to all order. The leading order contribution to the uncertainty in the
calculations arise from the quadruple excitations, which, in MBPT, first
appear at the third order of perturbation.
\subsection{Linearized CCSDT cluster equations}
A simplified approximation which incorporates most of the important the
many-body effects is the linearized RCCSDT. In this approximation we
only consider terms which are zeroth and first order in the cluster
operators. The importance of the linearized cluster equations is that, to
solve the RCCSDT equations iteratively, we take the solutions
as the initial values. The $T_1^{(0)}$, $T_2^{(0)}$ and
$T_3^{(0)}$ cluster equations, as described in our previous work, are then
\begin{widetext}
\begin{eqnarray}
\label{lccsdt_s}
& & \sum_{bq} g^{bp}_{aq}t_b^q + \frac{1}{2}\sum_{bcq}g^{bc}_{qa}
(t_{bc}^{qp} - t_{bc}^{pq})
+ \sum_{bqr}g^{bp}_{qr}(t_{ba}^{qr} - t_{ab}^{qr})
+ \frac{1}{2}\sum_{bcqr} (g^{bc}_{qr} - g^{bc}_{rq})t_{abc}^{pqr}
+ \left (\varepsilon_p - \varepsilon_a \right ) t_a^p= 0, \\
& & \sum_{r}g^{pq}_{ar}t_b^r - \sum_{c}g^{pc}_{ab}t_c^q
+ \sum_{cd}g^{cd}_{ab}t_{cd}^{pq} + \sum_{rs}g^{pq}_{rs}t_{ab}^{rs}
- \sum_{cr} \bigg [ g^{cp}_{ar} t_{cb}^{rq}
+ g^{pc}_{rb}t_{ac}^{rq} + \frac{1}{2} g^{pc}_{ar}(t_{cb}^{rq}
- t_{bc}^{rq}) \bigg ] + \sum_{rcs}(g^{rs}_{cq}
- g^{sr}_{cq}) t_{acb}^{prs}
\nonumber \\
\label{lccsdt_d}
& & + \frac{1}{2} \sum_{rcd}(g^{rb}_{cd}
- g^{rb}_{dc})t_{acd}^{prq} + \left ( \begin{array}{c}
p \leftrightarrow q \\
a \leftrightarrow c
\end{array}\right )
+ \left (\varepsilon_p + \varepsilon_q
- \varepsilon_a - \varepsilon_b\right )t_{ab}^{pq} + g_{ab}^{pq}= 0, \\
& & \sum_{s} g^{qr}_{sc}t_{ab}^{ps} + \sum_{d} g^{dr}_{bc}t_{ad}^{pq}
+ \sum_{ds} \bigg [ g^{as}_{pd}\left ( t_{dbc}^{sqr}
+ t_{bdc}^{sqr} \right ) + g^{sb}_{pd}t_{adc}^{sdr}
+ g^{as}_{dp}t_{dbc}^{sqr} \bigg ]
+ \sum_{st} g^{st}_{pq}t_{abc}^{str} + \sum_{de} g^{ab}_{de}t_{dec}^{pqr}
+ \left ( \begin{array}{c}
p \leftrightarrow q \leftrightarrow r\\
a \leftrightarrow b \leftrightarrow c
\end{array}\right )
\nonumber \\
& & + \left (\varepsilon_p + \varepsilon_q + \varepsilon_r- \varepsilon_a
- \varepsilon_b - \varepsilon_c\right ) t_{abc}^{pqr}= 0.
\label{lccsdt_t}
\end{eqnarray}
\end{widetext}
where, $\varepsilon_i$ is the orbital energy of the $i$th orbital,
$i\leftrightarrow j$ represents permutation of the two indexes and
$g_{ij}^{kl} = \langle kl|1/r_{12} + g^{B}(r_{12})|ij\rangle$ is the matrix
element of the two-electron interaction Hamiltonian. For the
cluster amplitudes $t_{abc}^{pqr}$, we use the representation introduced in
our previous work \cite{chattopadhyay-14a}. The representation is symmetric
with respect to the interchange of orbital indexes and reduces the number of
terms in the equations. So, in the cluster equations, only classes of
contractions based on the number of hole (particle) are considered or terms
with unique topology of the Goldstone diagrams are considered in the
equations. Another equivalent representation of $t_{abc}^{pqr}$ with a
different multipole structure is given in the work of Derevianko and
collaborators \cite{derevianko-08}.
The Eqs. (\ref{lccsdt_s}-\ref{lccsdt_t}) are in terms of the matrix elements
of the two-electron interactions. Another representation which is suitable
for atomic or ionic systems, and consistent with the expressions in
properties calculations is to write the equations in terms of reduced matrix
elements. For this consider the matrix element of the
electron-electron Coulomb interaction, following the standard multipole
decomposition \cite{lindgren-86,johnson-07,grant-10}
\begin{eqnarray}
\langle pq|\frac{1}{r_{12}}|ab\rangle & = &\sum_k \sum_q
\left ( \begin{array}{ccc}
j_p & k & j_a \\
-m_p & q & m_a \\
\end{array} \right )
\left ( \begin{array}{ccc}
j_q & k & j_b \\
-m_q & -q & m_b \\
\end{array} \right ) \nonumber \\
&& \times X^k_{\rm C}(pqab),
\end{eqnarray}
where, we have followed the notations in Ref. \cite{grant-10}. In the
above expression $X^k_{\rm C}(pqab)$ is the reduced matrix element or the part
of the matrix element which is independent of the magnetic quantum numbers. It
is defined as
\begin{eqnarray}
X^k_{\rm C}(pqab) &= & \{j_p,j_a,k\}\{j_q,j_b,k\}
\Pi^e(\kappa_p \kappa_a k) \Pi^e(\kappa_q \kappa_b k) \;\;\;\;\;
\nonumber \\
& & (-1)^k \langle j_p||\mathbf{C}^k||j_a\rangle
\langle j_q||\mathbf{C}^k||j_b\rangle R^k_{\rm C}(pqab),
\end{eqnarray}
where, $\{j_i,j_j,k\} $ is the triangular condition,
$\Pi^e(\kappa_i \kappa_j k)$ is the parity condition that $l_i + l_j + k$ must
be even, $\mathbf{C}^k$ is a c-tensor and $ R^k_{\rm C}(pqab)$ is the radial
part of the matrix element. The matrix elements of the Breit iteration,
$g^{\rm B}(r_{12}) $, may also be written in a similar form. For this
let $X^k_{\rm B}(pqab)$ represents the reduced matrix element of
$g^{\rm B}(r_{12}) $ and as a compact notation define
\begin{equation}
g_{ab,k}^{pq} = X^k_{\rm C}(pqab) + X^k_{\rm B}(pqab),
\end{equation}
as the two-electron reduced matrix element corresponding to the multipole $k$.
Based on this definition, the cluster amplitudes must also be defined in
terms of the multipole structure, and we use the notation $t_{ab,k}^{pq}$ to
represent the component of $t_{ab}^{pq}$ with multipole $k$. The cluster
amplitude equations Eqs. (\ref{lccsdt_s}-\ref{lccsdt_t}) are then in terms of
reduced matrix elements. To examine the multipole structure of $T_2^{(0)}$,
consider the approximation based on the first order in many-body perturbation
theory (MBPT). The cluster amplitude is then
\begin{equation}
t_{ab,k}^{pq} \approx \frac{g_{ab,k}^{pq}}{(\varepsilon_p + \varepsilon_q
- \varepsilon_a - \varepsilon_b)}.
\end{equation}
It must be mentioned here that this is also the expression we use as the
initial guess to solve the cluster equations iteratively using a method like
Jacobi. In this case, the multipoles $k$ of the cluster amplitudes are
identical to the two-electron interactions. The triple cluster amplitudes
$t_{abc}^{pqr} $, however, involves three multipoles and details related to
the multipole representation are discussed in our previous
work \cite{chattopadhyay-14a}. A similar description on the cluster equations
with the CCSD approximation, in terms of reduced matrix elements, is presented
in Ref. \cite{blundell-89}. The reference also provides detailed expressions
of the angular factors corresponding to each term in the cluster equation.
Adopting the notations defined here and representations of $t_{abc}^{pqr} $
discussed in our previous work \cite{chattopadhyay-14a}, we use
$t_{abc,l_1l_2l_3}^{pqr}$ to represent the cluster amplitudes
of $T_3^{(0)}$ in terms of reduced matrix elements. Where, $l_i$s are the
multipoles in the representation of the $T_3^{(0)}$ cluster
amplitudes \cite{chattopadhyay-14a}. The cluster equations are then
\begin{widetext}
\begin{eqnarray}
& & \sum_{bq} \left ( A_1 g_{aq,0}^{pb}
- A_2 g_{aq,0}^{bp}
\right ) t_b^q
+ \sum_{bcqk_1}g^{bc}_{qa,k_1} \left (
A_3t_{bc,k_1}^{qp}
- \sum_{k_2}
A_4t_{bc,k_2}^{pq} \right )
+ \sum_{bqrk_1}g_{qr,k_1}^{bp}\left (
A_5t_{ba,k_1}^{qr}
- \sum_{k_2}
A_6t_{ab,k_2}^{qr} \right )
\nonumber \\
& &+ \frac{1}{2}\sum_{bcqrl_1} \left (
A_7g_{bc,l_1}^{qr}
-
A_8g^{rq}_{bc,l_1} \right ) t_{abc,0l_1l_1}^{pqr}
+ \left (\tilde{\varepsilon}_p
- \tilde{\varepsilon}_a \right ) t_a^p= 0,
\label{lccsdt_s} \\
& & \sum_{r}
B_1g_{ar,k}^{pq}t_b^r
- \sum_{c}
B_2g_{ab,k}^{pc}t_c^q
+ \sum_{cdk_1k_2}
B_3g_{ab,k_1}^{cd}t_{cd,k_2}^{pq}
+ \sum_{rsk_1k_2}
B_4g_{rs,k_1}^{pq}t_{ab,k_2}^{rs}
- \sum_{cr} \left (
B_5g_{ar,k_1}^{cp} t_{cb,k}^{rq}
+ \sum_{k_1k_2}
B_6 g_{rb,k_1}^{pc}t_{ac,k_2}^{rq} \right .
\nonumber \\
& &+ \left .
B_7g_{ar,k}^{pc}t_{cb,k}^{rq}
-\sum_{k_1}
B_8g_{ar,k}^{pc}t_{bc,k_1}^{rq} \right )
+ \sum_{rcsl_1l_2}\left (
B_9g_{rs,l_1}^{cq}
- \sum_{k_1}
B_{10}g_{sr,k_1}^{cq}\right )
t_{acb,kl_1l_2}^{prs}
+ \frac{1}{2} \sum_{rcdl_1l_2}\left (
B_{11} g_{rb,l_1}^{cd} \right .
\nonumber \\
& & - \left . \sum_{k_1}
B_{12}g_{rb,k_1}^{dc} \right )
t_{acd,kl_1l_2}^{prq}
+ \left ( \begin{array}{c}
p \leftrightarrow q \\
a \leftrightarrow c
\end{array}\right )
+ \left (\tilde{\varepsilon}_p
+ \tilde{\varepsilon}_q
- \tilde{\varepsilon}_a
- \tilde{\varepsilon}_b\right )t_{ab,k}^{pq}
+ g_{ab,k}^{pq}= 0,
\label{lccsdt_d} \\
& & \sum_{s} C_1
g_{sc,l_3}^{qr}t_{ab,l_1}^{ps}
+ \sum_{d} C_2
g_{bc,l_3}^{dr}t_{ad,l_1}^{pq}
+ \sum_{ds} \bigg [ g_{as,l_1}^{pd}\left ( C_3
t_{dbc,l_1l_2l_3}^{sqr} \right .
+ \left . C_4
t_{bdc,m_1m_2l_3}^{sqr} \right )
+ \sum_{m_1m_2k}C_5
g_{sb,k}^{pd}t_{adc,m_1m_2l_3}^{sdr}
\nonumber \\
& & + \sum_{k}C_6
g_{as,k}^{dp}t_{dbc,l_1l_2l_3}^{sqr} \bigg ]
+ \sum_{st}\sum_{m_1m_2k} C_7
g_{st,k}^{pq}t_{abc,m_1m_2l_3}^{str}
+ \sum_{de}\sum_{m_1m_2k} C_8
g_{ab,k}^{de}t_{dec,m_1m_2l_3}^{pqr}
+ \left ( \begin{array}{c}
p \leftrightarrow q \leftrightarrow r\\
a \leftrightarrow b \leftrightarrow c
\end{array}\right )
\nonumber \\
& & + \left (\tilde{\varepsilon}_p
+ \tilde{\varepsilon}_q
+ \tilde{\varepsilon}_r
- \tilde{\varepsilon}_a
- \tilde{\varepsilon}_b
- \tilde{\varepsilon}_c\right ) t_{abc}^{pqr}= 0,
\label{lccsdt_t}
\end{eqnarray}
\end{widetext}
where $A_i$, $B_i$ and $C_i$ are the angular factors given in Appendix
\ref{app_a}-\ref{app_c}, and
$\tilde{\varepsilon}_i=\epsilon/\sqrt{[j_i]}$ with $[j_i] = 2j_i + 1$. These
are the cluster amplitude equations we solve in the LCCSDT theory.
\subsection{Linearized PRCC equations}
The details of the Goldstone diagrams and the corresponding algebraic
expressions for the PRCC theory with CCSD approximation are discussed in one
of our previous works \cite{chattopadhyay-12b}. In a subsequent work
\cite{chattopadhyay-13b} we also described the linearized PRCC (LPRCC)
equations obtained from the approximation
$ \left [\bar{H}^{\rm DC}_{\rm N},\mathbf{T}^{(1)}\right ]
\approx \left [H^{\rm DC}_{\rm N},\mathbf{T}^{(1)}\right ]$ and
$\bar{H}_{\rm int} \approx \mathbf{D} + \left[\mathbf{D},T^{(0)}\right ]$,
where $\bar{H}_{\rm int} = \exp (-T^{(0)})H_{\rm int}\exp(T^{(0)})$.
The eigenvalue equation in the PRCC theory is then
\begin{equation}
\left [H_{\rm N}^{\rm DCB},\mathbf{T}^{(1)}\right ] |\Phi_0\rangle
= \bigg ( \mathbf{D} + \left [\mathbf{D},T^{(0)}\right ]
\bigg )|\Phi_0\rangle. \;\;\;
\label{prcc_eq2}
\end{equation}
The cluster operator equations are as given in ref. \cite{chattopadhyay-13b}.
However, in terms of the cluster amplitudes, the equation for the
$\mathbf{T}^{(1)}_1$ cluster amplitudes is
\begin{eqnarray}
& & \mathbf{d}^{p}_{a} + \sum_q \mathbf{d}^{p}_{q} t^q_a
- \sum_b \mathbf{d}^{b}_{a} t^p_b + \sum_{bq}\Big ( \mathbf{d}^{b}_{q}
\tilde t^{qp}_{ba} + \tilde {g}^{bp}_{qa} \bm{\tau}_b^q \Big )
\nonumber \\
& & + \sum_{bqr}{\tilde g}^{bp}_{qr} \bm{\tau}_{ba}^{qr}
- \sum_{bcq} g_{qa}^{bc} \tilde{\bm{\tau}}_{bc}^{qp}
+ \left ( \varepsilon _p - \varepsilon _a \right ) \bm{\tau}_a^p = 0,
\label{psing_eqn}
\end{eqnarray}
where $\mathbf{d}^{j}_{i} = \langle j|\mathbf{d}|i\rangle$ is the matrix
element of the dipole operator,
$\tilde{g}_{ij}^{kl} = g_{ij}^{kl} - g_{ji}^{kl}\equiv g_{ij}^{kl}-g_{ij}^{lk}$
is the antisymmetrized matrix element of the two-body interaction and
similarly, $\tilde{\tau} $ is the antisymmetrized perturbed cluster
amplitudes. The Goldstone diagrams arising from the terms in the
equation are given in Fig. \ref{psing_diag}. Similarly, the LPRCC equation for
the $\mathbf{T}^{(1)}_2$ cluster amplitudes is
\begin{eqnarray}
& & \bigg [ \sum_r \Big ( \mathbf{d}_r^pt_{ab}^{rq}
+ g_{rb}^{pq}\bm{\tau}_a^r \Big ) - \sum_c \Big ( \mathbf{d}^{c}_{a}
t_{cb}^{pq} + g_{ab}^{cq} \bm{\tau}_c^p \Big )
+ \sum_{rc} \Big ( g_{ar}^{pc} \tilde{\bm{\tau}}_{cb}^{rq}
\nonumber \\
& & - g_{rb}^{pc} \bm{\tau}_{ac}^{rq} - g_{ar}^{cp} \bm{\tau}_{cb}^{rq}
\Big ) \bigg ] + \left [ \begin{array}{c}
p\leftrightarrow q \\
a\leftrightarrow b
\end{array} \right ]
+ \sum_{rs} g_{rs}^{pq} \bm{\tau}_{ab}^{rs}
+ \sum_{cd} g_{ab}^{cd} \bm{\tau}_{cd}^{pq}
\nonumber \\
& & + \left ( \varepsilon_p
+ \varepsilon_q - \varepsilon_a - \varepsilon_b\right )
\bm{\tau}_{ab}^{pq} = 0,
\label{pdbl_eqn}
\end{eqnarray}
where $\bigl( \begin{smallmatrix}p\leftrightarrow q \\ a\leftrightarrow b
\end{smallmatrix} \bigr )$ represents terms similar to those in
$[\cdots] $ but with the combined permutations $p\leftrightarrow q$ and
$a\leftrightarrow b$. The Goldstone diagrams arising from the terms
in the above equation are shown in Fig. \ref{pdbl_diag}. However, as
discussed earlier in the case of LCCSDT, it is more appropriate to write the
cluster amplitude equations in terms of the reduced matrix elements.
For this we define the cluster amplitude of $\mathbf{T}_1^{(1)}$ as
$\bm{\tau}_{a,1}^b$, where the bold face is to indicate that the cluster
amplitude correspond to a rank one operator and subscript `1' is to indicate
the rank of the operator.
As mentioned earlier, the PRCC theory is general and
applicable to perturbations with operators of any rank in the electron
sector. So, for other forms of perturbations, the index `1' may be replaced
with the appropriate rank.
This definition effectively subsumes the reduced
matrix element of the $c$-tensor in the definition of $\mathbf{T}_1^{(1)}$
given in Eq. (\ref{pt1_def}). Similarly, cluster amplitude of
$\mathbf{T}_2^{(1)}$ is
defined as $\bm{\tau}_{ab,l_1l_2}^{pq}$, where $l_1$ and $l_2$ are the
ranks of the $c$-tensor operators coupled to a rank one operator. With this
definition reduced matrix elements of the $c$-tensor part of the
representation in Eq. (\ref{pt2_def}) is incorporated to the definition of
$\bm{\tau}_{ab,l_1l_2}^{pq}$. Following similar procedure as in LCCSDT, the
linearized PRCC equations of the cluster amplitudes $\bm{\tau}_{a,1}^p$ and
$\bm{\tau}_{ab,l_1l_2}^{pq}$ in terms of reduced matrix elements are
\begin{widetext}
\begin{eqnarray}
& & \mathbf{d}_{a,1}^{p}
+ \sum_q {\cal A}_1\mathbf{d}_{q,1}^{p}t^q_a
- \sum_b {\cal A}_2\mathbf{d}_{a,1}^{b}t^p_b
+ \sum_{bq}\mathbf{d}_{q,1}^{b}\left (
{\cal A}_3 t_{ba,1}^{qp}
- \sum_{k} {\cal A}_4 t_{ab,k}^{qp} \right )
+ \sum_{bq} \bm{\tau}_{b,1}^q\left (
{\cal A}_5 g_{aq,1}^{pb}
- \sum_{k} {\cal A}_6 g_{aq,k}^{bp} \right)
\nonumber \\
& & + \sum_{bqr}\sum_{m_1 m_2}\bm{\tau}_{ba,m_1m_2}^{qr}
\left ( {\cal A}_7 g_{rq,m_2}^{pb}
- \sum_{k} {\cal A}_8 g_{qr,k}^{pb} \right )
- \sum_{bcq} \sum_{m_1 m_2} \left (
{\cal A}_9 g_{aq,m_2}^{cb} \bm{\tau}_{cb,m_1m_2}^{pq}
- \sum_{k}{\cal A}_{10} g_{aq,k}^{cb}
\bm{\tau}_{bc,m_1m_2}^{pq}
\right )
\nonumber \\
& & + \left ( \tilde{\varepsilon} _p
- \tilde{\varepsilon} _a \right ) \bm{\tau}_a^p = 0,
\label{lprcc_s}\\
& & \Bigg (
\sum_{r} {\cal B}_1 \mathbf{d}^{p}_{r} t^{ab,l_2}_{rq}
- \sum_{c} {\cal B}_2 \mathbf{d}^{c}_{a} t_{cb,l_2}^{pq}
+ \sum_{r} {\cal B}_3 g_{rb,l_2}^{pq} \bm{\tau}^r_a
- \sum_{c} {\cal B}_4 g_{ab,l_2}^{cq}\bm{\tau}^p_c
+ \sum_{rc} {\cal B}_5 g_{ar,l_1}^{pc}
\bm{\tau}_{cb,l_1l_2}^{rq}
- \sum_{rc} \sum_{m_1m_2} {\cal B}_6 g_{ar,l_1}^{pc}
\bm{\tau}_{bc,m_1m_2}^{rq}
\nonumber \\
& & - \sum_{rc} \sum_{km_1m_2} {\cal B}_7 g_{rb,k}^{pc}
\bm{\tau}_{ac,m_1m_2}^{rq}
- \sum_{rck} {\cal B}_8 g_{ar,k}^{cp}
\bm{\tau}_{cb,l_1l_2}^{rq}
\Bigg )
+ \Bigg ( \begin{array}{c}
p\leftrightarrow q \\
a\leftrightarrow b
\end{array} \Bigg )
+ \sum_{rs}\sum_{km_1m_2} {\cal B}_9 g_{rs,k}^{pq}
\bm{\tau}_{ab,m_1m_2}^{rs}
\nonumber \\
& & + \sum_{cd} \sum_{km_1m_2}{\cal B}_{10}g_{ab,k}^{cd}
\bm{\tau}_{cd,m_1m_2}^{pq}
+ \left ( \tilde{\varepsilon}_p
+ \tilde{\varepsilon}_q
- \tilde{\varepsilon}_a
- \tilde{\varepsilon}_b\right )\bm{\tau}_{ab,l_1l_2}^{pq}= 0,
\label{lprcc_d}
\end{eqnarray}
\end{widetext}
where ${\cal A} $ and ${\cal B} $ are the angular coefficients listed in the
Appendix \ref{app_d}-\ref{app_e} and
$\mathbf{d}_{i,1}^{j} = \langle j||\mathbf{d}||i\rangle $ is the
reduced matrix element of the electric dipole operator. In the above
equations, unlike in Eqs. (\ref{psing_eqn}) and (\ref{pdbl_eqn}), each of the
terms are written separately without symmetrization. This is essential as the
direct and exchange diagrams have different angular factors and summation
indexes.
\begin{figure}[h]
\includegraphics[width=8.5cm]{psingles}
\caption{Goldstone diagrams which contribute to the $\mathbf{T}_1^{(1)}$
equation in the LPRCC approximation. The diagrams (a-c), and (j)
arise from ${H}_{\rm N}\mathbf{T}^{(1)}_1$,
(d-i) arise from ${H}_{\rm N}\mathbf{T}^{(1)}_2$, and (k-o) arise
from $\mathbf{d}\mathbf{T}^{(1)}$. The dashed lines ending with a
circle ($\circ$) and filled circle ($\bullet$) correspond to
interactions associated with the single-body part of $H_N$ and
$H_{\rm int}$, respectively. The vertexes with undulating line
and a short vertical stump represent $\mathbf{T}_1^{(1)}$ and
$\mathbf{T}_2^{(1)}$, respectively.
}
\label{psing_diag}
\end{figure}
Although we include $T_3^{(0)}$ in the calculations of the unperturbed
cluster equations, in the PRCC theory computations we restrict to single and
double approximation. The reasons for this are the large number of cluster
amplitudes and a rather involved angular integration for the diagrams
associated with $\mathbf{T}_3^{(1)}$. We, however, consider the contributions
from approximate $\mathbf{T}_3^{(1)}$ obtained through perturbative
calculations. For this we consider the dominant perturbative term, and the
details are provided in the next Section.
\begin{figure}[h]
\includegraphics[width=8.5cm]{pdoubles}
\caption{Goldstone diagrams which contribute to the $\mathbf{T}_2^{(1)}$
in the LPRCC approximation. The diagrams (a-b), (c-j), and
(k-l) arise from ${H}_{\rm N}\mathbf{T}^{(1)}_1$,
${H}_{\rm N}\mathbf{T}^{(1)}_2$, and ${\mathbf{d}}\mathbf{T}^{(1)}_2$,
respectively. The dashed lines ending with a circle ($\circ$) and
filled circle ($\bullet$) correspond to interactions associated with
the single-body part of $H_N$ and $H_{\rm int}$, respectively.
The vertexes with undulating line and a short vertical stump
represent $\mathbf{T}_1^{(1)}$ and $\mathbf{T}_2^{(1)}$, respectively.
}
\label{pdbl_diag}
\end{figure}
\section{Dipole Polarizability}
\subsection{Expression of $\alpha$ in PRCC}
The electric dipole polarizability of the ground state of a closed-shell
atom is given by
\begin{equation}
\alpha = -2 \sum_{I} \frac
{\langle \Psi_0|\mathbf D|\Psi_I\rangle \langle \Psi_I|\mathbf D|
\Psi_0\rangle}{E_0 - E_I},
\end{equation}
where $|\Psi_I \rangle $ are the intermediate atomic states and $E_I$ is the
energy of the atomic state. Considering that the ground state of a
closed-shell atom or ion is even parity, $|\Psi_I \rangle $ must be odd
parity states as $\mathbf{D}$ is an odd parity operator. The above expression
of $\alpha$ in terms of the PRCC theory is
\begin{equation}
\alpha = -\frac{\langle \Phi_0|\mathbf{T}^{(1)\dagger}\bar{\mathbf{D}} +
\bar{\mathbf{D}}\mathbf{T}^{(1)}|\Phi_0\rangle}{\langle\Psi_0|\Psi_0\rangle},
\end{equation}
where, $\bar{\mathbf{D}} = e^{{T}^{(0)\dagger}}\mathbf{D} e^{T^{(0)}}$,
represents the unitary transformed electric dipole operator and
$\langle\Psi_0|\Psi_0\rangle$ is the normalization factor.
Following the derivations presented in our previous works
\cite{chattopadhyay-13a,chattopadhyay-14a}, retaining terms
up to quadratic in cluster operators, we can write
\begin{eqnarray}
\alpha & \approx & \frac{1}{\cal N}\langle\Phi_0|
\mathbf{T}_1^{(1)\dagger}\mathbf{D} + \mathbf{D}\mathbf{T}_1^{(1)}
+ \mathbf{T}_1^{(1)\dagger}\mathbf{D}T_1^{(0)}
+ T_1^{(0)\dagger}\mathbf{D}\mathbf{T}_1^{(1)}\nonumber \\
&& + \mathbf{T}_2^{(1)\dagger}\mathbf{D}T_1^{(0)}
+ T_1^{(0)\dagger}\mathbf{D}\mathbf{T}_2^{(1)}
+ \mathbf{T}_1^{(1)\dagger}\mathbf{D}T_2^{(0)}\nonumber \\
&& + T_2^{(0)\dagger}\mathbf{D}\mathbf{T}_1^{(1)}
+ \mathbf{T}_2^{(1)\dagger}\mathbf{D}T_2^{(0)}
+ T_2^{(0)\dagger}\mathbf{D}\mathbf{T}_2^{(1)}
|\Phi_0\rangle,
\label{exp_alpha}
\end{eqnarray}
where ${\cal N} = \langle\Phi_0|\exp[T^{(0)\dagger}]\exp[T^{(0)}]
|\Phi_0\rangle$ is the normalization factor, which involves a non-terminating
series of contractions between ${T^{(0)}}^\dagger $ and $T^{(0)} $.
In the present work we use
${\cal N} \approx \langle\Phi_0|T_1^{(0)\dagger}T_1^{(0)} +
T_2^{(0)\dagger}T_2^{(0)}|\Phi_0\rangle$. It must be mentioned here that,
as discussed in our previous work \cite{chattopadhyay-14a}, the expression
of $\alpha$ involves only connected diagrams and the normalization
factor is essential. From the above expression of $\alpha$,
an evident advantage of calculation using PRCC theory is the absence of
summation over $|\Psi_I\rangle $. The summation is subsumed in the
evaluation of the $\mathbf{T}^{(1)}$ in a natural way. This is
one of the key advantage of using PRCC theory.
\begin{figure}[h]
\includegraphics[width=8.5cm]{t3_ptr}
\caption{Goldstone diagrams of approximate $\mathbf{T}_3^{(1)}$ obtained
from perturbing $\mathbf{T}_2^{(1)}$ with one order of the
electron-electron interaction $g=1/r_{12} + g^{\rm B}_{12}$,
represented by dashed lines in the diagrams.
}
\label{t3_ptr}
\end{figure}
\subsection{Perturbative $\mathbf{T}_3^{(1)}$ and $\alpha$}
To obtain the dominant contributions from the triple excitation cluster
operators in PRCC, $\mathbf{T}_3^{(1)}$, we consider the perturbative
approximation. In this scheme $\mathbf{T}_3^{(1)}$ is approximated as a
first order perturbation to $\mathbf{T}_2^{(1)}$, and it accommodates the
leading order terms in the cluster amplitude equations of $\mathbf{T}_3^{(1)}$.
There are two diagrams in this approximation and are shown in
Fig. \ref{t3_ptr}, and these combine to give the perturbative triple
excitation cluster amplitude
\begin{equation}
\bm{\tau}_{abc}^{pqr} \approx \frac{1}{\Delta\epsilon_{pqr}^{abc} }
\Big ( \sum_s\bm{\tau}_{ab}^{ps}g^{qr}_{sc}
- \sum_d \bm{\tau}_{ad}^{pq}g^{dr}_{bc} \Big ),
\end{equation}
where, $\Delta\epsilon_{pqr}^{abc}= \epsilon_p + \epsilon_q + \epsilon_r -
\epsilon_a - \epsilon_b - \epsilon_c$, and as defined earlier
$g_{ij}^{kl} = \langle kl| 1/r_{12} + g^{\rm B}_{12}|ij\rangle $. The first
and second term on the right hand side of the above equation correspond to the
Goldstone diagrams in Fig. \ref{t3_ptr}(a) and (b), respectively. Each of these
diagrams, after contraction with $T_2^{(0)\dagger}$ and $\mathbf{D}$,
generate sixteen diagrams of $\alpha$ each. For example, the set of the
sixteen diagrams arising from the perturbative $\mathbf{T}_3^{(1)}$
represented by Fig. \ref{t3_ptr}(a) are shown in Fig. \ref{t3_ptrpp}.
The other term associated with $\mathbf{T}_3^{(1)}$ which contributes to
$\alpha$ is $T_1^{(0)\dagger} T_1^{(0)\dagger}\mathbf{D}\mathbf{T}_3^{(1)}$.
We, however, neglect this as it is second order in $T_1^{(0)\dagger}$ and
expect the contribution to be smaller than
$T_1^{(0)\dagger}\mathbf{D}\mathbf{T}_2^{(1)}$, which as we shall discuss
later has the smallest contribution in the expression of $\alpha$
in Eq. (\ref{exp_alpha}). Here after, the values of $\alpha$ obtained with
the inclusion of perturbative $\mathbf{T}_3^{(1)}$ are referred to as
PRCC(T).
\begin{figure}[h]
\includegraphics[width=8.5cm]{t3_ptrpp}
\caption{Diagrams of $\alpha$ which arise from,
$T_2^{(0)\dagger}\mathbf{D}\mathbf{T}_3^{(1)}$, which represents
the perturbative $\mathbf{T}_3^{(1)}$ contracted with
$T_2^{(0)\dagger}$ and $\mathbf{D}$. The $\mathbf{T}_3^{(1)}$,
considered in the diagrams, is obtained from particle-particle
contraction of $\mathbf{T}_2^{(1)}$ and electron-electron interaction
Hamiltonian $g=1/r_{12} + g^{\rm B}_{12}$, represented by dashed
lines in the diagrams.
}
\label{t3_ptrpp}
\end{figure}
\section{Computational details}
\subsection{Basis set and nuclear density}
We use Gaussian type orbitals (GTOs) \cite{mohanty-91}, and the details
relevant to the use of GTOs in RCC and PRCC are described in our previous
works \cite{chattopadhyay-12a, chattopadhyay-13a}. The GTOs are finite basis
set orbitals and are the linear combinations of Gaussian type functions (GTFs).
The exponents of the GTFs are defined in terms of two parameters $\alpha_{0}$
and $\beta$. We
consider even tempered basis set, or in other words, different $\alpha_{0}$
and $\beta$ for orbitals of each $j$. We also use kinetic balance condition
\cite{stanton-84} to obtain small components of the orbitals from the
large component. Further more, it is appropriate to incorporate Breit
interaction \cite{quiney-91} in the generation of GTOs as the
present study includes Hg, a high $Z$ atom. For this the works of
Quiney \cite{quiney-03} and Mohanty \cite{mohanty-91}, and their collaborators
are excellent references. Keeping in view the implementations general and
incorporating mathematically intricate interaction Hamiltonians, for example,
the Uehling potential, we generate the GTOs on a grid \cite{chaudhuri-99}
with $V^{\rm N}$ potential. The basis parameters $\alpha_{0}$ and $\beta$
are optimized by matching the orbital and self-consistent field (SCF)
energies obtained from GRASP2K \cite{jonsson-13} with the Dirac-Coulomb
Hamiltonian. The values of the optimized parameters of Zn, Cd and Hg are
listed in Table. \ref{basis}.
\begin{table}[h]
\caption{The $\alpha_0$ and $\beta$ parameters for the $s$, $p$
and $d$ orbitals of the even tempered GTO basis used in the
present calculations.}
\label{basis}
\begin{tabular}{cccccccc}
\hline
\hline
Atom & \multicolumn{2}{c}{$s$} & \multicolumn{2}{c}{$p$} &
\multicolumn{2}{c}{$d$} \\
& $\alpha_{0}$ & $\beta$ & $\alpha_{0}$ & $\beta$
& $\alpha_{0}$ & $\beta$ \\
\hline
Zn &\, 0.0385 &\, 2.045 &\, 0.1095 &\, 2.035 &\, 0.0091 &\, 2.010 \\
Cd &\, 0.0505 &\, 2.101 &\, 0.0775 &\, 1.985 &\, 0.0340 &\, 1.950 \\
Hg &\, 0.0505 &\, 2.045 &\, 0.1019 &\, 2.223 &\, 0.0380 &\, 2.050 \\
\hline
\end{tabular}
\end{table}
The SCF energies $E_{\rm SCF}$ obtained with the optimized basis
parameters are listed in Table. \ref{scf}. It is evident from
the table that $E_{\rm SCF}$ from the GTOs are in very good agreement with
the results of GRASP2K, which solves the Dirac-Hartree-Fock equations
numerically. An important step in generating the orbitals with
GRASP2K is, we use the Hartree-Fock orbitals \cite{fischer-87} as the
starting values of GRASP2K to improve convergence. As mentioned earlier,
we also compare the orbital energies for basis parameter optimization.
The details of these comparisons are presented and described in the
results and discussions section.
\begin{table}[h]
\caption{The Dirac-Coulomb SCF energies $E_{\rm SCF}$ of Zn, Cd and Hg
obtained from GRASP2K \cite{jonsson-13} and using Gaussian
type orbitals are listed. The Breit interaction corrections
to SCF energy $\Delta{E}^{\rm SCF}_{\rm Br}$ are computed
using the Gaussian type orbitals. All the values are in
atomic units (hartree).
}
\label{scf}
\begin{center}
\begin{tabular}{ldddd}
\hline
Atom &
\multicolumn{2}{c}{$E_{\rm SCF}$} &
\multicolumn{2}{c}{$\Delta{E}^{\rm SCF}_{\rm Br}$} \\
\hline
&\multicolumn{1}{c}{GTO} &
\multicolumn{1}{c}{GRASP2K} &
\multicolumn{1}{c}{Present} &
\multicolumn{1}{c}{Ref. \cite{ishikawa-94}} \\
\hline
Zn & -1794.6127 & -1794.6127 & -0.7610 & -0.7610 \\
Cd & -5593.3188 & -5593.3184 & -3.8389 & -3.8389 \\
Hg & -19648.8243 & -19648.8580 & -22.6328 & -22.6325 \\
\hline
\end{tabular}
\end{center}
\end{table}
To generate the nuclear potential $V_N(r)$, we use two-parameter
finite size Fermi density distribution of the nucleus
\begin{equation}
\rho_{\rm nuc}(r) = \frac{\rho_0}{1 + e^{(r-c)/a}},
\end{equation}
where, $a=t 4\ln(3)$. The parameter $c$ is the half charge radius so that
$\rho_{\rm nuc}(c) = {\rho_0}/{2}$ and $t$ is the skin thickness.
Using the orbital basis set, we can then solve the RCC and PRCC equations
with standard linear algebra method. For efficient parallel implementation
we solve the equations iteratively using Jacobi method. It is, however,
a method with slow convergence, so employ direct inversion in the
iterated subspace (DIIS) \cite{pulay-80} to improve convergence.
\subsection{Breit and vacuum polarization corrections}
In the present work, we use the general expressions of Breit interaction
integrals listed in the work of Grant \cite{grant-76}. To examine the
corrections to orbital energies arising from the Breit interactions, we
generate the orbitals as solutions of two slightly different
single particle equations. In the first case, the orbitals $|\psi_i \rangle$
are computed with the Dirac-Hartree-Fock (DHF) potential and solutions
of the equation,
\begin{equation}
\left ( h_{0} + U_{\rm DHF} \right ) |\psi_i \rangle =
\epsilon_i |\psi_i \rangle, \nonumber
\end{equation}
where,
$h_0 = c\bm{\alpha}\cdot\mathbf{p} +(\beta - 1)c^2 -V_{\rm N}(\mathbf{r})$ is
the single particle part of Dirac-Coulomb Hamiltonian, $|\psi_i \rangle$ is a
four component orbital and $\epsilon_i$ is the corresponding eigenvalue.
The DHF potential in the above equation is defined as
\begin{equation}
U_{\rm DHF}|\psi_i \rangle = \sum_c^{\rm core} \left [
\langle \psi_c|\frac{1}{r_{12}}\left (1-P_{12} \right )|\psi_c\rangle
|\psi_i\rangle \right ],
\label{dhf_pot}
\end{equation}
where, $P_{12}$ is the permutation operator to represent the exchange
integral, $c$ represents core orbitals and `core' indicates sum over all
the core orbitals. This implies that the core orbitals are solutions of
a set of coupled integro-differential equations and solved using
self-consistent-field (SCF) methods. In the second case, we compute
the orbitals $|\psi'_i \rangle$ with the Dirac-Hartree-Fock-Breit (DHFB)
potential. The orbitals are then the solutions of the single particle
equation
\begin{equation}
\left ( h_{0} + U_{\rm DHFB} \right) |\psi'_i \rangle =
\epsilon'_i |\psi '_i \rangle, \nonumber
\end{equation}
where, $\epsilon'_i$ is the eigenvalue with the DHFB potential, and
$U_{\rm DHFB}$ is obtained by adding $g^{\rm B}_{12}$ to the central
potential in Eq. (\ref{dhf_pot}). From the solutions we define the
correction to orbital energies due to Breit interaction as
\begin{equation}
\Delta\epsilon_{{\rm Br}(i)} = \epsilon' _i - \epsilon_i.
\end{equation}
In a similar way, we also compute the correction due to Uehling potential
$\Delta\epsilon_{\rm Ueh}$.
From the two sets of the orbitals, we define two many-particle ground
state reference $|\Phi_0\rangle$ and $|\Phi'_0\rangle$, which
are determinantal states consisting of $|\psi_c \rangle$ and
$|\psi'_c \rangle$ orbitals, respectively. Based on these states, the
SCF energy correction due to Breit interaction is
\begin{equation}
\Delta E_{\rm Br}^{\rm SCF} = \langle\Phi'_0|H^{\rm DCB}|\Phi'_0\rangle
- \langle\Phi_0|H^{\rm DC}|\Phi_0\rangle,
\end{equation}
where, $H^{\rm DC}$ is the Dirac-Coulomb Hamiltonian: the Hamiltonian
$H^{\rm DCB}$ defined in Eq. (\ref{ham_dcb}) without the Breit interaction.
The values of $\Delta E_{\rm Br}^{\rm SCF}$ for Zn, Cd and Hg are listed in
Table. \ref{scf}, and are near perfect match with the values reported in
a previous work \cite{ishikawa-94}. This is another important comparison
which validates the choice of the optimized basis set parameters used in the
present study. In the results and discussions section, we present
$\Delta\epsilon_{\rm Br}$ of Zn, Cd and Hg orbitals. For the first two atoms,
Zn and Cd, we were unable to get previous results from the literature.
However, for Hg a previous work \cite{lindroth-89a} has provided the values of
$\Delta\epsilon_{\rm Br}$, and our results are in excellent agreement with
those values.
Another way to quantify the effect of Breit interaction is to calculate
the first order correction to the SCF energy as
\begin{equation}
\langle H^{\rm B}\rangle_{\rm DF} = \langle\Phi_0|\sum_{i<j}g^{\rm B}(r_{ij})
|\Phi_0\rangle.
\end{equation}
In a previous work \cite{chattopadhyay-12b}, we have reported
$\langle H^{\rm B}\rangle_{\rm DF}$ for the noble gas atoms, and computations
were based on the compact expressions of Breit interaction integrals listed in
the work of Grant and McKenzie \cite{grant-80}. The computation of
$\langle H^{\rm B}\rangle_{\rm DF}$ is well suited for testing the
implementation of Breit interactions. In the present work, as we have
incorporated Breit interaction in the GTO generation and coupled-cluster
codes, we give our results of $\Delta E_{\rm Br}^{\rm SCF}$ and
$\Delta\epsilon_{\rm Br}$, but not the values of
$\langle H^{\rm B}\rangle_{\rm DF}$. It must be mentioned here that, among
the previous works on Breit interactions, there is another approach to
evaluate the Breit interaction matrix elements reported in the work of Mann
and Johnson \cite{mann-71}. It is based on the coupling of the Dirac
matrices with the angular part of the orbitals. In contrast, the expressions
of Grant and collaborators, which we have used, are based on the expansion
of $g^{\rm B}(r_{12})$ as linear combination of irreducible tensor operators.
\section{Results and Discussions}
The elements of the group IIB studied in the present work, have filled
$ns$ orbitals as valence shells and in this regard, similar to the neutral
alkaline-earth-metal atoms. There is, however, an important difference:
in the group IIB elements the filled $(n-1)d$ shells are the highest energy
core orbitals and we can expect significant contribution to the correlation
effects from the electrons in the $(n-1)d$ shell. This is indeed the case
and is reflected in the identification of the occupied orbitals with
dominant contributions to the leading order (LO) term,
$\mathbf{T}_1^{(1)\dagger}\mathbf{D} + \text{H.c.}$,
in $\alpha$. We also examine the trends in the contribution from
Breit-interaction to the energies of the occupied orbitals. For better
description the results for each of the elements (Zn, Cd and Hg) are
discussed separately. All the values of $\alpha$ are in atomic units, that
is in units of $a_0^3$, where $a_0$ is the Bohr radius.
\begin{table}[h]
\caption{Orbital energies of Zn and Cd obtained from GRASP2K
\cite{jonsson-13} and Gaussian type orbitals in atomic units
(hartree). Here [x] represents multiplication by ${10^x}$.}
\label{orbe_zncd}
\begin{tabular}{lddD..{3.7}D..{3.7}}
\hline
{Orbital}
& \multicolumn{1}{c}{\textrm{GRASP2K}} &
\multicolumn{1}{c}{\textrm{DC}} &
\multicolumn{1}{c}{\text{$\Delta\epsilon_{\rm Br}$}} &
\multicolumn{1}{c}{\text{$\Delta\epsilon_{\rm Ueh}$}} \\
\hline
\multicolumn{5}{c}{Zn}\\
$1s_{1/2}$ & -357.7486 & -357.7486 & 4.364[-1] &-2.174[-2] \\
$2s_{1/2}$ & -45.3461 & -45.3461 & 3.129[-2] &-2.125[-3] \\
$2p_{1/2}$ & -39.7403 & -39.7402 & 5.524[-2] & 1.724[-4] \\
$2p_{3/2}$ & -38.8513 & -38.8513 & 3.586[-2] & 1.859[-4] \\
$3s_{1/2}$ & -5.8000 & -5.7999 & 3.301[-3] &-3.113[-4] \\
$3p_{1/2}$ & -3.9579 & -3.9578 & 5.896[-3] & 3.201[-5] \\
$3p_{3/2}$ & -3.8372 & -3.8371 & 3.135[-3] & 3.419[-5] \\
$3d_{3/2}$ & -0.7709 & -0.7709 & 2.015[-4] & 2.470[-5] \\
$3d_{5/2}$ & -0.7547 & -0.7547 &-8.255[-4] & 2.453[-5] \\
$4s_{1/2}$ & -0.2986 & -0.2986 & 1.251[-4] &-1.080[-5] \\
\multicolumn{5}{c}{Cd}\\
$1s_{1/2}$ & -987.3591 & -987.3580 & 2.017 &-1.519[-1] \\
$2s_{1/2}$ & -149.8044 & -149.8032 & 1.810[-1] &-1.702[-2] \\
$2p_{1/2}$ & -139.0231 & -139.0218 & 3.117[-1] & 5.952[-4] \\
$2p_{3/2}$ & -131.9158 & -131.9145 & 2.109[-1] & 1.006[-3] \\
$3s_{1/2}$ & -29.3222 & -29.3212 & 2.465[-2] &-3.239[-3] \\
$3p_{1/2}$ & -24.9552 & -24.9541 & 4.581[-2] & 1.674[-4] \\
$3p_{3/2}$ & -23.6459 & -23.6451 & 2.743[-2] & 2.552[-4] \\
$3d_{3/2}$ & -16.0009 & -16.0001 & 1.231[-2] & 2.389[-4] \\
$3d_{5/2}$ & -15.7383 & -15.7374 & 4.173[-3] & 2.344[-4] \\
$4s_{1/2}$ & -4.7469 & -4.7460 & 3.487[-3] &-5.810[-4] \\
$4p_{1/2}$ & -3.2707 & -3.2698 & 6.390[-3] & 5.666[-5] \\
$4p_{3/2}$ & -3.0461 & -3.0451 & 3.141[-3] & 7.209[-5] \\
$4d_{3/2}$ & -0.7383 & -0.7374 & 8.961[-5] & 5.591[-5] \\
$4d_{5/2}$ & -0.7089 & -0.7080 &-8.777[-4] & 5.506[-5] \\
$5s_{1/2}$ & -0.2814 & -0.2810 & 1.930[-4] &-3.273[-5] \\
\hline
\end{tabular}
\end{table}
\subsection{Zn}
The corrections to the orbitals energies $\Delta\epsilon_{\rm Br}$
and $\Delta\epsilon_{\rm Ue}$ arising from Breit-interaction and
Uehling potential, respectively, are listed in Table. \ref{orbe_zncd}. From the
table it is evident that the Breit-interaction tends to {\em relax} the
orbitals as $\Delta\epsilon_{\rm Br}$ is positive in all the cases except
$3d_{5/2}$. For the latter, $\Delta\epsilon_{{\rm Br}(3d_{5/2})}$, is negative
and indicates contraction of the orbital. In absolute terms the value of
$-8.255\times 10^{-4}$ hartree for $\Delta\epsilon_{{\rm Br}(3d_{5/2})}$ is
small but the magnitude is larger than $\Delta\epsilon_{{\rm Br}(3d_{3/2})}$.
As to be expected, the deeper core orbitals or orbitals with lower principal
quantum number $n$ have larger $\Delta\epsilon_{\rm Br}$ and there is a three
orders of magnitude difference between the values of $\Delta\epsilon_{\rm Br}$
for $1s$ and $4s$.
\begin{table}[h]
\caption{Convergence pattern of $\alpha$ for Zn and Cd as function of
the basis set size. The values of $\alpha$ are in atomic units
( $a_0^3$).}
\label{conv_pat}
\begin{tabular}{lcc}
\hline
No. of orbitals & Basis size & $\alpha $ \\
\hline
\multicolumn{3}{c}{Zn}\\
113 & $(15s, 13p, 11d, 9f, 9g, 7h) $ & 38.722 \\
135 & $(17s, 15p, 15d, 10f, 10g, 9h) $ & 38.717 \\
153 & $(19s, 17p, 17d, 11f, 11g, 11h) $ & 38.716 \\
171 & $(21s, 19p, 19d, 13f, 13g, 11h) $ & 38.716 \\
\multicolumn{3}{c}{Cd}\\
99 & $(15s, 12p, 11d, 7f, 6g, 6h) $ & 49.421 \\
121 & $(17s, 14p, 13d, 9f, 8g, 8h) $ & 49.135 \\
143 & $(19s, 16p, 15d, 11f, 10g, 10h) $ & 49.113 \\
165 & $(21s, 18p, 17d, 13f, 12g, 12h) $ & 49.112 \\
\multicolumn{3}{c}{Hg}\\
112 & $(12s, 11p, 11d, 11f, 9g, 8h) $ & 33.513 \\
134 & $(14s, 13p, 13d, 13f, 11g, 10h) $ & 33.499 \\
167 & $(17s, 16p, 16d, 16f, 14g, 13h) $ & 33.499 \\
178 & $(18s, 17p, 17d, 17f, 15g, 14h) $ & 33.499 \\
\hline
\end{tabular}
\end{table}
The energy correction arising from the Uehling potential
$\Delta\epsilon_{\rm Ue} $ are also listed in Table. \ref{orbe_zncd}. It
is evident that Uehling potential tends to contract the $s$ orbitals
as $\Delta\epsilon_{\rm Ue}$ of these orbitals are negative. On the other
hand, the occupied orbitals of other symmetries ($p$ and $d$) {\em relax}
and are indicated by the positive values of $\Delta\epsilon_{\rm Ue} $. This
trend is similar to the results of doubly ionized alkaline-earth-metals
Mg$^{2+}$, Ca$^{2+}$, Sr$^{2+}$, and Ba$^{2+}$ reported in our previous
work \cite{chattopadhyay-13b}. In terms of magnitude, the values of
$\Delta\epsilon_{\rm Ue} $ are on average an order of magnitude smaller than
$\Delta\epsilon_{\rm Br}$.
From Table. \ref{orbe_zncd}, it is evident that the basis set parameters
reproduces the numerical values of the orbital energies, obtained using
GRASP2K \cite{jonsson-13}, to an accuracy of $10^{-4}$ hatree or lower.
To determine the optimal orbital basis set, we compute $\alpha$ with
increasing basis size and the results are listed in Table. \ref{conv_pat}.
From the table, we observe convergence of $\alpha$ up to $10^{-3}$ a.u. with
a basis set of 171 orbitals. Based on the results, we choose the set with
135 orbitals as the optimal one and use it for more detailed studies.
In Table. \ref{pol_group2b} the converged values of $\alpha$ along with the
previous theoretical results and experimental data are listed for comparison.
From the table it is evident that our result of 38.72 is in very good agreement
with the experimental value of 38.8(8). Among the previous theoretical
results, the results from configuration interaction with a semi-empirical
core-polarization potential (CICP) \cite{ye-PRA-08} is on the lower side.
There are two other theoretical results based on coupled-cluster theory.
The first \cite{goebel-96} is using non-relativistic Hamiltonian with
finite field approach, where as the second \cite{yashpal-14} uses
Dirac-Coulomb Hamiltonian with the external electric field treated as a
perturbation. In both the works, the contributions from triple excitations
are included perturbatively. Compared to the experimental value, the results
from the first work \cite{goebel-96} is on the higher side, but the
result from the second work \cite{yashpal-14} is close to the experimental
value. The method used in ref. \cite{yashpal-14} is similar, in the way
the external field is treated as a perturbation and computation of a second set
of cluster amplitudes, to PRCC. However, our result is in
better agreement with the experimental value. This may be on account of two
important factors: inclusion of Breit-interaction in the atomic Hamiltonian
and computation of $T_3^{(0)}$ without perturbative approximations.
With the inclusion of perturbative $\mathbf{T}_3^{(1)}$,
result listed as PRCC(T) in Table.\ref{pol_group2b}, our
result is in excellent agreement with the experimental data.
The term wise contribution to $\alpha$ in Eq. (\ref{exp_alpha}) are listed
in Table. \ref{cont_term}. From the table the LO contribution arises from
$\mathbf{T}_1^{(1)\dagger}\mathbf{D} + \text{H.c.}$ and is larger than
the total value of $\alpha$. This is, perhaps, not surprising as the LO term
subsumes the Dirac-Hartree-Fock contribution and core-polarization effects.
The next to leading order (NLO) is
$\mathbf{T}_1^{(1)\dagger}\mathbf{D}T_1^{(0)} + \text{H.c.}$,
and opposite in phase to the LO. A similar phase relation between the LO and
NLO was observed in our previous work on noble gas \cite{chattopadhyay-12b}
and alkaline-Earth-metal \cite{chattopadhyay-14a} atoms. Among the remaining
terms, the contribution from
$\mathbf{T}_1^{(1)\dagger}\mathbf{D}T_2^{(0)} + \text{H.c.}$ is similar
in value and phase to the NLO term.
The sub-shell wise contributions from the LO term, as mentioned earlier is
the sum of $\mathbf{T}_1^{(1)\dagger}\mathbf{D}$ and its hermitian
conjugate, are listed in Table. \ref{result_t1d}. From the table, the
valence sub-shell $4s_{1/2}$ is the most dominant, and followed by
$3d_{5/2}$. Both the sub-shell contributions have same phase, and together
accounts for more than 99\% of the LO term.
\begin{table}[h]
\caption{Static dipole polarizability $\alpha$ of Zn, Cd and Hg
in atomic units ($a_0^3$). }
\label{pol_group2b}
\begin{center}
\begin{tabular}{ldcdc}
\hline
\multicolumn{1}{c}{Atom} & \multicolumn{1}{c}{Present} &
\multicolumn{1}{c}{Method} & \multicolumn{1}{c}{Previous Works} &
\multicolumn{1}{c}{Method} \\
\hline
\hline
$\rm{Zn}$ & 38.72 & PRCC & 38.12 \text{\cite{ye-PRA-08}}
& CICP \\
& 38.76 & PRCC(T) & 38.5 \text{\cite{rosenkrantz-80}}
& MCSCF \\
& & & 38.4 \text{\cite{roos-05}}
& CASPT2 \\
& & & 37.86 \text{\cite{kello-95}}
& CCSD(T) \\
& & & 38.01 \text{\cite{seth-97}}
& CCSD(T) \\
& & & 39.2(8) \text{\cite{goebel-96}}
& CCSD(T) \\
& & & 38.666(35) \text{\cite{yashpal-14}}
& RCCSD$_p$T \\
& & & 38.8(8) \text{\cite{goebel-96}}
& Expt. \\
& & & 38.92 \footnotemark[1] & Expt. \\
$\rm{Cd}$ & 49.11 & PRCC & 44.63 \text{\cite{ye-PRA-08}}
& CICP \\
& 49.20 & PRCC(T) & 46.9 \text{\cite{roos-05}}
& CASPT2 \\
& & & 47.63 \text{\cite{kello-95}}
& CCSD(T) \\
& & & 46.25 \text{\cite{seth-97}}
& CCSD(T) \\
& & & 45.856(42) \text{\cite{yashpal-14}}
& RCCSD$_p$T \\
& & & 49.65(1.47) \text{\cite{goebel-95}}
& Expt. \\
& & & 49.50 \footnotemark[1] & Expt. \\
& & & 50.0(2.8) \footnotemark[2] & Expt. \\
$\rm{Hg}$ & 33.50 & PRCC & 31.32 \text{\cite{ye-PRA-08}}
& CICP \\
& 33.59 & PRCC(T) & 33.3 \text{\cite{roos-05}}
& CASPT2 \\
& & & 33.44 \text{\cite{schwerdtfeger-94}}
& QCISD(T) \\
& & & 31.82 \text{\cite{kello-95}}
& CCSD(T) \\
& & & 34.42 \text{\cite{seth-97}}
& CCSD(T) \\
& & & 34.15 \text{\cite{pershina-08}}
& CCSD(T) \\
& & & 33.6 \text{\cite{hachisu-08}}
& CI + MBPT\\
& & & 33.7(1.3) \footnotemark[3] & Expt. \\
& & & 33.75 \footnotemark[4] & Expt. \\
& & & 33.91(34) \text{\cite{goebel-96a}}
& Expt. \\
\hline
\hline
\end{tabular}
\end{center}
\footnotetext[1]{Reference \cite{qiao-12} based on experimental data
in Ref. \cite{goebel-95, goebel-96}.}
\footnotetext[2]{Reference \cite{goebel-95} based on the refractive
index data in Ref. \cite{cuthbertson-1908}.}
\footnotetext[3]{Reference \cite{goebel-96a} based on the dielectric
data in Ref. \cite{wusthoff-36}.}
\footnotetext[4]{Reference \cite{tang-08} based on the experimental data
in Ref. \cite{goebel-96a}.}
\end{table}
\subsection{Cd}
The corrections to the orbitals energies $\Delta\epsilon_{\rm Br}$
and $\Delta\epsilon_{\rm Ue}$ arising from Breit-interaction and
Uehling potential, respectively, are listed in Table. \ref{orbe_zncd}.
From the table it is evident that like in Zn $\Delta\epsilon_{\rm Br}$ of
the $4d_{5/2}$, the sub-shell next to the valence, is negative. Over all the
general trend in the corrections is very similar to the case of Zn, except
that the magnitude of the corrections are one order higher. There is,
however, one noticeable change in the relative values of
$\Delta\epsilon_{\rm Ue}$ for the $p_{1/2}$ and $p_{3/2}$ orbitals. In
the case of Zn, $\Delta\epsilon_{{\rm Ue}(mp_{1/2})}
\approx \Delta\epsilon_{{\rm Ue}(mp_{3/2})}$ (with $m=2,3$), but in the Cd,
$\Delta\epsilon_{{\rm Ue}(mp_{1/2})}$ is about a factor of two smaller than
$\Delta\epsilon_{{\rm Ue}(mp_{3/2})}$. This indicates an enhanced effect of
the Uehling potential or vacuum polarization potential to the inner $p_{1/2}$
orbitals with higher nuclear charge $Z$. It is an expected trend as the
$p_{1/2}$ orbitals contract with higher $Z$ due to larger relativistic
corrections, and the inner orbitals contract more as the correction is
larger.
\begin{table}[h]
\caption{Contribution to $\alpha $ from different terms and their
hermitian conjugates in the PRCC theory in atomic units
($a_0^3$).}
\label{cont_term}
\begin{center}
\begin{tabular}{lddd}
\hline
Terms + h.c. & \multicolumn{1}{r}{$\rm{Zn}$}
& \multicolumn{1}{r}{$\rm{Cd}$}
& \multicolumn{1}{r}{$\rm{Hg}$} \\
\hline
$\mathbf{T}_1^{(1)\dagger}\mathbf{D} $
& 45.590 & 61.456 & 41.927 \\
$\mathbf{T}_1{^{(1)\dagger}}\mathbf{D}T_2^{(0)} $
& -1.850 & -3.128 & -2.724 \\
$\mathbf{T}_2{^{(1)\dagger}}\mathbf{D}T_2^{(0)} $
& 1.364 & 2.060 & 1.504 \\
$\mathbf{T}_1{^{(1)\dagger}}\mathbf{D}T_1^{(0)} $
& -1.901 & -3.808 & -1.583 \\
$\mathbf{T}_2{^{(1)\dagger}}\mathbf{D}T_1^{(0)} $
& 0.081 & 0.243 & 0.091 \\
Normalization & 1.118 & 1.157 & 1.171 \\
Total & 38.716 & 49.112 & 33.499 \\
\hline
\end{tabular}
\end{center}
\end{table}
Like in the case of Zn, orbital energies of Cd corresponding to the
GTOs and numerical results from GRASP2K \cite{jonsson-13} are listed in
the Table. \ref{orbe_zncd}. It is evident that the basis parameters chosen
for the Cd basis matches the orbital energies with the numerical results to
within $10^{-4} - 10^{-3}$ hatrees. On comparison, on an average the
agreement is in the case of Zn an order of magnitude better. This is on
account of the larger number of occupied orbitals Cd, which increases the
parameters of optimization. Coming to the results of $\alpha$, from
Table. \ref{conv_pat}, we find that $\alpha$ converges to
$\approx 10^{-3}$ a. u. with a basis set of 165 orbitals. However, considering
the number of cluster amplitudes, we take the basis set consisting of
143 orbitals for further computations. It must be mentioned that, with this
basis set the convergence of $\alpha$ is $\approx 10^{-2}$ a. u..
From the results listed in Table. \ref{pol_group2b}, it is evident
that there is a variation in the previous results from coupled-cluster theory.
There are three previous theoretical works on the computation of $\alpha$
using coupled-cluster theory \cite{kello-95,seth-97,yashpal-14}. However,
each of these use different types of basis sets, Ref. \cite{kello-95} and
\cite{seth-97} are based on optimization with polarization potential and
pseudo-potential, respectively. In terms of the theory and type of basis
functions, the methods we have used in the present work is very similar to
Ref. \cite{yashpal-14}. There is, however, noticeable difference between
the two results, and this may be due to difference in the methods at various
stages of computations. For the present work, as described earlier, we have
provided detailed information about the basis set parameters, and convergence
of $\alpha$ with the basis size. It must be emphasized that our
result for $\alpha$ is closest to the experimental value.
The agreement with the experimental data improves with the
inclusion of perturbative $\mathbf{T}_3^{(1)}$, the result listed as
PRCC(T) in Table. \ref{pol_group2b}.
The term wise contribution to $\alpha$ as listed in Table. \ref{cont_term}
has the same trend, albeit larger values, as in Zn.
Coming to the sub-shell contributions to
the LO term, from the values listed in Table. \ref{result_t1d} the
pattern is similar to Zn: the dominant contribution arises from the valence
sub-shell $5s_{1/2}$, and followed by $4d_{5/2}$, the sub-shell below the
valence. However, compared to Zn, the dominant and next contribution in Cd
are $\approx 25$ \% and $\approx 47$\% larger, respectively.
\begin{table}[h]
\caption{Four leading contributions to
$\{ \mathbf{T}_1^{(1)\dagger}\mathbf{D} \}$ to $\alpha $
in terms of the core spin-orbitals in atomic units ($a_0^3$). }
\label{result_t1d}
\begin{center}
\begin{tabular}{rrr}
\hline
\multicolumn{1}{c}{Zn} & \multicolumn{1}{c}{Cd}
& \multicolumn{1}{c}{Hg} \\ \hline
22.244 (4$s_{1/2}$) & 29.771 (5$s_{1/2}$) & 17.768 (6$s_{1/2}$) \\
0.362 (3$d_{5/2}$) & 0.678 (4$d_{5/2}$) & 2.239 (5$d_{5/2}$) \\
0.193 (3$d_{3/2}$) & 0.340 (4$d_{3/2}$) & 0.965 (5$d_{3/2}$) \\
-0.001 (3$s_{1/2}$) & -0.004 (4$p_{3/2}$) & -0.009 (5$p_{3/2}$) \\
\hline
\end{tabular}
\end{center}
\end{table}
Concerning the experimental results, there is slight variation of the
experimental uncertainty listed in the literature. In the original
experimental work of Goebel and Hohm \cite{goebel-95}, the $\alpha$ of Cd is
reported as $49.65\pm1.46\pm0.16$ a.u. Based on this result the experimental
value is listed as $49.65\pm1.46$, $49.65 (1.49)$ and $49.65\pm1.62$ in
Ref. \cite{schwerdtfeger-06}, \cite{mitroy-10}, and \cite{hohm-12},
respectively. However, the quadrature of the uncertainties
reported in Ref. \cite{goebel-95} gives the result $49.65(1.47)$, the value
listed in Table. \ref{pol_group2b} of the present work. This is a minor issue
and does not impact on the experimental results. We have mentioned this to
explain the difference in the experimental result of Cd listed in
Table. \ref{pol_group2b} from the previous works, namely
Ref. \cite{schwerdtfeger-06}, \cite{mitroy-10} and \cite{hohm-12}.
One issue which require some consideration is the consistent lower values of
$\alpha$ reported in the previous theoretical works when compared to the
experimental data. A comprehensive overview of the experimental results
indicates the value of 49.50 reported by Qiao and collaborators \cite{qiao-12}
based on the experimental data of Goebel and Hohm \cite{goebel-95}, we believe,
is robust and reliable. This observation is based on
three important considerations. First, the Wolfsohn's three term expression
\cite{wolfsohn-33} used in Ref. \cite{qiao-12}, to calculate $\alpha$ from the
frequency dependent
dipole polarizability $\alpha(\omega)$, is an improvement over the three
term Cauchy expansion used in Ref. \cite{goebel-95}. Second, the value
50.0(2.8) reported in Ref. \cite{goebel-95}, based on the refractive index
data from the work of Cuthbertson and Metcalfe \cite{cuthbertson-1908}, is
consistent with the results in Ref. \cite{goebel-95,qiao-12}. Finally, in the
recent work of Hohm and Thakker \cite{hohm-12}, using a fitting function
with second ionization energy and Waber-Cromer radius \cite{waber-65}
as parameters, they
arrive at the value of $\alpha$ for Cd as 50.72. This is very closed to the
experimental values and must be given weightage as the values of
$\alpha$ reported in Ref. \cite{hohm-12}, except for Hf, Pd and Hg, are in
good agreement with the reliable theoretical and experimental results.
So, there is consistency in the experimental, and semi-empirical results
reported in the literature. This indicates the genesis of the lower
theoretical results in the previous works must lie within the theoretical
means and methods employed.
Returning to the wide variation in the theoretical results, the possible
reason for this could be, as evident from Table. \ref{pol_group2b} Cd has
the largest value of $\alpha$ among the group IIb elements. In addition,
$Z$ of Cd lies in the domain where relativistic effects begin to have an
importance. So, in Cd, the relativistic and electron-correlation effects are
inter-related strongly, as a result the properties which depend on electron
correlation effects are sensitive to the choice of the basis set. One
indication of this is the difference between the Hartree-Fock and CCSD(T)
results of the $\alpha$ from the relativistic computations. From
Ref. \cite{seth-97}, this is found to be 17.12 which is larger than the
corresponding values of 12.18 and 10.36 for Zn and Hg, respectively. This
demonstrates the importance of the relativistic and correlation effects.
\begin{table*}
\caption{Orbital energies of Hg obtained from GRASP2K
\cite{jonsson-13} and Gaussian type orbitals in atomic units.
The quantities $\Delta{E}_{\rm Br}$ and $\Delta{E}_{\rm Ueh}$
are the orbital energy corrections arising from the Breit
interaction and Uehling potential, respectively. In the
table, [x] represents multiplication by ${10^x}$. All the
values are in atomic units (hartree).
}
\label{orb_hg}
\begin{ruledtabular}
\begin{tabular}{lddddd}
{Orbital}
& \multicolumn{1}{c}{\text{GRASP2K}} &
\multicolumn{1}{c}{\text{DC}} &
\multicolumn{2}{c}{\text{$\Delta{E}_{\rm Br}$}} &
\multicolumn{1}{c}{\text{$\Delta{E}_{\rm Ueh}$}} \\
\hline
& & & \multicolumn{1}{c}{\text{Present}}&
\multicolumn{1}{c}{\text{Ref. \cite{lindroth-89a}}} & \\
\hline
$1s_{1/2}$ & -3074.226\,002 & -3074.235\,257 & 10.963\,407
& 10.96 & -1.557\,141 \\
$2s_{1/2}$ & -550.251\,032 & -550.254\,927 & 1.229\,461
& 1.230 & -2.206\,016[-1] \\
$2p_{1/2}$ & -526.854\,793 & -526.857\,122 & 2.067\,249
& 2.067 & -1.415\,186[-2] \\
$2p_{3/2}$ & -455.156\,786 & -455.159\,068 & 1.304\,845
& 1.305 & 7.499\,950[-3] \\
$3s_{1/2}$ & -133.113\,168 & -133.116\,535 & 2.275\,130[-1]
& 2.276[-1] & -5.013\,463[-2] \\
$3p_{1/2}$ & -122.639\,005 & -122.640\,349 & 3.933\,351[-1]
& 3.933[-1] & -3.454\,750[-3] \\
$3p_{3/2}$ & -106.545\,242 & -106.546\,285 & 2.346\,877[-1]
& 2.347[-1] & 2.184\,390[-3] \\
$3d_{3/2}$ & -89.436\,975 & -89.440\,259 & 1.708\,149[-1]
& 1.708[-1] & 2.426\,999[-3] \\
$3d_{5/2}$ & -86.020\,282 & -86.023\,564 & 1.098\,651[-1]
& 1.098[-1] & 2.298\,568[-3] \\
$4s_{1/2}$ & -30.648\,324 & -30.649\,589 & 4.665\,828[-2]
& 4.667[-2] & -1.258\,914[-2] \\
$4p_{1/2}$ & -26.124\,024 & -26.123\,690 & 8.337\,968[-2]
& 8.339[-2] & -7.139\,890[-4] \\
$4p_{3/2}$ & -22.188\,555 & -22.188\,057 & 4.359\,830[-2]
& 4.360[-2] & 7.141\,810[-4] \\
$4d_{3/2}$ & -14.796\,757 & -14.797\,894 & 2.297\,811[-2]
& 2.297[-2] & 7.153\,100[-4] \\
$4d_{5/2}$ & -14.052\,597 & -14.053\,659 & 9.563\,165[-3]
& 9.554[-3] & 6.841\,500[-4] \\
$4f_{5/2}$ & -4.472\,939 & -4.472\,953 & -5.808\,097[-3]
& -5.816[-3] & 5.019\,090[-4] \\
$4f_{7/2}$ & -4.311\,769 & -4.311\,745 & -1.148\,315[-2]
& -1.150[-2] & 4.923\,107[-4] \\
$5s_{1/2}$ & -5.103\,103 & -5.103\,080 & 7.030\,344[-3]
& 7.033[-3] & -2.389\,679[-3] \\
$5p_{1/2}$ & -3.537\,946 & -3.537\,438 & 1.212\,951[-2]
& 1.213[-2] & 3.017\,200[-6] \\
$5p_{3/2}$ & -2.842\,014 & -2.841\,487 & 4.829\,281[-3]
& 4.828[-3] & 2.641\,881[-4] \\
$5d_{3/2}$ & -0.650\,063 & -0.649\,907 & 2.431\,914[-4]
& 2.394[-4] & 2.060\,225[-4] \\
$5d_{5/2}$ & -0.574\,649 & -0.574\,475 & -1.088\,398[-3]
& -1.093[-3] & 1.954\,800[-4] \\
$6s_{1/2}$ & -0.328\,036 & -0.327\,943 & 4.584\,067[-4]
& 4.575[-4] & -2.026\,796[-4] \\
\end{tabular}
\end{ruledtabular}
\end{table*}
\subsection{Hg}
The results of Hg deserve detailed discussions as the current work is
precursor to a refined recalculation of the Hg atomic EDM \cite{latha-09}.
Like in the previous cases, the orbital
energies of Hg and corrections are listed in Table. \ref{orb_hg}. From the
table it is evident that the values of $\Delta\epsilon_{\rm Br}$ from the
current work are in excellent agreement with the results reported
in Ref. \cite{lindroth-89a}. One noticeable change in the trend of
$\Delta\epsilon_{\rm Br}$ is the
negative values of $\Delta\epsilon_{{\rm Ue}(4f_{5/2})}$ and
$\Delta\epsilon_{{\rm Ue}(4f_{7/2})}$. In comparison,
$\Delta\epsilon_{\rm Br}$ is negative for $3d_{5/2}$ and $4d_{5/2}$ in
Zn and Cd, respectively. The results seem to indicate that the outermost
sub-shell with $j\geqslant5/2$ have negative $\Delta\epsilon_{\rm Br}$, which
could be on account of the larger weight factor $(2j+1)$ associated with higher
$j$ in the exchange two-electron integrals. The reason behind this remark is,
only the exchange integrals contribute to the $\Delta\epsilon_{\rm Br}$ in
closed-shell atoms and ions.
The Uehling potential corrections to the orbitals energies exhibit one
marked change compared to Zn and Cd. In Hg, the values of
$\Delta\epsilon_{{\rm Ue}(mp_{1/2})}$ with $ m = 2, 3, 4$ are negative. A
similar result was reported for the case of Ra$^{2+}$ in our previous work
on doubly ionized alkaline-earth-metal atoms \cite{chattopadhyay-13b}. There
is, however, one minor but important difference. In the case of Ra$^{2+}$ the
$\Delta\epsilon_{\rm Ue}$ is negative for all the $p_{1/2}$ orbitals. Whereas
in Hg, $5p_{1/2}$ orbital, the outermost $p_{1/2}$ orbital, has positive
$\Delta\epsilon_{\rm Ue}$. We attribute this to the larger relativistic
effects in Ra$^{2+}$ due to the stronger nuclear potential. Coming to the
basis set parameters, the values we have chosen generates orbitals with
energies within $ 10^{-4}-10^{-3}$ hartree of the numerical orbital energies.
The PRCC computations with excitations from all the core sub-shells of Hg
generate cluster amplitudes in excess of $10^7 $ when the basis size
is $\sim 160$. The computation of $\alpha$, then, requires thousands of
hours of compute time, and detailed studies on the convergence properties is
unfeasible (with our existing facilities). To mitigate this computational
conundrum we restrict the cluster amplitudes to excitations from the
$(4-6)s, (4-5)p, (4-5)d$, and $4f$ core sub-shells. From the results listed
in Table. \ref{conv_pat}, the $\alpha$ of Hg converges to 33.499 with a
basis size of 134 orbitals.
Among the previous theoretical results, three are based on coupled-cluster
theory, and we discuss these in some detail. Consider first the CCSD(T)
results of Kello and Sadlej \cite{kello-95}, it is obtained with
a polarized basis set, and correlating the $5d^{10}6s^{2}$ electrons.
So, it is effectively 12 electron coupled-cluster calculations with
relativistic corrections through the mass-velocity operator. Their result is
lower than ours, and below the experimental data as well. They also mention
that $\alpha$ decreases to 31.24 when the computations are done with larger
number of correlated electrons, namely, $5s^25p^65d^{10}6s^{2}$. So, the
primary reason for the difference may be the form of the
relativistic effects. The second result is based on the CCSD(T) work of Seth
and collaborators \cite{seth-97} using a basis set generated with an
optimized quasirelativistic pseudopotential \cite{schwerdtfeger-86}.
Their result is close to the experimental value, but on the higher side.
The estimate of the contributions from the triple excitation
is 0.84, which is smaller than the value 1.43 listed in the work of
Kello and Sadlej \cite{kello-95}. This indicates that the contribution from
the triple excitation depends on the nature of basis set and form of the
effective interaction to account for relativistic corrections. This is
perhaps not surprising as the electron correlation effects subsumed through
the cluster operators depend on the nature of the basis functions. The third
or the last previous work \cite{pershina-08} on $\alpha$ of Hg with CCSD(T)
is the closest, in terms of theoretical approach, to our present work. The
computations are based on the Dirac-Coulomb Hamiltonian, and their result
is within the experimental uncertainty. In summary, there is a variation
in the trend of the previous CCSD(T) results. The first \cite{kello-95}
and second \cite{seth-97} reports values which are below and above all the
experimental data, respectively. The result of the third work
\cite{pershina-08} is consistent with the experimental results. It must also
be mentioned that all of these three previous works are based on finite
field method.
In the present work, as mentioned earlier, we use the Dirac-Coulomb-Breit
atomic Hamiltonian. So, the Breit interaction is an additional relativistic
effect considered in the present work compared to the previous coupled-cluster
works. We must, however, add that there are other relativistic effects like
frequency dependent transverse photon interaction not included in the present
work. Our result of 33.50 is close, but below the experimental uncertainty of
the most recent work \cite{goebel-96a}. With the inclusion of perturbative
$\mathbf{T}_3^{(1)}$ we get 33.59, this improves the agreement with
experimental data. Among the previous works, the results
based on QCISD \cite{schwerdtfeger-94} and CI-MBPT \cite{hachisu-08} are in
very good agreement with our result. In the latter case an important point is,
the basis set is generated with $V^{N-1}$ potential. Whereas all the other
previous works and ours are with basis generated using $V^N$ potential.
Considering that the results from the recent works
\cite{seth-97,pershina-08,hachisu-08}, and the present work are with different
methods, the relative variance of the results ($\approx 0.6$\%) is low. This
demonstrates the methods do consolidate important relativistic and many-body
effects correctly. From this we can infer that the basis set, and
PRCC(T) theory used in the present work is well suited
for precision computation of properties like atomic electric dipole moment.
The term wise contribution, as listed in Table. \ref{cont_term},
Hg exhibits a noticeable change in the trend. The NLO contribution arises from
$\mathbf{T}_1^{(1)\dagger}\mathbf{D}T_2^{(0)} + \text{H.c.}$, where as
it is
$\mathbf{T}_1^{(1)\dagger}\mathbf{D}T_1^{(0)} + \text{H.c.}$ in Zn and Cd. We
attribute this to the electron-correlation effects associated with the
electrons in $5d$ shell, which enhances the cluster amplitude of
$T_2^{(0)}$. This is also reflected in the pattern of the core sub-shell
contribution to the LO term,
where there is a marked change in the trend compared to Zn and Cd. The
contribution from the valence shell, $6s_{1/2}$, is $\approx 40$\% smaller
than the valence sub-shell contribution in Cd. However, the contribution
from the next core sub-shell $5d_{5/2}$ is more than double of
$4d_{5/2}$ in Cd. This is on account of the relativistic contraction of
the $6s_{1/2}$ radial wavefunction.
For Hg, two experimental results are available in the literature. First
is based on the data of dielectric constant reported in Ref.
\cite{wusthoff-36}, and the other is based on the recent experimental
measurement of Goebel and Hohm \cite{goebel-96a}. The two results are in
very good agreement. There is another result \cite{tang-08} derived from the
experimental data of Ref. \cite{goebel-96a} using the three term expression
of Wolfsohn \cite{wolfsohn-33}. The reanalysis is in view of the findings
in Ref. \cite{salek-05} and \cite{gaston-06}, which report the need
for eight or more terms, compared to three in Ref. \cite{goebel-96a}, in
the Cauchy expansion of frequency dependent polarizability to obtain
converged moments.
\subsection{Uncertainty estimates}
We have identified different sources of uncertainties in the present work.
These arise from various approximations at different stages of the
RCC and PRCC computations. The first two sources of uncertainties are
associated with the truncation of the basis set, and consideration of cluster
operators up to $T_3^{(0)}$ in the RCC theory. These are, however, negligible
as we consider a basis set which gives converged results of $\alpha$.
The third source of uncertainty is the incomplete consideration of
$\mathbf{T}^{(1)}_3$ as we include it perturbatively. To estimate an upper
bound on this uncertainty, consider the case of Hg, where the
contribution from perturbative $\mathbf{T}^{(1)}_3$ is $\approx 0.3$\%, and is
the largest among the three atoms studied. Since the perturbative treatment
is considering the most dominant term, we can assume an uncertainty of
$\approx 0.3$\% as the upper bound arising from the remaining contributions
from $\mathbf{T}^{(1)}_3$. The fourth source of uncertainty is the
truncation in the expression of $\alpha$ in Eq. (\ref{exp_alpha}), in which
we retain terms up to second order in cluster operators. In one of our previous
works \cite{mani-10}, we have shown the contribution from the third and higher
order terms in cluster amplitudes is negligible. So, the uncertainty from
this can also be neglected. The last two sources of uncertainties are
associated with the frequency dependent Breit interaction, and violation of
no-virtual-pair approximation. In our previous work \cite{chattopadhyay-14a},
we had estimated the upper bound on the contribution from frequency dependent
Breit interaction to be 0.13\% for Ra. For the present work too, as Ra has
higher $Z$ than Hg, we consider this as the upper bound on the uncertainty
arising from frequency dependent Breit interaction. As the systems under study
are neutral atoms the contribution from the latter, violation of
no-virtual-pair approximation, is negligible. Combining these, we estimate
the uncertainty in the results of Zn and Cd to below 0.5\%. For Hg, an
additional source of uncertainty is the restriction of excitations from the
core sub-shells $(4-6)s$, $(4-5)p$, $(4-5)d$ and $4f$ in the converged
basis set. Based on the computations with smaller basis set, but with
excitations from all the core sub-shells, the upper bound on the uncertainty
of the Hg results is $1.0$\%.
\section{Conclusion}
We have computed the $\alpha$ of Zn, Cd and Hg, the elements of the
groupIIB, using PRCC and our results are in very good agreement with the
experimental data. Among the three elements, our result of Cd is of
significance as ours is the only theoretical result consistent with the
experimental data. Based on the analysis of available experimental data, we
conclude that $\alpha$ of Cd reported by Qiao and
collaborators \cite{qiao-12} is reliable. We attribute the lower values
reported in the previous theoretical works to the choice of basis set, and
the interplay of relativistic corrections with electron correlation effects.
This is in contrast to the case of Zn and Hg, where the electron
correlation, and relativistic corrections are predominant effects,
respectively.
In the PRCC sector, we have considered the triple excitation cluster operator
through the dominant contribution from the perturbative $\mathbf{T}_3^{(1)}$,
and included it in the computation of $\alpha$. This brings the level of
electron correlation effects, in terms of excited state, in PRCC theory
on par with the RCCSDT theory we have developed and used.
The present work is based on use of Dirac-Coulomb-Breit atomic Hamiltonian.
In addition, we also consider the corrections from the Uehling potential,
the leading order term in the vacuum polarization effects. So, we incorporate
relativistic effects, albeit incomplete, better than the previous theoretical
works. The relativistic effects left out in the present work include
self-energy corrections, frequency dependent transverse photon interaction
and Wichmann-Kroll potential. We shall examine these in detail in future
works, and may be essential to reduce the uncertainties to below 0.5\% in the
properties calculations of high $Z$ elements like Hg.
An important highlight associated with an integral part of the
Hamiltonian we use, Breit interaction, is the orbital energy correction
associated with it. Our results are in excellent agreement with the previous
results we could find in the literature, that is for Hg. This, we consider,
as a reliable validation of our implementation of Breit interactions.
In future works, we shall report the application of PRCC theory to
one- and two-valence systems. For which we have reported the results with
unperturbed RCCSD theory \cite{mani-10,mani-11}.
\begin{acknowledgements}
We thank Arko Roy and Kuldeep Suthar for useful discussions. The
results presented in the paper are based on the computations using the 3TFLOP
HPC Cluster at Physical Research Laboratory, Ahmedabad.
\end{acknowledgements}
|
\section{Introduction}
\label{Section-Introduction}
The aim of this article is to prove the following result.
\begin{theorem}[Theorem \ref{algebraic-family-hyperelliptic-curve-hp-theorem}]
\label{Theorem-The-main-theorem}
Let $n > 5$ be an integer such that $n \not\equiv 0 \pmod{4}$. Then there is an algebraic family $\cC_t$ of hyperelliptic curves of genus $n$ such that $\cC_t$ is a counterexample to the Hasse principle explained by the Brauer-Manin obstruction for all $t \in \bQ$. Furthermore, $\cC_t$ contains no zero-cycles of odd degree over $\bQ$ for every $t \in \bQ$.
\end{theorem}
Recall from \cite{poonen1} that an \textit{algebraic} family of curves is a family of curves depending on a parameter $T$ such that substituting any rational number for $T$ results in a smooth curve over $\bQ$.
A smooth geometrically irreducible curve $\cC$ over $\bQ$ is said to \textit{satisfy the Hasse principle} if the everywhere local solvability of $\cC$ is equivalent to the global solvability of $\cC$. In more concrete terms, this means that
\begin{align*}
\cC(\bQ) \ne \emptyset \; \; \text{if and only if} \; \; \cC(\bQ_p) \ne \emptyset \; \; \text{for every prime $p$ including $p = \infty$}.
\end{align*}
If $\cC$ has points locally everywhere but has no rational points, we say that $\cC$ is a \textit{counterexample to the Hasse principle}. Furthermore, if we also have $\cC(\bA_{\bQ})^{\Br} = \emptyset$, we say that $\cC$ is a \textit{counterexample to the Hasse principle explained by the Brauer-Manin obstruction}. The Hasse principle fails in general. The first counterexamples of genus one curves to the Hasse principle were discovered by Lind \cite{lind} in $1940$ and independently shortly thereafter by Reichardt \cite{reichardt}. For a basic introduction to the Brauer-Manin obstruction, see \cite{poonen3} and \cite{skorobogatov}.
Let us relate Theorem \ref{Theorem-The-main-theorem} to existing results in literature. For $n = 1$, Colliot-Th\'el\`ene and Poonen \cite{colliot-thelene-poonen} showed how to produce one parameter families of curves of genus one violating the Hasse principle. Poonen \cite{poonen1} explicitly constructed an algebraic family of genus one cubic curves violating the Hasse principle using the general method developed in \cite{colliot-thelene-poonen}. It is not known that there exists an algebraic family of curves of genus $n$ violating the Hasse principle for all $n \ge 2$.
Here, as throughout the article, we say that a smooth geometrically irreducible variety $\cV$ over $\bQ$ is said to \textit{satisfy HP1} if it is a counterexample to the Hasse principle explained by the Brauer-Manin obstruction, and that a smooth geometrically irreducible variety $\cV$ over $\bQ$ is said to \textit{satisfy HP2} if it contains no zero-cycles of odd degree over $\bQ$.
Coray and Manoil \cite{coray-manoil} showed that for each positive integer $n \ge 2$, the smooth projective model of the affine curve defined by
\begin{align}
\label{the-Coray-Manoil-smooth-model-equation}
z^2 = 605\cdot 10^6x^{2n + 2} + (18x^2 - 4400)(45x^2- 8800)
\end{align}
satisfies HP1 and HP2. The Coray-Manoil family of curves is the first family of hyperelliptic curves of varying genus that satisfies HP1 and HP2. Although Coray and Manoil restricted themselves to constructing only one hyperelliptic curve of genus $n$ satisfying HP1 and HP2 for each integer $n \ge 2$, it seems plausible that their approach can be modified to produce algebraic families of hyperelliptic curves of arbitrary genus satisfying HP1 and HP2. Since we will follow the approach of Coray and Manoil with some modifications to prove Theorem \ref{Theorem-The-main-theorem}, we briefly recall their main ideas of constructing the family $(\ref{the-Coray-Manoil-smooth-model-equation})$.
Colliot-Th\'el\`ene, Coray and Sansuc \cite{colliot-thelene-coray-sansuc} proved that the threefold $\cY_{(5, 1, 1)}$ in $\bP^5_{\bQ}$ defined by
\begin{align*}
\cY_{(5, 1, 1)} :
\begin{cases}
u_1^2 - 5v_1^2 &= 2xy \\
u_2^2 - 5v_2^2 &= 2(x + 20y)(x + 25y)
\end{cases}
\end{align*}
satisfies HP1 and HP2. Building on this result, Coray and Manoil \cite{coray-manoil} introduced a geometric construction of hyperelliptic curves that allows to smoothly embed the family of curves defined by $(\ref{the-Coray-Manoil-smooth-model-equation})$ into the threefold $\cY_{(5, 1, 1)}$. It follows immediately from functoriality that the Coray-Manoil family of curves satisfies HP1 and HP2.
In order to generalize the result of Coray and Manoil, we first construct a family of threefolds in $\bP^5_{\bQ}$ that satisfies HP1 and HP2 has the threefold $\cY_{(5, 1, 1)}$ as its member. The construction of such threefolds is achieved building on that of the threefold $\cY_{(5, 1, 1)}$. In order to show that the Brauer-Manin obstruction for these threefolds is non-empty, we also need to show the existence of infinitely many primes $p$ and $q$ satisfying certain quadratic equations. We do this by calling on the result of Iwaniec \cite{iwaniec} that a quadratic polynomial in two variables represents infinitely many primes. Since the existence of certain threefolds in $\bP^5_{\bQ}$ satisfying HP1 and HP2 is of interest in its own right, we state this result here.
\begin{theorem}
\label{Theorem-The-second-main-theorem}
Let $p$ be a prime such that $p \equiv 5 \pmod{8}$ and $3$ is quadratic non-residue in $\bF_p^{\times}$. Then there exist infinitely many couples $(b, d) \in \bZ^2$ such that any smooth and proper $\bQ$-model $\cZ$ of the smooth $\bQ$-variety $\cX$ in $\bA^5_{\bQ}$ defined by
\begin{align*}
\begin{cases}
0 \ne u_1^2 - pv_1^2 &= 2x \\
0 \ne u_2^2 - pv_2^2 &= 2(x + 4pb^2)(x + p^2d^2)
\end{cases}
\end{align*}
satisfies HP1 and HP2.
\end{theorem}
The next step is to choose a family of hyperelliptic curves of arbitrary genus that can be smoothly embed into the family of threefolds in Theorem \ref{Theorem-The-second-main-theorem} using the geometric construction of Coray and Manoil. For each $n \ge 2$, we define a family of hyperelliptic curves of genus $n$ of the shape
\begin{align}
\label{Equation-The-equation-of-hyperelliptic-curves-in-the-introduction}
z^2 = p\alpha^2Q^2 x^{2n + 2} + (2b^2Px^2 + \beta Q)(d^2pPx^2 + 2\beta Q),
\end{align}
where $\alpha, \beta, \gamma$ are certain rational numbers, and $P, Q$ depend on $\alpha, \beta, \gamma, p, b, d$. In order to apply the geometric construction of hyperelliptic curves of Coray and Manoil, the polynomials on the right-hand side of $(\ref{Equation-The-equation-of-hyperelliptic-curves-in-the-introduction})$ are required to be separable.
In order to smoothly embed these hyperelliptic curves into the threefolds in Theorem \ref{Theorem-The-second-main-theorem}, we impose certain conditions on $\alpha, \beta, \gamma$ such that these rational numbers satisfy certain local congruences and certain conics in $\bP^2_{\bQ}$ constructed from sextuples $(p, b, d, \alpha, \beta, \gamma)$ possess at least one non-trivial rational point. Lemma \ref{infinitude-alpha-beta-gamma-lemma} and Lemma \ref{Lemma-the-infinitude-of-p-b-d-alpha-beta-gamma-satisfying-A1-A5-and-B1} show that there are infinitely many sextuples $(p, b, d, \alpha, \beta, \gamma)$ satisfying these conditions. For any such sextuple $(p, b, d, \alpha, \beta, \gamma)$, it follows from functoriality and Theorem \ref{Theorem-The-second-main-theorem} that the family of hyperelliptic curves of genus $n$ defined by $(\ref{Equation-The-equation-of-hyperelliptic-curves-in-the-introduction})$ satisfies HP1 and HP2 for each $n \ge 2$.
In the last step, the main difficulty is to show the existence of rational functions in $\bQ(T)$ that parameterize rational numbers $\alpha, \beta, \gamma$ such that for each integer $n$, substituting any rational number for $T$ in the polynomials on the right-hand side of $(\ref{Equation-The-equation-of-hyperelliptic-curves-in-the-introduction})$ results in a separable polynomial of degree $2n + 2$ over $\bQ$. We do this by calling on a \textit{separability criterion} of the author \cite{dq-separable-polynomials} that will be reviewed in Section \ref{algebraic-families-hyperelliptic-curve-hp-section}.
After this article was finished, the author learned that Bhargava, Gross and Wang \cite{bhargava-gross-wang} showed that for any integer $n \ge 1$, there is a positive proportion of everywhere locally solvable hyperelliptic curves over $\bQ$ of genus $n$ that have no points over any number field of odd degree over $\bQ$. The main theorem of this article describes an explicit algebraic family of such curves of genus $n$ with $\gcd(n, 4) = 1$ and $n > 5$.
\section{The Hasse principle for certain threefolds in $\bP^5_{\bQ}$}
\label{threefold-hp-section}
In this section, we will construct families of threefolds satisfying HP1 and HP2. We begin by stating some lemmas that we will need in the proof of the main results throughout the paper.
\begin{lemma}$(\text{\cite[Lemma 4.8]{coray-manoil}})$
\label{functoriality-azumaya-lemma}
Let $k$ be a number field, and let $\cV_1$ and $\cV_2$ be (proper) $k$-varieties. Assume that there is a $k$-morphism $\alpha : \cV_1 \rightarrow \cV_2$ and $\cV_2(\bA_k)^{\mathrm{Br}} = \emptyset$. Then $\cV_1(\bA_k)^{\mathrm{Br}} = \emptyset$.
\end{lemma}
\begin{lemma}$(\text{\cite[Proposition 6.4]{corn-PLMS}})$
\label{bright-corn-lemma}
Let $\cX$ be a smooth $F$-variety. Let $L/F$ be a cyclic extension, and let $F(\cX)$ be the function field of $\cX$. Let $f$ be an element of $F(\cX)$, and let $\cX_L = \cX \times_F L$. Then the class of the cyclic algebra $\left(L/F, f\right) \in \Br(F(\cX))$ lies in the image of the inclusion $\Br(\cX) \hookrightarrow \Br(F(\cX))$ if and only if $\mathrm{div}(f) = \mathrm{Norm}_{L/F}(D)$ for some $D \in \mathrm{Div}(\cX_L)$.
\end{lemma}
\begin{remark}
\label{Remark-Sketching-a-proof-the-Bright-Corn-lemma-in-the-appendix}
We will sketch a proof of Lemma \ref{bright-corn-lemma} in Subsection \ref{Subsection-Proof-of-the-Bright-Corn-Lemma}.
\end{remark}
\begin{lemma}$(\text{Lang-Nishimura, \cite[Lemme 3.1.1, p. 164]{colliot-thelene-coray-sansuc}})$
\label{lang-nishimura-lemma}
Let $F$ be a field, and let $\cX$ be an integral $F$-variety. Let $\cY$ be a proper $F$-variety, and let $f : \cX \longrightarrow \cY$ be an $F$-rational map. If $\cX(F)$ contains a regular $F$-point, then $\cY(F)$ is non-empty. In particular, the condition $\cX(F) \ne \emptyset$ is an $F$-birational invariant in the category of smooth, proper and integral $F$-varieties $\cX$.
\end{lemma}
The following result describes how to construct certain Azumaya algebras on certain threefolds.
\begin{lemma}
\label{azumaya-threefold-lemma}
Let $p$ be a prime such that $p \equiv 5 \pmod{8}$. Assume that the following are true.
\begin{itemize}
\item [(A1)] $3$ is a quadratic non-residue in $\bF_p^{\times}$.
\item [(B)] there exists a couple $(b, c) \in \bZ^2$ such that $\gcd(b, c) = 1$, $b \not\equiv 0 \pmod{p}$ and $q := |pc - 4b^2|$ is either $1$ or an odd power of an odd prime. Here $|\cdot|$ denotes the absolute value in $\bQ$. Furthermore, if $b \equiv 0 \pmod{3}$, then $c \equiv 2 \pmod{3}$.
\end{itemize}
Let $\cV$ be a smooth, proper $\bQ$-model of the smooth $\bQ$-variety $\cU$ in $\bA^5_{\bQ}$ defined by
\begin{equation}
\label{eqn-U}
\cU:
\begin{cases}
0 \ne u_1^2 - pv_1^2 = 2x \\
0 \ne u_2^2 - pv_2^2 = 2(x + 4pb^2)(x + p^2c).
\end{cases}
\end{equation}
Let $\bQ(\cV)$ be the function field of $\cV$, and let $\cA$ be the class of the quaternion algebra $\left(p, x + 4pb^2\right)$. Then $\cA$ is
an Azumaya algebra of $\cV$, that is, $\cA$ belongs to the subgroup $\Br(\cV)$ of $\Br(\bQ(\cV))$.
\end{lemma}
\begin{proof}
Let $K = \bQ(\sqrt{p})$, and let $\Gamma$ be the divisor defined over $\bQ(\sqrt{p})$ and lying on $\cV$ defined by
\begin{align*}
\Gamma : f := x + 4pb^2 = 0, \; u_2 - \sqrt{p}v_2 = 0, \; u_1^2 - pv_1^2 = -8pb^2.
\end{align*}
Let $\sigma$ be a generator of $\text{Gal}(K/\bQ)$. We see that
\begin{align*}
\sigma\Gamma : x + 4pb^2 = 0, \; u_2 + \sqrt{p}v_2 = 0, \; u_1^2 - pv_1^2 = -8pb^2.
\end{align*}
Hence it follows that
\begin{align*}
\mathrm{div}(f) = \Gamma + \sigma \Gamma.
\end{align*}
By Lemma \ref{bright-corn-lemma}, we deduce that $\cA$ is in the image of $\Br(\cV) \hookrightarrow \Br(\bQ(\cV))$.
\end{proof}
\begin{lemma}
\label{existence-of-a}
Let $p$ be a prime such that $p \equiv 5 \pmod{8}$. Assume that conditions $(A1)$ and $(B)$ in Lemma \ref{azumaya-threefold-lemma} are true. Then there exists a non-zero integer $a$ such that
\begin{equation}
\label{gcd-a-b-c}
\gcd\left((a^2 + 2pb^2)(2a^2 + p^2c), 3(2b^2 + pc)\right) = 1.
\end{equation}
\end{lemma}
\begin{proof}
Assume that $H_1 := 2b^2 + pc = \prod_{i = 1}^{m}l_i^{\alpha_i}$, where $l_i$ are distinct primes and $\alpha_i \in \bZ_{>0}$. Note that since $q = |pc - 4b^2|$ is either $1$ or an odd power of an odd prime, $c$ is odd. Thus $H_1$ is odd, and therefore $l_i \ne 2$ for each $1 \le i \le m$. We also have that $l_i \ne p$ for each $1 \le i \le m$; otherwise, $l_i = p$ for some integer $1 \le i \le m$. Since $2b^2 + pc \equiv 0 \pmod{l_i}$ and $l_i = p$, it follows that $b \equiv 0 \pmod{p}$, which is a contradiction. We consider the following cases.
$\star$ \textit{Case 1. $b \equiv 0 \pmod{3}$ and $c \not\equiv 0 \pmod{3}$.}
By assumption $(B)$, one knows that $c \equiv 2 \pmod{3}$. Define $a := \prod_{i = 1}^{m}l_i$. We contend that $a$ satisfies $(\ref{gcd-a-b-c})$. Indeed, we have that $l_i \ne 3$ for each $1 \le i \le m$; otherwise, $l_i = 3$ for some integer $1 \le i \le m$. Since $b \equiv 0 \pmod{3}$ and $p \ne 3$, it follows that $c \equiv 0 \pmod{3}$, which is a contradiction.
Let $H_2 := a^2 + 2pb^2$ and $H_3 := 2a^2 + p^2c$. We see that $a^2 = \prod_{i = 1}^{m}l_i^2 \equiv 1 \pmod{3}$. Since $p \ne 3$, we deduce that
$H_2 \equiv 1 \pmod{3}$ and $H_3 \equiv 2 + c \equiv 1 \pmod{3}$, and thus $H_2H_3 \equiv 1 \pmod{3}$.
Suppose that $l_j$ divides $H_2$ for some integer $1 \le j \le m$. Since $a = \prod_{i = 1}^{m}l_i \equiv 0 \pmod{l_j}$, it follows that
$b \equiv 0 \pmod{l_j}$. Thus $c \equiv 0 \pmod{l_j}$, which is a contradiction to $(B)$.
Suppose that $l_j$ divides $H_3$ for some integer $1 \le j \le m$. Since $a = \prod_{i = 1}^{m}l_i \equiv 0 \pmod{l_j}$ and $l_j \ne p$, it follows that
$c \equiv 0 \pmod{l_j}$. Hence $b \equiv 0 \pmod{l_j}$, which is a contradiction to $(B)$. Therefore, in any event, $(\ref{gcd-a-b-c})$ holds.
$\star$ \textit{Case 2. $b \not\equiv 0 \pmod{3}$ and $c \equiv 0 \pmod{3}$.}
Let $a := \prod_{i = 1}^{m}l_i$. By $(A1)$, we know that $p \equiv 2 \pmod{3}$. Hence repeating in the same manner as in \textit{Case 1}, we deduce that $(\ref{gcd-a-b-c})$ holds.
$\star$ \textit{Case 3. $b \not\equiv 0 \pmod{3}$ and $c \not\equiv 0 \pmod{3}$.}
Let $a := 3\prod_{i = 1}^{m}l_i$. The same arguments as in \textit{Case 1} show that $(\ref{gcd-a-b-c})$ holds.
\end{proof}
Following the techniques in the proof of Proposition 7.1 in \cite{colliot-thelene-coray-sansuc}, we now prove the main theorem in this section.
\begin{theorem}
\label{general-threefold-hp-theorem}
We maintain the same notation as in Lemma \ref{azumaya-threefold-lemma}. Let $p$ be a prime such that $p \equiv 5 \pmod{8}$. Assume further that $(A1)$ and $(B)$ are true. Let $\cU$ and $\cV$ be the $\bQ$-varieties defined by $(\ref{eqn-U})$ in Lemma \ref{azumaya-threefold-lemma}. Let $\cT$ be the singular $\bQ$-variety in $\bP^5_{\bQ}$ defined by
\begin{equation}
\label{eqn-T}
\cT:
\begin{cases}
u_1^2 - pv_1^2 &= 2xy \\
u_2^2 - pv_2^2 &= 2(x + 4pb^2y)(x + p^2cy).
\end{cases}
\end{equation}
Then $\cU$, $\cV$ and $\cT$ satisfy HP1 and HP2.
\end{theorem}
\begin{proof}
$\bullet$ \textit{Step 1. $\cU(\bQ) = \cT(\bQ)$.}
It is clear that $\cU(\bQ) \subseteq \cT(\bQ)$. Hence it suffices to prove that
\begin{align*}
\cT(\bQ) \subseteq \cU(\bQ).
\end{align*}
Assume that there is a point $P := (x : y : u_1 : v_1 : u_2 : v_2) \in \cT(\bQ)$. Suppose first that $y = 0$. It follows from $(\ref{eqn-T})$ that
\begin{equation}
\label{y=0-prove-U-T-equal}
\begin{cases}
u_1^2 - pv_1^2 = 0 \\
u_2^2 - pv_2^2 = 2x^2.
\end{cases}
\end{equation}
We see from the first equation of $(\ref{y=0-prove-U-T-equal})$ that $u_1 = v_1 = 0$. If $x = 0$, then the second equation of $(\ref{y=0-prove-U-T-equal})$ implies that
$u_2 = v_2 = 0$, which is a contradiction. Hence $x \ne 0$, and thus $2 = \left(\dfrac{u_2}{x}\right)^2 - p\left(\dfrac{v_2}{x}\right)^2$. Hence $2$ is the norm of an element in $\bQ(\sqrt{p})^{\times}$, and therefore $2$ is the norm of an element in $\bQ_p(\sqrt{p})^{\times}$. Hence it follows that the local Hilbert symbol $(2, p)_p$ is $1$. On the other hand, using Theorem $5.2.7$ in \cite[page 296]{cohen} and since $p \equiv 5 \pmod{8}$, we deduce that
\begin{align*}
(2, p)_p = \left(\dfrac{2}{p}\right) = -1,
\end{align*}
which is a contradiction.
Now we assume that $y \ne 0$, and with no loss of generality, assume further that $y = 1$. It follows from $(\ref{eqn-T})$ that
\begin{equation}
\label{y=1-prove-U-T-equal}
\begin{cases}
u_1^2 - pv_1^2 = 2x \\
u_2^2 - pv_2^2 = 2(x + 4pb^2)(x + p^2c).
\end{cases}
\end{equation}
We consider the following cases.
$\star$ \textbf{Case 1.} $x = 0$.
The second equation of $(\ref{y=1-prove-U-T-equal})$ implies that $u_2^2 - pv_2^2 = 8p^3b^2c$. Thus $8p^3b^2c$ is the norm of an element in $\bQ_2(\sqrt{p})^{\times}$, and hence the local Hilbert symbol $(8p^3b^2c, p)_2$ is $1$. Since $q = |pc - 4b^2|$ is either $1$ or an odd power of an odd prime, $c$ is odd. Hence $v_2(8p^3b^2c) = 3 + 2v_2(b)$ which is an odd integer. Using Theorem $5.2.7$ in \cite[page 296]{cohen}, we deduce that
\begin{align*}
(8p^3b^2c, p)_2 = \left(\dfrac{p}{2}\right) = -1,
\end{align*}
which is a contradiction.
$\star$ \textbf{Case 2.} $x = -4pb^2$.
It follows from $(\ref{eqn-T})$ that
\begin{equation*}
u_1^2 - pv_1^2 = -8pb^2.
\end{equation*}
Using the same arguments as in \textbf{Case 1}, we deduce that $-8pb^2$ is not the norm of any element in $\bQ_2(\sqrt{p})^{\times}$, which is a contradiction to the last identity.
$\star$ \textbf{Case 3.} $x = -p^2c$.
It follows from $(\ref{eqn-T})$ that
\begin{equation*}
u_1^2 - pv_1^2 = -2p^2c.
\end{equation*}
Using the same arguments as in \textbf{Case 1}, we deduce that $-2p^2c$ is not the norm of any element in $\bQ_2(\sqrt{p})^{\times}$, which is a contradiction to the last identity.
Therefore, in any event, we have shown that if the point $P := (x : y : u_1 : v_1 : u_2 : v_2)$ belongs to $\cT(\bQ)$, then $y = 1$, $x \ne 0$, $x + 4pb^2 \ne 0$ and $x + p^2c \ne 0$. In other words, the point $P$ satisfies
\begin{equation*}
\begin{cases}
0 \ne u_1^2 - pv_1^2 = 2x \\
0 \ne u_2^2 - pv_2^2 = 2(x + 4pb^2)(x + p^2c),
\end{cases}
\end{equation*}
and thus $P \in \cU(\bQ)$. Therefore $\cU(\bQ) = \cT(\bQ)$.
$\bullet$ \textit{Step 2. $\cU, \cV$ and $\cT$ are everywhere locally solvable.}
We now prove that $\cU, \cV$ and $\cT$ are everywhere locally solvable. By Lemma \ref{lang-nishimura-lemma}, it suffices to prove that $\cU$ is everywhere locally solvable. Recall that by Lemma \ref{existence-of-a}, there is a non-zero integer $a$ such that
\begin{equation*}
\gcd\left((a^2 + 2pb^2)(2a^2 + p^2c), 3(2b^2 + pc)\right) = 1.
\end{equation*}
Hence it suffices to consider the following cases.
$\star$ \textbf{Case I.} $l$ is a prime such that $l \ne p$ and $\gcd(l, (a^2 + 2pb^2)(2a^2 + p^2c)) = 1$.
Let $x = 2a^2$. Since $2x = 4a^2$ is a square in $\bZ$, we see that the local Hilbert symbol $(2x, p)_l$ satisfies
\begin{align*}
(2x, p)_l = (4a^2, p)_l = 1.
\end{align*}
Thus $2x$ is the norm of an element in $\bQ_l(\sqrt{p})^{\times}$. We see that
\begin{equation*}
v_l\left(2(x + 4pb^2)(x + p^2c)\right) = v_l\left(4(a^2 + 2pb^2)(2a^2 + p^2c)\right) = 2v_l(2) + v_l((a^2 + 2pb^2)(2a^2 + p^2c)) = 2v_l(2).
\end{equation*}
Hence using Theorem $5.2.7$ in \cite[page 296]{cohen}, we deduce that the local Hilbert symbol $(2(x + 4pb^2)(x + p^2c), p)_l$ satisfies
\begin{align*}
(2(x + 4pb^2)(x + p^2c), p)_l = 1
\end{align*}
Thus $2(x + 4pb^2)(x + p^2c)$ is the norm of an element in $\bQ_l(\sqrt{p})^{\times}$. Therefore $\cU$ is locally solvable at $l$.
$\star$ \textbf{Case II. } $l$ is a prime such that $\gcd(l,3(2b^2 + pc)) = 1$. Note that $p$ is among these primes.
Assume first that $l = p$, and set $x = 2pb^2$. We see that
\begin{align*}
2x &= p(2b)^2, \\
2(x + 4pb^2)(x + p^2c) &= p^2(12b^2)(2b^2 + pc).
\end{align*}
Note that $(2b)^2 \not\equiv 0 \pmod{p}$ and $(12b^2)(2b^2 + pc) \equiv 6(2b^2)^2 \not\equiv 0 \pmod{p}$. Hence using Theorem $5.2.7$ in \cite[page 296]{cohen}, we deduce that the local Hilbert symbol $(2x, p)_p$ satisfies
\begin{align*}
(2x, p)_p = (-1)^{(p - 1)/2}\left(\dfrac{(2b)^2}{p}\right) = 1.
\end{align*}
Hence $2x$ is the norm of an element in $\bQ_p\left(\sqrt{p}\right)$.
By $(A1)$, we know that $6$ is quadratic residue in $\bF_p^{\times}$. Since $(12b^2)(2b^2 + pc) \equiv 6(2b^2)^2 \pmod{p}$, we see that $(12b^2)(2b^2 + pc)$ is a quadratic residue in $\bF_p^{\times}$. Thus using the same arguments as above, we deduce that
\begin{align*}
\left(2(x + 4pb^2)(x + p^2c), p\right)_p = \left(p^2(12b^2)(2b^2 + pc), p\right)_p = 1.
\end{align*}
Therefore $2(x + 4pb^2)(x + p^2c)$ is the norm of an element in $\bQ_p\left(\sqrt{p}\right)$. Hence $\cU$ is locally solvable at $p$.
Suppose that $l \ne p$, and set $x = 2pb^2$. We see that
\begin{align*}
v_l(2x) &= v_l(4pb^2) = v_l(p) + 2v_l(2b) = 2v_l(2b), \\
v_l(2(x + 4pb^2)(x + p^2c)) &= v_l(p^2(12b^2)(2b^2 + pc)) = 2v_l(2b) + v_l(3(2b^2 + pc)) = 2v_l(2b).
\end{align*}
Hence using the same arguments as in \textbf{Case I}, we deduce that $\cU$ is locally solvable at $l$.
Therefore it follows from \textbf{Case I} and \textbf{Case II} that $\cU$ is everywhere locally solvable, and thus $\cU, \cV$ and $\cT$ are everywhere locally solvable.
$\bullet$ \textit{Step 3. $\cV$ satisfies HP1.}
We will prove that $\cV(\mathbb{A}_{\bQ})^{\Br} = \emptyset$. Let $\bQ(\cV)$ be the function field of $\cV$, and let $\cA$ be the class of quaternion algebra $(p, x + 4pb^2)$ in $\Br(\bQ(\cV))$. It follows from Lemma \ref{azumaya-threefold-lemma} that $\cA$ is an Azumaya algebra of $\cV$. We will prove that for any $P_l \in \cV(\bQ_l)$,
\begin{equation}
\label{inv-value}
\text{inv}_l(\cA(P_l)) =
\begin{cases}
0 & \text{if $l \ne 2$,}
\\
1/2 & \text{if $l = 2$.}
\end{cases}
\end{equation}
Since $\cV$ is smooth, we know that $\cU(\bQ_l)$ is $l$-adically dense in $\cV(\bQ_l)$. It is well-known \cite[Lemma 3.2]{viray} that $\text{inv}_l(\cA(P_l))$ is a continuous function on $\cV(\bQ_l)$ with the $l$-adic topology. Hence it suffices to prove $(\ref{inv-value})$ for $P_l \in \cU(\bQ_l)$.
Suppose that $l = \infty$ or $l$ is an odd prime such that $l \ne p$ and $p$ is a square in $\bQ_l^{\times}$. We see that $p \in \bQ_l^{2, \times}$, and hence the local Hilbert symbol $(p, t)_l$ is $1$ for any $t \in \bQ_l^{\times}$. Thus $\text{inv}_l(\cA(P_l))$ is $0$.
Suppose that $l$ is an odd prime such that $l \ne p$ and $p$ is not a square in $\bQ_l^{\times}$. Let $P_l \in \cU(\bQ_l)$, and let $x = x(P_l)$. It follows from equations $(\ref{eqn-U})$ and Theorem $5.2.7$ in \cite[page 296]{cohen} that $v_l(x)$ and $v_l((x + 4pb^2)(x + p^2c))$ are even, and hence the sum $v_l(x + 4pb^2) + v_l(x + p^2c)$ is even. Assume first that $v_l(x) < 0$. We deduce that $v_l(x + 4pb^2) = v_l(x)$, and hence it is even. Suppose now that $v_l(x) \ge 0$. We then see that $v_l(x + 4pb^2) \ge 0$ and $v_l(x + p^2c) \ge 0$. We contend that at least one of the last two numbers is zero. Otherwise, since $x \in \bZ_l$, one sees that $x + 4pb^2 \equiv 0 \pmod{l}$ and $x + p^2c \equiv 0 \pmod{l}$. Hence $l$ divides $p(pc - 4b^2)$, and thus by condition $(B)$, we deduce that $l$ divides $pq$. If $q$ is $1$, then $l = p$, which is a contradiction. Hence with no loss of generality, we might assume that $q$ is an odd power of an odd prime, say $q_1^{2m + 1}$ for some odd prime $q_1$ and $m \in \bZ_{\ge 0}$. It follows that $l = q_1$. By condition $(B)$, we know that $q = q_1^{2m + 1} \equiv \pm 4b^2 \not\equiv 0 \pmod{p}$. Hence
\begin{align*}
l = q_1 \equiv \pm \left(\dfrac{2b}{q_1^m}\right)^2 \pmod{p}.
\end{align*}
Since $-1$ is a square in $\bF_p^{\times}$, it follows from the congruence above that $l$ is a square in $\bF_p^{\times}$. By the quadratic reciprocity law, $p$ is a square in $\bQ_l^{\times}$, which is a contradiction. Since the sum $v_l(x + 4pb^2) + v_l(x + p^2c)$ is even and at least one of the two summands is even, we deduce that each of them is even. Hence using Theorem $5.2.7$ in \cite[page 296]{cohen}, we deduce that the local Hilbert symbol $(p, x + 4pb^2)_l$ is $1$. Therefore $\text{inv}_l(\cA(P_l))$ is $0$.
Suppose that $l = p$. Let $P_p \in \cU(\bQ_p)$ and $x = x(P_p)$. Since the local Hilbert symbol $(p, 2)_p$ is $-1$, we deduce from $(\ref{eqn-U})$ and Theorem $5.2.7$ in \cite[page 296]{cohen} that
\begin{align}
\label{congruence-relation}
\begin{cases}
x = p^n\alpha, \;\text{with} \;n \in \bZ, \alpha \in \bZ_p^{\times} \; \text{and} \; \left(\dfrac{\alpha}{p}\right) = -1, \\
(x + 4pb^2)(x + p^2c) = p^m\beta \; \text{with} \; m \in \bZ, \beta \in \bZ_p^{\times} \; \text{and} \; \left(\dfrac{\beta}{p}\right) = -1.
\end{cases}
\end{align}
Assume that $n \le 0$. We see that $p^{-n}x \equiv \alpha \pmod{p}$. Hence $p^{-n}(x + 4pb^2) \equiv \alpha \pmod{p}$ and
$p^{-n}(x + p^2c) \equiv \alpha \pmod{p}$. Thus the product of the two last congruences contradicts the second equation of $(\ref{congruence-relation})$. Hence with no loss of generality, we might assume that $n \ge 1$. Assume first that $n = 1$. We deduce that $p^{-1}x \equiv \alpha \pmod{p}$, and hence $p^{-1}(x + p^2c) = p^{-1}x + pc \equiv \alpha \pmod{p}$.
Thus by $(\ref{congruence-relation})$, there exists an integer $k \in \bZ$ such that $p^{k}(x + 4pb^2) \equiv \beta \alpha^{-1} \pmod{p}$. We see that $\left(\dfrac{\beta \alpha^{-1}}{p}\right) = 1$. Hence using Theorem $5.2.7$ in \cite[page 296]{cohen}, we deduce that the local Hilbert symbol $\left(p, x + 4pb^2\right)_p$ satisfies
\begin{align*}
\left(p, x + 4pb^2\right)_p = \left(\dfrac{\beta \alpha^{-1}}{p}\right) = 1.
\end{align*}
Therefore $\text{inv}_p(\cA(P_p))$ is $0$.
Suppose now that $n \ge 2$. We see that
\begin{equation*}
p^{-1}(x + 4pb^2) = p^{n - 1}\alpha + 4b^2 \equiv 4b^2 \pmod{p}.
\end{equation*}
Hence using the same arguments as above, we deduce that the local Hilbert symbol $\left(p, x + 4pb^2\right)_p$ is $1$, and thus $\text{inv}_p(\cA(P_p))$ equals $0$.
Therefore, in any event, we see that $\text{inv}_p(\cA(P_p)) = 0$.
Suppose that $l = 2$. Let $P_2 \in \cU(\bQ_2)$, and let $x = x(P_2)$. Since the local Hilbert symbol $(p, 2)_2$ satisfies
\begin{align*}
(p, 2)_2 = \left(\dfrac{p}{2}\right) = -1,
\end{align*}
we deduce from $(\ref{eqn-U})$ and Theorem $5.2.7$ in \cite[page 296]{cohen} that
\begin{align*}
(p, x)_2 = \left(p, (x + 4pb^2)(x + p^2c)\right) = -1.
\end{align*}
Hence $v_2(x)$ and $v_2((x + 4pb^2)(x + p^2c))$ are odd, and thus the sum $v_2(x + 4pb^2) + v_2(x + p^2c)$ is odd. We contend that $v_2(x) \ge 0$. Otherwise, we deduce that
\begin{align*}
v_2(x + 4pb^2) + v_2(x + p^2c) = 2v_2(x),
\end{align*}
which is a contradiction since the left-hand side is odd whereas the right-hand side is even. Since $v_2(x)$ is odd and $v_2(x) \ge 0$, we see that $v_2(x) \ge 1$. Since $c$ is odd, it follows that $v_2(p^2c) = 0$. Hence
$v_2(x + p^2c) = v_2(p^2c) = 0$, and thus $v_2(x + 4pb^2)$ is odd. Since $p \equiv 5 \pmod{8}$, the local Hilbert symbol $(p, x + 4pb^2)_2$ satisfies
\begin{align*}
(p, x + 4pb^2)_2 = \left(\dfrac{p}{2}\right) = -1.
\end{align*}
Therefore $\text{inv}_2(\cA(P_2))$ equals $1/2$.
Therefore, in any event, $\sum_{l}\text{inv}_l\cA(P_l) = 1/2$ for any $(P_l)_{l} \in \cV(\bA_{\bQ})$. Thus $\cV(\bA_{\bQ})^{\Br} = \emptyset$.
$\star$ \textit{Step 4. $\cU$ and $\cT$ satisfy HP1.}
For any point $P_l \in \cU(\bQ_l)$, let $x = x(P_l)$. By the definition of $\cU$, we see that $x + 4pb^2$ is nonzero. By what we have proved in \textit{Step 3}, we know that the local Hilbert symbol $(p, x + 4pb^2)_l$ satisfies
\begin{align*}
(p, x + 4pb^2)_l =
\begin{cases}
1 \; \; &\text{if $l \ne 2$,} \\
- 1\; \; &\text{if $l = 2$.}
\end{cases}
\end{align*}
Hence it follows that $x + 4pb^2$ is the norm of an element of $\bQ_l(\sqrt{p})$ for every $l \ne 2$ including $l = \infty$, and that $x + 4pb^2$ is not a local norm of any element of $\bQ_2(\sqrt{p})$. Thus we deduce that
\begin{align}
\label{Equation-The-product-of-local-Hilbert-symbols-equals-minus-1-in-Theorem-about-no-rational-points-on-the-threefolds}
\prod_{\substack{l}}(p, x + 4pb^2)_l = -1,
\end{align}
where the product is taken over every prime $l$ including $l = \infty$. Therefore it follows from the product formula \cite[Theorem 5.3.1]{cohen} that $\cU(\bQ)$ is empty; otherwise there exists a rational point $P \in \cU(\bQ)$. Thus the element $x + 4pb^2$ is in $\bQ^{\times}$, where $x = x(P)$. Hence by the product formula, we see that
\begin{align*}
\prod_{\substack{l}}(p, x + 4pb^2)_l = 1,
\end{align*}
which is a contradiction to $(\ref{Equation-The-product-of-local-Hilbert-symbols-equals-minus-1-in-Theorem-about-no-rational-points-on-the-threefolds})$. Hence $\cU$ satisfies HP1, and it thus follows from \textit{Step 1} that $\cT$ satisfies HP1.
$\star$ \textit{Step 5. $\cU, \cV$ and $\cT$ satisfy HP2.}
For this contention, note that since $\cT(\bQ) = \emptyset$, it follows from the Amer-Brumer theorem \cite{amer} \cite{brumer} that $\cT$ does not contain any zero-cycle of odd degree over $\bQ$. Thus $\cU, \cV$ and $\cT$ satisfy HP2, and hence our contention follows.
\end{proof}
The following result plays a key role in constructing algebraic families of curves satisfying HP1 and HP2.
\begin{theorem}
\label{threefold-hp-theorem}
Let $p$ be a prime such that $p \equiv 5 \pmod{8}$. Assume $(A1)$, and assume further that the following is true.
\begin{itemize}
\item [(A2)] there exists a couple $(b, d)$ of integers such that $b, d$ are odd, $b \not\equiv 0 \pmod{3}$, $b \not\equiv 0 \pmod{p}$ and $q := |pd^2 - 4b^2|$ is either $1$ or an odd prime.
\end{itemize}
Let $\cZ$ be a smooth and proper $\bQ$-model of the smooth $\bQ$-variety $\cX$ in $\bA^5_{\bQ}$ defined by
\begin{equation}
\label{eqn-X}
\cX :
\begin{cases}
0 \ne u_1^2 - pv_1^2 &= 2x \\
0 \ne u_2^2 - pv_2^2 &= 2(x + 4pb^2)(x + p^2d^2).
\end{cases}
\end{equation}
Let $\cY \subset \bP^5_{\bQ}$ be the singular $\bQ$-variety defined by
\begin{equation}
\label{eqn-Y}
\cY :
\begin{cases}
u_1^2 - pv_1^2 &= 2xy \\
u_2^2 - pv_2^2 &= 2(x + 4pb^2y)(x + p^2d^2y).
\end{cases}
\end{equation}
Then $\cX$, $\cY$ and $\cZ$ satisfy HP1 and HP2.
\end{theorem}
\begin{remark}
In Section \ref{infinitude-threefold-section}, we will prove that there are infinitely many triples $(p, b, d)$ satisfying $(A1)$ and $(A2)$.
\end{remark}
\begin{proof}
Let $c = d^2$. We contend that the couple $(b, c)$ satisfies $(B)$ in Lemma \ref{azumaya-threefold-lemma}. Indeed, we note that $\gcd(b, d) = 1$; otherwise, there exists an odd prime
$l$ such that $b = lb_1$ and $d = ld_1$ for some integers $b_1, d_1 \in \bZ$. Hence $q := l_1^2|pd_1^2 - 4b_1^2|$, which is a contradiction to $(A2)$. Thus $\gcd(b, d) = 1$, and it follows that $\gcd(b, c) = 1$.
We know that $q = |pc - 4b^2|$ is either $1$ or an odd prime, and that $b \not\equiv 0 \pmod{3}$ and $b \not\equiv 0 \pmod{p}$. Hence the couple $(b, c)$ satisfies $(B)$ in Lemma \ref{azumaya-threefold-lemma}. Thus by Theorem \ref{general-threefold-hp-theorem}, we deduce that $\cX$, $\cY$ and $\cZ$ satisfy HP1 and HP2.
\end{proof}
\section{Infinitude of the triples $(p, b, d)$.}
\label{infinitude-threefold-section}
In this section, we will prove that there are infinitely many triples $(p, b, d)$ satisfying $(A1)$ and $(A2)$ in Theorem \ref{threefold-hp-theorem}. We begin by recalling a theorem of Iwaniec's.
Let $P(x,y)$ be a quadratic polynomial in two variables $x$ and $y$. We say that $P$ $\textit{depends essentially on two
variables}$ if $\tfrac{\partial{P}}{\partial{x}}$ and $\tfrac{\partial{P}}{\partial{y}}$ are linearly independent as elements of
the $\bQ$-vector space $\bQ[x, y]$.
\begin{theorem}$(\text{Iwaniec, \cite[p.$443$]{iwaniec}})$
\label{iwaniec-thm}
Let $P(x,y) = ax^2 + bxy + cy^2 + ex + fy + g$ be a quadratic polynomial defined over $\bQ$, and assume that the following are true.
\begin{itemize}
\item[(i)] $a$, $b$, $c$, $e$, $f$, $g$ are in $\bZ$ and $\gcd(a,b,c,e,f,g) = 1$.
\item[(ii)] $P(x,y)$ is irreducible in $\bQ[x,y]$, and represents arbitrarily large odd numbers and depends essentially on two
variables.
\item[(iii)] $D = af^2 - bef + ce^2 + (b^2 - 4ac)g = 0$ or $\Delta = b^2 - 4ac$ is a perfect square.
\end{itemize}
Then
\begin{equation*}
Nlog^{-1}N \ll \sum_{\substack{p \le N, \, p = P(x,y)\\ p \; prime}}1.
\end{equation*}
\end{theorem}
We now prove the main lemma in this section.
\begin{lemma}
\label{threefolds-infinitude-triple-lemma}
Let $p$ be a prime such that $p \equiv 5 \pmod{8}$, and assume that $3$ is a quadratic non-residue in $\bF_p^{\times}$. Then there are infinitely many triples $(p, b, d)$ satisfying $(A1)$ and $(A2)$ in Theorem \ref{threefold-hp-theorem}.
\end{lemma}
\begin{proof}
Let $b_0 \in \bZ$ such that $b_0 \not\equiv 0 \pmod{3}$, $b_0 \not\equiv 0 \pmod{p}$ and $b_0$ is odd. Let $P(x, y) \in \bQ[x, y]$ be a polynomial in two variables $x, y$ defined by
\begin{equation*}
P(x, y) := p(2x + 1)^2 - 4(6py + b_0)^2.
\end{equation*}
Expanding $P(x, y)$ in the form of $Ax^2 + Bxy + Cy^2 + Ex + Fy + G$, we know that
\begin{equation*}
A := 4p, \; B := 0, \; C := -144p^2, \; E := 4p, \; F := -48pb_0, \; G := p - 4b_0^2 .
\end{equation*}
Hence we deduce that $AF^2 - BEF + CE^2 + (B^2 - 4AC)G = 0$ and $\gcd(A, B, C, E, F, G) = 1$. Thus $P$ satisfies the conditions in Theorem \ref{iwaniec-thm}. Therefore there are infinitely many primes $q$ such that $q = P(x, y)$. Upon letting $b = 6py + b_0$ and $d = 2x + 1$, we see that the triple $(p, b, d)$ satisfies $(A1)$ and $(A2)$.
\end{proof}
\begin{remark}
\label{Remark-We-only-need-a-special-case-of-Iwaniec-theorem-which-is-due-to-Hecke}
In the proof of Lemma \ref{threefolds-infinitude-triple-lemma}, we only need to use a special case of Iwaniec's theorem that quadratic polynomials in two variables with zero discriminant represent infinitely many primes. As remarked in \cite[page 436]{iwaniec}, the proof of this special case can be obtained by using relatively classical ideas going back to Hecke \cite{hecke-Primes-representations}.
\end{remark}
\begin{example}
Let $(p, b, d) = (5, 1, 1)$. We see that the triple $(p, b, d)$ satisfies $(A1)$ and $(A2)$. Let $\cY_{(5, 1, 1)}$ be the singular $\bQ$-threefold in $\bP^5_{\bQ}$ defined by
\begin{align*}
\cY_{(5, 1, 1)} :
\begin{cases}
u_1^2 - 5v_1^2 = 2xy \\
u_2^2 - 5v_2^2 = 2(x + 20y)(x + 25y).
\end{cases}
\end{align*}
By Theorem \ref{threefold-hp-theorem}, $\cY_{(5, 1, 1)}$ satisfies HP1 and HP2. The threefold $\cY_{(5, 1, 1)}$ is
the well-known Colliot-Th\'el\`ene-Coray-Sansuc threefold \cite[Proposition 7.1, p. 186]{colliot-thelene-coray-sansuc}.
\end{example}
\begin{example}
Let $(p, b, d) = (29, 1, 3)$. We see that
\begin{equation*}
q = |pd^2 - 4b^2| = |29\cdot 3^2 - 4\cdot 1^2| = 257,
\end{equation*}
which is an odd prime. Hence $(29, 1, 3)$ satisfies $(A1)$ and $(A2)$. Let $\cY_{(29, 1, 3)}$ be the singular $\bQ$-threefold in $\bP^5_{\bQ}$ defined by
\begin{align*}
\cY_{(29, 1, 3)} :
\begin{cases}
u_1^2 - 29v_1^2 = 2xy \\
u_2^2 - 29v_2^2 = 2(x + 116y)(x + 7569y).
\end{cases}
\end{align*}
By Theorem \ref{threefold-hp-theorem}, $\cY_{(29, 1, 3)}$ satisfies HP1 and HP2.
\end{example}
\section{Hyperelliptic curves violating the Hasse principle}
\label{hyperelliptic-curve-hp-section}
In this section, we give a sufficient condition under which for each integer $n \ge 2$ and $n \not\equiv 0 \pmod{4}$, there exist hyperelliptic curves of genus $n$ that lie on the threefolds $\cY$ in Theorem \ref{threefold-hp-theorem} and satisfy HP1 and HP2. The sufficient condition is in terms of the existence of certain sextuples $(p, b, d, \alpha, \beta, \gamma)$, and obtained using the geometric construction of hyperelliptic curves due to Coray and Manoil \cite[Proposition 4.2]{coray-manoil}.
\begin{theorem}
\label{hyperelliptic-curves-theorem}
Let $p$ be a prime such that $p \equiv 5 \pmod{8}$, and let $(p, b, d) \in \bZ^3$ be a triple of integers satisfying $(A1)$ and $(A2)$ in Theorem \ref{threefold-hp-theorem}. Let $n$ be an integer such that $n \ge 2$, and let $(\alpha, \beta, \gamma) \in \bQ^3$ be a triple of rational numbers such that $\alpha \beta \gamma \ne 0$. Assume further that the following are true.
\begin{itemize}
\item [(A3)]
\begin{equation}
\label{eqn-P}
P := p\alpha^2 + 2\beta^2 - 2p\gamma^2 \ne 0,
\end{equation}
\begin{equation}
\label{eqn-Q}
Q := 4bdp\gamma - 4b^2\beta - d^2p\beta \ne 0,
\end{equation}
and the conic $\cQ_1 \subset \bP^2_{\bQ}$ defined by
\begin{equation*}
\cQ_1: pU^2 - V^2 - (\beta PQ)T^2 = 0
\end{equation*}
has a point $(u, v, t) \in \bZ^3$ with $uvt \ne 0$ and $\gcd(u, v, t) = 1$.
\item [(S)] the polynomial $P_{p, b, d, \alpha, \beta, \gamma}(x) \in \bQ[x]$ defined by
\begin{align*}
P_{p, b, d, \alpha, \beta, \gamma}(x) := p\alpha^2Q^2 x^{2n + 2} + (2b^2Px^2 + \beta Q)(d^2pPx^2 + 2\beta Q)
\end{align*}
is separable; that is, $P_{p, b, d, \alpha, \beta, \gamma}(x)$ has exactly $2n + 2$ distinct roots in $\bC$.
\end{itemize}
Let $\cC$ be the smooth projective model of the affine curve defined by
\begin{equation}
\label{eqn-hyperelliptic-curve-C}
\cC: z^2 = p\alpha^2Q^2 x^{2n + 2} + (2b^2Px^2 + \beta Q)(d^2pPx^2 + 2\beta Q).
\end{equation}
Then $\cC(\bQ_l) \ne \emptyset$ for every prime $l \ne 2, p$ and $\cC(\bA_{\bQ})^{\Br} = \emptyset$. Furthermore $\cC$ satisfies HP2.
\end{theorem}
\begin{proof}
The proof of Theorem \ref{hyperelliptic-curves-theorem} presented below follows closely from that of Proposition \cite[Proposition 4.2]{coray-manoil}. We begin by recalling the geometric construction of hyperelliptic curves due to Coray and Manoil.
Let $\cC_{a} \subset \bA_K^2$ be the affine curve defined by $z^2 = P(x)$, where $P(x)$ is a separable polynomial of degree $2n + 2$ and $K$ is a number field. Recall from \cite[Chapter II, Exercise 2.14]{silverman-EC} that the smooth projective model of $\cC_{a}$ can be described as the closure of
the image of $\cC_{a}$ under the mapping
\begin{align*}
\cC_{a} &\longrightarrow \bP^{n + 2}_K \\
(x, z) &\mapsto \left(1, x, \ldots, x^{n + 1}, z\right).
\end{align*}
Following \cite[Proposition 4.2]{coray-manoil}, we will index the coordinates of $\bP^{n+2}_K$ in such a way that $z_i$ corresponds to $x^i$ for $0 \le i \le n + 1$ and $z_{n + 2}$ corresponds to $z$.
Using the above arguments, we deduce from $(\ref{eqn-hyperelliptic-curve-C})$ that $\cC$ can be smoothly embedded into the intersection of quadrics defined by
\begin{align}
\label{quadrics-eqns}
\begin{cases}
z_{n + 2}^2 &= p\alpha^2Q^2z_{n+1}^2 + (2b^2Pz_2 + \beta Qz_0)(d^2pPz_2 + 2\beta Qz_0) \\
z_1^2 &= z_2z_0.
\end{cases}
\end{align}
Recall that $(u, v, t) \in \bZ^3$ is the point on the conic $\cQ_1$ defined in $(A3)$ that is assumed to exist. Upon letting
\begin{align*}
\begin{cases}
z_0 &= \dfrac{1}{\beta Q}x \\
z_1 &= \dfrac{t}{u}u_1 \\
z_2 &= \dfrac{2p}{P}y \\
z_{n + 1} &= \dfrac{1}{\alpha Q}v_2 \\
z_{n+2} &= u_2,
\end{cases}
\end{align*}
we deduce from $(\ref{quadrics-eqns})$ that
\begin{align}
\label{dpsurface}
\begin{cases}
\dfrac{\beta PQt^2}{pu^2}u_1^2 &= 2xy \\
u_2^2 - pv_2^2 &= 2(x + 4pb^2y)(x + p^2d^2y).
\end{cases}
\end{align}
We see that $(\ref{dpsurface})$ defines a singular del Pezzo surface $\cD \subseteq \bP^4_{\bQ}$. We contend that $\cD(\bA_{\bQ})^{\text{Br}} = \emptyset$ and $\cD$ does not contain any zero-cycle of odd degree over $\bQ$. Indeed, upon letting
\begin{equation*}
v_1 = \dfrac{v}{pu}u_1,
\end{equation*}
we deduce from the first equation of $(\ref{dpsurface})$ and $(A3)$ that
\begin{equation*}
u_1^2 - pv_1^2 = u_1^2 - p\dfrac{v^2}{p^2u^2}u_1^2 = \dfrac{\beta PQt^2}{pu^2}u_1^2 = 2xy.
\end{equation*}
Therefore $\cD$ is a hyperplane section of the threefold $\cY$ in Theorem \ref{threefold-hp-theorem}. Hence there exists a sequence of $\bQ$-morphisms
\begin{align*}
\cC \longrightarrow \cD \longrightarrow \cY.
\end{align*}
Hence it follows from Lemma \ref{functoriality-azumaya-lemma} and Theorem \ref{threefold-hp-theorem} that $\cD(\bA_{\bQ})^{\text{Br}} = \emptyset$. Thus $\cC(\bA_{\bQ})^{\text{Br}} = \emptyset$. Furthermore since $\cY$ does not contain any zero-cycle of odd degree over $\bQ$, so do $\cC$ and $\cD$.
We now prove that $\cC$ is locally solvable at primes $l$ with $l \ne 2, p$. We consider the following cases.
$\star$ \textbf{Case I.} \textit{$l = \infty$ or $l$ is an odd prime such that $l \ne p$ and $\left(\dfrac{p}{l}\right) = 1$.}
We know that the curve $\cC^{\ast}$ defined by
\begin{equation*}
\cC^{\ast} : z^2 = p\alpha^2Q^2 x^{2n + 2} + y^{2n - 2}(2b^2Px^2 + \beta Qy^2)(d^2pPx^2 + 2\beta Qy^2)
\end{equation*}
is an open subscheme of $\cC$. We see that $P_{\infty} = (x :y : z) = (1 : 0 : \sqrt{p}\alpha Q)$ belongs to $\cC^{\ast}(\bQ_l) \subset \cC(\bQ_l) $, and hence $\cC$ is locally solvable at $l$.
$\star$ \textbf{Case II.} \textit{$l$ is an odd prime such that $\left(\dfrac{2}{l}\right) = 1$.}
It follows from $(\ref{eqn-hyperelliptic-curve-C})$ that the point $P_1 = (x, z) = (0, \sqrt{2}\beta Q)$ belongs to $\cC(\bQ_l)$.
$\star$ \textbf{Case III.} \textit{$l$ is an odd prime such that $l \ne p$ and $\left(\dfrac{2p}{l}\right) = 1$.}
Let $F(x, z)$ be the defining polynomial of $\cC$ defined by
\begin{equation*}
F(x, z) = p\alpha^2Q^2 x^{2n + 2} + (2b^2Px^2 + \beta Q)(d^2pPx^2 + 2\beta Q) - z^2.
\end{equation*}
We see that
\begin{equation*}
F\left(1, \sqrt{2p}(\gamma Q + bdP)\right) = (p\alpha^2Q^2 + 2p(bdP)^2 + 4b^2\beta PQ + \beta pPQd^2 + 2\beta^2 Q^2) - 2p(\gamma Q + bdP)^2
\end{equation*}
Hence it follows from $(\ref{eqn-P})$ and $(\ref{eqn-Q})$ that
\begin{equation*}
p\alpha^2Q^2 + 4b^2\beta PQ + \beta pPQd^2 + 2\beta^2 Q^2 = 2p\gamma^2Q^2 + 4p(\gamma Q)(bdP).
\end{equation*}
Thus
\begin{equation*}
p\alpha^2Q^2 + 2p(bdP)^2 + 4b^2\beta PQ + \beta pPQd^2 + 2\beta^2 Q^2 = 2p(\gamma Q + bdP)^2.
\end{equation*}
Hence we deduce that $F(1,\sqrt{2p}(\gamma Q + bdP)) = 0$, and therefore the point $P_2 = (1,\sqrt{2p}(\gamma Q + bdP))$ belongs to $\cC(\bQ_l)$.
Thus, in any event, $\cC$ is locally solvable at primes $l$ with $l \ne 2, p$, which proves our contention.
\end{proof}
\begin{remark}
Theorem \ref{hyperelliptic-curves-theorem} constructs hyperelliptic curves of genus at least two such that they satisfy HP2 and all conditions in HP1 except local solvability at $2$ and $p$. The rest of this section presents certain sufficient conditions for which those hyperelliptic curves arising from Theorem \ref{hyperelliptic-curves-theorem} are locally solvable at $2$ and $p$, and hence satisfy HP1 and HP2.
\end{remark}
\begin{lemma}
\label{p-divisibility-beta-lemma}
Let $p$ be a prime such that $p \equiv 5 \pmod{8}$, and let $(b, d) \in \bZ^3$ be a couple of integers satisfying $(A1)$ and $(A2)$ in Theorem \ref{threefold-hp-theorem}. Assume that there is a triple $(\alpha, \beta, \gamma) \in \bQ^3$ satisfying $(A3)$ in Theorem \ref{hyperelliptic-curves-theorem}, and assume further that $\alpha, \beta, \gamma \in \bZ_p$. Then there is a rational number $\bar{\beta} \in \bQ$ such that $\beta = p\bar{\beta}$ and $\bar{\beta} \in \bZ_p$.
\end{lemma}
\begin{proof}
Let $\cQ_1$ be the conic defined in $(A3)$. Assume that $(u, v, t) \in \bZ^3$ belongs to $\cQ_1(\bQ)$ such that $uvt \ne 0$ and $\gcd(u, v, t) = 1$. We see that
\begin{equation*}
pu^2 - v^2 - \beta PQt^2 = 0,
\end{equation*}
where $P$ and $Q$ are defined by $(\ref{eqn-P})$ and $(\ref{eqn-Q})$, respectively. Taking the identity above modulo $p$, it follows that
\begin{equation*}
v^2 \equiv 8b^2\beta^4t^2 \pmod{p}.
\end{equation*}
Since $2$ is a quadratic non-residue in $\bF_p^{\times}$ and $b \not\equiv 0 \pmod{p}$, we deduce from the congruence above that
\begin{equation*}
v \equiv \beta t \equiv 0 \pmod{p}.
\end{equation*}
Assume that $\beta \not\equiv 0 \pmod{p}$. Then $v \equiv t \equiv 0 \pmod{p}$, and hence $v = pv_1$ and $t = pt_1$ for some integers $v_1, t_1$. Substituting $v$ and $t$ into the defining equation of the conic $\cQ_1$, we get
\begin{equation*}
u^2 - pv_1^2 - p\beta PQ t_1^2 = 0,
\end{equation*}
and hence it follows that $p$ divides $u$. Thus $p$ divides $\gcd(u, v, t)$, which is a contradiction. Therefore there is a rational number $\bar{\beta} \in \bQ$ such that
$\beta = p\bar{\beta}$ and $\bar{\beta} \in \bZ_p$.
\end{proof}
\begin{remark}
\label{p-divisibility-P-Q-remark}
By Lemma \ref{p-divisibility-beta-lemma}, one knows that if $(\alpha, \beta, \gamma) \in \bQ^3$ satisfies $(A3)$ and $\alpha, \beta, \gamma \in \bZ_p$, then there is a rational number $\bar{\beta}$ such that $\beta = p\bar{\beta}$ and $\bar{\beta} \in \bZ_p$. Hence one sees that $P = pP_1$ and $Q= pQ_1$, where
\begin{align*}
P_1 &:= \alpha^2 + 2p\bar{\beta}^2 - 2\gamma^2, \\
Q_1 &:= 4bd\gamma - 4b^2\bar{\beta} - d^2p\bar{\beta}.
\end{align*}
We also see that $P_1$ and $Q_1$ belong to $\bZ_p$.
\end{remark}
In the proofs of Corollary \ref{hyperelliptic-curves-hp-corollary} and Corollary \ref{n-even-hyperelliptic-curve-hp-corollary} below, we will use Hensel's lemma to deduce local solvability at primes $2$ and $p$. For the sake of self-containedness, we recall the statement of Hensel's lemma.
\begin{theorem}
\label{Theorem-Hensel-lemma}
$(\text{Hensel's lemma, \cite[Theorem 3, Section 5.2]{borevich-shafarevich}})$
Let $p$ be a prime. Let $F(x_1, x_2, \ldots, x_n) \in \bZ_p[x_1, x_2, \ldots, x_n]$ be a polynomial whose coefficients are $p$-adic integers. Let $\delta \ge 0$ be a nonnegative integer. Assume that there are $p$-adic integers $a_1, a_2, \ldots, a_n$ such that for some integer $1 \le k \le n$, we have
\begin{align*}
F(a_1, a_2, \ldots, a_n) &\equiv 0 \pmod{p^{2\delta + 1}}, \\
\dfrac{\partial F}{\partial x_k}(a_1, a_2, \ldots, a_n) &\equiv 0 \pmod{p^{\delta}}, \\
\dfrac{\partial F}{\partial x_k}(a_1, a_2, \ldots, a_n) &\not\equiv 0 \pmod{p^{\delta + 1}}.
\end{align*}
Then there exist $p$-adic integers $\theta_1, \theta_2, \ldots, \theta_n$ such that $F(\theta_1, \theta_2, \ldots, \theta_n) = 0$.
\end{theorem}
The following result provides a sufficient condition under which certain hyperelliptic curves of odd genus satisfy HP1 and HP2.
\begin{corollary}
\label{hyperelliptic-curves-hp-corollary}
We maintain the same notation and assumptions as in Theorem \ref{hyperelliptic-curves-theorem}. Assume $(A1)-(A3)$ and $(S)$ in Theorem \ref{hyperelliptic-curves-theorem}. Assume further that the following are true.
\begin{itemize}
\item [(A4)] $\alpha, \beta, \gamma \in \bZ_2^{\times}$, $\alpha, \gamma, d \in \bZ_p^{\times}$ and $\beta \in \bZ_p$.
\item [(A5)] $\gamma Q_1 + bdP_1 \equiv 0 \pmod{p^2}$, where $\bar{\beta}$, $P_1$ and $Q_1$ are defined as in Remark \ref{p-divisibility-P-Q-remark}.
\item [(A6)] $n \not\equiv -2\left(\dfrac{\gamma}{\alpha}\right)^2 \pmod{p}$, $n \ge 3$ and $n$ is odd.
\end{itemize}
Let $\cC$ be the smooth projective model of the affine curve defined by $(\ref{eqn-hyperelliptic-curve-C})$ in Theorem \ref{hyperelliptic-curves-theorem}. Then $\cC$ satisfies HP1 and HP2.
\end{corollary}
\begin{proof}
By Theorem \ref{hyperelliptic-curves-theorem}, it suffices to prove that $\cC$ is locally solvable at $2$ and $p$.
$\star$ \textit{Step 1. $\cC$ is locally solvable at $p$.}
We will use Theorem \ref{Theorem-Hensel-lemma} with the exponent $\delta = 3$ to prove local solvability of $\cC$ at $p$. We consider the following system of equations
\begin{align}
\label{hensel-p}
\begin{cases}
F(x , z) &= p\alpha^2Q^2 x^{2n + 2} + (2b^2Px^2 + \beta Q)(d^2pPx^2 + 2\beta Q) - z^2 \equiv 0 \pmod{p^7} \\
\dfrac{\partial{F}}{\partial{x}}(x, z) &= (2n + 2)p\alpha^2 Q^2x^{2n + 1} + 4b^2Px(d^2pPx^2 + 2\beta Q) + 2d^2pPx(2b^2Px^2 + \beta Q) \equiv 0 \pmod{p^3} \\
\dfrac{\partial{F}}{\partial{x}}(x, z) &\not\equiv 0 \pmod{p^4}.
\end{cases}
\end{align}
Repeating the same arguments as in \textbf{Case III} of the proof of Theorem \ref{hyperelliptic-curves-theorem}, we deduce that
\begin{equation*}
F(1, 0) = 2p(\gamma Q + bdP)^2.
\end{equation*}
By Remark \ref{p-divisibility-P-Q-remark}, one knows that $P = pP_1$ and $Q = pQ_1$. Hence
\begin{equation}
\label{eqn-F}
F(1, 0) = 2p^3\left(\gamma Q_1 + bdP_1\right)^2.
\end{equation}
Thus it follows from $(A5)$ and $(\ref{eqn-F})$ that $F(1, 0) \equiv 0 \pmod{p^7}$. On the other hand, we see that
\begin{equation}
\label{der-F-eqn}
\dfrac{\partial F}{\partial x}(1, 0) = p^3\left((2n + 2)\alpha^2Q_1^2 + 4b^2P_1(d^2P_1 + 2\bar{\beta}Q_1) + 2d^2P_1(2b^2P_1 + p\bar{\beta}Q_1)\right).
\end{equation}
Since $\alpha, \bar{\beta}, \gamma$ and $P_1, Q_1$ are in $\bZ_p$, one obtains that
\begin{equation*}
\dfrac{\partial F}{\partial x}(1, 0) \equiv 0 \pmod{p^3}.
\end{equation*}
Assume that
\begin{equation}
\label{p^4-der-F-eqn}
\dfrac{1}{p^3}\left(\dfrac{\partial F}{\partial x}(1, 0)\right) \equiv 0 \pmod{p}.
\end{equation}
Since $\gamma \in \bZ_p^{\times}$, it follows from $(A5)$ that
\begin{equation}
\label{relation-PQ}
Q_1 \equiv -\dfrac{bd}{\gamma}P_1 \pmod{p}.
\end{equation}
Upon replacing $Q_1$ by $-\dfrac{bd}{\gamma}P_1$ in $(\ref{p^4-der-F-eqn})$, we deduce that
\begin{equation*}
\dfrac{2P_1^2bd}{\gamma^2}\left((n + 1)\alpha^2bd + \gamma(4bd\gamma - 4b^2\bar{\beta} - d^2p\bar{\beta}) \right) \equiv 0 \pmod{p}.
\end{equation*}
Thus it follows from the equation of $Q_1$ in Remark \ref{p-divisibility-P-Q-remark} that
\begin{equation*}
\dfrac{2P_1^2bd}{\gamma^2}\left((n + 1)\alpha^2bd + \gamma Q_1\right) \equiv 0 \pmod{p}.
\end{equation*}
Note that $P_1 \in \bZ_p^{\times}$; otherwise, we deduce from the equation of $P_1$ in Remark \ref{p-divisibility-P-Q-remark} that
\begin{equation*}
\alpha^2 - 2\gamma^2 \equiv P_1 \equiv 0 \pmod{p}.
\end{equation*}
Since $\alpha, \gamma \in \bZ_p^{\times}$, it follows from the congruence above that $2 \equiv \left(\dfrac{\alpha}{\gamma}\right)^2 \pmod{p}$, which is a contradiction to the fact that
$p \equiv 5 \pmod{8}$. Thus $P_1 \in \bZ_p^{\times}$. Since $2, b, d, \gamma$ and $P_1$ are in $\bZ_p^{\times}$, we obtain that
\begin{equation*}
(n + 1)\alpha^2bd + \gamma Q_1 \equiv 0 \pmod{p}.
\end{equation*}
Since $\gamma Q_1 \equiv -bdP_1 \pmod{p}$ and $b, d \in \bZ_p^{\times}$, we deduce from the last congruence that
\begin{align*}
(n + 1)\alpha^2 \equiv P_1 \equiv (\alpha^2 - 2\gamma^2) \pmod{p}.
\end{align*}
Since $\alpha, \gamma \in \bZ_p^{\times}$, it follows that $n \equiv -2\left(\dfrac{\gamma}{\alpha}\right)^2 \pmod{p}$, which is a contradiction to $(A6)$. Thus the system $(\ref{hensel-p})$ has a solution $(x, z) = (1, 0)$. By Hensel's lemma, $\cC$ is locally solvable at $p$.
$\star$ \textit{Step 2. $\cC$ is locally solvable at $2$.}
We will use Theorem \ref{Theorem-Hensel-lemma} with the exponent $\delta = 1$ to prove local solvability of $\cC$ at $2$. We consider the following system of equations
\begin{align}
\label{hensel-2}
\begin{cases}
F(x,z) &\equiv 0 \pmod{2^3} \\
\dfrac{\partial{F}}{\partial{x}}(x,z) &\equiv 0 \pmod{2} \\
\dfrac{\partial{F}}{\partial{x}}(x,z) &\not\equiv 0 \pmod{2^2}.
\end{cases}
\end{align}
We see from $(\ref{eqn-F})$ and the equations of $P_1$ and $Q_1$ in Remark \ref{p-divisibility-P-Q-remark} that
\begin{equation*}
F(1, 0) = 2p^3\left(\gamma(4bd\gamma - 4b^2\bar{\beta} - d^2p\bar{\beta}) + bd(\alpha^2 + 2p\bar{\beta}^2 - 2\gamma^2)\right)^2.
\end{equation*}
Since $\beta$ is in $\bZ_2^{\times}$ and $p \ne 2$, we see that $\bar{\beta}$ is also in $\bZ_2^{\times}$. Since $b, d, p, \alpha, \bar{\beta}, \gamma \in \bZ_2^{\times}$, we see that
\begin{equation*}
-d^2p \bar{\beta} \gamma + bd\alpha^2 \equiv 0 \pmod{2}.
\end{equation*}
Let $v_2$ denote the $2$-adic valuation. We see that
\begin{align*}
v_2\left(\gamma(4bd\gamma - 4b^2\bar{\beta} - d^2p\bar{\beta}) + bd(\alpha^2 + 2p\bar{\beta}^2 - 2\gamma^2)\right) &= v_2\left((4\gamma(bd\gamma - b^2\bar{\beta}) + 2bd(p\bar{\beta}^2 - \gamma^2)) + (-d^2p \bar{\beta} \gamma + bd\alpha^2)\right) \\
&\ge \min\left(v_2(4\gamma(bd\gamma - b^2\bar{\beta}) + 2bd(p\bar{\beta}^2 - \gamma^2)), v_2(-d^2p \bar{\beta} \gamma + bd\alpha^2)\right) \\
&\ge 1.
\end{align*}
Hence $F(1, 0) \equiv 0 \pmod{2^3}$. On the other hand, we know from $(\ref{der-F-eqn})$ that $\dfrac{\partial{F}}{\partial{x}}(1, 0) \equiv 0 \pmod{2}$. Since $n$ is odd, $(2n + 2) \equiv 2(n + 1) \equiv 0 \pmod{2^2}$. Hence it follows from $(\ref{der-F-eqn})$ that
\begin{equation*}
\dfrac{\partial F}{\partial x}(1, 0) \equiv 2d^2p^4\bar{\beta}P_1Q_1 \pmod{2^2}.
\end{equation*}
By $(A4)$ and the equations of $P_1$ and $Q_1$ in Remark \ref{p-divisibility-P-Q-remark}, we know that
\begin{equation*}
d^2p^4\bar{\beta}P_1Q_1 \not\equiv 0 \pmod{2}.
\end{equation*}
Hence we deduce that $\dfrac{\partial F}{\partial x}(1, 0) \not\equiv 0 \pmod{2^2}$. Thus the system $(\ref{hensel-2})$ has a solution $(x, z) = (1, 0)$. By Hensel's lemma, $\cC$ is locally solvable at $2$, and hence our contention follows.
\end{proof}
\begin{remark}
\label{local-solubility-at-2-p-remark}
Assuming $(A1), (A2), (A3), (A5)$ and $(S)$. Following closely the proof of Corollary \ref{hyperelliptic-curves-hp-corollary}, we note that the following are true.
\begin{itemize}
\item [(1)] if $\alpha, \beta, \gamma \in \bZ_2^{\times}$ and $n$ is odd, then $\cC$ is locally solvable at $2$.
\item [(2)] if $\alpha, \gamma, d \in \bZ_p^{\times}$, $\beta \in \bZ_p$, $n \ge 2$ and $n \not\equiv -2\left(\dfrac{\gamma}{\alpha}\right)^2 \pmod{p}$, then $\cC$ is locally solvable at $p$.
\end{itemize}
\end{remark}
We now prove a sufficient condition under which certain hyperelliptic curves of genus $n \equiv 2 \pmod{4}$ satisfy HP1 and HP2.
\begin{corollary}
\label{n-even-hyperelliptic-curve-hp-corollary}
We maintain the same notation as in Theorem \ref{hyperelliptic-curves-theorem} and Corollary \ref{hyperelliptic-curves-hp-corollary}. Assume $(A1)-(A5)$ and $(S)$. Assume further that the following are true.
\begin{itemize}
\item [(B1)] $bd - \bar{\beta}\gamma \equiv 0 \pmod{4}$.
\item [(B2)] $n \not\equiv -2\left(\dfrac{\gamma}{\alpha}\right)^2 \pmod{p}$, $n \ge 2$ and $n \equiv 2 \pmod{4}$.
\end{itemize}
Let $\cC$ be the smooth projective model defined by $(\ref{eqn-hyperelliptic-curve-C})$ in Theorem \ref{hyperelliptic-curves-theorem}. Then $\cC$ satisfies HP1 and HP2.
\end{corollary}
\begin{proof}
By Theorem \ref{hyperelliptic-curves-theorem} and Remark \ref{local-solubility-at-2-p-remark}, it suffices to prove that $\cC$ is locally solvable at $2$. We will use Theorem \ref{Theorem-Hensel-lemma} with the exponent $\delta = 2$ to prove local solvability of $\cC$ at $2$. We consider the following
system of equations
\begin{align}
\label{locally-solvable-2-hensel}
\begin{cases}
F(x , z) &\equiv 0 \pmod{2^5} \\
\dfrac{\partial{F}}{\partial{x}}(x, z) &\equiv 0 \pmod{2^2} \\
\dfrac{\partial{F}}{\partial{x}}(x, z) &\not\equiv 0 \pmod{2^3},
\end{cases}
\end{align}
where $F(x, z)$ denotes the polynomial in variables $x, z$ defined in $(\ref{hensel-p})$. Since $\alpha \in \bZ_2^{\times}$, we know that $\alpha \equiv 1 \pmod{4}$ or $\alpha \equiv 3 \pmod{4}$. Hence $\alpha^2 \equiv 1 \pmod{4}$. Similarly we know that $\bar{\beta}^2, \gamma^2, b^2, d^2 \equiv 1 \pmod{4}$. Since $p \equiv 5 \pmod{8}$, it follows that
\begin{align*}
P_1 &\equiv 1 \pmod{4}, \\
Q_1 &\equiv -\bar{\beta} \pmod{4}.
\end{align*}
By $(B1)$, we know that
\begin{equation*}
\gamma Q_1 + bd P_1 \equiv bd - \bar{\beta}\gamma \equiv 0 \pmod{4},
\end{equation*}
and hence we deduce from $(\ref{eqn-F})$ that $F(1, 0) \equiv 0 \pmod{2^5}$.
Since $n \equiv 2 \pmod{4}$, there is a non-negative integer $l$ such that $n = 4l + 2$. We know that
\begin{equation*}
4b^2P_1(d^2P_1 + 2\bar{\beta}Q_1) + 2d^2P_1(2b^2P_1 + p\bar{\beta}Q_1) = 8b^2d^2P_1^2 + 8b^2\bar{\beta}P_1Q_1 + 2pd^2\bar{\beta}P_1Q_1.
\end{equation*}
Hence it follows from $(\ref{der-F-eqn})$ that
\begin{equation*}
\dfrac{\partial F}{\partial x}(1, 0) \equiv 2\alpha^2Q_1^2 + 2pd^2\bar{\beta}P_1Q_1 \equiv 2 -2\bar{\beta}^2 \equiv 0 \pmod{2^2}.
\end{equation*}
Similarly one sees that
\begin{equation*}
\dfrac{\partial F}{\partial x}(1, 0) \equiv 5(8l + 6)\alpha^2Q_1^2 + 10pd^2\bar{\beta}P_1Q_1 \pmod{2^3}.
\end{equation*}
Since $\alpha, \bar{\beta}, \gamma, b, d \in \bZ_2^{\times}$, we deduce that $\alpha^2, \bar{\beta}^2, \gamma^2, b^2, d^2 \equiv 1 \pmod{2^3}$. Since $p \equiv 5 \pmod{2^3}$ and $bd\gamma - b^2\bar{\beta} \equiv 0 \pmod{2}$, it follows from the equations of $P_1$ and $Q_1$ in Remark \ref{p-divisibility-P-Q-remark} that
\begin{align*}
P_1 &\equiv 1 \pmod{2^3}, \\
Q_1 &\equiv 4(bd\gamma - b^2\bar{\beta}) - 5\bar{\beta} \equiv -5\bar{\beta} \pmod{2^3}.
\end{align*}
Thus we see that
\begin{equation*}
\dfrac{\partial F}{\partial x}(1, 0) \equiv 30 - 250\bar{\beta}^2 \equiv 4 \not\equiv 0 \pmod{2^3}.
\end{equation*}
Therefore the system $(\ref{locally-solvable-2-hensel})$ has a solution $(x, z) = (1, 0)$. By Hensel's lemma, $\cC$ is locally solvable at $2$, which proves our contention.
\end{proof}
\section{Infinitude of the sextuples $(p, b, d, \alpha, \beta, \gamma)$}
\label{infinitude-hyperelliptic-curve-hp-section}
By Corollary \ref{hyperelliptic-curves-hp-corollary} and Corollary \ref{n-even-hyperelliptic-curve-hp-corollary}, we know that in order to construct algebraic families of hyperelliptic curves satisfying HP1 and HP2, we need to find certain sextuples of rational functions in $\bQ(T)$ that parameterize sextuples $(p, b, d, \alpha, \beta, \gamma)$ satisfying $(A1)-(A5)$, $(S)$ and $(B1)$. In this section, we will show how to produce infinitely many sextuples $(p, b, d, \alpha, \beta, \gamma)$ satisfying $(A1)-(A5)$ and $(B1)$ from the known ones.
\begin{lemma}
\label{infinitude-alpha-beta-gamma-lemma}
Let $(p, b, d)$ be a triple of integers satisfying $(A1)$ and $(A2)$. Assume that there is a triple $(\alpha_0, \beta_0, \gamma_0) \in \bQ^3$ satisfying $(A3)$, $(A4)$, $(A5)$ and $(B1)$.
Let $(u_0, v_0, t_0) \in \bZ^3$ be a point on the conic $\cQ_1^{(\alpha_0, \beta_0, \gamma_0)}$ such that $u_0v_0t_0 \ne 0$ and $\gcd(u_0, v_0, t_0) = 1$, where the conic
$\cQ_1^{(\alpha_0, \beta_0, \gamma_0)}$ is defined by
\begin{equation*}
\cQ_1^{(\alpha_0, \beta_0, \gamma_0)} : pU^2 - V^2 - \beta_0P_0Q_0T^2 = 0
\end{equation*}
with
\begin{align*}
P_0 &= p\alpha_0^2 + 2\beta_0^2 - 2p\gamma_0^2, \\
Q_0 &= 4bdp\gamma_0 - 4b^2\beta_0 - d^2p\beta_0.
\end{align*}
Let $A, B \in \bQ$ be rational numbers, and assume that the following are true.
\begin{itemize}
\item [(C1)] $A, B \in \bZ_2$ and $B^2 - pA^2 \in \bZ_2^{\times}$.
\item [(C2)] $A \in \bZ_p$ and $B \in \bZ_p^{\times}$.
\item [(C3)] $u := u_0 + AC \ne 0$ and $v := v_0 + BC \ne 0$, where
\begin{equation}
\label{eqn-C}
C := \dfrac{2pu_0A - 2v_0B - 4p^3\alpha_0\beta_0t_0^2Q_0}{B^2 - pA^2 + 4p^5\beta_0t_0^2Q_0}.
\end{equation}
\end{itemize}
Define
\begin{align*}
\begin{cases}
\alpha &:= \alpha_0 + 2p^2C \\
\beta &:= \beta_0 \\
\gamma &:= \gamma_0.
\end{cases}
\end{align*}
Then the triple $(\alpha, \beta, \gamma) \in \bQ^3$ satisfies $(A3)$, $(A4)$, $(A5)$ and $(B1)$.
\end{lemma}
\begin{remark}
\label{Remark-The-geometric-motivation-for-parametrizing-the-triples-alpha-beta-gamma}
In order to use Theorem \ref{hyperelliptic-curves-theorem} to show the existence of algebraic families of hyperelliptic curves satisfying HP1 and HP2, one of the crucial steps is to describe a parametrization of triples $(\alpha, \beta, \gamma)$ such that the conics associated to these triples in $(A3)$ of Theorem \ref{hyperelliptic-curves-theorem} has a non-trivial rational point. Assuming the existence of one triple $(\alpha_0, \beta_0, \gamma_0)$ satisfying $(A3)-(A5)$ and $(B1)$, Lemma \ref{infinitude-alpha-beta-gamma-lemma} shows how to construct families of triples $(\alpha, \beta, \gamma)$ satisfying the same conditions as the triple $(\alpha_0, \beta_0, \gamma_0)$.
\end{remark}
\begin{proof}
We first prove that $(\alpha, \beta, \gamma)$ satisfies $(A4)$. Since $A \in \bZ_p$, $B \in \bZ_p^{\times}$ and the triple $(\alpha_0, \beta_0, \gamma_0)$ satisfies $(A4)$, it follows that
$B^2 - pA^2 + 4p^5\beta_0t_0^2Q_0 \in \bZ_p^{\times}$. Hence by $(\ref{eqn-C})$ and $(C2)$, we see that $C \in \bZ_p$. Thus $\alpha = \alpha_0 + 2p^2C \in \bZ_p$. Hence it follows that
\begin{equation*}
\alpha \equiv \alpha_0 \not\equiv 0 \pmod{p},
\end{equation*}
which proves that $\alpha \in \bZ_p^{\times}$. By assumption, one knows that the triple $(\alpha_0, \beta_0, \gamma_0)$ satisfies $(A4)$. Since $\beta = \beta_0$ and $\gamma = \gamma_0$, we deduce that $\beta, \gamma \in \bZ_2^{\times}$, $\beta \in \bZ_p$ and $\gamma, d \in \bZ_p^{\times}$. Hence it remains to prove that $\alpha \in \bZ_2^{\times}$. By assumptions and $(C1)$, we know that $Q_0 \in \bZ_2$ and $B^2 - pA^2 \in \bZ_2^{\times}$. Hence it follows that
\begin{align*}
B^2 - pA^2 + 4p^5\beta_0t_0^2Q_0 \equiv B^2 - pA^2 \not\equiv 0 \pmod{2}.
\end{align*}
Thus $B^2 - pA^2 + 4p^5\beta_0t_0^2Q_0 \in \bZ_2^{\times}$, and hence we deduce that $C \in \bZ_2$. Thus we see that
\begin{align*}
\alpha = \alpha_0 + 2p^2C \equiv \alpha_0 \not\equiv 0 \pmod{2}.
\end{align*}
Therefore $\alpha \in \bZ_2^{\times}$, and hence $(\alpha, \beta, \gamma)$ satisfies $(A4)$.
Now we prove that $(\alpha, \beta, \gamma)$ satisfies $(A3)$. By what we have proved above, we know that $\alpha, \beta, \gamma \in \bZ_2^{\times}$. This implies that $\alpha, \beta, \gamma \ne 0$. Let $P$ and $Q$ be the rational numbers defined by $(\ref{eqn-P})$ and $(\ref{eqn-Q})$, respectively. One knows that $Q = Q_0 \ne 0$. Since $\alpha, \beta, \gamma \in \bZ_2^{\times}$, it follows that $P \in \bZ_2$. Hence we deduce that
\begin{equation*}
P \equiv p\alpha^2 \not\equiv 0 \pmod{2},
\end{equation*}
which proves that $P \in \bZ_2^{\times}$. Note that $P \ne 0$ since $P \in \bZ_2^{\times}$.
Let $\cQ_1 \subset \bP^2_{\bQ}$ be the conic defined by
\begin{equation*}
\cQ_1 :pU^2 - V^2 - \beta PQ T^2= 0.
\end{equation*}
We prove that the point $\P := (u, v, t) \in \bQ^3$ belongs to $\cQ_1(\bQ)$, where $u$ and $v$ are defined in $(C3)$ and $t := t_0$. Indeed, since $\beta = \beta_0$, $\gamma = \gamma_0$ and $Q = Q_0$, we deduce from $(\ref{eqn-P})$ that
\begin{align*}
-\beta PQt^2 = -\beta_0t_0^2Q_0(p(\alpha_0 + 2p^2C)^2 + 2\beta_0^2 - 2p\gamma_0^2) = -(4p^5\beta_0t_0^2Q_0)C^2 - (4p^3\alpha_0\beta_0t_0^2Q_0)C - (\beta_0P_0Q_0)t_0^2.
\end{align*}
Hence
\begin{align*}
pu^2 - v^2 -\beta PQt^2 &= p(u_0 + AC)^2 - (v_0 + BC)^2 -(4p^5\beta_0t_0^2Q_0)C^2 - (4p^3\alpha_0\beta_0t_0^2Q_0)C - (\beta_0P_0Q_0)t_0^2 \\
&= \left(pA^2 - B^2 - 4p^5\beta_0t_0^2Q_0 \right)C^2 + \left(2pu_0A - 2v_0B - 4p^3\alpha_0\beta_0t_0^2Q_0\right)C + (pu_0^2 - v_0^2 - \beta_0P_0Q_0t_0^2).
\end{align*}
Since $(u_0, v_0, t_0)$ belongs to $\cQ_1^{(\alpha_0, \beta_0, \gamma_0)}(\bQ)$, we see that
\begin{equation*}
pu_0^2 - v_0^2 - \beta_0P_0Q_0t_0^2 = 0.
\end{equation*}
Hence it follows from from $(\ref{eqn-C})$ that
\begin{align*}
pu^2 - v^2 -\beta PQt^2 = \left(pA^2 - B^2 - 4p^5\beta_0t_0^2Q_0 \right)C^2 + \left(2pu_0A - 2v_0B - 4p^3\alpha_0\beta_0t_0^2Q_0\right)C = 0.
\end{align*}
Thus $\P \in \cQ_1(\bQ)$. Since $\cQ_1$ is a nonsingular conic in $\bP^2_{\bQ}$, $\cQ_1(\bQ) \ne \emptyset$ and $uvt \ne 0$, it follows that $(\alpha, \beta, \gamma)$ satisfies $(A3)$.
We now prove that $(\alpha, \beta, \gamma)$ satisfies $(A5)$. Indeed, we have shown that $(\alpha, \beta, \gamma)$ satisfies $(A3), (A4)$. This implies that $\alpha, \beta, \gamma \in \bZ_p$. By Lemma \ref{p-divisibility-beta-lemma}, we know that there is a rational number $\bar{\beta} \in \bQ$ such that $\beta = p\bar{\beta}$ and $\bar{\beta} \in \bZ_p$. Similarly, since $(\alpha_0, \beta_0, \gamma_0)$ satisfies $(A3)$ and $(A4)$, there is a rational number $\bar{\beta}_0$ such that $\beta_0 = p\bar{\beta}_0$ and $\bar{\beta}_0 \in \bZ_p$. Since $\beta = \beta_0$, we deduce that $\bar{\beta} = \bar{\beta}_0$.
Let $P_1$ and $Q_1$ be the rational numbers defined in Remark \ref{p-divisibility-P-Q-remark} and
let $P_1^{(0)}$ and $Q_1^{(0)}$ be the rational numbers defined by the same equations as $P_1$, $Q_1$, respectively
in Remark \ref{p-divisibility-P-Q-remark} with $(\alpha_0, \bar{\beta_0}, \gamma_0)$ in the role of $(\alpha, \bar{\beta}, \gamma)$. By assumption, one knows that the triple $(\alpha_0, \beta_0, \gamma_0)$ satisfies $(A5)$, that is,
\begin{equation*}
\gamma_0 Q_1^{(0)} + bdP_1^{(0)} \equiv 0 \pmod{p^2}.
\end{equation*}
We will prove that
\begin{equation*}
\gamma Q_1 + bdP_1 \equiv 0 \pmod{p^2}.
\end{equation*}
Indeed, one can check that
\begin{align*}
P_1 = \alpha^2 + 2p\bar{\beta}^2 - 2\gamma^2 = 4p^4C^2 + 4p^2\alpha_0C + P_1^{(0)}
\end{align*}
and $Q_1 = Q_1^{(0)}$. Since $\alpha, \bar{\beta}, \gamma$ are in $\bZ_p$, we deduce that $P_1 \in \bZ_p$. Recall that $C \in \bZ_p$. Hence
\begin{align*}
P_1 = 4p^4C^2 + 4p^2\alpha_0C + P_1^{(0)} \equiv P_1^{(0)} \pmod{p^2},
\end{align*}
and thus we deduce that
\begin{equation*}
\gamma Q_1 + bd P_1 \equiv \gamma_0 Q_1^{(0)} + bdP_1^{(0)} \equiv 0 \pmod{p^2}.
\end{equation*}
Therefore $(\alpha, \beta, \gamma)$ satisfies $(A5)$.
Finally, since $(\alpha_0, \beta_0, \gamma_0)$ satisfies $(B1)$, we see that
\begin{align*}
bd - \bar{\beta}\gamma = bd -\bar{\beta}_0\gamma_0 \equiv 0 \pmod{4}.
\end{align*}
Thus $(\alpha, \beta, \gamma)$ satisfies $(B1)$, which proves our contention.
\end{proof}
\begin{lemma}
\label{infinitude-(A,B)-lemma}
Let $(p, b, d)$ be a triple of integers satisfying $(A1)$ and $(A2)$. Assume that there is a triple $(\alpha_0, \beta_0, \gamma_0) \in \bQ^3$ satisfying $(A3)$, $(A4)$, $(A5)$ and $(B1)$.
Let $(u_0, v_0, t_0) \in \bZ^3$ be a point on the conic $\cQ_1^{(\alpha_0, \beta_0, \gamma_0)}$ such that $u_0v_0t_0 \ne 0$ and $\gcd(u_0, v_0, t_0) = 1$, where $P_0$, $Q_0$ and the conic
$\cQ_1^{(\alpha_0, \beta_0, \gamma_0)}$ are defined as in Lemma \ref{infinitude-alpha-beta-gamma-lemma}. Let $\I$ be the set defined by
\begin{align*}
\I := \left\{(A, B) \in \bQ^2: (A, B) \; \; \text{satisfies} \; \; (C1), (C2), (C3) \; \; \text{in Lemma \ref{infinitude-alpha-beta-gamma-lemma}} \right\}.
\end{align*}
Then $\I$ is of infinite cardinality.
\end{lemma}
\begin{proof}
Let $B_0$ be an integer such that $\gcd(B_0, 2p) = 1$. For each $x \in \bZ$, define $B = 2px + B_0$. We see that $B \in \bZ_2^{\times}$ and $B \in \bZ_p^{\times}$. The latter implies that $B \ne 0$. Let $A = 0$, and let $C$ be the rational number defined by $(\ref{eqn-C})$. Define
\begin{align*}
u &:= u_0 + AC = u_0, \\
v &:= v_0 + BC.
\end{align*}
By assumption, we know that $u = u_0 \ne 0$. Assume that $v = 0$. Since $B \ne 0$, it follows from $(\ref{eqn-C})$ and the definition of $v$ that
\begin{equation*}
C = -\dfrac{v_0}{B} = \dfrac{- 2v_0B - 4p^3\alpha_0\beta_0t_0^2Q_0}{B^2 + 4p^5\beta_0t_0^2Q_0}.
\end{equation*}
Hence we deduce that $B$ is a zero of the quadratic polynomial $\cB(T) \in \bQ[T]$, where $\cB(T)$ is defined by
\begin{equation}
\label{B(T)-polynomial}
\cB(T) := v_0T^2 + (4p^3\alpha_0\beta_0t_0^2Q_0)T - 4p^5\beta_0v_0t_0^2Q_0.
\end{equation}
Hence upon letting $T_1$ and $T_2$ be the zeros of $\cB(T)$, we deduce that $(0, B)$ satisfies $(C3)$ in Lemma \ref{infinitude-alpha-beta-gamma-lemma} if and only if $B \ne T_1$ and $B \ne T_2$. The latter holds if and only if $x \ne \dfrac{T_1 - B_0}{2p}$ and $x \ne \dfrac{T_2 - B_0}{2p}$. This implies that if $T_1, T_2 \not\in \bZ$, then $(0, B)$ automatically satisfies $(C3)$ for any integer $x \in \bZ$. Furthermore we see that $B^2 - pA^2 = B^2 \in \bZ_2^{\times}$. Hence $(0, B)$ satisfies $(C1)$ and $(C2)$. Thus $\J$ is a subset of $\I$, where $\J$ is defined by
\begin{align*}
\J := \left\{(0, B) : x \in \bZ, \; \; x \ne \dfrac{T_1 - B_0}{2p} \; \; \text{and} \; \; x \ne \dfrac{T_2 - B_0}{2p} \right\}.
\end{align*}
Since $\J$ is of infinite cardinality, so is $\I$. Hence our contention follows.
\end{proof}
Using Lemma \ref{infinitude-alpha-beta-gamma-lemma} and Lemma \ref{infinitude-(A,B)-lemma}, we prove the main result in this section.
\begin{lemma}
\label{Lemma-the-infinitude-of-p-b-d-alpha-beta-gamma-satisfying-A1-A5-and-B1}
There are infinitely many sextuples $(p, b, d, \alpha, \beta, \gamma)$ satisfying $(A1)-(A5)$ and $(B1)$.
\end{lemma}
\begin{proof}
Assume that there is a sextuple $(p, b, d, \alpha_0, \beta_0, \gamma_0)$ satisfying $(A1)-(A5)$ and $(B1)$. Let $(u_0, v_0, t_0) \in \bZ^3$ be a point on the conic $\cQ_1^{(\alpha_0, \beta_0, \gamma_0)}$ such that $u_0v_0t_0 \ne 0$ and $\gcd(u_0, v_0, t_0) = 1$, where $P_0$, $Q_0$ and the conic $\cQ_1^{(\alpha_0, \beta_0, \gamma_0)}$ are defined as in Lemma \ref{infinitude-alpha-beta-gamma-lemma}. Let $\J$ be the set defined in the proof of Lemma \ref{infinitude-(A,B)-lemma}. We construct an infinite sequence $(0, B_n)_{n \in \bZ_{\ge 0}}$ of elements of $\J$ as follows.
Let $(0, B_1)$ be an arbitrary element of $\J$, and assume that the elements $(0, B_i)$ of $\J$ with $1 \le i \le n$ are already constructed.
Since $\J$ is infinite, we can choose an element $(0, B_{n + 1})$ of $\J$ such that $B_{n + 1} \ne B_i$ for $1 \le i \le n$ and $B_{n + 1}$
is not a zero of any of the polynomials $\H_i(T)$ for $1 \le i \le n$, where for each $1 \le i \le n$, $\H_i(T)$ is defined by
\begin{align}
\label{Hi-polynomial}
\H_i(T) = (v_0B_i + 2p^3\alpha_0\beta_0t_0^2Q_0)T + 2p^3\alpha_0\beta_0t_0^2Q_0B_i - 4p^5\beta_0v_0t_0^2Q_0 \in \bQ[T].
\end{align}
Indeed, we see that $2p^3\alpha_0\beta_0t_0^2Q_0B_i - 4p^5\beta_0v_0t_0^2Q_0 \ne 0$ for every $1 \le i \le n$; otherwise, there is an integer $1 \le i \le n$ such that
\begin{equation*}
\alpha_0B_i = 2p^2v_0.
\end{equation*}
Hence $\alpha_0B_i \not\in \bZ_p^{\times}$, which is a contradiction since $\alpha_0$ and $B_i$ are in $\bZ_p^{\times}$. Hence $\H_i(T)$ is nonzero and of degree at most $1$ for each $1 \le i \le n$. Thus $\H_i(T)$ has at most one zero in $\bZ$ for each $1 \le i \le n$; hence, excluding these $n$ zeros (if existing) and the integers $B_i$ for $1 \le i \le n$ out of the infinite set $\J$, one can choose an element $(0, B_{n+1})$ as desired. Therefore we have inductively constructed an infinite sequence $\left\{(0, B_{n})\right\}_{n \ge 1}$ of elements of $\J$. We contend that for any two distinct members $(0, B_m)$ and $(0, B_n)$ of the sequence with $m < n$, the triples $(\alpha_m, \beta_0, \gamma_0)$ and $(\alpha_n, \beta_0, \gamma_0)$ are distinct, that is, $\alpha_m \ne \alpha_n$, where
\begin{align*}
\alpha_m &:= \alpha_0 + 2p^2C_{(m)}, \\
\alpha_n &:= \alpha_0 + 2p^2C_{(n)},
\end{align*}
and $C_{(m)}$, $C_{(n)}$ are defined as in $(\ref{eqn-C})$ with $(0, B_m)$ and $(0, B_n)$ in the role of $(A, B)$, respectively. Assume the contrary, that is, $\alpha_m = \alpha_n$. It follows that
\begin{equation*}
\dfrac{- 2v_0B_m - 4p^3\alpha_0\beta_0t_0^2Q_0}{B_m^2 + 4p^5\beta_0t_0^2Q_0} = C_{(m)} = C_{(n)} = \dfrac{- 2v_0B_n - 4p^3\alpha_0\beta_0t_0^2Q_0}{B_n^2 + 4p^5\beta_0t_0^2Q_0}.
\end{equation*}
Hence we deduce that
\begin{equation*}
2(B_n - B_m)\left(\left(v_0B_m + 2p^3\alpha_0\beta_0t_0^2Q_0\right)B_n + 2p^3\alpha_0\beta_0t_0^2Q_0B_m - 4p^5\beta_0v_0t_0^2Q_0\right) = 0.
\end{equation*}
Since $B_n \ne B_m$, we deduce that $B_n$ is a zero of $\H_m(T)$, where $\H_m(T)$ is defined by $(\ref{Hi-polynomial})$, which is a contradiction to the choice of $B_n$. Thus we have shown that there are infinitely many sextuples $(p, b, d, \alpha, \beta, \gamma)$ satisfying $(A1)-(A5)$ and $(B1)$ provided that there exists one sextuple $(p, b, d, \alpha_0, \beta_0, \gamma_0)$ satisfying $(A1)-(A5)$ and $(B1)$. On the other hand, in the proof of part $(i)$ of Theorem \ref{algebraic-family-hyperelliptic-curve-hp-theorem} below, we will show that the sextuple
$(p, b, d, \alpha_0, \beta_0, \gamma_0) = (29, 1, 3, 7, 261, 15)$ satisfies $(A1)-(A5)$ and $(B1)$, and hence our contention follows.
\end{proof}
\section{Algebraic families of hyperelliptic curves violating the Hasse principle}
\label{algebraic-families-hyperelliptic-curve-hp-section}
Let $n$ be an integer such that $n > 5$ and $n \not\equiv 0 \pmod{4}$. In this section, using the results in the last section, we will show how to construct algebraic families of hyperelliptic curves of genus $n$ satisfying HP1 and HP2. We begin by proving the following useful lemma.
\begin{lemma}
\label{algebraic-family-hyperelliptic-curve-lemma}
Let $\S$ be a finite set of primes, and let $\G(t) \in \bQ(t)$ be a nonzero rational function. Let $\Z$ be the finite set of rational zeros and poles of $\G(t)$, that is, $\Z$ consists of the rational numbers $z \in \bQ$ for which $\G(z)$ is either zero or infinity. For any $z \in \Z$, let $a_z, b_z$ be integers such that $b_z \ne 0$, $\gcd(a_z, b_z) = 1$ and $z = a_z/b_z$. Assume that the following is true.
\begin{itemize}
\item [(D)] let $z$ be any element in $\Z$ such that $a_z \ne 0$. Then $a_z \not\equiv 0 \pmod{l}$ for each prime $l \in \S$.
\end{itemize}
Then there is a rational function $\F(t) \in \bQ(t)$ such that the following are true.
\begin{itemize}
\item [(1)] $\F(t) \in \bZ_l^{\times}$ for each prime $l \in \S$ and each $t \in \bQ$, and
\item [(2)] $\G(\F(t))$ is defined (that is, not infinity) and nonzero for each $t \in \bQ$.
\end{itemize}
\end{lemma}
\begin{proof}
We consider the following two cases.
$\star$ \textit{Case 1. $\Z$ is non-empty.}
By the Chinese Remainder Theorem, there exists an integer $\epsilon$ such that $\epsilon \equiv 2 \pmod{4}$ and $\epsilon$ is a quadratic non-residue in $\bF_l^{\times}$ for each odd prime $l \in \S$ with $l \ne 2$. Let $p_0$ be an odd prime such that the following are true.
\begin{itemize}
\item [(i)] $p_0 \not\in \S$;
\item [(ii)] $b_z \not\equiv 0 \pmod{p_0}$ for every $z \in \Z$; and
\item [(iii)] let $z$ be any element in $\Z$ such that $a_z \ne 0$. Then $a_z \not\equiv 0 \pmod{p_0}$.
\end{itemize}
For each $z \in \Z$, we define
\begin{equation}
\label{algebraic-families-Di-equation}
D_z := p_0b_z\text{sign}(a_z) \prod_{\substack{w \in \Z\setminus \{z\}}}\max(1, |a_w|) \in \bZ,
\end{equation}
where $\text{sign}(\cdot)$ denotes the usual sign function of $\bR$, that is, $\text{sign}(x) = 1$ if $x \ge 0$, and $\text{sign}(x) = -1$ if $x < 0$. We see that $|D_z| \ge p_0 \ge 3$ for each $z \in \Z$. This implies that $|D_z - 1| \ge 1$ for every $z \in \Z$. We will prove that the rational function $\F(t) \in \bQ(t)$ defined by
\begin{equation}
\label{rational-function-F(t)-nonempty-case}
\F(t) := \left(p_0\prod_{z \in \Z}\max(1, |a_z|)\right)\left(1 + \dfrac{4\left(\prod_{l \in \S, l \ne 2}l \right)\left(\prod_{z \in \Z}(D_z - 1)\right)}{t^2 - p_0^2\epsilon} \right)
\end{equation}
satisfies $(1)$ and $(2)$ in Lemma \ref{algebraic-family-hyperelliptic-curve-lemma}. Indeed, let $t = \dfrac{t_1}{t_2}$, where $t_1, t_2 \in \bZ$, $t_2 \ne 0$ and $\gcd(t_1, t_2) = 1$. For each prime $l$, denote by $v_l$ the $l$-adic valuation of $\bQ_l$. For each prime $l \in \S$ with $l \ne 2$, one knows that
\begin{align*}
v_l\left(\dfrac{1}{t^2 - p_0^2\epsilon}\right) = v_l(t_2^2) - v_l(t_1^2 - t_2^2p_0^2\epsilon).
\end{align*}
Assume that $t_1^2 - t_2^2p_0^2\epsilon \equiv 0 \pmod{l}$. Since $p_0 \ne l$ and $\epsilon$ is a quadratic non-residue in $\bF_l^{\times}$, it follows that $t_1 \equiv t_2 \equiv 0 \pmod{l}$, which is a contradiction. Hence we deduce that $v_l(t_1^2 - t_2^2p_0^2\epsilon) = 0$. Thus we see that
\begin{align*}
v_l\left(\dfrac{1}{t^2 - p_0^2\epsilon}\right) = v_l(t_2^2) \ge 0.
\end{align*}
Therefore $\dfrac{1}{t^2 - p_0^2\epsilon} \in \bZ_l$, and hence we deduce that
\begin{align*}
1 + \dfrac{4\left(\prod_{l \in \S, l \ne 2}l \right)\left(\prod_{z \in \Z}(D_z - 1)\right)}{t^2 - p_0^2\epsilon} \in 1 + l\bZ_l.
\end{align*}
By assumption $(D)$ and the choice of $p_0$, one also knows that $p_0\prod_{z \in \Z}\max(1, |a_z|) \in \bZ_l^{\times}$. Hence it follows that for each prime $l \in \S$ with $l \ne 2$, $\F(t) \in \bZ_l^{\times}$ for every $t \in \bQ$. Thus we have shown that if $2 \not\in \S$, then $\F(t)$ satisfies $(1)$ in Lemma \ref{algebraic-family-hyperelliptic-curve-lemma}. Hence it remains to show that if $2 \in \cS$, then $\F(t) \in \bZ_2^{\times}$ for every $t \in \bQ$.
Let first assume that $t_1$ is even. Hence $t_2$ is odd, and hence one sees that $t_1^2 - t_2^2p_0^2\epsilon \equiv 2 \pmod{4}$. Thus $v_2(t_1^2 - t_2^2p_0^2\epsilon) = 1$. Hence it follows that
\begin{align*}
v_2\left(\dfrac{2}{t^2 - p_0^2\epsilon}\right) = 1 + v_2(t_2^2) - v_2(t_1^2 - t_2^2p_0^2\epsilon) = 0,
\end{align*}
which implies that $\dfrac{2}{t^2 - p_0^2\epsilon} \in \bZ_2$.
Now assume that $t_1$ is odd. Since $\epsilon$ is even, one sees that $t_1^2 - t_2^2p_0^2\epsilon$ is odd. Hence it follows that
\begin{align*}
v_2\left(\dfrac{2}{t^2 - p_0^2\epsilon}\right) = 1 + v_2(t_2^2) - v_2(t_1^2 - t_2^2p_0^2\epsilon) = 1 + v_2(t_2^2) \ge 1.
\end{align*}
Thus we have shown that $\dfrac{2}{t^2 - p_0^2\epsilon} \in \bZ_2$. By the definition of $\F(t)$ and assumption $(D)$, we deduce that $\F(t) \in \bZ_2^{\times}$. Hence $\F(t)$ satisfies $(1)$ in Lemma \ref{algebraic-family-hyperelliptic-curve-lemma}.
Now we prove that $\F(t)$ satisfies $(2)$ in Lemma \ref{algebraic-family-hyperelliptic-curve-lemma}. Since $z$ is a rational zero or pole of $\G(t)$ for each $z \in \Z$, we deduce that $\G(\F(t))$ is defined (that is, not infinity) and nonzero for each $t \in \bQ$ if $\F(t) \ne z$ for every $z \in \Z$ and every $t \in \bQ$.
Assume that there is a rational number $t \in \bQ$ such that $\F(t) = z$ for some $z = a_z/b_z \in \Z$. We consider the following two subcases.
$\bullet$ \textit{Subcase 1. $a_z \ne 0$.}
We see that $\max(1, |a_z|) = |a_z|$. Hence it follows that
\begin{align*}
D_z\left(1 + \dfrac{4\left(\prod_{l \in \S, l \ne 2}l \right)\left(\prod_{w \in \Z}(D_w - 1)\right)}{t^2 - p_0^2\epsilon} \right) = 1.
\end{align*}
Upon multiplying by $t^2 - p_0^2\epsilon$ both sides of the identity above and simplifying, we deduce that
\begin{align*}
t^2 = p_0^2\epsilon - 4D_z\left(\prod_{\substack{l \in \S, l \ne 2}}l \right)\left(\prod_{\substack{w \in \Z\setminus \{z\}}}(D_w - 1)\right).
\end{align*}
Hence it follows from $(\ref{algebraic-families-Di-equation})$ that
\begin{align*}
t^2 = p_0\left(p_0\epsilon - 4b_z\text{sign}(a_z)\left(\prod_{w \in \Z\setminus \{z\}}\max(1, |a_w|)\right)\left(\prod_{l \in \S, l \ne 2}l \right)
\left(\prod_{w \in \Z\setminus \{z\}}(D_w - 1)\right)\right).
\end{align*}
This implies that $t \in \bZ$ and $t \equiv 0 \pmod{p_0}$. Hence $v_{p_0}(t^2) = 2v_{p_0}(t) \ge 2$. Thus we deduce that
\begin{align*}
p_0\epsilon - 4b_z\text{sign}(a_z)\left(\prod_{w \in \Z\setminus \{z\}}\max(1, |a_w|)\right)\left(\prod_{l \in \S, l \ne 2}l \right)
\left(\prod_{w \in \Z\setminus \{z\}}(D_w - 1)\right) \equiv 0 \pmod{p_0}.
\end{align*}
Hence
\begin{align*}
4b_z\text{sign}(a_z)\left(\prod_{w \in \Z\setminus \{z\}}\max(1, |a_w|)\right)\left(\prod_{l \in \S, l \ne 2}l \right)
\left(\prod_{w \in \Z\setminus \{z\}}(D_w - 1)\right) \equiv 0 \pmod{p_0}.
\end{align*}
By $(\ref{algebraic-families-Di-equation})$, one knows that $D_w \equiv 0 \pmod{p_0}$ for every $w \in \Z$. Hence
\begin{align*}
\prod_{w \in \Z\setminus \{z\}}(D_w - 1) \equiv (-1)^{m - 1} \pmod{p_0}.
\end{align*}
Thus we deduce that
\begin{align*}
(-1)^{m - 1}4b_z\text{sign}(a_z)\left(\prod_{w \in \Z\setminus \{z\}}\max(1, |a_w|)\right)\left(\prod_{l \in \S, l \ne 2}l \right) \equiv 0 \pmod{p_0},
\end{align*}
which is a contradiction to the choice of $p_0$. Therefore $\F(t) \ne z$ for every $t \in \bQ$.
$\bullet$ \textit{Subcase 2. $a_z = 0$.}
We see that $\F(t) = \dfrac{a_z}{b_z} = 0$. Hence we deduce from the definition of $\F(t)$ that
\begin{align*}
t^2 = p_0^2\epsilon - 4\left(\prod_{l \in \S, l \ne 2}l \right)\left(\prod_{w \in \Z}(D_w - 1)\right).
\end{align*}
This implies that $t \in \bZ$. Hence we deduce that
\begin{align*}
t^2 = p_0^2\epsilon \pmod{l}
\end{align*}
for each prime $l \in \S$ with $l \ne 2$. Since $\epsilon$ is a quadratic non-residue in $\bF_l^{\times}$, it follows that $t \equiv p_0 \equiv 0 \pmod{l}$, which is a contradiction to the choice of $p_0$. Thus, in any event, $\F(t) \ne z$ for every $t \in \bQ$. Therefore $\F(t)$ satisfies $(2)$ in Lemma \ref{algebraic-family-hyperelliptic-curve-lemma}.
$\star$ \textit{Case 2. $\Z = \emptyset$.}
In this case, let $\epsilon$ be the same as in \textit{Case 1}, and let $p_0$ be an odd prime such that $p_0 \not\in \S$. Let $\F(t) \in \bQ(t)$ be the rational function defined by
\begin{equation}
\label{rational-function-F(t)-empty-case}
\F(t) := 1 + \dfrac{4\left(\prod_{l \in \S, l \ne 2}l \right)}{t^2 - p_0^2\epsilon}.
\end{equation}
Using the same arguments as in \textit{Case 1}, one can show that $\F(t)$ satisfies $(1)$ and $(2)$ in Lemma \ref{algebraic-family-hyperelliptic-curve-lemma}.
\end{proof}
\begin{lemma}
\label{Lemma-Rational-functions-of-degree-4-belong-to-q-Z-q}
Let $\D(t) \in \bQ(t)$ be a nonzero rational function of the form
\begin{align*}
\D(t) := \dfrac{at^4 + bt^2 + c}{dt^4 + et^2 + f},
\end{align*}
where $a, b, c, d, e, f$ are integers. Let $q$ be an odd prime. Assume that there exists an integer $t_0$ such that
\begin{align*}
at_0^4 + bt_0^2 + c \equiv 0 \pmod{q},
\end{align*}
\begin{align*}
at_0^4 + bt_0^2 + c \not\equiv 0 \pmod{q^2},
\end{align*}
and
\begin{align*}
dt_0^4 + et_0^2 + f \not\equiv 0 \pmod{q}.
\end{align*}
Then there exists a rational function $\Gamma(t) \in \bQ(t)$ such that for every $t \in \bQ$, $\D(\Gamma(t))$ belongs to $q\bZ_q$, but does not belong to $q^2\bZ_q$.
\end{lemma}
\begin{proof}
Let $\epsilon$ be an integer such that $\epsilon$ is a quadratic non-residue in $\bF_q^{\times}$. Let $q_0$ be an odd prime such that $q_0 \ne q$. We will show that the rational function $\Gamma(t) \in \bQ(t)$ defined by
\begin{align}
\label{Definition-The-definition-of-Gamma-function}
\Gamma(t) := t_0 + \dfrac{q^2}{t^2 - q_0^2\epsilon}
\end{align}
satisfies the assertions in Lemma \ref{Lemma-Rational-functions-of-degree-4-belong-to-q-Z-q}.
Since $\epsilon$ is not a square in $\bF_q^{\times}$, it follows that $t^2 - q_0^2\epsilon$ is non-zero for each $t \in \bQ$, and hence $\Gamma(t)$ is well-defined for every $t \in \bQ$.
We now prove that $\Gamma(t)$ belongs to $t_0 + q^2\bZ_q$ for each $t \in \bQ$. Indeed, take any rational number $t$, and write $t = \dfrac{t_1}{t_2}$, where $t_1, t_2$ are integers such that $t_2 \ne 0$ and $\gcd(t_1, t_2) = 1$. We see that
\begin{align*}
v_q\left(\dfrac{1}{t^2 - q_0^2\epsilon}\right) = v_q\left(\dfrac{t_2^2}{t_1^2 - q_0^2\epsilon t_2^2}\right) = v_q(t_2^2) - v_q(t_1^2 - q_0^2\epsilon t_2^2).
\end{align*}
If $t_2 \equiv 0 \pmod{q}$, then it follows that $t_1 \not\equiv 0 \pmod{q}$. Hence we deduce that
\begin{align*}
v_q(t_1^2 - q_0^2\epsilon t_2^2) = \min\left(v_q(t_1^2), v_q(q_0^2\epsilon t_2^2)\right) = \min\left(0, v_q(t_2^2)\right) = 0,
\end{align*}
and thus
\begin{align*}
v_q\left(\dfrac{1}{t^2 - q_0^2\epsilon}\right) = v_q(t_2^2) - v_q(t_1^2 - q_0^2\epsilon t_2^2) = 2v_q(t_2) \ge 2.
\end{align*}
Therefore $\dfrac{1}{t^2 - q_0^2\epsilon}$ belongs to $\bZ_q$, and hence it follows from $(\ref{Definition-The-definition-of-Gamma-function})$ that $\Gamma(t)$ belongs to $t_0 + q^2\bZ_q$.
If $t_2 \not\equiv 0 \pmod{q}$, it follows that $v_q(t_2^2) = 0$. We contend that $t_1^2 - q_0^2\epsilon t_2^2 \not\equiv 0 \pmod{q}$. Assume the contrary, that is, $t_1^2 - q_0^2\epsilon t_2^2 \equiv 0 \pmod{q}$. Since $t_2 \not\equiv 0 \pmod{q}$ and $q_0 \ne q$, we deduce that
\begin{align*}
\epsilon \equiv \left(\dfrac{t_1}{q_0t_2}\right)^2 \pmod{q},
\end{align*}
which contradicts the choice of $\epsilon$. This contradiction establishes that $t_1^2 - q_0^2\epsilon t_2^2 \not\equiv 0 \pmod{q}$, and thus
\begin{align*}
v_q\left(\dfrac{1}{t^2 - q_0^2\epsilon}\right) = v_q(t_2^2) - v_q(t_1^2 - q_0^2\epsilon t_2^2) = 0.
\end{align*}
Therefore $\dfrac{1}{t^2 - q_0^2\epsilon}$ belongs to $\bZ_q^{\times}$, and hence it follows from $(\ref{Definition-The-definition-of-Gamma-function})$ that $\Gamma(t)$ belongs to $t_0 + q^2\bZ_q$.
Since $\Gamma(t)$ belongs to $t_0 + q^2\bZ_q$, we see that
\begin{align*}
a(\Gamma(t))^4 + b(\Gamma(t))^2 + c \equiv at_0^4 + bt_0^2 + c \equiv 0 \pmod{q},
\end{align*}
\begin{align*}
a(\Gamma(t))^4 + b(\Gamma(t))^2 + c \equiv at_0^4 + bt_0^2 + c \not\equiv 0 \pmod{q^2},
\end{align*}
and
\begin{align*}
d(\Gamma(t))^4 + e(\Gamma(t))^2 + f \equiv dt_0^4 + et_0^2 + f \not\equiv 0 \pmod{q}
\end{align*}
for each $t \in \bQ$. The last congruence shows that
\begin{align*}
\dfrac{1}{c(\Gamma(t))^4 + d(\Gamma(t))^2 + e}
\end{align*}
belongs to $\bZ_q^{\times}$, and hence we deduce that for every $t \in \bQ$,
\begin{align*}
\D(\Gamma(t)) = \dfrac{a(\Gamma(t))^4 + b(\Gamma(t))^2 + c}{d(\Gamma(t))^4 + e(\Gamma(t))^2 + f}
\end{align*}
belongs to $q\bZ_q$, but does not belong to $q^2\bZ_q$. Thus our contention follows.
\end{proof}
The following two examples will be used in the proof of the main theorem in this section.
\begin{example}
\label{Example-The-first-rational-function-D1-related-to-the-construction-of-algebraic-families}
Let $\D_1(T) \in \bQ(T)$ be the rational function defined by
\begin{align}
\label{Definition-The-rational-function-D1-T}
\D_1(T) := \frac{45588894173298T^4 - 1641200890885920T^2 + 14770814323798008}{-5477180725633679T^4 + 197178506122812676T^2 - 1774606555105302716},
\end{align}
and define
\begin{align}
\label{Definition-The-rational-function-D1-ast-T}
\D_1^{\ast}(T) : = 7 + 1682\D_1(T) = \frac{-38340254920051483T^4 + 1380250355610428708T^2 -12422263806891130444}{5477180725633679T^4 - 197178506122812676T^2 + 1774606555105302716}.
\end{align}
Let $T_0 = 0$, and let $q = 31$. Since
\begin{align*}
12422263806891130444 = 2^2\cdot7^3\cdot 31 \cdot 433\cdot 3299 \cdot 10589 \cdot 19309,
\end{align*}
it follows that
\begin{align*}
v_q(-12422263806891130444) = v_{31}(-12422263806891130444) = 1,
\end{align*}
and hence we deduce that
\begin{align*}
&-38340254920051483T_0^4 + 1380250355610428708T_0^2 - 12422263806891130444 \\
&= -12422263806891130444 \equiv 0 \pmod{q}
\end{align*}
and
\begin{align*}
&-38340254920051483T_0^4 + 1380250355610428708T_0^2 - 12422263806891130444 \\
&= -12422263806891130444 \not\equiv 0 \pmod{q^2}.
\end{align*}
Since
\begin{align*}
1774606555105302716 = 2^2 \cdot 7^2 \cdot 47 \cdot 192640746320593,
\end{align*}
we see that
\begin{align*}
&5477180725633679T_0^4 - 197178506122812676T_0^2 + 1774606555105302716 \\
&= 1774606555105302716 \not\equiv 0 \pmod{q}.
\end{align*}
Let $\epsilon = 3$, and let $q_0 = 5$. Following the proof of Lemma \ref{Lemma-Rational-functions-of-degree-4-belong-to-q-Z-q}, we define the rational function $\Gamma_1(T) \in \bQ(T)$ by $(\ref{Definition-The-definition-of-Gamma-function})$, that is,
\begin{align}
\label{Definition-The-definition-of-the-rational-function-Gamma1}
\Gamma_1(T) := T_0 + \dfrac{q^2}{T^2 - q_0^2\epsilon} = \dfrac{961}{T^2 - 75}.
\end{align}
Applying Lemma \ref{Lemma-Rational-functions-of-degree-4-belong-to-q-Z-q} with $\D_1^{\ast}(T)$ in the role of $\D(t)$, we deduce that the rational function $\D_1^{\ast}(\Gamma_1(T))$ belongs to $31\bZ_{31}$, but does not belong to $31^2\bZ_{31}$, where
\begin{align}
\label{Definition-The-rational-function-D-1-ast-of-Gamma-1}
\D_1^{\ast}(\Gamma_1(T)) = \frac{\Sigma_{1, 1}(T)}{\Sigma_{1, 2}(T)}
\end{align}
with
\begin{align}
\label{Equation-Sigma-1-1}
\Sigma_{1, 1}(T) &= -12422263806891130444 T^8 + 3726679142067339133200 T^6 \\
&+ 855438785181123078355868 T^4 - 170240958125426027001880200 T^2 \nonumber \\
&- 25922975674046723162225380003 \nonumber
\end{align}
and
\begin{align}
\label{Equation-Sigma-1-2}
\Sigma_{1, 2}(T) &= 1774606555105302716 T^8 - 532381966531590814800 T^6 \\
&- 122205519918242118687196 T^4 + 24320125111216714469579400 T^2 \nonumber \\
&+ 3703283999134302153081910439. \nonumber
\end{align}
\end{example}
\begin{example}
\label{Example-The-second-rational-function-D2-related-to-the-construction-of-algebraic-families}
Let $\D_2(T) \in \bQ(T)$ be the rational function defined by
\begin{align}
\label{Definition-The-rational-function-D2-T}
\D_2(T) := \frac{-64380401708754T^4 + 2317693623118880T^2 - 20859235062503544}{407097080892401T^4 - 14655494912126204T^2 + 131899454209147204},
\end{align}
and define
\begin{align}
\label{Definition-The-rational-function-D2-ast-T}
\D_2^{\ast}(T) : = 133 + 1682\D_2(T) = \frac{-54143923915434895T^4 + 1949179850773171028T^2 - 17542605965314382876}{407097080892401T^4 - 14655494912126204T^2 + 131899454209147204}.
\end{align}
Let $T_0 = 0$, and let $q = 11$. Since
\begin{align*}
17542605965314382876 = 2^2 \cdot 7 \cdot 11 \cdot 56956512874397347,
\end{align*}
it follows that
\begin{align*}
-54143923915434895T_0^4 + 1949179850773171028T_0^2 - 17542605965314382876 \equiv 0 \pmod{11}
\end{align*}
but
\begin{align*}
-54143923915434895T_0^4 + 1949179850773171028T_0^2 - 17542605965314382876 \not\equiv 0 \pmod{11^2}.
\end{align*}
Since
\begin{align*}
131899454209147204 \equiv 8 \not\equiv 0 \pmod{11},
\end{align*}
we deduce that
\begin{align*}
407097080892401T^4 - 14655494912126204T^2 + 131899454209147204 \not\equiv 0 \pmod{7}.
\end{align*}
Let $\epsilon = 7$, and let $q_0 = 3$. Following the proof of Lemma \ref{Lemma-Rational-functions-of-degree-4-belong-to-q-Z-q}, we define the rational function $\Gamma_2(T) \in \bQ(T)$ by $(\ref{Definition-The-definition-of-Gamma-function})$, that is,
\begin{align}
\label{Definition-The-definition-of-the-rational-function-Gamma2}
\Gamma_2(T) := T_0 + \dfrac{q^2}{T^2 - q_0^2\epsilon} = \dfrac{121}{T^2 - 63}.
\end{align}
Applying Lemma \ref{Lemma-Rational-functions-of-degree-4-belong-to-q-Z-q} with $\D_2^{\ast}(T)$ in the role of $\D(t)$, we deduce that the rational function $\D_2^{\ast}(\Gamma_2(T))$ belongs to $11\bZ_{11}$, but does not belong to $11^2\bZ_{11}$, where
\begin{align}
\label{Definition-The-rational-function-D-2-ast-of-Gamma-2}
\D_2^{\ast}(\Gamma_2(T)) = \frac{\Sigma_{2, 1}(T)}{\Sigma_{2, 2}(T)}
\end{align}
with
\begin{align}
\label{Equation-Sigma-2-1}
\Sigma_{2, 1}(T) &= -17542605965314382876 T^8 + 4420736703259224484752 T^6 \\
&-389221676262826716788116 T^4 + 13950123258644442355341240 T^2 \nonumber \\
&-174687125980796870729105719 \nonumber
\end{align}
and
\begin{align}
\label{Equation-Sigma-2-2}
\Sigma_{2, 2}(T) &= 131899454209147204 T^8 - 33238662460705095408 T^6 \\
&+ 2926482501528191763292 T^4 - 104888292579475114826088 T^2 \nonumber \\
&+ 1313439132893945928914009. \nonumber
\end{align}
\end{example}
We recall the following separability criterion \cite{dq-separable-polynomials}.
\begin{theorem}
\label{the-criterion-for-separability-of-certain-polynomials-theorem}
Let $n, m, k$ be positive integers, and let $a, b, c, d, e$ be rational numbers such that $a \ne 0$. Let $p$ be an odd prime such that $a, b, c, d, e$ belong to $\bZ_p$ and $a \equiv 0 \pmod{p}$. Let $F(x) \in \bQ[x]$ be the polynomial defined by
\begin{align}
\label{the-polynomial-F-equation}
F(x) := ax^{2n + 2} + (bx^{2m} + c)(dx^{2k} + e).
\end{align}
Define
\begin{align*}
n_1 &:= (m + k)(v_p(a) - v_p(bd)) + m + k - 1, \\
n_2 &:= (m + k)(v_p(a) - v_p(b)) + m - 1, \\
n_3 &:= (m + k)(v_p(a) - v_p(d)) + k - 1, \\
n_4 &:= (m + k)v_p(a) - 1, \\
n_5 &:= v_p(a) - v_p(bd) + m + k - 1.
\end{align*}
Suppose that the following are true:
\begin{itemize}
\item [(S1)] $n > m + k - 1$ and $n > \max(n_1, n_2, n_3, n_4, n_5)$.
\item [(S2)] $ce \not\equiv 0 \pmod{p}$, $km \not\equiv 0 \pmod{p}$, and $b^ke^m + (-1)^{m + k + 1}c^kd^m \not\equiv 0 \pmod{p}$.
\end{itemize}
Then $F$ is separable: that is, it has exactly $2n + 2$ distinct roots in $\bC$.
\end{theorem}
\begin{remark}
Theorem \ref{the-criterion-for-separability-of-certain-polynomials-theorem} is a very mild generalization of Theorem 2.1 in \cite{dq-separable-polynomials}. The only difference between Theorem \ref{the-criterion-for-separability-of-certain-polynomials-theorem} and \cite[Theorem 2.1]{dq-separable-polynomials} is that in \cite[Theorem 2.1]{dq-separable-polynomials}, the author assumed that $a, b, c, d, e$ are integers, whereas in Theorem \ref{the-criterion-for-separability-of-certain-polynomials-theorem} presented here, we only need the assumption that $a, b, c, d, e$ belong to $\bZ_p$. Upon looking closely the proof of \cite[Theorem 2.1]{dq-separable-polynomials}, we see that it is sufficient to assume that $a, b, c, d, e$ are in $\bZ_p$, and hence Theorem \ref{the-criterion-for-separability-of-certain-polynomials-theorem} follows immediately from \cite[Theorem 2.1]{dq-separable-polynomials}.
\end{remark}
Using Theorem \ref{the-criterion-for-separability-of-certain-polynomials-theorem}, we prove the following corollaries that play a key role in constructing algebraic families of curves violating the Hasse principle.
\begin{corollary}
\label{Corollary-The-separability-of-the-first-algebraic-family-of-hyperelliptic-curves}
We maintain the same notation as in Example \ref{Example-The-first-rational-function-D1-related-to-the-construction-of-algebraic-families}. Let $\D_1(T)$, $\D_1^{\ast}(T), \Gamma_1(T) \in \bQ(T)$ be the rational functions defined by $(\ref{Definition-The-rational-function-D1-T})$, $(\ref{Definition-The-rational-function-D1-ast-T})$, $(\ref{Definition-The-definition-of-the-rational-function-Gamma1})$, respectively in Example \ref{Example-The-first-rational-function-D1-related-to-the-construction-of-algebraic-families}. Let $n$ be a positive integer such that $n > 5$. For each rational number $T \in \bQ$, let $\cP_{1, T}(x) \in \bQ[x]$ be the polynomial of degree $2n + 2$ given by
\begin{align}
\label{Definition-The-definition-of-the-polynomial-P-1-T}
&\cP_{1, T}(x) := 118579927725(\D_1^{\ast}(\Gamma_1(T)))^2x^{2n + 2} \\
&+ \left(2\left(29(\D_1^{\ast}(\Gamma_1(T)))^2 + 123192\right)x^2 - 16689645\right)\left(261\left(29(\D_1^{\ast}(\Gamma_1(T)))^2 + 123192\right)x^2 - 33379290\right), \nonumber
\end{align}
where the composition $\D_1^{\ast}(\Gamma_1(T))$ of $\D_1^{\ast}(T)$ and $\Gamma_1(T)$ is given by $(\ref{Definition-The-rational-function-D-1-ast-of-Gamma-1})$ in Example \ref{Example-The-first-rational-function-D1-related-to-the-construction-of-algebraic-families}. Then for each $T \in \bQ$, the polynomial $\cP_{1, T}(x)$ is separable: that is, it has exactly $2n + 2$ distinct roots in $\bC$.
\end{corollary}
\begin{proof}
Throughout the proof, we maintain the same notation as in Theorem \ref{the-criterion-for-separability-of-certain-polynomials-theorem}. Take any rational number $T \in \bQ$, and define
\begin{align*}
a &:= 118579927725(\D_1^{\ast}(\Gamma_1(T)))^2, \\
b &:= 2\left(29(\D_1^{\ast}(\Gamma_1(T)))^2 + 123192\right), \\
c &:= -16689645, \\
d &:= 261\left(29(\D_1^{\ast}(\Gamma_1(T)))^2 + 123192\right), \\
e &:= -33379290.
\end{align*}
Let $p = 31$, and let $m = k = 1$. Since $118579927725 \equiv 27 \not\equiv 0 \pmod{31}$, $123192 \equiv 29 \not\equiv 0 \pmod{31}$, it follows from Example \ref{Example-The-first-rational-function-D1-related-to-the-construction-of-algebraic-families} that
\begin{align*}
v_{p}(a) = v_{31}\left( 118579927725(\D_1^{\ast}(\Gamma_1(T)))^2 \right) = 2v_{31}\left(\D_1^{\ast}(\Gamma_1(T)) \right) = 2,
\end{align*}
\begin{align*}
v_p(b) = v_{31}\left(2\left(29(\D_1^{\ast}(\Gamma_1(T)))^2 + 123192\right)\right) = v_{31}(123192) = 0,
\end{align*}
and
\begin{align*}
v_p(d) = v_{31}\left(261\left(29(\D_1^{\ast}(\Gamma_1(T)))^2 + 123192\right)\right) = v_{31}(123192) = 0.
\end{align*}
We see that
\begin{align*}
n_1 &:= (m + k)(v_p(a) - v_p(bd)) + m + k - 1 = 2v_p(a) + 1 = 5, \\
n_2 &:= (m + k)(v_p(a) - v_p(b)) + m - 1 = 2v_p(a) = 4, \\
n_3 &:= (m + k)(v_p(a) - v_p(d)) + k - 1 = 2v_p(a) = 4, \\
n_4 &:= (m + k)v_p(a) - 1 = 2v_p(a) - 1 = 3, \\
n_5 &:= v_p(a) - v_p(bd) + m + k - 1 = 2 + 1 = 3,
\end{align*}
and hence
\begin{align*}
\max(n_1, n_2, n_3, n_4, n_5) = 5.
\end{align*}
By assumption, we know that
\begin{align*}
n > 5 = \max(n_1, n_2, n_3, n_4, n_5),
\end{align*}
and hence condition $(S1)$ in Theorem \ref{the-criterion-for-separability-of-certain-polynomials-theorem} is satisfied.
It is obvious that $km = 1 \not\equiv 0 \pmod{31}$ and
\begin{align*}
ce = (-16689645)\cdot (-33379290) \equiv 25 \not\equiv 0 \pmod{31}.
\end{align*}
Furthermore, since $\D_1^{\ast}(\Gamma_1(T))$ belongs to $31\bZ_{31}$, we deduce that
\begin{align*}
b^ke^m + (-1)^{m + k + 1}c^kd^m &= be - cd \\
&= \left(2\left(29(\D_1^{\ast}(\Gamma_1(T)))^2 + 123192\right)\right)(-33379290) \\
&- (-16689645)\left(261\left(29(\D_1^{\ast}(\Gamma_1(T)))^2 + 123192\right)\right) \\
&\equiv 2\cdot 123192 \cdot (-33379290) + 16689645 \cdot 261 \cdot 123192 \\
&\equiv 12 \not\equiv 0 \pmod{31}.
\end{align*}
Therefore condition $(S2)$ in Theorem \ref{the-criterion-for-separability-of-certain-polynomials-theorem} is satisfied, and hence the polynomial $\cP_{1, T}(x)$ is separable. Since $T$ is an arbitrary rational number, our contention follows.
\end{proof}
\begin{corollary}
\label{Corollary-The-separability-of-the-second-algebraic-family-of-hyperelliptic-curves}
We maintain the same notation as in Example \ref{Example-The-second-rational-function-D2-related-to-the-construction-of-algebraic-families}. Let $\D_2(T)$, $\D_2^{\ast}(T), \Gamma_2(T) \in \bQ(T)$ be the rational functions defined by $(\ref{Definition-The-rational-function-D2-T})$, $(\ref{Definition-The-rational-function-D2-ast-T})$, $(\ref{Definition-The-definition-of-the-rational-function-Gamma2})$, respectively in Example \ref{Example-The-second-rational-function-D2-related-to-the-construction-of-algebraic-families}. Let $n$ be a positive integer such that $n > 5$. For each rational number $T \in \bQ$, let $\cP_{2, T}(x) \in \bQ[x]$ be the polynomial of degree $2n + 2$ given by
\begin{align}
\label{Definition-The-definition-of-the-polynomial-P-2-T}
&\cP_{2, T}(x) := 84898109(\D_2^{\ast}(\Gamma_2(T)))^2 x^{2n + 2} \\
&+ \left(2\left(29(\D_2^{\ast}(\Gamma_2(T)))^2 - 40600\right)x^2 + 49619\right)\left(261\left(29(\D_2^{\ast}(\Gamma_2(T)))^2 - 40600\right)x^2 + 99238\right),
\end{align}
where the composition $\D_2^{\ast}(\Gamma_2(T))$ of $\D_2^{\ast}(T)$ and $\Gamma_2(T)$ is given by $(\ref{Definition-The-rational-function-D-2-ast-of-Gamma-2})$ in Example \ref{Example-The-second-rational-function-D2-related-to-the-construction-of-algebraic-families}. Then for each $T \in \bQ$, the polynomial $\cP_{2, T}(x)$ is separable: that is, it has exactly $2n + 2$ distinct roots in $\bC$.
\end{corollary}
\begin{proof}
Throughout the proof, we maintain the same notation as in Theorem \ref{the-criterion-for-separability-of-certain-polynomials-theorem}. Take any rational number $T \in \bQ$, and define
\begin{align*}
a &:= 84898109(\D_2^{\ast}(\Gamma_2(T)))^2, \\
b &:= 2\left(29(\D_2^{\ast}(\Gamma_2(T)))^2 - 40600\right), \\
c &:= 49619, \\
d &:= 261\left(29(\D_2^{\ast}(\Gamma_2(T)))^2 - 40600\right), \\
e &:= 99238.
\end{align*}
Let $p = 11$, and let $m = k = 1$. Since $84898109 \equiv 10 \not\equiv 0 \pmod{11}$, $40600 \equiv 10 \not\equiv 0 \pmod{11}$, it follows from Example \ref{Example-The-second-rational-function-D2-related-to-the-construction-of-algebraic-families} that
\begin{align*}
v_{p}(a) = v_{11}\left(84898109(\D_2^{\ast}(\Gamma_2(T)))^2 \right) = 2v_{11}\left(\D_2^{\ast}(\Gamma_2(T))\right) = 2,
\end{align*}
\begin{align*}
v_p(b) = v_{11}\left(2\left(29(\D_2^{\ast}(\Gamma_2(T)))^2 - 40600\right)\right) = v_{11}(40600) = 0,
\end{align*}
and
\begin{align*}
v_p(d) = v_{31}\left(261\left(29(\D_2^{\ast}(\Gamma_2(T)))^2 - 40600\right)\right) = v_{11}(40600) = 0.
\end{align*}
We see that
\begin{align*}
n_1 &:= (m + k)(v_p(a) - v_p(bd)) + m + k - 1 = 2v_p(a) + 1 = 5, \\
n_2 &:= (m + k)(v_p(a) - v_p(b)) + m - 1 = 2v_p(a) = 4, \\
n_3 &:= (m + k)(v_p(a) - v_p(d)) + k - 1 = 2v_p(a) = 4, \\
n_4 &:= (m + k)v_p(a) - 1 = 2v_p(a) - 1 = 3, \\
n_5 &:= v_p(a) - v_p(bd) + m + k - 1 = 2 + 1 = 3,
\end{align*}
and hence
\begin{align*}
\max(n_1, n_2, n_3, n_4, n_5) = 5.
\end{align*}
By assumption, we know that
\begin{align*}
n > 5 = \max(n_1, n_2, n_3, n_4, n_5),
\end{align*}
and hence condition $(S1)$ in Theorem \ref{the-criterion-for-separability-of-certain-polynomials-theorem} is satisfied.
It is obvious that $km = 1 \not\equiv 0 \pmod{11}$ and
\begin{align*}
ce = 49619 \cdot 99238 \equiv 8 \not\equiv 0 \pmod{11}.
\end{align*}
Since $\D_2^{\ast}(\Gamma_2(T))$ belongs to $11\bZ_{11}$, we deduce that
\begin{align*}
b^ke^m + (-1)^{m + k + 1}c^kd^m &= be - cd \\
&= \left(2\left(29(\D_2^{\ast}(\Gamma_2(T)))^2 - 40600\right)\right)(99238) \\
&- (49619)\left(261\left(29(\D_2^{\ast}(\Gamma_2(T)))^2 - 40600\right)\right) \\
&\equiv 2\cdot (-40600) \cdot 99238 - (49619) \cdot 261 \cdot (-40600) \\
&\equiv 8 \not\equiv 0 \pmod{11}.
\end{align*}
Therefore condition $(S2)$ in Theorem \ref{the-criterion-for-separability-of-certain-polynomials-theorem} is satisfied, and hence the polynomial $\cP_{2, T}(x)$ is separable. Since $T$ is an arbitrary rational number, our contention follows.
\end{proof}
For the rest of this section, let
\begin{align*}
\A_1 := \left\{n \in \bZ: n > 5, \; \; n \not\equiv 0 \pmod{4} \; \; \text{and} \; \; n \not\equiv 21 \pmod{29} \right\},
\end{align*}
and
\begin{align*}
\A_2 := \left\{n \in \bZ: n > 5, \; \; n \not\equiv 0 \pmod{4} \; \; \text{and} \; \; n \not\equiv 8 \pmod{29} \right\}.
\end{align*}
We see that
\begin{align}
\label{Equation-the-union-of-A1-and-A2-is-the-set-of-all-n-greater-than-5-and-not-equiv-to-zero-mod-4}
\A_1 \cup \A_2 = \left\{n \in \bZ: n > 5 \; \; \text{and} \; \; n \not\equiv 0 \pmod{4} \right\}.
\end{align}
We now prove the main theorem in this section.
\begin{theorem}
\label{algebraic-family-hyperelliptic-curve-hp-theorem}
For each $n \in \A_1$ and each rational number $T \in \bQ$, let $\cP_{1, T}(x) \in \bQ[x]$ be the polynomial of degree $2n + 2$ defined by $(\ref{Definition-The-definition-of-the-polynomial-P-1-T})$ in Corollary \ref{Corollary-The-separability-of-the-first-algebraic-family-of-hyperelliptic-curves}. For each $n \in \A_2$ and each rational number $T \in \bQ$, let $\cP_{2, T}(x) \in \bQ[x]$ be the polynomial of degree $2n + 2$ defined by $(\ref{Definition-The-definition-of-the-polynomial-P-2-T})$ in Corollary \ref{Corollary-The-separability-of-the-second-algebraic-family-of-hyperelliptic-curves}. Then
\begin{itemize}
\item [(i)] For each $n \in \A_1$ and each rational number $T \in \bQ$, the hyperelliptic curve $\cC_{n, T, (29, 1, 3)}^{(7, 261, 15)}$ of genus $n$ satisfies HP1 and HP2, where $\cC_{n, T, (29, 1, 3)}^{(7, 261, 15)}$ is the smooth projective model of the affine curve defined by
\begin{align*}
\cC_{n, T, (29, 1, 3)}^{(7, 261, 15)} : z^2 = \cP_{1, T}(x).
\end{align*}
\item [(ii)] For each $n \in \A_2$ and each rational number $T \in \bQ$, the hyperelliptic curve $\cC_{n, T, (29, 1, 3)}^{(133, 29, 27)}$ of genus $n$ satisfies HP1 and HP2, where $\cC_{n, T, (29, 1, 3)}^{(133, 29, 27)}$ is the smooth projective model of the affine curve defined by
\begin{align*}
\cC_{n, T, (29, 1, 3)}^{(133, 29, 27)} &: z^2 = \cP_{2, T}(x).
\end{align*}
\end{itemize}
\end{theorem}
\begin{remark}
\label{Remark-The-above-theorem-implies-the-main-theorem-in-the-paper}
By $(\ref{Equation-the-union-of-A1-and-A2-is-the-set-of-all-n-greater-than-5-and-not-equiv-to-zero-mod-4})$ and Theorem \ref{algebraic-family-hyperelliptic-curve-hp-theorem}, we see that Theorem \ref{Theorem-The-main-theorem} follows immediately.
\end{remark}
\begin{proof}
Throughout the proof of Theorem \ref{algebraic-family-hyperelliptic-curve-hp-theorem}, we will use the same notation as in Theorem \ref{hyperelliptic-curves-theorem} and Lemma \ref{infinitude-alpha-beta-gamma-lemma}. We first prove that part $(i)$ of Theorem \ref{algebraic-family-hyperelliptic-curve-hp-theorem} holds.
Let $n$ be any integer such that $n \in \A_1$. Let $(p, b, d, \alpha_0, \beta_0, \gamma_0) = (29, 1, 3, 7, 261, 15)$. We see that $\bar{\beta_0} = 9$. One can check that the sextuple $(p, b, d, \alpha_0, \beta_0, \gamma_0)$ satisfies $(A1)-(A5)$ and $(B1)$. Indeed, $(A1), (A2), (A4)$, $(A5)$ and $(B1)$ are obvious. It remains to prove that $(p, b, d, \alpha_0, \beta_0, \gamma_0)$ satisfies $(A3)$. By $(\ref{eqn-P})$ and $(\ref{eqn-Q})$, we know that
\begin{align*}
P_0 &= 124613, \\
Q_0 &= -63945.
\end{align*}
The conic $\cQ_1^{(7, 261, 15)}$ in $(A3)$ of Theorem \ref{hyperelliptic-curves-theorem} defined by
\begin{equation*}
\cQ_1^{(7, 261, 15)} : 29U^2 - V^2 + 2079746732385T^2 = 0
\end{equation*}
has a point $(u_0, v_0, t_0) = (166257, 3020031, 2)$, and hence $(p, b, d, \alpha_0, \beta_0, \gamma_0)$ satisfies $(A3)$.
Let $\S := \{2, 29\}$, and let $\C_1(T)$ be the rational function in $\bQ(T)$ defined by the same equation $(\ref{eqn-C})$ of $C$ with $(0, T)$ in the role of $(A, B)$, that is,
\begin{align*}
\C_1(T) := \dfrac{- 2v_0T - 4p^3\alpha_0\beta_0t_0^2Q_0}{T^2 + 4p^5\beta_0t_0^2Q_0} = \dfrac{-6040062T + 45588900213360}{T^2 - 5477180725633680}.
\end{align*}
Let $\G_1(T) \in \bQ(T)$ be the rational function defined by
\begin{align}
\label{G1(T)-rational-function-eqn}
\G_1(T) = v_0 + T\C_1(T) = \frac{-3020031T^2 + 45588900213360T - 16541255584016208244080}{T^2 - 5477180725633680}.
\end{align}
Since the numerator and denominator of $\G_1(T)$ are irreducible polynomials over $\bQ$, the set $\Z_1$ of rational zeros and poles of $\G_1(T)$ is empty. Hence applying Lemma \ref{algebraic-family-hyperelliptic-curve-lemma} for the triple $\left(\S, \G_1(T), \Z_1 \right)$, we know that $\F_1(T)$ satisfies $(1)$ and $(2)$ in Lemma \ref{algebraic-family-hyperelliptic-curve-lemma}, where $\F_1(T)$ is the rational function defined by $(\ref{rational-function-F(t)-empty-case})$ in \textit{Case 2} of Lemma \ref{algebraic-family-hyperelliptic-curve-lemma} with $(p_0, \epsilon) = (3, 2)$ and $\left(\S, \G_1(T), \Z_1 \right)$ in the role of $(\S, \G(T), \Z)$, that is,
\begin{align*}
\F_1(T) := 1 + \dfrac{4\left(\prod_{l \in \S, l \ne 2}l \right)}{T^2 - p_0^2\epsilon} = 1 + \dfrac{116}{T^2 - 18} = \dfrac{T^2 + 98}{T^2 - 18}.
\end{align*}
Let $\Gamma_1(T) \in \bQ(T)$ be the rational function defined by $(\ref{Definition-The-definition-of-the-rational-function-Gamma1})$ in Example \ref{Example-The-first-rational-function-D1-related-to-the-construction-of-algebraic-families}. Recall that
\begin{align*}
\Gamma_1(T) := \dfrac{961}{T^2 - 75}.
\end{align*}
It is known that $\Gamma_1(T)$ is well-defined for each rational number $T \in \bQ$.
Let $(A, B) = (0, \F_1(\Gamma_1(T)))$. By Lemma \ref{algebraic-family-hyperelliptic-curve-lemma}, we know that $(0, \F_1(T))$ satisfies $(C1)$ and $(C2)$ in Lemma \ref{infinitude-alpha-beta-gamma-lemma} for each $T \in \bQ$. Thus it follows that $(A, B) = (0, \F_1(\Gamma_1(T)))$ satisfies $(C1)$ and $(C2)$ in Lemma \ref{infinitude-alpha-beta-gamma-lemma} for each $T \in \bQ$.
Let $\D_1(\Gamma_1(T))$ be the rational function in $\bQ(T)$ defined by the same equation $(\ref{eqn-C})$ of $C$ with $(0, \F_1(\Gamma_1(T)))$ in the role of $(A, B)$, that is,
\begin{align}
\label{D1-eqn-algebraic-family}
&\D_1(\Gamma_1(T)) := \C_1(\F_1(\Gamma_1(T))) \\
&= \frac{45588894173298(\Gamma_1(T))^4 - 1641200890885920(\Gamma_1(T))^2 + 14770814323798008}{-5477180725633679(\Gamma_1(T))^4 + 197178506122812676(\Gamma_1(T))^2 - 1774606555105302716}. \nonumber
\end{align}
Note that $(0, \F_1(\Gamma_1(T)))$ satisfies $(C1)$ and $(C2)$. Hence using the same arguments as in the proof of Lemma \ref{infinitude-alpha-beta-gamma-lemma}, one knows that $\D_1(\Gamma_1(T)) \in \bZ_{29}$ for each $T \in \bQ$.
We see that
\begin{align*}
u := u_0 + A\D_1(\Gamma_1(T)) = u_0 = 166257 \ne 0
\end{align*}
for each $T \in \bQ$. Furthermore it follows from part $(2)$ of Lemma \ref{algebraic-family-hyperelliptic-curve-lemma} that
\begin{align*}
v = v_0 + B\D_1(\Gamma_1(T)) = \G_1(\F_1(\Gamma_1(T)))
\end{align*}
is defined (that is, not infinity) and non-zero for each $T \in \bQ$. Hence $(0, \F_1(\Gamma_1(T)))$ satisfies $(C3)$ in Lemma \ref{infinitude-alpha-beta-gamma-lemma} for each $T \in \bQ$.
Set
\begin{align*}
\alpha &:= \alpha_0 + 2p^2\D_1(\Gamma_1(T)) = 7 + 1682\D_1(\Gamma_1(T)) = \D_1^{\ast}(\Gamma_1(T)), \\
\beta &:= \beta_0 = 261, \\
\gamma &:= \gamma_0 = 15,
\end{align*}
where $\D_1^{\ast}(\Gamma_1(T))$ is defined by $(\ref{Definition-The-rational-function-D-1-ast-of-Gamma-1})$ in Example \ref{Example-The-first-rational-function-D1-related-to-the-construction-of-algebraic-families}. Recall from $(\ref{Definition-The-rational-function-D-1-ast-of-Gamma-1})$ in Example \ref{Example-The-first-rational-function-D1-related-to-the-construction-of-algebraic-families} that
\begin{align*}
\D_1^{\ast}(\Gamma_1(T)) = \frac{\Sigma_{1, 1}(T)}{\Sigma_{1, 2}(T)},
\end{align*}
where $\Sigma_{1, 1}(T), \Sigma_{1, 2}(T)$ are defined by $(\ref{Equation-Sigma-1-1})$, $(\ref{Equation-Sigma-1-2})$, respectively.
By Lemma \ref{infinitude-alpha-beta-gamma-lemma}, we know that $(\alpha, \beta, \gamma)$ satisfies $(A1)-(A5)$ and $(B1)$ for each $T \in \bQ$. By $(\ref{eqn-P})$ and $(\ref{eqn-Q})$, we know that
\begin{align*}
P &= 29(7 + 1682\D_1(\Gamma_1(T)))^2 + 123192 = 29(\D_1^{\ast}(\Gamma_1(T)))^2 + 123192, \\
Q &= Q_0 = -63945.
\end{align*}
It is not difficult to see that for each $T \in \bQ$, the curve $\cC_{n, T, (29, 1, 3)}^{(7, 261, 15)}$ defined in part $(i)$ of Theorem \ref{algebraic-family-hyperelliptic-curve-hp-theorem} is the smooth projective model of the affine curve defined by $(\ref{eqn-hyperelliptic-curve-C})$.
By Corollary \ref{Corollary-The-separability-of-the-first-algebraic-family-of-hyperelliptic-curves}, we know that $\cP_{1, T}(x)$ is separable for each $T \in \bQ$, and hence we deduce that condition $(S)$ in Theorem \ref{hyperelliptic-curves-theorem} is true. Since $\D_1(\Gamma_1(T)) \in \bZ_{29}$ for each $T \in \bQ$, we see that
\begin{align*}
-2\left(\dfrac{\gamma}{\alpha}\right)^2 \equiv 21 \pmod{29}.
\end{align*}
Since $n \in \A_1$, we deduce that
\begin{align*}
n \not\equiv -2\left(\dfrac{\gamma}{\alpha}\right)^2 \pmod{29},
\end{align*}
and thus $(A6)$ holds if $n$ is odd and $(B2)$ holds if $n \equiv 2 \pmod{4}$. By Corollary \ref{hyperelliptic-curves-hp-corollary} and Corollary \ref{n-even-hyperelliptic-curve-hp-corollary}, we deduce that for each $n \in \A_1$ and each $T \in \bQ$, $\cC_{n, T, (29, 1, 3)}^{(7, 261, 15)}$ satisfies HP1 and HP2.
We now prove that part $(ii)$ of Theorem \ref{algebraic-family-hyperelliptic-curve-hp-theorem} is true. We will use the same notation as in the proof of part $(i)$ as long as they do not cause any confusion. We will use the same arguments as in the proof of part $(i)$ to construct an algebraic family of hyperelliptic curves of genus $n$ satisfying HP1 and HP2 for each $n \in \A_2$.
Let $n$ be any integer such that $n \in \A_2$. Let $(p, b, d, \alpha_0, \beta_0, \gamma_0) = (29, 1, 3, 133, 29, 27)$. We see that $\bar{\beta_0} = 1$. One can check that the sextuple $(p, b, d, \alpha_0, \beta_0, \gamma_0)$ satisfies $(A1)-(A5)$ and $(B1)$. Indeed $(A1), (A2), (A4), (A5)$ and $(B1)$ are obvious. It remains to prove that the sextuple satisfies $(A3)$. By $(\ref{eqn-P})$ and $(\ref{eqn-Q})$, we know that
\begin{align*}
P_0 &= 472381, \\
Q_0 &= 1711.
\end{align*}
The conic $\cQ_1^{(133, 29, 27)}$ in $(A3)$ of Theorem \ref{hyperelliptic-curves-theorem} defined by
\begin{equation*}
\cQ_1^{(133, 29, 27)} : 29U^2 - V^2 - 23439072839T^2 = 0
\end{equation*}
has a point $(u_0, v_0, t_0) = (728799, 3613777, 10)$, and thus $(p, b, d, \alpha_0, \beta_0, \gamma_0)$ satisfies $(A3)$.
Let $\S := \{2, 29\}$, and let $\C_2(T) \in \bQ(T)$ be the rational function defined by the same equation $(\ref{eqn-C})$ of $C$ with $(0, T)$ in the role of $(A, B)$, that is,
\begin{align*}
\C_2(T) = \dfrac{- 2v_0T - 4p^3\alpha_0\beta_0t_0^2Q_0}{T^2 + 4p^5\beta_0t_0^2Q_0} = \dfrac{-7227554T - 64380394481200}{T^2 + 407097080892400}.
\end{align*}
Let $\G_2(T) \in \bQ(T)$ be the rational function defined by
\begin{equation}
\label{G2(T)-rational-function-eqn}
\G_2(T) = v_0 + T\C_2(T) = \frac{-3613777T^2 - 64380394481200T + 1471158067696094594800}{T^2 + 407097080892400}.
\end{equation}
Since the numerator and denominator of $\G_2(T)$ are irreducible polynomials over $\bQ$, the set $\Z_2$ of rational zeroes and poles of $\G_2(T)$ is empty. Hence applying Lemma \ref{algebraic-family-hyperelliptic-curve-lemma} for the triple $\left(\S, \G_2(T), \Z_2 \right)$, we know that $\F_2(T)$ satisfies $(1)$ and $(2)$ in Lemma \ref{algebraic-family-hyperelliptic-curve-lemma}, where $\F_2(T)$ is the rational function defined by $(\ref{rational-function-F(t)-empty-case})$ in \textit{Case 2} of Lemma \ref{algebraic-family-hyperelliptic-curve-lemma} with $(p_0, \epsilon) = (3, 2)$ and $\left(\S, \G_2(T), \Z_2 \right)$ in the role of $(\S, \G(T), \Z)$, that is,
\begin{align*}
\F_2(T) := 1 + \dfrac{4\left(\prod_{l \in \S, l \ne 2}l \right)}{T^2 - p_0^2\epsilon} = 1 + \dfrac{116}{T^2 - 18} = \dfrac{T^2 + 98}{T^2 - 18}.
\end{align*}
Let $\Gamma_2(T) \in \bQ(T)$ be the rational function defined by $(\ref{Definition-The-definition-of-the-rational-function-Gamma2})$ in Example \ref{Example-The-second-rational-function-D2-related-to-the-construction-of-algebraic-families}. Recall that
\begin{align*}
\Gamma_2(T) := \dfrac{121}{T^2 - 63}.
\end{align*}
It is known that $\Gamma_2(T)$ is well-defined for each rational number $T \in \bQ$.
Let $(A, B) = (0, \F_2(\Gamma_2(T)))$. By Lemma \ref{algebraic-family-hyperelliptic-curve-lemma}, we know that $(0, \F_2(T))$ satisfies $(C1)$ and $(C2)$ in Lemma \ref{infinitude-alpha-beta-gamma-lemma} for each $T \in \bQ$. Thus it follows that $(A, B) = (0, \F_2(\Gamma_2(T)))$ satisfies $(C1)$ and $(C2)$ in Lemma \ref{infinitude-alpha-beta-gamma-lemma} for each $T \in \bQ$.
Let $\D_2(\Gamma_2(T))$ be the rational function in $\bQ(T)$ defined by the same equation $(\ref{eqn-C})$ of $C$ with $(0, \F_2(\Gamma_2(T)))$ in the role of $(A, B)$, that is,
\begin{align}
\label{D2-eqn-algebraic-family}
&\D_2(\Gamma_2(T)) := \C_2(\F_2(\Gamma_2(T))) \\
&= \frac{-64380401708754 (\Gamma_2(T))^4 + 2317693623118880 (\Gamma_2(T))^2 - 20859235062503544}{407097080892401 (\Gamma_2(T))^4 - 14655494912126204 (\Gamma_2(T))^2 + 131899454209147204}. \nonumber
\end{align}
Note that $(0, \F_2(\Gamma_2(T)))$ satisfies $(C1)$ and $(C2)$. Hence using the same arguments as in the proof of Lemma \ref{infinitude-alpha-beta-gamma-lemma}, one knows that $\D_2(\Gamma_2(T)) \in \bZ_{29}$ for each $T \in \bQ$.
We see that
\begin{align*}
u := u_0 + A\D_1(\Gamma_1(T)) = u_0 = 728799 \ne 0
\end{align*}
for each $T \in \bQ$. Furthermore it follows from $(2)$ of Lemma \ref{algebraic-family-hyperelliptic-curve-lemma} that
\begin{align*}
v = v_0 + B\D_2(\Gamma_2(T)) = \G_2(\F_2(\Gamma_2(T)))
\end{align*}
is defined (that is, not infinity) and non-zero for each $T \in \bQ$. Hence $(0, \F_2(\Gamma_2(T)))$ satisfies $(C3)$ in Lemma \ref{infinitude-alpha-beta-gamma-lemma} for each $T \in \bQ$.
Set
\begin{align*}
\alpha &:= \alpha_0 + 2p^2\D_2(\Gamma_2(T)) = 133 + 1682\D_2(\Gamma_2(T)) = \D_2^{\ast}(\Gamma_2(T)), \\
\beta &:= \beta_0 = 29, \\
\gamma &:= \gamma_0 = 27,
\end{align*}
where $\D_2^{\ast}(\Gamma_2(T))$ is defined by $(\ref{Definition-The-rational-function-D-2-ast-of-Gamma-2})$ in Example \ref{Example-The-second-rational-function-D2-related-to-the-construction-of-algebraic-families}. Recall from $(\ref{Definition-The-rational-function-D-2-ast-of-Gamma-2})$ in Example \ref{Example-The-second-rational-function-D2-related-to-the-construction-of-algebraic-families} that
\begin{align*}
\D_2^{\ast}(\Gamma_2(T)) = \frac{\Sigma_{2, 1}(T)}{\Sigma_{2, 2}(T)},
\end{align*}
where $\Sigma_{2, 1}(T), \Sigma_{2, 2}(T)$ are defined by $(\ref{Equation-Sigma-2-1})$, $(\ref{Equation-Sigma-2-2})$, respectively.
By Lemma \ref{infinitude-alpha-beta-gamma-lemma}, we know that $(\alpha, \beta, \gamma)$ satisfies $(A1)-(A5)$ and $(B1)$ for each $T \in \bQ$. By $(\ref{eqn-P})$ and $(\ref{eqn-Q})$, we know that
\begin{align*}
P &= 29(133 + 1682\D_2(T))^2 - 40600 = 29(\D_2^{\ast}(\Gamma_2(T)))^2 - 40600, \\
Q &= Q_0 = 1711.
\end{align*}
It is not difficult to see that for each $T \in \bQ$, the curve $\cC_{n, T, (29, 1, 3)}^{(133, 29, 27)}$ defined in part $(ii)$ of Theorem \ref{algebraic-family-hyperelliptic-curve-hp-theorem} is the smooth projective model of the affine curve defined by $(\ref{eqn-hyperelliptic-curve-C})$.
By Corollary \ref{Corollary-The-separability-of-the-second-algebraic-family-of-hyperelliptic-curves}, we know that $\cP_{2, T}(x)$ is separable for each $T \in \bQ$, and hence we deduce that condition $(S)$ in Theorem \ref{hyperelliptic-curves-theorem} is true. Since $\D_2(\Gamma_2(T)) \in \bZ_{29}$ for each $T \in \bQ$, we see that
\begin{align*}
-2\left(\dfrac{\gamma}{\alpha}\right)^2 \equiv 8 \pmod{29}.
\end{align*}
Since $n \in \A_2$, we deduce that
\begin{align*}
n \not\equiv -2\left(\dfrac{\gamma}{\alpha}\right)^2 \pmod{29},
\end{align*}
and thus $(A6)$ holds if $n$ is odd, and $(B2)$ holds if $n \equiv 2 \pmod{4}$. By Corollary \ref{hyperelliptic-curves-hp-corollary} and Corollary \ref{n-even-hyperelliptic-curve-hp-corollary}, we deduce that for each $n \in \A_2$ and each $T \in \bQ$, $\cC_{n, T, (29, 1, 3)}^{(133, 29, 27)}$ satisfies HP1 and HP2.
\end{proof}
\begin{remark}
\label{Remark-Existence-of-algebraic-families-of-curves-of-genus-n-with-n-divisible-by-4}
It seems that the approach used in this paper to construct algebraic families of hyperelliptic curves of genus $n$ violating the Hasse principle does not work if $n \equiv 0 \pmod{4}$. The reason is that the hyperelliptic curves defined in Theorem \ref{hyperelliptic-curves-theorem} might fail to be locally solvable at $2$ if the genus of these curves is divisible by four, which in turn comes from the fact that the hyperelliptic curves in Theorem \ref{hyperelliptic-curves-theorem} lie on the threefolds $\cY$ in Theorem \ref{threefold-hp-theorem} and the local Azumaya invariant at $2$ of the threefolds $\cY$ is $1/2$.
In \cite{dq-Algebraic-families-of-curves-with-n-divisible-by-4}, the author constructs threefolds parameterized by the primes $p \equiv 1 \pmod{8}$ that are counterexamples to the Hasse principle. Furthermore it is shown in \cite{dq-Algebraic-families-of-curves-with-n-divisible-by-4} that there exists an odd prime $q$ such that the local Azumaya algebra at $q$ of the threefolds is $1/2$ and the local Azumaya algebra invariant at $l$ of the threefolds is zero for any prime $l \ne q$. For any hyperelliptic curve $\cC$ of genus $n$ lying on the threefolds in \cite{dq-Algebraic-families-of-curves-with-n-divisible-by-4}, since the local Azumaya algebra invariant at $2$ of the threefolds is zero, the local solvability at $2$ of $\cC$ should not depend on whether $n$ is divisible by $4$. Hence using the approach in this paper, it seems promising that there should exist algebraic families of curves of arbitrary genus violating the Hasse principle. The author hopes to finish this work in the near future.
\end{remark}
\section{Appendix}
\subsection{Proof of Lemma \ref{bright-corn-lemma}}
\label{Subsection-Proof-of-the-Bright-Corn-Lemma}
In this subsection, we sketch a proof of Lemma \ref{bright-corn-lemma}. The proof presented here follows closely from \cite[Proposition 4.17]{bright-Thesis} and \cite[Proposition 2.2.3]{corn-PLMS}. We begin by recalling some facts about the Tate cohomology groups.
Let $G$ be a finite group, and let $A$ be a $G$-module. The \textit{Tate cohomology groups $\widehat{H}^n(G, A)$ with negative and positive exponents} \cite[Chapter VIII]{serre-Local-Fields} are defined by
\begin{align*}
\widehat{H}^n(G, A) :=
\begin{cases}
H^n(G, A) \; &\text{if $n \ge 1$,}\\
A^G/NA \; &\text{if $n = 0$,} \\
_{N}A/I_GA \; &\text{if $n = -1$,} \\
H_{n - 1}(G, A) \; &\text{if $n \le - 2$,}
\end{cases}
\end{align*}
where $A^G$ is the submodule of $A$ consisting of the elements fixed by $G$, $N : A \rightarrow A$ is the endomorphism given by $Na = \sum_{g \in G}ga$ for all $a \in A$, $_{N}A$ is the kernel of $N$, and $I_G$ is the \textit{augmentation ideal} of the group algebra $\bZ[G]$, i.e., it consists of all linear combinations of the elements $g - 1$ with $g \in G$.
We now proceed to prove Lemma \ref{bright-corn-lemma}. We have the exact sequence
\begin{align}
\label{Equation-The-exact-sequence-of-Div-and-Pic}
0 \rightarrow F(\cX_L)^{\times}/L^{\times} \rightarrow \text{Div}(\cX_L) \rightarrow \text{Pic}(\cX_L) \rightarrow 0,
\end{align}
where $\cX_L = \cX \times_{F} L$ and $F(\cX_L)$ denotes the function field of $\cX_L$. Let $\text{Gal}(L/F)$ denote the Galois group of the extension $L/F$. It is well-known \cite[Corollary to Proposition 6, page 133]{serre-Local-Fields} that the Tate cohomology groups with exponents $n$ depend only on the parity of $n$. Hence it follows from the exact sequence $(\ref{Equation-The-exact-sequence-of-Div-and-Pic})$ that we have a commutative diagram
\begin{align}
\label{Equation-The-commutative-diagram-of-Div-and-Pic}
\begin{CD}
\widehat{H}^{-1}(\text{Gal}(L/F), \text{Pic}(\cX_L)) @>>> \widehat{H}^{0}(\text{Gal}(L/F), F(\cX_L)^{\times}/L^{\times}) @>>> \widehat{H}^{0}(\text{Gal}(L/F), \text{Div}(\cX_L)) \\
@VVV @VVV @VVV \\
H^{1}(\text{Gal}(L/F), \text{Pic}(\cX_L)) @>>> H^{2}(\text{Gal}(L/F), F(\cX_L)^{\times}/L^{\times}) @>>> H^{2}(\text{Gal}(L/F), \text{Div}(\cX_L)),
\end{CD}
\end{align}
where the vertical arrows are isomorphisms. By \cite[Chapter V]{deuring-Algebra}, we know that the cyclic algebra $(L/F, f)$ corresponds to the element $f \in H^{2}(\text{Gal}(L/F), F(\cX_L)^{\times}/L^{\times})$, and it thus follows from the diagram $(\ref{Equation-The-commutative-diagram-of-Div-and-Pic})$ that $(L/F, f)$ corresponds to the element $f \in \widehat{H}^{0}(\text{Gal}(L/F), F(\cX_L)^{\times}/L^{\times})$.
By the definition of the Tate cohomology groups, we see that
\begin{align*}
\widehat{H}^{0}(\text{Gal}(L/F), \text{Div}(\cX_L)) = \text{Div}(\cX)/N \text{Div}(\cX_L).
\end{align*}
Thus our contentions follow immediately from the commutative diagram $(\ref{Equation-The-commutative-diagram-of-Div-and-Pic})$.
\section*{Acknowledgements}
I was supported by a postdoctoral fellowship in the Department of Mathematics at University of British Columbia during the time when the paper was prepared.
|
\section{Introduction}
For all terms and definitions, not defined specifically in this paper, we refer to \cite{FH}. Unless mentioned otherwise, all graphs considered here are simple, finite and have no isolated vertices.
The researches on graph labeling problems commenced with the introduction of the concept of number valuations of graphs in \cite{AR1}. Since then, the studies on graph labeling have contributed significantly to the researches in graph theory and associated fields. Graph labeling problems have numerous theoretical and practical applications. Many types of graph labelings are surveyed and listed in \cite{JAG}.
Motivated from certain problems related to social interactions and social networks, Acharya introduced, in \cite{BDA1}, the notion of set-valuation of graphs analogous to the number valuations of graphs. For a non-empty ground set $X$, finite or infinite, the {\em set-valuation} or {\em set-labeling} of a given graph $G$ is an injective function $f:V(G) \to \mathcal{P}(X)$, where $\mathcal{P}(X)$ is the power set of the set $X$.
Also, in \cite{BDA1}, a {\em set-indexer} of a graph $G$ is defined as an injective set-valued function $f:V(G) \to \mathcal{P}(X)$ such that the function $f^{\ast}:E(G)\to \mathcal{P}(X)-\{\emptyset\}$ defined by $f^{\ast}(uv) = f(u ){\ast} f(v)$ for every $uv{\in} E(G)$ is also injective, where $\mathcal{P}(X)$ is the set of all subsets of $X$ and $\ast$ is a binary operation on sets.
Taking the symmetric difference of two sets as the operation between two set-labels of the vertices of $G$, the following theorem was proved in \cite{BDA1}.
\begin{theorem}\label{T-SIG}
\cite{BDA1} Every graph has a set-indexer.
\end{theorem}
Let $A$ and $B$ be two sets. The {\em sum set} of $A$ and $B$ is denoted by $A+B$ and is defined as $A+B=\{a+b : a\in A, b\in B\}$. For any positive integer $n$, an $n$-integral multiple of a set $A$, denoted by $n.A$ and is defined as $n.A=\{na : a\in A\}$. We denote the cardinality of a set $A$ by $|A|$.
In this work, we review the studies on the graphs which admit a particular type of set-labeling called integer additive set-labeling, defined in terms of the sum sets of given sets of non-negative integers.
\section{On Integer Additive Set-Labeled Graphs}
During a personal communication with the second author, Acharya introduced the notion of integer additive set-labeling, using the concept of sum sets of two sets of non-negative integers, as follows.
\begin{definition}{\rm
Let $\mathbb{N}_0$ be the set of all non-negative integers and let $\mathcal{P}(\mathbb{N}_0)$ be its power set. A function $f:V(G)\to \mathcal{P}(\mathbb{N}_0)$, whose associated function $f^+(uv):E(G)\to \mathcal{P}(\mathbb{N}_0)$ is defined by $f^+(uv)=f(u)+f(v)$, is said to be an {\em integer additive set-labeling} if it is injective. A graph $G$ which admits an integer additive set-labeling is called an {\em integer additive set-labeled graph} or {\em integer additive set-valued graph}. }
\end{definition}
The notion of integer additive set-indexers of graphs was first appeared in \cite{GA} as follows.
\begin{definition}\label{DEFN-1}{\rm
An {\em integer additive set-indexer} (IASI) is defined as an injective function $f:V(G)\to \mathcal{P}(\mathbb{N}_0)$ such that the induced function $f^+:E(G) \to \mathcal{P}(\mathbb{N}_0)$ defined by $f^+ (uv) = f(u)+ f(v)=\{a+b: a \in f(u), b \in f(v)\}$ is also injective. A graph $G$ which admits an IASI is called an integer additive set-indexed graph.}
\end{definition}
In this discussion, we denote the associated function of an IASL $f$, defined over the edge set of $G$ by $f^+$.
\begin{definition}\label{DEFN-1a}{\rm
An IASL (or IASI) is said to be {\em $k$-uniform} if $|f^+(e)| = k$ for all $e\in E(G)$. That is, a connected graph $G$ is said to have a $k$-uniform IASL if all of its edges have the same set-indexing number $k$.}
\end{definition}
If either $f(u)$ or $f(v)$ is countably infinite, then their sum set will also be countably infinite and hence the study of the cardinality of the sum set $f(u)+f(v)$ becomes countably infinite. Hence, we restrict our discussion to finite sets of non-negative integers.
Certain studies about integer additive set-indexed graphs have been done in \cite{TMKA}, \cite{GS1}, \cite{GS2} and \cite{GS0}. The following are the important terms and definitions made in these papers.
The cardinality of the set-label of an element (vertex or edge) of a graph $G$ is called the {\em set-indexing number} of that element. If the set-labels of all vertices of $G$ have the same cardinality, then the vertex set $V(G)$ is said to be {\em uniformly set-indexed}. A graph is said to admit {\em $(k,l)$-completely uniform IASI (or IASL)} $f$ if $f$ is a $k$-uniform IASL (or IASI) and $V(G)$ is $l$-uniformly set-indexed.
Analogous to Theorem \ref{T-SIG}, we have proved the following theorem.
\begin{theorem}\label{T-IASL1}
Every graph has an integer additive set-labeling.
\end{theorem}
The proof follows from the fact that the sum set of two sets of non-negative integers is again a set of non-negative integers.
The following theorem was proved in \cite{GS0} by choosing the set-labels of the vertices of a given graph $G$ in a suitable manner.
\begin{theorem}\label{T-IASI1a}
\cite{GS0} Every graph has an integer additive set-indexer.
\end{theorem}
The sum sets of different pairs of sets non-negative integers need not be distinct in all cases. But we can establish the existence of an IASI for any given graph $G$ if we choose the set-labels of the vertices of $G$ suitably in such a way that no two edges of $G$ have the same set-indexing number.
\section{IASLs of Certain Graph Operations}
We know that if $A$ and $B$ are two sets of non-negative integers, then their sum set $A+B$ is also a set of non-negative integers. That is, if $A,B\subset \mathbb{N}_0$, then $A+B\subset \mathbb{N}_0$. This property leads us to the following results on certain operations and products of IASL graphs.
\noindent First, we consider the union of two IASL graphs.
\begin{theorem}
The union of two IASL (or IASI) graphs also admits (induced) IASL (or IASI).
\end{theorem}
\begin{proof}
Let $G_1$ and $G_2$ be two graphs that admit the IASLs $f_1$ and $f_2$ respectively, chosen in such a way that $f_1$ and $f_2$ are the same for the common vertices of $G_1$ and $G_2$, if any. Then, define a function $f$ on $G_1\cup G_2$, such that $f=f_1$ for all vertices in $G_1$ and $f=f_2$ for all vertices in $G_2$ such that the associated function $f^+$ of $f$ is defined as $f^+(uv)=f_1(u)+f_1(v)$ if $uv\in E(G_1)$ and $f^+(uv)=f_2(u)+f_2(v)$ if $uv\in E(G_2)$. Clearly, $f$ is an IASL of $G_1\cup G_2$.
\end{proof}
\noindent Next, in the following theorem, we check whether the join of two IASL graphs is an IASL graph.
\begin{theorem}
The join of two IASL (or IASI) graphs also admits (induced) IASL (or IASI).
\end{theorem}
\begin{proof}
Let $G_1$ and $G_2$ be two graphs that admit the IASLs $f_1$ and $f_2$ respectively. Consider the join $G=G_1+G_2$. Define a function $f$ on $G$ in such a way that $f=f_1$ for all vertices in $G_1$ and $f=f_2$ for all vertices in $G_2$ and the associated function $f^+:E(G)\to \mathcal{P}(\mathbb{N}_0)$ of $f$ is defined as
\begin{equation}
f^+(uv)=
\begin{cases}
f_1(u)+f_1(v) & ~\text{if} ~ u,v\in V(G_1)\\
f_2(u)+f_2(v) & ~\text{if} ~ u,v\in V(G_2)\\
f_1(u)+f_2(v) & ~\text{if} ~ u\in V(G_1), v\in V(G_2).
\end{cases}
\end{equation}
Clearly, $f$ is an IASL of $G_1+G_2$.
\end{proof}
\begin{theorem}
The complement of an IASL graph is also an IASL graph.
\end{theorem}
\begin{proof}
The proof follows from the fact that a graph $G$ and its complement $\overline{G}$ have the same set of vertices and hence have the same set of set-labels. That is, an IASL $f$ defined on a graph $G$ is an IASL of the complement of $G$ also.
\end{proof}
\noindent Now, recall the definitions of the certain fundamental products of two graphs given in \cite{HIS}.
Let $G_1(V_1,E_1)$ and $G_2(V_2,E_2)$ be two graphs. Then, the {\em Cartesian product} of $G_1$ and $G_2$, denoted by $G_1\Box G_2$, is the graph with vertex set $V_1\times V_2$ defined as follows. Let $u=(u_1, u_2)$ and $v=(v_1,v_2)$ be two points in $V_1\times V_2$. Then, $u$ and $v$ are adjacent in $G_1\Box G_2$ whenever [$u_1=v_1$ and $u_2$ is adjacent to $v_2$] or [$u_2=v_2$ and $u_1$ is adjacent to $v_1$].
The {\em direct product} of two graphs $G_1$ and $G_2$, is the graph whose vertex set is $V(G_1)\times V(G_2)$ and for which the vertices $(u,v)$ and $(u',v')$ are adjacent if $uu'\in E(G_1)$ and $vv'\in E(G_2)$. The direct product of $G_1$ and $G_2$ is denoted by $G_1\times G_2$.
The {\em strong product} of two graphs $G_1$ and $G_2$ is the graph, denoted by $G_1\boxtimes G_2$, whose vertex set is $V(G_1) \times V(G_2)$ and for which the vertices $(u,v)$ and $(u',v')$ are adjacent if $[uu'\in E(G_1)~\text{and}~ v=v']$ or $[u=u' ~ \text{and}~vv'\in E(G_2)]$ or $[uu'\in E(G_1)$ and $vv'\in E(G_2)]$.
\cite{IK} The {\em lexicographic product} or {\em composition} of two graphs $G_1$ and $G_2$ is the graph, denoted by $G_1\circ G_2$, is the graph whose vertex set $V(G_1)\times V(G_2)$ and for two vertices $(u,v)$ and $(u',v')$ are adjacent if $[uu'\in E(G_1)]$ or $[u=u'~\text{and}~ vv'\in E(G_2)]$.
\noindent Now we proceed to verify whether these products of two IASL graphs admit IASLs.
\begin{theorem}
The Cartesian product of two IASL graphs also admits an IASL.
\end{theorem}
\begin{proof}
Let $u_1,u_2,\ldots, u_m$ be the vertices of $G_1$ and $v_1,v_2,\ldots, v_n$ be the vertices of $G_2$. Also let $f_1$, $f_2$ be the IASLs defined on $G_1$ and $G_2$ respectively. Now, define a function $f$ on $G_1\Box G_2$ such that $f(u_i,v_j)=f_1(u_1)+f_2(v_j)$ where the associated function $f^+:(E(G_1\Box G_2))\to \mathcal{P}(\mathbb{N}_0)$ is defined by $f^+((u_i,v_j),(u_r,v_s))=f(u_i,v_j)+f(u_r,v_s)$.
Therefore, $f$ is an IASL on $G_1\Box G_2$.
\end{proof}
\noindent Using the definition of $f$ for the vertices and the associated function $f^+$ for the edges, we can establish the existence of IASLs for the other fundamental graph products also as stated in the following results.
\begin{theorem}
The direct product of two IASL graphs also admits an IASL.
\end{theorem}
\begin{theorem}
The strong product of two IASL graphs also admits an IASL.
\end{theorem}
\begin{theorem}
The lexicographic product of two IASL graphs is also an IASL graph.
\end{theorem}
\noindent Next, we consider certain products of graphs in which we take finite number copies of any one of the two graphs and attach these copies to the vertices of the other graph using certain rules.
The most popuar one among such graph products, is the {\em corona} of two graphs which is defined in \cite{FFH} as follows.
The {\em corona} of two graphs $G_1$ and $G_2$, denoted by $G_1\odot G_2$, is the graph obtained by taking $|V(G_1)|$ copies of $G_2$ and then joining the $i$-th point of $G_1$ to every point of the $i$-th copy of $G_2$.
Another commonly used graph product in various literature, is the {\em rooted product} of two graphs, which is defined in \cite{GM} as follows.
The {\em rooted product} of a graph $G_1$ on $n_1$ vertices and rooted graph $G_2$ on $n_2$ vertices, denoted by $G_1\circ G_2$, is defined as the graph obtained by taking $n_1$ copies of $G_2$, and for every vertex $v_i$ of $G_1$ and identifying $v_i$ with the root node of the $i$-th copy of $G_2$.
The admissibility of an IASL by these products of two IASL graphs is established in the following theorem.
\begin{theorem}
The corona of two IASL graphs also admits an IASL.
\end{theorem}
\begin{proof}
Let $u_1, u_2,\ldots, u_m$ be the vertices of $G_1$ and let $v_1, v_2,\ldots, v_n$ be the vertices of $G_2$. For $1\le i\le m$ and $1\le j\le n$, let $v_{ij}$ be the $j$-th vertex of the $i$-th copy of $G_2$. Now, label the vertex $v_{ij}$ of $i$-th copy $G_2$ corresponding to vertex $v_j$ of $G_2$ by the set $f_i(v_{ij})=i.f(v_j)$. Let the set-label of an edge in $G$ be the sum set of the set-labels of its end vertices. Clearly, this labeling is an IASL on $G_1\odot G_2$.
\end{proof}
\begin{theorem}
The rooted product of two IASL graphs is also an IASL graph.
\end{theorem}
\begin{proof}
In the rooted product $G_1\circ G_2$, for $1\le i\le |V(G_1)|$, the roots of the $i$-th copy of $G_2$ is identified with the $i$-th vertex of $G_1$. Label these identified vertices by the corresponding set-labels of $G_1$. All other vertices in the different copies of $G_2$ by distinct integral multiples of the set-labels of corresponding vertices of $G_2$. Also, let the set-labels of edges in $G_1\circ G_2$ be the sum set of the set-labels of their end vertices. Clearly, this set-labeling is an IASL on $G_1\circ G_2$.
\end{proof}
\section{IASLs of Certain Associated Graphs}
An IASL of a graph $G$ induces an IASL to certain other graphs associated to $G$. The induction property of a IASL of a graph $G$ to various associated graphs of $G$ is studied in \cite{GS0}. The major concepts and findings in this paper are mentioned in this section. More over, we establish some new results in this area.
An obvious result about an IASI $f$ of a given graph $G$ is its hereditary nature. The following theorem shows that an IASI of a given graph induces an IASI on any subgraph of $G$.
\begin{proposition}\label{P-IASI2}
\cite{GS0} If a given graph $G$ admits an IASL (or IASI) $f$, then any subgraph $H$ of $G$ also admits an IASI $f^*$, the restriction of $f$ to $V(H)$. That is, the existence of an IASL for a graph $G$ is a hereditary property.
\end{proposition}
If we replace an element (a vertex or an edge) of $G$ by another element (a vertex or an edge) which is not in $G$ to form an associated graph structure, say $G'$, it is customary that the common elements of the graph $G$ and the associated graph $G'$ preserve the same set-labels and the newly introduced elements assume the same set-label of the corresponding deleted elements of $G$. That is, the IASI defined on $G$ induces an IASI for the graph $G'$. The IASI, thus defined for the newly formed graph $G'$ is called an {\em induced IASI} of $G'$.
This induction property of integer additive set-labeling holds not only for the subgraphs of an IASI graph $G$, but also for certain other graph structures associated with $G$ such as minors, topological reductions etc.
In the following results, we consider only the IASL of the associated graph which are induced from the IASL of the graph $G$ concerned.
By an {\em edge contraction} in $G$, we mean an edge, say $e$, is removed and its two incident vertices, $u$ and $v$, are merged into a new vertex $w$, where the edges incident to $w$ each correspond to an edge incident to either $u$ or $v$.
The existence of an IASI for the graph obtained by contracting finite edges of an IASL graph $G$ is established in the following theorem.
\begin{proposition}\label{P-GEC}
\cite{GS0} Let $G$ be an IASI graph and let $e$ be an edge of $G$. Then, $G\circ e$ admits an IASI.
\end{proposition}
An undirected graph $H$ is called a {\em minor} of a given graph $G$ if $H$ can be formed from $G$ by deleting edges and vertices and by contracting edges.
\begin{proposition}
Let $G$ be an IASI graph and let $H$ be a minor of $G$. Then, $H$ admits an IASI.
\end{proposition}
\noindent The proof of this result is an immediate consequence of Proposition \ref{P-IASI2} and \ref{P-GEC}.
Let $G$ be a connected graph and let $v$ be a vertex of $G$ with $d(v)=2$. Then, $v$ is adjacent to two vertices $u$ and $w$ in $G$. If $u$ and $v$ are non-adjacent vertices in $G$, then delete $v$ from $G$ and add the edge $uw$ to $G-\{v\}$. This operation is called an {\em elementary topological reduction} on $G$.
Then, we have
\begin{proposition}\label{P-IASI3a}
\cite{GS0} Let $G$ be a graph which admits an IASI. Then any graph $G'$, obtained by applying finite number of elementary topological reductions on $G$, admits an IASI.
\end{proposition}
A {\em subdivision} or an {\em expansion} of a graph is the reverse operation of an elementary topological reduction. Under this operation, we introduce a new vertex, say $w$, to an edge $uv$ in $G$. Thus, the vertex $w$ replaces the edge $uv$ and two edges $uw$ and $wv$ are introduced to $G-uv\cup \{w\}$. Two graphs $G$ and $G'$ are said to be {\em homeomorphic} if there is a graph isomorphism from some subdivision of $G$ to some subdivision of $G'$.
The following theorem establishes the existence of an induced IASL for a graph $G'$ which is homeomorphic to a given IASL graph $G$.
\begin{proposition}\label{P-IASI3b}
Let $G$ be an IASL (or IASI) graph. Then, any subdivision $G'$ of $G$ admits an (induced) IASL (or IASI).
\end{proposition}
\begin{proof}
Let $f$ be an IASL defined on a given graph $G$ and let $u,v$ be two adjacent vertices in $G$. Introduce a new vertex $w$ on the edge $uv$ so that $uv$ is subdivided to two edges $uw$ and $wv$. Let $G'=[G-(uv)]\cup \{uw, vw\}$. Define a function $g:V(G')\to \mathcal{P}(\mathbb{N}_0)$ such that $g(v)= f(v)$ if $v\in V(G)$ and $g(w)=f^+(uv)$.
Clearly, $g$ is an IASL on $G'$. This result can be extended to a graph obtained by finite number of subdivisions on $G$. There fore, any subdivision of an IASL graph admits an (induced) IASL.
\end{proof}
Invoking Proposition \ref{P-IASI3b}, the admissibility of an IASL by a graph that is homeomorphic to a given IASL graph.
\begin{proposition}
Let $G$ be an IASL (or IASI) graph and let $H$ be a a graph that is homeomorphic to $G$. Then, $H$ admits an IASL (or IASI).
\end{proposition}
\begin{proof}
Assume that the graphs $G$ and $H$ are homeomorphic graphs. Let $G'$ and $H'$ be the isomorphic subdivisions of $G$ and $H$ respectively. Let $f$ be an IASL defined on $G$. By Proposition \ref{P-IASI3b}, the subdivision $G'$ admits an IASL, say $f'$ induced by $f$. Since $G'\cong H'$, define a function $g:V(H')\to \mathcal{P}(\mathbb{N}_0)$ by $g(v')=f(v)$, where $v'$ is the vertex in $H'$ corresponding to the vertex $v$ of $G'$. That is, $g$ assigns the same set-labels of the vertices of $G'$ to the corresponding vertices of $H'$. Therefore, $g(V(H'))=f(V(G'))$. Since $H$ can be considered as a graph obtained by applying a finite number of topological reductions on $H'$, by Proposition \ref{P-IASI3a}, $g$ induces an IASL on $H$. Therefore, $H$ is an IASL graph.
\end{proof}
The existence of integer additive set-labelings for certain other graphs associated to the given graph $G$, which are not obtained by replacing some elements of $G$ by certain other elements not in $G$, have been established in \cite{GS0}.
An important graph of this kind which is associated to a given graph $G$ is the line graph of $G$ which is defined as follows.
For a given graph $G$, its {\em line graph} $L(G)$ is a graph such that each vertex of $L(G)$ represents an edge of $G$ and two vertices of $L(G)$ are adjacent if and only if their corresponding edges in $G$ incident on a common vertex in $G$. The following theorem establishes the existence of IASI by the line graph of a graph.
\begin{theorem}\label{T-IASI-LG}
\cite{GS0} If a graph $G$ admits IASI, say $f$, then its line graph $L(G)$ also admits an (induced) IASI.
\end{theorem}
Another graph that is associated to a given graph is its total graph which is defined as follows.
The {\em total graph} of a graph $G$ is the graph, denoted by $T(G)$, is the graph having the property that a one-to one correspondence can be defined between its points and the elements (vertices and edges) of $G$ such that two points of $T(G)$ are adjacent if and only if the corresponding elements of $G$ are adjacent (either if both elements are edges or if both elements are vertices) or they are incident (if one element is an edge and the other is a vertex).
The following theorem establishes the existence of an IASI for the total graph of a graph.
\begin{theorem}\label{T-IASI-TG}
\cite{GS0} If a graph $G$ admits IASI, say $f$, then its total graph $T(G)$ also admits an (induced) IASI.
\end{theorem}
So far, we discussed about the integer additive set-labeling of graphs with respect to the countably infinite set $\mathbb{N}_0$. Can we choose a finite set $X$ as the ground set for labeling the vertices of a graph $G$? This is possible only when the ground set $X$ has sufficiently large cardinality. Hence, the study about the cardinality of the ground set $X$ arises much interest.
The minimum cardinality of the ground set that is required for set-labeling a graph $G$ is called the {\em set-indexing number} of $G$. The following theorem determines the set-indexing number of a given graph $G$.
\begin{theorem}\label{P-FinSet}
\cite{GS0} Let $X$ be a finite set of non-negative integers and let $f:V(G)\to 2^X-\{\emptyset\}$ be an IASI on $G$, which has $n$ vertices. Then, $X$ has at least $\lceil log_2(n+1)\rceil$ elements.
\end{theorem}
The following theorem is special case of the above theorem when the vertices of $G$ have the same set-indexing number.
\begin{theorem}\label{P-UniSet}
\cite{GS0} Let $X$ be a finite set of non-negative integers and let $f:V(G)\to 2^X-\{\emptyset\}$ be an IASI on $G$, which has $n$ vertices, such that $V(G)$ is $l$-uniformly set-indexed. Then, $n\le \binom{|X|}{l}$.
\end{theorem}
We have discussed the cardinality of the ground set $X$ for labeling the vertices of a given graph so that it admits an IASL. The cardinality of the set-labels of the elements of $G$ is also worth studying. Some studies in this area are reviewed in the following section.
\section{Cardinality of the Set-Labels of IASL Graphs}
Certain studies on set-indexing numbers of the elements of the IASI-graphs have been done in \cite{GS1}, \cite{GS2} and \cite{GS0}. The set-theoretic foundations of the IASLs , established in these studies, are reviewed in this section.
Let $A$ and $B$ be two non-empty sets of non-negative integers. Then the ordered pairs $(a,b)$ and $(c,d)$ in $A\times B$ is said to be {\em compatible} if $a+b=c+d$. If $(a,b)$ and $(c,d)$ are compatible, then we write $(a,b)\sim (c,d)$. Clearly, $\sim$ is an equivalence relation.
A {\em compatible class} with respect to an integer $k$ is the subset of $A\times B$ defined by $\{(a,b)\in A\times B:a+b=k\}$ and is denoted by $\mathsf{C}_k$.
All the compatibility classes need not have the same number of elements but can not exceed a certain number. It can be noted that exactly one representative element of each compatibility class $\mathsf{C}_k$ of $f(u)\times f(v)$ contributes an element to the set-label of the corresponding edge $uv$ and all other $r_k-1$ elements are neglected and we call these elements {\em neglected elements} of the class $\mathsf{C}_k$. The number of neglected elements in the set-label of an edge $uv$ with respect to the set $f(u)\times f(v)$ is called the {\em neglecting number} of that edge.
The compatibility classes which contain the maximum possible number of elements are called {\em saturated classes}. The compatibility class that contains maximum elements is called a {\em maximal compatibility class}. Hence, we have
\begin{proposition}\label{P-CardCC}
\cite{GS0} The cardinality of a saturated class in $A\times B$ is $n$, where $n=\min(|A|,|B|)$.
\end{proposition}
The number of distinct compatibility classes in $A\times B$ is called the {\em compatibility index} of the pair of sets $(A,B)$ and is denoted by $\mho_{(A,B)}$. Hence, we have the following lemma.
\begin{lemma}\label{L-3}
Let $f$ be an IASI of a graph $G$ and $u,v$ be two vertices of $G$. The set-indexing number of an edge $e=uv$ of $G$ is given by $|f^+(e)|=|f(u)|+|f(v)|=\mho_{(f(u),f(v))}$
\end{lemma}
The following result establishes the bounds for the set-indexing number of the edges of an IASI graph, in terms of the set-indexing numbers of their end vertices.
\begin{lemma}\label{L-0}
\cite{GS1} For an IASI $f$ of a graph $G$, $\max(|f(u)|,|f(v)|) \le |f^+(uv)|= |f(u)+f(v)| \le |f(u)| |f(v)|$, where $u,v \in V(G)$.
\end{lemma}
\begin{theorem}\label{T-Card}
Let $f$ be an IASI of a graph $G$ and let $u$ and $v$ be two adjacent vertices of $G$. Let $|f(u)|=m$ and $|f(v)|=n$. Then, $|f^+(uv)|=mn-r$, where $r$ is the number of neglected elements in $f(u)\times f(v)$.
\end{theorem}
An interesting question that arises in this context is about a necessary and sufficient condition for an IASL of a given graph $G$ to be a uniform IASL of $G$. The following result provides a platform for finding the conditions for a given graph $G$ to admit a uniform IASL.
\begin{proposition}
Two adjacent edges $e_i$ and $e_j$ of a graph $G$ have the same set-indexing number if and only if the set-indexing number of the common vertex of these edges is the quotient when the difference of the neglecting numbers of the edges is divided by the differences of set-indexing numbers of the end vertices that are not common to these edges.
\end{proposition}
\begin{proof}
Let $e_i=uv_i$ and $e_j=uv_j$ be two adjacent edges in the given graph $G$. Let $f$ be an IASL defined on $G$. Also, let $|f(u)|=m$, $|f(v_i)|=n_i$ and $|f(v_j)|=n_j$. Then, the set-indexing number of $e_i$ is $mn_i-r_i$ and that of $e_j$ is $mn_j-r_j$ where $r_i$ and $r_j$ are neglecting numbers of $e_i$ and $e_j$ respectively.
Assume that $e_i$ and $e_j$ have the same set-indexing number. Then, $mn_i-r_i=mn_j-r_j$. That is, $m=\frac{r_i-r_j}{n_i-n_j}$.
Conversely, assume that the set-indexing number of the common vertex of these edges is the quotient when the difference of the neglecting numbers of the edges is divided by the differences of set-indexing numbers of the end vertices that are not common to these edges. That is, $m=\frac{r_i-r_j}{n_i-n_j}$. Hence, $mn_i-r_i=mn_j-r_j$.
\noindent Therefore, the edges have the same set-indexing number.
\end{proof}
\noindent Due to the above proposition, we establish a necessary and sufficient condition for an IASL of a given graph to be a uniform IASL of $G$.
\begin{theorem}
Let $G$ be a graph that admits an IASL $f$. Then, $f$ is a uniform IASL of $G$ if and only if the set-indexing number of the common vertex of any two adjacent edges is the ratio of the difference between their neglecting numbers to the difference between the set-indexing numbers of their distinct end vertices.
\end{theorem}
In the following theorem, we prove a necessary and sufficient condition for an IASL of a graph $G$, which assigns uniform set-labels to the vertices of $G$, to be a uniform IASL of $G$.
\begin{theorem}
Let the graph $G$ admits an IASL $f$ under which all vertices of $G$ are uniformly set-labeled. Then, $f$ is a uniform IASL if and only if all edges of $G$ have the same neglecting number.
\end{theorem}
\begin{proof}
Let $V(G)$ be $l$-uniformly set-labeled. Then, for any vertex $u$ in $G$, $|f(u)|=l$ and $f^+(e_i)f^+(uv)=l^2-r_i$.
Let $f$ be a uniform IASI of $G$. Then, for any two adjacent edges $e_i=uv_i$ and $e_j=uv_j$ of $G$, we have
\begin{eqnarray*}
|f^+(e_i)|& = &|f^+(e_j)|\\
\implies l^2-r_i & = & l^2-r_j\\
\implies r_i-r_j & = & 0\\
\implies r_i & = & r_j.
\end{eqnarray*}
Conversely, assume that all the edges of $G$ have the same neglecting number, say $r$. Therefore, for any edge $e=uv$ of $G$, $|f^+(e)|=|f(u)|\,|f(v)|-r$. Since the graph $V(G)$ is $l$-uniformly set-labeled, $|f(v)|=l$ for all $v\in V(G)$. Therefore, $|f^+(e)|=l^2-r$ for all edges $e\in E(G)$. That is, $f$ is a uniform IASL.
\end{proof}
A necessary and sufficient condition for a graph to have a $2$-uniform IASI is proved in \cite{TMKA} as follows.
\begin{theorem}
\cite{TMKA} A graph $G$ admits a $2$-uniform IASI if and only if $G$ is bipartite.
\end{theorem}
In view of Lemma \ref{L-0} the following definitions are introduced in \cite{GS1} and \cite{GS2}.
\begin{definition}\label{D-WIASI}{\rm
\cite{GS1} A {\em weak IASI} is an IASI $f$ whose associated function is defined as $|f^+(uv)|=\max(|f(u)|,|f(v)|)$ for all $u,v\in V(G)$. A graph which admits a weak IASI may be called a {\em weak IASI graph}. A weak IASI is said to be {\em weakly uniform IASI} if $|f^+(uv)|=k$, for all $u,v\in V(G)$ and for some positive integer $k$.}
\end{definition}
\begin{definition}\label{D-SIASI}{\rm
\cite{GS2} An IASI $f$ is called a {\em strong IASI} if $|f^+(uv)|=|f(u)| |f(v)|$ for all $u,v\in V(G)$. A graph which admits a strong IASI may be called a {\em strong IASI graph}. A strong IASI is said to be {\em strongly uniform IASI} if $|f^+(uv)|=k$, for all $u,v\in V(G)$ and for some positive integer $k$.}
\end{definition}
Form the above definitions it follows that if $f;V(G)\to \mathcal{P}(\mathbb{N}_0)$ is a weak (or strong) IASI defined on a graph $G$, then for each adjacent pair of vertices $u$ and $v$ of $G$, each compatibility class of the pair of set-labels $f(u)$ and $f(v)$ is a trivial class.
Let $G$ be an IASI graph. Then, the following result for $G$ was proved in \cite{GS1}.
\begin{lemma}\label{L-IASI01}
\cite{GS1} Let $f:V(G)\to \mathcal{P}(\mathbb{N}_0)$ be an IASI defined on a graph $G$. Let $u$ and $v$ be any two adjacent vertices in $G$. Then, $|f^+(uv)|=\max(|f(u)|,|f(v)|)$ if and only if either $|f(u)|=1$ or $|f(v)|=1$.
\end{lemma}
This lemma leads us to the following result.
\begin{theorem}
An IASI $f$ of a graph $G$ is a weak IASI of $G$ if and only if at least one end vertex of any edge in $G$ has the set-indexing number $1$.
\end{theorem}
The following theorem describes a necessary and sufficient condition for a graph to admit a weakly uniform IASI.
\begin{theorem}
A connected graph $G$ admits a weakly uniform IASI $f$ if and only if $G$ is a bipartite graph.
\end{theorem}
The necessary and sufficient condition for the sum set of two sets to have maximum cardinality is established in the following result.
\begin{lemma}\label{L-RDS}
\cite{GS2} Let $A$, $B$ be two non-empty subsets of $\mathbb{N}_0$. Let $D_A$ and $D_B$ be the sets of all differences between two elements in the sets $A$ and $B$ respectively. Then, $|A+B|=|A|.|B|$ if and only if $D_A$ and $D_B$ are disjoint.
\end{lemma}
\noindent The sets $D_A$ of all differences between two elements of set $A$ is called the {\em difference set} of $A$.
In view of the above lemma, a necessary and sufficient condition for an IASI to be a strong IASI is established in \cite{GS2} as follows.
\begin{theorem}
An IASI $f$ of a graph $G$ is a strong IASI of $G$ if and only if the difference sets of the set-labels of any two adjacent vertices in $G$ are disjoint sets.
\end{theorem}
The following theorem describes a necessary and sufficient condition for a graph to admit a strongly uniform IASI.
\begin{theorem}
A connected graph $G$ admits a strongly $k$-uniform IASI $f$ if and only if $G$ is a bipartite graph or $f$ is a $(k,l)$-completely uniform IASI of $G$, where $k=l^2$.
\end{theorem}
\section{Conclusion}
So far, we have reviewed the studies about integer additive set-labelings and integer additive set-indexers of certain graphs and their characteristics. In this paper, we also propose some new results in this area. Certain problems regarding the admissibility of integer additive set-indexers by various other graph classes are still open.
More properties and characteristics of different types of IASLs and IASIs, both uniform and non-uniform, are yet to be investigated. There are some more open problems regarding necessary and sufficient conditions for various graphs and graph classes to admit certain IASIs.
\section*{Acknowledgement}
The authors dedicate this work to the memory of Professor Belamannu Devadas Acharya who introduced the concept of integer additive set-indexers of graphs.
|
\section{Introduction}
\label{sec:intro}
The first phase of the LHC experiments has given two important messages: a scalar resonance closely resembling the Standard Model Higgs boson
has been discovered, and new physics beyond the Standard Model has not been found. The latter result would imply that new states or effects beyond the Standard Model predictions may be much more difficult to spot at the LHC than we previously thought. In fact, very strong bounds have been posed on {\it easy--catch} models, like the
constrained version of the Minimal Supersymmetric Standard Model~\cite{atlas:susytwiki,cms:susytwiki}.
Many theorists have therefore recently turned their attention on a more signature-based strategy,
focusing on unusual final states which are difficult to detect or have not been considered yet by experimental collaborations.
One such final state that has been gaining popularity among
phenomenologists~\cite{Berger:1999zt,delAguila:1999ac,Berger:2000zk,Desai:2010sq,%
Davoudiasl:2011fj,Andrea:2011ws,Kamenik:2011nb,Dong:2011rh,Wang:2011uxa,Fuks:2012im,Kumar:2013jgb,Alvarez:2013jqa,Agram:2013wda,Fuks:2014uka,Ng:2014pqa} and
experimentalists~\cite{Aaltonen:2012ek,CMS:2014hba,Khachatryan:2014uma,Aad:2014wza}
is the monotop signature: a single top quark produced in association with a large amount of missing energy.
Although the production of this final state is very suppressed in the Standard Model,
it is however not easy to obtain this kind of events in realistic and complete models of new physics.
Two main production mechanisms can lead to a monotop state~\cite{Andrea:2011ws,Agram:2013wda}, arising either from
the resonant production of a coloured bosonic state which further decays into a
top quark plus an invisible neutral fermion; or via the production of a single top quark
in association with an invisible boson that has flavour-changing couplings to top and light quarks.
Examples of the first class of models include $R$-parity violating supersymmetry, where the
produced resonance is a top squark decaying into a top plus a long-lived
neutralino~\cite{Berger:1999zt,Berger:2000zk,Desai:2010sq,Fuks:2012im}.
The second class of models has
been described in scenarios of dark matter from a hidden sector that couples to the Standard Model
via flavour-violating couplings of a bosonic mediator~\cite{Davoudiasl:2010am,Davoudiasl:2011fj,Kamenik:2011nb,Alvarez:2013jqa}.
In general, monotop signatures can be however generated by other processes
involving, for instance, the $t$-channel exchange of a new particle, different spin
assigments for the new states or higher-spin tensors. Motivated by the
setups currently under study experimentally, this work is limited
to the case of a spin-0 or spin-1 state that can be either
exchanged in the $s$-channel (named ``resonant'' in the following) or produced in
association with a top quark via flavour-changing interactions
(named ``non-resonant'').
All such models can be described in terms of a simple Lagrangian~\cite{Andrea:2011ws}, which contains all the possible
couplings giving rise to a monotop signal. A very
general analysis of this framework can be found in Ref.~\cite{Agram:2013wda}, the limiting
case of higher-dimensional operators has been discussed in Ref.~\cite{Dong:2011rh},
while monotop production via flavour-changing interactions of quarks with
an invisible $Z$-boson has been detailed in Ref.~\cite{delAguila:1999ac}.
Although this simple description has the advantage of being complete, it has the drawback of containing too many free parameters to be efficiently scanned by an experimental search.
Furthermore, the included couplings do not respect the symmetries of the Standard Model,
as they are intended to describe the model dynamics after the breaking of the electroweak symmetry.
In this way, this approach ignores other interactions needed to restore gauge invariance which can
give rise to new physics signals in different search channels, the latter possibly
implying stronger constraints on the parameters of the model than the monotop search itself.
In this work, we revisit the parametrisation originally proposed in
Ref.~\cite{Andrea:2011ws} by paying particular attention to the embedding of the
Lagrangian description into $SU(2)_L \times U(1)_Y$ invariant operators.
We therefore present a set of minimal effective Lagrangians, extending the Standard Model with gauge-invariant operators. Our
approach allows us to restrict the number of ``interesting'' scenarios, \textit{i.e.},
the cases where the monotop signal is genuinely the main signal of new physics to be expected at
the LHC. Equivalently, this reduces the number of free parameters to a manageable number.
Finally, we discuss in detail how the effective model could be completed in order to
guarantee that the missing energy particle produced in association with the top quark
is indeed either long-lived or decaying into invisible states.
The rest of this work is organized as follows. In Section~\ref{sec:modelbuilding},
we present how to construct gauge-invariant effective models for the monotop signal.
We consider separately the resonant case, where the mediator is a scalar or vector boson, and the non resonant case where the top is produced in association with a scalar or vector via flavour-changing couplings. We also discuss how the effective Lagrangians can be matched to the simple monotop descriptions of Ref.~\cite{Andrea:2011ws}.
We then focus on non-resonant scenarios which turn out to be less
``standard'' and investigate, in Section~\ref{sec:nonrespheno}, the conditions under which the
invisible state is effectively
invisible, and other experimental observations, which can further constrain
the model. Our conclusions are presented in Section~\ref{sec:conclusions}.
\section{Gauge-invariant effective Lagrangians for monotop production}\label{sec:modelbuilding}
\subsection{Resonant monotop production}
\label{sec:lres}
In the first class of scenarios yielding the production of
a monotop system at colliders that we consider, the produced top quark recoils against
an invisible fermionic state $\chi$.
Being singly produced, the $\chi$ particle cannot be stable, thus it is either
long-lived or it decays into a pair of stable particles. In each case, it has to be
electrically-neutral and a
colour-singlet.
In resonant monotop production, both final-state particles arise from the decay of a heavy scalar $\varphi$ or vector $X$ field, lying
in the fundamental representation of $SU(3)_c$, that is produced in the $s$-channel
from the fusion of two down-type (anti-)quarks.
The parton-level process that we want to focus on is therefore
\begin{eqnarray}}% can be used as {equation} or {eqnarray
d d \to \varphi, X_\mu \to \bar{t} \chi\,.
\end{eqnarray}
In order to understand how the scalar or vector mediators transform under the full
Stanard Model symmetries, it is crucial to define the chirality of the quarks it couples too.
In the following we will analyse separately the scalar and vector case in detail.
\subsubsection*{Spin-0 mediator}
The initial state consists of a pair of down-type quarks which form a spin-0 state. The related
Lorentz scalar fermionic bilinear reads $\bar{\psi} \psi$, which can be written as
$\bar{\psi}_L \psi_R + \bar{\psi}_R \psi_L$. This implies that the two quarks
have opposite chiralities. Recalling that the charge conjugate of the right-handed quark $d^C_R$
is left-handed while the one of the left-handed quark $d^C_L$ is right-handed, we can define two
independent initial states, one with explicit right-handed and one with left-handed chirality indices.
According to the Standard Model gauge symmetries $SU(3)_c$, $SU(2)_L$ and $U(1)_Y$, the two states transform as:
\begin{equation}\bsp
\bar{d}_R^{C\; i} d_R^j = ({\utilde{\bf \bar{3}}}, {\utilde{\bf 1}}, -2/3)\,; \qquad
\bar{d}_L^{C\; i} d_L^j = ({\utilde{\bf \bar{3}}}, {\utilde{\bf 3}}, 1/3) \,.
\esp\end{equation}
In our notation, an undertilde indicates a representation under a non-Abelian gauge symmetry, and the indices $i,j$ refer to flavour.
Since the diquark states are made of identical quarks and fermions are
anticommuting quantities, the corresponding wavefunctions need to be
antisymmetric under the exchange of the quark fields.
The exchange of the flavour indices is therefore forced to be antisymmetric too,
since the one of the spin and colour indices (which we do not explicit for
simplicity) are antisymmetric and the one of the triplet (adjoint) representation
of SU(2)$_L$ is symmetric (for the left-handed quark setup).
From the representations shown above, the right and left-handed quarks cannot couple to the same scalar and two different objects must thus be introduced,
\begin{eqnarray}}% can be used as {equation} or {eqnarray
\varphi_s = ({\utilde{\bf 3}}, {\utilde{\bf 1}}, 2/3)\,; \qquad \varphi_t = ({\utilde{\bf 3}}, {\utilde{\bf 3}}, -1/3) \equiv
\left( \begin{array}{c} \varphi_t^{2/3} \\ \varphi_t^{-1/3} \\ \varphi_t^{-4/3} \end{array} \right)\,,
\label{eq:scafields}\end{eqnarray}
where the subscript $_s$ and $_t$ refer to singlet and triplet of SU(2)$_L$.
The operators containing the interactions needed for monotop production
can be written as
\begin{eqnarray}}% can be used as {equation} or {eqnarray
\lambda_s\ \varphi_s\, \bar{d}_R^C d_R + \lambda_t\ \varphi_t \,\bar{q}_L^C q_L
+ {\rm h.c.} \ ,
\label{eq:lresGauge}\end{eqnarray}
where $q_L$ is the left-handed doublet, and $\lambda_{s,t}$ are antisymmetric matrices in flavour space.
In the second term, it is the component of the triplet with electric charge $\pm 2/3$ that couples eventually to the top.
We can now repeat the same analysis for the final state. We first assume,
for minimality, that the invisible fermion $\chi$ is a singlet under
the Standard Model symmetries.
In this case, the final state system representation under the
Standard Model gauge group can be
\begin{equation}\bsp
\chi_R t_R = ({\utilde{\bf 3}}, {\utilde{\bf 1}}, 2/3)\,; \qquad
\chi_L t_L = ({\utilde{\bf 3}}, {\utilde{\bf 2}}, 1/6) \,.
\esp\end{equation}
In order to allow for a $\chi$-coupling to a left-handed top quark,
we therefore need to introduce an extra scalar field $\varphi_d$, compared
to Eq.~\eqref{eq:scafields}, which transforms as a doublet of
$SU(2)_L$.
The operators relevant for monotop production can then be written as
\begin{eqnarray}}% can be used as {equation} or {eqnarray
y_s\ \varphi_s^\dagger \, \bar{\chi} t_R + y_d\ \varphi_d^\dagger \bar{\chi} q_L
+ {\rm h.c.}
\end{eqnarray}
The initial and final state can consequently only be connected via an
$SU(2)_L$-singlet field $\varphi_s$ that couples to right-handed quarks.
This scenario being minimal, it will therefore be considered in the rest of
this work. For completeness, let us mention that
non-minimal models with several additional scalars could also
be constructed. In these setups, the new scalar fields are allowed to mix
after electroweak symmetry breaking through couplings to the
Brout-Englert-Higgs field
$\phi_H = ({\utilde{\bf 1}}, {\utilde{\bf 2}}, 1/2)$,
\begin{eqnarray}}% can be used as {equation} or {eqnarray
\mu_t\ \phi_H^\dagger \varphi_t^\dagger \varphi_d+
\mu_d\ \varphi_s \varphi_d^\dagger \phi_H^\dagger + {\rm h.c.}
\end{eqnarray}
The resulting mass
splitting is nevertheless constrained to be small by the perturbativity of the
couplings~\cite{Jack:1982sr} and corrections to the $S$ and $T$
parameters~\cite{Li:1992dt,Bhattacharyya:1993zy}.
We have also imposed that the $\chi$-field has the same quantum numbers
as a right-handed neutrino, so that it could potentially mix
with neutrinos. This mixing is however strongly constrained
by proton-decay processes like $p \to \pi^+/K^+ \nu$. In this case,
the contribution of the box-diagram-induced subprocess
$d u \to \bar{d}/\bar{s} \nu$ (through a $\varphi_s$ and $W^+$
exchange) has indeed to be
maintained small. There is however a way to evade the bound by
preventing $\chi$ from mixing with neutrinos (by assigning it, \textit{e.g.}, a
baryon number), unless $\chi$ is lighter than the proton.
We have considered so far models where the invisible fermion
$\chi$ is a Standard Model gauge singlet. Another option would be
to assume that $\chi$ is the neutral component of a non-trivial
$SU(2)_L$ multiplet. For instance, one could choose (with $\sigma_2$ being the
second Pauli matrix)
\begin{equation}\bsp
\chi_d = ({\utilde{\bf 1}}, {\utilde{\bf 2}}, 1/2) \equiv
\begin{pmatrix} \chi^+ \\ \chi^0 \end{pmatrix}
\qquad &\ \Rightarrow \qquad
\chi_L t_L = ({\utilde{\bf 3}}, {\utilde{\bf 1}}, 2/3) \ ,\\
\chi^\prime_d = i \sigma_2 \chi_d^\ast = ({\utilde{\bf 1}}, {\utilde{\bf 2}}, -1/2) \equiv
\begin{pmatrix} \chi^0 \\ \chi^- \end{pmatrix}
\qquad &\ \Rightarrow \qquad
\chi_L t_L = ({\utilde{\bf 3}}, {\utilde{\bf 3}}, -1/3) \ ,\\
\chi_t = ({\utilde{\bf 1}}, {\utilde{\bf 3}}, -1) \equiv
\begin{pmatrix}\chi_0 \\ \chi^- \\ \chi^{--} \end{pmatrix}
\qquad &\ \Rightarrow \qquad
\chi_R t_R = ({\utilde{\bf 3}}, {\utilde{\bf 3}}, -1/3) \ .
\esp\end{equation}
For the sake of the example, we focus on the first option where monotop
systems can be produced via the production and the decay of
a $\varphi_s$ resonance.
The presence of charged degrees of freedom in the $\chi$ multiplet
allows one to constrain this scenario
by other sources like, for instance, single bottom production
$\bar d_R \bar d_R \to \varphi_s \to b \chi^+$.
The mass splitting between the neutral and charged component of
the $\chi_d$ doublet being generated by electroweak loop-diagrams
(unless they mix to other fermions),
the decay of the charged $\chi^+$ field is driven
by the two competing channels $\chi^+ \to \bar b_L \bar d_R^i \bar d_R^j$
(mediated by $\varphi_s$) and $\chi^+ \to \chi_0 W^*$ (with a very
off-shell $W$-boson).
As a result, the $\chi^+$ particle is in general long-lived, which
is heavily constrained by current
LHC searches~\cite{Aad:2013pqd,Chatrchyan:2013oca}.
Furthermore, $\chi_d$ has the same quantum numbers as the
lepton doublets of the Standard Model so that these fields can mix,
which induces the proton decay modes $p \to \pi^+/K^+ \nu$ and
$p \to \pi^0/K^0 e^+$ similarly as
described above. Consequently, it turns out that
scenarios where the invisible $\chi$ fermion is one of the components
of a larger $SU(2)_L$ multiplet are unlikely to be realized.
Although we will ignore those non-minimal
scenarios, their complete analysis is however in order,
which goes beyond the scope of
this work focusing on monotop production only.
\subsubsection*{Spin-1 mediator}
We now turn to cases where monotop systems are produced from the decay of
a spin-$1$ resonance $X$. The related Lorentz vector fermionic bilinear
is given by $\bar{\psi}\gamma_\mu \psi$,
which can be written as $\bar{\psi}_L \gamma_\mu \psi_L + \bar{\psi}_R \gamma_\mu \psi_R$.
This implies that the two quarks have the same chirality. In order to build a scalar invariant, vector fields have
to couple to these spinors of the same chiralities. Using the same properties of the charge conjugation used in the notations for the scalar case,
the possible couplings of the $X$-field to down-type quarks are then of the form
\begin{eqnarray}}% can be used as {equation} or {eqnarray
\lambda_V^1\ X^\mu \bar{d}_L^C \gamma_\mu d_R +
{\rm h.c.} \ ,
\end{eqnarray}
where we denote the coupling strength by $\lambda_V$. In order for such
couplings to be $SU(2)_L$-invariant, the $X$-boson must belong
to a weak doublet with hypercharge $1/6$,
\begin{eqnarray}}% can be used as {equation} or {eqnarray
X_\mu = ({\utilde{\bf 3}}, {\utilde{\bf 2}}, 1/6) \equiv
\left( \begin{array}{c} X_\mu^{2/3} \\ X_\mu^{-1/3} \end{array} \right)\ .
\end{eqnarray}
Turning to the final state, we begin with the fact that
the $X$-field defined above
has the quantum numbers of a left-handed quark doublet. It
can consequently couple to a left-handed top quark and
a singlet field $\chi$,
\begin{eqnarray}}% can be used as {equation} or {eqnarray
\lambda_V^2 X_\mu \bar{q}_L \gamma^\mu \chi + {\rm h.c.}
\end{eqnarray}
Enforcing weak isospin gauge invariance implies thus, in addition to
the interaction relevant for the production of a monotop
state, the presence of the
interaction of a left-handed bottom quark to the second
component of the $X$-doublet. This however induces the fast decay of the
neutral $\chi$ fermion via an off-shell $X$-state,
\begin{eqnarray}}% can be used as {equation} or {eqnarray
\chi \to b_L (X_\mu^{-1/3})^* \to b_L u_L d_R
\quad(\text{or}\ \ \bar b_L \bar u_L \bar d_R \text{ if }
\chi\text{ is a Majorana fermion})\ ,
\end{eqnarray}
so that this model does not predict any monotop signal.
We therefore move on with a second option for linking the $\chi$-$t$ monotop
system to the $X$-boson by considering right-handed quarks.
A coupling to right-handed quarks can be obtained if the fermion $\chi$ belongs to an $SU(2)_L$ doublet with hypercharge $1/2$, $\chi_d = ({\utilde{\bf 1}}, {\utilde{\bf 2}}, 1/2)$,
\begin{eqnarray}}% can be used as {equation} or {eqnarray
\lambda_V^3 X_\mu \bar{t}_R \gamma^\mu \chi_d + {\rm h.c.}
\end{eqnarray}
This model however contains a charged fermion that can be produced in association with a top via the vector of charge $-1/3$ and is therefore likely to be constrained
by channels different from the monotop one.
More complicated non-minimal possibilities could also be considered but will
be ignored from this work for the same reason: they are
strongly constrained by other processes and their analysis
must account for these other channels, in addition to the monotop signature.
\subsubsection*{Summary for the resonant channel}
From the analysis of gauge invariant effective Lagrangians performed in this
subsection, we have shown that the chirality of both the initial down-type
quarks and the final-state top quark are correlated with the quantum numbers
allowed for the bosonic mediator and the invisible fermion. In
this work, we focus on the minimal model in terms of field content
and interactions. In this case, the setup
that predicts monotop production at the LHC as its main signature (and that
is thus not constrained by any other observable) contains a scalar mediator
and a new fermion that are both gauge-singlet and couple to right-handed quarks.
The effective Lagrangian reads
\begin{equation}\bsp
\mathcal{L}_{\rm eff.} = {\cal L}_{\rm kin}(\varphi_s,\chi) +
\lambda_s^{ij} \ \varphi_s\, \bar{d}_{R,i}^C d_{R,j}
+ y_s\ \varphi_s^\dagger \, \bar{\chi} t_R + {\rm h.c.} \ ,
\esp \label{eq:effres}\end{equation}
where ${\cal L}_{\rm kin}$ contains gauge interaction, kinetic and mass terms
for the new fields, and the other terms focus on their interactions with
the Standard Model quarks.
This can be compared to the notation of Ref.~\cite{Andrea:2011ws}
where the Lagrangian describing the same
scenario is written as
\begin{equation}\bsp
\mathcal{L}_{\rm res} &\ = {\cal L}_{\rm kin}(\varphi_s,\chi)
+ \Big( \varphi\, \bar{d}^C_i \Big[ (a_{SR}^q)^{ij} + (b_{SR}^q)^{ij}\gamma^5\Big] d_j
+ \varphi\, \bar{t} \Big[ a_{SR}^{1/2} + b^{1/2}_{SR} \gamma^5 \Big] \chi + {\rm h.c.} \Big) \ ,
\esp\label{eq:lres}\end{equation}
flavour indices being noted by $i,j$ and colour indices being omitted for clarity.
The two Lagrangians are related by
\begin{eqnarray}}% can be used as {equation} or {eqnarray
a_{SR}^q = b_{SR}^q = \lambda_s/2\, , \qquad a_{SR}^{1/2} = b^{1/2}_{SR} = y_s^*/2\,.
\end{eqnarray}
As already mentioned, the couplings of the scalar to the down quarks are antisymmetric under the exchange of the flavour
indices. Consequently, parton density effects enhance the
production mode $d s \to \varphi^\ast$ at hadron colliders (with the relevant
coupling strengths being non-vanishing), as already
pointed out in previous works~\cite{Andrea:2011ws,Fuks:2012im,Agram:2013wda}.
\subsection{Non-resonant monotop production}
\label{sec:lnonres}
In the second class of scenarios implying the production of monotop states, the
top quark is produced in association with an invisible bosonic field that couples
in a flavour-changing way to top and light up-type (up or charm) quarks.
The bosonic state is in general not stable since it
couples to quarks. A missing energy signature is therefore enforced by requiring
these fields either to be long-lived so that they decay outside of the detector,
or to decay predominantly into a pair of additional neutral stable particles. In particular,
the latter possibility has been proposed in the framework of flavourful dark
matter models~\cite{Kamenik:2011nb}, where the extra boson ($\phi$ or $V_\mu$) is
a mediator of the interactions of the dark matter candidate with the Standard
Model particles.
The main issue with this class of models is to ensure that the new boson leads to
a missing energy signature in a detector. In this work, we address it by
assuming that the $\phi/V$ field dominantly decays into a pair of dark matter
candidate particles.\footnote{One could
also consider that neither the boson nor its decay products are stable, but instead
long-lived. Although this is a viable assumption, this implies further
complications in the building of the model. We therefore stick with the minimal
case.} In this case, extra constraints arise
from the requirement that the particle the boson decays into is a good candidate
for dark matter, or that at least it does not overpopulate the
Universe.
\subsubsection*{Spin-0 mediator}
As already stated in Section~\ref{sec:lres}, the interactions of a scalar field
to quarks involve both the right-handed and left-handed components of the
fermions.
Consequently, the scalar $\phi$ field must transform as a doublet of $SU(2)_L$
with an hypercharge quantum number of $1/2$,
\begin{eqnarray}}% can be used as {equation} or {eqnarray
\phi =({\utilde{\bf 1}}, {\utilde{\bf 2}},1/2)
\equiv \left( \begin{array}{c} \phi^+\\ \phi^0\end{array} \right) \ .
\end{eqnarray}
The coupling in the effective Lagrangian can be written as
\begin{eqnarray}}% can be used as {equation} or {eqnarray
y^{ij}\ \phi \, \bar{q}_{L,i} u_{R,j} + {\rm h.c.} \label{eq:phinr}
\end{eqnarray}
where $i,j$ span over the quark flavours.
The new scalar has the same quantum numbers as the Brout-Englert-Higgs field, thus one can also write a coupling to the right-handed down-type quarks and mix $\phi$ with the Higgs with the potential harm of generating a
non-vanishing vacuum expectation value. The presence alone of couplings to both
up-type and down-type quarks already generates dangerous flavour-changing effects.
Nevertheless, we assume, in a first step, that the only extra coupling
with respect to the Standard Model is the one of
Eq.~\eqref{eq:phinr} and will show, in the following,
that it is already hard to construct a phenomenologically viable
model. Gauge invariance implies the
presence of interactions between the charged component field $\phi^+$ and
quarks, so that the $\phi^+$ field always promptly decays into
two-body final states, $\phi^+ \to u\bar{b}$ or $t\bar{d}$. Analogously, the
neutral component $\phi^0$ could also decay into an associated particle pair
comprised of a top and an up quark, $\phi^0 \to u\bar t+ t \bar u$, as well as into
a three-body final state via the exchange of a virtual charged scalar
field\footnote{Due to reasons already stated in Section~\ref{sec:lres},
the $\phi^+$ and $\phi^0$ states are assumed to have similar
masses.}, $\phi^0 \to W^- [\phi^+]^* + W^+ [\phi^-]^*
\to W^- \bar{b} u + W^+ b \bar{u}$ or $W^- \bar{d} t + W^+ d \bar{t}$.
All these decay channels
are however assumed to be negligible when compared to
a decay into a pair of dark matter particles. In this case, no minimal coupling
to a single stable state is achievable since $\phi$ is a doublet of
$SU(2)_L$, and one
must design an interaction of the $\phi$ state to two extra fields whose
combination forms a doublet of $SU(2)_L$. If we restrict ourselves to $\phi^0$-decays into
fermionic particles, the most minimal option is given by the Lagrangian
\begin{eqnarray}}% can be used as {equation} or {eqnarray
{\cal L}_{\phi-{\rm decay}} = y_{\chi}\, \phi \bar{\chi_d} \chi_s + {\rm h.c.} \ ,
\end{eqnarray}
where $\chi_s$ is an electroweak singlet and $\chi_d$ a weak doublet with
an hypercharge of $1/2$.
This term induces decays of both components of $\phi$
\begin{eqnarray}}% can be used as {equation} or {eqnarray
\phi_0 \to \chi_s \chi_d^0
\qquad\text{and}\qquad
\phi^+ \to \chi_d^+ \chi_s \to [W^+]^* \chi_d^0 \chi_s\ ,
\end{eqnarray}
the charged component $\chi_d^+$ being taken heavier than, but close in mass to,
the neutral component $\chi_d^0$ so that both neutral fields $\chi_s$ and
$\chi_d$ can be seen as viable dark matter candidates.
As a consequence of this non-minimal dark sector of the model, monotop
production via flavour-changing interactions of up-type quarks with a new
invisible scalar field will always be accompanied by an extra single top
production mode
\begin{eqnarray}}% can be used as {equation} or {eqnarray
p p \to t \phi^- \to t \chi_d^0 \chi_s^0 [W^-]^*\ .
\label{eq:LFCSnewproc}\end{eqnarray}
The nature and magnitude of the associated effects are very
benchmark dependent. For instance, a small mass splitting between the
component fields of $\chi$ leads to very soft $W$-boson decay products, so
that the process of Eq.~\eqref{eq:LFCSnewproc} would imply new contributions
to monotop production. On the other hand, in the case of larger mass splittings,
related new physics scenarios feature an LHC signature comprised of a single top
quark and an isolated lepton.
Nevertheless, we choose to keep the focus on minimal models, and therefore ignore, in the rest of this work,
scenarios where monotop states are produced from flavour-changing
interactions of up-type quarks with a
scalar particle mediating dark matter couplings to the Standard Model.
\subsubsection*{Spin-1 mediator}
When the mediator is a vector boson $V$, one can design very simple models since
it can be singlet under the electroweak group. In this setup,
the associated couplings involve either
right-handed or left-handed quarks and take the form
\begin{eqnarray}}% can be used as {equation} or {eqnarray
\Big(a_R^{ij}\ V_\mu \bar{u}_{R,i} \gamma^\mu u_{R,j} +
a_L^{ij}\ V_\mu (\bar{u}_{L,i} \gamma^\mu u_{L,j} + \bar{d}_{L,i} \gamma^\mu d_{L,j}) +
{\rm h.c.} \Big) \ ,
\label{eq:spin1coup} \end{eqnarray}
where the $a_{L,R}$ parameters denote the strengths of the interactions
of the $V$-field with the quarks. As in the rest of this section, we
restrict ourselves to interactions focusing on
the monotop hadroproduction modes, so that only couplings involving
the third generation are assumed to be present. Moreover, our analysis
will focus on monotop production modes enhanced by parton densities. As a
consequence, interactions to second generation quarks are ignored.
Furthermore, being a singlet, $V$ can mix via a kinetic term with the hypercharge gauge boson in the Standard
Model: such mixing will in turn generate couplings of $V$ with all the quarks and leptons proportionally to their hypercharge.
As a result, bounds on the new particle $V$
can be derived from many new production and decay processes, like for intance,
its Drell-Yan production followed by a dilepton decay.
For this reason, we ignore possible kinetic mixing of the $V$ field
in the following.
The Lagrangian terms of Eq.~\eqref{eq:spin1coup} open various decay channels
for the $V$-field. Recalling that only couplings involving
the first and third generation quarks have been retained,
non-vanishing left-handed couplings allow the mediator to promptly
decay into jets initiated by down-type quarks,
$V \to b \bar{d} + d \bar{b}$.
Next, the importance of the decays into top and
up quarks (this time both in the context of left-handed and right-handed couplings) depends
on the mass hierarchy between the mediator and the top quark, the tree-level decay
$V \to t \bar{u} + u \bar{t}$ being only allowed when $m_V > m_t$. Furthermore, when
$m_V < m_t$, a triangle loop-diagram involving a $W$-boson could also
contribute to the
decay of the $V$-field into a pair of jets, $V \to d_i \bar{d}_j$. Finally, when
$m_W < m_V < m_t$, the three-body decay channel $V \to b W^+ \bar{u} + \bar{b} W^- u$ is open,
mediated by a virtual top quark.
A monotop signal is thus expected only when the $V$-field is invisible and
dominantly decays into a pair of dark matter particles.
Since $V$ is an electroweak singlet, the associated couplings can be written,
in the case of fermionic dark matter, as
\begin{eqnarray}}% can be used as {equation} or {eqnarray
{\cal L}_{V-{\rm decay}} =
V_\mu \Big(g_{R \chi}\; \bar{\chi}_R \gamma^\mu \chi_R +
g_{L\chi} \; \bar{\chi}_L \gamma^\mu \chi_L \Big)\ ,
\end{eqnarray}
where $\chi$ is a Dirac fermion, singlet under the Standard Model gauge
symmetries. The consistency of the model, \textit{i.e.}, the requirement that
$V$ always mainly decays into a pair of $\chi$-fields and not into one
of the above-mentioned visible decay modes, implies constraints on the
Lagrangian parameters. They will be studied in details in the next section, together with
other requirements that can be applied to viable
non-resonant monotop scenarios.
For completeness, one can also couple the $V$-boson to
left-handed quarks in (non-minimal) scenarios where it lies
in a triplet of $SU(2)_L$, $V_t = ({\utilde{\bf 1}}, {\utilde{\bf 3}}, 0)$.
In this case, it is however difficult to build couplings to a
minimal dark matter sector, the simplest case being the one of a
fermionic doublet of $SU(2)_L$. This also predicts the existence
of a charged component that can be produced, at the LHC, in association
with a single top quark or that can give rise to
a monobottom signature (for small mass gaps among
the vector degrees of freedom).
Following the minimality principle, we will not consider this case any further.
\subsubsection*{Summary for the non-resonant channel}
Summarising all the considerations above, the minimal gauge-invariant Lagrangian
yielding monotop production in the flavour-changing mode is given by
\begin{equation}\bsp
\mathcal{L}_{\rm non-res} &\ = {\cal L}_{\rm kin}(V,\chi)
+
V_\mu \Big(g_{R \chi}\; \bar{\chi}_R \gamma^\mu \chi_R +
g_{L\chi}\; \bar{\chi}_L \gamma^\mu \chi_L \Big)\\ &\ +
\Big(a_R^{ij}\ V_\mu \bar{u}_{R,i} \gamma^\mu u_{R,j} +
a_L^{ij}\ V_\mu (\bar{u}_{L,i} \gamma^\mu u_{L,j} + \bar{d}_{L,i} \gamma^\mu d_{L,j}) +
{\rm h.c.} \Big)\ ,
\esp\label{eq:lnonresfinal}\end{equation}
where the first term contains kinetic, mass and gauge interaction terms
for the $V$ and $\chi$ fields. In the notations of
Ref.~\cite{Andrea:2011ws,Agram:2013wda}, the second line of the Lagrangian
reads
\begin{equation}\bsp
\mathcal{L}_{\rm non-res} = &\ {\cal L}_{\rm kin} + \Big( V_\mu\, \bar{u}_i \Big[ (a^1_{FC})^{ij} \gamma^\mu + (b^1_{FC})^{ij} \gamma^\mu \gamma^5 \Big] u_j + {\rm h.c.} \Big) \ ,
\esp\label{eq:lnonres}\end{equation}
so that the two parameterizations are related by
\begin{equation}
a_{FC}^1 = \frac{a_R + a_L}{2}\,, \quad b_{FC}^1 = \frac{a_R - a_L}{2} \ .
\end{equation}
The two parameter bases
are therefore equivalent, although gauge invariance imposes
that non-vanishing left-handed couplings relevant for monotop
production are accompanied by
interactions with left-handed down-type quarks too.
Since such interactions also enable the production of mono($b$-)jet final
states,
a full analysis of this scenario should account for monojet search results.
In the following, we will mainly focus on the case $a_L = 0$ unless specified.
Note also that in general $\chi$ may be a Majorana fermion, however this is a less likely situation as, in order to couple to $V$, the dark matter candidate is expected to carry a $U(1)$ charge.
In the following, we limit ourselves to the Dirac case, but the results in the Majorana case are qualitatively similar.
\section{Monotop phenomenology specific to non-resonant models}
\label{sec:nonrespheno}
Some features of the resonant models mediated by a scalar, like the lifetime of the invisible fermion produced in association
with the top quark, have been studied in details in
Ref.~\cite{Wang:2011uxa}. In the following, we therefore focus on various
features of non-resonant spin-1 models by studying the effective lifetime of the invisible
vector, associated single top signals, and the dark matter relic density.
We separately consider two regions of the parameter space which have very different
phenomenology: the case where the mediator is lighter than the top quark
(its mass $m_V$ being smaller than the top mass $m_t$) and the case where it
is heavier, with $m_V > m_t$.
On the basis of the minimality argument employed in the
previous section, we also restrict ourselves to the case where
monotop production is the only expected new physics signature of the
model. We therefore set $a_L = 0$ in the effective Lagrangian of
Eq.~\eqref{eq:lnonresfinal}. As stated above, non-vanishing
$a_L$ values imply mono($b$-)jets production, and the associated
constraints may be predominant. The corresponding
detailed study is postponed to future work.
\subsection{Mediators heavier than the top quark}
We first start with the scenario of heavy mediators: the mediator $V$ is not long lived as it can
always decay into a top quark.
Including in the model a $V$-decay channel into an invisible state to be
considered as a dark matter candidate
is thus always necessary.
Focusing on the minimal case, we study below the interesting interplays between the
requirement that the invisible channel dominates and bounds originating
from the relic density of the dark matter candidate.
\subsubsection{Tree-level decays of the mediator}
\label{sec:LOdecays}
When the $V$-boson is heavier than the top quark, it can decay into
either a pair of down-type quarks, an
associated pair comprised of a top quark and a lighter quark or a pair of dark
matter particles, as already discussed in Section~\ref{sec:lnonres}. Since the
first two decay modes are driven by the same interaction vertices allowing one
for monotop production, we need to make sure that the invisible decay channel
always dominates.
As we focus on the couplings to up and top quarks, we set all the couplings $a_{R,L}^{ij} = 0$ except for $a_{R,L}^{13}$ and $a_{R,L}^{31}$.
The relevant partial widths are given by\footnote{The results have been checked using the decay module
of {\sc FeynRules}~\cite{Alwall:2014bza}.}
\begin{equation}\bsp
\Gamma (V \!\to\! b \bar{d} \!+\! \bar{b} d) = &\ \frac{m_V}{4 \pi} |a_L^{31} + a_L^{13,\ast}|^2 \ , \\
\Gamma (V \!\to\! t \bar{u} \!+\! \bar{t} u) = &\
\frac{m_V}{4 \pi} |a_R^{31} + a_R^{13,\ast}|^2 \left(1-\frac{m_t^2}{m_V^2} \right) \left( 1-\frac{m_t^2}{2 m_V^2} - \frac{m_t^4}{2 m_V^4} \right)\ , \\
\Gamma (V \!\to\! \chi \chi) = &\
\frac{m_V}{24 \pi} \sqrt{1\!-\! 4 \frac{m_\chi^2}{m_V^2}}\ \Bigg[
\Big(|g_{L\chi}|^2\!+\!|g_{R\chi}|^2\Big) \Big(1\!-\!\frac{m_\chi^2}{m_V^2}
\Big) +
\frac{6 m_\chi^2}{m_V^2} \Re \{g_{L\chi} g_{R\chi}^*\} \Bigg]\ ,
\esp\label{eq:widths}\end{equation}
where we neglect all quark masses but the top mass. In addition,
we denote by $m_\chi$ the mass of the dark matter candidate.
In the minimal scenario where only right-handed couplings are present, $a_L = 0$, the decay to light quarks vanishes and we are left with two decay channels above the top threshold.
For future convenience, we define $a_R = a_R^{31} + a_R^{31,\ast}$.
In this set-up, we study typical
constraints that can be imposed on ratios of the $g_{L\chi}$, $g_{R\chi}$ and $a_R$
parameters when they are all assumed to be real quantities. Since ratios of
branching ratios are equivalent to ratios of partial widths, we use this latter
quantity and show, in Figure~\ref{fig:V2body}, the maximum value of the
$a_R$ coupling
strength in units of the $\chi V$ coupling that ensures the $V$-field to decay
invisibly in at least 99\% of the cases. In the left panel of the figure, we
consider scenarios where $g_{R\chi}$ vanishes (the same result holds for vanishing
$g_{L\chi}$), while
in the right panel of the figure, we assume vector-like couplings, $g_{L\chi} = g_{R\chi} = g_{V\chi}$. In general,
the coupling to the top quark $a_R$ (that is responsible for the monotop signal) has to be quite small compared to the coupling to the dark matter candidate in order for the mediator $V$ to be invisible, unless the mass of the mediator $V$ is close to the top mass.
On the contrary, if the mass of $V$ is close to the $\chi \chi$ threshold, the invisible decay modes are suppressed.
This study shows that it is not straightforward to have $V$ to decay invisibly, and this constraint may play an important role in the interpretation of the signal, especially when associated with the study of the properties of $\chi$ as a dark matter candidate. We study more in detail this question in the next subsection.
\begin{figure}
\centering
\includegraphics[width=.47\columnwidth]{gLchimin}
\includegraphics[width=.47\columnwidth]{gVchimin}
\caption{Maximum value of $a_R$ necessary to enforce the mediator $V$ to decay
invisibly in 99\% of the cases. We focus on scenarios where the couplings
of the mediator to dark matter are chiral with $g_{R\chi}=0$ (or
$g_{L\chi}=0$) in the left panel, and vector with $g_{L\chi} = g_{R\chi} =
g_{V\chi}$ in the right panel. The four curves correspond to $m_\chi = 5$, $75$, $100$ and $150$~GeV from
the lower to the upper ones in each figure.} \label{fig:V2body}
\end{figure}
Similar conclusions would hold in less minimal models, like the one with a left-handed coupling $a_L$ where the decay to down-type quarks is open and dominant also below the top threshold.
\subsubsection{Dark matter constraints} \label{sec:DMheavy}
We have seen that, in order to avoid visible decays of the mediator $V$, it has to be
coupled to a stable particle $\chi$ and the decay $V \to \chi \chi$ must always dominate.
If $\chi$ is stable, and if the model is minimal in the sense that $V$ is the only mediator of interactions between the dark sector and the Standard Model, then the only annihilation process that will determine the thermal relic abundance of $\chi$ is $\chi \chi \to V \to t \bar{u}$ and $\bar{t} u$.
Such process is proportional to the same coupling that gives rise to the monotop signature at the LHC, and also to the coupling of $V$ to dark matter.
By studying the relic abundance of $\chi$ one can therefore derive interesting constraints on the couplings, especially when imposing that the relic abundance is smaller than the measured density of dark matter.
Those restrictions can in principle always be evaded by assuming that there are additional
mediators, or that $\chi$ is not a stable particle but rather a long-lived one that decays on cosmological time scales.
In the rest of the section, we nevertheless focus on the minimal case of $\chi$ being the only
dark matter candidate.
As the relic abundance decreases with increasing annihilation cross sections, one can calculate a lower bound on the product of $a_R$
with the couplings of $V$ to the dark matter by requiring that the relic abundance is equal or smaller than the measured one. Values of the couplings below the bound would be excluded as the stable particle would overpopulate the Universe. The bound has been computed by implementing the model described by the
Lagrangian of Eq.~\eqref{eq:lnonresfinal} in {\sc CalcHep}~\cite{Belyaev:2012qa}.
For the calculation of the relic abundance, we used the usual approximate formulas deriving from an analytic solution of the Boltzmann equation (see Ref.~\cite{Kolb:1990vq} for more details):
\begin{eqnarray}}% can be used as {equation} or {eqnarray
\Omega_{DM} h^2 = \frac{1.04 \cdot 10^9}{M_{Pl}} \frac{x_F}{\sqrt{g_\ast}} \frac{1}{\langle \sigma v\rangle}
\end{eqnarray}
where $x_F = m_{\chi}/T_F$ and the freeze-out temperature is $T_F \sim 25$~GeV, $g_\ast = 92$ is the number of relativistic degrees of freedom at freeze-out, and all dimensionful quantities are in GeV.
We consider, for concreteness, a vectorial model with $g_{L\chi} = g_{R\chi} = g_{V\chi}$.
The results of the calculation are shown in Figure~\ref{fig:gVaR}, where we present
the lower bound on $a_R \times g_{V\chi}$ as a function of the mediator mass $m_V$ and the dark matter mass $m_\chi$.
We restrict ourselves to values of the $\chi$ mass above the top threshold, $2 m_\chi > m_t$, so that a two-body process is kinematically allowed.
Below the top threshold, the dark matter candidate can only annihilate into three-body final
states or via loop-induced processes, so that the annihilation cross section is too small and the $\chi$ particle overpopulates the Universe.
The figure shows that the product of couplings is bound to be larger than about $0.1$, with the lower bound increasing towards the top threshold as the phase space closes down, and becomes smaller towards the $V$ threshold $2 m_\chi = m_V$ where the resonant $V$ exchange enhances the annihilation. We recall that the $V$-boson mass must be at least twice as large as the
dark matter candidate mass to allow invisible decays for $V$. The corresponding regions
of the parameter space are tagged as kinematically inaccessible.
\begin{figure}[tb]
\begin{center}
\includegraphics[width=7cm]{uppergVaR}
\end{center}
\caption{Lower bound on $g_{V\chi} \times a_R$ from the dark matter relic abundance as a function of $m_V$ and $m_\chi$.} \label{fig:gVaR}
\end{figure}
This result, very interesting \textit{per se}, can be combined with other constraints to better determine
the viable regions of the parameter space of the model.
The requirement that the invisible $V$-decay dominates has allowed us, in
Section~\ref{sec:LOdecays}, to calculate a lower bound on the ratio $g_{V\chi}/a_R$ which
depends on the mediator and dark matter masses (see Figure~\ref{fig:V2body}).
Multiplying it with the limits derived from the relic abundance predictions, we
extract a lower bound on $g_{V\chi}$ independently of the value of $a_R$: the results are
shown in Figure~\ref{fig:gV}.
The lower bound on $g_{V\chi}$ is found to grow with smaller values of the $\chi$ mass.
Moreover, near the top threshold, it reaches values well above unity, tending hence to the
non-perturbative regime.
\begin{figure}[tb]
\begin{center}
\includegraphics[width=7cm]{uppergV}
\end{center}
\caption{Lower bound on $g_{V\chi}$ obtained combining the dark matter relic abundance
constraints with the requirement that the mediator $V$ decays invisibly in 99\% of the cases.}
\label{fig:gV}
\end{figure}
Under the assumption that $\chi$ is the only dark matter candidate of the theory, we can further
restrict our analysis to parameter space regions where the values of the couplings are such that
the bound from the dark matter abundance is saturated. We first reinterpret,
as a function of the masses, the limits calculated in the CMS monotop search~\cite{CMS:2014hba}
by accounting for an invisible branching ratio of the mediator that may not be 100\%.
Next, we correlate these to the dark matter results:
for increasing values of $a_R$, the coupling $g_{V\chi}$ has to be smaller to
satisfy the dark matter constraints. This indicates that an enhancement of the $tV$ production
rate (by increasing $a_R$) is accompanied by a reduction of the invisible branching
ratio of $V$, which possibly reduces the production cross section of monotop
systems.
A general bound on $a_R$ can be obtained using the relation
\begin{eqnarray}}% can be used as {equation} or {eqnarray
a_R^2 \times \frac{k^2/a_R^2 \tilde{\Gamma}_{\chi\chi}}{k^2/a_R^2 \tilde{\Gamma}_{\chi\chi} + a_R^2 \tilde{\Gamma}_{tu}} \leq a_{R-CMS}^2 \ ,
\end{eqnarray}
where $\tilde{\Gamma}$ denote the partial widths into $\chi \chi$ and $t u$ final states given
by Eq.~\eqref{eq:widths} stripped by the coupling strengths, $a_{R-CMS}$ is the upper bound on
$a_R$ derived from the CMS analysis that assumes that $V$ decays are always invisible, and
$k$ is the lower bound on $g_{V\chi} \times a_R$ deduced from the dark matter relic abundance
in Figure~\ref{fig:gVaR}.
On the left panel of Figure~\ref{fig:DM}, we extract the bound on $a_{R-CMS}$ from
the CMS analysis of Ref.~\cite{CMS:2014hba}.
Inverting the above equation, bounds on $a_R$ for a $\chi$ particle saturating the dark
matter relic abundance can then be rewritten as
\begin{equation}\bsp
&\ a_R^2 \leq \frac{k^2 \tilde{\Gamma}_{\chi\chi}}{2 a_{R-CMS}^2 \tilde{\Gamma}_{tu}} \left( 1 - \sqrt{1-4 \frac{a_{R-CMS}^4 \tilde{\Gamma}_{tu}}{k^2 \tilde{\Gamma}_{\chi\chi}}} \right) \ , \\
\quad\text{or}\quad
&\ a_R^2 \geq \frac{k^2 \tilde{\Gamma}_{\chi\chi}}{2 a_{R-CMS}^2 \tilde{\Gamma}_{tu}} \left( 1 + \sqrt{1-4 \frac{a_{R-CMS}^4 \tilde{\Gamma}_{tu}}{k^2 \tilde{\Gamma}_{\chi\chi}}} \right)\,.
\esp\label{eq:arineq}\end{equation}
The result is shown on the right panel of Figure~\ref{fig:DM}.
Above the blue curve, the argument of the square root is negative and the inequalities of
Eq.~\eqref{eq:arineq} have no solution, therefore there is no bound that can be applied on $a_R$.
Below the blue line, near the top threshold, the dark matter constraint requires larger couplings and therefore larger monotop rates are allowed, thus a bound on $a_R$ can
be calculated. Naturally, larger portions of the parameter space are expected
to be covered with the upcoming run II of the LHC.
\begin{figure}[tb]
\begin{center}
\includegraphics[width=7cm]{boundCMS}
\includegraphics[width=7cm]{uppergDM}
\end{center}
\caption{Reinterpretation of the CMS monotop limits of Ref.~\cite{CMS:2014hba}
in terms of $a_R$ (left) for a 100\% invisible mediator. The results are then used to determine the viable regions of the
parameter space when enforcing dark matter and LHC constraints as shown by
Eq.~\eqref{eq:arineq} (right). The region above the blue line is found not to be bounded by
current searches. Below the blue line, limits on $a_R$ deduced from the left panel of the
figure are in order.} \label{fig:DM}
\end{figure}
The region where the monotop signal is suppressed can have interesting additional features.
The boson $V$ may dominantly decay into top and lighter quarks, yielding at the same time
a signature comprised of same-sign top quark pairs ($t V \to t t \bar{u}$) and extra
contributions to top-antitop production ($t V \to t \bar{t} u$) that may be difficult
to observe due to the overwhelming $t\bar{t}$ Standard Model background. These extra channels
deserve a particular attention, in particular in upcoming data from LHC collisions
at $\sqrt{s}=13$~TeV.
\subsection{Mediators lighter than the top quark}
\label{sec:phenolightv}
When the spin-1 mediator $V$ is lighter than the top quark, its possible decay
modes into a top and a lighter quark are kinematically forbidden.
At tree-level, in the minimal scenario $a_L = 0$, $V$ can therefore only decay into a multibody final state
such as $V \to u \bar{b} W^-$ or $\bar{u} b W^+$, where the $W$-boson is
virtual when $m_V < m_W$ ($m_W$ denoting the $W$-boson mass)~\footnote{When $a_L \neq 0$, the decays into a pair of down quark will always dominate.}. In
this mass range, loop-induced decays must however be considered too. For instance,
a triangle loop-diagram with a $W$-boson exchange
generates couplings to down-type quarks, thus opening a dijet decay channel.
As the decay channels in this region are either kinematically or loop-suppressed, one may wonder
whether $V$ may be long-lived without the need
for an additional invisible decay channel.
Another interesting property of this mass region is that a new decay of the top quark
is allowed, $t \to u V$, and extra constraints on monotop scenarios could
therefore be extracted
from, \textit{e.g.}, top width measurements or the analysis of $t \bar{t}$ events when one
of the top quarks decays into a jet plus missing energy.
\subsubsection{Loop-induced and multibody decays of the mediator}
\label{sec:loopdecays}
Light mediators, below the top mass threshold, may decay dominantly
into two jets via loop-induced interactions.
The structure of the loop crucially depends on the chirality of the monotop couplings: for instance, in the case $a_L \neq 0$, the loop is divergent,
which signals that a tree-level coupling to down-type quarks is necessary for the theory to be consistently renormalizable.
This result confirms the necessity to consider an effective model fully invariant under the electroweak symmetry.
In the minimal case $a_L=0$, the loop contributions turn out to be finite.
Since weak interactions are left-handed, the
chiralities of the quarks involved in these diagrams must be flipped, which
implies that the loop-induced couplings are proportional to the product of the
up and top masses $m_u m_t$. Contrary to setups where monotops are produced from
left-handed interactions of the mediator with quarks, the loop-induced $Vd_Ld_L$
couplings are thus finite, in line with the fact that no associated
counterterm appears after renormalization. The interaction strength reads, in
the limit of small light quark masses,
\begin{eqnarray}}% can be used as {equation} or {eqnarray
g_{Vd^id^j}^{\rm 1-loop} (a_R)= \frac{\alpha a_R}{4 \pi s_W^2} \frac{m_u}{m_t} (V_{ud^i}^* V_{td^j} + V_{td^i}^* V_{ud^j}) \tilde{c}_0\ ,
\end{eqnarray}
where $\alpha$ stands for the electromagnetic coupling constant, $s_W$ for the
sine of the weak mixing angle and $V_{ij}$ for the elements of the CKM matrix.
In addition, the loop factor
\begin{eqnarray}}% can be used as {equation} or {eqnarray
\tilde{c}_0 = m_t^2\ C_0 (p_1, -(p_1+p_2); m_W, m_t, 0)
\end{eqnarray}
depends on the Passarino-Veltman three-point function $C_0$ where $p_1$ and
$p_2$ are the momenta of the external down-type quarks.
We can therefore calculate the partial width associated with the
decay $V\to\bar{d}^id^j$ which reads, after summing over
all down-type quark flavours,
\begin{eqnarray}}% can be used as {equation} or {eqnarray
\Gamma (V \to jj) = \frac{\alpha^2 a_R^2}{64 \pi^3 s_W^4} \frac{m_V m_u^2}{m_t^2} |\tilde{c}_0|^2\ .
\label{eq:loopwidth}\end{eqnarray}
We observe that it exhibits both a loop-suppression and a
$(m_u/m_t)^2$ factor, so that it is expected to be numerically small.
\begin{figure}
\centering
\includegraphics[width=.47\columnwidth]{GammaVloop}
\includegraphics[width=.47\columnwidth]{aRVmaxloop}
\caption{Partial decay width associated with the loop-induced (blue)
and multibody (green) decay of the
$V$-field into down-type quarks as a function of the $V$-boson mass, as
given by Eq.~\eqref{eq:loopwidth} and returned by {\sc MadWidth} (left panel). We consider
scenarios in which $a_R = 0.04$ and $m_V$ is kept smaller than the top mass.
The results are translated, in the right panel, as a bound on $a_R$ that
ensures that the $V$-boson has a decay length of at least 50 m.}
\label{fig:V2bodyloop}
\end{figure}
In Figure~\ref{fig:V2bodyloop}, we show the partial width in Eq.~\eqref{eq:loopwidth} as a function of the mediator mass for $a_R = 0.04$ (left panel, blue curve).
We compare this result to preditions for three-body and four-body decays (left panel, green) as
calculated by {\sc MadWidth}~\cite{Alwall:2014bza}, which turn out to be dominant upon the entire
mass range.
On the right panel of the figure, the partial width is translated as an upper bound on the value of $a_R$ in order for $V$ to have a mean decay length of at least 50 metres so that it is long-lived enough to decay outside of typical hadron collider detectors.
The figure shows that the lifetime of $V$ would be long enough only for extremely small
values of the coupling $a_R$ that will challenge the possible
observation of a monotop signal at the LHC by reducing the associated production cross section.
The only way out is thus to extent the theoretical framework so that invisible decay channels
are enabled, as in the previous section.
It should also be mentioned that above the $W$-boson threshold, a tree-level three-body decay is kinematically open, which further shortens the decay length of $V$. Finally, in cases where the model features a coupling of the $V$-field to
top and charm quarks, the
partial width of Eq.~\eqref{eq:loopwidth} would exhibit an enhancement
proportional to $(m_c/m_u)^2$.
In summary even for monotop scenarios in which the mediator cannot decay into a top quark,
its lifetime is generally too short and one needs to complete the model by adding a decay channel into an invisible state.
Although the class of minimal scenarios described in this section features a light extra
vector boson, the setup is compatible with current Tevatron and LHC bounds on
monotop production as the latter are always derived under the assumption of very
large coupling values of ${\cal O}(0.1)$~\cite{Aaltonen:2012ek,CMS:2014hba}.
They could however be constrained by other observations, as will be shown in
the next subsections.
\subsubsection{Single top constraints on monotop scenarios}
\begin{figure}
\centering
\includegraphics[width=.5\columnwidth]{Gammatop}
\caption{Partial width associated with the $t\to V u$ decay mode of the
top quark as a function of the $V$-boson mass and the $a_R$ coupling. We show
curves for a partial width of 3, 1, 0.5, 0.1 and 0.01~GeV. The solid black curve corresponds to the upper bound on $a_R$ from a partial width less than $3$~GeV, roughly corresponding to the direct measurement in Ref.~\cite{Aaltonen:2013kna}.}
\label{fig:GammaTop}
\end{figure}
Motivated by minimality principles, we have discussed, in the previous section,
appealing monotop scenarios in which the mediator $V$ is lighter than the top
quark. In this case, the
former couples to up and top quarks via right-handed couplings and one needs to add an invisible decay channel to potential dark matter particles $\chi$ to guarantee a monotop signature,
unless the coupling strength $a_R$ is very small. On different
grounds, these scenarios feature a new decay channel for the top quark,
$t \to u V$. This observation can be used to further restrict the viable regions
of the parameter space by imposing that new
physics contributions to the top width do not challenge the measured value of
$1.10<\Gamma_t < 4.05$~GeV~\cite{Aaltonen:2013kna}, from direct measurements at CDF.
A more precise measurement, which takes the value $\Gamma_t = 2.0 \pm 0.5$~GeV~\cite{Beringer:1900zz}, can also be obtained by fitting the single-top measurements.
However, the latter does not apply in our scenario where new physics contributions to single-top can arise.
Assuming a good agreement between the Standard Model expectation and the top width measurement, therefore, the partial
width $\Gamma(t\to V u)$ can thus be enforced to be of at most 3~GeV. On
Figure~\ref{fig:GammaTop}, we present the dependence of this partial width on the
coupling $a_R$ and the mediator mass $m_V$. We observe that for couplings smaller
than 0.01, new physics effects in the top width are predicted to be very small,
except when the mediator is almost massless. This consequently disfavours such
setups in which the mediator is very light, even in cases with coupling strengths
of ${\cal O}(0.001)$.
Kinematically allowed $t\to Vu$ decays also imply that monotop events can be
issued from the production of a top-antitop pair when one of the top quarks decays
into a $V$-boson and a light quark,
\begin{eqnarray}}% can be used as {equation} or {eqnarray
pp \to t \bar{t} \to t \bar{u} V \qquad\text{or}\qquad
pp \to t \bar{t} \to \bar t u V \ .
\end{eqnarray}
This process induces additional contributions to the production of a monotop system ($tV$ or
$\bar t V$) in association with an additional jet, a signature already
accounted for in the LHC monotop analysis of Ref.~\cite{CMS:2014hba}.
How much this new channel will contribute to the monotop signal depends on the cuts employed
in the experimental analysis. However, due to the large $t\bar{t}$ cross section,
these effects cannot be neglected.
Complementary constraints on this channel could be deduced from Standard Model single top
analyses whose signal regions could capture monotop events as above. For instance,
both CMS~\cite{Khachatryan:2014iya} and ATLAS~\cite{Aad:2012ux} have analyses dedicated to the measurement of the single top cross section in
the $t$-channel which contain
a region that could be populated by monotop events as
above.\footnote{Other single top analyses could be
considered. However, they in general use multivariate techniques that cannot be
employed in the reinterpretation framework pursued below.}
In the CMS analysis, events are selected by requiring one
single isolated electron or muon and exactly two jets, one of them being
$b$-tagged. The background is reduced by requiring an important amount of missing
energy and by imposing that the transverse mass computed after combining the
lepton transverse momentum with the missing transverse momentum is large.
A final selection is preformed by means of an advanced multivariate technique.
We have nevertheless to ignore this last step of the selection as the
amount of information provided in the experimental publication is not sufficient
for satisfactorily recasting it (see Ref.~\cite{Kraml:2012sg} for more
information on this aspect).
\begin{figure}
\centering
\includegraphics[width=.5\columnwidth]{cms12038}
\caption{Parameter space region excluded (the area above the red curve)
at the 95\% confidence level by the
CMS single top analysis of Ref.~\cite{Khachatryan:2014iya} before the multivariate selection, compared to the exclusion from the top width~\cite{Aaltonen:2013kna} (black line), and the CMS monotop search~\cite{CMS:2014hba} (blue line).}
\label{fig:cms12038}
\end{figure}
We simulate our new physics signal by using the monotop model~\cite{Andrea:2011ws}
implemented in the {\sc FeynRules} package~\cite{Christensen:2008py,Alloul:2013bka},
tuning the model parameters to the setup of Eq.~\eqref{eq:lnonresfinal}, so that
we can export the model to a UFO library~\cite{Degrande:2011ua} that is then
linked to
{\sc MadGraph5\_aMC@NLO}~\cite{Alwall:2014hca}. The generated parton-level events
have subsequently been processed by {\sc Pythia}~\cite{Sjostrand:2006za} for
parton showering and hadronization and by {\sc Delphes}~\cite{deFavereau:2013fsa}
for detector simulation, making use of the recent `MA5Tune'~\cite{Dumont:2014tja}
of the CMS detector description of {\sc Delphes}. The CMS analysis of
Ref.~\cite{Khachatryan:2014iya} has finally been implemented in the
{\sc MadAnalysis5} framework~\cite{Conte:2012fm,Conte:2014zja}, which has allowed
us to derive exclusion bounds at the 95\% confidence level in the $(m_V,a_R)$
plane, as shown on Figure~\ref{fig:cms12038}.
The figure also shows the constraint from the top width, and from the dedicated CMs monotop search~\cite{CMS:2014hba}.
The monotop search is currently more sensitive.
However, the bound from the single top is a rough estimate, and the bound may be much stronger once the full analysis, including
the multivariate selection, is taken into account.
Nevertheless, our result shows that the constraints from single top searches can play an important role in constraining monotop scenarios.
The only region in parameter space where the monotop search will always be dominant is the region where the vector is close in mass to the top, because the single top channel will be suppressed by phase space while the monotop signal is not.
\subsubsection{Dark matter constraints}
We have argued that, even for mediator masses below the top threshold, an invisible decay channel is typically needed in order for the monotop signature to be present.
The simplest way out is to couple $V$ to a fermionic stable dark matter candidate $\chi$.
However, in a minimal scenario where $V$ is the only mediator for the interactions of the dark matter candidate, one needs to ask whether the relic abundance of $\chi$ is enough to fulfill the bounds from observations.
Below the top threshold, the main annihilation process $\chi \chi \to V \to t \bar{u}$ and $\bar{t} u$ is kinematically forbidden,
so that the annihilation of dark matter particles can only proceed to a
three-body or four-body final state (via a virtual top quark), or via loop-diagrams $\chi \chi \to V \to d_i \bar{d}_j$.
As discussed in Section~\ref{sec:loopdecays}, the loop contributions
are suppressed by the mass of the light up-type quark that the mediator couples to.
In Section~\ref{sec:DMheavy}, we have shown
that in the region where two-body final states are allowed, the relic abundance requires the couplings to be fairly large, therefore in the light mass region
the $\chi \chi$ annihilation rate is, without doubts, too slow for the stable particle $\chi$ not to overpopulate the Universe.
One possible way-out is to turn on the $a_L$ coupling. In this case, we open back a two body decay $\chi \chi \to b \bar{d}$ and $\bar{b} d$, and similar numerical results as in Section~\ref{sec:DMheavy} at the price of a less minimal scenario. A second possibility is to have a very small $a_R$ so that $V$ is long lived at the price of suppressing the monotop signal beyond any hope of detectability, or complicating the dark sector so that $\chi$ is not the dark matter candidate.
Following this argument, we can state that the minimal monotop scenario is excluded by dark matter relic abundance constraints when the mediator is lighter that the top.
\section{Conclusions}
\label{sec:conclusions}
Monotop final states comprised of a single top quark produced in association with missing
energy can be a striking sign of new physics at the LHC.
The main production mechanisms can be divided into two classes: resonant production, where a
heavy coloured boson is first produced in the $s$-channel and further decays via its
couplings to a single top quark and an invisible neutral fermion, and non-resonant production
where the top quark is produced in association with an invisible boson that couples to top and
up (or charm) quarks.
A complete and model independent parametrisation of the two channels has been provided in
Ref.~\cite{Andrea:2011ws}.
In the present work, we have revisited this description by embedding the effective interactions
in an SU(2)$_L \times$ U(1)$_Y$ invariant formalism.
In doing so, we have shown that, depending on the chirality of the tops, a complete model contains necessarily extra states and couplings that may spoil the monotop signal, or add more new physics signatures that should be studied in association with the monotop one.
We have identified two minimal setups. In the first case,
a scalar field is resonantly produced by the fusion of a pair of down-type quarks and
couples to a right-handed top quark and a new
invisible fermion, like a right-handed stop in $R$-parity violating supersymmetry. In the second case, a vector
state couples to right-handed top and up quarks and decays dominantly into new invisible fields,
like in models of dark matter where the dark sector couples to the Standard Model via a
flavour-sensitive mediator.
We have further investigated the phenomenology of the second class of models that can be split
into two subclasses, depending on the mass of the mediator.
For mediators lighter than the top quark, their visible decay modes are either loop-suppressed
or phase-space-suppressed, or both. Nevertheless, one always needs to add (and tune the couplings
of) an invisible
field to prevent the mediator from decaying inside a typical hadron collider detector as
this would otherwise spoil the monotop signature originally motivating the model.
An important feature of these scenarios is that they allow for the top quark to decay into the
mediator and an extra jet. This feature can enhance the monotop production rate, as the monotop
system can be produced in association with an extra jet from
$t \bar{t}$ events when one of the top quarks decays in the exotic channel. Such events
could also be searched for in standard typical single-top searches, as they are
expected to populate signal regions of associated analyses.
We have indeed observed that a CMS analysis of single top events could imply significant
constraints on the mediator couplings, competitive and sometimes
stronger than those obtained from monotop searches.
Scenarios with a mediator mass above the top threshold have a very different phenomenology
as the mediator decays significantly into top quarks and jets. One needs a large coupling
to the invisible sector in order to preserve the monotop signature.
Describing the dark sector with a new fermion $\chi$, we have found that the latter could be
a viable dark matter candidate if heavier than half the top quark mass, with a correct relic abundance
driven by its annihilation via an $s$-channel mediator into a top and an up quark.
We have used relic abundance constraints to derive lower bounds on the product of the couplings
of the mediator to quarks and to the dark matter candidate.
We have then further restricted the monotop parameter space by combining cosmological
and collider results and enforcing the mediator to decay mostly invisibly.
We have found that the issue of the perturbativity of the model could be raised
for dark matter masses close to the top mass and that the parameter space
turns out to be largely constrained when the $\chi$ fermion is demanded to reproduce the
observed relic density. However, a large portion of the parameter space is still
left unconstrained by current data and future experimental results are in order,
in particular analyzing a same-sign top quark pair final state arising from the visible decays of the mediator.
\acknowledgments
We thank T.~Theveneaux-Pelzer, J.~Donini and F.~Maltoni for useful discussions and comments. We acknowledge
support by the Theory-LHC France initiative of CNRS/IN2P3.
G.C. and A.D. acknowledge partial support from the Labex-LIO (Lyon Institute of Origins) under grant ANR-10-LABX-66
and FRAMA (FR3127, F\'ed\'eration de Recherche ``Andr\'e Marie Amp\`ere"). A.D. is partially supported by Institut
Universitaire de France.
\bibliographystyle{JHEP}
|
\section{Introduction}
\label{intro}
Elastography is an imaging technology based on biomechanical contrast;
among its current clinical applications are early detection of skin, breast and prostate cancer,
detection of liver cirrhosis, and characterization of artherosclerotic plaque in vascular imaging
(see for instance \cite{Doy12,ParDoyRub11,WojFarWebThoFis10,AigPalSchoLebJun11,WanChanHunEngTun09,WooRomLerPanBro06,BisPatParCicRic10}).
Typically, elastography is implemented as an \emph{on top imaging method}
to various existing imaging techniques, such as ultrasound imaging (see for instance \cite{LerParHolGraWaa88,OphCesPonYazLi91}),
magnetic resonance imaging (see for instance \cite{MutLomRosGreMan95,ManOliDreMahKru01})
or optical coherence tomography (see for instance \cite{SunStaYan11,NahBauRouBoc13}).
With all these techniques, it is possible to visualize momentum images,
from which mechanical displacement $\b u$ can be calculated, which forms the basis of clinical examinations.
For motion estimation in \emph{ultrasound elastography (USE)}, \emph{optical coherence elastography (OCE)}
and in certain variants of \emph{magnetic resonance elastography (MRE)},
common techniques are optical flow and motion tracking algorithms \cite{ParDoyRub11,BohGeiAndGebTra00,Schm98,PraBha14,PriMcV92,FuChuiTeoKob11};
in USE and OCE, these are specifically referred to as \emph{speckle tracking methods}.
Speckle tracking can only be realized if the imaging data contains a high amount of correlated pattern information.
This is the predominant structure in ultrasound imaging.
Photoacoustic imaging is an emerging functional and morphological imaging technology,
which, for instance, is particularly suited for imaging of vascular systems \cite{Bea11,NusSlePal14,LiWan09}.
Opposed to ultrasound imaging, photoacoustic imaging is considered to reveal little speckle patterns \cite{LiWan06},
which is considered an advantage for imaging but a disadvantage for elastography.
Passive coupling of photoacoustic imaging and elastography has been reported in \cite{EmeAglShaSetSco04},
where the contrast of photoacoustic imaging, ultrasound, and US-elastography has been fused (see also \cite[sec.4.9]{ParDoyRub11}).
Active coupling of photoacoustic and elastography has not been reported so far.
The reason for that is that motion estimation and speckle tracking
cannot be implemented reliably because of homogeneous regions in monospectral photoacoustic imaging,
which do not allow for detection of microlocal displacements.
In this paper, we provide a mathematically founded way of introducing speckles in photoacoustic imaging data.
Theoretically, photoacoustic imaging is based on the assumption
that the whole frequency spectrum can be measured with the detectors.
Common ultrasound imaging, on the contrary, is operating with a fixed single frequency mode.
This superficial comparison motivates us to investigate, using \cite{Hal11,HalSchZan09a}, the effect band-limited measurements have on the inversion.
In fact, as we show by mathematical consideration,
the use of band-limited data enforces speckling-like patterns in the reconstructions.
Our suggested approach then consists then of carefully choosing a frequency band of measurements and back-projecting these data.
Because these data is speckled, it can be used to support tracking and optical flow techniques for displacement estimations.
The structure of the article is as follows:
We first review the principles of elastography in section~\ref{sec:Elastography}.
In section~\ref{sec:PAI} we review the principles of photoacoustic imaging.
Then, in section~\ref{sec:Texture}, we describe the methods to create texture patterns in photoacoustic imaging.
In section~\ref{sec:MotionEst}, the methods of motion estimation for photoacoustic elastography are described, and
in section~\ref{sec:Experiments}, we show the results of imaging experiments.
The paper ends with a discussion (section~\ref{sec:Discussion}).
\section{Elastographic imaging}
\label{sec:Elastography}
In this section we explain the basic principles of elastography. In theory, elastography can be implemented on top of any imaging technique.
Below, we review mathematical models which are used for qualitative elastography.
\subsection{Experiments and measurement principle}
According to \cite{Doy12}, elastography consists of the following consecutive steps:
\label{subsec:Elastography}
\begin{enumerate}
\item The specimen is exposed to a mechanical source. Imaging is performed before and during source exposition.
\item {\bf Qualitative elastography:} From the images the tissue displacement $\b u$ is determined.
\item {\bf Quantitative elastography:} Mechanical properties are computed from the displacement $\b u$.
\end{enumerate}
In the literature there have been documented various ways to perturb the tissue,
such as quasi-static, transient and time-harmonic excitation.
In this paper we focus on qualitative elastography in the quasi-static case, which is reviewed below.
\subsection{Quasi-static qualitative elastography}
\label{sec:Quasi_static_elast}
Although it is theoretically possible to perform quantitative imaging all at once,
in practice, qualitative imaging is performed beforehand.
Depending on the used modalities different models are used for qualitative elastography
(see for instance \cite{ParDoyRub11}):
We start from images $f(\b x,t)$, which are recorded before and during the mechanical excitation.
These images can be B-scan data in US-imaging, MRI magnitude images, OCT images,
or in principle, images from any modality \cite{ParDoyRub11,WasMig04}.
A common model then is to assume that
\begin{equation}
\label{eq:Advektion}
f(\b x(t),t)= const.
\end{equation}
for every time $t$, assuming that the intensities are transported along the trajectories $\b x(t)$. according to the vector field $\b u(\b x)$.
A model such as \eqref{eq:Advektion} can serve as a basis for an image registration model to recover the displacement $\b u=\dot{\b x}(t)$ from $f$ (as in \cite{LedKybDesSanSuh05}
for detection of the movement of the heart). For smaller displacements typically encountered in elastography,
the constraint \eqref{eq:Advektion} can be linearized
\begin{equation}
\label{eq:OptFlow}
\nabla f \cdot \b u + f_t = 0\;,
\end{equation}
which can serve as a basis for inversion.
In a \emph{quasi-static} experiment, there are two images:
before and after the mechanical excitation from the exterior,
which we denote as $f_1(\b x)=f(\b x,t_1)$ and $f_2(\b x)=f(\b x,t_2)$.
We are calculating the spatial dependent flow $\b u(\b x)$ only.
In this case we are solving the semi-continuous equation:
\begin{equation}
\label{eq:OptFlow_sc}
\nabla f_1 \cdot \b u + (f_2-f_1) = 0\;.
\end{equation}
The equation is underdetermined for $\b u$ and therefore regularization has to be involved for stable solution.
Typically, there are some constraints here as regularization as in the Horn-Schunck model \cite{HorSchu81}:
that is
\begin{equation}
\label{eq:HS}
\b u = \argmin_{\mathbf{v}} \Norm{\nabla_{\bf x} f \cdot \mathbf{v} + f_t}^2_{L^2(\Omega)} + \lambda \int_\Omega \abs{\nabla_{\b x} \mathbf{v}}^2 \, \mathrm{d}\mathbf{x}
\end{equation}
Other choices of regularization and data terms are possible.
An alternative to the optical flow approach is block matching \cite{BohGeiAndGebTra00}. Here, one assumes that the displacement is constant in defined regions; using a target block, one compares the image patterns in subsequent frames by using a correlation measure.
We emphasize that all these techniques assume that texture is present in the image.
In MRI, it was observed that part of tissue motion is invisible in magnitude images because of homogeneous regions.
To overcome this limitation, artificial tags have been introduced in the image \cite{PriMcV92,FuChuiTeoKob11}.
These make motion estimation possible in regions where no intensity is initially present.
In ultrasound imaging and optical coherence tomography, texture is provided by patterns in the images referred to as speckle.
These are correlated texture patterns which provide a signature of the points.
Therefore, motion estimation techniques in USE or OCE are sometimes comprehensively denoted as \emph{speckle tracking} algorithms \cite{RevMirMcn05,ZakQinMau10}. Often, the term \emph{speckle tracking} is only used for the block-matching-type algorithms \cite{BohGeiAndGebTra00,PanGaoTaoLiuBai14}.
In the next section, we review photoacoustic imaging and its image model.
In section \ref{sec:Texture}, we investigate how the band-limitation
effect will create such a speckle-like texture pattern in photoacoustic image data.
\section{Photoacoustic imaging}
\label{sec:PAI}
Photoacoustic imaging (PAI) is among the most prominent coupled-physics techniques \cite{ArrSch12}.
It operates with laser excitation and records acoustic pressure, as the coupled modality.
We first review the imaging formation in PAI.
\subsection{Mathematical modeling}
\label{sec:2}
Commonly, in Photoacoustics, the wave equation is used to describe the propagation of the acoustic pressure $p$:
\begin{align}\label{eq:wav_gen}
\begin{aligned}
p_{tt}-\Delta_x p\,&=\,I_t f,\qquad\text{in }\mathbb R^n\times(0,T],\\
p\,&=\,0,\qquad\text{in }\mathbb R^n\times(-\infty,0).\\
\end{aligned}
\end{align}
The function $I$ models the laser excitation and is usually considered a time dependent $\delta$-distribution.
The function $f$ represents the capability of the medium to transfer electromagnetic waves into pressure waves;
$f$ is material dependent and is visualized in photoacoustic imaging.
Details of deduction of \eqref{eq:wav_gen} from the Euler equations and the
diffusion equation of thermodynamics can be found for instance in \cite{SchGraGroHalLen09}.
If we assume the excitation to be perfectly focused in time (that is $I(t)=\delta(t)$),
equation (\ref{eq:wav_gen}) can be reformulated as a homogeneous initial value problem \cite{SchGraGroHalLen09}
\begin{align}\label{eq:wav_delta}
\begin{aligned}
p_{tt}-\Delta_x p\,&=\,0,\qquad\text{in }\mathbb R^n\times(0,\infty),\\
p(t=0)\,&=\,f,\qquad\text{in }\mathbb R^n,\\
p_t(t=0)\,&=\,0,\qquad\text{in }\mathbb R^n.
\end{aligned}
\end{align}
This (direct) problem is well-posed under suitable smoothness assumptions on $f$ (see, e.g., \cite{Eva98}).
We denote by
\begin{align}\label{eq:pai1a}
\begin{aligned}
\mathcal Pf({\bm{x}},t)\,=\,p({\bm{x}},t),\quad {\bm{x}}\in\mathbb R^n,~t\in(0,\infty)\,,
\end{aligned}
\end{align}
the operator that maps the initial pressure $f$ to the solution of \eqref{eq:wav_delta}.
\begin{remark}\label{rem:extension}
Since we want to apply a convolution to our solution $p$,
we have to extend it to negative values of $t$ in a way that the wave equation \eqref{eq:wav_delta} is still fulfilled.
We distinguish the causal extension $\mathcal Pf=0$ for $t< 0$ (that we denote again with the letter $\mathcal P$),
and the even extension
\begin{align}\label{eq:PAI_even}
\mathcal{P}_\text{even}f({\bm{x}},t):=\left\{
\begin{array}{cll}
&\mathcal{P} f({\bm{x}},t),\quad &t \geq 0,\\
&\mathcal{P} f({\bm{x}},-t),\quad &t<0.
\end{array}\right.
\end{align}
\end{remark}
\subsection{Photoacoustic imaging as an inverse problem}
\label{sec:3}
In Photoacoustics, we assume the pressure to be measured on a surface $\Gamma$ over time.
The inverse problem now consists of reconstructing the initial pressure $f$ in \eqref{eq:wav_delta} by these data,
ideally given as trace of the solution on $\Gamma$.
For the sake of simplicity of notation, we are denoting this operator by
\begin{align}\label{eq:pai1b}
\begin{aligned}
\mathcal Pf\,=\,p\big|_{\Gamma\times(0,\infty)}
\end{aligned}
\end{align}
as well. Here, $\mathcal P$ is mapping $f$ to the trace of the solution $p$ of \eqref{eq:wav_delta} at the surface $\Gamma$.
The \emph{Photoacoustic inverse problem} consists in solving equation \eqref{eq:pai1b} for $f$.
This problem obtains a unique solution, provided $\Gamma$ is a so-called uniqueness set
(for a review over existing results see \cite{KucKun08}).
These uniqueness sets contain the case of a closed measurement surface surrounding the Photoacoustic source.
For some of the most important simple geometrical shapes of closed manifolds $\Gamma$,
there exist analytical reconstruction formulae of series expansion and/or filtered\- backprojection
type (see again \cite{KucKun08} and the references therein,
for instance \cite{Nor80,NorLin81,Faw85,Ram85,Nil97,Pal04,FinRak05,FinHalRak07,ElbSchSchu12}).
This paper focuses on the case where $\Gamma$ is a sphere in $\mathbb R^2$ (circle) or $\mathbb R^3$.
For photoacoustic reconstruction, we make use of the explicit filtered
back\-projection formulas established in \cite{FinRak05,FinHalRak07}.
Since we will have to deal with initial sources not necessarily of compact support,
we remark that a result in \cite{AgrBerKuc96} guarantees injectivity of the
photoacoustic problem provided certain integrability conditions on the source hold.
Particularly, the photoacoustic mapping is injective if and only if the source is
$L^p$-integrable on the entire space, where $p\leq 2n/(n-1)$.
\begin{figure*}
\centering
\includegraphics[width=10cm]{speckled_point.png}
\caption{Point source (left) and textured reconstruction (right)}
\label{fig:2}
\end{figure*}
\section{Photoacoustics with band-limited data}
\label{sec:Texture}
We create speckle patterns computationally from photoacoustic data using band-limited measurements
for backprojection and approximating the initial source $f$.
To be more precise, instead of measuring the exact trace of the solution of \eqref{eq:wav_delta} at $\Gamma$,
we instead assume to measure the bandlimited data $m=\phi*_t p$.
The mathematical background is an application of some results by Haltmeier \cite[Lmm.~3.1]{Hal11}
(see also~\cite{HalSchZan09a,HalZan10}) to convolution kernels which do not
necessarily have compact support. Before we state the theorem, we define the Radon und Fourier transform:
\begin{definition}
The Radon transform $\mathcal R\varphi(\theta,s)$ maps $\varphi({\bm{x}})$ to its integrals
over hyperplanes in $\mathbb R^n$ with distance $s\in\mathbb R$ to the origin
and unit normal vector $\theta\in S^{n-1}$. Namely,
\begin{align}\label{eq:Radon}
\mathcal R\varphi(\theta,s)\,=\,\int_{\theta\cdot{\bm{y}}=s}\varphi({\bm{y}}) d{\bm{y}}.
\end{align}
In the case $n=1$, the Radon transform corresponds to the absolute value of the function.
In the case where $\varphi$ is rotationally symmetric, the Radon transform
$\mathcal R\varphi$ is independent of $\theta$. We can therefore write
\begin{equation}
\label{eq:Kleinvieh}
\mathcal R\varphi(\theta,s)=\phi(s)
\end{equation}
for a suitable, even function $\phi:\mathbb R\to\mathbb R$.
\end{definition}
\begin{definition}
The $n$-dimensional Fourier transform $\widehat\varphi({\bm{\kappa}})$ of $\varphi$ is defined as
\begin{align}\label{eq:Fourier}
\widehat\varphi({\bm{\kappa}})\,=\,\int_{\mathbb R^n} \varphi({\bm{y}})e^{-\mathrm{i}{\bm{y}}\cdot{\bm{\kappa}}}d{\bm{y}}\;.
\end{align}
If not stated differently, the Fourier transform of a time-dependent function
$q({\bm{x}},t)$ is with respect to the time variable,~i.e.
\begin{align*}\label{eq:Fourier}
\widehat q({\bm{x}},\kappa)\,=\,\int_{\mathbb R} q({\bm{x}},t)e^{-\mathrm{i} t\kappa}dt\;.
\end{align*}
\end{definition}
Now we are ready to state the theorem:
\begin{theorem}\label{thm:conv_thm}
Let $p=\mathcal P f$ be a solution of \eqref{eq:pai1a} with initial pressure
$f\in C^\infty_c(\mathbb R^n)$.
Furthermore, assume that $\,\Psi \in L^p(\mathbb R^n)$, for some $p$ such that $1\leq p < n/(n-1)$,
is a radially symmetric convolution kernel, i.e., $\Psi({\bm{x}})\,=\,\psi(|{\bm{x}}|)$. Then
\begin{equation}
\label{eq:Haltmeier}
(\mathcal{P}(\Psi *_{\bm{x}} f))({\bm{x}}, t)
\,=\, ( \mathcal{R} \Psi *_t \mathcal{P}_{\mathrm{even}} f) ({\bm{x}}, t) \,,
\end{equation}
for all ${\bm{x}} \in\mathbb R^n$, $t>0$. The convolution acts with respect to the
${\bm{x}}$-variable on the left hand side and with respect to the $t$-variable
on the right hand side of (\ref{eq:Haltmeier}), respectively.
\end{theorem}
\begin{proof}
First we note that $\mathcal P(\Psi*_{\bm{x}} f)\in L^p(\mathbb R^n)$:
Since we have
\[\Psi*_{\bm{x}}\Delta f\,=\,\Delta(\Psi*_{\bm{x}} f),\quad\text{in }\mathbb R^n\times(0,\infty)\,,\]
we immediately conclude that
\begin{align}\label{eq:Conv_space_wav}
\mathcal P(\Psi*_{\bm{x}} f) = \Psi*_{\bm{x}} \mathcal P f,\quad\text{in }\mathbb R^n\times(0,\infty)\;.
\end{align}
Young's inequality ensures that
\[L^p(\mathbb R^n)*L^1(\mathbb R^n)\,\subseteq\,L^p(\mathbb R^n)\,,\]
so that
\[{\mathcal P}(\Psi *_{\bm{x}} f) = \Psi *_{\bm{x}} \mathcal Pf \in L^p(\mathbb R^n)\,,\]
since $\mathcal Pf\in C^\infty_c(\mathbb R^n)\subseteq L^1(\mathbb R^n)$ for every $t>0$.
The rest of the proof is essentially the same as in \cite[Lemma 3.1]{Hal11}:
We start proving the result for one spatial dimension, i.e. $n=1$,
and write here $x$ instead of ${\bm{x}}$ for the spatial variable.
To avoid confusion later on, we write $\bar{\mathcal P}$ instead of $\mathcal P$
for the wave operator in one dimension. Using D'Alembert's formula it follows that
\begin{align*}
\begin{aligned}
&\bar{\mathcal P}\left(\Psi*_x f\right)(x,t)\\
&\,=\,\frac12\left(\int_{\mathbb R}\psi(|y|)f(x-t-y)dy\,+\,\int_{\mathbb R}\psi(|y|)f(x+t-y) dy\right)\\
& \forall\, t > 0\;.
\end{aligned}
\end{align*}
Then, by substituting $y$ by $-y$ in the first integral it follows that for all $t > 0$,
\begin{align}
\label{eq:psi}
\begin{aligned}
&\bar{\mathcal P}\left(\Psi*_x f\right)(x,t)\\
&\,=\,\frac12\left(\int_{\mathbb R} \psi(|y|)\big(f(x-(t-y))\,+\,f(x+(t-y))\big) dy \right)\;.
\end{aligned}
\end{align}
According to D'Alembert's formula,
$$ \frac12 \big(f(x-(t-y))\,+\,f(x+(t-y))\big) = \bar{\mathcal P} f(x,t-y) \text{ for } t-y > 0\;.$$
Due to our choice of extension to negative times in \eqref{eq:PAI_even}, we also have
$$ \frac12 \big(f(x-(t-y))\,+\,f(x+(t-y))\big) = \bar{\mathcal P}_\text{even}f(x,t-y) \text{ for all } t\;.$$
Therefore, it follows from \eqref{eq:psi} that
\begin{align}
\label{eq:conv_thm1D}
\begin{aligned}
\bar{\mathcal P}(\Psi*_x f)(x;t)\,&=\,\int_{\mathbb R} \psi(|y|) \bar{\mathcal P}_\text{even} f(x,t-y) dy\;.
\end{aligned}
\end{align}
Using the definition of the Radon transform in {1D} it follows then that
$$ \bar{\mathcal P}(\Psi*_x f)(x;t) = \int_{\mathbb R} \mathcal{R} \psi(y) \bar{\mathcal P}_\text{even} f(x,t-y) dy\;.$$
For $n>1$ we note that for all $h \in C^2(\mathbb R^n)$,
\begin{align*}
(\mathcal{R} \Delta h)(\theta,s)\,=\,(\partial_s^2 \mathcal{R}h)(\theta,s)\,, \quad \text{ for all } s \geq 0\,,
\end{align*}
see e.g.~\cite[p.3]{Hel11}.
By applying the Radon transform to the wave equation, we can therefore conclude
\begin{align}\label{eq:Wave_Radon}
\left(\mathcal{R}\mathcal{P} f\right)(\theta,s;t)\,=\,\left(\bar{\mathcal P} \mathcal{R}f\right)(\theta,s;t)\;.
\end{align}
Now we use the convolution theorem of the Radon transform (see e.g. \cite{Nat01}), which states that
\begin{align}\label{eq:Conv_Thm_Radon}
\mathcal{R}(f*_{\bm{x}} g)(\theta,s)=(\mathcal{R}f*_s \mathcal{R}g)(\theta,s),
\end{align}
where the convolution on the left-hand is $n$-dimensional, whereas the convolution on the right-hand side is taken in one dimension.
By applying (\ref{eq:Conv_space_wav}), (\ref{eq:Conv_Thm_Radon}) and (\ref{eq:Wave_Radon}), (\ref{eq:Conv_space_wav}) and (\ref{eq:conv_thm1D}), and again (\ref{eq:Wave_Radon}), it follows that:
\begin{align*}
(\mathcal{R} \mathcal{P} )(\Psi*_{\bm{x}} f)\,&=\,\mathcal R\Psi*_s \bar{\mathcal P} \mathcal{R} f\,=\,(\mathcal{R}\Psi)*_t(\bar{\mathcal P}_\text{even} \mathcal{R}f)\\
&=\,(\mathcal{R}\Psi)*_t(\mathcal{R}\mathcal{P_\text{even}} f).
\end{align*}
The term $\mathcal R\Psi=\phi(|\cdot|)$ on the right-hand-side is independent of $\theta$ due to the rotational symmetry of $\Psi$. We therefore can write (see also Remark \ref{rem:extension}):
\begin{align}
\label{eq:Radon_equ}
(\mathcal{R} \mathcal{P} )(\Psi*_{\bm{x}} f)(\theta,s;t)\,&=\,\mathcal{R}\big(\phi(|\cdot|)*_t\mathcal{P} f\big)(\theta,s;t).
\end{align}
Due to our choice of $p$, the Radon transform is injective on $L^p(\mathbb R^n)$ (see \cite{Sol87}). From \eqref{eq:Radon_equ}, we therefore derive \eqref{eq:Haltmeier}.
\qed
\end{proof}
Theorem \ref{thm:conv_thm} is the main ingredient to relate the convolved measurement data with a convolution of $f$. Note, however, that on the right-hand side of statement \eqref{eq:Haltmeier}, the quantity $\mathcal P_\text{even}f$ appears, whereas our measurements give only knowledge of $\phi*\mathcal P f$ as in \eqref{eq:pai1b}. For application of Theorem~\ref{thm:conv_thm} to our case of bandlimited data, we therefore need a relation between $\mathcal P_\text{even}f$ and $\phi*\mathcal P f$, which is provided by the following corollary.
\begin{corollary}\label{cor:conv_thm}
Let $\Psi:\mathbb R^n \to \mathbb R$ be radially symmetric and $\Psi\in L^p(\mathbb R^n)$, for some $1\leq p < n/(n-1)$, and let $\mathcal R\Psi$ be represented as
\[\mathcal{R}\Psi(\theta,s)\,=\,\Phi(\theta,|s|)\,=\,\phi(|s|)\,,\]
for an even function $\phi:\mathbb R\to\mathbb R$ as in \eqref{eq:Kleinvieh}.
Moreover, let our measurements be given by
\[m({\bm{x}};t)=(\phi*_t \mathcal{P} f)({\bm{x}};t)\quad \text{on }\Gamma\times(0,\infty)\;.\]
Then, the function
\begin{equation}
\label{eq:m_even}
m_\text{even}({\bm{x}},t):=(\phi*_t P_\text{even}f)({\bm{x}},t) \quad \text{on }\Gamma\times(0,\infty)\,,
\end{equation} where ${\bm{x}}\in\Gamma$, $t\in(0,\infty)$, can be computed analytically from the causal measurement data $m({\bm{x}},t)$. Furthermore,
\begin{equation}
\label{eq:Haltmeier2}
m_\text{even}({\bm{x}};t)\,=\,(\mathcal{P}(\Psi *_{\bm{x}} f))({\bm{x}}, t) \quad \text{on }\Gamma\times(0,\infty)\;.
\end{equation}
\end{corollary}
\begin{proof}
Let $p$ denote the solution of \eqref{eq:wav_delta}. Since $p({\bm{x}};t)$ is real-valued, it follows that
\[\widehat p({\bm{x}},-\kappa)\,=\,\overline{\widehat{p}({\bm{x}},\kappa)}\,\qquad \forall x \in \mathbb R^n, \kappa \in \mathbb R\,,\]
and therefore
\[\widehat{\mathcal{P}_\text{even}f}({\bm{x}},\kappa)\,=\, 2\,{\rm Re\,}(\hat{p}({\bm{x}},\kappa))\,\qquad \forall {\bm{x}} \in \mathbb R^n, \kappa \in \mathbb R\,.\]
Thus, from \eqref{eq:m_even} it follows that
\begin{equation*}
\begin{aligned}
\widehat {m_{\text{even}}}({\bm{x}},\kappa)\,
&=\,\widehat \phi(\kappa)\widehat {\mathcal{P}_\text{even}f}({\bm{x}},\kappa)\\
&=\,\widehat \phi(\kappa)\,2\,{\rm Re\,}\left(\widehat {\mathcal{P} f}({\bm{x}},\kappa)\right)\\
&=\,2\,{\rm Re\,}\left(\widehat \phi(\kappa)\,\widehat {\mathcal{P} f}({\bm{x}},\kappa)\right)\\
&=\,2\,{\rm Re\,}\left(\widehat m({\bm{x}},\kappa)\right)
\qquad \forall {\bm{x}}\in\Gamma,\,\forall \kappa\in\mathbb R\,,
\end{aligned}
\end{equation*}
where in the third equality we use that $\widehat{\phi}$ is real-valued, since $\phi$ is a real-valued and even function.
The function $m_\text{even}=\phi*_t \mathcal{P}_\text{even} f$ is of the appropriate form
to apply (\ref{eq:Haltmeier}), which allows us to derive (\ref{eq:Haltmeier2}).
\qed\end{proof}
Corollary \ref{cor:conv_thm} gives a simple relation between causal and even data convolved in time.
Using Theorem \ref{thm:conv_thm}, the even data can be related to an initial source
convolved in space by a point-spread function (PSF) that is given in terms of an inverse Radon transform
of the radially symmetric extension of the impulse-response function (IRF), previously denoted by the letter $\phi$.
\section{PAI elastography using texture information}
\label{sec:MotionEst}
The results of Section \ref{sec:Texture} give the theoretical description for the influence of using band-limited data in the photoacoustic reconstruction.
In the following subsection, we describe how to find pairs of filter functions $\phi$ and $\Psi$ in practice.
Moreover, we give an example of a pair of oscillating functions, that we use in what follows to create speckle-like patterns on photoacoustic images.
The rest of the paper will treat the case of two spatial dimensions.
Since the theoretical considerations from Section \ref{sec:Texture} are valid in any spatial dimension,
the application to 3D images works in complete analogy to the two-dimensional case described below.
\subsection{Speckle generation in 2D Photoacoustics}
We assume to measure the bandpass data
\[m=\phi*_t \mathcal P f\,,\]
where we choose $\phi$ as follows: The time-domain equivalent of the bandpass
\begin{align}\label{eq:bandpass_speckle}
\widehat \phi({\bm{x}},\kappa)=\chi_{[\kappa_\text{min},\kappa_\text{max}]}(|\kappa|)
\end{align}
is given by the IRF
\begin{align}\label{eq:IRF_speckle}
\begin{aligned}
\phi(t)=\,\frac{\cos(\kappa_0 t)}{2a}\mathrm{sinc}\left(\frac{t}{4\pi a}\right),
\end{aligned}
\end{align}
where $2a=\kappa_\text{max}-\kappa_\text{min}$ is the bandwidth and $\kappa_0=\kappa_\text{min}+a$ is the center frequency of our window.
Note that with the formulation above, we cover the cases where our detector measures the described signal
as well as the case where we manipulate the data by a post-processing step.
Now assume that we have computed $m_\text{even}$ as described in Corollary \ref{cor:conv_thm}.
The results of Section \ref{sec:Texture} in principle describe the relationship of the filter in time
and a resulting filter in space. But since we actually want to compute this space filter explicitly,
it is convenient to make use of the so-called \emph{Fourier-slice} theorem for the Radon transform \cite{Hel11,Sol87},
that relates the Fourier transform of the Radon transform (in radial direction)
to the Fourier transform of the image (in all spatial dimensions).
In 2D, the Radon transform of a radially symmetric function
is nothing else than the \emph{Abel transform} \cite{Nat01}.
The Fourier-transform of a two dimensional radially symmetric function is the so-called \emph{Hankel transform}.
The 2D version of the Fourier slice theorem for radially symmetric functions
is therefore often written as
\[
\mathcal F\mathcal A = \mathcal H,
\]
also called the \emph{FHA-cycle}.
In order to find the corresponding $\Psi$ to the function $\phi$ in \eqref{eq:bandpass_speckle},
we make use of the tables in literature describing important pairs of the above mentioned Fourier,
Abel and Hankel transforms (see,~e.g.,~\cite{Bra78,Obe90}). In fact, for our given hard bandpass \eqref{eq:bandpass_speckle},
it suffices to compute its (inverse) Hankel transform, so that the corresponding point-spread function $\Psi$ is given by
\begin{align}\label{eq:PSF_speckle}
\begin{aligned}
\Psi({\bm{x}})= \,\frac{2\pi}{|{\bm{x}}|}\left[(\kappa_\text{max}) J_1\left((\kappa_\text{max})|{\bm{x}}|\right)-(\kappa_\text{min})J_1\left((\kappa_\text{min})|{\bm{x}}|\right)\right]\;,
\end{aligned}
\end{align}
where $J_1$ is the first-kind Bessel function of order $1$.
By using an asymptotic estimate of $J_1$ for large arguments,
it is easy to check that $\Psi\in L^p(\mathbb R^2)$ iff $p>4/3$,
which means that $\Psi$ fulfils the integrability requirements demanded in Corollary~\ref{cor:conv_thm}.
This ensures that the result actually applies to the used filter.
Our suggested approach for texture generation then is this: We choose $\kappa_\text{max}$ and $\kappa_\text{min}$ to determine the IRF $\phi$. Then we compute $m_\text{even}$ and solve the photoacoustic inverse problem with data $m_\text{even}$. Theorem~\ref{thm:conv_thm} then ensures that this yields the perturbed reconstruction
\begin{equation}
\label{eq:TextureConv}
f *_{\bm{x}} \Psi
\end{equation}
With the right choice of $\kappa_\text{min}$ and $\kappa_\text{max}$, this is a natural candidate for a textured variant of the photoacoustic contrast in the initial pressure $f$.
In Figure \ref{fig:2}, a point source and its photoacoustic reconstruction from band-limited data (i.e., data convolved with the IRF in (\ref{eq:IRF_speckle})) are shown. The oscillations introduced by the present band-limited photoacoustic reconstruction method introduce additional texture on the image. The use of this texture in estimating the optical flow between two photoacoustic images is investigated in the following sections.
\subsection{Principle of PAI elastography}
In the previous subsection, we introduced a texture method for photoacoustic images. We now will study how motion estimation can be performed and amended by adding texture to photoacoustic images.
We emphasize that the initial pressure $f$ introduced in \eqref{eq:wav_delta} in the photoacoustic forward problem is spatially varying and can either represent the image before (i.e., $f_1$) or after ($f_2$) mechanical deformation as described in Section \ref{sec:Quasi_static_elast}.
The main concept in the proposed method of \emph{photoacoustic elastography} is to perform the following steps in the first step in section \ref{subsec:Elastography}:
\begin{abclist}
\item record a PAI image $f_1$ using the texture-generating method
\item perturb the tissue using a mechanical source
\item record the perturbed configuration $f_2$ using the texture-generation method
\end{abclist}
We will now estimate the displacement $\b u$ as in the second step in section \ref{subsec:Elastography}.
In the following, we evaluate the motion estimation using the Horn-Schunck model \eqref{eq:HS} with or without speckle generation.
\FloatBarrier
\section{Experiments}
\label{sec:Experiments}
There are many different varieties of experiments one can perform. In this section, we present a first selection, using structures which contain homogeneous regions, similar to vascular structures.
\subsection{Simulations}
We simulate photoacoustic forward data using the k-wave toolbox \cite{TreCox10}. For reconstruction, we use a filtered back-projection algorithm. Displacement vector fields have been simulated using the FEM and mesh-generating packages GetDP and Gmsh \cite{DulGeuHenLeg13,GeuRem09}.
\begin{figure*}[h!]
\subfloat[Visualized mask]{
\includegraphics[width=4.5cm]{Exp1Mask.png}
\label{fig:Exp1M}}
\subfloat[Ground truth]{
\includegraphics[width=4.5cm]{Exp1GroundTruthOnImage.png}}
\\
\subfloat[none]{
\includegraphics[width=4.5cm]{Exp1noneOnImage.png}
\label{fig:Exp1n}}
\subfloat[Gauss $\alpha=0.3$ ]{
\includegraphics[width=4.5cm]{Exp1Gauss03OnImage.png}}
\\
\subfloat[Band $\kappa_\text{min}=0.4$]{
\includegraphics[width=4.5cm]{Exp1Band04OnImage.png}}
\subfloat[Band $\kappa_\text{min}=1.8$]{
\includegraphics[width=4.5cm]{Exp1Band18OnImage.png}}
\\
\subfloat[Angular error]{
\includegraphics[width=4.5cm]{Exp1AngularErrors.png}}
\subfloat[Distance error]{
\includegraphics[width=4.5cm]{Exp1DistanceErrors.png}
\label{fig:Exp1DiagrB}}
\caption{(Experiment 1) (c)-(f): Computed vector fields for $\lambda=12.5893$. (g), (h): different error measures for regularization parameters $10^{0.5}\leq\lambda\leq 10^{2.5}$,
full line: original data; dashed line: band-limitation texture $\kappa=0.4$}
\label{fig:Exp1}
\end{figure*}
\begin{figure*}[h!]
\subfloat[Visualized mask]{
\includegraphics[width=4.5cm]{Exp2Mask.png}
\label{fig:Exp2M}}
\subfloat[Ground truth]{
\includegraphics[width=4.5cm]{Exp2GroundTruthOnImage.png}}
\\
\subfloat[none]{
\includegraphics[width=4.5cm]{Exp2noneOnImage.png}
\label{fig:Exp2n}}
\subfloat[Gauss 0.3 ]{
\includegraphics[width=4.5cm]{Exp2Gauss03OnImage.png}}
\\
\subfloat[Band 0.4]{
\includegraphics[width=4.5cm]{Exp2Band04OnImage.png}}
\subfloat[Band 1.8]{
\includegraphics[width=4.5cm]{Exp2Band18OnImage.png}}
\\
\subfloat[Angular error]{
\includegraphics[width=4.5cm]{Exp2AngularErrors.png}}
\subfloat[Distance error]{
\includegraphics[width=4.5cm]{Exp2DistanceErrors.png}
\label{fig:Exp2DiagrB}}
\caption{(Experiment 2) (c)-(f): Computed vector fields for $\lambda=12.5893$. (g), (h): different error measures for regularization parameters $10^{0.5}\leq\lambda\leq 10^{2.5}$,
full line: original data; dashed line: band-limitation texture $\kappa=1.8$}
\label{fig:Exp2}
\end{figure*}
\begin{table*}[b]
\caption{Experiments 1 and 2: Error analysis}
\label{tab:Exp12}
\subfloat[Rigid experiment, $\lambda=12.5893$ (see Figure \ref{fig:Exp1})]{
\begin{tabular}{lcccc}
\hline\noalign{\smallskip}
Texture Mode & AAE & AEEabs & AEErel & Warping \\
\noalign{\smallskip}\hline\noalign{\smallskip}
none & 0.2499 & 0.0445 & 3.3232 & 0.3739 \\
Gauss 0.3 & 0.2499 & 0.0445 & 3.3237 & 0.3797 \\
Band 0.4 & 0.1880 & 0.0101 & 0.7548 & 0.0737 \\
Band 1.8 & 0.2382 & 0.0114 & 0.8546 & 0.0345 \\
\noalign{\smallskip}\hline
\end{tabular}}\hfill
\subfloat[Non-rigid experiment, $\lambda=12.5893$ (see Figure \ref{fig:Exp2})]{
\label{tab:Exp2}
\begin{tabular}{lcccc}
\hline\noalign{\smallskip}
Texture Mode & AAE & AEEabs & AEErel & Warping \\
\noalign{\smallskip}\hline\noalign{\smallskip}
none & 0.1780 & 0.0617 & 2.9980 & 0.4886 \\
Gauss 0.3 & 0.1784 & 0.0617 & 2.9990 & 0.4923 \\
Band 0.4 & 0.0928 & 0.0135 & 0.6595 & 0.1189 \\
Band 1.8 & 0.0447 & 0.0190 & 0.9232 & 0.0494 \\
\noalign{\smallskip}\hline
\end{tabular}}
\end{table*}
\FloatBarrier
\subsection{Material, displacement and parameters}
The synthetic material was chosen to exhibit homogeneous regions surrounded by edges. In each experiment, we evaluated a rigid deformation and a non-rigid deformation.
In Experiments 1 and 2, we use a tree structure designed by Brian Hurshman and licensed under CC BY 3.0\footnote{http://thenounproject.com/term/tree/16622/}.
\subsection{Texture Modes}
The purpose of experiments is to evaluate the influence of the texture pattern introduced by \eqref{eq:TextureConv} to the images, using the basic filtered-backprojection reconstruction of the image.
Letting $\kappa_\text{max}=10$, we choose different values of choose $\kappa_\text{min}$ and record the error with respect to the measures defined below \ref{subsec:Validation}.
On the one hand, these are compared with the motion reconstructions from the unperturbed images. On the other hand, we compare the pictures to the following \emph{Gaussian texture} explained in the sequel.
We will produce the Gaussian texture in the image as follows: We take an image $f(\b x)$ and apply Gaussian noise to the image $f$, producing a reference image
\begin{equation}
\label{eq:TextureApplication}
f_1(\b x)=f(\b x)+\alpha\;r(\b x),
\end{equation}
where $\alpha$ is a constant and $r(\b x)$ is a noise function governed by Gaussian noise. Then we warp the textured image using a certain vector field $\b u$, producing the reference image $f_1$ and the target image
\begin{equation}
\label{eq:Warping}
f_2(\b x)=f_1(\b x+\b u(\b x)).
\end{equation}
At last, we compute the optical flow between $f_1$ and $f_2$ and then determine whether the computed field is approximately the correct motion field.
We emphasize that due to \eqref{eq:TextureApplication} and \eqref{eq:Warping}, the artificial texture pattern thus introduced behaves like a material characteristic which is advected by the vector field $\b u$.
In contrast to that, the in using the texture method~\eqref{eq:TextureConv}, we have recourse to the principle outlined in section \ref{sec:MotionEst}: to arrive at the target image $f_2$
\begin{itemize}
\item (sec. \ref{sec:MotionEst}, step b) the mechanical deformation is applied
\item (sec. \ref{sec:MotionEst}, step c) the texture method is applied.
\end{itemize}
\subsection{Validation}
\label{subsec:Validation}
The field which is computed with the optical flow algorithm should approximately match the correct motion field. In order to study how PAI and textured PAI images behave under mechanical deformations, we adopt the following validation procedure:
\paragraph{Synthectic Data verification}
\begin{itemize}
\item Choose a particular vector field $\b u_0$, as well as a reference image $f_1$
\item Compute the warped image $f_2$ by interpolation, i.e. $f_2= f_1(\b x + \b u_0(\b x))$
\item Compute the optical flow $\b u(\b x)$ from $f_1$ and $f_2$
\item Compare the result $\b u$ against the ground-truth vector field $\b u_0$
\end{itemize}
\paragraph{Error measures}
To compare the computed flows produced to the ground truth field, we use the angular and distance error, and to assess the prediction quality of the flow, we calculate the warping error. To define these error measures, write
\[
\begin{aligned}
\b u_0(\b x) &= r_0(\b x)\; e^{i\varphi_0(\b x)}\\
\b u(\b x) &= r(\b x)\; e^{i\varphi(\b x)}.
\end{aligned}
\]
Then we define the
\begin{itemize}
\item average angular error (AAE) \[
\int_\Omega|\varphi(\b x) - \varphi_0(\b x)|d\b x\]
\item average endpoint error (AEE) \[\int_\Omega \|\b u - \b u_0\| d\b x\]
\item average relative endpoint error (AEErel)
\[\int_\Omega \frac{1}{\|\b u_0\|}\|\b u - \b u_0\| d\b x\]
\item warping error
\[
\int_\Omega \|f_2(\b x) - f_1(\b x + \b u(\b x))\| d\b x.
\]
\end{itemize}
\section{Discussion}
\label{sec:Discussion}
As mentioned in the introduction, elastography often relies on speckle tracking methods, including correlation techniques and optical flow. It is clear that such methods have a problem with homogeneous regions. As for the optical flow, this can be seen from \eqref{eq:OptFlow}, where the data term for homogeneous regions gives no information.
In Experiments 1-4, we used several pieces of synthetic data showing homogeneous regions and investigated the effect of the homogeneity in several regions of the data (see Figs. \ref{fig:Exp1M}-\ref{fig:Exp6M}.
The visualization of the computed motion fields in Figs. \ref{fig:Exp1n}, \ref{fig:Exp2n}, \ref{fig:Exp5n} and \ref{fig:Exp6n} shows aberrations from the respective ground truth fields. Comparing the values for the angular, distance and warping errors in Table \ref{tab:Exp12} to \ref{tab:Exp56} shows these aberrations, if one restricts to the untextured original images.
We then applied the texture generation methods introduced in Section \ref{sec:MotionEst}. The results in section \ref{sec:Experiments} show that addition of texture is able to alleviate this problem of homogeneous regions to a considerable amount. The effect shows up in the different error types.
For the specimens we used, the angular error decreases about 20-30 \% compared to the original error, and in extreme cases the decrease is as high as 75 \% (as seen from Table \ref{tab:Exp2}). As seen from Figs. \ref{fig:Exp1DiagrB}-\ref{fig:Exp6DiagrB}, where the errors were plotted as a function of the regularization parameters, the distance error and in the warping error reach their minimum in the textured variant at lower regularization values than the original data. In this context, we note that the angular error is a monotonically increasing function of the regularization parameters in the cases we investigated. In some cases (as seen from Fig. \ref{fig:Exp1DiagrB} and Fig. \ref{fig:Exp5DiagrB}), the textured versions give also a lower distance error for the optimal regularization value; in other cases, with the motion estimation we used, the distance error is about the same magnitude as in the original versions.
The optimum frequency windows for the texture generating method also seem to differ for the rigid and the non-rigid deformations we used. Whereas for the rigid deformations, the window with $\kappa_\text{min}=0.4$ gave better results, the non-rigid deformations gave better results with $\kappa_\text{min}=1.8$.
The effect of adding texture seems to come from a filling-in-effect in the optical flow equation \eqref{eq:OptFlow}. Although the regularization term is responsible for such an interpolation usually, here this filling-in-effect originates from the data term; the function $\Psi$ seems to propgate the information from within the objects out across the edges and boundaries. This seems also to alleviate the aperture problem in optical flow, as the new texture creates also new gradients around edges. This may account for the lessening of the angular error.
Overall, the results point at the phenomenon that an effect which has deteriorating the image quality in one contrast (here the photoacoustic contrast) can have an advantageous effect on another contrast (here the mechanical contrast, which is inherent in the displacement~$\b u$).
\section{Conclusion}
\label{sec:Conclusion}
We studied the topic of texture generation in photoacoustics, and applied bandwidth filter techniques for generating such texture in the reconstructed images. This kind of texture was mathematically characterized. Then we tested an application of the PAI texture for elastography purposes. It turned out that the texture generation technique has the potential to fill in otherwise untextured regions. The displacements can be better measured then, making photoacoustic elastography viable.
\begin{figure*}[h!]
\subfloat[Visualized mask]{
\includegraphics[width=4.5cm]{Exp5Mask.png}
\label{fig:Exp5M}}
\subfloat[Ground truth]{
\includegraphics[width=4.5cm]{Exp5GroundTruthOnImage.png}}
\\
\subfloat[none]{
\includegraphics[width=4.5cm]{Exp5noneOnImage.png}
\label{fig:Exp5n}}
\subfloat[Gauss $\alpha=0.3$ ]{
\includegraphics[width=4.5cm]{Exp5Gauss03OnImage.png}}
\\
\subfloat[Band $\kappa_\text{min}=0.4$]{
\includegraphics[width=4.5cm]{Exp5Band04OnImage.png}}
\subfloat[Band $\kappa_\text{min}=1.8$]{
\includegraphics[width=4.5cm]{Exp5Band18OnImage.png}}
\\
\subfloat[Angular error]{
\includegraphics[width=4.5cm]{Exp5AngularErrors.png}}
\subfloat[Distance error]{
\includegraphics[width=4.5cm]{Exp5DistanceErrors.png}
\label{fig:Exp5DiagrB}}
\caption{(Experiment 3) (c)-(f): Computed vector fields for $\lambda=11.2202$. (g), (h): different error measures for regularization parameters $10^{0.5}\leq\lambda\leq 10^{2.5}$,
full line: original data; dashed line: band-limitation texture $\kappa=0.4$}
\label{fig:Exp5}
\end{figure*}
\begin{figure*}[h!]
\subfloat[Visualized mask]{
\includegraphics[width=4.5cm]{Exp6Mask.png}
\label{fig:Exp6M}}
\subfloat[Ground truth]{
\includegraphics[width=4.5cm]{Exp6GroundTruthOnImage.png}}
\\
\subfloat[none]{
\includegraphics[width=4.5cm]{Exp6noneOnImage.png}
\label{fig:Exp6n}}
\subfloat[Gauss 0.3 ]{
\includegraphics[width=4.5cm]{Exp6Gauss03OnImage.png}}
\\
\subfloat[Band 0.4]{
\includegraphics[width=4.5cm]{Exp6Band04OnImage.png}}
\subfloat[Band 1.8]{
\includegraphics[width=4.5cm]{Exp6Band18OnImage.png}}
\\
\subfloat[Angular error]{
\includegraphics[width=4.5cm]{Exp6AngularErrors.png}}
\subfloat[Distance error]{
\includegraphics[width=4.5cm]{Exp6DistanceErrors.png}
\label{fig:Exp6DiagrB}}
\caption{(Experiment 4) (c)-(f): Computed vector fields for $\lambda=11.2202$. (g), (h): different error measures for regularization parameters $10^{0.5}\leq\lambda\leq 10^{2.5}$,
full line: original data; dashed line: band-limitation texture $\kappa=1.8$}
\label{fig:Exp6}
\end{figure*}
\begin{table*}[b]
\caption{Experiments 3 and 4: Error analysis}
\label{tab:Exp56}
\subfloat[Rigid experiment, $\lambda=11.2202$ (see Figure \ref{fig:Exp5})]{
\begin{tabular}{lcccc}
\hline\noalign{\smallskip}
Texture Mode & AAE & AEEabs & AEErel & Warping \\
\noalign{\smallskip}\hline\noalign{\smallskip}
none & 0.1089 & 0.0635 & 4.7460 & 0.4382 \\
Gauss 0.3 & 0.1090 & 0.0635 & 4.7465 & 0.4430 \\
Band 0.4 & 0.0840 & 0.0211 & 1.5807 & 0.1753 \\
Band 1.8 & 0.1344 & 0.0094 & 0.7024 & 0.0533 \\
\noalign{\smallskip}\hline
\end{tabular}}\hfill
\subfloat[Non-rigid experiment, $\lambda=11.2202$ (see Figure \ref{fig:Exp6})]{
\begin{tabular}{lcccc}
\hline\noalign{\smallskip}
Texture Mode & AAE & AEEabs & AEErel & Warping \\
\noalign{\smallskip}\hline\noalign{\smallskip}
none & 0.2962 & 0.0112 & 1.3676 & 0.2053 \\
Gauss 0.3 & 0.2956 & 0.0112 & 1.3680 & 0.2127 \\
Band 0.4 & 0.3251 & 0.0026 & 0.3236 & 0.0191 \\
Band 1.8 & 0.2115 & 0.0043 & 0.5228 & 0.0216 \\
\noalign{\smallskip}\hline
\end{tabular}}
\end{table*}
\FloatBarrier
\paragraph{Acknowledgements.}
We thank Joyce McLaughlin, Paul Beard and Ben Cox for helpful discussions and acknowledge support from the Austrian Science Fund (FWF) in projects S10505-N20 and P26687-N25.
\def$'${$'$}
\providecommand{\noopsort}[1]{}\def\ocirc#1{\ifmmode\setbox0=\hbox{$#1$}\dimen0=\ht0
\advance\dimen0 by1pt\rlap{\hbox to\wd0{\hss\raise\dimen0
\hbox{\hskip.2em$\scriptscriptstyle\circ$}\hss}}#1\else {\accent"17 #1}\fi}
\def$'${$'$}
|
\section{Introduction}
The observed correlations between the mass of a supermassive black
hole (SMBH) and the properties of its host galaxy bulge
\citep[e.g.][]{mag98,gul09} imply that SMBH and bulge growth may be
linked. To maintain the SMBH-bulge relations over a galaxy's long
evolutionary history, rapid SMBH accretion in active galactic nucleus
(AGN) phases probably coincide with periods of high star formation
rate in the host galaxy. Indeed, rapidly accreting AGNs are observed
to be predominantly located in the rapidly star-forming galaxies, from
the local universe \citep{kau03b,tru13b} to higher redshifts
\citep{mul12,har12,chen13,ros13a,ros13b}. But the detailed physical
processes behind the coupled growth of SMBHs and galaxies remain
mysterious.
A key difficulty for any AGN/galaxy coevolution model is in
efficiently funneling gas down to the SMBH sphere of influence.
Gas-rich major mergers of galaxies are an effective means to
accomplish this, simultaneously (or near-simultaneously) igniting both
a starburst and a luminous AGN \citep{san88,dim05,hop06}. However
many observations indicate that AGNs do not prefer merger remnant
hosts \citep{gro05,gab09, cis11, koc12}, and major mergers are likely
to fuel only nearby AGNs \citep{koss10,ell11} or the rare population
of very luminous quasars \citep{tru11c,tre12}.
\begin{figure*}[t]
\begin{center}
\epsscale{1.1}
{\plotone{clumpiness.eps}}
\end{center}
\figcaption{The diagram used for visual classification of clumpiness.
The classification grid includes 9 template galaxies, each with
images in the ACS F606W (rest-frame UV) and WFC3 F160W (rest-frame
optical). For each galaxy in GOODS-S, classifiers chose the
template galaxy which best matched in both bands (with a focus on
the F606W, where clumps are most visible). We sum both axes and
average over the 5 classifiers to assign each galaxy a ``total
clumpiness'' $C$ from 0 (top left) to 4 (bottom right), motivated by
theoretical simulations which demonstrate that patchiness might
simply be clumpiness reprocessed by dust. Clumpy galaxies have $C
\geq 2$, smooth galaxies have $C \leq 1$, and intermediate galaxies
have $1<C<2$.
\label{fig:classification}}
\end{figure*}
Violent disk instabilities might provide another solution to the gas
inflow problem \citep{bou11}. These models are supported by the
observation that massive galaxies at $z \sim 2$ frequently have most
of their star formation in massive ($10^8$-$10^9$ $M_\odot$,
$100-500$~pc) clumps within irregular galaxies
\citep{cow95,elm04,rav06,guo12}, unlike the much smoother morphologies
of similar-mass star-forming galaxies in the local universe. The
kinematics of these clumps indicate that they are not accreted as
minor mergers, but instead form as in-situ gravitational instabilities
\citep{sha08,bou08,for09,gen11,man14}. The high gas fractions
observed in $z \sim 2$ galaxies \citep{tac10,dad10}, likely accreted
as cold gas along filaments of the cosmic web
\citep{ker05,dek06,dek09a}, would naturally result in highly turbulent
star-forming clumps \citep{her89,shlos93,dek09b}.
Violent disk instabilities might lead to clumps rapidly falling into
the center of a galaxy and efficiently fueling a luminous AGN
\citep{bou11}. This remains a point of debate, however, as most
simulations lack the resolution and detailed physics to follow the gas
all the way into a galaxy's center. Observations indicate that
star-forming clumps at $z \sim 2$ typically make up only a small
fraction of their galaxy's stellar mass, suggesting that they may be
too short-lived to migrate all the way into a galaxy center
\citep{wuyts12}. Indeed, simulations show that clumps might be
destroyed by feedback \citep{hop12,hop13} or exhausted by star
formation \citep{for14} before reaching the nuclear AGN. On the other
hand, other simulations find that clumps are sufficiently long-lived
to form central bulges and fuel AGNs even with feedback and mass loss
\citep{bou14}, and the gas between clumps, rather than the clumps
themselves, may be more important for large-scale inflows anyway
\citep[e.g.][]{dek13}. In support of AGN fueling from violent disk
instabilities, \citet{bou12} found direct observational evidence for a
higher AGN fraction in a small sample (14) of $z \sim 0.7$ clumpy
galaxies compared to smooth (non-clumpy) extended disks matched in
stellar mass and redshift.
Here we extend the search for a connection between clumps and AGNs to
$1.3<z<2.4$, using data from the Cosmic Assembly Near-Infrared Legacy
Survey \citep[CANDELS][]{candels,candels2}. This redshift range is
particularly interesting because it represents the peak of both cosmic
star formation and the AGN luminosity function
\citep[e.g.][]{cosmicsfr,aird10}. Cold gas inflows and resultant
violent disk instabilities are also most prominent in simulations at
$z \sim 2$ \citep{ker05,dek09b}. And in contrast to the local
universe, AGNs at $z>1$ may be unique in growing before a host galaxy
bulge develops, with evidence from both evolution in BH-bulge
relations \citep{peng06,jahnke09,ben11,cis11} and from stacked
detection of AGN signatures in low-mass disk galaxies
\citep{tru11b,xue12}. The gravitational torques provided by violent
disk instabilities provide a potential way to fuel an AGN in a disk
galaxy even while the bulge is still under construction.
Section 2 describes the combination of Hubble Space Telescope ({\it
HST}) Wide Field Camera 3 (WFC3) near-infrared (IR) imaging and
spectroscopy used to construct and study our samples of $z \sim 2$
galaxies with clumpy, smooth, and intermediate morphologies. We
identify AGNs using X-ray emission, emission-line diagnostics, and
spatially resolved emission-line ratios: as Section 3 describes, this
includes a broad range of AGN properties and avoids contamination by
shock-dominated galaxies. We put these selection techniques together
in Section 4, and demonstrate that clumpy galaxies are not more likely
to host AGNs than smoother galaxies. Section 5 concludes with a
discussion of what these results mean for AGN/galaxy coevolution at $z
\sim 2$. Throughout the paper we use a standard $\Lambda$CDM
cosmology with $h_0=0.7$.
\begin{deluxetable*}{ccrrrrrr}
\tablecolumns{8}
\tablecaption{Galaxy Properties\label{tbl:catalog}}
\tablehead{
\colhead{ID} &
\colhead{Morphology} &
\colhead{RA} &
\colhead{Dec} &
\colhead{$z$} &
\colhead{$f(\hbox{{\rm H}$\beta$})$} &
\colhead{$f(\hbox{[{\rm O}\kern 0.1em{\sc iii}]})$} &
\colhead{$\log(M_*)$} \\
\colhead{-} &
\colhead{-} &
\colhead{(deg)} &
\colhead{(J2000)} &
\colhead{-} &
\colhead{$10^{-18}$~erg~s$^{-1}$~cm$^{-2}$} &
\colhead{$10^{-18}$~erg~s$^{-1}$~cm$^{-2}$} &
\colhead{$(\log(M_{\odot}))$} }
\startdata
7502 & clumpy & 53.07221603 & -27.84033012 & 1.61 & $99.95\pm28.56$ & $83.47\pm24.51$ & 9.63 \\
7897 & clumpy & 53.14831924 & -27.83686638 & 1.90 & $23.49\pm12.22$ & $27.09\pm8.42$ & 9.82 \\
7907 & clumpy & 53.15573502 & -27.83712769 & 1.39 & $18.95\pm13.73$ & $169.79\pm21.69$ & 9.67 \\
7930 & clumpy & 53.15451050 & -27.83648491 & 2.03 & $14.20\pm10.43$ & $43.64\pm10.26$ & 9.41 \\
8206\tablenotemark{a} & clumpy & 53.14358902 & -27.83471107 & 1.99 & $10.82\pm5.73$ & $39.97\pm10.54$ & 9.85 \\
5023 & smooth & 53.10656738 & -27.86481476 & 1.90 & $<3.31$ & $18.62\pm5.92$ & 9.74 \\
5728 & smooth & 53.08922577 & -27.85734558 & 2.03 & $16.80\pm7.95$ & $79.73\pm11.31$ & 10.37 \\
6278\tablenotemark{b} & smooth & 53.06018448 & -27.85304642 & 1.54 & $<2.73$ & $240.61\pm21.63$ & 10.79 \\
6678 & smooth & 53.07507324 & -27.84802055 & 1.73 & $37.84\pm7.38$ & $21.91\pm6.77$ & 10.14 \\
7806 & smooth & 53.16970062 & -27.83803558 & 1.94 & $8.18\pm7.71$ & $38.97\pm11.10$ & 10.37 \\
5552 & intermed & 53.06779480 & -27.85925293 & 1.55 & $15.67\pm14.78$ & $73.40\pm16.57$ & 9.81 \\
6895 & intermed & 53.09495926 & -27.84582710 & 1.56 & $<4.50$ & $49.29\pm8.75$ & 9.44 \\
7952 & intermed & 53.18961334 & -27.83626175 & 2.09 & $39.79\pm13.92$ & $52.87\pm9.57$ & 9.89 \\
8124 & intermed & 53.10293198 & -27.83484268 & 1.38 & $55.60\pm21.62$ & $102.88\pm15.12$ & 9.18 \\
8307 & intermed & 53.07443237 & -27.83325577 & 1.54 & $<6.95$ & $55.32\pm10.17$ & 9.26 \enddata
\tablecomments{Weak \hbox{{\rm H}$\beta$}\ lines detected below a 1$\sigma$ threshold
are treated as upper limits (using the 1$\sigma$ error as the
limit). The full catalog of 44 clumpy galaxies, 41 smooth
galaxies, and 35 intermediate galaxies appears as a
machine-readable table in the electronic version.}
\tablenotetext{a}{X-ray detected, but in the soft band only and
consistent with emission from a star-forming galaxy \citep{xue11}.}
\tablenotetext{b}{X-ray detected and classified as an AGN by
\citet{xue11}.}
\end{deluxetable*}
\section{Observational Data}
We select a sample of 44 clumpy galaxies from the Great Observatories
Origins Deep Survey South \citep[GOODS-S][]{goods} region of CANDELS.
For comparison, we also construct mass-matched samples of 41 smooth
(non-clumpy) and 35 intermediate galaxies. All galaxies have $H<24$
(to ensure reliable classification of clumpiness) and have $\hbox{[{\rm O}\kern 0.1em{\sc iii}]}$
detected at the 3$\sigma$ level (for reliable AGN line ratio
diagnostics) in the redshift range $1.3<z<2.4$. The effective limits
on star formation rate (SFR) and stellar mass ($M_*$) caused by these
flux limits are discussed in Section 2.3. Each morphology category
has a median redshift of $z=1.85$. Redshifts and emission line
measurements come from {\it HST}/WFC3 grism spectroscopy taken by the
3D-HST survey \citep{3dhst}.
The observational data are described below, with particular attention
to the methods for clumpiness classification and AGN identification.
The derived data for the clumpy, smooth, and intermediate galaxies are
presented in Table \ref{tbl:catalog}.
\subsection{Visual Morphologies: Clumpy, Smooth, and Intermediate}
\begin{figure}[t]
\epsscale{1.15}
{\plotone{clumpcompare.eps}}
\figcaption{Comparison between the automated number of clumps from
\citet{guo14} and the visual clumpiness used in this work. There is
a loose correlation between the two measures of clumpiness, matching
the $\sim$75\% agreement between automated and visual methods found
by \citet{guo14}.
\label{fig:clumpcompare}}
\end{figure}
Our samples of clumpy, smooth, and intermediate galaxies come from the
4-epoch CANDELS GOODS-S visual classification catalog \citep{kar14}.
Each galaxy has high-resolution {\it HST} imaging in the rest-frame UV
from ACS F606W and F850LP and in the rest-frame optical from WFC3
F125W and F160W \citep{candels,candels2}. Galaxies were classified on
the clumpiness/patchiness grid shown in Figure
\ref{fig:classification}. Here ``clumpiness'' is a measure of the
number of compact knots, while ``patchiness'' refers to the more
diffuse irregularities within galaxies.
We use visual classification because automated selection of clumps and
patchiness remains a difficult computational problem
\citep[e.g.][]{guo12}. Inspectors were instructed to focus on the
bluest passband for clumpiness classification, as clumps are typically
most evident in in the rest-UV \citep{guo12}. All galaxies in our
sample (from the 4-epoch catalog of \citealt{kar14}) were inspected by
at least five classifiers, and we combine the different
classifications into a single averaged result for each galaxy.
\begin{figure*}[t]
\epsscale{1.15}
{\plotone{images.eps}}
\figcaption{Images of four clumpy, four smooth, and four
intermediate-morphology galaxies, chosen to be representative in
redshift and stellar mass (stellar masses given in each panel are in
units of solar mass). Each thumbnail is 10$\arcsec$ on a side and
is a color-composite using the {\it HST}/ACS $i$ and WFC3 $JH$ band
observations.
\label{fig:images}}
\end{figure*}
We use both clumpiness and patchiness as indicators of violent disk
instabilities within a galaxy. This is motivated by the presence of
both compact/round clumps and diffuse/elongated structures in
simulations of $z \sim 2$ galaxies undergoing violent disk
instabilities \citep{cev12,man14}. In addition, if star-forming
clumps are enshrouded by dust, their emission would likely be
reprocessed in a patchy morphology. Observations show that galaxies
appearing clumpy in the UV are merely patchy morphologies in the
near-infrared: adding dust to such a system would cause it to appear
patchy in the UV rather than clumpy. We also note that \citet{guo14}
find better agreement between automated clump-finding and visual
classification when patchiness is included in the visual clumpiness.
Therefore we combine both axes in Figure \ref{fig:classification} for
the identification of clumpy galaxies.
The classifications on each axis are summed to give each galaxy a
clumpiness+patchiness value ranging from 0 to 4. Here 0 represents
the top left (no clumpiness, no patchiness) and 4 represents the lower
right (3+ clumps and maximally patchy). The individual classification
results are averaged together to assign ``total clumpiness'' parameter
$C$, which we use to define the galaxies most and least likely to be
dominated by violent disk instabilities. We define $C \geq 2$
galaxies as ``clumpy,'' $C<1$ galaxies as ``smooth,'' and $1 \leq C<2$
galaxies as ``intermediate.'' These divisions are chosen such that
each category has roughly the same number of galaxies, for ease of
comparing AGN fractions across morphologies.
We compare the visual classification $C$ with the automated number of
clumps from \citep{guo14} in Figure \ref{fig:clumpcompare}. There is
a loose correlation between the two methods: the visually-identified
smooth galaxies tend to have the fewest automated number of clumps,
while the visually clumpy galaxies typically have the most automated
number of clumps. There is some scatter due to the imperfection of
both methods. \citet{guo14} include a more detailed comparison of
automated clump-finding with the same visual classifications used
here, finding $\sim$75\% agreement between the two methods (see their
Appendix).
Figure \ref{fig:images} shows {\it HST} $iJH$ (rest UV-optical) color
composite images for several representative clumpy, smooth, and
intermediate galaxies in our sample. Images of all galaxies are
additionally shown in the Appendix (Figures \ref{fig:clumpyimages},
\ref{fig:smoothimages}, and \ref{fig:intimages}).
\subsection{HST/WFC3 G141 Slitless Grism}
The GOODS-S region of CANDELS has near-complete spectroscopic coverage
in the near-IR from publicly available 2-orbit {\it HST}/WFC3 G141
grism observations taken by the 3D-HST survey \citep{3dhst}. For
redshifts and line measurements of individual galaxies, the data were
reduced using the {\tt aXe} software \citep[][available at {\tt
http://axe.stsci.edu/axe/}]{kum09} to produce 2D and 1D wavelength-
and flux-calibrated spectra at $1.1<\lambda<1.7\mu$m. We used the
{\tt specpro} IDL software\footnote{The specpro package is available
at {\tt http://specpro.caltech.edu}} \citep{specpro} to visually
inspect and determine cross-correlation redshifts for the samples of
clumpy, smooth, and intermediate galaxies. Spectra with significant
($>$10\%) contamination (from neighboring objects) in the \hbox{{\rm H}$\beta$}+\hbox{[{\rm O}\kern 0.1em{\sc iii}]}\
region are rejected from each sample: this occurs for about $\sim$15\%
of galaxies.
\begin{figure}[t]
\epsscale{1.1}
{\plotone{linefits.eps}}
\figcaption{Example line fits to the \hbox{{\rm H}$\beta$}+\hbox{[{\rm O}\kern 0.1em{\sc iii}]}\ regions. The blended
lines are simultaneously fit by three Gaussians, with the \hbox{{\rm H}$\beta$}\ flux
measured directly from the \hbox{{\rm H}$\beta$}\ Gaussian fit, and the
\hbox{[{\rm O}\kern 0.1em{\sc iii}]}$\lambda$5007 line computed as $3/4$ of the two \hbox{[{\rm O}\kern 0.1em{\sc iii}]}\
Gaussians. The top row shows spectra of two clumpy galaxies, the
middle row is smooth galaxies, and the bottom row is intermediate
galaxies. Both 19908 and 16985 have weak \hbox{{\rm H}$\beta$}\ emission detected
below 1$\sigma$ significance, and for these galaxies the 1$\sigma$
errors in \hbox{{\rm H}$\beta$}\ are treated as upper limits.
\label{fig:linefits}}
\end{figure}
Line fluxes and ratios of \hbox{{\rm H}$\beta$}\ and \hbox{[{\rm O}\kern 0.1em{\sc iii}]}$\lambda$5007\AA\ were also
measured from the {\tt aXe}-reduced 3D-HST spectra. The low
resolution of the WFC3 slitless grism ($R \simeq 130$ and
46.5\AA/pixel, worse for extended sources) means that the \hbox{{\rm H}$\beta$}\ and
\hbox{[{\rm O}\kern 0.1em{\sc iii}]}$\lambda$4959,5007\AA\ lines are somewhat blended. We
simultaneously fit all three lines with Gaussians, constraining each
line to be fit within (rest-frame) 50\AA\ ($\sim$2-3 pixels) of the
line center. We do not constrain the width of each line because the
spatial broadening of the slitless grism frequently causes unusual
profiles which can differ for each line. The \hbox{{\rm H}$\beta$}\ line flux is
measured directly from the Gaussian fit, while the \hbox{[{\rm O}\kern 0.1em{\sc iii}]}$\lambda$5007
flux is measured as $3/4$ of the total flux from both blended \hbox{[{\rm O}\kern 0.1em{\sc iii}]}\
Gaussians \citep{sto00}. If we instead fix the two $\hbox{[{\rm O}\kern 0.1em{\sc iii}]}$ Gaussians
to have total fluxes with a 1:3 ratio, the measured line fluxes do not
significantly change. Examples of the line fits are shown in Figure
\ref{fig:linefits}.
Uncertainties in the line ratios are computed by re-fitting continua
and Gaussians on 10000 realizations of the resampled data. Our sample
includes only galaxies with \hbox{[{\rm O}\kern 0.1em{\sc iii}]}$\lambda$5007\AA\ fluxes measured at
the 3$\sigma$ level. We do not set any requirement on the $\hbox{{\rm H}$\beta$}$ line
flux measurements, and $\hbox{{\rm H}$\beta$}$ line fluxes less than the 1$\sigma$ error
are treated as upper limits (using the 1$\sigma$ error as the limit).
In general, the unique asymmetric shape of the blended $\hbox{[{\rm O}\kern 0.1em{\sc iii}]}$ lines
enables secure redshifts even when $\hbox{{\rm H}$\beta$}$ is poorly detected.
To study spatially resolved line ratios, we used separate reductions
of the WFC3 G141 data from the 3D-HST pipeline, as described in
\citet{3dhst,bra13}. The 3D-HST reduction method produces superior
spatially resolved spectra because it interlaces rather than drizzles:
this mitigates the correlated noise associated with drizzling in the
final high-resolution ($0\farcs06$/pixel) combined 2D spectra.
\begin{figure}[t]
\epsscale{1.15}
{\plotone{sedfits.eps}}
\figcaption{Best-fit models to the SEDs of the same galaxies shown in
Figure \ref{fig:linefits}. Stellar masses are computed from these
best-fit \citet{bru03} models. The SED of the galaxy 16985 (which
is an X-ray AGN) has red near-IR colors suggestive of an AGN
\citep[e.g.][]{don12}: it is the only galaxy in the sample which
might have significant AGN contribution to the SED.
\label{fig:sedfits}}
\end{figure}
\subsection{Stellar Masses and Star Formation Rates}
We calculate stellar masses and star formation rates for the clumpy,
smooth, and intermediate galaxies using the extensive UV / optical /
IR photometry in GOODS-S, with 18 bands including 8 with
high-resolution {\it HST} imaging \citep{guo13}. First, the
spectroscopic redshift from the {\it HST}/WFC3 grism is used to shift
the observed photometry to the rest-frame. We then fit the rest-frame
spectral energy distribution (SED) with \citet{bru03} models that
include a \citet{cha03} initial mass function, exponentially declining
star-formation histories, and a \citet{cal01} extinction law. Stellar
mass is given by the best-fit model. Example SED fits (for the same
galaxies shown in Figure \ref{fig:linefits} are shown in Figure
\ref{fig:sedfits}.
\begin{figure}[t]
\epsscale{1.15}
{\plotone{sfrmass.eps}}
\figcaption{{\it Top:} Histograms showing stellar mass distribution
for each sample. The smooth and intermediate galaxies are chosen to
have the same mass distribution as the clumpy galaxies. {\it
Bottom:} Star formation rate vs. stellar mass for the clumpy,
smooth, and intermediate galaxies studied in this work (colored
points). The contours show the larger population of galaxies with
$1.3<z_{\rm phot}<2.4$ in the same GOODS-S region. While the SFR is
estimated from the \citet{wuyts11} UV+IR method, it is correlated
with the $\hbox{[{\rm O}\kern 0.1em{\sc iii}]}$ emission line flux, and the observed SFR$\gtrsim
10$~M$_\odot$~yr$^{-1}$ limit is a consequence of the 3$\sigma$
\hbox{[{\rm O}\kern 0.1em{\sc iii}]}\ detection limit. Meanwhile the $H<24$ selection causes the
$M_* \gtrsim 10^9M_\odot$ limit.
\label{fig:sfrmass}}
\end{figure}
There is some evidence that $z \sim 2$ galaxies are more likely to
have constant star-formation histories rather than the exponentially
declining models used here. In practice, our best-fit SEDs generally
have $\tau$ values which are very long (parameterizing the star
formation history as $SFR \sim e^{-t/\tau}$). We also tested the
effects of forcing a constant star formation history, and found the
masses to change by only $\lesssim$0.1~dex in all cases.
Meanwhile star formation rates (SFRs) are calculated following the
method of \citet{wuyts11} when mid- and far-IR data are available, and
SED-fitting following the method of \citet{bar13} otherwise. IR
emission from an intrinsically luminous AGN might contaminate the IR
with a hot dust bump \citep[effectively observed as red IRAC colors,
e.g.][]{don12}, influencing both the stellar mass and SFR estimates.
This occurs for one of our objects: the clumpy galaxy and X-ray AGN
16985, shown in Figure \ref{fig:sedfits}. The remainder of the sample
have well-fit galaxy SEDs without evidence for significant AGN
contribution. This is unsurprising, as only intrinsically luminous
($log(L_X) \gtrsim 44$) AGNs tend to have hot dust IR emission which
dominates over the galaxy emission \citep{tru11a,don12}.
\begin{figure*}[t]
\epsscale{1.15}
{\plotone{bpt.eps}}
\figcaption{The mass-excitation (MEx) diagnostic \citep{jun11} of
$\hbox{[{\rm O}\kern 0.1em{\sc iii}]}/\hbox{{\rm H}$\beta$}$ line ratio versus stellar mass. Panels from left to
right show clumpy galaxies, smooth galaxies, and
intermediate-morphology galaxies. Cyan symbols indicate galaxies
detected in X-rays: filled symbols are X-ray AGN, while open cyan
symbols are consistent with X-ray galaxies (i.e., with X-rays
powered by non-AGN processes). For comparison, each panel also
shows a $z \sim 0$ galaxy sample from the SDSS. The fraction of MEx
AGN is slightly higher for the clumpy galaxies (29\% AGN) than for
the samples of smooth (27\% AGN) and intermediate (20\% AGN)
galaxies. However we show in Sections 3.2 and 3.3 that this is
likely due to extended high-ionization phenomena (e.g., clumpy star
formation) rather than nuclear AGN.
\label{fig:bpt}}
\end{figure*}
Figure \ref{fig:sfrmass} (top panel) shows the mass distributions for
each of the clumpy, smooth, and intermediate galaxy samples. The
three samples are selected to be matched in stellar mass: starting
with the clumpy galaxy sample, we constructed the smooth and
intermediate galaxy samples to have a similar mass distribution. All
three samples have median and mean $M_* = 10^{9.8}M_\odot$.
Our galaxies are plotted in $SFR-M_*$ space in Figure
\ref{fig:sfrmass} (bottom panel). For comparison we also show the SFR
and $M_*$ of the larger population of GOODS-S galaxies with
photometric redshifts calculated by \citet{dah13} in the same
$1.3<z<2.4$ redshift range. Our $z \sim 1.85$ galaxies are limited to
$M_* \gtrsim 10^9 M_\odot$ by the $H<24$ selection criterion. The
requirement for \hbox{[{\rm O}\kern 0.1em{\sc iii}]}\ detection at the 3$\sigma$ level imposes a flux
limit of $f(\hbox{[{\rm O}\kern 0.1em{\sc iii}]}) \gtrsim 5 \times 10^{-17}$~erg~s$^{-1}$~cm$^{-2}$
in the 2-orbit WFC3 grism data. Although we estimate SFR from the SED
\citep[following][]{wuyts11} rather than from emission lines, this
$\hbox{[{\rm O}\kern 0.1em{\sc iii}]}$ line flux translates roughly to the observed SFR limit of
SFR$\gtrsim 10$~M$_\odot$~yr$^{-1}$ (assuming $f(\hbox{{\rm H}$\alpha$}) \sim f(\hbox{[{\rm O}\kern 0.1em{\sc iii}]})$
and using the \citealp{ken98} relation). For the lower-mass half of
the sample ($M_*<10^{9.8}M_\odot$), this SFR limit restricts the
galaxies to be starbursting and above the star-forming ``main
sequence'' \citep[see also][]{noe07,dad07}. On the other hand the
higher-mass half of the sample ($M_*>10^{9.8}M_\odot$) generally lies
on the main sequence for star-forming galaxies at $1.3<z<2.4$. The
only notable exception is a few high-mass smooth galaxies below the
main sequence which might be quenching.
\subsection{X-ray Data}
The CANDELS GOODS-S field also contains 4~Ms of {\it Chandra} X-ray
data \citep{xue11}. The 4~Ms depth corresponds to an on-axis flux
limit of $3.2 \times 10^{-17}$~erg~$s^{-1}$~cm$^{-2}$, which
corresponds to a luminosity limit of $L_X>10^{41.9}$~erg~s$^{-1}$ at
our sample's median redshift of $z=1.85$. We use the classifications
of \citet{xue11} to separate X-ray galaxies (with X-ray detections
consistent with the $L_X-SFR$ relation, \citealt{leh10,min14}) from
the more luminous and hard-spectrum AGNs. The X-ray data for
undetected sources in each morphology category are also stacked, as
discussed in Section 3.3.
\section{AGNs in Clumpy Galaxies?}
We compare the AGN fraction of clumpy galaxies with the AGN fraction
among the mass-matched smooth and intermediate galaxies. AGN
detection is accomplished in three different ways: with standard line
ratio selection via the ``mass-excitation'' method \citep{jun11}, with
spatially resolved line ratios, and with X-rays (both for individual
sources and stacks of each morphological category).
\subsection{Mass Excitation Diagnostic for AGNs}
It has long been known that the high-ionization emission of AGNs
results in a different emission line signature than observed in
typical \hbox{{\rm H}\kern 0.1em{\sc ii}}\ regions associated with star formation (SF): in
particular, AGN narrow line regions tend to exhibit higher ratios of
collisionally excited ``forbidden'' lines to hydrogen recombination
lines \citep{seyf43,ost65}. The classic \citet[][BPT]{bpt81} and
\citet[][VO87]{vo87} AGN/SF diagnostics use the ratios of
$f(\hbox{[{\rm O}\kern 0.1em{\sc iii}]}\lambda5007)/f(\hbox{{\rm H}$\beta$})$ vs. $f(\hbox{[{\rm N}\kern 0.1em{\sc ii}]}\lambda6584)/f(\hbox{{\rm H}$\alpha$})$ or
$f(\hbox{[{\rm S}\kern 0.1em{\sc ii}]}\lambda6718+6731)/f(\hbox{{\rm H}$\alpha$})$: the small wavelength separation of
each line pair means that these ratios are effectively insensitive to
reddening.
The low resolution of the WFC3 grism means that the \hbox{[{\rm N}\kern 0.1em{\sc ii}]}\ and \hbox{{\rm H}$\alpha$}\
lines are not resolved from one another, and so the standard BPT or
VO87 diagnostics cannot be used on our data. Instead we use the
``mass-excitation'' (MEx) method \citep{jun11}, which uses the
$\hbox{[{\rm O}\kern 0.1em{\sc iii}]}/\hbox{{\rm H}$\beta$}$ ratio with the stellar mass $M_*$ to separate AGN and SF
galaxies. Functionally, the MEx method uses the correlation between
mass and metallicity \citep[e.g.][]{tre04} to separate low-metallicity
galaxies from AGN, both of which exhibit similarly high $\hbox{[{\rm O}\kern 0.1em{\sc iii}]}/\hbox{{\rm H}$\beta$}$
ratios. Compared to X-ray or infrared surveys, line ratio selection
is often more sensitive to obscured and moderately accreting AGN, but
it also less reliable \citep[e.g.][]{jun14}. There is also some
debate whether line ratio diagnostics remain applicable in $z>1$
galaxies at all, since higher redshift galaxies may have different gas
properties \citep{liu08,bri08,kew13}, changing their line ratios in
the absence of AGN \citep[but for counterarguments,
see][]{wri10,tru11b,tru13a}. \citet{jun14} also demonstrate that the
different selection effects in high- and low-redshift samples have
important effects on their observed MEx distributions. For this
reason we avoid comparing the $z \sim 1.5$ galaxy samples with local
objects. Instead we simply compare clumpy, smooth, and intermediate
galaxies all within the same redshift range.
Figure \ref{fig:bpt} shows the MEx diagram for the clumpy, smooth, and
intermediate galaxies. The solid line in each panel shows the
empirical line from \citet{jun14} dividing AGN and SF galaxies,
adjusted for the redshift ($z \sim 1.85$) and line flux limit
($f(\hbox{[{\rm O}\kern 0.1em{\sc iii}]}) \gtrsim 5 \times 10^{-17}$~erg~s$^{-1}$~cm$^{-2}$):
\begin{equation}
y = 0.375/(m-10.4)+1.14~{\rm if}~m \le 9.88
\end{equation}
\begin{equation}
y = 290.2 - 76.34m + 6.69m^2 - 0.1955m^3~{\rm otherwise.}
\end{equation}
Here $y=\log(\hbox{[{\rm O}\kern 0.1em{\sc iii}]}/\hbox{{\rm H}$\beta$})$ and $m = \log(M_*/M_\odot)-0.51$. Assuming
that all galaxies above this line are AGNs, clumpy galaxies seem to
have the highest AGN fraction (12/44, 27\%), followed by smooth
galaxies (11/41, 27\%) and intermediate galaxies (7/35, 20\%).
However the simple binary classification of counting objects above and
below the AGN/SF line is not entirely appropriate, given that most
galaxies have emission lines with composite contribution from both
SMBH accretion and \hbox{{\rm H}\kern 0.1em{\sc ii}}\ regions. The MEx diagram is much more suited
to a probabilistic classification approach, as introduced by
\citet{jun11}. The MEx probabilities, updated for use at $z>1$ by
\citet{jun14} and assuming an error of 0.2~dex in $M_*$, indicate that
$39_{-6}^{+8}\%$ of clumpy, $36_{-6}^{+8}\%$ of smooth, and
$35_{-7}^{+9}\%$ of intermediate galaxies are AGN-dominated. Given
the uncertainty of the MEx diagram at $z>1$
\citep[e.g.][]{kew13,new14,jun14}, these probabilities do not
necessarily represent absolute AGN fractions, but they remain useful
for comparing relative AGN fractions among the three morphology
classes. The three AGN fractions are all consistent with one another,
and the clumpy galaxies have only marginally ($<$1$\sigma$) more AGNs
than smoother galaxies. In the next two subsections we show that the
higher line ratios in clumpy galaxies are likely an effect of high
ionization in extended regions (due to shocks or dense star-forming
clumps) rather than nuclear AGN.
\subsection{Spatially Resolved Line Ratios}
The traditional BPT and MEx methods use line ratios integrated over
the entire galaxy, and thus may be diluted by star formation and/or
affected by non-nuclear ionization that has nothing to do with AGN
activity. In particular, the violent disk instabilities of clumpy
galaxies may lead to high-velocity shocks or dense $\hbox{{\rm H}\kern 0.1em{\sc ii}}$ regions
which can mimic AGN-like line ratios \citep[e.g.][]{rich11,new14}.
For this reason we use the spatial resolution of the WFC3 grism
spectroscopy to go beyond integrated line ratios and investigate line
ratio gradients, following \citet{wri10,tru11b}. Spatially resolved
line ratios can reveal if AGNs selected by integrated line ratios are
the product of nuclear AGNs or extended phenomena (like shocks or
high-ionization star formation).
\begin{figure}[t]
\epsscale{1.15}
{\plotone{stackimages.eps}}
\figcaption{Stacked two-dimensional {\it HST}/WFC3 grism spectra for
clumpy, smooth, and intermediate-morphology galaxies. The top three
spectra are stacks of all galaxies in each morphology class, while
the bottom three are stacks of MEx AGNs only. The emission line
sizes in both the dispersion and cross-dispersion directions are a
result of galaxy size rather than velocity due to the low resolution
($R \sim 300$) of the slitless grism. The stacked data provide
sufficient signal-to-noise to separately extract nuclear and
extended one-dimensional spectra and line ratios.
\label{fig:stackimages}}
\end{figure}
The two-dimensional (2D) spectra of individual sources generally lack
the signal-to-noise for well-measured spatially resolved line ratios,
so we stack spectra of galaxies in each of the clumpy, smooth, and
intermediate galaxy samples. Two stacks are constructed for each
morphology class: one with all galaxies, and another using only
galaxies classified as AGNs on the MEx diagram in Section 3.1.
Centers of galaxies are defined by the SExtractor \citep{sextractor}
coordinates in the F160W detection image, translated to a spectral
trace in the dispersed G141 2D spectrum using well-calibrated
polynomials from {\tt aXe} which are typically accurate to
$\lesssim$0.2~pixels. The image centroid and resultant spectral trace
may not coincide with the brightest (continuum or emission line) flux
of a galaxy, particularly for clumpy galaxies: we investigate
centering on the brightest flux rather than image centroids in Section
4.1. Each galaxy spectrum is also normalized by its total $\hbox{[{\rm O}\kern 0.1em{\sc iii}]}$
flux so that the stack is equal-weighted and not dominated by the most
$\hbox{[{\rm O}\kern 0.1em{\sc iii}]}$-luminous AGNs. (This normalization means that the stacked
line ratios are slightly lower than the mean of the individual
ratios.) The stacked 2D spectra of all galaxies in each morphology
class are shown in Figure \ref{fig:stackimages}. It is immediately
evident that the clumpy galaxies are typically larger than the
intermediate or smooth galaxies: for a galaxy to be classified as
clumpy, it must be large enough for the clumps to be resolved in the
{\it HST} data. Still, the smooth and intermediate galaxies are not
so small as to be unresolved, and they have enough signal beyond the
central three pixels to have well-measured extended (non-nuclear)
spectra. We return to a discussion of the different sizes of the
clumpy and smooth galaxies in Section 4.2.
We separately extract nuclear and extended one-dimensional (1D)
spectra from the stacked data, shown in Figure \ref{fig:stackspec}.
Here ``nuclear'' is defined as the central 3 pixels, corresponding to
$0\farcs18$ across and a $\sim$1~kpc radius at $1.3<z<2.4$. The
``extended'' spectra are extracted from pixels 3-6 on either side of
the trace, translating to an annulus extending radially from
$\sim$2-4~kpc. It is possible that some emission from the AGN narrow
line region extends into the extended aperture \citep{hai13,vdl13},
and the $\sim$1~kpc nuclear region is also likely to include some
galaxy starlight. Although our regions will not perfectly disentangle
AGN and galaxy light, the nuclear region preferentially includes AGN
light and the extended region includes more galaxy starlight. Thus
the spatially resolved line ratios are likely to be significantly more
sensitive than the integrated line ratios.
\begin{figure}[t]
\epsscale{1.15}
{\plotone{stackspec.eps}}
\figcaption{Nuclear and extended 1D spectra from the stacked {\it
HST}/WFC3 grism data for clumpy, smooth, and
intermediate-morphology galaxies. The top three panels are the
stacked spectra of all galaxies, while the bottom three are for
galaxies classified as MEx AGNs. The $\hbox{{\rm H}$\beta$}$ and $\hbox{[{\rm O}\kern 0.1em{\sc iii}]}$ line centers
are shown by the dashed lines, and the orange and cyan lines give
the best-fit Gaussians to the nuclear and extended 1D spectra.
\label{fig:stackspec}}
\end{figure}
\begin{figure*}[ht]
\epsscale{1.15}
{\plotone{bpt_inout.eps}}
\figcaption{The nuclear and extended line ratios of stacked clumpy,
smooth, and intermediate morphology galaxies. These line ratios are
placed in the MEx diagram using the median $M_*$ of each stack, with
a small artificial offset between the nuclear and extended points.
The integrated values for individual galaxies, seen in Figure
\ref{fig:bpt}, are shown in gray. Similar nuclear and extended line
ratios are seen in all three panels, suggesting that none of the
morphology classes have dominant nuclear AGN populations.
\label{fig:bpt_inout}}
\end{figure*}
\begin{figure*}[ht]
\epsscale{1.15}
{\plotone{bpt_inoutagn.eps}}
\figcaption{The nuclear and extended line ratios of the galaxies
identified as AGN on the MEx diagram, stacked in each of the clumpy,
smooth, and intermediate morphology classes. As in Figure
\ref{fig:bpt_inout}, we place the line ratios in the MEx diagram
using the median $M_*$ of each stack (with a small artificial offset
between the nuclear and extended points). Smooth and intermediate
MEx-classified AGNs have significantly higher nuclear ratios,
suggesting that they are indeed likely to host AGN in their centers.
However clumpy MEx AGNs have only marginally higher line ratios in
their nuclear region, and much of their AGN-like $\hbox{[{\rm O}\kern 0.1em{\sc iii}]}/\hbox{{\rm H}$\beta$}$ ratios
comes from extended phenomena like shocks or high-ionization star
formation. The spatially resolved line ratios of stacked MEx AGN in
each morphology class suggest that clumpy galaxies may actually be
less likely to host nuclear AGNs than smoother galaxies.
\label{fig:bpt_inoutagn}}
\end{figure*}
\begin{deluxetable}{lrr}
\tablecolumns{3}
\tablecaption{Stacked Spectra Nuclear and Extended $\log(\hbox{[{\rm O}\kern 0.1em{\sc iii}]}/\hbox{{\rm H}$\beta$})$
\label{tbl:stackratios}}
\tablehead{
\colhead{Stack} &
\colhead{Nuclear} &
\colhead{Extended} }
\startdata
All Clumpy & $3.5\pm0.8$ & $4.5\pm1.2$ \\
All Smooth & $3.2\pm0.4$ & $2.6\pm1.2$ \\
All Intermediate & $3.9\pm1.1$ & $3.3\pm1.3$ \\
MEx AGN Clumpy & $7.5\pm1.5$ & $7.1\pm2.0$ \\
MEx AGN Smooth & $8.3\pm1.4$ & $5.8\pm1.3$ \\
MEx AGN Intermediate & $11.8\pm2.1$ & $9.8\pm2.3$ \\
\hbox{[{\rm O}\kern 0.1em{\sc iii}]}-centered Clumpy AGNs & $5.5\pm1.2$ & $10.5\pm1.8$
\enddata
\end{deluxetable}
The \hbox{[{\rm O}\kern 0.1em{\sc iii}]}\ and \hbox{{\rm H}$\beta$}\ emission lines are measured from each nuclear and
extended spectrum following the same method as in Section 2.2: three
Gaussians were simultaneously fit to the continuum-subtracted spectra,
with the \hbox{[{\rm O}\kern 0.1em{\sc iii}]}$\lambda$5007 flux given as 3/4 of the total
\hbox{[{\rm O}\kern 0.1em{\sc iii}]}$\lambda$4959+5007 emission. Figure \ref{fig:bpt_inout} shows
the nuclear and extended $\hbox{[{\rm O}\kern 0.1em{\sc iii}]}/\hbox{{\rm H}$\beta$}$ ratios for the stacks of all
clumpy, smooth, and intermediate galaxies in the MEx diagram (using
the median stellar mass of each sample with small offsets between the
points). Figure \ref{fig:bpt_inoutagn} shows the nuclear and extended
line ratios using the stacked spectra of MEx AGN only. The nuclear
and extended line ratio measurements for each stack are also presented
in Table \ref{tbl:stackratios}.
In the stacks of all galaxies in Figure \ref{fig:bpt_inout}, the
nuclear and extended spectra are nearly indistinguishable. This
indicates that there is not a dominant AGN population in any of the
morphology classes. Nuclear AGNs begin to marginally emerge in Figure
\ref{fig:bpt_inoutagn} in the stacked spectra of MEx AGNs among smooth
galaxies. This implies that AGN classification using the updated
\citet{jun14} MEx method is effective for smooth galaxies at $z \sim
2$, with their stacked spectra indicating nuclear AGN. The clumpy
galaxies classified as MEx AGN, however, have essentially no
difference in nuclear and extended $\hbox{[{\rm O}\kern 0.1em{\sc iii}]}/\hbox{{\rm H}$\beta$}$ ratios. The high
$\hbox{[{\rm O}\kern 0.1em{\sc iii}]}/\hbox{{\rm H}$\beta$}$ ratios observed in the integrated spectra of clumpy
galaxies are probably caused by extended phenomena like shocks or
high-ionization star formation, both of which are likely in dense star
forming clumps \citep{kew13,new14,rich14}. Spatial line ratios
indicate that clumpy galaxies may actually be \textit{less} likely to
host nuclear AGNs than smooth or intermediate galaxies. We further
investigate the possibility of off-nuclear AGNs in Section 4.1.
\subsection{X-ray AGN}
X-rays tend to be the most efficient and least contaminated indicator
of AGN activity, and so we also use the X-ray data to quantify the AGN
fraction among clumpy, smooth, and intermediate morphology galaxies.
All galaxies are matched to the \citet{xue11} 4~Ms {\it Chandra}
catalog. In total, 5/44 clumpy, 4/41 smooth, and 2/35 intermediate
galaxies have X-ray counterparts. X-ray detection is not sufficient
for AGN classification, however, as luminous starburst
(SFR$\sim$300~M$_\odot$/yr) galaxies can have enough X-ray binaries to
exceed an integrated luminosity of $L_X \sim 10^{42}$
\citep{leh10,min14}. \citet{xue11} classify the GOODS-S X-ray sources
into galaxies consistent with pure star formation or systems likely to
be AGN on the basis of their spectral shape, X-ray to optical flux
ratio, and X-ray luminosity. These classifications reveal that 2/48
clumpy, 3/41 smooth, and 2/35 intermediate galaxies are likely to be
X-ray AGNs. The detection fractions and full-band luminosities of the
X-ray AGNs are broadly consistent across the morphology categories,
ranging from $10^{42}-10^{44}$~erg~s$^{-1}$. X-ray data for the
detected galaxies, as well as for the stacked non-detections
(discussed below), are presented in Table \ref{tbl:xray}.
\begin{deluxetable*}{crrrrc}
\tablecolumns{6}
\tablecaption{X-ray Data: Detections and Stacked Non-detections\label{tbl:xray}}
\tablehead{
\colhead{Galaxy ID} &
\colhead{Redshift} &
\colhead{$f$(0.5-2~keV)} &
\colhead{$f$(2-8~keV)} &
\colhead{$f$(0.5-8~keV)} &
\colhead{Classification} \\
\colhead{} &
\colhead{} &
\colhead{(erg~s$^{-1}$~cm$^{-2}$)} &
\colhead{(erg~s$^{-1}$~cm$^{-2}$)} &
\colhead{(erg~s$^{-1}$~cm$^{-2}$)} &
\colhead{} }
\startdata
8206 & 1.99 & $2.24 \times 10^{-17}$ & $<1.02 \times 10^{-16}$ & $<7.13 \times 10^{-17}$ & Galaxy \\
8409 & 1.82 & $2.26 \times 10^{-17}$ & $<1.21 \times 10^{-16}$ & $ 5.18 \times 10^{-17}$ & AGN \\
9258 & 2.02 & $1.64 \times 10^{-17}$ & $<1.05 \times 10^{-16}$ & $<6.58 \times 10^{-17}$ & Galaxy \\
9474 & 2.12 & $2.59 \times 10^{-17}$ & $<1.39 \times 10^{-16}$ & $ 7.37 \times 10^{-17}$ & Galaxy \\
10493 & 1.83 & $8.00 \times 10^{-17}$ & $2.11 \times 10^{-16}$ & $ 2.89 \times 10^{-16}$ & AGN \\
6278 & 1.54 & $4.70 \times 10^{-17}$ & $6.58 \times 10^{-16}$ & $ 6.75 \times 10^{-16}$ & AGN \\
9493 & 1.47 & $2.72 \times 10^{-17}$ & $<1.14 \times 10^{-16}$ & $ 6.64 \times 10^{-17}$ & Galaxy \\
10650 & 1.31 & $1.28 \times 10^{-16}$ & $1.33 \times 10^{-16}$ & $ 2.56 \times 10^{-16}$ & AGN \\
16985 & 1.73 & $2.99 \times 10^{-16}$ & $3.95 \times 10^{-15}$ & $ 4.03 \times 10^{-15}$ & AGN \\
9245 & 2.07 & $1.59 \times 10^{-17}$ & $9.04 \times 10^{-17}$ & $ 8.01 \times 10^{-17}$ & AGN \\
18315 & 2.32 & $1.57 \times 10^{-16}$ & $6.75 \times 10^{-16}$ & $ 8.18 \times 10^{-16}$ & AGN \\
\hline
Stacked Clumpy & -- & $4.3\pm0.9 \times 10^{-18}$ & $<1.1 \times 10^{-17}$ & $9.2\pm2.8 \times 10^{-18}$ \\
Stacked Smooth & -- & $4.0\pm0.9 \times 10^{-18}$ & $<1.6 \times 10^{-17}$ & $8.8\pm2.9 \times 10^{-18}$ \\
Stacked Intermediate & -- & $<4.5 \times 10^{-18}$ & $<9.3 \times
10^{-18}$ & $<9.1 \times 10^{-18}$ \\
\hline
Stacked Clumpy MEx AGN & -- & $5.2\pm1.7 \times 10^{-18}$ & $<2.4 \times 10^{-17}$ & $13.9\pm5.1 \times 10^{-18}$ \\
Stacked Smooth MEx AGN & -- & $<5.2 \times 10^{-18}$ & $<2.6 \times 10^{-17}$ & $<17.7 \times 10^{-18}$ \\
Stacked Intermediate MEx AGN & -- & $<4.6 \times 10^{-18}$ & $<3.3 \times 10^{-17}$ & $<15.7 \times 10^{-18}$
\enddata
\tablecomments{X-ray fluxes and AGN/galaxy classifications from
\citet{xue11}. X-ray stacking following the method in \citet{luo11}.}
\end{deluxetable*}
Figure \ref{fig:sfrxray} presents the $L_X-SFR$ relationship for X-ray
detected sources, for both hard-band (2--8~keV) and soft-band
(0.5--2~keV) luminosities. Also shown is the $L_X-SFR$ relation for
star-forming galaxies without AGN, $\log(L_X)=[39.6\pm0.4]+\log(SFR)$
\citep{min14}, with $L_X$ the full-band luminosity in erg/s and $SFR$
in M$_\odot$/yr. This relation is translated to the soft and hard
bands by assuming a spectral index of $\Gamma=1.8$: the resultant
relation is consistent with the hard-band $L_X-SFR/M_*$ relation of
\citet{leh10}. A few X-ray sources which are undetected in the hard
band lie near the soft-band $L_X-SFR$ relation and might not host AGN,
but most have sufficient X-ray luminosities to be robustly considered
as having X-ray emission powered by an AGN.
\begin{figure}[t]
\epsscale{1.15}
{\plotone{sfrxray.eps}}
\figcaption{Hard-band (top panel) and soft-band (bottom panel) X-ray
luminosities versus star formation rates, for both individually
detected galaxies (solid points) and stacked non-detections (open
points) among the three morphology types. The dashed lines indicate
the $L_X-SFR$ relation for star-forming galaxies without AGN from
\citet{min14}, and the gray shaded region indicates the relation's
1$\sigma$ errors. Most individually detected sources have
sufficiently high (hard) X-ray detections to be AGN, although the
stacked data suggest only weak AGN emission in the X-ray
non-detections. In both detected sources and stacked data, there
are no significant differences in X-ray AGN likelihood data between
clumpy, smooth, and intermediate morphology galaxies.
\label{fig:sfrxray}}
\end{figure}
We also performed X-ray stacking of the three samples in the full
band, soft band (0.5--2 keV), and hard band using the 4~Ms CDF-S data
\citep{xue11}. X-ray detected galaxies and galaxies near any detected
X-ray source (within twice the soft-band 90\% encircled-energy
aperture radius of the X-ray source) were excluded from the stacking.
The final clumpy, smooth, and intermediate samples used in the
stacking contain 40, 38, and 33 galaxies, respectively. We adopted
the same stacking procedure as described in Section 3.1 of
\citet{luo11}. Briefly, we extracted total (source plus background)
counts for each galaxy within a 3\arcsec-diameter circular aperture
centered on its optical position. The corresponding background counts
within this aperture were determined with a Monte Carlo approach which
randomly (avoiding known X-ray sources) places 1000 apertures within a
1\arcmin-radius circle of the optical position to measure the mean
background. The total counts ($S$) and background counts ($B$) for
the stacked sample were derived by summing the counts for individual
sources. The net source counts are then given by $(S-B)$, and the S/N
is calculated as $(S-B)/\sqrt{B}$. A stacked signal is considered
undetected if it has a binomial no-source probability of $P_{\rm
b}>0.01$ (equivalent to ${\rm S/N}\lesssim2.6\sigma$): for these
sources we derive 90\% confidence upper limits using the Bayesian
method of \citet{kra91}. Aperture corrections were applied when
converting the source count rates to fluxes and luminosities.
The stacked luminosities are shown with the median SFR of each subset
in Figure \ref{fig:sfrxray}. The small number of objects in each
morphology subset means that none are well-detected in the hard band,
and only the clumpy and smooth stacks are detected in the soft-band.
Due to these poor hard band detections, the hardness ratios are
unconstrained in all three stacks. The stacked X-ray luminosities
(and upper limits) for all three samples lie only marginally above the
$L_X-SFR$ line for SF galaxies, indicating a weak or sub-dominant AGN
contribution in the X-ray undetected galaxies. This is consistent
with the location of most galaxies in the star-forming region of the
MEx diagram (Figure \ref{fig:bpt}), and with the lack of nuclear AGN
in the resolved line ratios among all galaxies (Figure
\ref{fig:bpt_inout}). We also stack the MEx AGNs in each morphology
class, but the numbers of objects are too small to make any meaningful
conclusions.
Clumpy, smooth, and intermediate galaxies all have similar detected
X-ray AGN fractions ($\sim$5\%) and stacked X-ray luminosities for
undetected sources ($\log(L_X) \sim 41.4$). In agreement with the
spatially resolved line ratios, the X-ray data show essentially no
differences for clumpy, smooth, and intermediate morphology galaxies
at $z \sim 2$.
\section{Discussion}
Both spatially resolved line ratios and X-ray data (detections and
stacked non-detections) indicate that there is little or no difference
in AGN fraction between clumpy galaxies and those with smoother
morphologies. While integrated line ratios in the MEx method suggest
a slightly higher AGN fraction among clumpy galaxies, the lack of a
nuclear AGN signature suggests that this is due to extended phenomena
like dense star-forming clumps rather than black hole growth.
Galaxies at $1.3<z<2.4$ in each morphology category have similar AGN
fractions of $\sim$5--30\%, depending on whether X-ray data or
spatially resolved line ratios are used for AGN selection.
We discuss several implications of these results below.
\subsection{Off-Nuclear AGN in Clumpy Galaxies?}
In Section 3.2 we argued that the high $\hbox{[{\rm O}\kern 0.1em{\sc iii}]}/\hbox{{\rm H}$\beta$}$ ratio in the
extended regions of clumpy galaxies indicates shocks or
high-ionization \hbox{{\rm H}\kern 0.1em{\sc ii}}\ regions rather than AGN activity \citep[see
also][]{rich11,new14,rich14}. However, it is instead possible that
some clumpy galaxies host off-nuclear AGN
\citep[e.g.][]{schaw11,schaw14}. We test for the presence of
off-nuclear AGN by creating a 2D stack of clumpy galaxies classified
as MEx-AGNs, centering on the $\hbox{[{\rm O}\kern 0.1em{\sc iii}]}$ emission lines rather than the
continuum trace. The brightest knot of $\hbox{[{\rm O}\kern 0.1em{\sc iii}]}$ emission would be most
likely to correspond to any off-nuclear AGN\footnote{Our assumption
that the brightest $\hbox{[{\rm O}\kern 0.1em{\sc iii}]}$ knot corresponds to an AGN is inaccurate
if there is large-scale dust asymmetrically obscuring an AGN
narrow-line region.}. The ``nuclear'' region of the
$\hbox{[{\rm O}\kern 0.1em{\sc iii}]}$-centered stack corresponds to the brightest $\hbox{[{\rm O}\kern 0.1em{\sc iii}]}$ knots
rather than the galaxy center, and should therefore be where
off-nuclear AGNs are most likely to reside.
\begin{figure}[t]
\epsscale{1.15}
{\plotone{bpt_linectr.eps}}
\figcaption{Nuclear and extended line ratios of clumpy galaxies
classified as MEx AGNs. The left panel uses a stacked spectrum from
galaxies centered on their continuum trace, while the right panel
tests for the presence of off-nuclear AGNs with a stacked spectrum
constructed by centering each galaxy on the brightest knot of
$\hbox{[{\rm O}\kern 0.1em{\sc iii}]}$ emission. There is no evidence for a large population of
off-nuclear AGNs in clumpy galaxies, which would have manifested as
a higher ``nuclear'' line ratio in the $\hbox{[{\rm O}\kern 0.1em{\sc iii}]}$-centered stack.
\label{fig:bpt_linectr}}
\end{figure}
Figure \ref{fig:bpt_linectr} compares the nuclear and extended
$\hbox{[{\rm O}\kern 0.1em{\sc iii}]}/\hbox{{\rm H}$\beta$}$ ratios for the continuum-centered stack with the
$\hbox{[{\rm O}\kern 0.1em{\sc iii}]}$-centered stack of clumpy galaxies. The right panel suggests
that the brightest $\hbox{[{\rm O}\kern 0.1em{\sc iii}]}$ knots are somewhat {\textit less} likely to
host AGNs than the galaxy centers. The lack of widespread off-nuclear
AGNs is also supported by the X-ray data, which indicate that clumpy
galaxies do not host a greater number of X-ray AGNs (whether nuclear
or off-nuclear) than smoother mass-matched galaxies. We conclude that
there is unlikely to be a large population of off-nuclear AGNs in
clumpy galaxies misclassified by the spatially resolved line ratio
method.
\subsection{Comparison to Lower Redshift}
At first glance, our $z \sim 1.85$ result is in apparent contrast to
the preference for AGN in clumpy galaxies at $z \sim 0.7$ observed by
\citet{bou12}. Our clumpy galaxy sample has very similar images and
morphologies to theirs, as seen in a comparison of Figure
\ref{fig:images} with Figure 4 of \citet{bou12}. And although we
study a much larger sample (three times more clumpy galaxies),
\citet{bou12} performed a comprehensive bootstrap analysis which
indicates that sample size is very unlikely to lead to the difference.
Our studies differ greatly, however, in the comparison samples of
smooth galaxies.
\citet{bou12} required that both their clumpy and smooth galaxy
samples be extended disks. Meanwhile we impose no requirements on
morphology beyond the clumpiness classification, and Figure
\ref{fig:sersic} demonstrates that our $z \sim 2$ clumpy and smooth
galaxies are quite different in size and \citet{sersic} index. These
quantities were measured using GALFIT \citep{galfit} and presented by
\citet{vdw12}. Clumpy galaxies are much more likely to be extended
disks, with a median semi-major axis of $0\farcs46$ and 41/44 galaxies
having disk-dominated S\'{e}rsic indices ($n<2.5$). The smooth
galaxies are still mostly disk-dominated, as 30/41 have $n<2.5$, but
they are significantly smaller (with a median semi-major axis of
$0\farcs20$) and have higher S\'{e}rsic indices.
\begin{figure}[t]
\epsscale{1.15}
{\plotone{sersic.eps}}
\figcaption{Sersic index versus size (semi-major axis) for the clumpy,
smooth, and intermediate galaxies. The smooth and intermediate
galaxies are generally smaller than the clumpy galaxies, and tend to
have more prominent bulges (although $>$70\% are still
disk-dominated with $n<2.5$).
\label{fig:sersic}}
\end{figure}
It is interesting that so few smooth extended disks are observed in
CANDELS and 3D-HST. The data are sensitive to the $z \sim 2$
star-forming main sequence at $M_*/M_\odot>10^{9.8}$ (Figure
\ref{fig:sfrmass}), and so if smooth extended disks exist they must be
low-mass and/or weakly star-forming. \citet{kas12} and \citet{ee14}
argue that ordered disks are rare in $z>1$ galaxies of any mass,
suggesting that the lack of smooth extended disks at $z \sim 2$ is a
consequence of galaxy evolution rather than selection effects.
Regardless, the lack of smooth extended disks at $z \sim 2$ means we
cannot perform the same comparison between clumpy and smooth extended
disks as \citet{bou12} at $z \sim 0.7$. Instead we conclude that
clumpy extended galaxies have the same AGN fraction as more compact
and higher-Sersic smooth galaxies matched in stellar mass.
\subsection{Fueling AGNs at $z \sim 2$}
Our results demonstrate that clumpy galaxies at $z \sim 2$ are no more
likely to host AGNs than are smooth galaxies of the same stellar mass
and similar star formation rate. What does this indicate about AGN
fueling modes at $z \sim 2$?
\citet{dek14} argue that compact star-forming galaxies (also called
blue nuggets) might actually be constructed by a history of violent
disk instabilities. Our smooth galaxies, which tend to be more
compact and bulgy than the clumpy extended disks, have sizes and
Sersic indices consistent with this category \citep[see
also][]{cev14}. The observed similarity in AGN fraction between $z
\sim 2$ clumpy and smooth galaxies would then indicate that nuclear
inflow onto the AGN must persist until after the galaxy becomes stable
and the star-forming clumps are no longer visible. In this scenario
the violent disk instabilities are responsible for building a gas
reservoir in the galaxy's center. Since the actual nuclear inflow
onto the AGN persists into the smooth compact galaxy phase, it must
not depend on violent instabilities in the larger disk. So while the
gas reservoir is deposited by violent disk instabilities, the AGN
fueling within $\sim$1~kpc must instead be driven by secular
processes.
On the other hand, the disk-dominated nature of the smooth galaxies
might imply that they are unlikely to have experienced violent mixing
in the past from extreme disk instabilities (or major mergers). It is
also not clear that the smooth galaxies are older than the clumpy
galaxies, given their similar position in the SFR-M* plane (excepting
the most massive galaxies: see Figure \ref{fig:sfrmass}). This might
imply that even the large-scale fueling occurs via secular processes,
perhaps due to stochastic turbulence from the high gas fractions which
are common in $z \sim 2$ galaxies \citep{tac10}. Indeed,
\citet{hop14} predict that this turbulence-driven stochastic fueling
is the dominant mode of SMBH growth for $z \sim 2$ galaxies with
$M_{BH}<10^{7}M_\odot$ (or $M_*<10^{10}M_\odot$): the same mass range
occupied by much of our sample. Assuming that the smooth and clumpy
morphologies in our sample trace different fueling modes, then our
data best support the scenario of \citet{bel13}, where AGN growth
depends not on fueling mode, but on gas fraction.
\section{Summary}
We use spatially resolved line ratios and X-ray data to demonstrate
that $z \sim 1.85$ clumpy galaxies are no more likely to host AGNs
than are smoother galaxies at the same redshift. While integrated
line ratios indicate a higher AGN fraction in clumpy galaxies, the
spatially resolved emission lines show that this is likely due to
extended phenomena (like shocks or high-ionization $\hbox{{\rm H}\kern 0.1em{\sc ii}}$ regions)
rather than nuclear (or off-nuclear) AGNs. While the smooth galaxies
have the same masses and nearly the same star formation rates, they
are somewhat smaller and not quite as disk-dominated as the clumpy
galaxies. Whatever process drives AGN fueling in the smooth galaxies
is just as efficient as are violent disk instabilities in clumpy
galaxies at $z \sim 2$.
\acknowledgements
We thank Frederic Bournaud for valuable discussion which significantly
improved this work. Support was provided by NASA through Hubble
Fellowship grant \#51330 awarded by the Space Telescope Science
Institute, which is operated by the Association of Universities for
Research in Astronomy, Inc., for NASA under contract NAS 5-26555. JRT
and the authors from UCSC also acknowledge support form NASA HST
grants GO-12060.10-A and AR-12822.03, Chandra grant G08-9129A, and NSF
grant AST-0808133. SJ acknowledges financial support from the
E.C. through an ERC grant StG-257720. This work made use of the
Rainbow Cosmological Surveys Database, which is operated by the
Universidad Complutense de Madrid (UCM), partnered with the University
of California Observatories at Santa Cruz (UCO/Lick, UCSC).
|
\section{Introduction: single emitters close to interfaces}
A single atom or molecule in the vicinity of a planar layered medium represents a paradigmatic system in near field physics \cite{Novotny1997}. On-chip antenna design e.g., as well as surface-enhanced spectroscopy \cite{Moskovits1985}, require a quantitative analysis of the dipole radiation and energy distribution for such geometries.
This study is then essential in the quest for efficient light-matter interfaces, on which optical sensing and photon-based communication protocols strongly rely \cite{Homola1999,OBrien2009}.
Experiments performed by Drexhage in 1970 \cite{Drexhage1970} already showed a clear modification of emitter decay rate and radiation pattern as a function of the distance to a reflecting interface. Classically, this is understood as a feedback effect of the reflected complex field on the molecule itself. Quantum mechanically, vacuum fluctuations depend on the problem boundary conditions and affect the density of states. The radiative decay rate results modified, as well as the efficiency of all those processes which can be described by the exchange of virtual photons. This is the case e.g. of a fundamental process in nature, i.e.\ photosynthesis, relying on the energy transfer between different chromophores.
Molecules close to surfaces are also ideal probes of local effects and fields, as they are affected by the local environment at the nm scale, i.e.\ on the order of their physical size \cite{Moerner1999, Michaelis1999, Veerman1999}. Furthermore, similarly to what happens in cavity quantum electrodynamics, strong interactions of quantum emitters (QE) or absorbers with guided modes at the interface between different media can lead to cooperative effects \cite{Martin-Cano2010} and entanglement phenomena \cite{tudela2011, tudela2014}.
A QE placed in the vicinity of a surface with complex electric permittivity can decay into three different channels \cite{Amos1997}: excitation of surface plasmons that propagate along the surface, radiation of photons into the far field, and dissipation through Ohmic losses. The absolute and relative probability of these processes depends on the specifications of the investigated system. Among other materials, graphene is now very well known for its unique optical, electronic and mechanical properties, which are a result of its gapless band structure and locally linear dispersion relation \cite{Neto2009}. From a fundamental point of view, the effects occurring for the case of a quantum emitter in close proximity to such a purely two-dimensional material are still largely unexplored, although they have the potential to set a new scenario for the physics of strong light-matter interaction \cite{Koppens2011}. In a broad frequency range, i.e.\ for energies above twice the Fermi energy and for distances within a quarter of the wavelength, the main mechanism of energy relaxation is non-radiative dipole-dipole resonant energy transfer \cite{Chen2010}. In particular, graphene's unequaled conductivity gives rise to a very strong polarizability, resulting in an efficient energy transfer from donor molecules. As a consequence, graphene's use in the realm of functional materials, e.g.\ as an extraordinary energy sink for photodetection, seems advantageous. Based on the near-field interaction between biomolecules and graphene, a wealth of sensing applications have been proposed \cite{Zhang2011, Wang2011g}. More recently, even quantum position sensors have been suggested, relying e.g.\ on the Casimir effect exerted by graphene on a nearby two level system \cite{Muschik2013}. Similar quantum technologies have been discussed in the literature with the purpose of detecting and manipulating mechanical degrees of freedom of nano-oscillators \cite{Puller2013, Arcizet2011}.
In this letter we provide a first proof of principle for a graphene nano-ruler, based on the measurement of the energy-transfer rate between single organic molecules and a graphene monolayer.
\section{Nanoscopic rulers by optical means}
\begin{figure}
\begin{center}
\includegraphics[width=\textwidth]{setup2D+concept-h.pdf}
\caption{(a) Artistic picture of our test sample, where a DBT:anthracene crystal is interfaced with a graphene monolayer sheet. Depending on the relative distance between molecule and graphene, the excited state decays with different branching ratios via radiative emission or energy transfer to graphene excitons. (b) Sketch of the experimental setup, combining Raman spectroscopy and single-molecule microscopy. Flippable Mirrors (FM) allow to switch between pulsed Ti:Sapphire and a continuous diode-laser (\SI{640}{\nm}) excitation. Fluorescence or Raman signal, collected through a Spatial (SF) and Long-Pass Filter (LP), can be analyzed by means of an electron-multiplied Charged-Coupled-Device camera (CCD), a spectrometer (SPT), avalanche photodiodes (APD) arranged in a Hanbury-Brown-Twiss configuration (HBT), or coupled to a Time Correlated Single Photon Counting card (TCSPC) for lifetime measurements.}
\label{setup}\end{center}
\end{figure}
In the near-field range, plasmonic rulers have been proposed \cite{Jain2007, Jain2007b} and employed to measure e.g.\ nuclease activity \cite{Liu2006}, to follow dimer assembly and DNA hybridization \cite{Sonnichsen2005}. In the simplest geometry, two metal nano particles, integrated into the sample as a probe, interact with each other, yielding a shift of the plasmonic resonance which obeys a $1/d^3$ distance dependence. Different schemes helped improving sensitivity (up to about \SIrange{10}{15}{\nm}) or maximum range (\SI{40}{\nm} \cite{Seelig2007}), by means of nanoparticle-induced lifetime modification. The promise of plasmonic rulers, however, has been partially compromised by a lack of universality, as the actual scaling laws typically depend on the nano-particle shape \cite{Tabor2008}. Our method represents a key extension of another ruler species, based on fluorescence-resonant-energy-transfer (FRET) \cite{Foerster1948, Stryer1978, Lakowicz2006}, which may be employed to measure distances well beyond \SI{10}{\nm}.
In the near field, the energy transfer rate between a donor and an acceptor dipole scales as $d^{-6}$ in free space. When the acceptor takes the form of a surface of dipoles, then integration over all possible transfer sites yields a $d^{-4}$ dependence, whereas transfer to the bulk shows a $d^{-3}$ behavior \cite{Amos1997}. Recent experiments with single emitters \cite{Tisler2013} and ensemble of emitters \cite{Gaudreau2013} have confirmed the predicted \cite{Swathi2009, Gomez-Santos2011} $d^{-4}$ distance dependence of the non-radiative transfer rate to monolayer graphene. The magnitude of such coupling, enhanced with respect to other lossy materials, is described by universal parameters (such as the fine structure constant) so that the relative distance of an object --- in particular that of a fluorescent molecule --- could be accurately determined (within few \si{\nm}), by measuring the emitter decay rate with respect to vacuum. An artistic view of our concept for a graphene nanoruler is depicted in the top left panel of figure \ref{setup}. At the single emitter level, a quantitative analysis for the decay rate modification was still missing to date.
\section{DBT single emitters}
Dibenzoterrylene molecules embedded in thin anthracene crystals (about \SI{40}{\nm}) have been employed as probes in this validation test for a graphene nanoruler. Such study represents likewise an exploratory experiment for the development of a graphene quantum nanoposition-sensor, as described in \cite{Muschik2013}.
The solid-state quantum system that we use combines the high oscillator strength and brightness of organic fluorescent dye molecules with the photostability of single-photon sources, such as inorganic quantum dots or color centers in diamond \cite{Toninelli2010}. As the anthracene crystalline matrix, which acts as a shield to oxygen diffusion, is only few-tens-of-nm-thick, coupling to external photonic structures is envisioned and has indeed been shown in \cite{Toninelli2010a}. Its impact on quantum-optics experiments is due to a narrow emission into the zero-phonon line around \SI{785}{\nm} which, at cryogenic temperatures, is not subject to dephasing and, as a result, is only limited to about \SI{40}{\MHz} by the excited-state lifetime \cite{Nicolet2007, Trebbia2009}.
\begin{figure}
\begin{center}
\includegraphics[width=0.8\columnwidth]{DBTcharacterization.pdf}
\caption{(a) AFM image of anthracene crystals grown by spin-casting on a glass substrate. Crystal edges are clearly visible and, in the inset, a line section cut through the black line shows a thickness of \SI{30}{\nm} with low surface roughness ($\approx\SI{1}{\nm}$). (b) Confocal fluorescence scan of DBT molecules embedded in thin anthracene crystals: typical concentrations and count rates. Evidence of emission from single molecule excitation is given by measurements of the intensity autocorrelation function, such as the one displayed in panel (c). Here, photon anti-bunching from single-molecule emission is displayed in normalized units, without any background correction, and fitted with the function $ y = C (1-b\exp(-|t-t_0|/\tau)) $, yielding a coincidence reduction at $\tau=\SI{0}{\ns}$ of 93\%. (d) Raman spectrum of the CVD pristine graphene sample, obtained with a CW solid state diode laser (\SI{640}{\nm}), exhibiting the characteristic lines at $\approx\SI{1581}{\cm^{-1}}$ and $\approx\SI{2640}{\cm^{-1}}$. The single Lorentzian profile of the 2D band and the respective positions for G and 2D bands show unambiguously graphene monolayer signature.}
\label{DBTcharacterization}\end{center}
\end{figure}
For the purpose of this work, exploiting a broad-band effect, we operate at room temperature and perform a statistical analysis of lifetimes on 150 DBT molecules, whose distances from a single graphene sheet vary from few to \SI{80}{\nm}. To fabricate the DBT-doped crystals, we use a \si{\nano\Molar} solution of DBT, prepared by mixing a solution of DBT in toluene of known concentration with a solution of anthracene in diethyl ether with a concentration of \SI{2.5}{\mg\per\ml}. The crystals are then obtained by spin-casting a \SI{20}{\micro\litre}-droplet of this solution on different substrates, including a reference glass cover slide. Following this protocol, crystals with clear-cut facets and thicknesses ranging from 20 to \SI{80}{\nm} are formed, embedding DBT at single-molecule concentration. The sample surface roughness and thickness were studied by means of Atomic Force Microscopy (AFM); a typical recorded image is shown in figure \ref{DBTcharacterization}a, displaying an average crystal thickness of \SI{30}{\nm}, and a typical surface roughness of the order of \SI{1}{\nm}. DBT is known to be hosted as an impurity in such matrices, but with a well-defined orientation, so as to minimize Gibbs free energy (figure \ref{DBTcharacterization}) \cite{Nicolet2007}. In figure \ref{DBTcharacterization}b, we show a fluorescence image of a DBT:anthracene sample deposited on silica, similar in characteristics to the one measured by AFM. The map is obtained by scanning the sample in the confocal microscope configuration, sketched in figure \ref{setup}. A diode laser at \SI{767}{\nm} (typical intensity is \SI{60}{\kW\per\square\cm}) is used in excitation and the emitted red-shifted fluorescence is collected through a 1.4-N.A. oil-immersion objective after wavelength selection. Bright spots in figure \ref{DBTcharacterization}b correspond to single molecule emission, as the second-order autocorrelation function of panel c) suggests. By reconstructing the histogram of the time intervals between start and stop events, recorded by two avalanche photodiodes (APDs) arranged in a Hanbury-Brown-Twiss geometry (see figure \ref{setup}), we approximate the intensity autocorrelation function for short time periods. We find $g^2(0)\simeq 0.2$ which is a clear indication of sub-poissonian photon statistics from a single photon source \cite{Loudon2000}. Having access to the single molecule behavior also allows for the probing of local properties such as Fermi energy spatial variation.
In order to observe the relaxation dynamics we employ a Ti:Sapphire laser optimized to emit pulses \SI{200}{\fs} long around a wavelength of \SI{767}{\nm}, with a repetition rate of \SI{81.2}{\MHz}. The histogram of photon arrival times upon laser triggering is recorded via a PicoHarp Time-Correlated Single-Photon-Counting card. The acquired signal is well fitted by the convolution between system response function (gray curve in figure \ref{scan}) and a single exponential decay, associated to the depopulation of a single excited state. In absence of external loss channels, the quantum efficiency of our emitter is close to unity and the inverse of the population relaxation time decay is solely due to the radiative decay rate ($\simeq$\SI{25}{\MHz}).
\section{Graphene monolayer}
The graphene monolayers under investigation have been fabricated by chemical vapor deposition (CVD) on copper.
Millimeter-size of polycrystalline graphene is then transferred on a microscope slide and annealed in H$_2$-Ar (1:5) \SI{300}{\celsius} during 3h.
The presence of a monolayer has been verified by performing Raman spectroscopy \cite{Ferrari2003} on the sample before the DBT:anthracene solution was spincasted.
Raman spectroscopy of graphitic systems is a non-invasive, fast and complete tool to investigate doping, stress and structural properties. In particular, the two main Raman features of graphene, G and 2D bands, provide informations about the number of layers, the intrinsic graphene doping or structural defects.
The optical setup for the analysis of the Stokes red-shifted bands is standard for confocal Raman spectroscopy, employing a 100X Objective (N.A.=0.7).
The source is a continuous wave solid state diode laser, with central emission wavelength around \SI{640}{\nm} and TEM00 spatial mode. The excitation laser line is cleaned with a narrow (\SI{2}{\nm}) bandpass filter, whereas the Rayleigh line in detection is filtered out by means of a longpass filter.
The spectrometer used to record Raman spectra has a spectral resolution of \SI{3}{\cm^{-1}}. As shown in figure \ref{DBTcharacterization}d, the respective positions and shape for G and 2D bands are clear signature of a graphene monolayer \cite{Saito2011, Ferrari2006, Berciaud2013}.
The G band position is measured at \SI{1581}{\cm^{-1}} and has pure lorentzian shape.
The profile of the Raman 2D band at \SI{2640}{\cm^{-1}} is lorentzian and shows correct agreement with the expected line shape for monolayer graphene \cite{Ferrari2006}.
\section{Lifetime measurements}
We here prove that the measurement of the decay rate is a tool to accurately pinpoint an emitter position away from a graphene interface. Specifically, we show how the lifetime distribution from a collection of 150 DBT molecules close to a graphene monolayer only depends on universal parameters, besides their position distribution. As lifetime measurements are not affected by the instrumental collection efficiency, whether geometrical or intrinsic, the system shows promising characteristics as a nanoscopic ruler.
We consider a semi-classical model along the lines of what is described in \cite{Blanco2004} and already applied in \cite{Gaudreau2013} for the energy transfer, taking place between an ensemble of molecules and the two-dimensional material.
The decay rate of a molecule at a distance $d$ from a graphene film ($\Gamma_\mathrm{g}$) is calculated as the power radiated by a classical dipole, placed in a semi-infinite medium facing the absorbing material. $\Gamma$ is derived by integration over parallel wave vectors, considering the total field as a result of interference between the dipole emission and its Fresnel reflection. A particularly elegant form describes $\Gamma_\mathrm{g}$: in the \SI{15}{\nm}-range, using graphene DC conductivity $\sigma=e^2/4\hbar$, for a parallel oriented dipole we find:
\begin{equation}
\Gamma_\mathrm{g}/\Gamma_\mathrm{ant}\simeq1+\frac{9\alpha}{256\pi^3(\epsilon_\mathrm{ant}+\epsilon_\mathrm{sub})^2}\left(\frac{\lambda_0}{d}\right)^4
\label{transfer_eq}\end{equation}
where $\Gamma_\mathrm{ant}$ is the decay rate in the bulk medium, $\alpha$ is the fine structure constant, $\lambda_0$ is the free-space emission wavelength, $\epsilon_\mathrm{ant}$ and $\epsilon_\mathrm{sub}$ are the permittivities of the anthracene hosting medium and of the substrate supporting graphene, respectively. Besides the characteristic $d^{-4}$ dependence, inherent to the system dimensionality, the transfer mechanism appears to be well described by universal parameters, independent of the specific realization or the experimental setup. It is worth noting that the theoretical results are in agreement with a full quantum optical analysis \cite{Glauber1991}.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.8\textwidth]{fastnslowdecay+fit+scan.pdf}
\caption{(a) Time-resolved measurements of the fluorescence decay for single DBT molecules, placed at different distances from a graphene layer. Solid black lines are fit to the experimental data (colored on line), with the convolution between the system response function (light gray) and a single-exponential-decay function. The longer lifetime of molecule B with respect to A is associated to a brighter signal, as displayed in panel (b). Here, a close-up of a bigger confocal fluorescence scan is reported, suggesting the typical emitter concentration and signal to noise ratios.}
\label{scan}\end{center}
\end{figure}
The decay rate of molecules coupled and uncoupled to graphene have been compared in terms of excited-state lifetime measurements. The analysis was performed in a systematic way, starting from a scan similar to the one in figure \ref{DBTcharacterization}b and defining an intensity threshold for the faintest detectable molecule, according to a minimum signal to noise ratio (S/N) of 3. On each selected molecule, we then measured the relaxation dynamics and extracted a value for the excited-state lifetime. Figure \ref{scan}a shows particularly clean single-exponential decays of DBT fluorescence, in the vicinity of low-doped graphene. Such signals can be associated to a single optically active system, i.e.\ with no contribution of the host matrix, and with a simple level structure \cite{Toninelli2010}. Decay times were derived from the best fit with a convolution between the system response function (gray solid line in figure \ref{scan}) and a single-exponential decay.
The proximity to a graphene monolayer is clearly reflected in lifetime measurements: a molecule, say molecule A in the example, has been characterized by a short lifetime due to efficient energy transfer to the graphene sheet, also resulting in a fluorescence quenching. On the other hand, molecule B appears brighter and with a longer lifetime, and is supposedly further away from the graphene. Shorter lifetimes are due to the non-radiative decay rate enhancement (expressed by equation \ref{transfer_eq}), which results in a decreased quantum efficiency, hence in fluorescence quenching. This effect has even been exploited to image a graphene monolayer \cite{Treossi2009}, or to characterize the coupling efficiency between emitters and graphene, (see \cite{Chen2010}, quenching factor 70).
\begin{figure}
\centering
\subfloat
{
\includegraphics[width=0.4\columnwidth]{Lifetimes_MZ+Gbnb_Out.pdf}
\label{fig:LifetimesOut}
}
\qquad
\subfloat
{
\includegraphics[width=0.4\columnwidth]{Lifetimes_MZ+Gbnb.pdf}
\label{fig:Lifetimes}
}
\subfloat
{
\includegraphics[width=0.4\columnwidth]{PositionDistribution_MZ+Gbnb.pdf}
\label{fig:Position}
}
\qquad
\subfloat
{
\includegraphics[width=0.4\columnwidth]{PositionUncertainty.pdf}
\label{fig:PositionUncertainty}
}
\caption{Probability density function (PDF) of single-molecule excited-state lifetimes on a reference sample (a) and on a graphene monolayer sheet (b). Histograms represent experimental values, while blue solid lines result from the theoretical models discussed in the text. A PDF of a normal distribution (solid line), centered around \SI{4.2}{\ns} well reproduces the intrinsic spread in lifetimes of DBT molecules in a thin anthracene film. The PDF for lifetimes on graphene requires information on the molecule position distribution, for which we assume the PDF reported in panel (c). The expected potential accuracy in operating our system to measure distances is plotted in panel (d) vs the relative distance.}
\label{Lifetimes}
\end{figure}
As a reference sample we have considered a \ce{SiO2} substrate and collected measurements for 75 molecules.
We observe therein a symmetric distribution of lifetimes around \SI{4.1}{\ns} with $\sigma\simeq \SI{0.4}{\ns}$ (see figure \ref{fig:LifetimesOut}), which cannot be fully accounted for by simple interface effects, acting differently depending on the distance to the surface. In fact, considering the in-plane orientation of DBT molecules \cite{Toninelli2010} and calculating the spread in lifetime for a \SI{40}{\nm}-thick crystal, following the discussion for a multilayer in \cite{Amos1997}, we estimate a 10\% total lifetime variation, resulting in approximately half the experimental spread. Local environment and edge effects contribute to determine the observed variation of lifetimes for DBT, as already suggested in \cite{Kreiter2002, Rogobete2004}. Such mechanisms are reasonably accounted for by a Gaussian distribution (blue solid line in figure \ref{fig:LifetimesOut}).
When molecules are instead brought in close proximity to a mono-atomic carbon layer by spin-casting a DBT-doped anthracene thin film onto a graphene sheet, the lifetime distribution is strongly affected, becoming asymmetric with a long tail for short lifetimes. In particular, the average lifetime is shorter (\SI{3.7}{\ns}), as a new non-radiative decay channel has opened up. In the histogram shown in figure \ref{fig:Lifetimes} the bin size is given by the time resolution of the setup, amounting to \SI{400}{\ps}, after deconvolution with the instrumental response function.
In order to test the model with our actual system, we make an educated guess for the molecule position distribution, insert this assumption into eq.\ \ref{transfer_eq} and compare the obtained probability density function with the experimental data. Let us assume that DBT molecules are homogeneously distributed inside anthracene crystals, whose thicknesses are on average \SI{50\pm10}{\nm}, according to AFM measurements. A cutoff on the distribution is introduced, accounting for the before-mentioned quenching of molecules, which makes detection difficult for distances to graphene closer than \SI{20}{\nm}. In figure \ref{fig:Position}, the resulting probability density function (PDF) of the molecule position distribution is reported. As for $\Gamma_\mathrm{ant}$, we consider gaussian-distributed values, as suggested by the fit to our experimental results on the reference sample (see figure \ref{fig:LifetimesOut}). The expected PDF for the molecule lifetimes on graphene is calculated and plotted (blue solid line) for comparison to the experimental data histogram, shown in figure \ref{fig:Lifetimes}. The agreement clearly confirms the validity of the model, whose simple final formulation enables our system as a tool for position measurement at the nanoscale. In figure \ref{fig:PositionUncertainty} the uncertainty in position measurement is shown for the specific case of DBT molecules in anthracene. Here we have only included the effect of the intrinsic DBT:anthracene lifetime spread and assumed an ideal setup, i.e.\ no cutoff due to detection is considered. We observe that the distance to the graphene interface of DBT molecules can be determined with an accuracy below \SI{5}{\nm}, for distances smaller than \SI{30}{\nm}.
In figure \ref{fig:PositionUncertainty} the uncertainty in position measurement is shown for the specific case of DBT molecules in anthracene. Here we have only included the effect of the intrinsic DBT:anthracene lifetime spread and assumed an ideal setup, i.e.\ no cutoff due to detection is considered. We observe that the distance to the graphene interface of DBT molecules can be determined with an accuracy below \SI{5}{\nm}, for distances smaller than \SI{30}{\nm}.
Given the lifetime distribution for molecules on glass and on graphene, we can estimate the probability to measure a transfer efficiency $\eta$ higher than 40\%, which is the maximum value reported in literature for single emitters \cite{Liu2012,Tisler2013}. According to the equation $\eta=1-\frac{\tau_\mathrm{g}}{\tau_\mathrm{ref}}$, which holds when the intrinsic quantum yield amounts to 1, with $\tau_\mathrm{g}$ and $\tau_\mathrm{ref}$ being respectively the lifetime measured with and without graphene \cite{Tisler2013}, we find that in our case 12\% of the measured molecules have experienced an energy transfer efficiency higher than 40\%. Given a typical S/N for a single molecule equal to about 15 and a minimum detectable lifetime of about \SI{1}{\ns}, we do not expect (probability smaller then 0.1\%) to observe molecules with a transfer more efficient than $\simeq 70\%$. Note that such numbers are only determined by instrumental issues, such as minimum detectable lifetime and S/N. Therefore they do not represent absolute limitations for the proposed nano ruler. Finally we can estimate a maximum measured transfer efficiency from a single DBT molecule to graphene equal to $(61\pm 21)\%$. The uncertainty is estimated taking into account the fluctuations in the reference value and the precision of lifetime measurements.
\section{Conclusions}
In conclusion, we have presented a full statistical study of the coupling between single long-lived organic molecules and a graphene monolayer sheet. We have reported the highest --- to our knowledge --- ever measured transfer efficiency from single emitters to graphene, amounting to $(61\pm 21)\%$. The molecule excited-state lifetime is strongly affected by the presence of the monoatomic carbon layer, because of its two-dimensionality, high conductivity and gapless dispersion relation. As a result, we can detect a FRET-like effect to distances well beyond the characteristic \SI{10}{\nm} of standard acceptor-donor energy transfer. The semi-classical model discussed in \cite{Gomez-Santos2011}, yielding a universal $d^{-4}$ dependence of the coupling efficiency, was successfully verified in the near-field range against a statistical distribution of the molecule lifetimes. The presented investigation on our favorable DBT:anthracene platform constitutes a first proof of principle for a graphene-based nano ruler, where ideally the distance to a surface can be measured by extracting the lifetime of a well-referenced single emitter, serving as a marker. In the near future, the use of single emitters will be essential to focus on local effects such as mapping the local Fermi energy in graphene, useful for electron-transport engineering in graphene based devices. Dibenzoterrylene molecules, emitting single photons on demand in the near infrared, are also particularly promising candidates to launch deterministic single plasmons into heavily-doped graphene.\\
This work was supported by the Seventh Framework Programme for Research of the European
Comission, under FET-Open grant MALICIA (265522) and by MIUR under FIRB Futuro in Ricerca project HYTEQ. F.\ K.\ acknowledges support by the Fundacio Cellex Barcelona, the ERC Career integration grant 294056 (GRANOP), the ERC starting grant 307806 (CarbonLight) and support by the E.\ C.\ under Graphene Flagship (contract no.\ CNECT-ICT-604391). We thank M.\ Inguscio, D.\ S.\ Wiersma, M.\ Bellini and M.\ Gurioli for fruitful discussions and continuous support. We wish to thank also B.\ Tiribilli at the ISC-CNR (Florence) for the realization of the AFM measurements and M.\ Santoro (LENS, Florence) for the Raman characterization of graphene samples.
\vspace{2 cm}
\section*{References}
\addcontentsline{toc}{section}{References}
\bibliographystyle{unsrt}
|
\section{\label{sec:intro}Introduction}
In the seminal paper of Bak and Sneppen, a Self-Organized Critical (SOC)
model \cite{Bak} is introduced to describe the ecological co-evolution of
interacting species. The success of the model in reproducing the punctuated
equilibrium behavior proposed by Gould \cite{Gould}, and already observed
in the fossil records \cite{Raup}, has attracted many authors to study the
model, and variations of it, through several approaches ranging from
simulation \cite{Grassberger},\cite{Rios}
to the renormalization group \cite{Marsili}, \cite{Mikeska}. Bak and Sneppen's
model has found interesting applications in economic studies \cite{Ausloos}, \cite{Yamano}, bacterial evolution \cite{Bose} and even optimization
problems \cite{Stefan},\cite{Boettcher}.
The Theory of Punctuated Equilibrium emerged as an opposition to Phyletic gradualism, which is a theory of speciation that states evolution occurs uniformly and by the steady and gradual transformation of whole lineages, so that no clear line of demarcation exists between an ancestral species and a descendant one. Punctuated equilibrium, on the contrary, states that evolutionary change takes place in short periods of time tied to speciation and extinction events, separated by large time periods of evolutionary quiescence, called stasis. Evidence for these ideas has been found in the fossil record of bryozoans \cite{Cheetham}. This record shows that the first individuals appeared about 140 million years ago, remain unchanged for its first 40 million years (stasis). After that, an explosion of diversification is observed, followed by another period of stasis. Other well known events, observed in the fossil record and explained by Punctuated Equilibrium are the extinction of dinosaurs, about 50 million years ago, or the huge number and sudden emergence of new species during the Cambrian period, in the Paleozoic Era, about 500 million years ago, called the Cambrian Explosion.
On the other hand, the study of wealth and income distributions in society, is a very important and fundamental area of research for practical and theoretical reasons to social scientists, economists, econophysicists, sociologist, philosophers, etc. and also concerns to politicians, government administrators, international bankers,and surely to national security agencies from many countries and of course to every common citizen. Although questions on the origin and causes of inequality are very old; attempts to answer them have been not very successful, even if many ideas have been proposed to understand and solve the problem. Between these ideas, we can mention the following: difference in religious ethics, lack of a qualified workforce, dependence on external technology, low level of internal savings, non-equilibrium between exports and imports, low cognitive and schooling skills of population, level of corruption and quality of democracy, capital's rate of return exceeding rate of output and income, and many more \cite{Weber, Ranis,Lewis,Chenery,Taylor,Hanushek,Hilman,Hongyi,Acemoglu, Piketty}.
Even more, large scale social and ideological experiments, intended and implemented by force, to solve the inequality problem by centralization of economy, have failed spectacularly with a terrible prior and posterior cost in human suffering and lives, human rights violations, famines, waste of economic resources, political and economical instability, ``hot'' and ``cold'' wars, immigration waves, etc.
The important fact is that currently the extreme economic inequality problem seems is not any more only restricted to the beforehand called ``Third world countries'', but is also becoming a big concern and serious problem in developed economies, where the social and wealth gap between the low-medium income segments of population and the richer one, has been recently increasing fast and systematically~\footnote{Although some economists and policy designers do not make any distinction between inequality and poverty, they are different issues. A society or country can be quite equal with a high number of very poor people or vice versa.} (for an extensive and polemic discussion on this topic see \cite{Piketty}).
The first empirical studies to understand wealth distribution were made by Pareto \cite{pareto}, who proposed that the wealth and income distributions obey an universal power law. Subsequent studies have shown that this is not the case for the whole range of population wealth values. Mandelbrot \cite{mandelbrot} proposed that the Pareto conjecture only holds at the higher values of wealth and income. The initial part (low wealth or income) of the distribution has been identified with the Gibbs distribution \cite{chakrabarti3,chakrabarti5}, while the middle part, according to Gibrat \cite{gibrat}, takes the form of a log-normal distribution.
Recently, due to great advances in Complexity Sciences and computing power new ways to model and understand social and economic systems have emerged. Between the most important and well known applications of this computational methods we can mention the use of multi-agent based models to investigate the problem of wealth distribution \cite{chakrabarti5,Dragulescu,Bouchaud,Chatt,scalas,scalas3}
In this work, by using a multi-agent computer methodology, we explore the effect of introducing the extremal dynamics of the already mentioned Bak-Sneppen model, on the wealth distribution produced by two very simple economic exchange models and study their corresponding Gini indices.
In particular, we focus our attention in two well known toy-models of economic interactions that have been used extensively due to their simplicity, such as the so called ``Yard-Sale" (YS) and
``Theft and Fraud" (TF) models \cite{Hayes}. Although these two models have the
advantage of their simple rules for analysis and simulation, they are over simplified, toy model versions of a real economy and they do not produce realistic wealth distributions. For this reason, several authors have made some refinements to introduce and model more realistic situations, such as the introduction of savings \cite{chakrabarti}, changing the probability of winning according to the
relative wealth of the traders \cite{Sihna}, allowing the agents to go into debt \cite{Dragulescu} and by the introduction of altruistic behavior \cite{Trigaux, achach}. Interesting and deep textbook discussions of multi-agents exchange models in the context of the present work are \cite{Scalasbook} and \cite{Chakabook}.
\section{Model Description}
We can treat an economy in its simplest form as an interchange of wealth
between pairs of economic entities (people, companies, countries, etc.), named our ``agents" at successive instants of time. Every time two agents interact, wealth flows from one to the other according to some rule. In the case of the YS model,
the winner takes a random fraction of wealth from the
poorer player, while in the TF case,
the winner takes a random fraction of the loser's wealth.
There is not production or consumption of wealth in these models,
no taxes, savings, etc. Under these circumstances, the YS model produces a
collapse in the economy: all the wealth ends in the hands of a single
agent, a phenomenon know as extreme wealth condensation. On the other hand, the TF model
does not collapse, but leads to a wealth distribution given by the Gibbs
distribution~\cite{Dragulescu1,chakrabarti5,ispolatov}. In particular, the introduction of punctuated equilibrium in these two
models has not been studied in depth, and therefore, in this paper we
investigate the effect that punctuated equilibrium behavior has on the
dynamics of the models and the changes that it can produce on the
distribution of the wealth.
As discussed in \cite{Hayes}, after a sufficient number of interactions take place in the YS model, a single agent ends up with (almost) all the money in the system.
On the other hand, the TF model produces a distribution with the
majority of the agents ending with a wealth close to the average, and no
agent becomes extremely rich.
Note that in the TF case, a very poor agent that interacts with a rich one,
can become rich if he wins the bet and the fraction is enough high, a situation that is not expected to happen in the real world, unless we are dealing with illegal activities, hence the name of the model.
These two outcomes from the YS and TF models actually do not reflect what
really happens in a modern economy, where the wealth distribution takes
an exponential distribution form for the poor and medium class sectors
of the population, and a Pareto distribution for the richer individuals
\cite{Estevez,Wright}.
\subsection{Simulation implementation}
Our simulation runs in the following way: $N$ agents are arranged on an
one dimensional lattice with periodic boundary conditions. This lattice is not important in any trading activity between agents, however it will be very important and necessary to keep track of every agent's first neighbors, in order to apply EP rules to introduce wealth mutations in a subsequent step of our simulation as explained below. Initially a certain amount of money $M$, is distributed equally to all agents in such a way that
$\Sigma_{i=1}^N m_i= M$. The system is closed, meaning that the total amount of money in the system, $M$, is always constant (i.e. no production) and the number of agents $N$ remains unchanged (neither dead, birth or migration of agents is allowed). At each time step, two agents are randomly chosen and they interact according to the YS or the TF model. However, with probability $\rho$, the interaction will be ruled by punctuated equilibrium (PE) dynamics instead of YS or TF. Then, another pair of agents are chosen and the process is repeated $K$ times, which constitutes one Monte Carlo step (MCS). In our simulations and to obtain wealth distributions showed in next section a typical number for $K = 10^6$ and every agent starts the simulation owning 100 monetary units.
Punctuated equilibrium (PE) is introduced in both models in the following
way: locate the poorest agent, lets say agent $k$ and assign new values of wealth to agents $k-1$, $k$ and $k+1$ at random, but taking care that their combined wealth does not change. Extinction of agents is not allowed in order to maintain the overall number of agents $N$ constant.
In the case of the YS or TF rules, two agents $i$ and $j$ are randomly
chosen at time $t$. The winner, which is also chosen at random, takes an amount $T$
of money from the loser.
The traders' wealth $w_i$ and $w_j$ at time $t+1$, assuming that agent $i$ is the winner and agent $j$ is the loser, will be
\begin{eqnarray}
& & w_i(t+1)=w_i(t)+T\ ,\\
& & w_j(t+1)=w_j(t)-T\ ,
\end{eqnarray}
Where $w_j$ is the wealth of the poorest agent for the YS model and $w_j$ is the wealth of the loser for the TF model.
\noindent
The amount $T$ of the wealth that changes hands in the bet is defined as
\begin{equation}
T = \alpha MIN (w_i(t), w_j(t)),
\end{equation}
\noindent
for the YS model and
\begin{equation}
T = \alpha w_j(t),
\end{equation}
\noindent
for the TF model, assuming that agent $j$ is the loser, where the parameter
$\alpha$ is a random number from an uniform distribution
in the interval [0,1].
The inequality in the final wealth distribution of the system can be quantified using the Gini index
\cite{Berrebi}, defined by
\begin{equation}
G=\frac{\Sigma^N_{i=1} \Sigma^N_{j=1} |x_i-x_j|}{2N^2 \mu},
\end{equation}
\noindent
where $\mu$ is the average wealth, and $x_i$ and $x_j$ represent the wealth
of agents $i$ and $j$ respectively. A perfect distribution of wealth, where
everybody has the same amount of money will give a value $G=0$. The other extreme, where one individual owns all the money has a Gini value of 1.
\section{Simulations Results}
\subsection{Gini Index Analysis}
We first consider the Gini index in the YS model as a function of time, for several
values of the PE ``perturbation" $\rho$. The results are shown in Figure \ref{giniYS1400}, where one can see that for small enough values of $\rho$, the final result is the same as with the ``pure" YS model, that is, $G(t)$ reaches the value of 1 as time increases (the economy collapses in a state where a single agent has all the money). However, there is a critical value of $\rho_c$ for which the system does not collapse and $G$ takes an asymptotic value less than 1. As the system size increases, the perturbation necessary to get the system out of collapse increases, as can be seen in Figure \ref{bothGiniI}, where the asymptotic value of $G$ is shown as function of $\rho$ for several system sizes $N$. Same figure shows the results for the TF model. In this model the effect of the perturbation only decreases the asymptotic value of $G$.
\begin{figure}[htb!]
\begin{center}
\resizebox{0.80\textwidth}{!}{%
\includegraphics[angle=-90]{plotginiYS1400aN.eps}}
\caption{Gini index $G(t)$ as a function of time for the YS Model for
$N=1400$ agents. We set up the strength of the perturbation $\rho$ to the values $\rho$=0.6, 0.65, 0.7, 0.75, 0.8, 0.85, 0.9, 0.95, 0.99. In this
figure the upper curve corresponds to $\rho$=0.6, and the bottom one
corresponds to $\rho$=0.99. We observe that as $\rho$ increases, the curves tend to reach the $G =1$ region more slowly. For $\rho \ge 0.8$ the curves avoid completely that region.}
\label{giniYS1400}
\end{center}
\end{figure}
\begin{figure}[htb!]
\begin{center}
\resizebox{0.80\textwidth}{!}{%
\includegraphics[angle=-90]{bothGiniIndexN.eps}}
\caption{Gini index $G(p)$ as a function of perturbation for our two models: (a) YS model: $G(\rho)$ curves plotted for a different increasing number of agents $N$ indicated in the figure. (b) TF model: In this case, we use lower case $n$ to indicate the number of agents. $G(\rho)$ curves were plotted for $n$=128, 256, 512, 1024 values.}
\label{bothGiniI}
\end{center}
\end{figure}
\subsection{Wealth Distribution}
Figure \ref{cumulative2048} shows in a log scale for the YS model
the wealth cumulative distribution function (CDF), that gives
the probability $P$ that an agent chosen at random will have a wealth greater than $w$. Here, $N=2048$ agents and were simulated the values for the perturbation $\rho$=0.78, 0.79, 0.80, 0.84, 0.90, 0.99.
\begin{figure}[htb!]
\begin{center}
\resizebox{0.80\textwidth}{!}{%
\includegraphics[angle=-90]{cumulativeplot2048NNNNN.eps}}
\caption{Wealth Cumulative Distribution Function (CDF) in a log-log scale for the YS model. i.e. the probability $P$ that an agent chosen at random will have a wealth greater than or equal to $w$. We show simulations with $N=2048$ agents and different increasing values of $\rho \geq \rho_c$. We observe that as $\rho$ increases, wealth distribution becomes fairer. Power law fits for curves determined by $\rho = 0.78$, 0.79 and 0.80 are displayed. Corresponding exponents are $\alpha_1$ = 0.332, $\alpha_2$ = 0.729 and $\alpha_3$ = 0.80. From second fit, we select $\rho_c$ = 0.79. See table \ref{tab:PowerFits}. Curve for $\rho_3$ = 0.80 decays asymptotically as a power law fit, and it is the closer to the observed in real data.}
\label{cumulative2048}
\end{center}
\end{figure}
From this figure \ref{cumulative2048}, we can see that, as the perturbation $\rho$ increases, the probability of finding richer agents decreases, giving us a ``fairer'' distribution of wealth, compared to lower perturbations where only higher values of wealth are found.
We also observe in same figure that for the values of $\rho=0.78$ and $0.79$ it is possible to fit very well these curves with a power law model with the form $F(w) = cw^{-\alpha}$. Parameters $c$ and $\alpha$ of these fits are displayed in table \ref{tab:PowerFits}
\begin{table}[!htb]
\begin{center}
\caption{Wealth CDF power law fit parameters for values of $\rho= 0.78$ and 0.79. NDF denotes the number of degrees of freedom.}
\label{tab:PowerFits}
\begin{tabular}{lccccr}
\hline
\hline
$\rho$ &$c$ &$\alpha$ & $\chi^2$/NDF& Selected Fit \\
\hline
\noalign{\smallskip}\hline\noalign{\smallskip}
0.78 &0.029 $\pm$ 0.062 &0.332 $\pm$ 0.062 &32.49/998& No\\
0.79 & 2.550 $\pm$ 0.024& 0.729 $\pm$ 0.002 & 4.956/998& Yes\\
\hline
\hline
\end{tabular}
\end{center}
\end{table}
Again, from figure \ref{cumulative2048} and value of $\chi^2$/NDF in table \ref{tab:PowerFits}, we can see that the best power law fit corresponds to $\rho_c$ = 0.79, with a exponent $\alpha$ = 0.729. This confirms our observation of a phase transition around $\rho_c = 0.79$ described by the analysis of the behavior of the Gini index as a function of time for different $\rho$ values and displayed in figure \ref{giniYS1400}.
For $\rho = 0.80$ the corresponding wealth CDF decays asymptotically as a power law with an exponent $\alpha_3$ = 1.898 $\pm$ 0.002. The fit was performed in the region $w > 14764$ on the biggest 647 observations $\chi^2$/NDF obtained is 0.1685/645. This exponent has a value enough close to the observed in real wealth cumulative distribution data that is approximately 2~\cite{Dragulescu1}
. Interestingly, the more realistic wealth distribution obtained does not corresponds to the value of $\rho_c$ (a pure power law distribution), but to a value close to it, $\rho=0.80$ which is an like real asymptotic power law distribution with an Pareto exponent close to, but slightly lower than 2.
\subsection{System Size effect}
After being established that the maximum wealth depends on the perturbation, we now proceed to investigate the effect of the system size on our simulations. This is shown in Figure \ref{linescumulaN}. We divided the Gini index from Figure 2 by the system size $N$ and then normalized it to 100. We can see that its behavior is independent of $N$.
\begin{figure}[htb!]
\begin{center}
\resizebox{0.80\textwidth}{!}{%
\includegraphics{linescumulativeNN.eps}}
\caption{Gini index divided by the agents number N and scaled to 100 as a function of the perturbation for our two models: (a) YS model: figure shows this scaled Gini index for five different number of agents $N$ with values indicated in the figure (b) TF model: Scaled Gini index for three different number of agents$ n$=256, 512, 1024 in the inset. Again we have used lower case $n$ to indicate the number of agents for TF case.}
\label{linescumulaN}
\end{center}
\end{figure}
For the YS model, in Figure \ref{cumulativeC2048} we show that the critical Gini index $G_c$, does not depend on the number of agents $N$ involved in our simulations. $G_c$ fluctuates around a mean value of $<G_c> = 0.6162 \pm 0251$. The mean value $<G_c>$ was obtained by fitting a constant horizontal straight line to our data of $G_c$ values for a different number $N$ of agents. Same numerical value results from averaging directly our $G_c$ values until the fourth position after the decimal point ($10^{-4}$). This is a good result because if $G_c$ were dependent on $N$, the wealth distribution for the YS model should change on the number of agents, something that should not happen in a good simulation
Continuing with the YS model, in the upper inset of same Figure \ref{cumulativeC2048} we plot the critical $\rho_c$ as a function of $log(N)$. We can see as was pointed out above, at the beginning of this section, that the critical perturbation increases very slowly with the system size. In fact, since we have defined $\rho$ as the probability of switching on or not PE in the interactions between agents, its numerical value can not be larger than one. For the extreme case of``an infinite system size'' we should have $\rho_c\rightarrow 1$, case where nothing can be done to avoid that the system collapses increasing the strength of the perturbation $\rho$. Of course this extreme case does not represent any real economic system, since that although they can be constituted of a very big number of agents, their number is always finite.
\begin{figure}[htb!]
\begin{center}
\resizebox{0.85\textwidth}{!}{%
\includegraphics{GcRhovslogNNFinal.eps}}
\caption{YS model. Gini index threshold $G_c$ as a function of the number of agents $N$. We can see that $G_c$ seems to become independent of $N$. For $N > 256$ It fluctuates around its mean value $<G_c>$. Red broken line was obtained by a $\chi^2$ fit procedure applied to the 8 last data points. We also show in the inset the plot $(1-\rho_c)$ vs $log(N)$. It can be observed that $(1-\rho_c) \rightarrow 1$ very slowly and asymptotically}
\label{cumulativeC2048}
\end{center}
\end{figure}
\section{Conclusions}
In this work, we analyze the TF and YS models under the influence of the punctuated equilibrium dynamics, which is introduced as a perturbation $\rho$ determined by the probability of applying or not PE in agents exchange of money. Although TF model is weakly affected by the introduction of PE, for the case of YS and for perturbations $\rho >0.8$ , the asymptotic Gini index becomes different to one, meaning that the perturbation avoids the collapse in the economy where a single agent takes all money, re-allocating it between agents in a way that a lower inequality in the distribution of wealth is observed. Even more, a phase transition around a critical value $\rho_c = 0.79$ is observed . For this critical value of $\rho$, the corresponding wealth distribution displays a power law decay with an exponent $\alpha=0.729 \pm 0.002$. For a value of $\rho$ = 0.80 the wealth distribution decays asymptotically as a power law with an exponent $\alpha_3$ = 1.898 $\pm$ 0.002, consistent with the values observed in real data.
The resulting wealth distributions can be tuned to different values of the Gini index, adjusting the perturbation. A corollary from this, that is possible to extend to real economics is that a fairer wealth distribution is only attainable by means of a mechanism of wealth re-distribution, as by example taxes in real life. The implementation of different re-distribution mechanisms or determining what is the optimum proportion of individual income to taxing or re-distribute are problems that also could be explored through agents modeling methodology.
Even if the evolutionary economic approach presented here seem very naive and far from representing real economic phenomena, it has precisely the virtue that preserving the simplicity of YS and TF models, produces results more realistic than the obtained by the original models without PE dynamics included, such as a finite wealth distribution decaying as a power law. This is important for the agents model theory and implementation goal of constructing a minimum agents model for real economic and social systems. Also, and in our opinion, our results are important for the most ambitious and long term end of constructing a real microeconomics theory, in the sense of statistical physics, if possible \cite{Scalas2}. We believe that an initial and main way of attacking successfully this difficult problem is through the intensive use of multi-agent simulations techniques, followed by a formalization of results and techniques emerged from this approach. For a nice discussion on these issues in the context of diverse agents models and SOC, see \cite{Jeldtoft}, chapter 5.
Besides, other studies have been made where the simple YS and TF models are extended including effects of savings, of course taxes and other mechanisms in order to make them more realistic. We believe that our approach using the extremal dynamics of the Bak-Sneppen model which has the effect of re-allocating wealth between agents, is simpler and yields similar results, a feature that can be used to investigate particular economic phenomena with a simpler model. This can benefit both simulation and analytic studies. Any way, the study of many-body real world systems, through the study of computational models is not an easy task, considering that sometimes we do not understand in deep those models. In our case, it would be very interesting to attack the same problem presented here using a more formal approach as a Mean Field Model. This is matter of a future work.
\section*{Acknowledgments}
This work has been supported by SEP-Conacyt (M\'exico) under project grant numbers 155492 and 135297. We also thank PROMEP and SNI (M\'exico) for continued support. Some of the analyses of this paper were performed using ROOT \cite{root}. We thank pertinent comments and suggestions of E. Scalas.
|
\section{\indexname}]%
\thispagestyle{plain}\parindent\z@
\parskip\z@ \@plus .3\p@\relax
\let\item\@idxitem}
{\if@restonecol\onecolumn\else\clearpage\fi}
\makeatother
\makeatletter
\let\o@verbatim\verbatim
\def\verbatim{%
\ifhmode\unskip\par\fi
\ifx\@currsize\normalsize
\small
\fi
\o@verbatim
}
\renewcommand \verbatim@font {%
\normalfont \ttfamily
\catcode`\<=\active
\catcode`\>=\active
}
\RequirePackage{shortvrb}
\MakeShortVerb{\|}
\begingroup
\catcode`\<=\active
\catcode`\>=\active
\gdef<{\@ifnextchar<\@lt\@meta}
\gdef>{\@ifnextchar>\@gt\@gtr@err}
\gdef\@meta#1>{\m{#1}}
\gdef\@lt<{\char`\<}
\gdef\@gt>{\char`\>}
\endgroup
\def\@gtr@err{%
\ClassError{ltxguide}{%
Isolated \protect>%
}{%
In this document class, \protect<...\protect>
is used to indicate a parameter.\MessageBreak
I've just found a \protect> on its own.
Perhaps you meant to type \protect>\protect>?
}%
}
\def\verbatim@nolig@list{\do\`\do\,\do\'\do\-}
\newcommand{\m}[1]{\mbox{$\langle$\it #1\/$\rangle$}}
\renewcommand{\arg}[1]{{\tt\string{}\m{#1}{\tt\string}}}
\newcommand{\oarg}[1]{{\tt
}\m{#1}{\t
]}}
\def\cmd#1{\cs{\expandafter\cmd@to@cs\string#1}}
\def\cmd@to@cs#1#2{\char\number`#2\relax}
\DeclareRobustCommand\cs[1]{\texttt{\char`\\#1}}
\newcommand*{\file}[1]{\texttt{#1}}%
\def\GetFileInfo#1{%
\def\filename{#1}%
\def\@tempb##1 ##2 ##3\relax##4\relax{%
\def\filedate{##1}%
\def\fileversion{##2}%
\def\fileinfo{##3}}%
\edef\@tempa{\csname ver@#1\endcsname}%
\expandafter\@tempb\@tempa\relax? ? \relax\relax}
\def{}
\def{}
\def{}
\makeatother
\title{%
Probing the Mysteries of the X-Ray Binary 4U 1210-64\\
with ASM, PCA, MAXI, BAT and Suzaku
}%
\title{\today}
\author{Joel~B.~Coley\altaffilmark{1,2},
Robin~H.~D.~Corbet\altaffilmark{1,2},
Koji~Mukai\altaffilmark{1,2},
Katja~Pottschmidt\altaffilmark{1,3}}
\email{<EMAIL>}
\altaffiltext{1}{University of Maryland Baltimore County, 1000 Hilltop Cir, Baltimore, MD 21250 USA}
\altaffiltext{2}{CRESST/Mail Code 662, X-ray Astrophysics Laboratory, NASA Goddard Space Flight Center, Greenbelt, MD 20771, USA}
\altaffiltext{3}{CRESST/Mail Code 661, X-ray Astroparticle Physics Laboratory, NASA Goddard Space Flight Center, Greenbelt, MD 20771, USA}
\expandafter\GetFileInfo\expandafter{\jobname.tex}%
\begin{document}
\begin{abstract}
4U 1210-64 has been postulated to be a High-Mass X-ray Binary powered by the Be mechanism. X-ray observations { with} \textsl{Suzaku}, the \textsl{ISS} Monitor of All-sky X-ray Image (MAXI) and the { \textsl{Rossi X-ray Timing Explorer}} Proportional Counter Array (PCA) and All Sky Monitor (ASM) provide detailed temporal and spectral { information} on this poorly understood source. Long term ASM and MAXI observations show distinct high and low states and the presence of a 6.7101$\pm$0.0005\,day modulation, { interpreted as the orbital period}. Folded light curves reveal a sharp dip, { interpreted as an eclipse. To determine the nature of the mass donor}, the predicted eclipse half-angle was calculated as a function of inclination angle for several stellar spectral types. The eclipse half-angle is not consistent with a mass donor of spectral type B5 V; { however, stars with spectral types B0 V or B0-5 III are possible. The best-fit spectral model consists of a power law with index $\Gamma$=1.85$^{+0.04}_{-0.05}$ and a high-energy cutoff at 5.5$\pm$0.2\,keV modified by an absorber that fully covers the source as well as partially covering absorption.} Emission lines from S XVI K$\alpha$, Fe K$\alpha$, Fe XXV K$\alpha$ and Fe XXVI K$\alpha$ were observed in the \textsl{Suzaku} spectra. Out-of-eclipse, Fe K$\alpha$ line flux was strongly correlated with unabsorbed continuum flux, { indicating} that the Fe I emission is the result of fluorescence of cold dense material near the compact object. The Fe I feature is not detected during eclipse, further supporting an origin close to the compact object.
\end{abstract}
\section{Introduction}
4U 1210-64 is an unusual X-ray Binary (XRB) first detected by the \textsl{Uhuru} satellite in 1978 \citep{1978ApJS...38..357F}. \citet{1978ApJS...38..357F} found that 4U 1210-64 appears to be a variable source with a mean flux of 8.9$\times$10$^{-11}$\,erg cm$^{-2}$ s$^{-1}$ in the 2-6\,keV band. Subsequent detections of 4U 1210-64 include the \textsl{Einstein} \citep{1992ApJS...80..257E} and \textsl{EXOSAT} \citep{1999A&AS..134..287R} slew surveys at fluxes 1.7$\times$10$^{-11}$\,erg cm$^{-2}$ s$^{-1}$ and $\sim$7$\times$10$^{-10}$\,erg cm$^{-2}$ s$^{-1}$ in the 0.16--3.5\,keV and 1--8\,keV bands, respectively; the wide-field cameras (WFC) on board \textsl{BeppoSAX} \citep{2007A&A...472..705V}, \textsl{INTEGRAL}/IBIS \citep{2003A&A...411L.131U} and the Burst Alert Telescope (BAT) on board the \textsl{Swift} observatory \citep{2010ApJS..186..378T, 2010A&A...510A..48C} at fluxes 2.57$\times$10$^{-10}$\,erg cm$^{-2}$ s$^{-1}$, 1.1$\times$10$^{-11}$\,erg cm$^{-2}$ s$^{-1}$ and $\sim$2$\times$10$^{-11}$\,erg cm$^{-2}$ s$^{-1}$ in the 2--10\,keV, 20--100\,keV and 14--195\,keV bands, respectively.
A soft X-ray counterpart at coordinates (J2000) RA=12:13:14.7, Dec=-64:52:31 with a position uncertainty of $\sim$4$\arcsec$ was revealed \citep{2007ATel.1253....1R} in observations conducted with the X-ray telescope \citep[XRT,][]{2005SSRv..120..165B} on board \textsl{Swift}. \citet{2007ATel.1253....1R} modeled the X-ray spectrum of 4U 1210-64 using a continuum consisting of a power law with hard photon index and an emission feature at 6.7\,keV with an equivalent width (EQW) of $\sim$400 eV. These observations resulted in a classification of 4U 1210-64 as an intermediate polar cataclysmic variable \citep[CV,][]{2007ATel.1253....1R}.
Observations using the \textsl{Swift}/XRT in late 2006 and early 2008 confirmed the presence of an emission feature at 6.7\,keV, thought to be Fe XXV \citep{2010ApJ...511..A48}. The emission feature, observed to be prominent when the source is at intermediate flux levels, was not seen when the source entered periods of low or high flux \citep{2010ApJ...511..A48}. At intermediate flux levels, the EQW of the emission line was observed to be $\sim$1.6\,keV. This very large EQW is suggestive of a blend of Fe lines. While \citet{2007ATel.1253....1R} suggested that the spectral properties indicate that 4U 1210-64 is a CV, \citet{2010ApJ...511..A48} proposed that the stellar remnant in the system is a neutron star based on the very large EQW and variability of the emission feature. Under the assumption of a power law continuum and a distance of $\sim$2.8\,kpc \citep{2009A&A...495..121M}, the high state X-ray luminosity of 4U 1210-64 was found to be 1.9$\times$10$^{35}$ erg s$^{-1}$ in the 2--10\,keV band \citep{2010ApJ...511..A48}. This exceeds the typical luminosities observed in CVs by a factor of $\sim$1--2 orders of magnitude \citep{2006MNRAS.372..224B, 2008A&A...489.1121R, 2009A&A...496..121B}. A blackbody soft excess with a temperature of $\sim$1.5\,keV was found in 4U 1210-64, which implies that the accretor is more compact than a white dwarf \citep{2010ApJ...511..A48}. \citet{2010ApJ...511..A48} concluded that the presence of the soft excess provides compelling evidence against a CV interpretation of the system.
The optical counterpart of 4U 1210-64 was observed { on MJD\,54529.3} using the 1.5\,m Cerro
Tololo Interamerican Observatory (CTIO) in Chile \citep{2009A&A...495..121M}. Based on its
optical spectrum, \citet{2009A&A...495..121M} proposed that the spectral class of the mass donor
is B5 V. The features observed in the optical spectrum consist of Balmer series lines in
absorption and emission of neutral helium, singly ionized helium and a blend of doubly ionized
nitrogen and carbon. The EQW of the H$\alpha$ line was observed to be approximately 2.5\,${\rm \AA}$ \citep{2009A&A...495..121M}. This is less than half the expected value for a B5 V star,
which suggests that emission features are present. While an early-type mass donor suggests that
4U 1210-64 is a High-Mass X-ray Binary, the presence of a B5 V main-sequence star in a HMXB would
be unprecedented \citep{1998A&A...338..505N}. The majority of Be X-ray binaries (BeXBs) host
primaries with spectral properties that range from late O to early B type stars. This will be
discussed in further detail in Section 4.
A 6.7 day orbital period was discovered using data from the \textsl{Rossi X-ray Timing Explorer} All-Sky Monitor (ASM) \citep{2008ATel.1861....1C}. In addition, \citet{2008ATel.1861....1C} report three main states of the system along with the possibility of an eclipse. We investigate these results in more detail in Section 3.1 using additional data from ASM, the Monitor of All-Sky X-ray Image (MAXI) on board the \textsl{International Space Station (ISS)} and Burst Alert Telescope (BAT) instruments on board the \textsl{Swift} spacecraft.
This paper is structured in the following order: \textsl{Suzaku}, MAXI, PCA and ASM observations are presented in Section 2; Section 3 focuses on the results of the X-ray campaign and Section 4 presents a discussion of the results. The conclusions are outlined in Section 5. If not stated otherwise, the uncertainties and limits presented in the paper are at the 1$\sigma$ confidence level.
\section{Observations and Data Analysis}
The observations outlined below consist of data collected during a two day \textsl{Suzaku} observation of 4U 1210-64 (2010 Dec. 23--25), pointed observations using the \textsl{RXTE} Proportional Counter Array (PCA) as well as long-term observations of the system using the following all-sky monitors: MAXI, ASM, and BAT.
\subsection{RXTE}
\subsubsection{ASM}
The ASM on board RXTE \citep{1996ApJ...469L..33L} consisted of three coded-aperture Scanning Shadow Cameras (SSCs), each containing a position-sensitive proportional counter with a field-of-view (FOV) of 6$\arcdeg{\times}$90$\arcdeg$ FWHM \citep{1997asxo.proc...29R}. ASM scanned approximately 80$\%$ of the sky per spacecraft orbit ($\sim$90 minutes) with a 90\,s time resolution (``dwells") \citep{1997asxo.proc...29R, 1996ApJ...469L..33L}. The FOV of the cameras is fixed on the sky for the duration of a ``dwell". Sensitive to energies in the 1.5--12\,keV band, ASM observed 4U 1210-64 from MJD 50087 to 55924.
The light curves were retrieved from the ASM/\textsl{RXTE} database\footnote{http://xte.mit.edu/ASM\_lc.html} managed by MIT, which includes ``dwell-by-dwell" light curves and daily averaged light curves. The light curves are divided into three energy bands: 1.5--3\,keV, 3--5\,keV and 5--12\,keV. Over the entire energy range (1.5--12\,keV), the Crab produces approximately 75.5\,counts s$^{-1}$ \citep{1997asxo.proc...29R, 1996ApJ...469L..33L}. ``Blank field" observations of regions at high Galactic latitudes indicate that a systematic uncertainty of $\sim$0.1\,counts\,s$^{-1}$ must be taken into account \citep{1997asxo.proc...29R, 1996ApJ...469L..33L}. We used the ``dwell-by-dwell" light curves in our analysis (see Section 3.1).
\subsubsection{PCA}
\label{PCA Description}
Consisting of five proportional counter units (PCUs), the PCA was sensitive to X-rays in the energy band 2--60\,keV. The total effective area of the instrument was $\sim$6500\,cm$^2$ with a FOV at FWHM of 1$\arcdeg$ \citep{1996SPIE.2808...59J}. { At 6.0\,keV, the energy resolution FWHM in the PCA is $\sim$1\,keV \citep{2006ApJS..163..401J}.}
4U 1210-64 was observed 65 times between MJD\,54804.0--54842.0. Individual observations typically lasted for 2--4\,ks but were as short as 1.4\,ks and as long as 45.6\,ks. Dataset IDs were 93455-01 (23 observations) and 94409-01 (42 observations). The majority of the observations were performed during a single spacecraft orbit with no interruptions; longer ones spanned multiple orbits, with interruptions due to Earth occultations and/or SAA passages. Of the 5 PCUs, only PCU2 was used consistently for these observations, so we have opted to analyze only data taken with PCU2. The Crab produces $\sim$2000--2400\,counts s$^{-1}$ in the PCU2 top layer\footnote{http://www.sternwarte.uni-erlangen.de/wilms/rxte/}.
Each detected event was recorded in different ways by the on-board experiment data system (EDS). We analyzed Standard2 mode data, producing 129-channel spectra every 16\,seconds, for spectral analysis and for low-frequency timing analysis. We generated the ``faint'' model background\footnote{http://heasarc.nasa.gov/docs/xte/recipes/pcabackest.html}, since the source mostly stayed below 40\,counts s$^{-1}$ PCU$^{-1}$. For high-frequency timing analysis, we generated light curves in 10\,ms bins using the GoodXenon event mode data, without background subtraction. Spectral data including background subtraction were reduced and analyzed using the standard screening criteria \citep{2006ApJS..163..401J}.
\subsection{MAXI}
The MAXI instrument is an X-ray slit camera sensitive to energies 0.5--30\,keV \citep{2010SPIE.7732E..26M}. MAXI consists of two types of slit cameras, the Gas Slit Camera (GSC) and the Solid-state Slit Camera (SSC), which observe the X-ray variability of over 1000 sources over every ISS orbit of approximately 92 minutes \citep{2010SPIE.7732E..26M}. Of the two cameras, we make use of the GSC data that are routinely made available by the MAXI team.
Covering an energy range of 2--30\,keV, the GSC consists of twelve one-dimensional position sensitive proportional counters (PSPC), which make six camera units \citep{2010SPIE.7732E..26M}. The overall FOV is a 160$\arcdeg{\times}$3$\arcdeg$ slit, which scans both the horizon and zenith directions \citep{2010SPIE.7732E..26M}.
We analyzed MAXI data obtained between MJD 55061.5--56394.5. Light curves of energies 2--4\,keV, 4--10\,keV and 10--20\,keV were retrieved from the data available in the MAXI RIKEN database\footnote{http://maxi.riken.jp/top/}. The 2--4\,keV and 4--10\,keV light curves were subsequently co-added for comparison with the light curves produced by the ASM.
\subsection{Swift}
The BAT on board the \textsl{Swift} spacecraft is a hard X-ray telescope operating in the 15--150\,keV energy band \citep{2005SSRv..120..143B}. The detector is composed of CdZnTe where the detecting area and field of view (FOV) are 5240\,cm$^2$ and 1.4\,sr (half-coded), respectively \citep{2005SSRv..120..143B}. The BAT provides an all-sky hard X-ray survey with a sensitivity of $\sim$2\,mCrab \citep{2005SSRv..120..143B}. The Crab produces $\sim$0.045\,counts s$^{-1}$ over the 14--195\,keV energy band.
We analyzed BAT data obtained during the time period MJD 55152--56141. Light curves were retrieved using \citet{2013ApJS..209...14K}'s extraction of the BAT data available on the NASA GSFC HEASARC website\footnote{http://heasarc.gsfc.nasa.gov/docs/swift/results/transients/}, which includes orbital and daily-averaged light curves. We used the orbital light curves in the 15--50\,keV energy band in our analysis (see Section 3).
\subsection{Suzaku}
The \textsl{Suzaku} observation of 4U 1210-64 took place in 2010 December 23--25 (MJD 55553.16--55555.15) with an exposure time of $\sim$80\,ks (ObsID 405045010). Data, collected using the X-ray Imaging Spectrometer (XIS) and Hard X-ray Detector (HXD) instruments, were reduced and analyzed using the standard criteria defined in the ABC Guide\footnote{http://heasarc.gsfc.nasa.gov/docs/suzaku/analysis/abc/}. These procedures are described below.
\subsubsection{XIS data}
\label{XIS description}
The \textsl{Suzaku} XIS suite consists of four X-ray imaging telescopes each fitted with a CCD chip covering a region 17.$\arcmin$8$\times$17.$\arcmin$8 \citep{2007PASJ...59S..23K, 2007PASJ...59S...1M}. Three of the chips are front-illuminated -- XIS-0, XIS-2, and XIS-3. The energy resolution FWHM in the front-illuminated XIS instruments is $\sim$130\,eV at 6.0\,keV. Sensitive to X-rays ranging from 0.6--10\,keV, the effective area of the front-illuminated XIS instruments is 330\,cm$^2$ at 1.5\,keV. XIS-2 failed in late 2006 due to a charge leak. The back-illuminated XIS instrument, XIS-1, is characterized by a greater sensitivity to X-rays in the energy band between 0.2--6\,keV.
4U 1210-64 was observed in full window mode with a data readout of 8\,s. Data collected using the XIS were reduced and screened using the HEAsoft v.6.13 package and calibration files dated 2013 September 08 (XIS) and 2011 June 30 (XRT) implementing the procedures defined in the \textsl{Suzaku} ABC Guide. The data were reprocessed with the FTOOL \texttt{aepipeline} using the standard criteria to apply the newest calibration and default screening criteria. The XIS exposures in the 3$\times$3 and 5$\times$5 event modes were combined using \texttt{XSELECT}. Circular regions of radius 3.9$\arcmin$ centered on the source and offset from the source were selected to distinguish between photons originating from the source and those originating from the background. The light curves were binned at 16\,s.
Response matrices were generated using the \texttt{xisrmfgen} and \texttt{xissimarfgen} FTOOL packages. Pileup was taken into consideration in the region files, where the central pixels were shown to be affected. As a result, the inner parts of the point spread function (PSF) were removed using two overlapping rectangular shaped regions to reduce pile-up to $\lesssim$1$\%$ using the FTOOLs \texttt{aeattcor2} and \texttt{pileest}. \texttt{Aeattcor2} creates an improved attitude file, which is then applied to the event file. The FTOOL \texttt{pileest} was then used on the improved event file, which provides a rough estimate of the degree of pile-up.
Data in the spectral file produced by \texttt{XSELECT} were further processed using the FTOOL \texttt{GRPPHA}. \texttt{GRPPHA} is designed to define the binning, quality flags and systematic errors of the spectra and used the bad quality flag to further eliminate bad data from the PHA file. Bins were grouped to ensure a mininum of 20 counts in each in the XIS spectra. The spectrum was analyzed using \texttt{XSPEC v12.7.1d}. { To avoid poorly calibrated Si and Au features, only the energy ranges 0.5--1.7\,keV and 2.1--9.2\,keV were considered} \citep{2011ApJ...728...13N}. We will discuss the spectral analysis and results in Section 3.4.
\subsubsection{HXD data}
The HXD is a non-imaging X-ray spectrometer consisting of 64 silicon PIN diodes as well as the Gadolinium silicate crystal (GSO) instruments \citep{2007PASJ...59S...1M}. Our analysis only considers data collected by the PIN diodes because 4U 1210-64 is not bright enough for GSO analysis. The calibration files used in the analysis of the HXD data were dated 2010 December 6.
The HXD-PIN consists of 16 identical (4$\times$4) detector units surrounded by 20 anti-coincidence counters. The usable energy range for the HXD PIN diodes is between 15 and 70\,keV \citep{2007PASJ...59S...1M, 2007PASJ...59S..35T}. The effective area of the HXD instrument is $\sim$160\,cm$^2$ at 20\,keV \citep{2007PASJ...59S...1M, 2007PASJ...59S..35T}. For this specific observation, 4U 1210-64 is not easily detectable at energies exceeding $\sim$30\,keV. As a result, the analyzed part of the spectrum is between 15--30\,keV.
The HXD-PIN spectral data were extracted and reduced using the ``cleaned" event files in the \texttt{hxd/event\_cl} directory and the \texttt{hxdpinxbpi} FTOOL package, respectively. The FTOOL \texttt{hxdpinsbpi} automatically runs the tasks outlined as follows. Good Time Intervals were calculated using the Non X-ray Background (NXB)\footnote{ftp://legacy.gsfc.nasa.gov/suzaku/data/background/pinnxb\_ver2.0\_tuned/} data overlapping in time with the GTI of the observation. The source and NXB spectra are extracted using \texttt{hxdpinxbpi}. The rate in the NXB event file is scaled by a factor of 10 to account for Poisson errors. As a result, the exposure time of the derived background spectra and light curves must be increased by a factor of 10. Since the NXB spectrum does not include the Cosmic X-ray Background (CXB), a simulated CXB spectrum was produced using the parameters determined by \citet{1987IAUS..124..611B}. The total PIN background spectrum is the sum of the NXB and simulated CXB spectra, which were added together using \texttt{addspec}. The net count rates for the NXB and CXB spectra are 0.2336$\pm$0.0006 and 0.0153$\pm$0.0001\,counts s$^{-1}$, respectively. The source spectrum is dead time corrected by $\sim$4--5$\%$.
\section{Results}
\subsection{Long Term Temporal Analysis}
Using data acquired from the ASM, MAXI, and BAT instruments, we produced light curves of 4U 1210-64 to investigate long-term variability of the source (see Figure~\ref{Rebinned Light Curves}). Data produced by ASM, MAXI and BAT span periods of $\sim$16\,years, $\sim$3.3\,years, and $\sim$5\,years respectively.
\begin{figure}
\centerline{\includegraphics[width=3in]{2-3-14_rebinned-lightcurves.pdf}}
\figcaption[2-3-14_rebinned-lightcurves.pdf]{
Long term light curves of 4U 1210-64 produced by BAT in the 15--150\,keV band (top), MAXI in the 2--10\,keV band (middle) and ASM in the 1.5--12\,keV band (bottom) show a two year time overlap between ASM and MAXI and a five\,year time overlap between ASM and BAT. The ASM, MAXI and BAT light curves use 21\,day, 70\,day and 70\,day time bins, respectively. { The times of the PCA} and \textsl{Suzaku} observations are indicated by { the green shaded region} and blue dashed line, respectively. The ASM light curve (bottom) is fit with an asymmetric ``step and ramp" function (see solid red line), which models the long-term behavior of the system.
\label{Rebinned Light Curves}
}
\end{figure}
\subsubsection{ASM Temporal Analysis}
\label{ASM Temporal Analysis}
Three distinct states of the system, two active phases and a quiescent phase, are seen in the ASM
light curve (see Figure~\ref{Rebinned Light Curves}, bottom). To parameterize the states observed
in the system, the light curve was fit using an asymmetric ``step-and-ramp" function (see Figure~
\ref{Rebinned Light Curves}, bottom). The parameters in this model are as follows: the times
corresponding to the start of the transition between the first active state to the quiescent \
state, $T_1$, the start of the transition between the quiescent state and the second active state,
$T_2$, the transition time between state 1 and state 2, $\Delta T_1$, the transition time between
state 2 and state 3, $\Delta T_2$, the count rates during both active states,
$C_{\rm act1}$ and
$C_{\rm act2}$, and the count rate during the quiescent
state, $C_{\rm quies}$. The model parameters
are reported in Table~\ref{Long Term Behavior}. The duration of the quiescent phase was found to
be 2506$^{+36}_{-26}$\,days (6.19$^{+0.10}_{-0.07}$\,yr) using the following equation:
\begin{equation}
\Delta T_{\rm quies}=T_2-(T_1+\Delta T_1)
\end{equation}
\begin{deluxetable}{rrrrr}
\tablecolumns{3}
\tablewidth{0pc}
\tablecaption{Best-fit parameters for the states observed in 4U 1210-64 \label{Long Term Behavior}}
\tablehead{
\colhead{Model Parameter} & \colhead{ASM}}
\startdata
$T_1$ (MJD) & 50703$^{+13}_{-5}$ \\
$\Delta T_1$ (days) & 135$^{+32}_{-22}$ \\
$T_1$+$\Delta T_1$ (MJD) & 50838$^{+35}_{-22}$ \\
$C_{\rm act1}$ (Counts s$^{-1}$) & 0.63$\pm$0.02 \\
$C_{\rm act2}$ (Counts s$^{-1}$) & 0.45$^{+0.02}_{-0.01}$ \\
$T_2$ (MJD) & 53444$^{+11}_{-13}$ \\
$\Delta T_2$ (days) & 109$^{+16}_{-21}$ \\
$T_2$+$\Delta T_2$ (MJD) & 53452$^{+19}_{-25}$ \\
$C_{\rm quies}$ (Counts s$^{-1}$) & 0.136$\pm$0.009 \\
\tableline
$\Delta T_{\rm quies}$ (days) & 2506$^{+36}_{-26}$ \\
\tableline
$\chi^2_\nu$ (dof) & 1.19 (222) \\
\enddata
\label{ASM State Table}
\end{deluxetable}
A power spectrum was used to search for periodicities in temporal data \citep{1982ApJ...263..835S, 2008ATel.1861....1C}. Power spectra, produced using the MAXI and ASM light curves at a time resolution of one dwell (90\,s), show the presence of a 6.7101$\pm$0.0005\,day peak (see Figure~\ref{Power Spectrum}, bottom). The false-alarm-probability \citep[FAP;][]{1982ApJ...263..835S} in the ASM light curve is $\sim$10$^{-9}$. We interpret this peak as the orbital period of the system.
\begin{figure}
\centerline{\includegraphics[width=3in]{6-9-14_power-spectra.pdf}}
\figcaption[6-9-14_power-spectra.pdf]{
Power spectra produced by BAT (top), MAXI (middle) and ASM (bottom) with the 99.9$\%$ and 99.999$\%$ confidence intervals shown. { The BAT (top), MAXI (middle) and ASM (bottom) data are in the 15--50\,keV, 2--20\,keV and 1.5-12\,keV bands, respectively.} The $\sim$70 day precession period of the ISS is also seen in the MAXI power spectrum.
\label{Power Spectrum}
}
\end{figure}
{ For this paper, we use an ephemeris based on the orbital period observed in the power spectrum (see Figure~\ref{Folded Light Curves}, bottom) and time of mid-eclipse: $P$=6.7101$\pm$0.0005\,days and $T0$=MJD 54001.8$\pm$0.2. To parameterize the orbital modulation of the system, the ``dwell-by-dwell" ASM lightcurves were folded using the orbital period.}
\begin{figure}
\centerline{\includegraphics[width=3in]{8-12-13_sinusoidal-lightcurves.pdf}}
\figcaption[8-12-13_sinusoidal-lightcurves.pdf]{
BAT (top), MAXI (middle) and ASM (bottom) lightcurves folded on the orbital period using 20 bins (BAT) and 80 bins (ASM and MAXI). The shaded region indicates the orbital phases of the system for the duration of the \textsl{Suzaku} observation. The sharp dip in flux between phases $\phi{\sim}$-0.04 and $\phi{\sim}$0.03 is interpreted as an eclipse. The solid red line is the sinusoidal fit to the folded light curves using the ephemeris defined in Section 3.1.1.
\label{Folded Light Curves}
}
\end{figure}
\subsubsection{MAXI Temporal Analysis}
\label{MAXI Temporal Analysis}
We confirmed the orbital period using MAXI. We folded the MAXI light curve over the orbital period using the ephemeris described above (see Figure~\ref{Folded Light Curves}, middle), showing that the binned folded light curve strongly agrees with that produced by ASM. Compared to the ASM power spectrum, more statistical noise is apparent in the power spectrum produced using MAXI data (see Figure~\ref{Power Spectrum}, middle). A ``blind search" for the period of the system yields a FAP of $\sim$10$^{-3}$. Considering only the period derived by the ASM data, a single trial search, the FAP is reduced to $\sim$10$^{-6}$.
\subsubsection{BAT Temporal Analysis}
The modulation interpreted as the 6.7 day orbital period was not detected in the power spectrum produced by the BAT instrument (see Figure~\ref{Power Spectrum}, top). Comparisons between the BAT power spectrum and ASM power spectrum indicate that BAT does not have the sensitivity to detect the 6.7 day modulation at the level seen in ASM and MAXI respectively (see Figure~\ref{Power Spectrum}, top). We define the modulation depth as ($Count_{\rm max}$$-$$Count_{\rm min}$)$/Mean Count$.
We folded the BAT light curve on the orbital period using the ephemeris defined by the ASM (see Section 3.1.1 and Figure~\ref{Folded Light Curves}, top), showing that the binned folded light curve is consistent with that produced by ASM and MAXI with a possible indication of an eclipse centered at $\phi=$0.
\subsection{Eclipse Profile}
\label{Eclipse Profile}
The folded light curves show the presence of a sharp dip between orbital phases $\phi{\sim}$-0.04 and $\phi{\sim}$0.03, which is suggestive of an eclipse. { The source emission does not reach 0\,counts s$^{-1}$. We interpret this dip as an eclipse since the feature is persistent over many years of data. The feature is seen at the same orbital phase in States 1 and 3 (see Table~\ref{ASM State Table}), which are separated by $\gtrsim$6\,years. The rapid ingress and egress requires obscuration by clearly defined boundaries that are suggestive of an object such as the mass donor in the system.}
The eclipse was modeled using a symmetric ``step and ramp" function (see Figure~\ref{Eclipse Plot}) where the intensities are assumed to remain constant before ingress, during eclipse and after egress and follow a linear trend during the ingress and egress transitions. The parameters in this model are as follows: the phases corresponding to the start of ingress as well as egress, $\phi_{\rm ing}$ and $\phi_{\rm eg}$, the duration of ingress and egress, $\Delta{\phi}$, the count rates before ingress and after egress, $C$, and the count rate during eclipse, $C_{\rm ecl}$. The symmetric nature of the model ensures that the duration of ingress and egress, $\Delta{\phi}$, as well as the count rate before ingress and after egress, $C$, are equal. The eclipse model parameters are reported in Table~\ref{Step and Ramp}. The eclipse duration and mid-eclipse are calculated using Equations~\ref{Half Angle Equation} and~\ref{Mid Eclipse Equation}.
\begin{equation}
\label{Half Angle Equation}
\Delta{\phi}_{\rm ecl}=\phi_{\rm egr}-(\phi_{\rm ing}+\Delta{\phi})
\end{equation}
\begin{equation}
\label{Mid Eclipse Equation}
\phi_{\rm mid}=\frac{1}{2}(\phi_{\rm egr}+(\phi_{\rm ing}+\Delta{\phi}))
\end{equation}
Since the Suzaku observation begins at MJD 55553.1, we express the time of mid-eclipse, $T_{\rm mid}$, at an epoch closest to the Suzaku observation (see Table~\ref{Step and Ramp}). The eclipse duration, time of mid-eclipse, and eclipse half-angle ($\Delta{\phi}_{\rm ecl}$$\times$180$\degr$) from fitting the ASM and MAXI folded light curves are reported in Table~\ref{Step and Ramp}. The fits of both the ASM and MAXI data, which use 150 and 88 bins respectively, indicate that the mid-eclipse times and eclipse durations are in agreement at the 1 $\sigma$ level. While all model parameters were free for the ASM and MAXI fits, the phases corresponding to the start of ingress as well as egress and the transition duration in the BAT fit, which uses 96 bins, were frozen to the weighted average of the ASM and MAXI values.
\begin{figure}
\centerline{\includegraphics[width=3in]{6-11-14_eclipse-profile.pdf}}
\figcaption[6-11-14_eclipse-profile.pdf]{
BAT (top), MAXI (middle) and ASM (bottom) folded light curves are fit with a symmetric ``step and ramp" function, which models the eclipse. { The BAT (top), MAXI (middle) and ASM (bottom) data are in the 15--50\,keV, 2--20\,keV and 1.5-12\,keV bands, respectively.}
\label{Eclipse Plot}
}
\end{figure}
\begin{deluxetable}{ccccc}
\tablecolumns{3}
\tablewidth{0pc}
\tablecaption{4U 1210-64 Eclipse Model Parameters}
\tablehead{
\colhead{Model Parameter} & \colhead{ASM} & \colhead{MAXI} & \colhead{Combined} & \colhead{BAT$^a$} \\
\colhead{} & \colhead{(1.5--12\,keV)} & \colhead{(2--20\,keV)} & \colhead{} & \colhead{(15-50\,keV)}}
\startdata
$\phi_{\rm ing}$ & -0.047$^{+0.004}_{-0.003}$ & -0.044$\pm$0.002 & -0.045$\pm$0.002 & \nodata \\
$\Delta{\phi}$ & 0.014$^{+0.001}_{-0.006}$ & 0.013$\pm$0.002 & 0.014$^{+0.001}_{-0.002}$ & \nodata \\
$C$ & 0.54$\pm$0.01$^b$ & 1.97$\pm$0.03$^c$ & \nodata & 5.5$\pm$0.4$^d$ \\
$\phi_{\rm egr}$ & 0.033$\pm$0.005 & 0.031$\pm$0.002 & 0.031$\pm$0.002 & \nodata \\
$C_{\rm ecl}$ & 0.34$^{+0.02}_{-0.03}$$^b$ & 1.11$\pm$0.07$^c$ & \nodata & 1.0$\pm$1.4$^d$ \\
\tableline
$\Delta{\phi}_{\rm ecl}$ & 0.066$^{+0.007}_{-0.008}$ & 0.062$\pm$0.004 & 0.062$\pm$0.003 & \nodata \\
$T_{\rm mid}^e$ & 55553.099$\pm$0.004 & 55553.098$\pm$0.002 & 55553.098$\pm$0.002 & \nodata \\
$\Theta_{\rm e}^f$ & 11.9$^{+1.3}_{-1.5}$ & 11.1$^{+0.6}_{-0.7}$ & 11.2$\pm$0.6 & \nodata \\
\tableline
$\chi^2_\nu$ (dof) & 0.92(146) & 1.21(84) & \nodata & 0.93(96) \\
\enddata
\tablecomments{\\*
$^a$ The $\phi_{{\rm ing}}$, $\Delta{\phi}$, $\phi_{\rm egr}$ parameters in the BAT fit are frozen to the weighted average of the ASM and MAXI values. \\*
$^b$ Units are counts s$^{-1}$. \\*
$^c$ Units are 10$^{-2}$ counts cm$^{-2}$ s$^{-1}$. \\*
$^d$ Units are 10$^{-5}$ counts cm$^{-2}$ s$^{-1}$. \\*
$^e$ Units are MJD. \\*
$^f$ Units are degrees.}
\label{Step and Ramp}
\end{deluxetable}
\subsection{Short Term Temporal Analysis}
\subsubsection{Suzaku}
\label{Suzaku Short Term Temporal Analysis}
The two-day \textsl{Suzaku} observation in 2010 Dec. took place during the phase of transition from the minimum to the maximum of the 6.7\,day modulation (see Figure~\ref{Folded Light Curves}). We binned the \textsl{Suzaku} light curves to a resolution of 16\,s to investigate short-term variability in the system, shown in the light curve produced by the sum of the XIS-0 and XIS-3 light curves (see Figure~\ref{Suzaku Light Curve}).
\begin{figure}
\centerline{\includegraphics[width=3in]{12-12-2013_hardness-ratio.pdf}}
\figcaption[12-12-2013_hardness-ratio.pdf]{
The weighted average of the \textsl{Suzaku} XIS0 and XIS3 light curves using bin sizes of 16\,s (top). The top axis indicates the orbital phase of the system. The shaded region indicates the egress start and end times calculated by the ``step-and-ramp" function. The hardness ratio (bottom), which uses bin sizes of 1000\,s, is defined as ($C_{\rm hard}$-$C_{\rm soft}$)/($C_{\rm hard}$+$C_{\rm soft}$), where the soft and hard energy bands are 0.5--4\,keV and 4--10\,keV, respectively.
\label{Suzaku Light Curve}
}
\end{figure}
The two-day \textsl{Suzaku} observation revealed large variations in flux indicative of significant variability beyond the orbital modulation. The modulation depth between the peak in the light curve and the mean count rate is on the order of 140$\%$.
We divided the light curve into two energy bands where the soft band is defined between energies 0.5--4\,keV, characterized by the count rate $C_{\rm soft}$, and the hard band is between 4--10\,keV, characterized by the count rate $C_{\rm hard}$. Using the definition in Equation~\ref{Hardness Equation}, we produced a hardness ratio (see Figure~\ref{Suzaku Light Curve}, bottom) binned to a resolution of $\sim$1000\,s, where a soft spectrum is indicated by negative values and a hard spectrum is indicated by positive. The hardness ratio was also plotted against the count rate in the full energy band of 0.5--10\,keV to search for a correlation between the hardness ratio and source intensity (see Figure~\ref{Hardness Correlation}). { Throughout the paper, we use the weighted Pearson correlation coefficient, r \citep[e.g][]{2003drea.book.....B}.} { We only take the data observed out-of-eclipse into account since the phenomenology is different from the data observed during eclipse (see Table~\ref{Suzaku Spectral Parameters}). Out-of-eclipse, we} found a positive correlation between the hardness ratio and source intensity (r=0.69, p$\lesssim$10$^{-6}$), which will be interpreted in Section~\ref{What is the origin of the variability in the high-state?}.
\begin{equation}
\label{Hardness Equation}
HR=(C_{\rm hard}-C_{\rm soft})/(C_{\rm hard}+C_{\rm soft})
\end{equation}
\begin{figure}
\centerline{\includegraphics[width=3in,angle=-90]{february18-2014_hardness_correlation.pdf}}
\figcaption[february18-2014_hardness_correlation.pdf]{
The hardness ratio as defined in { Figure 5} vs. the full energy band of 0.5--10\,keV of the \textsl{Suzaku} light curve. The red dashed line indicates the best linear fit { to the points outside the eclipse}. { The blue points indicate the data collected during eclipse.}
\label{Hardness Correlation}
}
\end{figure}
\subsubsection{PCA}
\label{PCA Temporal}
\begin{figure}
\centerline{\includegraphics[width=3in]{may6-2014_PCA_hardness-MJD.pdf}}
\figcaption[may6-2014_PCA_hardness-MJD.pdf]{
Long-term light curve of 4U1210-64 during the RXTE PCA campaign (top). We calculated the average and standard deviation of the PCU2 count rate for each spacecraft orbit. Also indicated (shaded regions) are the eclipse intervals { due to the companion. Hardness ratios of the PCA light curves (middle and bottom), using bin sizes of 0.2\,d., where the energy bands for $C_{\rm 1}$, $C_{\rm 2}$ and $C_{\rm 3}$ are 2.5--6\,keV, 6--10\,keV and 10--20\,keV, respectively.}
\label{PCA Light Curve}
}
\end{figure}
{ To present an overview of the RXTE observations, we calculated the average, background-subtracted count rate of 4U 1210-64 during each spacecraft orbit and plotted this against time in Figure~\ref{PCA Light Curve}. We divided the PCA light curves into three energy bands defined between between energies 2.5--6\,keV, 6--10\,keV and 10--20\,keV, which are characterized by count rates $C_{\rm 1}$, $C_{\rm 2}$ and $C_{\rm 3}$, respectively. Using the definition in Equation~\ref{Hardness Equation}, we produced hardness ratios binned to a resolution of $\sim$0.2\,d (see Figure~\ref{PCA Light Curve}, middle and bottom). To search for correlations between the hardness ratios in each energy band, we produced a color-color diagram (see Figure 8) between the two hardness ratios to examine the correlation between them. A strong correlation was found in the color-color diagram (r=0.79,p$\lesssim$10$^{-6}$.)}
\begin{figure}
\centerline{\includegraphics[width=3in,angle=-90]{may6-2014_PCA-color.pdf}}
\figcaption[may6-2014_PCA-color.pdf]{
PCA color-color diagram comparing soft color vs. hard color as defined in Figure~\ref{PCA Light Curve}. The red line indicates the best linear fit to the points outside the eclipse. The blue points indicate the data collected during eclipse.
\label{PCA Color Color Diagram}
}
\end{figure}
{ We performed a Fourier transform of all background-subtracted \textsl{RXTE} PCA Standard2 mode data together, up to a frequency of 100\,cycles per day \citep{1982ApJ...263..835S}. The power spectrum is dominated by low frequency (see Figure~\ref{Low Frequency}), ``red" noise, and the orbital period was not detected in this data set. This is likely because of a patchy orbital phase coverage.}
Note that Scargle's method is strictly correct only in judging the FAP of the highest peak mixed in with otherwise frequency-independent, ``white", noise \citep{1982ApJ...263..835S}. This is clearly not the case here. { To improve our search for a spin period, we analyzed the relationship between power and frequency in log-log space to estimate and remove the amount of low-frequency red noise present in the power spectrum \citep{2005A&A...431..391V}. A quadratic fit was found to give a reasonable approximation to the continuum noise level. We subtracted the quadratic fit from the logarithm of the power spectrum along with a constant value of 0.25068 to account for the bias due to the $\chi^2$ distribution of the power spectrum \citep{2005A&A...431..391V}}. { The only statistically significant feature in the power spectrum is an artifact caused by a group of peaks around 15 cycles per day near the spacecraft orbital period (see Figure~\ref{Low Frequency}).}
\begin{figure}
\centerline{\includegraphics[width=3in,angle=-90]{4-8-14_PCA_powerspectrum.pdf}}
\figcaption[4-8-14_PCA_powerspectrum.pdf]{
{ Scargle's power spectrum of the PCA data up to a frequency of 100 cycles per day with the estimated continuum noise component removed (top). The uncorrected power spectrum (bottom) is dominated by strong low frequency variability. The orbital period (see Sections~\ref{ASM Temporal Analysis} and~\ref{MAXI Temporal Analysis}) is indicated by the blue dashed line. The red short dashed line is the power corresponding to 1$\%$ FAP, and the long dashed line 0.1$\%$ FAP.}
\label{Low Frequency}
}
\end{figure}
We created light curves binned to 10\,ms for each GoodXenon event mode file, covering no more than a single spacecraft orbit. { The estimated low-frequency red noise was removed using the procedures in \citet{2005A&A...431..391V}, similar to what we describe above. A function that is quadratic for frequencies below $\sim$10$^{-2}$\,Hz and constant above this was found to give a reasonable approximation to the continuum noise level found in the log-log plot between power and frequency.} The highest peak in the resulting power spectra (see Figure~\ref{Short Power}) was often found at the lowest frequency ($<$0.01\,mHz), presumably resulting from both source and background variability on $>$100\,s time scales. In 5 cases, additional higher frequency ($>$10\,Hz), apparently significant (FAP less than 0.1$\%$) peaks were also found. However, these turned out to be related to the low frequency peak: the removal of the low frequency sinusoid also removed these peaks. We conclude that these were artifacts, created by some (unknown) combination of the sampling pattern, the precise characteristics of the low frequency variability, and possibly also the numerical limitations of the particular implementation of Scargle's algorithm that we used \citep{1982ApJ...263..835S}. Other than these, the strongest high frequency peaks had FAP between 1$\%$ and 0.1$\%$ (note that the number of trials for each Fourier transform is taken into account, but not the fact that we analyzed 84 independent light curves), and none of these candidate frequencies repeated in multiple observations. We conclude that we did not detect the spin period of the compact object in 4U 1210-64. Scaling from the amplitudes of the highest peaks, we estimate that a sinusoidal modulation with an amplitude of 8$\%$ of the mean flux would have been detectable at 99.9$\%$ significance.
\begin{figure}
\centerline{\includegraphics[width=3in]{april8-2014_Scargle-powerspectra.pdf}}
\figcaption[april8-2014_Scargle-powerspectra.pdf]{
Typical PCA Power spectra for Obs ID 99009-01-01-17 (top), Obs ID 99009-01-01-22 (second from top), Obs ID 99009-01-23-00 (third panel) and Obs ID 99009-01-03-02 (bottom). { The frequency range is from 5.73$\times$10$^{-4}$--50\,Hz.} The long dashed line is the power corresponding to 1$\%$ FAP, and the short dashed line the 0.1$\%$ FAP. The highest peak in the Obs ID 99009-01-01-22 power spectrum is found to be between the 99$\%$ and 99.9$\%$ confidence intervals (second from top).
\label{Short Power}
}
\end{figure}
\subsection{Spectral Analysis}
{ The X-ray spectral data from the \textsl{Suzaku} and PCA observations of 4U 1210-64 were analyzed using the package \texttt{XSPEC v12.8.0}. We made use of the \texttt{XSPEC} convolution model \texttt{cflux} to calculate the fluxes and associated errors of 4U 1210-64.}
\subsubsection{Suzaku Spectral Analysis}
\label{Suzaku Spectral Analysis}
To fit the \textsl{Suzaku} spectra, we used several models: a power law, thermal bremsstrahlung, a power law modified with a high energy cutoff (see Figure~\ref{Suzaku Spectra}), and emission due to collisionally-ionized diffuse gas \citep[\texttt{APEC} in \texttt{XSPEC},][]{2012ApJ...756..128F}. All models were modified by a partially covering absorber in addition to an absorber that fully covers the source using the \citet{1992ApJ...400..699B} cross sections and \citet{2000ApJ...542..914W} abundances.
{ The model that provides a good fit to the data is a power law modified by a high energy cutoff (reported in Table~\ref{Suzaku Spectral Parameters}). We find that the neutral hydrogen column densities for fully covered and partially covering absorption are $N_{\rm H}$=0.70$\pm$0.01$\times$10$^{22}$ and $N_{\rm H}$=6.7$^{+0.3}_{-0.4}\times$10$^{22}$\,atoms cm$^{-2}$ respectively with a partial covering fraction of 0.36$\pm$0.03. The measured values of the fully covered absorber are comparable to the Galactic H I values reported by the Leiden/Argentine/Bonn survey \citep{2005A&A...440..775K} and in the review by \citet{1990ARA&A..28..215D}, which are 8.16$\times$10$^{21}$ and 9.37$\times$10$^{21}$\,atoms cm$^{-2}$, respectively. Therefore, we assume that the fully covered absorber is interstellar in origin unless otherwise noted (see Section~\ref{What is the origin of the variability in the high-state?} for treatment in eclipse). A good fit does not require an additional blackbody to the power law component, even though such a soft excess was seen in the \textsl{INTEGRAL} data by \citet{2010ApJ...511..A48}. Furthermore, we do not detect any cyclotron lines in the Suzaku spectra.}
Emission features in the Fe K$\alpha$ region were detected at 6.4\,keV, 6.7\,keV and 6.97\,keV; which were modeled using a Gaussian centered on the peak of the lines (see Figure~\ref{Emission Residuals}). We interpret these features as Fe K$\alpha$, Fe XXV K$\alpha$ and Fe XXVI K$\alpha$ respectively (see Figure~\ref{Emission Residuals}). In addition, an emission line was observed at 2.6\,keV, which we interpret as S XVI K$\alpha$ (see Figure~\ref{Emission Residuals}).
\begin{figure}
\centerline{\includegraphics[width=2.5in,angle=-90]{11-27-13_cutoffpl-chisq.pdf}}
\figcaption[11-27-13_cutoffpl-chisq.pdf]{
One of the best fit models for 4U 1210-64 where XIS0$+$3, XIS1 and HXD/PIN are indicated by the black, red and green data/models, respectively. This consists of a continuum comprised of a power law with a high energy cutoff and four emission lines composed of S XVI K$\alpha$, Fe K$\alpha$ , Fe XXV K$\alpha$ and Fe XXVI K$\alpha$ . The continuum is absorbed by partial covering absorption and fully covering absorption.
\label{Suzaku Spectra}
}
\end{figure}
\begin{deluxetable}{ccc}
\tablecolumns{3}
\tablewidth{0pc}
\tablecaption{X-ray spectral parameters for 4U 1210-64}
\tablehead{
\colhead{Model Parameter} & \colhead{Cutoff power law} & \colhead{{ Cutoff} power law} \\
\colhead{} & \colhead{{ Out-of-Eclipse}} & \colhead{Eclipse}}
\startdata
$\chi_\nu^2$ (dof) & { 1.08 (2553)} & 0.81 (382) \\
Cutoff Energy (keV) & { 5.5$\pm$0.1} & \nodata \\
Folding Energy (keV) & { 12$\pm$1} & \nodata \\
Phabs $N_{\rm H}$ ($\times$10$^{22}$\,atoms cm$^{-2}$)& { 0.70$\pm$0.01} & 0.94$\pm$0.08 \\
Pcfabs $N_{\rm H}$ ($\times$10$^{22}$\,atoms cm$^{-2}$)& 6.7$^{+0.3}_{-0.4}$ & 11$^{+2}_{-1}$ \\
Covering Fraction & 0.36$\pm$0.03 & 0.80$^{+0.04}_{-0.05}$ \\
$\Gamma$ & { 1.80$^{+0.04}_{-0.05}$} & 2.9$\pm$0.2 \\
Normalization ($\times$10$^{-2}$) & { 2.9$\pm$0.2} & 1.8$^{+0.7}_{-0.5}$ \\
\tableline
S XVI Energy (keV) & { 2.62$\pm$0.02} & 2.58$\pm$0.05 \\
S XVI Width ($\sigma_{\rm SXVI}$) & { 0.1$^b$} & 0.1$^b$ \\
Normalization ($\times$10$^{-3}$\,photons cm$^{-2}$ s$^{-1}$) & { 0.11$\pm$0.02} & 0.11$^{+0.06}_{-0.05}$ \\
S XVI EQW (eV) & { 20$\pm$3} & 51$^{+26}_{-22}$ \\
S XVI Flux ($\times$10$^{-13}$\,erg cm$^{-2}$ s$^{-1}$) & { 0.54$\pm$0.09} & 0.4$\pm$0.2 \\
\tableline
Fe I Energy (keV) & { 6.39$\pm$0.01} & 6.4$^a$ \\
Fe I Width ($\sigma_{\rm Fe I}$) & { 0.1$^b$} & 0.1$^b$ \\
Normalization ($\times$10$^{-3}$\,photons cm$^{-2}$ s$^{-1}$) & { 0.093$\pm$0.007} & 0.009$^c$ \\
Fe I EQW (eV) & { 77$^{+7}_{-5}$} & 122$^c$ \\
Fe I Flux ($\times$10$^{-13}$\,erg cm$^{-2}$ s$^{-1}$) & { 1.02$\pm$0.08} & 0.7$^c$ \\
\tableline
Fe XXV Energy (keV) & { 6.684$\pm$0.008} & 6.68$\pm$0.04 \\
Fe XXV Width ($\sigma_{\rm Fe XXV}$) & { 0.1$^b$} & 0.1$^b$ \\
Normalization ($\times$10$^{-3}$\,photons cm$^{-2}$ s$^{-1}$) & { 0.185$\pm$0.008} & 0.043$\pm$0.009 \\
Fe XXV EQW (eV) & { 144$\pm$7} & 392$^{+160}_{-70}$ \\
Fe XXV Flux ($\times$10$^{-13}$\,erg cm$^{-2}$ s$^{-1}$) & { 1.25$\pm$0.06} & 1.5$\pm$0.3 \\
\tableline
Fe XXVI Energy (keV) & { 6.970$^{+0.007}_{-0.005}$} & 6.98$\pm$0.05 \\
Fe XXVI Width ($\sigma_{\rm Fe XXVI}$) & { 0.1$^b$} & 0.1$^b$ \\
Normalization ($\times$10$^{-3}$\,photons cm$^{-2}$ s$^{-1}$) & { 0.219$^{+0.008}_{-0.009}$} & 0.035$\pm$0.009 \\
Fe XXVI EQW (eV) & { 199$^{+9}_{-8}$} & 319$^{+160}_{-74}$ \\
Fe XXVI Flux ($\times$10$^{-13}$\,erg cm$^{-2}$ s$^{-1}$) & { 1.16$\pm$0.05} & 1.5$\pm$0.4 \\
\tableline
Absorbed Flux ($\times$10$^{-10}$\,erg cm$^{-2}$ s$^{-1}$) & { 1.024$\pm$0.003} & 0.108$\pm$0.004 \\
Unabsorbed Flux ($\times$10$^{-10}$\,erg cm$^{-2}$ s$^{-1}$) & { 1.73$^{+0.06}_{-0.05}$} & 0.16$\pm$0.01 \\
\enddata
\tablecomments{\\*
$^a$ The energy is frozen because we can only obtain an upper limit. \\*
$^b$ The natural width was frozen to 0.1\,keV, the resolution of the XIS instruments. \\*
$^c$ The upper limit for the parameters (90$\%$ confidence interval) associated with the Fe I line is reported during eclipse.}
\label{Suzaku Spectral Parameters}
\end{deluxetable}
We investigated temporal dependence of the spectral parameters, comparing the parameters during eclipse to those out-of-eclipse. The out-of-eclipse region was subdivided into several 10\,ks intervals to further investigate the temporal variability of the spectral parameters. { Since 4U 1210-64 is not easily detectable in the HXD-PIN for much of the observation, we only considered the XIS spectra for the analysis of time dependent changes in spectral parameters. A good fit to the spectra does not require a high-energy cutoff although it is needed for the time-averaged out-of-eclipse spectrum for which we analyzed the well-exposed HXD-PIN spectrum along with the XIS data. We choose to model the spectra using a power law with a high-energy cutoff frozen at the values of the time-averaged spectra (see Table~\ref{Time Dependent Cutoff Spectrum}).} The Fe I line was not detected during eclipse (see Table~\ref{Suzaku Spectral Parameters}). We therefore derived an upper limit for the strength of the line.
\begin{deluxetable}{cccccccccccccc}
\rotate
\tablecolumns{14}
\tabletypesize{\small}
\tablewidth{0in}
\tablecaption{Spectral parameters of 4U 1210-64 from the \textsl{Suzaku} observation using the power law continuum with the High Energy Cutoff$^a$}
\tablehead{
\colhead{Time$^b$} & \colhead{Phabs} & \colhead{Pcfabs} & \colhead{Cvr} & \colhead{$\Gamma$} & \colhead{EQW} & \colhead{EQW} & \colhead{EQW} & \colhead{EQW} & \colhead{Flux} & \colhead{$F_{\rm unabs}$} & \colhead{$\chi^2_\nu$} \\
\colhead{(ks)} & \colhead{$N_{\rm H}$} & \colhead{$N_{\rm H}$} & \colhead{} & \colhead{} & \colhead{S XVI} & \colhead{Fe I} & \colhead{Fe XXV} & \colhead{Fe XXVI} & \colhead{Ratio$^c$} & \colhead{(10$^{-10}$ erg cm$^{-2}$ s$^{-1}$)} & \colhead{(d.o.f)} \\
\colhead{} & \colhead{(10$^{22}$ cm$^{-2}$)} & \colhead{(10$^{22}$ cm$^{-2}$)} & \colhead{} & \colhead{} & \colhead{(eV)} & \colhead{(eV)} & \colhead{(eV)} & \colhead{(eV)} & \colhead{} & \colhead{}}
\startdata
0-22.5 & 0.93$\pm$0.08 & 11$^{+2}_{-1}$ & 0.80$^{+0.04}_{-0.05}$ & 2.9$\pm$0.2 & 51$^{+26}_{-22}$ & 70$^d$ & 420$^{+162}_{-76}$ & 364$^{+156}_{-83}$ & 1.0$\pm$0.3 & 0.53$\pm$0.03 & 0.82(382) \\
30-40 & 0.76$\pm$0.06 & 8$\pm$2 & 0.49$^{+0.07}_{-0.08}$ & 1.9$\pm$0.1 & 48$^e$ & 118$^{+50}_{-36}$ & 231$^{+53}_{-40}$ & 147$^{+47}_{-37}$ & 1.3$^{+0.5}_{-0.4}$ & 1.0$\pm$0.1 & 0.92(499) \\
40-50 & 0.71$^{+0.04}_{-0.05}$ & 4$\pm$1 & 0.33$\pm$0.05 & 1.69$^{+0.07}_{-0.06}$ & 16$^e$ & 55$^{+19}_{-20}$ & 100$^{+28}_{-25}$ & 174$^{+39}_{-32}$ & 1.2$\pm$0.3 & 2.1$\pm$0.1 & 0.96(1146) \\
50-60 & 0.69$\pm$0.03 & 5.7$\pm$0.9 & 0.38$\pm$0.04 & 1.76$\pm$0.05 & 24$^{+7}_{-8}$ & 67$^{+14}_{-13}$ & 100$^{+18}_{-15}$ & 171$^{+21}_{-19}$ & 1.0$\pm$0.2 & 2.5$\pm$0.1 & 0.97(2030) \\
60-70 & 0.75$\pm$0.02 & 6.3$\pm$0.7 & 0.43$\pm$0.04 & 1.94$\pm$0.05 & 18$\pm$8 & 65$^{+16}_{-13}$ & 121$^{+19}_{-16}$ & 193$^{+23}_{-19}$ & 1.0$\pm$0.2 & 3.0$^{+0.2}_{-0.1}$ & 0.92(2094) \\
70-80 & 0.76$\pm$0.03 & 6.6$\pm$0.9 & 0.41$\pm$0.04 & 1.97$\pm$0.06 & 18$\pm$9 & 50$^{+17}_{-15}$ & 205$^{+26}_{-21}$ & 206$^{+27}_{-25}$ & 1.2$\pm$0.2 & 2.1$\pm$0.1 & 0.97(1572) \\
80-90 & 0.68$^{+0.04}_{-0.05}$ & 5$\pm$2 & 0.32$\pm$0.07 & 1.84$\pm$0.09 & 23$^{+12}_{-13}$ & 59$^{+26}_{-22}$ & 158$^{+33}_{-26}$ & 263$^{+47}_{-39}$ & 1.2$\pm$0.3 & 1.3$\pm$0.1 & 0.96(1103) \\
90-100 & 0.72$^f$ & 6$\pm$1 & 0.35$^{+0.08}_{-0.09}$ & 1.95$^{+0.08}_{-0.09}$ & 42$^{+16}_{-15}$ & 129$^{+31}_{-29}$ & 203$^{+37}_{-31}$ & 264$^{+46}_{-41}$ & 1.0$^{+0.3}_{-0.2}$ & 0.82$^{+0.07}_{-0.06}$ & 0.94(809) \\
100-110 & 0.67$\pm$0.05 & 12$\pm$3 & 0.3$\pm$0.1 & 1.6$\pm$0.1 & 26$^{+17}_{-18}$ & 108$^{+33}_{-27}$ & 207$^{+46}_{-31}$ & 174$^{+45}_{-31}$ & 1.1$^{+0.4}_{-0.3}$ & 1.2$\pm$0.1 & 0.91(631) \\
110-120 & 0.70$\pm$0.04 & 9$\pm$1 & 0.42$^{+0.06}_{-0.07}$ & 1.69$\pm$0.08 & 15$^{+12}_{-13}$ & 88$^{+23}_{-20}$ & 110$^{+25}_{-20}$ & 168$^{+32}_{-27}$ & 1.2$\pm$0.3 & 2.8$\pm$0.2 & 0.96(1022) \\
120-130 & 0.66$^{+0.04}_{-0.05}$ & 5$\pm$2 & 0.27$\pm$0.07 & 1.73$\pm$0.08 & 20$^{+11}_{-12}$ & 31$^{+21}_{-20}$ & 155$^{+31}_{-25}$ & 215$^{+38}_{-30}$ & 1.0$\pm$0.3 & 2.1$^{+0.2}_{-0.1}$ & 0.96(1021) \\
130-140 & 0.70$\pm$0.03 & 9$\pm$1 & 0.37$\pm$0.07 & 1.86$^{+0.09}_{-0.08}$ & 38$^{+13}_{-12}$ & 26$^{+21}_{-20}$ & 247$^{+37}_{-28}$ & 206$^{+34}_{-25}$ & 1.0$^{+0.3}_{-0.2}$ & 1.4$\pm$0.1 & 1.07(1036) \\
140-150 & 0.66$\pm$0.03 & 7$\pm$1 & 0.30$\pm$0.06 & 1.66$\pm$0.06 & 17$^e$ & 101$^{+18}_{-17}$ & 126$^{+21}_{-16}$ & 190$^{+24}_{-23}$ & 1.1$\pm$0.2 & 1.9$\pm$0.1 & 0.95(1616) \\
150-160 & 0.72$\pm$0.03 & 7$\pm$1 & 0.40$^{+0.06}_{-0.07}$ & 1.89$\pm$0.08 & 16$\pm$12 & 78$^{+23}_{-20}$ & 198$^{+34}_{-25}$ & 231$^{+33}_{-28}$ & 1.1$^{+0.3}_{-0.2}$ & 1.4$\pm$0.1 & 0.98(1108) \\
160-170 & 0.63$^f$ & 16$^{+13}_{-5}$ & 0.23$^{+0.08}_{-0.09}$ & 1.64$^{+0.08}_{-0.07}$ & 61$\pm$16 & 161$^{+34}_{-30}$ & 164$^{+32}_{-27}$ & 242$^{+44}_{-36}$ & 1.1$\pm$0.3 & 0.70$^{+0.05}_{-0.04}$ & 0.99(786) \\
\tableline
\enddata
\tablecomments{The best-fit parameters for the eclipse and out-of-eclipse, which is subdivided into 10 ks time intervals. \\*
$^a$ The cutoff and folding energies of the high energy cutoff are frozen to the best fit values of the out-of-eclipse spectrum: 5.5\,keV and 12.0\,keV, respectively. \\*
$^b$ Time is relative to the start of the \textsl{Suzaku} observation MJD 55553.16. \\*
{ $^c$ The ratio of the strengths of the Fe XXV line with respect to the Fe XXVI line.} \\*
$^d$ The upper limit for the EQW of the Fe I line at the 90$\%$ confidence interval is reported during eclipse. \\*
$^e$ The upper limit for the EQW of the S XVI line at the 90$\%$ confidence interval is reported at the time intervals 30--40\,ks, 40--50\,ks, and 140--150\,ks. \\*
$^f$ The fully covered absorber is frozen to the value that was extracted using the power law model.}
\label{Time Dependent Cutoff Spectrum}
\end{deluxetable}
\begin{figure}
\centerline{\includegraphics[width=3.5in]{august20-2013_Total-Emission-Lines.pdf}}
\figcaption[august20-2013_Total-Emission-Lines.pdf]{
The XIS-0$+$3 spectrum of 4U 1210-64 in the 2.3--3.5\,keV (top) and 5.0--7.8\,keV bands (third panel) to illustrate the S XVI and Fe K$\alpha$ emission lines along with the best fit model. The normalization of the lines was set to 0 in the model. Residuals are plotted in the second from top and bottom panels.
\label{Emission Residuals}
}
\end{figure}
\begin{figure}
\centerline{\includegraphics[width=3.5in]{3-13-14_highecut_FeKalpha-flux.pdf}}
\figcaption[3-13-14_highecut_FeKalpha-flux.pdf]{
Flux of the Fe I (top), Fe XXV (middle) and Fe XXVI (bottom) lines vs. the unabsorbed continuum flux in the 7.1--9.0\,keV band in logarithmic units. An upper limit at the 90$\%$ confidence interval is shown for the Fe I line during eclipse. The red and blue lines indicate the best fit for constant and power law models respectively. The correlation coefficients, r, are as follows: r=0.82, r=0.81 and r=0.87 for the Fe I, Fe XXV and Fe XXVI lines, respectively. Using a logarithmic parameter space, we measured the slope of the Fe K$\alpha$ line flux versus continuum flux--m=1.0$\pm$0.1, m=0.6$\pm$0.1 and m=0.7$\pm$0.1 for the Fe I, Fe XXV and Fe XXVI lines, respectively. { The blue points indicate the data collected during eclipse.}
\label{Flux Flux}
}
\end{figure}
In order to place constraints on the origin of the Fe K$\alpha$ lines, we compared the flux of the
Fe K$\alpha$ lines with the unabsorbed continuum flux in the 7.1--9.0\,keV band (see Figure~\ref
{Flux Flux}). The flux of the Fe I line is found to decrease by a factor of $\sim$15 during
eclipse (see Table~\ref{Time Dependent Cutoff Spectrum}). Additionally, we found that the flux of
Fe XXV and Fe XXVI decreases by a factor of approximately 3 during the eclipse phase (see Table~
\ref{Time Dependent Cutoff Spectrum}). Within measurement errors, the flux of the Fe K$\alpha$
lines { observed out-of-eclipse} was found to follow the flux of the continuum. The
correlation coefficients, r, are as follows: r=0.82, r=0.81 and r=0.87 for the Fe I, Fe XXV and Fe
XXVI lines, respectively. Using a logarithmic parameter space, we measured the slope of the Fe K$
\alpha$ line flux { observed out-of-eclipse} versus continuum flux--m=1.0$\pm$0.1, m=0.6$\pm
$0.1 and m=0.7$\pm$0.1 for the Fe I, Fe XXV and Fe XXVI lines, respectively.
{ To constrain the state of the plasma, we calculated the flux ratio between the Fe XXV and Fe XXVI emission features. We defined the flux ratio as the flux of the Fe XXV line divided by the flux of the Fe XXVI line. No change was found in the flux ratio between the Fe XXV and Fe XXVI emission features and the continuum flux in the 7.1--9.0\,keV band (see Table~\ref{Time Dependent Cutoff Spectrum}).}
In contrast to the Fe K$\alpha$ lines, the S XVI K$\alpha$ line was not consistently detected throughout the duration of the \textsl{Suzaku} observation (see Table~\ref{Time Dependent Cutoff Spectrum}). During the time intervals specified in Table~\ref{Time Dependent Cutoff Spectrum}, we calculated an upper limit of the EQW and flux of the S XVI line at the 90$\%$ confidence interval. We also searched for correlations between the continuum spectral parameters (the power law { with the high energy cutoff} modified with fully covered and partial covering absorption) with respect to the 7.1--9.0\,keV continuum flux (see Table~\ref{Time Dependent Cutoff Spectrum}). We found no clear correlation between the parameter values and flux.
{ In addition, we found that a bremsstrahlung model { with a temperature of 7.2$\pm$0.2\,keV} also fits the \textsl{Suzaku} data reasonably well. However, this model is found to provide unsatisfactory fits to the PCA spectra where $\chi^2_\nu$ was found to exceed 2 (see Section~\ref{PCA Spectral Description}). Therefore, the power law with the high energy cutoff model is the preferred description of 4U 1210-64.}
\subsubsection{PCA Spectral Analysis}
\label{PCA Spectral Description}
We analyzed the PCA spectral data of 4U 1210-64 from MJD 54804--54842, using the models described in Section~\ref{Suzaku Spectral Analysis}. The best-fit model is a power law with a high energy cutoff (see Figure~\ref{PCA Spectra}, top and middle). Due to the lack of low energy response (see Section~\ref{PCA Description}), { the absorption that is observed in the \textsl{Suzaku} spectra cannot be accurately measured using the PCA (see Section~\ref{Suzaku Spectral Analysis}). Therefore, we froze $N_{\rm H}$ to the out-of-eclipse value determined with \textsl{Suzaku}.}
{ The power law index, $\Gamma$, and cutoff energies found with the PCA are consistent with those determined with the \textsl{Suzaku} analysis. However, we note that the folding energy obtained with the PCA differs somewhat from that obtained with \textsl{Suzaku}. We also note that the folding energy was found to be variable among different time segments of PCA data.}
{ Due to the $\sim$10 times lower spectral resolution of the PCA compared with \textsl{Suzaku} (see Sections~\ref{PCA Description} and~\ref{XIS description}, respectively), we detected only a broad emission feature between energies 6.4--6.7\,keV in the Fe K$\alpha$ region, which was modeled using a single Gaussian (see Figure~\ref{PCA Spectra}, bottom). Furthermore, the S XVI K$\alpha$ feature was not detected in the PCA spectra, which is expected due to the low sensitivity of the PCA at 2.62\,keV.}
\begin{figure}
\centerline{\includegraphics[width=3.5in]{6-13-14_Total-PCAspectrum.pdf}}
\figcaption[6-13-14_Total-PCAspectrum.pdf]{
{ A typical PCA spectrum (MJD 54804) and best fit model (top panel). This consists of a continuum comprised of a power law with a high energy cutoff and a broad Fe K$\alpha$ emission line. Residuals of the best fit model are plotted in the middle panel. To illustrate the Fe K$\alpha$ emission lines along with the best fit model, the normalization of the line was set to 0. Residuals are plotted in the bottom panel.}
\label{PCA Spectra}
}
\end{figure}
\begin{deluxetable}{cccccccccccc}
\rotate
\tablecolumns{12}
\tabletypesize{\scriptsize}
\tablewidth{10.7in}
\tablecaption{Spectral parameters of 4U 1210-64 from the PCA observations}
\tablehead{
\colhead{Time} & \colhead{Phabs} & \colhead{Pcfabs} & \colhead{Cvr} & \colhead{High Energy} & \colhead{Folding} & \colhead{Power Law} & \colhead{Power Law} & \colhead{Fe K$\alpha$} & \colhead{EQW} & \colhead{$F_{\rm unabs}$} & \colhead{$\chi^2_\nu$} \\
\colhead{(MJD)} & \colhead{$N_{\rm H}$} & \colhead{$N_{\rm H}$} & \colhead{} & \colhead{Cutoff} & \colhead{Energy} &\colhead{$\Gamma$} & \colhead{Normalization} & \colhead{Energy} & \colhead{Fe K$\alpha$} & \colhead{(10$^{-10}$ erg cm$^{-2}$ s$^{-1}$)} & \colhead{(d.o.f)} \\
\colhead{} & \colhead{(10$^{22}$ cm$^{-2}$)} & \colhead{(10$^{22}$ cm$^{-2}$)} & \colhead{} & \colhead{(keV)} & \colhead{(keV)} & \colhead{} & \colhead{(10$^{-2}$)} & \colhead{(keV)} & \colhead{(eV)} & \colhead{} & \colhead{}}
\startdata
54804 & 0.70$^a$ & 6.74$^a$ & 0.36$^a$ & 6.0$\pm$0.1 & 5.6$\pm$0.1 & 1.50$\pm$0.03 & 7.8$\pm$0.3 & 6.66$^{+0.01}_{-0.02}$ & 228$^{+26}_{-18}$ & 7.4$\pm$0.2 & 0.82(35) \\
54812 & 0.70$^a$ & 6.74$^a$ & 0.36$^a$ & 7.4$\pm$0.2 & 5.5$\pm$0.4 & 1.96$\pm$0.04 & 4.6$\pm$0.3 & 6.65$^{+0.01}_{-0.07}$ & 414$^{+52}_{-39}$ & 1.71$\pm$0.04 & 0.49(35) \\
54813 & 0.70$^a$ & 6.74$^a$ & 0.36$^a$ & 6.6$\pm$0.3 & 7.0$^{+0.5}_{-0.4}$ & 1.70$\pm$0.05 & 3.2$\pm$0.2 & 6.59$^{+0.08}_{-0.02}$ & 306$^{+30}_{-36}$ & 2.01$\pm$0.06 & 0.71(35) \\
54814 & 0.70$^a$ & 6.74$^a$ & 0.36$^a$ & 6.5$\pm$0.2 & 5.9$\pm$0.3 & 1.70$\pm$0.04 & 4.6$^{+0.3}_{-0.2}$ & 6.66$\pm$0.09 & 159$^{+27}_{-32}$ & 2.86$\pm$0.08 & 0.35(35) \\
54815 & 0.70$^a$ & 6.74$^a$ & 0.36$^a$ & 6.3$\pm$0.2 & 6.8$^{+0.2}_{-0.3}$ & 1.70$\pm$0.03 & 5.0$\pm$0.2 & 6.662$^{+0.004}_{-0.085}$ & 250$^{+32}_{-23}$ & 3.08$^{+0.07}_{-0.06}$ & 0.74(35) \\
54816 & 0.70$^a$ & 6.74$^a$ & 0.36$^a$ & 5.9$\pm$0.2 & 6.9$^{+0.3}_{-0.2}$ & 1.43$\pm$0.03 & 4.2$\pm$0.2 & 6.61$^{+0.03}_{-0.04}$ & 297$^{+37}_{-36}$ & 4.6$\pm$0.1 & 1.00(34) \\
54817 & 0.70$^a$ & 6.74$^a$ & 0.36$^a$ & 6.1$\pm$0.2 & 6.2$^{+0.4}_{-0.3}$ & 1.51$\pm$0.05 & 3.5$\pm$0.2 & 6.56$^{+0.04}_{-0.05}$ & 318$^{+47}_{-46}$ & 3.3$\pm$0.1 & 0.51(34) \\
54818 & 0.70$^a$ & 6.74$^a$ & 0.36$^a$ & 7.6$^{+0.7}_{-1.0}$ & 10$^{+3}_{-2}$ & 2.01$^{+0.08}_{-0.11}$ & 1.7$^{+0.2}_{-0.3}$ & 6.486$^{+0.002}_{-0.085}$ & 909$^{+175}_{-135}$ & 0.60$^{+0.04}_{-0.02}$ & 0.55(35) \\
54821 & 0.70$^a$ & 6.74$^a$ & 0.36$^a$ & 7.6$^{+0.2}_{-0.3}$ & 4.7$^{+0.6}_{-0.5}$ & 2.20$\pm$0.05 & 4.3$\pm$0.3 & 6.669$^{+0.089}_{-0.001}$ & 502$^{+51}_{-62}$ & 1.00$\pm$0.02 & 0.70(35) \\
54822{ $^b$} & 0.94 & 10.73 & 0.80 & 8.1$\pm$0.2 & 8.7$^{+0.8}_{-0.7}$ & 2.56$^{+0.02}_{-0.03}$ & 13.5$\pm$0.5 & 6.67$\pm$0.01 & 440$^{+26}_{-39}$ & 1.63$\pm$0.01 & 1.17(35) \\
54826 & 0.70$^a$ & 6.74$^a$ & 0.36$^a$ & 6.5$\pm$0.1 & 7.4$\pm$0.2 & 1.63$\pm$0.02 & 7.1$^{+0.3}_{-0.2}$ & 6.63$^{+0.04}_{-0.05}$ & 160$^{+16}_{-14}$ & 5.01$\pm$0.08 & 1.00(35) \\
54827 & 0.70$^a$ & 6.74$^a$ & 0.36$^a$ & 6.7$\pm$0.2 & 6.5$^{+0.4}_{-0.3}$ & 1.79$\pm$0.04 & 5.4$\pm$0.3 & 6.67$\pm$0.06 & 306$^{+23}_{-34}$ & 2.79$^{+0.07}_{-0.06}$ & 0.83(35) \\
54828 & 0.70$^a$ & 6.74$^a$ & 0.36$^a$ & 7.3$\pm$0.2 & 6.9$\pm$0.4 & 1.86$\pm$0.04 & 5.3$\pm$0.3 & 6.63$^{+0.02}_{-0.07}$ & 305$^{+39}_{-27}$ & 2.41$\pm$0.05 & 1.07(35) \\
54829{ $^b$} & 0.94 & 10.73 & 0.80 & 8.5$\pm$0.4 & 11$\pm$2 & 2.62$\pm$0.03 & 18.1$\pm$0.9 & 6.48$^{+0.01}_{-0.08}$ & 485$^{+56}_{-45}$ & 1.98$\pm$0.02 & 1.69(35) \\
54830 & 0.70$^a$ & 6.74$^a$ & 0.36$^a$ & 6.6{ $^c$} & 6.6{ $^c$} & 1.82$\pm$0.08 & 0.7$\pm$0.1 & 6.51$^{+0.06}_{-0.03}$ & 696$\pm$105 & 0.36$\pm$0.01 & 0.87(37) \\
54831 & 0.70$^a$ & 6.74$^a$ & 0.36$^a$ & 8.1$\pm$0.5 & 10$^{+3}_{-2}$ & 2.12$\pm$0.06 & 3.0$^{+0.2}_{-0.3}$ & 6.59$\pm$0.03 & 1061$^{+159}_{-131}$ & 0.85$\pm$0.02 & 0.81(34) \\
54832 & 0.70$^a$ & 6.74$^a$ & 0.36$^a$ & 7.0$\pm$0.3 & 7.1$^{+0.7}_{-0.6}$ & 1.96$\pm$0.05 & 2.9$\pm$0.2 & 6.54$^{+0.04}_{-0.05}$ & 568$^{+56}_{-39}$ & 1.07$\pm$0.02 & 0.77(35) \\
54834 & 0.70$^a$ & 6.74$^a$ & 0.36$^a$ & 6.6$\pm$0.2 & 7.5$\pm$0.2 & 1.55$\pm$0.03 & 4.1$\pm$0.2 & 6.49$\pm$0.04 & 267$^{+15}_{-20}$ & 3.43$\pm$0.07 & 1.07(35) \\
54835 & 0.70$^a$ & 6.74$^a$ & 0.36$^a$ & 6.7$^{+0.1}_{-0.2}$ & 7.2$^{+0.3}_{-0.2}$ & 1.75$\pm$0.03 & 4.5$\pm$0.2 & 6.66$\pm$0.09 & 148$^{+19}_{-16}$ & 2.53$\pm$0.06 & 0.84(35) \\
54836 & 0.70$^a$ & 6.74$^a$ & 0.36$^a$ & 7.8$^{+0.2}_{-0.3}$ & 9.8$^{+0.5}_{-0.4}$ & 1.58$\pm$0.03 & 9.4$^{+0.4}_{-0.3}$ & 6.488$^{+0.086}_{-0.002}$ & 312$^{+36}_{-41}$ & 7.5$\pm$0.2 & 0.79(35) \\
54837 & 0.70$^a$ & 6.74$^a$ & 0.36$^a$ & 6.4$\pm$0.1 & 6.3$\pm$0.1 & 1.73$\pm$0.02 & 5.6$\pm$0.2 & 6.68$^{+0.07}_{-0.01}$ & 236$^{+14}_{-17}$ & 3.26$\pm$0.05 & 0.95(35) \\
54839 & 0.70$^a$ & 6.74$^a$ & 0.36$^a$ & 6.6{ $^c$} & 6.6{$^c$} & 1.7$\pm$0.2 & 0.41$^{+0.12}_{-0.09}$ & 6.4$\pm$0.1 & 570$^{+202}_{-182}$ & 0.27$\pm$0.02 & 0.94(37) \\
54840 & 0.70$^a$ & 6.74$^a$ & 0.36$^a$ & 6.6$^{+0.6}_{-0.3}$ & 7.8$\pm$0.6 & 1.90$^{+0.03}_{-0.04}$ & 3.8$^{+0.4}_{-0.2}$ & 6.73$^{+0.04}_{-0.13}$ & 362$^{+48}_{-47}$ & 1.59$\pm$0.05 & 0.55(35) \\
54842{ $^b$} & 0.94 & 10.73 & 0.80 & 7.9{ $^d$} & 8.0{ $^d$} & 2.5$\pm$0.1 & 3.3$^{+0.8}_{-0.6}$ & 6.9$\pm$0.2 & 529$^{+217}_{-211}$ & 0.50$\pm$0.03 & 0.55(37) \\
\tableline
\enddata
\tablecomments{The best-fit parameters for the PCA spectra. \\*
$^a$ Out-of-eclipse, the fully covered absorber and partially covered absorber are frozen to the best fit values of the \textsl{Suzaku} out-of-eclipse spectrum (see Table~\ref{Suzaku Spectral Parameters}) \\*
$^b$ Time intervals when 4U 1210-64 is in eclipse. $N_{\rm H}$ is frozen to the best fit values of the \textsl{Suzaku} eclipse spectrum (see Table~\ref{Suzaku Spectral Parameters}) \\*
$^c$ The Cutoff Energy and Folding Energy parameters at these time intervals are frozen to the weighted average of the out-of-eclipse spectra. \\*
$^d$ The Cutoff Energy and Folding Energy parameters at these time intervals are frozen to the weighted average of the eclipse spectra.}
\label{PCA Blend Spectrum}
\end{deluxetable}
{ The temporal dependence of the spectral parameters was investigated (see Table~\ref{PCA Blend Spectrum}). We searched for correlations between the continuum spectral parameters (folding energy, the high energy cutoff and the power law index) with respect to the 2.5--20\,keV continuum flux (see Figure~\ref{PCA Spectral Parameters}). While no clear correlation between the folding energy and flux was found, the high energy cutoff and power law index are anti-correlated in respect to the continuum flux. The correlation coefficients (r) are: -0.69 and -0.81 for the high energy cutoff and power law index, respectively.}
\begin{figure}
\centerline{\includegraphics[width=3.2in]{june16-2014_PCAspectrum_SpectralParameters.pdf}}
\figcaption[june16-2014_PCAspectrum_SpectralParameters.pdf]{
{ Power Law Index (top), Folding Energy (middle) and High Energy Cutoff Energy (bottom) for the PCA spectra vs. the continuum flux in the 2.5--20\,keV band in logarithmic units. The red lines indicate the best fit for the power law model. The correlation coefficients (r) are: -0.72 and r=-0.80 for the high energy cutoff and power law index, respectively. The blue points indicate the data collected during eclipse.}
\label{PCA Spectral Parameters}
}
\end{figure}
\section{Discussion}
In our analysis of 4U 1210-64, we found the presence of an eclipse, long and short-term variability, and an Fe K$\alpha$ emission complex. Below, we discuss constraints on the mass donor based on the eclipse half-angle, the nature of the compact object, the variability found in the system, and the mechanism responsible for the Fe K$\alpha$ emission seen in 4U 1210-64.
The source emission does not reach 0\,counts s$^{-1}$ in the folded light curves (see Section~\ref{Eclipse Profile}). Residual emission was also found in the \textsl{Suzaku} and PCA { observations} (see Tables~\ref{Suzaku Spectral Parameters}--\ref{PCA Blend Spectrum}). { While residual emisison has been attributed to a dust-scattering in some other HMXBs \citep[Cen X-3; Vela X-1; OAO 1657-415,][]{1991MNRAS.251...76D,1994ApJ...436L...5W,2006MNRAS.367.1147A}, the residual emission found in 4U 1210-64 is seen at much higher levels compared to the out-of-eclipse emission, and is found in both MAXI (2--20\,keV) and PCA (2.5--20\,keV). A dust-scattering halo is predominantyly a soft X-ray phenomenon: Given the intersteller $N_{\rm H}$ (see Table~\ref{Suzaku Spectral Parameters}) we infer perhaps 10-20$\%$ of the out-of-eclipse flux at 1 keV may be in a dust-scattering halo \citep{1995A&A...293..889P}, but a much smaller fraction in the MAXI and the PCA band. Since we detect Fe XXV and Fe XXVI lines in eclipse, indicating the presence of an extended region of ionized gas, we believe it's plausible that Compton scattering and reprocessing can account for the residual flux in eclipse \citep{2006ApJ...651..421W}.}
\subsection{Constraints on the Mass-Donor}
\label{Constraints on the Mass-Donor}
Our analysis of the ASM, MAXI and BAT folded light curves reveals the presence of a sharp dip, which is suggestive of an eclipse (see Section~\ref{Eclipse Profile}).
\subsubsection{Eclipse Half-Angle Constraints}
X-ray Binaries that are eclipsing have an eclipse duration (see Section~\ref{Eclipse Profile}) that is dependent on the radius of the mass donor, inclination angle of the system and the
orbital separation of the components. Using the observed orbital period and Kepler's third law,
the phenomenon can be written in terms of the sum of the donor star and compact object masses,
which stipulates that the eclipse half-angle, $\Theta_{\rm e}$, can now be expressed in terms of
the radius, inclination and masses of the components. { In one set of calculations, we
assume a 1.4\,$M_\sun$ compact object which may be appropriate for an accreting neutron star.}
The region allowed by the measured eclipse half-angle for 4U 1210-64 in the Mass-Radius plot is
shown in Figure~\ref{Stellar Constraints} (shaded region). Its inclination is constrained between
edge-on orbits (left boundary of the shaded region in Figure~\ref{Stellar Constraints}) and close
to face-on orbits (the right boundary of the shaded region in Figure~\ref{Stellar Constraints}).
We can attach additional constraints assuming that the mass donor underfills the Roche lobe
radius, which is dependent on the mass ratio of the system and the orbital separation (see Figure~
\ref{Stellar Constraints}, bottom boundary). To calculate the eclipse half-angle and the Roche-
lobe radius, we used Equation E.4 in \citet{1996PhDT........96C} and Equation 2 in \citet
{1983ApJ...268..368E}, respectively. We additionally calculated the minimum inclination angle of
the system, $i_{\rm min}$, that is consistent with the measured eclipse half-angle (see Table~\ref
{Primary Parameters} and Figure~\ref{Half Angle}). An intriguing result is that the duration of
the observed eclipse is inconsistent with the proposed B5 V spectral type (see Figure~\ref{Half
Angle}). A B5 V has a mass of 5.9\,$M_\sun$ according to \citet{2006ima..book.....C} and \citet
{2000asqu.book.....A}. For a donor star of 5.9\,$M_\sun$ to satisfy the eclipse half-angle
constraint, we calculated the radius must exceed 5.37\,$R_\sun$. This is clearly larger than the
radii reported in \citet{2006ima..book.....C} and \citet{2000asqu.book.....A}, which is 4.1\,$R_
\sun$ and 3.9\,$R_\sun$, respectively (see Table~\ref{Primary Parameters}). Therefore, the radius
of a B5 V is too small to satisfy our observed eclipse duration (see Figure~\ref{Stellar
Constraints}).
The eclipse half-angle was calculated as a function of inclination angle for other B-type stars--B0 V, B5 III, B0 III, B5 I and B0 I (see Figure~\ref{Half Angle}). The eclipse half-angle was found to be consistent with a main-sequence star of spectral class B0 V only at high inclination angles (see Table~\ref{Primary Parameters}). We also consider intermediate and late spectral types in our analysis (see Figures~\ref{Stellar Constraints} and~\ref{Half Angle}). These will be discussed in Section~\ref{Intermediate-Mass}.
{ We also calculated the eclipse half-angle as a function of the inclination angle for these stars under the assumption of more massive compact objects. The results are presented in Figures~\ref{Stellar Constraints} and~\ref{Half Angle} for the scenario of a 1.9\,$M_\sun$ neutron star, which is one of the highest known masses for neutron stars in XRBs \citep{2012ARNPS..62..485L}. Our results remained the same for more substantial mass donors. However, the results require slightly higher inclination angles for intermediate and late spectral types (see Figure~\ref{Stellar Constraints}).}
The spectral type of the mass donor places an additional constraint on the distance of 4U 1210-64. Under the assumption that the R-band magnitude ($m_{\rm R}$) and extinction in the V-band ($A_{\rm V}$) are magnitudes 13.9 and { 3.3} for a B5 V classification \citep{2009A&A...495..121M}, the distance and average X-ray luminosity of the source are found to be $\sim$2.8\,kpc and 1.79$\pm$0.02 $\times$10$^{35}$\,erg s$^{-1}$, respectively (see Table~\ref{Primary Parameters}). { We calculated the extinction in the V-band ($A_{\rm V}$) using Equation 1 in \citet{2009MNRAS.400.2050G} and the measured neutral hydrogen column densities for the fully covered absorber (see Table~\ref{Time Dependent Cutoff Spectrum}). $A_{\rm V}$ was found to be 3.2$\pm$0.1 for the power law with high energy cutoff model.}
The distance and average X-ray luminosity of 4U 1210-64 assuming the aforementioned spectral types is reported in Table~\ref{Primary Parameters} using the values for $M_{\rm V}$ and $A_{\rm R}$ obtained from \citet{2006ima..book.....C}. A B0 V star places the system at an estimated distance of { $\sim$9.5\,kpc} away from the Sun, indicating that 4U 1210-64 could be located in the Carina arm (approximately 10\,kpc). A supergiant classification places 4U 1210-64 at a galactocentric distance exceeding $\sim$26\,kpc, which is outside the Galaxy. Therefore, the possibility of a supergiant must be excluded. { Since \citet{2009A&A...495..121M}'s previous classification must also be excluded due to the observed eclipse duration}, it is possible that the mass donor could be an early B-type giant or an early F-type giant (see Section~\ref{Intermediate-Mass}). Main-sequence stars with the exception of very early types and very high inclination angles are also excluded.
\begin{deluxetable}{ccccccccccc}
\rotate
\tablecolumns{11}
\tablewidth{0pc}
\tablecaption{ }
\tablehead{
\colhead{Spectral Type} & \colhead{$M/M_\sun$} & \colhead{$R/R_\sun$} & \colhead{$R_{\rm L}$$/R_\sun$$^a$} & \colhead{$M_{\rm V}$} & { $(B-R)_{\rm 0}$$^b$} & { $E(B-R)$$^b$} & \colhead{$i_{\rm min}$$\degr$$^c$} & \colhead{$d_{\rm sun}$$^d$} & \colhead{$d_{\rm gal}$$^e$} & \colhead{$L_{\rm xavg}$} \\
\colhead{} & \colhead{} & \colhead{} & \colhead{} & \colhead{} & \colhead{} & \colhead{} & \colhead{} & \colhead{(kpc)} & \colhead{(kpc)} & \colhead{($\times$10$^{35}$\,erg s$^{-1}$)}}
\startdata
\textsl{B5 V$^f$} & \textsl{5.9} & \textsl{4.1} & 13.7 & \textsl{-1.2} & { -0.21} & { 1.71} & \nodata$^g$ & $\sim$2.8 & $\sim$7.4& $\sim$1.6 \\
\textsl{B5 V$^f$} & \textbf{\textsl{5.9}} & \textbf{\textsl{3.9}} & 13.7 & \textsl{-1.2} & { -0.21} & { 1.71} & \nodata$^g$ & $\sim$2.8 & $\sim$7.4& $\sim$1.6 \\
\textsl{B0 V} & \textsl{18} & \textsl{8.4} & 23.1 & \textsl{-4.0} & { -0.40} & { 1.90} & 84 & $\sim$9.8 & $\sim$9.3& $\sim$20.0 \\
\textsl{B5 III} & \textsl{7.0} & \textsl{6.3} & 14.9 & \textsl{-2.2} & { -0.21} & { 1.71} & 85 & $\sim$4.1 & $\sim$7.3& $\sim$3.4 \\
\textsl{B0 III} & \textsl{20} & \textsl{13} & 24.6 & \textsl{-5.1} & { -0.34} & { 1.84} & 75 & $\sim$15 & $\sim$13& $\sim$44.6 \\
\textsl{B5 I} & \textsl{20} & \textsl{41} & 24.6 & \textsl{-6.2} & { -0.10} & { 1.60} & 9 & $\sim$29 & $\sim$26& $\sim$175 \\
\textsl{B0 I} & \textsl{25} & \textsl{25} & 27.2 & \textsl{-6.4} & { -0.36} & { 1.86} & 57 & $\sim$31 & $\sim$28 & $\sim$197 \\
\textsl{F0 III} & \textsl{2.0} & \textsl{5.0} & 7.7 & \textsl{1.5} & { 0.45} & { 1.05} & 82 & $\sim$0.9 & $\sim$7.9 & $\sim$0.2 \\
\textsl{G0 III} & \textsl{1.0} & \textsl{5.7}$^h$ & 5.2$^h$ & \textsl{1.0} & { 1.11} & { 0.39} & 77 & $\sim$1.3 & $\sim$7.8 & $\sim$0.3 \\
\enddata
\tablecomments{The values in italics are obtained from \citet{2006ima..book.....C} and \citet{2009A&A...495..121M}. \\*
The values in both italics and bold are obtained from \citet{2000asqu.book.....A} in comparison with those in \citet{2006ima..book.....C}. \\*
$^a$ The definition for the Roche lobe, $R_{\rm L}$, as given in \citet{1983ApJ...268..368E}, assuming $M_{\rm NS}$ is 1.4 $M_\sun$. \\*
$^b$ The value for $(B-R)_{\rm 0}$ was calculated using $(B-V)_{\rm 0}$ and $(R-V)_{\rm 0}$ published in \citet{1994MNRAS.270..229W}. $E(B-R)$ is found by subtracting $(B-R)_{\rm 0}$ from the observed $B-R$ (see Section~\ref{Does the Proposed Mass Donor Spectral Type Agree with Masetti's Optical Spectrum?}) \\*
$^c$ The minimum inclination angle of the system that is consistent with the measured eclipse half-angle. \\*
$^d$ The distance the object is from the Sun. \\*
$^e$ Galactocentric distance of 4U 1210-64. \\*
$^f$ \citet{2009A&A...495..121M}'s proposed spectral type and distance of the object. \\*
$^g$ A B5 V classification is inconsistent with the observed eclipse half-angle. \\*
$^i$ A G0 III classification significantly overfills the Roche-lobe assuming a 1.4 $M_\sun$ compact object.}
\label{Primary Parameters}
\end{deluxetable}
\begin{figure}
\centerline{\includegraphics[width=3.5in]{march12-2014_mass-constraints.pdf}}
\figcaption[march12-2014_mass-constraints.pdf]{
Log-log plot of stellar masses as a function of stellar radii. The shaded region, derived using Equation E.4 in \citet{1996PhDT........96C} for all possible inclination angles, indicates the allowed spectral types that satisfy the eclipse observed in the BAT, MAXI, and ASM folded light curves provided that the compact object is 1.4 $M_\sun$ (see Figures~\ref{Folded Light Curves} and~\ref{Eclipse Plot}, respectively). The black dotted lines show the allowed spectral types that satisfy the eclipse for a compact object of 1.9 $M_\sun$ (see Section~\ref{Constraints on the Mass-Donor}). Stellar masses and radii are given in Table~\ref{Primary Parameters}.
\label{Stellar Constraints}
}
\end{figure}
\begin{figure}
\centerline{\includegraphics[width=3in]{4-4-2014_weighted_eclipse_param.pdf}}
\figcaption[4-4-2014_weighted_eclipse_param.pdf]{
The black curves show predicted eclipse half angle as a function of inclination for stars with the indicated spectral types. The red and black dashed lines indicate the half angle and its estimated error as measured by ASM and MAXI (see Section~\ref{Constraints on the Mass-Donor}). We assume a neutron star mass of 1.4 $M_\sun$ (top) and of mass 1.9 $M_\sun$ (bottom) and typical masses and radii for the assumed companion spectral type (see Table~\ref{Primary Parameters}).
\label{Half Angle}
}
\end{figure}
\subsubsection{Does the Proposed Mass Donor Spectral Type Agree with Masetti's Optical Spectrum?}
\label{Does the Proposed Mass Donor Spectral Type Agree with Masetti's Optical Spectrum?}
{ The duration of the eclipse along with the constraint that the mass donor must underfill the Roche lobe allows for the possibility of several different spectral types (see Sections~\ref{Constraints on the Mass-Donor} and~\ref{Intermediate-Mass}). We compared the expected optical spectra for each proposed mass donor with \citet{2009A&A...495..121M}'s optical spectrum (the top-right panel in Figure 4) to place an additional constraint on the nature of the mass donor. Since the 4000-5000${\rm \AA}$ region is compressed in \citet{2009A&A...495..121M}'s broadband spectrum, there are possible caveats in the identification of spectral features to correctly classify the mass donor.}
{ The features observed in \citet{2009A&A...495..121M}'s optical spectrum include absorbed Balmer series lines and the emission of neutral helium, singly ionized helium and a blend of doubly ionized nitrogen and carbon. The optical spectra of B-type stars are expected to show the absorption of neutral helium and H$\alpha$ lines \citep[][and references therein]{2006ima..book.....C}. Singly ionized calcium features at $\sim$3900${\rm \AA}$ become dominant in F-type stars \citep[][and references therein]{2006ima..book.....C}, which would be difficult to detect in \citet{2009A&A...495..121M}'s broadband spectrum.} { The stellar luminosity type can also in principle be determined using spectral lines, which could lead to distinguishing a B III or B V type. For B5 stars, the ratio of Si II to Si III as well as the Al III and Fe III lines can be used to determine luminosity type \citep{2009ssc..book.....G}. Unfortunately, the existing optical spectra are not suitable in detecting these effects.}
The observed value of the $B-R$ color of 1.5\footnote{http://www.iasfbo.inaf.it/\~masetti/IGR/main.html} was compared with the intrinsic $(B-R)_{\rm 0}$ for the proposed mass donors (see Table~\ref{Primary Parameters}). Calculating the difference between the observed $B-R$ and the intrinsic $(B-R)_{\rm 0}$, we found the reddening values $E(B-R)$ for each proposed spectral type for the mass donor (see Table~\ref{Primary Parameters}). We calculated the reddening in the $B-V$ band, $E(B-V)$, using Equation 1 in \citet{2009MNRAS.400.2050G} and the measured neutral hydrogen column densities for fully covered absorber (see Table~\ref{Time Dependent Cutoff Spectrum}). We converted $E(B-V)$ to $E(B-R)$ using Table 3 in \citet{1985ApJ...288..618R}. We found $E(B-R)$ to be 1.82$\pm$0.08 for the power law with high energy cutoff model, where the reddening was found to be consistent with main-sequence and giant B-type stars but was inconsistent with F-types. Since a B5 V classification does not satisfy the eclipse half-angle, the optical features might indicate that the mass donor is a B-type giant. We note the optical spectra of B-type giants would be difficult to distinguish from that of main-sequence B-type stars because of the low-resolution of \citet{2009A&A...495..121M}'s broadband spectrum. F-type giants cannot be completely excluded due to systematic effects which prevent a definite determination. For example, we assumed that the fully covered $N_{\rm H}$ is entirely interstellar in origin (see Section~\ref{Suzaku Spectral Analysis}). { While this is the simplest interpretation of the data, we cannot exclude the possibility that the fully covered $N_{\rm H}$ is due to a combination of intrinsic and interstellar absorbers. According to our calculation, if $\sim$40$\%$ of the measured fully covered $N_{\rm H}$ is intrisic to the source and the rest interstellar, then the inferred $E(B-R)$ would be consistent with an F-type mass donor.}
{ Finally, we considered the possibility that the apparent spectral type is affected by heating by the radiation of the X-ray source. We first calculated the flux of 4U 1210-64 using both \textsl{Suzaku} and the PCA. Out-of-eclipse, the flux of 4U 1210-64 was found to be $\sim$2$\times$10$^{-10}$\,erg cm$^{-2}$ s$^{-1}$ for the power law model with a high energy cutoff model in the \textsl{Suzaku} data. The flux in the PCA is reported in Table~\ref{PCA Blend Spectrum} ranging from 0.29$\pm$0.03$\times$10$^{-10}$\,erg cm$^{-2}$ s$^{-1}$ to 8.9$\pm$0.4$\times$10$^{-10}$\,erg cm$^{-2}$ s$^{-1}$. We converted the apparent magnitude in the R band \citep{2009A&A...495..121M} into an optical flux using \citet{1979PASP...91..589B}. The optical flux in the R band was found to be 0.08\,Jy, which can be converted to $\sim$2$\times$10$^{-7}$\,erg cm$^{-2}$ s$^{-1}$. The flux ratio $F_{X}/F_{opt}$ was found to be $\sim$10$^{-4}$, which is much smaller than what is observed in systems were irradiation is important \citep[e.g. $F_{X}/F_{opt}$ was found to be $\sim$10$^{2}$ in Her X-1;][]{1983ARA&A..21...13B}. We conclude that irradiation effects are negligible in 4U 1210-64.}
\subsubsection{Is 4U 1210-64 an Intermediate-Mass X-ray Binary?}
\label{Intermediate-Mass}
{ In addition to the possibility that the donor star is a B-type giant, the duration of the eclipse suggests that an Intermediate-Mass classification cannot be ruled out}. We found the eclipse half angle is consistent with F0 and G0 giants at inclination angles exceeding 79$\degr$ and 70$\degr$, respectively (see Table~\ref{Primary Parameters}). Using the values for $M_{\rm V}$ and $A_{\rm R}$ published in \citet{2006ima..book.....C} and \citet{2009A&A...495..121M}, the distance of 4U 1210-64 was found to be $\sim$0.9\,kpc and $\sim$1.3\,kpc for a mass donor of spectral type F0 III and G0 III, respectively; which places 4U 1210-64 at a luminosity of $\sim$10$^{34}$\,erg s$^{-1}$ (see Table~\ref{Primary Parameters}). Other intermediate XRBs that host F-type stars include Cyg X-2 and Her X-1 \citep[e.g.][]{1995exru.book.....S}, which have considerably higher luminosities on the order of $\sim$10$^{37}$\,erg s$^{-1}$. The luminosities calculated for the F0 III and G0 III spectral types still exceed that of cataclysmic variables (see Section~\ref{What is the Nature of the Compact Object?}).
The possibility that the mass donor in 4U 1210-64 is an intermediate or late-type star hints at the presence of an accretion disk. XRBs that host intermediate- and low-mass stars accrete matter through Roche-lobe overflow \citep[see Equation 2,][]{1983ApJ...268..368E}. The Roche-lobe places an additional constraint on the spectral type of the mass donor in 4U 1210-64. While a spectral type of G0 III satisfies the observed eclipse half-angle, the Roche-lobe would be significantly overfilled (see Table~\ref{Primary Parameters}).
Finally, we discuss caveats in our hypothesis that the mass donor is an Intermediate-Mass star. No strong disk component was found, which is expected in the spectra of Intermediate-Mass X-ray Binaries \citep[e.g.][]{1995exru.book.....S}. We note that the reddening, $E(B-R)$, was found to be consistent with main-sequence and giant B-type stars but was inconsistent with { F-type and G-type stars} (see Section~\ref{Does the Proposed Mass Donor Spectral Type Agree with Masetti's Optical Spectrum?}). While $E(B-R)$ is apparently inconsistent with F-type stars, the possibility of systematic effects prevents excluding F-type stars as the possible mass donor { (see Section~\ref{Does the Proposed Mass Donor Spectral Type Agree with Masetti's Optical Spectrum?})}. However, we can confidently exclude G-type stars due to the additional constraint that the Roche-lobe is not significantly overfilled.
\subsection{What is the Nature of the Compact Object?}
\label{What is the Nature of the Compact Object?}
The nature of the compact object present in 4U 1210-64 remains ambiguous. An analysis of the ASM, MAXI and PCA power spectra shows that no pulsation period could be identified. { The PCA power spectra, which cover the range of 860\,s--38\,d (see Figure~\ref{Low Frequency}) and 10\,ms--14\,minutes (see Figure~\ref{Short Power}), are dominated by red noise, which could compromise our search for the pulsation period. While we removed the low-frequency noise from the power spectra (see Section~\ref{PCA Temporal}), the PCA power spectrum covering the range of 860\,s--38\,d was still compromised due to the orbital period of \textsl{RXTE} (see Figure~\ref{Low Frequency})}. A pulsation period would provide a clear indication that the compact object is a neutron star. Spectral results so far have also been inconclusive. Cyclotron lines, which would have proved a neutron star explanation for the compact object, are absent in the \textsl{Suzaku} spectra. Additionally, { the \textsl{Suzaku} and PCA data suggest the continuum can be modeled using an absorbed cutoff power law where the high-energy cutoff is 5.5$\pm$0.2\,keV}. Since a firm identification of the nature of the compact object in 4U 1210-64 has so far proven elusive, we compare our findings to systems where the compact object is known.
We first discuss the possible scenario that the compact object present in 4U 1210-64 is a black hole. Observations show that the exponential cut-off energy in HMXBs that host black holes exceeds 60\,keV \citep{2009ApJ...694..344T}, sharply contrasting with the 5.5$\pm$0.2\,keV cut-off observed in the \textsl{Suzaku} spectra. Additionally, \textsl{INTEGRAL} observations reveal the presence of soft excess in 4U 1210-64 \citep{2010ApJ...511..A48}, which is characteristic of HMXBs that host neutron stars \citep{2004ApJ...614..881H}. The low high-energy cutoff suggests that a black hole explanation of the compact object is unlikely.
We also consider the possibility that the compact object could be an accreting white dwarf. The luminosities observed in cataclysmic variables depend on the magnetic nature of the white dwarf, which affects the mode of accretion. The most luminous sub-type of CV, the intermediate polars, were found to be on the order of 10$^{31}$--4$\times$10$^{33}$\,erg s$^{-1}$ \citep{2009A&A...496..121B}. In comparison, the luminosities calculated for 4U 1210-64 exceed the above result by at least 1--2 orders of magnitude (see Table~\ref{Primary Parameters}). Assuming a bremsstrahlung fit, \citet{2009A&A...496..121B} found the temperatures ($kT_{\rm brems}$) of intermediate polars are on the order of 10--40\,keV, which differs from the value observed in 4U 1210-64 { (see Section~\ref{Suzaku Spectral Analysis})}. Since a bremsstrahlung model was found to be an unsatisfactory fit to the PCA data, we conclude that the CV explanation is unlikely.
Finally, we discuss the possibility that 4U 1210-64 contains a neutron star. Several geometries have been proposed to describe the apparent lack of a signal corresponding to the pulsation period. One possibility is a co-alignment of the magnetic and spin axes of the neutron star \citep[e.g.][]{2010ApJ...719..451B}. A second explanation suggests that throughout the rotation the accretion beam points in our direction \citep[e.g.][]{2010ApJ...719..451B}. The absence of a well defined pulsation period could also be explained by a weak magnetic field. Another possibility is that the compact object in 4U 1210-64 is a slowly rotating neutron star \citep[e.g. 2S 0114+650,][and references therein]{2008MNRAS.389..608F}.
\subsection{What is the physical process responsible for the low state observed in the ASM data?}
\label{What is the physical process responsible for the low state observed in the ASM data?}
The ASM data reveal the presence of three distinct system states as previously noted by \citet{2008ATel.1861....1C}. These are two active states and one low state we interpret as quiescence (see Section~\ref{ASM Temporal Analysis}). This long term variability is suggestive of a variable accretion rate.
We first discuss the possibility that 4U 1210-64 is powered by the Be mechanism. { In BeXBs the compact object accretes material from the circumstellar decretion disc of a rapidly rotating main-sequence or giant B-type star.} If the system is a BeXB, changes in the circumstellar disc around the Be star could also explain the period of low activity. Observations indicate that bright and faint states might correspond to the formation and dissipation of the circumstellar disc \citep{2010MNRAS.401...55R}. The time-scale for the development and disappearance of a circumstellar disc is typically on the order of 3--7\,years, which is consistent with the ASM data. One notable BeXB is SAX J2103.5+4545, which consists of a neutron star in a 12.7\,day orbit around a B0Ve star \citep{2010MNRAS.401...55R}. The timescale for the development and disappearance of the circumstellar disc is 1--2\,years, possibly due to the short orbital period \citep{2010MNRAS.401...55R}. { While a BeXB explanation supports the presence of high and low states observed in the ASM data, the presence of an eclipse of the compact object is inconsistent with most main-sequence B-type stars, where main-sequence stars later than a B0 do not satisfy the observed eclipse half angle (see Figures~\ref{Stellar Constraints}--~\ref{Half Angle}). This is not surprising due to the smaller radius of the mass donors observed in most BeXBs. We note; however, that B-type giants would satisfy the observed eclipse duration at high inclination angles (see Section~\ref{Constraints on the Mass-Donor}). Additionally, the Balmer lines, particularly the H$\alpha$ line, were found to be in absorption during \citet{2009A&A...495..121M}'s optical campaign of 4U 1210-64 (MJD\,54529.3). Since Masetti's observation occured during a high state of the system (see Table~\ref{ASM State Table} and Section~\ref{ASM Temporal Analysis}), we would expect to see H$\alpha$ in emission \citep{2010ApJS..187..228S}.}
{ We also consider the possibility that the state transitions originate due to a mechanism similar to what is observed in Black Hole Candidates (BHCs). Multiple states are observed in BHCs, which are defined as soft/high, intermediate and hard/low \citep[Cygnus X-1; GX 339-4,][]{2013A&A...554A..88G,2006smqw.confE...1N,2010MNRAS.403...61D}. The high-energy cutoff in BHCs exceeds 60\,keV \citep{2009ApJ...694..344T}, which is in variance with our results obtained with \textsl{Suzaku} and PCA. Additionally, a disk blackbody is required to model the low energy spectra of BHCs in the soft/high state \citep{1973A&A....24..337S,2006smqw.confE...1N}, which is not seen in the \textsl{Suzaku} and PCA analysis of 4U 1210-64.}
Finally, we discuss the possibility that the mode of accretion in 4U 1210-64 is Roche-lobe overflow, which primarily occurs in both Intermediate and Low-Mass X-ray Binaries. { We compare the long-term behavior of 4U 1210-64 to that seen in soft X-ray transients, NS-LMXBs that host at least two different states \citep[e.g. Aql X-1;][]{2013arXiv1308.6091S}}. In soft X-ray transients, the physical mechanism that could lead to a reduction of intensity is changes in the accretion rate, $\dot{M}$ \citep{2013arXiv1308.6091S}. The luminosity is significantly reduced when the accretion rate is low. Since the magnetic field of the majority of NS-LMXBs is weak, the propeller effect becomes important when the accretion rate, $\dot{M}$, is low \citep{2013arXiv1308.6091S}. The long-term behavior of 4U 1210-64 differs from soft X-ray transients. While there are extended low states \citep{2013PASJ...65...26M,2013arXiv1308.6091S}, the high states are shorter and brighter than what is observed in 4U 1210-64 { \citep{2014MNRAS.438.2634C}}.
{ No mechanisms described above appear to be consistent with 4U 1210-64. While there are uncertainties in the spectral classification, we note that the behavior of Intermediate-Mass X-ray Binaries is not well known. Therefore, if 4U 1210-64 is a member of this class, unusual variability may be possible.}
\subsection{What is the origin of the variability in the high-state?}
\label{What is the origin of the variability in the high-state?}
Our analysis of the \textsl{Suzaku} { and PCA} data reveals the presence of strong variability in the light curves ({ see Figures~\ref{Suzaku Light Curve} and~\ref{PCA Light Curve})}. { In the \textsl{Suzaku} data, ``flares" were observed to reach nearly 1.4 times the mean count rate (i.e. a modulation depth of 140$\%$)}. { The variability is even stronger in the PCA data, where the modulation depth was found to be 330$\%$. A reduced count rate was found in the PCA light curve between $\sim$MJD\,54830--54833, which is outside the eclipse (see Figure~\ref{PCA Light Curve}).} The unabsorbed flux of 4U 1210-64 was found to vary by a factor of $\sim$25 over the course of both the \textsl{Suzaku} and { PCA observations (see Sections~\ref{Suzaku Spectral Analysis} and~\ref{PCA Spectral Description})}.
{ The large variability in the Suzaku and PCA light curves could be attributed to several different physical processes, all resulting in changes in the accretion rate, $\dot{M}$. A positive correlation between the \textsl{Suzaku} hardness ratio and continuum flux was found (see Figure~\ref{Hardness Correlation}), which provides evidence against a strong wind. This is further strengthened by a decrease in the \textsl{Suzaku} hardness ratio during the egress phase of the observation. Since the system is eclipsing, a strong wind should lead to an increase in absorption and thus the X-ray hardness ratio during the ingress and egress phases of the orbit \citep{1988ApJ...324..974C,2013A&A...554A..37D,2008ApJ...675.1487S}. The folded MAXI and ASM light curves (see Figure~\ref{Folded Hardness Ratios}, top and bottom) provide additional evidence against it. The increase in the folded MAXI hardness ratio is modest (Figure~\ref{Folded Hardness Ratios}, top), indicating the possible presence of a tenuous wind but not of a typical HMXB wind \citep[e.g. Vela X-1; Cen X-3,][]{2013A&A...554A..37D,2008ApJ...675.1487S}. Such behavior is not seen in the folded hardness ratio produced by the ASM, which is possibly due to the low count rate (see Figure~\ref{Folded Hardness Ratios}, bottom). The observed $N_{\rm H}$ could possibly originate in an accretion stream \citep[e.g. Cygnus X-1,][and references therein]{2009ApJ...690..330H}.}
\begin{figure}
\centerline{\includegraphics[width=3.5in]{november06-2013_hardness-ratios.pdf}}
\figcaption[november06-2013_hardness-ratios.pdf]{
Hardness ratios of the folded MAXI (top) and ASM (bottom) light curves. The hardness ratio is defined as $C_{\rm hard}$-$C_{\rm soft}$/$C_{\rm hard}$+$C_{\rm soft}$, where the soft and hard energy bands are defined as 2--4\,keV and 4--10\,keV and 1.5--5\,keV and 5-12\,keV for the MAXI and ASM light curves, respectively. The eclipse is indicated by the shaded regions.
\label{Folded Hardness Ratios}
}
\end{figure}
Magnetic and centrifugal barriers have been proposed to inhibit the accretion process in XRBs on hourly timescales \citep{2008arXiv0811.0995B}, where the timescale is consistent with the variability in the light curve. { This could explain the reduced count rate in the PCA lightcurve (see Figure~\ref{PCA Light Curve}, top).} Another mechanism that could lead to { such a reduction} is the formation and dissipation of an unstable accretion disk \citep[e.g.][and references therein]{2011ApJ...727...59B}.
\subsection{Emission Lines}
The analysis of spectral data produced by \textsl{Suzaku} reveals the presence of emission lines at energies 2.6\,keV, 6.4\,keV, 6.7\,keV and 6.97\,keV. We interpret the emission lines as S XVI K$\alpha$, Fe K$\alpha$, Fe XXV K$\alpha$ and Fe XXVI K$\alpha$, respectively. Below we discuss the mechanisms that are proposed to explain the emission lines seen in 4U 1210-64.
\subsubsection{Fe K$\alpha$ Emission}
A 6.4\,keV emission line has been shown to be present in many XRBs \citep[e.g. Vela X-1, 4U 1700-377, Cen X-3 and 4U 1822-37;][] {2002ApJ...564L..21S,2005A&A...432..999V,2011ApJ...737...79N,2013arXiv1311.4618S}. The origin of the 6.4\,keV emission line is due to neutral Fe or Fe in a low ionization state. Unless otherwise stated, we assume that Fe I emission is responsible for the 6.4\,keV emission line.
Our analysis of the Fe I line indicates that the line flux tracks the flux in the continuum in the 7.1--9.0\,keV band when 4U 1210-64 is out-of-eclipse and is not detected in eclipse (see Section ~\ref{Suzaku Spectral Analysis}). { This is suggestive that the region responsible for the Fe I emission is close to the compact object. The accretion mechanism in 4U 1210-64 differs from the X-ray excited wind observed in the HMXB Cen X-3 \citep{1993ApJ...403..322D}, but the Fe K$\alpha$ emission region is similar in both objects. For instance, the Fe I emission line in Cen X-3 was observed to be weakest during the eclipse phase \citep{1996PASJ...48..425E,2012BASI...40..503N}. Because of these similarities, we compare 4U 1210-64 with Cen X-3 to understand the Fe K$\alpha$ emission feature}.
Possible mechanisms that have been suggested to cause the fluorescence of cold material include a plasma layer at the surface of the Alfven shell and an optically thick accretion disc \citep{1980A&A....87..330B,1996PASJ...48..425E}. The flux of Fe I was found to decrease by more than an order of magnitude during eclipse, which is a comparable to the change of flux observed in the continuum. This is further indication that the origin of the Fe I emission is close to the compact object. The slope of the Fe I flux versus the continuum flux is near unity, which shows that the Fe I emission is in agreement with fluorescence \citep{2012A&A...547A.103N,2013A&A...551A...1R}.
We also consider the mechanism responsible for the Fe XXV and Fe XXVI emission features. A possible correlation between the flux of both the Fe XXV and Fe XXVI emission features with respect to the continuum flux in the 7.1--9.0\,keV band was found, which shows that the presence of Fe XXV and Fe XXVI increases as the continuum flux increases (see Section 3.4). We note that the slope between the continuum flux and the flux of the lines is significantly less than 1, which we interpret as a possible sign that an increasing part of the medium might be completely ionized.
To place constraints on the state of the plasma, we analyzed the flux ratio between the Fe XXV and Fe XXVI emission features and the continuum flux in the 7.1--9.0\,keV band. No change was found in the flux ratio between the Fe XXV and Fe XXVI emission features and the continuum flux in the 7.1--9.0\,keV band (see Table~\ref{Time Dependent Cutoff Spectrum}), which indicates that the { Fe XXV and Fe XXVI features possibly originate in the same region of a structured medium, where the} ionization state can be independent of luminosity \citep{1996PASJ...48..425E}, which is in agreement with our result of a constant flux ratio as a function of luminosity.
Fe XXVI and Fe XXVI are likely due to photoionization. Recombination followed by electron cascade transitions is present in systems such as SMC X-1 \citep{2001ApJ...563L.139V}, Cen X-3 \citep{2012A&A...547A.103N,2013A&A...551A...1R}, Vela X-1 \citep{2004AJ....127.2310G}, 4U 1700-37 \citep{1996ApJ...468L..33L} and 4U 1822-37 \citep{2013arXiv1311.4618S}. Emission features are more prominent during eclipse since direct emission from the compact object irradiating the accretion stream is no longer visible. The photoionization mechanism must originate in regions of low density since the range of luminosities we inferred in 4U 1210-64 (see Table~\ref{Primary Parameters}) is significantly lower than what is observed in systems such as Cen X-3 and SMC X-1 \citep[$\sim$10$^{37}$\,erg s$^{-1}$,][]{2012BASI...40..503N,2001ApJ...563L.139V}.
The very large EQWs of the Fe XXV and Fe XXVI emission features in eclipse (see Section 3.4) are consistent with an origin due to the reprocessing of photons in the accretion stream. In comparison, the EQW of the Fe XXV and Fe XXVI emission features observed in the HMXB Cen X-3 is largest during eclipse and tends to decrease as the continuum flux increases \citep{2012BASI...40..503N}. \citet{2012BASI...40..503N} show that the region responsible for the Fe XXV and Fe XXVI emission observed in Cen X-3 is extended and is comparable to the size of the mass donor in the system, which is likely what we observe in 4U 1210-64.
\subsubsection{The presence of S XVI in 4U 1210-64}
Different ionization species of low to mid-Z elements such as S are present in eclipsing XRBs \citep[e.g. Vela X-1, 4U 1700-377 and LMC X-4,][]{2002ApJ...564L..21S, 2005A&A...432..999V, 2010ApJ...720.1202H}. While near neutral fluorescent lines in addition to highly ionized species of emission lines were observed in SGXBs such as Vela X-1 \citep{2002ApJ...564L..21S}, the \textsl{Suzaku} spectra of 4U 1210-64 reveal only the highly ionized species of S XVI.
It has been shown that photoionization and radiative recombination are responsible for the presence of hydrogen-like species of S in absorbed XRBs \citep{2004ApJ...600..358I}. The S XVI K$\alpha$ emission features only appeared during part of the \textsl{Suzaku} observation (see Table~\ref{Time Dependent Cutoff Spectrum}). As a result, we could not measure the temporal variability of the EQW or fluxes of S XVI.
\section{Conclusion}
{ 4U 1210-64, for which we determined an orbital period of 6.7101$\pm$0.0005\,days, is a unique XRB. The companion star was previously proposed to have a spectral type of B5 V. We found that a B5 V classification does not satisfy the eclipse half-angle, compelling evidence against a B5 V classification. 4U 1210-64’s spectral features seem to indicate that the mass donor could be a B0 V or B0-5 III star. A Be-type accretion mechanism, with most BeXBs hosting primaries of spectral type late O to B2 \citep{1998A&A...338..505N}, is unlikely, since these systems usually have longer periods and are transients. A supergiant classification must be excluded since the implied distance would put the object outside the Galaxy. F-type giants also satisfy the constraints imposed by the eclipse half-angle and Lagrange point, L1, where Roche-lobe transfer would be expected to occur. To further constrain the spectral type of the mass donor, the reddening values $E(B-R)$ were calculated for the possible spectral types. The reddening was found to be consistent with main sequence and giant B-type stars but was inconsistent with F-types. Due to the uncertainties in the conversion between the $N_{\rm H}$ and $E(B-R)$, we cannot completely exclude an F-type mass donor.}
4U 1210-64 hosts a compact object that remains ambiguous in nature. No signs of pulsations or cyclotron features were found in our analysis of 4U 1210-64, which would prove that the compact object is a neutron star. The spectral properties of the continuum strongly contrast with those typically seen in black hole candidates (BHC) as well as CVs. { In particular, a disk blackbody is required to model the low energy spectra of BHCs in the soft/high state while CVs are typically fit with a bremsstrahlung model.} While the nature of the compact object has proven elusive, a neutron star with a weak magnetic field possibly aligned with the spin axis is consistent with the lack of pulsations and cyclotron features.
Emission lines at 2.62\,keV, 6.41\,keV, 6.7\,keV and 6.97\,keV were all clearly detected in the \textsl{Suzaku} spectra, which we interpret as S XVI K$\alpha$, Fe K$\alpha$, Fe XXV K$\alpha$ and Fe XXVI K$\alpha$, respectively. The flux of the Fe K$\alpha$ lines closely tracks the flux of the unabsorbed continuum. We found a linear relationship between the flux of Fe I vs. the continuum, which shows that the most probable origin of the Fe I line is fluorescence of cold and dense material close to the compact object. An origin close to the compact object is further supported by the fact that Fe I is not clearly detected during eclipse. The slopes of the relationship between the logarithm of the Fe XXV and Fe XXVI flux versus the logarithm continuum possibly show that an increasing part of the medium might be completely ionized as the flux increases.
{ Strong variability was found in the \textsl{Suzaku} and PCA observations. The out-of-eclipse flux was found to be 1.73$^{+0.06}_{-0.05}$$\times$10$^{-10}$\,erg s$^{-1}$ cm$^{-2}$, which implies a luminosity $\sim$10$^{34}$-$\times$10$^{36}$\,erg s$^{-1}$. The variability was found to be a factor of 25 in both the \textsl{Suzaku} and PCA observations. A positive correlation was seen in both the \textsl{Suzaku} hardness-intensity diagram and the PCA color-color diagram, which provides evidence against a strong wind. In eclipsing X-ray binaries, a strong wind should lead to an increase in absorption and thus the X-ray hardness ratio during the ingress and egress phases of the orbit. Additional evidence from the folded MAXI and ASM light curves suggests that the observed $N_{\rm H}$ could originate possibly originate in an accretion stream.}
{ 4U 1210-64 appears to be a NS-HMXB but conclusive evidence remains to be found. Additional multi-wavelength observations are required to achieve a full understanding of the source.}
\acknowledgements
We thank Drs. Tim Kallman, Vanessa McBride and Joern Wilms for useful discussion. We also thank the anonymous referee for useful comments.
|
\section{Introduction}
We study the $N$-dimensional $(N \geq 2)$ Navier-Stokes equations (NSE) and magnetohydrodynamics (MHD) system defined respectively as follows:
\begin{subequations}
\begin{align}
&\frac{du}{dt} + (u\cdot\nabla) u + \nabla \pi = \nu \Delta u,\\
&\nabla \cdot u = 0, \hspace{3mm} u(x,0) = u_{0}(x),
\end{align}
\end{subequations}
\begin{subequations}
\begin{align}
&\frac{du}{dt} + (u\cdot\nabla) u + \nabla \pi = \nu \Delta u + (b\cdot\nabla) b,\\
&\frac{db}{dt} + (u\cdot\nabla) b = \eta \Delta b + (b\cdot\nabla)u,\\
&\nabla \cdot u = \nabla\cdot b = 0, \hspace{3mm} (u,b)(x,0) = (u_{0}, b_{0})(x),
\end{align}
\end{subequations}
where $u = (u_{1}, \hdots, u_{N}): \mathbb{R}^{N} \times \mathbb{R}^{+} \mapsto \mathbb{R}^{N}, b = (b_{1}, \hdots, b_{N}): \mathbb{R}^{N} \times \mathbb{R}^{+} \mapsto \mathbb{R}^{N}, \pi: \mathbb{R}^{N} \times \mathbb{R}^{+} \mapsto \mathbb{R}$ represent the velocity vector field, magnetic vector field and pressure scalar field respectively. We denote by the parameters $\nu, \eta \geq 0$ the viscosity and magnetic diffusivity respectively. Hereafter, we also denote $\frac{d}{dt}$ by $\partial_{t}$ and $\frac{d}{dx_{i}}$ by $\partial_{i}, i = 1, \hdots, N$ and by $\nabla_{i,j}$ the gradient vector field with $\partial_{i}, \partial_{j}$ on the $i$-th, $j$-th component respectively and zero elsewhere and $\Delta_{i,j}$ the sum of second derivatives in the $i$-th and $j$-th directions , e.g. $\nabla_{1,2} = (\partial_{1}, \partial_{2}, 0, \hdots, 0), \Delta_{1,2} = \sum_{k=1}^{2}\partial_{kk}^{2}.$
The importance and difficulty of the global regularity issue of the solution to these two systems are well known. In short, this is because the systems are both energy-supercritical in any dimension bigger than two even with $\nu, \eta > 0$. Indeed, e.g. for the MHD system, taking $L^{2}$-inner products with $(u,b)$ on (2a)-(2b) respectively and integrating in time lead to
\begin{equation}
\sup_{t \in [0,T]} (\lVert u\rVert_{L^{2}}^{2} + \lVert b\rVert_{L^{2}}^{2})(t) + \int_{0}^{T} \lVert \nabla u\rVert_{L^{2}}^{2} + \lVert \nabla b\rVert_{L^{2}}^{2} d\tau \leq \lVert u_{0} \rVert_{L^{2}}^{2} + \lVert b_{0} \rVert_{L^{2}}^{2}.
\end{equation}
On the other hand, it can be shown that if $(u,b)(x,t)$ solves the system (2a)-(2c), then so does $(u_{\lambda}, b_{\lambda})(x,t) \triangleq \lambda (u,b)(\lambda x, \lambda^{2} t)$. A direct computation shows that
\begin{equation*}
\lVert u_{\lambda}(x,t) \rVert_{L^{2}}^{2} + \lVert b_{\lambda}(x,t) \rVert_{L^{2}}^{2} = \lambda^{2-N}(\lVert u(x, \lambda^{2} t) \rVert_{L^{2}}^{2} + \lVert b(x, \lambda^{2} t)\rVert_{L^{2}}^{2}).
\end{equation*}
We call an equation with a scaling symmetry critical when the strongest norm for which an \emph{a priori } estimate is available is scaling-invariant. Thus, it is standard to classify the two-dimensional NSE and the MHD system as energy-critical while for any dimension higher, energy-supercritical; in fact, it can be considered that the supercriticality increases in dimension.
In two-dimensional case with $\nu, \eta > 0$, the authors in [22, 26] have shown the uniqueness of the solution to the NSE and the MHD system respectively. In fact, in the two-dimensional case due to the simplicity of the form after taking curls, when the dissipative and diffusive terms are replaced by fractional Laplacians, their powers may be reduced furthermore below one; we refer interested readers to [34] for the NSE with $\nu = 0$, [6] and references found therein for the MHD system. In any dimension strictly higher than two, the problem concerning the global regularity of the strong solution and the uniqueness of the weak solution to both systems remain open and hence much effort has been devoted to provide criterion so that they hold. We now review some of them, emphasizing on those of most relevance to the current manuscript.
Initiated by the author in [27], it has been established that if a weak solution $u$ of the NSE with $\nu > 0$ satisfies
\begin{equation}
u \in L^{r}(0, T; L^{p}(\mathbb{R}^{N})), \hspace{3mm} \frac{N}{p} + \frac{2}{r} \leq 1, \hspace{3mm} p \in (N, \infty],
\end{equation}
then $u$ is smooth (see [9, 11] for the endpoint case). In [2], the author showed that if $u$ solves the NSE (1a)-(1b) with $\nu > 0$ and
\begin{equation}
\nabla u \in L^{r}(0, T; L^{p}(\mathbb{R}^{N})), \hspace{2mm} N \geq 3, \hspace{2mm} \frac{N}{p} + \frac{2}{r} = 2, \hspace{2mm} 1 < r \leq \min\{2, \frac{N}{N-2}\},
\end{equation}
then $u$ is a regular solution. For the MHD system, the authors in [15, 37] independently showed that the sufficient condition for the regularity of the solution pair $(u,b)$ to the MHD system (2a)-(2c) may be reduced to just $u$. For many more important results in this direction of research, all of which we cannot list here, we refer to the prominent work of [1, 14] and references found therein. We do mention that the author in [38] showed that only in case $N = 3, 4$, $u$, the solution to the NSE (1a)-(1b) with $\nu > 0$, is regular and unique if
\begin{equation}
\nabla \pi \in L^{r}(0,T; L^{p}(\mathbb{R}^{N})), \hspace{3mm} \frac{N}{p} + \frac{2}{r} \leq 3, \hspace{3mm} \frac{N}{3} \leq p \leq \infty.
\end{equation}
We emphasize that the norm $\lVert \cdot \rVert_{L_{T}^{r}L_{x}^{p}}$ in (4) is scaling invariant precisely when $\frac{N}{p} + \frac{2}{r} = 1$; i.e.
\begin{equation*}
\int_{0}^{T} \lVert u_{\lambda}(x,t) \rVert_{L^{p}}^{r} dt = \int_{0}^{\lambda^{2} T} \lVert u(x,t)\rVert_{L^{p}}^{r} dt \hspace{1mm} \text{ if and only if } \hspace{1mm} \frac{N}{p} + \frac{2}{r} = 1,
\end{equation*}
where $u_{\lambda}(x,t) = \lambda u(\lambda x, \lambda^{2} t)$, and similarly for the norm in (5) at the endpoint of $2$.
We now survey some component reduction results of such criterion. The authors in [20] showed that if $u$ solves the NSE with $N = 3, \nu > 0$ and
\begin{align}
u_{3} \in& L^{r}(0, T; L^{p}(\mathbb{R}^{3})), \hspace{3mm} \frac{3}{p} + \frac{2}{r} \leq \frac{5}{8}, \hspace{3mm} r \in [\frac{54}{23}, \frac{18}{5}],\\
\text{ or } \nabla u_{3} \in& L^{r}(0, T; L^{p}(\mathbb{R}^{3})), \hspace{3mm} \frac{3}{p} + \frac{2}{r} \leq \frac{11}{6}, \hspace{3mm} r \in [\frac{24}{5}, \infty],\nonumber
\end{align}
then the solution is regular (see also [3, 39] for similar results on $u_{3}, \nabla u_{3}$). For the MHD system, in particular the authors in [17] showed that if $(u,b)$ solves (2a)-(2c) with $N = 3, \nu, \eta > 0$ and
\begin{equation}
u_{3}, b \in L^{r}(0, T; L^{p}(\mathbb{R}^{3})), \hspace{3mm} \frac{3}{p} + \frac{2}{r} \leq \frac{3}{4} + \frac{1}{2p}, \hspace{3mm} p > \frac{10}{3},
\end{equation}
then the solution pair $(u,b)$ remains smooth for all time. In [29], the author reduced this constraint on $u_{3}, b$ to $u_{3}, b_{1}, b_{2}$ in special cases making use of the special structure of (2b). For more interesting component reduction results of the regularity criterion, we refer to e.g. [4, 5, 12, 16, 21, 24, 28, 30, 36]. In particular, the authors in [7] obtained a regularity criterion for the three-dimensional NSE in terms of only $u_{3}$ in a scaling-invariant norm, although no longer $L_{T}^{r}L_{x}^{p}$-space (see also [33]). In relevance to our discussion below, we already emphasize that every component reduction result listed here is of the case $N = 3$.
We now motivate the study of (1a)-(1b), (2a)-(2c) in fourth dimension specifically. It has been realized by many mathematicians working in the research direction of the NSE that the dimension four deserves special attention (see e.g. Section 4 [18]). The significance of the fourth dimension for the NSE (and six-dimensional stationary NSE) has motivated much investigation in the research direction of partial regularity theory (see e.g. [8, 10, 25]); we also recall (6) which holds only for $N = 3, 4$. In fact, fourth dimension being a certain threshold to the component reduction regularity criteria can be seen clearly as follows. To the best of the author's knowledge, all such component reduction results to the systems (1a)-(1b) and (2a)-(2c) are obtained through an $H^{1}$-estimate. Due to Lemma 2.3, higher regularity follows once we show that the solution e.g. $u$ in the case of the NSE (1a)-(1b) satisfies $\int_{0}^{T} \lVert \nabla u\rVert_{L^{N}(\mathbb{R}^{N})}^{2}d\tau < \infty$. This implies that because $H^{1}(\mathbb{R}^{N}) \hookrightarrow L^{N}(\mathbb{R}^{N})$ only for $N =2, 3, 4$ but not $N > 4$ by Sobolev embedding, $H^{1}$-bound, from which $u\in L^{2}(0, T; H^{2}(\mathbb{R}^{N}))$ follows from the dissipative term, is sufficient for higher regularity only if $N = 2, 3, 4$. Thus, in dimension strictly higher than four, one needs to bound beyond $H^{1}$-norm; however, because the decomposition of the non-linear terms is the most important ingredient of component reduction results (see Proposition 3.1), this will complicate the proof significantly. To the best of the author's knowledge, component reduction results for dimension strictly larger than three does not exist in the literature.
Let us also discuss the two major obstacles in extending the component reduction results of regularity criteria from dimension three to four. In the case of the NSE (1a)-(1b) with $N = 3, \nu > 0$, the standard procedure to obtain a criteria in terms of $u_{3}$ may be to, e.g. first estimate every partial derivative except the last and hence $\lVert \nabla_{1,2} u\rVert_{L^{2}}$ and in this process separate $u_{3}$ in the non-linear term:
\begin{equation}
\int (u\cdot\nabla) u \cdot\Delta_{1,2} u dx \leq c \int \lvert u_{3} \rvert \lvert \nabla u\rvert \lvert \nabla\nabla_{1,2} u\rvert dx
\end{equation}
where $\nabla_{1,2} = (\partial_{1}, \partial_{2}, 0), \Delta_{1,2} = \sum_{k=1}^{2}\partial_{kk}^{2}$ (cf. [20] Lemma 2.3). Thereafter, upon a full gradient and hence an $H^{1}$-estimate, on the non-linear term one separates $\lvert \nabla_{1,2} u\rvert$:
\begin{equation}
\int (u\cdot\nabla) u \cdot\Delta u dx \leq c\int \lvert \nabla_{1,2} u\rvert \lvert \nabla u\rvert^{2} dx
\end{equation}
(cf. [39]) so that the $\lVert \nabla_{1,2} u\rVert_{L^{2}}$-estimate may be applied in (10). In the case $N =4$, it seems difficult to separate $u_{3}$ or even $u_{3}$ and $u_{4}$ in $\int (u\cdot\nabla) u \cdot \Delta_{1,2,3}u dx$. Our first key observation is that we can separate $u_{3}, u_{4}$ from $\int (u\cdot\nabla) u \cdot \Delta_{1,2}u dx$ (See Proposition 3.1). However, this leaves two other directions, namely $x_{3}, x_{4}$, instead of only one in contrast to the case $N = 3$ and disables us to obtain an inequality analogous to (10) upon the full $H^{1}$-estimate due to a sum of this type:
\begin{align*}
\sum_{j=1}^{4}\sum_{i,k=3}^{4}\int \partial_{k}u_{i}\partial_{i}u_{j}\partial_{k}u_{j} dx
\end{align*}
(see (43)). We observe that in the three-dimensional case, $i,j$ and $k$ sum up to only 3 so that using $\nabla\cdot u = 0$ from (1b), one may deduce
\begin{align*}
\sum_{j=1}^{3}\sum_{i,k=3}^{3}\int \partial_{k}u_{i}\partial_{i}u_{j}\partial_{k}u_{j} dx =& \sum_{j=1}^{3} \int \partial_{3}u_{3}\partial_{3}u_{j}\partial_{3}u_{j} dx \\
=& -\sum_{j=1}^{3} \int (\partial_{1}u_{1} + \partial_{2}u_{2})\partial_{3}u_{j}\partial_{3}u_{j}dx
\end{align*}
and hence (10) follows. However, in the four-dimensional case, there are cross-terms such as $\partial_{3}u_{4}$ which disables us to reach (10). Our second key observation is that the non-linear term may be seen as an operator as a sum of
\begin{equation*}
u\cdot\nabla = \sum_{i=1}^{4}u_{i}\partial_{i} = \sum_{i=1}^{2}u_{i}\partial_{i} + \sum_{i=3}^{4}u_{i}\partial_{i}
\end{equation*}
so that in the first sum, the $\nabla_{1,2}$-estimate may be applied while in the second, use our hypothesis on $u_{3}, u_{4}$ (see (43) and also (46)).
We now present our results:
\begin{theorem}
Let $N = 4$ and
\begin{equation}
u \in C([0,T); H^{s}(\mathbb{R}^{4})) \cap L^{2}([0,T); H^{s+1}(\mathbb{R}^{4}))
\end{equation}
be the solution to the NSE (1a)-(1b) for a given $u_{0} \in H^{s}(\mathbb{R}^{4}), s > 4$. Suppose $u_{3}, u_{4}$ with their corresponding $p_{i}, r_{i}, i = 3, 4$ satisfy the following roles of $f$:
\begin{equation}
\int_{0}^{T} \lVert f\rVert_{L^{p_{i}}}^{r_{i}} d\tau \leq c, \hspace{3mm} \frac{4}{p_{i}} + \frac{2}{r_{i}} \leq \frac{1}{p_{i}} + \frac{1}{2}, \hspace{3mm} 6 < p_{i} \leq \infty,
\end{equation}
or $\sup_{t\in [0, T]}\lVert f(t) \rVert_{L^{6}}$ being sufficiently small. Then $u$ remains in the same regularity class (11) on $[0, T']$ for some $T'> T$.
\end{theorem}
\begin{theorem}
Let $N = 4$ and $u$ in the regularity class of (11) be the solution to the NSE (1a)-(1b) for a given $u_{0} \in H^{s}(\mathbb{R}^{4}), s > 4$. Suppose $\nabla u_{3}, \nabla u_{4}$ with their corresponding $p_{i}, r_{i}, i = 3, 4$ satisfy the following roles of $f$:
\begin{equation}
\int_{0}^{T} \lVert f\rVert_{L^{p_{i}}}^{r_{i}} d\tau \leq c, \hspace{3mm} \frac{4}{p_{i}} + \frac{2}{r_{i}} \leq \begin{cases}
\frac{5}{4} + \frac{1}{p_{i}}, &\text{ if } \frac{12}{5} < p_{i} \leq 4\\
1 + \frac{2}{p_{i}}, &\text{ if } 4 < p_{i} \leq \infty
\end{cases}, \hspace{3mm}
\end{equation}
or $\sup_{t\in [0, T]}\lVert f(t) \rVert_{L^{\frac{12}{5}}}$ being sufficiently small. Then $u$ remains in the same regularity class (11) on $[0, T']$ for some $T'> T$.
\end{theorem}
\begin{theorem}
Let $N = 4$ and
\begin{equation}
u, b \in C([0,T); H^{s}(\mathbb{R}^{4})) \cap L^{2}([0,T); H^{s+1}(\mathbb{R}^{4}))
\end{equation}
be the solution pair to the MHD system (2a)-(2c) for a given $u_{0}, b_{0} \in H^{s}(\mathbb{R}^{4}), s > 4$. Suppose $u_{3}, u_{4}, b$ with their corresponding $p_{i}, r_{i}, i = 3, 4, b$ satisfy the following roles of $f$:
\begin{equation}
\int_{0}^{T} \lVert f\rVert_{L^{p_{i}}}^{r_{i}} d\tau \leq c, \hspace{3mm} \frac{4}{p_{i}} + \frac{2}{r_{i}} \leq \frac{1}{p_{i}} + \frac{1}{2}, \hspace{3mm} 6 < p_{i} \leq \infty,
\end{equation}
or $\sup_{t\in [0, T]}\lVert f(t) \rVert_{L^{6}}$ being sufficiently small. Then $u,b$ remain in the same regularity class (14) on $[0, T']$ for some $T'> T$.
\end{theorem}
\begin{theorem}
Let $N = 4$ and $u,b$ in the regularity class of (14) be the solution pair to the MHD system (2a)-(2c) for a given $u_{0}, b_{0} \in H^{s}(\mathbb{R}^{4}), s > 4$. Suppose $\nabla u_{3}, \nabla u_{4}, \nabla b$ with their corresponding $p_{i}, r_{i}, i = 3, 4, b$ satisfy the following roles of $f$:
\begin{equation}
\int_{0}^{T} \lVert f\rVert_{L^{p_{i}}}^{r_{i}} d\tau \leq c, \hspace{1mm} \frac{4}{p_{i}} + \frac{2}{r_{i}} \leq \begin{cases}
\frac{5}{4} + \frac{1}{p_{i}}, &\text{ if } \frac{12}{5} < p_{i} \leq 4\\
1 + \frac{2}{p_{i}}, &\text{ if } 4 < p_{i} \leq \infty
\end{cases}, \hspace{1mm}
\end{equation}
or $\sup_{t\in [0, T]}\lVert f (t)\rVert_{L^{\frac{12}{5}}}$ being sufficiently small. Then $u,b$ remain in the same regularity class (14) on $[0, T']$ for some $T'> T$.
\end{theorem}
\begin{theorem}
Let $N = 4$ and $u$ in the regularity class of (11) be the solution to the NSE (1a)-(1b) for a given $u_{0} \in H^{s}(\mathbb{R}^{4}), s > 4$. Suppose $\partial_{3} \pi, \partial_{4}\pi$ with their corresponding $p_{i}, r_{i}, i = 3, 4$ satisfy the following roles of $f$:
\begin{equation}
\int_{0}^{T} \lVert f\rVert_{L^{p_{i}}}^{r_{i}} d\tau \leq c, \hspace{3mm} \frac{4}{p_{i}} + \frac{2}{r_{i}} < \frac{8}{3}, \hspace{3mm} \frac{12}{7} < p_{i} < 6.
\end{equation}
Then $u$ remains in the same regularity class (11) on $[0, T']$ for some $T'> T$.
\end{theorem}
\begin{remark}
\begin{enumerate}
\item In comparison of Theorem 1.1 with (4), Theorem 1.2 with (5), Theorem 1.5 with (6), we may consider the results of this manuscript as component reduction of many previous work. Moreover, in comparison of Theorems 1.1 and 1.2 with (7), Theorem 1.3 with (8), we may consider the results of this manuscript as four-dimension extension of many previous work in three-dimension.
\item The Lemma 2.3 of [20] has found much applications, e.g. in the study on the anisotropic NSE (e.g. [35]). We note that our Proposition 3.1 can be readily generalized further to any $\mathbb{R}^{N}, N \geq 3$; we chose to state the case $N =4$ for the simplicity of presentation.
\item In [32], the author showed that for dimensions $N = 3, 4, 5$, $N$-many component regularity criteria may be reduced to $(N-1)$ many components for the generalized MHD system following the method in [28]; the results in [32] and this manuscript do not cover each other. In [31] the author also obtained a regularity criteria of $N$-dimensional porous media equation governed by Darcy's law in terms of one partial derivative of the scalar-valued solution. The method in [31] cannot be applied to (1a)-(1b), (2a)-(2c).
\end{enumerate}
\end{remark}
In the Preliminaries section, we set up notations and state key facts. Local theory is well-known (cf. [23]); hence, by the standard argument of continuation of local theory, we only need to obtain $H^{s}$-bounds. We present the proofs of Theorems 1.3, 1.4 and 1.5. Because the NSE is the MHD system at $b \equiv 0$, the proofs of Theorem 1.3 and 1.4 immediately deduce Theorems 1.1 and 1.2 respectively. Thereafter, we conclude with a brief further discussion.
\section{Preliminaries}
Throughout the rest of the manuscript, we shall assume $\nu, \eta = 1$ for simplicity. For brevity, we write $\int f$ for $\int_{\mathbb{R}^{N}} f(x) dx$ and $A \lesssim_{a,b} B$ when there exists a constant $c \geq 0$ of significant dependence only on $a, b$ such that $A \leq c B$, similarly $A \approx_{a,b} B$ in case $A = cB$. We denote the fractional Laplacian operator $\Lambda^{s} \triangleq (-\Delta)^{\frac{s}{2}}$ and
\begin{align*}
&W(t) \triangleq (\lVert \nabla_{1,2} u\rVert_{L^{2}}^{2} + \lVert \nabla_{1,2}b\rVert_{L^{2}}^{2})(t), \hspace{8mm} X(t) \triangleq (\lVert \nabla u\rVert_{L^{2}}^{2} + \lVert \nabla b\rVert_{L^{2}}^{2})(t),\\
&Y(t) \triangleq (\lVert \nabla\nabla_{1,2} u\rVert_{L^{2}}^{2} + \lVert \nabla\nabla_{1,2} b\rVert_{L^{2}}^{2})(t) , \hspace{3mm} Z(t) \triangleq (\lVert \Delta u\rVert_{L^{2}}^{2} + \lVert \Delta b\rVert_{L^{2}}^{2})(t).
\end{align*}
The following is a special case of Troisi's inequality (cf. [13]). The proof of the case $N = 3$ in the Appendix of [5] can be readily generalized to the case $N = 4$:
\begin{lemma}
Let $f \in C_{0}^{\infty} (\mathbb{R}^{4})$. Then
\begin{equation}
\lVert f\rVert_{L^{4}} \lesssim \lVert \partial_{1} f\rVert_{L^{2}}^{\frac{1}{4}} \lVert \partial_{2} f\rVert_{L^{2}}^{\frac{1}{4}} \lVert \partial_{3} f\rVert_{L^{2}}^{\frac{1}{4}} \lVert \partial_{4} f\rVert_{L^{2}}^{\frac{1}{4}}.
\end{equation}
\end{lemma}
We will use the following elementary inequality frequently:
\begin{equation}
(a+b)^{p} \leq 2^{p}(a^{p} + b^{p}), \hspace{3mm} \text{for } 0 \leq p < \infty \text{ and } a, b \geq 0.
\end{equation}
We will also use the following commutator estimate to prove another lemma concerning higher regularity:
\begin{lemma}
(cf. [19]) Let $f,g$ be smooth such that $\nabla f \in L^{p_{1}}, \Lambda^{s-1}g \in L^{p_{2}}, \Lambda^{s}f \in L^{p_{3}}, g \in L^{p_{4}}, p \in (1,\infty), \frac{1}{p} = \frac{1}{p_{1}}+\frac{1}{p_{2}} = \frac{1}{p_{3}} + \frac{1}{p_{4}}, p_{2}, p_{3} \in (1, \infty), s > 0.$ Then
\begin{equation*}
\lVert \Lambda^{s}(fg) - f\Lambda^{s}g\rVert_{L^{p}} \lesssim (\lVert \nabla f\rVert_{L^{p_{1}}}\lVert \Lambda^{s-1}g\rVert_{L^{p_{2}}} + \lVert \Lambda^{s}f\rVert_{L^{p_{3}}}\lVert g\rVert_{L^{p_{4}}}).
\end{equation*}
\end{lemma}
An immediate application of Lemma 2.2 gives the following result:
\begin{lemma}
Let $(u,b)$ be the solution to the MHD system (2a)-(2c) in $[0,T]$ with $u_{0}, b_{0} \in H^{s}(\mathbb{R}^{N}), N \geq 3, s > 2 + \frac{N}{2}$. Then if $\int_{0}^{T} \lVert \nabla u\rVert_{L^{N}}^{2} + \lVert \nabla b\rVert_{L^{N}}^{2} d\tau \lesssim 1$, then
\begin{equation*}
\sup_{t\in [0,T]} (\lVert \Lambda^{s}u\rVert_{L^{2}}^{2} + \lVert \Lambda^{s} b\rVert_{L^{2}}^{2})(t) + \int_{0}^{T} \lVert \Lambda^{s}\nabla u\rVert_{L^{2}}^{2} + \lVert \Lambda^{s} \nabla b\rVert_{L^{2}}^{2}d\tau \lesssim 1.
\end{equation*}
\end{lemma}
\begin{proof}
This is a standard computation; we sketch it for completeness. We apply $\Lambda^{s}$ on (2a)-(2b), take $L^{2}$-inner products with $\Lambda^{s}u, \Lambda^{s} b$ respectively to obtain
\begin{align*}
& \frac{1}{2} \partial_{t} (\lVert \Lambda^{s} u\rVert_{L^{2}}^{2} + \lVert \Lambda^{s} b\rVert_{L^{2}}^{2}) + \lVert \Lambda^{s} \nabla u\rVert_{L^{2}}^{2} + \lVert \Lambda^{s} \nabla b\rVert_{L^{2}}^{2}\\
=& -\int [\Lambda^{s} ((u\cdot\nabla) u) - u\cdot\nabla \Lambda^{s} u]\cdot\Lambda^{s} u - \int [\Lambda^{s} ((u\cdot\nabla) b) - u\cdot\nabla \Lambda^{s} b] \cdot\Lambda^{s} b\\
&+ \int [\Lambda^{s} ((b\cdot\nabla) b) - b\cdot\nabla \Lambda^{s} b]\cdot\Lambda^{s} u + \int [\Lambda^{s} ((b\cdot\nabla) u) - b\cdot\nabla \Lambda^{s} u] \cdot\Lambda^{s} b\\
\lesssim &(\lVert \nabla u\rVert_{L^{N}} + \lVert \nabla b\rVert_{L^{N}})(\lVert \Lambda^{s} u\rVert_{L^{2}} + \lVert \Lambda^{s} b\rVert_{L^{2}}) (\lVert \Lambda^{s} \nabla u\rVert_{L^{2}} + \lVert \Lambda^{s} \nabla b\rVert_{L^{2}})\\
\leq& \frac{1}{2}(\lVert \Lambda^{s} \nabla u\rVert_{L^{2}}^{2} + \lVert \Lambda^{s} \nabla b\rVert_{L^{2}}^{2}) + c(\lVert \nabla u\rVert_{L^{N}}^{2} + \lVert \nabla b\rVert_{L^{N}}^{2})(\lVert \Lambda^{s} u\rVert_{L^{2}}^{2} + \lVert \Lambda^{s} b\rVert_{L^{2}}^{2})
\end{align*}
by H$\ddot{o}$lder's inequalities, Lemma 2.2, Sobolev embedding of $\dot{H}^{1}(\mathbb{R}^{N}) \hookrightarrow L^{\frac{2N}{N-2}}(\mathbb{R}^{N})$, Young's inequalities and (19). Thus, after absorbing, Gronwall's inequality completes the proof of Lemma 2.3.
\end{proof}
Due to Lemma 2.3, the proof of our theorems are complete once we obtain $H^{1}$-bound.
\section{Proof of Theorem 1.3}
\subsection{$\lVert \nabla_{1,2} u\rVert_{L^{2}}^{2} + \lVert \nabla_{1,2} b\rVert_{L^{2}}^{2}$-estimate}
We first prove an important decomposition which we present as a proposition:
\begin{proposition}
Let $N = 4$ and $(u,b)$ be the solution pair to the MHD system (2a)-(2c). Then
\begin{align}
&\int (u\cdot\nabla) u \cdot \Delta_{1,2} u + (u\cdot\nabla) b \cdot \Delta_{1,2} b - (b\cdot\nabla) b \cdot \Delta_{1,2} u - (b\cdot\nabla) u \cdot \Delta_{1,2} b\nonumber \\
\lesssim& \int (\lvert u_{3} \rvert + \lvert u_{4}\rvert) \lvert \nabla u\rvert \lvert \nabla\nabla_{1,2} u\rvert + \lvert b\rvert (\lvert \nabla u\rvert + \lvert \nabla b\rvert)(\lvert \nabla\nabla_{1,2} u\rvert + \lvert \nabla\nabla_{1,2} b\rvert).
\end{align}
Moreover,
\begin{align}
&\int (u\cdot\nabla) u \cdot \Delta_{1,2} u + (u\cdot\nabla) b \cdot \Delta_{1,2} b - (b\cdot\nabla) b \cdot \Delta_{1,2} u - (b\cdot\nabla) u \cdot \Delta_{1,2} b \nonumber \\
\lesssim& \int (\lvert \nabla u_{3} \rvert + \lvert \nabla u_{4} \rvert) \lvert \nabla_{1,2} u\rvert \lvert \nabla u\rvert + \lvert \nabla b\rvert \lvert \nabla_{1,2} b\rvert \lvert \nabla u\rvert.
\end{align}
\end{proposition}
\begin{proof}
We write components-wise and integrate by parts to obtain
\begin{align}
&\int (u\cdot\nabla) u \cdot\Delta_{1,2} u\\
=& -\sum_{i,j=1}^{4}\sum_{k=1}^{2}\int\partial_{k}u_{i}\partial_{i}u_{j}\partial_{k}u_{j}\nonumber\\
=& -\sum_{j=1}^{4}\sum_{i,k=1}^{2}\int\partial_{k}u_{i}\partial_{i}u_{j}\partial_{k}u_{j} - \sum_{i=3}^{4}\sum_{j=1}^{4}\sum_{k=1}^{2}\int\partial_{k}u_{i}\partial_{i}u_{j}\partial_{k}u_{j}\nonumber\\
=& -\sum_{i,j,k=1}^{2}\int\partial_{k}u_{i}\partial_{i}u_{j}\partial_{k}u_{j} - \sum_{j=3}^{4}\sum_{i,k=1}^{2}\int\partial_{k}u_{i}\partial_{i}u_{j}\partial_{k}u_{j} - \sum_{i=3}^{4}\sum_{j=1}^{4}\sum_{k=1}^{2}\int\partial_{k}u_{i}\partial_{i}u_{j}\partial_{k}u_{j}.\nonumber
\end{align}
For the second and third integrals of (22), we integrate by parts to obtain
\begin{align}
&-\sum_{j=3}^{4}\sum_{i,k=1}^{2}\int\partial_{k}u_{i}\partial_{i}u_{j}\partial_{k}u_{j} - \sum_{i=3}^{4}\sum_{j=1}^{4}\sum_{k=1}^{2}\int\partial_{k}u_{i}\partial_{i}u_{j}\partial_{k}u_{j}\\
=& \sum_{j=3}^{4}\sum_{i,k=1}^{2} \int u_{j} \partial_{i}(\partial_{k}u_{i}\partial_{k}u_{j}) + \sum_{i=3}^{4}\sum_{j=1}^{4}\sum_{k=1}^{2} \int u_{i} \partial_{k}(\partial_{i}u_{j}\partial_{k}u_{j})\nonumber\\
\lesssim& \int (\lvert u_{3}\rvert + \lvert u_{4}\rvert) \lvert \nabla u\rvert \lvert \nabla\nabla_{1,2} u\rvert.\nonumber
\end{align}
On the other hand, we write the first integral of (22) explicitly
\begin{align}
&-\sum_{i,j,k=1}^{2}\int \partial_{k}u_{i}\partial_{i}u_{j}\partial_{k}u_{j}\\
=& -\int(\partial_{1}u_{1})^{3} + \partial_{2}u_{1}\partial_{1}u_{1}\partial_{2}u_{1} + \partial_{1}u_{1}\partial_{1}u_{2}\partial_{1}u_{2} + \partial_{2}u_{1}\partial_{1}u_{2}\partial_{2}u_{2}\nonumber\\
&+ \partial_{1}u_{2} \partial_{2}u_{1}\partial_{1}u_{1} + \partial_{2}u_{2}\partial_{2}u_{1}\partial_{2}u_{1} + \partial_{1}u_{2}\partial_{2}u_{2}\partial_{1}u_{2} + (\partial_{2}u_{2})^{3} \triangleq \sum_{i=1}^{8} I_{i}.\nonumber
\end{align}
We combine and use the incompressibility condition of $u$ to obtain
\begin{align}
&I_{1} + I_{8} = -\int (\partial_{1}u_{1})^{3} + (\partial_{2}u_{2})^{3}\\
=& \int (\partial_{1}u_{1})^{2}\partial_{2}u_{2} + (\partial_{1}u_{1})^{2} (\partial_{3}u_{3} + \partial_{4}u_{4})+ (\partial_{2}u_{2})^{2}\partial_{1}u_{1} + (\partial_{2}u_{2})^{2}(\partial_{3}u_{3} + \partial_{4}u_{4}).\nonumber
\end{align}
We combine the first and third terms to obtain
\begin{align*}
\int (\partial_{1}u_{1})^{2} \partial_{2}u_{2} + (\partial_{2}u_{2})^{2} \partial_{1}u_{1}
=- \int \partial_{1}u_{1}\partial_{2}u_{2}(\partial_{3}u_{3} + \partial_{4}u_{4})
\end{align*}
so that we may continue (25) by
\begin{align}
I_{1} + I_{8} =& -\int \partial_{1}u_{1}\partial_{2}u_{2}(\partial_{3}u_{3} + \partial_{4}u_{4})\\
&+ \int (\partial_{1}u_{1})^{2} (\partial_{3}u_{3} + \partial_{4}u_{4}) + (\partial_{2}u_{2})^{2}(\partial_{3}u_{3} + \partial_{4}u_{4})\nonumber\\
=& \int u_{3} \partial_{3} (\partial_{1}u_{1}\partial_{2}u_{2}) + u_{4}\partial_{4}(\partial_{1}u_{1}\partial_{2}u_{2})\nonumber\\
&- \int u_{3} \partial_{3}[(\partial_{1}u_{1})^{2} + (\partial_{2}u_{2})^{2}] + u_{4} \partial_{4} [(\partial_{1}u_{1})^{2} + (\partial_{2}u_{2})^{2}]\nonumber\\
\lesssim& \int (\lvert u_{3}\rvert + \lvert u_{4}\rvert) \lvert \nabla u\rvert \lvert \nabla\nabla_{1,2} u\rvert.\nonumber
\end{align}
Similarly,
\begin{align}
I_{2} + I_{6} =& -\int \partial_{2}u_{1}\partial_{1}u_{1}\partial_{2}u_{1} + \partial_{2}u_{2}\partial_{2}u_{1}\partial_{2}u_{1}\\
=& \int (\partial_{2}u_{1})^{2} (\partial_{3}u_{3} + \partial_{4}u_{4}) \lesssim \int (\lvert u_{3}\rvert + \lvert u_{4}\rvert) \lvert \nabla u\rvert \lvert \nabla\nabla_{1,2}u\rvert,\nonumber
\end{align}
\begin{align}
I_{3} + I_{7} =& -\int \partial_{1}u_{1}\partial_{1}u_{2}\partial_{1}u_{2} + \partial_{1}u_{2}\partial_{2}u_{2}\partial_{1}u_{2}\\
=& \int (\partial_{1}u_{2})^{2} (\partial_{3}u_{3} + \partial_{4}u_{4}) \lesssim \int (\lvert u_{3}\rvert + \lvert u_{4}\rvert) \lvert \nabla u\rvert \lvert \nabla\nabla_{1,2} u\rvert,\nonumber
\end{align}
\begin{align}
I_{4} + I_{5} =& -\int \partial_{2}u_{1} \partial_{1}u_{2} \partial_{2}u_{2} + \partial_{1}u_{2}\partial_{2}u_{1} \partial_{1}u_{1}\\
=& \int \partial_{2}u_{1} \partial_{1}u_{2} (\partial_{3}u_{3} + \partial_{4}u_{4}) \lesssim \int (\lvert u_{3} \rvert + \lvert u_{4} \rvert) \lvert \nabla u\rvert \lvert \nabla \nabla_{1,2} u\rvert.\nonumber
\end{align}
Next, we may estimate the other three terms as follows:
\begin{align}
&\int (u\cdot\nabla) b \cdot \Delta_{1,2} b - (b\cdot\nabla) b \cdot \Delta_{1,2} u - (b\cdot\nabla) u \cdot \Delta_{1,2} b\\
=& -\sum_{i,j=1}^{4}\sum_{k=1}^{2}\int \partial_{k}u_{i} \partial_{i}b_{j}\partial_{k}b_{j} + \sum_{i,j=1}^{4}\sum_{k=1}^{2}\int \partial_{k}b_{i}\partial_{i}b_{j}\partial_{k}u_{j} + \partial_{k}b_{i}\partial_{i}u_{j}\partial_{k}b_{j}\nonumber\\
=& \sum_{i,j=1}^{4}\sum_{k=1}^{2} \int \partial_{k}u_{i}b_{j} \partial_{ik}^{2} b_{j} - \sum_{i,j=1}^{4} \sum_{k=1}^{2} \int \partial_{k}b_{i} b_{j} \partial_{ik}^{2}u_{j} + b_{i}\partial_{k}(\partial_{i}u_{j}\partial_{k}b_{j})\nonumber\\
\lesssim& \int \lvert b\rvert (\lvert \nabla u\rvert + \lvert \nabla b\rvert) (\lvert \nabla\nabla_{1,2} u\rvert + \lvert \nabla\nabla_{1,2} b\rvert).\nonumber
\end{align}
Applying (26)-(29) in (24), considering (22), (23) and (30) we obtain (20). Now we go back to (22) and estimate the second and third integrals by
\begin{align}
&-\sum_{j=3}^{4} \sum_{i,k=1}^{2} \int \partial_{k}u_{i}\partial_{i}u_{j} \partial_{k}u_{j} - \sum_{i=3}^{4} \sum_{j=1}^{4}\sum_{k=1}^{2}\int \partial_{k}u_{i}\partial_{i}u_{j}\partial_{k}u_{j}\\
\lesssim& \sum_{j=3}^{4}\sum_{k=1}^{2}\int \lvert \partial_{k} u \rvert \lvert \nabla u_{j} \rvert \lvert \partial_{k} u\rvert + \sum_{i=3}^{4}\sum_{k=1}^{2}\int \lvert \nabla u_{i} \rvert \lvert \nabla u\rvert \lvert \partial_{k} u\rvert \nonumber\\
\lesssim& \int (\lvert \nabla u_{3} \rvert + \lvert \nabla u_{4} \rvert) \lvert \nabla_{1,2} u\rvert \lvert \nabla u\rvert\nonumber
\end{align}
whereas continuing from (26),
\begin{align}
I_{1} + I_{8} =& -\int \partial_{1}u_{1} \partial_{2}u_{2} (\partial_{3} u_{3} + \partial_{4} u_{4}) + \left((\partial_{1}u_{1})^{2} + (\partial_{2}u_{2})^{2}\right) (\partial_{3}u_{3} + \partial_{4} u_{4})\\
\lesssim& \int \lvert \nabla_{1,2} u\rvert^{2} ( \lvert \partial_{3}u_{3} \rvert + \lvert \partial_{4}u_{4}\rvert),\nonumber
\end{align}
continuing from (27),
\begin{align}
I_{2} + I_{6} = \int (\partial_{2}u_{1})^{2}(\partial_{3}u_{3} + \partial_{4}u_{4}) \lesssim \int \lvert \nabla_{1,2} u\rvert^{2}(\lvert \partial_{3}u_{3} \rvert + \lvert \partial_{4}u_{4}\rvert),
\end{align}
continuing from (28),
\begin{align}
I_{3} + I_{7} = \int (\partial_{1}u_{2})^{2}(\partial_{3} u_{3} + \partial_{4}u_{4})
\lesssim \int \lvert \nabla_{1,2} u\rvert^{2}(\lvert \partial_{3}u_{3} \rvert + \lvert \partial_{4}u_{4}\rvert),
\end{align}
and continuing from (29),
\begin{align}
I_{4} + I_{5} = \int \partial_{2}u_{1}\partial_{1}u_{2}(\partial_{3}u_{3} + \partial_{4}u_{4})
\lesssim \int \lvert \nabla_{1,2} u\rvert^{2}(\lvert \partial_{3}u_{3} \rvert + \lvert \partial_{4}u_{4}\rvert).
\end{align}
Thus, considering (31)-(35) in (22), we have shown
\begin{equation}
\int (u\cdot\nabla) u \cdot \Delta_{1,2} u \lesssim \int (\lvert \nabla u_{3} \rvert + \lvert \nabla u_{4} \rvert) \lvert \nabla_{1,2} u\rvert \lvert \nabla u\rvert.
\end{equation}
Next, we estimate continuing from (30)
\begin{align}
&\int (u\cdot\nabla) b \cdot \Delta_{1,2} b - (b\cdot\nabla) b \cdot \Delta_{1,2} u - (b\cdot\nabla) u \cdot \Delta_{1,2}b\\
=& -\sum_{i,j=1}^{4}\sum_{k=1}^{2}\int \partial_{k}u_{i} \partial_{i}b_{j}\partial_{k}b_{j} - \partial_{k}b_{i}\partial_{i}b_{j}\partial_{k}u_{j} - \partial_{k}b_{i}\partial_{i}u_{j}\partial_{k}b_{j}\lesssim \int \lvert \nabla b \rvert \lvert \nabla_{1,2} b\rvert \lvert \nabla u\rvert.\nonumber
\end{align}
Considering (36) and (37), we obtain (21). This completes the proof of Proposition 3.1.
\end{proof}
With this proposition, we now obtain our first estimate:
\begin{proposition}
Let $N =4$ and $(u,b)$ be the solution pair to the MHD system (2a)-(2c) that satisfies the hypothesis of Theorem 1.3. Then $\forall \hspace{1mm} t \in (0, T], p_{i} \in [6, \infty]$,
\begin{align*}
&\sup_{\tau \in [0, t]} W(\tau) + \int_{0}^{t}Y(\tau) d\tau\\
\leq& W(0) + c\sum_{i=3}^{4}\int_{0}^{t} \lVert u_{i} \rVert_{L^{p_{i}}}^{\frac{2p_{i}}{p_{i} - 2}}X^{\frac{p_{i} - 4}{p_{i} - 2}}(\tau)Z^{\frac{2}{p_{i} - 2}}(\tau)+ \lVert b\rVert_{L^{p_{b}}}^{\frac{2p_{b}}{p_{b} - 2}}X^{\frac{p_{b} - 4}{p_{b} - 2}}(\tau)Z^{\frac{2}{p_{b} - 2}}(\tau)d\tau
\end{align*}
with the usual convention at the case $p_{i} = \infty, i = 3 , 4, b$; i.e. $\frac{2p_{i}}{p_{i} - 2} = 2, \frac{p_{i} - 4}{p_{i} - 2} = 1, \frac{2}{p_{i} - 2} = 0$.
\end{proposition}
\begin{proof}
We treat the case $6 \leq p_{i} < \infty \hspace{1mm} \forall \hspace{1mm} i = 3, 4, b$ first. We take $L^{2}$-inner products on (2a)-(2b) with $-\Delta_{1,2}u, -\Delta_{1,2} b$ respectively to obtain in sum
\begin{align}
&\frac{1}{2} \partial_{t} W(t) + Y(t)\\
\lesssim& \sum_{i=3}^{4} \int \lvert u_{i} \rvert \lvert \nabla u\rvert \lvert \nabla\nabla_{1,2} u\rvert + \lvert b\rvert (\lvert \nabla u\rvert + \lvert \nabla b\rvert)(\lvert \nabla\nabla_{1,2} u\rvert + \lvert \nabla\nabla_{1,2} b\rvert) \triangleq II_{1} + II_{2} \nonumber
\end{align}
by (20). Now we estimate
\begin{align}
II_{1} \approx& \sum_{i=3}^{4} \int \lvert u_{i} \rvert \lvert \nabla u\rvert \lvert \nabla\nabla_{1,2} u\rvert\lesssim \sum_{i=3}^{4} \lVert u_{i} \rVert_{L^{p_{i}}} \lVert \nabla u\rVert_{L^{\frac{2p_{i}}{p_{i}-2}}} \lVert \nabla\nabla_{1,2} u\rVert_{L^{2}}\\
\lesssim& \sum_{i=3}^{4}\lVert u_{i} \rVert_{L^{p_{i}}} \lVert \nabla u\rVert_{L^{2}}^{\frac{p_{i}-4}{p_{i}}} \lVert \nabla u\rVert_{L^{4}}^{\frac{4}{p_{i}}} \lVert \nabla\nabla_{1,2} u\rVert_{L^{2}}\nonumber\\
\lesssim&\sum_{i=3}^{4} \lVert u_{i} \rVert_{L^{p_{i}}} \lVert \nabla u\rVert_{L^{2}}^{\frac{p_{i}-4}{p_{i}}} \lVert \nabla\nabla_{1,2} u\rVert_{L^{2}}^{\frac{2}{p_{i}} + 1} \lVert \Delta u\rVert_{L^{2}}^{\frac{2}{p_{i}}} \nonumber\\
\leq& \frac{1}{4} \lVert \nabla\nabla_{1,2} u\rVert_{L^{2}}^{2} + c\sum_{i=3}^{4}\lVert u_{i} \rVert_{L^{p_{i}}}^{\frac{2p_{i}}{p_{i} -2}} X^{\frac{p_{i} - 4}{p_{i} - 2}}(t)Z^{\frac{2}{p_{i} -2}}(t)\nonumber
\end{align}
by H$\ddot{o}$lder's and interpolation inequalities, (18) and Young's inequalities. Similarly,
\begin{align}
II_{2} \approx& \int \lvert b\rvert ( \lvert \nabla u\rvert + \lvert \nabla b\rvert) ( \lvert \nabla\nabla_{1,2} u \rvert + \lvert \nabla\nabla_{1,2} b\rvert)\\
\lesssim& \lVert b\rVert_{L^{p_{b}}} (\lVert \nabla u\rVert_{L^{2}}^{\frac{p_{b} - 4}{p_{b}}} + \lVert \nabla b\rVert_{L^{2}}^{\frac{p_{b} - 4}{p_{b}}})(\lVert \nabla u\rVert_{L^{4}}^{\frac{4}{p_{b}}} + \lVert \nabla b\rVert_{L^{4}}^{\frac{4}{p_{b}}}) (\lVert \nabla\nabla_{1,2} u\rVert_{L^{2}} + \lVert \nabla\nabla_{1,2} b\rVert_{L^{2}})\nonumber\\
\lesssim& \lVert b\rVert_{L^{p_{b}}} X^{\frac{p_{b} - 4}{2p_{b}}}(t)(\lVert \nabla\nabla_{1,2} u\rVert_{L^{2}}^{\frac{2}{p_{b}}} \lVert \Delta u\rVert_{L^{2}}^{\frac{2}{p_{b}}} + \lVert \nabla\nabla_{1,2} b\rVert_{L^{2}}^{\frac{2}{p_{b}}}\lVert \Delta b\rVert_{L^{2}}^{\frac{2}{p_{b}}})Y^{\frac{1}{2}}(t)\nonumber\\
\leq& \frac{1}{4} Y(t) + c\lVert b\rVert_{L^{p_{b}}}^{\frac{2p_{b}}{p_{b} - 2}}X^{\frac{p_{b} - 4}{p_{b} - 2}}(t)Z^{\frac{2}{p_{b} - 2}}(t)\nonumber
\end{align}
by H$\ddot{o}$lder's and interpolation inequalities, (19), (18) and Young's inequality. In sum of (39) and (40) in (38), after absorbing and integrating over time $[0,t], t \in (0, T]$, we obtain the desired result in case $6 \leq p_{i} < \infty$. In case, $p_{i} = \infty$, the estimate is in fact simpler: we have
\begin{align*}
II_{1} \lesssim& \sum_{i=3}^{4} \lVert u_{i} \rVert_{L^{\infty}} \lVert \nabla u\rVert_{L^{2}} \lVert \nabla\nabla_{1,2} u\rVert_{L^{2}}
\leq \frac{1}{4} \lVert \nabla\nabla_{1,2} u\rVert_{L^{2}}^{2} + c\sum_{i=3}^{4}\lVert u_{i} \rVert_{L^{\infty}}^{2} \lVert \nabla u\rVert_{L^{2}}^{2},\\
II_{2} \lesssim& \lVert b\rVert_{L^{\infty}} (\lVert \nabla u\rVert_{L^{2}} + \lVert \nabla b\rVert_{L^{2}})(\lVert \nabla\nabla_{1,2} u\rVert_{L^{2}} + \lVert \nabla\nabla_{1,2} b\rVert_{L^{2}})
\leq \frac{1}{4}Y(t) + c\lVert b\rVert_{L^{\infty}}^{2} X(t).
\end{align*}
Thus, in case $p_{i} = \infty$, Proposition 3.2 holds with $\frac{2p_{i}}{p_{i} -2} = 2, \frac{p_{i} -4}{p_{i} -2} = 1, \frac{2}{p_{i} -2} = 0$.
\end{proof}
\subsection{$\lVert \nabla u\rVert_{L^{2}}^{2} + \lVert \nabla b\rVert_{L^{2}}^{2}$-estimate}
The next important step of the proof is to make use of the $\lVert \nabla_{1,2} u\rVert_{L^{2}}^{2} + \lVert \nabla_{1,2} b\rVert_{L^{2}}^{2}$-estimate to obtain the bound on $\lVert \nabla u\rVert_{L^{2}}^{2} + \lVert \nabla b\rVert_{L^{2}}^{2}$, which requires another key decomposition (see (43), (46)).
\begin{proposition}
Let $N =4$ and $(u,b)$ be the solution pair to the MHD system (2a)-(2c) that satisfies the hypothesis of Theorem 1.3. Then
\begin{align*}
\sup_{t \in [0,T]} X(t) + \int_{0}^{T} Z(\tau) d\tau \lesssim 1.
\end{align*}
\end{proposition}
\begin{proof}
Firstly, we assume $6 \leq p_{i} < \infty$ again. We take $L^{2}$-inner products on (2a)-(2b) with $(-\Delta u, -\Delta b)$ respectively to obtain
\begin{align}
&\frac{1}{2} \partial_{t} X(t) + Z(t)\\
=& \int (u\cdot\nabla) u \cdot \Delta_{1,2} u + (u\cdot\nabla) u \cdot \Delta_{3,4} u + (u\cdot\nabla) b \cdot \Delta_{1,2} b + (u\cdot\nabla) b \cdot \Delta_{3,4} b \nonumber\\
& - (b\cdot\nabla) b \cdot \Delta_{1,2} u - (b\cdot\nabla) b \cdot \Delta_{3,4} u -(b\cdot\nabla) u \cdot \Delta_{1,2} b - (b\cdot\nabla) u \cdot \Delta_{3,4} b\triangleq \sum_{i=1}^{8} III_{i}.\nonumber
\end{align}
From (38)-(40), we already have the estimates of
\begin{align}
&III_{1} + III_{3} + III_{5} + III_{7}\lesssim II_{1} + II_{2}\\
\lesssim& \sum_{i=3}^{4}\lVert u_{i} \rVert_{L^{p_{i}}} \lVert \nabla u\rVert_{L^{2}}^{\frac{p_{i} -4}{p_{i}}} \lVert \nabla \nabla_{1,2} u\rVert_{L^{2}}^{\frac{2+p_{i}}{p_{i}}} \lVert \Delta u\rVert_{L^{2}}^{\frac{2}{p_{i}}} + \lVert b\rVert_{L^{p_{b}}} X^{\frac{p_{b} - 4}{2p_{b}}}(t)Z^{\frac{1}{p_{b}}}(t)Y^{\frac{1}{p_{b}} + \frac{1}{2}}(t)\nonumber\\
\leq& \frac{1}{16} Z(t) + c\sum_{i=3}^{4}(\lVert u_{i} \rVert_{L^{p_{i}}}^{\frac{2p_{i}}{p_{i} - 4}} + \lVert b\rVert_{L^{p_{b}}}^{\frac{2p_{b}}{p_{b} - 4}})X(t)\nonumber
\end{align}
by Young's inequalities. Next, we work on $III_{2}$, which we first integrate by parts and decompose as follows:
\begin{align}
III_{2} =&\int (u\cdot\nabla) u \cdot \Delta_{3,4} u= -\sum_{i,j=1}^{4}\sum_{k=3}^{4}\int \partial_{k}u_{i}\partial_{i}u_{j}\partial_{k}u_{j}\\
=& -\sum_{i=1}^{2}\sum_{j=1}^{4}\sum_{k=3}^{4}\int \partial_{k}u_{i}\partial_{i}u_{j}\partial_{k}u_{j} - \sum_{j=1}^{4}\sum_{i,k=3}^{4}\int\partial_{k}u_{i}\partial_{i}u_{j}\partial_{k}u_{j}\nonumber\\
=& -\sum_{i=1}^{2}\sum_{j=1}^{4}\sum_{k=3}^{4}\int\partial_{k}u_{i}\partial_{i}u_{j}\partial_{k}u_{j} + \sum_{j=1}^{4}\sum_{i,k=3}^{4}\int u_{i}\partial_{k}(\partial_{i}u_{j}\partial_{k}u_{j})\nonumber\\
\lesssim& \int \lvert \nabla u\rvert^{2} \lvert \nabla_{1,2} u\rvert + \sum_{i=3}^{4} \int \lvert u_{i} \rvert \lvert \nabla u\rvert \lvert \nabla^{2} u\rvert \triangleq IV_{1} + IV_{2}.\nonumber
\end{align}
We estimate
\begin{align}
IV_{1} \approx& \int \lvert \nabla_{1,2} u\rvert \lvert \nabla u\rvert^{2} \lesssim \lVert \nabla_{1,2} u\rVert_{L^{2}} \lVert \nabla u\rVert_{L^{4}}^{2}\\
\lesssim& \lVert \nabla_{1,2} u\rVert_{L^{2}} \lVert \nabla\nabla_{1,2} u\rVert_{L^{2}} \lVert \Delta u\rVert_{L^{2}}\lesssim W^{\frac{1}{2}}(t)Y^{\frac{1}{2}}(t) Z^{\frac{1}{2}}(t)\nonumber
\end{align}
by H$\ddot{o}$lder's inequalities and (18). On the other hand,
\begin{align}
IV_{2} \lesssim& \sum_{i=3}^{4}\lVert u_{i} \rVert_{L^{p_{i}}} \lVert \nabla u\rVert_{L^{\frac{2p_{i}}{p_{i} -2}}} \lVert \nabla^{2} u\rVert_{L^{2}}\\
\lesssim& \sum_{i=3}^{4}\lVert u_{i}\rVert_{L^{p_{i}}} \lVert \nabla u\rVert_{L^{2}}^{1-\frac{4}{p_{i}}} \lVert \Delta u\rVert_{L^{2}}^{1+\frac{4}{p_{i}}} \leq \frac{1}{16}Z(t) + c\sum_{i=3}^{4}\lVert u_{i} \rVert_{L^{p_{i}}}^{\frac{2p_{i}}{p_{i} -4}} X(t) \nonumber
\end{align}
by H$\ddot{o}$lder's, Gagliardo-Nirenberg and Young's inequalities. Next, again we carefully decompose
\begin{align}
III_{4} =& \int (u\cdot\nabla) b \cdot \Delta_{3,4} b = -\sum_{i,j=1}^{4}\sum_{k=3}^{4}\int \partial_{k}u_{i}\partial_{i}b_{j}\partial_{k}b_{j} \\
=& -\sum_{i=1}^{2}\sum_{j=1}^{4}\sum_{k=3}^{4}\int \partial_{k}u_{i}\partial_{i}b_{j}\partial_{k}b_{j} - \sum_{j=1}^{4}\sum_{i,k=3}^{4}\int \partial_{k}u_{i}\partial_{i}b_{j}\partial_{k}b_{j}\nonumber\\
=& -\sum_{i=1}^{2}\sum_{j=1}^{4}\sum_{k=3}^{4}\int \partial_{k}u_{i}\partial_{i}b_{j}\partial_{k}b_{j} + \sum_{j=1}^{4}\sum_{i,k=3}^{4}\int u_{i}\partial_{k} (\partial_{i}b_{j}\partial_{k}b_{j})\nonumber\\
\lesssim&\int \lvert \nabla u\rvert \lvert \nabla_{1,2} b\rvert \lvert \nabla b\rvert + \sum_{i=3}^{4} \int \lvert u_{i}\rvert \lvert \nabla b\rvert \lvert \nabla^{2} b\rvert \triangleq IV_{3} + IV_{4}.\nonumber
\end{align}
We estimate
\begin{align}
IV_{3} \approx& \int \lvert \nabla u\rvert \lvert \nabla_{1,2} b\rvert \lvert \nabla b\rvert \lesssim \lVert \nabla_{1,2} b\rVert_{L^{2}} \lVert \nabla u\rVert_{L^{4}} \lVert \nabla b\rVert_{L^{4}}\\
\lesssim& \lVert \nabla_{1,2} b\rVert_{L^{2}} \lVert \nabla\nabla_{1,2} u\rVert_{L^{2}}^{\frac{1}{2}} \lVert \nabla\nabla_{1,2} b\rVert_{L^{2}}^{\frac{1}{2}} \lVert \Delta u\rVert_{L^{2}}^{\frac{1}{2}} \lVert \Delta b\rVert_{L^{2}}^{\frac{1}{2}}\lesssim W^{\frac{1}{2}}(t) Y^{\frac{1}{2}}(t) Z^{\frac{1}{2}}(t)\nonumber
\end{align}
by H$\ddot{o}$lder's inequalities, (18) and Young's inequalities. On the other hand, we estimate similarly to $IV_{2}$ in (45),
\begin{align}
IV_{4} \lesssim& \sum_{i=3}^{4} \lVert u_{i} \rVert_{L^{p_{i}}} \lVert \nabla b\rVert_{L^{\frac{2p_{i}}{p_{i} -2}}} \lVert \nabla^{2} b\rVert_{L^{2}}\\
\lesssim& \sum_{i=3}^{4} \lVert u_{i}\rVert_{L^{p_{i}}} \lVert \nabla b\rVert_{L^{2}}^{1-\frac{4}{p_{i}}} \lVert \Delta b\rVert_{L^{2}}^{1+\frac{4}{p_{i}}} \leq \frac{1}{16} Z(t) + c\sum_{i=3}^{4} \lVert u_{i} \rVert_{L^{p_{i}}}^{\frac{2p_{i}}{p_{i} -4}} X(t) \nonumber
\end{align}
by H$\ddot{o}$lder's, Gagliardo-Nirenberg and Young's inequalities. Finally, similarly to $IV_{2}$ in (45) again
\begin{align}
III_{6} + III_{8} =& -\int (b\cdot\nabla) b \cdot \Delta_{3,4} u + (b\cdot\nabla) u \cdot \Delta_{3,4} b\\
\lesssim& \lVert b\rVert_{L^{p_{b}}} (\lVert \nabla b\rVert_{L^{2}}^{1-\frac{4}{p_{b}}} + \lVert \nabla u\rVert_{L^{2}}^{1-\frac{4}{p_{b}}}) (\lVert \Delta u\rVert_{L^{2}}^{1+ \frac{4}{p_{b}}} + \lVert \Delta b\rVert_{L^{2}}^{1+ \frac{4}{p_{b}}})\nonumber\\
\leq& \frac{1}{16} Z(t) + c\lVert b\rVert_{L^{p_{b}}}^{\frac{2p_{b}}{p_{b} - 4}}X(t)\nonumber
\end{align}
by H$\ddot{o}$lder's, Gagliardo-Nirenberg and Young's inequalities. Thus, applying (42)-(49) in (41), we obtain after absorbing
\begin{align}
\frac{1}{2} \partial_{t} X + \frac{1}{2}Z(t)
\lesssim \sum_{i=3}^{4}(\lVert u_{i} \rVert_{L^{p_{i}}}^{\frac{2p_{i}}{p_{i} - 4}} + \lVert b\rVert_{L^{p_{b}}}^{\frac{2p_{b}}{p_{b} - 4}})X(t) + W^{\frac{1}{2}}(t)Y^{\frac{1}{2}}(t)Z^{\frac{1}{2}}(t).
\end{align}
Now we assume $6 < p_{i} < \infty$. Integrating over $[0,t], t \in (0, T]$, we obtain
\begin{align*}
&X(t) + \int_{0}^{t} Z(\tau) d\tau\\
\leq& X(0) + c\sum_{i=3}^{4}\int_{0}^{t} (\lVert u_{i} \rVert_{L^{p_{i}}}^{\frac{2p_{i}}{p_{i} - 4}} + \lVert b\rVert_{L^{p_{b}}}^{\frac{2p_{b}}{p_{b} - 4}})X(\tau) d\tau + c\int_{0}^{t} W^{\frac{1}{2}}(\tau) Y^{\frac{1}{2}}(\tau) Z^{\frac{1}{2}}(\tau) d\tau.
\end{align*}
We focus only on the last integral which we bound by a constant multiples of
\begin{align*}
&\sup_{\tau \in [0,t]} W^{\frac{1}{2}}(\tau) \left(\int_{0}^{t} Y(\tau) d\tau\right)^{\frac{1}{2}} \left(\int_{0}^{t} Z(\tau) d\tau\right)^{\frac{1}{2}}\\
\lesssim& \left(W(0) + \sum_{i=3}^{4}\int_{0}^{t} \lVert u_{i} \rVert_{L^{p_{i}}}^{\frac{2p_{i}}{p_{i} - 2}}X^{\frac{p_{i} - 4}{p_{i} - 2}}(\tau)Z^{\frac{2}{p_{i} - 2}}(\tau) + \lVert b\rVert_{L^{p_{b}}}^{\frac{2p_{b}}{p_{b} - 2}}X^{\frac{p_{b} - 4}{p_{b} - 2}}(\tau)Z^{\frac{2}{p_{b} - 2}}(\tau)d\tau \right)\\
&\times \left(\int_{0}^{t} Z(\tau) d\tau\right)^{\frac{1}{2}}\\
\lesssim& \left(\int_{0}^{t} Z(\tau)d\tau\right)^{\frac{1}{2}} +\sum_{i=3}^{4}\left(\int_{0}^{t} \lVert u_{i} \rVert_{L^{p_{i}}}^{\frac{2p_{i}}{p_{i} -4}} X(\tau) d\tau\right)^{\frac{p_{i} -4}{p_{i} -2}} \left(\int_{0}^{t} Z(\tau) d\tau\right)^{\frac{p_{i} + 2}{2(p_{i} -2)}}\\
&+ \left(\int_{0}^{t} \lVert b \rVert_{L^{p_{b}}}^{\frac{2p_{b}}{p_{b} -4}} X(\tau) d\tau\right)^{\frac{p_{b} -4}{p_{b} -2}} \left(\int_{0}^{t} Z(\tau) d\tau\right)^{\frac{p_{b} + 2}{2(p_{b} -2)}}\\
\leq& \frac{1}{2}\int_{0}^{t} Z(\tau) d\tau + c(1 + \sum_{i=3}^{4}\left(\int_{0}^{t} \lVert u_{i} \rVert_{L^{p_{i}}}^{\frac{4p_{i}}{p_{i} - 6}}X(\tau) d\tau\right) \left(\int_{0}^{t} X(\tau) d\tau\right)^{\frac{p_{i} - 2}{p_{i} - 6}}\\
& \hspace{25mm} + \left(\int_{0}^{t} \lVert b \rVert_{L^{p_{b}}}^{\frac{4p_{b}}{p_{b} - 6}}X(\tau) d\tau\right) \left(\int_{0}^{t} X(\tau) d\tau\right)^{\frac{p_{b} - 2}{p_{b} - 6}})\\
\leq& \frac{1}{2} \int_{0}^{t} Z(\tau) d\tau + c\left(1+ \sum_{i=3}^{4}\int_{0}^{t} (\lVert u_{i}\rVert_{L^{p_{i}}}^{\frac{4p_{i}}{p_{i} -6}} + \lVert b\rVert_{L^{p_{b}}}^{\frac{4p_{b}}{p_{b} -6}}) X(\tau) d\tau \right)
\end{align*}
by H$\ddot{o}$lder's inequalities, Proposition 3.2, Young's inequalities and (3). After absorbing, Gronwall's inequality implies the desired result in case $6 < p_{i} < \infty, r_{i} < \infty$.
We now consider the case $p_{i} = \infty$, assuming for the simplicity of presentation that $p_{3} = p_{4} = p_{b} = \infty$. Firstly, we could have computed in contrast to (42), (43), (46) and (49) respectively
\begin{align}
&III_{1} + III_{3} + III_{5} + III_{7}\lesssim II_{1} + II_{2}\\
\lesssim& \sum_{i=3}^{4} \lVert u_{i} \rVert_{L^{\infty}} \lVert \nabla u\rVert_{L^{2}} \lVert \Delta u\rVert_{L^{2}} + \lVert b\rVert_{L^{\infty}}(\lVert \nabla u\rVert_{L^{2}} + \lVert \nabla b\rVert_{L^{2}}) (\lVert \Delta u\rVert_{L^{2}} + \lVert \Delta b\rVert_{L^{2}})\nonumber\\
\leq& \frac{1}{16} Z(t) + c\sum_{i=3}^{4}(\lVert u_{i} \rVert_{L^{\infty}}^{2} + \lVert b\rVert_{L^{\infty}}^{2})X(t),\nonumber
\end{align}
\begin{align}
III_{2} \leq& IV_{1} + IV_{2} \\
\lesssim& \lVert \nabla_{1,2} u\rVert_{L^{2}} \lVert \nabla u\rVert_{L^{4}}^{2} + \sum_{i=3}^{4}\lVert u_{i} \rVert_{L^{\infty}}\lVert \nabla u\rVert_{L^{2}} \lVert \nabla^{2} u\rVert_{L^{2}}\nonumber\\
\leq& \frac{1}{16} Z(t) + c\left(W^{\frac{1}{2}}(t)Y^{\frac{1}{2}}(t)Z^{\frac{1}{2}}(t) + \sum_{i=3}^{4}\lVert u_{i} \rVert_{L^{\infty}}^{2}X(t)\right),\nonumber
\end{align}
\begin{align}
III_{4} \lesssim& IV_{3} + IV_{4}\\
\lesssim& \lVert \nabla u\rVert_{L^{4}} \lVert \nabla_{1,2} b\rVert_{L^{2}} \lVert \nabla b\rVert_{L^{4}} + \sum_{i=3}^{4}\lVert u_{i} \rVert_{L^{\infty}}\lVert \nabla b\rVert_{L^{2}} \lVert \Delta b\rVert_{L^{2}}\nonumber\\
\lesssim& \lVert \nabla_{1,2} b\rVert_{L^{2}} \lVert \nabla\nabla_{1,2} u\rVert_{L^{2}}^{\frac{1}{2}} \lVert \Delta u\rVert_{L^{2}}^{\frac{1}{2}} \lVert \nabla\nabla_{1,2} b\rVert_{L^{2}}^{\frac{1}{2}} \lVert \Delta b\rVert_{L^{2}}^{\frac{1}{2}} + \sum_{i=3}^{4}\lVert u_{i} \rVert_{L^{\infty}}\lVert \nabla b\rVert_{L^{2}} \lVert \Delta b\rVert_{L^{2}}\nonumber\\
\leq& \frac{1}{16} Z(t) + c\left(W^{\frac{1}{2}}(t)Y^{\frac{1}{2}}(t) Z^{\frac{1}{2}}(t) + \sum_{i=3}^{4}\lVert u_{i} \rVert_{L^{\infty}}^{2} X(t)\right),\nonumber
\end{align}
\begin{align}
III_{6} + III_{8}
\lesssim& \lVert b\rVert_{L^{\infty}} (\lVert \nabla b\rVert_{L^{2}} + \lVert \nabla u\rVert_{L^{2}})(\lVert \nabla^{2} u\rVert_{L^{2}} + \lVert \nabla^{2} b\rVert_{L^{2}})\\
\leq& \frac{1}{16} Z(t) + c\lVert b\rVert_{L^{\infty}}^{2} X(t)\nonumber
\end{align}
all by H$\ddot{o}$lder's and Young's inequalities and (18) only in (52) and (53). Thus applying (51)-(54) in (41), absorbing and integrating in time $[0, t]$, we obtain
\begin{align*}
&X(t) + \frac{3}{2}\int_{0}^{t}Z(\tau) d\tau\\
\leq& X(0) + c\sum_{i=3}^{4}\int_{0}^{t} (\lVert u_{i} \rVert_{L^{\infty}}^{2} + \lVert b\rVert_{L^{\infty}}^{2})X(\tau) d\tau\\
& \hspace{8mm} + c\sup_{\tau \in [0,t]} W^{\frac{1}{2}}(\tau) \left(\int_{0}^{t} Y(\tau) d\tau\right)^{\frac{1}{2}} \left(\int_{0}^{t} Z(\tau) d\tau \right)^{\frac{1}{2}}\\
\leq& \frac{1}{2} \int_{0}^{t}Z(\tau) d\tau + c \sum_{i=3}^{4}\int_{0}^{t} (\lVert u_{i} \rVert_{L^{\infty}}^{2}+ \lVert b\rVert_{L^{\infty}}^{2})X(\tau) d\tau\\
+&c\left(W(0) + \sum_{i=3}^{4}\int_{0}^{t} (\lVert u_{i} \rVert_{L^{\infty}}^{2} + \lVert b\rVert_{L^{\infty}}^{2}) X(\tau) d\tau)\right)^{2}\\
\leq& \frac{1}{2} \int_{0}^{t}Z(\tau) d\tau\\
&+c\left( \sum_{i=3}^{4}\int_{0}^{t} (\lVert u_{i} \rVert_{L^{\infty}}^{2} + \lVert b\rVert_{L^{\infty}}^{2})X(\tau) d\tau + 1+ \sum_{i=3}^{4}\int_{0}^{t} (\lVert u_{i} \rVert_{L^{\infty}}^{4} + \lVert b\rVert_{L^{\infty}}^{4}) X(\tau) d\tau \right)
\end{align*}
by H$\ddot{o}$lder's inequality, Proposition 3.2, Young's inequality, (19) and (3). This completes the proof in case $p_{i} = \infty$.
We now prove the second statement of Theorem 1.3, namely the smallness result when $p_{i} = 6, r_{i} = \infty$. For simplicity of presentation, we assume $p_{i} = 6 \hspace{1mm} \forall \hspace{1mm} i = 3, 4, b$. We integrate in time on (50) to obtain
\begin{align*}
&X(t) + \int_{0}^{t} Z(\tau) d\tau\\
\leq& X(0) + c(\sum_{i=3}^{4}\sup_{\tau \in [0,t]} (\lVert u_{i} \rVert_{L^{6}}^{6} + \lVert b\rVert_{L^{6}}^{6})(\tau) \int_{0}^{t} X(\tau)d\tau + \sup_{\tau \in [0,t]} W^{\frac{1}{2}}(t) \left(\int_{0}^{t} Y(\tau) d\tau\right)^{\frac{1}{2}}\\
& \hspace{10mm} \times\left(\int_{0}^{t}Z(\tau) d\tau\right)^{\frac{1}{2}})\\
\lesssim& 1 + \left(W(0) + \sum_{i=3}^{4}\int_{0}^{t} (\lVert u_{i} \rVert_{L^{6}}^{3} + \lVert b\rVert_{L^{6}}^{3}) X^{\frac{1}{2}}(\tau) Z^{\frac{1}{2}}(\tau) d\tau \right) \left(\int_{0}^{t} Z(\tau) d\tau \right)^{\frac{1}{2}}\\
\lesssim& 1 + \left(\int_{0}^{t} Z(\tau) d\tau \right)^{\frac{1}{2}} + \sum_{i=3}^{4}\sup_{\tau \in [0,t]} (\lVert u_{i} \rVert_{L^{6}}^{3} + \lVert b\rVert_{L^{6}}^{3})(\tau) \left(\int_{0}^{t} X(\tau) d\tau\right)^{\frac{1}{2}} \left(\int_{0}^{t} Z(\tau) d\tau\right)\\
\leq& \frac{1}{2} \int_{0}^{t}Z(\tau) d\tau + c
\end{align*}
for $\sum_{i=3}^{4}\sup_{\tau \in [0,t]} (\lVert u_{i} \rVert_{L^{6}}^{3} + \lVert b\rVert_{L^{6}}^{3})(\tau)$ sufficiently small where we used H$\ddot{o}$lder's inequality, Proposition 3.2, Young's inequality and (3). Absorbing, Gronwall's inequality completes the proof of Theorem 1.3.
\end{proof}
\section{Proof of Theorem 1.4}
We assume for simplicity of presentation that $\forall \hspace{1mm} i = 3, 4, b, p_{i} \in [\frac{12}{5}, 4]$ or $p_{i} \in [4, \infty]$. A combination of mixed cases can be obtained following the proofs below.
\begin{proposition}
Let $N =4$ and $(u,b)$ be the solution pair to the MHD system (2a)-(2c) that satisfies the hypothesis of Theorem 1.4. Then $\forall \hspace{1mm} t \in (0, T]$,
\begin{align*}
&\sup_{\tau \in [0, t]} W(\tau) + \int_{0}^{t}Y(\tau) d\tau\\
\leq&
\begin{cases}
W(0) + c\sum_{i=3}^{4} \int_{0}^{t} \lVert \nabla u_{i} \rVert_{L^{p_{i}}}^{\frac{4p_{i}}{3p_{i} -4}} X^{\frac{4(p_{i} -2)}{3p_{i} -4}}(\tau) Z^{\frac{4-p_{i}}{3p_{i} -4}}(\tau)\\
\hspace{30mm} + \lVert \nabla b\rVert_{L^{p_{b}}}^{\frac{4p_{b}}{3p_{b} -4}} X^{\frac{4(p_{b} -2)}{3p_{b} - 4}} (\tau)Z^{\frac{4-p_{b}}{3p_{b} -4}}(\tau) d\tau, & \text{ if } p_{i} \in [\frac{12}{5}, 4],\\
W(0) + c\sum_{i=3}^{4} \int_{0}^{t} (\lVert \nabla u_{i} \rVert_{L^{p_{i}}}^{\frac{p_{i}}{p_{i} -2}} + \lVert \nabla b\rVert_{L^{p_{b}}}^{\frac{p_{b}}{p_{b} -2}})X(\tau) d\tau, & \text{ if } p_{i} \in [4, \infty],
\end{cases}
\end{align*}
with the usual convention at $p_{i} = \infty, i = 3, 4, b$; i.e. $\frac{p_{i}}{p_{i} - 2} = 1$.
\end{proposition}
\begin{proof}
We first assume $p_{i} \in \left[\frac{12}{5}, 4\right]$. We take $L^{2}$-inner products of (2a)-(2b) with $-\Delta_{1,2} u, -\Delta_{1,2} b$ respectively and estimate
\begin{align}
\frac{1}{2}\partial_{t} W(t) + Y(t)
\lesssim \sum_{i=3}^{4} \int \lvert \nabla u_{i} \rvert\lvert \nabla_{1,2}u \rvert \lvert \nabla u\rvert + \lvert \nabla b\rvert \lvert \nabla_{1,2} b\rvert \lvert \nabla u\rvert
\end{align}
by (21). Now we estimate
\begin{align}
\sum_{i=3}^{4}\int \lvert \nabla u_{i} \rvert\lvert \nabla_{1,2} u\rvert \lvert \nabla u\rvert \lesssim& \sum_{i=3}^{4} \lVert \nabla u_{i} \rVert_{L^{p_{i}}} \lVert \nabla_{1,2} u\rVert_{L^{4}} \lVert \nabla u\rVert_{L^{\frac{4p_{i}}{3p_{i} -4}}} \\
\lesssim& \sum_{i=3}^{4} \lVert \nabla u_{i} \rVert_{L^{p_{i}}} \lVert \nabla\nabla_{1,2} u\rVert_{L^{2}} \lVert \nabla u\rVert_{L^{2}}^{2(\frac{p_{i} -2}{p_{i}})} \lVert \nabla u\rVert_{L^{4}}^{\frac{4-p_{i}}{p_{i}}}\nonumber\\
\lesssim& \sum_{i=3}^{4} \lVert \nabla u_{i} \rVert_{L^{p_{i}}} \lVert \nabla\nabla_{1,2} u\rVert_{L^{2}}^{\frac{4+p_{i}}{2p_{i}}} \lVert \nabla u\rVert_{L^{2}}^{2(\frac{p_{i} -2}{p_{i}})} \lVert \Delta u\rVert_{L^{2}}^{\frac{4-p_{i}}{2p_{i}}}\nonumber\\
\leq& \frac{1}{4} Y(t) + c\sum_{i=3}^{4} \lVert \nabla u_{i} \rVert_{L^{p_{i}}}^{\frac{4p_{i}}{3p_{i} - 4}}X^{\frac{4(p_{i} - 2)}{3p_{i} - 4}}(t)Z^{\frac{4-p_{i}}{3p_{i} - 4}}(t)\nonumber
\end{align}
by H$\ddot{o}$lder's inequalities, Sobolev embedding of $\dot{H}^{1}(\mathbb{R}^{4}) \hookrightarrow L^{4}(\mathbb{R}^{4})$, interpolation inequality, (18) and Young's inequality. Similarly, we obtain
\begin{align}
\int \lvert \nabla b \rvert \lvert \nabla_{1,2} b\rvert \lvert \nabla u\rvert
\lesssim& \lVert \nabla b\rVert_{L^{p_{b}}}\lVert \nabla\nabla_{1,2} b\rVert_{L^{2}} \lVert \nabla u\rVert_{L^{\frac{4p_{b}}{3p_{b} - 4}}}\\
\lesssim& \lVert \nabla b\rVert_{L^{p_{b}}} Y^{\frac{1}{2}}(t) \lVert \nabla u\rVert_{L^{2}}^{2(\frac{p_{b} - 2}{p_{b}})}\lVert \nabla u\rVert_{L^{4}}^{\frac{4-p_{b}}{p_{b}}}\nonumber\\
\lesssim& \lVert \nabla b\rVert_{L^{p_{b}}}Y^{\frac{1}{2}}(t) X^{\frac{p_{b} -2}{p_{b}}} (t) \lVert \nabla\nabla_{1,2} u\rVert_{L^{2}}^{\frac{4-p_{b}}{2p_{b}}}\lVert \Delta u\rVert_{L^{2}}^{\frac{4-p_{b}}{2p_{b}}}\nonumber\\
\leq& \frac{1}{4} Y(t) + c\lVert \nabla b\rVert_{L^{p_{b}}}^{\frac{4p_{b}}{3p_{b} - 4}}X^{\frac{4(p_{b} -2)}{3p_{b} - 4}}(t) Z^{\frac{4-p_{b}}{3p_{b} - 4}}(t).\nonumber
\end{align}
With (56) and (57) applied to (55), absorbing and integrating in time lead to
\begin{align}
&W(t) + \int_{0}^{t} Y(\tau) d\tau\\
\leq& W(0)\nonumber\\
&+c\sum_{i=3}^{4} \int_{0}^{t} \lVert \nabla u_{i} \rVert_{L^{p_{i}}}^{\frac{4p_{i}}{3p_{i} -4}}X^{\frac{4(p_{i} -2)}{3p_{i} -4}}(\tau) Z^{\frac{4-p_{i}}{3p_{i} -4}}(\tau) + \lVert \nabla b\rVert_{L^{p_{b}}}^{\frac{4p_{b}}{3p_{b} -4}} X^{\frac{4(p_{b} -2)}{3p_{b} - 4}}(\tau) Z^{\frac{4-p_{b}}{3p_{b} -4}}(\tau) d\tau.\nonumber
\end{align}
We now work on the case $4 < p_{i} < \infty$:
\begin{align}
\sum_{i=3}^{4}\int \lvert \nabla u_{i} \rvert \lvert \nabla_{1,2} u\rvert \lvert \nabla u\rvert \lesssim& \sum_{i=3}^{4} \lVert \nabla u_{i} \rVert_{L^{p_{i}}} \lVert \nabla_{1,2} u\rVert_{L^{\frac{2p_{i}}{p_{i} - 2}}}\lVert \nabla u\rVert_{L^{2}}\\
\lesssim& \sum_{i=3}^{4} \lVert \nabla u_{i} \rVert_{L^{p_{i}}} \lVert \nabla_{1,2} u\rVert_{L^{2}}^{\frac{p_{i} - 4}{p_{i}}}\lVert \nabla_{1,2} u\rVert_{L^{4}}^{\frac{4}{p_{i}}}\lVert \nabla u\rVert_{L^{2}}\nonumber\\
\lesssim& \sum_{i=3}^{4} \lVert \nabla u_{i} \rVert_{L^{p_{i}}} \lVert \nabla_{1,2} u\rVert_{L^{2}}^{\frac{p_{i} - 4}{p_{i}}}\lVert \nabla\nabla_{1,2} u\rVert_{L^{2}}^{\frac{4}{p_{i}}}\lVert \nabla u\rVert_{L^{2}}\nonumber\\
\leq& \frac{1}{4}Y(t) + c\sum_{i=3}^{4} \lVert \nabla u_{i}\rVert_{L^{p_{i}}}^{\frac{p_{i}}{p_{i} - 2}}X(t)\nonumber
\end{align}
by H$\ddot{o}$lder's and interpolation inequalities, Sobolev embedding of $\dot{H}^{1}(\mathbb{R}^{4})\hookrightarrow L^{4}(\mathbb{R}^{4})$ and Young's inequality. Similarly, we estimate
\begin{align}
\int \lvert \nabla b\rvert \lvert \nabla_{1,2} b\rvert \lvert \nabla u\rvert
\lesssim& \lVert \nabla b\rVert_{L^{p_{b}}} \lVert \nabla_{1,2} b\rVert_{L^{2}}^{\frac{p_{b} -4}{p_{b}}} \lVert \nabla_{1,2} b\rVert_{L^{4}}^{\frac{4}{p_{b}}} \lVert \nabla u\rVert_{L^{2}}\\
\lesssim& \lVert \nabla b\rVert_{L^{p_{b}}} \lVert \nabla_{1,2} b\rVert_{L^{2}}^{\frac{p_{b} -4}{p_{b}}} \lVert\nabla \nabla_{1,2} b\rVert_{L^{2}}^{\frac{4}{p_{b}}} \lVert \nabla u\rVert_{L^{2}}\nonumber\\
\leq& \frac{1}{4} Y(t) + c\lVert \nabla b\rVert_{L^{p_{b}}}^{\frac{p_{b}}{p_{b} - 2}}X(t).\nonumber
\end{align}
We apply (59) and (60) in (55), absorb and integrate in time to obtain
\begin{align}
W(t) + \int_{0}^{t} Y(\tau) d\tau
\leq W(0) + c\sum_{i=3}^{4} \int_{0}^{t} (\lVert \nabla u_{i} \rVert_{L^{p_{i}}}^{\frac{p_{i}}{p_{i} -2}} + \lVert \nabla b\rVert_{L^{p_{b}}}^{\frac{p_{b}}{p_{b} -2}}) X(\tau) d\tau.
\end{align}
The case $p_{i} = \infty$ requires only a standard modification as done in the proof of Theorem 1.3; that is,
\begin{align*}
\sum_{i=3}^{4}\int \lvert \nabla u_{i} \rvert \lvert \nabla_{1,2} u\rvert \lvert \nabla u\rvert
\lesssim \sum_{i=3}^{4} \lVert \nabla u_{i} \rVert_{L^{\infty}} \lVert \nabla_{1,2} u\rVert_{L^{2}} \lVert \nabla u\rVert_{L^{2}} \lesssim \sum_{i=3}^{4} \lVert \nabla u_{i} \rVert_{L^{\infty}} X(t),
\end{align*}
\begin{align*}
\int \lvert \nabla b\rvert \lvert \nabla_{1,2} b\rvert \lvert \nabla u\rvert
\lesssim \lVert \nabla b\rVert_{L^{\infty}} \lVert \nabla_{1,2} b\rVert_{L^{2}} \lVert \nabla u\rVert_{L^{2}} \lesssim \lVert \nabla b\rVert_{L^{\infty}}X(t)
\end{align*}
so that summing and integrating in time leads to the desired result. This completes the proof of Proposition 4.1.
\end{proof}
\begin{proposition}
Let $N =4$ and $(u,b)$ be the solution pair to the MHD system (2a)-(2c) that satisfies the hypothesis of Theorem 1.4. Then
\begin{align*}
\sup_{t \in [0,T]} X(t) + \int_{0}^{T} Z(\tau) d\tau \lesssim 1.
\end{align*}
\end{proposition}
\begin{proof}
Similarly to the proof of Theorem 1.3, we estimate from (41). For $p_{i} \in [\frac{12}{5}, 4]$, we continue our estimate from (55), (56) and (57) to obtain
\begin{align}
&III_{1} +III_{3} + III_{5} + III_{7}\\
\lesssim& \sum_{i=3}^{4} \lVert \nabla u_{i} \rVert_{L^{p_{i}}} \lVert \nabla\nabla_{1,2} u\rVert_{L^{2}}^{\frac{4+p_{i}}{2p_{i}}}\lVert \nabla u\rVert_{L^{2}}^{2(\frac{p_{i} - 2}{p_{i}})}\lVert \Delta u\rVert_{L^{2}}^{\frac{4-p_{i}}{2p_{i}}}\nonumber\\
&+ \lVert \nabla b\rVert_{L^{p_{b}}} Y^{\frac{4+p_{b}}{4p_{b}}}(t) X^{\frac{p_{b} - 2}{p_{b}}}(t)\lVert \Delta u\rVert_{L^{2}}^{\frac{4-p_{b}}{2p_{b}}}\nonumber\\
\leq& \frac{1}{16} Z(t) + c\sum_{i=3}^{4} \left(\lVert \nabla u_{i} \rVert_{L^{p_{i}}}^{\frac{p_{i}}{p_{i} - 2}} + \lVert \nabla b\rVert_{L^{p_{b}}}^{\frac{p_{b}}{p_{b} -2}} \right)X(t)\nonumber
\end{align}
by Young's inequality. We now decompose integrating by parts
\begin{align}
III_{2}
=& -\sum_{i,j=1}^{4}\sum_{k=3}^{4}\int \partial_{k}u_{i}\partial_{i}u_{j} \partial_{k}u_{j}\\
=& -\sum_{i=1}^{2}\sum_{j=1}^{4}\sum_{k=3}^{4}\int \partial_{k}u_{i}\partial_{i}u_{j}\partial_{k}u_{j} - \sum_{j=1}^{4}\sum_{i,k=3}^{4}\int \partial_{k}u_{i}\partial_{i}u_{j}\partial_{k}u_{j}\nonumber\\
\lesssim& \int \lvert \nabla u\rvert^{2} \lvert \nabla_{1,2} u\rvert + \sum_{i=3}^{4} \int \lvert \nabla u_{i} \rvert \lvert \nabla u\rvert^{2} \triangleq V_{1} + V_{2}\nonumber
\end{align}
where $V_{1}$ is estimated identically as $IV_{1}$ in (44) while we estimate
\begin{align}
V_{2} \lesssim& \sum_{i=3}^{4} \lVert \nabla u_{i} \rVert_{L^{p_{i}}} \lVert \nabla u\rVert_{L^{\frac{2p_{i}}{p_{i} - 1}}}^{2} \\
\lesssim& \sum_{i=3}^{4} \lVert \nabla u_{i} \rVert_{L^{p_{i}}} \lVert \nabla u\rVert_{L^{2}}^{2(\frac{p_{i} - 2}{p_{i}})}\lVert \Delta u\rVert_{L^{2}}^{2(\frac{2}{p_{i}})}
\leq \frac{1}{16} Z(t) + c\sum_{i=3}^{4} \lVert \nabla u_{i} \rVert_{L^{p_{i}}}^{\frac{p_{i}}{p_{i} - 2}}X(t)\nonumber
\end{align}
by H$\ddot{o}$lder's, Gagliardo-Nirenberg and Young's inequalities. Next, we decompose
\begin{align}
III_{4} =& -\sum_{i,j=1}^{4}\sum_{k=3}^{4}\int \partial_{k}u_{i} \partial_{i}b_{j}\partial_{k}b_{j} \\
=& -\sum_{i=1}^{2}\sum_{j=1}^{4}\sum_{k=3}^{4}\int \partial_{k}u_{i}\partial_{i}b_{j}\partial_{k}b_{j} - \sum_{j=1}^{4}\sum_{i,k=3}^{4}\int \partial_{k}u_{i}\partial_{i}b_{j}\partial_{k}b_{j}\nonumber\\
\lesssim& \int \lvert \nabla u\rvert \lvert \nabla_{1,2} b\rvert \lvert \nabla b\rvert + \sum_{i=3}^{4} \int \lvert \nabla u_{i} \rvert \lvert \nabla b\rvert^{2} \triangleq V_{3} + V_{4}\nonumber
\end{align}
where we estimate $V_{3}$ as $IV_{3}$ in (47) while same estimate of $V_{2}$ in (64) lead to
\begin{align}
V_{4} \lesssim& \sum_{i=3}^{4} \lVert \nabla u_{i} \rVert_{L^{p_{i}}} \lVert \nabla b\rVert_{L^{\frac{2p_{i}}{p_{i} - 1}}}^{2}\\
\lesssim& \sum_{i=3}^{4} \lVert \nabla u_{i} \rVert_{L^{p_{i}}} \lVert \nabla b\rVert_{L^{2}}^{2(\frac{p_{i} - 2}{p_{i}})}\lVert \Delta b\rVert_{L^{2}}^{2(\frac{2}{p_{i}})} \leq \frac{1}{16} Z(t) + c\sum_{i=3}^{4} \lVert \nabla u_{i} \rVert_{L^{p_{i}}}^{\frac{p_{i}}{p_{i} - 2}}X(t).\nonumber
\end{align}
Finally,
\begin{align}
III_{6} + III_{8} =& \sum_{i,j=1}^{4}\sum_{k=3}^{4}\int \partial_{k}b_{i}\partial_{i}b_{j}\partial_{k}u_{j} + \partial_{k}b_{i}\partial_{i}u_{j}\partial_{k}b_{j}\\
\lesssim& \int \lvert \nabla b\rvert^{2} \lvert \nabla u\rvert\nonumber\\
\lesssim& \lVert \nabla b\rVert_{L^{p_{b}}}\lVert \nabla b\rVert_{L^{\frac{2p_{b}}{p_{b} - 1}}}\lVert \nabla u\rVert_{L^{\frac{2p_{b}}{p_{b} - 1}}}\nonumber\\
\lesssim& \lVert \nabla b\rVert_{L^{p_{b}}} \lVert \nabla b\rVert_{L^{2}}^{\frac{p_{b} - 2}{p_{b}}}\lVert \Delta b\rVert_{L^{2}}^{\frac{2}{p_{b}}} \lVert \nabla u\rVert_{L^{2}}^{\frac{p_{b} -2}{p_{b}}} \lVert \Delta u\rVert_{L^{2}}^{\frac{2}{p_{b}}}\nonumber\\
\leq& \frac{1}{16}Z(t) + c\lVert \nabla b\rVert_{L^{p_{b}}}^{\frac{p_{b}}{p_{b} - 2}}X(t) \nonumber
\end{align}
by H$\ddot{o}$lder's, Gagliardo-Nirenberg and Young's inequalities. Thus, we obtain by applying (62)-(67) in (41), absorbing and integrating in time,
\begin{align}
&X(t) + \frac{3}{2}\int_{0}^{t} Z(\tau) d\tau\\
\lesssim& 1 + \sum_{i=3}^{4} \int_{0}^{t} (\lVert \nabla u_{i} \rVert_{L^{p_{i}}}^{\frac{p_{i}}{p_{i} - 2}} + \lVert \nabla b\rVert_{L^{p_{b}}}^{\frac{p_{b}}{p_{b} - 2}})X(\tau) d\tau\nonumber\\
&+ \sup_{\tau \in [0,t]} W^{\frac{1}{2}}(\tau) \left(\int_{0}^{t} Y(\tau) d\tau\right)^{\frac{1}{2}} \left(\int_{0}^{t} Z(\tau) d\tau\right)^{\frac{1}{2}}\nonumber
\end{align}
where we also used H$\ddot{o}$lder's inequality. Now we assume $p_{i} \in (\frac{12}{5}, 4]$. For the last term only, we bound it by a constant multiples of
\begin{align*}
&\left(1 + \sum_{i=3}^{4} \int_{0}^{t} \lVert \nabla u_{i} \rVert_{L^{p_{i}}}^{\frac{4p_{i}}{3p_{i} -4}} X^{\frac{4(p_{i} -2)}{3p_{i} -4}}(\tau) Z^{\frac{4-p_{i}}{3p_{i} -4}}(\tau) + \lVert \nabla b\rVert_{L^{p_{b}}}^{\frac{4p_{b}}{3p_{b} -4}} X^{\frac{4(p_{b} -2)}{3p_{b} - 4}}(\tau) Z^{\frac{4-p_{b}}{3p_{b} -4}}(\tau) d\tau\right)\\
&\times \left(\int_{0}^{t}Z(\tau) d\tau\right)^{\frac{1}{2}}\\
\lesssim& \left(\int_{0}^{t}Z(\tau) d\tau\right)^{\frac{1}{2}} + \sum_{i=3}^{4} \left(\int_{0}^{t} \lVert \nabla u_{i} \rVert_{L^{p_{i}}}^{\frac{p_{i}}{p_{i} - 2}} X(\tau) d\tau\right)^{\frac{4(p_{i} - 2)}{3p_{i} - 4}}\left(\int_{0}^{t} Z(\tau) d\tau \right)^{\frac{4+p_{i}}{2(3p_{i} - 4)}}\\
&+ \left(\int_{0}^{t} \lVert \nabla b\rVert_{L^{p_{b}}}^{\frac{p_{b}}{p_{b} - 2}}X(\tau) d\tau\right)^{\frac{4(p_{b} - 2)}{3p_{b} - 4}}\left(\int_{0}^{t} Z(\tau) d\tau\right)^{\frac{4+p_{b}}{2(3p_{b} - 4)}}\\
\leq& \frac{1}{2}\int_{0}^{t}Z(\tau) d\tau\\
&+ c\left(1+ \sum_{i=3}^{4} \left(\int_{0}^{t} \lVert \nabla u_{i} \rVert_{L^{p_{i}}}^{\frac{p_{i}}{p_{i} - 2}}X(\tau) d\tau\right)^{\frac{8(p_{i} - 2)}{5p_{i} - 12}} + \left(\int_{0}^{t} \lVert \nabla b\rVert_{L^{p_{b}}}^{\frac{p_{b}}{p_{b} - 2}}X(\tau) d\tau \right)^{\frac{8(p_{b} - 2)}{5p_{b} - 12}}\right)\\
\leq& \frac{1}{2}\int_{0}^{t}Z(\tau) d\tau + c\left(1+ \sum_{i=3}^{4} \left(\int_{0}^{t} \lVert \nabla u_{i}\rVert_{L^{p_{i}}}^{\frac{8p_{i}}{5p_{i} - 12}}X(\tau) d\tau\right) + \left(\int_{0}^{t} \lVert \nabla b\rVert_{L^{p_{b}}}^{\frac{8p_{b}}{5p_{b} - 12}}X(\tau) d\tau\right)\right)
\end{align*}
due to Proposition 4.1, H$\ddot{o}$lder's and Young's inequalities and (3).
Next, we consider the case $4 < p_{i} < \infty$. We restart from (41) where we continue our estimates from (55), (59) and (60) to obtain
\begin{align}
&III_{1} + III_{3} + III_{5} + III_{7} \\
\lesssim& \sum_{i=3}^{4} \lVert \nabla u_{i} \rVert_{L^{p_{i}}} \lVert \nabla_{1,2} u\rVert_{L^{2}}^{\frac{p_{i} - 4}{p_{i}}}\lVert \nabla\nabla_{1,2} u\rVert_{L^{2}}^{\frac{4}{p_{i}}}\lVert \nabla u\rVert_{L^{2}}\nonumber\\
&+ \lVert \nabla b\rVert_{L^{p_{b}}}\lVert \nabla_{1,2} b\rVert_{L^{2}}^{\frac{p_{b} -4}{p_{b}}} \lVert \nabla\nabla_{1,2} b\rVert_{L^{2}}^{\frac{4}{p_{b}}} \lVert \nabla u\rVert_{L^{2}}\nonumber\\
\leq& \frac{1}{16} Z(t) + c\sum_{i=3}^{4} (\lVert \nabla u_{i} \rVert_{L^{p_{i}}}^{\frac{p_{i}}{p_{i} -2}} + \lVert \nabla b\rVert_{L^{p_{b}}}^{\frac{p_{b}}{p_{b} -2}} )X(t) \nonumber
\end{align}
by Young's inequality. The rest of the estimates of $III_{2}, III_{4}, III_{6}, III_{8}$ all go through as in the case $p_{i} \in [\frac{12}{5}, 4]$. Indeed, continuing from (63), we bound $III_{2} \lesssim V_{1} + V_{2}$ where $V_{1}$ is estimated as $IV_{1}$ in (44) and $V_{2}$ is estimated identically as (64). The estimates of $III_{4}$ also goes through as in (65): $III_{4} \lesssim V_{3} + V_{4}$ where $V_{3}$ is estimated as $IV_{3}$ in (47) and $V_{4}$ in (66). Finally, we use the estimate of $III_{6} + III_{8}$ in (67). Thus, in sum, after absorbing, integrating in time, we obtain
\begin{align*}
X(t) + \frac{3}{2}\int_{0}^{t}Z(\tau) d\tau
\leq X(0) &+ c\sum_{i=3}^{4}\int_{0}^{t} (\lVert \nabla u_{i} \rVert_{L^{p_{i}}}^{\frac{p_{i}}{p_{i} -2}} + \lVert \nabla b\rVert_{L^{p_{b}}}^{\frac{p_{b}}{p_{b} - 2}})X(\tau) d\tau \\
&+ c\sup_{\tau \in [0,t]} W^{\frac{1}{2}}(\tau) \left(\int_{0}^{t} Y(\tau) d\tau\right)^{\frac{1}{2}} \left(\int_{0}^{t} Z(\tau) d\tau\right)^{\frac{1}{2}}
\end{align*}
by H$\ddot{o}$lder's inequality. We bound the last term by
\begin{align*}
&c(W(0) + \sum_{i=3}^{4} \int_{0}^{t} (\lVert \nabla u_{i} \rVert_{L^{p_{i}}}^{\frac{p_{i}}{p_{i} - 2}} + \lVert \nabla b\rVert_{L^{p_{b}}}^{\frac{p_{b}}{p_{b} - 2}})X(\tau) d\tau \left(\int_{0}^{t} Z(\tau) d\tau\right)^{\frac{1}{2}}\\
\leq& \frac{1}{2}\int_{0}^{t} Z(\tau) d\tau + c\left(1+ \sum_{i=3}^{4}\left(\int_{0}^{t} (\lVert \nabla u_{i} \rVert_{L^{p_{i}}}^{\frac{p_{i}}{p_{i} - 2}} + \lVert \nabla b\rVert_{L^{p_{b}}}^{\frac{p_{b}}{p_{b} - 2}})X(\tau) d\tau\right)^{2} \right)\\
\leq& \frac{1}{2}\int_{0}^{t} Z(\tau) d\tau + c\left(1+ \sum_{i=3}^{4}\int_{0}^{t} (\lVert \nabla u_{i} \rVert_{L^{p_{i}}}^{\frac{2p_{i}}{p_{i} - 2}} + \lVert \nabla b\rVert_{L^{p_{b}}}^{\frac{2p_{b}}{p_{b} - 2}})X(\tau) d\tau\right)
\end{align*}
by Proposition 4.1, Young's and H$\ddot{o}$lder's inequalities and (3). After absorbing, Gronwall's inequality implies the desired result. We now consider the case $p_{i} = \infty$. For simplicity, we assume $p_{i} = \infty \forall \hspace{1mm} i = 3, 4, b$. We continue from (41) where we estimate in contrast to (69),
\begin{align*}
&III_{1} + III_{3} + III_{5} + III_{7}\\
\lesssim& \sum_{i=3}^{4} \int \lvert \nabla u_{i} \rvert \lvert \nabla_{1,2} u\rvert \lvert \nabla u\rvert + \lvert \nabla b\rvert \lvert \nabla_{1,2} b\rvert \lvert \nabla u\rvert \lesssim \sum_{i=3}^{4} (\lVert \nabla u_{i} \rVert_{L^{\infty}} + \lVert \nabla b\rVert_{L^{\infty}})X(t)
\end{align*}
due to (55), H$\ddot{o}$lder's and Young's inequalities. Moreover, from $III_{2} \lesssim V_{1} + V_{2}$ of (63), we estimate $V_{1}$ is estimated as $IV_{1}$ in (44) and
\begin{equation*}
V_{2} \approx \sum_{i=3}^{4}\int \lvert \nabla u_{i} \rvert \lvert \nabla u\rvert^{2} \lesssim \sum_{i=3}^{4} \lVert \nabla u_{i} \rVert_{L^{\infty}} \lVert \nabla u\rVert_{L^{2}}^{2}.
\end{equation*}
Moreover, from $III_{4} \lesssim V_{3} + V_{4}$ of (65), we have $V_{3}$ estimated as $IV_{3}$ in (47) while
\begin{equation*}
V_{4} \approx \sum_{i=3}^{4}\int \lvert \nabla u_{i} \rvert \lvert \nabla b\rvert^{2} \lesssim \sum_{i=3}^{4} \lVert \nabla u_{i} \rVert_{L^{\infty}} \lVert \nabla b\rVert_{L^{2}}^{2}.
\end{equation*}
Finally, continuing our estimate from (67),
\begin{align*}
III_{6} + III_{8} \lesssim \int \lvert \nabla b\rvert^{2} \lvert \nabla u\rvert
\lesssim \lVert \nabla b\rVert_{L^{\infty}}\lVert \nabla b\rVert_{L^{2}} \lVert \nabla u\rVert_{L^{2}}
\lesssim \lVert \nabla b\rVert_{L^{\infty}} X(t).
\end{align*}
In sum, integrating in time we obtain
\begin{align*}
&X(t) + 2\int_{0}^{t} Z(\tau) d\tau\\
\lesssim& X(0) + \sum_{i=3}^{4}\int (\lVert \nabla u_{i} \rVert_{L^{\infty}} + \lVert \nabla b\rVert_{L^{\infty}})X(\tau) d\tau\\
&+ \sup_{\tau \in [0,t]} W^{\frac{1}{2}}(\tau) \left(\int_{0}^{t} Y(\tau) d\tau\right)^{\frac{1}{2}} \left(\int_{0}^{t} Z(\tau) d\tau\right)^{\frac{1}{2}}\\
\lesssim& X(0)+ \sum_{i=3}^{4}\int (\lVert \nabla u_{i} \rVert_{L^{\infty}} + \lVert \nabla b\rVert_{L^{\infty}})X(\tau) d\tau\\
&+\left(W(0) + \sum_{i=3}^{4} \int_{0}^{t} (\lVert \nabla u_{i} \rVert_{L^{\infty}} + \lVert \nabla b\rVert_{L^{\infty}}) X(\tau) d\tau\right)\left(\int_{0}^{t} Z(\tau) d\tau \right)^{\frac{1}{2}}\\
\leq& \int_{0}^{t} Z(\tau) d\tau + c \left(1+ \sum_{i=3}^{4} (\int_{0}^{t} \lVert \nabla u_{i} \rVert_{L^{\infty}}^{2} + \lVert \nabla b\rVert_{L^{\infty}}^{2})X(\tau) d\tau)\right)
\end{align*}
by H$\ddot{o}$lder's inequality, Proposition 4.1, Young's inequality and (3).
Finally, we prove the smallness result in the case $p_{i} = \frac{12}{5}, r_{i} = \infty$, for which for simplicity of presentation, we assume $r_{i} = \infty, p_{i} = \frac{12}{5}, \forall \hspace{1mm} i = 3, 4, b$. From (68),
\begin{align*}
&X(t) + \frac{3}{2}\int_{0}^{t}Z(\tau) d\tau\\
\lesssim& X(0) + \sum_{i=3}^{4}\int_{0}^{t} (\lVert \nabla u_{i} \rVert_{L^{\frac{12}{5}}}^{6} + \lVert \nabla b\rVert_{L^{\frac{12}{5}}}^{6})X(\tau) d\tau\\
&+ \left(W(0) + \sum_{i=3}^{4}\int_{0}^{t} (\lVert \nabla u_{i} \rVert_{L^{\frac{12}{5}}}^{3} + \lVert \nabla b\rVert_{L^{\frac{12}{5}}}^{3}) X^{\frac{1}{2}}(\tau) Z^{\frac{1}{2}}(\tau) d\tau \right) \left(\int_{0}^{t} Z(\tau) d\tau\right)^{\frac{1}{2}}\\
\leq& \frac{1}{4}\int_{0}^{t}Z(\tau) d\tau + c\sum_{i=3}^{4}\int_{0}^{t} (\lVert \nabla u_{i} \rVert_{L^{\frac{12}{5}}}^{6} + \lVert \nabla b\rVert_{L^{\frac{12}{5}}}^{6})X(\tau) d\tau\\
&+ c\left(1+ \sum_{i=3}^{4} (\int_{0}^{t} (\lVert \nabla u_{i} \rVert_{L^{\frac{12}{5}}}^{3} + \lVert \nabla b\rVert_{L^{\frac{12}{5}}}^{3}) X^{\frac{1}{2}}(\tau) Z^{\frac{1}{2}}(\tau) d\tau )^{2}\right)\\
\leq& \frac{1}{4} \int_{0}^{t}Z(\tau) d\tau + c\sum_{i=3}^{4}\sup_{\tau \in [0,t]} (\lVert \nabla u_{i}\rVert_{L^{\frac{12}{5}}}^{6} + \lVert \nabla b\rVert_{L^{\frac{12}{5}}}^{6})(\tau) \int_{0}^{t}X(\tau) d\tau\\
&+ c\left(1+ \sum_{i=3}^{4}\sup_{\tau \in [0,t]} (\lVert \nabla u_{i} \rVert_{L^{\frac{12}{5}}}^{6} + \lVert \nabla b\rVert_{L^{\frac{12}{5}}}^{6})(\tau) \int_{0}^{t}X(\tau) d\tau \int_{0}^{t} Z(\tau) d\tau \right)\\
\leq& \frac{1}{2} \int_{0}^{t} Z(\tau) d\tau + c
\end{align*}
for $\sum_{i=3}^{4}\sup_{t \in [0,T]} (\lVert \nabla u_{i} \rVert_{L^{\frac{12}{5}}}^{6} + \lVert \nabla b\rVert_{L^{\frac{12}{5}}}^{6})(t)$ sufficiently small where we used H$\ddot{o}$lder's inequality, Proposition 4.1, Young's inequality, (19) and (3). This completes the proof of Theorem 1.4.
\end{proof}
\section{Proof of Theorem 1.5}
We fix $q_{i} \in (\frac{12}{7}, 6)$ and then $p_{i} = 6 + \epsilon$ for $\epsilon > 0$ sufficiently small so that $\frac{2(6+\epsilon)}{(6+\epsilon) + 1} < q_{i}$ and also $q_{i} < 6 < p_{i}$. This implies that $\forall \hspace{1mm} \epsilon > 0$ sufficiently small, we have $q_{i} \in (\frac{2p_{i}}{p_{i} + 1}, p_{i})$. Now we multiply the $i$-th component of (1a) with $\lvert u_{i} \rvert^{p_{i} -2} u_{i}$, integrate in space to obtain
\begin{align*}
\frac{1}{p_{i}} \partial_{t} \lVert u_{i} \rVert_{L^{p_{i}}}^{p_{i}} + c(p_{i}) \lVert u_{i} \rVert_{L^{2p_{i}}}^{p_{i}}
\lesssim& \lVert \partial_{i} \pi \rVert_{L^{q_{i}}} \lVert u_{i} \rVert_{L^{\frac{(p_{i} - 1) q_{i}}{q_{i} - 1}}}^{p_{i} - 1}\\
\lesssim& \lVert \partial_{i} \pi \rVert_{L^{q_{i}}} \lVert u_{i} \rVert_{L^{p_{i}}}^{\frac{p_{i}q_{i} - 2p_{i} + q_{i}}{q_{i}}}\lVert u_{i} \rVert_{L^{2p_{i}}}^{\frac{2(p_{i} - q_{i})}{q_{i}}}\nonumber\\
\leq& \frac{c(p_{i})}{2} \lVert u_{i} \rVert_{L^{2p_{i}}}^{p_{i}} + c\lVert \partial_{i} \pi \rVert_{L^{q_{i}}}^{\frac{p_{i}q_{i}}{p_{i}q_{i} - 2p_{i} + 2q_{i}}}\lVert u_{i} \rVert_{L^{p_{i}}}^{p_{i}(\frac{p_{i}q_{i} - 2p_{i} + q_{i}}{p_{i}q_{i} - 2p_{i} + 2q_{i}})}\nonumber
\end{align*}
where we used the lower bound estimate on the dissipative term of
\begin{align*}
c(p_{i})\lVert u_{i} \rVert_{L^{2p_{i}}}^{p_{i}} \approx& \lVert \lvert u_{i} \rvert^{\frac{p_{i}}{2}}\rVert_{L^{4}}^{2} \lesssim \lVert \lvert u_{i} \rvert^{\frac{p_{i}}{2}} \rVert_{\dot{H}^{1}}^{2} \approx \frac{(p_{i} - 1)4}{p_{i}^{2}}\int \lvert \nabla \lvert u_{i} \rvert^{\frac{p_{i}}{2}} \rvert^{2} = -\int \Delta u \lvert u_{i} \rvert^{p_{i} -2} u_{i}
\end{align*}
for some constant $c(p_{i})$ that depends on $p_{i}$, H$\ddot{o}$lder's, interpolation and Young's inequalities. We absorb and obtain
\begin{align*}
\frac{1}{p_{i}} \partial_{t} \lVert u_{i} \rVert_{L^{p_{i}}}^{p_{i}} + \frac{c(p_{i})}{2} \lVert u_{i} \rVert_{L^{2p_{i}}}^{p_{i}} \lesssim \lVert \partial_{i} \pi \rVert_{L^{q_{i}}}^{\frac{p_{i}q_{i}}{p_{i}q_{i} - 2p_{i} + 2q_{i}}}(1 + \lVert u_{i} \rVert_{L^{p_{i}}}^{p_{i}})
\end{align*}
by Young's inequality. By hypothesis of Theorem 1.5 and Gronwall's inequality, $\forall \hspace{1mm} \epsilon > 0$ sufficiently small we have $\sum_{i=3}^{4} \sup_{t \in [0,T]} \lVert u_{i} \rVert_{L^{p_{i}}}(t) \lesssim 1$ where $p_{i} = 6 + \epsilon$. By Theorem 1.1, the proof of Theorem 1.5 is complete.
\section{Further Discussion}
There are many results that exist for the regularity criteria component reduction theory of the three-dimensional NSE and the MHD system that we may look forward to being generalized to the four-dimensional case. We remark however that some of such results did not seem readily generalizable. We also note that in order to reduce our two-component regularity criterion for the four-dimensional NSE to one component or to extend it to higher dimension such as five, it seems to require a new approach.
\section{Acknowledgment}
The author expresses gratitude to Professor Jiahong Wu and Professor David Ullrich for their teaching and Professor Vladimir Sverak for helpful comments on the presentation of the manuscript.
|
\section{Introduction}
There is six times more mass in dark matter than baryonic matter
\footnote[2]{Throughout this review we adopt the common convention of
astronomers to refer to all standard model particles, including leptons,
as baryons.} in our Universe \citep{WMAP9,Planck2013}. For decades it has
been assumed that, because dark matter is so much more common than baryons,
dark matter dominates the gravity in the Universe, and that wherever the
dark matter is, baryons must follow. This assumption led galaxy theorists
to make predictions for the formation of galaxies using dark matter only,
neglecting baryonic physics, despite the fact that galaxies like our own
Milky Way are baryon-dominated within their inner $\sim$10kpc. In doing
so, a number of discrepancies between galaxy formation theory and
observations were identified, particularly on ``small scales,'' i.e., in
small galaxies and in the central regions of galaxies \citep{PrimackAnnalen}.
To address these problems, alternative models to the standard Cold Dark
Matter (CDM) have been explored. Recent interest has surged in Warm Dark
Matter (WDM) and Self-Interacting Dark Matter (SIDM) models as the favorite
alternatives to CDM amongst galaxy theorists.
However, there has also been a recent reconsideration of the importance
of baryonic physics in solving CDM's small scale problems. Observationally,
it is clear that energy feedback from stars and black holes operates to
alter the evolution of galaxies. For example, the existence of gas outflows
(``winds'') from galaxies seems to be ubiquitous at high redshift
\citep{Veilleux2005}.
Energetic feedback from stars (in the form of radiation pressure from young,
massive stars, momentum injection by the winds of the same stars, and supernovae)
has long been included in galaxy simulations, but only recently have simulations
achieved sufficiently high resolution to deposit this feedback in localized
regions. Localized feedback dramatically impacts the evolution of the galaxy
\citep{Governato2010, Guedes2011, Christensen2012, Agertz2013, Aumer2013,
Hopkins2013}, and drives the ubiquitous winds that we observe.
The processes that drive galaxy winds also have a dramatic impact on the
dark matter structure of galaxies. Feedback from stars can push the dark
matter out of the central $\sim$kpc by generating a repeated fluctuation
in the potential wells of galaxies \citep{Navarro1996, Read2005, deSouza2011,
Pontzen2012, Teyssier2013, diCintio2014}.
This result reconciles the dark matter density profile predicted in CDM
that is steeply rising toward the center \citep[``cuspy,''][]{Navarro1997a,
Springel2008, Navarro2010} with observations which instead prefer a
shallower density slope or even a constant dark matter density ``core''
\citep{vandenbosch2000, deblok2001, deblok2002, Simon2003, Swaters2003,
Weldrake2003, Kuzio2006, Gentile2007, Spano2008, Trachternach2008,
deblok2008, Oh2011}. Hence, one of the seemingly intractable problems
plaguing CDM theory is now thought to be potentially solved by a careful
consideration of the impact of baryonic physics \citep{Pontzen2014}.
Galaxy winds also solve another problem within CDM galaxy formation theory: the
existence of bulgeless disk galaxies. Galaxies are thought to obtain their
angular momentum through large-scale tidal torques \citep{Peebles1969,
White1984, Barnes1987, Quinn1992}. Gas and dark matter start with the
same angular momentum distribution \citep{vdBosch2002}, with a tail
of low angular momentum material that is expected to settle at the center
of galaxies. Low angular momentum gas should go on to form stars, forming
large bulges, at odds with observed small or non-existent stellar bulges
\citep{vdBosch2001, Dutton2009b}. Galaxy winds naturally arise from the
region where most star formation is occurring, in dense galaxy centers
where low angular momentum gas resides. Hence, winds naturally drive low
angular momentum material from galaxies \citep{Brook2011}, and can create
bulgeless disk galaxies \citep{Governato2010, Teyssier2013}, solving another
of CDM's small scale problems.
One of the oldest problems and one of the newest problems facing CDM galaxy
formation theory both relate to the satellites that orbit around our Milky Way
galaxy. First, simulations predict that there should be many more satellites
than we observe \citep{Moore1999, Klypin1999}. Many of these satellites are expected
to be ``dark,'' unable to have formed stars due to photoevaporation of their gas
when the Universe was re-ionized \citep{Quinn1996, Thoul1996, Barkana1999, Gnedin2000,
Okamoto2008}, though this process alone may not be enough to bring the predicted
number of massive, luminous satellites into agreement with observations
\citep{Brooks2013}. Furthermore, even if we could get the number of satellites
correct, there exists a population of satellites in simulations run without
baryons that are much more dense than we observe \citep{Boylan-kolchin2011,
Boylan-kolchin2012, Tollerud2012, Collins2014}. This latter problem is also
known as the ``Too Big to Fail'' problem because the simulated satellites are
too massive to have failed to form stars, yet we do not observe them. Again,
recent high resolution simulations have shown that baryonic effects may reconcile
both of these predictions with observations \citep{Zolotov2012, diCintio2013,
Arraki2014}. The primary physics at work is the fact that gas, unlike dark matter,
can cool. In a simulation with baryons, this cooled component adds more mass to the
center of the parent halo, creating stronger tidal forces that strip mass from the
satellite galaxies \citep{Penarrubia2010}. This enhanced tidal stripping reduces the
mass of the satellites, bringing the kinematic predictions in line with the
observational data \citep{Brooks2014}. The presence of the disk in a baryonic
simulation (which doesn't exist in a dark matter-only run because dark matter
cannot dissipate) will also fully destroy roughly 1/3 of the most massive
satellites, reducing the number of luminous, surviving satellites so that
it is consistent with observations \citep{Brooks2013, Brooks2014}.
The lesson we have learned from these studies is that baryons have
the potential to alter our expectations for the structure of dark matter
halos that form within CDM. While CDM does an excellent job of describing
the large scale structure of the Universe \citep{Hlozek2012}, we can no
longer neglect the influence of baryons when considering small scales.
It is important to note that, of the problems listed above, the creation
of bulgeless galaxies cannot be solved by any correction to the dark matter
model. Only baryonic feedback is able to explain the loss of low angular
momentum baryons from galaxies.
The fact that galaxy winds offer a single, unified solution to the existence
of both bulgeless disks and dark matter cores is tantalizing evidence that
these two problems are intimately tied together. Despite this, modifications
to the dark matter model are still being pursued as another possible explanation
for the existence of dark matter cores. Given the fact that baryonic physics
cannot be neglected (and is in fact essential to solve at least one
problem in CDM galaxy formation theory), the challenge for theorists is
to first understand the role of baryons within any viable dark matter model.
Thus, it is becoming clear that using dark matter-only simulations leads to
biased predictions for the distribution of dark matter in galaxies.
The recent successes in modeling the baryonic component of galaxies have
allowed theorists, {\it for the first time}, to realistically model dwarf
galaxies \citep{Governato2010, Governato2012, Shen2013, Trujillo-Gomez2013,
Brooks2014}. Hence, simulators are finally in a position to be able to make
predictions for observations that include the effect of baryons on galaxy
evolution.
Our advances in understanding baryonic physics require that we re-evaluate
current observational data with a new perspective. If baryons
alter the evolution of dark matter halos, what are the real limits of the
currently favored models? The goal of this review is to cast a critical
eye on the observations in light of our favored models. What are the
successes and failures of the models? What are potential paths forward to
break degeneracies and to rule out models?
In what follows, I will assume that 100\% of the dark matter follows a
given model. I do not discuss the possibility of mixed
models, (e.g., dark matter as a mix of both CDM and WDM, or WDM with
self-interactions, or that some fraction of the dark matter is dissipative).
There are currently a few intriguing signals that may be interpreted
as an indirect detection of dark matter. These results will be reviewed
in Section~\ref{section2},
but much of the power of observations lies in the ability to constrain
our favored models. I will discuss the two popular models already
mentioned, WDM and SIDM in Sections~\ref{section3} and \ref{section4},
respectively. The current observational constraints are
already hinting that WDM cannot be warm enough to substantially
differentiate it from CDM. SIDM offers a more tantalizing path forward,
and I will highlight future theoretical and observational probes to
test SIDM models.
\section{Indirect Detection}
\label{section2}
Before discussing the properties of galaxies that constrain the dark
matter model, I first discuss the evidence for a more straightforward
astrophysical signal. Two possible paths may lead to detectable
standard model particles indicative of the presence of dark matter.
The first path is annihilation of dark matter, and the second is decay
of a dark matter particle. Annihilation of dark matter in the Universe
today is an expected signal of a favored candidate for a CDM particle:
the Weakly Interacting Massive Particle (WIMP). In the WIMP model, dark
matter particles are a thermal relic that ``froze out'' of equilibrium
in the early Universe. Freeze-out occurs when the rate of annihilation
between dark matter particles is outpaced by the Hubble expansion
\citep{Gondolo1991, kolbturner}. After freeze-out, annihilation does not
significantly decrease the WIMP number density, but will continue at low
rates in the Universe today. Annihilation is particularly likely to happen
in the densest regions of the Universe, i.e., in galaxies, and particularly
in high density galaxy centers.
Despite its popularity in the theory community, it is by no means certain that
dark matter is a WIMP-like thermal relic \citep{Feng2008}. For example, if there
is a primordial excess of WIMP particles over their anti-particle pairs (as
there seems to have been with baryons to anti-baryons), the WIMPs and anti-WIMPs
may continue to annihilate until nearly all the anti-particles are
eliminated.\footnote[3]{This requires an annihilation cross-section larger than
that of a typical CDM WIMP \citep{Buckley2011b}.} The particle excess that is
left over is the relic dark matter density in the Universe today. This option
has become known as ``asymmetric dark matter'' \citep{Nussinov1985, Barr1990,
Kaplan1992, Kribs2010, Buckley2011}. As there is no particle to annihilate
against today, indirect detection from annihilation is not expected in such
models. However, a second type of indirect signal is possible if the dark matter
particles instead decay. The lifetime of such decay must be very long
\citep[$\gtrsim 10^{26}$ seconds][]{Ibarra2013}, but such models can be
constructed. Decaying dark matter has even been suggested as a solution to
CDM's small scale problems \citep[e.g.,][]{Wang2014}.
The spectral signatures of annihilating or decaying dark matter can vary
greatly between theoretical models. Annihilation of two dark matter particles
into two photons would result in a spectral line of gamma rays with energy
equal to the dark matter mass. Alternatively, annihilation could proceed into
Standard Model quarks, leptons, or $W/Z$ bosons, which provide a continuum
of gamma ray energies through their decays and bremsstrahlung. Decay, on the
other hand, is typically expected to result in a photon with an energy that
is half of the mass of the dark matter particle, yielding an emission line at
a specific wavelength rather than a spectrum. The morphology of any indirect
signal can be used to distinguish the two options (annihilation or decay), as
annihilation is proportional to the dark matter density squared, while decay
is only proportional to the density itself.
As of this writing, there is an exciting hint of an annihilation spectrum
seen from our Galactic Center. Likewise, there are also two unidentified
lines, one at 130 GeV and the other at 3.5 keV, that are being discussed as
possible indications of dark matter.
An excess distribution of gamma-rays from the Galactic Center has been
seen in {\it Fermi} Gamma-Ray Space Telescope data that is not attributable
to any known or understood source \citep{Hooper2011, Daylan2014}. While
the Galactic Center is a complicated place, full of baryonic physics that
can contribute gamma-rays \citep{Abazajian2012, Boyarsky2011}, this excess
is roughly spherical in shape and extends out to at least $\sim$10$^{\circ}$,
and likely to several kpc \citep{Daylan2014, Hooper2013, Huang2013}. It is
seen after subtraction of a model for the gas disk, and of known gamma-ray
point sources. While originally suggested to be a population of pulsars, the
extended distribution seems to rule out this possibility.\footnote[4]{Note
that this excess should also not be confused with the Fermi bubbles
\citep{Su2010}, which are likely caused by a past energetic event such as
accretion in the Galactic Center.} The excess can be fit by a dark matter
density distribution that follows a ``cuspy'' profile that scales as
$\rho \propto r^{-\gamma}$, with $\gamma \sim$1.2 \citep{Daylan2014}.
Unfortunately, no other searches for excess gamma-rays due to dark matter
annihilation have yet revealed a signal. Despite the nearness of the
Magellanic Clouds, they are gas-rich, making a gamma-ray signal from dark
matter annihilation difficult to extract from the signal of cosmic rays
interacting with the galactic interstellar medium \citep{Tasitsiomi2004,
fermilmc}. The most popular place to search for annihilating dark matter
is in the dwarf spheroidal
galaxies of the Milky Way, as they are gas-free and dark matter-dominated
\citep{Strigari2007, Strigari2008, Kuhlen2008, Kuhlen2010}. To date, no
significant detection has been found, and the dwarfs yield an upper limit
that place bounds on the WIMP model \citep{Geringer2011, fermidsph, fermidsph2,
hessdsph, hessdsph1}. However, the
dark matter densities of the dwarf spheroidals may be too low to be
detected with current measurements \citep{Walker2011, Cholis2012}. A
better hope for detection would be to find a signal from a faint,
as yet undiscovered dwarf that happens to be relatively nearby so that the
flux in gamma-rays is large \citep{He2013}. Again, a search for such a signal
in gamma-rays has not revealed any conclusive targets \citep{Buckley2010,
Belikov2012, Hooper2012}. The Dark Energy Survey (DES) offers the best
hope of identifying such a dwarf in the near future, as it will be the
first to survey the southern Galactic hemisphere for faint dwarfs
\citep{He2013}.
{\it Fermi} data is also the source of the tentative 130 GeV line
\citep{Weniger2012, Tempel2012}. This line has been suggested to be an
instrumental line \citep{Whiteson2012, Whiteson2013}, but so far no one has
been able to conclusively demonstrate this \citep{Finkbeiner2013}. Because
{\it Fermi} is an all sky survey, one might expect the significance of
this line to increase with time with more data if it is truly due to
a dark matter source. Instead, the significance has fluctuated
\citep{fermi130}. At the moment, there are no other gamma-ray telescopes
within this energy range that can test whether the line may be instrumental.
Given the intriguing nature of this line and the inability to rule out
systematic effects, {\it Fermi} has recently altered its survey strategy
to spend more time on the Galactic Center in an attempt to better understand
whether this line is related to dark matter. Intriguingly, $\sim$100 GeV
has long been favored as the WIMP mass, as a WIMP model with this mass
provides a natural fit to the relic dark matter density in the Universe
after freeze-out \citep{Gondolo1991, kolbturner}.
As this review was being written, another possible line associated with
decaying dark matter was observed, in the x-ray at 3.5 keV. This line
has been seen both in individual objects \citep[the Andromeda galaxy
and the Perseus galaxy cluster,][]{Boyarsky2014}, and the stacked spectrum of
73 clusters \citep{Bulbul2014}. Most of the data comes from the {\it XMM}
x-ray telescope, though \citet{Bulbul2014} also searched for it in
{\it Chandra} data. The same line was detected in the {\it Chandra} data
for the Perseus cluster, consistent with the {\it XMM} flux, but it was not
detected in {\it Chandra} data for the Virgo cluster. Unlike the 130 GeV
line, a line at 3.5 keV can be searched for with multiple current telescopes
(both {\it Chandra} and two separate detectors on {\it XMM}, but also
{\it Suzaku}), allowing to test if the detection is an instrumental line.
The stacked analysis in \citet{Bulbul2014} already argues against an
instrumental line, as the varying redshifts of the sources should wash out any
instrumental feature. Note that in this case, the mass of the dark matter
particle would be 7 keV, making it a WDM particle candidate. Sterile
neutrinos are a popular candidate for WDM. I will discuss in the next
section whether a 7 keV sterile neutrino is consistent with observed
galaxy properties.
\section{WDM}
\label{section3}
WDM is usually invoked to explain a lack of low mass halos that are a generic
prediction of CDM (see Fig.~\ref{fig1}). This problem extends beyond just the
missing satellites problem \citep{Moore1999, Klypin1999} and into the field
\citep{Klypin2014}. In CDM, the mass function of dark matter halos increases
toward smaller mass halos. Using a characteristic velocity at a given halo
mass, the velocity function, $n(V) \propto V^{\alpha}$ rises toward small halos
with $\alpha \sim-3$. The HI {\sc alfalfa} survey \citep{Giovanelli2005} has allowed
for a test of the CDM velocity function to lower masses than previous optical
surveys. The velocity function measured by the {\sc alfalfa} survey is much more
shallow ($\alpha \sim-0.8$) than CDM predicts \citep{Papastergis2011}. The
shallower slope is better described by the halo mass function predicted
in WDM models \citep{Schneider2013}. Likewise, a semi-analytic model of galaxy
formation in a WDM scenario is a better fit to observed central
and satellite luminosity functions \citep{Menci2012, Nierenberg2013}.
While CDM has power all the way down to very small scales \citep[e.g., Earth
mass halos,][]{Anderhalden2013}, the higher streaming velocities of WDM at high
redshift prevent it from initially collapsing into small halos with shallow
gravitational wells
\citep{Bode2001}. The halo mass at which WDM can begin to gravitationally
coalesce is set by the mass (and hence velocity) of the WDM particle. Thus,
quantifying the amount of small scale structure in the Universe can place
constraints on the mass of a WDM particle. Early studies to determine
whether dark matter was hot or cold showed that relativistic dark matter
(e.g., neutrinos) would erase structures up to tens of Mpc, yet we see
structure on smaller scales in the Universe \citep[and it is well described
by the CDM power spectrum,][]{White1983}. Given the success of CDM in
describing the observed power on large scales \citep{Hlozek2012} while
failing on small scales, WDM can be thought of as the Goldilocks solution.
A popular candidate for a WDM particle is the sterile neutrino, or a
right-handed neutrino. In the standard model, all fermions are expected
to come in both left and right-handed varieties. The left-handed neutrino
participates in weak interactions, while the right-handed neutrino does
not (hence, it is sterile), making it difficult to detect. While the canonical
CDM candidate, the WIMP, is expected to be a thermal relic of the early
Universe, it is very difficult to devise a scenario in which the sterile
neutrino is a thermal relic, and an alternative scenario must be invoked
\citep{Dodelson1994}. However, the transfer function (the modification to
the power spectrum) resulting from these alternative models has the same
shape as a thermal production mechanism. This allows the mass of the
sterile neutrino to be directly compared to the mass of a thermal relic
\citep{Colombi1996}. If the source of the 3.5 keV line mentioned above
was a thermal relic, it would have a mass in the range of 1.5 -- 3.0 keV
\citep{Abazajian2014}. In what follows, I will quote the equivalent thermal
relic WDM mass for comparison to CDM models.
There are multiple independent observations that can constrain the WDM
mass, e.g., phase-space constraints \citep{Boyarsky2009, Horiuchi2014},
gravitational lensing \citep{Miranda2007}, satellite abundance
\citep{Maccio2010, Polisensky2011, Anderhalden2013, Horiuchi2014}, the
amount of small scale structure in the Lyman-$\alpha$ forest
\citep{Viel2006, Seljak2006, Viel2008}, and the earliest epoch of star
formation \citep{Barkana2001, Mesinger2005, deSouza2013,
Pacucci2013}. Some of the tightest constraints on the WDM mass come
from observations of the Lyman-$\alpha$ forest at 2.5 $< z <$ 5.5
interpreted using hydrodynamical simulations. For many years these
results suggested that a WDM particle with mass $>$ 1 keV was allowed
\citep{Viel2006, Seljak2006, Viel2008}. However, a recent update set
new limits on the WDM mass to be $>$ 3.3 keV at the 2$\sigma$ level
\citep{Viel2013}. Hence, the limits from the Lyman-$\alpha$ forest may
be at odds with the tentative 3.5 keV x-ray line depending on the thermal
relic equivalent mass \citep{Abazajian2014}.
\subsection{WDM as a solution to CDM's small scale problems}
The erasure of substructure in WDM has the ability to solve at least one
of the major problems in CDM, the missing satellites problem \citep{Moore1999,
Klypin1999}. In recent years, a number of new satellites have been detected
at fainter luminosities \citep{Willman2005, SatData9, SatData11,
SatKinematics, SatData16, SatData18, SatData19}. All of these
``ultra-faint'' dwarfs have been detected in the {\it Sloan Digital Sky
Survey} (SDSS). Accounting for the footprint and magnitude limits of SDSS
suggests that there may be hundreds of faint galaxies orbiting the Milky
Way that remain undetected \citep{Willman2004, SatKinematics, Tollerud2008,
Walsh2009}.
The existence of hundreds of ultra-faint galaxies requires that the mass
of the WDM candidate would need to be greater than $\sim$2 keV
\citep{Maccio2010, Polisensky2011, Anderhalden2013, Horiuchi2014}.
\begin{figure*}
\center
\includegraphics[width=0.90\textwidth]{h2003_compressed.ps}
\caption{
A 10$^{10}$ M$_{\odot}$ halo run with SIDM (left), CDM (middle), and WDM (right).
These simulations are dark matter only. Each image is 50kpc across. Color
corresponds to density. The lowest densities (blue) are 100$\rho_{crit}$, and
the highest densities (white) are 10$^6$$\rho_{crit}$.
The SIDM model (left) has been run with $\sigma = 2$ cm$^2$/g (Fry et al., in
prep). Note the more spherical shape of the halo compared to the CDM run, as well
as the lower densities reached at the very center ($\sim10^5$$\rho_{crit}$).
The WDM model (right) has been run with the power spectrum corresponding to a
2 keV thermal relic mass.}
\label{fig1}
\end{figure*}
Because the phase--space density of dark matter should never be higher
than its initial density at decoupling, it was originally suggested that
WDM might naturally lead to the existence of matter cores in galaxies
\citep{tremainegunn, Dalcanton2001, Boyarsky2009}. However, WDM cannot explain the large
cores that we observe in galaxies (e.g., $\sim$1kpc core in a dwarf
galaxy with stellar mass of 10$^8$ M$_{\odot}$) without violating the
mass limits imposed by other observational constraints. To create a
1kpc dark matter core, the mass of the WDM particle would need to be
$\sim$0.1 keV \citep{Maccio2012}, a low mass which is already ruled out
by both the Lyman-$\alpha$ forest and the amount of substructure observed
around the Milky Way. At 2 keV, roughly the lower limit allowed by the
abundance of substructure, the core size drops below 10pc. Hence, if WDM is on the
order of $\sim$2 keV, then a separate mechanism for creating dark matter
cores in galaxies is still required. Energetic feedback from supernovae
provides a natural mechanism for dark matter core creation within the
allowed WDM mass range \citep{Pontzen2012,Governato2012}.
WDM has also been invoked to solve the Too Big to Fail problem found in
the Milky Way and M31 satellites. The central densities of halos should
be lower in WDM models because structure formation occurs later. In CDM,
small structures form first, but in WDM models this smallest structure is
wiped out, causing structure formation to be delayed compared to CDM
\citep{Lovell2012}. It has been established that the concentration
of a halo is related to formation time, with earlier forming halos being
more concentrated than later forming halos \citep{Wechsler2002, Zhao2003}.
Hence, WDM halos are less concentrated, and there is less mass enclosed
at a fixed radius \citep{Lovell2014}. In the case of the dwarf spheroidals,
the mass is measured most robustly at the half light radii, which are
typically $\lesssim$1kpc for the luminous dwarfs \citep{McConnachie2012}.
The mass enclosed at these small radii is sufficiently lowered to solve
the Too Big to Fail Problem \citep{Lovell2012}, which requires masses to
be lower by a factor of $\sim$2-4. However, to fully solve the problem
with no other contributing solution, the mass of the WDM particle cannot
be larger than $\sim$2 keV \citep{Schneider2013}. In other words, tension
exists between the allowed mass range for WDM from the Lyman-$\alpha$ forest
($>$ 3.3 keV) and the mass range required in order to solve the problems
of the satellites. If the dark matter mass is indeed above $\sim$3 keV,
then an additional process is still required to bring the masses of the
luminous satellites in line with observations. Fortunately, a baryonic
solution (enhanced tidal stripping in the presence of a disk) exists
that could solve this problem \citep{Penarrubia2010, Arraki2014, Brooks2014}.
\subsection{Future Prospects}
If further investigation of the 3.5 keV x-ray line proves that it is
difficult to explain as something other than dark matter, the mass of
this WDM particle needs to be reconciled with other observational
constraints. The thermal relic equivalent mass of an originating sterile
neutrino \citep[1.5-3.0 keV,][]{Abazajian2014} is already in tension with the
limits set by the Lyman-$\alpha$ forest, suggesting that we would need to
re-evaluate our interpretation of the hydrodynamic simulations used to
place the Lyman-$\alpha$ forest constraints. Further, a 1.5 keV WDM particle
is hard to reconcile with the number of ultra-faint halos already observed
around the Milky Way, though $\sim$3 keV is not. A 1.5 keV WDM particle
would suggest that we have been biased in finding faint dwarfs, so that our
extrapolations to the full number not yet detected are overly generous.
A more complete census of the number of ultra-faint galaxies and their
distribution on the sky is required before this can be reconciled.
Fortunately, there are a number of upcoming surveys ({\it Skymapper},
DES, and LSST) that should be able to inventory hundreds of faint satellites
if they exist.
A liberal reading of the observational constraints suggests that a minimum
mass of $\sim$2 keV is allowed, but a more conservative reading of the
Lyman-$\alpha$ forest limits suggests an even heavier particle. Even 2 keV
is broadly consistent with the number of satellites around the Milky Way, and
anything heavier is nearly indistinguishable from CDM in terms of the amount
of small-scale structure formed. In fact, assuming a WDM particle of 2 keV,
there is very little difference in the resulting structure of any individual
galaxy between CDM and WDM. The concentration -- mass
relation for WDM dark matter halos is essentially identical to CDM in
this mass range \citep{Schneider2012}.
When baryons are added, a slight contraction of the dark matter halo
is seen in the CDM case compared to the WDM case \citep{Herpich2014,
Lovell2014}. \citet{Herpich2014} simulated Milky Way-mass and
smaller galaxies in both CDM and WDM models that include baryons. They
attribute halo contraction in CDM to the existence of subhalos that
drive disk instabilities, causing gas to flow the center of galaxies
and leading to contraction. By the same argument, the WDM simulations
with baryons have less star formation at $z < 1$ due to a lack of
subhalo induced instabilities.
If the only change between the CDM+baryon and allowed WDM+baryon models
is in slightly less concentrated galaxies and slightly lower star formation
rates at low $z$, then it will be extremely difficult to disentangle
WDM from baryons. In fact, \citet{Herpich2014} and \cite{Governato2014}
demonstrated that the resulting change in the star formation history and
concentration of a galaxy is more sensitive to the details of star formation
than it is to the range of allowed WDM masses.
Given the lack of evidence for a WDM particle at low $z$, it appears
that the best route to constrain the WDM model further is to probe the
faintest structures at high $z$ to quantify the formation times of the
smallest halos. While a number of studies have already attempted to
use high $z$ star formation to constrain the WDM particle mass
\citep{Barkana2001, Mesinger2005, deSouza2013}, astronomers are now
pursuing a series of observations that will allow us to probe to fainter
structure than ever before. These observations use lensing clusters to
identify magnified galaxies at high $z$. The Cluster Lensing And
Supernovae survey with Hubble \citep[CLASH][]{Postman2012} has already
idenfied two candidate galaxies at $z \gtrsim 10$ \citep{Zheng2012,
Coe2013}.\footnote[5]{Note that a combinination of {\it HST} surveys
have also identified six candidate galaxies at $z \sim10$ without
lensing \citep{Bouwens2014}.} \citet{Pacucci2013} recently argued
that a very high number density of halos is required to be seen in
the small volumes that the lensing studies are probing. Two
candidate galaxies at $z \gtrsim 10$ restricts the WDM mass to be
heavier than 1 keV. However, the ongoing {\it Hubble Space Telescope}
Frontier Fields observations\footnote[6]{www.stsci.edu/hst/campaigns/frontier-fields}
will push $\sim$3 magnitudes deeper. These lensing observations should
put stronger bounds on the allowed WDM mass. If additional $z \sim10$
galaxy candidates are identified and confirmed, WDM is likely to be ruled
out.
\section{SIDM}
\label{section4}
SIDM is usually invoked to solve the cusp/core problem in CDM.
Motivation behind a model for SIDM can be found in examining the
standard model of particle physics. Given the number of particles
that exist, it seems natural to ask ourselves if dark matter may be
more complicated than we tend to assume. Is dark matter another particle with
small interactions with the standard model (like a sterile neutrino)? Or
might the ``dark sector'' contain a similarly complex model that contains
multiple particles? In a more complex model, there may be a mediator
particle that can be exchanged by the dark matter. In galaxies, this
particle exhange would occur most frequently where dark matter is most
closely packed together, i.e., in the center of galaxies. The exchange
redistributes the energy of the dark matter particles. Assuming an
elastic scattering of particles, the overall effect is to heat the
inner regions of galaxies so that particles move outward, transforming
a cuspy inner density profile into a cored profile \citep{Spergel2000}.
The redistribution of dark matter in SIDM models has several other
observational implications in addition to core creation. The scattering
tends to equalize the velocity of particles, leading to a constant velocity
dispersion profile within the scale radius of the galaxy (where
collisions are relatively frequent), rather than an increasing profile
as predicted by CDM \citep{Vogelsberger2012, Rocha2013}. The redistribution
also acts to transform a triaxial dark matter halo into a more spherical
distribution (see Fig.~\ref{fig1}).
The predicted circularization of the halo shapes led to a quick dismissal
of the SIDM model when it was first invoked in the early 2000's. Core
sizes observed in dwarf galaxies can be used to set constraints on the
cross section for interactions, $\sigma$ in cm$^2$/g, in SIDM. These
same limits, when applied to clusters, suggested very circularized halos
\citep{Yoshida2000}. Maps of clusters showed that they were much more
elliptical than predicted by SIDM \citep{Miralda2002}. Because of these
results, SIDM was generally neglected for the following decade. However,
the question of halo shapes has been revisited recently.
\citet{Peter2013} demonstrated that the $\sigma$ values required
to match dwarf galaxies do not lead to enough change in the halo
shapes of clusters to significantly distinguish them from CDM.
There are currently two models for SIDM being explored. The simplest
case posits that, no matter the relative velocities of the two dark
matter particles, there is a constant cross-section for interaction.
On the other hand, it is not unreasonable
to assume that it becomes easier for dark matter particles to scatter
as their relative velocities become smaller. Introducing a Yukawa
potential to the model \citep{Loeb2011} leads to a velocity-dependent
cross-section. This has the additional benefit of leading to core
formation in dwarf galaxies, while not altering the shapes of the
larger cluster halos where velocities are larger, and avoiding the
earlier problems posed by a constant-velocity cross-section.
\subsection{SIDM as a solution to CDM's small scale problems}
As discussed above, SIDM is invoked to solve the cusp/core problem
in CDM \citep{Spergel2000}. Assuming a constant-velocity cross-section,
a minimum $\sigma > 0.1$ cm$^2$/g is required in order to achieve the
large cores we see in dwarf galaxies \citep{Loeb2011}. This is bounded
by observations on the massive end, where the shapes of cluster halos
suggest $\sigma < 1$ cm$^2$/g \citep{Peter2013, Rocha2013,
Vogelsberger2012}.
SIDM does not do nearly as well as WDM at solving the problems with
the Milky Way's satellites, though. Around the time that early studies
of SIDM were pointing to overly spherical cluster shapes, other authors
noted that subhalos that traveled through the dense regions of their
parent halos should experience interactions that could lead to easier
disruption of the satellites. Some authors concluded that the large
number of observed subhalos in clusters was also a strike against SIDM
\citep{Gnedin2001}. Again, more recent work shows that the number of
disrupted halos is small enough that the discrepancy in subhalo numbers
between SIDM and CDM would be hard to detect \citep{Rocha2013}.
When comparing CDM and SIDM simulations that both neglect baryons,
the surviving subhalo mass function for elastic scattering models in
the range of allowed $\sigma$ values (0.1 to 1 cm$^2$/g) is identical
\citep{Zavala2013}. Thus, SIDM does not solve the missing satellites
problem.
While not significantly reducing the number of satellites, it has been
suggested that SIDM may help alleviate the Too Big to Fail problem.
\citet{Rocha2013} suggested a constant-velocity model with
$\sigma = 0.1$ would create large enough cores to lower the central
densities of satellites to bring them into line with observations.
However, this was based on an extrapolation of simulation
results below their resolution limits. \citet{Zavala2013}
instead showed that $\sigma = 0.1$ could not reduce the central
masses of the satellites enough to match observations. They
suggested a minimum $\sigma > 0.6$ cm$^2$/g is necessary to
alleviate the Too Big to Fail problem. However, even this value
would not fully explain Fornax. As one of the brightest satellites,
Fornax is expected to have been formed in one of the largest subhalos.
It's observed low velocity dispersion cannot be fully fit by the
0.6 cm$^2$/g model. This tension suggests that even with core
creation, some additional mechanism is still necessary to reduce the
densities of the most luminous satellites enough to match observations.
Again, a baryonic solution exists that could solve this problem
\citep{Penarrubia2010, Zolotov2012, Arraki2014, Brooks2014}.
\subsection{Future Prospects}
Initial analytic work suggests that the effect of baryons may
substantially alter the predictions from SIDM models that neglect baryons.
\citet{Kaplinghat2014} found that contraction of the baryons in a dark
matter halo will shrink the size of the dark matter core formed by
scatterings in SIDM. For the Milky Way, the core size when neglecting
baryons can be as large as the scale radius, $\sim$20kpc. Contraction
of baryons shinks the core to 0.5kpc \citep{Kaplinghat2014}. This is a
dramatic difference that needs to be confirmed by SIDM simulations
that include baryons. Because supernova feedback can also lead to
dark matter core creation \citep{Pontzen2012}, the effects of SIDM,
contraction, and supernovae will all need to be carefully understood
and disentangled.
To date, \citet{Vogelsberger2014} are the only group to have
published simulations of SIDM that include baryons. However, the
simulations do not include baryonic feedback that leads to dark matter
core creation, so the combined influence of core creation from both
supernovae and SIDM scatterings has not been examined. Despite this, their
model already demonstrates that the stellar component may be altered from
the CDM case, suggesting that the stellar distribution of galaxies may
allow us to probe the dark matter content. More simulations,
particularly with feedback that independently leads to cores, are
necessary to explore these trends further.
It is also critical to note that all of the current bounds on $\sigma$
have been derived by comparing dark matter-only SIDM simulations to
observations. If baryons lead to a dramatic change in the central
regions of galaxies compared to dark matter-only SIDM models, then all
of the current bounds will need to be re-examined. This is particularly
true in massive galaxies and clusters. Clusters of galaxies with masses
$> 10^{14}$ M$_{\odot}$ have scale radii $\sim$150 kpc. If core size is
comparable to the scale radius, core sizes this large are already ruled
out \citep{Rocha2013}. However, might baryons shrink the core size in
clusters to an allowed size? Recent measurements of brightest cluster
galaxies have found evidence for cores, but on the scales of a few kpc
to several tens of kpc \citep{Newman2013}. It is difficult for baryonic
physics to explain core sizes of tens of kpc. Might these core sizes
instead be indicative of SIDM with baryonic contraction? Better modeling
is required to answer this question.
Galaxies more massive than the Milky Way are dominated by baryons in their
central regions, making it difficult to put tight constraints on the dark
matter profile given uncertainties in removing the baryon contribution. This
makes low mass dwarf galaxies the more ideal place to test SIDM models, as
they are dark matter-dominated and the complications of baryons are minimized.
Dark matter-dominated dwarfs already outline a clear prediction to identify
SIDM from CDM: even if baryonic physics
can create dark matter cores in galaxies, it will do so in a distinctly
different mass regime from SIDM. In the allowed velocity-dependent models
of \citet{Vogelsberger2012}, or in the constant-velocity models with
$\sigma \sim1$ cm$^2$/g, even halos as small as Draco (with stellar
mass 3$\times$10$^5$ M$_{\odot}$) have large dark matter cores. This
is not true in CDM+baryon models. The creation of a core in baryonic
models is tied to the amount of energy that has been injected, i.e.,
to the amount of stars that have formed. Halos the size of Draco are
too small to have had enough star formation to create a large core
\citep{Garrison-kimmel2013}. The exact scaling of core size and mass
will depend on how well stellar/supernovae feedback couples to the ISM.
Assuming a coupling of 40\%, \citet{Penarrubia2012} showed that roughly
10$^7$ M$_{\odot}$ in stars is necessary to create kpc-sized cores.
Adopting this standard, it implies that kpc-sized cores cannot be
created in halos as faint as Draco. If such large cores were to be
identified in these faint halos, it would be strong evidence for SIDM.
Unfortunately, determining whether such faint halos have cores is
a daunting observational challenge. It has been claimed that Draco has
both a core \citep{Wolf2012} and a cusp \citep{Jardel2013}. The results are not
only disparate for Draco, but even for the more massive and luminous dwarf
Spheroidal satellites \citep{Strigari2010, Walker2011b, Hayashi2012, Jardel2012,
Breddels2013b}. The interpretation is most sensitive to assumptions about the
anisotropy of the stellar orbits, an unknown \citep{Evans2009, Battaglia2013,
Richardson2013, Richardson2013b}.
Regardless of the {\it slope} of the dark
matter density profile in these dwarfs, there is mounting observational evidence
that the {\it normalization} of the dark matter density is lower than predicted by
CDM, i.e., that dwarf galaxies have lower masses than predicted within a given
radius. For dwarf satellites, this may be caused by tidal stripping while
orbiting around the parent halo's disk \citep{Brooks2014}. However, even field
dwarf galaxies that should not have been influenced by tidal stripping seem to
have lower masses than predicted by CDM. Abundance matching of stellar masses
to halo masses suggests that galaxies in the stellar mass range below
10$^7$ M$_{\odot}$ have rotational velocities consistently lower than expected
in CDM \citep{Ferrero2012, Garrison-Kimmel2014, Papastergis2014}. If this trend cannot be be
explained by baryonic physics, then again SIDM would provide a natural explanation.
\newpage
\section{Conclusions}
Baryonic phyiscs has been shown to be able to solve all of the problems
of galaxy formation within CDM that are highlighted in this review:
(1) the cusp/core problem \citep{Navarro1996, Read2005, deSouza2011, Governato2010,
Pontzen2012, Teyssier2013, diCintio2014}, (2) the existence of
bulgeless disk galaxies \citep{Governato2010, Brook2011, Teyssier2013},
(3) the missing satellites problem \citep{Brooks2013}, and (4) the ``Too
Big to Fail'' problem \citep{Zolotov2012, Arraki2014, Brooks2014}.
Only baryons have the potential to solve all of these problems together.
Of the two alternative models to CDM that are discussed in this article,
neither can solve these problems simultaneously. WDM may alleviate the
problems in the satellites, but cannot create dark matter cores. SIDM can
create dark matter cores, but cannot alleviate the satellite problems.
Importantly, neither WDM nor SIDM can create bulgeless disk galaxies
without baryonic feedback.
Because baryons have been shown to so dramatically alter the evolution
of the dark matter structure in the center of galaxies and in satellites,
it is clear that dark matter-only simulations cannot be used to make accurate
predictions on small scales. Future preditions for galaxy formation
{\it in any model} must consider the role of baryons. This review
has highlighted the future prospects of constraining two popular
dark matter models as an alternative to CDM.
The current limits on the mass of a WDM particle are relatively heavy,
$\gtrsim$2 keV. At 2 keV, structure formation in WDM is nearly
indistinguishable from CDM. Theorists have already begun to include
baryons in predictions for WDM, but the main difference is that less
star formation occurs in WDM models \citep{Herpich2014, Governato2014}. Current simulations
are more sensitive to the star formation prescription than they are to
the mass of the WDM particle. Hence, identifying WDM from CDM based on
simulation predictions requires a better understanding of star formation
than we currently have. Rather, the best path forward for ruling out
or favoring WDM is through observations. Further observations of the
tentative 3.5 keV x-ray line, and the amount of star formation at
$z > 10$, are currently the optimal observations to pursue.
Simulations of galaxies formed with SIDM that include baryons are needed.
While analytic
predictions are beginning to appear \citep{Kaplinghat2014}, their
dramatic predictions need to be confirmed. If baryons are as important
as claimed in reducing SIDM core sizes, the bounds on $\sigma$ will need
to be re-evaluated. On the theoretical side, progress will be made
utilizing simulations over a range of galaxy masses. Presumably the
scaling relations of galaxies might show systematic differences between
SIDM and CDM, allowing the model to be constrained. On the observational
side, the existence of kpc-sized cores in galaxies with less than
10$^7$ M$_{\odot}$ in stellar mass would favor SIDM models. These
faint galaxies already show hints of being less massive than predicted
by CDM \citep{Ferrero2012, Papastergis2014}. Hence, understanding the mass distributions
in these faint field galaxies is the immediate best observational test
of SIDM.
|
\section{INTRODUCTION}
\label{Intr}
The magnetization dynamics in ferromagnetic materials with strong exchange interaction between atomic magnetic moments is well described by the phenomenological Landau-Lifshitz (LL) or Landau-Lifshitz-Gilbert (LLG) equation.\cite{LaLi1, Gilb} Although the damping term in the LLG equation is physically more relevant than the corresponding term in the LL equation,\cite{Kiku} these equations are mathematically equivalent and equally efficient. According to them, the local magnetization undergoes damped precessional motion about the local effective magnetic field, which usually includes the exchange and anisotropy fields, the externally applied magnetic field and the magnetostatic field. Except in the simplest cases (see, e.g., Refs.~[\onlinecite{ KIK, Lak, BBI}] and references therein), the LLG equation (because of the equivalence of the LL and LLG equations, we refer only to the LLG one) can not be solved exactly. Moreover, since the magnetostatic field depends on the magnetization distribution, the LLG equation must be solved together with the magnetostatic Maxwell equations, i.e., the LLG equation is in general not closed. In metallic materials, the effective field includes also the magnetic field of eddy currents that are induced by changing in time the external magnetic field and the magnetization direction. As a consequence, the LLG equation must be supplemented by the full system of Maxwell's equations. This coupled system of the LLG and Maxwell equations can be solved numerically using, e.g., advanced methods discussed in Refs.~[\onlinecite{SlBa, Cimr, LeTr}].
It is well known\cite{Brow1} that if the ferromagnetic particles are sufficiently small (with the particles size of the order of $10^{2}$ nanometers or less) then the uniform magnetization is energetically preferable. These nanoparticles exhibit unique properties and have many current and potential applications, e.g., in data storage,\cite{Ross, MTMA, Kiki} spintronics\cite{WABD, ZFS} and biomedicine.\cite{PCJD, Ferr, LaLe, LFPR} If, in addition, the nanoparticles are ellipsoidal, when the internal magnetostatic field is strictly uniform,\cite{Brow2} or if in non-ellipsoidal nanoparticles this field is assumed to be uniform and known, then the LLG equation becomes closed. This equation, which due to the uniform magnetization reduces to the first order vector differential equation, is a valuable tool for studying the magnetization dynamics in dielectric nanostructures. In particular, it was used to study the periodic and quasiperiodic regimes of the magnetization precession,\cite{BSM, BMS1, DLHT, BMS} chaotic magnetization dynamics,\cite{APC, VaPo, BPSV, LBGP} precessional magnetization switching,\cite{BWH, BFH, KaRu, SMB, SuWa} thermal\cite{Brow3, GaLa, CKW, DLH, DPL} and many other effects.
In conducting nanostructures the LLG equation becomes unclosed because of the presence of the magnetic field of eddy currents. Stimulated by potential applications, the temporal evolution of the (non-uniform) magnetization and eddy currents in some nanostructures has been studied by numerically solving the coupled system of the LLG and Maxwell equations. In particular, this approach was used to analyze the process of magnetization reversal in metallic nanocubes.\cite{ToLo, ToMa, HrKi} It was shown that, similarly to the case of domain walls moving in metallic ferromagnets, eddy currents act on the reversal process as an additional damping parameter in the LLG equation. Arising from Faraday's law of induction, this result seems to be quite general. However, even in the simplest case of a uniformly magnetized spherical nanoparticle, considered in Ref.~[\onlinecite{MaLo}], the problem of reducing the coupled system of the LLG and Maxwell equations to the closed LLG equation has not been solved completely. In particular, the authors determined the magnetic field of eddy currents only in the nanoparticle centre and used it to find the contribution of eddy currents to the damping parameter. But because the current-induced magnetic field inside the nanoparticle is strongly non-uniform, this contribution differs significantly from the exact one (see below).
In this paper, we consider a more general model in which the nanoparticle and its environment have different magnetic susceptibilities, provide an exact analytical solution of the quasi-static Maxwell's equations, and analyze in detail the role of eddy currents in the magnetization dynamics. The paper is organized as follows. In Sec.~\ref{Model}, we describe the model and introduce the coupled system of the LLG and Maxwell equations. By solving Maxwell's equations in the quasi-static approximation, we determine the induced electric field and the magnetic field of eddy currents in Secs.~\ref{Electr} and \ref{Magn}, respectively. The closed LLG equation, which accounts for the effects of eddy currents in the magnetization dynamics, is derived in Sec.~\ref{Eff}. In the same section, we demonstrate the importance of eddy currents for the correct description of precessional switching of magnetization in metallic nanoparticles. Finally, our findings are summarized in Sec.~\ref{Concl}.
\section{MODEL AND BASIC EQUATIONS}
\label{Model}
We consider a single-domain ferromagnetic particle of radius $a$ which is characterized by the electric conductivity $\sigma$ and magnetic susceptibility $\mu_{1}$. It is also assumed that the particle is electrically neutral and is embedded in a dielectric matrix, whose magnetic susceptibility equals $\mu_{2}$, and the origin of the Cartesian coordinate system $xyz$ is assumed to be located at the centre of the particle. If the exchange energy between neighboring spins significantly exceeds the magnetic energy $W$ of the particle, then the particle magnetization $\mathbf{M} = \mathbf{M}(t)$ changes with time in such a way that $|\mathbf{M}| = M = \mathrm{const}$. In this case, the dynamics of $\mathbf{M}$ can be described by the LLG equation\cite{Gilb}
\begin{equation}
\dot{\mathbf{M}} = -\gamma \mathbf{M} \times
\bm{\mathcal{H}}_{\mathrm{eff}}
+ \frac{\alpha}{M}\, \mathbf{M} \times
\dot{\mathbf{M}}.
\label{LLG}
\end{equation}
Here, $\gamma(>0)$ is the gyromagnetic ratio, $\alpha(>0)$ is the Gilbert damping parameter, the overdot denotes differentiation with respect to time, the cross stands for the cross (vector) product, and $\bm{\mathcal{H}}_{\mathrm{eff}}$ is the total effective magnetic field acting on the magnetization vector.
Due to the particle conductivity, it is convenient to represent the total effective field as a sum of two terms: $\bm{\mathcal{H}}_{\mathrm{eff}} = \mathbf{H}_{\mathrm{eff}} + \overline{\mathbf{H}}$, where $\mathbf{H}_{ \mathrm{ eff}} = -(1/V) \partial W/ \partial \mathbf{M}$ is the effective magnetic field at $\sigma=0$ and $\overline{\mathbf{H}} = (1/V) \int_{V} \mathbf{H} d\mathbf{r}$ is the averaged (over the particle volume $V=4\pi a^{3}/3$) magnetic field of eddy currents. In particular, if the particle is magnetically uniaxial then
\begin{equation}
\mathbf{H}_{\mathrm{eff}} = \frac{H_{a}}{M}(
\mathbf{M} \cdot \mathbf{e}_{a})\mathbf{e}_{a}
+ \mathbf{H}_{1}^{(0)}.
\label{H_eff}
\end{equation}
Here, $H_{a}$ is the anisotropy field, $\mathbf{e}_{a}$ is the unit vector along the anisotropy axis, the dot denotes the dot (scalar) product, and $\mathbf{H }_{1} ^{(0)}$ is the magnetic field inside the particle. By solving the magnetostatic equations for a given geometry, it can be easily shown\cite{BaTo} that
\begin{equation}
\mathbf{H}^{(0)}_{1} = \kappa\mathbf{H}_{0} -
\frac{4\pi \kappa}{3\mu_{2}}\mathbf{M}
\label{H^0_1}
\end{equation}
with
\begin{equation}
\kappa = \frac{3\mu_{2}}{\mu_{1} + 2\mu_{2}}.
\label{kappa}
\end{equation}
The first term on the right-hand side of Eq.~(\ref{H^0_1}) is the uniform magnetic field induced by the external magnetic field $\mathbf{H}_{0} = \mathbf{H}_{0}(t)$, and the second term represents the demagnetization field induced by the magnetization.
An important feature of Eq.~(\ref{LLG}) is that it is not closed. This is because the magnetic field $\mathbf{H} = \mathbf{H} (\mathbf{r},t)$ of eddy currents itself depends on the magnetization $\mathbf{M}$. Therefore, Eq.~(\ref{LLG}) in the case of conducting particles must be solved together with the Maxwell equations. In the quasi-static approximation, these equations (in CGS units) can be written as follows\cite{LaLi2}:
\begin{subequations}\label{EH}
\begin{eqnarray}
&\displaystyle \nabla \times \mathbf{E}_{l}
= -\frac{1}{c} \frac{\partial}{\partial t}
\mathbf{B}_{l}, \qquad
\nabla \cdot \mathbf{E}_{l} = 0,
\label{eqs_E_l}
\\[4pt]
&\displaystyle \nabla \times \mathbf{H}_{l}
= \frac{4\pi \sigma_{l}}{c} \mathbf{E}_{l},
\qquad \nabla \cdot \mathbf{H}_{l} = 0.
\label{eqs_H_l}
\end{eqnarray}
\end{subequations}
Here, $\mathbf{E}_{l} = \mathbf{E}_{l}(\mathbf{r},t)$ is the induced electric field, the indexes $l=1$ and $l=2$ denote the corresponding quantities inside ($r = |\mathbf{r}| <a$) and outside ($r>a$) the particle, respectively, $\sigma_{l} = \sigma \delta_{1l}$, $\delta_{1l}$ is the Kronecker delta, $c$ is the light speed in vacuum, $\nabla \times$ is the curl, $\nabla \cdot$ is the divergence, and $\mathbf{B}_{l} = \mathbf{B}_{l}(\mathbf{r},t)$ is the magnetic induction. In addition to the conditions under which Eqs.~(\ref{EH}) hold, we use also the condition (for details, see Sec.~\ref{Eff}) that permits us to use the magnetostatic approximation for finding $\mathbf{B}_{1}$ and $\mathbf{B}_{2}$. According to Ref.~\onlinecite{BaTo}, in this approximation $\mathbf{B}_{1} = \mu_{1} \mathbf{H}^{(0)}_{1} + 4\pi \mathbf{M}$ and $\mathbf{B}_{2} = \mu_{2} \mathbf{H}^{(0)}_{2}$, where
\begin{equation}
\mathbf{H}^{(0)}_{2} = \mathbf{H}_{0} -
\frac{\mathbf{m}^{(0)}}{r^{3}} +
\frac{3(\mathbf{m}^{(0)} \cdot
\mathbf{r}) \mathbf{r}}{r^{5}}
\label{H^0_2}
\end{equation}
represents the magnetic field outside the particle and
\begin{equation}
\mathbf{m}^{(0)} = \frac{a^{3} \kappa}{3\mu_{2}}
[(\mu_{1} - \mu_{2})\mathbf{H}_{0} + 4\pi \mathbf{M}]
\label{m^0}
\end{equation}
is the particle magnetic moment induced by the external magnetic field and magnetization. Note also that the tangential and normal components of the vectors $\mathbf{E}_{l}$ and $\mathbf{H}_{l}$ (denoted by the indices $\tau$ and $n$, respectively) must satisfy the following boundary conditions:
\begin{subequations}\label{bc_EH}
\begin{eqnarray}
&\mathbf{E}_{1\tau} = \mathbf{E}_{2\tau}, \quad
E_{2n} = 0,
\label{bc_E}
\\[4pt]
&\mathbf{H}_{1\tau} = \mathbf{H}_{2\tau}, \quad
\mu_{1}H_{1n} = \mu_{2}H_{2n}.
\label{bc_H}
\end{eqnarray}
\end{subequations}
\section{INDUCED ELECTRIC FIELD}
\label{Electr}
According to Eq.~(\ref{H^0_1}), the magnetic induction $\mathbf{B}_{1}$ in the magnetostatic approximation reads
\begin{equation}
\mathbf{B}_{1} = \kappa\Big(\mu_{1}\mathbf{H}_{0}
+ \frac{8\pi}{3} \mathbf{M} \Big).
\label{B1}
\end{equation}
Because $\mathbf{B}_{1}$ does not depend on $\mathbf{r}$, we seek the induced electric field $\mathbf{E}_{1}$ in the form $\mathbf{E}_{1} = \mathbf{a}(t) \times \mathbf{r}$. For this representation of $\mathbf{E}_{1}$, the second equation in (\ref{eqs_E_l}) holds identically and the first one yields
\begin{equation}
\mathbf{E}_{1} = -\frac{\kappa}{2c} \Big( \mu_{1}
\dot{\mathbf{H}}_{0} + \frac{8\pi}{3}
\dot{\mathbf{M}} \Big)\! \times \mathbf{r}.
\label{E1}
\end{equation}
It is this electric field which induces eddy currents of density $\mathbf{J} = \sigma \mathbf{E}_{1}$ inside the particle.
Similarly, from Eqs.~(\ref{H^0_2}) and (\ref{m^0}), for the magnetic induction outside the particle we obtain
\begin{eqnarray}
\mathbf{B}_{2} \!\!&=&\!\! \mu_{2}\mathbf{H}_{0}
- \frac{(\mu_{1} - \mu_{2})\kappa V}{4\pi r^{5}}
[r^{2}\mathbf{H}_{0} - 3(\mathbf{H}_{0}\cdot
\mathbf{r})\mathbf{r}]
\nonumber\\ [4pt]
&&\!\! - \frac{\kappa V}{r^{5}} [r^{2}\mathbf{M}
- 3(\mathbf{M}\cdot \mathbf{r})\mathbf{r}].
\label{B2}
\end{eqnarray}
This suggests to seek the induced electric field at $r>a$ in the form
\begin{equation}
\mathbf{E}_{2} = [u(r)\dot{\mathbf{H}}_{0}
+ v(r)\dot{\mathbf{M}}] \times \mathbf{r},
\label{E2a}
\end{equation}
where the functions $u(r)$ and $v(r)$ should be determined. As before, equation $\nabla \cdot \mathbf{E}_{2} = 0$ is satisfied identically and, in accordance with Eq.~(\ref{E2a}), the curl of $\mathbf{E}_{2}$ is given by
\begin{eqnarray}
\nabla \times \mathbf{E}_{2} \!\!&=&\!\! [ru'(r)
+ 2u(r)]\dot{\mathbf{H}}_{0} - \frac{u'(r)}{r}
(\dot{\mathbf{H}}_{0} \cdot \mathbf{r})\mathbf{r}
\nonumber\\ [4pt]
&&\!\! + [rv'(r) + 2v(r)]\dot{\mathbf{M}} -
\frac{v'(r)}{r}(\dot{\mathbf{M}} \cdot \mathbf{r})
\mathbf{r}. \quad
\label{curlE2}
\end{eqnarray}
Then, using Eq.~(\ref{B2}) and equating the right-hand sides of Eq.~(\ref{curlE2}) and equation $\nabla \times \mathbf{E}_{2} = - (1/c) \dot{ \mathbf{B}}_{2}$, we make sure that $u(r)$ must satisfy the equations
\begin{eqnarray}
&\displaystyle ru'(r) + 2u(r) = -\frac{\mu_{2}}{c}
+ \frac{(\mu_{1} - \mu_{2})\kappa V}{4\pi cr^{3}}, &
\nonumber\\ [4pt]
&\displaystyle u'(r) = \frac{3(\mu_{1} - \mu_{2})
\kappa V}{4\pi cr^{4}}&
\label{eq_u}
\end{eqnarray}
and $v(r)$ the equations
\begin{equation}
rv'(r) + 2v(r) = \frac{\kappa V}{cr^{3}},
\quad v'(r) = \frac{3\kappa V}{cr^{4}}
\label{eq_v}
\end{equation}
(the prime denotes differentiation with respect to $r$). These equations are easily solved,
\begin{equation}
u(r) = -\frac{\mu_{2}}{2c} -\frac{(\mu_{1}
- \mu_{2})\kappa V}{4\pi cr^{3}},
\quad v(r) = -\frac{\kappa V}{cr^{3}},
\label{u,v}
\end{equation}
and from Eq.~(\ref{E2a}) one finds the induced electric field outside the particle
\begin{equation}
\mathbf{E}_{2} = - \frac{\kappa}{2c}
\Big[\Big( \frac{\mu_{2}} {\kappa} +
\frac{(\mu_{1} - \mu_{2})V}{2\pi r^{3}}
\Big) \dot{\mathbf{H}}_{0} + \frac{2V}
{r^{3}} \dot{\mathbf{M}}\Big]\! \times
\mathbf{r}.
\label{E2}
\end{equation}
Using Eqs.~(\ref{E1}) and (\ref{E2}), it is not difficult to verify that the boundary conditions (\ref{bc_E}) are identically fulfilled. It should also be emphasized that since the quasi-static approximation is used, Eq.~(\ref{E2}) correctly describes the induced electric field at distances not too far from the particle surface. But, as follows from Eq.~(\ref{eqs_H_l}) (see also below), this fact does not affect the magnetic field of eddy currents both inside and outside the particle.
\section{MAGNETIC FIELD OF EDDY CURRENTS}
\label{Magn}
The induced electric field (\ref{E1}) shows that the magnetic field of eddy currents can be represented in the form
\begin{equation}
\mathbf{H}_{l} = f_{l}(r)\dot{\mathbf{H}}_{0} +
g_{l}(r)(\dot{\mathbf{H}}_{0}\cdot \mathbf{r})
\mathbf{r} + p_{l}(r)\dot{\mathbf{M}} + q_{l}(r)
(\dot{\mathbf{M}}\cdot \mathbf{r})\mathbf{r}
\label{Hl}
\end{equation}
with so far unknown functions $f_{l}(r)$, $g_{l}(r)$, $p_{l}(r)$, and $q_{l}(r)$. From this it follows straightforwardly that
\begin{eqnarray}
\nabla \times \mathbf{H}_{l} \!\!&=&\!\!
-[f'_{l}(r)/r - g_{l}(r)] \dot{\mathbf{H}}_{0}
\times \mathbf{r}
\nonumber\\ [4pt]
&&\!\! -[p'_{l}(r)/r - q_{l}(r)]\dot{\mathbf{M}}
\times \mathbf{r}
\label{curlHl}
\end{eqnarray}
and
\begin{eqnarray}
\nabla \cdot \mathbf{H}_{l} \!\!&=&\!\!
[f'_{l}(r)/r + rg'_{l}(r) + 4g_{l}(r)]
\dot{\mathbf{H}}_{0} \cdot \mathbf{r}
\nonumber\\ [4pt]
&&\!\! + [p'_{l}(r)/r + rq'_{l}(r) +
4q_{l}(r)]\dot{\mathbf{M}} \cdot
\mathbf{r}.
\label{divHl}
\end{eqnarray}
Using these formulas, the first equation in (\ref{eqs_H_l}) yields
\begin{eqnarray}
&\displaystyle f'_{l}(r) - rg_{l}(r) = \frac{2\pi
\sigma \kappa\mu_{1}}{c^{2}}r \delta_{1l},&
\nonumber\\ [4pt]
&\displaystyle p'_{l}(r) - rq_{l}(r) = \frac{16
\pi^{2} \sigma \kappa}{3c^{2}}r \delta_{1l}&
\label{eqs1}
\end{eqnarray}
and the second equation in (\ref{eqs_H_l}) reduces to
\begin{eqnarray}
&\displaystyle f'_{l}(r) + r^{2}g'_{l}(r) +
4rg_{l}(r) = 0,&
\nonumber\\ [6pt]
&\displaystyle p'_{l}(r) + r^{2}q'_{l}(r) +
4rq_{l}(r) = 0.&
\label{eqs2}
\end{eqnarray}
Considering the region inside the particle (when $l=1$) and assuming that the physically reasonable condition $|\mathbf{H}_{1}| < \infty$ holds, from Eqs.~(\ref{eqs1}) and (\ref{eqs2}) one obtains
\begin{eqnarray}
&\displaystyle f_{1}(r) = \phi - 2r^{2}g_{1}(r),
\quad g_{1}(r) = - \frac{3\tau_{\sigma}}{2\kappa a^{2}},&
\nonumber\\ [4pt]
&\displaystyle p_{1}(r) = \psi - 2r^{2}q_{1}(r),
\quad q_{1}(r) = - \frac{4\pi \tau_{\sigma}}{\kappa
\mu_{1} a^{2}},&\quad
\label{fgpq1}
\end{eqnarray}
where $\phi$ and $\psi$ are constants of integration and
\begin{equation}
\tau_{\sigma} = \frac{4\pi \sigma \kappa^{2}
a^{2}\mu_{1}}{15c^{2}}
\label{def_tau}
\end{equation}
is the characteristic time. It is not difficult to verify that outside the particle (when $l=2$) the solution of Eqs.~(\ref{eqs1}) and (\ref{eqs2}), which satisfies the natural condition $|\mathbf{H} _{2}| \to 0$ as $r \to \infty$, is given by
\begin{eqnarray}
&\displaystyle f_{2}(r) = -\frac{\nu}{3r^{3}},
\quad g_{2}(r) = \frac{\nu}{r^{5}},&
\nonumber\\ [4pt]
&\displaystyle p_{2}(r) = -\frac{\epsilon}{3r^{3}},
\quad q_{2}(r) = \frac{\epsilon}{r^{5}}.&
\label{fgpq2}
\end{eqnarray}
To determine the constants of integration $\phi$, $\psi$, $\nu$ and $\epsilon$, we use the boundary conditions (\ref{bc_H}). Taking into account that $\mathbf{H}_{1,2 \tau} = \mathbf{e}_{n} \times (\mathbf{H}_{1,2} \times \mathbf{e}_{n})|_{r=a}$ and $H_{1,2 n} = \mathbf{H}_{1,2} \cdot \mathbf{e}_{n} |_{r=a}$ ($\mathbf{e}_{n} = \mathbf{r}/r$), these boundary conditions together with Eqs.~(\ref{Hl}), (\ref{fgpq1}) and (\ref{fgpq2}) lead to the following system of algebraic equations:
\begin{eqnarray}
&\displaystyle \phi + \frac{\nu}{3a^{3}} = -
\frac{3\tau_{ \sigma}}{\kappa}, \quad \psi +
\frac{\epsilon}{3a^{3}} = - \frac{4\pi \tau_{
\sigma}} {\kappa \mu_{1}},&
\nonumber\\ [4pt]
&\displaystyle \phi - \frac{2\mu_{2}\nu}{3
\mu_{1}a^{3}} = - \frac{3\tau_{\sigma}}
{2\kappa}, \quad \psi - \frac{2\mu_{2}
\epsilon}{3\mu_{1}a^{3}} = - \frac{4\pi
\tau_{\sigma}}{\kappa \mu_{1}}.&\quad
\label{syst}
\end{eqnarray}
Solving it with respect to the mentioned constants of integration, one gets
\begin{eqnarray}
&\displaystyle \phi = -\Big(1 + \frac{3}
{2\kappa}\Big)\tau_{\sigma}, \quad
\nu = - \frac{3\mu_{1}}{2\mu_{2}} a^{3}
\tau_{\sigma},&
\nonumber\\ [4pt]
&\displaystyle \psi = - \frac{8\pi}{3\mu_{1}}
\Big(1 + \frac{3}{2\kappa}\Big)\tau_{\sigma},
\quad \epsilon = - \frac{4\pi}{\mu_{2}}
a^{3}\tau_{\sigma}.&\quad
\label{par}
\end{eqnarray}
Thus, collecting the above results, we find the magnetic field of eddy currents inside the particle
\begin{equation}
\mathbf{H}_{1} = \frac{\mu_{2}}{\kappa
\mu_{1}}\Big[\Big(3 + 2\kappa - 6\frac{r^{2}}
{a^{2}} \Big)\frac{\mathbf{m}}{a^{3}}
+ \frac{3}{a^{5}}(\mathbf{m}\cdot
\mathbf{r})\mathbf{r}\Big]
\label{H1}
\end{equation}
and outside the particle
\begin{equation}
\mathbf{H}_{2} = -\frac{\mathbf{m}}{r^{3}}
+ \frac{3}{r^{5}}(\mathbf{m}\cdot \mathbf{r})
\mathbf{r},
\label{H2}
\end{equation}
where
\begin{equation}
\mathbf{m} = -\frac{a^{3} \tau_{\sigma}}
{2\mu_{2}} \Big( \mu_{1}\dot{\mathbf{H}}_{0}
+ \frac{8\pi}{3} \dot{\mathbf{M}} \Big)
\label{m}
\end{equation}
is the magnetic moment of the particle induced by eddy currents. For illustration, in Fig.~\ref{fig1} we show the lines of the induced electric field inside the particle and the magnetic field of eddy currents.
\begin{figure}
\centering
\includegraphics[totalheight=6cm]{fig1.eps}
\caption{\label{fig1} (Color online) Plots of the electric
and magnetic field lines for $\mu_{1} = \mu_{2} = 1$. Since
according to Eqs.~(\ref{E1}) and (\ref{m}) $\mathbf{E}_{1}
= (\kappa \mu_{2}/ca^{3} \tau_{\sigma})\mathbf{m} \times
\mathbf{r}$, the lines of the induced electric field are
circular and lie in planes perpendicular to the vector
$\mathbf{m}$. The lines of the magnetic field $\mathbf{H}$
of eddy currents, determined by Eqs.~(\ref{H1}) and
(\ref{H2}), are shown in the plane of the figure.}
\end{figure}
\section{CLOSED LLG EQUATION FOR CONDUCTING NANOPARTICLES}
\label{Eff}
Now, we are ready to derive the closed LLG equation describing the magnetization dynamics in conducting nanoparticles. For this we need to calculate the average magnetic field $\overline{\mathbf{H}} = (1/V)\int_{V} \mathbf{H}_{1} d\mathbf{r}$ of eddy currents. Using Eq.~(\ref{H1}) and simple formulas
\begin{equation}
\frac{1}{V} \int_{V} r^{2} d\mathbf{r} =
\frac{3a^{2}}{5}, \quad
\frac{1}{V} \int_{V} (\mathbf{m}\cdot
\mathbf{r})\mathbf{r} d\mathbf{r} =
\frac{a^{2}}{5}\mathbf{m},
\nonumber
\label{int}
\end{equation}
we obtain $\overline{\mathbf{H}} = (2\mu_{2}/ \mu_{1}a^{3}) \mathbf{m}$, and so the total effective magnetic field acting on the magnetization becomes
\begin{equation}
\bm{\mathcal{H}}_{\mathrm{eff}} = \mathbf{H}_{
\mathrm{eff}} - \tau_{\sigma} \dot{\mathbf{H}}_{0} -
\frac{8\pi\tau_{\sigma}}{3\mu_{1}}\dot{\mathbf{M}}.
\label{mean_H}
\end{equation}
With this result, the unclosed LLG equation (\ref{LLG}) reduces to the closed one
\begin{equation}
\dot{\mathbf{M}} = -\gamma \mathbf{M} \times
\tilde{\mathbf{H}}_{\mathrm{eff}} + \frac{
\tilde{\alpha}}{M}\, \mathbf{M} \times
\dot{\mathbf{M}},
\label{eff_LLG}
\end{equation}
where $\tilde{\mathbf{H}}_{\mathrm{eff}} = \mathbf{H}_{\mathrm{eff}} - \tau_{\sigma} \dot{\mathbf{H}}_{0}$, $\tilde{\alpha} = \alpha + \alpha_{\sigma}$, and
\begin{equation}
\alpha_{\sigma} = \frac{8\pi \gamma M\tau_{
\sigma}} {3\mu_{1}}.
\label{relax}
\end{equation}
Thus, the magnetization dynamics in ferromagnetic metal particles can be described by the closed LLG equation (\ref{eff_LLG}). In this equation, which is the main result of this paper, the effects of particle conductivity are accounted by two terms. The first, $-\tau_{\sigma} \dot{\mathbf{H}}_{0}$, can be considered as an additional external magnetic field and the second, $\alpha_{\sigma}$, as an additional damping parameter. Both these terms arise from eddy currents inside the particle. However, while the first term results from eddy currents induced by changing the external magnetic field, the second term results from eddy currents induced by changing the magnetization direction. We note also that since $|\mathbf{H}_{1} |_{r=0}/ |\overline{\mathbf{H}}| = 2 + \mu_{1}/2\mu_{2}$, the exact result (\ref{relax}) is $5/2$ times less (at $\mu_{1} = \mu_{2} =1$) than that obtained in Ref.~[\onlinecite{MaLo}].
To clarify the importance of these terms in the magnetization dynamics, we first analyze the conditions under which Eq.~(\ref{eff_LLG}) is derived. In our model, we consider spherical ferromagnetic particles that are assumed to be single-domain. According to the Brown's fundamental theorem,\cite{Brow1} the single-domain state in these particles is energetically favorable if the particle radius $a$ is less than the critical radius $a_{\mathrm{cr}}$, which usually ranges from a few tens to a few hundreds of nanometers.
Next, the quasi-static approximation of Maxwell's equations implies\cite{LaLi2} that (a) the wavelength of the electromagnetic field is much larger than the particle size, (b) the displacement current $(1/4\pi) \partial \mathbf{E}_{1}/ \partial t$ is much smaller than the conduction current $\sigma \mathbf{E}_{1}$, and (c) the electric conductivity $\sigma$ and magnetic susceptibilities $\mu_{l}$ in the Maxwell equations (\ref{EH}) are the same as in the static case. Introducing the characteristic angular frequency $\omega$ of the electromagnetic field, the first two conditions can be written as $\omega \ll c/a$ and $\omega \ll \sigma$, respectively. The third condition for the conductivity is satisfied if the field period greatly exceeds the electron mean-free time $\tau_{0}$, i.e., if $\omega \tau_{0} \ll 1$. Since $\max{a} = a_{\mathrm{cr}}$, $\max{a_{\mathrm{ cr}}} \sim 10^{-5}\, \mathrm{cm}$, for good conducting metals $\sigma \sim 10^{17} - 10^{18}\, \mathrm{s}^{-1}$, and $\tau_{0}$ at room temperature is of the order of $10^{-13}\, \mathrm{s}$, one can make sure that these three conditions of quasi-stationarity are equivalent to the single condition $\omega \tau_{0} \ll 1$. It should be noted, however, that the magnetic susceptibilities $\mu_{l}$ tend to $1$ as $\omega$ increases and the difference between $\mu_{l}$ and $1$ can vanish at $\omega \tau_{0} \ll 1$.\cite{LaLi2} In this case, it is necessary to replace $\mu_{l}$ by $1$ in all formulas obtained within the quasi-static approximation.
Finally, let us discuss the conditions under which the magnetostatic approximation can be used to determine the magnetic induction inside and outside the particle. It is clear from the previous analysis that this approximation is valid if $\mu_{l} |\mathbf{H}_{l}| \ll |\mathbf{B}_{l}|$. Using formulas (\ref{B1}) and (\ref{B2}) for the magnetic induction in the magnetostatic approximation and formulas (\ref{H1}) and (\ref{H2}) for the magnetic field of eddy currents, it can be easily verified that this inequality holds if $\omega \tau_{\sigma} \ll 1$. Hence, collecting all the conditions, we may conclude that the LLG equation (\ref{eff_LLG}) is valid if $a< a_{\mathrm{cr}}$ and $\omega \max\{ \tau_{\sigma}, \tau_{0} \} \ll 1$.
Since, according to Eqs.~(\ref{def_tau}) and (\ref{relax}), $\alpha_{\sigma} \sim \tau_{\sigma}$ and $ \tau_{\sigma} \sim a^{2}$, the influence of eddy currents on the magnetization dynamics increases with increasing the particle size and reaches the maximum at $a \sim a_{\mathrm{cr}}$. To estimate the parameters $\tau_{\sigma}$ and $\alpha_{\sigma}$ in this case, we assume that $a = 10^{-5}\, \mathrm{cm}$, $\sigma = 10^{18}\, \mathrm{s}^{-1}$, $4\pi M = 2 \times 10^{4}\, \mathrm{G} $, $\gamma = 1.76 \times 10^{7}\, \mathrm{s}^{-1} \mathrm{G} ^{-1}$, and $\mu_{1} = \mu_{2} = 1$. Then, from Eqs.~(\ref{def_tau}) and (\ref{relax}), one gets $\tau_{\sigma} \approx 9.31 \times 10^{-14}\, \mathrm{s}$ and $\alpha_{\sigma} \approx 2.18 \times 10^{-2}$. Due to the smallness of the characteristic time $\tau_{\sigma}$ and limitation of $\omega$, the term $-\tau_{\sigma} \dot{\mathbf{H} }_{0}$ can usually be neglected compared to the external magnetic field $\mathbf{H}_{0}$ and, as a consequence, the effective magnetic field $\tilde{\mathbf{H}}_{\mathrm{eff}}$ in Eq.~(\ref{eff_LLG}) can be replaced by $\mathbf{H}_{\mathrm{eff}}$. As to the additional damping parameter $\alpha_{\sigma}$, its influence on the magnetization dynamics is, in general, not negligible and it is most pronounced when $\alpha \lesssim \alpha_{\sigma}$.
To illustrate the role of $\alpha_{\sigma}$, we consider the behavior of the magnetization in the time-dependent magnetic field
\begin{equation}
\mathbf{H}_{0}(t) = H_{0}\mathbf{e}_{x}
\left\{\!\! \begin{array}{cl}
t/\tau,
& t \leq \tau
\\ [4pt]
1,
& t > \tau.
\end{array}
\right.
\label{H_0}
\end{equation}
Assuming that $\mathbf{e}_{a} = \mathbf{e}_{z}$, from Eqs.~(\ref{H_eff}) and (\ref{H^0_1}) at $\mu_{1} = \mu_{2} = 1$ we obtain
\begin{equation}
\mathbf{H}_{\mathrm{eff}} = \frac{H_{a}}{M}(
\mathbf{M} \cdot \mathbf{e}_{z})\mathbf{e}_{z}
+ \mathbf{H}_{0}(t).
\label{H_eff2}
\end{equation}
[The demagnetization field $-(4\pi/3) \mathbf{M}$ in the right-hand side of Eq.~(\ref{H_eff2}) is omitted because it does not affect the magnetization dynamics.] Since the main features of the magnetization dynamics in the reference case has already been studied in the context of precessional switching of magnetization,\cite{BWH, BFH, KaRu, SMB, SuWa} here we only intend to show that the time behavior of $\mathbf{M}$ in dielectric and metallic nanoparticles can be qualitatively differen.
Replacing $\tilde{\mathbf{H}}_{ \mathrm{eff}}$ by $\mathbf{H}_{ \mathrm{eff}}$, taking the cross product of both sides of Eq.~(\ref{eff_LLG}) with $\mathbf{M}$ and using the condition $\dot{\mathbf{M}} \cdot \mathbf{M} = 0$, the LLG equation (\ref{eff_LLG}) can easily be reduced to the LL equation
\begin{equation}
\dot{\mathbf{M}} = -\frac{\gamma}{1 + \tilde{
\alpha}^{2}} \mathbf{M} \times \mathbf{H}
_{\mathrm{eff}} - \frac{\tilde{\alpha} \gamma}
{(1 + \tilde{\alpha}^{2})M}\, \mathbf{M} \times
(\mathbf{M} \times \mathbf{H}_{\mathrm{eff}}),
\label{eff_LL}
\end{equation}
which is more convenient for numerical solution. Using the initial condition $\mathbf{M}(0) = M\mathbf{e}_{z}$, we solved this equation at $H_{a} = 5 \times 10^{3}\, \mathrm{Oe}$, $\tau = 10^{-12}\, \mathrm{s}$, $\alpha = 2 \times 10^{-2}$, $h = H_{0}/ H_{a} = 0.52$ and other parameters given above. The trajectories of $\mathbf{M}$ in the plane $(\eta_{x}, \eta_{z})$, where $\eta_{x,z} = M_{x,z}/M$, are shown in Fig.~\ref{fig2}. The magnetization dynamics for $\sigma=0$ is represented by the trajectory (a) that begins at the point with coordinates $\eta_{x} = 0$ and $\eta_{z} = 1$ (at $t=0$) and ends at the point A with coordinates $\eta_{x} = h$ and $\eta_{z} = -\sqrt{1 - h^{2}}$ (at $t = \infty$). Since $\eta_{z}$ at $t=0$ and $t=\infty$ has different signs, the magnetization switching occurs (the time at which $\eta_{z} =0$ approximately equals $7.29 \times 10^{-11} \, \mathrm{s}$). In contrast, the magnetization dynamics for $\sigma \neq 0$ is represented by the trajectory (b) that ends at the point B with coordinates $\eta_{x} = h$ and $\eta_{z} = \sqrt{1 - h^{2}}$, i.e., there is no magnetization switching in this case. This explicitly shows that eddy currents in ferromagnetic metal nanoparticles can significantly affect the magnetization dynamics.
\begin{figure}
\centering
\includegraphics[totalheight=6cm]{fig2.eps}
\caption{\label{fig2} (Color online) Trajectories of
the reduced magnetization $\bm{\upeta} = \mathbf{M}/M$
in the $(\eta_{x}, \eta_{z})$ plane. The time dependence
of $\bm{\upeta}$ in dielectric and metallic nanoparticles
is represented by the trajectories (a) and (b),
respectively. The qualitative difference between these
trajectories results solely from the action of the
magnetic field of eddy currents.}
\end{figure}
\section{CONCLUSIONS}
\label{Concl}
We have developed an analytical model to describe the magnetization dynamics in ferromagnetic metal nanoparticles. It is based on the coupled system of the Landau-Lifshitz-Gilbert (LLG) equation for the magnetization, in which the effective magnetic field contains the magnetic field of eddy currents, and Maxwell's equations for the electromagnetic field induced by the external magnetic field and magnetization. We have analytically solved the Maxwell equations in the quasi-static approximation and have determined the magnetic field of eddy currents averaged over the particle volume. Using this result, we have derived the closed LLG equation describing the magnetization dynamics in metallic nanoparticles.
This equation contains two additional terms in comparison with the LLG equation that describes the magnetization dynamics in dielectric nanoparticles. The first term arises from eddy currents induced by changing the external magnetic field and is represented as an additional external magnetic field. In contrast, the second term results from eddy currents induced by changing the magnetization and is accounted for by an additional dumping parameter. We have shown that while the additional external magnetic field can be neglected in most cases, the additional dumped parameter may strongly influence the magnetization dynamics in relatively large metallic nanoparticles. This has been demonstrated by considering the precessional switching of magnetization in metallic and dielectric nanoparticles.
\section*{ACKNOWLEDGMENTS}
We are grateful to the Ministry of Education and Science of Ukraine for financial support under Grant No.~0112U001383.
|
\chapter{Conclusions}
\label{chapter:conclusion}
In this thesis we have studied the quasistatic processes of small driven systems described by stochastic Markovian dynamics.
Specifically, we have considered (1) Markov jump processes with continuous time, which in physical reality can describe ratchets, various semi-classical models of molecular motors, etc.,
(2) diffusion, both underdamped or overdamped, as describing e.g. colloid particle moving in a rotationally driven medium.
In both cases the time evolution of such systems can be described by Kolmogorov generators, which provide unified framework, see chapter \ref{chapter:models}.
After a brief recollection of well known facts for these systems in equilibrium, see chapter \ref{chapter:equilibrium_stat_phys},
we focused on the analysis of the heat and work for thermodynamic processes in the quasistatic limit, which is the main topic of this theses.
\section{Quasistatic limit for non-equilibrium systems}
\subsection{Non-potential force driven systems}
The first class of systems described mainly in chapter \ref{chapter:non-equilibrium_thermodynamics} are systems driven out of equilibrium by the action of non-potential force, or by the action of multiple thermal or particle baths.
For these systems we provide an analytical formalism for the quasistatic expansion of the mean values of work and heat.
We have shown that the mean heat and work in the quasistatic limit can be naturally decomposed into what we call the ``housekeeping'' and ``reversible'' components and we have obtained explicit formulas for their computation,
see equations \eqref{def:housekeeping_work}, \eqref{def:reversible_work}, \eqref{def:reversible_heat}.
The diverging ``housekeeping'' component of the heat and work is related to the steady dissipation of the system out of equilibrium and as such it is symmetric with respect to protocol reversal $\Theta \alpha$,
while the finite ``reversible'' component is antisymmetric with respect to protocol reversal and also independent of the protocol parametrization and thus geometric.
In particular, in equilibrium the ``reversible'' component coincides with the total heat or work.
We have also formulated the energy balance equation \eqref{equ:first_law_noneq} stating that the change of the internal energy represented as the mean value of energy of the system is given by the total ``reversible'' heat and work to the system,
i.e. the ``housekeeping'' components do not contribute to the energy balance.
In this sense the ``reversible'' work and heat are a natural extension of the equilibrium reversible work and heat.
Further, we have introduced a generalized heat capacity as the ``reversible'' heat produced along the quasistatic change of the temperature of the thermal bath \eqref{def:generalized_heat_capacity}.
As demonstrated on various models the generalized heat capacity has a particularly rich behaviour out of equilibrium.
The most surprising yet not fully understood is the possibility of the generalized heat capacity to be negative as seen in subsections \ref{ssec:two_level_nonp}, \ref{ssec:three_level_nonp}, \ref{ssec:diffusion_on_plane}.
Although the negative generalized heat capacity occurs in two- and three-level systems close to the population inversion, the actual effect appears to be much more subtle, cf. subsections \ref{ssec:two_level_nonp} and \ref{ssec:three_level_nonp}.
In case of the diffusion the negative heat capacity typically occurs when the increase of temperature pushes the system towards smaller dissipation, cf. \ref{ssec:diffusion_on_plane}.
\subsection{Periodically driven systems}
In chapter \ref{chapter:periodically_driven_systems} we have developed the extension of the formalism of quasistatic processes to periodically driven systems, which was not previously discussed in literature.
Our starting point was the application of the Floquet theory to remove the explicit time dependence from the forward Kolmogorov generator by using the Fourier picture.
That helped us to cast the problem into similar framework as in chapter \ref{chapter:non-equilibrium_thermodynamics} and to identify the generalized
``reversible'' and ``housekeeping'' components of the mean value of the total heat and work, see equations \eqref{equ:generalized_reversible_work}, \eqref{equ:generalized_housekeeping_work}, \eqref{equ:generalized_housekeeping_heat} and \eqref{equ:generalized_reversible_heat}.
These components have similar properties as those discussed in the previous subsection such as the ``reversible'' component being again independent of parametrization of the protocol $\alpha$ or the ``housekeeping'' component being generically divergent in the quasistatic limit.
However there also arise new fundamental interpretational problems as the generalized ``housekeeping'' component contains terms of the comparable magnitude as the ``reversible'' component, reflecting the initial and final state within the time period.
These terms can heavily oscillate in the quasistatic limit and hence they can introduce an additional uncertainty to the experimental accessibility of the ``reversible'' components.
We have also studied the generalized heat capacity on particular models as the means to analyse the behaviour of the ``reversible'' component of the heat.
In these models we have seen the generalized heat capacity may become negative in the intermediate regime where the relaxation time is comparable to the period of driving, while it is generally positive in either singular or slow driving limit.
\subsection{Generalized Clausius relation}
We have also discussed the possibility of generalization of the Clausius relation \eqref{equ:Clausius_relation} to non-equilibrium systems.
It has been argued that the Clausius relation or some straightforward generalization of that is not valid in general out of equilibrium.
Nevertheless if the steady state is characterized by the Boltzmann-like distribution in the McLennan form \eqref{equ:generalized_Boltzmann} then the generalized Clausius relation is verified; the system presented in subsection \ref{ssec:diffusion_quadratic_potential} is exactly such an example.
What physical conditions cause the steady state to be represented by such a distribution is still an open question as well as whether there exists any weaker condition.
The only case where the McLennan structure of the stationary distribution is generally verified is the close-to-equilibrium regime, see section \ref{sec:first_order_expansion}.
\section{Markovianness in time-scale separation}
The chapter \ref{chapter:slow-fast_coupling} was dedicated to provide a link between the concept of quasistatic changes of some external parameters and the time-scale separation,
as well as to show how one can obtain an effective Markovian dynamics for slow degrees of freedom on their specific time-scale by projecting out the fast degrees of freedom, in a way which is more precise that standard arguments in literature.
In the first section we provided a heuristic analysis when we can consider the system to be quasistatically driven.
After giving some heuristic arguments on the connection to previously discussed quasistatic processes,
we have studied in detail the Markovian structure of the effective dynamics for large time-scale separation.
In particular, the Markovianness has been proven to hold true up to the first order in the expansion around infinite separation limit.
\section{Open problems}
We finish by providing a number of open question that have emerged during the work and which have not yet been mentioned before.
\subsection{Generalization of Nernst theorem}
A natural question is whether the Nernst theorem representing the (equilibrium) Third law allows for an extension beyond equilibrium.
In our systems it might correspond to the conjecture that the heat capacity with the temperature of all thermal baths acting on the system going to zero also tends to zero.
We have seen on several examples that there are systems in which such a generalization of the Nernst theorem indeed holds true, cf. subsections \ref{ssec:two_level_nonp}, \ref{ssec:three_level_nonp}, as well as other examples where it does not, cf. subsections \ref{ssec:diffusion_on_plane}, \ref{ssec:merry-go-around}, \ref{ssec:diffusion_on_the_ring}.
While there is no surprise that the conjectured generalization of the Nernst theorem is not valid for diffusion, its invalidity even for system with finite number of state might sound surprising and it needs to be further understood, cf. subsection \ref{ssec:merry-go-around}.
For driven diffusion on the ring we have seen that the zero-temperature transition from the limit fixed point phase to the limit cycle phase is accompanied with the transition from the equilibrium-like to a new and highly nontrivial heat capacity asymptotics, cf. subsection \ref{ssec:diffusion_on_the_ring}.
Even a richer collection of different low-temperature patterns have been obtained for the driven 2D diffusion, cf. subsection \ref{ssec:diffusion_on_plane}.
It remains to be seen on a more general basis in what precise sense the emergence of diverging or negative low-temperature heat capacity patterns characterize the low-temperature steady states.
There are still no definite answers why in some finite systems the Nernst theorem holds true whether in others does not or what are the exact conditions for the zero temperature phase transition to occurs.
\subsection{Generalized quasistatic response functions}
The heat capacity has been introduced in order to quantify the reversible component of heat along the specific quasistatic process with temperature as the only time-dependent parameter.
Analogously, we can analyze other quasistatic processes in terms of generalized (quasistatic) response functions like the compressibility, thermal expansion coefficient etc.
Beyond obvious questions about their possible anomalous behavior in far-from-equilibrium regimes, still a more fundamental issue arises:
Are those mutually related in a way similar to the equilibrium Mayer of Maxwell relations?
Clearly, this question has much to do with generalizations of the Clausius relation and the existence of non-equilibrium entropy.
Within the heat-renormalization scheme adopted in this work, these question remain largely open and will require further research.
\chapter{Introduction}
The statistical thermodynamics of non-equilibrium processes, as originated already by the founders of modern thermodynamic theory including L. Boltzmann, J. C. Maxwell or J. W. Gibbs, has experienced a large revival in last decades.
This is mainly due to a great recent progress in micro- and nano-technologies which demands development of reliable theoretical methods to describe physical systems which are far from thermal equilibrium and in which fluctuations play a crucial role.
This has led to an application of stochastic techniques in a fresh and deeper way, with a better understanding of the role of time-reversal symmetry and its breaking as formulated in terms of the global and local detailed balance conditions, to the discovery of exact symmetry relations obeyed by the fluctuation even very far from equilibrium etc.
The thermodynamics of open and typically rather small systems exhibiting non-negligible fluctuations has been coined the name ``stochastic thermodynamics'' \cite{Seifert2008,Seifert2012,Sekimoto1997,Speck2004,Sagawa2011,Komatsu2008-2,Komatsu2009,Esposito2010}.
Beyond the recently extensively studied properties of fluctuations in such open systems, there is a more conservative line of thoughts which tries to link the stochastic thermodynamics back to the standard thermodynamic framework as represented on the macroscopic level by the famous laws of thermodynamics \cite{Komatsu2008,Komatsu2010}.
The main motivation is to formulate, if possible, general laws valid for large classes of thermodynamic process under inherently non-equilibrium conditions, similarly like this had been done with great success within the equilibrium framework.
In contrast to purely thermodynamic concepts, the stochastic thermodynamics provides a far more reliable, systematic and in a good sense ``microscopic'' approach that older theories coming usually under the general name ``non-equilibrium thermodynamics''.
Even restricting only to expected (or mean) values of relevant thermodynamic observables like work and heat, important and natural questions arise which have not yet been sufficiently answered: How to construct the fundamental elements of a standard thermodynamic theory, in particular the notion of quasistatic (= slow in a specific mathematical limit) processes above the inherently dissipative background (making the work or heat diverge in that limit)?
Apparently, some ``renormalization'' scheme is needed here which many not be (and it is not!) unique, in general.
Further, can the Clausius equality expressing the Second law for quasistatic processes be generalized to at least some large enough class of non-equilibrium systems?
Intimately related questions are whether the operationally defined Clausius entropy has any natural non-equilibrium counterpart, whether there exist general relations between non-equilibrium response quantities in a way analogous to the Maxwell relations, what is the low-temperature asymptotics of (slow) far-from-equilibrium processes etc.
It is exactly this sort of questions which are addressed in this thesis.
Although we are still very far from providing complete answers, it is hoped that the results obtained in this work might provide another small step towards building a general and practicable scheme to describe strongly non-equilibrium systems and to reliably predict their properties.
\section{Outline of the thesis}
Introduction and the necessary mathematical formalism used throughout the thesis is given in chapter \ref{chapter:models}.
Then in chapter \ref{chapter:equilibrium_stat_phys} we explain the central notion of quasistatic limit within the framework of equilibrium stochastic thermodynamics.
The main results of this work can be found in chapter \ref{chapter:non-equilibrium_thermodynamics} (quasistatic non-equilibrium process driven by non-potential forces),
chapter \ref{chapter:periodically_driven_systems} (a generalization to systems driven by time-periodic forces) and chapter \ref{chapter:slow-fast_coupling} (the application of our methods to the problem of time scale separation).
Finally, chapter \ref{chapter:conclusion} summarizes our results and shortly discusses open problems.
In the chapter \ref{chapter:models} we introduce the models widely used further within this thesis along with the necessary mathematical formalism.
We define the general Markovian time evolution for an open system which we later specify as continuous-time Markov jump processes and diffusion.
In order to describe the diffusions we review the Wiener process and consequently the It\^{o} and Stratonovich calculus and show some of its main results.
The chapter \ref{chapter:equilibrium_stat_phys} presents basic elements of (mostly equilibrium) stochastic thermodynamics.
We start with the introduction to Sekimoto stochastic energetics and its relation to the first law of thermodynamics.
We define the local heat production and the local power which will play a crucial role later in the construction of the so called quasi-potential.
Then we introduce the concepts of global and local detailed balance as the consequence of microscopic time reversibility
and review the Crooks relation and the Jarzynski equality.
We also show how the Second law inequality follows from the Crooks fluctuation relation.
The chapter \ref{chapter:non-equilibrium_thermodynamics} presents our first results.
We start with the discussion of the consequences of the global detailed balance condition breaking
and follow with the generalization of the stochastic energetics out of equilibrium.
We decompose the quasistatic mean value of work and heat to what we will call a ``reversible'' \eqref{def:reversible_work} and ``housekeeping'' \eqref{def:housekeeping_work} work and heat and discuss their properties.
As our first result we show how the energetics on the level of mean values is solely governed by finite, geometric ``reversible'' components of the heat and work \eqref{equ:first_law_noneq}.
We define the generalized heat capacity as the ``reversible'' heat associated with the quasistatic change of the temperature of the attached thermal bath \eqref{def:generalized_heat_capacity} and discuss its behaviour in various examples and show that it can be also negative, which is in contrast to its equilibrium counterpart.
Our next result states that as the consequence of McLennan theorem the generalized version of the Clausius relation is verified in the close to equilibrium regime.
We finish the chapter by brief illustration of low temperature behaviour of the heat capacity and conclude why no simple generalization of the Nernst theorem even for systems with finite number of non-degenerate states is to be expected.
In the chapter \ref{chapter:periodically_driven_systems} we study the behaviour of periodically driven systems in the quasistatic process, which was not previously studied in literature.
By using the Floquet theory we are able again to identify the ``reversible'' and ``housekeeping'' components of the mean values of the total work and heat, see e.g. \eqref{equ:generalized_reversible_work}.
We also define the generalized heat capacity and study its behaviour in various examples in order to better understand the behaviour of the ``reversible'' heat.
In these examples we find that the generalized heat capacity becomes negative as it approaches intermediate regime where the relaxation time is comparable to the period of driving.
We also discuss some fundamental interpretation problems as the ``housekeeping'' component contain a term which in general does not converge in the quasistatic limit.
In chapter \ref{chapter:slow-fast_coupling} we analyse the separation of time scales and the possibility to describe the evolution of such system by an effective Markovian dynamics.
We briefly discuss the time evolution on the short time scale which in a special case is equivalent to system undergoing the quasistatic process.
On the long time scale we provide more rigorous derivation of the Markovian dynamics and provide some estimates and bounds on precision of the Markovian approximation.
\chapter{Kolmogorov generators for Markov jump processes}
\label{ap:jump_processes_generators}
\index{jump processes}
\index{Kolmogorov generator!forward}
In this appendix we derive the forward and backward Kolmogorov generator from the conditional path measure for Markov jump processes.
Our starting point is the definition of the forward Kolmogorov generator \eqref{def:forward_generator}, first introduced in chapter \ref{chapter:models} section \ref{sec:Markov_processes},
\[
{\mathcal L}^*_t [\mu_t] (x) =
\lim_{\Delta t \rightarrow 0^+} \frac{1}{\Delta t} \left( \int {\mathrm d} y \; \mu_{t-\Delta t}(y) \int \cprob[{(t-\Delta t,t]}]{\omega}{X_{t-\Delta t}=y} \; \delta_{X_t}(x) - \mu_{t-\Delta t}(x) \right) ,
\]
in terms of the conditional probability measure ${\mathrm d} \mathbb P$, in this case, for Markov jump processes.
The conditional path probability measure for Markov jump processes \eqref{def:jump_process_path_probability},
was first introduced in chapter \ref{chapter:models} section \ref{sec:jump_processes}, and was defined as
\[
\cprob[{(0,T]}]{\omega}{X_0=x_0}
= \exp \left[ - \int\limits_0^T {\mathrm d} s \; \lambda_s(x_s) \right]
\prod_{i=1}^{k} \rate[t]{x_{t_i^-}}{x_{t_i}} \; {\mathrm d} t_i ,
\]
where $\rate[t]{x}{y}$ denotes transition rate from $x$ to $y$ at time $t$, $x_{s^-}$ the state right before the jump at time $s$,
$\lambda_t(x)$ is the escape rate from the state $x$ at time $t$ and $t_i$ denotes jump times.
We start by expanding the first term in the definition of the forward Kolmogorov generator \eqref{def:forward_generator} in terms of $\Delta t$
\begin{multline*}
\int \cprob[{(t-\Delta t,t]}]{\omega}{X_{t-\Delta t}=y} \; \delta_{X_t}(x) =
\delta_{xy} \exp\biggl[- \int\limits_{t-\Delta t}^{t} {\mathrm d} s \; \lambda_s(y) \biggr] + \\
+ (1 - \delta_{xy} ) \int\limits_{t-\Delta t}^t {\mathrm d} s \; \rate[s]{y}{x}
\exp\biggl[ - \int\limits_{t-\Delta t}^s {\mathrm d} u \; \lambda_u(y) - \int\limits_s^t {\mathrm d} u \; \lambda_u(x) \biggr] + \err[2]{\Delta t} ,
\end{multline*}
where we have used the fact that the only contributions up to the linear order in $\Delta t$ are from trajectories without any jump (the first term)
or with one jump (the second term).
Combining this result with the definition of the forward Kolmogorov generator above and taking the limit we obtain
\begin{align*}
{\mathcal L}^*_t [\mu_t] (x)
&= \sum\limits_{y \neq x} \mu_{t}(y) \rate[t]{y}{x} - \mu_t(x) \lambda_t(x) \\
&= \sum\limits_{y \neq x} \left[ \mu_{t}(y) \rate[t]{y}{x} - \mu_t(x) \rate[t]{x}{y} \right]
\end{align*}
where we have also used the definition of the escape rate \eqref{def:escape_rate}.
The backward Kolmogorov generator\index{Kolmogorov generator!backward} \eqref{def:backward_generator} is then defined by the forward Kolmogorov generator as
\[
\int {\mathrm d} x \; {\mathcal L}_t[A](x) \, \mu_t(x) = \int {\mathrm d} x \; A(x) \, {\mathcal L}^*_t [\mu_t] (x) .
\]
Inserting the forward Kolmogorov generator into this definition we obtain
\begin{align*}
\sum\limits_x {\mathcal L}_t[A](x) \, \mu_t(x)
&= \sum\limits_x A(x) \sum\limits_{y \neq x} \left[ \mu_t(y) \rate[t]{y}{x} - \mu_t(x) \rate[t]{x}{y} \right], \\
&= \sum\limits_x \mu_t(x) \sum\limits_{y \neq x} \left[ A(y) \rate[t]{x}{y} - A(x) \rate[t]{x}{y} \right] .
\end{align*}
Hence by comparison we obtain the backward Kolmogorov generator
\[
{\mathcal L}_t[A](x) = \sum\limits_{y \neq x} \rate[t]{x}{y} \left[ A(y) - A(x) \right] .
\]
\subsubsection{Poděkování}
Tato práce vznikala pod záštitou Fyzikálního ústavu Akademie věd ČR, v.v.i.,
které patří velký dík za příjemné a motivující pracovní prostředí.
V první řadě bych chtěl poděkovat svému vedoucímu Karlu Netočnému za velmi podnětné diskuze a za to,
že jsem právě díky němu měl možnost proniknout alespoň trochu do tajů statistické fyziky a termodynamiky.
Nesmím však opomenout jeho velkou trpělivost, obětavost a ochotu pomoci, za což mu také patří velký dík.
Chci také poděkovat kolegům z kanceláře Pavlovi Augustinskému a Vladovi Pokornému za podnětné diskuze a příjemně strávený čas v kanceláři.
Určitě velký dík patří skupině okolo profesora Maese v Belgické Lovani za jejich přijetí, mnoho podnětných dotazů a diskuzí a v neposlední řadě za plodnou spolupráci.
Jmenovitě pak především Christianu Maesovi, Eliranovi, Simi a Winnymu.
\end{otherlanguage}
\newpage
\cleardoublepage
\vglue 0pt plus 1fill
\noindent
I declare that I carried out this doctoral thesis independently, and only with the cited sources, literature and other professional sources.
\medskip\noindent
I understand that my work relates to the rights and obligations under the Act No.~121/2000 Coll., the Copyright Act, as amended, in particular the fact that the Charles University in Prague has the right to conclude a license agreement on the use of this work as a school work pursuant to Section~60 paragraph~1 of the Copyright Act.
\vspace{10mm}
\hbox{\hbox to 0.5\hsize{%
In Prague 5$^\text{th}$ September 2013
\hss}\hbox to 0.5\hsize{%
Jiří Pešek
\hss}}
\vspace{20mm}
\newpage
\cleardoublepage
\vbox to 0.5\vsize{
\setlength\parindent{0mm}
\setlength\parskip{5mm}
\begin{otherlanguage}{czech}
\noindent{\bf Název práce:}
Tepelné procesy v nerovnovážných stochastických systémech \\
{\bf Autor:}
Jiří Pešek \\
{\bf Katedra:}
Ústav teoretické fyziky MFF UK \\
{\bf Vedoucí disertační práce:}
Mgr. Karel Netočný, Ph.D., Fyzikální ústav AV ČR, v.v.i. \\
{\bf Abstrakt:} \\
Tato disertační práce je věnována teoretickému studiu pomalých ter\-mo\-dy\-na\-mi\-ckých procesů v nerovnovážných stochastických systémech.
Jejím hlavním výsledkem je fyzikálně a matematicky konzistentní konstrukce relevantních termodynamických veličin v kvazistatické limitě pro širokou třídu nerovnovážných modelů.
Jako aplikaci obecných metod zavádí přirozené nerovnovážně zobecnění tepelné kapacity a detailně analyzuje jeho vlastnosti, včetně anomálního chování daleko od rovnováhy.
Vyvinuté metody jsou dále použity na příbuzný problém oddělování časových škál, kde u\-mo\-žňu\-jí přesněji popisovat efektivní dynamiku pomalých i rychlých stupňů volnosti. \\ \\
{\bf Klíčová slova:}
stochastická termodynamika, kvazistatické procesy, stacionární stavy
\end{otherlanguage}
\vss}\nobreak\vbox to 0.49\vsize{
\setlength\parindent{0mm}
\setlength\parskip{5mm}
\noindent{\bf Title:}
Heat processes in non-equilibrium stochastic systems \\
{\bf Author:}
Jiří Pešek \\
{\bf Department:}
Institute of Theoretical Physics \\
{\bf Supervisor:}
Mgr. Karel Netočný, Ph.D., Institute of Physics ASCR, v.v.i. \\
{\bf Abstract:} \\
This thesis is devoted to the theoretical study of slow thermodynamic processes in non-equilibrium stochastic systems. Its main result is a physically and mathematically consistent construction of relevant thermodynamic quantities in the quasistatic limit for a large class of non-equilibrium models. As an application of general methods a natural non-equilibrium generalization of heat capacity is introduced and its properties are analyzed in detail, including an anomalous far-from-equilibrium behavior. The developed methods are further applied to the related problem of time-scale separation where they enable to describe the effective dynamics of both slow and fast degrees of freedom in a more precise way. \\ \\
{\bf Keywords:}
stochastic thermodynamics, quasistatic process, steady state
\vss}
\newpage
\cleardoublepage
\addtocontents{toc}{\protect\thispagestyle{empty}}
\addtocontents{lof}{\protect\thispagestyle{empty}}
\addtocontents{lot}{\protect\thispagestyle{empty}}
\tableofcontents
\cleardoublepage
\pagestyle{fancy}
\setcounter{page}{1}
\pagenumbering{arabic}
\fancyhead[RO,LE]{\thepage}
\fancyhead[RE]{\slshape\small\MakeUppercase{\chaptername~\thechapter}}
\fancyhead[LO]{\slshape\small\leftmark}
\fancyfoot[L,C,R]{}
\renewcommand{\chaptermark}[1]{\markboth{#1}{}}
\input{Introduction.tex}
\input{Stochastic_models_of_open_systems.tex}
\input{Stochastic_equilibrium_thermodynamics.tex}
\input{Stochastic_non-equilibrium_thermodynamics.tex}
\input{Quasistatic_transformations_of_periodically_driven_systems.tex}
\input{Slow-fast_coupling.tex}
\input{Conclusions.tex}
\section{Alternative definitions}
The definition of forward \eqref{def:forward_pseudoinverse} and backward \eqref{def:backward_pseudoinverse} pseudoinverse is not suitable for practical computations.
In this section we will provide other two equivalent more convenient expressions.
The first alternative definition of forward Kolmogorov generator is based on the spectral decomposition of the forward Kolmogorov generator (on the assumption it exists)
\[
{\mathcal L}^* [\mu] (x) = \sum_{i>0} \lambda_i \, {\mathcal P}^*_i [\mu](x) ,
\]
where ${\mathcal P}^*_i$ is the projection to the eigenstate with the eigenvalue of $\lambda_i$.
Because we also assume that the system starting from an arbitrary state $\mu(x)$ will reach the stationary state as $t \to \infty$, the zero eigenvalue has to be non-degenerate and all the nonzero eigenvalues has to have negative real part $\Re \lambda_i < 0$.
By default we denote by the index $0$ the projection to the zero eigenvalue, i.e. the projection to the stationary state $\rho(x)$, defined as
\begin{equation}
{\mathcal P}^*_0 [\mu](x) = \rho(x) \int {\mathrm d} \Gamma(y) \; \mu(y) .
\label{def:projector}
\end{equation}
From the definition it can be directly verified that $\mathcal P^*_0$ is indeed a projection to the stationary state
\begin{gather*}
{\mathcal P}^*_0 [ \rho ] (x) = \rho(x) , \\
\left( {\mathcal P}^*_0 \right)^2 = {\mathcal P}^*_0 .
\end{gather*}
Using the linearity of the forward Kolmogorov generator and the normalization condition \eqref{equ:normalization_condition} we also obtain
\[
{\mathcal L}^* {\mathcal P}^*_0 = {\mathcal P}^*_0 {\mathcal L}^* = 0 .
\]
Using the orthogonality of projections, ${\mathcal P}^*_i {\mathcal P}^*_j = 0 \quad \forall i \neq j$, we obtain the first alternative expression for the forward pseudoinverse
\begin{equation}
\frac{1}{{\mathcal L}^*}[\mu] (x)
= - \int\limits_0^\infty {\mathrm d} t \; \sum_{i>0} {\mathrm e}^{\lambda_i t} {\mathcal P}^*_i [\mu] (x)
= \sum_{i>0} \frac{1}{\lambda_i} {\mathcal P}^*_0 [\mu](x) .
\label{equ:forward_pseudoinverse_spectral}
\end{equation}
This expression has proved to be particularly convenient for the case of models with quadratic potential, where the eigenvectors for low powers in position and momentum can be explicitly found, e.g. see sections \ref{sec:pseudoinverse_underdamped} and \ref{sec:pseudoinverse_overdamped} .
The second alternative expression for the forward pseudoinverse is by including the projection $\mathcal P^*_0$ into the forward Kolmogorov generator making thus the total operator invertible and subtracting the projection afterwards
\[
\left[ {\mathcal L}^* + {\mathcal P}^*_0 \right]^{-1} - {\mathcal P}^*_0 = \left[ \frac{1}{1} {\mathcal P}^*_0 + \sum_{i>0} \frac{1}{\lambda_i} {\mathcal P}^*_i \right] - {\mathcal P}^*_0 = \frac{1}{{\mathcal L}^*} ,
\]
which on the other hand proved to be useful in numerical calculations in case of models describing Markov jump processes.
These expressions are also valid for backward pseudoinverse
\begin{align*}
\frac{1}{{\mathcal L}}[A](x) &= \sum_{i>0} \frac{1}{\lambda_i} {\mathcal P}_i [A] (x), \\
\frac{1}{{\mathcal L}}[A](x) &= \frac{1}{{\mathcal L} + {\mathcal P}_0} [A] (x) - {\mathcal P}_0 [A] (x) ,
\end{align*}
where projections $\mathcal P_i$ are related to the projections $\mathcal P^*_i$ by
\[
\int {\mathrm d} \Gamma(x) \; A(x) \, {\mathcal P}^*_i [ \mu ] (x) = \left\langle {\mathcal P}_i [A] \right\rangle_\mu ,
\]
from where we conclude that the projection to the zero eigenvalue $\mathcal P_0$ corresponds to taking the stationary mean value of the observable $A$
\[
{\mathcal P}_0 [A] (x) = \left\langle A \right\rangle_\rho .
\]
\section{Identities for the forward pseudoinverse}
In this section we will show some of the identities for the forward pseudoinverse and also show that the forward pseudoinverse is in fact the Drazin pseudoinverse of the forward Kolmogorov generator.
The first identity states that the forward pseudoinverse applied to the stationary state $\rho$ is zero, which can be obtained directly from the definition
\begin{equation}
\frac{1}{{\mathcal L}^*}[\rho](x) = \int\limits_0^\infty {\mathrm d} t \; \left[ \rho(x) - {\mathrm e}^{t {\mathcal L}^*}[\rho](x) \right] = 0 .
\label{equ:pseudoinverse_stationary_state}
\end{equation}
Because the forward Kolmogorov generator and the integral are linear operators, the forward pseudoinverse is also a linear operator
\begin{multline*}
\frac{1}{{\mathcal L}^*}[\mu+\nu](x)
= \int\limits_0^\infty {\mathrm d} t \; \left[ \rho(x) \int {\mathrm d} \Gamma(y) \; \left( \mu(y) + \nu(y) \right) + {\mathrm e}^{t {\mathcal L}^*} [ \mu + \nu ] (x) \right] = \\
= \int\limits_0^\infty {\mathrm d} t \; \left[ \rho(x) \int {\mathrm d} \Gamma(y) \; \mu(y) + {\mathrm e}^{t {\mathcal L}^*} [ \mu ] (x) \right]
+ \int\limits_0^\infty {\mathrm d} t \; \left[ \rho(x) \int {\mathrm d} \Gamma(y) \; \nu(y) + {\mathrm e}^{t {\mathcal L}^*} [ \nu ] (x) \right] = \\
= \frac{1}{{\mathcal L}^*}[\mu](x) + \frac{1}{{\mathcal L}^*}[\nu](x) .
\end{multline*}
If we apply the forward Kolmogorov generator to the forward pseudoinverse and use the linearity we obtain
\[
{\mathcal L}^* \frac{1}{{\mathcal L}^*} [\mu] (x) = - \int\limits_0^\infty {\mathrm d} t \; \partial_t {\mathrm e}^{t {\mathcal L}^*} [\mu] (x) = - \left[ \mu_t(x) \right]_{t=0}^\infty = \mu(x) - \rho(x).
\]
The same result can also be obtained by application of the forward pseudoinverse to the forward Kolmogorov generator,
however in this case it is the consequence of normalization of the probability distribution \eqref{equ:normalization_condition}
\[
\frac{1}{{\mathcal L}^*} {\mathcal L}^* [\mu] (x) = - \int\limits_0^\infty {\mathrm d} t \; {\mathrm e}^{t {\mathcal L}^*} {\mathcal L}^* [\mu] (x) = \mu(x) - \rho(x) .
\]
Putting all these facts together we can see that it is also valid
\begin{equation}
\begin{gathered}
\frac{1}{{\mathcal L}^*} {\mathcal L}^* [\mu] (x) = {\mathcal L}^* \frac{1}{{\mathcal L}^*} [\mu] (x), \\
\left({\mathcal L}^*\right)^n \frac{1}{{\mathcal L}^*} [\mu] (x) = \left( {\mathcal L}^* \right)^{n-1} [\mu] (x) , \qquad n > 2 \\
\left(\frac{1}{{\mathcal L}^*}\right)^n {\mathcal L}^* [\mu] (x) = \left( \frac{1}{{\mathcal L}^*} \right)^{n-1} [\mu] (x) . \qquad n > 2
\end{gathered}
\label{equ:forward_pseudoinverse_Drazin}
\end{equation}
These properties also uniquely characterizes the Drazin pseudoinverse\index{Drazin pseudoinverse}, for further details see \cite{Drazin1958}.
\section{Identities for backward pseudoinverse}
To show that the backward pseudoinverse is also a Drazin pseudoinverse, we need to show for that same set of identities \eqref{equ:forward_pseudoinverse_Drazin} is valid.
By putting the definition of the backward Kolmogorov generator \eqref{def:backward_generator} together with the normalization condition \eqref{equ:normalization_condition} we see that for every state $\mu$ and every constant $c$,
\[
\left\langle {\mathcal L} [c] \right\rangle_\mu = c \int {\mathrm d} \Gamma(x) \; {\mathcal L}^* [ \mu ] (x) = 0 ,
\]
which leads to the conclusion that the constant function $c(x)\equiv c$ is invariant under time evolution
\[
{\mathcal L}[c](x) = 0.
\]
Hence constant functions in case of the backward Kolmogorov plays the role of stationary state in case the forward Kolmogorov generator.
From there immediately follows that also the backward pseudoinverse applied to constant function $c$ is zero
\[
\frac{1}{{\mathcal L}}[c](x) = \int\limits_0^\infty {\mathrm d} t \; \left[ c - {\mathrm e}^{t {\mathcal L}}[c](x) \right] = 0 .
\]
Because the mean value does not depend on configuration $x$ it is also valid that
\[
{\mathcal L} \frac{1}{{\mathcal L}} [A] (x)
= - \int\limits_0^\infty {\mathrm d} t \; \partial_t {\mathrm e}^{t {\mathcal L}} [A] (x)
= - \left[ \left\langle {\mathrm e}^{t {\mathcal L}} [A] \right\rangle_{\delta_x} \right]_{t=0}^\infty
= A(x) - \left\langle A \right\rangle_\rho .
\]
The stationary mean value of the observable to which the backward Kolmogorov pseudoinverse is applied is also zero
\begin{equation}
\left\langle {\mathcal L}[A] \right\rangle_\rho = \int {\mathrm d} \Gamma(x) \; A(x) \, {\mathcal L}^*[\rho] (x) = 0
\label{equ:generator_stationary_mean_value}
\end{equation}
In a similar fashion from \eqref{equ:pseudoinverse_stationary_state} follows
\begin{equation}
\left\langle \frac{1}{{\mathcal L}} [A] \right\rangle_\rho = \int {\mathrm d} \Gamma(x) \; A(x) \, \frac{1}{{\mathcal L}^*}[\rho](x) = 0 .
\label{equ:backward_pseudoinverse_zero_mean}
\end{equation}
By using \eqref{equ:generator_stationary_mean_value} we can conclude
\begin{equation}
\frac{1}{{\mathcal L}} {\mathcal L} [A] (x)
= - \int\limits_0^\infty {\mathrm d} t \; {\mathrm e}^{t {\mathcal L}} {\mathcal L} [A] (x)
= A(x) - \left\langle A \right\rangle_\rho ,
\label{equ:backward_pseudoinverse_identity}
\end{equation}
hence the backward Kolmogorov generator and the backward pseudoinverse are again commutating.
By putting all these facts together we obtain the identities
\begin{gather*}
\frac{1}{{\mathcal L}} {\mathcal L} [\mu] (x) = {\mathcal L} \frac{1}{{\mathcal L}} [\mu] (x), \\
\left({\mathcal L}\right)^n \frac{1}{{\mathcal L}} [\mu] (x) = \left( {\mathcal L} \right)^{n-1} [\mu] (x), \qquad n > 2 \\
\left(\frac{1}{{\mathcal L}}\right)^n {\mathcal L} [\mu] (x) = \left( \frac{1}{{\mathcal L}} \right)^{n-1} [\mu] (x), \qquad n > 2
\end{gather*}
we can again see that the backward pseudoinverse is the Drazin pseudoinverse.
\chapter{Generator pseudoinverse}
\label{ap:pseudoinverse}
\input{pseudoinverse2.tex}
\input{spectral_properties_of_generators.tex}
\chapter{Quasistatic transformations of periodically driven systems}
\label{chapter:periodically_driven_systems}
In previous chapter we have described the behaviour of various physical systems with broken detailed balance undergoing quasistatic transformations of external parameters.
The global detailed balance in the previous chapter was broken by either the effect of non-potential forces or by attaching the system to multiple thermal or particle baths.
In this chapter we further extend the formalism of quasistatic processes to also describe systems driven by periodic forces.
We will study the quasistatic transformations connecting periodic steady states and corresponding to slow changes of arbitrary system parameters, except for the period $\tau$ of the driving which in the sequel is always assumed constant.
Typical physical example of such system is diffusion of ions in the presence of periodical electric field.
The first part of this chapter will be more mathematical as we will introduce the description of these systems,
first in terms of explicitly time dependent quantities,
then by using the results of Floquet theory in terms of Fourier components.
The second part of this chapter will then focus on how to generalize thermodynamics to these systems,
followed by an analysis of specific limiting cases and finished by numerical results for some particular examples.
\section{Description of periodically driven system}
The time evolution of a system driven by periodical external forces is again characterized by the forward Kolmogorov generator\index{Kolmogorov generator!forward} ${\mathcal L}^*_t$
\begin{equation}
\partial_t \mu_t = {\mathcal L}^*_t \left[ \mu_t \right],
\label{equ:time_evolution_2}
\end{equation} \index{time evolution!periodical driving}
where the generator does not depend on time only via external parameters $\alpha(t)$ but it also explicitly depends on time, ${\mathcal L}^*_t \equiv {\mathcal L}^*_{t;\alpha(t)}$.
We assume that for fixed external parameters $\alpha$ the generator is periodical with period $\tau$
\[
{\mathcal L}^*_{t+\tau;\alpha} = {\mathcal L}^*_{t;\alpha} ,
\]
where the period does not depend on $\alpha$, $\tau(\alpha) \equiv \tau$.
As we will see this assumption is essential, because it will allow us to reformulate the problem by using the Floquet theory
and thus effectively remove the explicit time dependence.
\subsection{The Floquet theory}
\label{ssec:Floquet_theory}
The main result of the Floquet theory is the \emph{Floquet theorem}\index{Floquet theorem} \cite{Ward:Floquet_theory},
which states that every solution of \eqref{equ:time_evolution_2} can be written in the form
\begin{equation}
\mu_t(x) = \sum_k {\mathrm e}^{-\lambda_k t} p^k_t(x),
\label{equ:Floquet_theorem}
\end{equation}
where $p^k_t(x)$ are real $\tau$-periodic functions, $p^k_{t+\tau}(x) = p^k_t(x)$, and $\lambda_k$ are non-negative numbers,
which can be interpreted as the reciprocal of relaxation times associated with each mode.
Furthermore the number of linearly independent periodic functions $p^k_t(x)$ is given by the dimension of the space, for proof see \cite{Ward:Floquet_theory}.
Although the Floquet theory gives us the explicit form of the solution, it gives us no simple algorithm how to find the $\tau$-periodic functions nor the relaxation times $1/\lambda_k$.
To provide such an algorithm we use the periodicity of the functions $p^k_t(x)$ and the forward Kolmogorov generator ${\mathcal L}^*_{t;\alpha}$ to expand them in the Fourier series
and we will construct the generalized Kolmogorov generator describing the time evolution of Fourier components.
We will further show that $p^k_t(x)$ correspond to eigenvectors of generalized forward Kolmogorov generator associated with real eigenvalues.
\subsubsection{Generalized forward Kolmogorov generator}
We start with the construction of the generalized forward Kolmogorov generator by expansion of the periodic functions $p^k_t(x)$ to Fourier series
\begin{equation}
p^k_t(x) = \sum\limits_{n \in {\mathbb Z}} {\mathrm e}^{\frac{2 \pi {\mathrm i} n t}{\tau}} \hat{p}^k_n(x),
\label{equ:Fourier_decomposition_p}
\end{equation}
where by $\hat{p}^k_n(x)$ we denote the Fourier components.
The fact that functions $p^k_t(x)$ are real implies a symmetry
\[
\hat{p}^k_{-n}(x) = \overline{\hat{p}^k_n(x)},
\]
where $\bar{f}$ denotes the complex conjugation of $f$.
Inserting the decomposition \eqref{equ:Fourier_decomposition_p} into \eqref{equ:Floquet_theorem} we obtain Fourier expansion of probability density
\begin{equation}
\mu_t(x) = \sum\limits_{n \in {\mathbb Z}} {\mathrm e}^{\frac{2 \pi {\mathrm i} n t}{\tau}} \sum_k {\mathrm e}^{- \lambda_k t} \hat{p}^k_n(x)
= \sum\limits_{n \in {\mathbb Z}} {\mathrm e}^{\frac{2 \pi {\mathrm i} t}{\tau}} \hat{\mu}_{n;t}(x) ,
\label{equ:mu_decomposition}
\end{equation}
where
\[
\hat{\mu}_{n;t}(x) = \sum_k {\mathrm e}^{- \lambda_k t} \hat{p}^k_n (x)
\]
denote time-dependent Fourier components of the probability density $\mu_t(x)$.
Negative-order Fourier components of probability density are also by definition connected to the positive ones by the symmetry
\begin{equation}
\hat{\mu}_{-n;t}(x) = \overline{\hat{\mu}_{n;t}(x)}.
\label{equ:complex_conjugation_of_density}
\end{equation}
Similarly we expand the forward Kolmogorov generator
\[
{\mathcal L}^*_{t;\alpha} \left[ \mu \right] (x) = \sum\limits_{n \in {\mathbb Z}} {\mathrm e}^{\frac{2 \pi {\mathrm i} n t}{\tau}} \hat{\mathcal L}^*_n \left[ \mu \right] (x),
\]
where $\hat{\mathcal L}^*_n$ denotes the Fourier component which also obeys
\begin{equation}
\overline{\hat{\mathcal L}^*_{-n}} = \hat{\mathcal L}^*_n .
\label{equ:complex_conjugation_of_generator}
\end{equation}
By putting all these decompositions together into the time evolution equation \eqref{equ:time_evolution_2} we obtain
\[
0 = \sum\limits_{n \in {\mathbb Z}} {\mathrm e}^{\frac{2 \pi {\mathrm i} n t}{\tau}}
\left[ \partial_t \hat\mu_{n;t} (x) - \sum\limits_{m \in {\mathbb Z}} \hat{\mathcal L}^*_{n-m} \left[ \hat\mu_{m;t} \right] (x) + \frac{2 \pi {\mathrm i} n t}{\tau} \hat\mu_{n;t} (x) \right] ,
\]
which has to be valid for each possible time $t$, hence each Fourier component can be considered as independent and henceforth the equation itself is equivalent to the set of equations
\begin{equation}
\partial_t \hat\mu_{n;t} (x) = \sum\limits_{m \in {\mathbb Z}} \hat{\mathcal L}^*_{n-m} \left[ \hat\mu_{m;t} \right] (x) - \frac{2 \pi {\mathrm i} n t}{\tau} \hat\mu_{n;t} (x) .
\label{equ:time_evolution_of_Fourier_components}
\end{equation}
Notice that equation for $(-n)$-th component and $n$-th component are connected by complex conjugation,
which preserve the symmetry \eqref{equ:complex_conjugation_of_density} during the time evolution.
By introduction of \emph{vectors of Fourier components},
\[
\hat{\vec\mu}_t (x) = \left( \dots , \hat\mu_{-1;t} (x), \hat\mu_{0;t} (x), \hat\mu_{1;t} (x), \dots \right),
\]
we can rewrite the set of equations \eqref{equ:time_evolution_of_Fourier_components} into the form of \emph{generalized time evolution}\index{time evolution!periodical driving}
\begin{equation}
\partial_t \hat{\vec\mu}_t (x) = {\hat{\mathscr L}}^*_\alpha \left[ \hat{\vec\mu}_t \right] (x) ,
\label{equ:generalized_time_evolution}
\end{equation}
where ${\hat{\mathscr L}}^*_\alpha$ denotes the \emph{generalized forward Kolmogorov generator}\index{Kolmogorov generator!forward!generalized}
\begin{equation}
\left[ {\hat{\mathscr L}}^*_\alpha \right]_{mn} = \hat{\mathcal L}^*_{m-n} - \frac{2 \pi {\mathrm i} m}{\tau} \delta_{mn} .
\label{equ:generalized_forward_generator}
\end{equation}
Notice that the generalized forward Kolmogorov generator is \emph{not explicitly time-dependent} and the only time-dependence left lies in $\alpha$.
Also notice that the generalized forward Kolmogorov generator is not self adjoint
\[
\overline{\left[ {\hat{\mathscr L}}^*_\alpha \right]}_{nm} = \left[ {\hat{\mathscr L}}^*_\alpha \right]_{mn} - \frac{4 \pi {\mathrm i} m}{\tau} \delta_{mn} ,
\]
neither is symmetric
\[
\left[ {\hat{\mathscr L}}^*_\alpha \right]_{nm} \neq \left[ {\hat{\mathscr L}}^*_\alpha \right]_{mn} .
\]
\subsubsection{Spectral properties of generalized Kolmogorov generator}
As was noted before we want to show that vectors $\hat{\vec{p}}^k(x)$ are eigenvectors with real eigenvalue of the generalized Kolmogorov generator.
We start by inserting the decomposition \eqref{equ:mu_decomposition} into the time evolution equation \eqref{equ:generalized_time_evolution}
\begin{equation}
0 = \sum\limits_k \left( {\hat{\mathscr L}}^*_\alpha \left[ \hat{\vec{p}}^k \right] (x) + \lambda_k \hat{\vec{p}}^k (x) \right) {\mathrm e}^{-\lambda_k t} ,
\label{equ:generalized_time_evolution_for_p}
\end{equation}
where by $\hat{\vec{p}}^k (x)$ we denote the vector of Fourier components.
One of the statements of the Floquet theorem is that the periodic functions are linearly independent,
which ensures that also Fourier components and so the vectors of Fourier components are linearly independent.
Also all the operators in the equation \eqref{equ:generalized_time_evolution_for_p} are linear, which yields to
\[
{\hat{\mathscr L}}^*_\alpha \left[ \hat{\vec{p}}^k \right] (x) = - \lambda_k \hat{\vec{p}}^k (x) ,
\]
where we recognize the eigenvector problem.
Thus every $p^k_t(x)$ gives us an eigenvector $\vec{p}^k(x)$ of the generalized forward Kolmogorov generator ${\hat{\mathscr L}}^*$ with the real eigenvalue $-\lambda_k$.
If we shift the eigenvector's $\hat{\vec{p}}^k$ components by $m$
\begin{equation}
\left[ \hat{\vec{q}}^{k;m} (x) \right]_n = \hat{p}^k_{m+n} (x),
\label{equ:shifted_state}
\end{equation}
we again obtain an eigenvector although with shifted eigenvalue $- \lambda_k + 2 \pi {\mathrm i} m/\tau$
\begin{multline*}
\left[ {\hat{\mathscr L}}^*_\alpha \left[ \hat{\vec{q}}^{k;m} \right] (x) \right]_l
= \sum_{j \in {\mathbb Z}} \hat{\mathcal L}^*_{l+m-j} [ \hat{p}^k_j ] (x) - \frac{2 \pi {\mathrm i} l}{\tau} \hat{p}^k_{l+m} (x) = \\
= - \lambda_k \hat{p}^k_{l+m}(x) + \frac{2 \pi {\mathrm i} m}{\tau} \hat{p}^k_{l+m} (x)
= \left( - \lambda_k + \frac{2 \pi {\mathrm i} m}{\tau} \right) \left[ \hat{\vec{q}}^{k;m} \right]_l .
\end{multline*}
Thus we can generate an infinite number of eigenvectors from a single one.
Hence it is reasonable to assume that the maximum number of eigenvectors with real eigenvalues corresponds dimension of the space $x$, i.e. it corresponds to the count of states at fixed time.
From where it follows that all linear independent solutions \eqref{equ:Floquet_theorem} can be found as the eigenvectors corresponding to real eigenvalues.
e.g. for two-level system we have two eigenvectors with real eigenvalues, one presumably being steady state thus corresponding to the zero eigenvalue.
We can also conclude that the steady state corresponds to the steady state of the generalized time evolution equation \eqref{equ:generalized_time_evolution},
\begin{equation}
0 = {\hat{\mathscr L}}^* \left[ \hat{\vec{\rho}} \right] (x) .
\label{equ:generalized_stationary_distribution}
\end{equation}
Let us also notice here that by shifting the steady state \eqref{equ:shifted_state} we obtain an infinite number of eigenvectors with zero real part.
\subsubsection{Time-dependence in external parameters}
Up to now we have considered the external parameters $\alpha$ to be fixed.
In systems driven by non-potential forces the change from fixed external parameters to external parameters explicitly dependent on the time
occurred only in the change
\begin{align*}
\alpha &\to \alpha(t) &
&\Longrightarrow &
\partial_t\mu_t (x) = {\mathcal L}^*_\alpha [ \mu_t ] (x) &
\to \partial_t \mu_t (x) = {\mathcal L}^*_{\alpha(t)} [ \mu_t ] (x) ,
\end{align*}
which proved to be crucial for further considerations in the quasistatic limit.
Although the same can be done on the level of the time dependent forward Kolmogorov generator ${\mathcal L}^*_{t;\alpha(t)}$ even in systems driven by periodic forces,
it is not clear whether we can assume the same to be true also for generalized forward Kolmogorov generator.
In order to ensure the same behaviour of the generalized forward Kolmogorov,
\begin{align*}
\alpha &\to \alpha(t) &
&\Longrightarrow &
\partial_t\hat{\vec{\mu}}_t (x) = {\hat{\mathscr L}}^*_\alpha [ \hat{\vec{\mu}}_t ] (x) &
\to \partial_t \hat{\vec{\mu}}_t (x) = {\hat{\mathscr L}}^*_{\alpha(t)} [ \hat{\vec{\mu}}_t ] (x),
\end{align*}
we need to assume that we are able to trace each eigenvector as we change the external parameters, i.e.
\begin{align*}
\lim_{t' \to t } \hat{\vec{p}}^k_{\alpha(t')}(x) &= \hat{\vec{p}}^k_{\alpha(t)}(x) &
&\Longleftrightarrow &
\lim_{t' \to t } p^k_{t',\alpha(t')}(x) &= p^k_{t,\alpha(t)}(x)
\end{align*}
and also
\[
\lim_{t' \to t} \lambda^k_{\alpha(t')} = \lambda^k_{\alpha(t)} .
\]
\subsubsection{Initial condition}
We start by the construction of the initial state in the generalized time evolution \eqref{equ:generalized_time_evolution}.
Our starting point is the assumption that the functions $p^k_t(x)$ are linearly independent in an arbitrary time $t$,
hence the initial state described by the probability distribution $\mu_{t_0}(x)$ can be decomposed into these functions
\[
\mu_{t_0}(x) = \sum_k \beta_k {\mathrm e}^{- t_0 \lambda_k} p^k_{t_0}(x),
\]
where $\beta_k$ represent real coefficients in the decomposition.
This decomposition enables us to rewrite the initial state in the form of the vector of Fourier components by
\[
\hat{\vec{\mu}}_{t_0}(x) = \sum_k \beta_k \hat{\vec{p}}^k (x) .
\]
We can see that the determination of the initial condition suitable for the generalized time evolution is at least as hard as to solve the original time evolution \eqref{equ:time_evolution_2} with the initial condition $\mu_{t_0}(x)$,
because in order to obtain the initial state we need to determine the basis of $\hat{\vec{p}}^k(x)$ as well as its eigenvalues $\lambda_k$
and hence at that point we are able to directly describe the solution of the time evolution at arbitrary time starting from the arbitrary initial condition.
Let us note here, that our aim is not to provide any simple method how to solve the time evolution with periodical driving,
but rather to reformulate it as an explicitly time-independent evolution at cost of enlarging the state space.
Moreover, we will see that the ``reversible'' component of the heat and work in the quasistatic limit will again be independent of the initial condition.
\subsubsection{The generalized backward Kolmogorov generator}
In the chapter \ref{chapter:non-equilibrium_thermodynamics} we have introduced the backward Kolmogorov generator as the adjoint generator to the forward Kolmogorov generator.
We would like to define the generalized backward Kolmogorov generator in the similar fashion.
We start by representing the mean value of some observable $A_t$ which can also periodically depend on time in terms of Fourier components
\[
\left\langle A_t \right\rangle_{\mu_t} = \int {\mathrm d} \Gamma(x) \; A_t(x) \, \mu_t(x)
= \sum_{n \in {\mathbb Z}} {\mathrm e}^{\frac{2 \pi {\mathrm i} n t}{\tau}} \int {\mathrm d} \Gamma(x) \; \left[ \hat{\vec{A}}(x) * \hat{\vec{\mu}}_t(x) \right]_n ,
\]
where $\hat{\vec{a}} * \hat{\vec{b}}$ is a shorthand for discrete convolution
\begin{equation}
[\hat{\vec{a}} * \hat{\vec{b}}]_n = \sum_{m \in {\mathbb Z}} \hat{a}_{n-m} \hat{b}_m .
\label{def:konvolution}
\end{equation}
The steady state is now time-periodic and so is the mean value of an arbitrary (possibly periodic as well) observable $A_t$.
Then the time derivative of the mean value of $A_t$ is
\begin{align*}
\partial_t \left\langle A_t \right\rangle_{\mu_t}
=& \sum_{n \in {\mathbb Z}} \frac{2 \pi {\mathrm i} n}{\tau} {\mathrm e}^{\frac{2 \pi {\mathrm i} n t}{\tau}} \int {\mathrm d} \Gamma(x) \; \left[ \hat{\vec{A}}(x) * \hat{\vec{\mu}}_t(x) \right]_n + \\
&+ \sum_{n \in {\mathbb Z}} {\mathrm e}^{\frac{2 \pi {\mathrm i} n t}{\tau}} \int {\mathrm d} \Gamma(x) \; \left[ \hat{\vec{A}}(x) * {\hat{\mathscr L}}^*[\hat{\vec{\mu}}_t](x) \right]_n , \\
=& \sum_{n \in {\mathbb Z}} {\mathrm e}^{\frac{2 \pi {\mathrm i} n t}{\tau}} \int {\mathrm d} \Gamma(x) \; \left[ \hat{\Omega} \cdot \left( \hat{\vec{A}}(x) * \hat{\vec{\mu}}_t(x) \right) + \hat{\vec{A}}(x) * {\hat{\mathscr L}}^* [ \hat{\vec{\mu}}_t ] (x) \right]_n ,
\end{align*}
where we have introduced \emph{the frequency matrix}\index{frequency matrix}
\begin{equation}
\left[ \hat{\Omega} \right]_{mn} = \frac{2 \pi {\mathrm i} n}{\tau} \delta_{mn} .
\label{def:frequency_matrix}
\end{equation}
We see that the time derivative of the mean value consists of two terms,
the first one reflecting the periodic nature of the stationary distribution (together with the possible periodicity of the observable itself),
and the second containing the relaxation of the probability distribution towards the steady state $\rho_t$.
Notice that in the steady state only the first term contributes
\[
\partial_t \left\langle A_t \right\rangle_{\rho_t}
= \sum_{n \in {\mathbb Z}} {\mathrm e}^{\frac{2 \pi {\mathrm i} n t}{\tau}} \int {\mathrm d} \Gamma(x) \; \left[ \hat{\Omega} \cdot \left( \hat{\vec{A}}(x) * \hat{\vec{\rho}}(x) \right) \right]_n .
\]
By using the definition of backward Kolmogorov generator \eqref{def:backward_generator} the time derivative can also be expressed as
\[
\partial_t \left\langle A_t \right\rangle_{\mu_t}
= \sum_{n \in {\mathbb Z}} {\mathrm e}^{\frac{2 \pi {\mathrm i} n t}{\tau}} \int {\mathrm d} \Gamma(x) \; \sum_{m \in {\mathbb Z}} \left[ \hat{\vec{\mu}}_t(x) \right]_{n-m}
\left\{ \sum_{k \in {\mathbb Z}} \hat{\mathcal L}_{m-k} [ \hat{A}_k ] (x) + \frac{2 \pi {\mathrm i} m}{\tau} \hat{A}_m(x) \right\} ,
\]
where we can identify \emph{the generalized backward Kolmogorov generator}\index{backward Kolmogorov generator!generalized}
\begin{equation}
\left[ {\hat{\mathscr L}} [ \hat{\vec{A}} ] (x) \right]_n = \sum_{m \in {\mathbb Z}} \hat{\mathcal L}_{n-m} [ \hat{A}_m ] (x) + \frac{2 \pi {\mathrm i} n}{\tau} \hat{A}_n(x) ,
\label{def:generalized_backward_generator}
\end{equation}
which yields to more compact expression
\[
\partial_t \left\langle A_t \right\rangle_{\mu_t}
= \sum_{n \in {\mathbb Z}} {\mathrm e}^{\frac{2 \pi {\mathrm i} n t}{\tau}} \int {\mathrm d} \Gamma(x) \; \left[ \hat{\vec{\mu}}_t (x) * {\hat{\mathscr L}}[ \hat{\vec{A}} ] (x) \right]_n .
\]
We can see that when the observable is periodic function of time, i.e. there is no explicit aperiodic time dependence for example in external parameters,
the time derivative of the mean value is solely given by the generalized backward Kolmogorov generator,
this corresponds in case when there is no periodical driving to the situation when the observable does not depend on time at all,
then there also the time derivative of the mean value is solely given by backward Kolmogorov generator.
Notice that from the linear independence of Fourier components we also obtained the relation between generalized forward and backward generators
\begin{equation}
\int {\mathrm d} \Gamma(x) \; \hat{\vec{\mu}}_t (x) * {\hat{\mathscr L}}[ \hat{\vec{A}} ] (x) =
\int {\mathrm d} \Gamma(x) \; \left\{ \hat{\Omega} \cdot \left( \hat{\vec{A}}(x) * \hat{\vec{\mu}}_t(x) \right) + \hat{\vec{A}}(x) * {\hat{\mathscr L}}^* [ \hat{\vec{\mu}}_t ] (x) \right\} .
\label{equ:generalized_generators_identity}
\end{equation}
\section{Quasistatic processes}
In the previous section we have developed a mathematical description of the time evolution, which reformulates the problem with explicit time dependence of the forward Kolmogorov generator in terms of the generalized forward Kolmogorov generator acting on vector of components in the Fourier basis.
Since that we have formally obtained the time evolution equation, which does not explicitly depends on time \eqref{equ:generalized_time_evolution},
hence we can follow the same pattern as in section \ref{sec:quasistatic_processes_noneq} in order to study the heat and the work in the quasistatic limit.
\subsection{Probability distribution}
We start with the Fourier components of the probability distribution on configurations,
where we assume that the system is at any time close to the stationary state,
\[
\hat{\vec{\mu}}_t (x) = \hat{\vec{\rho}}_{\alpha(\epsilon t)} (x) + \Delta \hat{\vec{\mu}}_{\epsilon t}(x) .
\]
By expanding the time evolution equation \eqref{equ:generalized_time_evolution} up to the first order in $\epsilon$ we immediately obtain the equation for the correction $\Delta \hat{\vec{\mu}}$
\[
\epsilon \dot\alpha(\epsilon t) \cdot \left. \nabla_\alpha \hat{\vec{\rho}}_\alpha (x) \right|_{\alpha=\alpha(\epsilon t)}
= {\hat{\mathscr L}}^*_{\alpha(\epsilon t)} [ \Delta \hat{\vec{\mu}}_{\epsilon t} ] (x) + \err[2]{\epsilon} ,
\]
which can be again formally solved
\begin{equation}
\Delta \hat{\vec{\mu}}_t(x) = \epsilon \dot\alpha(\epsilon t) \cdot \frac{1}{{\hat{\mathscr L}}^*_{\alpha(\epsilon t)}} \left[ \left. \nabla_\alpha \hat{\vec{\rho}}_\alpha \right|_{\alpha=\alpha(\epsilon t)} \right] (x)
+ \err[2]{\epsilon}
\label{equ:quasistatic_expansion_generalized_density}
\end{equation}
with \emph{the generalized forward pseudoinverse}\index{forward pseudoinverse!generalized}
\begin{equation}
\frac{1}{{\hat{\mathscr L}}^*_\alpha} [ \hat{\vec{\mu}} ] (x) = \lim_{k \to \infty} \int\limits_0^{k \tau} {\mathrm d} s \; \left\{ {\hat{\mathscr P}}^*_{0;\alpha} [ \hat{\vec{\mu}} ] (x) - {\mathrm e}^{s {\hat{\mathscr L}}^*_\alpha} [ \hat{\vec{\mu}} ] (x) \right\} ,
\label{def:generalized_forward_pseudoinverse}
\end{equation}
where by ${\hat{\mathscr P}}^*_{0;\alpha}$ we denote the projection to the steady state $\hat{\vec{\rho}}_\alpha$ and we integrate over multiples of periods, i.e. the $k$ is an integer $k \in \mathbb N$.
Notice that in the case the argument of pseudoinverse is physically acceptable solution, i.e. it is the decomposition to Fourier series of real function and hence does not contain shifted steady states,
the pseudoinverse is simply given by the integral from $0$ to $\infty$.
The integration over the multiple of periods is there to provide a correct definition for unphysical arguments, where we need to suppress indeterminate oscillating integrals
from the imaginary eigenvalues of shifted steady states.
\subsubsection{Stationary projection}
In the section \ref{sec:quasistatic_processes_noneq} the projector to the steady state was given simply by the stationary distribution multiplied by the normalization of distribution \eqref{def:projector},
in this case in general we expect the stationary projection to have a similar form of the product of the steady state and some normalization
\[
\left[ \hat{\hat{\mathscr P}}^*_{0;\alpha} [ \hat{\vec{\mu}} ] (x) \right]_n = \left[ \hat{\vec{\rho}}_\alpha (x) \right]_n \sum_{m \in \mathcal I} \int {\mathrm d} \Gamma(y) \; \hat\mu_m(y) ,
\]
where $\mathcal I$ is an index set, which yet need to be determined.
The projection has to obey
\begin{align*}
\hat{\hat{\mathscr P}}^*_{0;\alpha} [\hat{\vec{\rho}}_\alpha ] (x) &= \hat{\vec{\rho}}_\alpha (x) , && \\
\hat{\hat{\mathscr P}}^*_{0;\alpha} [\hat{\vec{\mu}}_\alpha ] (x) &= 0 , & \forall \hat{\vec{\mu}} (x) & \text{ linearly independent of } \hat{\vec{\rho}}(x) .
\end{align*}
For the first condition to be valid the index $0$ has to member of the index set $\mathcal I$,
because only the zeroth component of the probability distribution is normalized to unity, while the other components are normalized to zero.
For the second condition,
we recall that shifted eigenvector \eqref{equ:shifted_state} of the generalized forward Kolmogorov is again an eigenvector of the generalized forward Kolmogorov generator which is linearly independent of the original one.
If we now take the steady state $\hat{\vec{\rho}}_\alpha(x)$ as the basis for generating the shifted solution,
we can see that the component normalized to unity also shifts, and hence for the second condition on the projector to be valid, any other component than zero cannot contribute.
Hence the projector has to be defined as
\[
\left[ \hat{\hat{\mathscr P}}^*_{0;\alpha} [ \hat{\vec{\mu}} ] (x) \right]_n = \left[ \hat{\vec{\rho}}_\alpha (x) \right]_n \int {\mathrm d} \Gamma(y) \; \hat\mu_0(y) .
\]
\subsection{Heat and work}
Now we can proceed to the quasistatic expansion of path observables as heat and work.
Although we will proceed along the same lines as in chapter \ref{chapter:non-equilibrium_thermodynamics}, hence we would expect to obtain similar results,
we will see that the system driven by periodic action of external forces differs in several aspects from what we have introduced before.
\subsubsection{Local power and local heat production}
The first step is to represent the local power and the local heat production in terms of Fourier components.
In case there are no non-potential forces acting on the system the local power is given by the total time derivative of the energy $E_{t;\alpha}(x)$
\begin{equation}
w_{t;\alpha(\epsilon t)}(x) = \frac{{\mathrm d} E_{t;\alpha(\epsilon t)}(x)}{{\mathrm d} t}
= \epsilon \dot\alpha(\epsilon t) \cdot \left. \nabla_\alpha E_{t;\alpha}(x) \right|_{\alpha=\alpha(\epsilon t)} + \partial_t E_{t;\alpha(\epsilon t)}(x) ,
\label{def:local_power_t}
\end{equation}
where the first term correspond to the action of external forces, the second term corresponds to the action of driving forces.
Fourier components of the local power are then obtained by expanding the explicit time dependence in to the Fourier series
\[
\left[ \hat{\vec{w}}_{\alpha(\epsilon t)}(x) \right]_n
= \epsilon \dot\alpha(\epsilon t) \cdot \left. \nabla_\alpha \left[ \hat{\vec{E}}_\alpha(x) \right]_n \right|_{\alpha=\alpha(\epsilon t)} + \frac{2 \pi {\mathrm i} n}{\tau} \left[ \hat{\vec{E}}_{\alpha(\epsilon t)}(x) \right]_n ,
\]
which by using the definition of frequency matrix \eqref{def:frequency_matrix} can be written more compactly as
\begin{equation}
\hat{\vec{w}}_{\alpha(\epsilon t)}(x)
= \epsilon \dot\alpha(\epsilon t) \cdot \left. \nabla_\alpha \hat{\vec{E}}_\alpha(x) \right|_{\alpha=\alpha(\epsilon t)} + \hat{\Omega} \cdot \hat{\vec{E}}_{\alpha(\epsilon t)}(x) ,
\label{def:work_nonp_gen}
\end{equation}
where the first term represents again the work of external forces, or in terminology of the chapter \ref{chapter:non-equilibrium_thermodynamics} potential forces $\hat{\vec{w}}^\text{pot}_\alpha(x)$, and the second term corresponds to the work of driving forces $\hat{\vec{w}}^\text{drv}_\alpha(x)$, which cannot be described in terms of the derivative of the potential along an external parameters hence corresponding to the work of non-potential forces.
Similarly the local heat production given by
\begin{equation}
q_{t;\alpha(\epsilon t)}(x) = {\mathcal L}^*_{t;\alpha(\epsilon t)} \left[ E_{t;\alpha(\epsilon t)} \right] (x)
\label{def:local_heat_production_t}
\end{equation}
is in terms of Fourier components represented by
\[
\left[ \hat{\vec{q}}_{\alpha(\epsilon t)}(x) \right]_n
= \sum_{m \in {\mathbb Z}} \hat{{\mathcal L}}_{n-m;\alpha(\epsilon t)} \left[ \left[\hat{\vec{E}}_{\alpha(\epsilon t)}\right]_m \right] (x)
= \left[ {\hat{\mathscr L}}_{\alpha(\epsilon t)} \left[ \hat{\vec{E}}_{\alpha(\epsilon t)} \right] (x) \right]_n
- \frac{2 \pi {\mathrm i} n}{\tau} \left[ \hat{\vec{E}}_{\alpha(\epsilon t)} (x) \right]_n ,
\]
where we have used the definition of the generalized backward Kolmogorov generator \eqref{def:generalized_backward_generator},
or in more compact form as
\begin{equation}
\hat{\vec{q}}_{\alpha(\epsilon t)}(x)
= {\hat{\mathscr L}}_{\alpha(\epsilon t)} \left[ \hat{\vec{E}}_{\alpha(\epsilon t)} \right] (x) - \hat{\Omega} \cdot \hat{\vec{E}}_{\alpha(\epsilon t)} (x) .
\label{def:heat_gen}
\end{equation}
\subsubsection{Steady currents}
Having the local power and local heat production we can verify that there is a steady energy current through the system.
The mean steady local power is given by
\begin{align*}
\left\langle w_{t;\alpha} \right\rangle_{\rho_{t;\alpha}}
&= \sum_{n \in {\mathbb Z}} {\mathrm e}^{\frac{2 \pi {\mathrm i} n t}{\tau}} \int {\mathrm d} \Gamma(x) \; \left[ \hat{\vec{w}}_\alpha(x) * \hat{\vec{\rho}}_\alpha(x) \right]_n \\
&= \sum_{n \in {\mathbb Z}} {\mathrm e}^{\frac{2 \pi {\mathrm i} n t}{\tau}} \int {\mathrm d} \Gamma(x) \; \left[ \left( \hat{\Omega} \cdot \hat{\vec{E}}_\alpha (x) \right) * \hat{\vec{\rho}}_\alpha (x) \right]_n ,
\end{align*}
and similarly we obtain the mean steady local heat production as
\begin{multline*}
\left\langle q_{t;\alpha} \right\rangle_{\rho_{t;\alpha}}
= \sum_{n \in {\mathbb Z}} {\mathrm e}^{\frac{2 \pi {\mathrm i} n t}{\tau}} \int {\mathrm d} \Gamma(x) \; \left[ \hat{\vec{q}}_\alpha (x) * \hat{\vec{\rho}}_\alpha (x) \right]_n = \\
= \sum_{n \in {\mathbb Z}} {\mathrm e}^{\frac{2 \pi {\mathrm i} n t}{\tau}} \int {\mathrm d} \Gamma(x) \; \left[ \hat{\Omega} \cdot \left( \hat{\vec{E}}_\alpha (x) * \hat{\vec{\rho}}_\alpha (x) \right) - \left( \hat{\Omega} \cdot \hat{\vec{E}}_\alpha (x) \right) * \hat{\vec{\rho}}_\alpha (x) \right]_n = \\
= \sum_{n \in {\mathbb Z}} {\mathrm e}^{\frac{2 \pi {\mathrm i} n t}{\tau}} \int {\mathrm d} \Gamma(x) \; \left[ \hat{\vec{E}}_\alpha (x) * \left( \hat{\Omega} \cdot \hat{\vec{\rho}}_\alpha (x) \right) \right]_n ,
\end{multline*}
where we have used the identity \eqref{equ:generalized_generators_identity}, the fact the $\hat{\vec{\rho}}$ is steady state and the distributivity of the frequency matrix \eqref{def:frequency_matrix}
\begin{equation}
\hat{\Omega} \cdot \left( \hat{\vec{A}} * \hat{\vec{B}} \right)
= \left( \hat{\Omega} \cdot \hat{\vec{A}} \right) * \hat{\vec{B}}
+ \hat{\vec{A}} * \left( \hat{\Omega} \cdot \hat{\vec{B}} \right) .
\label{equ:spectral_matrix_property}
\end{equation}
We can see that together the mean local heat production and the mean local power sum up to the time derivative of the mean steady value of the energy, as one would have suspected,
\begin{multline*}
\partial_t \left\langle E_t \right\rangle_{\rho_{t;\alpha}}
= \sum_{n \in {\mathbb Z}} {\mathrm e}^{\frac{2 \pi {\mathrm i} n t}{\tau}} \int {\mathrm d} \Gamma(x) \; \left[ \hat{\Omega} \cdot \left( \hat{\vec{E}}_\alpha (x) * \hat{\vec{\rho}}_\alpha (x) \right) \right]_n = \\
= \sum_{n \in {\mathbb Z}} {\mathrm e}^{\frac{2 \pi {\mathrm i} n t}{\tau}} \int {\mathrm d} \Gamma(x) \; \left[ \left( \hat{\Omega} \cdot \hat{\vec{E}}_\alpha (x) \right) * \hat{\vec{\rho}}_\alpha (x)
+ \hat{\vec{E}}_\alpha (x) * \left( \hat{\Omega} \cdot \hat{\vec{\rho}}_\alpha (x) \right) \right]_n .
\end{multline*}
From there we can also see that the time-average change of the energy over the period $\tau$ is equal to zero, while the average local power and heat is in general not,
hence there is the steady energy current through the system even on the scales longer than the period.
This is due to the fact that external periodic force is responsible for the changes of the total energy, hence it is constantly powering the system,
while the same amount of energy need to be dissipated from the system to the thermal bath in order to maintain the steady state.
\subsubsection{Quasistatic expansion}
The first step in obtaining the quasistatic expansion of the mean work or heat is then rewriting it in terms of Fourier components,
where we will take the advantage of the characterization of the total work by the local power
\begin{multline*}
{\mathcal W}(\alpha(\epsilon t))
= \int\limits_0^{\frac{T}{\epsilon}} {\mathrm d} t \; \left\langle w_{t;\alpha(\epsilon t)}(x) \right\rangle_{\mu_t} = \\
= \int\limits_0^{\frac{T}{\epsilon}} {\mathrm d} t \; \int {\mathrm d} \Gamma(x) \; \sum_{n \in {\mathbb Z}} {\mathrm e}^{\frac{2 \pi {\mathrm i} n t}{\tau}} \left[ \hat{\vec{w}}_{\alpha(\epsilon t)} (x) \right]_n
\sum_{m \in {\mathbb Z}} {\mathrm e}^{\frac{2 \pi {\mathrm i} m t}{\tau}} \left[ \hat{\vec{\mu}}_{t;\alpha(\epsilon t)} (x) \right]_m = \\
= \sum_{n \in {\mathbb Z}} \int\limits_0^{\frac{T}{\epsilon}} {\mathrm d} t \; {\mathrm e}^{\frac{2 \pi {\mathrm i} n t}{\tau}} \int {\mathrm d} \Gamma(x) \;
\left[ \hat{\vec{w}}_{\alpha(\epsilon t)} (x) * \hat{\vec{\mu}}_{t;\alpha(\epsilon t)} (x) \right]_n .
\end{multline*}
Now we can insert the quasistatic expansion of the probability distribution \eqref{equ:quasistatic_expansion_generalized_density} and obtain
\begin{multline*}
{\mathcal W}(\alpha(\epsilon t))
= \sum_{n \in {\mathbb Z}} \int\limits_0^{\frac{T}{\epsilon}} {\mathrm d} t \; {\mathrm e}^{\frac{2 \pi {\mathrm i} n t}{\tau}} \int {\mathrm d} \Gamma(x) \;
\left[ \hat{\vec{w}}_{\alpha(\epsilon t)} (x) * \hat{\vec{\rho}}_{\alpha(\epsilon t)} (x) \right]_n + \\
+ \sum_{n \in {\mathbb Z}} \int\limits_0^{\frac{T}{\epsilon}} {\mathrm d} t \; \epsilon \, \dot\alpha(\epsilon t) \cdot {\mathrm e}^{\frac{2 \pi {\mathrm i} n t}{\tau}} \int {\mathrm d} \Gamma(x) \;
\left[ \hat{\vec{w}}_{\alpha(\epsilon t)} (x) * \frac{1}{{\hat{\mathscr L}}^*_{\alpha(\epsilon t)}} \left[ \left. \nabla_\alpha \hat{\vec{\rho}}_\alpha \right|_{\alpha=\alpha(\epsilon t)} \right] (x) \right]_n + \\
+ \err{\epsilon} ,
\end{multline*}
where the first term is the integrated mean local steady power which contains an extensive in time component of the work and hence in the leading order is proportional to $1/\epsilon$,
while the second term is finite, which can be better seen after the substitution $\epsilon t \to t$.
By inserting the definition of the work \eqref{def:work_nonp_gen} we can separate the terms according the power of $\epsilon$ and we obtain
\begin{multline*}
{\mathcal W}(\alpha(\epsilon t))
= \int\limits_0^{\frac{T}{\epsilon}} {\mathrm d} t \; \left\langle w^\text{drv}_{t;\alpha(\epsilon t)} \right\rangle_{\rho_{\alpha(\epsilon t)}}
+ \sum_{n \in {\mathbb Z}} \int\limits_0^T {\mathrm d} t \; \dot\alpha(t) \cdot {\mathrm e}^{\frac{2 \pi {\mathrm i} n t}{\epsilon \tau}} \int {\mathrm d} \Gamma(x) \times \\
\times \left[ \left. \nabla_\alpha \hat{\vec{E}}_\alpha (x) \right|_{\alpha=\alpha(\epsilon t)} * \hat{\vec{\rho}}_{\alpha(t)} (x) + \hat{\vec{w}}^\text{drv}_{\alpha(t)} (x) * \frac{1}{{\hat{\mathscr L}}^*_{\alpha(t)}} \left[ \left. \nabla_\alpha \hat{\vec{\rho}}_\alpha \right|_{\alpha=\alpha(t)} \right] (x) \right]_n + \\
+ \err{\epsilon} .
\end{multline*}
Moreover if we assume that the all quantities in the second term are smooth with respect to time,
we can apply the Riemann-Lebesgue lemma and see that only the zeroth Fourier component is preserved up to the linear order in $\epsilon$,
and while the zeroth component depends on time only via external parameters $\alpha$ it can be represented as the geometric integral
\begin{multline*}
{\mathcal W}(\alpha(\epsilon t))
= \int\limits_0^{\frac{T}{\epsilon}} {\mathrm d} t \; \left\langle w^\text{drv}_{t;\alpha(\epsilon t)} \right\rangle_{\rho_{\alpha(\epsilon t)}} + \\
+ \int {\mathrm d} \alpha \cdot \int {\mathrm d} \Gamma(x) \;
\left[ \nabla_\alpha \hat{\vec{E}}_\alpha (x) * \hat{\vec{\rho}}_\alpha (x) +\hat{\vec{w}}^\text{drv}_\alpha (x) * \frac{1}{{\hat{\mathscr L}}^*_\alpha} \left[ \nabla_\alpha \hat{\vec{\rho}}_\alpha \right] (x) \right]_0
+ \err{\epsilon} ,
\end{multline*}
where we will again identify the ``housekeeping'' component of the work\index{housekeeping work} as the first term and the ``reversible''\index{reversible work} component of work as the second geometric term.
Directly from the definition we can see that the ``housekeeping'' component is
\begin{enumerate}
\item extensive in time in the leading order,
\item present even if the external parameters $\alpha$ are kept constant,
\item invariant with respect to external protocol reversal $\alpha(t) \to \alpha(T-t)$ if $T/\epsilon$ is multiple of period~$\tau$,
\item equivalent to the ``housekeeping'' component \eqref{def:housekeeping_work} in case there are no periodical forces acting on the system,
\end{enumerate}
while the ``reversible'' component is
\begin{enumerate}
\item geometric in the sense that it is independent of parametrization of the trajectory of external parameters,
\item finite,
\item antisymmetric with respect to the external protocol reversal $\alpha(t) \to \alpha(T-t)$,
\item equivalent to the ``reversible'' component \eqref{def:reversible_work} in case there are no periodical forces acting on the system,
\end{enumerate}
which justifies our choices.
\subsubsection{The ``reversible'' work and the generalized backward pseudoinverse}
Up to now we didn't try to provide any physical description of the generalized forward pseudoinverse \eqref{def:generalized_forward_pseudoinverse} present in the ``reversible'' component of the total work
as we have done for other types of driving in the section \ref{sec:quasistatic_processes_noneq}.
This is due to the fact that in general we cannot associate the generalized backward pseudoinverse with the generalized forward pseudoinverse due to the extra term in the relation \eqref{equ:generalized_generators_identity} if compared with \eqref{def:backward_generator}.
However in the ``reversible'' component of the total work there is only the zeroth Fourier component present, hence the relation \eqref{equ:generalized_generators_identity} simplifies to
\[
\int {\mathrm d} \Gamma(x) \; \left[ \hat{\vec{\mu}}_t (x) * {\hat{\mathscr L}}[ \hat{\vec{A}} ] (x) \right]_0 =
\int {\mathrm d} \Gamma(x) \; \left[ \hat{\vec{A}}(x) * {\hat{\mathscr L}}^* [ \hat{\vec{\mu}}_t ] (x) \right]_0 .
\]
In this particular case we can effectively define \emph{the generalized backward pseudoinverse}\index{backward pseudoinverse!generalized} as
\begin{equation}
\frac{1}{{\hat{\mathscr L}}} [ \hat{\vec{w}} ] (x) = \int\limits_0^\infty {\mathrm d} t \; \left\{ \hat{\vec{1}}_0 \int {\mathrm d} \Gamma(y) \; \left[ \hat{\vec{w}} (y) * \hat{\vec{\rho}} (y) \right]_0 - {\mathrm e}^{t {\hat{\mathscr L}}} [ \hat{\vec{w}} ] (x) \right\} ,
\label{def:generalized_backward_pseudoinverse}
\end{equation}
where the vector $[ \hat{\vec{1}}_m ]_n = \delta_{mn}$ select the $m$-th Fourier component.
Because the zeroth component is obtained by time-averaging over the period, the generalized backward pseudoinverse can be equivalently defined as
\[
\frac{1}{{\hat{\mathscr L}}} [ \hat{\vec{w}} ] (x) = \lim_{k \to \infty} \int\limits_0^{k \tau} {\mathrm d} t \; \left\{ \hat{\vec{1}}_0 \left\langle w_t \right\rangle_{\rho_t} - {\mathrm e}^{t {\hat{\mathscr L}}} [ \hat{\vec{w}} ] (x) \right\} ,
\]
which corresponds to the additional work done on the system to the steady work production along the relaxation process starting from the configuration $x$,
when taking the limit in such a way, that we suppress the oscillations.
If we also introduce the notation for time-averaged mean values with respect to the steady state
\[
\frac{1}{\tau} \int\limits_0^\tau {\mathrm d} t \; \left\langle A_t \right\rangle_{\rho_t}
= \int {\mathrm d} \Gamma(x) \; \left[ \hat{\vec{A}}(x) * \hat{\vec{\rho}}(x) \right]_0
\equiv \left\llangle \hat{\vec{A}} \right\rrangle_{\hat{\vec{\rho}}} ,
\]
the ``reversible'' work can be written more compactly as
\begin{equation}
{\mathcal W}^\text{rev}(\alpha) = \int {\mathrm d}\alpha \cdot \left\llangle \nabla_\alpha \left( \hat{\vec{E}}_\alpha - \frac{1}{{\hat{\mathscr L}}_\alpha} [ \hat{\vec{w}}^\text{drv}_\alpha ] \right) \right\rrangle_{\hat{\vec{\rho}}_\alpha} .
\label{equ:generalized_reversible_work}
\end{equation}
\subsubsection{The properties of the ``housekeeping'' work}
We have seen that the ``reversible'' component of work is well defined in the limit for all rescaled smooth time evolutions.
As we will see this is not always true for the ``housekeeping'' component of the work.
At first we will rewrite the ``housekeeping'' work in terms of Fourier components with rescaled time scale
\begin{equation}
{\mathcal W}^\text{kh}(\alpha(t)) = \frac{1}{\epsilon} \int\limits_0^T {\mathrm d} t \; \sum_{n \in {\mathbb Z}} {\mathrm e}^{\frac{2 \pi {\mathrm i} n t}{\epsilon \tau}} \int {\mathrm d} \Gamma(x) \; \left[ \hat{\vec{w}}^\text{drv}_{\alpha(t)} (x) * \hat{\vec{\rho}}_{\alpha(t)}(x) \right]_n .
\label{equ:generalized_housekeeping_work}
\end{equation}
Then we treat the zero component separately from the rest of the terms and rewrite it to the form more suitable for further analysis
\begin{multline}
{\mathcal W}^\text{kh}(\alpha(t))
= \frac{1}{\epsilon} \int\limits_0^T {\mathrm d} t \; \left\llangle \hat{\vec{w}}^\text{drv}_{\alpha(t)} \right\rrangle_{\hat{\vec{\rho}}_{\alpha(t)}} - \\
- \sum_{n \in {\mathbb Z} \setminus \{0\} } \frac{{\mathrm i} \tau}{2 \pi n} \Biggl\{ {\mathrm e}^{\frac{2 \pi {\mathrm i} n T}{\epsilon \tau}} \int {\mathrm d} \Gamma(x) \; \left[ \hat{\vec{w}}^\text{drv}_{\alpha(T)} (x) * \hat{\vec{\rho}}_{\alpha(T)}(x) \right]_n - \\
- \int {\mathrm d} \Gamma(x) \; \left[ \hat{\vec{w}}^\text{drv}_{\alpha(0)} (x) * \hat{\vec{\rho}}_{\alpha(0)}(x) \right]_n \Biggr\} + \\
+ \int\limits_0^T {\mathrm d} t \; \sum_{n \in {\mathbb Z} \setminus \{0\} } \frac{{\mathrm i} \tau}{2 \pi n} {\mathrm e}^{\frac{2 \pi {\mathrm i} n t}{\epsilon \tau}} \int {\mathrm d} \Gamma(x) \; \partial_t \left[ \hat{\vec{w}}^\text{drv}_{\alpha(t)} (x) * \hat{\vec{\rho}}_{\alpha(t)}(x) \right]_n .
\label{equ:generalized_housekeeping_work_rew}
\end{multline}
We can see that in the quasistatic limit $\epsilon \to 0^+$ the last term tends to go to zero because of Riemann-Lebesgue lemma,
the first term is responsible for the divergence of the ``housekeeping'' component of the work in the quasistatic limit.
The second term has the most interesting properties,
in general it is of the same order of the ``reversible'' component and hence in principle it can be comparable,
it is also responsible for the ``housekeeping'' component not being invariant with respect of the external parameters trajectory reversal $\alpha (t) \to \alpha(T-t)$ unless taken the limit over the multiples of period $\tau$.
Moreover the second term heavily depends on the initial and final condition in the sense that it is oscillating and hence does not converge in general in the quasistatic limit.
From physical point of view this term corresponds to the choice of the initial and final ``phase'' or in other words to the choice of the initial and final time within the scope of one period.
It also means that the second term then can play an important role in the evaluation of experimental results.
To illustrate it consider two different experimental setups, in whose the quasistatic limit is approached by extending the time interval,
in the first setup we assume we are able to determine the times inside the period quite precisely and the fluctuations are small across ale runs of the experiment, then the second term can contribute significantly.
In the second setup lets assume that we are not able to control at all the initial and final time within the scope of one period,
then we can assume the ``phases'' are uniformly distributed and then if we average over the ``phases'' before taking the quasistatic limit, we can see that the second term does not contribute.
\subsection{Quasistatic energetics}
In previous chapter \ref{chapter:non-equilibrium_thermodynamics} we have obtained the firs law of thermodynamics by analyzing the difference of the mean value of the total energy in the quasistatic limit.
Here we choose a bit different approach, we start with combination of ``reversible'' components only and then discuss how it is related to the change of energy and compare it with sum of the full quasistatic expansion of heat and work together.
In the previous subsection we have mainly discussed the quasistatic expansion of the work,
however in a similar fashion we can also obtain the ``housekeeping'' and ``reversible'' component of the heat
\begin{align}
{\mathcal Q}^\text{hk}(\alpha) &= \int\limits_0^{\frac{T}{\epsilon}} {\mathrm d} t \; \left\langle q_{t;\alpha(\epsilon t)} \right\rangle_{\rho_{t;\alpha(\epsilon t)}} , \label{equ:generalized_housekeeping_heat} \\
{\mathcal Q}^\text{rev}(\alpha) &= - \int {\mathrm d} \alpha \cdot \left\llangle \nabla_\alpha \frac{1}{{\hat{\mathscr L}}_\alpha}[\hat{\vec{q}}_\alpha] \right\rrangle_{\hat{\vec{\rho}}_\alpha} \label{equ:generalized_reversible_heat} .
\end{align}
If we combine the ``reversible'' components and apply the definitions of the local heat production \eqref{def:heat_gen} and the local power \eqref{def:work_nonp_gen} we obtain
\begin{multline*}
{\mathcal W}^\text{rev}(\alpha) + {\mathcal Q}^\text{rev}(\alpha)
= \int {\mathrm d}\alpha \cdot \left\llangle \nabla_\alpha \left( \hat{\vec{E}}_\alpha - \frac{1}{{\hat{\mathscr L}}_\alpha} [ \hat{\vec{w}}^\text{drv}_\alpha + \hat{\vec{q}}_\alpha ] \right) \right\rrangle_{\hat{\vec{\rho}}_\alpha} = \\
= \int {\mathrm d}\alpha \cdot \left\llangle \nabla_\alpha \left( \hat{\vec{E}}_\alpha - \frac{1}{{\hat{\mathscr L}}_\alpha} {\hat{\mathscr L}}_\alpha [ \hat{\vec{E}}_\alpha ] \right) \right\rrangle_{\hat{\vec{\rho}}_\alpha} ,
\end{multline*}
which corresponds to the change of time-averaged mean value of the energy
\begin{equation}
{\mathrm d} \llangle \hat{\vec{E}}_\alpha \rrangle_{\hat{\vec{\rho}}_\alpha}
= {\mathchar'26\mkern-11mu {\mathrm d}} {\mathcal W}^\text{rev}(\alpha) + {\mathchar'26\mkern-11mu {\mathrm d}} {\mathcal Q}^\text{rev}(\alpha) .
\label{equ:first_law_t}
\end{equation}
This equation we will consider to be a candidate for the generalized first law of thermodynamics\index{first law of thermodynamics!generalized}, with ${\mathcal U}(\alpha) = \llangle \hat{\vec{E}}_\alpha \rrangle_{\hat{\vec{\rho}}_\alpha}$ being a generalizes internal energy\index{internal energy!generalized}.
Notice that in case there is no periodical time dependence the generalized internal energy is already equal to the internal energy introduced in chapter \ref{chapter:non-equilibrium_thermodynamics} and hence the \eqref{equ:first_law_t} corresponds to the equilibrium first law of thermodynamics \eqref{equ:first_law_eq}.
On the other hand if we took the total work and the total heat in the quasistatic expansion we obtain the difference of the total mean energy as a consequence of \eqref{def:local_power_t} and \eqref{def:local_heat_production_t}
\[
{\mathcal Q}(\alpha) + {\mathcal W}(\alpha)
= \int\limits_0^{\frac{T}{\epsilon}} \frac{{\mathrm d} \left\langle E_{t;\alpha(\epsilon t)} \right\rangle_{\mu_t} }{{\mathrm d} t}
= \left\langle E_{\frac{T}{\epsilon};\alpha(T)} \right\rangle_{\mu_{\frac{T}{\epsilon}}} - \left\langle E_{0;\alpha(0)} \right\rangle_{\mu_0} ,
\]
which in general oscillates in the quasistatic limit $\epsilon \to 0^+$.
As a consequence the quasistatic limit of the difference of mean values of energies is not defined.
This reflects the fact that we are not able to uniquely associate a fixed energy with the steady state as it is a periodic function of time.
However we have seen that from thermodynamical point of view it is reasonable to associate the internal energy with the mean steady energy averaged over the period,
as the time-averaged quantity represents the overall energy present in the system independent of the fluctuations on the level of single period
hence the equation \eqref{equ:first_law_t} represents the first law of thermodynamics in systems driven by periodical force.
\subsubsection{Heat capacity}
The fact that the ``reversible'' component of the heat does not depend on the exact parametrization of the path of external parameters as well as its antisymmetry with respect to the trajectory inverse $\alpha(t) \to \alpha(T-t)$,
enables us to extend the definition of the generalized heat capacity\index{generalized heat capacity!periodically driven systems} from the section \ref{sec:quasistatic_processes_noneq} to systems with periodical driving.
We define the generalized heat capacity as
\[
C = \frac{{\mathchar'26\mkern-11mu {\mathrm d}} {\mathcal Q}^\text{rev} (\alpha)}{{\mathrm d} T}
= - \left\llangle \partial_T \frac{1}{{\hat{\mathscr L}}_\alpha}[\hat{\vec{q}}_\alpha] \right\rrangle_{\hat{\vec{\rho}}_\alpha} ,
\]
which is equivalent in case there are present no non-potential forces to
\[
C = \partial_T \left\llangle \hat{\vec{E}}_\alpha \right\rrangle_{\hat{\vec{\rho}}_\alpha} - \left\llangle \partial_T \hat{\vec{E}}_\alpha \right\rrangle_{\hat{\vec{\rho}}_\alpha}
+ \left\llangle \partial_T \frac{1}{{\hat{\mathscr L}}_\alpha} \left[ \hat{\Omega} \cdot \hat{\vec{E}}_\alpha \right] \right\rrangle_{\hat{\vec{\rho}}_\alpha} ,
\]
where the first two terms are presented also in the equilibrium thus the third term being genuine non-equilibrium contribution.
If we introduce a quasi-potential
\[
\hat{\vec{V}}_\alpha (x) = \hat{\vec{E}}_\alpha(x) - \frac{1}{{\hat{\mathscr L}}_\alpha} \left[ \hat{\Omega} \cdot \hat{\vec{E}}_\alpha \right] (x) ,
\]
the heat capacity is then given in the similar fashion as in equilibrium although there is quasi-potential instead of the total energy
\[
C = \partial_T \left\llangle \hat{\vec{V}}_\alpha \right\rrangle_{\hat{\vec{\rho}}_\alpha} - \left\llangle \partial_T \hat{\vec{V}}_\alpha \right\rrangle_{\hat{\vec{\rho}}_\alpha} ,
\]
notice that the steady state does not necessarily need to function of the quasi-potential and hence the generalized heat capacity is not necessarily positive.
\subsection{Clausius relation}
Again as in other cases of non-equilibrium driving the second law of thermodynamics \eqref{equ:second_law} poses no limitations to the ``reversible'' component of the heat as the ``housekeeping'' component diverge in the quasistatic limit.
Although there is no direct consequence of the second law of thermodynamics, one can be still interested in which special cases there is a generalized Clausius relation relating the heat with the entropy production.
In section \eqref{sec:quasistatic_processes_noneq} we have seen that in the case of the system is driven out of equilibrium by attaching the system to multiple thermal baths or by action on non-potential forces the sufficient condition has been that the steady state is function of pseudo-potential.
The same argument is here a bit complicated by the explicit time dependence of all quantities.
To avoid these complications we represent the ``reversible'' component of the heat by time dependent quantities
\[
{\mathchar'26\mkern-11mu {\mathrm d}} {\mathcal Q}^\text{rev} = {\mathrm d} \alpha \cdot \frac{1}{\tau} \int\limits_0^\tau {\mathrm d} t \; \left[ \nabla_\alpha \left\langle V_{t;\alpha} \right\rangle_{\rho_{t;\alpha}} - \left\langle \nabla_\alpha V_{t;\alpha} \right\rangle_{\rho_{t;\alpha}} \right] ,
\]
where the time dependent potential is constructed as $V_{t;\alpha} (x) = \sum_{n \in {\mathbb Z}} \left[ \hat{\vec{V}}_\alpha(x) \right]_n \exp \frac{2 \pi {\mathrm i} n t}{\tau}$.
If we now assume the steady state probability distribution is given as
\[
\rho_{t;\alpha} (x) = \frac{1}{Z_{t;\alpha}} {\mathrm e}^{-\beta(\alpha) V_{t;\alpha} (x)},
\]
where $\beta(\alpha)$ is an arbitrary function of external parameters representing the generalized temperature,
the ``reversible'' component of the heat is given by
\[
{\mathchar'26\mkern-11mu {\mathrm d}} {\mathcal Q}^\text{rev} = {\mathrm d} \alpha \cdot \frac{1}{\tau \beta(\alpha)} \int\limits_0^\tau {\mathrm d} t \; \nabla_\alpha \left\langle - \ln \rho_{t;\alpha} \right\rangle_{\rho_{t;\alpha}} ,
\]
where we can recognize the generalized Clausius relation and where the generalized thermodynamical entropy is given by the time-averaged Shannon entropy
\[
{\mathcal S}(\alpha) = - \frac{1}{\tau} \int\limits_0^\tau {\mathrm d} t \; \left\langle \ln \rho_{t;\alpha} \right\rangle_{\rho_{t;\alpha}} .
\]
\section{Slow driving limit}
\label{ssec:slow}
\index{slow driving limit}
We have seen that in fact we have three time scales present in the system.
The first time scale is given by the evolution of external parameters, the second by the period of the driving and the third by the typical relaxation time of the system $\tau_\text{relax}$.
While in the quasistatic limit the evolution of the external parameters has to be always on the longest time scale
\begin{align*}
\frac{T}{\epsilon} &\gg \tau , &
\frac{T}{\epsilon} &\gg \tau_\text{relax} \quad \forall k ,
\end{align*}
the other two time scales can be arbitrary,
i.e. we can have $\tau > \tau_\text{relax}$ as well as $\tau < \tau_\text{relax}$.
The two limiting cases then being the slow driving limit $\tau \gg \tau_\text{relax}$ and the singular driving limit $\tau \ll \tau_\text{relax}$.
In this section we will focus on the slow driving limit while in the other section we will analyse the singular driving limit.
The slow driving limit is the limiting case when the period of the driving forces is much larger then the relaxation time of the system, however is is still much smaller than the characteristic time of the change of external parameters $\alpha$.
Technically we characterize the slow driving regime by the small scaling parameter $\eta \ll 1$,
which scales the period of the driving $\tau \to \tau/\eta $,
and which in the slow driving limit goes to zero $\eta \to 0^+$ while scaling the period of the driving up to the infinity.
To avoid dealing with infinities we also rescale the time $t'=t/\eta$, which correspond to observing the system on the time scale of the characteristic relaxation time.
Then on this particular time scale the time evolution of the system is described by
\[
\eta \, \partial_{t'} \mu_{t'}(x) = {\mathcal L}^*_{t'} [\mu_{t'}] (x),
\]
which yields to the generalized forward Kolmogorov generator being
\begin{equation}
\left[ {\hat{\mathscr L}}^*_\eta \left[\hat{\vec{\mu}}\right] (x) \right]_n
= \sum_{m \in {\mathbb Z}} \hat{{\mathcal L}}^*_{n-m} \left[ \hat{\mu}_m \right] (x) - \eta \frac{2 \pi {\mathrm i} n}{\tau} \hat{\mu}_n (x) .
\label{equ:generator_slow}
\end{equation}
We can see that the effect of the slow driving regime is reduced to formal substitution of the spectral matrix $\hat{\Omega} \to \hat{\Omega}' = \eta \hat{\Omega}$ in the definition of generalized forward Kolmogorov generator \eqref{equ:generalized_forward_generator},
where now the spectral matrix $\hat{\Omega}'$ is the small parameter.
The same can be done for the local power \eqref{def:work_nonp_gen} and the local heat production \eqref{def:heat_gen}, from where we obtain
\begin{align*}
\hat{\vec{w}}_{\alpha(t)}^\eta (x) &= \dot\alpha(t) \cdot \left. \nabla_\alpha \hat{\vec{E}}_\alpha(x) \right|_{\alpha=\alpha(t)} + \eta \, \hat{\Omega} \cdot \hat{\vec{E}}_{\alpha(t)}(x), \\
\hat{\vec{q}}_{\alpha(t)}^\eta (x) &= {\hat{\mathscr L}}_\eta \left[ \hat{\vec{E}}_{\alpha(t)} \right] (x) - \eta \, \hat{\Omega} \cdot \hat{\vec{E}}_{\alpha(t)}(x).
\end{align*}
\subsection{Steady state}
In the slow driving limit one expects the steady state of the system to be such that at each time it is in the steady state of the dynamics corresponding to that particular time as the driving is so slow that it provides the system enough time to relax there.
To prove this proposition we start need to assume that for sufficiently small $\eta$ the steady state can be expanded into the power series
\[
\hat{\vec{\rho}}^\eta (x) = \sum_{i=0}^\infty \eta^i \, \hat{\vec{\rho}}^{(i)} (x) .
\]
Then by inserting this assumption to the steady state condition \eqref{equ:generalized_stationary_distribution}
we obtain a set of equations determining the steady state
\begin{align*}
\sum_{m \in {\mathbb Z}} \hat{{\mathcal L}}^*_{n-m} \left[ \hat{\rho}^{(0)}_m \right] (x) &= 0, \\
\sum_{m \in {\mathbb Z}} \hat{{\mathcal L}}^*_{n-m} \left[ \hat{\rho}^{(i+1)}_m \right] (x) &= \frac{2 \pi {\mathrm i} n}{\tau} \hat{\rho}^{(i)}_n (x) , \qquad \forall i>0 .
\end{align*}
From where we can see that the leading order component $\hat{\vec{\rho}}^{(0)}$ given by the first equation is equivalent to the Fourier series of the solution of the equation
\[
{\mathcal L}^*_t \left[ \rho^{(0)}_t \right] (x) = 0 ,
\]
i.e. the leading order of the steady state is really given by steady states \eqref{equ:stationary_condition} at each time with respect to fixed dynamics at that particular time.
In case the system is attached to the single thermal bath at constant inverse temperature and when there are no non-potential forces acting on the system, the driving is present only in the explicit time dependence of the potential
and hence the leading term is given by the Boltzmann or Maxwell-Boltzmann distribution with time dependent potential
\[
\rho^{(0)}_t (x) = \frac{1}{Z_t} {\mathrm e}^{- \beta E_t(x) } .
\]
The corrections beyond the slow driving limit are then obtained by application of the recurrent relation
\begin{equation}
\hat{\vec{\rho}}^{(n+1)} (x) = \frac{1}{{\hat{\mathscr L}}^*_0}\left[ \hat\Omega \cdot \hat{\vec{\rho}}^{(n)} \right] (x) ,
\label{equ:recurent_relation_steady_state}
\end{equation}
where we have denoted by
\[
{\hat{\mathscr L}}^*_0 [\hat{\vec{\mu}}] (x) = \sum_{m \in {\mathbb Z}} {\mathcal L}^*_{n-m} [ \hat{\mu}_m ] (x)
\]
the generalized forward Kolmogorov generator in the slow driving limit $\eta \to 0^+$,
and where $1/{\hat{\mathscr L}}^*_0$ is the pseudoinverse \eqref{def:generalized_forward_pseudoinverse} although this time with the projector to the zero eigenvalue subspace is given by
\[
{\hat{\mathscr P}}^*_0 [\hat{\vec{\mu}}] (x) = \sum_{n \in {\mathbb Z}} \hat{\vec{\rho}}^{(0)}_{+n} (x) \int {\mathrm d} \Gamma(y) \; \hat\mu_n (y) ,
\]
where by $\hat{\vec{\rho}}^{(0)}_{+n}$ we denote shifted zeroth component \eqref{equ:shifted_state}.
This is due to the fact that no only the zeroth component of the steady state is the eigenvector corresponding to the zero eigenvalue of this limiting generator ${\hat{\mathscr L}}^*_0 [ \hat{\vec{\rho}}^{(0)} ] = 0$,
but also the vectors obtained by shifting the zero component are also the eigenvectors to the zero eigenvalue in the slow driving limit.
I.e. the zero eigenvalue is degenerate and hence we need to project out the whole subspace in order for pseudoinverse to exist.
We can conclude that the steady state in the slow driving regime is then explicitly given by a series
\[
\hat{\vec{\rho}}i^\eta (x) = \hat{\vec{\rho}}^{(0)} (x) + \sum_{n=1}^\infty \eta^n \; \underbrace{\frac{1}{{\hat{\mathscr L}}^*_0} \biggl[ \hat\Omega \cdot \frac{1}{{\hat{\mathscr L}}^*_0} \biggl[ \hat{\Omega} \cdot \dots \cdot \frac{1}{{\hat{\mathscr L}}^*_0} \biggl[ \hat{\Omega}}_{n \times} \cdot \hat{\vec{\rho}}^{(0)} \biggr] \dots \biggr] \biggr] (x) .
\]
\subsection{The heat and work}
Having the expansion of the steady state around the slow driving limit we can proceed further to investigate the behaviour of the ``reversible'' work and heat in the slow driving limit.
We start with the ``reversible'' work which we already expand up to the first order in $\eta$
\begin{multline*}
{\mathcal W}^\text{rev} (\alpha)
= \int {\mathrm d} \alpha \cdot \left\llangle \nabla_\alpha \hat{\vec{E}}_\alpha \right\rrangle_{\hat{\vec{\rho}}^{(0)}_\alpha} + \\
+ \eta \left\{ \int {\mathrm d} \alpha \cdot \left\llangle \nabla_\alpha \hat{\vec{E}}_\alpha \right\rrangle_{\hat{\vec{\rho}}^{(1)}_\alpha}
- \int {\mathrm d} \alpha \cdot \left\llangle \nabla_\alpha \frac{1}{{\hat{\mathscr L}}_{\alpha;\eta}} \left[ \hat{\Omega} \cdot \hat{\vec{E}}_\alpha \right] \right\rrangle_{\hat{\vec{\rho}}^{(0)}_\alpha} \right\}
+ \err[2]{\eta} ,
\end{multline*}
from where we can see that the leading order in the expansion is an equilibrium-like term,
which in case when there are no non-potential forces correspond to the work done in equilibrium averaged over the period $\tau$.
Moreover if the trajectory over the external parameters is isothermal it corresponds to the change of the equilibrium free energy $\mathcal F$ averaged over the period
\[
\left. {\mathchar'26\mkern-11mu {\mathrm d}} {\mathcal W}^\text{rev}_0 (\alpha) \right|_{T=\text{const.}}
= {\mathrm d} \alpha \cdot \frac{1}{\tau} \int\limits_0^\tau {\mathrm d} t \; \nabla_\alpha {\mathcal F}_{\alpha;t}
= {\mathrm d} \alpha \cdot \frac{1}{\tau} \int\limits_0^\tau {\mathrm d} t \; \left[ - \frac{1}{\beta} \nabla_\alpha \ln Z_{\alpha;t} \right] .
\]
In order to refine the first order correction to more compact form we expand the pseudoinverse in $\eta$ in a similar fashion as in the section \ref{sec:first_order_expansion} and up to the first order we obtain
\begin{multline*}
{\mathcal W}^\text{rev} (\alpha)
= \int {\mathrm d} \alpha \cdot \left\llangle \nabla_\alpha \hat{\vec{E}}_\alpha \right\rrangle_{\hat{\vec{\rho}}^{(0)}_\alpha} + \\
- \eta \int {\mathrm d} \alpha \cdot \left\llangle \hat{\Omega} \cdot \frac{1}{{\hat{\mathscr L}}_{\alpha;0}} \left[ \nabla_\alpha \hat{\vec{E}}_\alpha \right]
+ \nabla_\alpha \frac{1}{{\hat{\mathscr L}}_{\alpha;0}} \left[ \hat{\Omega} \cdot \hat{\vec{E}}_\alpha \right] \right\rrangle_{\hat{\vec{\rho}}^{(0)}_\alpha}
+ \err[2]{\eta} ,
\end{multline*}
where we have explicitly used the expression for the first order correction in the steady state \eqref{equ:recurent_relation_steady_state} and the definition of generalized backward Kolmogorov generator \eqref{def:generalized_backward_pseudoinverse} along with the behaviour of the spectral matrix $\hat{\Omega}$ under the mean value \eqref{equ:spectral_matrix_property},
from where we can see that the first order correction is proportional to the frequency of the driving or to be more precise to the ration of the characteristic relaxation time and the period of driving.
Similarly for the ``reversible'' heat we obtain
\begin{multline*}
{\mathcal Q}^\text{rev} (\alpha)
= \int {\mathrm d} \alpha \cdot \left[ \nabla_\alpha \left\llangle E_\alpha \right\rrangle_{\hat{\vec{\rho}}^{(0)}} - \left\llangle \nabla_\alpha E_\alpha \right\rrangle_{\hat{\vec{\rho}}^{(0)}} \right] + \\
- \int {\mathrm d} \alpha \cdot \left\{ \eta \left\llangle \hat{\Omega} \cdot \frac{1}{{\hat{\mathscr L}}_{\alpha;0}} \left[ \nabla_\alpha \hat{\vec{E}}_\alpha \right] \right\rrangle_{\hat{\vec{\rho}}_\alpha^{(0)}}
+ \nabla_\alpha \left\llangle \frac{1}{{\hat{\mathscr L}}_{\alpha;\eta}} {\hat{\mathscr L}}_{\alpha;0} \left[ \hat{\vec{E}}_\alpha \right] \right\rrangle_{\hat{\vec{\rho}}^{(0)}} \right\}
+ \err[2]{\eta} ,
\end{multline*}
where we can see the first term being again equilibrium-like as it is determined only by the energy function $\hat{\vec{E}}_\alpha$
and from where it follows that the heat generalized heat capacity in the slow driving limit in case there are no non-potential forces corresponds to the equilibrium heat capacity averaged over the period of the driving $\tau$
\[
C_\alpha = \frac{{\mathchar'26\mkern-11mu {\mathrm d}} {\mathcal Q}^\text{rev}(\alpha)}{{\mathrm d} T} = \frac{1}{\tau} \int\limits_0^\tau {\mathrm d} t \; C_{\alpha;t} + \err{\eta} .
\]
Notice that the generalized heat capacity is positive as in the leading order it is given by the time average of equilibrium heat capacities, which are known to be positive \cite{Callen1985}.
At last let us remark, that in case there are no non-potential forces acting on the system the extensive part of the ``housekeeping'' heat and work over the time interval $T$ are given by
\[
{\mathcal W}^\text{hk} (\alpha) = - {\mathcal Q}^\text{hk} (\alpha) = \frac{\eta}{\epsilon} \int\limits_0^T {\mathrm d} t \; \left\llangle \hat{\Omega} \cdot \hat{\vec{E}}_{\alpha(t)} \right\rrangle_{\hat{\vec{\rho}}_{\alpha(t);\eta}} ,
\]
which magnitude is of order of $\eta$.
What need to noticed is that although this term is of order $\eta$ it cannot be in slow driving limit neglected, because the external parameters are evolving on much longer time-scale, $\epsilon \ll \eta$.
\section{Singular driving limit}
\label{ssec:fast}
\index{singular driving limit}
Another limit can be obtained, when the system is driven so fast, that the relaxation time $\tau_R$ of the system is much longer than the period of the driving $\tau$.
The limit when the ratio of relaxation time by period of driving is going to infinity is called a \emph{singular driving limit}, $\tau_R/\tau \rightarrow \infty$.
In this section we want to inspect the system in the singular driving limit as well as obtain first non-zero corrections.
In usual physical situation we have the period of the external force under control not the relaxation time, hence we induce the singular driving limit in the system by scaling the period to zero.
However this approach has serious disadvantage from mathematical point of view.
Especially to avoid dealing with terms of $1/\tau$ in \eqref{equ:generalized_forward_generator} and corresponding quantities,
we rather rescale both the physical time $t$ and the period $\tau$ of the driving by the same scaling factor $\eta \ll 1$,
\begin{align}
\tau \longrightarrow \eta \tau, && t \longrightarrow \eta t,
\label{equ:rescaling}
\end{align}
thus fixating the period and effectively sending the relaxation time to infinity.
The singular driving limit is then obtained by taking the limit $\eta \rightarrow 0^+$.
Notice that the hierarchy of time scales in the singular driving limit follows
\[
\tau \ll \frac{\tau_{\text{relax}}}{\eta} \ll \frac{T}{\varepsilon},
\]
where by $T/\varepsilon$ we denote the characteristic time scale of the quasistatic transformations.
This setup yields to the same time evolution as in \eqref{equ:generalized_time_evolution} with $\eta$-dependent generator
\[
\left[ {\hat{\mathscr L}}^*_\eta \right]_{mn} = \eta \hat{\mathcal L}^*_{m-n} - \frac{2 \pi {\mathrm i} m}{\tau} \delta_{mn} ,
\]
compare with \eqref{equ:generalized_forward_generator} or with \eqref{equ:generator_slow}.
Although the scaling also affects other aspects of the system like local heat power,
for now we will solely focus on the consequences of this particular structure of generalized forward generator.
\subsection{Steady state}
At first we will inspect the stationary distribution and its corrections.
The stationary distribution is given by \eqref{equ:generalized_stationary_distribution}, which in this particular case yields to set of equations
\begin{equation}
0 = \eta \sum\limits_{n \in {\mathbb Z} } \hat{\mathcal L}^*_{m-n} \left[ \hat\rho_n \right] (x) - \frac{2 \pi {\mathrm i} m}{\tau} \hat\rho_m (x).
\label{equ:fast_stationary_distribution}
\end{equation}
To proceed further we expect that the typical physical system in the singular driving limit does not ``feel'' the oscillations of the driving force
and that the oscillations became more important as the system is further from singular driving limit.
This expectation leads us to the assumption that the stationary distribution is analytical in $\eta$,
and hence we can expand the stationary distribution in powers of $\eta$ with zeroth leading order
\[
\hat\rho (x) = \hat\rho^{(0)} (x) + \eta \hat\rho^{(1)} (x) + \eta^2 \hat\rho^{(2)} (x) + \dots .
\]
Inserting the expansion of the stationary distribution to the \eqref{equ:fast_stationary_distribution}
and inspecting the zeroth order for $m \neq 0$ we immediately find that the non-zero components of the zeroth order contribution vanish
\[
\forall n \neq 0 : \quad \hat\rho^{(0)}_n (x) = 0,
\]
i.e. only the zeroth component of the stationary distribution is essentially non-zero.
Applying the same technique but inspecting higher orders for $m \neq 0$ we also obtain a recurrent relation for non-zero components of the stationary distribution
\begin{equation}
\forall n \neq 0, k \in \mathbb N : \quad \hat\rho^{(k+1)}_n (x) = - \frac{{\mathrm i} \tau}{2 \pi n} \sum\limits_{m \in {\mathbb Z}} \hat{\mathcal L}^*_{n-m} \left[ \hat\rho^{(k)}_m \right] (x).
\label{equ:fast_stationary_density_nonzero}
\end{equation}
The only remaining independent set of equations left in \eqref{equ:fast_stationary_distribution} corresponds to $m=0$,
\[
\forall k \in \mathbb N : \quad 0 = \sum\limits_{n \in {\mathbb Z}} \hat{\mathcal L}^*_{-n} \left[ \hat\rho^{(k)}_{n} \right] (x),
\]
which determine the zeroth component.
In the zeroth order we obtain the zeroth component as a solution of equation
\begin{equation}
0 = \hat{\mathcal L}^*_0 \left[ \hat\rho^{(0)}_0 \right] (x) ,
\label{equ:steady_state_fast}
\end{equation}
while for higher order contribution of the zeroth component we obtain a recurrent relation
\begin{equation}
\hat\rho^{(k)}_0 (x) = - \frac{1}{\hat{\mathcal L}^*_0} \sum\limits_{n \neq 0} \hat{\mathcal L}^*_{-n} \left[ \hat\rho^{(k)}_n \right] (x) ,
\label{equ:fast_stationary_density_zero}
\end{equation}
where $1/\hat{\mathcal L}^*_0$ is the pseudo-inverse\index{pseudo-inverse} \eqref{def:forward_pseudoinverse} as in chapter \ref{chapter:non-equilibrium_thermodynamics} with $\hat{\mathcal L}^*_0$ as the explicitly time independent generator, i.e.
\[
\frac{1}{\hat{\mathcal L}^*_0} \left[ \mu \right] = \int\limits_0^\infty {\mathrm d} s \; \left\{ \hat\rho^{(0)}_0 (x) \int {\mathrm d} y \; \mu(y) - {\mathrm e}^{s \hat{\mathcal L}^*_0} \left[ \mu \right] (x) \right\}.
\]
We can see, that in the singular driving limit the system's steady state is explicitly time independent as expected, $\rho_t (x) \equiv \hat\rho^{(0)}_0 (x)$,
and is given by time-averaged generator,
\[
0 = \frac{1}{\tau} \int\limits_0^\tau {\mathrm d} t \; {\mathcal L}^*_t [ \rho ] = \hat{\mathcal L}^*_0 [ \rho ].
\]
Notice that even in the case the periodic time dependence of the system is given only by the time periodic potential,
the stationary distribution with respect to time-averaged generator is not necessarily the stationary distribution obtained by implying the time-averaged potential,
as we will briefly discuss further.
We can also obtain corrections beyond the singular driving limit by using the recurrent relations \eqref{equ:fast_stationary_density_nonzero} and \eqref{equ:fast_stationary_density_zero}.
Combining these equations together we can explicitly obtain the first order correction
\begin{align*}
\hat\rho^{(1)}_{n \neq 0} (x) &= - \frac{{\mathrm i} \tau}{2 \pi n} \hat{\mathcal L}^*_n \left[ \hat\rho^{(0)}_0 \right] (x), &
\hat\rho^{(1)}_0 (x) &= \frac{{\mathrm i} \tau}{2 \pi} \frac{1}{\hat{\mathcal L}^*_0} \left[ \sum\limits_{n \neq 0} \frac{1}{n} \hat{\mathcal L}^*_{-n} \hat{\mathcal L}^*_n \left[ \hat\rho^{(0)}_0 \right] \right] (x),
\end{align*}
where the zeroth component of the first order correction can be also rewritten using \eqref{equ:complex_conjugation_of_generator} as
\[
\hat\rho^{(1)}_0 (x) = - \frac{\tau}{\pi} \frac{1}{\hat{\mathcal L}^*_0} \left[ \sum\limits_{n = 1}^\infty \frac{1}{n} \, \Im \left( \overline{\hat{\mathcal L}^*_n} \hat{\mathcal L}^*_n \right) \left[ \hat\rho^{(0)}_0 \right] \right] (x),
\]
where $\Im$ denotes the imaginary part and $\bar{x}$ denotes the complex conjugation.
\subsection{Generalized pseudo-inverse}
Before we proceed further to the analysis of the work and heat in the singular driving limit,
we need to investigate the behaviour of the pseudoinverse in the quasistatic limit.
The basic problem here is that the limit of the forward Kolmogorov generator $\lim_{\eta \to 0^+} {\hat{\mathscr L}}^*_\eta $ does not evolve towards any steady state
and hence the generalized forward pseudoinverse \eqref{def:generalized_forward_pseudoinverse} as well as generalized backward pseudoinverse \eqref{def:generalized_backward_pseudoinverse}
cannot be defined in that way as the integrals in definitions does not converge.
To at least partially avoid the problem, let us then define the regularized \emph{backward} generator as
\[
\left[ {\hat{\mathscr L}}_\epsilon [\hat{\vec{A}} ] (x) \right]_n = \left[ \hat\Omega \cdot \hat{\vec{A}} (x) \right]_n +
\begin{cases}
\epsilon \hat{\mathcal L}_0 \left[ \hat{A}_0 \right] (x) & n = 0 \\
- \epsilon \hat{A}_n (x) & n \neq 0
\end{cases} ,
\]
which in limit $\epsilon \to 0^+$ corresponds to the limit of the generalized forward Kolmogorov generator,
otherwise ensures the observable to converge toward its steady mean value averaged over the period $\tau$ in the singular driving limit.
Then the generalized pseudoinverse in the limit is given as the limit of the pseudoinverse for regularized generator
\[
\left[ \frac{1}{{\hat{\mathscr L}}_\epsilon} [ \hat{\vec{A}} ] (x) \right]_n
= \begin{cases}
\displaystyle \frac{1}{\epsilon \, \hat{{\mathcal L}}_0 } \left[ \hat{A}_0 \right] (x) & n = 0 \\ \\
\displaystyle \frac{\tau}{2 \pi {\mathrm i} n - \tau \epsilon} \hat{A}_n (x) & n \neq 0
\end{cases} ,
\]
which diverges in case of the zeroth component while it converge for any other component.
Although the zeroth component of the pseudoinverse of regularized generator,
in the ``reversible'' component of the work and the heat the pseudoinverse is applied to the quantity which zeroth components is zero $A_0(x) = 0$.
\subsection{Heat and work}
In previous subsections we have analysed the structure of the steady state in the singular driving limit and prepared the regularized backward pseudoinverse as the leading order of the generalized backward generator in the singular driving expansion.
Now we will proceed to the expansion of the work and heat in the singular driving limit.
We start with the analysis of the local power and the local heat production.
The local power in the singular driving limit is given by the equation \eqref{def:work_nonp_gen}
\[
\hat{\vec{w}}_{\alpha(t)} (x) = \dot\alpha(t) \cdot \left. \nabla_\alpha \hat{\vec{E}}_\alpha (x) \right|_{\alpha = \alpha(t)} + \hat{\Omega} \cdot \hat{\vec{E}}_{\alpha(t)} ,
\]
as it is independent of the backward Kolmogorov generator,
on the other hand we can deduce that the local heat production \eqref{def:heat_gen} is entirely of the first order in $\eta$
\[
\left[ \hat{\vec{q}}_\alpha (x) \right]_n = \left[ {\hat{\mathscr L}}_\alpha [ \hat{\vec{E}}_\alpha ] (x) \right]_n - \left[ \hat{\Omega} \cdot \hat{\vec{E}}_\alpha (x) \right]_n
= \eta \sum\limits_{m \in {\mathbb Z}} \hat{\mathcal L}_{n-m;\alpha} [ \hat{E}_{m;\alpha} ] (x) .
\]
The leading order of the ``reversible'' work is then obtained by application of the regularized pseudoinverse on the second term in the local power which yields to
\begin{multline*}
{\mathcal W}^\text{rev} (\alpha) = \int {\mathrm d} \alpha \cdot \left\llangle \nabla_\alpha \left( \hat{\vec{E}}_\alpha - \frac{1}{{\hat{\mathscr L}}_\alpha} \left[ \hat{\Omega} \cdot \hat{\vec{E}}_\alpha \right] \right) \right\rrangle_{\hat{\vec{\rho}}_{\alpha}} = \\
= \int {\mathrm d} \alpha \cdot \left\langle \nabla_\alpha \hat{E}_{0;\alpha} \right\rangle_{\hat{\rho}^{(0)}_0} + \err{\eta} .
\end{multline*}
Similarly for the heat we have
\begin{multline*}
{\mathcal Q}^\text{rev} (\alpha) = \int {\mathrm d} \alpha \cdot \left\{ \nabla_\alpha \left\llangle \hat{\vec{E}}_\alpha \right\rrangle_{\hat{\vec{\rho}}_{\alpha}} - \left\llangle \nabla_\alpha \left( \hat{\vec{E}}_\alpha - \frac{1}{{\hat{\mathscr L}}_\alpha} \left[ \hat{\Omega} \cdot \hat{\vec{E}}_\alpha \right] \right) \right\rrangle_{\hat{\vec{\rho}}_{\alpha}} \right\} = \\
= \int {\mathrm d} \alpha \cdot \left\{ \nabla_\alpha \left\langle \hat{E}_{0;\alpha} \right\rangle_{\hat{\rho}^{(0)}_0} - \left\langle \nabla_\alpha \hat{E}_{0;\alpha} \right\rangle_{\hat{\rho}^{(0)}_0} \right\} + \err{\eta} ,
\end{multline*}
from where under the assumption that the energy function $\hat{E}_0$ is independent of temperature $T$ we obtain the generalized heat capacity as
\[
C = \frac{{\mathrm d}}{{\mathrm d} T} \left\langle \hat{E}_{0;\alpha} \right\rangle_{\hat{\rho}^{(0)}_0} + \err{\eta} .
\]
Notice that we have obtained in the leading order the generalized heat capacity to be given by the temperature derivative of the internal energy.
Moreover in case the driving of the system is given by periodic time dependence of the energy function and if the generator depends on the potential linearly we obtain the equilibrium heat capacity with respect to the energy function averaged over the period $\tau$ and thus the positivity of the generalized heat capacity is guaranteed.
Notice that also in this case the strongly non-equilibrium behaviour of the system can be expected only in the intermediate regime where the period of driving $\tau$ and the typical relaxation time are comparable.
Although the Kolmogorov generators for diffusions \eqref{equ:forward_generator_underdamped} and \eqref{equ:time_evolution_overdamped} are linear in the potential,
the Kolmogorov generator for jump processes is in general not \eqref{equ:forward_Kolmogorov_gen_jump},
thus the generalized heat capacity in case of jump processes in the singular driving limit does not necessarily correspond to any equilibrium heat capacity.
The extensive part of the ``housekeeping'' component of the total heat and the total work simplifies in the singular driving limit to
\[
{\mathcal W}^\text{hk} (\alpha(t)) = - {\mathcal Q}^\text{hk} (\alpha(t))
= \frac{1}{\epsilon} \int\limits_0^T {\mathrm d} t \; \left\llangle \hat{\Omega} \cdot \hat{\vec{E}}_{\alpha(t)} \right\rrangle_{\hat{\vec{\rho}}_{\alpha(t)}}
= \frac{\eta}{\epsilon} \int\limits_0^T {\mathrm d} t \; \left\llangle \hat{\Omega} \cdot \hat{\vec{E}}_{\alpha(t)} \right\rrangle_{\hat{\vec{\rho}}^{(1)}_{\alpha(t)}} + \err[2]{\eta},
\]
where we have used the fact that in the leading order only the zeroth component of the steady state is non-zero while the zeroth component of the local power of the driving forces is zero, thus effectively canceling each other out.
Again as in the slow driving limit, although the extensive parts of ``housekeeping'' heat and work are of linear order in $\eta$ they are in general not negligible as the external parameters are again changed on much longer time scale.
\section{Example: Periodically driven two-level models}
\label{sec:two_level_models_t}
\index{discrete model!two-level}
\index{discrete model!two-level!time dependent}
In this section we provide a simple example of periodically driven system with only two configurations and for two physically different scenarios.
The first scenario corresponds, e.g., to an "incoherent" hoping between two levels in a quantum dot.
In the second scenario the two configurations can represent a pair of metastable states of a complex system, mutually separated by an energy barrier.
In both cases the local detailed balance condition related the time-dependent transition rates to the time-dependent energy levels.
We mostly provide numerical results only.
\subsection{Scenario A: no barrier}
\begin{figure}[ht]
\caption{Illustration of the two-level model with periodical driving.}
\label{pic:two_level_illustration}
\begin{center}
\includegraphics[width=.8\textwidth,height=!]{two_level.pdf}
\end{center}
\end{figure}
The standard two-level model, see subsection \ref{ssec:two_level}, is determined by energy levels,
which in this case periodically depends on the time
\[
E_t = \begin{bmatrix}
\epsilon_0 + \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \\
- \epsilon_0 - \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} ,
\end{bmatrix}
\]
where $2 \epsilon_0$ is the basic gap between the states and $\Delta \epsilon$ is the amplitude of the energy levels movement.
The periodic transition rates are then given by
\[
\rate[t]{x}{y} = \psi \exp\left[ -\frac{\beta}{2} \left( E_t(y) -E_t(x) \right) \right].
\]
Notice that they obey detailed balance condition \eqref{equ:global_detailed_balance_jump} at each time $t$.
Given the transition rates we can construct the forward Kolmogorov generator, which also explicitly depends on time
\[
{\mathcal L}^*_t = \begin{bmatrix}
- \psi \exp \left[ \beta \left( \epsilon_0 + \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \right) \right] &
\psi \exp \left[ - \beta \left( \epsilon_0 + \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \right) \right] \\
\psi \exp \left[ \beta \left( \epsilon_0 + \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \right) \right] &
- \psi \exp \left[ - \beta \left( \epsilon_0 + \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \right) \right] \\
\end{bmatrix}
\]
and equivalently similarly we can define the backward Kolmogorov generator, which is in this case equivalent to the transpose of matrix representation of forward Kolmogorov generator.
\[
{\mathcal L}_t = \left( {\mathcal L}^*_t \right)^T = \begin{bmatrix}
- \psi \exp \left[ \beta \left( \epsilon_0 + \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \right) \right] &
\psi \exp \left[ \beta \left( \epsilon_0 + \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \right) \right] \\
\psi \exp \left[ - \beta \left( \epsilon_0 + \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \right) \right] &
- \psi \exp \left[ - \beta \left( \epsilon_0 + \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \right) \right] \\
\end{bmatrix} .
\]
Before we show on this particular example the construction of the generalized forward Kolmogorov generator and local power in terms of Fourier components we have to obtain the local power and the local heat production in the time domain first.
The non-potential part of the local power is given by the time derivative of the total energy with external parameters fixed, in this particular case $\Delta \epsilon$ and $\epsilon_0$ is sufficient
\[
w^\text{drv}_t =
\dot{E_t} = \frac{\Delta \epsilon \pi}{\tau} \sin \left( \frac{2 \pi t}{\tau} \right) \cdot
\begin{bmatrix}
- 1 \\
1
\end{bmatrix} ,
\]
whether the local heat production is then given by \eqref{def:local_heat_production_t}
\[
q_t = {\mathcal L}_t \left[ E_t \right] = 2 \psi \left( \epsilon_0 + \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \right) \cdot
\begin{bmatrix}
- \exp \left[ \beta \left( \epsilon_0 + \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \right) \right] \\
\exp \left[ - \beta \left( \epsilon_0 + \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \right) \right]
\end{bmatrix} .
\]
Now we can proceed further by finding the Fourier components of the energy
\begin{align*}
\left[ \hat{E} \right]_0 &= \epsilon_0 \cdot
\begin{bmatrix}
1 \\
- 1
\end{bmatrix} &
\left[ \hat{E} \right]_{\pm 1} &= \frac{\Delta \epsilon}{4} \cdot
\begin{bmatrix}
1 \\
- 1
\end{bmatrix}
\end{align*}
The local power in terms of Fourier components is
\begin{align*}
\left[ \hat{\vec{w}} \right]_{-1} &= -\frac{{\mathrm i} \pi \Delta \epsilon}{2 \tau} \begin{pmatrix} 1 \\ -1 \end{pmatrix} &&&
\left[ \hat{\vec{w}} \right]_1 &= \frac{{\mathrm i} \pi \Delta \epsilon}{2 \tau} \begin{pmatrix} 1 \\ -1 \end{pmatrix} &&&
\left[ \hat{\vec{w}} \right]_n &= \begin{pmatrix} 0 \\ 0 \end{pmatrix} && n \neq \pm 1
\end{align*}
while the local heat power is represented by
\[
\left[ \hat{\vec{q}} \right]_n = \frac{{\mathrm i}^n \psi \Delta \epsilon}{2}
\begin{pmatrix}
J_{n+1} \left( \frac{{\mathrm i} \beta \Delta \epsilon}{2} \right) + J_{n-1} \left( \frac{{\mathrm i} \beta \Delta \epsilon}{2} \right) \\
J_{n+1} \left( - \frac{{\mathrm i} \beta \Delta \epsilon}{2} \right) + J_{n-1} \left( - \frac{{\mathrm i} \beta \Delta \epsilon}{2} \right)
\end{pmatrix},
\]
where $J_n$ is Bessel function of the first order.
Similarly one can obtain diagonal
\[
\left[{\hat{\mathscr L}}^* \right]_{nn} = {\mathrm i}^n \psi
\begin{pmatrix}
\frac{2 \pi {\mathrm i} n}{\tau} - J_0 \left( \frac{{\mathrm i} \beta \Delta \epsilon}{2} \right) & J_0 \left( - \frac{{\mathrm i} \beta \Delta \epsilon}{2} \right) \\
J_0 \left( \frac{{\mathrm i} \beta \Delta \epsilon}{2} \right) & \frac{2 \pi {\mathrm i} n}{\tau} - J_0 \left( - \frac{{\mathrm i} \beta \Delta \epsilon}{2} \right)
\end{pmatrix}
\]
as well as of-diagonal elements of effective generator ${\hat{\mathscr L}}$
\begin{align*}
\left[{\hat{\mathscr L}}^* \right]_{nm} &= {\mathrm i}^{n-m} \psi
\begin{pmatrix}
- J_{n-m} \left( \frac{{\mathrm i} \beta \Delta \epsilon}{2} \right) & J_{n-m} \left( - \frac{{\mathrm i} \beta \Delta \epsilon}{2} \right) \\
J_{n-m} \left( \frac{{\mathrm i} \beta \Delta \epsilon}{2} \right) & - J_{n-m} \left( - \frac{{\mathrm i} \beta \Delta \epsilon}{2} \right)
\end{pmatrix}
&& n \neq m.
\end{align*}
One can then easily check that relations \eqref{def:heat_gen} and \eqref{def:work_nonp_gen} are valid.
\begin{figure}[ht]
\caption{Heat capacity of the driven two-level system as a function of inverse temperature $\beta$ and driving frequency $\frac{1}{\tau}$ on the logarithmic scale, with the average relaxation time denoted by the green dashed line.
$\psi=25$, $\Delta \epsilon = 1$, $\epsilon_0 = 1/8$
}
\label{pic:two_level_C_beta}
\begin{center}
\includegraphics[width=.95\textwidth,height=!]{2lvl-C_beta_.pdf}
\end{center}
\end{figure}
The generalized heat capacity $C$ is numerically evaluated with respect to inverse temperature $\beta$ and frequency of the driving $1/\tau$ on exponential scale, see figure~\ref{pic:two_level_C_beta}.
As one can see, the system exhibits slow driving limit behaviour as well as singular driving limit behaviour, depending on the driving frequency with respect to characteristic relaxation time $\tau_{relax}$.
Although we cannot determine characteristic relaxation time analytically, we can heuristically estimate the temperature dependence by analysing the relaxation time for each fixed time, i.e. fixed value of the potential.
The heuristic estimate of the characteristic time may be obtained by looking at the typical magnitude of the relaxation times corresponding to the dynamics "frozen" at each time $t$, which has the form
\begin{align*}
\ln \left[ \frac{1}{\tau_{relax} (t)} \right]
&= \ln \left[ k_t ( 1 \rightarrow 2 ) + k_t ( 2 \rightarrow 1 ) \right] \\
&= \ln \psi + \ln \left[ 2 \cosh \left( \beta \left( \epsilon_0 + \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \right) \right) \right].
\end{align*}
In low temperature region $\beta \gg 1$ the logarithm of relaxation time can be easily approximated by linear dependence on temperature
\[
\ln \left[ \frac{1}{\tau_{relax} (t)} \right]
\sim \beta \left| \epsilon_0 + \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \right| \ge 0 ,
\]
while in the high temperature region $\beta \ll 1$ the temperature dependence of logarithm of characteristic relaxation time vanish
\[
\ln \left[ \frac{1}{\tau_{relax} (t)} \right]
= \ln \psi + \left( \epsilon_0 + \frac{\Delta \epsilon}{2} \cos \left( \frac{2 \pi t}{\tau} \right) \right)^2 \beta^2 + \err[4]{\beta} ,
\]
which is in good qualitative agreement with results shown in the figure \ref{pic:two_level_C_beta}.
Notice also that the region with negative generalized heat capacity is in the vicinity of the characteristic relaxation time, hence it is in the vicinity of the transition regime between fast and slow driving.
\subsection{Scenario B: with barrier}
\begin{figure}[ht]
\caption{Illustration of the two-level model with barrier with periodical driving.}
\label{pic:two_level_barrier_illustration}
\begin{center}
\includegraphics[width=.8\textwidth,height=!]{two_level_barrier.pdf}
\end{center}
\end{figure}
The main difference between these two models is the presence of the barrier in this model, hence energy levels are the same as in previous case
\[
E_t = \begin{bmatrix}
\epsilon_0 + \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \\
- \epsilon_0 - \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau}
\end{bmatrix}
\]
while the transition rates are given by
\[
\rate[t]{x}{y} = \psi \exp\left[ - \beta \left( \Delta - E_t(x) \right) \right],
\]
where $\Delta$ is the height of barrier, also one can verify that even this choice of transition rates obey the detailed balance condition \eqref{equ:global_detailed_balance_jump}.
Given the transition rates we can again construct the forward Kolmogorov generator
\[
{\mathcal L}^*_t = \begin{bmatrix}
- \psi \exp \left[ - \beta \left( \Delta - \epsilon_0 - \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \right) \right] &
\psi \exp \left[ - \beta \left( \Delta + \epsilon_0 + \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \right) \right] \\
\psi \exp \left[ - \beta \left( \Delta - \epsilon_0 - \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \right) \right] &
- \psi \exp \left[ - \beta \left( \Delta + \epsilon_0 + \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \right) \right] \\
\end{bmatrix}
\]
and equivalently similarly we can define the backward Kolmogorov generator, which is in this case equivalent to the transposed matrix of the forward Kolmogorov generator
\[
{\mathcal L}_t = \left( {\mathcal L}^*_t \right)^T = \psi {\mathrm e}^{- \beta \Delta} \begin{bmatrix}
- \exp \left[ \beta \left( \epsilon_0 + \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \right) \right] &
\exp \left[ \beta \left( \epsilon_0 + \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \right) \right] \\
\exp \left[ - \beta \left( \epsilon_0 + \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \right) \right] &
- \exp \left[ - \beta \left( \epsilon_0 + \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \right) \right] \\
\end{bmatrix} .
\]
We can see that the local power is the same
\[
w_t =
\dot{E_t} = \frac{\Delta \epsilon \pi}{\tau} \sin \left( \frac{2 \pi t}{\tau} \right) \cdot
\begin{bmatrix}
- 1 \\
1
\end{bmatrix}
\]
although what differs is the expression for the local heat production
\[
q_t = {\mathcal L}_t \left[ E_t \right] = 2 \psi {\mathrm e}^{-\beta \Delta} \left( \epsilon_0 + \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \right) \cdot
\begin{bmatrix}
- \exp \left[ \beta \left( \epsilon_0 + \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \right) \right] \\
\exp \left[ - \beta \left( \epsilon_0 + \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \right) \right]
\end{bmatrix} .
\]
\begin{figure}[ht]
\caption{Heat capacity of the driven two-level system with barrier as a function of inverse temperature $\beta$ and driving frequency $\frac{1}{\tau}$, with the mean relaxation time denoted by the green dashed line. }
\label{pic:two_level_barrier_C_beta}
\begin{center}
\includegraphics[width=.95\textwidth,height=!]{2lvl-barrier-C_beta_.pdf}
\end{center}
\end{figure}
We again evaluate numerically the generalized heat capacity $C$ with respect to inverse temperature $\beta$ and frequency of the driving $1/\tau$ on exponential scale, see figure~\ref{pic:two_level_barrier_C_beta}.
As one can see, the system exhibits slow driving limit behaviour as well as singular driving limit behaviour, depending on the driving frequency with respect to characteristic relaxation time $\tau_{relax}$.
Moreover we can see that the system again has a rich behaviour on the time scale comparable with the typical relaxation time.
Again, we are not able to determine characteristic relaxation time analytically,
however we can heuristically estimate the temperature dependence by analysing the relaxation time for each fixed time, i.e. fixed value of potential.
Again, the characteristic time of the dynamics can be estimated via the relaxation times of the "frozen" system,
\begin{align*}
\ln \left[ \frac{1}{\tau_{relax} (t)} \right]
&= \ln \left[ k_t ( 1 \rightarrow 2 ) + k_t ( 2 \rightarrow 1 ) \right] \\
&= \ln \psi - \beta \Delta + \ln \left[ 2 \cosh \left( \beta \left( \epsilon_0 + \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \right) \right) \right].
\end{align*}
In low temperature region $\beta \gg 1$ the logarithm of relaxation time can be easily approximated by linear dependence on temperature, approximatly
\[
\ln \left[ \frac{1}{\tau_{relax} (t)} \right]
\sim \beta \left[ \left| \epsilon_0 + \frac{\Delta \epsilon}{2} \cos \frac{2 \pi t}{\tau} \right| - \Delta \right] \le 0 ,
\]
while in the high temperature region $\beta \ll 1$ the temperature dependence of logarithm of characteristic relaxation time vanish
\[
\ln \left[ \frac{1}{\tau_{relax} (t)} \right]
= \ln \psi - \beta \Delta + \left( \epsilon_0 + \frac{\Delta \epsilon}{2} \cos \left( \frac{2 \pi t}{\tau} \right) \right)^2 \beta^2 + \err[4]{\beta}.
\]
We can see that there is an additional term when compared with classical model which largely influence the transition region.
\section{Conclusions}
We have developed the theory of quasistatic processes in systems driven by periodical action of force.
Our starting point were results of the Floquet theory \eqref{equ:Floquet_theorem}, which we used to reformulate the problem such that the explicit time dependence is removed \eqref{equ:generalized_time_evolution}.
Then we have been able to show that also in this type of driving the quasistatic heat and work can be decomposed to so-called ``housekeeping'' part \eqref{equ:generalized_housekeeping_work} or \eqref{equ:generalized_housekeeping_heat} responsible for maintaining the steady state out of equilibrium
and to the ``reversible'' part \eqref{equ:generalized_reversible_work} or \eqref{equ:generalized_reversible_heat} which is of geometrical nature and corresponds to what we in classical thermodynamics call heat and work, which is our main result of this chapter.
While ``housekeeping'' components of the work \eqref{equ:generalized_housekeeping_work_rew} and heat are in general extensive in time, it also contains fluctuating term tightly bounded with the initial and the terminal condition, which in general does not need to converge in the quasistatic limit and hence can affect the result of potential measurements.
Similarly to the systems driven out by non-potential forces or by attaching the system to multiple thermal bath presented in the chapter \ref{chapter:non-equilibrium_thermodynamics},
the systems driven by periodical forces has also a rich non-equilibrium behaviour such as the general non-validity of the Clausius relation, negative generalized heat capacities,
as was shown on two level models, see section \ref{sec:two_level_models_t}.
However as our results for slow, see subsection \ref{ssec:slow}, and singular, see subsection \ref{ssec:fast}, driving limit suggests,
this behaviour is typically limited to the intermediate regime where the period of the driving is comparable to the relaxation time, as was also illustrated in the aforementioned models.
The article summarizing these results is still in preparation \cite{result3}.
\chapter{Slow-fast coupling}
\label{chapter:slow-fast_coupling}
Until now we have been trying to describe thermodynamical processes of the system in contact with several thermal and chemical baths and in general in the presence of external non-potential force fields,
e.g. electromagnetic field, which time-dependence is periodical.
To be more precise, we have been studying Markovian systems in environment described by few fully manageable external parameters,
which entirely determine the Markovian dynamics of the system as well as the steady state of the system and which were subjects of the quasistatic transformations.
Although these conditions seem to be very general, not every possible physical situation fits into this description even in the simplified case of quasistatic processes in the vicinity of steady states.
For example if we are interested only in the properties and behaviour of the subsystem of a more complex system,
the evolution of such system in general depends on all degrees of freedom of the composite system, which are basically intractable.
However in some cases we are able to find an effective Markovian evolution for such subsystem, so they can be treated in the same manner as above mentioned systems with few external parameters.
The basic questions are in which systems there exists an effective Markovian description of the time evolution and how to properly construct the description from the description of more complex systems?
In this chapter we will focus on these questions in a particular class of composite systems with so called \emph{separation of timescales}.\index{separation of timescales}
We speak about the separation of timescales, when we can sort the degrees of freedom of a complex system into two or more distinct sets characterized by typical relaxation times.
Moreover we demand that the typical relaxation times $\tau_i$ of each set greatly differs, $\tau_1 \gg \tau_2 \gg \dots \gg \tau_k$.
We will restrict ourselves further on to the case when all degrees of freedom can be sorted only to two of such sets.
The degrees of freedom associated with the long (short) typical relaxation time we call ``slow'' (``fast'').
A typical example of such a system is a heavy particle in the bath of light ones, where the degrees of freedom associated with \emph{heavy particle} are considered to be \emph{``slow''},
i.e. it takes much more time to relax the heavy particle after a perturbation of the environment than to relax light particles.
The dynamical parameters, which we are typically able to control in the experiment, describe the average density of light particles and their local temperature,
however it is difficult to control the feedback between the slow and fast particles as well as local fluctuation of density.
In the first section we will focus on the dynamics of the ``fast'' degrees of freedom on their respective timescale,
while in the second section we will investigate the dynamics of the ``slow'' degrees of freedom again on their respective timescale.
In the context of our example the first section focuses on the dynamics of light particles under the influence of quasistatically moving slow particle,
while in the second section we will describe the movement of the heavy particle in the continuum of the light ones, which keeps nearby a steady state.
Through this chapter we will be again using the framework of Kolmogorov generators firstly introduced in chapter \ref{chapter:models} section \ref{sec:Markov_processes},
where we describe our full system by probabilistic density $\mu_t(x_S,x_F)$,
where by $x_S$ ($x_F$) we explicitly denote the dependence on the ``slow'' (``fast'') degrees of freedom.
We also assume that the time evolution of the full system is Markovian and given by \eqref{equ:Markov_stochastic_time_evolution}
\[
\partial_t \mu_t(x_S,x_F) = {\mathcal L}^*_\epsilon \left[ \mu_t \right] (x_S,x_F)
\]
where the total forward Kolmogorov generator ${\mathcal L}^*_\epsilon$ acts on the overall system and does not explicitly depend on time,
and $\epsilon$ denotes the ratio of the ``fast'' and ``slow'' degrees of freedom's relaxation times ($\tau_F$ and $\tau_S$), $\epsilon = \tau_F/\tau_S$.
The formal solution of such evolution is given by
\begin{equation}
\mu_t (x_F,x_S) = {\mathrm e}^{ \left( t-t_0 \right) {\mathcal L}^*_\epsilon } \left[ \mu_{t_0} \right] (x_S,x_F).
\label{equ:full_solution}
\end{equation}
\section{Autonomous dynamics for fast degrees of freedom}
\label{sec:dynamics_for_fast}
The first situation, which we want to analyse, is the behaviour of the system on the timescale which is comparable to the typical relaxation times of the ``fast'' degrees of freedom.
In that case we would expect to see almost no time evolution of the ``slow'' degrees of freedom.
Hence we presumably expect to obtain a situation very similar to the quasistatic limit, which is a central topic of this thesis, where the ``slow'' degrees of freedom represent the external parameters, although in this case they don't have a prescribed trajectory, but rather evolve on their own.
To show more precisely in what sense the proposition holds true,
we start with the decomposition of the total forward Kolmogorov generator to the part evolving just the ``slow'' degrees of freedom ${\mathcal L}^*_S$ and the part responsible for the time evolution of ``fast'' degrees of freedom ${\mathcal L}^*_F$.
To have a clear distinction between what are ``fast'' and what are ``slow'' degrees of freedom, we assume that they are entangled only by the interaction potential,
i.e. in the forward Kolmogorov generator for ``fast'' degrees of freedom the ``slow'' degrees of freedom occurs only as parameters and vice versa.
To simplify the analysis we introduce the parameter $\epsilon$ which explicitly characterizes the timescale separation.
Hence the time evolution of the joint probability distribution \eqref{equ:Markov_stochastic_time_evolution} is given by
\[
\partial_t \mu_t (x_S,x_F) = \epsilon {\mathcal L}^*_S [\mu_t] (x_S,x_F) + {\mathcal L}^*_F[\mu_t](x_S,x_F) ,
\]
where the only epsilon dependence lies in the pre-factor.
Our main aim is to obtain the time evolution equation for the marginal probability distribution on fast degrees of freedom $\widetilde\mu(x_F)$.
We start with the decomposition of the joint probability distribution to the conditional probability distribution of the ``fast'' degrees of freedom conditioned to the particular configuration of the ``slow'' degrees of freedom $\nu_t(x_F|x_S)$ and the marginal probability distribution of ``slow'' degrees of freedom $\sigma_t(x_S)$,
\[
\mu_t(x_S,x_F) = \nu_t (x_F | x_S) \, \sigma_{\epsilon t} (x_S) ,
\]
where we have also assumed that the appropriate timescale of the ``slow'' degrees of freedom is $\epsilon t$, i.e. the marginal distribution of ``slow'' degrees of freedom $\sigma$ is function in $\epsilon t$.
The time evolution equation is then given by
\begin{multline*}
\partial_t \nu_t (x_F|x_S) \, \sigma_{\epsilon t}(x_S) + \epsilon \, \nu_t(x_F|x_S) \, \left. \partial_s \sigma_s(x_S) \right|_{s=\epsilon t} = \\
= \sigma_{\epsilon t}(x_S) \, {\mathcal L}^*_F[\nu_t](x_F|x_S) + \epsilon \, {\mathcal L}^*_S\left[\nu_t \, \sigma_{\epsilon t}\right](x_S,x_F) ,
\end{multline*}
from which we can obtain a time evolution equation for marginal distribution of the ``slow'' degrees of freedom $\sigma(x_S)$ by integration over the ``fast'' degrees of freedom
\begin{equation}
\epsilon \left. \partial_s \sigma_s (x_S) \right|_{s = \epsilon t} = \epsilon \int {\mathrm d} \Gamma(x_F) \; {\mathcal L}^*_S\left[ \nu_t \, \sigma_{\epsilon t} \right](x_S,x_F) ,
\label{equ:time_evolution_slow}
\end{equation}
where we also assumed that the normalization condition \eqref{equ:normalization_condition} is valid also for the forward Kolmogorov generator responsible for time evolution of ``fast'' degrees of freedom ${\mathcal L}^*_F$.
Notice that we have assumed that the dependence of the forward Kolmogorov generator for ``slow'' degrees freedom ${\mathcal L}^*_S$ on fast degrees of freedom $x_F$ lies only in the interaction potential,
hence the time evolution equation \eqref{equ:time_evolution_slow} depends on the averaged potential $\langle V \rangle_{\nu} (x_S)$.
Similarly we can obtain the time evolution equation for marginal distribution on ``fast'' degrees of freedom $\widetilde{\mu}(x_F)$
\[
\partial_t \widetilde{\mu}_t(x_F)
= \left\langle {\mathcal L}^*_F[\nu_t](x_F|x_S) \right\rangle_{\sigma_{\epsilon t}} + \epsilon \int {\mathrm d} \Gamma(x_S) \; {\mathcal L}^*_S\left[ \nu_t \, \sigma_{\epsilon t} \right](x_S,x_F) ,
\]
where we can see two contributions, the first term is the time evolution of the ``fast'' degrees of freedom, where the ``slow'' degrees of freedom occurs only as parameters.
The second term is correction corresponding to the time evolution of ``slow'' degrees of freedom.
If we assume that the feedback of the action of ``slow'' degrees of freedom is small, i.e. the interaction potential in \eqref{equ:time_evolution_slow} is at least of first order in $\epsilon$,
then up to the zeroth order in epsilon the time evolution of marginal distribution of ``slow'' degrees of freedom is independent of the distribution of ``fast'' degrees of freedom,
i.e. we have autonomous dynamics for ``slow'' degrees of freedom in the leading order
\[
\left. \partial_s \sigma_s (x_S) \right|_{s = \epsilon t} = {\mathcal L}^*_S\left[ \sigma_s \right](x_S) + \err{\epsilon} .
\]
Hence the time evolution of the marginal distribution on fast degrees of freedom can also be considered autonomous,
in the manner that we consider the ``slow'' degrees of freedom as time dependent parameters of the dynamics.
If the dynamics of the ``slow'' degrees of freedom is deterministic
\[
\sigma_t (x_S) = \delta(x_S - x_S(t)) \equiv \delta_{x_S(t)}(x_S)
\]
then the time evolution equation of the marginal distribution of ``fast'' degrees of freedom simplifies to
\[
\partial_t \widetilde{\mu}_t(x_F)
= \left\langle {\mathcal L}^*_F[\nu_t](x_F|x_S) \right\rangle_{\sigma_{\epsilon t}} + \err[2]{\epsilon},
\]
where we have also used that in case the ${\mathcal L}^*_S$ is independent of $x_F$ in the leading order,
hence it has to obey the normalization condition \eqref{equ:normalization_condition} as being the forward Kolmogorov generator for autonomous time evolution up to the same order
and so forth the first order term on the right hand side is zero.
The left hand side can be rewritten as
\begin{multline*}
\partial_t \widetilde{\mu}_t(x_F)
= \partial_t \int {\mathrm d} \Gamma(x_S) \nu_t(x_F|x_S) \, \delta(x_S - x_S(\epsilon t)) = \\
= \int {\mathrm d} \Gamma(x_S) \left[ \partial_t \nu_t(x_F|x_S) \, \delta(x_S - x_S(\epsilon t)) + \nu_t(x_F|x_S) \, \partial_t \delta(x_S - x_S(\epsilon t)) \right] = \\
= \partial_t \nu_t(x_F|x_S(\epsilon t)) + \epsilon \, \dot{x}_S(\epsilon t) \cdot \left. \nabla_{x_S} \nu_t(x_F|x_S) \right|_{x_S = x_S(\epsilon t)} ,
\end{multline*}
which provides a more precise formulation of our claim in introduction.
We have seen that up to the zeroth order there is autonomous dynamics for ``slow'' degrees of freedom under the assumption that the time evolution of the ``slow'' degrees of freedom depends on the ``fast'' ones only through interaction potential and the feedback is weak, i.e. of order $\epsilon$.
We have also seen that under the same assumption the time evolution of ``fast'' degrees of freedom depends only on ``slow'' degrees of freedom parametrically.
Moreover if we also assume that the dynamics of ``slow'' degrees of freedom is deterministic,
we obtain the quasistatic expansion of the conditional distribution of ``fast'' degrees of freedom on ``slow'' ones up to the first order, as in chapter \ref{chapter:non-equilibrium_thermodynamics} section \ref{sec:quasistatic_processes_noneq},
where we can identify the ``slow'' degrees of freedom with external parameters, whose time dependence is in this given implicitly be evolution equation \eqref{equ:time_evolution_slow}.
\section{Autonomous dynamics for slow degrees of freedom}
\label{sec:dynamics_for_slow}
In this section we describe the time-evolution of slow degrees of freedom on their appropriate timescale.
Although Berglund and Gentz \cite{Berglund2006} provided a rigorous and thorough analysis of slow-fast coupling in case the dynamics is deterministic,
the extension of these ideas to stochastic dynamics isn't treated in such a rigorous manner,
see for example \cite{Elimination_of_fast_stochastic_variables-Thomas}.
We provide an alternative more exact derivation of the effective Markovian dynamics for ``slow'' degrees of freedom.
Our method enables us to systematically quantify the corrections due to the slow-fast coupling,
and also to estimate their propagation in time.
We start by using Dyson-like expansion to provide zeroth order contribution of the effective Kolmogorov generator for slow degrees of freedom, see subsection \ref{ssec:slow-fast_zeroth_order}.
Furthermore by expanding the Dyson-like series in the parameter characterizing the separation of timescales $\epsilon$ we show that the effective dynamics is Markovian up to the first order with effective generator being \eqref{equ:effective_time_evolution_slow}.
At the end of the section we illustrate the results on examples.
In order to simplify the investigation we consider a specific form of $\epsilon$ dependency, namely
\[
{\mathcal L}^*_\epsilon = {\mathcal L}^*_S + \frac{1}{\epsilon} {\mathcal L}^*_F,
\]
where ${\mathcal L}^*_S$ is the evolution of slow degrees of freedom given the fast ones fixed and ${\mathcal L}^*_F$ is the opposite, i.e. the evolution of fast degrees of freedom given the slow ones.
This particular decomposition corresponds to independent dynamics for fast and slow degrees of freedom apart,
hence the interaction between the fast and slow degrees of freedom is provided only by mutual dependence of some parameters, usually in the form of interaction potential.
Also the $\epsilon$ dependency corresponds to the situation when the fast degrees of freedom are more active,
e.g. in case of jump processes with continuous time we scale the symmetric part of jump rates or in case of diffusion we scale the \emph{mobility}\index{mobility}.
We further assume that there exists a unique stationary distribution on fast degrees of freedom $\rho_F(x_S,x_F)$ for arbitrary yet fixed configuration of slow degrees of freedom $x_S$
\[
{\mathcal L}^*_F [ \rho_F ] (x_S,x_F) = 0 ,
\]
and that the evolution up to the infinite time of fast degrees of freedom
is equivalent to the projection ${\mathcal P}^*_0$ of the fast degrees of freedom to the steady state $\rho_F$
\begin{equation}
\lim_{t \rightarrow \infty} {\mathrm e}^{ t {\mathcal L}^*_F } [ \mu ] = {\mathcal P}^*_0 [ \mu ] = \rho_F(x_S,x_F) \int {\mathrm d} y \; \mu(x_S,y) .
\label{equ:projection}
\end{equation}
Our starting point is the formal solution \eqref{equ:full_solution}.
To find an appropriate timescale at which there exists an autonomous dynamics for slow degrees of freedom,
we slice the total time interval $t$ to time steps of length $\Delta t$
\[
{\mathrm e}^{ t \left( {\mathcal L}^*_S + \frac{1}{\epsilon} {\mathcal L}^*_F \right) } [\mu_0] = \left( {\mathrm e}^{ \Delta t \left( {\mathcal L}^*_S + \frac{1}{\epsilon} {\mathcal L}^*_F \right) } \right)^{\frac{t}{\Delta t}} [\mu_0]
\]
and we choose the time step $\Delta t \ll \tau_S $ in a suitable way that we can simultaneously expand in both $\Delta t$ and $\epsilon$.
To separate the contributions of fast and slow degrees of freedom we use the Dyson-like expansion
\begin{equation}
{\mathrm e}^{\Delta t \left( {\mathcal L}^*_S + \frac{1}{\epsilon} {\mathcal L}^*_F \right) }
= {\mathrm e}^{ \frac{\Delta t}{\epsilon} {\mathcal L}^*_F }
+ \int\limits_0^{\Delta t} {\mathrm d} s \; {\mathrm e}^{ \frac{ \left( \Delta t - s \right) }{\epsilon} {\mathcal L}^*_F} {\mathcal L}^*_S {\mathrm e}^{ s \left( {\mathcal L}^*_S + \frac{1}{\epsilon} {\mathcal L}^*_F \right) }
\label{equ:dyson_expansion_noniterated}
\end{equation}
recursively.
So for the evolution over the single time step $\Delta t$ we obtain a series
\begin{multline}
{\mathrm e}^{\Delta t \left( {\mathcal L}^*_S + \frac{1}{\epsilon} {\mathcal L}^*_F \right) }
= {\mathrm e}^{ \frac{\Delta t}{\epsilon} {\mathcal L}^*_F } + \\
+ \sum\limits_{n=1}^\infty \quad \idotsint\limits_{\Delta t \ge t_1 \ge \dots \ge t_n \ge t_{n+1} \equiv 0 } {\mathrm d} t_1 \dots {\mathrm d} t_n \;
{\mathrm e}^{ \frac{ \left( \Delta t - t_1 \right) }{\epsilon} {\mathcal L}^*_F} \prod\limits_{k=1}^n \left[ {\mathcal L}^*_S {\mathrm e}^{ \left( t_k - t_{k+1} \right) \frac{1}{\epsilon} {\mathcal L}^*_F } \right],
\label{equ:dyson_expansion}
\end{multline}
where exponential terms have only the evolution of the fast dynamics as the argument.
\subsection{Infinite timescale separation}
\label{ssec:slow-fast_zeroth_order}
In order to describe the time evolution of slow degrees of freedom even in the zeroth order by effective Markovian dynamics, we need to observe the system on an appropriate timescale.
Which corresponds to the $\Delta t$ being much larger than the characteristic time of fast degrees of freedom $\Delta t \gg \tau_F$
and doesn't scale with fast degrees of freedom, i.e.
\begin{equation}
\lim_{\epsilon \to 0^+} \frac{\epsilon}{\Delta t} = 0 .
\label{equ:lower_bound}
\end{equation}
Obtaining the zeroth order contribution is quite straightforward, we take the limit $\epsilon \rightarrow 0^+$ of the Dyson-like expansion \eqref{equ:dyson_expansion}
\begin{equation}
\lim_{\epsilon \rightarrow 0} {\mathrm e}^{\Delta t \left( {\mathcal L}^*_S + \frac{1}{\epsilon} {\mathcal L}^*_F \right) }
= {\mathcal P}^*_0 + \sum\limits_{n=1}^\infty \frac{\left(\Delta t\right)^n}{n!} {\mathcal P}^*_0 \left[ {\mathcal L}^*_S {\mathcal P}^*_0 \right]^n
= {\mathrm e}^{\Delta t {\mathcal P}^*_0 {\mathcal L}^*_S {\mathcal P}^*_0} {\mathcal P}^*_0 ,
\label{equ:zeroth_order_limit_slow}
\end{equation}
where we have used the asymptotic of the time evolution on the fast degrees of freedom \eqref{equ:projection} along with the assumption \eqref{equ:lower_bound}.
By introducing the marginal distribution of the slow degrees of freedom $\nu_t$
\[
\nu_t (x_S) = \int {\mathrm d} x_F \; \mu_t(x_S,x_F)
\]
we can indeed show that the time evolution is Markovian.
We show that the marginal distribution after the time step $\Delta t$ depends only on the marginal distribution in the initial time
using the limit \eqref{equ:zeroth_order_limit_slow}
\[
\nu_{t+\Delta t} (x_S) = \int {\mathrm d} x_F \; {\mathrm e}^{\Delta t {\mathcal P}^*_0 {\mathcal L}^*_S {\mathcal P}^*_0} {\mathcal P}^*_0 \left[ \mu_t \right] \left( x_S , x_F \right)
= {\mathrm e}^{\Delta t \left\langle {\mathcal L}^*_S \right\rangle_{\rho_F} } \left[ \nu_t \right] \left( x_S \right) .
\]
And hence by taking the limit $\Delta t \rightarrow 0^+$ we obtain a Markovian time evolution equation
\begin{equation}
\partial_t \nu_t (x_S) = \left\langle {\mathcal L}^*_S \right\rangle_{\rho_F} \left[ \nu_t \right] (x_S) .
\label{equ:zeroth_order_slow}
\end{equation}
One can see that the leading order describes the situation where the fast degrees of freedom can already be considered relaxed to the steady state
given by fixed values of the external parameters but also by the fixed values slow degrees of freedom.
However the assumption \eqref{equ:lower_bound} alone is not sufficient to obtain higher order contributions, where we would likely to see some feedback from the fast degrees of freedom.
\subsection{First-order correction}
In order to find corrections up to higher order of epsilon let us assume that the generator of fast degrees of freedom has a spectral decomposition
\begin{equation}
{\mathcal L}^*_F = - \sum_{i>0} \lambda_i P_i^*,
\label{equ:spectral_projection}
\end{equation}
where eigenvalues has positive real part $\Re \lambda_i > 0$.
Similarly the time evolution of the fast degrees of freedom alone can also be written as the decomposition to the same eigenstates
\[
{\mathrm e}^{\frac{s}{\epsilon} {\mathcal L}^*_F} = {\mathcal P}^*_0 + \sum_{i>0} {\mathrm e}^{- \frac{s \lambda_i}{\epsilon}} {\mathcal P}^*_i.
\]
Using this decomposition in the Dyson-like series \eqref{equ:dyson_expansion} we can again collect all the zeroth order terms in $\epsilon$,
hence all the remaining terms are at least of order $\epsilon$.
While we don't expect an existence of an effective Markovian dynamics behind the first order corrections in $\epsilon$ we include only those.
Finally we can neglect all the terms containing the exponential term $ {\mathrm e}^{ - \frac{\Delta t}{\epsilon} \lambda_i }$, which are much smaller because of the condition \eqref{equ:lower_bound}.
All these considerations yield to
\begin{multline}
{\mathrm e}^{\Delta t \left( {\mathcal L}^*_S + \frac{1}{\epsilon} {\mathcal L}^*_F \right) }
= {\mathrm e}^{ \Delta t \, {\mathcal P}^*_0 {\mathcal L}^*_S {\mathcal P}^*_0} {\mathcal P}^*_0
- \epsilon \sum\limits_{n=0}^{\infty} \frac{\left( \Delta t \right)^n}{n!} \sum\limits_{k=0}^{n+1} \left( {\mathcal P}^*_0 {\mathcal L}^*_S \right)^k \frac{1}{{\mathcal L}^*_F} \left( {\mathcal L}^*_S {\mathcal P}^*_0 \right)^{n+1-k} + \\
+ \err{ {\mathrm e}^{- \frac{\Delta t}{\epsilon} \min\limits_{i>0} \lambda_i}, \epsilon^2} ,
\label{equ:dyson_first_order}
\end{multline}
where $1/{\mathcal L}^*_F$ denotes the pseudoinverse\index{pseudoinverse}
\begin{equation}
\frac{1}{{\mathcal L}^*_F} = \int\limits_0^\infty {\mathrm d} s \; \left[ {\mathcal P}^*_0 - {\mathrm e}^{s {\mathcal L}^*_F} \right] = - \sum\limits_{i>0} \frac{1}{\lambda_i} {\mathcal P}^*_i ,
\label{equ:pseudoinverse_fast}
\end{equation}
introduced in chapter \ref{chapter:non-equilibrium_thermodynamics} as \eqref{def:forward_pseudoinverse},
which emerges here as the consequence of the asymptotic behaviour of the time evolution in the memory kernel
\begin{multline*}
\int\limits_0^t {\mathrm d} s \; {\mathrm e}^{\frac{s}{\epsilon} {\mathcal L}^*_F} = t {\mathcal P}^*_0 + \epsilon \int\limits_0^{\frac{t}{\epsilon}} {\mathrm d} u \; \left( {\mathrm e}^{u {\mathcal L}^*_F} - {\mathcal P}^*_0 \right) \longrightarrow \\
\longrightarrow t {\mathcal P}^*_0 + \epsilon \int\limits_0^\infty {\mathrm d} u \; \left( {\mathrm e}^{u {\mathcal L}^*_F} - {\mathcal P}^*_0 \right) + \err[2]{\epsilon}= t {\mathcal P}^*_0 - \epsilon \frac{1}{{\mathcal L}^*_F} + \err[2]{\epsilon}
\end{multline*}
By rearranging all the terms in \eqref{equ:dyson_first_order} we obtain more compact expression for the expansion of the Dyson-like series up to first order in $\epsilon$
\begin{multline}
{\mathrm e}^{\Delta t \left( {\mathcal L}^*_S + \frac{1}{\epsilon} {\mathcal L}^*_F \right) }
= \exp \left[ \Delta t \, {\mathcal P}^*_0 \left( {\mathcal L}^*_S - \epsilon {\mathcal L}^*_S \frac{1}{{\mathcal L}^*_F} {\mathcal L}^*_S \right) {\mathcal P}^*_0 \right] {\mathcal P}^*_0 - \\
- \epsilon \left[ \frac{1}{{\mathcal L}^*_F} {\mathcal L}^*_S {\mathcal P}^*_0 {\mathrm e}^{ \Delta t \, {\mathcal P}^*_0 {\mathcal L}^*_S {\mathcal P}^*_0} + {\mathrm e}^{ \Delta t \, {\mathcal P}^*_0 {\mathcal L}^*_S {\mathcal P}^*_0} {\mathcal P}^*_0 {\mathcal L}^*_S \frac{1}{{\mathcal L}^*_F} \right]
+ \err{ {\mathrm e}^{- \frac{\Delta t}{\epsilon} \min\limits_{i>0} \lambda_i}, \epsilon^2}.
\label{equ:dyson_first_order_reordered}
\end{multline}
One can see that up to this point we haven't assume anything about the total probability density.
The basic idea of the separation of timescale is that in every time instance the system can be considered to be nearby the state with fast degrees of freedom sampled from a stationary distribution
\begin{equation}
\mu_t (x_S,x_F ) = \rho_F(x_S,x_F) \nu_t (x_S) + \Delta \mu_t (x_S,x_F),
\label{equ:prob_density_structure}
\end{equation}
where $\Delta \mu_t$ is expected to be small
and by $\nu_t$ we denote the marginal distribution of the slow degrees of freedom hence
\[
\int {\mathrm d} x_F \; \Delta \mu_s (x_S,x_F) = 0 .
\]
Furthermore we assume that the correction is order $\epsilon$.
To check if the proposed structure of the total probability density is conserved during the time evolution,
we evolve the total probability density in the form \eqref{equ:prob_density_structure} using \eqref{equ:dyson_first_order_reordered} and expanding it only up to the first order
\begin{multline*}
\mu_{t+\Delta t}
={\mathrm e}^{\Delta t \left( {\mathcal L}^*_S + \frac{1}{\epsilon} {\mathcal L}^*_F \right) } \left[ \mu_t \right]
= \rho_F \exp \left\{ \Delta t \, \left\langle \left( {\mathcal L}^*_S - \epsilon {\mathcal L}^*_S \frac{1}{{\mathcal L}^*_F} {\mathcal L}^*_S \right) \right\rangle_{\rho_F} \right\} \left[ \nu_t \right] - \\
- \epsilon \frac{1}{{\mathcal L}^*_F} {\mathcal L}^*_S \left[ \rho_F \exp \left\{ \Delta t \, \left\langle {\mathcal L}^*_S \right\rangle_{\rho_F} \right\} \left[ \nu_t \right] \right]
+ \err{ {\mathrm e}^{- \frac{\Delta t}{\epsilon} \min\limits_{i>0} \lambda_i}, \epsilon^2}.
\end{multline*}
We can see that the structure \eqref{equ:prob_density_structure} is really preserved by having
\[
\nu_{t+\Delta t} (x_S) = \exp \left\{ \Delta t \, \left\langle \left( {\mathcal L}^*_S - \epsilon {\mathcal L}^*_S \frac{1}{{\mathcal L}^*_F} {\mathcal L}^*_S \right) \right\rangle_{\rho_F} \right\} \left[ \nu_t \right]
\]
and
\begin{equation}
\Delta \mu_{t+\Delta t} = - \epsilon \frac{1}{{\mathcal L}^*_F} {\mathcal L}^*_S \left[ \rho_F \exp \left\{ \Delta t \, \left\langle {\mathcal L}^*_S \right\rangle_{\rho_F} \right\} \left[ \nu_t \right] \right],
\label{equ:difference_exact_vs_markovian}
\end{equation}
where one can see that the correction is of the first order in $\epsilon$ and also does not depend on the previous correction $\Delta \mu_t$ at that specific order.
The assumption on the correction being small is equivalent to be in a state close to steady on the fast degrees of freedom and also that the magnitude of the correction is bounded.
Next we prove that the correction remain small in this sense during the time evolution.
To prove this proposition we need to assume that the dynamics is converging to its steady state $\rho$ on the arbitrary large timescale
\begin{equation}
\left\| \left( {\mathrm e}^{t \left( {\mathcal L}^*_S + \frac{1}{\epsilon} {\mathcal L}^*_F \right) } - {\mathcal P}^* \right) [ \mu ] \right\|_1 \le \left\| \mu \right\|_1 ,
\label{equ:contraction}
\end{equation}
where $\|\cdot\|_1$ is the $L^1$ norm, ${\mathcal P}^*$ is the projection to the full-system steady state
\[
{\mathcal P}^* = \rho \iint {\mathrm d} x_S \, {\mathrm d} x_F .
\]
Then the correction after a single time-step
\[
\left\| {\mathrm e}^{\Delta t \left( {\mathcal L}^*_S + \frac{1}{\epsilon} {\mathcal L}^*_F \right) } [ \Delta \mu ] \right\|_1
\le \left\| {\mathrm e}^{\frac{\Delta t}{\epsilon} {\mathcal L}^*_F} [ \Delta \mu ] \right\|_1
+ \left\| \int\limits_0^{\Delta t} {\mathrm d} s \; {\mathrm e}^{\left(\Delta t - s \right) \left( {\mathcal L}^*_S + \frac{1}{\epsilon} {\mathcal L}^*_F \right) } {\mathcal L}^*_S {\mathrm e}^{\frac{s}{\epsilon} {\mathcal L}^*_F } [ \Delta \mu ] \right\|_1,
\]
where we've already used the Dyson-like expansion \eqref{equ:dyson_expansion_noniterated} and the triangle inequality.
Using the spectral decomposition for evolution of the fast degrees of freedom~\eqref{equ:spectral_projection} and estimating the integral by its norm we obtain
\[
\left\| {\mathrm e}^{\Delta t \left( {\mathcal L}^*_S + \frac{1}{\epsilon} {\mathcal L}^*_F \right) } [ \Delta \mu ] \right\|_1
\le \left\| \sum\limits_{i>0} {\mathrm e}^{- \frac{\Delta t}{\epsilon} \lambda_i} {\mathcal P}^*_i [ \Delta \mu ] \right\|_1
+ \int\limits_0^{\Delta t} {\mathrm d} s \; \left\| {\mathrm e}^{\left(\Delta t - s \right) \left( {\mathcal L}^*_S + \frac{1}{\epsilon} {\mathcal L}^*_F \right) } {\mathcal L}^*_S {\mathrm e}^{\frac{s}{\epsilon} {\mathcal L}^*_F } [ \Delta \mu ] \right\|_1,
\]
using the assumption \eqref{equ:contraction} together with the fact that the evolution on slow degrees of freedom conserves the probability $\int {\mathrm d} x_S \; {\mathcal L}^*_S$ we get
\[
\left\| {\mathrm e}^{\Delta t \left( {\mathcal L}^*_S + \frac{1}{\epsilon} {\mathcal L}^*_F \right) } [ \Delta \mu ] \right\|_1
\le {\mathrm e}^{- \frac{\Delta t}{\epsilon} \lambda_g } \left\| \Delta \mu \right\|_1
+ \int\limits_0^{\Delta t} {\mathrm d} s \; \left\| {\mathcal L}^*_S {\mathrm e}^{\frac{s}{\epsilon} {\mathcal L}^*_F } [ \Delta \mu ] \right\|_1,
\]
where we also estimated the first term in terms of spectral gap
\[
\lambda_g = \min_{i > 0 } \Re \lambda_i .
\]
Using the supremal norm of ${\mathcal L}^*_S$ and the estimate on the exponential we have
\[
\left\| {\mathrm e}^{\Delta t \left( {\mathcal L}^*_S + \frac{1}{\epsilon} {\mathcal L}^*_F \right) } [ \Delta \mu ] \right\|_1
\le {\mathrm e}^{- \frac{\Delta t}{\epsilon} \lambda_g } \left\| \Delta \mu \right\|_1
+ \left\| {\mathcal L}^*_S \right\|_\infty \int\limits_0^{\Delta t} {\mathrm d} s \; {\mathrm e}^{- \frac{s}{\epsilon} \lambda_g } \left\| \Delta \mu \right\|_1,
\]
and by integration and using the fact $\frac{\epsilon}{\lambda_g}$ is of order $\tau_F$ we obtain
\begin{equation}
\left\| {\mathrm e}^{\Delta t \left( {\mathcal L}^*_S + \frac{1}{\epsilon} {\mathcal L}^*_F \right) } [ \Delta \mu ] \right\|_1
\le {\mathrm e}^{- \frac{\Delta t}{\tau_F} } \left( 1 - \tau_F \left\| {\mathcal L}^*_S \right\|_\infty \right) \left\| \Delta \mu \right\|_1
+ \tau_F \left\| {\mathcal L}^*_S \right\|_\infty \left\| \Delta \mu \right\|_1 .
\label{equ:difference_time_evolution}
\end{equation}
From this we can see that the assumption $\Delta t \gg \tau_F$ makes the first term negligible and the second term is of order of $\epsilon$,
as the consequence of the norm $\left\| {\mathcal L}^*_S \right\|_\infty$ being of order $1/\tau_S$ and also $\tau_F/\tau_S$ being of order $\epsilon$,
so after one iteration the correction is already of order epsilon.
Putting all the results together,
we have verified that at the timescale much longer than characteristic time of the fast degrees of freedom $\Delta t \gg \tau_F$ we have an autonomous dynamics for slow degrees of freedom,
which for small time steps $\tau_F \ll \Delta t \ll \tau_S$ is effectively governed by
\begin{equation}
\partial_t \nu_t \approx \frac{ \nu_{t+\Delta t} - \nu_t }{\Delta t} = \left\langle {\mathcal L}^*_S - \epsilon {\mathcal L}^*_S \frac{1}{{\mathcal L}^*_F} {\mathcal L}^*_S \right\rangle_{\rho_F} [ \nu_t ] .
\label{equ:effective_time_evolution_slow}
\end{equation}
The correction to the full probability density is of order $\epsilon$ and is determined only by the distribution on the slow degrees of freedom.
Up to the first order, if the evolution on slow degrees of freedom is smooth, the memory term
\[
\Delta \mu_t = - \epsilon \frac{1}{{\mathcal L}^*_F} {\mathcal L}^*_S \left[ \rho_F \lim_{s \rightarrow t^-} \nu_s \right]
\]
is not present.
Furthermore we can see that the correction diminishes with time.
\subsection{Example I: Diffusive particles entangled by harmonic potential}
To demonstrate our theory we introduce a simple model of two diffusive particles coupled by harmonic interaction potential with the fast particle trapped in an optical trap.
Both slow $(R,P)$ and fast $(x,p)$ particles has mass $m$ and diffuse through the same environment with inverse temperature $\beta$ and friction $\gamma$.
The only coupling between them is by interaction harmonic potential $\frac{1}{2} \omega^2 (x-R)^2$, moreover the fast variable is trapped by the optical trap with effective potential $\frac{1}{2} k (x-X_0)^2$, where $X_0$ is the center of optical trap and $k$ is the strength of the trap.
The time evolution is then introduced by Kolmogorov generators for the fast and slow particles
\begin{align*}
{\mathcal L}^*_F [ \mu ]
&= - \frac{p}{m} \partial_x \mu + \left[ \omega^2 (x-R) + k (x-X_0) \right] \partial_p \mu + \frac{\gamma}{m} \partial_p \left( p \mu \right) + \frac{\gamma}{\beta} \partial_p^2 \mu, \\
{\mathcal L}^*_S [ \mu ] &= - \frac{P}{m} \partial_R \mu + \omega^2 (R-x) \partial_P \mu + \frac{\gamma}{m} \partial_P \left( P \mu \right) + \frac{\gamma}{\beta} \partial_P^2 \mu.
\end{align*}
One can easily find the stationary distribution of the fast particle conditioned on the slow one fixed
\[
\rho_F(x,p) = \frac{1}{Z_R} \exp \left[ - \beta \left( \frac{p^2}{2m} + \frac{1}{2} \omega^2 (x-R)^2 + \frac{1}{2} k \left(x-X_0\right)^2 \right) \right],
\]
where $Z_R$ is the partition function
\begin{multline*}
Z_R = \iint {\mathrm d} x \, {\mathrm d} p \; \exp \left[ - \beta \left( \frac{p^2}{2m} + \frac{1}{2} \omega^2 (x-R)^2 + \frac{1}{2} k \left(x-X_0\right)^2 \right) \right] = \\
= \frac{2 \pi}{\beta} \sqrt{\frac{m}{\omega^2+k}} \exp \left[ \frac{1}{2} k \omega^2 (R-X_0)^2 \right].
\end{multline*}
The correction term in the effective generator \eqref{equ:effective_time_evolution_slow} has the form $ {\mathcal L}^*_S (1/{\mathcal L}^*_F) {\mathcal L}^*_S [ \rho_F \nu ]$ so we start by applying ${\mathcal L}^*_S$
\[
{\mathcal L}^*_S [ \rho_F \nu ] = \rho_F {\mathcal L}^*_S [ \nu ] - \frac{\beta \omega^2 P}{m} \left( x - \frac{\omega^2 R + k X_0}{\omega^2 + k} \right) \rho_F \nu.
\]
The computation of pseudoinverse is in general difficult, however at least for this particular example the pseudoinverse can be computed explicitly, see appendix \ref{sec:pseudoinverse_underdamped} for further details, we obtain
\[
\frac{1}{{\mathcal L}^*_F} {\mathcal L}^*_S [ \rho_F \nu ] = - \frac{\omega^2}{\omega^2 + k} \left( \partial_P \nu + \frac{\beta P}{m} \nu \right)
\left[ p - \gamma \left( x - \frac{\omega^2 R + k X_0}{\omega^2 + k} \right) \right] \rho_F .
\]
The last step is to apply the ${\mathcal L}^*_S$ and integrate over fast degrees of freedoms $x$ and $p$
\[
\iint {\mathrm d} x \, {\mathrm d} p \;
{\mathcal L}^*_S \frac{1}{{\mathcal L}^*_F} {\mathcal L}^*_S [ \rho_F \nu ] = - \frac{\omega^4}{\left(\omega^2 + k\right)^2} \left[ \frac{\gamma}{m} \partial_P \left( P \nu \right) + \frac{\gamma}{\beta} \partial_P^2 \nu \right].
\]
Therefore we obtain the effective generator \eqref{equ:effective_time_evolution_slow} for slow degrees of freedom
\[
{\mathcal L}^*_{{\mathrm{eff}}} [\nu] = - \frac{P}{m} \partial_R \nu + \frac{ k \omega^2 }{\omega^2 + k} \left(R - X_0\right) \partial_P \nu + \left( 1 + \epsilon \frac{\omega^4}{\left(\omega^2 + k\right)^2} \right) \left[ \frac{\gamma}{m} \partial_P \left(P \nu \right) + \frac{\gamma}{\beta} \partial_P^2 \nu \right],
\]
where the slow particle is again diffusive in effective potential and the first order correction presents itself as the correction of the friction
\[
\gamma_{{\mathrm{eff}}} = \gamma \left( 1 + \epsilon \frac{\omega^4}{\left(\omega^2 + k \right)^2} \right) .
\]
\subsection{Example II: Overdamped diffusion}
Overdamped regime can be seen as another example of separation of fast and slow degrees freedom.
In this case the fast variable is the momentum $p$, the slow is the position $x$ and the small parameter is $\epsilon = 1/\gamma$.
We will further assume that the diffusion occurs in non-homogeneous environment with spatially dependent inverse temperature $\beta(x)$.
Based on these considerations, we split the generator \eqref{equ:forward_generator_underdamped} for underdamped diffusion into two parts
\begin{align*}
{\mathcal L}^*_S [\mu] &= - \frac{p}{m} \partial_x \mu - F \partial_p \mu, \\
{\mathcal L}^*_F [\mu] &= \frac{1}{m} \partial_p \left( p \mu \right) + \frac{1}{\beta} \partial_p^2 \mu,
\end{align*}
where the second one describes the relaxation dynamics of momentum.
The stationary distribution for the fast degree of freedom conditioned on the slow one is
\[
\rho(p,x)=\frac{1}{Z(x)} {\mathrm e}^{- \beta(x) \frac{p^2}{2m}}.
\]
We will proceed with the same protocol to determine the effective generator as in the previous example.
First we use the stationary distribution of fast degrees of freedom to determine the structure the first term in the correction
\[
{\mathcal L}^*_S [ \nu \rho ] = - \left[ \partial_x \nu + \left( \frac{\partial_x \beta}{2 \beta} - \beta F \right) \nu \right] \frac{p}{m} \rho + \left( \partial_x \beta \, \nu \right) \frac{p^3}{2 m^2} \rho,
\]
upon which we apply the pseudoinverse,
see appendix \ref{sec:pseudoinverse_overdamped} in particular \eqref{equ:pseudoinverse2_const} to \eqref{equ:pseudoinverse2_p3} for further details
\[
\frac{1}{{\mathcal L}^*_F} {\mathcal L}^*_S [ \nu \rho] = - \frac{p^3}{6m} \left( \partial_x \beta \, \nu \right) \rho + \left[ \partial_x \nu - \left( \frac{\partial_x \beta}{2 \beta} + \beta F \right) \nu \right] p \rho
\]
to obtain the total correction
\[
\int {\mathrm d} p \; {\mathcal L}^*_S \frac{1}{{\mathcal L}^*_F} {\mathcal L}^*_S [ \nu \rho ] = \partial_x \left[ \left( F - k_B \partial_x T \right) \nu - \frac{1}{\beta} \partial_x \nu \right].
\]
While the average of the ${\mathcal L}^*_S$ is zero, the only contribution to the effective generator is the first order correction
\[
{\mathcal L}^*_{{\mathrm{eff}}} [\nu] = - \frac{1}{\gamma} \partial_x \left[ \left( F - k_B \partial_x T \right) \nu - k_B T \partial_x \nu \right],
\]
which is in agreement with Smoluchowski equation \eqref{equ:time_evolution_overdamped} for diffusion in inhomogeneous environment \cite{Inhomogeneous_diffusion1-vanKampen,Inhomogeneous_diffusion2-vanKampen}.
In case the force is conservative with potential $V(x)$, the stationary distribution can be easily determined
\[
\nu_{st} (x) = \frac{1}{Z} \beta(x) \exp\left[ - \int\limits {\mathrm d} x \; \beta(x) V'(x) \right],
\]
which corresponds to the zero steady current $j=0$.
In case the boundaries are not placed in infinity but rather restrict the diffusion to the finite interval $[x_0,x_1]$, we obtain another stationary solution this time with a non-zero steady current $j$
\[
\nu_{st} (x) = j \beta(x) \int\limits_{x_0}^x {\mathrm d} y \; \exp\left[ - \int\limits_{y}^x {\mathrm d} z \; \beta(z) V'(z) \right] .
\]
Notice that the steady current $j$ represents here the normalization condition on the probability distribution and hence is fully determined by the potential $V(x)$ and the temperature profile $\beta(x)$.
We can see that the temperature gradient acts here as an additional force pushing the particles to the colder regions.
We can better understand this effect in the underdamped case, where it is caused by the thermalization of the kinetic energy.
As a consequence of the equipartition theorem in the hot spot the average kinetic energy of the particle is larger than in its surroundings and so the typical momentum.
This means that also the pressure is larger there, which induce the flux of the particles out of the hot spot.
While the physical reality does not depend on the level of our description, hence in the overdamped diffusion, which is a time coarse grained description of the underdamped diffusion, such effect has to be present too.
Moreover if we interpret the whole force as the gradient of the total energy
\[
E(x) = V(x) + \frac{1}{2} k_B T(x) ,
\]
where we can see that the kinetic term is represented by the mean value of the kinetic energy according to \emph{the equipartition theorem}.
This means that this is the correct total energy of the particle undergoing the overdamped diffusion, which does not neglect the coarse grained kinetic energy.
However in the homogeneous system the term is in most cases the additional term can be neglected, because it does not depend on position thus only shifts the total energy.
It only manifests itself as the constant contribution to the heat capacity.
\section{Conclusions}
In the first section we have briefly discussed the possibility of the autonomous dynamics for ``fast'' degrees of freedom on their appropriate timescale.
The main result of that particular section is the observation
that the autonomous dynamics on the characteristic timescale of ``fast'' degrees of freedom can be in some cases associated with quasistatic process, where the autonomously time evolved ``slow'' degrees of freedom act as external parameters.
In the second section we have discussed the possibility of Markovian autonomous dynamics on the level of ``slow'' degrees of freedom on their proper timescale.
We have shown that up to the first order in the parameter characterizing the timescale separation the dynamics is Markovian and can be effectively described by the effective forward Kolmogorov generator \eqref{equ:effective_time_evolution_slow},
where the time derivative $\lim_{\Delta t \to 0^+} \Delta f_t/\Delta t$ has to be taken in the sense that $\tau_S \gg \Delta t \gg \tau_F$.
What is new is that we have also obtained an exact expression how to determine the difference between the effective Markovian solution and the actual solution \eqref{equ:difference_exact_vs_markovian}
and an estimate how it behaves under the time evolution \eqref{equ:difference_time_evolution}.
At the end we have illustrated the results on several models,
most notably we have obtained a Smoluchovski equation describing the overdamped diffusion in the inhomogeneous medium as the first order correction.
We can also see that the overdamped diffusion is the limiting case of the underdamped diffusion in case of fast thermalization of the momentum and kinetic energy in comparison with the slow relaxation of positions.
Moreover notice that the Smoluchovski equation is in fact the first order correction.
\section{Underdamped diffusion with harmonic potential}
\label{sec:pseudoinverse_underdamped}
The forward Kolmogorov generator for underdamped diffusion in harmonic potential has the form
\[
{\mathcal L}^* [ \mu ] = - \frac{p}{m} \partial_x \mu + \omega^2 (x-x_0) \partial_p \mu + \frac{\gamma}{m} \partial_p \left( p \mu \right) + \frac{\gamma}{\beta} \partial_p^2 \mu,
\]
where $x$ and $p$ are the degrees corresponding degrees of freedom, $x_0$ is the center of the harmonic potential and $\omega^2$ is its strength.
The friction $\gamma$ and inverse temperature $\beta$ of the environment are constant in the whole volume.
One can easily find the stationary state
\[
\rho = \frac{1}{Z} \exp \left[ - \beta \left( \frac{p^2}{2m} + \frac{1}{2} \omega^2 \left(x-x_0\right)^2 \right) \right],
\]
which also corresponds to the eigenvector to the eigenvalue zero.
The task to find all the others eigenvalues and eigenvectors is hard,
however for the purpose of computing the pseudoinverse of the linear correction by \eqref{equ:forward_pseudoinverse_spectral} it is not necessary to have them all.
The quadratic nature of the generator implies that any application of it does not increase the order of the polynomial in $p$ neither in $x$.
From these considerations we expect the eigenvector to be a linear combination of $x-x_0$ and $p$.
We are looking for the solution of the equation for eigenvectors
\[
{\mathcal L}^*\left[ \left( C (x-x_0) + D p \right) \rho \right] = \lambda \left( C (x-x_0) + D p \right) \rho ,
\]
which can be expressed as a set of algebraic equations
\begin{align*}
\lambda D &= - \left( \frac{C}{m} + \frac{\gamma D}{m} \right), \\
\lambda C &= \omega^2 D .
\end{align*}
There are two independent solutions with eigenvalues
\[
\lambda_{1,2} = \frac{ - \gamma \pm \sqrt{ \gamma^2 - 4 m \omega^2 }}{2 m}
\]
and corresponding coefficients
\begin{align*}
C_{1,2} &= \frac{\omega^2}{\sqrt{\omega^4 + \lambda_{1,2}^2}}, &
D_{1,2} &= \frac{\lambda_{1,2}}{\sqrt{\omega^4 + \lambda_{1,2}^2}}.
\end{align*}
Using these results, we compute the pseudoinverse up to the linear order corrections
\begin{subequations}
\begin{align}
\frac{1}{{\mathcal L}^*} \left[ \rho \right] &= 0, \label{equ:pseudoinverse_const} \\
\frac{1}{{\mathcal L}^*} \left[ (x-x_0) \rho \right] &= - \frac{\gamma}{\omega^2} (x-x_0) \rho + \frac{1}{\omega^2} p \rho, \label{equ:pseudoinverse_x} \\
\frac{1}{{\mathcal L}^*} \left[ p \rho \right] &= - m (x-x_0) \rho. \label{equ:pseudoinverse_p}
\end{align}
\label{equ:pseudoinverse_underdamped}
\end{subequations}
\section{From underdamped to overdamped diffusion}
\label{sec:pseudoinverse_overdamped}
To obtain the overdamped diffusion we assume that the velocity and in our case the momentum $p$ is the fast variable, which is predominantly governed by
\[
{\mathcal L}^*[\mu] = \frac{1}{m} \partial_p \left( p \mu \right) + \frac{1}{\beta(x)} \partial_p^2 \mu .
\]
The stationary distribution of the momentum $p$ with respect to fixed position $x$ is the Maxwell distribution corresponding to the inverse temperature $\beta(x)$
\[
\rho(x,p) = \frac{1}{Z(x)} {\mathrm e}^{- \beta(x) \frac{p^2}{2m} }.
\]
In general it is valid that
\begin{align*}
{\mathcal L}^*[ p^k \rho ] &= \frac{k}{\beta(x)} \partial_p \left( p^{k-1} \rho \right) \\
&= - \frac{k}{m} p^k \rho + \frac{k (k-1)}{\beta(x)} p^{k-2} \rho ,
\end{align*}
which suggests that eigenvectors will be polynomials times stationary density $\rho$.
For our purpose we need the first four, which are listed bellow.
\begin{align*}
{\mathcal L}^*[ \rho ] &= 0 \\
{\mathcal L}^*[p \rho] &= - \frac{1}{m} p \rho \\
{\mathcal L}^*\left[\left( p^2 - \frac{m}{\beta(x)} \right) \rho \right] &= - \frac{2}{m} \left( p^2 - \frac{m}{\beta(x)} \right) \rho \\
{\mathcal L}^*\left[\left( p^3 - \frac{3m}{\beta(x)} p \right) \rho \right] &= - \frac{3}{m} \left( p^3 - \frac{3m}{\beta(x)} p \right) \rho
\end{align*}
Using them we obtain
\begin{subequations}
\begin{align}
\frac{1}{{\mathcal L}^*} \left[ \rho \right] &= 0 , \label{equ:pseudoinverse2_const} \\
\frac{1}{{\mathcal L}^*} \left[ p \rho \right] &= - m p \rho , \label{equ:pseudoinverse2_p} \\
\frac{1}{{\mathcal L}^*} \left[ p^2 \rho \right] &= - \frac{m}{2} \left( p^2 - \frac{m}{\beta(x)} \right) \rho , \label{equ:pseudoinverse2_p2}\\
\frac{1}{{\mathcal L}^*} \left[ p^3 \rho \right] &= - \frac{m}{3} \left( p^3 + \frac{6m}{\beta(x)} p \right) \rho . \label{equ:pseudoinverse2_p3}
\end{align}
\label{equ:pseudoinverse_overdamped}
\end{subequations}
\chapter{Stochastic calculus}
\label{ap:stochastic_calculus}
In this appendix we sketch some of the proofs of statements given in chapter \ref{chapter:models} subsection \ref{ssec:stochastic_calculus}.
We will mostly follow the text \cite{Evans2001}, where one can find an additional details.
\section{Mean value of It\^{o} integral}
\label{sec:mean_value_Ito}
We have stated that the mean value of the Ito stochastic integral is zero \eqref{equ:Ito_zero_mean_value}.
We start the proof with the definition of the integral \eqref{def:Ito_integral}
and distribute the mean value throughout the sum
\[
\left\langle \int\limits_0^T \vec{f}(\vec{W}_t,t) \cdot {\mathrm d} \vec{W}_t \right\rangle_{\mu_0}
= \lim\limits_{N \to \infty} \sum\limits_{i=0}^{N-1}
\left\langle \vec{f}(\vec{W}_{t_i},t_i) \cdot \left( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \right) \right\rangle_{\mu_0} ,
\]
where $\vec{W}_t$ is the Wiener process,
$N$ is the number of points in the partition of the time interval $[0,T]$
and $t_i$ denotes the respective times of the points of the partitioning.
We use the definition of the conditional mean value to extract the vector field $\vec{f}(\vec{W}_{t_i},t_i)$ from the mean value
\begin{multline*}
\left\langle \vec{f}(\vec{W}_{t_i},t_i) \cdot \left( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \right) \right\rangle_{\mu_0} = \\
= \int\limits_{{\mathbb R}^d} {\mathrm d}^d \vec{x} \; \vec{f}(\vec{x},t_i) \cdot \left\langle \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \middle| \vec{W}_{t_i} = \vec{x} \right\rangle_{\mu_0} \, \cprob{ \vec{W}_{t_i} = \vec{x} }{\mu_0} .
\end{multline*}
Now we use the Markov property of the Wiener process,
i.e. the future displacement of the position does not depend on the history up to this time,
to express the conditional expectation by the expectation over all realizations of the Wiener process with the new initial condition $\delta_{\vec{x}}$ at time $t_i$
\[
\left\langle \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \middle| \vec{W}_{t_i} = \vec{x} \right\rangle_{\mu_0}
= \left\langle \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \right\rangle_{\delta_\vec{x}} .
\]
Using the fact that the mean value of the displacement of position along the Wiener process is zero \eqref{equ:Wiener_properties} concludes the statement.
\section{Covariance of It\^{o} integrals}
\label{sec:covariance_Ito}
The second statement connected the covariance of the It\^{o} stochastic integrals \eqref{equ:Ito_covariance} with the time integral of the covariance.
In this case we start by dividing it to three terms
\begin{multline*}
\left\langle \int\limits_0^T \vec{f}(\vec{W}_t,t) \cdot {\mathrm d} \vec{W}_t \int\limits_0^T \vec{g}(\vec{W}_t,t)
\cdot {\mathrm d} \vec{W}_t \right\rangle_{\mu_0} = \\
= \underbrace{ \sum_{i=0}^{N-1} \sum_{j=i+1}^{N-1} \left\langle \left( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \right)
\cdot \vec{f}(\vec{W}_{t_i},t_i) \, \vec{g}(\vec{W}_{t_j},t_j)
\cdot \left( \vec{W}_{t_{j+1}} - \vec{W}_{t_j} \right) \right\rangle_{\mu_0} }_\text{A} + \\
+ \underbrace{ \sum_{i=0}^{N-1} \sum_{j=0}^{i-1} \left\langle \left( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \right)
\cdot \vec{f}(\vec{W}_{t_i},t_i) \, \vec{g}(\vec{W}_{t_j},t_j)
\cdot \left( \vec{W}_{t_{j+1}} - \vec{W}_{t_j} \right) \right\rangle_{\mu_0} }_\text{B} + \\
+ \underbrace{ \sum_{i=0}^{N-1} \left\langle \left( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \right)
\cdot \vec{f}(\vec{W}_{t_i},t_i) \, \vec{g}(\vec{W}_{t_i},t_i)
\cdot \left( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \right) \right\rangle_{\mu_0} }_\text{C}
\end{multline*}
taking the limit $N \to \infty$ afterwards.
By following the same line of thoughts as in the previous case see that parts A and B are zero.
The remaining term C can be rearranged in the similar fashion using the conditional mean value
\begin{multline*}
\left\langle \left( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \right)
\cdot \vec{f}(\vec{W}_{t_i},t_i) \, \vec{g}(\vec{W}_{t_i},t_i)
\cdot \left( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \right) \right\rangle_{\mu_0} = \\
\begin{aligned}
= \int\limits_{{\mathbb R}^d} {\mathrm d}^d \vec{x} \; & \cprob{\vec{W}_{t_i} = \vec{x}}{\mu_0} \times \\
& \times \vec{f}(\vec{x},t_i)
\cdot \left\langle \left( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \right) \left( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \right) \middle| \vec{W}_{t_i} = \vec{x} \right\rangle_{\mu_0}
\cdot \vec{g}(\vec{x},t_i)
\end{aligned}
\end{multline*}
and using the fact that the variance is proportional to the time interval while the mean value is zero \eqref{equ:Wiener_properties} we obtain
\begin{multline*}
\left\langle \left( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \right)
\cdot \vec{f}(\vec{W}_{t_i},t_i) \, \vec{g}(\vec{W}_{t_i},t_i)
\cdot \left( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \right) \right\rangle_{\mu_0} = \\
= \int\limits_{{\mathbb R}^d} {\mathrm d}^d \vec{x} \; \vec{f}(\vec{x},t_i)
\cdot \vec{g}(\vec{x},t_i) \left( t_{i+1} - t_i \right)
\cprob{\vec{W}_{t_i} = \vec{x}}{\mu_0} = \\
= \left\langle \vec{f}(\vec{W}_{t_i},t_i) \cdot \vec{g}(\vec{W}_{t_i},t_i) \right\rangle_{\mu_0} \left( t_{i+1} - t_i \right) ,
\end{multline*}
which concludes the proof.
\section{Riemann sum of square of displacement}
\label{sec:Riemann_sum}
Another statement connected the sum of square of displacements with the length of the time interval.
Here we sketch the proof of the more general version \eqref{equ:pre_Ito_lemma}.
We will show that the variance of the difference between the Riemann sum representing the integral and the sum of square displacement which represents the mean value tends to vanish with $N \to \infty$,
\[
\sigma^2 = \left\langle \left( \sum\limits_{i=0}^{N-1} f(\vec{W}_{t_i},t_i) \left[ \left( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \right)^2 - \left( t_{i+1} - t_i \right) \right] \right)^2 \right\rangle_{\mu_0} .
\]
We start again by dividing the product of the sum into three terms,
where by following the same pattern the terms corresponding to $i \neq j$ are equal to zero,
which leave us with
\[
\sigma^2
= \sum\limits_{i=0}^{N-1} \left\langle f^2(\vec{W}_{t_i},t_i) \left[ \left( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \right)^2 - \left( t_{i+1} - t_i \right) \right]^2 \right\rangle_{\mu_0} ,
\]
which can be rewritten using the conditional probability to
\begin{multline*}
\left\langle f^2(\vec{W}_{t_i},t_i) \left[ \left( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \right)^2 - \left( t_{i+1} - t_i \right) \right]^2 \right\rangle_{\mu_0} = \\
\begin{aligned}
=& \int\limits_{{\mathbb R}^d} {\mathrm d}^d \vec{x} \; \cprob{\vec{W}_{t_i} = \vec{x}}{\mu_0} \, f^2(\vec{x},t_i) \times \\
& \times \left[ \left\langle \left( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \right)^4 - 2 \left( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \right)^2 \left( t_{i+1} - t_i \right) \middle| \vec{W}_{t_i} = \vec{x} \right\rangle_{\mu_0} + \left( t_{i+1} - t_i \right)^2 \right] .
\end{aligned}
\end{multline*}
By using the Gaussian properties of the Wiener process \eqref{equ:Wiener_distribution} we conclude the proof.
\section{It\^{o} lemma}
\label{sec:Ito_lemma}
To prove the It\^{o} lemma \eqref{equ:total_differential_Ito} we need to associate the difference in the result $Y(\vec{W}_T,T) - Y(\vec{W}_0,0)$ with its integral representation.
Lets take an arbitrary partition $t_i$ of the time interval $[0,T]$ then it is valid
\[
Y(\vec{W}_T,T) - Y(\vec{W}_0,0) = \sum\limits_{i=0}^{N-1} \left[ Y(\vec{W}_{t_{i+1}},t_{i+1}) - Y(\vec{W}_{t_i},t_i) \right] .
\]
To obtain terms similar to Riemann sum we expand the particular differences in both arguments to Taylor series we obtain
\begin{multline*}
Y(\vec{W}_T,T) - Y(\vec{W}_0,0)
= \sum\limits_{i=0}^{N-1} \Biggl[ \left. \partial_t Y(\vec{W}_{t_i},t) \right|_{t=t_i} ( t_{i+1} - t_i ) + \\
+ \left. \nabla_\vec{x} Y(\vec{x},t_i) \right|_{\vec{x} = \vec{W}_{t_i}} \cdot ( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} ) + \\
+ \frac{1}{2} ( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} ) \cdot \left. \nabla^2_\vec{x} Y(\vec{x},t_i) \right|_{\vec{x} = \vec{W}_{t_i}} \cdot ( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} )
+ \dots \Biggr] .
\end{multline*}
In the limit $N \to \infty$ one can see the first two terms correspond to Riemann sums for the Riemann integral and It\^{o} integral \eqref{def:Ito_integral}.
The third term in the sum under the same limit $N \to \infty$ according to \eqref{equ:pre_Ito_lemma} converges almost surely to
\[
\lim_{N \to \infty} \sum\limits_{i=0}^{N-1} ( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} ) \cdot \left. \nabla^2_\vec{x} Y(\vec{x},t_i) \right|_{\vec{x} = \vec{W}_{t_i}} \cdot ( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} )
= \int\limits_0^T {\mathrm d} t \; \left. \Delta_\vec{x} Y(\vec{x},t) \right|_{\vec{x} = \vec{W}_t} ,
\]
while other terms in the sum converge almost surely to zero by a similar reasoning as in section \ref{sec:Riemann_sum}.
As an example we will show that the term corresponding to the mixed derivative tends to go to zero by proving that the mean value as well as the variance goes to zero in the limit $N \to \infty$.
To obtain the mean value we distribute the sum at first and then use that the displacement of the Wiener process is independent on the history \eqref{equ:Wiener_combination} and has zero mean value \eqref{equ:Wiener_properties}
\begin{multline*}
\left\langle
\sum\limits_{i=0}^{N-1} ( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} ) \cdot \left. \vec{\nabla}_\vec{x} \partial_t Y(\vec{x},t) \right|_{t=t_i,\vec{x} = \vec{W}_{t_i}} ( t_{i+1} - t_i ) \right\rangle_{\mu_0} = \\
= \sum\limits_{i=0}^{N-1} \left\langle ( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} ) \cdot \left. \vec{\nabla}_\vec{x} \partial_t Y(\vec{x},t) \right|_{t=t_i,\vec{x} = \vec{W}_{t_i}} \right\rangle_{\mu_0} ( t_{i+1} - t_i ) = \\
= \sum\limits_{i=0}^{N-1} ( t_{i+1} - t_i ) \int\limits_{{\mathbb R}^d} {\mathrm d}^d \vec{x} \; \left\langle \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \right\rangle_{\delta_\vec{x}} \cdot \left. \vec{\nabla}_\vec{x} \partial_t Y(\vec{x},t) \right|_{t=t_i} \cprob{\vec{W}_{t_i}=\vec{x}}{\mu_0} = 0.
\end{multline*}
By similar reasons as in section \ref{sec:covariance_Ito} in case of variance the only contributing terms are quadratic,
then we proceed in the similar fashion as for the mean value
\begin{multline*}
\Biggl\langle
\sum\limits_{j=0}^{N-1} ( \vec{W}_{t_{j+1}} - \vec{W}_{t_j} ) \cdot \left. \vec{\nabla}_\vec{x} \partial_t Y(\vec{x},t) \right|_{t=t_j,\vec{x} = \vec{W}_{t_j}} ( t_{j+1} - t_j ) \times \\
\times \sum\limits_{i=0}^{N-1} ( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} ) \cdot \left. \vec{\nabla}_\vec{x} \partial_t Y(\vec{x},t) \right|_{t=t_i,\vec{x} = \vec{W}_{t_i}} ( t_{i+1} - t_i ) \Biggr\rangle_{\mu_0} = \\
= \sum\limits_{i=0}^{N-1} \left\langle \left[ ( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} ) \cdot \left. \vec{\nabla}_\vec{x} \partial_t Y(\vec{x},t) \right|_{t=t_i,\vec{x} = \vec{W}_{t_i}} \right]^2 \right\rangle_{\mu_0} ( t_{i+1} - t_i )^2 = \\
\begin{split}
= \sum\limits_{i=0}^{N-1} ( t_{i+1} - t_i )^2 \int\limits_{{\mathbb R}^d} {\mathrm d}^d \vec{x} \; \left. \vec{\nabla}_\vec{x} \partial_t Y(\vec{x},t) \right|_{t=t_i} \cdot \left\langle \left( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \right) \left( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \right) \right\rangle_{\delta_\vec{x}} \cdot \quad \\
\cdot \left. \vec{\nabla}_\vec{x} \partial_t Y(\vec{x},t) \right|_{t=t_i} \cprob{\vec{W}_{t_i}=\vec{x}}{\mu_0} = \end{split} \\
= \sum\limits_{i=0}^{N-1} ( t_{i+1} - t_i )^3 \left\langle \left\| \left. \vec{\nabla}_\vec{x} \partial_t Y(\vec{x},t) \right|_{t=t_i,\vec{x}=\vec{W}_{t_i}} \right\|^2 \right\rangle_{\mu_0} ,
\end{multline*}
where $\|\cdot\|$ denotes the Euclid norm.
If we assume that the time step is uniformly bounded
\[
\forall i : \qquad t_{i+1} - t_i < \frac{C T}{N}
\]
and that the mean value in the Riemann sum is on the interval $[0,T]$ bounded
then we can make estimate for variance
\begin{multline*}
\Biggl\langle
\sum\limits_{j=0}^{N-1} ( \vec{W}_{t_{j+1}} - \vec{W}_{t_j} ) \cdot \left. \vec{\nabla}_\vec{x} \partial_t Y(\vec{x},t) \right|_{t=t_j,\vec{x} = \vec{W}_{t_j}} ( t_{j+1} - t_j ) \times \\
\times \sum\limits_{i=0}^{N-1} ( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} ) \cdot \left. \vec{\nabla}_\vec{x} \partial_t Y(\vec{x},t) \right|_{t=t_i,\vec{x} = \vec{W}_{t_i}} ( t_{i+1} - t_i ) \Biggr\rangle_{\mu_0} < \\
< \frac{C^3 T^3}{N^2} \max_{t_i} \left\langle \left\| \left. \vec{\nabla}_\vec{x} \partial_t Y(\vec{x},t) \right|_{t=t_i,\vec{x}=\vec{W}_{t_i}} \right\|^2 \right\rangle_{\mu_0} \xrightarrow{N \to \infty} 0 ,
\end{multline*}
which concludes the statement that higher order contributions vanish.
Putting all these partial results altogether we obtain that the difference can be represented as
\[
Y(\vec{W}_T,T) - Y(\vec{W}_0,0) = \int\limits_0^T {\mathrm d} t \; \left[ \partial_t Y(\vec{W}_{t_i},t) + \frac{1}{2} \left. \Delta_\vec{x} Y(\vec{x},t) \right|_{\vec{x}=\vec{W}_t} \right] + \int\limits_0^T {\mathrm d} \vec{W}_t \cdot \left. \nabla_\vec{x} Y(\vec{x},t) \right|_{\vec{x}=\vec{W}_t} ,
\]
which together with \eqref{def:stochastic_differential_equation} concludes the proof.
\section{Comparison of It\^{o} and Stratonovich integral}
\label{sec:Ito_vs_Stratonovich}
We have provided the relation between the Stratonovich and It\^{o} integral \eqref{equ:relation_Ito_Stratonovich}.
To prove it we start with the expansion of the argument of the Stratonovich integral to the Taylor series up to the first order in the position in a similar fashion to the previous section, while higher orders are almost surely zero
\begin{multline*}
\int\limits_0^T \vec{f}(\vec{W}_t,t) \circ {\mathrm d} \vec{W}_t = \lim\limits_{N \to \infty} \sum\limits_{i=0}^{N-1} \vec{f}(\vec{W}_{\tau_i},\tau_i) \cdot \left[\vec{W}_{t_{i+1}} - \vec{W}_{t_i}\right] \\
= \lim\limits_{N \to \infty} \sum\limits_{i=0}^{N-1} \left[ \vec{f}(\vec{W}_{t_i},t_i) + \frac{1}{2} \left( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \right) \cdot \left. \vec{\nabla}_\vec{x} \vec{f}(\vec{x},t_i) \right|_{\vec{x}=\vec{W}_{t_i}} \right] \cdot \left[\vec{W}_{t_{i+1}} - \vec{W}_{t_i}\right] .
\end{multline*}
Using again the \eqref{equ:pre_Ito_lemma} gives us the desired relation.
\chapter{Stochastic equilibrium thermodynamics}
\label{chapter:equilibrium_stat_phys}
There are several different approaches to classical equilibrium physics, for example standard textbook by Callen \cite{Callen1985} follows the argument of equal probability of microstates
while for example Balian \cite{Balian1982} prefers information theory approach.
In this chapter we will rehearse an alternative way how to build an equilibrium statistical physics more suitable for mesoscopic stochastic systems.
Our starting point will be the concepts of conservation of the total energy on the microscopic level and the local and global detailed balance.
We will use these principles to obtain the first law of thermodynamics for stochastic time evolution and to derive Crooks fluctuation relation as a weak formulation of the second law.
\section{First law of thermodynamics}
\index{first law of thermodynamics}
\label{sec:first_law}
The first law of thermodynamics is the mesoscopic version of the conservation law of total energy.
In our systems we assume that the law of the total energy conservation is in general obeyed by the underlying microscopic dynamics for the total system,
hence the energy of the system can only be altered by the interaction with a thermal bath or by the change of external conditions $\alpha$.
Sekimoto associated \emph{the heat}\index{heat} $Q_{\alpha(t)}(\omega)$ with the energy dissipated to thermal baths and \emph{the work}\index{work} $W_{\alpha(t)}(\omega)$ with the energy change by varying the external conditions on the microscopic level for every possible microscopic path $\omega$ and thus introduced the concept of \emph{stochastic energetics}\index{stochastic energetics} \cite{Sekimoto1997,Sekimoto1998,Sekimoto2007,Seifert2008}
\begin{equation}
E_{\alpha(T)}(x_T) - E_{\alpha(0)}(x_0) = Q^{\text{sys}}_{\alpha(t)}(\omega) + W^{\text{sys}}_{\alpha(t)}(\omega),
\label{def:first_law_path}
\end{equation}
where $E_\alpha(x)$ is the energy of the system in configuration $x$ under external conditions $\alpha$
and where $Q^{\text{sys}}_{\alpha(t)}(\omega)$ is the heat dissipated to the system and $W^{\text{sys}}_{\alpha(t)}(\omega)$ is the work done on the system along the path $\omega$ from time $0$ to time $T$ with external conditions driven by protocol $\alpha(t)$.
The heat and work strongly depend on the actual model,
however usually the heat is the energy associated with the change of the configuration,
which for example in case of Markov jump processes is given by
\[
Q^{\text{sys}}_{\alpha(t)}(\omega) = \sum_{t_i} \left[ E_{\alpha(t_i)}(x_{t_i}) - E_{\alpha(t_i)}(x^-_{t_i}) \right],
\]
while the work is associated with the action of the external agent via \emph{the work}
\begin{equation}
W^{\text{sys}}_{\alpha(t)}(\omega) = \int {\mathrm d} t \; \dot\alpha(t) \cdot \left. \nabla_\alpha E_\alpha(x_t) \right|_{\alpha=\alpha(t)} .
\label{def:work_external_parameters}
\end{equation}
For more examples see subsections \ref{ssec:work_heat_jump} and \ref{ssec:work_heat_diffusion}.
We can take advantage of the Markov property of time evolution and define \emph{the local power}\index{local power} $w_{\alpha(t)}(x)$ as an work done on the system per unit time if the system is starting from the initial configuration $x$ and freely evolve for timespan $s$
\begin{equation}
w_{\alpha(t)}(x) = \lim_{s \to 0^+} \frac{1}{s} \left\langle W^{\text{sys}}_{\alpha(t)}\left(\omega^{(0,s]}\right) \right\rangle^{\alpha(t)}_{\delta_x}
\label{def:local_power}
\end{equation}
and similarly we define \emph{the local heat production}\index{local heat production}
\begin{equation}
q_{\alpha(t)}(x) = \lim_{s \to 0^+} \frac{1}{s} \left\langle Q^{\text{sys}}_{\alpha(t)}\left(\omega^{(0,s]}\right) \right\rangle^{\alpha(t)}_{\delta_x} .
\label{def:local_heat_production}
\end{equation}
Notice that in the most general case both of these quantities are functionals on the protocol of external parameters $\alpha(t)$,
however because the mean value of the work and the heat are smooth in most cases with respect to protocol $\alpha(t)$ it depends only on the actual value of external parameters $\alpha$ and its first time-derivative $\dot\alpha$.
A direct consequence is, that \emph{the mean total work} ${\mathcal W}_{\mu_0}(\alpha(t))$ and analogously \emph{the mean total heat} is equivalent to integrated mean value of the local power with respect to actual state $\mu_t$ at time $t$
when starting the time evolution from the initial state $\mu_0$
\begin{align}
{\mathcal W}_{\mu_0}(\alpha(t)) &= \left\langle W^{\text{sys}}_{\alpha(t)}(\omega) \right\rangle^{\alpha(t)}_{\mu_0} = \int\limits_0^T {\mathrm d} t \; \left\langle w_{\alpha(t)} \right\rangle_{\mu_t} , \label{equ:total_local_work_relation} \\
{\mathcal Q}_{\mu_0}(\alpha(t)) &= \left\langle Q^{\text{sys}}_{\alpha(t)}(\omega) \right\rangle^{\alpha(t)}_{\mu_0} = \int\limits_0^T {\mathrm d} t \; \left\langle q_{\alpha(t)} \right\rangle_{\mu_t}. \label{equ:total_local_heat_relation}
\end{align}
While we have relation between the change of the energy and work and heat for each possible realization of the path $\omega$ \eqref{def:first_law_path},
the similar relation has to be valid also on the level of mean values over all possible realizations of path $\omega$
\begin{equation}
{\mathcal U}_{\mu_T}(\alpha(T)) - {\mathcal U}_{\mu_0}(\alpha(0)) = {\mathcal W}_{\mu_0}(\alpha(t)) + {\mathcal Q}_{\mu_0}(\alpha(t)) ,
\label{equ:first_law_eq}
\end{equation}
where we have defined \emph{the internal energy}\index{internal energy} ${\mathcal U}_{\mu}(\alpha) = \left\langle E_\alpha \right\rangle_\mu $,
and which we associate with \emph{the first law of thermodynamics}\index{first law of thermodynamics}, as we will see later.
From here we can also conclude that the mean local power and mean local heat production sums up to the change of the internal energy
\begin{equation}
\partial_t {\mathcal U}_{\mu_t}(\alpha(t)) = \left\langle w_{\alpha(t)} + q_{\alpha(t)} \right\rangle_{\mu_t} .
\label{equ:first_law_rates}
\end{equation}
An alternative approach to local power and local heat production can be by starting from the equation \eqref{equ:first_law_rates} and applying \eqref{equ:time_derivative_observable}
\begin{equation}
\partial_t \left\langle E_{\alpha(t)} \right\rangle_\mu = \left\langle \partial_t E_{\alpha(t)} + {\mathcal L}_{\alpha(t)} \left[ E_{\alpha(t)} \right] \right\rangle_\mu ,
\label{equ:time_evolution_energy}
\end{equation}
where we can identify the local power as the part connected with the change of external parameters
\begin{equation}
w_{\alpha(t)}(x) = \partial_t E_{\alpha(t)}(x) = \dot\alpha(t) \cdot \left. \nabla_\alpha E_\alpha(x) \right|_{\alpha(t)}
\label{equ:local_power}
\end{equation}
which can be also obtained from the definition \eqref{def:local_power} with the work defined by \eqref{def:work_external_parameters},
while the local heat production is the part corresponding to the change of the configuration of the system
\begin{equation}
q_{\alpha(t)}(x) = {\mathcal L}_{\alpha(t)}\left[ E_{\alpha(t)} \right] (x) ,
\label{equ:local_heat_production}
\end{equation}
which is associated with the energy transfer to thermal bath.
\subsection{Work and heat in the Markov jump processes}
\label{ssec:work_heat_jump}
We now provide examples of definitions of the heat and work and their local production in two major classes of models starting with Markov jump processes,
where the system is in contact with single thermal bath and where the only time dependence lies in the protocol of the external parameters $\alpha(t)$.
The simplest quantity to define in jump processes is the heat.
The heat\index{heat!Markov jump processes} is the energy exchange associated with the change of configuration of the system, which occurs at jump points, hence we have
\[
Q_{\alpha(t)}(\omega) = \sum_i \left[ E_{\alpha(t_i)}(x_{t_i}) - E_{\alpha(t_i)}(x_{t^-_i}) \right] ,
\]
where we sum over all jump times $t_i$ and where $x_{t^-_i}$ again denotes the configuration just before the jump.
To obtain the local heat production\index{local heat production!Markov jump processes} we can use either definition \eqref{def:local_heat_production} or rather use the equation \eqref{equ:local_heat_production} along with the backward Kolmogorov generator for jump processes \eqref{equ:backward_Kolmogorov_gen_jump}
\[
q_\alpha(x) = \sum_y \rate[\alpha]{x}{y} \left[ E(y) - E(x) \right] .
\]
Notice that in case of Markov jump processes the local heat production depends only on the actual values of external parameters $\alpha$ but not on the velocity of its change.
Under such conditions the work\index{work!Markov jump processes} is solely associated with the change of the energy due to the change of external parameters $\alpha(t)$, hence we obtain
\[
W_{\alpha(t)}(\omega) = \sum_i \left[ E_{\alpha(t_{i+1})}(x_{t_i}) - E_{\alpha(t_i)}(x_{t_i}) \right] ,
\]
where we sum over all time intervals $[t_i,t_{i+1}]$ in the path $\omega$ where the configuration $x_{t_i}$ holds.
The local power\index{local power!Markov jump processes} in this case is directly obtained from the equation \eqref{equ:local_power}
\[
w_{\alpha(t)}(x) = \dot\alpha(t) \cdot \left. \nabla_\alpha E_\alpha(x) \right|_{\alpha=\alpha(t)}.
\]
\subsection{Work and heat in diffusion}
\label{ssec:work_heat_diffusion}
In diffusion we can determine the local power and the local heat production by decomposing the time evolution of the mean total energy \eqref{equ:time_evolution_energy} to the part which depends on the change of external parameters $\alpha$ \eqref{equ:local_power} and the part which depends on the change of state \eqref{equ:local_heat_production}.
However we can also obtain them directly from the pure mechanical considerations.
In this section we will show that these two approaches are consistent.
\subsubsection{Underdamped diffusion}
In the underdamped diffusion the total force acting on the system \eqref{equ:momentum_underdamped} is
\[
\vec{F}_\alpha(\vec{q},\vec{p}) = - \vec{\nabla}_\vec{q} U_\alpha (\vec{q}) - \frac{\gamma_\alpha(\vec{q})}{m} \vec{p} + \sqrt{\frac{2 \gamma_\alpha(\vec{q})}{\beta}} \frac{{\mathrm d} \vec{W}_t}{{\mathrm d} t} ,
\]
where the first term is the interaction with the potential $U_\alpha(\vec{q})$,
the second term corresponds to the friction
and the third term is the thermal random force acting on the particle,
where by ${\mathrm d} \vec{W}_t / {\mathrm d} t$ we denote \emph{the white noise}\index{white noise}.
We we will handle the white noise formally in this subsection as the formal time derivative of the Wiener process, although it does not exists,
because the Wiener process is in fact nowhere differentiable \cite{Evans2001}.
The last two terms make together what we will later call heat.
In case there is no change in external parameters it is know from the classical mechanics, that the work done by action of the total forces corresponds to the change of the kinetic energy
\[
\Delta T = \int \vec{F}_\alpha(\vec{q}_t,\vec{p}_t) \circ {\mathrm d} \vec{q}_t
= \int \left[ - \vec{\nabla}_\vec{q} U_\alpha (\vec{q}) - \frac{\gamma_\alpha(\vec{q})}{m} \vec{p} + \sqrt{\frac{2 \gamma_\alpha(\vec{q})}{\beta}} \frac{{\mathrm d} \vec{W}_t}{{\mathrm d} t} \right] \circ {\mathrm d} \vec{q}_t ,
\]
where $\circ$ denotes the Stratonovich integral \eqref{def:Stratonovich_integral}.
The choice of the Stratonovich stochastic integral here is due to the antisymmetry with respect to the time inversion \eqref{equ:antisymmetry_Stratonovich}.
The first term in the work corresponds to the infinitesimal change of the potential ${\mathrm d} U_\alpha(\vec{q})$
in the Stratonovich sense \eqref{equ:total_differential_Stratonovich}
and hence can be integrated out and we obtain that the change of the total energy is given by thermal forces, i.e. friction and random force
\begin{equation}
\Delta \left[ \frac{\vec{p}_t^2}{2m} + U_{\alpha}(\vec{q}) \right] = \int \left[ - \frac{\gamma_\alpha(\vec{q}_t)}{m} \vec{p}_t + \sqrt{\frac{2 \gamma_\alpha(\vec{q})}{\beta}} \frac{{\mathrm d} \vec{W}_t}{{\mathrm d} t} \right] \circ {\mathrm d} \vec{q}_t ,
\label{equ:energy_balance_underdamped}
\end{equation}
where we have written the kinetic energy explicitly.
From thermodynamics we know that the change of total energy in case there is no action of external forces, i.e. no work done by the change of external parameters $\alpha$, corresponds to the heat, hence we define the heat in the underdamped diffusion as
\[
Q(\omega) = \int \left[ - \frac{\gamma_\alpha(\vec{q}_t)}{m} \vec{p}_t + \sqrt{\frac{2 \gamma_\alpha(\vec{q})}{\beta}} \frac{{\mathrm d} \vec{W}_t}{{\mathrm d} t} \right] \circ {\mathrm d} \vec{q}_t .
\]
To obtain the local heat production we need to express the mean total heat as the time integral.
In this case we will treat each of the two terms separately.
The first term does not explicitly depend on the Wiener process, it depends on it only implicitly through the momentum $\vec{p}_t$,
hence by applying the time evolution equation for the position \eqref{def:position_underdamped} we obtain an ordinary time integral
\[
\left\langle Q^{(1)}_{\alpha(t)}(\omega) \right\rangle_{\mu_0}
= - \int \left\langle \left[ \frac{\gamma_{\alpha(t)}(\vec{q}_t)}{m} \vec{p}_t \right] \cdot \frac{\vec{p}_t}{m} \right\rangle_{\mu_t} \; {\mathrm d} t ,
\]
from where we can directly obtain the local of that particular part of the total force
\[
q^{(1)}_\alpha(\vec{q},\vec{p}) = - \frac{\gamma_\alpha(\vec{q})}{m^2} \vec{p}^2 .
\]
In the mean value for the random force
\begin{multline*}
\left\langle Q^{(2)}_{\alpha(t)}(\omega) \right\rangle_{\mu_0}
= \left\langle \int \sqrt{\frac{2 \gamma_{\alpha(t)}(\vec{q}_t)}{\beta}} \frac{{\mathrm d} \vec{W}_t}{{\mathrm d} t} \circ {\mathrm d}{\vec{q}_t} \right\rangle_{\mu_0} = \\
= \left\langle \int \sqrt{\frac{2 \gamma_{\alpha(t)}(\vec{q}_t)}{\beta}} \frac{{\mathrm d} \vec{q}_t}{{\mathrm d} t} \circ {\mathrm d}{\vec{W}_t} \right\rangle_{\mu_0}
= \left\langle \int \sqrt{\frac{2 \gamma_{\alpha(t)}(\vec{q}_t)}{\beta}} \frac{\vec{p}_t}{m} \circ {\mathrm d}{\vec{W}_t} \right\rangle_{\mu_0}
\end{multline*}
we have obtained the Stratonovich integral of a function of the momentum with respect to the Wiener process.
At first we rewrite the Stratonovich stochastic integral by using the relation between Stratonovich and It\^{o} integral \eqref{equ:relation_Ito_Stratonovich}
\[
\left\langle Q^{(2)}_{\alpha(t)}(\omega) \right\rangle_{\mu_0} =
\left\langle \int \sqrt{\frac{2 \gamma_{\alpha(t)}(\vec{q}_t)}{\beta}} \frac{\vec{p}_t}{m} \cdot {\mathrm d}{\vec{W}_t} \right\rangle_{\mu_0}
+ \left\langle \int \frac{\gamma_{\alpha(t)}(\vec{q}_t)}{\beta} \frac{d}{m} \; {\mathrm d} t \right\rangle_{\mu_0} ,
\]
where $d$ is the dimension of the momentum space $\vec{p} \in {\mathbb R}^d$.
Then we used that the mean value of the It\^{o} integral is zero \eqref{equ:Ito_zero_mean_value}
and we have obtain the local heat production from the random force term as
\[
q^{(2)}_\alpha(\vec{q},\vec{p})
= \frac{d \gamma_\alpha(\vec{q})}{m \beta} .
\]
Putting these two contribution together we obtain the total local power generated by the total local heat produced by the particle
\[
q_\alpha(\vec{q},\vec{p}) = - \frac{2 \gamma_\alpha(\vec{q})}{m} \left[ \frac{\vec{p}^2}{2m} - \frac{d}{2 \beta} \right],
\]
which is proportional to the difference between actual kinetic energy and the mean kinetic energy from the equipartition theorem.
In case there is also the change of the external parameters, there is an additional term in the total energy balance equation \eqref{equ:energy_balance_underdamped},
which corresponds to the action of external forces
\[
\Delta \left[ \frac{\vec{p}_t^2}{2m} + U_{\alpha(t)}(\vec{q}_t) \right]
= \int {\mathrm d} t \; \dot\alpha(t) \cdot \left. \nabla_\alpha U_\alpha(\vec{q}_t) \right|_{\alpha=\alpha(t)} + Q(\omega) ,
\]
and which defines the work \eqref{def:work_external_parameters}.
\subsubsection{Overdamped diffusion}
We again start with the case when the external parameters are kept constant by analysing the energy balance equation.
In overdamped diffusion the total force acting on the system is given by \eqref{def:position_overdamped}
\[
\vec{F}_\alpha(\vec{q}) = - \vec{\nabla}_\vec{q} U_\alpha(\vec{q}) + \frac{1}{\beta} \chi^{-1}(\vec{q}) \cdot \left( \vec{\nabla}_\vec{q} \cdot \chi(\vec{q}) \right) + \sqrt{\frac{2}{\beta \chi(\vec{q})}} \cdot \frac{{\mathrm d} \vec{W}_t}{{\mathrm d} t} ,
\]
where we have used the Einstein relation to represent the expression in terms of mobility matrix $\chi$ (and its inverse $\chi^{-1}$) and the inverse temperature.
In the overdamped diffusion there is no notion of the kinetic energy, hence the energy balance equation leads to
\[
0 = \int \vec{F}_\alpha(\vec{q}_t) \circ \vec{q}_t
= \int \left[ - \vec{\nabla}_\vec{q} U_\alpha(\vec{q}) + \frac{1}{\beta} \chi^{-1}(\vec{q}) \cdot \left( \vec{\nabla}_\vec{q} \cdot \chi(\vec{q}) \right) + \sqrt{\frac{2}{\beta \chi(\vec{q})}} \cdot \frac{{\mathrm d} \vec{W}_t}{{\mathrm d} t} \right] \circ {\mathrm d} \vec{q}_t ,
\]
where the first term can be again as the total differential in the Stratonovich sense \eqref{equ:total_differential_Stratonovich} of the potential $U_\alpha(\vec{q}_t)$
\[
\Delta U_\alpha (\vec{q}_t) = \int \left[ \frac{1}{\beta} \chi^{-1}(\vec{q}) \cdot \left( \vec{\nabla}_\vec{q} \cdot \chi(\vec{q}) \right) + \sqrt{\frac{2}{\beta \chi(\vec{q})}} \cdot \frac{{\mathrm d} \vec{W}_t}{{\mathrm d} t} \right] \circ {\mathrm d} \vec{q}_t ,
\]
where the right hand side represent the heat
\[
Q(\omega) = \int \left[ \frac{1}{\beta} \chi^{-1}(\vec{q}) \cdot \left( \vec{\nabla}_\vec{q} \cdot \chi(\vec{q}) \right) + \sqrt{\frac{2}{\beta \chi(\vec{q})}} \cdot \frac{{\mathrm d} \vec{W}_t}{{\mathrm d} t} \right] \circ {\mathrm d} \vec{q}_t .
\]
To obtain the local heat production we need again to obtain the mean value of the heat in the form of the integral over time.
However in this case it is more convenient to use it equivalence to the energy change.
We start from
\[
\left\langle Q(\omega) \right\rangle_{\mu_0} = \left\langle \int \left. \vec{\nabla}_\vec{q} U_\alpha (\vec{q}) \right|_{\vec{q}=\vec{q}_t} \circ {\mathrm d} \vec{q}_t \right\rangle_{\mu_0},
\]
where we apply the time evolution equation in the Stratonovich form \eqref{def:position_overdamped}
\begin{multline*}
\left\langle Q(\omega) \right\rangle_{\mu_0}
= \int \left\langle \vec{\nabla}_\vec{q} U_\alpha(\vec{q}) \cdot \left[ - \chi_\alpha(\vec{q}) \cdot \vec{\nabla}_\vec{q} U_\alpha(\vec{q}) + \frac{1}{2} \vec{\nabla}_\vec{q} \cdot D_\alpha(\vec{q}) \right] \right\rangle_{\mu_t} \; {\mathrm d} t + \\
+ \left\langle \int \vec{\nabla}_\vec{q} U_\alpha(\vec{q}_t) \cdot \sqrt{2 D_\alpha(\vec{q}_t)} \circ {\mathrm d} \vec{W}_t \right\rangle_{\mu_0} .
\end{multline*}
Now we can apply the relation between It\^{o} and Stratonovich integral \eqref{equ:relation_Ito_Stratonovich}
\begin{multline*}
\left\langle Q(\omega) \right\rangle_{\mu_0}
= \int \left\langle \vec{\nabla}_\vec{q} U_\alpha(\vec{q}) \cdot \left[ - \chi_\alpha(\vec{q}) \cdot \vec{\nabla}_\vec{q} U_\alpha(\vec{q}) + \frac{1}{2} \vec{\nabla}_\vec{q} \cdot D_\alpha(\vec{q}) \right] \right\rangle_{\mu_t} \; {\mathrm d} t + \\
+ \frac{1}{2} \int \left\langle \tr \left[ \sqrt{2 D_\alpha(\vec{q})} \, \vec{\nabla}_\vec{q} \cdot \left( \sqrt{2 D_\alpha(\vec{q})} \cdot \vec{\nabla}_\vec{q} U_\alpha(\vec{q}) \right) \right] \right\rangle_{\mu_t} \; {\mathrm d} t + \\
+ \left\langle \int \vec{\nabla}_\vec{q} U_\alpha(\vec{q}_t) \cdot \sqrt{2 D_\alpha(\vec{q}_t)} \cdot {\mathrm d} \vec{W}_t \right\rangle_{\mu_0} .
\end{multline*}
From where immediately follows
\[
q_\alpha(\vec{q}) = - \vec{\nabla}_\vec{q} U_\alpha(\vec{q}) \cdot \vec{\chi}_\alpha(\vec{q}) \cdot \vec{\nabla}_\vec{q} U_\alpha(\vec{q}) + \vec{\nabla}_{\vec{q}} \cdot \left[ \vec{D}_\alpha(\vec{q}) \cdot \vec{\nabla}_\vec{q} U_\alpha(\vec{q}) \right] .
\]
The reasoning in case when also external parameters $\alpha$ are time dependent is the same as for the underdamped diffusion, hence the work is also defined as
\[
W(\omega) = \int {\mathrm d} t \; \dot\alpha(t) \cdot \left. \nabla_\alpha U_\alpha(\vec{q}_t) \right|_{\alpha=\alpha(t)} .
\]
\subsubsection{Common properties}
We have verified that in both cases when we are in equilibrium, where there are no non-potential forces, the local heat production up to the sign reads
\begin{equation}
{\mathcal L}_\alpha \left[ E_\alpha \right] = q_\alpha .
\label{equ:local_power_potential_underdamped}
\end{equation}
In non-equilibrium there can be in principle non-potential forces acting to the system.
While the mechanical work dissipated by these forces can be considered as heat dissipated to the thermal bath at infinite temperature,
because the stationary distribution is uniform with respect to the non-potential forces only, which corresponds to $\beta \to 0^+$ in Boltzmann distribution,
we rather consider the system attached to the ``work'' reservoir and hence the local power for non-potential forces is given by the mean value of mechanical work.
We can see that work and heat is then distinguished purely by convention,
which in non-equilibrium situation leads to non-uniques of certain quantities, as we will see in the chapter \ref{chapter:non-equilibrium_thermodynamics}.
\section{Global and local detailed balance condition}
\label{sec:detailed_balance}
While investigating the equilibrium of the continuous time Markov jump processes we have introduced the concept of \emph{global detailed balance condition}\index{detailed balance condition!global},
see subsection \ref{ssec:detailed_balance_jump}.
The global detailed balance condition connects probabilities of transition forth and back with the equilibrium occupation probabilities under \emph{constant external conditions}.
In general the statement can be written as
\begin{equation}
\frac{ \dcprob[{(0,T]}][\alpha]{\omega}{X_0 = x_0 } }{ \dcprob[{(0,T]}][\pi \alpha]{\Theta \omega}{X_0 = \pi x_T }} = \frac{{\mathrm d} \rho_{\pi \alpha}(\pi x_T)}{{\mathrm d} \rho_\alpha(x_0)},
\label{def:global_detailed_balance}
\end{equation}
where $\alpha$ denotes the set of \emph{external parameters} and its \emph{fixed values},
$\omega$ denotes the path starting from configuration $x_0$ and ending in configuration $x_T$,
${\mathrm d} \rho_\alpha$ denotes the equilibrium probability measure given external conditions $\alpha$,
$\pi$ is \emph{kinematic inversion}\index{kinematic inversion}, which changes the sign of the momentum, impulse momentum and similar quantities
\begin{equation}
\pi: \quad (\vec{x},\vec{p}) \to (\vec{x},-\vec{p}),
\label{def:kinematic_inversion}
\end{equation}
or in case it is applied to external parameters it changes the sign of the magnetic fields and similar quantities
\[
\pi: \quad (\vec{E},\vec{B}) \to (\vec{E},-\vec{B}) ,
\]
and $\Theta \omega^{[0,T]}$ denotes \emph{time-reversed path}\index{path!time reversed} and is defined as
\begin{equation}
\Theta \omega^{[0,T]} = \left\{ \pi X_{T-t} \middle| X_t \in \omega^{[0,T]} \right\}.
\label{def:time_reversal}
\end{equation}
Notice that the path probability measure ${\mathrm d} \mathbb P$ is a conditional probability with respect to initial configuration and values of external parameters.
If there are no momentum-like degrees of freedom in the system, like the typical case of Markov jump processes, the kinematic inversion reduces to identity, $\pi \equiv {\mathrm {id}}$,
and hence the time reversal simplifies to what was introduced in \eqref{equ:time_reversal_jump}.
Moreover by examining the paths with only one transition in case of continuous time Markov process one can obtain the global detailed balance condition \eqref{equ:global_detailed_balance_jump}
as a result of the definition \eqref{def:global_detailed_balance}.
We also introduced the principle of \emph{local detailed balance}\index{detailed balance condition!local} at the end of subsection \ref{ssec:equlibrium_systems_jump},
which in general states that \emph{any} difference between the probability of transition forth and back can always be associated with \emph{entropy}\index{entropy} production in the environment
\begin{equation}
k_B \ln \frac{ \dcprob[{(0,T]}][\alpha(t)]{\omega}{X_0 = x_0 } }{ \dcprob[{(0,T]}][\Theta \alpha (t)]{\Theta \omega}{X_0 = \pi x_T }} = S^{\text{bath}}_{\alpha(t)}( \omega ),
\label{def:local_detailed_balance}
\end{equation}
where $\alpha(t)$ denotes the protocol describing the changes of external parameters, $\alpha(t) : \quad [0,T] \to \alpha$,
$S^{\text{bath}}_{\alpha(t)}(\omega)$ denotes the entropy production in the environment consisting of thermal and particle baths along the path $\omega$ given the protocol $\alpha(t)$ and $\Theta \omega$ again describes its time reversal \eqref{def:time_reversal}.
The $k_B$ is the \emph{Boltzmann constant}\index{Boltzmann constant} determining the physical dimension of entropy, which we will set to one in the rest of the text, $k_B \equiv 1$.
Although both these principles are consequence of \emph{microscopic time-reversibility}\index{time-reversibility},
we can see the fundamental difference between them in the fact that the local detailed balance condition connects path probabilities with entropy production,
while the global detailed balance condition is a mathematical statement about invariance of the stationary density with respect to time reversal,
which in case the configurations are invariant with respect to kinematic inversion leads to the condition of zero probabilistic currents along each possible transition in equilibrium.
Moreover we consider the local detailed balance to be a more general principle, which is in general valid even out of equilibrium.
\subsubsection{Boltzmann distribution}
\index{Boltzmann distribution}
The Boltzmann distribution is the direct consequence of the simultaneous validity of the global \eqref{def:global_detailed_balance} and local \eqref{def:local_detailed_balance} detailed balance condition.
To simplify the situation we assume that system is attached to single thermal bath at inverse temperature $\beta$ and all force are given by potential $E_\alpha(x)$, which is invariant with respect to kinematic inversion.
Moreover we restrict ourselves to the case when we hold the external parameters fixed $\alpha(t) \equiv \alpha$.
As the first step we can see that the probability distribution of configurations in steady state is given by the entropy production in thermal bath
\[
\frac{ \rho_{\pi \alpha}(\pi x_T) }{\rho_{\alpha}(x_0)} = \exp \left[ S^{\text{bath}}_{\alpha} (\omega ) \right] ,
\]
from where it immediately follows that the entropy production does not depend on the choice of the path $\omega$.
Using also the fact that the entropy production in the thermal bath can be associated with heat transfered to the thermal bath from the system
\[
S^{\text{bath}}_{\alpha}(\omega) = \beta Q^{\text{bath}}_{\alpha}(\omega) = - \beta Q^{\text{sys}}_{\alpha}(\omega) ,
\]
along with the fact that there is no work done while the external parameters are fixed \eqref{equ:local_power} and the fact that the change of the energy in case no work is done on the system can be associated with the heat \eqref{def:first_law_path}, we obtain the Boltzmann distribution
\[
\frac{ \rho_{\pi \alpha}(\pi x_T) }{\rho_{\alpha}(x_0)}
= \exp \left[ S^{\text{sys}}_\alpha(x_0) - S^{\text{sys}}_{\pi \alpha} (\pi x_T) \right]
= \exp \left[ \beta \left( E_\alpha(x_0) - E_\alpha(x_T) \right) \right] .
\]
\subsection{Local detailed balance as a consequence of time-reversibility}
The reason why we can assume the validity of local detailed balance in most non-equilibrium situations is that the local detailed balance condition on mesoscopic scale is a consequence of time-reversibility of the underlying microscopic evolution.
This claim was proved by Maes and Neto\v{c}n\'{y} in \cite{Maes2002} and later refined by Maes in \cite{Maes2003}, whom we will follow in this text.
In order to make arguments as clear as possible, we will restrict ourselves only to systems with classical Hamiltonian dynamics with fixed external conditions as an underlying dynamics,
which simulates a mesoscopic open system in contact with single thermal path.
We also restrict ourselves to the case where each configuration of the system corresponds to the single microstate of the mesoscopic system alone.
Microstates of such system are points in the phase space $\Omega$ upon which the Hamiltonian dynamic acts.
The time-evolution from the time $0$ to time $T$ is represented by automorphism $\varphi_T$ on $\Omega$,
\[
x_T = \varphi_T (x_0),
\]
where $x_0$ is the initial microstate at time $0$ and $x_T$ is the final microstate.
The time evolution has to be a semi-group
\[
\varphi_{T_2} \circ \varphi_{T_1} = \varphi_{T_1+T_2}
\]
and in case of Hamiltonian dynamics it also preserves the \emph{Liouville measure}\index{Liouville measure} $\Gamma$,
\begin{equation}
\Gamma \circ \varphi_T = \Gamma.
\label{equ:sym_of_Liouville_measure_time_evol}
\end{equation}
The microscopic reversibility of Hamiltonian dynamics states that
if we reverse all the velocities of all particles, i.e. we apply kinematic inversion \eqref{def:kinematic_inversion} to the microstate,
then let the system evolve over some period $T$ and then again inverse the velocities, we obtain a state before the time period $T$,
\[
\pi \circ \varphi_T \circ \pi = \varphi_{-T} .
\]
With regard of kinematic inversion one can also observe that the Liouville measure is invariant with respect to kinematic inversion,
\begin{equation}
\Gamma \circ \pi = \Gamma .
\label{equ:sym_of_Liouville_measure_kinematic}
\end{equation}
On the mesoscopic level of description the full information about the system is not accessible, hence we are not able to determine the exact microstate of the system,
rather we represent the state of the system by its configuration on the mesoscopic level.
We associate the configuration of the system with \emph{macrostate}\index{state!macro-} which corresponds to the specific values $a_i$ of macroscopic observables $A_i$, which are accessible.
We denote by $\vec{A}$ the \emph{complete collection of the macroscopic observables} determining the macrostate $\vec{a}=\{\dots,a_i,\dots\}$.
From the microscopic point of view macrostate is set of microstates with the same value of macroscopic observables,
\[
\vec{A}^{-1}(\vec{a}) = \left\{ x \middle| x \in \Omega \land \vec{A} (x) = \vec{a} \right\} ,
\]
Because observables are functions, recall \eqref{def:observable}, each microstate also fully determines the macrostate, which belongs to $\vec{A}(x) = \vec{a}$.
Moreover we assume that macroscopic observables commute with kinematic inversion, $\vec{A} \circ \pi = \pi \circ \vec{A}$ and also $\pi \vec{A}^{-1}(\vec{a}) = \vec{A}^{-1}( \pi \vec{a})$.
Up to now we introduced the underlying microscopic dynamics and macrostates.
Furthermore we also need to introduce entropy to our description.
We define \emph{the entropy}\index{entropy} for each macrostate $\vec{a}$ by variational principle as an extremal Shannon entropy \eqref{def:Shannon_entropy}
\begin{equation}
S(\vec{a}) = \sup_{\mu} \left[ - \int\limits_\Omega {\mathrm d} \Gamma(x) \; \mu(x) \ln \mu(x) \right] ,
\label{def:Shannon_entropy_extr}
\end{equation}
where $\mu$ denotes the probabilistic density of microstates representing the given macrostate $\vec{a}$.
The supremum is taken over all probability densities $\mu$ under the conditions
\begin{align*}
\int\limits_\Omega {\mathrm d} \Gamma(x) \; \mu(x) &= 1 , & \supp \mu &= \vec{A}^{-1}(\vec{a}),
\end{align*}
where the first condition is the normalization and the second ensures that all microstates represents the given macrostate and hence
\[
\int\limits_\Omega {\mathrm d} \Gamma(x) \; \mu(x) \, \vec{A}(x) = \vec{a} .
\]
We can see, that the supremal probability distribution $\mu^*_{\vec{a}}$
\[
S(\vec{a}) = - \int\limits_\Omega {\mathrm d} \Gamma(x) \; \mu^*_{\vec{a}}(x) \ln \mu^*_{\vec{a}}(x)
\]
is uniform in $\vec{A}^{-1}(\vec{a})$ and hence the entropy of macrostate corresponds to the Boltzmann entropy
\[
S(\vec{a}) = \ln \Gamma\left( \vec{A}^{-1}(\vec{a}) \right) .
\]
Notice that the supremal probability distribution obeys the symmetry with respect to kinematic inversion
\begin{equation}
\mu^*_{\pi \vec{a}}(\pi x) = \mu^*_{\vec{a}}(x) .
\label{equ:sym_supremal_dist}
\end{equation}
For detailed discussion about the definition of entropy see \cite{Maes2003}.
Although we are not able to determine the exact microstate of the system we still can be asking
what is the probability that we will observe a \emph{sequence of macrostates} while starting the evolution from a given \emph{macrostate} $\vec{a}_0$.
The probability of observing the sequence of macrostates $\omega=\{\vec{a}_t | t \in [0,T] \}$ when starting from macrostate $\vec{a}_0$ is given by
\[
\cprob[{(0,T]}]{\omega}{X_0=\vec{a}_0}
= \frac{ \Gamma\left(\bigcap\limits_{t \in [0,T]} \varphi_{-t} \left( \vec{A}^{-1}(\vec{a}_t) \right) \right) }{ \Gamma\left( \vec{A}^{-1}(\vec{a}_0) \right) },
\]
where the denominator is the probability of observing macrostate $\vec{a}_0$ and the numerator corresponds to probability of observing the sequence of macrostates $\omega$.
If we compare it with the probability of observing the reversed sequence of macroscopic observables $\Theta \omega$ starting from macrostate $\pi \vec{a}_T$ we obtain the local detailed balance
\[
\ln \frac{\cprob[{(0,T]}]{\omega}{X_0=\vec{a}_0}}{\cprob[{(0,T]}]{\Theta \omega}{X_0 = \pi \vec{a}_T}} = \ln \Gamma\left( \vec{A}^{-1}(\vec{a}_T) \right) - \ln \Gamma\left( \vec{A}^{-1}(\vec{a}_0) \right) = S^{\text{bath}}(\vec{a}_T) - S^{\text{bath}}(\vec{a}_0),
\]
where we have used the invariance of the Liouville measure with respect to time evolution \eqref{equ:sym_of_Liouville_measure_time_evol} and kinematic inversion \eqref{equ:sym_of_Liouville_measure_kinematic},
and also the symmetry \eqref{equ:sym_supremal_dist}.
Notice that the change of Boltzmann entropy corresponds to the entropy change in the thermal bath, because each configuration corresponds to the single microstate of the observed system,
but also to multiple microstates realizing the bath with inverse temperature $\beta$.
Also notice that in this particular case we have not changed the external parameters of the system, so in this case we are in the transient regime.
\subsection{Local detailed balance for Markov jump processes}
In the case of Markov jump process another line of thoughts can be followed.
In subsection \ref{ssec:detailed_balance_jump} we have derived the local detailed balance condition for systems in equilibrium as a consequence of global detailed balance condition and Boltzmann distribution of microstates.
In general case we assume that each jump, i.e. the change of microstate is associated with exchange of either particle or heat with a bath.
Furthermore we assume the baths in equilibrium and independent.
In order to simplify the discussion we restrict ourselves only to the case of the system attached to multiple thermal baths.
The independence of thermal baths ensures that the probability that two transitions coincide is zero and hence each transition in the system is associated with \emph{single} bath.
The independence of thermal baths also ensures that the transition rates associated with the given thermal bath are independent while the other thermal baths are connected to the system or not.
Also along the single transition the system changes it's energy which is transfered to the corresponding reservoir and is directly associated with the production of the entropy there.
Following these considerations we can conclude the local detailed balance condition is still even outside of equilibrium
\[
\ln \frac{\rate[i]{x}{y}}{\rate[i]{y}{x}} = \beta_i \left( E(x) - E(y) \right) = S^{\text{bath}}(y) - S^{\text{bath}}(x) ,
\]
where $i$ denotes the thermal bath with inverse temperature $\beta_i$.
\section{Jarzynski and Crooks equalities}
In last twenty years the principle of local detailed balance condition or more generally the microscopical reversibility proved to be essential in developing fluctuation symmetries.
First fluctuation symmetry was observed in numerical simulations of deterministic Hamiltonian evolution by Evans et al. \cite{Evans1993},
who also together with Searles provided first proof for such systems \cite{Evans1994}.
In 1997 Jarzynski proved another fluctuation theorem the Jarzynski equality \cite{Jarzynski1997,Crooks1998},
which later Crooks showed as a special case of more general Crooks fluctuation relation \cite{Crooks1999,Crooks2000}.
The first experimental verification of these relations were provided by Collin et al. in \cite{Ritort_ver_of_Crooks}.
For review on the fluctuation theorems see \cite{Seifert2012}.
\emph{The Crooks fluctuation relation}\index{fluctuation relation!Crooks} states that the probability of the increase of the \emph{total entropy} $\Sigma$ along the given protocol of external parameters $\alpha(t)$
is exponentially larger then probability of the decrease of the total entropy $-\Sigma$ along the time reversed protocol $\Theta \alpha(t)$,
\begin{equation}
\frac{\prob[{[0,T]}][\alpha(t)]{S^{\text{tot}}(\omega)=\Sigma}}{\prob[{[0,T]}][\Theta \alpha(t)]{S^{\text{tot}}(\omega)=-\Sigma}} = {\mathrm e}^\Sigma ,
\label{equ:Crooks_relation}
\end{equation}
where \emph{the total entropy production}\index{entropy!total} consists of the change of the entropy of the system and of the entropy production in thermal bath
\begin{equation}
S^{\text{tot}}_{\alpha(t)}(\omega) = \ln \mu_{\alpha(0)}(x_0) - \ln \mu_{\pi \alpha(T)}(\pi x_T) + S^{\text{bath}}_{\alpha(t)}(\omega) ,
\label{def:total_entropy}
\end{equation}
where $\mu_{\alpha(0)}(x_0)$ is the initial state under the external conditions $\alpha(0)$
and $\mu_{\pi \alpha(T)}(x_T)$ is the initial state for time-reversed process.
Notice that the total entropy production is also antisymmetric with respect to kinematic inversion
\begin{equation}
S^{\text{tot}}_{\Theta \alpha(t)}(\Theta \omega) = - S^{\text{tot}}_{\alpha(t)}(\omega) .
\label{equ:antisymmetry_total_entropy}
\end{equation}
As was stated before the proof of the Crooks fluctuation relation is centered around the local detailed balance condition \eqref{def:local_detailed_balance}.
The probability to observe the total entropy production $\Sigma$ is given by
\begin{align*}
\prob[{[0,T]}][\alpha(t)]{ S^{\text{tot}}(\omega) = \Sigma }
&= \int \dprob[{[0,T]}][\alpha(t)]{\omega} \; \delta\left( S^{\text{tot}}_{\alpha(t)}(\omega) - \Sigma \right) \\
&= \int {\mathrm d} \mu_{\alpha(0)} (x_0) \, \dcprob[{(0,T]}][\alpha(t)]{\omega}{X_0 = x_0} \; \delta\left( S^{\text{tot}}_{\alpha(t)}(\omega) - \Sigma \right) ,
\end{align*}
where $\delta(\cdot)$ denotes the Dirac's delta function and where we have already extracted the initial condition from path probability.
Inserting now the local detail balance condition and switching between the initial and final condition we obtain
\begin{multline*}
\prob[{[0,T]}][\alpha(t)]{ S^{\text{tot}}(\omega) = \Sigma }
= \int {\mathrm d} \mu_{\pi \alpha(T)}(\pi x_T) \, \dcprob[{(0,T]}][\Theta \alpha(t)]{\Theta \omega}{X_0 = \pi x_T} \; \times \\
\times {\mathrm e}^{S_{\alpha(t)}(\omega)} \frac{\mu_{\alpha(0)}(x_0)}{\mu_{\pi \alpha(T)}(\pi x_T)} \, \delta\left( S^{\text{tot}}_{\alpha(t)}(\omega) - \Sigma \right) .
\end{multline*}
Using the definition of total entropy \eqref{def:total_entropy} and its antisymmetry \eqref{equ:antisymmetry_total_entropy} along with the definition of Dirac's delta function concludes the proof.
\paragraph{Transient fluctuation relation}
The direct application of Crooks relation yields to the well known \emph{transient fluctuation symmetry}\index{fluctuation relation!transient-} for entropy
\begin{multline}
\left\langle {\mathrm e}^{-S^{\text{tot}}_{\alpha(t)}(\omega)} \right\rangle_{\mu_{\alpha(0)}}^{\alpha(t)}
= \int {\mathrm d} \Sigma \; \left\langle \delta\left(S^{\text{tot}}_{\alpha(t)}(\omega)-\Sigma\right) \right\rangle_{\mu_{\alpha(0)}}^{\alpha(t)} \, {\mathrm e}^{-\Sigma} = \\
= \int {\mathrm d} \Sigma \; {\mathrm e}^{-\Sigma} \, \prob[{[0,T]}][\alpha(t)]{S^{\text{tot}}(\omega)=\Sigma} = \\
= \int {\mathrm d} \Sigma \; \prob[{[0,T]}][\Theta \alpha(t)]{S^{\text{tot}}(\omega)=-\Sigma}
= 1 ,
\label{equ:transient_fluctuation_relation}
\end{multline}
where by $\langle \cdot \rangle_{\mu_{\alpha(0)}}^{\alpha(t)}$ we denote the mean value over all paths $\omega$ with the initial condition given by $\mu_{\alpha(0)}(x)$ and with the protocol of external parameters $\alpha(t)$.
\paragraph{Jarzynski equality}
When the system is connected to single thermal bath at constant inverse temperature $\beta$ and when the protocol $\alpha(t)$ connects two \emph{equilibrium} states then the total entropy reads
\[
S^{\text{tot}}_{\alpha(t)}(\omega) = \beta \left[ F_{\alpha(0)} - F_{\pi \alpha(T)} - E_{\alpha(0)}(x_0) + E_{\pi \alpha(T)}(\pi x_T) \right] + S^{\text{bath}}_{\alpha(t)}(\omega),
\]
where we have used the fact that the probability distribution of configurations in equilibrium in contact with single thermal bath is given by Boltzmann distribution
\[
\rho_\alpha(x) = \frac{1}{Z_\alpha} {\mathrm e}^{- \beta E_\alpha(x)} = {\mathrm e}^{\beta \left( {\mathcal F}_\alpha - E_\alpha(x) \right)},
\]
where ${\mathcal F}_\alpha$ denotes the equilibrium free energy and $E_\alpha(x)$ is the potential describing the energy landscape of the system, see e.g. \cite{Callen1985}.
We also assume that the potential is symmetric with respect to kinematic inversion $E_{\pi \alpha(T)}(\pi x_T) = E_{\alpha(T)}(x_T)$ as well as the free energy is $ {\mathcal F}_{\pi \alpha(T)} = {\mathcal F}_{\alpha(T)} $.
Using also the fact that the entropy production in the thermal bath can be associated with heat transfered to the thermal bath from the system
\[
S^{\text{bath}}_{\alpha(t)}(\omega) = \beta Q^{\text{bath}}_{\alpha(t)}(\omega) = - \beta Q^{\text{sys}}_{\alpha(t)}(\omega)
\]
and the law of conservation of energy \eqref{def:first_law_path},
we obtain \emph{the Jarzynski equality}\index{fluctuation relation!Jarzynski equality}
\begin{equation}
\left\langle {\mathrm e}^{- \beta W^{\text{sys}}_{\alpha(t)}(\omega)} \right\rangle_{\mu_{\alpha(0)}}^{\alpha(t)} = {\mathrm e}^{- \beta \left( {\mathcal F}_{\alpha(T)} - {\mathcal F}_{\alpha(0)} \right)}
\label{equ:Jarzynski_equality}
\end{equation}
as a special case of the transient fluctuation relation \eqref{equ:transient_fluctuation_relation}.
\section{Second law inequality}
\index{second law of thermodynamics}
One of the most prominent results of the nineteenth century was the formulation of the second law of thermodynamics.
One of those formulations of the second law states, that along any equilibrium process the total entropy is non-decreasing quantity, \cite{Callen1985}.
Within our framework we cannot say such a definite statement without any other assumptions.
In most general case the Crooks relation \eqref{equ:Crooks_relation} states that the probability of the increase of the total entropy is exponentially larger than the probability of decrease by the same amount
and hence can be considered as \emph{a weak formulation of the second law of thermodynamics}\index{second law of thermodynamics!weak formulation}.
One of the consequences of the Crooks relation is the transient fluctuation relation \eqref{equ:transient_fluctuation_relation},
from which we can prove that also the mean total entropy production is non-decreasing quantity by using the \emph{Jensen's inequality}\index{Jensen inequality}
\begin{equation}
\exp \left[ - \left\langle S^{\text{tot}}_{\alpha(t)}(\omega) \right\rangle_{\mu_{\alpha(0)}}^{\alpha(t)} \right] \le \left\langle {\mathrm e}^{ - S^{\text{tot}}_{\alpha(t)}(\omega)} \right\rangle_{\mu_{\alpha(0)}}^{\alpha(t)} = 1
\quad \Longrightarrow \quad
\left\langle S^{\text{tot}}_{\alpha(t)}(\omega) \right\rangle_{\mu_{\alpha(0)}}^{\alpha(t)} \ge 0 .
\label{equ:second_law}
\end{equation}
The main feature of this particular formulation of the second law of thermodynamics is that it holds true for any initial condition $\mu$ and for an arbitrary protocol $\alpha(t)$ to the contrary of the classical thermodynamical formulation of the second law as seen for example in \cite{Callen1985}.
Using the definition of total entropy \eqref{def:total_entropy} we can see that the mean total entropy production consists of the difference of Shannon entropies \eqref{def:Shannon_entropy} for system and the entropy production in thermal bath
\[
S(\mu_{\pi \alpha(T)}) - S(\mu_{\alpha(0)}) + \left\langle S^{\text{bath}}_{\alpha(t)}(\omega) \right\rangle_{\mu_0} \ge 0 ,
\]
which in case of the equilibrium process, which connects states described by Boltzmann distribution can be rewritten to the well known form of the second law of thermodynamics
\[
{\mathcal W}^{\text{sys}}(\alpha(t)) \ge {\mathcal F}_{\alpha(T)} - {\mathcal F}_{\alpha(0)} ,
\]
where $\mathcal F_\alpha$ denotes again the equilibrium free energy and ${\mathcal W}^\text{sys}(\alpha(t)$ is the mean value of the total work \eqref{equ:total_local_work_relation}.
\section{Quasistatic processes}
\label{sec:quasistatic_processes_eq}
Up to now we have been discussing a general processes while we haven't made any specific assumptions about the trajectory in the external parameters space $\vec{\alpha}(t)$ nor about the system involved.
At first we will restrict ourselves to equilibrium systems, hence we consider all the forces in the system to be of potential nature and also to be attached only to single thermal reservoir.
The non-equilibrium quasistatic processes will be discussed in later chapters, while they are the main object of interest of this work.
In equilibrium thermodynamics the equilibrium or in other words quasistatic processes play a prominent role in the definition of thermodynamic entropy or thermodynamical potentials in general.
The quasistatic process in general can be defined as the limiting process when we rescale the velocity of the changes of external parameters $\dot{\alpha}$ to zero, while preserving the path
\begin{align}
\alpha(t) & \longrightarrow \alpha(\epsilon t) , &
t \in [0,T] & \longrightarrow t \in \left[ 0, \frac{T}{\epsilon} \right] ,
\label{equ:time_rescaling}
\end{align}
where $\epsilon$ denotes the scaling parameter characterizing the magnitude of the velocity $\epsilon \, \dot{\alpha} (\epsilon t)$.
The quasistatic process is then obtain by taking the limit $\epsilon \to 0^+$.
It is reasonable to further assume that the state of the system at any time $\mu_t$ during the quasistatic process is in the vicinity of the corresponding equilibrium state $\rho_{\alpha(\epsilon t)}$,
also it is reasonable to assume that the difference is diminished with $\epsilon \to 0^+$, while the system has more time to relax closer to equilibrium.
Moreover we can also assume that the difference between the equilibrium state and an actual state has to change on the same time scale for it not to exceed bounds as time goes on,
so in general we assume the solution of the time evolution \eqref{equ:Markov_stochastic_time_evolution} to be
\[
\mu_t (x) = \rho_{\alpha(\epsilon t)} (x) + \epsilon \, \Delta \mu_{\epsilon t} (x) ,
\]
where $\Delta \mu_t$ denotes the difference.
The time evolution equation \eqref{equ:Markov_stochastic_time_evolution} then gives us the prescription how to determine the $\Delta \mu_t$
\[
\epsilon \, \dot\alpha(\epsilon t) \cdot \left. \nabla_\alpha \rho_\alpha (x) \right|_{\alpha=\alpha(\epsilon t)} + \epsilon^2 \left. \partial_s \Delta \mu_s (x) \right|_{s=\epsilon t}
= {\mathcal L}^*_{\alpha(\epsilon t)} \left[ \rho_{\alpha(\epsilon t)} + \epsilon \, \Delta \mu_{\epsilon t} \right] (x) ,
\]
which up to the first order in $\epsilon$ leads to
\begin{equation}
\dot\alpha(\epsilon t) \cdot \left. \nabla_\alpha \rho_\alpha (x) \right|_{\alpha=\alpha(\epsilon t)} = {\mathcal L}^*_{\alpha(\epsilon t)} \left[ \Delta \mu_{\epsilon t} \right] (x) ,
\label{equ:equation_first_order_density_correction}
\end{equation}
where we have used that the equilibrium state $\rho_\alpha$ is stationary under the equilibrium dynamic.
When we have determined the structure of the state of the system undergoing the quasistatic process we can proceed further.
In the quasistatic limit we can directly see that the mean value of the total work in equilibrium \eqref{equ:local_power} is the \emph{geometric} integral over the path of external parameters
\begin{align}
{\mathcal W}(\alpha) = \lim_{\epsilon \to 0^+} {\mathcal W}_{\rho_{\alpha(0)}}(\alpha(\epsilon t))
&= \lim_{\epsilon \to 0^+} \int\limits_0^{\frac{T}{\epsilon}} {\mathrm d} t \; \left\langle \epsilon \, \dot{\alpha}(\epsilon t) \cdot \left. \nabla_\alpha E_\alpha (x) \right|_{\alpha=\alpha(\epsilon t)} \right\rangle_{\mu_t} \nonumber \\
&= \int {\mathrm d} \alpha \cdot \left\langle \nabla_\alpha E_\alpha \right\rangle_{\rho_\alpha} ,
\label{equ:quasistatic_work_eq}
\end{align}
where $\langle \nabla_\alpha E_\alpha \rangle_{\rho_\alpha}$ is \emph{the thermodynamic force}\index{thermodynamic force}.
As an example let us consider Heisenberg model in the presence of the constant external magnetic field $\vec{H}$ with the Hamiltonian
\[
H(\{\vec{\sigma}_i\}) = H_{\text{int}}(\{\vec{\sigma}_i\}) + g \vec{H} \cdot \sum_i \vec{\sigma}_i ,
\]
where $H_{\text{int}}(\{\vec{\sigma}_i\})$ describes the interaction of the spins independent of the external field.
One can see that the thermodynamic force corresponding to the quasistatic change of external magnetic field is the magnetization in agreement with thermodynamics
\[
{\mathcal W}(\alpha) = - \int {\mathrm d} \vec{H} \cdot \vec{M} .
\]
For the heat the situation is bit more complicated.
We start from the definition \eqref{equ:local_heat_production} by transferring the backward Kolmogorov generator from the potential to the total probability density by using \eqref{def:backward_generator},
then we can finally apply the first order correction \eqref{equ:equation_first_order_density_correction}
\begin{multline*}
{\mathcal Q}(\alpha) = \lim_{\epsilon \to 0^+} {\mathcal Q}_{\rho_{\alpha(0)}}(\alpha(\epsilon t))
= \lim_{\epsilon \to 0^+} \int\limits_0^{\frac{T}{\epsilon}} {\mathrm d} t \; \left\langle {\mathcal L}_{\alpha(\epsilon t)} \left[ E_{\alpha_{\epsilon t}} \right] \right\rangle_{\mu_t} = \\
= \lim_{\epsilon \to 0^+} \int\limits_0^{\frac{T}{\epsilon}} {\mathrm d} t \; \epsilon \, \dot\alpha(\epsilon t) \cdot \int {\mathrm d} \Gamma(x) \; \left. \nabla_\alpha \rho_\alpha(x) \right|_{\alpha=\alpha(\epsilon t)} E_{\alpha_{\epsilon t}}(x) = \\
= \int {\mathrm d} \alpha \cdot \int {\mathrm d} \Gamma(x) \; \nabla_\alpha \rho_\alpha(x) \, E_\alpha(x) .
\end{multline*}
From where we can see that in the quasistatic limit \emph{the first law of thermodynamics}\index{first law of thermodynamics} can be written as
\begin{multline*}
\int {\mathrm d} \alpha \cdot \nabla_\alpha \left\langle E_\alpha \right\rangle_{\rho_\alpha} = \\
= \int {\mathrm d} \alpha \cdot \left\langle \nabla_\alpha E_\alpha \right\rangle_{\rho_\alpha}
+ \int {\mathrm d} \alpha \cdot \int {\mathrm d} \Gamma(x) \; E_\alpha(x) \, \nabla_\alpha \rho_\alpha(x) = \\
= {\mathcal W}(\alpha) + {\mathcal Q}(\alpha).
\end{multline*}
In the case the equilibrium state is described by Boltzmann distribution at inverse temperature $\beta$ the mean value of the total heat can be expressed by covariance
\[
{\mathcal Q}(\alpha) = \int {\mathrm d} \alpha \cdot \left[ \left\langle \nabla_\alpha \left( \beta E_\alpha \right) \right\rangle_{\rho_\alpha} \left\langle E_\alpha \right\rangle_{\rho_\alpha}
- \left\langle \nabla_\alpha \left( \beta E_\alpha \right) E_\alpha \right\rangle_{\rho_\alpha} \right] ,
\]
or we can obtain \emph{the second law of thermodynamics}\index{second law of thermodynamics} in the form of relation between the total heat $\mathcal Q$ and a total differential of entropy with the temperature $1/\beta$ being the Lagrange multiplier
\begin{equation}
{\mathcal Q}(\alpha)
= - \int {\mathrm d} \alpha \cdot \frac{1}{\beta} \nabla_\alpha \left\langle \ln \rho_\alpha \right\rangle_{\rho_\alpha}
= \int \frac{1}{\beta} \; {\mathrm d} S(\alpha)
\label{equ:Clausius_relation}
\end{equation}
where $S(\alpha)$ is \emph{the Shanon entropy}\index{entropy!Shanon}
\begin{equation}
S(\alpha) = - \left\langle \ln \rho_\alpha \right\rangle_{\rho_\alpha} ,
\label{def:Shannon_entropy}
\end{equation}
and $\rho_\alpha$ is the Boltzmann distribution.
\subsection{Quasistatic work fluctuations}
The total work and the total heat are random quantities and as such one can be interested in the behaviour of the fluctuations in the quasistatic limit.
Although the problem is still open,
some partial results based upon the general validity of the Crooks relation \eqref{equ:Crooks_relation} and Jarzynski equality \eqref{equ:Jarzynski_equality} can be obtained.
Namely the consequence of the validity of the Jarzynski equality in the quasistatic limit is that the difference between the free energy and the mean value of the total work is given by the variance of the total work in the leading order.
To sketch the proof we start from the Jarzynski equality in the form
\[
{\mathrm e}^{\beta \left( \left\langle W(\omega) \right\rangle - \Delta {\mathcal F} \right) } = \left\langle {\mathrm e}^{-\beta \left[ W(\omega) - \left\langle W(\omega) \right\rangle \right] } \right\rangle ,
\]
where we assume that in the quasistatic limit the mean value of the total work is close to the free energy difference as well as the probability distribution of the total work is in the quasistatic limit narrow and hence $W(\omega) - \left\langle W(\omega) \right\rangle$ can be considered as small parameter.
The proposition is then obtained by expanding the terms up to the leading order
\[
\left\langle W(\omega) \right\rangle - \Delta {\mathcal F} \approx \frac{\beta}{2} \left\langle \left( W(\omega) - \left\langle W(\omega) \right\rangle \right)^2 \right\rangle ,
\]
from where it follows that the probability distribution of the total work has the zero variance in the quasistatic limit.
Notice that it does not necessarily ensures the probability distribution to be asymptotically Gaussian,
although in case of overdamped diffusion Speck and Seifert \cite{Speck2004} proved that the distribution of the total work is asymptotically in the quasistatic limit Gaussian.
\chapter{Stochastic models of open systems}
\label{chapter:models}
For several centuries physicists have been trying to describe the real world using mathematics.
During that time period they have established procedures how to describe any real physical system.
Although the procedure can differ in technical aspects depending on the application,
in general it always goes along similar lines.
First we usually determine the set of all possible states of the system with respect to the level of description and observables of interest
e.g. for mechanical particle system we determine all possible positions and momentums, for molecules we determine all possible conformations,
then we relate these states to physical observables
and finally we describe the time evolution, i.e. how the state changes in the course of time.
In this chapter we will introduce the mathematical formalism necessary to describe open thermodynamic systems in terms of continuous-time Markov stochastic processes,
our presentation is rather informal as it is not the goal of this work to provide an exact mathematical treatment of the subject, cf. \cite{Feller2008}.
We introduce some of the models and examples used later in this thesis.
\section{Deterministic processes}
In classical mechanics we describe the \emph{state}\index{state} $x$ of the system by a complete collection of dynamical observables associated with all independent degrees of freedom.
For example in the case of single particle we need only to know its position and momentum $x=(\vec{q},\vec{p})$ to fully determine its state.
Similarly mathematical pendulum is efficiently described by the angle and impulse momentum $x=(\varphi,l)$,
or ratchet with $n$ tooths described by the position $x \in \mathbb Z_n$ of the pawl.
All possible states, i.e. configurations of the system, together make a \emph{configuration space}\index{configuration space} $\Omega$.
In this context physical \emph{observables}\index{observable!state}, which depends on the state in which the physical system is, are defined as functions
\begin{equation}
A: \Omega \longrightarrow {\mathbb R},
\label{def:observable}
\end{equation}
where $A$ denotes some physical quantity.
For example in the case of single particle its position is observable defined as $\vec{Q}(x)=\vec{q}$, similarly its velocity is $\vec{V}(x)= \frac{\vec{p}}{m}$
and also energy can be considered as an observable $E(x) = \frac{\vec{p}^2}{2m} + U(\vec{q})$, where $U$ is a potential in which presence the particle is.
Physical systems evolve with time and we denote $x_t$ the state of the system at time $t$.
One particular realization of the time evolution then creates a \emph{path}\index{path} $\omega$ in configuration space $\Omega$
\[
\omega = \left\{ x_t \, \middle| \, t \in {\mathbb R} \right\}.
\]
The path can also be restricted over the time interval of our interest e.g. $[0,T]$, which we denote by
\[
\omega^{[0,T]} = \left\{ x_t \, \middle| \, 0 \le t \le T \right\} .
\]
Introducing the path of the physical system enables us to describe another class of observables, so called \emph{path observables}\index{observable!path},
which do not depend only on the physical states of the system but also on its history.
Typical representatives are heat and work, which we will define later, or first passage times $T_y(\omega^{[0,\infty)}) = \min_{s: x_s = y} s $.
In deterministic mechanics the history before the time $t$ represented by the path $\omega^{(-\infty,t)}$ uniquely determines the state $x_t$ at time $t$,
i.e. the \emph{time evolution}\index{time evolution!deterministic} in deterministic case is defined as
\begin{equation}
x_t : \omega^{(-\infty,t)} \longrightarrow \Omega.
\label{def:mechanics}
\end{equation}
In most cases we usually don't need a complete history but only a part of it, e.g. $\omega^{[0,t)}$.
\subsubsection{Example: Hamiltonian mechanics}
\index{Hamiltonian mechanics}
A typical example of such deterministic time evolution is the Hamiltonian mechanics of one particle described by Hamilton equations
\begin{equation}
\begin{aligned}
\partial_t \vec{q}_t &= \left\{ \vec{q}_t , H_t\left(\vec{q}_t,\vec{p}_t\right) \right\}, \\
\partial_t \vec{p}_t &= \left\{ \vec{p}_t , H_t\left(\vec{q}_t,\vec{p}_t\right) \right\},
\end{aligned}
\label{equ:hamilton_equations}
\end{equation}
where $\{ \cdot , \cdot \}$ denotes Poisson bracket and $H_t(\vec{q},\vec{p})$ is time-dependent Hamiltonian.
The consequence of the Hamilton equations it that the full time evolution depends only on the initial condition, hence the dependence on the history is in a sense trivial.
By integrating these differential equations from initial state $x_0=(\vec{q}_0,\vec{p}_0)$ at time $0$ over the time interval $[0,t)$ we obtain
\begin{equation}
\begin{aligned}
\vec{q}_t &= \vec{q}_0 + \int\limits_0^t {\mathrm d} s \; \left\{ \vec{q}_s , H_s(\vec{q}_s,\vec{p}_s) \right\} , \\
\vec{p}_t &= \vec{p}_0 + \int\limits_0^t {\mathrm d} s \; \left\{ \vec{p}_s , H_s(\vec{q}_s,\vec{p}_s) \right\} ,
\end{aligned}
\label{equ:hamiltonian_mechanics}
\end{equation}
where we can recognize the dependence on the path $\omega^{[0,t)}$ more clearly.
Also the result \eqref{equ:hamiltonian_mechanics} more resembles the definition \eqref{def:mechanics} of the time evolution.
The Hamiltonian mechanics is memory-less in the sense that the state at an arbitrary time $t$ contains enough information for the prediction of the system's behavior in the future with respect to the time $t$.
When modeling open systems coupled to their environment, we are forced to abandon the deterministic Hamiltonian structure and to replace with a more general stochastic law, yet a proper identification of relevant degrees of freedom often allows us to retain the memory-less property.
Informally, this leads to the basic concept of \emph{Markovian dynamics}:
The stochastic dynamical system is Markovian whenever the future time evolution started from state $x_t$ at any time $t$ (the "present") is statistically independent of the way the state $x_t$ was prepared (the "past").
\section{Stochastic processes}
The detailed description of a large physical system in terms of all microscopic degrees of freedom becomes intractable since the number of available microscopic states grows exponentially with the number of degrees of freedom and due to the high complexity of the time evolution.
Fortunately, we usually do not need the full information about the system yet we still want to make estimates about the values of relevant observables, the probabilistic approach proves to be useful.
Within this approach we replace the notion of the state from microscopic level of description, now called \emph{microstate}\index{state!micro-}, by the \emph{configuration of the system}\index{state!configuration},
which corresponds to the maximum information about the real physical state in principle available on the coarse-grained level.
While the configuration is determined only by the maximum information available to us,
which does not necessarily corresponds to the full information contained on the microscopic level of description,
we can conclude that one configuration can correspond to several microstates.
Moreover in most cases we are not able to fully determine the exact configuration of the system either,
but we are at least able to determine in which of them the system is more or less likely to be.
Hence the physical state\index{state} is then described by probability distribution $\mu(x)$ over the configuration space $\Omega$,
which members $x \in \Omega$ are all possible configurations of the system.
Similarly as in the deterministic approach,
we also define state-dependent observables\index{observable!state} as in this case \emph{measurable} functions \eqref{def:observable} from configuration space $\Omega$ to results represented by real numbers ${\mathbb R}$,
as well as the path observables\index{observable!path} as \emph{measurable} functions on paths.
The remaining step is to adapt the time evolution\index{time evolution!stochastic} to fit into the current framework.
We cannot determine the exact outcome of the state anymore, hence we describe the time-evolution as a probability measure ${\mathrm d} \mathbb P$ on paths $\omega$ with normalization to unity
\[
\int \dprob{\omega} = 1 ,
\]
i.e. we assign a probability to each possible path, which also contain complete information about the initial state.
The system's state at time $t$ is represented by the marginal probability
\[
\mu_t (x) = \int \dprob{\omega} \; \delta_{X_t}(x),
\]
where $\delta(\cdot)$ is the Dirac delta function in continuous case or the Kronecker's delta in discrete case and $X_t$ denotes the configuration in time $t$ from the path $\omega$, $X_t \in \omega$.
Within this framework we can consider the deterministic time evolution as a special case in which each configuration is uniquely mapped on another.
Although we have described the time evolution in its entirety, we are usually interested only in its finite range.
To address this requirement we define the probability measure up to the time $t$ as
\[
\dprob[{(-\infty,t]}]{\omega^{(-\infty,t]}} = \int\limits_{\omega^{(-\infty,t]} \subset \varphi} \dprob{\varphi} ,
\]
where we integrate over all paths $\varphi$ containing the path $\omega^{(-\infty,t]}$,
by which we can also fully characterize the time evolution up to arbitrary time.
The state of the system at time $t$ is also determined by this probability measure
\[
\mu_t (x) = \int \dprob[{(-\infty,t]}]{\omega} \; \delta_{X_t}(x).
\]
To correctly predict the time-evolution we in general need to know the full history of the system, which is usually not accessible to us,
fortunately in most cases the dependency of the time evolution on the history weak enough so that in order to determine the plausibility of states at time $t$ we only need to know the path $\omega^{[t-T,t)}$ for not too large $T$.
In these cases we can restrict the description of the time evolution to the probability measure on more confined time interval $[t-T,t]$
\[
\dprob[{[t-T,t]}]{\omega^{[t-T,t]}} = \int\limits_{\omega^{[t-T,t]} \subset \varphi^{(-\infty,t]}} \dprob[{(-\infty,t]}]{\varphi^{(-\infty,t]}}
= \int\limits_{\omega^{[t-T,t]} \subset \varphi} \dprob{\varphi} .
\]
This also allows us to exclude the information about the initial state from the probability measure
and characterize the time-evolution by the conditional probability measure on the time interval $(t-T,t]$.
Hence the mean value over all paths of some path observable $A$ can equivalently be obtained as the mean value with respect to the conditional measure
conditioned upon the initial configuration and then by averaging over the initial condition
\begin{multline}
\int \dprob[{[t-T,t]}]{\omega} \; A(\omega) = \\
= \int {\mathrm d} y \; \mu_{t-T}(y) \int \dcprob[{(t-T,t]}]{\varphi}{X_{t-T}=y} \; A(\varphi \cup \{X_{t-T} = y \} ) .
\label{def:conditional_measure}
\end{multline}
Using this conditional probability the state at time $t$ can be evaluated as
\begin{equation}
\mu_t (x) = \int {\mathrm d} y \; \mu_{t-T}(y) \int \dcprob[{(t-T,t]}]{\omega}{X_{t-T}=y} \; \delta_{X_t}(x).
\label{equ:state_by_conditional_measure}
\end{equation}
\section{Markov continuous-time stochastic processes}
\label{sec:Markov_processes}
\index{Markov property}
\index{Markov stochastic process}
In context of stochastic processes the Markov processes are those, which have extremely weak dependence on the history.
Hence to obtain the full time evolution it is sufficient to know the conditional probability measure for arbitrarily short time.
As a consequence we can describe the full time evolution over the time interval $T$ by the time evolution over arbitrary number of shorter intervals.
This feature can be exploited in order to show that the conditional mean value of some observable at time $t > t_k > t_{k-1} > \dots > t_0 $ conditioned on several values $x_0, \dots , x_k$ in the past does depend only on the last value $x_k$,
\[
\left\langle X_t \middle| X_{t_k}=x_k , X_{t_{k-1}}=x_{k-1}, \dots , X_{t_0} = x_0 \right\rangle_{\mu_{t_0}} =
\left\langle X_t \middle| X_{t_k}=x_k \right\rangle_{\mu_{t_k}} .
\]
\subsection{Forward Kolmogorov generator}
\index{Kolmogorov generator!forward}
Another consequence of Markov property is the existence of autonomous time evolution for probability densities $\mu_t$ describing the state of the system,
which is in particular useful when we are interested only in estimates of state dependent observables.
However Markov property alone is not sufficient, we also need to \emph{assume} that the \emph{time evolution is smooth},
in particular that the difference between the state at time $t$ and time $t-\Delta t$ diminishes proportionally with $\Delta t$.
Only then we can define the \emph{forward Kolmogorov generator} as the linear operator on probability densities
\begin{multline}
{\mathcal L}^*_t [\mu_t] =
\lim_{\Delta t \rightarrow 0^+} \frac{1}{\Delta t} \Bigl[ \int {\mathrm d} y \; \mu_{t-\Delta t}(y) \times \\
\times \int \dcprob[{(t-\Delta t,t]}]{\omega}{X_{t-\Delta t}=y} \; \delta_{X_t}(x) - \mu_{t-\Delta t}(x) \Bigr] ,
\label{def:forward_generator}
\end{multline}
and the time evolution of the state is described by the differential equation
\begin{equation}
\partial_t \mu_t (x) = {\mathcal L}_t^* [\mu_t] (x) ,
\label{equ:Markov_stochastic_time_evolution}
\end{equation}
compare with \eqref{equ:state_by_conditional_measure}.
By integrating \eqref{equ:Markov_stochastic_time_evolution} over all possible states and by using the fact that the probability density $\mu_t$ has to be always normalized to unity,
we obtain a condition
\begin{equation}
\int {\mathrm d} x \; {\mathcal L}_t^* [\mu_t] (x) = 0 ,
\label{equ:normalization_condition}
\end{equation}
note that this condition can also be obtained directly from the definition of the forward Kolmogorov generator \eqref{def:forward_generator} and the normalization of probability densities.
By explicit integration of \eqref{equ:Markov_stochastic_time_evolution} we obtain formal solution
\[
\mu_t(x) = \overleftarrow{\exp} \left\{ \int\limits_{t_0}^t {\mathrm d} t \; {\mathcal L}^*_t \right\} [\mu_{t_0}](x),
\]
where $\overleftarrow{\exp}$ denotes time-ordered exponential.
\subsubsection{Example: Hamiltonian mechanics revisited}
\index{Hamiltonian mechanics}
To provide a simple example, we derive the forward Kolmogorov generator for the Hamiltonian mechanics.
Although the Hamiltonian mechanics is deterministic, it can also be described within the framework of stochastic time-evolution as was noted above.
The conditional probability measure is Dirac-like measure which gives zero for every trajectory which does not comply with solution of Hamilton equations \eqref{equ:hamiltonian_mechanics}.
Hence for the generator \eqref{def:forward_generator} we obtain
\[
{\mathcal L}^*_t [\mu_t] (\vec{q},\vec{p}) =
\lim_{\Delta t \rightarrow 0^+} \frac{1}{\Delta t}
\left[ \mu_{t-\Delta t} \left( \vec{q}_{t-\Delta t}(\vec{q},\vec{p},t) , \vec{p}_{t-\Delta t}(\vec{q},\vec{p},t) \right) - \mu_{t-\Delta t}(\vec{q},\vec{p}) \right] ,
\]
where $( \vec{q}_{t-\Delta t}(\vec{q},\vec{p},t), \vec{p}_{t-\Delta t}(\vec{p},t) )$ denotes the configuration of the particle at time $t-\Delta t$
starting from which we reach the configuration $x=(\vec{q},\vec{p})$ at time $t$.
Because the Hamiltonian mechanics is deterministic we can determine a past configuration of the system knowing the present configuration.
Using \eqref{equ:hamiltonian_mechanics} we obtain
\begin{align*}
\vec{q}_{t-\Delta t}(\vec{q},\vec{p},t) &= \vec{q} + \int\limits_t^{t-\Delta t} {\mathrm d} s \; \left\{ \vec{q}_s , H_s(\vec{q}_s,\vec{p}_s) \right\} , \\
\vec{p}_{t-\Delta t}(\vec{q},\vec{p},t) &= \vec{p} + \int\limits_t^{t-\Delta t} {\mathrm d} s \; \left\{ \vec{p}_s , H_s(\vec{q}_s,\vec{p}_s) \right\} .
\end{align*}
By expanding the distribution density in $\Delta t$ and inserting the solution above we obtain the \emph{Liouville's equation}\index{Liouville's equation}
\[
{\mathcal L}^*_t [\mu_t] (\vec{q},\vec{p}) = \left\{ H_t(\vec{q},\vec{p}) , \mu_t(\vec{q}, \vec{p}) \right\} .
\]
Notice that in the context of stochastic Markov processes the condition \eqref{equ:normalization_condition} corresponds to the Liouville's theorem.
\subsection{Spectral properties of forward Kolmogorov generator}
The condition \eqref{equ:normalization_condition} has several mathematical implications related to the spectrum of the generator.
The first one states that every \emph{eigenvector} $\nu_t$ of the generator ${\mathcal L}^*_t$ at fixed time $t$ corresponding \emph{to the non-zero eigenvalue} $\lambda_t$ \emph{has zero integral} over all configurations
\[
0 = \int {\mathrm d} x \; {\mathcal L}^*_t [\nu_t] (x) = \lambda_t \int {\mathrm d} x \; \nu_t (x).
\]
It means that all eigenvectors corresponding to non-zero eigenvalues are either zero up to the null set or have both negative and positive part.
This necessarily leads to the condition that the \emph{real part of all eigenvalues has to be non-positive},
otherwise the negative part of the eigenvector will grow over all bounds causing the probability distribution to be negative $\mu_t (x) < 0$.
In most of our applications we will also assume that if we fix the parameters of the generator at some fixed time $\widetilde{{\mathcal L}}^* = {\mathcal L}^*_s$ and let the system evolve with these fixed values,
i.e. using the generator $\widetilde{{\mathcal L}}^*$ instead of the original generator ${\mathcal L}^*_t$, then there always exists a steady state to which the system will tend to converge.
\subsection{Steady state}
\index{steady state}
The steady-sate of a Markov system described by the time-independent generator ${\mathcal L}^*_t={\mathcal L}^*$ is described by the stationary distribution which is obtained as
\begin{equation}
{\mathcal L}^* [ \rho ] (x) = 0 ,
\label{equ:stationary_condition}
\end{equation}
so the steady states $\rho$ lies in the kernel of the linear operator.
In general the notion of stationarity can be extended to periodical time evolutions,
where the steady state $\rho_t$, now explicitly depending on time $t$, is the \emph{periodic} solution of \eqref{equ:Markov_stochastic_time_evolution}.
\subsection{Backward Kolmogorov generator}
\index{Kolmogorov generator!backward}
Until now we have described the time evolution in terms of time-dependent probability distribution $\mu_t$.
Equivalently, it can be described as the time-evolution on observables (cf. the quantum-mechanical Heisenberg picture),
where we can go from Schrödinger picture (representation) to Heisenberg picture.
For this purpose we define the \emph{backward Kolmogorov generator} as the adjoint generator
\begin{equation}
\int {\mathrm d} x \; A(x) \, {\mathcal L}^*_t[\mu_t] (x) = \int {\mathrm d} x \; {\mathcal L}_t [A] (x) \, \mu_t (x) .
\label{def:backward_generator}
\end{equation}
The time evolution using the backward Kolmogorov generator is then characterized by
\[
{\mathrm d}_t A_t (x) = \partial_t A_t(x) + {\mathcal L}_t [A_t] (x) ,
\]
where by ${\mathrm d}_t$ we denote the total time derivative,
while by $\partial_t$ we denote the time derivative of explicit time-dependence of observable $A_t$ with respect of time.
The time derivative of the mean value of some explicitly time-dependent observable $A_t$ can be written using the backward Kolmogorov generator as
\begin{equation}
\partial_t \left\langle A_t \right\rangle_{\mu_t} = \left\langle \partial_t A_t + {\mathcal L}_t [ A_t ] \right\rangle_{\mu_t} .
\label{equ:time_derivative_observable}
\end{equation}
\section{Continuous-time jump process}
\label{sec:jump_processes}
\index{jump processes}
Continuous-time jump processes represent a large class of the Markov stochastic processes with countable many configurations.
They are used to model a huge variety of systems including lattice models, ratchets, chemical networks, biological systems and semi-classical description of quantum systems, etc.
From a mathematical point of view the configuration space $\Omega$ is a countable set with a counting measure and the probabilities $\mu_t(x)$ are densities with respect to that measure.
Hence integration over all possible configurations is represented as a summation
\[
\int {\mathrm d} x \longrightarrow \sum\limits_x .
\]
Typical paths in continuous-time jump processes consists of intervals where the configuration of the system does not change followed by the sudden change of the configuration (jump),
for illustration see figure \ref{pic:illustration_of_path_jump_process}.
\begin{figure}[t]
\caption{Illustration of several paths $\omega^{[0,3]}$ all beginning at configuration $1$ for a system with three configurations (or ``levels''). }
\label{pic:illustration_of_path_jump_process}
\begin{center}
\includegraphics[width=.8\textwidth,height=!]{ilustration.pdf}
\end{center}
\end{figure}
Moreover we define all paths as right-continuous, i.e. the configuration at jump-time is the same as the configuration after the jump.
By $x_{t^-}$ we denote the configuration just before the jump
\[
x_{t^-} = \lim\limits_{s \rightarrow t^-} x_s .
\]
The time evolution is fully characterized by \emph{transition rates}\index{transition rate} $\rate[t]{x}{y}$ defining the conditional probability measure on paths $\omega$
\begin{equation}
\dcprob[{(0,T]}]{\omega}{X_0=x_0}
= \exp \left[ - \int\limits_0^T {\mathrm d} s \; \lambda_s(x_s) \right]
\prod_{i=1}^{m} \rate[t]{x_{t_i^-}}{x_{t_i}} \; {\mathrm d} t_i ,
\label{def:jump_process_path_probability}
\end{equation}
where $m$ is the number of jumps in the path $\omega$, $x_0$ is the initial configuration of the path $\omega$ ($x_0 \in \omega$), $t_i$ are the jump times,
$\lambda_t(x)$ denotes the \emph{escape rate}\index{escape rate} from configuration $x$,
\begin{equation}
\lambda_t(x) = \sum\limits_{y \neq x} \rate[t]{x}{y} .
\label{def:escape_rate}
\end{equation}
By definition impossible transitions are characterized by zero transition rate, $\rate[t]{x}{y}=0$.
Because the time-evolution is Markovian,
we can alternatively describe the time evolution using forward \eqref{def:forward_generator}\index{Kolmogorov generator!forward!Markov jump process} and backward \eqref{def:backward_generator}\index{Kolmogorov generator!backward!Markov jump process} Kolmogorov generator
\begin{align}
{\mathcal L}_t^* [\mu_t] (x) &= \sum\limits_{y \neq x} \left[ \mu_t(y) \rate[t]{y}{x} - \mu_t(x) \rate[t]{x}{y} \right] ,
\label{equ:forward_Kolmogorov_gen_jump} \\
{\mathcal L}_t [A] (x) &= \sum\limits_{y \neq x} \rate[t]{x}{y} \left[ A(y) - A(x) \right] .
\label{equ:backward_Kolmogorov_gen_jump}
\end{align}
The derivation of Kolmogorov generators from the conditional path probability measure ${\mathrm d} \mathbb P$ is quite straightforward although technical
and is provided in appendix \ref{ap:jump_processes_generators}.
The introduction of the forward Kolmogorov generator yields to the \emph{master equation}\index{master equation}
\begin{equation}
\partial_t \mu_{t}(x) = \sum\limits_{y \neq x} \left[ \mu_t(y) \rate[t]{y}{x} - \mu_t(x) \rate[t]{x}{y} \right] ,
\label{equ:Master_equation}
\end{equation}
representing here the time evolution equation \eqref{equ:Markov_stochastic_time_evolution}.
\subsection{Probability currents}
\index{current!probabilistic}
In typical examples of continuous-time Markov jump processes like chemical networks, lattice models, etc.,
the jump is usually associated with transport or exchange of some quantity, e.g. electric charge, matter, with its surroundings.
To describe the currents associated with such exchanges we define the \emph{probability current}\index{current!probabilistic}
\begin{multline}
\curr[t]{x}{y} = \lim_{\Delta t \rightarrow 0^+} \frac{1}{\Delta t} \left[ \mu_{t-\Delta t} (x) \int \dcprob[{(t-\Delta t,t]}]{\omega}{X_{t-\Delta t} = x} \; \delta_{X_t}(y) - \right. \\
\left. - \mu_{t-\Delta t} (y) \int \dcprob[{(t-\Delta t,t]}]{\omega}{X_{t-\Delta t} = y} \; \delta_{X_t}(x) \right],
\label{def:current}
\end{multline}
which characterizes the exchange of probability along one possible transition $x \rightarrow y$ at one particular time $t$.
We can see, that by definition the current is antisymmetric
\[
\curr[t]{x}{y} = - \curr[t]{y}{x},
\]
and is zero along the impossible transitions.
\paragraph{Example: Particle current}
To provide a specific example, let us consider a transport of single charged particle over the lattice.
All possible configurations of the system are fully characterized by the position of the particle.
Hence the transition in the system corresponds to jump of the particle along the edge of the lattice.
Consider a situation when we do not know a precise position of the particle, but only a probability $\mu_t(x)$ that the particle is at the particular vertex $x$ at time $t$.
Then the average electrical current $\bar\jmath^{\,e}_{xy}(t)$ caused by the transition along the edge from $x$ to $y$ is the probability that the particle is at $x$
times the probability that the transition from $x$ to $y$ occurs times the charge $q$ of the particle.
Moreover the transition from $y$ to $x$ can also occur producing the counter-current, hence reducing the total current along the edge.
Altogether we obtain the average current along the edge from $x$ to $y$ being
\[
\bar\jmath^{\,e}_{xy}(t) = q \curr[t]{x}{y}.
\]
If we now consider having the huge amount of \emph{independent non-interacting} particles on the same lattice, then the effect of fluctuations will be largely reduced,
and we will actually observe a macroscopic electrical current $j^e_{xy}(t)$ along the edge from $x$ to $y$
\[
j^e_{xy}(t) = N \bar\jmath^{\,e}_{xy}(t) = q N \curr[t]{x}{y},
\]
where $N$ is the total number of particles on the lattice.
In case of Markov jump processes the probability current are characterized by transition rates $\rate[t]{x}{y}$ and the actual probability distribution $\mu_t$
\begin{equation}
\curr[t]{x}{y} = \mu_t(x) \rate[t]{x}{y} - \mu_t(y) \rate[t]{y}{x} .
\label{equ:current_MJP}
\end{equation}
Comparing this expression for probability current with the master equation \eqref{equ:Master_equation} we obtain
\begin{equation}
\partial_t \mu_t(x) = \sum_y \curr[t]{y}{x} ,
\label{equ:continuity}
\end{equation}
which is the \emph{continuity equation} for probabilities.
From the continuity equation \eqref{equ:continuity} we obtain Kirchhoff's-like law for probability currents in the steady state
\[
0 = \sum_y \curr{y}{x}
\]
as a consequence of the stationary condition \eqref{equ:stationary_condition} in case of the system with time-independent transition rates.
Notice that this condition does not necessarily ensures the probability currents to be zero in the steady state.
\subsection{Global detailed balance condition}
\label{ssec:detailed_balance_jump}
\index{detailed balance condition!global}
In nature there exists a large class of systems in which all currents vanish in steady state,
\begin{equation}
\curr{x}{y}=0,
\label{equ:zero_current}
\end{equation}
namely equilibrium systems, see examples below.
To be more specific, we restrict ourselves in this section to the class of systems with time-independent transition rates, $\rate[t]{x}{y} \equiv \rate{x}{y}$.
Under such assumptions the condition on all probability currents to be zero is equivalent to the \emph{global detailed balance condition}\index{detailed balance condition!global}
\begin{equation}
\frac{\rho(y)}{\rho(x)} = \frac{\rate{x}{y}}{\rate{y}{x}} ,
\label{equ:global_detailed_balance_jump}
\end{equation}
compare \eqref{equ:zero_current} and \eqref{equ:current_MJP}.
\footnote{The global detailed balance condition is also valid for systems with arbitrarily symmetric part $\psi(x,y)=\sqrt{\rate{x}{y} \, \rate{y}{x}}$ of the transition rates.}
The immediate consequence of global detailed balance condition is closely related to the time-reversal symmetry.
Namely if there exists a sequence of configurations $x_0, x_1, \dots, x_n \equiv x_0 $ such that the transitions from $x_i$ to $x_{i+1}$ are possible,
then also transitions in the opposite direction are possible and the rates obeys
\[
1 = \frac{\rate{x_0}{x_1} \, \rate{x_1}{x_2} \cdots \rate{x_{n-1}}{x_0} }{ \rate{x_0}{x_{n-1}} \cdots \rate{x_2}{x_1} \, \rate{x_1}{x_0}} .
\]
Put in words, the probability \eqref{def:jump_process_path_probability} of the closed path, $\omega^{(0,T]} : X_0 = X_T$, and its \emph{time reversal}\index{path!time reversed}
\begin{equation}
\Theta \omega^{(0,T]} = \left\{ X_{T-t} \middle| X_t \in \omega^{(0,T]} \right\}
\label{equ:time_reversal_jump}
\end{equation}
are equal.
\subsubsection{Equilibrium systems}
\label{ssec:equlibrium_systems_jump}
As was mentioned before, the equilibrium systems are the systems obeying the global detailed balance condition \eqref{equ:global_detailed_balance_jump},
which is ensured by the fact that in the equilibrium there are no macroscopic currents in the system.
We consider the typical example is the system with discrete energy levels $E(x)$ in equilibrium with a single thermal bath at the inverse temperature $\beta$.
The stationary distribution in this case is given by canonical distribution of such system
\[
\rho(x) = \frac{1}{Z} {\mathrm e}^{-\beta E(x)},
\]
where $Z$ is the partition function.
As a consequence the antisymmetric part of the transition rates is closely related to the energy exchanged with the thermal bath and hence to the entropy production in the bath along the transition
\[
\ln \frac{\rate{x}{y}}{\rate{y}{x}} = \beta \left[ E(x) - E(y) \right] = S^{\text{bath}}(x)-S^{\text{bath}}(y),
\]
where $S^{\text{bath}}(x)$ denotes the entropy of the \emph{thermal bath},
here as a function of the system's configuration $x$.
Another example is the system attached to a single thermal and particle bath at the inverse temperature $\beta$ as well as to a single particle bath at the chemical potential $\mu$.
The stationary distribution in this case is given by the grand-canonical distribution
\[
\rho(x) = \frac{1}{Z_G} {\mathrm e}^{-\beta \left( E(x) - \mu N(x) \right)},
\]
where $N(x)$ is the number of particles present in the system and $Z_G$ denotes the grand-canonical partition function.
And similarly to the closed system the antisymmetric part corresponds to the entropy production within the thermal bath
\[
\ln \frac{\rate{x}{y}}{\rate{y}{x}} = \beta \left[ E(x) - E(y) \right] - \beta \mu \left[ N(x) - N(y) \right] = S^{\text{bath}}(x)-S^{\text{bath}}(y) .
\]
We can conclude that in equilibrium the antisymmetric part of the transition rates corresponds to the entropy production in thermal bath associated with such transition in general,
which is known as the \emph{local detailed balance condition}\index{detailed balance condition!local}.
\subsection{Three-level model}
\label{ssec:three_level}
\index{discrete model!three-level}
\index{three-level model}
Our first model representing the class of Markov jump processes is the three-level model with fixed energy levels $E_0 \le E_1 \le E_2$, see figure \ref{fig:three_level}.
\begin{figure}[thb]
\caption{Three-level model with every transition possible and $E_0$ calibrated to zero.}
\label{fig:three_level}
\begin{center}
\includegraphics[width=.45\textwidth,height=!]{3lvl-ilustration.pdf}
\end{center}
\end{figure}
We consider the system to be connected to a single thermal bath at the inverse temperature $\beta$
and driven out of global detailed balance by the action of the non-potential external force performing work $W_{\text{ext}}$ along any jump in the direction $0 \to 2 \to 1 \to 0$.
Hence the transitions in the direction $0 \rightarrow 1 \rightarrow 2 \rightarrow 0$ are more likely to occur than those in the opposite direction, $ 0 \rightarrow 2 \rightarrow 1 \rightarrow 0$.
As a typical example of such system can be considered a simplified classical model for three-level laser.
The transition rates has to obey global detailed balance condition in case $W_{\text{ext}}=0$, hence
\begin{align*}
\begin{aligned}
\rate{0}{1} &= \psi_{\beta}(0,1) \, {\mathrm e}^{- \frac{\beta}{2} \left( E_1 - E_0 - W_{\text{ext}} \right) } , \\
\rate{1}{2} &= \psi_{\beta}(1,2) \, {\mathrm e}^{- \frac{\beta}{2} \left( E_2 - E_1 - W_{\text{ext}} \right) } , \\
\rate{2}{0} &= \psi_{\beta}(0,2) \, {\mathrm e}^{- \frac{\beta}{2} \left( E_0 - E_2 - W_{\text{ext}} \right) } ,
\end{aligned}
&&
\begin{aligned}
\rate{1}{0} &= \psi_{\beta}(0,1) \, {\mathrm e}^{ \frac{\beta}{2} \left( E_1 - E_0 - W_{\text{ext}} \right) } , \\
\rate{2}{1} &= \psi_{\beta}(1,2) \, {\mathrm e}^{ \frac{\beta}{2} \left( E_2 - E_1 - W_{\text{ext}} \right) } , \\
\rate{0}{2} &= \psi_{\beta}(0,2) \, {\mathrm e}^{ \frac{\beta}{2} \left( E_0 - E_2 - W_{\text{ext}} \right) } ,
\end{aligned}
\end{align*}
where $\psi_{\beta}(x,y)=\psi_{\beta}(y,x)$ denotes arbitrary symmetric parts in general dependent on $\beta$.
Having the transition rates, we can directly obtain the escape rates
\begin{align*}
\lambda(0) &=
{\mathrm e}^{ \frac{\beta}{2} E_0} \left[ \psi_{\beta}(0,1) {\mathrm e}^{- \frac{\beta}{2} \left( E_1 - W_{\text{ext}} \right) }
+ \psi_{\beta}(0,2) {\mathrm e}^{- \frac{\beta}{2} \left( E_2 + W_{\text{ext}} \right)} \right] , \\
\lambda(1) &=
{\mathrm e}^{ \frac{\beta}{2} E_1} \left[ \psi_{\beta}(0,1) {\mathrm e}^{- \frac{\beta}{2} \left( E_0 + W_{\text{ext}} \right) }
+ \psi_{\beta}(1,2) {\mathrm e}^{- \frac{\beta}{2} \left( E_2 - W_{\text{ext}} \right)} \right] , \\
\lambda(2) &=
{\mathrm e}^{ \frac{\beta}{2} E_2} \left[ \psi_{\beta}(0,2) {\mathrm e}^{- \frac{\beta}{2} \left( E_2 - W_{\text{ext}} \right) }
+ \psi_{\beta}(1,2) {\mathrm e}^{- \frac{\beta}{2} \left( E_1 + W_{\text{ext}} \right)} \right] .
\end{align*}
\subsection{Multichannel two-level model}
\label{ssec:two_level}
\index{discrete model!two-level}
Another way how to break the global detailed condition is by attaching the system to multiple thermal or particle baths.
In this subsection we introduce the simplest of such models, which consists of a two-level system attached to two independent thermal baths
in general at different inverse temperatures $\beta_+ > \beta_-$, see figure \ref{fig:two_level}.
\begin{figure}[tbh]
\caption{ Two-level model with two channels enabling transitions between configurations $0$ and~$1$, with the energy gap $\Delta E$. }
\label{fig:two_level}
\begin{center}
\includegraphics[width=.5\textwidth,height=!]{2lvl-ilustration.pdf}
\end{center}
\end{figure}
Having the system in contact with multiple thermal baths enforces us to modify the framework of continuous-time Markov jump processes (see beginning of the section \ref{sec:jump_processes})
to include also the information about which action of particular thermal bath caused which transition.
Therefore a single realization of the time evolution of the system in contact with multiple thermal baths cannot be fully described only by the path on configurations $\omega$,
because it does not contain such information about the transition.
We extend the definition of the path and consequently conditional path measure \eqref{def:jump_process_path_probability}
\[
\dcprob{\widetilde\omega}{X_0=x_0}
= \exp \left[ - \int\limits_0^T {\mathrm d} s \; \sum\limits_i \lambda_i(x_s) \right]
\prod_{i=1}^{m} \rate[(i)]{x_{t_i^-}}{x_{t_i}} \; {\mathrm d} t_i ,
\]
where $\lambda_i(x)$ denotes the escape rate of the $i$-th bath and $\rate[(i)]{x}{y}$ denotes the transition rate of the bath associated with $i$-th transition.
As a consequence we obtain a modified master equation \eqref{equ:Master_equation} for the time evolution of this particular system
\[
\partial_t \mu_t(x)
= \sum\limits_{y \neq x} \mu_t(y) \left[ \rate[+]{y}{x} + \rate[-]{y}{x} \right] - \mu_t(x) \left[ \lambda_+(x) + \lambda_-(x) \right] ,
\]
where $k_+$ ($k_-$) are transition rates corresponding to the contact with thermal bath with the inverse temperature $\beta_+$ ($\beta_-$),
and where it is also assumed that transitions governed by a single thermal bath obey detailed balance condition \eqref{equ:global_detailed_balance_jump}, hence
\begin{align*}
\rate[+]{0}{1} &= \psi_+ \, {\mathrm e}^{- \frac{1}{2} \beta_+ \Delta E } , &
\rate[-]{0}{1} &= \psi_- \, {\mathrm e}^{- \frac{1}{2} \beta_- \Delta E } , \\
\rate[+]{1}{0} &= \psi_+ \, {\mathrm e}^{ \frac{1}{2} \beta_+ \Delta E } , &
\rate[-]{1}{0} &= \psi_- \, {\mathrm e}^{ \frac{1}{2} \beta_- \Delta E } .
\end{align*}
We can also define a probability current \eqref{equ:current_MJP} induced by a particular thermal bath as
\begin{equation}
\curr[\pm]{x}{y} = \rate[\pm]{x}{y} \, \mu_t(x) - \rate[\pm]{y}{x} \, \mu_t(y) .
\label{equ:current_multiple_baths}
\end{equation}
A notable special case is when both baths are at the same temperature although having different symmetric parts of transition rates $\psi_+ \neq \psi_-$.
This situation is formally equivalent to the case when the system is connected to two thermal baths at the same temperature with the same symmetric part of transition rates $\psi$,
while the symmetry between the channels is broken by application of non-potential force $F$
\begin{align*}
\rate[+]{0}{1} &= \psi \, {\mathrm e}^{- \frac{\beta}{2} \left( \Delta E - F \right) } , &
\rate[-]{0}{1} &= \psi \, {\mathrm e}^{- \frac{\beta}{2} \left( \Delta E + F \right) } , \\
\rate[+]{1}{0} &= \psi \, {\mathrm e}^{ \frac{\beta}{2} \left( \Delta E - F \right) } , &
\rate[-]{1}{0} &= \psi \, {\mathrm e}^{ \frac{\beta}{2} \left( \Delta E + F \right) } ,
\end{align*}
where we set
\begin{align*}
\psi &= \sqrt{\psi_+ \, \psi_-}, &
F &= \frac{1}{\beta} \ln \frac{\psi_+}{\psi_-}.
\end{align*}
\section{Diffusive systems}
\index{diffusion}
Diffusion are continuous-space Markovian models often used to describe the transport of matter or energy through a homogeneous environment,
the typical examples being heat conduction or transport of diluted chemicals in liquids or gases.
The behaviour of such systems is in general quite different from the systems described by continuous time Markov jump processes,
although overdamped diffusion is the limiting case of discrete jump process on the lattice with properly rescaled time and space \cite{Itzykson_Drouffe:Statistical_field_theory}.
In this section we build the formalism necessary to describe diffusion.
At first we show the diffusion limit of the random walk and its connection to the Wiener process,
then we will introduce the stochastic calculus, namely It\^o and Stratonowich calculus \cite{Oksendal2003,Evans2001}, as an universal tool to describe diffusion.
We will conclude this section with an application of the introduced calculus to the under- and over-damped diffusion along with classical description of Josephson junction.
\subsection{Random walks and Wiener process}
\label{ssec:diffusion_limit}
Random walks are realizations of discrete time Markov process on some graph,
as such they can be considered as a discrete versions of diffusion.
They are also a simple tool to study a basic properties of diffusion or more precisely the properties of Wiener process,
which is fundamental for stochastic calculus and hence for stochastic approach to diffusion.
Let us consider a single \emph{random walker}\index{random walker} on the uniform regular $d$-dimensional hypercubic lattice with lattice constant $a$, see fig. \ref{fig:random_walk}
\begin{figure}[htb]
\caption{Single random walker on square lattice with all possible transitions.}
\label{fig:random_walk}
\begin{center}
\includegraphics[width=.4\textwidth,height=!]{grid.pdf}
\end{center}
\end{figure}
The random walker at the beginning of each time step $\Delta t$ starts to move from the site of the lattice designed by its position $\vec{x}$ along an edge in the direction $\pm i$
and ends the move at the end of the time step in the site with position $\vec{x}^{(\pm i)} = ( x_0 , \dots , x_{i-1} , x_i \pm 1 , x_{i+1} , \dots , x_{d-1} )$.
The process of random walker on uniform lattice can be then described by discrete time Markov jump process with time step $\Delta t$ and with the same constant transition probability $p$ for all transitions
\[
\cprob{\vec{X}_{t+\Delta t} = \vec{x}^{(\pm i)}}{\vec{X}_t = \vec{x}}
= \cprob{\vec{X}_{t+\Delta t} = \vec{x}}{\vec{X}_t = \vec{x}^{(\pm i)}}
= p,
\]
where the random variable $\vec{X}_t$ denotes the position at time $t$.
Typically if we speak about the random walker, the random walker is not allowed to rest, hence the transition probability on the square lattice can be directly derived from the dimension of the lattice
\[
p = \frac{1}{2d} .
\]
Before we introduce diffusion limit we show some of the properties of random walker.
At first we can conclude that in this setup the probability distribution of positions of the random walker after $k$ time steps when starting from the position $\vec{x}_0$ is given by multinomial distribution
\begin{align*}
\cprob{\vec{X}_{k \Delta t} = \vec{x}_0 + \Delta \vec{x} }{ \vec{X}_0 = \vec{x}_0 }
&= \sum\limits_{m^\pm_0, \dots , m^\pm_{d-1}} \binom{k}{m^+_0 \dots m^-_{d-1}} \left(\frac{1}{2d}\right)^{m^+_0} \dots \left(\frac{1}{2d}\right)^{m^-_{d-1}} \\
&= \frac{1}{(2d)^k} \sum\limits_{m^\pm_0, \dots , m^\pm_{d-1}} \binom{k}{m^+_0 \dots m^-_{d-1}} ,
\end{align*}
where we sum over all possible numbers of steps $m_i^\pm$ in the direction $\pm i$ such that we change the position by $\Delta \vec{x}$ and the total count of steps is $k$, i.e.
\begin{gather*}
m^\pm_i \in \{0, 1, \dots, k \} , \\
\Delta x_i = a \left( m^+_i - m^-_i \right) , \\
k = \sum\limits_{i=0}^{d-1} \sum\limits_{s \in \{+,-\}} x^s_i .
\end{gather*}
As a consequence of the symmetries of the lattice the probability distribution of the terminal position after $k$ steps is invariant with respect to mirroring the change of the position $\Delta \vec{x}$ along any axis.
Hence the mean position of the walker while starting from the position $\vec{0}$ at arbitrary time $t$ is zero
\begin{equation}
\left\langle \vec{X}_t \right\rangle_{\delta_{\vec{0}}}
= \frac{a}{(2d)^{\frac{t}{\Delta t}}} \sum\limits_{m^\pm_0, \dots, m^\pm_{d-1} } \binom{\frac{t}{\Delta t}}{m^\pm_0 \dots m^\pm_{d-1}} \left( m^+_0 - m^-_0 , \dots , m^+_{d-1} - m^-_{d-1} \right)
= \vec{0} ,
\label{equ:average_position}
\end{equation}
where we also sum over all possible positions after $t/\Delta t$ steps,
hence the imposed conditions now are
\begin{gather*}
m^\pm_i \in \left\{ 0, 1, \dots , \frac{t}{\Delta t} \right\} , \\
\frac{t}{\Delta t} = \sum\limits_{i=0}^{d-1} \sum\limits_{s \in \{+, -\}} m^s_i .
\end{gather*}
In order to investigate the variance of the position at time $t$ it is convenient to introduce the \emph{moment-generating function}\index{function!moment-generating}
\[
M_t(\vec{\alpha}) = \left\langle {\mathrm e}^{\vec{\alpha} \cdot \vec{X}_t } \right\rangle_{\delta_{\vec{0}}} ,
\]
which in this particular case corresponds to
\[
M_t(\vec{\alpha}) = \left[ \frac{1}{d} \sum\limits_{i=0}^{d-1} \cosh (\alpha_i a) \right]^{\frac{t}{\Delta t}} .
\]
The various moments of position are obtained by taking derivatives with respect to $\vec{\alpha}$ at $\vec{0}$.
By taking the gradient with respect to $\vec{\alpha}$ at point $\vec{0}$ we the result for mean position already obtained by taking notion of the symmetries of the lattice \eqref{equ:average_position}.
We obtain the variance of position by applying the Laplace operator with respect to $\vec{\alpha}$ at point $\vec{0}$
\begin{equation}
\left\langle \vec{X}_t^2 \right\rangle_{\delta_\vec{0}} = \left. \Delta_\vec{\alpha} M_t(\vec{\alpha}) \right|_{\vec{\alpha}=\vec{0}} = \frac{a^2 t}{\Delta t} ,
\label{equ:random_walk_variance}
\end{equation}
which is proportional to the number of steps $t/{\Delta t}$.
The fact that the variance is proportional to the length of time interval $t$ is universal for all diffusive processes
and we will see later, that this property ensures that the generator of diffusion depends only on the probability density and its first and second derivative, see subsections \ref{ssec:underdamped_diff} and \ref{ssec:overdamped_diff}.
\subsubsection{Wiener process}\index{Wiener process}
Wiener process is a limiting case of random walker when we tend to go with the time step to zero,
however in order to preserve the finite variance in position of the process \eqref{equ:random_walk_variance}
we also need to scale the space in such a manner that the quantity $a^2/\Delta t$ is preserved.
What we will call the Wiener process is the limiting process $\Delta t \to 0$ of the random walker with the choice
\[
\frac{a^2}{\Delta t} = d .
\]
Such choice is also called \emph{diffusion limit}\index{diffusion limit}.
Moreover by taking the limit of the moment-generating function
\[
M_t^{\text{Wiener}}(\vec{\alpha}) = \lim_{\Delta t \to 0} \left[ 1 + \frac{\Delta t}{2} \vec{\alpha}^2 + \err[2]{\Delta t} \right]^{\frac{t}{\Delta t}} = \exp\left[ \frac{1}{2} \vec{\alpha}^2 t \right] ,
\]
we see that the distribution of position displacement $\vec{x}-\vec{x}_0$ is Gaussian with the zero mean and variance equal to $\sqrt{t}$
\begin{equation}
\cprob{\vec{X}_t = \vec{x} }{\vec{X}_0 = \vec{x}_0 } = \frac{1}{\left(2 \pi t\right)^{\frac{d}{2}}} \exp \left[- \frac{\left( \vec{x} - \vec{x}_0 \right)^2}{2t} \right] .
\label{equ:Wiener_distribution}
\end{equation}
One of the most important properties of the Wiener process is that the combination of independent Wiener processes is also a Wiener process
\begin{multline}
\cprob{\vec{X}_{t+s} = \vec{x} }{\vec{X}_0 = \vec{x}_0 } = \\
= \int\limits_{{\mathbb R}^d} {\mathrm d}^d \vec{y} \; \cprob{\vec{X}_{t+s} = \vec{x} }{\vec{X}_s = \vec{y} } \, \cprob{\vec{X}_s = \vec{y} }{\vec{X}_0 = \vec{x}_0 } .
\label{equ:Wiener_combination}
\end{multline}
Notice also that it also means that the taking of the mean value over all realizations of the Wiener process can be replaced by taking the mean value over realizations of several independent subsequent Wiener processes.
Consequently it is valid that for arbitrary times $0<t'<t$ the displacement $\vec{X}_t - \vec{X}_{t'}$ is Gaussian distributed random variable with zero mean value and variance equal to the length of the time interval
\begin{align}
\left\langle \vec{X}_t - \vec{X}_{t'} \right\rangle_{\mu_0} &= \vec{0} , &
\left\langle \left( \vec{X}_t - \vec{X}_{t'} \right)^2 \right\rangle_{\mu_0} &= t-t' .
\label{equ:Wiener_properties}
\end{align}
One can also find that another consequence of \eqref{equ:Wiener_combination} is that the covariance is proportional to lesser of the times $t'<t$,
\[
\left\langle \vec{X}_t \vec{X}_{t'} \right\rangle_{\delta_\vec{0}}
= \underbrace{ \left\langle \left( \vec{X}_t - \vec{X}_{t'} \right) \vec{X}_{t'} \right\rangle_{\delta_\vec{0}} }_{=0} + \left\langle \vec{X}_{t'} \vec{X}_{t'} \right\rangle_{\delta_\vec{0}}
= t' { \mathbb I} .
\]
Another characteristic of the Wiener process is that it is almost surely continuous,
this can be seen as a result of the property \eqref{equ:Wiener_combination} and the fact that the distribution \eqref{equ:Wiener_distribution} converges to delta function as $t \to 0^+$.
\subsection{It\^{o} versus Stratonovich calculus}
\label{ssec:stochastic_calculus}
In physics the diffusion can usually be handled on several level of description.
On the macroscopic scale we usually use the fluid dynamics approach with diffusion or Fokker-Planck equation.
However on the microscopic scale we describe the diffusion as a set of classical particles under the influence of the random force.
These two approaches have to be compatible.
In this subsection we provide such a link in the form of stochastic calculus.
The stochastic calculus is a tool developed to solve such a type of equations.
In the first part of this subsection we will provide some definitions and results of the It\^{o} calculus and later compare it with the approach of Stratonovich.
Some of the details and proofs of some statements can be found in the appendix \ref{ap:stochastic_calculus},
for even more details see the standard textbooks \cite{Oksendal2003,Evans2001}.
\subsubsection{It\^{o} calculus}
The basic idea behind the It\^{o} calculus is that we can handle the \emph{Gaussian white noise}\index{white noise} represented by the infinitesimal increment of the Wiener process ${\mathrm d} W_t$ as $\sqrt{{\mathrm d} t}$.
Or informally
\[
{\mathrm d} W_t^2 \approx {\mathrm d} t .
\]
To make the statement more precise we start with the definition of the It\^{o} stochastic integral and show some of its properties.
The basic idea for the stochastic integral comes from the generalization of the Riemann integral to stochastic functions and variables.
We define \emph{the It\^{o} stochastic integral}\index{stochastic integral!It\^{o}}\index{It\^{o} integral} of the vector field $\vec{f}(\vec{W},t)$ along the $d$-dimensional Wiener process $\vec{W}_t$ as
\begin{equation}
\int\limits_0^T \vec{f}(\vec{W}_t,t) \cdot {\mathrm d} \vec{W}_t = \lim\limits_{N \to \infty} \sum\limits_{i=0}^{N-1} \vec{f}(\vec{W}_{t_i},t_i) \cdot \left[\vec{W}_{t_{i+1}} - \vec{W}_{t_i}\right] ,
\label{def:Ito_integral}
\end{equation}
where the times $0=t_0 < t_1 < \dots < t_N = T$ correspond to the partition of the time interval $[0,T]$
and where the limit of the Riemann sum is taken over the decreasing length of the time interval
\[
\lim_{N \to \infty} \max_i \left| t_{i+1} - t_i \right| = 0 .
\]
The result of the integral is also a random variable with the following properties.
The integral is linear in the integrand
\[
\int\limits_0^T \left[ \alpha \vec{f}(\vec{W}_t,t) + \beta \vec{g}(\vec{W}_t,t) \right] \cdot {\mathrm d} \vec{W}_t
= \alpha \int\limits_0^T \vec{f}(\vec{W}_t,t) \cdot {\mathrm d} \vec{W}_t
+ \beta \int\limits_0^T \vec{g}(\vec{W}_t,t) \cdot {\mathrm d} \vec{W}_t ,
\]
which can be seen directly from definition \eqref{def:Ito_integral}.
The mean value of the integral is zero when starting from arbitrary initial state $\mu_0$ at time $0$
\begin{equation}
\left\langle \int\limits_0^T \vec{f}(\vec{W}_t,t) \cdot {\mathrm d} \vec{W}_t \right\rangle_{\mu_0}
= 0 .
\label{equ:Ito_zero_mean_value}
\end{equation}
It is also valid that the covariance of integrals corresponds to the time integral of the covariance
\begin{equation}
\left\langle \int\limits_0^T \vec{f}(\vec{W}_t,t) \cdot {\mathrm d} \vec{W}_t \int\limits_0^T \vec{g}(\vec{W}_t,t) \cdot {\mathrm d} \vec{W}_t \right\rangle_{\mu_0}
= \int\limits_0^T \left\langle \vec{f}(\vec{W}_t,t) \cdot \vec{g}(\vec{W}_t,t) \right\rangle_{\mu_0} \; {\mathrm d} t .
\label{equ:Ito_covariance}
\end{equation}
For further details see appendix \ref{ap:stochastic_calculus} sections \ref{sec:mean_value_Ito} and \ref{sec:covariance_Ito}.
Moreover the Riemann sum of the square of displacements alone converges to the length of the time interval $T$ \emph{almost surely}\index{almost surely}, i.e. the probability of the sum not being the length of the time interval converges to zero,
\[
\lim_{N \to \infty} \sum\limits_{i=0}^{N-1} \left( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \right)^2
= \lim_{N \to \infty} \sum\limits_{i=0}^{N-1} \left( t_{i+1} - t_i \right)
= T \qquad \text{a.s.}
\]
For the sketch of the proof see appendix \ref{ap:stochastic_calculus} section \ref{sec:Riemann_sum}.
Its generalized version can be used to connect the time integral of the random function over the time interval with the corresponding Riemann sum
\begin{equation}
\lim_{N \to \infty} \sum\limits_{i=0}^{N-1} f(\vec{W}_{t_i},t_i) \left( \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \right)^2
= \int\limits_0^T {\mathrm d} t \; f(\vec{W}_t,t) \qquad \text{a.s.}
\label{equ:pre_Ito_lemma}
\end{equation}
We apply the theory of It\^{o} stochastic integral to solve the stochastic differential equations.
Within the standard linear first-order differential equation theory we find the formal solution of the differential equation by associating it with an appropriate integral representation
\[
\vec{\nabla} f(\vec{x}) = \vec{g}(\vec{x}) \qquad \Longrightarrow \qquad {\mathrm d} f(\vec{x}) = \vec{g}(\vec{x}) \cdot {\mathrm d} \vec{x} \qquad \Longrightarrow \qquad f(\vec{x}) = \int {\mathrm d} \vec{x} \cdot \vec{g}(\vec{x}) .
\]
In a similar fashion the formal solution $Y(\vec{W}_T,T)$ of \emph{the stochastic differential equation}\index{stochastic differential equation} in the form of It\^{o} total differential
\begin{equation}
{\mathrm d} Y(\vec{W}_t,t) = f(\vec{W}_t,t) \; {\mathrm d} t + \vec{g}(\vec{W}_t,t) \cdot {\mathrm d} \vec{W}_t
\label{def:stochastic_differential_equation}
\end{equation}
is given by its integral representation
\[
Y(\vec{W}_T,T) = Y(\vec{W}_0,0) + \int\limits_0^T {\mathrm d} t \; f(\vec{W}_t,t) + \int\limits_0^T {\mathrm d} \vec{W}_t \cdot \vec{g}(\vec{W}_t,t) .
\]
Notice that consequently the solution $Y(\vec{W}_T,T)$ of stochastic differential equation is \emph{the random function}\index{random function} of the Wiener process.
Up to now we were developing the theory in order to solve a given stochastic differential equation.
One can also be interested in how to determine the stochastic differential equation knowing the solution.
The answer is given in the form of the \emph{It\^{o} lemma}\index{It\^{o} lemma} which states that the corresponding differential equation to the solution $Y(\vec{W}_t,t)$ with the initial condition $Y(\vec{W}_0,0)$ is given by
\begin{equation}
{\mathrm d} Y(\vec{W}_t,t) = \left[ \partial_t Y(\vec{W}_t,t) + \frac{1}{2} \left. \Delta_\vec{x} Y(\vec{x},t) \right|_{\vec{x} = \vec{W}_t} \right] \; {\mathrm d} t + \left. \vec{\nabla}_\vec{x} Y(\vec{x},t) \right|_{\vec{x} = \vec{W}_t} \cdot {\mathrm d} \vec{W}_t .
\label{equ:total_differential_Ito}
\end{equation}
The It\^{o} lemma is a direct consequence of the expansion of the solution in both the Wiener process $\vec{W}_t$ and the explicit time dependence while applying \eqref{equ:pre_Ito_lemma}.
For further details see again appendix \ref{ap:stochastic_calculus} section \ref{sec:Ito_lemma}.
Notice that we usually do not describe physical systems with the differential equation in the form of the total differential \eqref{def:stochastic_differential_equation},
instead we are describing the system under the influence of the random force $\xi_t$,
e.g. \emph{Langevin equation}\index{Langevin equation} is usually given in the form
\[
m \ddot{\vec{x}}_t = \vec{f}(\vec{x}_t) + \vec{\xi}_t .
\]
If the random force corresponds to \emph{the white Gaussian noise}\index{white noise} the connection is quite straightforward,
we associate the random force with the Wiener process by $\vec{\xi}_t \; {\mathrm d} t = {\mathrm d} \vec{W}_t$.
\subsubsection{Stratonovich calculus}
The \emph{Stratonovich stochastic integral}\index{stochastic integral!Stratonovich}\index{Stratonovich integral} is defined as
\begin{equation}
\int\limits_0^T \vec{f}(\vec{W}_t,t) \circ {\mathrm d} \vec{W}_t = \lim\limits_{N \to \infty} \sum\limits_{i=0}^{N-1} \vec{f}(\vec{W}_{\tau_i},\tau_i) \cdot \left[\vec{W}_{t_{i+1}} - \vec{W}_{t_i}\right] ,
\label{def:Stratonovich_integral}
\end{equation}
where $t_i$ again corresponds to the partition times of the interval $[0,T]$ and $\tau_i$ is a midpoint of the corresponding interval, $\tau_i = ( t_{i+1} + t_i )/2$.
We can see that the difference between the Stratonovich calculus and the It\^{o} calculus in the choice of the point, where we evaluate the function.
In case of It\^{o} integral we evaluate the function at the beginning of the interval while in case of Stratonovich integral we evaluate it in the middle.
From there also follows the different behaviour of It\^{o} and Stratonovich integral with respect to \emph{time inversion}\index{time inversion}, see below.
To show this difference, we define the time reversal of the particular realization of the Wiener process as
\[
\vec{W}^\Theta_t = \vec{W}_{T-t} .
\]
Notice that such definition of the time reversal also corresponds to the substitution $s = T-t$.
Also notice that the time reversal of the Wiener process is also a Wiener process with the same properties,
which is due to the fact that the probability of the random walker going along the given path fort and back is the same.
Then if we take the It\^{o} integral over the particular realization of the Wiener process $\vec{W}_t$ and try to represent it with its time reversal $\vec{W}^\Theta_t$ we obtain
\begin{multline*}
\int\limits_0^T \vec{f}(\vec{W}_t) \cdot \vec{W}_t = \sum\limits_{i=0}^{N-1} \vec{f}(\vec{W}_{t_i}) \cdot \left[ \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \right] = \\
= - \sum\limits_{i=0}^{N-1} \vec{f}(\vec{W}^\Theta_{T-t_i}) \left[ \vec{W}^\Theta_{T-t_i} - \vec{W}^\Theta_{T-t_{i+1}} \right] = \\
= - \sum\limits_{j=0}^{N-1} \vec{f}(\vec{W}^\Theta_{s_{j+1}}) \left[ \vec{W}^\Theta_{s_{j+1}} - \vec{W}^\Theta_{s_j} \right] ,
\end{multline*}
where we have already changed the notation of the partition $s_i = T - t_{N-i} $.
In the result we have obtained the increment of the Wiener process multiplied by the function evaluated at the terminal time $s_{j+1}$,
hence the function and the increment of the Wiener process are no longer independent.
By substitution $s=T-t$ we obtain a new object which fundamentally differs from the definition of the It\^{o} integral,
which is usually called the \emph{backward It\^{o} integral}\index{stochastic integral!backward It\{o}}\index{backward It\^{o} integral}
\begin{multline*}
\int\limits_0^T \vec{f}(\vec{W}_t) \cdot {\mathrm d} \vec{W}_t = \sum\limits_{i=0}^{N-1} \vec{f}(\vec{W}_{t_i}) \cdot \left[ \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \right] \xrightarrow{s=T-t} \\
\xrightarrow{s=T-t} - \sum\limits_{j=0}^{N-1} \vec{f}(\vec{W}^\Theta_{t_{j+1}}) \cdot \left[ \vec{W}^\Theta_{t_{j+1}} - \vec{W}^\Theta_{t_j} \right]
= - \int\limits_0^T \vec{f}(\vec{W}^\Theta_t) \cdot {\mathrm d}_{\text{b}}\vec{W}^\Theta_t .
\end{multline*}
On the contrary if we proceed along the same lines with the Stratonovich integral we obtain again a Stratonovich integral
\begin{multline}
\int\limits_0^T \vec{f}(\vec{W}_t) \circ {\mathrm d} \vec{W}_t = \sum\limits_{i=0}^{N-1} \vec{f}(\vec{W}_{\tau_i}) \cdot \left[ \vec{W}_{t_{i+1}} - \vec{W}_{t_i} \right] \xrightarrow{s=T-t} \\
\xrightarrow{s=T-t} - \sum\limits_{j=0}^{N-1} \vec{f}(\vec{W}^\Theta_{\tau_j}) \cdot \left[ \vec{W}^\Theta_{t_{j+1}} - \vec{W}^\Theta_{t_j} \right]
= - \int\limits_0^T \vec{f}(\vec{W}^\Theta_t) \circ {\mathrm d} \vec{W}^\Theta_t ,
\label{equ:antisymmetry_Stratonovich}
\end{multline}
because the midpoint $\tau_i$ of the interval is preserved by the substitution $s=T-t$.
Although the It\^{o} and Stratonovich integrals behave differently under the time inversion, there is a way how to transform one to another
\begin{equation}
\int\limits_0^T \vec{f}(\vec{W}_t,t) \circ {\mathrm d} \vec{W}_t
= \int\limits_0^T \vec{f}(\vec{W}_t,t) \cdot {\mathrm d} \vec{W}_t
+ \frac{1}{2} \int\limits_0^T \left. \vec{\nabla}_{\vec{x}} \cdot \vec{f}(\vec{x},t) \right|_{\vec{x}=\vec{W}_t} \; {\mathrm d} t ,
\label{equ:relation_Ito_Stratonovich}
\end{equation}
where $\vec{\nabla} \cdot \vec{f}$ denotes the divergence of $\vec{f}$.
This relation can also be used to express the total differential of the random process \eqref{equ:total_differential_Ito}
\begin{equation}
{\mathrm d} Y(\vec{W}_t,t) = \partial_t Y(\vec{W}_t,t) \; {\mathrm d} t
+ \left. \vec{\nabla}_\vec{x} Y(\vec{x},t) \right|_{\vec{x}=\vec{W}_t} \circ {\mathrm d} \vec{W}_t
\label{equ:total_differential_Stratonovich}
\end{equation}
in terms of the Stratonovich calculus.
From this point of view the Stratonovich approach can be considered as a more natural.
\subsection{Underdamped diffusion}
\label{ssec:underdamped_diff}
\index{diffusion!underdamped}
Although underdamped diffusion describes a large class of physical systems,
the typical example of underdamped diffusion describes a heavy particle in an environment consisting of lighter particles, e.g. the droplet of oil in the water.
In order to effectively describe the movement of the heavy particle in such an environment it's impractical or almost impossible to keep the track of all light particles.
Let us assume that there are no long range interaction, hence light particles interact with the heavy particle mostly by collisions.
Collisions mainly cause two effects, the friction of the environment proportional to the velocity of the heavy particle characterized by a coefficient $\gamma(\vec{q})$,
which in general depends on the position of the heavy particle $\vec{q}$,
and the random force acting on the heavy particle described by Wiener process with an amplitude of $\sqrt{2\gamma(\vec{q_t})/\beta}$.
The movement of the heavy particle with mass $m$ then can be effectively described through set of stochastic differential equations
\begin{align}
{\mathrm d} \vec{q}_t &= \frac{\vec{p}_t}{m} \; {\mathrm d} t , \label{def:position_underdamped} \\
{\mathrm d} \vec{p}_t &= \left[ \vec{F}(\vec{q}_t,\vec{p}_t) - \frac{\gamma(\vec{q}_t)}{m} \vec{p}_t \right] \; {\mathrm d} t + \sqrt{\frac{2\gamma(\vec{q}_t)}{\beta}} \circ {\mathrm d} \vec{W}_t , \label{def:momentum_underdamped}
\end{align}
where $\vec{F}(\vec{q},\vec{p})$ describes the deterministic part of the force,
which in general depends on the position $\vec{q}$ and momentum $\vec{p}$ of the heavy particle
and $\beta$ is the inverse temperature of the environment.
Directly from definitions \eqref{def:position_underdamped} and \eqref{def:momentum_underdamped} we can see that the position explicitly depends only on time $\vec{q}_t(t)$ while the momentum also explicitly depends on the realization of the Wiener process $\vec{p}_t(\vec{W}_t,t)$.
Notice also that in most cases there is no difference between the It\^{o} or Stratonovich approaches in \eqref{def:momentum_underdamped},
because the friction coefficient $\gamma(\vec{q})$ does not depend on the momentum.
However it makes difference in case of the underdamped diffusion as we will see later.
Since the It\^{o} approach is easier to handle we will use it in the rest of this subsection, hence
\begin{equation}
{\mathrm d} \vec{p}_t = \left[ \vec{F}(\vec{q}_t,\vec{p}_t) - \frac{\gamma(\vec{q}_t)}{m} \vec{p}_t \right] \; {\mathrm d} t + \sqrt{\frac{2 \gamma(\vec{q}_t) }{\beta}} \cdot {\mathrm d} \vec{W}_t .
\label{equ:momentum_underdamped}
\end{equation}
The solution of the stochastic differential equation is also a random function, hence one can be interested in the probability distribution of positions and momentums at a particular time
\begin{equation}
\mu_t(\vec{q},\vec{p}) = \left\langle \delta(\vec{q}-\vec{q}_t) \, \delta(\vec{p}-\vec{p}_t) \right\rangle_{\mu_0} ,
\label{def:probability_distribution_underdamped}
\end{equation}
where we take the mean value over all possible realizations of the random process $(\vec{q}_t,\vec{p}_t)$ with the initial condition given by $\mu_0(\vec{q},\vec{p})$.
This approach proves to be useful in cases when we have set of experiments which we need to evaluate or if the initial position or momentum is indeterminate.
It can be shown, see below, that the autonomous time evolution of the probability density is governed by
\begin{multline}
\partial_t \mu_t(\vec{q},\vec{p})
= - \frac{\vec{p}}{m} \cdot \vec{\nabla}_{\vec{q}} \mu_t(\vec{q},\vec{p}) - \\
- \vec{\nabla}_{\vec{p}} \cdot \left[ \left( \vec{F}(\vec{q},\vec{p}) - \frac{\gamma(\vec{q})}{m} \vec{p} \right) \mu_t(\vec{q},\vec{p})
- \frac{\gamma(\vec{q})}{\beta} \vec{\nabla}_\vec{p} \mu_t(\vec{q},\vec{p}) \right] .
\label{equ:time_evolution_underdamped}
\end{multline}
We will use the proof of this statement to provide an example of the usage of previously introduced stochastic calculus.
We start by expanding the definition of the distribution \eqref{def:probability_distribution_underdamped} using the It\^{o} lemma \eqref{equ:total_differential_Ito}
\begin{multline*}
{\mathrm d} \mu_t(\vec{q},\vec{p})
= - {\mathrm d} t \; \left\langle \partial_t \vec{q}_t(t) \cdot \vec{\nabla}_\vec{q} \delta(\vec{q}-\vec{q}_t(t)) \, \delta(\vec{p}-\vec{p}_t(\vec{W}_t,t)) \right\rangle_{\mu_0} - \\
- {\mathrm d} t \; \left\langle \partial_t \vec{p}_t(\vec{W}_t,t) \cdot \vec{\nabla}_\vec{p} \delta(\vec{p}-\vec{p}_t(\vec{W}_t,t)) \, \delta(\vec{q}-\vec{q}_t(t)) \right\rangle_{\mu_0} + \\
- \left\langle {\mathrm d} \vec{W}_t \cdot \left. \nabla_{\vec{x}} \vec{p}_t(\vec{x},t) \right|_{\vec{x}=\vec{W}_t} \cdot \vec{\nabla}_\vec{p} \delta(\vec{p}-\vec{p}_t(\vec{W}_t,t)) \, \delta(\vec{q}-\vec{q}_t(t)) \right\rangle_{\mu_0} + \\
+ \frac{1}{2} {\mathrm d} t \; \tr \left\langle \left. \nabla_\vec{x} \vec{p}_t(\vec{x},t) \right|_{\vec{x}=\vec{W}_t} \cdot \vec{\nabla}^2_\vec{p} \delta(\vec{p}-\vec{p}_t(\vec{W}_t,t)) \cdot \left. \nabla_\vec{x} \vec{p}_t(\vec{x},t) \right|_{\vec{x}=\vec{W}_t} \, \delta(\vec{q}-\vec{q}_t(t)) \right\rangle_{\mu_0} ,
\end{multline*}
where we have explicitly denoted the dependencies on the time and the realization of the Wiener process.
While the dependence on the $\vec{q}$ and $\vec{p}$ is only in the argument of delta functions, we can pull the derivatives with respect to them in front of the mean values.
We also insert the terms from \eqref{def:position_underdamped} and \eqref{equ:momentum_underdamped} and obtain
\begin{multline*}
{\mathrm d} \mu_t(\vec{q},\vec{p})
= - {\mathrm d} t \; \vec{\nabla}_\vec{q} \cdot \left\langle \frac{\vec{p}_t}{m} \, \delta(\vec{q}-\vec{q}_t) \, \delta(\vec{p}-\vec{p}_t) \right\rangle_{\mu_0} - \\
- \vec{\nabla}_\vec{p} \cdot \left\langle \left[ \left( \vec{F}(\vec{q}_t,\vec{p}_t) - \frac{\gamma(\vec{q}_t) \vec{p}_t }{m} \right) \; {\mathrm d} t + \sqrt{\frac{2\gamma(\vec{q}_t)}{\beta}} \; {\mathrm d} \vec{W}_t \right] \delta(\vec{q}-\vec{q}_t) \, \delta(\vec{p}-\vec{p}_t) \right\rangle_{\mu_0} + \\
+ {\mathrm d} t \; \Delta_\vec{p} \left\langle \frac{\gamma(\vec{q}_t)}{\beta} \, \delta(\vec{q}-\vec{q}_t) \, \delta(\vec{p}-\vec{p}_t) \right\rangle_{\mu_0} .
\end{multline*}
Now using the fact that the mean value of It\^{o} integral is zero \eqref{equ:Ito_zero_mean_value} along with
\[
\left\langle f(\vec{q}_t,\vec{p}_t) \, \delta(\vec{q}-\vec{q}_t) \, \delta(\vec{p}-\vec{p}_t) \right\rangle_{\mu_0}
= f(\vec{q},\vec{p}) \left\langle \delta(\vec{q}-\vec{q}_t) \, \delta(\vec{p}-\vec{p}_t) \right\rangle_{\mu_0}
= f(\vec{q},\vec{p}) \mu_t(\vec{q},\vec{p})
\]
concludes the proof.
\subsubsection{Kolmogorov generators}
At the beginning of the section \ref{sec:Markov_processes} we have introduced Kolmogorov generators as an effective description of the time evolution of the system.
To provide the same level of description also for underdamped diffusion we use the time evolution equation \eqref{equ:time_evolution_underdamped} to define \emph{the forward Kolmogorov generator}\index{Kolmogorov generator!forward!underdamped diffusion} \eqref{equ:Markov_stochastic_time_evolution}
\begin{multline}
{\mathcal L}^* \left[ \mu \right] (\vec{q},\vec{p})
= - \frac{\vec{p}}{m} \cdot \vec{\nabla}_{\vec{q}} \mu(\vec{q},\vec{p}) - \\
- \vec{\nabla}_{\vec{p}} \cdot \left[ \left( \vec{F}(\vec{q},\vec{p}) - \frac{\gamma(\vec{q})}{m} \vec{p} \right) \mu(\vec{q},\vec{p})
- \frac{\gamma(\vec{q})}{\beta} \vec{\nabla}_\vec{p} \mu(\vec{q},\vec{p}) \right] .
\label{equ:forward_generator_underdamped}
\end{multline}
It is easy to check that the normalization condition on forward Kolmogorov generator \eqref{equ:normalization_condition} is valid under the assumption that the probability distribution vanish at the boundary $\|\vec{q}\| \to \infty$ or $\|\vec{p}\| \to \infty$.
From \eqref{def:backward_generator} we can also obtain \emph{the backward Kolmogorov generator}\index{Kolmogorov generator!backward!underdamped diffusion} for underdamped diffusion
\begin{multline}
{\mathcal L} \left[ A \right] (\vec{q},\vec{p})
= \frac{\vec{p}}{m} \cdot \vec{\nabla}_{\vec{q}} A(\vec{q},\vec{p}) + \\
+ \left( \vec{F}(\vec{q},\vec{p}) - \frac{\gamma(\vec{q})}{m} \vec{p} \right) \cdot \vec{\nabla}_{\vec{p}} A(\vec{q},\vec{p})
+ \frac{\gamma(\vec{q})}{\beta} \Delta_\vec{p} A(\vec{q},\vec{p}) .
\label{equ:backward_generator_underdamped}
\end{multline}
If we examine the forward Kolmogorov generator,
we can see that it has a structure of the generalized divergence,
hence the time evolution equation is similar continuity equation well known from the hydrodynamics.
If we define a generalized probability current
\[
\widetilde{\vec{\jmath}}(\vec{q},\vec{p}) = \begin{bmatrix}
\frac{\vec{p}}{m} \mu(\vec{q},\vec{p}) \\
\left( \vec{F}(\vec{q},\vec{p}) - \frac{\gamma(\vec{q})}{m} \vec{p} \right) \mu(\vec{q},\vec{p}) - \frac{\gamma(\vec{q})}{\beta} \vec{\nabla}_\vec{p} \mu(\vec{q},\vec{p})
\end{bmatrix}
\]
and a generalized gradient
\[
\widetilde{\vec{\nabla}} = \begin{bmatrix}
\vec{\nabla}_\vec{q} \\
\vec{\nabla}_\vec{p}
\end{bmatrix} ,
\]
then the time evolution equation \eqref{equ:time_evolution_underdamped} can be written as
\[
\partial_t \mu_t (\vec{q},\vec{p}) + \widetilde{\vec{\nabla}} \cdot \widetilde{\vec{\jmath}} (\vec{q},\vec{p}) = 0 .
\]
\subsubsection{Equilibrium}
In equilibrium the force $\vec{F}(\vec{q},\vec{p})$ acting on particle is given by potential $U(\vec{q})$, which depends only on the position $\vec{q}$ of the particle.
The stationary equilibrium distribution $\rho(\vec{q},\vec{p})$ is then given by \emph{the Maxwell-Boltzmann distribution}\index{Maxwell-Boltzmann distribution}
\[
\rho(\vec{q},\vec{p}) = \frac{1}{Z} \exp \left[ - \beta \left( \frac{\vec{p}^2}{2m} + U(\vec{q}) \right) \right]
\]
at the inverse temperature $\beta$.
As a consequence the generalized probabilistic current is given by
\[
\widetilde{\vec{\jmath}} = \begin{bmatrix}
\frac{\vec{p}}{m} \\
- \vec{\nabla}_\vec{q} U(\vec{q})
\end{bmatrix}
\rho(\vec{q},\vec{p}) .
\]
\subsection{Overdamped diffusion}
\label{ssec:overdamped_diff}
\index{diffusion!overdamped}
The overdamped diffusion is the limiting case of the underdamped diffusion, when the relaxation time for the velocities is much shorter then the relaxation time for positions,
which we will show in chapter \ref{chapter:slow-fast_coupling}.
Another possible approach is by taking the diffusion limit of the Markov jump process on the lattice with non-uniform jump probabilities along the edges \cite{Itzykson_Drouffe:Statistical_field_theory} as in the subsection \ref{ssec:diffusion_limit}.
For this purpose, we define \emph{the overdamped diffusion}\index{diffusion!overdamped} by the total differential in the Stratonovich form
\begin{equation}
{\mathrm d} \vec{q}_t = \left[ \vec{\chi}(\vec{q}_t) \cdot \vec{F}(\vec{q}_t) + \frac{1}{2} \vec{\nabla}_\vec{q} \cdot \vec{D}(\vec{q_t}) \right] \; {\mathrm d} t + \sqrt{ 2 \vec{D}(\vec{q}_t) } \circ {\mathrm d} \vec{W}_t,
\label{def:position_overdamped}
\end{equation}
which can be rewritten to the It\^{o} form by using the chain rule along with \eqref{equ:relation_Ito_Stratonovich}
\begin{equation}
{\mathrm d} \vec{q}_t = \left[ \vec{\chi}(\vec{q}_t) \cdot \vec{F}(\vec{q}_t) + \vec{\nabla}_\vec{q} \cdot \vec{D}(\vec{q_t}) \right] \; {\mathrm d} t + \sqrt{ 2 \vec{D}(\vec{q}_t) } \cdot {\mathrm d} \vec{W}_t,
\label{equ:position_overdamped}
\end{equation}
where ${\mathrm d} \vec{W}_t$ is again a multidimensional white noise
and mobility matrix $\vec{\chi}(\vec{q})$ is related to the diffusion matrix $\vec{D}(\vec{q})$ by \emph{the Einstein relation}\index{Einstein relation} $\vec{\chi}(\vec{q}) =\beta \vec{D}(\vec{q})$.
We can again derive the autonomous time evolution equation for probability density in the same manner as in the subsection \ref{ssec:underdamped_diff} in case of the underdamped diffusion directly from the stochastic differential equation \eqref{equ:position_overdamped},
which yields to the Fokker-Planck equation\index{Fokker-Planck equation}
\begin{equation}
\partial_t \mu_t(\vec{q}) =
- \vec{\nabla}_{\vec{q}} \cdot \left[ \vec{\chi}(\vec{q}) \cdot \vec{F}(\vec{q}) \; \mu_t(\vec{q}) - \vec{D}(\vec{q}) \cdot \vec{\nabla}_{\vec{q}} \mu_t(\vec{q}) \right] .
\label{equ:time_evolution_overdamped}
\end{equation}
\subsubsection{Kolmogorov generators}
We can again associate the forward Kolmogorov generator with the operator on the right side of the Fokker-Planck equation \eqref{equ:time_evolution_overdamped}.
Hence we define \emph{the forward Kolmogorov generator}\index{Kolmogorov generator!forward!overdamped diffusion} for overdamped diffusion as
\[
{\mathcal L}^* \left[ \mu \right] (\vec{q}) =
- \vec{\nabla}_{\vec{q}} \cdot \left[ \vec{\chi}(\vec{q}) \cdot \vec{F}(\vec{q}) \; \mu(\vec{q}) - \vec{D}(\vec{q}) \cdot \vec{\nabla}_{\vec{q}} \mu(\vec{q}) \right] .
\]
In analogy with the underdamped diffusion we also define \emph{the backward Kolmogorov generator}\index{Kolmogorov generator!backward!overdamped diffusion} for overdamped diffusion
\begin{equation}
{\mathcal L} \left[ A \right] (\vec{q}) =
\vec{F}(\vec{q}) \cdot \vec{\chi}(\vec{q}) \cdot \vec{\nabla}_{\vec{q}} A(\vec{q}) + \vec{\nabla}_{\vec{q}} \cdot \left[ \vec{D}(\vec{q}) \cdot \vec{\nabla}_{\vec{q}} A(\vec{q}) \right] .
\label{equ:bacward_generator_overdamped}
\end{equation}
In case of the overdamped diffusion the structure of the continuity equation in case of the time evolution equation is even more pronounced than in the underdamped case.
Let us define the probability current
\begin{equation}
\vec{j}(\vec{q}) = \left[ \vec{\chi}(\vec{q}) \cdot \vec{F}(\vec{q}) \; \mu(\vec{q}) - \vec{D}(\vec{q}) \cdot \vec{\nabla}_{\vec{q}} \mu(\vec{q}) \right] .
\label{def:current_overdamped}
\end{equation}
Then the Fokker-Planck equation \eqref{equ:time_evolution_overdamped} corresponds to
\[
\partial_t \mu_t(\vec{q}) + \vec{\nabla}_\vec{q} \cdot \vec{j}(\vec{q}) = 0 .
\]
\subsubsection{Equilibrium}
In equilibrium the force $\vec{F}(\vec{q})$ is again determined by the potential $U(\vec{q})$,
hence the stationary distribution in equilibrium again corresponds to \emph{the Boltzmann distribution}\index{Boltzmann distribution}
\[
\rho(\vec{q}) = \frac{1}{Z} \exp \left[ - \beta U(\vec{q}) \right]
\]
at the inverse temperature $\beta$.
Moreover one can easily check that the probability current \eqref{def:current_overdamped} in equilibrium is zero.
\chapter{Stochastic non-equilibrium thermodynamics}
\label{chapter:non-equilibrium_thermodynamics}
At the beginning of this chapter we will briefly discuss consequences of the breaking of the global detailed balance such as the non-uniques of the energy function on the system or ``gauge'' freedom.
Then we will review some of the approaches to the description of non-equilibrium stochastic systems, namely the approach by Hatano-Sasa \cite{Hatano2001} and Komatsu-Nakagawa-Sasa-Tasaki \cite{},
which will be followed by the introduction of our approach to the problem of non-equilibrium stochastic system in case the system is driven by non-potential forces or is attached to multiple thermal baths.
In that section we will also show our first results \cite{result1,result2}.
At the end we will demonstrate the results on several models.
\section{Global detailed balance breaking}
\label{sec:broken_global_detailed_balance}
In the previous chapter \ref{chapter:equilibrium_stat_phys} in the section \ref{sec:detailed_balance} we have introduced the concepts of the global \eqref{def:global_detailed_balance} and local \eqref{def:local_detailed_balance} detailed balance.
While the global detailed balance condition connects the probability of the path compared to the probability of its time reversal with probabilities of configurations in the steady state,
the local detailed balance condition relates the ratio of the probability of the path $\omega$ with the probability of its time reversal to the entropy production in thermal baths upon which the system is coupled to.
As a consequence the global detailed balance along with local detailed balance condition ensures that there are no macroscopic currents in the system in the steady state,
however this is no longer valid out of equilibrium.
As we assume that the underlying microscopic time evolution for the system together with the thermal baths is still time reversible, which is not necessarily valid for the system itself, thus ensuring the validity of local detailed balance condition and hence the global detailed balance condition has to be broken.
There are several natural ways how to drive system out of equilibrium.
The first one is the introduction of time-dependent driving to the system,
which causes that even in the steady state there is in general a current of energy of particles through the system as the system constantly changes its state to adjust itself to ever-changing external conditions.
In these systems the notion of stationarity is quite different from other cases, the steady state if exists is such periodical state of the system to which the system converge from arbitrary initial state,
if we assume the ergodicity of the dynamics.
While the steady state is periodical it is not necessary for the steady state to be also invariant with respect to the time reversal.
We will discuss a class of these systems in separately in chapter \ref{chapter:periodically_driven_systems}.
The second way is by attaching the system to multiple thermal or particle baths.
In general each thermal bath can be at different temperature and hence in the steady state we obtain a steady heat current through the system.
From mesoscopic point of view each possible transition between two configurations of the system can be associated with the action of any of the thermal baths.
In that case the full dynamics given by the collective action of all thermal baths together,
so in general the steady state is given by the balance of the probability currents,
hence global detailed balance is not necessarily valid for each particular transition associated with a single thermal bath.
The third way is introduction of non-potential force to the system, which may represent the effective mechanical action of surrounding media, e.g. rotational forces acting on colloidal particle in a suspension.
In such case there is alway work being done on the system, which has to be dissipated out of the system to the thermal bath.
From mesoscopic point of view it corresponds to the constant energy current through the system in the steady state.
As in the previous case the steady state is again determined by collective action of the thermal bath and with the non-potential force, which also means that there are non-zero probability currents present in the system.
Hence by the same reasoning as in previous case the global detailed balance is no longer valid.
Moreover by introducing the non-potential forces the energy function is also no longer unique.
In the situation, when there were no non-potential forces acting on the system, it was natural to associate the energy of each particular configuration $x$ with the potential of the total force.
However in the situation when there are non-potential forces acting on the system,
we have the freedom in dividing of the total force acting on the system to the non-potential and the potential components,
an arbitrary part of the potential force can be included in the non-potential force while the system's response, which is determined by the total force applied to the system, remains the same.
This \emph{``gauge'' invariance}\index{gauge invariance}
\begin{equation}
\begin{aligned}
\vec{F}_\text{pot} = - \nabla U \quad & \longrightarrow \quad \widetilde{\vec{F}}_\text{pot} = - \nabla U - \nabla V \\
\vec{F}_\text{nonpot} \quad & \longrightarrow \quad \widetilde{\vec{F}}_\text{nonpot} = \vec{F}_\text{nonpot} + \nabla V
\end{aligned}
\label{equ:gauge_symmetry}
\end{equation}
tells us that the potential, which we associated with the energy, is no longer unique and hence the energy function itself is also no longer unique.
\section{Stochastic energetics}
In chapter \ref{chapter:equilibrium_stat_phys} in section \ref{sec:first_law} we have introduced the concept of stochastic energetics \eqref{def:first_law_path}.
We have seen that the concept of stochastic energetics is microscopic version of the first law of thermodynamics.
Taking the stochastic energetics given we have obtained the first law of thermodynamics \eqref{equ:first_law_eq} on the level of mean values of the total work and the heat follows.
Moreover we have been able to fully characterize the mean values of the total work \eqref{equ:total_local_work_relation} and the heat \eqref{equ:total_local_heat_relation} by the local power \eqref{def:local_power} and the local heat production \eqref{def:local_heat_production}.
All of these relations are direct consequences of the dynamic being Markovian and the validity of stochastic energetics,
which is reasonable to assume are to be valid also out of equilibrium,
then we can consider them to be valid even out of equilibrium.
The only essential difference there is in the particular definition of the work and heat along the given path and consequently in expressions for local power and local heat production.
In genuine non-equilibrium situation there is in general an additional contribution to the local power from non-potential forces or from the explicit dependence of the potential on the time
\begin{equation}
w^{\text{tot}}_{\alpha(t)} = \dot\alpha(t) \cdot \left. \nabla_\alpha E_\alpha(x) \right|_{\alpha(t)} + w_{\alpha(t)}^{\text{nonp}}(x) .
\label{equ:total_local_power}
\end{equation}
In the situation where the system is connected to multiple thermal baths the situation become more complicated,
because we need to track down the heat dissipated to each thermal bath respectively $Q_{\alpha(t)}^i(\omega)$,
hence the total local heat production in all baths is the sum of all contributions over all baths
\[
q^{\text{tot}}_{\alpha(t)}(x) = \sum_i q^i_{\alpha(t)}(x) .
\]
Notice also that the local heat production for a single bath in case the system is attached to multiple thermal baths does not necessarily correspond to the local heat production when the bath is attached alone.
\section{Quasistatic processes}
\label{sec:quasistatic_processes_noneq}
In previous sections we have discussed several approaches to non-equilibrium phenomenona in small systems.
In our attempt to approach the issue we have studied the quasistatic processes connecting various steady states,
which can be considered as a natural extension of equilibrium processes.
In equilibrium the quasistatic or in other words equilibrium processes are essential in formulation of thermodynamics \cite{Callen1985},
e.g. the first law of thermodynamics states that the work done on the system along an arbitrary adiabatic equilibrium process corresponds to the change of the internal energy.
Also one of the equivalent variant of the second law of thermodynamics connects the heat production with the entropy production, see chapter \ref{chapter:equilibrium_stat_phys} for further details.
The fact that the work along any close adiabatic or isothermal trajectory is zero tells us about the existence of thermodynamical potentials,
in those cases the internal energy $\mathcal U$ and the free energy $\mathcal F$.
Similarly the fact that the heat along the closed isochoric trajectory is zero gives us enthalpy $\mathcal H$.
The basic question which we ask is, to which extent some of these properties are still valid out of the equilibrium?
Our approach is very similar to the approach of Komatsu-Nakagawa-Sasa-Tasaki,
the main difference is that we rather focus on the energetic of quasistatic processes on the level of mean values and on the properties of generalized response functions, namely the generalized heat capacity,
while their main aim was to find the non-equilibrium version of Clausius theorem.
Komatsu-Nakagawa-Sasa-Tasaki addressed the problem of the quasistatic limit by approximating the quasistatic process by a step process.
In this section we will show in full a newly developed approach to the quasistatic limit based on an extension of the standard adiabatic theorem.
which was already partially shown in the section \ref{sec:quasistatic_processes_eq}, where we have addressed the quasistatic processes in equilibrium.
In the quasistatic limit the mean values of path observables such as the total heat or work will prove to naturally decompose to the diverging ``housekeeping'' component and the finite ``reversible'' component.
Where the ``reversible'' component will be responsible for the energetics of the system, and thus enables us to define generalized response functions, namely generalized heat capacity.
We will briefly address the problem of general non-existence of the Clausius relation for the ``reversible'' components and its consequences.
We also discuss the consequences of the broken detailed balance condition such as non-uniques of the internal energy.
These results will be illustrated on the series of examples.
Most of these results can be found in \cite{result1,result2}.
\subsection{Probability distribution}
As the first step towards thermodynamics based on quasistatic processes we study the behaviour of the state of the system undergoing the quasistatic process.
To simplify the situation we will assume that further on the only explicit time dependence lies in external parameters $\alpha$,
hence the global detailed balance is broken either by attaching the system to multiple thermal bath or by the application of the non-potential force.
In equilibrium the external parameters are usually quantities like volume, temperature or pressure, however we are not limited only to those.
We only assume that the set of external parameters uniquely determines the steady state of the system $\rho_\alpha$ for all possible values of $\alpha$.
The quasistatic process is the limiting process where the speed of the changes of external parameters is scaled to zero along with the time duration going to infinity \eqref{equ:time_rescaling}
\begin{align*}
\alpha(t) &\to \alpha(\epsilon t) &
t \in [0,T] &\to t \in \left[ 0 , \frac{T}{\epsilon} \right] .
\end{align*}
The time evolution \eqref{equ:Markov_stochastic_time_evolution} of the state $\mu_t$ along the rescaled trajectory is given by
\begin{equation}
\partial_t \mu_{t;\alpha(\epsilon t)} (x) = {\mathcal L}^*_{\alpha(\epsilon t)} [\mu_{t;\alpha(\epsilon t)}] (x) ,
\label{equ:time_evolution_quasistatic}
\end{equation}
where we have explicitly denoted the dependence on external parameters $\alpha$.
Being $\epsilon$ sufficiently small we can imagine, that the actual state can be considered as a perturbation of the steady state,
as the system is in pursuit to reach the steady state.
We also assume that the small change of external parameters cannot cause a large change in the probability density neither in the stationary probability density,
so we can assume the perturbation itself varies on the same time scale as external parameters
\[
\mu_{t;\alpha(\epsilon t)} (x) = \rho_{\alpha(\epsilon t)} (x) + \epsilon \Delta \mu_{\epsilon t} (x) .
\]
By expanding the time evolution \eqref{equ:time_evolution_quasistatic} up to the first order in $\epsilon$, which will prove to be sufficient, we obtain the equation for the perturbation
\[
\dot\alpha(\epsilon t) \cdot \left. \nabla_\alpha \rho_\alpha (x) \right|_{\alpha=\alpha(\epsilon t)} = {\mathcal L}^*_{\alpha(\epsilon t)} \left[ \Delta \mu_t \right] (x) .
\]
Although the forward Kolmogorov generator is not in general invertible,
due to the fact that the steady state $\rho_\alpha$ corresponds to the zero eigenvalue of the forward Kolmogorov generator ${\mathcal L}^*_\alpha$,
in this particular case the solution can be found
\begin{equation}
\mu_{t;\alpha(\epsilon t)} (x) = \rho_{\alpha(\epsilon t)} (x)
+ \epsilon \dot\alpha (\epsilon t) \cdot \frac{1}{{\mathcal L}^*_{\alpha(\epsilon t)}} \left[ \left. \nabla_\alpha \rho_\alpha \right|_{\alpha=\alpha(\epsilon t)} \right] (x)
+ \err[2]{\epsilon} .
\label{equ:quasistatic_expansion_state}
\end{equation}
where we have introduced \emph{the forward pseudoinverse}\index{pseudoinverse!forward}
\begin{equation}
\frac{1}{{\mathcal L}^*} \left[ \mu \right] (x) = \int\limits_0^\infty {\mathrm d} t \; \left\{ \rho(x) \int {\mathrm d} \Gamma(y) \; \mu(y) - {\mathrm e}^{t {\mathcal L}^*} \left[ \mu \right] (x) \right\} .
\label{def:forward_pseudoinverse}
\end{equation}
From mathematician point of view the forward pseudoinverse is a linear operator which returns the argument of the forward Kolmogorov generator whenever can be found and zero otherwise,
i.e. the forward pseudoinverse is the extension of the inverse of the forward Kolmogorov generator from the subspace where the inverse exists up to the whole space.
In order to understand the pseudoinverse from the physical viewpoint,
we define at first \emph{the backward pseudoinverse}\index{pseudoinverse!backward} in a similar fashion how the backward Kolmogorov generator is related to the forward Kolmogorov generator \eqref{def:backward_generator}
\[
\int {\mathrm d} \Gamma(x) \; w(x) \frac{1}{{\mathcal L}^*}[\mu](x) = \left\langle \frac{1}{{\mathcal L}}[w] \right\rangle_\mu ,
\]
which is equivalent to the definition
\begin{equation}
\frac{1}{{\mathcal L}}[w](x) = \int\limits_0^\infty {\mathrm d} t \; \left[ \left\langle w \right\rangle_\rho - \left\langle {\mathrm e}^{t {\mathcal L}}[w] \right\rangle_{\delta_x} \right] .
\label{def:backward_pseudoinverse}
\end{equation}
Now let us consider the $w$ being the local power, then the first term corresponds to the steady production of the work and the later term to the total work production along the relaxation process starting from the configuration $x$,
hence the backward pseudoinverse can be interpreted as the transient or excess contribution to the total work, when the system relax from the initial configuration $x$ to the steady state $\rho$, see figure \ref{pic:backward_pseudoinverse}.
\begin{figure}[ht]
\caption{Illustration of the transient contribution of the total work represented by pseudoinverse when relaxing toward steady state.
The transient contribution is denoted by the filled area, while the total work also contains the hatched area.}
\begin{center}
\includegraphics[width=.7\textwidth,height=!]{backward_pseudoinverse.pdf}
\end{center}
\label{pic:backward_pseudoinverse}
\end{figure}
The forward pseudoinverse can be then interpreted as the same transient contribution although in this case represented on states.
\subsection{Quasistatic work and heat}
\label{ssec:quasistatic_heat_and_work}
To obtain the quasistatic expansion of the mean value of the total work up to the zeroth order in $\epsilon$
we use the fact that the total mean value of the work can be expressed in terms of the local power \eqref{equ:total_local_work_relation},
to which we apply the quasistatic expansion of the state \eqref{equ:quasistatic_expansion_state}
\begin{multline*}
{\mathcal W}(\alpha(t))
= \int\limits_0^{\frac{T}{\epsilon}} {\mathrm d} t \; \left\langle w^\epsilon_{\alpha(\epsilon t)} \right\rangle_{\mu_t}
= \int\limits_0^{\frac{T}{\epsilon}} {\mathrm d} t \; \left\langle w^\epsilon_{\alpha(\epsilon t)} \right\rangle_{\rho_{\alpha(\epsilon t)}} + \\
+ \int\limits_0^{\frac{T}{\epsilon}} {\mathrm d} t \; \epsilon \dot\alpha(\epsilon t) \cdot \int {\mathrm d} \Gamma(x) \; \frac{1}{{\mathcal L}^*_{\alpha(\epsilon t)}} \left[ \left. \nabla_\alpha \rho_\alpha \right|_{\alpha=\alpha(\epsilon t)} \right] (x) \, w^\epsilon_{\alpha(\epsilon t)} (x) + \err{\epsilon} = \\
= \frac{1}{\epsilon} \int\limits_0^T {\mathrm d} t \; \left\langle w^\epsilon_{\alpha(t)} \right\rangle_{\rho_{\alpha(t)}}
- \int\limits {\mathrm d} \alpha \cdot \left\langle \nabla_\alpha \frac{1}{{\mathcal L}_\alpha} [w^\epsilon_\alpha] \right\rangle_{\rho_\alpha}
+ \err{\epsilon} .
\end{multline*}
By inserting the total local power \eqref{equ:total_local_power}
\[
w^\epsilon_\alpha = \epsilon \dot \alpha \cdot \nabla_\alpha U_\alpha (x) + w^{\text{nonp}}_\alpha (x) ,
\]
where the non-potential part no longer depends on $\epsilon$,
we obtain a quasistatic expansion of the mean value of the total work
\[
{\mathcal W}(\alpha(t))
= \frac{1}{\epsilon} \int\limits_0^T {\mathrm d} t \; \left\langle w^\text{nonp}_{\alpha(t)} \right\rangle_{\rho_{\alpha(t)}}
+ \int\limits {\mathrm d} \alpha \cdot \left\langle \nabla_\alpha \left( U_\alpha - \frac{1}{{\mathcal L}_\alpha} [w^\text{nonp}_\alpha] \right) \right\rangle_{\rho_\alpha}
+ \err{\epsilon} .
\]
The first term sometimes referred to as \emph{``housekeeping''}\index{housekeeping work}\index{work!housekeeping} or steady component of the work \cite{result1} corresponds to the work necessary to maintain the system in the non-equilibrium state
and as such diverge in the quasistatic limit as the amount of the work necessary to maintain steady state over the infinite period of time goes to infinite.
In equilibrium there is no work supply necessary to maintain the equilibrium state,
hence there is no steady production of work and so forth this term is zero in equilibrium leading to the total work being determined only by the second term, which is usually finite.
The second term usually referred to as \emph{``reversible''}\index{reversible work}\index{work!reversible} or ``excess''\index{excess work}\index{work!excess} work \cite{result1,result2} is the additional work associated with transitions between steady states along the process.
We can see that these two components differs in some key aspects,
the ``housekeeping'' work is symmetric with respect to the trajectory reversal $\Theta \alpha$ while the ``reversible'' is antisymmetric.
This is also quite different from the equilibrium situation where in the quasistatic limit the total work and just the part of it is antisymmetric with respect to trajectory reversal.
The second difference lies in the dependence on the trajectory $\alpha$,
while the ``housekeeping'' component does not depend on the trajectory itself but rather on which steady states were visited and for how long,
the ``reversible'' component depends only on the ``shape'' of the trajectory thus being geometric, i.e. it does not depend on the actual parametrization of the trajectory.
To summarize, we can see that the ``housekeeping'' component is
\begin{enumerate}
\item extensive in time,
\item non-zero unless the global detailed balance condition is satisfied,
\item invariant with respect to the reversal of the external protocol $\alpha(t) \to \alpha(T-t) $,
\end{enumerate}
while the ``reversible'' component is
\begin{enumerate}
\item finite in the quasistatic limit,
\item geometric,
\item antisymmetric under the reversal of the external protocol $\alpha(t) \to \Theta \alpha(t)$.
\end{enumerate}
A typical experimentally easiest accessible quantity is the total work done on the system,
however we can use the properties listed above namely the behaviour under trajectory reversal to separate these components apart.
If the experimental setup can be made in such a way that we can measure the total work along some trajectory $\alpha(t)$ as well as along $\Theta \alpha(t)$ over the time interval $[0,T]$
then if the time interval $T$ is long enough to consider the process close to quasistatic, the ``housekeeping'' work can be estimated as the symmetric part of these works
\begin{equation}
{\mathcal W}^{hk}(\alpha)
\equiv \int\limits^T_0 {\mathrm d} t \; \left\langle w_{\alpha(t)} \right\rangle_{\rho_{\alpha(t)}}
= \frac{1}{2} \left[ {\mathcal W}(\alpha(t)) + {\mathcal W}(\Theta \alpha(t)) \right] + \err{\epsilon}
\label{def:housekeeping_work}
\end{equation}
and the ``reversible'' work is estimated as the antisymmetric part
\begin{equation}
{\mathcal W}^{rev}(\alpha)
\equiv \int {\mathrm d} \alpha \cdot \left\langle \nabla_\alpha \left( U_\alpha - \frac{1}{{\mathcal L}} [ w_\alpha ] \right) \right\rangle_{\rho_\alpha}
= \frac{1}{2} \left[ {\mathcal W}(\alpha(t)) - {\mathcal W}(\Theta \alpha(t)) \right] + \err{\epsilon}.
\label{def:reversible_work}
\end{equation}
Although these estimates corresponds to the ``housekeeping'' and the ``reversible'' work respectively in the quasistatic limit,
the fact that fluctuations of the total work usually diverge in the quasistatic limit along with the ``reversible'' work being \emph{non extensive} in time makes it hard to obtain,
for illustration see figure \ref{pic:work_decomposition}.
\paragraph{Example: Dragged particle}
To illustrate the experimental accessibility of the ``housekeeping'' and the ``reversible'' work we consider a numerical simulation of the underdamped diffusing particle in the optical trap in plane.
The optical trap can be simulated by the quadratic potential.
The system is driven out of equilibrium by dragging the particle by the optical trap in this particular case in circles with radius $R$ and period $\tau$.
Such system can be equivalently described by non-potential angular force acting on the particle.
The approximation of the quasistatic process is then achieved by periodically changing the radius $R$ on much longer time scale than the period of the driving $T \gg \tau$.
The first half of the period $T$ is the radius $R$ linearly increasing, while the other half is linearly decreasing.
Then if we compare the total mean work along such process with the same process shifted by half the period $T/2$, which can be considered as the reversed process,
we observe, see fig. \ref{pic:work_decomposition}, that in these two cases the lines showing the dependency of mean values of the total work on time almost coincide.
We can see that the difference is of much smaller magnitude than the mean values of the total work, compare scales on left and right $y$ axis.
The ``reversible'' work is then given as half of the difference of the mean values of the total heat over the half of the period $T/2$.
\begin{figure}[ht]
\caption{Demonstration of the experimental accessibility of the ``housekeeping'' and ``reversible'' work on the simulation of the system containing a single diffusing particle in the optical trap.
The parameters of the model in the simulation are set to be: mass of the particle $m=1$, friction coefficient $\gamma=1$, inverse temperature of the environment $\beta=0.1$, spring constant of the quadratic potential $k=10$, and the distance of the center of the potential well from the origin $R=5$, period of driving $\tau=20$.
Where the quasistatic process consists of the periodical changing the distance $R$ by 5\% with the period of $T=1280$.
The red line is the mean value of the total work over an statistical ensemble of $8192$ independent particles,
the green line is the same although shifted by half the period $T/2$
and the blue line denotes the actual time dependence of the difference between, which in the times $kT+T/2$, $k \in {\mathbb N}$, coincide with the ``reversible'' work.
}
\begin{center}
\includegraphics[width=.85\textwidth,height=!]{reversible_work.pdf}
\end{center}
\label{pic:work_decomposition}
\end{figure}
The same computation can be repeated also for the mean value of the total heat dissipated to the system from the $i$-th thermal bath
\[
{\mathcal Q}_i(\alpha(t))
= \frac{1}{\epsilon} \int\limits_0^T {\mathrm d} t \; \left\langle q^i_{\alpha(t)} \right\rangle_{\rho_{\alpha(t)}}
- \int\limits {\mathrm d} \alpha \cdot \left\langle \nabla_\alpha \frac{1}{{\mathcal L}_\alpha} [q^i_\alpha] \right\rangle_{\rho_\alpha}
+ \err{\epsilon} ,
\]
where we identify \emph{the ``housekeeping'' heat}\index{housekeeping heat}\index{heat!housekeeping}
\[
{\mathcal Q}_i^\text{hk} (\alpha) = \int\limits_0^T {\mathrm d} t \; \left\langle q^i_{\alpha(t)} \right\rangle_{\rho_{\alpha(t)}},
\]
and \emph{the reversible heat}\index{reversible heat}\index{heat!reversible}
\begin{equation}
{\mathcal Q}_i^\text{rev} (\alpha) = - \int {\mathrm d} \alpha \cdot \left\langle \nabla_\alpha \frac{1}{{\mathcal L}_\alpha} [ q^i_\alpha ] \right\rangle_{\rho_\alpha}.
\label{def:reversible_heat}
\end{equation}
The ``housekeeping'' and the ``reversible'' heat have the same properties as the corresponding works.
Again the ``housekeeping'' heat is to the heat necessary to maintain the steady states, which the system passed by.
It is also symmetric with respect to trajectory reversal $\Theta \alpha$, is is extensive with time and also present even in case there is no change of external parameters.
The ``reversible'' heat is antisymmetric with respect to trajectory reversal and is geometric.
\subsection{Generalized thermodynamics}
\label{ssec:generalized_thermodynamics}
\subsubsection{First law of thermodynamics}
One of the basic assumptions of our framework is that the underlying microscopic dynamic preserves the total energy on the microscopic level even under non-equilibrium conditions.
This means that the first law of thermodynamic for each particular path \eqref{def:first_law_path} as well as on the level of mean values \eqref{equ:first_law_eq} is still valid
\[
\left\langle E_{\alpha(T)} \right\rangle_{\rho_\alpha(T)} - \left\langle E_{\alpha(0)} \right\rangle_{\rho_\alpha(0)} = {\mathcal W}(\alpha) + \sum_i {\mathcal Q}_i(\alpha) ,
\]
where we sum over all baths attached to the system and the internal energy is represented by the steady mean value of the energy function ${\mathcal U}(\alpha) = \left\langle E_\alpha \right\rangle_{\rho_\alpha}$.
While the first law is valid for arbitrary trajectory $\alpha(t)$ it has to be also valid even in the quasistatic limit,
even so that the mean value of the total heat and work diverge in the quasistatic limit.
To be able to see that the diverging parts has to cancel each other out, we realize that in the steady state the total work done on the system is immediately dissipated to thermal baths
\[
0 = \left\langle w_\alpha \right\rangle_{\rho_\alpha} + \sum_i \left\langle q^i_\alpha \right\rangle_{\rho_\alpha} .
\]
From there we can see that ``housekeeping'' parts in the quasistatic expansion cancel each other out, thus effectively renormalizing the mean values of the total work and heat
\[
{\mathcal W}^\text{hk}(\alpha) + \sum_t {\mathcal Q}_i^\text{hk}(\alpha) =
\int\limits_0^{\frac{T}{\epsilon}} {\mathrm d} t \; \left[ \left\langle w_{\alpha(\epsilon t)} \right\rangle_{\rho_{\alpha(\epsilon t)}} + \sum_i \left\langle q^i_{\alpha(\epsilon t)} \right\rangle_{\rho_{\alpha(\epsilon t)}} \right]
= 0 .
\]
This means that in the quasistatic limit even out of equilibrium the energetics of the system is governed only by antisymmetric ``reversible'' components,
\begin{equation}
{\mathcal U}(\alpha(T)) - {\mathcal U}(\alpha(0)) = {\mathcal W}^\text{rev}(\alpha) + \sum_i {\mathcal Q}_i^\text{rev}(\alpha)
\label{equ:first_law_noneq}
\end{equation}
which we will call the non-equilibrium version of \emph{the first law of thermodynamics}\index{first law of thermodynamics!non-equilibrium} in the quasistatic limit.
While the reversible components are usually finite, we have then found out the natural way how to renormalise the heat and work to obtain thermodynamically relevant quantities.
The ``reversible'' heat and work are given by geometric integral on configuration space,
which we can use to introduce the differential version of the first law of thermodynamics \eqref{equ:first_law_noneq}
\[
{\mathrm d}_\alpha {\mathcal U}(\alpha) = {\mathchar'26\mkern-11mu {\mathrm d}} {\mathcal W}^\text{rev}(\alpha) + \sum_i {\mathchar'26\mkern-11mu {\mathrm d}} {\mathcal Q}^\text{rev}_i(\alpha) ,
\]
where ${\mathchar'26\mkern-11mu {\mathrm d}} {\mathcal Q}^\text{rev}_i (\alpha) $ and ${\mathchar'26\mkern-11mu {\mathrm d}} {\mathcal W}^\text{rev} (\alpha) $ denote \emph{inexact differentials}\index{inexact differential}
\begin{align*}
{\mathchar'26\mkern-11mu {\mathrm d}} {\mathcal W}^\text{rev} (\alpha) &= {\mathrm d} \alpha \cdot \left\langle \nabla_\alpha \left( E_\alpha - \frac{1}{{\mathcal L}_\alpha} [w^\text{nonp}_\alpha] \right) \right\rangle_{\rho_\alpha}, \\
{\mathchar'26\mkern-11mu {\mathrm d}} {\mathcal Q}^\text{rev}_i (\alpha) &= - {\mathrm d} \alpha \cdot \left\langle \nabla_\alpha \frac{1}{{\mathcal L}_\alpha} [q^i_\alpha] \right\rangle_{\rho_\alpha} .
\end{align*}
\subsubsection{Generalized Clausius relation}
There exist in literature various proposals how to extend the (equilibrium) Clausius relation to non-equilibrium domains.
A crucial point is that all these tentative generalizations heavily depend on the adopted scheme for the heat "renormalization", i.e., on the way how the finite component of the heat is actually defined.
Whereas in some renormalization schemes the generalized Clausius equality is proved to be a mathematical identity, cf. \cite{Hatano2001,Esposito2010,trepagnier2004}, within the (supposedly more physical) scheme adopted here it posses a non-trivial problem.
In chapter \ref{chapter:equilibrium_stat_phys} we have shown the Clausius equality \eqref{equ:Clausius_relation} is the consequence of the local detailed balance condition \eqref{def:local_detailed_balance}.
While we assume the local detailed balance condition to be still valid, one might naively expect the Clausius relation also to be valid.
The main problem however lies in the fact that the total heat and hence the total entropy production along the quasistatic process diverge.
This means that the second law \eqref{equ:second_law} \emph{does not} impose any condition on the ``reversible'' heat, which thus can in principle be arbitrary large.
In general there is no Clausius relation relating the ``reversible'' heat \eqref{def:reversible_heat} to the entropy production,
which means that the existence of thermodynamical potentials and corresponding Maxwell relations known from equilibrium thermodynamics, or their generalized versions, is no longer guaranteed.
In the special case when the steady state can be written as Boltzmann-like distribution
\begin{equation}
\rho_\alpha(x) = \frac{1}{Z_\alpha} \exp \left[ - \widetilde\beta(\alpha) \left( E_\alpha(x) - \frac{1}{{\mathcal L}_\alpha} [ w^\text{nonp}_\alpha ] (x) \right) \right]
\label{equ:generalized_Boltzmann}
\end{equation}
the Clausius relation can be again retrieved, where
\begin{equation}
V_\alpha (x) = E_\alpha - \frac{1}{{\mathcal L}_\alpha} [w^\text{nonp}_\alpha ]
\label{def:generalized_potential}
\end{equation}
is called \emph{the quasi-potential}\index{quasi-potential}
and $\widetilde\beta(\alpha)$ is an arbitrary function of external parameters representing \emph{the generalized inverse temperature}\index{generalized inverse temperature}.
An example of such system is the diffusion on 2D plane with non-potential force in the angular direction and the potential force in the radial direction, for details see subsection \ref{ssec:diffusion_on_plane}.
Another example is McLennan distribution in close to equilibrium situation, see section \ref{sec:first_order_expansion}.
The internal energy is thus given by the mean value of the quasi-potential
\[
{\mathcal U}(\alpha) = \left\langle E_\alpha \right\rangle_{\rho_\alpha}
= \left\langle E_\alpha - \frac{1}{{\mathcal L}_\alpha} [ w^\text{nonp}_\alpha ] \right\rangle_{\rho_\alpha}
= \left\langle V_\alpha \right\rangle_{\rho_\alpha} ,
\]
where we have used the identity \eqref{equ:backward_pseudoinverse_zero_mean} ,
and the ``reversible'' work is directly obtained from the definition \eqref{def:reversible_work}
\[
{\mathchar'26\mkern-11mu {\mathrm d}} {\mathcal W}^\text{rev} (\alpha) = {\mathrm d} \alpha \cdot \left\langle \nabla_\alpha V_\alpha \right\rangle_{\rho_\alpha} .
\]
From there we can obtain the total ``reversible'' heat by using the first law of thermodynamics \eqref{equ:first_law_noneq}
\begin{equation}
{\mathchar'26\mkern-11mu {\mathrm d}} {\mathcal Q}^\text{rev}_\text{tot} (\alpha) = {\mathrm d} \alpha \cdot \left[ \nabla_\alpha \left\langle V_\alpha \right\rangle_{\rho_\alpha} - \left\langle \nabla_\alpha V_\alpha \right\rangle_{\rho_\alpha} \right] .
\label{equ:heat_capacity_quasi-potential}
\end{equation}
We can again introduce the Shannon entropy of the system \index{entropy!Shannon}
\[
{\mathcal S} (\alpha) = - \left\langle \ln \rho_\alpha \right\rangle_{\rho_\alpha} ,
\]
which enables us to express the ``reversible'' heat as a generalized version of Clausius equilibrium relation, however in this case the Lagrange multiplier is generalized temperature $1/\widetilde\beta(\alpha)$,
which in general does not depend only on the temperature of single or multiple thermal baths, but can depend also on other external parameters
\[
{\mathchar'26\mkern-11mu {\mathrm d}} {\mathcal Q}^\text{rev}_\text{tot} (\alpha) = \frac{1}{\widetilde\beta(\alpha)} \; {\mathrm d} {\mathcal S}(\alpha) .
\]
\subsubsection{Quasistatic response functions}
In equilibrium thermodynamics \emph{response functions}\index{response function} such as heat capacity or compressibility are related to the quasistatic change of a particular thermodynamical potential along the infinitesimal change of some external parameter \cite{Callen1985}, e.g. heat capacity at constant volume is the quasistatic change of the internal energy with respect to temperature, the compressibility can be related to the quasistatic isothermal change of the free energy with respect to pressure.
Although in general there is no second law of thermodynamics out of equilibrium for ``reversible'' components and hence there are no thermodynamical potentials in general,
it can be still meaningful to define response functions outside of equilibrium.
Response functions like the isothermal compressibility can be defined the same way as in equilibrium, if in this case the pressure $p$ and volume $V$ are also defined out of equilibrium,
i.e. the isothermal compressibility is defined as
\[
\kappa_t = - \frac{1}{V} \left. \partial_p V \right|_{T_i=\text{const.}} ,
\]
where by $\left. \partial_p V \right|_{T=\text{const.}}$ we denote the quasistatic change of the volume with respect to the pressure,
while all temperatures $T_i$ of thermal baths attached to system are constant.
The key observation in case of the heat capacity and similar quantities is that the ``housekeeping'' heat (work) does not contribute to the change of internal energy \eqref{equ:first_law_noneq}.
Also the ``reversible'' heat is geometric and hence does not depend on the parametrization of the trajectory on external parameters $\alpha(t)$,
thus making the ``reversible'' heat the natural candidate to be used in the definition of the generalized heat capacity.
We define \emph{the generalized heat capacity}\index{heat capacity!generalized} as
\begin{equation}
C_i (\alpha) = \left. \frac{{\mathchar'26\mkern-11mu {\mathrm d}} Q^\text{rev}_i (\alpha) }{{\mathrm d} T_i} \right|_\alpha = - \left\langle \partial_{T_i} \frac{1}{{\mathcal L}_\alpha} [q^i_\alpha] \right\rangle_{\rho_\alpha} .
\label{def:generalized_heat_capacity}
\end{equation}
Although the heat capacity defined in such way coincide with the standard heat capacity in equilibrium, as will be shown later, the non-existence of the second law of thermodynamics causes that the generalized heat capacity can be in general negative, as will be shown on examples.
We have already stated that the general non-validity of the Clausius equality on the level of ``reversible'' components cause the nonexistence of thermodynamical potentials,
hence in general there are no relations between response functions as those of Mayer and Maxwell.
\subsubsection{Equilibrium case revisited}
The equilibrium thermodynamics presented in the chapter \ref{chapter:equilibrium_stat_phys} especially in the section \ref{sec:quasistatic_processes_eq} can be considered as a special case of presented framework.
To demonstrate it we derive again within our framework the results presented there by taking
\begin{align*}
w^\text{nonp}_\alpha (x) &=0 , &
q_\alpha (x) &= {\mathcal L}_\alpha [ E_\alpha ] (x) .
\end{align*}
From where it immediately follows that ``housekeeping'' components are zero.
Because the local power of non-equilibrium forces is zero, the ``reversible'' work is in this particular case given entirely by the change of the energy
\[
{\mathcal W}^\text{rev}(\alpha) = \int {\mathrm d} \alpha \cdot \left\langle \nabla_\alpha E_\alpha \right\rangle_{\rho_\alpha} .
\]
The ``reversible'' heat can be also simplified to
\[
{\mathcal Q}^\text{rev}(\alpha)
= - \int {\mathrm d} \alpha \cdot \left\langle \nabla_\alpha \frac{1}{{\mathcal L}_\alpha} {\mathcal L}_\alpha [ E_\alpha ] \right\rangle_{\rho_\alpha}
= \int {\mathrm d} \alpha \cdot \left[ \nabla_\alpha \left\langle E_\alpha \right\rangle_{\rho_\alpha} - \left\langle \nabla_\alpha E_\alpha \right\rangle_{\rho_\alpha} \right]
\]
by using the identity \eqref{equ:backward_pseudoinverse_identity}
\[
\frac{1}{{\mathcal L}} {\mathcal L} [A] (x) = A(x) - \langle A \rangle_\rho .
\]
If we also assume that the equilibrium state is characterized by the Boltzmann or the Maxwell-Boltzmann distribution
\[
\rho_\alpha (x) = \frac{1}{Z_\alpha} {\mathrm e}^{- \beta E_\alpha (x) }
\]
we obtain the ``reversible'' heat in the form of the Clausius relation
\begin{equation}
{\mathcal Q}^\text{rev}(\alpha)
= - \int {\mathrm d} \alpha \cdot \frac{1}{\beta} \nabla_\alpha \left\langle \ln \rho_\alpha \right\rangle_{\rho_\alpha}
= \int {\mathrm d} \alpha \cdot \frac{1}{\beta} \nabla_\alpha {\mathcal S}(\alpha)
= \int \frac{1}{\beta} \; {\mathrm d} {\mathcal S}(\alpha) ,
\label{equ:reversible_heat_equilibrium}
\end{equation}
where ${\mathcal S}(\alpha) = - \langle \ln \rho_\alpha \rangle_{\rho_\alpha} $ denotes again the Shannon entropy.
Because the total energy does not depend on temperature the generalized heat capacity \eqref{def:generalized_heat_capacity} then simplifies to
\[
C = \partial_T \left\langle E_\alpha \right\rangle_{\rho_\alpha} = \partial_T {\mathcal U}(\alpha) ,
\]
or by using expression \eqref{equ:reversible_heat_equilibrium} we obtain the standard definition
\[
C = T \, \partial_T {\mathcal S}(\alpha) .
\]
\subsection{``Gauge'' invariance}
\label{ssec:gauge_invariane}
\index{gauge transformation}
In the section \ref{sec:broken_global_detailed_balance} we have argued that the physics is invariant under the ``gauge'' transformation \eqref{equ:gauge_symmetry},
i.e. the physical results cannot depend on how we divide the total force to the potential and the non-potential component.
As we have discussed in case of diffusion in chapter \ref{chapter:equilibrium_stat_phys} subsection \ref{ssec:work_heat_diffusion} the non-potential local power is given by the action of non-potential forces on the microscopic level.
If the non-potential power has also a component which can be characterized by some potential $U_\alpha(x)$,
then the corresponding component of the local power is given by $w_\alpha(x) = - {\mathcal L}[U_\alpha](x)$, see \eqref{equ:local_power_potential_underdamped}.
From there we can see that the ``gauge'' transformation \eqref{equ:gauge_symmetry} in terms of the local power is described by the transformation
\begin{align*}
& E_\alpha(x) &
&\longrightarrow&
& E_\alpha(x) + U_\alpha(x) , \\
w^\text{pot}_{\alpha(t)}(x) &= \dot \alpha (t) \cdot \left. \nabla_\alpha E_\alpha(x) \right|_{\alpha = \alpha(t)} &
&\longrightarrow&
\widetilde{w}^\text{pot}_{\alpha(t)}(x) &= \dot \alpha (t) \cdot \left. \nabla_\alpha \left[ E_\alpha(x) + U_\alpha(x) \right] \right|_{\alpha = \alpha(t)} , \\
& w^\text{nonp}_{\alpha(t)}(x) &
&\longrightarrow&
\widetilde{w}^\text{nonp}_{\alpha(t)}(x) &= w^\text{nonp}_{\alpha(t)}(x) + {\mathcal L}_{\alpha(t)} [U_{\alpha(t)}] (x) .
\end{align*}
The first thing to notice is that Kolmogorov generators and hence the steady state are invariant with respect to the ``gauge'' transformation.
This is due to the fact, that the evolution of the system is determined by the total work and not by any particular decomposition to the potential and the non-potential force.
However how other physical quantities behave under the ``gauge'' transformation cannot be seen as easily.
In this subsection we will focus on the behaviour of the work, heat and internal energy under the transformation further on.
We can see that the internal energy is modified by the ``gauge'' transformation by the mean value of the additional potential $U_\alpha(x)$
\[
\widetilde{\mathcal U}(\alpha) = \left\langle E_\alpha + U_\alpha \right\rangle_{\rho_\alpha} = {\mathcal U}(\alpha) + \left\langle U_\alpha \right\rangle_{\rho_\alpha}
\]
It is also easy to see that the ``housekeeping'' work is invariant with respect to the ``gauge'' transformation
\[
\widetilde{\mathcal W}^\text{hk} (\alpha)
= \int\limits_0^T {\mathrm d} t \; \left\langle \widetilde{w}^\text{nonp}_{\alpha(t)} \right\rangle_{\rho_{\alpha(t)}}
= \int\limits_0^T {\mathrm d} t \; \left\langle {w}^\text{nonp}_{\alpha(t)} + {\mathcal L}_{\alpha(t)} [ U_{\alpha(t)} ] \right\rangle_{\rho_{\alpha(t)}}
= {\mathcal W}^\text{hk} (\alpha) ,
\]
on the other hand the ``reversible'' part of the total work is not invariant with respect to the ``gauge'' transformation
\begin{multline*}
{\mathchar'26\mkern-11mu {\mathrm d}} \widetilde{\mathcal W}^\text{rev}(\alpha)
= {\mathrm d} \alpha \cdot \left\langle \nabla_\alpha \left( E_\alpha + U_\alpha - \frac{1}{{\mathcal L}_\alpha} \left[ w^\text{nonp}_\alpha + {\mathcal L}_\alpha [ U_\alpha ] \right] \right) \right\rangle_{\rho_\alpha} = \\
= {\mathrm d} \alpha \cdot \left\langle \nabla_\alpha \left( E_\alpha + \left\langle U_\alpha \right\rangle_{\rho_\alpha} - \frac{1}{{\mathcal L}_\alpha} [ w^\text{nonp}_\alpha ] \right) \right\rangle_{\rho_\alpha} = \\
= {\mathchar'26\mkern-11mu {\mathrm d}} {\mathcal W}^\text{rev}(\alpha) + {\mathrm d} \alpha \cdot \nabla_\alpha \left\langle U_\alpha \right\rangle_{\rho_\alpha}
= {\mathchar'26\mkern-11mu {\mathrm d}} {\mathcal W}^\text{rev}(\alpha) + {\mathrm d} \widetilde{\mathcal U}(\alpha) - {\mathrm d} {\mathcal U}(\alpha) ,
\end{multline*}
where we have again used the identity \eqref{equ:backward_pseudoinverse_identity}.
We can see that the ``reversible'' work compensate the change of the internal energy under the ``gauge'' transformation.
From where we can see that the ``reversible'' heat is invariant with respect to ``gauge'' transformation
\[
{\mathchar'26\mkern-11mu {\mathrm d}} \widetilde{\mathcal Q}^\text{rev} (\alpha)
= {\mathrm d} \widetilde{\mathcal U} (\alpha) - {\mathchar'26\mkern-11mu {\mathrm d}} \widetilde{\mathcal W}^\text{rev} (\alpha)
= {\mathrm d} {\mathcal U} (\alpha) - {\mathchar'26\mkern-11mu {\mathrm d}} {\mathcal W}^\text{rev} (\alpha)
= {\mathchar'26\mkern-11mu {\mathrm d}} {\mathcal Q}^\text{rev} (\alpha) ,
\]
where we have used the first law of thermodynamics \eqref{equ:first_law_noneq}.
The invariance of the ``reversible'' heat is important because it tells us that the generalized heat capacity \eqref{def:generalized_heat_capacity} is also invariant with respect to the ``gauge'' transformation.
Alternatively the invariance of the ``reversible'' heat can also be obtained from the fact, that the ``gauge'' transformation affects only the local power, while the local heat production is preserved.
To be more precise the ``reversible'' heat depends only on the steady state, backward pseudoinverse and the local heat production, which are all invariant with respect to the ``gauge'' transform, which concludes the proof.
The same reasoning can be made also for the ``housekeeping'' heat, from where we conclude that the ``housekeeping'' heat is also invariant with respect to the ``gauge'' transform
\[
\widetilde{\mathcal Q}^\text{hk}(\alpha) = {\mathcal Q}^\text{hk}(\alpha) .
\]
To summarize the discussion,
we can see that the mean value of the total heat as well as its components are invariant with respect to the ``gauge'' transformation,
which means that the total entropy production associated with the heat does not depend on how we divide the total work to the non-potential and the potential component.
Similarly the ``housekeeping'' work is also invariant with respect to the ``gauge'' symmetry,
which reflects the fact, that the mean steady power, which is related to the physical state of the system, also does not depend on the choice of the non-potential work.
On the other hand the ``reversible'' work and the internal energy are not invariant with respect to the ``gauge'' symmetry as they are tightly bounded with the definition of the energy on the microscopic level.
Thus we have there a certain freedom how to define the internal energy, in the extreme case the gauge can be fixed in such a way that the internal energy is uniformly zero.
In equilibrium the notion of internal energy is also associated with the total energy which can be in principle extracted from the system by a quasistatic process.
This is however no longer true in non-equilibrium steady, while there is constant energy current through the system.
However we can asked slightly different question, how much \emph{additional} energy to the steady production are we able to extract from the system by any quasistatic process?
From this point of view the most natural gauge fixation is what is in electrodynamic called the Coulomb gauge\index{Coulomb gauge}, i.e. the divergence of the non-potential force is zero,
which we will use in most cases in this thesis.
The last remark is that although the quasi-potential \eqref{def:generalized_potential} is also affected by the ``gauge'' transformation
\begin{multline*}
\widetilde{V}_\alpha(x)
= E_\alpha(x) + U_\alpha(x) - \frac{1}{{\mathcal L}_\alpha} \left[ w^\text{nonp}_\alpha + {\mathcal L}_\alpha [ U_\alpha ] \right] (x) = \\
= E_\alpha(x) - \frac{1}{{\mathcal L}_\alpha} \left[ w^\text{nonp}_\alpha \right] (x) + \left\langle U_\alpha \right\rangle_{\rho_\alpha}
= V_\alpha(x) + \left\langle U_\alpha \right\rangle_{\rho_\alpha} ,
\end{multline*}
because the effect of the transformation is given by the uniform shift of all energy levels only,
the stationary distribution \eqref{equ:generalized_Boltzmann} as well as to the Shannon entropy are not affected by the ``gauge'' transformation.
\subsection{Example: Two-level model}
\label{ssec:two_level_nonp}
The first model we consider \cite{result2} is a system with two states `$0$' and `$1$' with energies $E(0)=0$ and $E(1) = \Delta E > 0$ that are connected by two distinct channels `$+$' and `$-$', see \ref{ssec:two_level}.
Each channel is associated with its respective thermal bath, in this particular case at the same inverse temperature $\beta$.
The asymmetry between the channels is provided by an additional driving force performing work
\[
W^\text{nonp}\left(0 \xrightarrow{\pm} 1\right) = \pm F.
\]
Hence, for the loop formed by the allowed transitions we have $W^\text{nonp}(0 \xrightarrow{+} 1 \xrightarrow{-} 0) = 2F$,
manifesting a non-potential character of the driving force.
From the expression for the heat \eqref{def:first_law_path}
\[
Q\left(0 \xrightarrow{\pm} 1\right) = \pm F - \Delta E
\]
we immediately see
that the case $F > \Delta E$ (respectively $-F > \Delta E$) corresponds to a strong non-equilibrium regime in which the system dissipates a positive amount of energy along both transitions in the loop $0 \xrightarrow{+} 1 \xrightarrow{-} 0$ (respectively its reversal).
Note that in this regime the (original) notion of energy gap separating both states and uniquely distinguishing between the ground and excited states becomes essentially meaningless.
In the most general case we can describe the system with transition rates
\begin{align*}
\rate[\pm]{0}{1} &= A \exp \left[ \pm \frac{\Phi}{2} + \beta \frac{\pm F - \Delta E}{2} \right] , \\
\rate[\pm]{1}{0} &= A \exp \left[ \pm \frac{\Phi}{2} - \beta \frac{\pm F - \Delta E}{2} \right] ,
\end{align*}
where $A$ is the common symmetric part of the transition rates, which does not depends on the inverse temperature $\beta$, thus setting an overall time-scale,
the $\Phi$ describes the direction-independent asymmetry between the symmetric parts of each channel.
Although in general both parameters $A$ and $\Phi$ can possibly depend on other parameters like $\beta$, $F$ and $\Delta E$ in any non-trivial way,
we assume that they does not depend on the inverse temperature $\beta$.
Such assumption isn't physically well motivated,
however we use it simplify the situation and to better separate the non-potential and the additional channel-asymmetry effects.
While $A$ only sets an overall time-scale, it can be mostly ignored, hence for convenience, we set $A=1$ and always assume $\Phi \ge 0$,
due to the symmetry of the dynamics to the dual exchange of $F \to -F$ and $\Phi \to - \Phi$.
\subsubsection{Steady state}
The steady state \eqref{equ:stationary_condition} coincides with steady state of the channel-unresolved two level system with the \emph{total} escape rates $\lambda(x) = \lambda_+(x) + \lambda_-(x)$
\begin{equation}
\frac{\rho(1)}{\rho(0)} = \frac{\lambda(0)}{\lambda(1)} = {\mathrm e}^{- \beta \Delta E} \frac{1 + \zeta}{1 - \zeta} ,
\label{equ:2lvl_stat_occup}
\end{equation}
where
\[
\zeta = \tanh \left( \frac{\Phi}{2} \right) \, \tanh \left( \frac{\beta F}{2} \right) .
\]
In case either of $F$ or $\Phi$ is zero we have a probability distribution corresponding to the Boltzmann equilibrium.
\paragraph{Stationary currents}
In the steady state we would expect to observe a non-zero constant probability current \eqref{equ:current_multiple_baths} in each particular channel
\[
\curr[+]{0}{1} = \curr[-]{1}{0} = \frac{ \sinh\left( \frac{\beta F}{2} \right) }{ \cosh \left( \frac{ \Phi }{2} \right) \, \cosh \left( \frac{\beta \Delta E}{2} \right) \left[ 1 - \zeta \tanh\left( \frac{\beta \Delta E}{2} \right) \right] } ,
\]
from where it is evident that only the case $F = 0$ corresponds to the equilibrium.
Notice that the current diverge when
\[
\tanh \left( \frac{\Phi}{2} \right) \, \tanh \left( \frac{\beta F}{2} \right) \, \tanh\left( \frac{\beta \Delta E}{2} \right) = 1.
\]
It can be shown that the steady rate of dissipation from thermal bath associated with the $+$ channel to the system is the steady rate of dissipation from the system to the thermal bath associated with $-$ channel
and it is also proportional to the steady probability current $\curr[+]{0}{j}$
\[
\left\langle q^+ \right\rangle_\rho = - \left\langle q^- \right\rangle_\rho = 2 F \curr[+]{0}{1} .
\]
In formula \eqref{equ:2lvl_stat_occup} we have noticed modifications with respect to the equilibrium Boltzmann statistics whenever $\Phi \neq 0$.
It agrees with our intuition that relative throttling of the `$-$' with respect to the `$+$' channel under a positive $F > 0$ tends to increase the occupancy of the ``excited'' state `1'.
Eventually in the limit $\Phi \to +\infty$ the `$-$' channel completely closes and the system is again found at thermal equilibrium ($\curr[+]{0}{1} = 0$) but now with the energy gap $\Delta E - F$.
The resulting population inversion for $F > \Delta E$ is a most simple example of gauge transformation \eqref{equ:gauge_symmetry},
applied here to easily deal with the driving forces when they become derivable from a potential.
From this point of view, the population inversion is a superficial concept here since $\Delta E - F = E(1) - E(0)$ is just the full energy gap after the transformation with $U(1) - U(0) = -F$ completely removing the driving force has been applied.
\paragraph{High temperature behaviour}
However, more important is how these simple observations carry over when both channels remain open to hold the system out of equilibrium:
one checks that for an arbitrarily weak channel asymmetry the population inversion $\rho(1) > \rho(0)$ still occurs whenever the driving force is strong enough, and for large enough (but finite) temperature $T=1/\beta$.
This follows from
\begin{equation}
\log \frac{\rho(1)}{\rho(0)} =
-\frac{{\Delta E} - F \tanh\left(\frac{\Phi}{2}\right)}{T} + \err{\frac{1}{T^2}}
\label{equ:2lvl_distribution_governing_potential}
\end{equation}
In contrast to the limiting case $\Phi \to +\infty$, the driving force now does \emph{not} derive from a potential and hence cannot be transformed out.
Nevertheless, the leading term in the high-temperature expansion~\eqref{equ:2lvl_distribution_governing_potential} suggests that ${\Delta E} - F \tanh\bigl(\frac{\Phi}{2}\bigr)$ may take over the role of an effective energy gap,
though we are now dealing with a genuine non-equilibrium system where the energy levels are ambiguously defined.
\paragraph{Low temperature behaviour}
In the low-temperature regime the relative occupation~\eqref{equ:2lvl_distribution_governing_potential} has the asymptotic
\[
\log \frac{\rho(1)}{\rho(0)} = -\frac{\Delta E}{T} + \Phi \sgn(F) +\err{T},
\]
showing that independently of the driving force there is no population inversion at zero temperature.
Nevertheless, the system undertakes a transition between ``insulator'' and ``conductive'' regimes at $F = \pm {\Delta E}$ as seen from the low-temperature current asymptotic,
\[
\curr[+]{0}{1} \simeq \sgn(F)\,e^{\frac{|F| - {\Delta E}}{2T} + \frac{\phi}{2} \sgn(F)}
\xrightarrow{T \to 0^+}
\begin{cases}
0 & \text{if } |F| < {\Delta E} \\
\pm\infty & \text{if } |F| > {\Delta E}
\end{cases}
\]
This can be understood by observing that in the low-driving (or insulator) regime, $|F| < {\Delta E}$, the state `$0$' remains a well defined ground state in the sense that in both channels
\[
\log\frac{\rate[\pm]{1}{0}}{\rate[\pm]{0}{1}} \to +\infty \quad\text{for}\quad T \to 0,
\]
whereas in the high-driving (or conductive) regime $F > {\Delta E}$ the system exhibits a limit cycle behavior,
\begin{align*}
\log\frac{\rate[-]{1}{0}}{\rate[-]{0}{1}} &\to +\infty , &
\log\frac{\rate[+]{1}{0}}{\rate[+]{0}{1}} &\to -\infty ;
\end{align*}
analogously for $-F > {\Delta E}$.
\subsubsection{Energetics}
In the pursuit of generalized heat capacity \eqref{def:generalized_heat_capacity},
we need to determine the reversible work and heat at first.
We have seen that the reversible work can be expressed in terms of the quasi-potential \eqref{def:generalized_potential},
which depends on the total energy and local power of non-potential forces.
Fixing the \emph{a priori} gauge the energy levels are given as $E(0) = 0$, $E(1) = {\Delta E}$
while the non-potential work being $W^\text{nonp}(0 \xrightarrow{\pm} 1) = \pm F$ and $W^\text{nonp}(1 \xrightarrow{\pm} 0) = \mp F$.
The local power can be obtained directly then from definition \eqref{def:local_power}
\begin{align*}
w^\text{nonp}(0)
&= W^\text{nonp}\left( 0 \xrightarrow{+} 1\right) \rate[+]{0}{1} + W^\text{nonp}\left( 0 \xrightarrow{-} 1\right) \rate[-]{0}{1} \\
&= F(\rate[+]{0}{1} - \rate[-]{0}{1}), \\ \\
w^\text{nonp}(1)
&= W^\text{nonp}\left( 1 \xrightarrow{+} 0\right) \rate[+]{1}{0} + W^\text{nonp}\left( 1 \xrightarrow{-} 0\right) \rate[-]{1}{0} \\
&= F(\rate[-]{1}{0} - \rate[+]{1}{0})
\end{align*}
and hence the second term of the quasi-potential $V(i) = E(i) - 1/{\mathcal L}[w^\text{nonp}](i)$ can be in general determined by looking for solution of ${\mathcal L}[\breve V](i) = w^\text{nonp} (i) - \langle w^\text{nonp} \rangle_\rho$.
The generalized heat capacity \eqref{def:generalized_heat_capacity} in terms of quasi-potential
\[
C_\text{noneq} = \partial_T \left\langle V \right\rangle_\rho - \left\langle \partial_T V \right\rangle_\rho
= \left\langle V \right\rangle_{\partial_T \rho}
\]
simplifies in case of two level model to
\[
C_\text{noneq}
= V(0) \, \partial_T \rho(0) + V(1) \, \partial_T \rho(1)
= \beta^2 \rho(0) \, \rho(1) \, \Delta V \, G,
\]
where $\beta$ is the inverse temperature, $\Delta V = V(1)-V(0)$ is the gap in the quasi-potential and the $G$ is shorthand for $G = \partial_\beta \log (\rho_0 / \rho_1)$.
We can see that to be able to determine the generalized heat capacity the knowledge of the gap in quasi-potential $\Delta V$ is sufficient.
As a result we obtain
\[
\Delta V = {\Delta E} + F \frac{\tanh\left(\frac{{\Delta E}}{2T}\right) \tanh\left(\frac{F}{2T}\right)
- \tanh\left(\frac{\Phi}{2}\right)}{1 - \tanh\left(\frac{{\Delta E}}{2T}\right)
\tanh\left(\frac{F}{2T}\right) \tanh\left(\frac{\Phi}{2}\right)} .
\]
\begin{figure}[ht]
\caption{The quasi-potential gap $\Delta V = V(1) - V(0)$ compared to the gap $G = \partial_\beta \ln [ \rho(0) / \rho(1) ]$ as a function of temperature $T=1/\beta$.
The particular choice of parameters is $U=1$ and $\Phi=3$. }
\begin{center}
\includegraphics[width=.9\textwidth,height=!]{2lvl-gap.pdf}
\end{center}
\label{pic:2lvl-gap}
\end{figure}
Note first that for $F = 0$ one gets $\Delta V = G = \Delta E$ and we obtain a well-known formula for the heat capacity of an equilibrium two-state model.
Away from equilibrium the picture becomes far more complicated since both energy-dimensional quantities $G$ and $\Delta V$ are now different and generally not related in a simple way.
Moreover, they can obtain opposite signs for large enough driving forces, which then results in negative values of the generalized heat capacity, see figures~\ref{pic:2lvl-gap}--\ref{pic:2lvl-C_T}.
Next we separately analyze three asymptotic regimes.
\begin{figure}[ht]
\caption{The temperature dependence of the heat capacity for sub-critical, critical and supercritical driving. The model parameters are $U = 1$ and $\Phi = 3$.}
\begin{center}
\includegraphics[width=.9\textwidth,height=!]{2lvl-C_T.pdf}
\end{center}
\label{pic:2lvl-C_T}
\end{figure}
\paragraph{High temperatures}
For large temperature values the quasi-potential gap is
\[
\Delta V = {\Delta E} - F \tanh(\Phi / 2) + \err{1/T^2},
\]
which coincides with the asymptotic of $G$ as obtained from equation~\eqref{equ:2lvl_distribution_governing_potential}.
Hence the heat capacity equals
\[
C_\text{neq} = \frac{\bigl[
{\Delta E} - F \tanh \bigl( \frac{\Phi}{2} \bigr) \bigr]^2}{4 T^2} +
\err{\frac{1}{T^3}}
\]
We see that $F^* \equiv {\Delta E} / \tanh(\Phi / 2)$ is a critical value of the driving, above which the system exhibits a population inversion and also the gap $\Delta V$ changes sign.
As a result, for any $F \neq F^*$ the heat capacity is asymptotically strictly positive and decaying as $1 / T^2$, i.e. similarly as in equilibrium.
Note that the asymptotic equality $\Delta V \simeq G$ remains true even for a driving force $F$ much larger than the model parameter ${\Delta E}$,
the original meaning of which as an energy gap then becomes meaningless.
Instead, there is another gauge that becomes natural here: by making the transformation~\eqref{equ:gauge_symmetry} with $U(0) = 0$ and $U(1) = -F \tanh(\Phi / 2)$,
we obtain ``renormalized'' energy levels with the gap $\widetilde{E}(1) - \widetilde{E}(0) = {\Delta E} - F \tanh(\Phi / 2)$,
which is directly seen in the leading asymptotic $\Delta V \simeq G \simeq \widetilde{E}(1) - \widetilde{E}(0)$.
After this transformation, the residual non-potential forces contribute to the heat capacity only by correction ${\mathrm o}(1 / T^2)$.
In this sense the full high-temperature regime away from the critical value $F^*$ is to be understood as essentially close to equilibrium,
but with the renormalized energy levels $\widetilde{E}(i)$ and the corresponding Boltzmann stationary distribution.
From this point of view the observed population inversion at high temperatures and strong driving is only an artifact of describing the model in terms of ``unphysical'' energy levels $E(i)$.
\paragraph{High temperatures --- critical}
We have seen that the value $F = F^*$ plays a special role since in this case the above gauge transformation leads to degenerate energy levels,
and therefore the heat capacity becomes zero up to order $1 / T^2$.
More detailed calculations reveal that both gaps $\Delta V$ and $G$ are of order $1/T^2$ which yields an anomalously fast-decaying heat capacity,
\begin{equation}
C_\text{neq} = \frac{{\Delta E}^6}{64 T^6 \sinh^4 \frac{\Phi}{2} }
+ {\mathrm o}\left(\frac{1}{T^6}\right)
\end{equation}
The existence of the high-temperature critical driving leads to the following subtle phenomenon:
there is a temperature curve $T = T^1(F)$ along which $\Delta V = 0$ and another one,
$T = T^2(F) > T^1(F)$, on which $G = 0$.
Both curves have the identical leading asymptotic, for $F > F^*$,
\[
\frac{1}{T^{1,2}(F)} = 2 \sinh \left( \frac{\Phi}{2} \right) \sqrt{\frac{2(F - F^*)}{{\Delta E}^3 \sinh \Phi}} + {\mathrm o}\left((F - F^*)^{\frac{1}{2}}\right) .
\]
On both curves the heat capacity vanishes and they form the boundary of a tiny region in the $(T,F)-$space inside which $C_\text{neq}$ exhibits negative values.
Its full dependence on the driving for a fixed intermediate temperature is depicted on figure~\ref{pic:2lvl-C_F} where the above mentioned region has been zoomed in.
Notice the negative values of the heat capacity for large $F$;
this is a strong-non-equilibrium effect and we may expect that no gauge transformation would significantly simplify the thermodynamic description in this region due to a strong temperature-dependence of the quasi-potential gap $\Delta V$.
\begin{figure}[ht]
\caption{Steady heat capacity as a function of the driving, with the parameters
${\Delta E} = 1$, $\Phi = 3$, and $T = 1$. The tiny region of negative heat capacity in the vicinity of the critical driving is zoomed in.}
\begin{center}
\includegraphics[width=.9\textwidth,height=!]{2lvl-C_F.pdf}
\end{center}
\label{pic:2lvl-C_F}
\end{figure}
\paragraph{Low temperatures}
The quasi-potential gap $\Delta V$ has the low-temperature asymptotic
\begin{equation}
\Delta V = {\Delta E} + |F| + \err{|F|\,{\mathrm e}^{- \beta \min\{{\Delta E},|F|\} }}
\label{equ:2lvl_gap_lowT}
\end{equation}
in which the temperature dependence emerges only in the exponentially small correction (along the limit $T \to 0^+$).
This suggests that it is appropriate to make the gauge transformation with $U(0) = 0$ and $U(1) = F$,
to define the ``renormalized'' energy levels $\widetilde{E}(0) = 0$ and $\widetilde{E}(1) = {\Delta E} + |F|$ through which the heat capacity gets the simplified approximate form
\[
C_\text{neq} \simeq \partial_T \langle \widetilde{E} \rangle_\rho ,
\]
i.e. with only a negligible contribution from the second term in~\eqref{equ:heat_capacity_quasi-potential}.
One checks by comparing with the exact result that this intuition is indeed correct.
The apparent disagreement between the low-temperature asymptotic of the gaps $\Delta V$ and $G$, cf.~\eqref{equ:2lvl_stat_occup} and \eqref{equ:2lvl_gap_lowT},
indicates that the low-temperature regime corresponds to strong non-equilibrium with non-Boltzmannian statistics.
Formally, it can be described by an effective temperature defined by
\[
\log (\rho_1 / \rho_0) = -(\widetilde{E}(1) - \widetilde{E}(0)) / T^\text{eff},
\]
explicitly $T^\text{eff} = T (1 + |F| / {\Delta E}) > T$.
Using the fact that $C_\text{neq} \simeq (1 + |F|/{\Delta E})\,\partial \langle \widetilde{E} \rangle_\rho / \partial T^\text{eff}$,
we can trace back the exponential decay of the heat capacity for $T \to 0^+$ to the exponential suppression of thermal excitation, which is analogous to the equilibrium third law of thermodynamics.
Since $\Delta V \simeq {\Delta E} + |F| > 0$, and recalling that our model exhibits no population inversion in the zero-temperature limit,
we conclude that the generalized heat capacity remains strictly positive at low temperatures.
In particular, it does not exhibit any transition at $F = {\Delta E}$ where the system undertakes a change between the ``insulator'' and the ``conductive'' transport regimes.
We finish this example by indicating how to extend the above approximate description of the low-temperature behavior to arbitrary temperatures and driving forces.
\emph{Formally} defining the effective temperature $T^\text{eff} = \Delta V / \log(\rho_0 / \rho_1)$,
we can write the reversible heat~\eqref{def:reversible_heat} in the form of a Clausius equality
\begin{align*}
{\mathchar'26\mkern-11mu {\mathrm d}} Q^\text{rev} &= T^\text{eff} {\mathrm d} S , &
S &= -\sum_{i=0,1} \rho(i) \log \rho(i)
\end{align*}
with $S$ the Shannon entropy of the stationary distribution $\rho$.
In this framework the heat capacity obtains the form $C_\text{neq} = T^\text{eff} {\mathrm d} S / {\mathrm d} T$.
In contrast with the above low-temperature regime, the effective temperature now becomes a nontrivial function of $T$; for example, it becomes zero on the critical line $T = T^1(F)$.
Obviously, such a representation in terms of a (single) effective temperature has no straightforward extension to models with a larger number of states, and significant modifications are needed.
Some extensions of the Clausius relation to non-equilibrium and its limitations have been studied in~\cite{Komatsu2008,Komatsu2009,Sagawa2011}.
\subsection{Example: Three-level model}
\label{ssec:three_level_nonp}
Now we consider a three-level version of the above model, introduced in subsection \ref{ssec:three_level} also presented in \cite{result2},
with states `$0$', `$1$' and `$2$' mutually connected by single channels.
The system is driven out of equilibrium by a force acting along the loop $0 \rightarrow 1 \rightarrow 2 \rightarrow 0$ and performing equal non-potential work,
\[
W^\text{nonp}_{0 \rightarrow 1} = W^\text{nonp}_{1 \rightarrow 2} = W^\text{nonp}_{2 \rightarrow 0} = W_\text{ext},
\]
along all those transitions, see
The dynamics are defined by the transition rates
\[
\rate{i}{i\pm} =
A_{i,i\pm}\,\exp \left[ \frac{\beta}{2} \left( E(i) - E(i\pm) \pm W_\text{ext} \right) \right]
\]
where $i+$ (respectively $i-$) is the succeeding (respectively the preceding) state along the oriented loop $0 \rightarrow 1 \rightarrow 2 \rightarrow 0$; e.g., $0+ = 1$, $0- = 2$ etc.
The prefactors $A_{i j} = A_{j i} > 0$ are symmetric in order to satisfy the local detailed balance condition~\eqref{equ:global_detailed_balance_jump} but arbitrary otherwise.
We only assume that they can be kept constant and independent of other parameters like the temperature or forces.
To be specific, we assume that $E(2) > E(1) > E(0) = 0$ and $W_\text{ext} > 0$.
The present model exhibits a rich collection of different zero-temperature phases which are summarized in table~\ref{table:3lvl}.
In particular, it demonstrates a population inversion between levels `$0$' and `$1$' in the case $E(2) > 2E(1)$ and $W_\text{ext} > E(1)$.
Later we will see that \emph{at} the ``critical'' driving $W_\text{ext} = E(1)$, where the zero-temperature population inversion occurs, the model exhibits an anomalous low-temperature behavior.
\begin{table}
\caption{Zero-temperature phases of the driven three-level model.
In all cases $\log(\rho(2)/\rho(0 \text{ or } 1)) \to -\infty$ for $T \to 0^+$.}
\label{table:3lvl}
\begin{center}
\shorthandoff{-}
\begin{tabular}{|cc|c|c|}
\hline
\multicolumn{2}{|c|}{Zero-temperature phases}& $\log (\rho_0 / \rho_1)$ & $J$ \\ \hline \hline
\multicolumn{1}{|c|}{\multirow{2}{*}{$E(2) < 2 E(1)$}} & $W_\text{ext} < E(1)$ & $+\infty$ & $0$ \\ \cline{2-4}
\multicolumn{1}{|c|}{}& $W_\text{ext} > E(1)$ & $+\infty$ & $+\infty$ \\ \hline
\multicolumn{1}{|c|}{\multirow{3}{*}{$E(2) > 2 E(1)$}} & $W_\text{ext} < E(1)$ & $+\infty$ & $0$ \\ \cline{2-4}
\multicolumn{1}{|c|}{}& $E(1) < W_\text{ext} < E(2) - E(1)$ & $-\infty$ & $0$ \\ \cline{2-4}
\multicolumn{1}{|c|}{}& $E(2) - E(1) < W_\text{ext} $ & $-\infty$ & $+\infty$ \\ \hline
\end{tabular}
\shorthandon{-}
\end{center}
\end{table}
For the three-level model we skip the detailed analysis and only concentrate on some new features that were not seen in the previous two-level example.
Therefore we consider the case $E(2) > 2E(1)$ in which the system exhibits a zero-temperature transition in the stationary occupations, see table \ref{table:3lvl}.
The generalized heat capacity in the $(T,W_\text{ext})-$plane is depicted in figure~\ref{pic:3lvl-iso} where we see that $(T=0,F=E(1))$ is an accumulation point of the curves of zero generalized heat capacity.
\begin{figure}[ht]
\caption{The heat capacity landscape in $(T,W_\text{ext})$-plane with curves of constant heat capacity.
The particular choice of parameters is $E(1) = 1$, $E(2) = 3$ and $A_{0,1} = 1$, $A_{1,2} =2$ and $A_{2,0}=4$.}
\begin{center}
\includegraphics[width=.9\textwidth,height=!]{3lvl-iso.pdf}
\end{center}
\label{pic:3lvl-iso}
\end{figure}
\begin{sidewaysfigure}
\caption{The properties three-level model in various regimes.}
\begin{center}
\begin{subfigure}[hb]{0.31\textheight}
\caption{Sub-critical regime.}
\begin{center}
\includegraphics[width=\textwidth,height=!]{3lvl-subcrit.pdf}
\end{center}
\label{pic:subcritical-3lvl}
\end{subfigure} ~
\begin{subfigure}[hb]{0.31\textheight}
\caption{Critical regime.}
\begin{center}
\includegraphics[width=\textwidth,height=!]{3lvl-crit.pdf}
\end{center}
\label{pic:critical-3lvl}
\end{subfigure} ~
\begin{subfigure}[hb]{0.31\textheight}
\caption{Super-critical regime.}
\begin{center}
\includegraphics[width=\textwidth,height=!]{3lvl-supcrit.pdf}
\end{center}
\label{pic:supcritical-3lvl}
\end{subfigure}
\end{center}
\label{pic:properties-3lvl}
\end{sidewaysfigure}
In order to better understand the behavior we look into the critical case $W_\text{ext} = E(1)$ in more detail and compare it with sub- and super-critical case, the results are summarized in figure~\ref{pic:properties-3lvl}.
In the critical scenario at zero temperature the states `$0$' and `$1$' degenerate into a single energy level, both in the sense of stationary occupations, $\rho(0) = \rho(1) = 1/2$, and in the sense of the quasi-potential, $V(0) = V(1)$.
Increasing the temperature, the occupation of state `$0$' also increases (in this way behaving like an excited state), whereas the quasi-potential satisfies $V(1) > V(0)$,
meaning that the degeneracy gets removed and a positive energy gap opens between states `$0$' (lower) and `$1$' (higher).
As a direct consequence of these opposite tendencies the generalized heat capacity becomes negative at $W_\text{ext} = E(1)$ and low temperatures,
with an anomalously fast decay to zero for $T \to 0^+$ due to the zero-temperature degeneracy of both states.
Note that although the presence of state `$3$' is essential for breaking the detailed balance and for the non-equilibrium features of our model, it does not directly enter low-temperature energetics.
It also does not substantially contribute to the heat capacity until high enough temperatures where its occupation becomes relevant.
\subsection{Example: Diffusion in 2D-plane}
\label{ssec:diffusion_on_plane}
As an example for diffusions let us consider the driven two-dimensional rotationally symmetric overdamped diffusion in an homogeneous environment at the inverse temperature $\beta=1/T$,
see subsection \ref{ssec:overdamped_diff} in chapter \ref{chapter:models} for additional details about overdamped diffusions,
which is described by time evolution \eqref{def:position_overdamped}
\[
{\mathrm d} {\vec{q}}_t = \left[ \vec{F}(\vec{q}_t) - \vec{\nabla} U(\vec{q}_t) \right] \; {\mathrm d} t + \left( \frac{2}{\beta} \right)^{\frac{1}{2}}\, {\mathrm d} \vec{W}_t ,
\]
where $U(\vec{q}) = \frac{\lambda}{2}\,\|\vec{q}\|^2$ is the quadratic potential
and $\vec{F}(\vec{q}) = v(\|\vec{q}\|)\,\vec{e}_\theta(\vec{q})$ is the angular driving force, where $\vec{e}_\theta(\vec{q})$ is the unit vector in the angular direction at the position $\vec{q}$.
Note that both forces are mutually orthogonal $\vec{F}(\vec{q}) \cdot \vec{\nabla}U(\vec{q}) = 0$.
Furthermore we will assume that the magnitude of the driving is proportional to the power of radius $v(r) = \kappa\, r^\alpha$ with some fixed exponent $\alpha > -1$.
Because of the angular symmetry of the system we will further proceed in the polar coordinates $(r,\theta)$.
The stationary distribution and the steady probability current \eqref{def:current_overdamped} are then obtained from the stationarity equation
\[
\partial_r \left(r j_r(r,\theta) \right)+ \partial_\theta j_\theta(r,\theta) = 0
\]
with the components or the probability current being
\begin{align*}
j_r(r,\theta) &= -\rho(r,\theta) \, \partial_r U(r) - T\,\partial_r \rho(r,\theta), \\
j_\theta(r,\theta) &= \rho(r,\theta)\, v(r) - \frac{T}{r}\,\partial_\theta \rho(r,\theta) .
\end{align*}
The explicit solution can be found
\begin{align*}
\rho(r,\theta) &= \frac{1}{Z}\,{\mathrm e}^{-\beta U(r)}, &
Z &= \frac{2\pi T}{\lambda} , &
\vec{j}(r,\theta) &\equiv (j_r,j_\theta) = (0, \rho(r,\theta)\, v(r)),
\end{align*}
where we can see that steady state has the same probability distribution as in equilibrium, which is due to the orthogonality of the potential $-\vec{\nabla} U$ and non-potential driving $\vec{F}$ forces.
Moreover the steady probability current is proportional only to the non-potential forces and as such it has only an angular component, thus creating a vortex around the origin.
As there is a non-zero steady probabilistic current, there exists a steady heat current from the system to the environment which is given by the mean value of the local power,
where the local power of the driving force $F$ is given by the function
\[
w(\vec{q}) = \vec{F}(\vec{q}) \cdot ( \vec{F}(\vec{q}) - \vec{\nabla} U(\vec{q}) ) + T \, \vec{\nabla} \cdot \vec{F}(\vec{q}) = v^2(\|\vec{q}\|) ,
\]
from where it follows that the mean steady dissipated power is
\[
\langle w \rangle_\rho \equiv \int\limits_0^\infty {\mathrm d} r \; \int\limits_0^{2\pi} {\mathrm d} \theta \; w(r,\theta) \, \rho(r,\theta) =
\Gamma(\alpha + 1)\,\kappa^2\,\left( \frac{2T}{\lambda} \right)^\alpha,
\]
where $\Gamma$ is the gamma function.
Notice that $\langle w \rangle_\rho$ diverges when $\alpha \to -1$.
Moreover the mean local power is increasing function of temperature $T$ for $\alpha > 0$ while it is decreasing for $-1 < \alpha < 0$,
this associated with the fact that the increase of the temperature always increase the probability of observation a particle further from the origin,
which is in case $\alpha > 0$ is associated with higher dissipation as the driving force increase with the radius, while in the other case $-1 < \alpha < 0$ the dissipation diminishes as the driving force is weaker.
\subsubsection{Non-equilibrium heat capacity}
Such dependence on the parameter $\alpha$ is also reflected on quasistatic response functions, namely the heat capacity.
The generalized heat capacity \eqref{def:generalized_heat_capacity} as defined via the quasistatic ``reversible'' heat \eqref{equ:heat_capacity_quasi-potential} is given by the general formula
\begin{align*}
C_F &= \partial_T \langle U \rangle_\rho + \Delta C_F, \\
\Delta C_F &= \left\langle \partial_T\,\frac{1}{{\mathcal L}}[w] \right\rangle_\rho ,
\end{align*}
where with the subscript $F$ we denote the that the heat capacity is given with respect to constant driving force, namely $\kappa = {\mathrm {const.}}$
For this particular model, the first ``classical'' term is in accordance with the equilibrium equipartition theorem $\partial_T \langle U \rangle_\rho = 1$ (using the convention $k_B \equiv 1$).
For the second genuinely non-equilibrium term we need to compute the function
$G(r,\theta) = \frac{1}{{\mathcal L}}[w](r,\theta)$ satisfying the equations \eqref{equ:backward_pseudoinverse_identity} and \eqref{equ:backward_pseudoinverse_zero_mean}
\begin{gather*}
\langle G \rangle_\rho = 0 , \\
{\mathcal L}[G](r,\theta) = {\mathcal L} \frac{1}{{\mathcal L}} [w] (r,\theta) \equiv w(r,\theta) - \langle w \rangle_\rho .
\end{gather*}
where the backward generator generator \eqref{equ:bacward_generator_overdamped} is in polar coordinate system given by
\begin{align*}
{\mathcal L} [A] (r,\theta)
=& - \partial_r U(r) \, \partial_r A(r,\theta) + \frac{v(r)}{r}\,\partial_\theta A(r,\theta) + \\
&+ \frac{T}{r}\, \partial_r \left( r \partial_r A(r,\theta) \right)
+ \frac{T}{r^2}\, \partial^2_\theta A(r,\theta) .
\end{align*}
We will assume that the angular symmetry of the system also ensures the pseudoinverse will be again symmetric $G(r,\theta) \equiv G(r)$, which yields to
\[
-U'(r)\,G'(r) + \frac{T}{r}\,(r\,G'(r))' = w(r) - \langle w \rangle_\rho ,
\]
where we used the notation $f' \equiv \partial_r f(r)$.
By changing the variable $G(r) = g(z)$, $z = \frac{\lambda}{2T}\,r^2$, it simplifies to
\begin{gather*}
2 z\, \dot{g}(z) = h(z) , \\
\int\limits_0^\infty {\mathrm d} z \; g(z) \, {\mathrm e}^{-z} = 0 ,
\end{gather*}
where we denote $\dot g(z) \equiv \partial_z g(z)$
and $h(z)$ satisfy
\begin{align*}
\lambda \left[ \dot{h}(z) - h(z) \right] &= w(r) - \langle w \rangle_\rho , \\
h(0) &= 0 .
\end{align*}
The latter equation has the solution
\[
h(z) = \frac{\langle w \rangle_\rho}{\lambda}\,\left[1 - \bar\Gamma(\alpha + 1,z)\,{\mathrm e}^z \right]
\]
where
\[
\bar\Gamma(s,z) = \Gamma^{-1}(s) \int\limits_z^\infty {\mathrm d} t \; t^{s-1}\,{\mathrm e}^{-t}
\]
is the regularized incomplete gamma function.
Finally, the non-equilibrium correction to the equilibrium heat capacity is
\begin{equation}
\Delta C_F
= \int\limits_0^\infty {\mathrm d} z \; \left[ \partial_T g(z) - \dot g(z)\, \frac{z}{T} \right] {\mathrm e}^{-z}
= -\frac{1}{2T}\,\int\limits_0^\infty {\mathrm d} z \; h(z)\,{\mathrm e}^{-z}
= \frac{\alpha\langle w \rangle_\rho}{2\lambda\,T}
\label{equ:2D_diffusion_noneq_heat_capacity}
\end{equation}
We observe that the non-equilibrium correction is linearly growing with the steady entropy production rate, $\langle w \rangle_\rho / T$, and therefore that it depends on temperature as $\err[{\alpha-1}]{T}$.
It is positive for $\alpha > 0$ and negative for $-1 < \alpha < 0$.
The latter means that the full generalized heat capacity $C_F$ becomes negative whenever $\alpha < 0$ and $v$ large enough or $T$ small enough.
\begin{figure}[ht]
\caption{Non-equilibrium correction for the heat capacity for different values of $\alpha$.}
\label{pic:heat_capacity_2d_diffusion}
\begin{center}
\includegraphics[width=.9\textwidth,height=!]{2D_symmetric.pdf}
\end{center}
\end{figure}
Instead of the heat capacity $C_F$ at constant driving, here meaning $\kappa = {\mathrm {const.}}$, we can consider the heat capacity at constant dissipation, $C_W$, as defined under the constraint $\langle w \rangle_\rho = {\mathrm {const.}}$.
For the model under consideration, $\Delta C_W = \Delta C_F$ and also $C_W = C_F$ as an immediate consequence of the driving-independence of the stationary distribution $\rho$.
Nevertheless, the constant-dissipation heat capacity has a different temperature dependence $\Delta C_W \propto 1/T$, independently of the exponent $\alpha$.
One also checks that the heat capacity essentially does not depend on the mobility as long as the latter is homogeneous and isotropic.
Indeed, assuming a more general (scalar but possibly temperature-dependent) mobility $\chi(T)$, the formula~\eqref{equ:2D_diffusion_noneq_heat_capacity} gets only slightly modified
\begin{equation}
\Delta C_F = \frac{\alpha \langle w \rangle_\rho}{2\lambda\,T\,\chi(T)} \propto T^{\alpha - 1}
\label{equ:2D_diffusion_noneq_heat_capacity_mobility}
\end{equation}
Since $D = T \chi$ coincides with the diffusion parameter, we can read the result in the way that $\Delta C_F$ comes from the mutual ratio between the injected/dissipated energy $\langle w \rangle_\rho t$ and $\lambda D t$ which is a typical scale of system's energy changes by the diffusion process, both within same small time interval $t$.
One possible physical interpretation of the non-equilibrium correction in this particular model comes out by rewriting the formula~\eqref{equ:2D_diffusion_noneq_heat_capacity_mobility} in the form
\[
\Delta C_F = \frac{1}{2\lambda}\,\partial_T \left[ \chi^{-1} \langle w \rangle_\rho \right] .
\]
Since the relaxation mechanism is (at least in the $r-$sector) determined by the equilibrium $\vec{F} \equiv 0$ process, its characteristic time of relaxation is $\tau = \chi \lambda$.
Hence, the heat capacity correction goes like $\Delta C_F \propto \partial_T [\tau \langle w \rangle_\rho ]$, with only a numerical proportionality factor $1/2$.
Here the quantity $\tau \langle w \rangle_\rho$ reads the total amount of dissipated energy within the time needed to run through a relaxation process --- in some sense, one can consider that it measures the energy `available in the non-equilibrium surroundings' the changes of which contribute to the renormalized heat exchange; a bit analogously as the $pV-$term in the equilibrium enthalpy.
In this interpretation of the negative sign of the non-equilibrium correction term appears whenever the increase of temperature moves the system into a lower dissipation regime or, more precisely, into the regime with a lower ``available energy'' from the non-equilibrium driving --- typically when higher temperature effectively corresponds to weaker non-equilibrium as measured on the relaxation time-scale.
\subsection{Example: Diffusion in quadratic potential}
\label{ssec:diffusion_quadratic_potential}
As an example of the system where the Clausius relation holds true even far of equilibrium is the particle in quadratic potential in homogeneous media undergoing the driven overdamped diffusion in 2D, see also \ref{ssec:overdamped_diff},
here if compared with the example \ref{ssec:diffusion_on_plane} the quadratic potential is not in the center of the vortex created by non-potential force, see picture \ref{pic:2D_perturb}.
\begin{figure}[ht]
\caption{Scheme of the 2D model with quadratic potential.}
\label{pic:2D_perturb}
\begin{center}
\includegraphics[width=.6\textwidth,height=!]{2D_perturb.pdf}
\end{center}
\end{figure}
In order to simplify the analysis we assume that the mobility matrix $\chi$ does not depend on the temperature of the environment, in particular we can set the mobility to be one, $\chi = \mathbb I$.
The system's dynamics is then fully determined by the quadratic potential
\[
U(\vec{q}) = \frac{1}{2} \omega^2 \left( \vec{q} - \vec{q}_0 \right)^2
\]
and by the driving force
\[
\vec{F} (\vec{q}) = F {\mathbb J} \cdot \vec{q}
\]
where the matrix $\mathbb J$ creates an orthogonal vector to $\vec{q}$ and hence is given
\[
{\mathbb J} = \begin{bmatrix}
0 & -1 \\
1 & 0
\end{bmatrix} .
\]
The total force is then given by
\[
\vec{G}(\vec{q}) = - \omega^2 \left( \vec{q} - \vec{q}_0 \right) + F {\mathbb J} \cdot \vec{q}
= - {\mathbb A} \cdot \left( \vec{q} - \widetilde{\vec{q}}_0 \right)
\]
where the matrix $\mathbb A$ and the vector $\widetilde{\vec{q}}_0$ are given by
\begin{align*}
{\mathbb A} &= \begin{pmatrix}
\omega^2 & F \\
-F & \omega^2
\end{pmatrix} , &
\widetilde{\vec{q}}_0 &= \omega^2 {\mathbb A}^{-1} \cdot \vec{q}_0 .
\end{align*}
Notice that the matrix $\mathbb A$ is symmetric ${\mathbb A}^T = \mathbb A$ if and only if there is no non-potential driving.
One can check that whenever there exists a positively definite matrix $\mathbb K$ solving
\[
\mathbb K \cdot \mathbb A + \mathbb A^T \cdot \mathbb K = 2 \mathbb K^2
\]
then the steady state \eqref{equ:stationary_condition} is given by
\[
\rho(\vec{q}) = \frac{1}{Z} {\mathrm e}^{-\frac{\beta}{2} \left( \vec{q} - \widetilde{\vec{q}}_0 \right) \cdot \mathbb K \cdot \left( \vec{q} - \widetilde{\vec{q}}_0 \right) } .
\]
In this particular case the matrix $\mathbb K$ is given by
\[
\mathbb K = \omega^2 \mathbb I.
\]
Similarly the quasi-potential \eqref{def:generalized_potential} has the quadratic form
\[
V(\vec{q}) = \frac{1}{2} \left( \vec{q} - \widetilde{\vec{q}}_0 \right) \cdot \mathbb U \cdot \left( \vec{q} - \widetilde{\vec{q}}_0 \right) ,
\]
where $\mathbb U$ solves the equation
\[
\mathbb U \cdot \mathbb A + \mathbb A^T \cdot \mathbb U = 2 \mathbb A^T \cdot \mathbb A,
\]
which in this particular case is
\[
\mathbb U = \frac{F^2 + \omega^4}{\omega^2} \mathbb I.
\]
We can see that in this particular case the energy function governing the steady state is the multiple of the quasi-potential,
hence the stationary distribution is Boltzmann-like \eqref{equ:generalized_Boltzmann} with generalized inverse temperature
\[
\widetilde{\beta}(\alpha) = \beta \frac{\omega^4}{\omega^4 + F^2}.
\]
Hence we can see that in this particular model the Clausius relation holds true for arbitrary choice of parameters.
\section{Near equilibrium regime}
\label{sec:first_order_expansion}
It is well known fact the analysis of most non-equilibrium systems simplifies when we approach the equilibrium.
In this section we will analyse systems in the vicinity of equilibrium by providing a systematic expansion of the ``reversible'' heat and work up to the first order in presumably small non-equilibrium driving.
The underlying bases for this expansion was provided by McLennan, when he showed that the steady state of the system in the vicinity of equilibrium is approximately Boltzmann-like \cite{McLennan1959}.
Later Komatsu and Nakagawa proved the close relation of the proposition of McLennan to the transient fluctuation theorem \cite{Komatsu2008-2}.
In this section we however mostly follow the work of Maes and Neto\v{c}n\'{y} \cite{Maes2010}.
As we have already seen the Boltzmann-like steady state with an appropriate quasi-potential is a sufficient condition to obtain an extended Clausius relation, see subsection \ref{ssec:quasistatic_heat_and_work}.
We will show that in the linear order around equilibrium the Clausius equation is indeed valid.
\subsection{Steady state}
In order to simplify the discussion we restrict ourselves only to systems without any explicit time dependence in contact with the single thermal bath without velocity-like degrees of freedom, e.g. overdamped diffusion and Markov jump processes,
driven out of equilibrium only by a non-potential force $\vec{F}$.
Under these condition the magnitude of such force $\|\vec{F}\|$ can be used as the small parameter measuring the strength of the non-equilibrium driving.
By setting the $\vec{F}=\vec{0}$ we obtain an equilibrium dynamics represented by the forward Kolmogorov generator ${\mathcal L}^*_0$ with the steady state $\rho_0$ given by the equilibrium Boltzmann distribution, i.e.
\begin{align*}
0 &= {\mathcal L}^*_0 [ \rho_0 ] (x) , &
\rho_0(x) &= \frac{1}{Z_0} {\mathrm e}^{-\beta E(x)} ,
\end{align*}
which provides us the leading order of the expansion.
We formally expand the forward Kolmogorov generator and the steady state around the equilibrium ones ${\mathcal L}^*_0$ and $\rho_0$ up to the first order in the magnitude of driving and obtain
\begin{align*}
{\mathcal L}^* [\mu](x) &= {\mathcal L}^*_0 [\mu](x) + {\mathcal L}^*_1 [\mu](x) + \err[2]{\|\vec{F}\|} , \\
\rho(x) &= \rho_0(x) + \rho_1(x) + \err[2]{\|\vec{F}\|} ,
\end{align*}
where by index $1$ are denoted first order corrections.
Notice that in case of the overdamped diffusion the expansion of the forward Kolmogorov generator up to the first order is exact,
because the forward Kolmogorov generator \eqref{equ:time_evolution_overdamped} is a linear function of the applied force.
The steady state up to the first order in the magnitude of the non-equilibrium driving is then given by the solution of the stationary condition \eqref{equ:stationary_condition}, i.e. by
\[
0 = {\mathcal L}^*[\rho] (x) = {\mathcal L}^*_0 [\rho_1] (x) + {\mathcal L}^*_1[ \rho_0] (x) + \err[2]{\|\vec{F}\|} ,
\]
from where the first order correction can be found to be
\begin{equation}
\rho_1 (x) = - \frac{1}{{\mathcal L}^*_0} {\mathcal L}^*_1 \left[ \rho_0 \right] (x) ,
\label{equ:first_order_density}
\end{equation}
where $1/{\mathcal L}^*_0$ is the forward pseudoinverse \eqref{def:forward_pseudoinverse}.
The first step in order to associate the correction \eqref{equ:first_order_density} with an observable quantities,
is to realize that it is valid
\[
{\mathcal L}^*_1 [ \rho_0 ] (x) = - \beta \, w_1(x) \, \rho_0(x) + \err[2]{\| \vec{F} \|} ,
\]
where the $w_1(x)$ is the first order contribution the local power of non-potential forces.
One can show this either directly for each particular model, e.g. for jump processes see \cite{Maes2010}, or one can follow the general heuristic argument.
The ${\mathcal L}^*_1 [\rho_0]$ can be considered to be the time evolution of the equilibrium probability distribution with respect to the full dynamics,
i.e. we consider it to represent an initial stage of the relaxation process towards steady state starting from the equilibrium.
As the probability density has to be normalized at each time, the change of the probability distribution for particular configuration $x$ has to be compensated by induced probability currents from $x$.
Each probability current can also be related to the heat flux associated with the transition and also to the probability that given configuration is going to be occupied.
On the other hand the heat flux is solely induced by the action of non-potential forces during the relaxation process, while the potential forces are balanced in equilibrium, and hence can be directly related to the local power of these forces.
Thus concluding the heuristic argument.
The first order correction \eqref{equ:first_order_density} is hence given by
\[
\rho_1 (x) = \beta \frac{1}{{\mathcal L}^*_0} [ w \rho_0 ] (x) .
\]
One of the consequences of the global detailed balance \eqref{def:global_detailed_balance} for the equilibrium dynamics is the symmetry in correlation function with respect to equilibrium state
\begin{multline*}
\left\langle {\mathrm e}^{t {\mathcal L}_0}[A] \, B \right\rangle_{\rho_0}
= \iint {\mathrm d} \rho_0(x_0) \, \dcprob[{(0,T]}]{\omega}{X_0=x_0} \; A(x_T) \, B(x_0) = \\
= \iint {\mathrm d} \rho_0(x_T) \, \dcprob[{(0,T]}]{\Theta \omega}{X_0=x_T} \; A(x_T) \, B(x_0) = \\
= \iint {\mathrm d} \rho_0(x'_0) \, \dcprob[{(0,T]}]{\omega}{X_0=x'_0} \; A(x'_0) \, B(x'_T)
= \left\langle A \, {\mathrm e}^{t {\mathcal L}_0} [ B ] \right\rangle_{\rho_0} ,
\end{multline*}
which yields to the identity
\[
{\mathrm e}^{t {\mathcal L}_0} [ A ] (x) \, \rho_0 (x) = {\mathrm e}^{t {\mathcal L}_0^*} \left[ A \rho_0 \right] (x)
\]
or equivalently to
\[
{\mathcal L}^*_0 [ A \, \rho_0 ] (x) = {\mathcal L}_0 [A] (x) \, \rho_0(x) .
\]
As a consequence of this identity we express the forward pseudoinverse in the first order correction \eqref{equ:first_order_density} in terms of backward pseudoinverse and obtain
\[
\rho_1 (x) = \beta \frac{1}{{\mathcal L}_0} [ w_1 ] (x) \, \rho_0(x) .
\]
By putting all the terms of the stationary distribution altogether we obtain
\begin{equation}
\rho(x) = \frac{1}{Z_0} \exp \left[ - \beta \left( E(x) - \frac{1}{{\mathcal L}_0} [ w_1 ] (x) \right) \right] + \err[2]{\|\vec{F}\|} ,
\label{equ:first_order_rho}
\end{equation}
where we can see that the stationary probability distribution up to the first order in non-equilibrium driving effectively corresponds to the equilibrium system with an energy function given by
\[
\widetilde{E}(x) = E(x) - \frac{1}{{\mathcal L}_0} [w_1] (x) .
\]
\subsection{Work and heat}
We have seen that the steady state is given by the stationary distribution \eqref{equ:first_order_rho},
where the effective energy function resembles the quasi-potential present in the definition of the reversible work \eqref{def:reversible_work}.
In order to verify whether these terms are the same or if there is some fundamental difference we need to expand the quasi-potential up to the first order in the magnitude of non-equilibrium driving.
For that purpose we introduce the Dyson like series for backward pseudoinverse, which reads
\begin{equation}
\frac{1}{{\mathcal L}} [A] (x) = \frac{1}{{\mathcal L}_0} [A] (x) - \frac{1}{{\mathcal L}_0} {\mathcal L}_1 \frac{1}{{\mathcal L}} [A] (x) + \left\langle \frac{1}{{\mathcal L}} [A] \right\rangle_{\rho_0} .
\label{equ:Dyson_expansion}
\end{equation}
Notice that the main difference between the Dyson series and our extension is the presence of the last term which is associated with the difference between the kernels of ${\mathcal L}$ and ${\mathcal L}_0$.
It is quite straightforward that up to the first order only the first order term of the local power contributes, hence the quasi-potential simplifies to
\[
V(x) = E(x) - \frac{1}{{\mathcal L}} [w] (x) = E(x) - \frac{1}{{\mathcal L}} [w_1] (x) + \err[2]{\|\vec{F}\|} .
\]
We apply the Dyson-like expansion \eqref{equ:Dyson_expansion} and obtain
\[
V(x) = E(x) - \frac{1}{{\mathcal L}_0} [w_1] (x) - \left\langle \frac{1}{{\mathcal L}} [w_1] \right\rangle_{\rho_0} + \err[2]{\|\vec{F}\|} .
\]
Obviously,
\begin{multline*}
\left\langle \frac{1}{{\mathcal L}} [w_1] \right\rangle_{\rho_0} = \left\langle \frac{1}{{\mathcal L}} [w_1] - \frac{1}{{\mathcal L}_0} [w_1 ] \right\rangle_{\rho_0} = \\
= \int\limits_0^\infty {\mathrm d} t \; \left\{ \left( \left\langle w_1 \right\rangle_\rho - \left\langle w_1 \right\rangle_{\rho_0} \right) - \left\langle \left( {\mathrm e}^{t {\mathcal L}} - {\mathrm e}^{t {\mathcal L}_0} \right) [ w_1] \right\rangle_{\rho_0} \right\} = \\
= - \int\limits_0^\infty {\mathrm d} t \; \left\langle \left( {\mathrm e}^{t {\mathcal L}} - {\mathrm e}^{t {\mathcal L}_0} \right) [ w_1] \right\rangle_{\rho_0} + \err[2]{\|\vec{F}\|} ,
\end{multline*}
moreover it is also valid that
\[
\left( {\mathrm e}^{t {\mathcal L}} - {\mathrm e}^{t {\mathcal L}_0} \right) [ w_1] (x) = \int\limits_0^t {\mathrm d} s \; {\mathrm e}^{(t-s) {\mathcal L}_0} {\mathcal L}_1 {\mathrm e}^{s {\mathcal L}} [w_1 ] (x) = \err[2]{\|\vec{F}\|}
\]
hence the term $\left\langle \frac{1}{{\mathcal L}} [A] \right\rangle_{\rho_0}$ is at least of second order in non-equilibrium driving.
We can conclude that up to the first order in the non-equilibrium driving the ``reversible'' work is given by the quasi-potential which corresponds to the effective energy $\widetilde{E}(x)$.
As we have already seen if the ``reversible'' work is given by the same quasi-potential as the steady state, we have shown that the generalized Clausius relation \eqref{equ:Clausius_relation} is valid up to the first order in non-potential driving around equilibrium.
In this case the temperature is the actual temperature of the attached thermal bath and Shannon entropy is given as
\[
{\mathcal S} = \beta \left\langle E - \frac{1}{{\mathcal L}_0} [ w_1 ] \right\rangle_{\rho} + \ln Z_0 .
\]
Consequently the generalized heat capacity \eqref{def:generalized_heat_capacity} is always positive and given by
\[
C = \frac{1}{T^2} \left[ \left\langle V^2 \right\rangle_\rho - \left( \left\langle V \right\rangle_\rho \right)^2 \right] + \err[2]{\|\vec{F}\|} .
\]
\section{Low-temperature behaviour}
At beginning of the 20th century Nernst and Planck \cite{Callen1985} provided the last building block of the classical thermodynamics in the form of so called Nernst theorem or more commonly as the third law of thermodynamics.
The third law states that the for temperature going to absolute zero also the entropy needs to converge to zero, i.e. close to zero temperature the isothermal and adiabatic processes became indistinguishable.
It's precisely the third law of thermodynamics that ensures that many quasistatic response functions go to zero at zero temperature, e.g. the heat capacity under any constraint, thus ensuring practical inaccessibility of the absolute zero.
Although the third law seems to be valid in real systems there are models which does not obey the third law, e.g. the classical ideal gas.
In equilibrium the breaking of the third law of thermodynamics is closely related to the degeneracy of the ground state usually associated with the simplified description of the real physical system, which is not adequate at very low temperatures.
There is no definite answer to the question
whether the Third law of thermodynamics has a meaningful non-equilibrium generalization, e.g. in terms of vanishing generalized heat capacity in the zero temperature limit.
We have seen that the answer in case of the two- and three-level model was positive.
However in this sections we will show examples which violates the third law even among the systems with finite number of configurations and discuss the reasons behind such behaviour.
Let us stress once more that the classification of systems according to their low temperature behaviour is still an open question.
\subsection{Example: ``Merry-go-round'' model}
\label{ssec:merry-go-around}
The first example which we will study is a discrete model with $N$ states on the ring denoted as $x=1,2,\ldots,N$ ($N \geq 3$ and $N+1 \equiv 1$) and an extra state connected to all others denoted by $\odot$, see figure \ref{pic:merry-go-round}.
\begin{figure}[ht]
\caption{Illustration of the ``Merry-go-round'' model for $N=6$ with denoted possible transitions.}
\label{pic:merry-go-round}
\begin{center}
\includegraphics[width=.5\textwidth,height=!]{marry-go-around.pdf}
\end{center}
\end{figure}
Such system can represent a driven ratchet with the possibility to hurdle over several tooths at once.
The dynamics is fully determined by transition rates
\begin{align*}
\rate{x}{x\pm 1} &= {\mathrm e}^{\pm \beta F / 2} , \\
\rate{\odot}{x} &= {\mathrm e}^{-\frac{\beta(\Delta + U)}{2}} , \\
\rate{x}{\odot} &= {\mathrm e}^{-\frac{\beta(\Delta - U)}{2}} ,
\end{align*}
where $F$ denotes the work associated with driving forces, $U$ is the energy associated with the middle state $\odot$ and $\Delta>0$ is the height of the barrier which separates the middle state from the ring.
Given these transition rates, the stationary distribution is given by
\begin{align*}
\rho(\odot) &= \left(1 + N {\mathrm e}^{-\beta U} \right)^{-1} , \\
\rho(x) &= \left(N + {\mathrm e}^{\beta U} \right)^{-1} ,
\end{align*}
from where we can see that it is independent of the driving force $F$ as well as the height of the barrier $\Delta$.
On the level of occupations, we can interpret the situation as to have the `ground state' $\odot$ and $N$ identical `excitations' with energy gap $U$.
Despite its simplicity, our model exhibits a low-temperature crossover between prevailing conductor- and insulator-like behavior.
We compare the steady probabilistic current on the ring
\[
\jmath \equiv \curr{x}{x+1}
= \rho(x)\,[\rate{x}{x+1} - \rate{x+1}{x}]
= \frac{{\mathrm e}^{\frac{\beta F}{2}} - {\mathrm e}^{-\frac{\beta F}{2}}}{N + {\mathrm e}^{\beta U}}
\]
with the overall activity of the system represented by the overall ``number of transitions per second''
\begin{align*}
{\mathcal A} &= \sum_{x=1}^N \rho(x) \left[\rate{x}{x+1} + \rate{x}{x-1} + \rate{x}{\odot}\right]
+ \rho(\odot) \rate{\odot}{x}
\\
&= \frac{{\mathrm e}^{-\frac{\beta\Delta}{2}} + N({\mathrm e}^{\frac{\beta F}{2}} + {\mathrm e}^{-\frac{\beta F}{2}} + {\mathrm e}^{\frac{\beta(U - \Delta)}{2}})}{N + {\mathrm e}^{\beta U}} ,
\end{align*}
i.e. we calculate the relative number of those transitions contributing to transport and dissipation
\[
\frac{N \jmath}{{\mathcal A}}
= \frac{N \left( {\mathrm e}^{\frac{\beta F}{2}} - {\mathrm e}^{-\frac{\beta F}{2}} \right)}{{\mathrm e}^{-\frac{\beta\Delta}{2}} + N \left( {\mathrm e}^{\frac{\beta F}{2}} + {\mathrm e}^{-\frac{\beta F}{2}} + {\mathrm e}^{\frac{\beta(U - \Delta)}{2}} \right)} .
\]
With $N$ fixed, this ratio tends along $\beta \to +\infty$
\begin{enumerate}
\item to zero provided $|F| < U - \Delta$ (\emph{insulator}),
\item to unity on the condition $0 \neq |F| > U - \Delta$ (\emph{conductor}); in particular, it is always conductor in the case $\Delta > U$.
\end{enumerate}
This result can be also concluded from the heuristic picture:
For $|F| < U - \Delta$ the state $\odot$ is a preferred successor, i.e. transition rate $\rate{x}{\odot}$ is by far larger then transition rates $\rate{x}{x\pm1}$,
for all states on the ring and hence whenever an excitation from ground state occurs it is most likely to be suppressed by de-excitation within short time interval with marginal contribution to the current on the ring.
On the other hand, for $|F| > U - \Delta$ the preferred successor lies on the ring for every state and hence an excitation from the ground state $\odot$ is typically followed by a large number of rounds over the ring before it eventually vanishes, like when on a fast-enough rotating merry-go-round, thus inducing a large current.
In order to study the behaviour of the generalized heat capacity \eqref{def:generalized_heat_capacity} in low temperature asymptotic we need to find the quasi-potential $V_0 = V(\odot)$, $V_1 = V(1) = \ldots = V(N)$,
i.e. we need to solve the equation ${\mathcal L}[ V ] = q - \langle q \rangle_\rho$ with the backward Kolmogorov generator \eqref{equ:backward_Kolmogorov_gen_jump} being
\begin{align*}
{\mathcal L}[V]_0 &\equiv {\mathcal L}[V](\odot) = N\,{\mathrm e}^{\frac{-\beta(\Delta + U)}{2}} (V_1 - V_0) , \\
{\mathcal L}[V]_1 &\equiv {\mathcal L}[V](1) = \ldots = {\mathcal L}[V](N) = -{\mathrm e}^{\frac{-\beta(\Delta - U)}{2}} (V_1 - V_0)
\end{align*}
and the local heat production
\begin{align*}
q_0 & \equiv q(\odot) = -N U {\mathrm e}^{\frac{-\beta(\Delta + U)}{2}} , \\
q_1 & \equiv q(1) = \ldots = q(N) = U {\mathrm e}^{-\frac{\beta(\Delta - U)}{2}} + F \left({\mathrm e}^{\frac{\beta F}{2}} - {\mathrm e}^{-\frac{\beta F}{2}} \right)
\end{align*}
which yields
\[
V_0 - V_1 = U + F {\mathrm e}^{\frac{\beta(\Delta - U)}{2}} \frac{{\mathrm e}^{\frac{\beta F}{2}} - {\mathrm e}^{-\frac{\beta F}{2}}}{1 + N {\mathrm e}^{-\beta U}} .
\]
Hence, $V_0 - V_1 = U + \err{ |F| {\mathrm e}^{\frac{\beta(|F| - U + \Delta)}{2}} }$ where the non-equilibrium correction is exponentially damped in $\beta$ for the insulator $(|F| < U - \Delta)$,
whereas it exponentially diverges in the conductor regime.
Heuristically, the asymptotics follows by observing that in the conduction regime the particle started anywhere in the ring typically performs $\err{ \rate{x}{x+1} / \rate{x}{\odot} } \approx {\mathrm e}^{\beta(|F| - U + \Delta)}$ jumps along the ring (each contributing to the heat by $|F|$) before it finally jumps to $\odot$.
Finally, the generalized heat capacity \eqref{def:generalized_heat_capacity} is
\[
C = - \left\langle \frac{\partial V}{\partial (\beta^{-1})} \right\rangle_\rho
= N\beta^2 (V_0 - V_1) U {\mathrm e}^{-\beta U} \rho^2(\odot)
\]
In the conduction regime it goes asymptotically like
$C \asymp {\mathrm e}^{\frac{\beta(|F| - 3U + \Delta)}{2}}$ and hence it exhibits exponential divergence whenever $0 \neq |F| > 3U - \Delta$.
In particular, for $\Delta > 3U$ we have exponential damping for $F = 0$ (detailed balance), whereas already for an arbitrarily small $F \neq 0$ the heat capacity becomes divergent, see summary in table \ref{tab:merry-go-round}.
\begin{table}
\caption{Two regimes of the ``merry-go-round'' model.}
\label{tab:merry-go-round}
\begin{center}
\begin{tabular}{l|l|c}
insulator &
$|F| < U - \Delta$ &
$C \asymp {\mathrm e}^{-\beta U}$ \\ \hline
conductor &
$|F| > U - \Delta$ &
$C \asymp {\mathrm e}^{\frac{\beta(|F| - 3U + \Delta)}{2}}$ if $F \neq 0$ \\
&
&
$ \phantom{C} \asymp {\mathrm e}^{-\beta U}$ \phantom{******} if $F = 0$
\end{tabular}
\end{center}
\end{table}
We have seen that this simple model exhibits a rich behaviour in the zero temperature limit, where there is not only the transition between insulator and conductor,
but also a transition within the conductor regime, where the heat capacity becomes divergent in the zero temperature limit.
While the transition between conducting and insulator regime is easy to explain on the level of physical reality,
we have no simple physical explanation of the transition within the conducting regime yet.
\subsection{Example: Driven diffusion on the ring}
\label{ssec:diffusion_on_the_ring}
As a second example exhibiting the non-trivial low temperature behaviour we will investigate the diffusion on one-dimensional ring described by the stochastic differential equation \eqref{def:position_overdamped}
\begin{align*}
{\mathrm d} q_t &= f(q_t) \; {\mathrm d} t + \sqrt{2T}\,{\mathrm d} W_t , &
q_t &\in {\mathbb S}_1
\end{align*}
where the total force $f(q) = F - U'(q)$ contains the constant driving force $F$ and the $\mathbb S_1 $ is the unit length interval with periodical boundaries representing the ring.
The forward \eqref{equ:time_evolution_overdamped} and backward \eqref{equ:bacward_generator_overdamped} Kolmogorov generators are then given by
\begin{align*}
{\mathcal L} [A](q) &= f(q) A'(q) + T A''(q) , \\
{\mathcal L}^* [\mu](q) &= - (f(q) \mu(q))' + T \mu''(q) .
\end{align*}
We want to compute the generalized heat capacity \eqref{def:generalized_heat_capacity}
\begin{align*}
C_F &= \int {\mathrm d} q \; \partial_T \rho(q) \, \left( U(q) - \frac{1}{{\mathcal L}}[w](q) \right) , &
w(q) &= F f(q) ,
\end{align*}
with the driving $F$ constant.
From the stationarity equation \eqref{equ:stationary_condition} and linearity of the froward Kolmogorov generator we have
\[
\left(\partial_T {\mathcal L}^*\right) [\rho] (q) + {\mathcal L}^* [\partial_T \rho] (q) = 0
\]
and by using the normalization condition
\[
\int {\mathrm d} q \; \partial_T \rho(q) = 0,
\]
we obtain
\[
\partial_T \rho (q) = -\frac{1}{{\mathcal L}^*} [\rho''] (q)
\]
and hence
\begin{equation}
C_F = \int {\mathrm d} q \; \left\{ F f(q) \left(\frac{1}{{\mathcal L}^*}\right)^2 [\rho''] (q) - U(q) \frac{1}{{\mathcal L}^*} [\rho''] (q) \right\} .
\label{equ:1D_heat_capacity}
\end{equation}
Before we proceed to the analysis of the behaviour of the system under low temperatures we define the notion of the stable point as whether they are present will prove to be crucial in low temperature limit.
We say that $q^*$ is a stable point of the dynamics if the total force in such a point is zero and if forces in the vicinity of the stable point points towards it,
i.e. $U'(q*) = F$ and $U''(q^*) > 0$.
Notice that in case there is no random thermal force acting on the particle,
the position of the particle placed in the stable point does not evolve in time.
However the presence of random thermal forces ensures in the stochastic dynamics that the particle will eventually leave the point.
\subsubsection{``Insulator'' regime}
The first case which we will study is the case when there is a single stable point in the system.
We skip a rigorous treatment of the problem and only give a simple heuristic argument.
On the assumption that there exist a unique stable point $q*$, the low-temperature dynamics of the system is localized and hence it can be well approximated by the effective diffusion dynamics in the quadratic potential $U''(q^*) (q-q^*)^2/2$ on the full real line.
Hence, it becomes asymptotically indistinguishable from an equilibrium dynamics under the quadratic potential force, for which the equipartition theorem yields $C = 1/2$.
More detailed analysis reveal corrections exponentially damped when $T -> 0$.
\subsubsection{``Conducting'' regime}
In case there is no stable point where the forces can be balanced there is always a steady probability current even in case of low temperatures,
hence we speak about the conducting regime of the system,
i.e. the conducting regime then occurs whenever $| f(q) | > 0$ along the entire circle.
In this case we can evaluate the heat capacity \eqref{equ:1D_heat_capacity} in the zero temperature limit explicitly
and we reveal a remarkably different behaviour with respect to the insulator regime.
Using that the zero-temperature stationary density equals $\rho_0(q) = j_0 / f(q)$, with $j_0$ the stationary current, and by employing the identity
\begin{multline*}
\frac{1}{{\mathcal L}^*_0}[\mu'](q) = -\frac{1}{f(q)} \left[ \mu(q) - \left( \int {\mathrm d} \bar{q} \; \frac{1}{f(\bar{q})} \right)^{-1} \int {\mathrm d} \widetilde{q} \; \frac{\mu(\widetilde{q})}{f(\widetilde{q})} \right] = \\
= -\frac{1}{f(q)} \left[ \mu(q) - j_0 \int {\mathrm d} \widetilde{q} \; \frac{\mu(\widetilde{q})}{f(\widetilde{q})} \right]
= - \frac{\mu(q) - \left\langle \mu \right\rangle_{\rho_0}}{f(q)},
\end{multline*}
where by ${\mathcal L}_0^*$ we denote the zero temperature limit of the forward Kolmogorov generator $\lim_{T \to 0^+} {\mathcal L}^*$,
along with the periodic boundary conditions we get
\[
\left(\frac{1}{{\mathcal L}_0^*}\right)^2[\rho_0''] (q)
= -\frac{1}{{\mathcal L}_0^*} \left[ \frac{(\rho_0^2)'}{2j_0} \right] (q)
= \frac{\rho_0^3(q)}{2 j_0^2} - \frac{\rho_0(q)}{2 j_0^2} \int {\mathrm d} \bar{q} \; \rho_0^3(\bar{q})
\]
and finally
\begin{align*}
C_F^{T=0} &= \frac{F}{2j_0} \, \left[ \int {\mathrm d} q \; \rho_0^2(q) - \int {\mathrm d} q \; \rho_0^3(q) \right]
-\frac{1}{2j_0} \int {\mathrm d} q \; U'(q) \, \rho_0^2(q) \\
&= \frac{1}{2} - \frac{F}{2j_0} \int {\mathrm d} q \; \rho_0^3(q) .
\end{align*}
We can see that there is a clear distinction between the conduction and insulator regime represented as a zero temperature phase transition at $F= \max_{q} | U'(q) |$, where the zero temperature heat capacity suddenly changes from the (finite) equilibrium value to a divergent pattern, see for example the figure \ref{pic:1D_ring_lowT}.
Moreover we can see that in the large driving limit, $|F| \to \infty$, the low temperature heat capacity converges to zero as the effect of increased temperature is overwhelmed by the driving force and hence the stationary distribution converge to the uniform distribution over the ring.
Notice also that this behaviour corresponds to the physical intuition that in the singular driving limit the thermal force is negligible to the driving force and thus the system does not react to the changes of temperature.
\begin{figure}[ht]
\caption{$C_F^{T=0}$ for $U(q) = \sin(2\pi q)$.}
\label{pic:1D_ring_lowT}
\begin{center}
\includegraphics[width=.9\textwidth,height=!]{1D_ring_lowT.pdf}
\end{center}
\end{figure}
Let us also remark that the formula~\eqref{equ:1D_heat_capacity} can be written in the ``invariant'' form for arbitrary $T$,
\[
C_F = \int {\mathrm d} q \; (f^2(q) + T f'(q))\,\left(\frac{1}{{\mathcal L}^*}\right)^2[\rho''] (q)
= \left\langle \left[ \left(\frac{1}{{\mathcal L}}\right)^2 [ f^2 + T f' ] \right]'' \right\rangle_\rho
\]
which manifests that the generalized heat capacity is independent of how the forces are decomposed into the potential and non-potential components;
one may also note that $f^2 + T f' = {\mathcal L}[ V ]$ whenever $f = V'$.
\section{Conclusion}
In this chapter we have studied quasistatic processes in non-equilibrium system driven mostly by non-potential forces.
Our main result states that the heat and work can be naturally decomposed in the quasistatic limit to what we call ``housekeeping'' and ``reversible'' components, see \eqref{def:reversible_heat}, \eqref{def:reversible_work}, \eqref{def:housekeeping_work}.
We have identified the diverging ``housekeeping'' components with the steady heat and work production which is a consequence of the system being out of equilibrium.
On the other hand we have shown that the ``reversible'' components are geometric in the sense that they don't depend on the actual parametrization of the external protocol $\alpha$.
We have also shown that the ``reversible'' component of total heat/work is a natural generalization of the equilibrium concept of reversible heat/work.
By construction the ``housekeeping'' and ``reversible'' parts behave differently under the protocol inversion $\alpha \mapsto \Theta \alpha$: while the ``housekeeping'' part is symmetric, the ``reversible'' part is antisymmetric.
This enables us in principle to experimentally distinguish these components of the total heat by considering the cycle process composed of the forward protocol $\alpha$ and the backward protocol $\Theta \alpha$.
Another result is that the sum of ``reversible'' heat and work corresponds to the change of the internal energy defined as the mean value of the energy function,
which is to be interpreted as a (renormalized) quasistatic form of the law of energy conservation, see \ref{ssec:generalized_thermodynamics}.
We have also shown that the ``reversible'' heat is invariant with respect to the ``gauge'' transformation, see \ref{ssec:gauge_invariane},
which also enabled us to define the generalized heat capacity \eqref{def:generalized_heat_capacity} as generalized quasistatic response function.
We have studied the behaviour of the generalized heat capacity in several models \ref{ssec:two_level_nonp}, \ref{ssec:three_level_nonp} and \ref{ssec:diffusion_on_plane},
and seen that in strong non-equilibrium situations the heat capacity can be also negative.
A drawback of the presented renormalization scheme for the (quasistatically divergent) heat and work is that the Clausius relation \eqref{equ:Clausius_relation} isn't valid in general,
which is in good agreement with the fact that the generalized heat capacity can have negative values.
Nevertheless up to the first order in non-equilibrium driving \ref{sec:first_order_expansion} one proves the Clausius relation to be still valid as a consequence of the McLennan theorem.
We have found the general sufficient condition \eqref{equ:generalized_Boltzmann} for the system to obey the Clausius relation beyond the close-to-equilibrium regime, the question whether we can extend the class of such systems any further is still open.
A related open question is whether there exist general relations between (quasistatic) response functions analogous to equilibrium Maxwell relations.
Another open question concerns the possible extensions of the Nernst theorem to non-equilibrium systems, which we studied only briefly.
We have provided examples exhibiting a zero temperature phase transition and even a divergent low-temperature asymptotics of the steady heat capacity, see \ref{ssec:diffusion_on_plane}, \ref{ssec:merry-go-around} and \ref{ssec:diffusion_on_the_ring},
however the nature of such low temperature behaviour is still not well understood.
Notice also that we have been focused on mean values of the heat and work only,
whereas the detailed structure of fluctuations in the non-equilibrium quasistatic regime still remains a largely unexplored field.
This is also related to the question on the role of non-equilibrium response functions in the stability analysis of non-equilibrium steady states.
Main results of this chapter were published in \cite{result1,result2}.
|
\section{Introduction}
Granulation of information \cite{INT:INT20390,pedrycz2008handbook,6399462,Ulu20133713,liang2006information} emerges as an essential data analysis paradigm.
Information used or acquired to describe an abstract/physical/social process is usually expressed in terms of data (experimental evidence).
Therefore, granulation of information usually translates to data granulation.
Granulation of data can be roughly described as the action of aggregating semantically and functionally similar elements of the available experimental evidence.
This is performed to achieve a higher-level data description, which is implemented in terms of information granules (IGs) \cite{pedrycz2013granular}.
IGs are sound data aggregates that are formally described by a suitable mathematical model. Many mathematical settings have been proposed so far in the related literature, such as intervals--hyperboxes, (higher order) fuzzy sets, rough sets, and shadowed sets \cite{pedrycz2008handbook}.
The synthesized IGs can be used for interpretability purposes \cite{Mencar20084585,mencar_interpretability} or they can be used as a computational component of a suitable intelligent system \cite{apolloni2008interpolating,pedrycz2013granular,gralg_2012,pedrycz2011granular,si_asoc_grc,ganivada2011fuzzy,Pedrycz20083720,zhang2008granular,melin2013review,lu2014modeling}.
In any case, the problem of designing effective and justifiable data granulation procedures (GPs) is of paramount importance \cite{Choi20092102,6247495,usit2_2012,livi_hooman_it2fs_fitk,Castillo20115590,pedrycz2008dynamic,huang2013information,song2014human}.
The principle of justifiable granularity (PJG) is a well-established guideline for the synthesis of IGs \cite{6459568,Pedrycz20134209,pedrycz_timeseries}.
The PJG states that granulation should be performed by finding the ``optimal'' compromise among two conflicting requirements: specificity and generality. In other terms, an IG modeling input data should be designed such that it retains only the essential information (it should be specific, conveying a specific semantic content) but, at the same time, it should cover a reasonable amount of information.
Since the PJG is conceived to provide an adaptive mechanism to the information granulation problem, it is not designed to directly offer a built-in mechanism to objectively evaluate the quality of the granulation itself.
To this end, it is necessary to rely on external performance measures to quantify and judge over the quality of an IG.
The uncertainty is a peculiar property of virtually every human action that involves reasoning, decision making, and perception \cite{utkin2007decision,yager1992decision}.
Modeling the uncertainty of the input data is an essential mission in data granulation.
In fact, any IG model is designed to handle and hence express the uncertainty through an appropriate formalism. How the uncertainty is embedded into the IG model depends, of course, on the specific mathematical setting used for the IG.
However, while the numerical quantification of the uncertainty pertaining a specific situation may change as we change the mathematical setting of the IG, the \textit{level} of uncertainty should remain the same.
In these terms, the principles of uncertainty \cite{Klir199515,uncertainty_klir_2001} offer a compelling guideline to implement and evaluate practical data granulation techniques.
In this paper, we elaborate a conceptual data granulation framework over the principles of uncertainty.
A preliminary version of the herein exposed ideas appeared in \cite{entropy_preserving_graph_it2}. Here we further elaborate over this preliminary work by providing a more extensive discussion of the framework, offering new experiments that demonstrate the different facets underlying such ideas.
In the proposed framework we idealize the uncertainty as an ``invariant'' property, to be preserved as much as possible during the granulation of the input data.
As a consequence, we are able to objectively quantify the effectiveness of the granulation, regardless of the input data representation and the adopted IG model.
This interpretation allows also to quantitatively judge on a common groundwork different data granulation techniques operating on the same data.
We provide a demonstration of these ideas by discussing a data granulation technique that generates type-2 fuzzy sets (T2FSs).
This is paper is structured as follows. Sec. \ref{sec:poui} introduces the principles of uncertainty.
Throughout Sec. \ref{sec:granulation} we introduce the proposed conceptual framework for data granulation.
In Sec. \ref{sec:t2-minsod} we present a procedure to generate T2FSs by means of the minimum sum of distances (MinSOD) technique.
In Sec. \ref{sec:experiments} we discuss the experiments and related results. Sec. \ref{sec:conclusions} concludes the paper.
We provide two appendices: \ref{sec:t2fs} introduces to the context of T2FSs, while \ref{sec:minsod} the MinSOD.
\section{The Principles of Uncertainty}
\label{sec:poui}
The principles of uncertainty have been introduced by \cite{Klir199515} two decades ago, with the aim of providing high-level guidelines to the development of well-justified methods for problem solving in presence of uncertainty.
Such principles elaborate over the ubiquitous concepts of uncertainty and information.
It is intuitive to understand that uncertainty and information are intimately related: the reduction of uncertainty is caused by gaining new information, and vice versa.
Three principles have been introduced (quotes are taken from \cite{Klir199515}):
\begin{enumerate}
\item Principle of minimum uncertainty: ``It facilitates the selection of meaningful alternatives from solution sets obtained by solving problems in which some of the initial information is inevitably reduced in the solutions to various degrees. By this principle, we should accept only those solutions in a given solution set for which the information reduction is as small as possible.'';
\item Principle of maximum uncertainty: ``This is reasoning in which conclusions are not entailed in the given premises. Using common sense, the principle may be expressed by the following requirement: in any ampliative inference, use all information available but make sure that no additional information is unwittingly added.'';
\item Principle of uncertainty invariance: ``The principle requires that the amount of uncertainty (and information) be preserved when a representation of uncertainty in one mathematical theory is transformed into its couterpart in another theory.''.
\end{enumerate}
A combination of the first and third principle provides a compelling guideline for the purpose of data granulation. In fact, granulation of information implies mapping some input data (experimental evidence) originating from a certain input domain, say $\mathcal{X}$, to a domain of IGs, say $\mathcal{Y}$.
We argue that, when performing such a mapping, the uncertainty, regardless of the adopted formal mathematical framework, should be considered as an invariant property to be preserved as much as possible.
\section{Data Granulation with the Principles of Uncertainty}
\label{sec:granulation}
In this section, we introduce the proposed data granulation framework.
Fig. \ref{fig:mapping} illustrates the data granulation process. A procedure of data granulation can be formalized as a mapping, $\phi(\cdot)$, among two domains: input domain, $\mathcal{X}$, and the output domain, $\mathcal{Y}$.
$\mathcal{X}$ is the domain of the input data, whereas $\mathcal{Y}$ is a domain of IGs (e.g., a domain of hyperboxes, fuzzy sets, shadowed sets, rough sets and so on).
In practice, $\phi(\cdot)$ is a formal procedure for mapping a finite input dataset $\mathcal{S}\in\mathcal{P}_{<\infty}(\mathcal{X})$ with an output IG, say $\widetilde{\mathcal{A}}\in\mathcal{Y}$, i.e., $\widetilde{\mathcal{A}}=\phi(\mathcal{S})$.
Please note that we used a special mapping, $\mathcal{P}_{<\infty}(\cdot)$, in the input domain to allow discussing about $\mathcal{S}$ in terms of ``element'' of the input domain; in the following $\mathcal{P}_{<\infty}(\mathcal{X})$ is assumed to return all $n$-subsets of $\mathcal{X}$, with $n$ finite.
Note that $\mathcal{Y}$, as well as $\widetilde{\mathcal{A}}$, should be denoted by making explicit reference to $\mathcal{X}$ and $\mathcal{S}$, respectively, since IGs depend on the input. However, if no confusion is possible, we will avoid such specifications.
\begin{figure}[ht!]
\centering
\includegraphics[viewport=0 0 449 123,scale=0.43,keepaspectratio=true]{./mapping}
\caption{Data granulation as a mapping, $\phi: \mathcal{P}_{<\infty}(\mathcal{X})\rightarrow\mathcal{Y}$.}
\label{fig:mapping}
\end{figure}
There are a number of important questions that should be answered: ``Is the mapping $\phi(\cdot)$ well-justified? Moreover, how do we asses objectively the quality of the mapping?''; ``Are there invariant properties that must be preserved in the transformation from $\mathcal{S}$ to $\widetilde{\mathcal{A}}$?''; ``Can we numerically quantify those properties?''; ``Given two GPs, are we able to affirm that one performs a better granulation than the other by considering the same experimental conditions?''.
Reasoning over those questions provides important motivations for the design and formal evaluation of information GPs.
IGs are semantically sound constructs that are synthesized to convey higher-level information with respect to (w.r.t) the data from which they are generated \cite{pedrycz2013granular}.
All models used in information granulation \cite{pedrycz2008handbook} are designed to realize a ``simplification'' of the input data. The simplification consists in aggregating data that are considered indistinguishable (also termed indiscernible) and functionally/semantically related. IGs are hence designed also to handle the uncertainty caused by this simplification.
How the uncertainty is handled by the IG model depends on the specific mathematical setting used to describe the IG \cite{uncertainty_klir_2001}.
However, it is a reasonable assumption that, regardless of the specific mathematical setting, two IGs with different models, but synthesized from the same input data, should convey a comparable uncertainty, i.e., they should agree at least on the ``level of uncertainty''. The same concept holds for the uncertainty measured in the input with the one measured in the resulting output IG.
In the following, we formalize a conceptual framework to design and evaluate specific implementations of the mapping $\phi(\cdot)$. We refer to the proposed framework as the Principle of Uncertainty Level Preservation (PULP).
Usually, $\mathcal{X}$ is a domain of non-granulated data, such as $\mathbb{R}^d$ vectors, sequences of objects, or graphs. However, $\mathcal{X}$ can be conceived also as a domain of IGs. In this case, since the role of $\phi(\cdot)$ is to provide an abstraction, $\mathcal{Y}$ must be a domain of higher-level IGs w.r.t. those of $\mathcal{X}$.
In the following, however, we will consider mappings from input domains of non-granulated data types only.
\subsection{Minimization of the Input--Output Uncertainty Difference}
\label{sec:input-output_min}
First and most important component of PULP is a measure to calculate the uncertainty.
Let $\hat{H}: \mathcal{P}_{<\infty}(\mathcal{X})\rightarrow\mathbb{R}^+$ and $\check{H}: \mathcal{Y}\rightarrow\mathbb{R}^+$ be, respectively, the uncertainty measures for the input and output domain.
Experimental evidence is usually collected in the form of a finite dataset, $\mathcal{S}$, containing $n=|\mathcal{S}|$ patterns/samples proper of the input domain, $\mathcal{X}$. To provide a better description of the properties of the data in $\mathcal{S}$, usually it is idealized an underlying data generating process, $P$, which actually generates instances of $\mathcal{S}$.
This abstract process can be characterized by a deterministic analytical model known in closed-form, a non-deterministic model that is assumed to provide a suitable description of $P$, or an unknown model. In the last case, which is the most common one, the only useful information that is available is the finite dataset, i.e., $\mathcal{S}$.
For all practical purposes, the dataset $\mathcal{S}$ is usually assumed to be representative of all important statistics of the underlying process $P$.
Therefore, in the following we center our discussion on $\mathcal{S}$.
Let $\hat{H}(\mathcal{S})$ and $\check{H}(\phi(\mathcal{S}))$ be, respectively, the uncertainty calculated for the input dataset and the synthesized IG.
According to the guidelines conceptualized in PULP, we define the granulation error (GE) as
\begin{equation}
\label{eq:ge}
\delta = \lVert\hat{H}(\mathcal{S}) - \psi(\check{H}(\phi(\mathcal{S})))\rVert,
\end{equation}
where $\lVert\cdot\rVert$ is a norm and $\psi: \mathbb{R}^+\rightarrow\mathbb{R}^+$ is a monotonically increasing function that it is used to map the different formalizations of the uncertainty (see Fig. \ref{fig:mapping_uncertainty} for an illustration). $\psi(\cdot)$ must be monotone increasing since, reasonably, if the input uncertainty increases, then the output uncertainty must increase as well, although the increments could be of a different extent.
Eq. \ref{eq:ge} provides a formal way to affirm that $\phi(\cdot)$ is characterized by a GE equal to $\delta$, which is important also to compare different procedures operating on the same data. Clearly, the lower the error the better the procedure.
\begin{figure}[ht!]
\centering
\includegraphics[viewport=0 0 449 123,scale=0.43,keepaspectratio=true]{./mapping_uncertainty}
\caption{Function $\psi(\cdot)$ that provides a bridge among the input and output quantifications of the uncertainty.}
\label{fig:mapping_uncertainty}
\end{figure}
From the theoretical viewpoint, the best GP, given $\mathcal{S}$, is the one that minimizes the GE,
\begin{equation}
\label{eq:opt_ig_procedure}
\phi(\cdot)^* = \operatornamewithlimits{arg\ min}_{\phi(\cdot)\in\Phi} \lVert\hat{H}(\mathcal{S}) - \psi(\check{H}(\phi(\mathcal{S})))\rVert^2,
\end{equation}
where $\Phi$ is the set of all GPs suitable for the task at hand.
However, closed-form expressions for either $\phi(\cdot)$ and $\psi(\cdot)$ are necessary to evaluate either (\ref{eq:ge}) and (\ref{eq:opt_ig_procedure}). Moreover, a definition for the search space, $\Phi$, is also necessary in the case of Eq. \ref{eq:opt_ig_procedure}.
In practice, GPs are usually implemented as algorithms that most of times depend on some parameters, say $p$.
This is explicitly formalized by writing $\phi(\cdot; p)$. Therefore, to evaluate the quality of the mapping provided by a specific GP, we propose to deal with following optimization problem:
\begin{equation}
\label{eq:opt_params}
\delta^* = \min_{p} \lVert \hat{H}(\mathcal{S}) - \psi(\check{H}(\phi(\mathcal{S}; p))) \rVert^2.
\end{equation}
The optimal solution to (\ref{eq:opt_params}) yields the minimum GE, $\delta^*$, achievable for $\phi(\cdot)$.
Also $\delta^*$ can be used to objectively judge over the quality of the granulation provided by $\phi(\cdot)$.
The definition of the function $\psi(\cdot)$ is an important problem to be addressed. Such a function plays the role of the transformation among the different formalisms used to handle the uncertainty in the input (data) and output (granule) domains.
Defining $\psi(\cdot)$ in closed-form might be difficult, although it could be possible in specific cases (see Sec. \ref{sec:anal} for an example).
If this is the case, and also $\phi(\cdot)$ is available in closed-form, Eq. \ref{eq:opt_params} can be solved directly.
Optimization strategies to deal with (\ref{eq:opt_params}) depend on many factors, such as the nature of the uncertainty measures ($\hat{H}(\cdot)$ and $\check{H(\cdot)}$), and most importantly the specific implementation of $\phi(\cdot)$.
Providing general guidelines to the design of the specific optimization strategy is beyond the scope of this paper.
A universally valid method to obtain $\psi(\cdot)$ from a given problem instance, is via a suitable best-fitting algorithm. This can be performed, for instance, by analyzing $m$ i.i.d. dataset instances, $\mathcal{S}_i, i=1, 2, ..., m$, sampled from the same underlying data generating process.
The approximation of $\psi(\cdot)$, say $f(\cdot)$, is hence determined by fitting $m$ pairs of numbers obtained by the respective evaluations of $\hat{H}(\mathcal{S})$ and $\check{H}(\phi(\mathcal{S}; p))$.
The approximation would be characterized by a fitting error, $\epsilon\geq0$, which depends on $m$ and on the non-linearity of the underlying relation among the formalizations of the input/output uncertainty.
$\epsilon$ can be used in place of $\delta^*$ when the analytical definition of $\psi(\cdot)$ is not available.
Notably, the GE induced by the best-fitting error is defined as
\begin{equation}
\label{eq:ge_epsilon}
\epsilon = \lVert \hat{H}(\mathcal{S}) - f^*(\check{H}(\phi(\mathcal{S}))) \rVert,
\end{equation}
where $f^*(\cdot)$ is the optimal best-fitting function, which is derived by searching for the parameters $p$ of $\phi(\cdot)$ that generate $m$ samplings yielding the minimum fitting error.
\subsection{Brief Qualitative Discussion on the Proposed Framework}
The main contribution of PULP is a built-in formal criterion that is exploitable to judge over results of data granulation.
This is possible via the analysis of a quantity, GE, which is defined as the difference among the uncertainty measured in the input with the one calculated in the resulting output IG. In fact, the uncertainty in PULP is intended as an invariant property, to be preserved as much as possible during the granulation.
This particular aspect, which offers also a diagnostic tool of practical importance, is not included in the PJG. In fact, the PJG provides a guideline to design GPs by considering the two conflicting requirements of specificity (essentiality) and coverage -- an IG should be synthesized by finding a problem-dependent compromise among suitable implementations of those two factors.
Judging over the quality of the granulation is thus possible only indirectly, via human interpretation, by calculating some quality index, or by considering the synthesized IG as the input/component of another system (e.g., by using the IG in a classification system, evaluating thus the quality of the granulation as the accuracy of the classification).
If the input data contain outliers, a procedure designed by following the guidelines offered PULP would reflect also the contribution provided by those specific ``degenerate'' patterns; if not previously removed. In fact, the mapping is evaluated according to the capability of preserving the input uncertainty in the output IG model.
On the other hand, a procedure implemented according to the PJG, would determine an essential subset of input patterns, synthesizing an IG without considering those degenerate data.
Determining which approach offers a better solution, however, is something that depends on the context of application and on the ultimate needs of the user -- PULP is not proposed as a ``replacement'' for the PJG.
\section{A Type-2 Fuzzy Set Membership Functions Elicitation Method based on the MinSOD}
\label{sec:t2-minsod}
In this section we present a practical implementation of the mapping $\phi(\cdot)$ that is based on the MinSOD (see Sec. \ref{sec:minsod}).
The idea is to equip the MinSOD with the capability of generating an IG $\widetilde{\mathcal{A}}$ from $\mathcal{S}$, modeled as a T2FS (an introduction is provided in \ref{sec:t2fs}).
In the following, we refer to such a MinSOD as T2-MinSOD. Algorithm \ref{alg:t2minsod_algorithm} delivers the relative pseudo-code.
Given a dataset $\mathcal{S}$ and a dissimilarity measure $d:\mathcal{X}\times\mathcal{X}\rightarrow\mathbb{R}^+$, the T2-MinSOD can be described as a pair $(\nu, \mu_{\widetilde{\mathcal{A}}}(\cdot))$, where $\nu\in\mathcal{S}$ is the MinSOD representative and $\mu_{\widetilde{\mathcal{A}}}(\cdot)$ is the fuzzy membership function characterizing $\widetilde{\mathcal{A}}$.
Focusing on the IT2FS case, the membership degree of each $x\in\mathcal{S}$ is an interval $\mu_{\widetilde{\mathcal{A}}}(x)=[\mathrm{LMF}_{\widetilde{\mathcal{A}}}(x), \mathrm{UMF}_{\widetilde{\mathcal{A}}}(x)]$ in $[0, 1]$, bounded by LMF and UMF.
The element $\nu$ can be understood a suitable representative of $\mathcal{S}$ (a prototype). In fact, when $\mathcal{X}=\mathbb{R}$, $\nu$ corresponds to the median of the input dataset (see Claim \ref{th:Minsod}).
We exploit this fact in a more general setting, that is, when $\mathcal{X}$ is a user-defined data domain.
This allows us to interpret $\nu$, regardless of the nature of $\mathcal{X}$, as a well-justified representative of $\mathcal{S}$.
Accordingly, the evaluation of the UMF at $\nu$ should be one, denoting full membership.
Analogously, UMF for all other elements in $\mathcal{S}\setminus\{\nu\}$ should be defined by considering the dissimilarity value w.r.t. $\nu$.
This choice is motivated by the fact that we are trying to represent the uncertainty of $\mathcal{S}$ by relying only on the dissimilarity values among its elements.
The UMF is hence defined as a function of the dissimilarity w.r.t. $\nu$:
\begin{equation}
\label{eq:umf}
\mathrm{UMF}_{\widetilde{\mathcal{A}}}(x)=u(d(x, \nu)).
\end{equation}
$u(\cdot)$ is a monotonically non-increasing function of the argument yielding values in $[0, 1]$.
Candidate functions for (\ref{eq:umf}) are the Gaussian, rational quadratic kernel, or a linear functional in $[0, 1]$, like $M-d(x, y)$, where $M$ is the maximum value assumed by $d(\cdot, \cdot)$.
To form an interval membership function, we need to generate also the LMF.
The interval width in an IT2FS quantifies the uncertainty in describing the membership degree of an input element: the wider the interval, the higher the uncertainty.
LMF is determined as follows,
\begin{equation}
\label{eq:lmf}
\mathrm{LMF}_{\widetilde{\mathcal{A}}}(x)=\max\{\mathrm{UMF}(x) - l(D(x, \cdot)), 0\}.
\end{equation}
LMF is formed by considering the difference among the UMF and a function, $l(\cdot)$, of the dissimilarity w.r.t. the whole dataset -- note that $D(x, \cdot)$ denotes the set of dissimilarity values of all elements in $\mathcal{S}$ w.r.t. $x$.
$l(\cdot)$ could be implemented such that to capture the extent of the intra-granule dissimilarity values distribution (e.g., via the average or standard deviation etc.).
In this way, the uncertainty expressed by T2-MinSOD increases along with the diversity of the input patterns.
\begin{algorithm}[h!]\scriptsize
\caption{T2-MinSOD algorithm.}
\label{alg:t2minsod_algorithm}
\begin{algorithmic}[1]
\REQUIRE Dataset $\mathcal{S}$, a dissimilarity measure $d(\cdot, \cdot)$, functions $u(\cdot)$ and $l(\cdot)$
\ENSURE A T2-MinSOD, $(\nu, \mu_{\widetilde{\mathcal{A}}}(\cdot))$
\STATE According to Eq. \ref{eq:minsod}, determine $\nu$ on $\mathcal{S}$ using $d(\cdot, \cdot)$
\STATE Generate the interval-membership function, $\mu_{\widetilde{\mathcal{A}}}(\cdot)$, in the following loop
\FOR{\textbf{each} $x$ in $\mathcal{S}$}
\STATE Compute $\mathrm{UMF}_{\widetilde{\mathcal{A}}}(x)$ using $u(\cdot)$ with Eq. \ref{eq:umf}
\STATE Compute $\mathrm{LMF}_{\widetilde{\mathcal{A}}}(x)$ using $l(\cdot)$ with Eq. \ref{eq:lmf}
\ENDFOR
\RETURN $(\nu, \mu_{\widetilde{\mathcal{A}}}(\cdot))$
\end{algorithmic}
\end{algorithm}
T2-MinSOD could be exploited in many practical ways. For example, by using it (i) as an IG, thus providing a solution to analyze (and interpret) the uncertainty of $\mathcal{S}$, and (ii) as a computational component of a suitable intelligent system, which operates through data aggregation (e.g., a clustering-based procedure).
The T2-MinSOD does not just construct an IT2FS membership function, since in fact it allows also to easily defuzzify the granule by considering the representative $\nu\in\mathcal{S}$.
\section{Experiments}
\label{sec:experiments}
In Sec. \ref{sec:anal}, we discuss an example in which the input--output uncertainty mapping is solvable analytically.
Successively, we provide experiments considering T2-MinSOD operating in two different input domains: (i) Euclidean space and (ii) a domain of labeled graphs.
The first experiment (\ref{sec:exp_mv}) provides us also the possibility to visualize the results.
In Sec. \ref{sec:in-out_pres} we demonstrate that the T2-MinSOD is capable of generating IT2FS models that preserve the input uncertainty with reasonable GEs.
The second experiment (\ref{sec:exp_lg}) is performed considering several datasets of labeled graphs.
Finally, we discuss an experiment where T2-MinSOD is used in the clustering context (\ref{sec:clustering}).
Results are presented by implementing $u(\cdot)$ in Eq. \ref{eq:umf} as a Gaussian kernel -- dependent on the width -- and $l(\cdot)$ of Eq. \ref{eq:lmf} as the average.
We generate IGs modeled as IT2FSs. We rely on uncertainty measures based on entropy (see Refs. \cite{Wu:2009:CSR:1502817.1503036,Wu_UncIT2_2007} for detailed discussions on related measures of uncertainty for IT2FSs).
In particular, the uncertainty of the generated IT2FSs is computed by evaluating the (normalized) fuzzy entropy formulation given by \cite{Burillo1996305},
\begin{equation}
\label{eq:it2fs_entropy}
\check{H}(\widetilde{\mathcal{A}}) = \left(\frac{1}{|\mathcal{S}|}\sum_{i=1}^{|\mathcal{S}|} \overline{\mu}_{\widetilde{\mathcal{A}}}(x_i) - \underline{\mu}_{\widetilde{\mathcal{A}}}(x_i)\right) \in [0, 1].
\end{equation}
Analogously, we will characterize the uncertainty of the input by a suitable entropy measure.
\subsection{A Problem Solvable Analytically}
\label{sec:anal}
Let $\mathcal{S}=\{x_1, x_2, ..., x_n\}$ be a dataset of $n$ elements sampled from a Gaussian data generating process (also called source).
In this case, we calculate the (Shannon) entropy in closed-form \cite{cover2006elements}:
\begin{equation}
\label{eq:Gaussian_source_entropy}
\hat{H}(\mathcal{S}) = 1/2\ln(2\pi e \sigma^2),
\end{equation}
where $\sigma$ is the variance that completely characterizes the source (the higher the variance, the higher the entropy). For the purpose of this example, let us assume $\sigma\in[0, 1]$. Note that Eq. \ref{eq:Gaussian_source_entropy} actually holds for a dataset $\mathcal{S}$ as $n\rightarrow\infty$.
Now, let us define $\phi(\cdot)$ as a mapping that takes $\mathcal{S}$ and generates an IT2FS, $\widetilde{\mathcal{A}}$, with $\mu_{\widetilde{\mathcal{A}}}(x_i)=[0, \sigma], \forall x_i\in\mathcal{S}$.
Therefore, by evaluating (\ref{eq:it2fs_entropy}) on $\widetilde{\mathcal{A}}$, we have $\check{H}(\phi(\mathcal{S}))=\sigma$.
The definition of the function $\psi(\cdot)$ to map the uncertainty is straightforward.
In fact, $\psi(\check{H}(\phi(\mathcal{S})))=\psi(\sigma)=1/2\ln(2\pi e \sigma^2)=\hat{H}(\mathcal{S})$.
As a consequence, we can evaluate Eq. \ref{eq:ge} directly, obtaining $\delta= 0$, as $n\rightarrow\infty$.
It is worth noting that, according to Eq. \ref{eq:opt_ig_procedure}, such a GP $\phi(\cdot)$ is optimal.
\subsection{Tests on Real-valued Vectors}
\label{sec:exp_mv}
To grasp the concept in a data-driven scenario, in Fig. \ref{fig:1dvec} we show a sample of 100 patterns distributed according to a 1-dimensional Gaussian distribution with zero mean and $\sigma=0.2$.
In Fig. \ref{fig:1dvec_it2_w} we show the width (as a red spike) of the interval membership calculated by T2-MinSOD that characterizes each input pattern. As desired, patterns closer to the center of the distribution have a shorter interval width.
In Fig. \ref{fig:1dvec_it2} we show a representation of the generated IT2FS. It is possible to clearly recognize the Gaussian shape for both LMF and UMF. Notably, the interval-valued memberships are distributed as expected: the inner parts closer to the center are characterized by a ticker interval that, considering both endpoints, it is also closer to one.
\begin{figure*}[ht!]
\centering
\subfigure[Interval membership values generated for the input 1-dimensional patterns.]{
\includegraphics[viewport=0 0 346 247,scale=0.55,keepaspectratio=true]{./1dvec_it2_w}
\label{fig:1dvec_it2_w}}
~
\subfigure[Generated IT2FS.]{
\includegraphics[viewport=0 0 346 247,scale=0.55,keepaspectratio=true]{./1dvec_it2}
\label{fig:1dvec_it2}}
\caption{Widths of the interval memberships assigned to the input 1-dimensional patterns (\ref{fig:1dvec_it2_w}) and a representation of the generated IT2FS (\ref{fig:1dvec_it2}).}
\label{fig:1dvec}
\end{figure*}
Fig. \ref{fig:2dvec} shows a dataset distributed according to a 2-dimensional Gaussian distribution, with zero mean and spherical covariance matrix controlled by $\sigma=0.2$. Fig. \ref{fig:2dvec_distr} shows the obtained configuration of the 100 sampled patterns -- the green pattern is the computed MinSOD element. Fig. \ref{fig:2dvec_it2} shows the LMF and UMF of the generated IT2FS. It is worth noting that, in this case, input patterns have no trivial ordering, which justifies the visualization (\ref{fig:2dvec_it2}) of the generated IT2FS.
\begin{figure*}[ht!]
\centering
\subfigure[Distribution of the 2-dimensional input vectors.]{
\includegraphics[viewport=0 0 334 238,scale=0.55,keepaspectratio=true]{./2dvec}
\label{fig:2dvec_distr}}
~
\subfigure[Generated IT2FS.]{
\includegraphics[viewport=0 0 348 247,scale=0.55,keepaspectratio=true]{./2dvec_it2}
\label{fig:2dvec_it2}}
\caption{Example of the IT2FS generation over 100 patterns extracted from a 2-dimensional Gaussian distribution.}
\label{fig:2dvec}
\end{figure*}
\subsection{Preservation of Input--Output Uncertainty}
\label{sec:in-out_pres}
Here we calculate, in terms of GE, the quality of the granulation performed by the proposed T2-MinSOD.
We perform the test by generating the dataset $\mathcal{S}$ using a unidimensional Gaussian source (see Eq. \ref{eq:Gaussian_source_entropy} for the entropy expression) and a unidimensional exponential source.
The (Shannon) entropy of the exponential distribution is given by
\begin{equation}
\label{eq:exp_source}
\hat{H}(\mathcal{S}) = 1 - \ln(\lambda),
\end{equation}
where $\lambda>0$ is the scale parameter. From Eq. \ref{eq:exp_source}, it is clear that the entropy decreases as $\lambda$ increases.
For closed-form entropy formulas of other well-known distributions we refer the reader to \cite{Zografos200571}.
First, we performed a batch of 10 experiments by changing the variance, $\sigma$, of the Gaussian source in an suitable range.
Notably, $\sigma$ is progressively selected in $[0.1, 0.55]$, with an increment step of 0.05.
Fig. \ref{fig:bestfit_Gauss} depicts the results of the linear best-fitting (in the figure denoted as a function $f(\cdot)$) among the obtained 10 pairs of source and output entropy values.
The high coefficient of determination ($R^2\simeq 0.98$) denotes a very good relation among the input--output uncertainty; in this case, a linear model is sufficient to map the input--output uncertainty.
The optimization (\ref{eq:ge_epsilon}) to search for the optimal best-fit (dependent only on the width of the Gaussian implementing Eq. \ref{eq:umf}) is performed with a linear search on the $[0.01, 5]$ range with a step-size of 0.01.
Testing of the optimal best-fitting is performed on a new dataset instance generated with $\sigma=0.325$; the test is repeated 10 times by using different random initializations.
We found that $\epsilon \simeq 0.02 \pm 0.005$, which can be considered as a good result, demonstrating thus that the T2-MinSOD is capable of preserving the input uncertainty in the output IG model with a reasonable GE.
We repeated the experiment with an exponential source (\ref{eq:exp_source}), by varying $\lambda$ in $[1.5, 2.4]$ with an increment step of 0.1. The linear best-fit is once again sufficient ($R^2\simeq0.88$) to model the input--output relation among the entropic characterizations of the uncertainty -- see Fig. \ref{fig:bestfit_Exp}.
\begin{figure*}[ht!]
\centering
\subfigure[Gaussian source.]{
\includegraphics[viewport=0 0 342 243,keepaspectratio,scale=0.55]{./bestfit_gauss}
\label{fig:bestfit_Gauss}}
~
\subfigure[Exponential source.]{
\includegraphics[viewport=0 0 347 243,scale=0.55,keepaspectratio=true]{./bestfit_exp}
\label{fig:bestfit_Exp}}
\caption{Best fittings calculated by considering Gaussian (\ref{fig:bestfit_Gauss}) and exponential (\ref{fig:bestfit_Exp}) sources.}
\label{fig:best_fit}
\end{figure*}
\subsection{Tests on Labeled Graphs}
\label{sec:exp_lg}
We demonstrate the modeling capability of the T2-MinSOD when operating in the labeled graphs domain.
We consider both synthetic labeled graphs \cite{seriation+gradis_lncs_2012,odse,gralg_2012} and the letter dataset of the IAM repository \cite{riesen+bunke2008}; those datasets are originally conceived for benchmarking graph classifiers.
In the first case, we consider four out of the 15 original datasets -- datasets contain graphs constructed as Markov chains of decreasing similarity, i.e., related classification problems are intended with decreasing difficulty.
In the latter case we consider the letter dataset with two level of distortions: low (Letter-L) and high (Letter-H) -- this dataset contains graphs representing digitalized letters drawn over the 2D plane.
The dissimilarity measure (\ref{eq:minsod}) is implemented as the graph coverage graph matching algorithm \cite{livi2012_pgm}.
Fig. \ref{fig:graphs} shows the interval widths calculated for the four synthetic datasets of graphs (denoted as DS-G-2, DS-G-6, DS-G-10, and DS-G-14, where DS-G-2 induces a harder classification problems than DS-G-6 and so on).
As expected, the entropy (\ref{eq:it2fs_entropy}) calculated from the IT2FS models is in agreement with the nature of the datasets. More difficult (in terms of recognition) datasets have higher entropy than easier datasets; harder datasets are usually characterized by a less regular pattern organization in the input space.
A similar results holds for Letter-L and Letter-H, shown, respectively, in Figs. \ref{fig:letter1} and \ref{fig:letter2}. The IT2FS model of the easier dataset (Letter-L) denotes less entropy.
\begin{figure*}[ht!]
\centering
\subfigure[DS-G-2. IT2FS entropy=0.232693.]{
\includegraphics[viewport=0 0 348 247,scale=0.5,keepaspectratio=true]{./it2_w_graphs_ds2}
\label{fig:ds-g-2}}
~
\subfigure[DS-G-6. IT2FS entropy=0.21941.]{
\includegraphics[viewport=0 0 348 247,scale=0.5,keepaspectratio=true]{./it2_w_graphs_ds6}
\label{fig:ds-g-6}}
\subfigure[DS-G-10. IT2FS entropy=0.203789.]{
\includegraphics[viewport=0 0 348 247,scale=0.5,keepaspectratio=true]{./it2_w_graphs_ds10}
\label{fig:ds-g-10}}
~
\subfigure[DS-G-14. IT2FS entropy=0.187489.]{
\includegraphics[viewport=0 0 348 247,scale=0.5,keepaspectratio=true]{./it2_w_graphs_ds14}
\label{fig:ds-g-14}}
\caption{Widths and related entropy of the generated interval membership functions (synthetic graphs).}
\label{fig:graphs}
\end{figure*}
\begin{figure*}[ht!]
\centering
\subfigure[Letter-L. IT2FS entropy=0.553205.]{
\includegraphics[viewport=0 0 339 247,scale=0.5,keepaspectratio=true]{./graph_iam_letter1}
\label{fig:letter1}}
~
\subfigure[Letter-H. IT2FS entropy=0.628804.]{
\includegraphics[viewport=0 0 339 247,scale=0.5,keepaspectratio=true]{./graph_iam_letter2}
\label{fig:letter2}}
\caption{Widths and related entropy of the generated interval membership functions (IAM letter datasets).}
\label{fig:graphs_iam}
\end{figure*}
\subsection{T2-MinSOD in Data Clustering}
\label{sec:clustering}
In this section we use the T2-MinSOD to model clusters of data generated with the well-known \textit{k}-means algorithm \cite{cagata,Jain:2010:DCY:1755267.1755654}.
We process the Iris dataset taken from the UCI repository \cite{Bache+Lichman:2013}.
The Iris dataset contains 150 patterns equally distributed in three classes, named ``Iris-setosa'', ``Iris-versicolor'', and ``Iris-virginica''.
From the analysis of first two components of the PCA shown in Fig. \ref{fig:pca_iris}, it is possible to understand that patterns of the ``Iris-setosa'' class (in red) are well-separated from the others, while those of the other two classes show little overlap (according to the PCA space).
This fact suggests us that the uncertainty of a suitable IG modeling patterns belonging to the ``Iris-setosa'' class should be lower than those calculated from IGs describing the other two classes.
To test this hypothesis, we executed the \textit{k}-means algorithm directly in the input space (no pre-processing of data) setting $k=3$, generating thus three IGs modeled by means of the T2-MinSOD.
Fig. \ref{fig:kmeans} shows the widths of generated interval memberships. It is important to note that the first cluster (\ref{fig:uci_iris_c0}) consists of 50 patterns, all belonging to the ``Iris-setosa'' class. The entropy of the related IT2FS model is $\simeq 0.67$.
The other two clusters (shown in Figs. \ref{fig:uci_iris_c1} and \ref{fig:uci_iris_c2}) contain an unbalanced number of patterns: respectively 38 and 62. As fuzzy entropy calculations show, the uncertainty of those two IT2FS models is greater than the one of the first cluster, which agrees with the (visual) information provided by the PCA.
This result suggests that the IT2FS models generated by means of T2-MinSOD convey also useful and reliable higher-level information, to be exploited for interpretability purposes.
\begin{figure}[ht!]
\centering
\includegraphics[viewport=0 0 343 241,scale=0.6,keepaspectratio=true]{./PCA_Iris}
\caption{First two components of the PCA of the Iris dataset.}
\label{fig:pca_iris}
\end{figure}
\begin{figure*}[ht!]
\centering
\subfigure[First cluster. IT2FS entropy=0.670161.]{
\includegraphics[viewport=0 0 345 247,scale=0.45,keepaspectratio=true]{./UCI_Iris_C0}
\label{fig:uci_iris_c0}}
~
\subfigure[Second cluster. IT2FS entropy=0.794648.]{
\includegraphics[viewport=0 0 345 247,scale=0.45,keepaspectratio=true]{./UCI_Iris_C1}
\label{fig:uci_iris_c1}}
~
\subfigure[Third cluster. IT2FS entropy=0.885963.]{
\includegraphics[viewport=0 0 345 247,scale=0.45,keepaspectratio=true]{./UCI_Iris_C2}
\label{fig:uci_iris_c2}}
\caption{Widths and related entropy of the generated interval membership functions for the clusters calculated by $k$-means over the Iris dataset.}
\label{fig:kmeans}
\end{figure*}
\section{Conclusions and Future Directions}
\label{sec:conclusions}
The process of data granulation can be abstracted as a mapping among some input domain and a suitable domain of information granules.
In this paper, we have presented a conceptual framework to help designing and evaluating data granulation procedures.
The framework, called PULP, is based on the principles of uncertainty introduced by Klir.
The main idea is to consider the uncertainty of the input data as an invariant property, to be preserved as much as possible in the model of the output IG. The difference among the input and output uncertainty has been defined as the granulation error. Such a quantity has been used to (i) objectively judge over the quality of the granulation and (ii) to provide a common groundwork to compare different granulation procedures operating over the same data.
To put this idea in practice, we introduced a data granulation technique based on the MinSOD.
The procedure, called T2-MinSOD, is able to generate an interval-valued membership function by relying on the information of the input data dissimilarity values only.
We analyzed this procedure by considering different input data types and experimental settings.
Results show that T2-MinSOD is interpretable and it is able to preserve the uncertainty of the input with reasonable granulation errors.
Performing data granulation by considering the uncertainty as an invariant property to be preserved during the granulation process allows to apply this conceptual framework regardless 1) of the specific input data representation formalism and 2) the mathematical setting used to define information granules.
Future works include the theoretical consolidation of PULP.
For instance, it may be interesting to study if the mapping $\phi(\cdot)$ is bijective. This formal property may suggest important facts in terms of IG interpretability. In addition, we will study the so-called ``denagranulation'' by exploiting the invertibility of $\phi(\cdot)$.
Finally, we will use PULP for the purpose of benchmarking different data granulation procedures operating over the same data.
|
\section{Introduction}\label{sec introduction}
A midpoint algebra is a pair $(A,\oplus)$ where $A$ is a set and $\oplus$ a binary operation satisfying the following axioms:
\[\begin{array}{rl}
\text{(idempotency)} & x\oplus x= x,\\
\text{(commutativity)} & x\oplus y= y \oplus x ,\\
\text{(cancellation)} & (\exists a\in A, x\oplus a= y \oplus a)\Rightarrow x=y,\\
\text{(mediality)}&(x\oplus y)\oplus (z\oplus w)=( x \oplus z)\oplus (y\oplus w) .
\end{array}\]
The main example, which attracted our attention to this study, is the unit interval $A=[0,1]$ with $x\oplus y=\frac{x+y}{2}$.
The category of midpoint algebras is not a Mal'tsev category. Indeed, the usual order relation on the unit interval, with the arithmetic mean as above, is clearly a subalgebra of the product $[0,1]\times [0,1]$, which is reflexive and transitive but not symmetric, contradicting the well known characterization of Mal'tsev categories \cite{CKP,Carboni-Lambek-Pedicchio,Carboni-Pedicchio-Pirovano}. Nevertheless, the category of midpoint algebras does have some interesting properties weaker than those of Mal'tsev categories, namely that each object admits at most one internal monoid structure for each choice of a unit. To illustrate this aspect we use another example. Let $(A,\oplus)$ be the midpoint algebra with $A=]0,1]$, the set of positive real numbers smaller or equal than one, and the operation\[a\oplus b=\frac{2ab}{a+b}.\] It is clear that if $a$ and $b$ are positive real numbers then $a\oplus b$ is also positive; it is also easy to see that if $a$ and $b$ are less or equal than one then $a\oplus b$ is less or equal than one. A simple way to see it is to observe that the condition $a\oplus b\leq 1$ is equivalent to the condition\[0\leq a(1-b)+b(1-a).\] Hence the operation is well defined and it is easily checked that $(A,\oplus)$ is a midpoint algebra.
The diagram\[\xymatrix{A\times A\ar[r]^{m}&A&\{1\}\ar[l]},\] with \[m(x,y)=\frac{xy}{x+y-xy},\] is an internal monoid. This internal monoid structure is uniquely determined by the unit element $1\in A$ and there is no internal monoid structure for any other choice of a unit element. Moreover, it is not an internal group as, for example, $\frac{1}{2}$ has no inverse since the equation \[x=1+x\] has no solution in the real numbers.
While proving some of the results presented in this paper, we have observed that, in many cases, the idempotency law could be dealt with in a separate way. This suggested to us the study of the more general structure of commutative cancellative medial magmas, which, for simplicity, we will refer to as ccm-magmas. The medial law has been widely studied over the last three quarter of a century. In the context of magmas (formerly known as groupoids) this law was also referred to as bisymmetric, entropic or transposition, among other names. It has been studied either from an algebraic point of view or from a more geometrical perspective, see for example \cite{Aczel,Escardo,Jezek,Kannappan,Kermit70,Toyoda}.
The open-closed unit interval $]0,1]$ used in the example above has several other important structures of ccm-magma: that of a midpoint algebra (with the arithmetic mean), and that of a commutative monoid with cancellation (with the usual multiplication), resulting in an interplay between different ccm-magmas on the same set. The study of such interactions is certainly worthy. Nevertheless, for the moment, we dedicate our attention to some important aspects of the internal structures in the category of ccm-magmas, such as internal monoid, internal group and internal relation.
This paper is organized as follows: Section \ref{sec preliminaries} recalls some categorical notations and concepts such as the one of an internal monoid; Section \ref{sec ccm-magmas} introduces the category of ccm-magmas, gives some examples, and some useful lemmas are proven; Section \ref{sec mal'tsev} observes that the category of ccm-magmas is weakly Mal'tsev and characterizes the existence of an internal monoid structure on a given object, for every choice of a unit element; Section \ref{sec monoids} explores this property further by considering the notions of $e$-expansive, $e$-symmetric and homogeneous ccm-magmas (a short list of simple examples and counter-examples is also presented at the end of the section); and finally, Section \ref{sec relations} studies some properties of internal relations, namely symmetry, transitivity, reflexivity and difunctionality, which are not visible in the case of Mal'tsev categories.
\section{Preliminaries}\label{sec preliminaries}
The basic notions and notations from Category Theory used in this paper can be found, for instance, in \cite{MacLane}. Let us just recall a few of them. If $\mathbb{C}$ is any category with finite limits, an internal monoid in $\mathbb{C}$ is a diagram of the shape\[\xymatrix{A\times A\ar[r]^{m}&A&1\ar[l]_{e}}\] in which $A$ is any object in $\mathbb{C}$, $m$ and $e$ are morphisms in $\mathbb{C}$, $1$ denotes the terminal object and the following diagram is commutative (where $!_A$ denotes the unique morphism from $A$ to the terminal object $1$).\[\xymatrix{A^3\ar[r]^{1_A\times m}\ar[d]_{m\times 1_A}&A^2\ar[d]_{m}&A\ar@{=}[ld]\ar[l]_{\langle e!_A,1_A\rangle}\ar[d]^{\langle 1_A,e!_A\rangle}\\A^2\ar[r]^{m}&A&A^2\ar[l]_{m}}\]
An internal monoid $(A,m,e)$ is an internal group (see for example the Appendix in \cite{BB}) if and only if the diagram \[\xymatrix{A&A\times A\ar[r]^{\pi_2}\ar[l]_{m}& A}\] is a product diagram (the morphism $\pi_2$ is the canonical second projection), in other words, for every two morphisms $u,v\colon{X\to A}$ there exists a unique morphism, represented as $u-v$, from $X$ to $A$ such that $$m\langle u-v,v \rangle=u.$$ The inverse of a generalized element $x\colon{X\to A}$ is clearly $e-x$ with $e\colon{1\to A}$ the unit element. This is simply another way to say that there is a morphism $t\colon{A\to A}$ such that \[m\langle 1_A,t\rangle=e!_A=m\langle t,1_A\rangle.\]
This paper restricts itself to quasi-varieties of universal algebra \cite{Sankappanavar}, that is, categories in which the objects are sets equipped with an arbitrary family of finite-arity operations, satisfying a collection of axioms which may be either expressed as identities or as implications. All the results to be proven about conditions of uniqueness are easily generalized to a category with a faithful functor into the category of sets, preserving finite limits. However, the results involving existence conditions depend on the context and do not necessarily hold in general.
The notion of internal ccm-magma is also a natural one to be considered. It would be interesting, for example, to study topological ccm-magmas, i.e. ccm-magmas internal to the category of topological spaces \cite{Escardo,Kermit68}.
A congruence on an object $A$ is a subalgebra of $A\times A$, which, if considered as a relation, is reflexive, transitive and symmetric (see e.g. \cite{BB}). We will also consider difunctional relations: a relation $R\subseteq X\times Y$ is difunctional if the implication \[xRy,zRy,zRw\Rightarrow xRw\]holds for all $x,z\in X$ and $y,w\in Y$.
A Mal'tsev category is characterized by the property that every reflexive internal relation is a congruence relation, or equivalently, by the property that every internal relation is difunctional (\cite{CKP,Carboni-Lambek-Pedicchio,Carboni-Pedicchio-Pirovano} see also \cite{BB}). As we will see, this property does not hold in the categories we are considering. However, these have a weaker property: some types of reflexive internal relations are still congruences, and some types of internal relations are automatically difunctional.
\section{Ccm-magmas}\label{sec ccm-magmas}
A ccm-magma (commutative cancellative medial magma) may be obtained from a midpoint algebra simply by not requiring the idempotency axiom.
\begin{definition}\label{middle algebra}
A \emph{ccm-magma} is an algebraic structure $(A,\oplus)$ with a binary operation $\oplus$ satisfying the following axioms:
\begin{enumerate}
\item[M1] $a\oplus b=b\oplus a$
\item[M2] $a\oplus c=b\oplus c\implies a=b$
\item[M3] $(a\oplus b)\oplus (c\oplus d)=(a\oplus c)\oplus (b\oplus d)$
\end{enumerate}
\end{definition}
A morphism of ccm-magmas is simply a homomorphism, that is, a map preserving the binary operation.
It is known that the category of commutative cancellative magmas is weakly Mal'tsev (\cite{NMF2012}). The third axiom has an important consequence illustrated in the following proposition.
\begin{proposition}\label{prop1} If $f_1,f_2\colon{A\to B}$ are two morphisms of ccm-magmas, then the map $f\colon{A\times A\to B}$ defined by \[f(x,y)=f_1(x)\oplus f_2(y)\] is also a morphism of ccm-magmas.
\end{proposition}
\begin{proof}
The result is an immediate consequence of the axiom (M3).
\end{proof}
The result of Proposition \ref{prop1} stands even if commutativity is not included (see also \cite{Stein}, Theorem 12.8); in this case, among all medial-like identities studied in \cite{Polonijo}, (M3) is the most appropriate in proving this result. In fact, as we will see in Section \ref{sec mal'tsev}, even the weakly Mal'tsev property may be proven without commutativity. This paper considers the commutativity axiom. In a future work we will investigate a more general structure where (M1) is not included.
A simple variation of the previous result shows that for any two ccm-magmas $X$ and $Y$, there is a canonical morphism from product into co-product \[X\times Y\to X+Y\] sending each pair $(x,y)\in X\times Y$ to the element $\iota_X(x)\oplus \iota_Y(y)\in X+Y$, with $\iota_X\colon{X\to X+Y}$ and $\iota_Y\colon{Y\to X+Y}$ the canonical inclusions into the co-product. More generally, for any two given morphisms $f_1\colon{X\to B}$ and $f_2\colon{Y\to B}$ there is an induced morphism, say $f=f_1\oplus f_2\colon{X\times Y\to B}$, even thought there are no canonical injections to insert $X$ and $Y$ in $X\times Y$.
Next we present a short list of examples build up from well-known algebraic structures.
\begin{example}\label{ex1} A short list of simple examples:
\begin{enumerate}
\item The open unit interval $]0,1[$ with the operation \[x\oplus y=\frac{x+y}{2}\quad\text{or}\quad x\oplus y=\sqrt{xy}\] is a ccm-magma.
\item Every midpoint algebra is a ccm-magma.
\item The set of natural numbers with the usual addition is a ccm-magma.
\item Every commutative semigroup with cancellation is a ccm-magma.
\item If $A$ is a field, then the formula\[x\oplus y=a(x+y)+b\] gives a ccm-magma for every choice of $a,b\in A$ with $a\neq 0$.
\item If $A$ is any ring, then the formula\[x\oplus y=a(x+y)+b\] gives a ccm-magma for every choice of $a,b\in A$, with $a$ an invertible element.
\item If $V$ is any real or complex vector space, then the formula \[x\oplus y=\alpha(x+y)+b\]gives a ccm-magma on $V$ for every choice of a scalar $\alpha\neq 0$ and for every $b\in A$.
\item If $R$ is a ring and $A$ an $R$-module, then the formula \[x\oplus y=\alpha(x+y)+b\]gives a ccm-magma on $A$ for every choice of an invertible scalar $\alpha\in R$ and any $b\in A$.
\item If $A$ is the set of positive real numbers, then the formula\[x\oplus y=\frac{axy}{x+y}\]gives a ccm-magma for every choice of $a\in A$.
\item If $(A,+,\times,\cdot,0,1)$ is a commutative and associative classical algebra over a commutative ring $R$, then the formula \[x\oplus y=\alpha\cdot(x+y)+\beta\cdot(x\times y)\] gives a ccm-magma on the set $$\{x\in A\mid \alpha\cdot 1+\beta\cdot x\, \text{is invertible}\}$$ for every choice of scalars $\alpha,\beta\in R$ with $\alpha=\alpha^2$.
\item Every loop with $a(bc)=c(ba)$ is a ccm-magma, see for instance \cite{Stein}.
\item If $(A,\oplus)$ is a ccm-magma and $g\colon{A\to A}$ is a monomorphism, then the formula \[(x,y)\mapsto g(x\oplus y)\oplus a\] gives a ccm-magma on $A$, for every choice of $a\in A$.
\end{enumerate}
\end{example}
The last example is obtained directly from the proposition below by letting $f=k=1_A$ and $h=g$.
\begin{proposition}\label{prop2} Let $(A,\oplus)$ be a ccm-magma and $f,g\colon{A\to B}$ any two maps. If $f$ and $g$ are injective and there are maps $h,k\colon{A\to B}$ such that \[ f(g(x\oplus y)\oplus z)=(h(x)\oplus h(y))\oplus k(z), \quad x,y,z\in A,\] then the formula \[(x,y)\mapsto g(f(x)\oplus f(y))\oplus a\] gives a ccm-magma for every choice of $a\in A$.
\end{proposition}
\begin{proof}
The result is an immediate consequence of the axioms, combined with the hypotheses on the given maps. Commutativity is immediate. Cancellation follows from $f$ and $g$ being injective. To prove the medial law, on the one hand we have \[f(g(fx\oplus fy)\oplus a)\oplus f(g(fz\oplus fw)\oplus a)\] which by our assumptions on $f$ and $g$ simplifies to \begin{equation}\label{eq1prop2}((h(fx)\oplus h(fy))\oplus k(a))\oplus ((h(fz)\oplus h(fw))\oplus k(a));\end{equation} while on the other hand we have \begin{equation*}
f(g(fx\oplus fz)\oplus a)\oplus f(g(fy\oplus fw)\oplus a),
\end{equation*} which by our assumptions on $f$ and $g$ simplifies to \begin{equation}\label{eq2prop2}
((h(fx)\oplus h(fz))\oplus k(a))\oplus ((h(fy)\oplus h(fw))\oplus k(a)).
\end{equation} Finally we observe that expressions $(\ref{eq1prop2})$ and $(\ref{eq2prop2})$ are equal by a simple manipulation of axiom (M3) on the original $\oplus$, which completes the proof.
\end{proof}
We end this section with two lemmas which will be used later on while proving that some internal relations, in the category of ccm-magmas, are automatically difunctional. This contrasts with the case of Mal'tsev categories in which every internal relation is difunctional, and every internal reflexive relation is always a congruence.
\begin{lemma}\label{lemma1} Let $f,g\colon{A\times A\to B}$ be two morphisms in the category of ccm-magmas such that $f(a,a)=g(a,a)$ for every $a\in A$. Then:
\begin{enumerate}
\item[(i)] $f(a,b)=g(a,b),f(b,c)=g(b,c)\Rightarrow f(a,c)=g(a,c)$;
\item[(ii)] $f(a,b)=g(a,b)\Rightarrow f(b,a)=g(b,a)$.
\end{enumerate}
\end{lemma}
\begin{proof}
(i) We will show that if $f(a,b)=g(a,b)$ and $f(b,c)=g(b,c)$ then we always have \begin{equation}\label{eq1lemma1}
f(a,c)\oplus f(b,b)=g(a,c)\oplus g(b,b)
\end{equation} and since $f(b,b)=g(b,b)$ we use (M2) and get $f(a,c)=g(a,c)$. To show (\ref{eq1lemma1}) we observe:
\begin{eqnarray*}
f(a,c)\oplus f(b,b)&=&f(a\oplus b,c\oplus b)\\
&=&f(a\oplus b,b\oplus c)\\
&=&f(a, b)\oplus f(b,c)\\
&=&g(a, b)\oplus g(b,c)\\
&=&g(a\oplus b,b\oplus c)\\
&=&g(a\oplus b,c\oplus b)\\
&=&g(a,c)\oplus g(b,b).
\end{eqnarray*}
(ii) Since $f(a,a)=g(a,a)$ for every $a\in A$, we always have:
\begin{eqnarray*}
f(b,a)\oplus f(a,b)&=&f(b\oplus a,a\oplus b)\\
&=&f(a\oplus b,a\oplus b)\\
&=&g(a\oplus b,a\oplus b)\\
&=&g(b\oplus a,a\oplus b)\\
&=&g(b,a)\oplus g(a,b)
\end{eqnarray*}
now, if $f(a,b)=g(a,b)$ then, using (M2) we obtain $f(b,a)=g(b,a)$ as desired.
\end{proof}
\begin{lemma}\label{lemma2} Let $f,g\colon{X\times Y\to B}$ be any two morphisms in the category of ccm-magmas. Then:
\[\left(\begin{matrix}f(x,y)=g(x,y)\\f(z,y)=g(z,y)\\f(z,w)=g(z,w)\end{matrix}\right) \Rightarrow f(x,w)=g(x,w).\]
\end{lemma}
\begin{proof}
Assuming $f(x,y)=g(x,y)$ and $f(z,w)=g(z,w)$ we have:
\begin{eqnarray*}
f(x,w)\oplus f(z,y)&=&f(x\oplus z,w\oplus y)\\
&=&f(x\oplus z,y\oplus w)\\
&=&f(x, y)\oplus f(z,w)\\
&=&g(x, y)\oplus g(z,w)\\
&=&g(x\oplus z,y\oplus w)\\
&=&g(x\oplus z,w\oplus y)\\
&=&g(x,w)\oplus g(z,y).
\end{eqnarray*}
Now, if $f(z,y)=g(z,y)$ then we conclude that $f(x,w)=g(x,w)$ as desired, using axiom (M2).
\end{proof}
\section{Mal'tsev and weakly Mal'tsev properties}\label{sec mal'tsev}
We have already observed that the category of midpoint algebras, and hence the one of ccm-magmas, is not a Mal'tsev category. It is a weakly Mal'tsev category as a result of ccm-magmas being a subcategory of commutative cancellative magmas, known to be weakly Mal'tsev. Nonetheless, we present an alternative proof using the medial law instead of commutativity.
\begin{proposition}The category of ccm-magmas is a weakly Mal'tsev category.
\end{proposition}
\begin{proof}
From \cite{NMF2012} we know that if there exits a ternary term $p(x,y,z)$ such that \[p(x,y,y)=p(y,y,x)\] and \[p(x,a,a)=p(y,a,a)\Rightarrow x=y\] then the category is weakly Mal'tsev. Indeed this is the case for every ccm-magma $(A,\oplus)$ if we define \[p(x,y,z)=(y\oplus x)\oplus(z\oplus y).\qedhere\]
\end{proof}
In a weakly Mal'tsev category any diagram of the shape
\begin{equation}\label{couniv}
\vcenter{\xymatrix@!0@=4em{A \ar@<.5ex>[r]^-{f} \ar[rd]_-{u} & B
\ar@<.5ex>[l]^-{r}
\ar@<-.5ex>[r]_-{s}
\ar[d]^-{v} & C \ar@<-.5ex>[l]_-{g} \ar[ld]^-{w}\\
& D}}
\end{equation}
with $fr=1_{B}=gs$ and $u r=v=w s$, induces a bigger diagram
\begin{equation}\label{kite}
\vcenter{\xymatrix@!0@=3em{ & C \ar@<.5ex>[ld]^-{e_2} \ar@<-.5ex>[rd]_-{g}
\ar@/^/[rrrd]^-{w} \\
A\times_{B}C \ar@<.5ex>[ru]^-{\pi_2}
\ar@<-.5ex>[rd]_-{\pi_1} && B \ar@<.5ex>[ld]^-{r} \ar@<-.5ex>[lu]_-{s}
\ar[rr]|-{v} && D\\
& A \ar@<.5ex>[ru]^-{f} \ar@<-.5ex>[lu]_-{e_1} \ar@/_/[urrr]_-{u}}}\end{equation}
in which there is at most one morphism, say $\theta\colon{A\times_B C\to D}$, from the pullback $(A\times_B C,\pi_1,\pi_2)$ of the split epimorphism $f$ along the split epimorphism $g$ to the object $D$ such that $\theta e_1=u$ and $\theta e_2=w$. Where $e_1$ and $e_2$ are the induced morphisms of the form\[e_1(a)=(a,sf(a)),\quad e_2(c)=(rg(c),c).\]
The following result states under which conditions a diagram such as (\ref{couniv}) induces a morphism such as $\theta$.
\begin{proposition} Given a diagram such as $(\ref{kite})$ in the category of ccm-magmas, there is a (unique) morphism $\theta\colon{A\times_B C\to D}$, such that $\theta e_1=u$ and $\theta e_2=w$, if and only if, the equation \begin{equation}\label{equation xvuw}x\oplus v(b)=u(a)\oplus w(c)\end{equation} has a solution $x\in D$ for every $a\in A$, $b\in B$ and $c\in C$ with $f(a)=b=g(c)$. When that is the case, $\theta(a,c)=x$.
\end{proposition}
\begin{proof}
Suppose there is $\theta\colon{A\times_B C\to D}$ such that $\theta(a,s(b))=u(a)$ and $\theta(r(b),c)=w(c)$, where $b=f(a)=g(c)$, then $x=\theta(a,c)$ is always a solution to the equation (\ref{equation xvuw}). Indeed, using axiom (M1), we observe that \begin{eqnarray*}
\theta(a,c)\oplus v(b)&=&\theta(a,c)\oplus \theta(r(b),s(b))\\
&=& \theta((a,c)\oplus(r(b),s(b)))\\
&=& \theta(a\oplus r(b), c\oplus s(b))\\
&=& \theta(a\oplus r(b), s(b)\oplus c)\\
&=& \theta(a,s(b))\oplus \theta(r(b),c)\\
&=& u(a)\oplus w(c).
\end{eqnarray*}
Conversely, if the equation (\ref{equation xvuw}) has a solution (which is unique by axiom (M2)) for every $(a,c)\in A\times_B C$, then we may define a map $\theta\colon{A\times_B C\to D}$ which assigns the unique solution of (\ref{equation xvuw}) to every pair $(a,c)\in A\times_B C$. It remains to show that this map is a homomorphism. By axiom (M2), it suffices to prove that \[(\theta(a,c)\oplus\theta(a',c'))\oplus v(b\oplus b')=\theta(a\oplus a',c\oplus c')\oplus v(b\oplus b')\] for every $a,a'\in A$ and $c,c'\in C$ where \[f(a)=g(c)=b\in B \text{ and } f(a')=g(c')=b\in B.\] Now, because $u$, $v$ and $w$ are homomorphisms and, using axiom (M3), we have \begin{eqnarray*}
(\theta(a,c)\oplus\theta(a',c'))\oplus v(b\oplus b')&=&(\theta(a,c)\oplus\theta(a',c'))\oplus (v(b)\oplus v(b'))\\
&=&(\theta(a,c)\oplus v(b))\oplus (\theta(a',c')\oplus v(b'))\\
&=&(u(a)\oplus w(c))\oplus (u(a')\oplus w(c'))\\
&=&(u(a)\oplus u(a'))\oplus (w(c)\oplus w(c'))\\
&=& u(a\oplus a')\oplus w(c\oplus c')\\
&=&\theta(a\oplus a',c\oplus c')\oplus v(b\oplus b')
\end{eqnarray*} as desired, which completes the proof.
\end{proof}
In particular, any internal reflexive graph admits, at most, one structure of internal category. This is easily seen from the above result by choosing $D=A=C$, $r=s=v$, and $u$ and $w$ to be the identity morphisms. Even more particularly, by choosing $B$ to be a singleton, and if $(A,\oplus)$ is a ccm-magma then, for every idempotent $e\in A$, there is, at most, one internal monoid structure on $A$ which is compatible with the binary operation, that is, there exists at most one monoid $(A,*_e,e)$ such that \begin{equation}\label{condition * compatible with oplus}
(x*_ey)\oplus(z*_e w)=(x\oplus z)*_e(y\oplus w).
\end{equation}
Note that $e\in A$ must be an idempotent, so that the inclusion $\{e\} \to A$ may be a homomorphism.
\begin{corollary}\label{corollary1} Let $(A,\oplus)$ be a ccm-magma and $e\in A$ an idempotent element in $A$. There is a (unique) internal monoid structure $(A,*_e,e)$ in $A$, if and only if the equation \[\theta\oplus e=x\oplus y\] has a solution $\theta=\theta(x,y)\in A$ for every $x,y\in A$. In that case $x*_e y$ is given by $\theta(x,y)$.
\end{corollary}
\begin{proof}
It follows from the previous proposition with $A=D=C$, $B=1$ the terminal algebra, $f$ and $g$ uniquely determined while $r=s=v$ send the unique element in $1$ to the chosen element $e$ in $A$ (which is a homomorphism as soon as $e\oplus e=e$), and $u=w$ is the identity morphism. The fact that the operation $*_e$ is associative and has a unit $e$ follows from general arguments used in \cite{NMF2008} but may also be demonstrated directly.
\end{proof}
A particular case is when every element $a\in A$ can be decomposed as ${a=x_1\oplus x_2}$, studied in \cite{Jezek93} as division groupoids. In this case the property of having an internal monoid structure with unit element $e$ is equivalent to asking for a solution to the equation $x\oplus e=a$. We will further study these properties in the next section.
\section{Internal monoids and internal groups}\label{sec monoids}
It is well known \cite{CKP,Carboni-Lambek-Pedicchio,Carboni-Pedicchio-Pirovano} that, in Mal'tsev categories, every internal monoid is necessarily and internal group. This property does not apply to ccm-magmas. In this section, we will study some sufficient conditions for the existence of an internal monoid or group structure within a ccm-magma with a chosen unit element, which is necessarily an idempotent. For that we introduce the following notions, which have already been considered in the literature for different purposes, see for example \cite{Aczel-book,Jezek}:
\begin{definition}\label{def_homogeneous} Let $(A,\oplus)$ be a ccm-magma and consider any element $e\in A$. We will say that:
\begin{enumerate}
\item[(i)] $A$ is $e$-expansive if for every $a\in A$ there exists $2_e(a)\in A$ such that $2_e(a)\oplus e=a$;
\item[(ii)] $A$ is $e$-symmetric if for every $a\in A$ there exists $-_e(a)\in A$ such that $-_e(a)\oplus a=e$;
\item[(iii)] $A$ is homogeneous if it is $e$-expansive (or $e$-symmetric) for every $e\in A$.
\end{enumerate}
\end{definition}
In fact, a homogeneous ccm-magma is the same as a commutative medial quasi\-group (see for instance \cite{Leibak}).
A sufficient condition for a ccm-magma to admit an internal monoid structure with an idempotent element $e$ as its unit is to be $e$-expansive. When that is the case, then the internal monoid is a group if and only if the algebra is $e$-symmetric. Moreover, if every element is an idempotent, that is, if we have a midpoint algebra, then it is $e$-expansive if and only if there exists a monoid structure over $e$. Also every internal monoid (or group) is commutative and admits cancellation.
\begin{proposition}\label{prop expansive implies has a monoid} Let $(A,\oplus)$ be a ccm-magma and consider any idempotent element $e\in A$. If $A$ is $e$-expansive then $(A,*_e,e)$ is a monoid with $$x*_ey=2_e(x\oplus y).$$ Moreover it is a group if and only if $A$ is $e$-symmetric.
\end{proposition}
\begin{proof} If $A$ is $e$-expansive then in particular $2_e(x\oplus y)$ is a solution to the equation \[\theta\oplus e=x\oplus y,\] for every $x,y\in A$. From Corollary \ref{corollary1} we may conclude that $(A,*_e,e)$ is an internal monoid. Now, if moreover $A$ is $e$-symmetric then $-_e(a)$ is the inverse of $a$, for every $a\in A$. Indeed we have that \[-_e(a)*_ea=e\]is equivalent (by (M2)) to\[(-_e(a)*_e a)\oplus e=e\oplus e\] which holds because \begin{eqnarray*}
(-_e(a)*_e a)\oplus e = 2_e(-_e(a)\oplus a)\oplus e=-_e(a)\oplus a=e=e\oplus e.
\end{eqnarray*}
Conversely, if $(A,*_e,e)$ is a group, then, for every $a\in A$, its symmetric element, say $a'\in A$, is a solution to the equation $x\oplus a=e$. Indeed, since $a'$ is such that $a'*_e a=e$, or equivalently $2_e(a'\oplus a)=e$, we use axiom (M2) and the assumption that $e$ is an idempotent to conclude that \[e=e\oplus e= 2_e(a'\oplus a)\oplus e=a'\oplus a.\qedhere\]
\end{proof}
In the case when the operation $\oplus$ has the geometrical meaning of midpoint, the formula $a*_e b=2_e(a\oplus b)$ is intuitively illustrated via the following diagram.
\[\xymatrix@!@=2em{a\ar@{-}[rr]\ar@{-}[dd]\ar@{-}[rd]&&a*_eb\ar@{-}[ld]\ar@{-}[dd]\\&a\oplus b \ar@{-}[ld]\ar@{-}[rd]\\e\ar@{-}[rr]&&b}\]
As a more concrete example, let $A$ be any real vector space. Then, by defining \[a\oplus b=\frac{1}{2}(a+b)\] we obtain a ccm-magma which is $0$-symmetric and $0$-expansive with the usual interpretation of $-a$ and $2a$, as illustrated for the particular case of the real line.
\[\xymatrix{ \ar@{-}[r] & (-a)\ar@{=}[d]\ar@{-}[r] & (-a)\oplus a\ar@{=}[d] \ar@{-}[r] & a \ar@{=}[d] \ar@{-}[r] & 2a\ar@{=}[d] \ar@{->}[r] &\\\ar@{-}[r] & (-a)\ar@{-}[r] & 0 \ar@{-}[r] & 0\oplus (2a) \ar@{-}[r] & 2a \ar@{->}[r] &}\]
More generally, for every $e\in A$, this structure of ccm-magma is $e$-symmetric and $e$-expansive, with $2_e(a)=2a-e$ and $-_e(a)=2e-a$. This fact is related to the affine transformation $x\mapsto x+e$.
We also notice that if a ccm-magma is $e$-expansive and $e$-symmetric, for some element $e$, then it is so for all elements, in other words it is homogeneous. This result will be used in the proof of Corollary \ref{corollary3}.
\begin{proposition}\label{prop_homogeneous} Let $(A,\oplus)$ be a ccm-magma and $e\in A$ an element in it. If it is $e$-expansive and $e$-symmetric then it is homogeneous.
\end{proposition}
\begin{proof}
We have to prove that for every $u,v\in A$, there exists a solution $x$ to the equation $x\oplus u=v$. Indeed,
\[\begin{array}{rcl}
x\oplus u=v&\Longleftrightarrow&
(x\oplus u)\oplus(e\oplus -_e(u))=v\oplus(e\oplus -_e(u))\\
&\Longleftrightarrow&
(x\oplus e)\oplus(u\oplus -_e(u))=v\oplus(e\oplus -_e(u))\\
&\Longleftrightarrow&
(x\oplus e)\oplus e=v\oplus(e\oplus -_e(u))\\
&\Longleftrightarrow&
x\oplus e=2_e(v\oplus(e\oplus -_e(u)))\\
&\Longleftrightarrow&
x=2_e(2_e(v\oplus(e\oplus -_e(u)))),
\end{array}\] which gives the desired solution to the equation.
\end{proof}
As already referred, every homogeneous ccm-magma is a commutative medial quasigroup. This means that for homogeneous ccm-magmas, Proposition \ref{prop expansive implies has a monoid} is a special case (commutative) of the well known Toyoda Theorem \cite{Toyoda}. This theorem has been generalized for medial magmas with cancellation (see for instance \cite{Jezek}), but the result is no longer comparable with the one of an internal monoid structure.
Restricting the study to homogeneous ccm-magmas has some advantages but it forces the category to be Mal'tsev (and hence there is no longer the distinction between internal monoid and internal group). Indeed, adapting the well-known formulas describing the category of quasigroups as a Mal'tsev category, say from \cite{Edward}, we can conclude that if a ccm-magma is expansive for every element then \[p(x,y,z)=2_{2_y(y)}(x)\oplus 2_y(z)\] is a Mal'tsev term.
We continue with another aspect of ccm-magmas which will be used in Proposition \ref{prop_iso_monoids}. If a given ccm-magma is expansive with respect to some idempotent element $e$ then the respective map $2_e$ is a homomorphism. In general, we have:
\begin{proposition}\label{prop_exp_symm_morphism}
Let $(A,\oplus)$ be a ccm-magma. If it is $u$-expansive and $v$-expansive for some $u,v\in A$, then it is also $(u\oplus v)$-expansive, and moreover, \[2_u(a)\oplus 2_v(b)=2_{u\oplus v}(a\oplus b).\]
\end{proposition}
\begin{proof}
Suppose there exists $2_{u\oplus v}(a\oplus b)$, then we necessarily have $$2_{u}(a)\oplus 2_{v}(b)=2_{u\oplus v}(a\oplus b)$$ because, by \emph{adding} $u\oplus v$ in each term, we obtain $a\oplus b$. It remains to prove that $2_{u\oplus v}$ exists. In other words we have to prove that for every $a\in A$, there is $x\in A$ such that $x\oplus (u\oplus v)=a$. We claim that \[x=2_u(2_u(a))\oplus 2_v(u)\]is the needed solution. Indeed,
\begin{eqnarray*}
(2_u(2_u(a))\oplus 2_v(u))\oplus(u\oplus v)&=&(2_u(2_u(a))\oplus u)\oplus(2_v(u)\oplus v)\\
&=&2_u(a)\oplus u\\
&=& a.
\end{eqnarray*}
So, we get $2_{u\oplus v}(a)=2_u(2_u(a))\oplus 2_v(u)=2_v(2_v(a))\oplus 2_u(v)$.
\end{proof}
The next result explains the connection between two induced monoid structures for two different idempotents $u$ and $v$.
\begin{proposition}\label{prop_iso_monoids} Let $(A,\oplus)$ be a ccm-magma which is $u$-expansive and $v$-expansive for some idempotent elements $u,v\in A$. The two monoid structures on $A$, induced by $u$ and $v$, are isomorphic. Moreover the two structures are related as follows: \[a*_u b= (a*_v b)*_u v.\]
\end{proposition}
\begin{proof}
Let $a*_u b=2_u(a\oplus b)$ and $a*_v b=2_v(a\oplus b)$ be the two monoid operations induced, respectively by $u$ and $v$, assuming $u$ and $v$ to be idempotent and $A$ to be $u$-expansive and $v$-expansive.
The map $f\colon{(A,*_u,u)\to (A,*_v,v)}$, such that $f(a)=2_u(a\oplus v)$, $a\in A$, is a monoid homomorphism. Clearly, the units are preserved, since \[f(u)=2_{u}(u\oplus v)=2_{u}(v\oplus u)=2_u(v)\oplus 2_u(u)=2_u(v)\oplus u=v.\] From
\begin{eqnarray*}
f(a*_u b)\oplus v&=&f(2_u(a\oplus b))\oplus v\\
&=&2_u(2_u(a\oplus b)\oplus v)\oplus v\\
&=&2_u(2_u(a\oplus b)\oplus v)\oplus (u\oplus 2_u(v))\\
&=&(2_u(2_u(a\oplus b))\oplus 2_u(v))\oplus (u\oplus 2_u(v))\\
&=&(2_u(2_u(a\oplus b))\oplus u)\oplus (2_u(v)\oplus 2_u(v))\\
&=&2_u(a\oplus b)\oplus (2_u(v\oplus v)
\\&=&2_u(a\oplus b)\oplus (2_u(v)\\
&=&2_u((a\oplus b)\oplus v)
\end{eqnarray*}
and
\begin{eqnarray*}
(f(a)*_v f(b))\oplus v&=&2_v(f(a)\oplus f(b))\oplus v\\
&=&f(a)\oplus f(b)\\
&=&2_u(a\oplus v)\oplus 2_u(b\oplus v)\\
&=&2_u((a\oplus v)\oplus (b\oplus v))\\
&=&2_u((a\oplus b)\oplus (v\oplus v))\\
&=&2_u((a\oplus b)\oplus v)
\end{eqnarray*}
we may conclude that $f(a*_u b)=f(a)*_v f(b)$. The inverse homomorphism of $f$ is $g\colon{(A,*_v,v)\to (A,*_u,u)}$ with $g(a)=2_v(a\oplus u)$. Indeed,
\[gf(a)=2_v(2_u(a\oplus v)\oplus u)=2_v(a\oplus v)=2_v(a)\oplus 2_v(v)=2_v(a)\oplus v=a\]
and similarly we prove $fg(a)=a$.
Finally, we prove $a*_u b= (a*_v b)*_u v$ by observing that \[(a*_u b)\oplus u=a\oplus b= ((a*_v b)*_u v)\oplus u.\qedhere\]
\end{proof}
In some cases, a ccm-magma $(A,\oplus)$ does not only admit an internal monoid structure over some idempotent element $e$, but also the structure itself is a monoid with unit element $e$, that is $a\oplus e=a$. This property is summarized in the following proposition.
\begin{proposition}\label{prop ccm-magma associative}
Let $(A,\oplus)$ be a ccm-magma and $e\in A$ an idempotent. The following conditions are equivalent:
\begin{enumerate}
\item[(i)] the operation $\oplus$ is associative;
\item[(ii)] the element $e\in A$ is a unit element for $\oplus$;
\item[(iii)] the ccm-magma is $e$-expansive with $2_e(a)=a$;
\item[(iv)] the structure $(A,\oplus, e)$ is an internal monoid.
\end{enumerate}
\end{proposition}
\begin{proof}
If the operation $\oplus$ is associative and $e\in A$ is idempotent then \[(a\oplus e)\oplus e=a\oplus (e\oplus e)=a\oplus e\] and hence $a\oplus e=a$.
If $e\in A$, an idempotent, is also a unit element for $\oplus$ then $2_e(a)=a$ by definition of $2_e$.
We already know that if the ccm-magma is $e$-expansive then $(A,*_e,e)$ is an internal monoid structure with $x*_e y =2_e(x\oplus y)$. When $2_e$ is the identity map, the operation $*_e$ is simply the original $\oplus$.
This proves $\text{(i)}\Rightarrow\text{(ii)}\Rightarrow \text{(iii)}\Rightarrow\text{(iv)}$, whereas $\text{(iv)}\Rightarrow \text{(i)}$ is obvious.
\end{proof}
Note also that if a ccm-magma $(A,\oplus)$ is associative, then it has, at most, one idempotent. Indeed, if $e_1$ and $e_2$ are two idempotents, and $\oplus$ is associative, then \[e_1\oplus (e_1\oplus e_2)=e_1\oplus e_2=e_2\oplus(e_2\oplus e_1)\]from which it follows that $e_1=e_2$.
In the case of midpoint algebras, that is, when every element is an idempotent, there is no distinction between having a monoid structure for some element $e\in A$ and being $e$-expansive.
\begin{proposition}\label{prop midpoint and expansive}
Let $(A,\oplus)$ be a midpoint algebra with $e\in A$. The following conditions are equivalent:
\begin{enumerate}
\item[(i)] there is an internal monoid structure over $e$;
\item[(ii)] for every $a,b\in A$, there is $x\in A$ such that $x\oplus e=a\oplus b$;
\item[(iii)] the midpoint algebra is $e$-expansive.
\end{enumerate}
\end{proposition}
\begin{proof} The two conditions (i) and (ii) are already equivalent in the context of ccm-magmas (Corollary \ref{corollary1}). Also, in the more general case of ccm-magmas, (iii) implies (i), as it is proven in Proposition \ref{prop expansive implies has a monoid}. We are left to prove that (i) implies (iii). Assuming an internal monoid structure $(A,*_e,e)$ and the fact that every element $a\in A$ is idempotent ($A$ is a midpoint algebra), we have \[2_e(a)=2_e(a\oplus a)=a*_e a.\qedhere\]
\end{proof}
If, in a midpoint algebra $(A,\oplus)$, there is an internal monoid structure $(A,*_e,e)$, then $*_e$ is distributive over $\oplus$, as it immediately follows from $(\ref{condition * compatible with oplus})$. This is not true in general for ccm-magmas.
\begin{proposition}
Let $(A,\oplus)$ be a ccm-magma with an internal monoid structure $(A,*_e,e)$. The following two conditions are equivalent:
\begin{enumerate}
\item[(i)] the ccm-magma is a midpoint algebra (every element $a\in A$ is an idempotent);
\item[(ii)] for every $x,y,z\in A$, \begin{equation}\label{distribituvity}
x*_e(y\oplus z)=(x*_e y)\oplus (x*_e z).
\end{equation}
\end{enumerate}
\end{proposition}
\begin{proof} If $(A,\oplus)$ is a midpoint algebra then $(\ref{distribituvity})$ follows from $(\ref{condition * compatible with oplus})$.
Conversely, if we have an internal monoid structure $(A,*_e,e)$ together with $(\ref{distribituvity})$, then every element $a\in A$ is an idempotent \[a\oplus a=(a*_e e)\oplus(a*_e e)=a*_e(e\oplus e)=a*_e e=a.\] Note that $e$ is an idempotent because it is the unit element of an internal monoid.
\end{proof}
\subsection*{Some examples illustrating the properties discussed in the previous results}
This section concludes with a short list of examples and counter-examples showing several particular cases and properties, which a given ccm-magma may or may not have for some specific choice of an idempotent element $e$ in it. First we observe that if the four basic properties in the left column on the table below are considered, then there are only six possible combinations between them, namely the ones expressed in the other columns and denoted by I to VI:
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
& I & II & III & IV & V & VI\\
\hline
$e$-expansive & yes & yes & no & no & no & no \\
\hline
$e$-symmetric & yes & no & yes & no & no & yes \\
\hline
internal monoid over $e$ & yes & yes & no & no & yes & yes \\
\hline
internal group over $e$ & yes & no & no & no & no & yes \\
\hline
\end{tabular}
\end{center}
The properties I to VI defined in the table above have an obvious interpretation. For example, a ccm-magma has property III if and only if it is $e$-symmetric but not $e$-expansive for a specific choice of $e$, it does not have an internal group structure over $e$ or even an internal monoid; while the ccm-magmas with property V have an internal monoid structure over some idempotent element $e$, they are not $e$-expansive nor $e$-symmetric, and consequently do not posses an internal group structure.
Now, combining the previous properties with the existence of one or more, none, or even all idempotents, and also with associativity, we observe the following list of ccm-magmas $(A,\oplus)$.
\begin{enumerate}
\item Ccm-magmas with no idempotent elements (in this case there is no interaction with properties I to VI from above):
\begin{enumerate}
\item Non-associative:
\begin{enumerate}
\item $a\oplus b=\frac{a+b}{2}+1$, $A=\mathbb{R}$
\item $a\oplus b=\frac{3ab}{a+b}$, $A=\mathbb{R}^{+}$
\item $a\oplus b=2(a+b)$, $A=\mathbb{R}^{+}$
\end{enumerate}
\item Associative:
\begin{enumerate}
\item $a\oplus b=a+b+1$, $A=[0,+\infty]$
\item $a\oplus b=\frac{ab}{a+b}$, $A=\mathbb{R}^{+}$ or $A=]0,1]$
\item $a\oplus b=a+b+ab$, $A=\mathbb{R}^{+}$
\item $a\oplus b=a+b$, $A=\mathbb{R}^{+}$
\item $a\oplus b=\frac{a+b}{1+ab}$, $A=]0,1[$
\item $a\oplus b=\log(\exp(a)+\exp(b))$, $A=\mathbb{R}$
\end{enumerate}
\end{enumerate}
\item Ccm-magmas with every element an idempotent, that is, midpoint algebras (due to Proposition \ref{prop midpoint and expansive} the properties V and VI do not apply):
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
Property & $a\oplus b$ & $A$ & $e\in A$\\
\hline
I & $\frac{a+b}{2}$ & $\mathbb{R}$ & every\\
\hline
I & $\sqrt[3]{\frac{a^3+b^3}{2}}$ & $\mathbb{R}$ & every\\
\hline
II & $\frac{a+b}{2}$ & $[0,+\infty[$ & $e=0$\\
\hline
II & $\frac{2ab}{a+b}$ & $]0,1]$ & $e=1$\\
\hline
III & $\frac{a+b}{2}$ & $[0,1]$ & $e=\frac{1}{2}$\\
\hline
III & $\frac{2ab}{a+b}$ & $]1,+\infty[$ & $e=2$\\
\hline
IV & $\frac{a+b}{2}$ & $\mathbb{R}^{+}$ & N/A\\
\hline
IV & $\frac{2ab}{a+b}$ & $\mathbb{R}^{+}$ & N/A\\
\hline
\end{tabular}
\end{center}
\item Ccm-magmas with at least one idempotent element which are not mid\-point-algebras (in this case again we distinguish between associative and non-associative and give examples of each one of the properties I to VI from above):
\begin{enumerate}
\item non-associative:
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
Property & $a\oplus b$ & $A$ & $e\in A$\\
\hline
I & $2(a+b)$ & $\mathbb{R}$ & $e=0$\\
\hline
I & $2\sqrt[3]{a^3+b^3}$ & $\mathbb{R}$ & $e=0$\\
\hline
II & $2(a+b)$ & $[0,+\infty[$ & $e=0$\\
\hline
III & $\frac{a+b}{3}$ & $[-1,1]$ & $e=0$\\
\hline
IV & $\frac{a+b}{3}$ & $[0,1]$ & $e=0$\\
\hline
V & $2(a+b)$ & $\mathbb{N}_{0}$ & $e=0$\\
\hline
VI & $2(a+b)$ & $\mathbb{Z}$ & $e=0$\\
\hline
\end{tabular}
\end{center}
\item associative (in this case we only distinguish properties I and II from above, as it follows from Proposition \ref{prop ccm-magma associative}).
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
Property & $a\oplus b$ & $A$ & $e\in A$\\
\hline
I & $a+b$ & $\mathbb{R}$ & $e=0$\\
\hline
II & $a+b$ & $[0,+\infty[$ & $e=0$\\
\hline
II & $a+b-ab$ & $[0,1[$ & $e=0$\\
\hline
\end{tabular}
\end{center}
\end{enumerate}
\item Finite ccm-magmas. Every finite ccm-magma is homogeneous. Some examples are as follows. The matrix \[A_2\left(\begin{array}{ccc} 1 & 3 & 2\\ 3 & 2 & 1\\ 2 & 1 & 3 \end{array}\right)\] shows an example of the \emph{multiplication table} for a non-associative ccm-magma with three idempotents. The matrix\[A_3
\left(\begin{array}{ccc} 2 & 1 & 3\\ 1 & 3 & 2\\ 3 & 2 & 1 \end{array}\right)\] shows an example of a non-associative ccm-magma with no idempotents. However there are no ccm-magmas with a finite and even number of idempotents. This is due to the fact that if $(A,\oplus)$ is a ccm-magma then the subset \mbox{$\{a\in A\mid a\oplus a=a\}$} is a subalgebra of ccm-magmas and there are no commutative idempotent quasigroups (homogeneous ccm-magmas) of even order \cite{Stein}.
\end{enumerate}
\section{Internal Relations}\label{sec relations}
As we have observed in the introduction, the category of ccm-magmas is not Mal'tsev, and so there is no hope of having every internal reflexive relation automatically as a congruence, nor having any internal relation as a difunctional one. Nevertheless, as it is shown in \cite{ZJ-NMF} a category is weakly Mal'tsev if and only if every strong relation is difunctional. In particular, if $f,g\colon{X\times Y\to B}$ are any two morphisms of ccm-magmas, then the relation $R\subseteq X\times Y$, defined by
\begin{equation}\label{eq_strongrel}
xRy\Leftrightarrow f(x,y)=g(x,y)\end{equation}
is a strong relation.
Hence the following results:
\begin{proposition}\label{prop_intrel1} Let $f,g\colon{A\times A\to B}$ be two morphisms in the category of ccm-magmas such that $f(a,a)=g(a,a)$ for every $a\in A$. The internal relation defined by \[xRy\Leftrightarrow f(x,y)=g(x,y)\] is a congruence.
\end{proposition}
\begin{proof}
It is an immediate consequence of Lemma \ref{lemma1}.
\end{proof}
\begin{proposition}\label{prop_intrel2} Let $f,g\colon{X\times Y\to B}$ be two morphisms in the category of ccm-magmas. The internal relation defined by \[xRy\Leftrightarrow f(x,y)=g(x,y)\] is difunctional.
\end{proposition}
\begin{proof}
It is an immediate consequence of Lemma \ref{lemma2}.
\end{proof}
We now present some results involving another special type of internal relations, namely the ones constructed from a subalgebra of a given ccm-magma.
\begin{proposition}\label{prop R rel items} Let $(A,\oplus)$ be a ccm-magma with a subalgebra $X\subseteq A$ and an idempotent element $e\in X$. The relation
\begin{equation}\label{relationsubalg}aRb\Leftrightarrow \exists x\in X, a\oplus e=x\oplus b\end{equation}
\begin{enumerate}
\item[(i)] is an internal relation;
\item[(ii)] is reflexive;
\item[(iii)] is transitive whenever $X$ has a (unique) internal monoid structure with $e\in X$ as its unit;
\item[(iv)] is symmetric if and only if $X$ is $e$-symmetric.
\end{enumerate}
\end{proposition}
\begin{proof}
(i) The relation is an internal relation if and only if $R\subseteq A\times A$ is a subalgebra of the product, or equivalently, if and only if\[aRb,a'Rb'\Rightarrow (a\oplus a')R(b\oplus b').\] It is the case because if there exist $x,x'\in X$ such that $a\oplus e=x\oplus b$ and $a'\oplus e=x'\oplus b'$ then we have \begin{eqnarray*}
(a\oplus a')\oplus e &=& (a\oplus a')\oplus (e\oplus e)\\
&=& (a\oplus e)\oplus (a'\oplus e)\\
&=& (x\oplus b)\oplus (x'\oplus b')\\
&=& (x\oplus x')\oplus (b\oplus b')\\
\end{eqnarray*}
showing that there is $(x\oplus x')\in X$ such that\[(a\oplus a')\oplus e=(x\oplus x')\oplus (b\oplus b'),\] and so $(a\oplus a')R(b\oplus b')$.
(ii) The reflexivity of $R$ follows from the observation that \mbox{$a\oplus e=e\oplus a$} and $e\in X$.
(iii) Suppose there is an internal monoid structure in $X$ with $e$ as its unit element. Corollary \ref{corollary1} tells us that $X$ has a monoid structure with $e$ as unit if and only if for every $x,y\in X$ there is $(x*_e y)\in X$ such that $(x*_e y)\oplus e=x\oplus y$. In this case, we prove transitivity by showing that if $aRb$ and $bRc$, that is \[a\oplus e=x\oplus b\quad,\quad b\oplus e=y\oplus c,\] then $aRc$, because \[a\oplus e=(x*_e y)\oplus c.\]
It is straightforward to prove the above equality, by composing with $(e\oplus e)$ and then using cancellation:
\begin{eqnarray*}
(a\oplus e)\oplus (e\oplus e)&=&(x\oplus b)\oplus (e\oplus e)\\
&=&(x\oplus e)\oplus (b\oplus e)\\
&=&(x\oplus e)\oplus (y\oplus c)\\
&=&(x\oplus y)\oplus (e\oplus c)\\
&=&((x*_e y)\oplus e)\oplus (c\oplus e)\\
&=&((x*_e y)\oplus c)\oplus (e\oplus e).
\end{eqnarray*}
(iv) Let us show that $R$ is symmetric if and only if $X$ is $e$-symmetric.
By definition of $R$, for every $x\in X$, we have $xRe$. Hence, if the relation is symmetric we also have $eRx$ from which we conclude the existence of $y\in X$ such that $e=e\oplus e=y\oplus x$. This shows that $X$ is $e$-symmetric. Conversely, if $X$ is $e$-symmetric then for every $x\in X$ there is $y\in X$ such that $x\oplus y=e$ and consequently the relation $R$ is symmetric. Indeed, given $aRb$, that is $a\oplus e=x\oplus b$ for some $x\in X$, we have, with $y=-_e(x)$: \[(y\oplus a)\oplus e=(y\oplus e)\oplus(a\oplus e)=(y\oplus x)\oplus(e\oplus b)=e\oplus(e\oplus b)\]from which we conclude \[y\oplus a=b\oplus e\] and so $bRa$.
\end{proof}
The condition on the existence of a monoid structure in item (iii) above is sufficient for the relation to be transitive but it is not necessary. Indeed we only have the operation $(x*_e y)$ well defined for certain pairs $(x,y)\in X\times X$, namely the ones for which given any $c\in A$ there are $a,b\in A$ such that $a=(y*_e c)$ and $b=x*_e (y*_e c)$.
This suggests the following necessary and sufficient condition for the transitivity of this type of internal relations.
\begin{proposition} Let $(A,\oplus)$ be a ccm-magma with a subalgebra $X\subseteq A$ and an idempotent element $e\in X$. The relation\[aRb\Leftrightarrow \exists x\in X, a\oplus e=x\oplus b\]
is transitive if and only if:\begin{quote}
for all $x,y\in X$ and $c\in A$, if there exist $a,b\in A$ such that\begin{equation}\label{eq used for transitivity}
a\oplus e=x\oplus b\quad\text{and}\quad b\oplus e=y\oplus c
\end{equation}then there is $z\in X$ such that $z\oplus e=x\oplus y$.
\end{quote}
\end{proposition}
\begin{proof}
Assume $R$ is transitive. If we have solutions $a$ and $b$ for the equations (\ref{eq used for transitivity}) then we also have $aRb$ and $bRc$ which, by transitivity, gives us the desired $z\in X$ such that $a\oplus e=z\oplus c$. It is now a simple calculation to check that $z\oplus e=x\oplus y$. Indeed we have
\begin{eqnarray*}
(z\oplus e)\oplus (c\oplus e)&=&(z\oplus c)\oplus (e\oplus e)\\
&=&(a\oplus e)\oplus (e\oplus e)\\
&=&(x\oplus b)\oplus (e\oplus e)\\
&=&(x\oplus e)\oplus (b\oplus e)\\
&=&(x\oplus e)\oplus (y\oplus c)\\
&=&(x\oplus y)\oplus (e\oplus c)
\end{eqnarray*}
and the result follows from cancellation.
Conversely, if $aRb$ and $bRc$ then we also have $x,y\in X$ as in $(\ref{eq used for transitivity})$ and hence there is an element $z\in X$ such that $z\oplus e=x\oplus y$ from which we conclude that $a\oplus e=z\oplus c$. Indeed
\begin{eqnarray*}
(a\oplus e)\oplus (e\oplus e)&=&(x\oplus b)\oplus (e\oplus e)\\
&=&(x\oplus e)\oplus (b\oplus e)\\
&=&(x\oplus e)\oplus (y\oplus c)\\
&=&(x\oplus y)\oplus (e\oplus c)\\
&=&(z\oplus e)\oplus (c\oplus e)\\
&=&(z\oplus c)\oplus (e\oplus e),
\end{eqnarray*}
this shows that $aRc$ and concludes the proof.
\end{proof}
When $A$ is $e$-expansive, item (iii) in Proposition \ref{prop R rel items} can now be reformulated so to relate the transitivity of $R$ with the property of $X$ being $e$-expansive.
\begin{corollary}\label{corollary2} Let $(A,\oplus)$ be a ccm-magma which is $e$-expansive for some idempotent element $e\in A$ and let $X$ be a subalgebra with $e\in X$. The relation \[aRb\Leftrightarrow \exists x\in X, a\oplus e=x\oplus b\] is transitive if and only if $X$ is $e$-expansive.
\end{corollary}
\begin{proof}
If $X$ is $e$-expansive then in particular it has an internal monoid structure with $x*y=2_e(x\oplus y)_X$. Hence the relation is transitive (Proposition \ref{prop R rel items}(iii)). Conversely, if the relation is transitive, then for all $x,y\in X$ the element $2_e(x\oplus y)\in A$ is in fact an element in $X$. Indeed, $2_e(x\oplus y)Ry$ and $yRe$ implies $2_e(x\oplus y)Re$ which is equivalent to $2_e(x\oplus y)\in X$. This shows that $X$ is $e$-expansive.
\end{proof}
Combining the two previous results on symmetry and transitivity with $X$ being $e$-expansive and $e$-symmetric we also observe:
\begin{corollary}\label{corollary3} Let $(A,\oplus)$ be a homogeneous ccm-magma with a subalgebra $X\subseteq A$ and an idempotent element $e\in X$. The relation \[aRb\Leftrightarrow \exists x\in X, a\oplus e=x\oplus b\] is a congruence, if and only if $X$ is homogeneous.
\end{corollary}
\begin{proof}
If $X$ is homogeneous then it is $e$-expansive with $2_e(x)\in X$ and it is $e$-symmetric with $-_e(x)=2_x(e)\in X$, for every $x\in X$. As a consequence the relation $R$ is transitive and symmetric, and hence it is a congruence, since it is always an internal reflexive relation.
Conversely, let us suppose $R$ is a congruence. By Corollary \ref{corollary2} and Proposition \ref{prop R rel items}(iv) we already know that $X$ is $e$-expansive and $e$-symmetric, hence the result in Proposition \ref{prop_homogeneous} concludes the proof.
\end{proof}
\section{Conclusion}
This work shows that the category of ccm-magmas admits several classifications for its objects. One possibility is to differentiate between those ccm-magmas admitting an internal monoid structure and those who don't. A ccm-magma $(A,\oplus)$ with a given idempotent, $e$, admits an internal monoid structure with $e$ as its unit if and only if the equation \[x\oplus e=a\oplus b\] has a solution $x$ for every $a,b$ in $A$. This condition is weaker than $A$ being $e$-expansive. However, the two conditions are equivalent when every element $a\in A$ can be decomposed as ${a=x_1\oplus x_2}$. This property is considered, for instance, in \cite{Jezek93}.
It is shown that every relation $R$ of the form (\ref{eq_strongrel}), constructed with the two homomorphisms $f$ and $g$, is necessarily a difunctional relation. This result is also a consequence of the fact that the relation $R$ is a strong relation and the category of ccm-magmas is weakly Mal'tsev. More generally, when $f$ and/or $g$ are not homomorphisms, it might happen that $R$ is still an internal relation but not a difunctional one. In a similar way, if $f$ and $g$ are as in Proposition \ref{prop_intrel1}, with $f(a,a)=g(a,a)$, but not homomorphisms, then we may have a reflexive internal relation which is not a congruence. For example, the relation $R$ in Proposition \ref{prop R rel items}, is equivalently defined as $aRb$ if and only if $f(a,b)=g(a,b)$, where, for all $a,b\in A$, $f(a,b)=a\oplus e$ while $g(a,b)=a\oplus e$ if there exists $x\in X$ such that $a\oplus e=x\oplus b$, otherwise $g(a,b)=e\oplus b$.
Every finite ccm-magma is necessarily homogeneous (Definition \ref{def_homogeneous}), and since axiom (M3) is weaker than associativity, these kinds of structures may be useful to the random generation of finite abelian groups. The procedure is very simple: randomly generate a ccm-magma $M$ with at least one idempotent, say $e$ (although this is only important if we are interested in internal structures), and then define
\begin{center}
\texttt{A(i,j)=find(M(:,e)==M(i,j))}
\end{center}
for every $i$ and $j$, in order to obtain a matrix $A$ with the multiplication table for an abelian group with $e$ as unit element.
The notion of ccm-magma may also be defined internally in every category with binary products (as it is done in \cite{Escardo} for midpoint algebras) and so, some interesting interactions at this level are also expected, especially for the case of topological ccm-magmas.
|
\section{Introduction}
Pattern recognition problems involving the processing of patterns belonging to one class only are quite common \cite{Kemmler2013,Tax19991191,Juszczak20091859,occ_sg_enricods__arxiv,Tax:2002:UOG:944790.944809,fuzzy_occ,NIPS2002_2163,wang2013position,bodesheim1991divergence,6619277,6186735,4049825,pimentel2014review,Ding2014313,6722892,bicego2009soft}.
The interest in this type of problems is both methodological and of application-oriented character. In fact, one-class classification problems could be used to deal with tasks involving recognition of outliers in data. On the application side, instead, there are many real-world scenarios in which it is possible to obtain (or design) patterns only for the so-called ``target'' class. As an instance, we may cite the problem of determining whether a given machine/device is not working properly.
Intuitively, patterns representing the correct functioning of the machine/device are ``trivial'' and ``not informative'', in the sense that anything that is different from the observed faults is by definition an instance of correct functioning.
As a consequence, in this case one would model only those patterns representing fault instances (see Fig. \ref{fig:occ_example}).
However, modeling explicitly only one side of the decision boundary implies a more difficult setting with respect to (w.r.t.) well-established multi-class problems. In particular, the method for evaluating the performance of any one-class classifier (OCC) should take into account the implicit uncertainty rooted in the resulting decisions \cite{one-class_survey__2010,Juszczak20091859,fuzzy_occ,wang2013position,6722892}.
\begin{figure}[ht!]
\centering
\includegraphics[bb=0 0 448 437,scale=0.3,keepaspectratio=true]{./occ_example}
\label{fig:occ_example}
\caption{2D space describing the status of a machine/device. Regions denoting instances of faults are depicted in red (colors online). The model of a OCC consists of those regions.}
\end{figure}
The one-class classification setting has been adopted in many real-world applications (for a review, see \cite{one-class_survey__2010,pimentel2014review}) such as the recognition of faults in smart electric grids \cite{occ_sg_enricods__arxiv}, Raman spectroscopy \cite{Kemmler201329}, events detection in videos \cite{piciarelli2008trajectory}, document classification \cite{Manevitz01one-classsvms}, medical imaging \cite{desir2012random}, and oil spill detection \cite{oilspill__2010}.
In this paper, we propose a novel OCC that is based on an interplay of different techniques. Our first objective is to make the proposed OCC applicable to any data type. To this end, we develop our approach on the basis of the dissimilarity space representation \cite{Duin2012826}. This choice, although increases the overall computational complexity, allows us to cover virtually any application context, regardless of the adopted representation for the input data (e.g., features, labeled graphs, etc.).
Then, we represent the embedded data in the dissimilarity space (DS) by a complete Euclidean graphs, whose vertices denote input patterns and the edges their mutual (normalized) Euclidean distance in the DS.
Such a graph allows us to (i) estimate the informativeness of the represented data and at the same time (ii) construct the decision regions (DRs) that we use to define the OCC model.
Additionally, representing the data in the DS by means of the Euclidean graph provides us a way to bypass all problems related to the high-dimensionality of the DS.
The informativeness of the embedded data is computed with the use of the $\alpha$-order R\'{e}nyi entropy, estimated by means of the entropic spanning graph technique proposed by \citet{gs:HeroEtAl2002}. In particular, we use the minimum spanning tree, which is further analyzed with the aim of inducing a partition of the graphs into suitably compact and separated clusters of vertices. We designed a fast graph partitioning algorithm based on the concept of modularity \cite{4358966}.
The derived DRs are then equipped with suitable membership functions \cite{Livi_ga_2013,rizzi2002}, which allow us to form both hard (Boolean) and soft decisions about the classification.
To benchmark the proposed OCC, we test two types of data: features and labeled graphs (also termed attributed graphs) based patterns.
We provide experiments and offer a comparative analysis for different datasets from well-known benchmarks, such as those coming from the UCI \cite{Bache+Lichman_2013} and IAM \cite{riesen+bunke2008} repositories.
The paper is structured as follows.
Section \ref{sec:background} offers an overview on the one-class classification context (Section \ref{sec:related_works}) by hence providing also a clear collocation for the proposed OCC system. Moreover, in Section \ref{sec:tech_background} we introduce the main technical background material used by the proposed OCC.
Successively, in Section \ref{sec:eocc} we present the details of the proposed OCC.
In Section \ref{sec:exps} we show and discuss the experimental evaluations.
Finally, in Section \ref{sec:conclusions} we draw our conclusions, providing also pointers to future directions.
\section{Related Works and Background Material}
\label{sec:background}
\subsection{A Review on One-Class Classification Methods}
\label{sec:related_works}
\citet{pimentel2014review} group the current one-class methods in five different categories: (i) probabilistic, (ii) distance-based, (iii) domain-based, (iv) reconstruction-based, and finally (v) information-theoretic techniques.
Probabilistic methods are focused on the reconstruction of the generative probability density function (PDF) underlying the data at hand. Such methods are further subdivided into the usual parametric and non-parametric classes, where the former include methods based on the identification of the optimal parameters describing a pre-defined statistical model, while the latter include methods based on the reconstruction of the PDF directly from the data.
Distance-based methods operate, essentially, by means of a suitable distance measure in the input space. Techniques of this category can be grouped into clustering-based and nearest neighbors based approaches.
Reconstruction-based methods include classical data-driven approaches, such as neural networks and subspace-based methods.
Domain-based methods revolve around the well-established Support Vector Data Description (SVDD) \cite{Tax19991191}. In this case, the objective is to model the target data via suitable decision regions/surfaces by optimizing a specific convex optimization problem.
Finally, information-theoretic methods rely on information measures such as entropy, divergence, and mutual information.
Intuitively, a non-target pattern is identified as one that alters significantly the information content of the data.
As stated before, an important class of OCCs revolves around the SVDD method proposed by \citet{Tax19991191}, which has been elaborated taking inspiration from the well-known support vector machine (SVM) \cite{Tax19991191,SchWilSmoShaetal00,wang2013position}.
The classification model of SVDD is defined in terms of hyper-spheres, which cover the training set data through an SVM-like optimization problem (the minimization of the sphere radiuses is enforced).
SVDD is particularly exploitable since it can be used jointly with positive definite kernel functions, which allow the generalization of the input domain.
\citet{SchWilSmoShaetal00} proposed an alternative approach to SVDD that employs a hyperplane, like in the conventional SVM case.
The hyperplane is positioned to separated the region of the input space containing patterns form the region containing no data.
Other more recent approaches include algorithms based on the minimum spanning tree (MST) \cite{Juszczak20091859}, Gaussian processes \cite{Kemmler2013}, and on Random forests \cite{Desir20133490}.
The reader is referred to Ref. \cite{pimentel2014review} for a comprehensive survey portraying the state-of-the-art on one-class classification methods.
According to the aforementioned OCC systems categorization, the herein proposed OCC could be collocated in the intersection among probabilistic, distance-based, and information-theoretic based approaches.
Notably, the proposed OCC exhibits some linkages with the system of \citet{Juszczak20091859}, in the sense that their solution relies on a MST. However, their approach is substantially different, since they do not use either the information-theoretic, fuzzy sets, or graph partitioning concepts that we exploit in this study. Moreover, our approach is dissimilarity-based, which opens a way to a multitude of applications in different areas. This last aspect recalls the OCC scheme by \citet{NIPS2002_2163}; however the authors use different techniques to design their system, which are based on linear programming and prototype selection.
Graph-based, and in particular minimum spanning tree based, general clustering algorithms are popular in the literature \cite{10.1109/TPAMI.2012.226,1432700,xu2002clustering,Galluccio201396,Galluccio:2012:GBK:2184924.2185067}, since, in fact, a graph provides a powerful data abstraction and a sound mathematical framework. However, to the best of our knowledge, the use of graph-based entropy estimation techniques for the design of the OCC model is missing in the OCC literature.
The utilization of fuzzy sets to model the DRs establishes another connection with the so-called fuzzy one-class classifiers \cite{fuzzy_occ,6613141,Utkin:2012:FOC:2213741.2433967}.
\subsection{Technical Background Material}
\label{sec:tech_background}
In the following subsections, we introduce the main concepts used in the OCC proposed in this paper: dissimilarity representation, entropy estimation, and modularity of a graph partition.
For more detailed discussions on dissimilarity representation we refer the reader to \cite{pkekalska+duin2005}; for graph-based entropy estimation to \cite{intrdim_shapes_hero,gs:HeroEtAl2002}; finally, for modularity of a graph partition to \cite{fortunato2010,4358966}.
\subsubsection{Dissimilarity Representation}
\label{sec:dr}
In the dissimilarity representation \cite{pkekalska+duin2005,odse,Duin2012826}, the elements of an input dataset $\mathcal{S}\subset\mathcal{X}$ are characterized by considering their pairwise dissimilarity values.
The key component is the definition of a nonnegative (bounded) dissimilarity measure $d: \mathcal{X}\times\mathcal{X}\rightarrow\mathbb{R}^{+}$, which is in charge of synthesizing all relevant commonalities among the input patterns of $\mathcal{S}$ into a single real-valued number.
A set of prototypes, $\mathcal{R}$, called representation set (RS), is used to develop the dissimilarity matrix $\mathbf{D}$, which is given as $D_{ij}=d(x_i, r_j)$, for every $x_i\in\mathcal{S}$ and $r_j\in\mathcal{R}$. Of course, the determination of the most suitable $\mathcal{R}$, with usually $\mathcal{R}\subset\mathcal{S}$, is an important objective. Different techniques are discussed by \citet{pkekalska+duin2005}, which span from prototype selection strategies to criteria related to the embedded data.
By using directly the rows of $\mathbf{D}$ as embedding vectors, we can obtain the so-called dissimilarity space representation (DSR). This is the fastest way to represent the input data as $\mathbb{R}^d$ vectors by starting from the dissimilarity values. In addition, any common algebraic structure can be defined on the dissimilarity space (DS), making this approach very flexible.
An important property of the DSR is that distances in the DS are just scaled by a factor equal to $\sqrt{|\mathcal{R}|}$ w.r.t. those in the input space (see \cite[Sec.~4.4.1]{pkekalska+duin2005} for details).
\subsubsection{Graph-based Entropy Estimation}
\label{sec:e_mst}
Let $X$ be a continuous random variable with PDF $p(\cdot)$. The $\alpha$-order R\'{e}nyi entropy measure is defined as
\begin{equation}
\label{eq:differential_entropy}
H_{\alpha}(X)=\frac{1}{1-\alpha}\log\left(\int p(x)^{\alpha}dx\right), \ \alpha\geq0, \alpha\neq1.
\end{equation}
Let us assume to have a data sample $X_{n}$ of $n$ i.i.d. realizations of $X$, with $\underline{\mathbf{x}}_i\in X_{n}\subset\mathbb{R}^{d}, i=1, 2, ..., n$, and $d\geq 2$. Let $G$ be the complete Euclidean graph constructed over $X_{n}$.
An edge $e_{ij}$ connecting $\underline{\mathbf{x}}_i$ and $\underline{\mathbf{x}}_j$ is weighted using a weight based on their distance, $|e_{ij}| = d_{2}(\underline{\mathbf{x}}_i, \underline{\mathbf{x}}_j)$.
The $\alpha$-order R\'{e}nyi entropy (\ref{eq:differential_entropy}) can be estimated according to a geometric interpretation of an entropic spanning graph of $G$.
Examples of such graphs used in the literature are the MST, \textit{k}-NN graph, Steiner tree, and TSP graph \cite{md_ent,Bonev2013214,pal_renyi_e_knn__2010,odse2__arxiv,4897236,neemuchwala2005image,oubel2005assessment,intrdim_shapes_hero,Hero_Asympt__1999}. In this paper, we will focus on the MST \cite{bonev__2008,neemuchwala2005image}.
Let $L_{\gamma}(G)$ be the \textit{weighted length} of a MST, $T$, connecting the $n$ points in $X_{n}$,
\begin{equation}
\label{eq:l_mst}
L_{\gamma}(G) = \displaystyle\sum_{e_{ij}\in T} |e_{ij}|^{\gamma} ,
\end{equation}
where $\gamma\in(0, d)$ is a user-defined parameter.
The R\'{e}nyi entropy of order $\alpha\in(0, 1)$, elaborated using the MST length (\ref{eq:l_mst}), is defined as follows:
\begin{equation}
\label{eq:rentropy_mst}
\hat{H}_{\alpha}(G) = \frac{d}{\gamma}\left[ \ln\left(\frac{L_{\gamma}(G)}{n^{\alpha}}\right) - \ln\left(\beta(L_{\gamma}(G), d)\right) \right],
\end{equation}
where $\alpha=(d-\gamma)/d$.
The $\beta(L_{\gamma}(G), d)$ term is a constant that is defined as $\beta(L_{\gamma}(G), d) \simeq \gamma/2\ln\left( d/2\pi e \right)$.
As a consequence of $G$, the entropy estimator (\ref{eq:rentropy_mst}) is suitable for processing high-dimensional input data.
\subsubsection{Modularity of a Graph Partition}
\label{sec:modularity}
A graph $G=(\mathcal{V}, \mathcal{E})$ is a pair of vertices and edges. An edge models a binary relation among two vertices. Normally, an edge either exists or does not exist. However, in the case of \textit{weighted} graphs, every edge $e_{ij}\in\mathcal{E}$ is associated with a real-valued number called the weight, $w_{ij}=w(e_{ij})$, which determines the strength of the relation. The weighted adjacency matrix $\mathbf{A}$ of $G$ is defined as $A_{ij}=w_{ij}$.
The degree of a vertex $v_i\in\mathcal{V}$ is defined as $\mathrm{deg}(v_i)=\sum_{j=1}^{|\mathcal{V}|} A_{ij}$.
We will consider weighted graphs with weights in $[0, 1]$. If not diversely specified, when we refer to the ``number'' of edges we actually refer to the sum of their weights.
A partition \cite{Livi_ga_2013}, $K(G)$, of order $k$ of a graph $G=(\mathcal{V}, \mathcal{E})$, is commonly intended as a partition of the vertex set $\mathcal{V}(G)$ into disjoint subsets (clusters, modules), $K(G)=\{\mathcal{C}_1, \mathcal{C}_2, ..., \mathcal{C}_k\}$.
A well-established measure to determine the quality of $K(G)$ is the so-called modularity measure \cite{brandes+gaertler+wagner2003,fortunato2010,1367-2630-10-5-053039,PhysRevE.71.046117,4358966,rosvall2007information}, which basically quantifies how well $K(G)$ groups the vertices of $G$ into compact and separated clusters.
Intuitively, in a graph a cluster of vertices is compact if the number of the intra-cluster edges is considerably greater than the one of the inter-cluster edges.
The modularity measure $Q(\cdot, \cdot)$ is formally defined as follows,
\begin{align}
\label{eq:modularity}
&Q(G, K(G)) = \frac{1}{2|\mathcal{E}(G)|}\sum_{i=1,j=1}^{k} \left( A_{ij}-\frac{\mathrm{deg}(v_i)\mathrm{deg}(v_j)}{2|\mathcal{E}(G)|} \right) \delta(\mathcal{C}_i, \mathcal{C}_j),
\end{align}
where the graph $G$ is assumed to be partitioned according to a given $K(G)$.
Eq. \ref{eq:modularity} can be conveniently rewritten in terms of edges only (\cite{4358966,1367-2630-10-5-053039}):
\begin{equation}
\label{eq:modularity2}
Q(G, K(G)) = \sum_{l=1}^{k} \left[ \frac{|\mathcal{E}(\mathcal{C}_l)|}{|\mathcal{E}(G)|}-\left( \frac{\mathrm{deg}(\mathcal{C}_l)}{2|\mathcal{E}(G)|} \right)^2 \right].
\end{equation}
In the above expression, $|\mathcal{E}(\mathcal{C}_l)|$ is the number of intra-cluster edges, and $\mathrm{deg}(\mathcal{C}_l)$ is the sum of degrees of the vertices in the l\textit{th} cluster (considering all edges, i.e., also those with one end-point outside $\mathcal{C}_i$).
Detailing the terms in (\ref{eq:modularity2}), we have:
\begin{align}
|\mathcal{E}(\mathcal{C}_l)| = \sum_{\{e_{ij}|\ v_i,v_j\in\mathcal{C}_l\}} w_{ij};\ \ \ \mathrm{deg}(\mathcal{C}_l) &= \sum_{v_i\in\mathcal{C}_l} \mathrm{deg}(v_i); \ \ \ |\mathcal{E}(G)| = \sum_{e_{ij}\in\mathcal{E}(G)} w_{ij}.
\end{align}
The modularity of a graph $G$ is equal to the modularity of the partition of $G$ that maximizes (\ref{eq:modularity2}).
Finding such an optimal partition is NP-complete \cite{4358966}, and therefore many heuristics has been proposed in the literature \cite{fortunato2010}.
It is well-known that the modularity (\ref{eq:modularity2}) assumes values in $[-1/2, 1]$ \cite{4358966} (the higher, the better).
In this paper, we consider the normalized term $M(G, K(G))\in[0, 1]$, defined as:
\begin{equation}
\label{eq:normalized_modularity}
M(G, K(G))=\log(3/2 + Q(G, K(G)))/\log(5/2).
\end{equation}
\section{The Proposed One-Class Classifier}
\label{sec:eocc}
Given an input dataset $\mathcal{S}\subset\mathcal{X}$, we design a one-class classifier that is applicable to any input domain, $\mathcal{X}$.
To fulfill such a requirement, we first embed the input data, $\mathcal{S}$, into an Euclidean space. Notably, we implement the embedding by constructing the DSR of $\mathcal{S}$ (see Sec. \ref{sec:dr}).
Let $d_{\mathrm{I}}: \mathcal{X}\times\mathcal{X}\rightarrow\mathbb{R}^+$ be a suitable dissimilarity measure.
We assume here that $d_{\mathrm{I}}(\cdot, \cdot)$ depends on some (numerical) parameters, say $p$, which alter the resulting view of the input data, $\mathcal{S}$, in the DS. The derived DSR of $\mathcal{S}$ is accordingly denoted with $D(\mathcal{S}, \mathcal{R}, p)$, where $\mathcal{R}\subseteq\mathcal{S}$ is the RS, and $p$ is the specific parameters instance of $d_{\mathrm{I}}(\cdot, \cdot)$.
The embedded data are then represented by constructing a complete graph $G=(\mathcal{V}, \mathcal{E})$.
Vertices $\mathcal{V}$ denote the patterns in $\mathcal{S}$, while edges denote their relations in terms of distance; each edge $e_{ij}\in\mathcal{E}$ has a weight given by $w_{ij}=d_2(v_i, v_j)$, where $d_2(\cdot, \cdot)$ is a suitable (normalized) Euclidean metric.
Graph representations are popular for high-dimensional patterns \cite{gs:HeroEtAl2002,oubel2005assessment}. Constructing $G$ allows us also to avoid the computational burden of determining the optimal RS, $\mathcal{R}$; the size of the RS corresponds to the dimensionality of the DS, and therefore usually a proper prototype selection method is used to consider the smallest but most informative subset of $\mathcal{R}$.
Since $d_2(\cdot, \cdot)$ is Euclidean, $G$ can be used also to estimate the $\alpha$-order R\'{e}nyi entropy \cite{intrdim_shapes_hero,gs:HeroEtAl2002} of the underlying data distribution through the computation of the entropic MST (see Sec. \ref{sec:e_mst}).
The entropy is a powerful mesoscopic data descriptor -- it is also a measure of the \textit{spread} of the data -- that we use together with other terms to guide the synthesis of the OCC model.
In the following, we denote with $G(p)$ the (complete) Euclidean graph constructed over $D(\mathcal{S}, \mathcal{R}, p)$, obtained by setting $d_{\mathrm{I}}(\cdot, \cdot)$ with the $p$ instances.
$G(p)$ provides an abstract framework in which we can develop the model, the related decision rules, and the synthesis of the OCC.
Modules (i.e., vertex clusters) of $G(p)$, which we denote as a collection $K(G(p))=\{\mathcal{C}_1, \mathcal{C}_2, ..., \mathcal{C}_k\}$, are considered as suitable DRs, i.e., the OCC model; we derive such DRs by exploiting the concept of modularity (see Sec. \ref{sec:modularity}).
We maximize the modularity (\ref{eq:modularity2}) of the graph by analyzing a partition derived directly on the entropic MST, which has to be computed in order to calculate the $\alpha$-order entropy.
The point here is that the vertex degrees in the MST are a scaled approximation of those in $G(p)$ -- highly central vertices in $G(p)$ remain well-connected also in the MST.
Let $T(G(p))$ be a MST of $G(p)$. By using $T(G(p))$, we can induce a partition $K(G(p))=\{\mathcal{C}_1, \mathcal{C}_2, ..., \mathcal{C}_k\}$ whose quality can be evaluated by considering all edges of $G(p)$.
Since the weights of $G(p)$ denote the pairwise normalized Euclidean distances in the DS, for the purpose of calculating the modularity (\ref{eq:modularity2}) of $K(G(p))$, we consider instead the quantity $\overline{w}_{ij}=1-w_{ij}$ (the lower the distance, the higher the contribution in terms of modularity).
In the one-class classification setting we model the target class only. Therefore, we need to conceive an inference mechanism that takes into account the implicit uncertainty in the definition of the DR boundaries.
To this aim, we equip each cluster $\mathcal{C}_i$ with a membership function \cite{pedrycz1990fuzzy,Livi_ga_2013,pedrycz1998introduction,kuncheva2000fuzzy}, which efficiently describes the uncertainty of the DR boundaries. In the following, we denote with $F(G(p))=\{\mathcal{F}_1, \mathcal{F}_2, ..., \mathcal{F}_k\}$ the fuzzy set-based DRs, which will be used during the test stage of the classifier to provide \textit{soft} decisions.
The aim of the synthesis is to optimize the OCC model by searching for the best $p$ instances. We provide two objective functions (related optimization problems are always intended to realize maximization).
The first one considers a linear convex combination of entropy and modularity, calculated on the training set $\mathcal{S}_{tr}\subset\mathcal{S}$ only; $\mathcal{S}_{tr}$ contains target patterns only.
The two terms are clearly in conflict. In fact, the entropy favors the general spread--separation of the data in the DS, while the modularity constraints the data to group into compact clusters of $G(p)$. This combination yields solutions that help magnifying the structure of the DRs in $G(p)$.
The second objective function is designed to train the OCC model by cross-validation, i.e., we effectively test the OCC model instances on a validation set, $\mathcal{S}_{vs}$, containing both target and non-target patterns.
The first approach is considerably faster for what concerns the training stage, although the second one provides a more effective solution in terms of test set recognition performance.
Fig. \ref{fig:bloch_scheme} provides a block scheme describing the proposed OCC, while Fig. \ref{fig:functioning} conveys the same information but using more intuitive illustrations.
\begin{figure}[ht!]
\centering
\includegraphics[bb=0 0 571 407,scale=0.4,keepaspectratio=true]{./block_scheme}
\caption{Block scheme of training and test stages of the OCC; first objective function is assumed. $\mathcal{S}_{tr}$ is used to synthesize the OCC model.
The optimal parameters, $p^{*}$, are used to generate the fuzzy model, $F(G(p^{*}))$.
$\mathcal{S}_{ts}$ is first embedded into the DS obtained by means of $p^{*}$, and successively it is tested. The OCC outputs both Boolean and membership grades to the target class.
}
\label{fig:bloch_scheme}
\end{figure}
\begin{figure}[ht!]
\centering
\subfigure[Input data mapped into the DS.]{
\includegraphics[bb=0 0 535 267,scale=0.33,keepaspectratio=true]{./functioning_1}
\label{fig:functioning_1}}
~
\subfigure[Entropic spanning graph construction (MST).]{
\includegraphics[bb=0 0 570 230,scale=0.33,keepaspectratio=true]{./functioning_2}
\label{fig:functioning_2}}
\subfigure[Fuzzy model of the partition derived from the MST.]{
\includegraphics[bb=0 0 578 306,scale=0.33,keepaspectratio=true]{./functioning_3}
\label{fig:functioning_3}}
\caption{Functioning of the proposed OCC; refer also to functional blocks in Fig. \ref{fig:bloch_scheme}.}
\label{fig:functioning}
\end{figure}
\subsection{OCC Model and Related Testing}
\label{sec:occ_test}
A test pattern is classified by considering the optimal parameters $p^*$, which yield the graph $G(p^*)$ constructed over $D(\mathcal{S}_{tr}, \mathcal{R}, p^*)$, and the derived hard $K(G(p^*))$ and fuzzy $F(G(p^*))$ partitions.
Each fuzzy set $\mathcal{F}_i\in F(G(p^*)), i=1, 2, ..., k$, forms a fuzzy DR that is described by its membership function $\mu_{\mathcal{F}_{i}}(\cdot)$ and by a quantity $\tau_{i}>0$. The membership function $\mu_{\mathcal{F}_i}(\cdot)$ is parametrized by $\tau_{i}$, which is explicitly denoted as $\mu_{\mathcal{F}_{i}}(\cdot; \tau_i)$.
The scalar $\tau_{i}$ is determined by considering a statistics of the $\mathcal{C}_i$ intra-cluster edge weights (e.g., average, standard deviation).
A test pattern $x\in\mathcal{S}_{ts}$ is first mapped to a dissimilarity vector $\underline{\mathbf{v}}$ by setting $d_{\mathrm{I}}(\cdot, \cdot)$ with $p^*$ and considering the dissimilarity w.r.t. $\mathcal{R}$.
We define the soft decision function (SDF), which outputs a continuous value in $[0, 1]$ quantifying the membership degree of a test pattern to the target class, as:
\begin{align}
\label{eq:soft_decision_func}
\mathrm{SDF}(\underline{\mathbf{v}}) = \displaystyle\bot_{i=1}^{k} \mu_{\mathcal{F}_{i}}(\underline{\mathbf{v}}; \tau_i).
\end{align}
In the above expression, $\bot$ is a t-conorn (e.g., the maximum) and $\mu_{\mathcal{F}_{i}}(\cdot; \tau_i)$ is the membership function synthesized during the training -- the membership is a function of the distance of $\underline{\mathbf{v}}$ w.r.t. the cluster representative, $R(\mathcal{C}_i)$. Please note that here we do not provide closed-form expressions for $\bot$ and $\mu_{\mathcal{F}_{i}}(\cdot; \tau_i)$, since those two factors are general and they can be implemented in different fashions.
Later in the experiments section we will specify the setting that we adopted in this study.
It is important to provide also a hard (Boolean) decision function (HDF) about the classification. To this end, we exploit the cluster extent, $\tau_i$, for defining the HDF. Let
\begin{equation}
\mathcal{F}_{\mathrm{max}} = \operatornamewithlimits{arg\ max}_{\mathcal{F}_i\in F(G(p^*))} \mu_{\mathcal{F}_{i}}(\underline{\mathbf{v}}; \tau_i)
\end{equation}
be the fuzzy set in which $\underline{\mathbf{v}}$ achieves the maximum membership degree; $\mathcal{C}_{\mathrm{max}}$ is the corresponding non-fuzzy DR.
Then,
\begin{equation}
\label{eq:hard_decision_1}
\mathrm{HDF}(\underline{\mathbf{v}}) =
\begin{cases}
1 & \mathrm{if}\ d_{2}(\underline{\mathbf{v}}, R(\mathcal{C}_{\mathrm{max}}))\leq \tau_{\mathrm{max}}, \\
0 & \mathrm{otherwise.}
\end{cases}
\end{equation}
Fig. \ref{fig:fuzzy_model_graph} provides a schematic view of the test of an embedded pattern, $\underline{\mathbf{v}}_{\mathrm{test}}$.
Testing of the OCC model is characterized by a computational complexity given by the embedding of $x$ and the computation of both SDF and HDF.
The embedding can be performed in $O(|\mathcal{S}|+D |\mathcal{R}|)$, where $|\mathcal{S}|$ is the linear cost of deriving $\mathcal{R}$ from the input (training) data (that might be constant in the case $\mathcal{R}=\mathcal{S}$), and $D$ denotes the computational complexity of the dissimilarity measure for the input data. SDF can be computed by considering the $k$ different membership degrees; the same holds for HDF.
\begin{figure}[ht!]
\centering
\includegraphics[bb=0 0 417 294,scale=0.55,keepaspectratio=true]{./FuzzyModel_Graph}
\caption{Fuzzy model constructed over $G(p)$. A test pattern, $\underline{\mathbf{v}}_{\mathrm{test}}$, is classified as non-target by the HDF, while considering the SDF it gets a membership degree, $\mu_{\mathcal{F}_i}(\underline{\mathbf{v}}_{\mathrm{test}}; \tau_i)$, to the target class.}
\label{fig:fuzzy_model_graph}
\end{figure}
\subsection{Synthesis of the OCC Model}
\label{sec:fast_synthesis}
We cast the synthesis of the proposed OCC as the optimization of $G(p)$ w.r.t. the parameters $p\in\mathcal{P}$ of $d_{\mathrm{I}}(\cdot, \cdot)$. Note that in this paper, we make the fair assumption that $\mathcal{P}=[0, 1]^u$, where $u$ is the number of parameters characterizing $d_{\mathrm{I}}(\cdot, \cdot)$.
The idea is to determine the best-performing parameters setting so that the following objective function is maximized:
\begin{equation}
\label{eq:obj_func}
\max_{p\in\mathcal{P}} \eta \hat{H}_{\alpha}(G(p)) + (1-\eta) M(G(p), K(G(p))),\ \eta\in[0, 1].
\end{equation}
We evaluate two conflicting quantities on $G(p)$: the entropy, $H(G(p))$, and the modularity, $M(G(p), K(G(p)))$.
The entropy term favors the construction of a DS and related graph, $G(p)$, such that the overall distance among the vertices/patterns is maximized (the higher the entropy, the higher the spread of the data). On the other hand, evaluating the modularity of the derived partitioning, $K(G(p))$, constrains the optimization to search for solutions that magnify also the ``community'' structure of the graph -- community is a term \cite{fortunato2010} that is used to denote a compact and populated cluster of vertices.
This results in a graph $G(p)$ that is suitably optimized to derive the DRs in terms of compact and separated clusters of vertices.
However, to evaluate the modularity we first need to generate a partition, $K(G(p))$. In the next section, we describe the algorithm that we designed to derive a partition of $G(p)$, whose aim is to quickly find a reasonable approximation of the optimal modularity of $G(p)$.
\subsubsection{Greedy Edge Pruning Approach to Graph Partitioning}
The following algorithm exploits the fact that we need to compute the MST, $T(G(p))$, of $G(p)$ to calculate the $\alpha$-order R\'{e}nyi entropy (see Sec. \ref{sec:e_mst}).
Once we have the MST, we form clusters on $T(G(p))$ iteratively by pruning (i.e., removing) those edges with higher Euclidean distance values.
By construction of the MST, those edges are likely to be ``bridges'' among well-separated components of $G(p)$. A MST of a graph with $n$ vertices has $n-1$ edges, and hence the pruning loop is repeated at most $n-1$ times.
Therefore, at iteration $i=0, 1, ..., n-1$, we partition the vertices of the MST into $k=i+1$ connected components. The connected components of the MST are used to derive a partition on $G(p)$, by considering exactly the same grouping of the vertices. However, in $G(p)$ we have full information of the edges, which we use to compute the modularity of the resulting partition (\ref{eq:modularity2}).
To terminate the procedure, we exploit the following greedy assumption. Since the MST is connected, by first removing edges with maximum weight, we will form the most interesting communities/clusters in terms of modularity.
As a consequence, if at iteration $i+1$ we get a modularity value lower than the one obtained at iteration $i$, we stop the algorithm, returning the last computed partition.
Algorithm \ref{alg:edge_pruning} delivers the pseudo-code of the herein described procedure.
\begin{algorithm}[h!]\footnotesize
\caption{Greedy edge pruning algorithm.}
\label{alg:edge_pruning}
\begin{algorithmic}[1]
\REQUIRE The MST $T(G(p))$ and the graph $G(p)$ with $n$ vertices
\ENSURE A partition $K(G(p))$ of $G(p)$
\STATE Set $i=0$ and best modularity $M_i=-1$
\STATE Let $L$ be a list with the $n-1$ edges of $T(G(p))$ in non-increasing order
\FOR{$i=1, 2, ..., n-1$}
\STATE Remove $i$th edge from $T(G(p))$ and determine the resulting connected components, $K(T(G(p)))_i$
\STATE Derive the corresponding partitioning $K(G(p))_i$ of $G(p)$ by considering the same vertex grouping of $K(T(G(p)))_i$
\STATE Set $M_i$ according to the evaluation of (\ref{eq:modularity2}) on $K(G(p))_i$
\IF{$M_i<M_{i-1}$}
\RETURN $K(G(p))_{i-1}$
\ENDIF
\ENDFOR
\RETURN $K(G(p))_{i}$
\end{algorithmic}
\end{algorithm}
\subsubsection{Analysis of Computational Complexity}
\label{sec:cca_eocc1}
The overall computational overhead of synthesizing the OCC using the herein explained approach is characterized by the sum of the following costs: (i) embedding, (ii) graph construction and entropy estimation, (iii) determination of graph partition, and finally (iv) the generation of the membership functions. The first three components must be considered into a suitable optimization loop, while the last one is performed only once at the end of the optimization cycle.
The dissimilarity representation of the training set $\mathcal{S}_{tr}, n=|\mathcal{S}_{tr}|$, costs $O(n|\mathcal{R}|D)$, where $D$ is the cost of the dissimilarity measure for the input data.
The second cost can be summarized as follows:
\begin{align}
\label{eq:rentropy_mst_complexity}
O\left( \frac{n(n-1)}{2}E + \frac{n(n-1)}{2}\times\log\left(\frac{n(n-1)}{2}\right) + (n-1) \right).
\end{align}
The first term standing in (\ref{eq:rentropy_mst_complexity}) accounts for the generation of $G(p)$, computing the respective Euclidean distances for the edge weights ($E$ is the cost).
The second term quantifies the cost involved in the MST computation using the well-known Kruskal's algorithm. The last term in (\ref{eq:rentropy_mst_complexity}) concerns the computation of the MST length.
The third cost is given by Algorithm \ref{alg:edge_pruning}.
The main cycle is repeated a maximum of $n-1$ times. At each iteration, we derive the connected components on $T(G(p))$, which costs $O(n-1)$. To induce the partitioning on $G(p)$, we simply cycle through the vertices, grouping them according to the connected components.
Eq. \ref{eq:modularity2} can be computed in $O(n(n-1))$. Putting all together, we have for Algorithm \ref{alg:edge_pruning}
\begin{equation}
O((n-1)\times( 2(n-1) + n(n-1) ),
\end{equation}
which is dominated by a cubic computational complexity in the number of training patterns.
The last cost (membership function elicitation) depends on the order of the derived best partition. In particular, for each cluster $\mathcal{C}_i$ we determine the membership function by first deriving $\tau_{\mathcal{C}_i}$. In the case average intra-cluster distances are considered, then the cost of this step is $O(2n)$.
The overall worst-case cost, considering as main parameter $n$, is given by the determination of the partition with best modularity. However, since we have designed the system to operate in the DS, the actual cost depends also on $D$, which may have a significant impact in case of complex input data types.
\subsection{Synthesis By Cross-validation}
\label{sec:synthesis_crossval}
Eq. \ref{eq:obj_func} defines an objective that does not take into account explicitly the recognition capability of the synthesized model.
A blind partitioning that derives $K(G(p))$ according to Algorithm \ref{alg:edge_pruning} might suffer from the problem of generating a too simple model, i.e., with too few DRs.
In fact, Algorithm \ref{alg:edge_pruning} constraints the partition to be formed only by those clusters that induce an well-defined community structure in $G(p)$. However, since in the one-class setting we synthesize the model on the target class only, a well-formed cluster/community structure in $G(p)$ may not be easy to identify, especially in hard problems.
By relaxing the imperative of finding the partition with maximum modularity, we can conceive another objective function that allows us to force the derivation of additional DRs.
The alternative objective function to be considered reads as,
\begin{align}
\label{eq:obj_func2}
\max_{p\in\mathcal{P}}\ &\beta P(\mathcal{S}_{tr}, \mathcal{S}_{vs}; F(G(p))) +& \\
\nonumber&(1-\beta) \left[\eta \hat{H}_{\alpha}(G(p)) + (1-\eta) M(G(p), K(G(p)))\right],
\end{align}
where $\eta,\beta\in[0, 1]$.
Eq. \ref{eq:obj_func2} takes explicitly into account a measure of recognition performance, $P(\mathcal{S}_{tr}, \mathcal{S}_{vs}; F(G(p)))$, achieved on a validation set, $\mathcal{S}_{vs}$.
This term is combined (again with a linear convex combination) with (\ref{eq:obj_func}). Therefore, the final model does not necessarily imply the best possible modularity, $M(G(p), K(G(p)))$, and entropy, $\hat{H}_{\alpha}(G(p))$, of $G(p)$, focusing instead on the solutions that perform better also in terms of recognition.
This choice, potentially, implies obtaining a more complex model (i.e., a partition characterized by more clusters, with lower overall modularity), although, as we will observe in the experiments, it usually provides also a more effective classification system on the test set.
Algorithm \ref{alg:crossval} shows the pseudo-code of the herein described training scheme.
DRs are derived by exploiting, basically, the same MST-based graph partitioning approach (Algorithm \ref{alg:edge_pruning}). In fact, DRs are derived incrementally, by first removing edges that are more likely to induce well-formed clusters (i.e., those edges with higher weights).
\begin{algorithm}[h!]\footnotesize
\caption{Training scheme by cross-validation.}
\label{alg:crossval}
\begin{algorithmic}[1]
\REQUIRE The training set $\mathcal{S}_{tr}$, the validation set $\mathcal{S}_{vs}$, and the parameters $p$
\ENSURE A fuzzy partition $F(G(p))$
\STATE Determine the DSR $D(\mathcal{S}_{tr}, \mathcal{R}, p)$ and $D(\mathcal{S}_{vs}, \mathcal{R}, p)$
\STATE Construct $G(p)$ over $D(\mathcal{S}_{tr}, \mathcal{R}, p)$
\STATE Estimate the entropy, $H(G(p))$, and get the MST, $T(G(p))$
\STATE Let $L$ be a list with the $n-1$ edges of $T(G(p))$ in non-increasing order
\STATE Set $P_{\mathrm{max}}=0$
\FOR{$i=1, 2, ..., n-1$}
\STATE Remove $i$th edge from $T(G(p))$. Determine the resulting connected components, $K(T(G(p)))_i$
\STATE Derive $K(G(p))_i$ of $G(p)$ by considering the same vertex grouping of $K(T(G(p)))_i$
\STATE Set $M_i$ according to the evaluation of (\ref{eq:modularity2}) on $K(G(p))_i$
\STATE Generate fuzzy model $F(G(p))_i$ from $K(G(p))_i$
\STATE Set $P_i$ as the evaluation of the objective function (\ref{eq:obj_func2}) on $\mathcal{S}_{vs}$
\IF{$P_i>P_{\mathrm{max}}$}
\STATE $P_{\mathrm{max}}=P_i$
\ENDIF
\ENDFOR
\RETURN $F(G(p))_{\mathrm{max}}$ corresponding to $P_{\mathrm{max}}$
\end{algorithmic}
\end{algorithm}
\subsubsection{Analysis of Computational Complexity}
\label{sec:cca_eocc2}
The training procedure described in Algorithm \ref{alg:crossval} implies a substantial change in the OCC scheme illustrated in Fig. \ref{fig:bloch_scheme}.
First, the modularity (\ref{eq:modularity2}) is computed always exactly $n-1$ times.
Additionally, the fuzzy model must be synthesized and then tested on $\mathcal{S}_{vs}$ inside the optimization cycle (see costs related to testing the model in Sec. \ref{sec:occ_test}). Those changes affect considerably the effective running time of the whole procedure, although the computational complexity, asymptotically, is not altered.
\section{Experimental Evaluation}
\label{sec:exps}
We start by describing the experimental setting and the considered performance measures (Sec. \ref{sec:pm_expsettin}).
Then, in Sec. \ref{sec:exp_synth} we provide some preliminary and explanatory tests on synthetically generated problems.
In Sec. \ref{sec:exp_uci} we perform experiments on different datasets of feature-based patterns taken from the UCI repository \cite{Bache+Lichman_2013}. Lastly, in Sec. \ref{sec:exp_iam}, we discuss the results obtained over different IAM datasets \cite{riesen+bunke2008}, containing patterns represented as labeled graphs.
\subsection{Experimental Setting and Performance Measures}
\label{sec:pm_expsettin}
All non-synthetic datasets considered in this paper are originally conceived for multi-class classification problems. Accordingly, we convert them to fit the one-class setting by selecting one class as the target class, and considering all other classes as non-target.
If not specified otherwise, we train the proposed OCC over the target patterns, $\mathcal{S}_{tr}$, while we test the model on all patterns, i.e., all targets not used during the training and all available non-target patterns.
In the following, we shorten the proposed system as EOCC (standing for Entropic One-Class Classifier). The variant operating with the training scheme described in Sec. \ref{sec:fast_synthesis} is denoted as EOCC-1, while EOCC-2 is used to refer to the variant operating as described in Sec. \ref{sec:synthesis_crossval}.
The global optimization is implemented by a genetic algorithm.
It performs roulette wheel selection, two-point crossover, and random mutation on the parameters characterizing the dissimilarity measure. In addition, the genetic algorithm implements an elitism strategy which automatically imports the fittest individual into the next population; we set the population size to 30 individuals and the mutation rate to 0.3. Convergence criteria is determined by combining a maximum number of iterations/evolutions (here set to 100) and a check that evaluates if the best fitness is not changed over the last ten evolutions. Such settings are determined by a preliminary tuning of the system.
Fuzzification of the vertex clusters is performed by generating Gaussian membership functions. The width/size of the Gaussian is set according to $\tau_{\mathcal{C}_i}$, which is computed as the intra-cluster average distance. Gaussian membership functions are symmetric and they can be described by a single parameter (i.e., the width/size). This fact is in agreement with the single-parameter description of the intra-cluster distances distribution.
The RS, $\mathcal{R}$, is defined equal to $\mathcal{S}_{tr}$ (no selection is performed at all). Note that performing the DSR of the input data is not necessary when processing feature-based patterns. However, we do embed also this type of data to provide a uniform view of the experimental results.
$\gamma$ in (\ref{eq:rentropy_mst}) is set to 0.02; $\eta,\beta$ in (\ref{eq:obj_func}) and (\ref{eq:obj_func2}) are both set to 0.5.
Again, those settings are determined by a preliminary fine-tuning stage.
Software is implemented in standard C++ by means of the SPARE library \cite{spare_graph_2013}. Tests have been executed on a machine running a 64-bit Linux OS, equipped with an Intel(R) Core(TM) i7-3930K CPU \@ 3.20GHz and 32 Gb of RAM.
When considering the SDF, we must rely on an appropriate test set performance measure that takes into account the ``scorings'' (in our case, the membership degrees to the target class) assigned to the test patterns.
In particular, in this paper we consider the Area Under the ROC Curve (AUC); AUC is computed according to Ref. \cite{Fawcett:2006:IRA:1159473.1159475}. AUC is a robust statistics that gives the average probability that a target pattern is ranked higher than a non-target one.
When using the HDF, instead, we evaluate the obtained confusion matrix by analyzing standard measures such as accuracy, recall, precision, F-measure and so on.
The performance measure $P(\cdot, \cdot, \cdot)$ in (\ref{eq:obj_func2}) is defined as the accuracy on the validation set.
All results are reported in terms of averages of ten different runs executed with different random seeds; we report standard deviations and analysis of statistical significance with t-test.
\subsection{Synthetic Data}
\label{sec:exp_synth}
Fig. \ref{fig:ds1_full} illustrates an example in which the target class is distributed in three well-separated spherical clusters. EOCC is trained on the target instances (in red) and it is tested on the green and blue instances (the green instances actually belong to the target class, while those in blue are non-target instances) -- see Fig \ref{fig:ds1}. This is a very simple instance and in fact the EOCC solves the problem without errors (considering both training schemes).
The obtained membership degrees to the target class are plotted in Fig. \ref{fig:ds1_memberships}.
The fact that EOCC synthesizes three DRs corresponds to the best solution in terms of modularity, as demonstrated in Fig. \ref{fig:ds1_modularity}, where the modularity value (\ref{eq:normalized_modularity}) of all possible partitions derivable from the MST are plotted -- note that here we consider a specific instance of the parameters of the input data dissimilarity measure (weighted Euclidean metric).
Finally, in Fig. \ref{fig:ds1_entropy-modularity} we show the entropy and modularity trends during the iterations of the optimization (\ref{eq:obj_func}). Since the problem is simple, the increments are numerically small, although the monotonic trend is clearly recognizable. It is worth noting the early convergence at the 30-th iteration, since in fact both entropy and modularity -- and hence the fitness (\ref{eq:obj_func}) -- are stuck at the same values.
Fig. \ref{fig:ds2_full} shows a more subtle problem. The target instances shown in red are the same as in the previous problem. Pattern instances used for testing the OCC are now centered at $x=0.5$, but with a very narrow variance of the y-axis. Target instances used for testing are in blue and violet, while those represented in green are non-target instances -- see Fig. \ref{fig:ds2}.
Results achieved with EOCC-1 are good (for SDF, the AUC is 0.98), although it commits some errors considering the HDF (accuracy is 0.84, since eight test patterns of the target class are misclassified as non-target; represented in violet in figure). Three clusters are synthesized during the training. Fig. \ref{fig:ds2_memberships} reports the membership degrees obtained by EOCC-1.
On the other hand, EOCC-2 performs better in terms of SDF (see Fig. \ref{fig:ds2_memberships_crossval}; AUC is one, since EOCC-2 assigns always higher membership degrees to the target patterns. Moreover, also hard decisions are better, since zero errors are committed. Please note that, for clarity purpose, the validation data used for EOCC-2 is not shown in those examples.
This test is exemplar for the different views that can be obtained by considering soft or binary decisions in the one-class classification setting. In fact, the SDF may still provide a consistent picture of the correct labeling of the tested patterns even in complex situations. Additionally, this test yields a first insight on the possible difference in terms of recognition performance among EOCC-1 and EOCC-2 -- EOCC-2 is optimized explicitly towards those solutions that perform better.
Finally, Fig. \ref{fig:ds3} shows two additional relevant properties of EOCC.
In Fig. \ref{fig:ds3_ob} the training/test target data are distributed in two highly separated clusters denoting high variance on the y-axis. By assuming the weighted Euclidean metric as dissimilarity measure, $p$ of (\ref{eq:obj_func}) would consist in a real-valued vector in $[0, 1]^2$. EOCC synthesizes a model with two DRs by finding the best-performing solution, $p^*$, equal to $[0.995, 0.238]$ (average of five runs).
This in fact corresponds to what we expected, since the x-coordinate plays the most important role in magnifying the separation (entropy), while the y-coordinate the compactness (modularity) of the partition.
In Fig. \ref{fig:ds3_isolated} we show a situation in which the training data contains an isolated pattern (shown in the upper-left corner in green). In our interpretation of the one-class classification setting, such a pattern is not an outlier, since in fact outliers should be identified during the test stage. As a consequence, EOCC synthesizes four DRs.
\begin{figure*}[ht!]
\centering
\subfigure[]{
\includegraphics[bb=0 0 341 241,scale=0.6,keepaspectratio=true]{./DS1}
\label{fig:ds1}}
~
\subfigure[]{
\includegraphics[bb=0 0 345 243,scale=0.6,keepaspectratio=true]{./DS1_memberships}
\label{fig:ds1_memberships}}
\subfigure[]{
\includegraphics[bb=0 0 348 243,scale=0.6,keepaspectratio=true]{./DS1_modularity}
\label{fig:ds1_modularity}}
~
\subfigure[]{
\includegraphics[bb=0 0 333 241,scale=0.6,keepaspectratio=true]{./DS1_entr-mod}
\label{fig:ds1_entropy-modularity}}
\caption{In Fig. \ref{fig:ds1}, the red instances are used only during the training, while those in green are used for testing. Non-target instances are represented in blue. Fig. \ref{fig:ds1_memberships} shows the calculated membership values to the target class assigned by the SDF. Fig. \ref{fig:ds1_modularity} shows the trend of the modularity of all possible partitions derivable from the entropic MST. A peak is recognizable at the expected best partition order, i.e., three. Finally, Fig. \ref{fig:ds1_entropy-modularity} shows the entropy and modularity trends over the iterations of the optimization (EOCC-1 is assumed here).}
\label{fig:ds1_full}
\end{figure*}
\begin{figure*}[ht!]
\centering
\subfigure[]{
\includegraphics[bb=0 0 342 241,scale=0.6,keepaspectratio=true]{./DS2}
\label{fig:ds2}}
\subfigure[]{
\includegraphics[bb=0 0 340 243,scale=0.6,keepaspectratio=true]{./DS2_memberships}
\label{fig:ds2_memberships}}
~
\subfigure[]{
\includegraphics[bb=0 0 339 243,scale=0.6,keepaspectratio=true]{./DS2_memberships_cv}
\label{fig:ds2_memberships_crossval}}
\caption{In Fig. \ref{fig:ds2}, both blue and violet test patterns belong to the target class. The blue ones are correctly classified, while those in violet are misclassified in terms of HDF. Fig. \ref{fig:ds2_memberships} and \ref{fig:ds2_memberships_crossval} show, respectively, the membership degrees assigned to the test patterns by EOCC-1 and EOCC-2.}
\label{fig:ds2_full}
\end{figure*}
\begin{figure*}[ht!]
\centering
\subfigure[]{
\includegraphics[bb=0 0 346 241,scale=0.6,keepaspectratio=true]{./DS3_ob}
\label{fig:ds3_ob}}
~
\subfigure[]{
\includegraphics[bb=0 0 341 241,scale=0.6,keepaspectratio=true]{./DS3_isolated}
\label{fig:ds3_isolated}}
\caption{Fig. \ref{fig:ds3_ob} depicts a situation where training/test target data shows high variance on the y-axis; the best-performing parameters, $p^*$, are selected accordingly. In Fig. \ref{fig:ds3_isolated} an example in which the training target data contains an isolated pattern (in green).}
\label{fig:ds3}
\end{figure*}
\subsection{Results on UCI Datasets}
\label{sec:exp_uci}
Tab. \ref{tab:uci_ds} presents the details of the herein considered UCI datasets \cite{Bache+Lichman_2013}.
\citet{Juszczak20091859} provide results of experimentations completed for the one-class setting, which we use for comparison in our study; where possible, missing results have been retrieved from \cite{OCC_results}. Please note that to provide a consistent comparison with Ref. \cite{Juszczak20091859}, we actually considered the versions of the UCI datasets downloaded from \cite{OCC_results}. We consider two versions of those data: (i) non-normalized (as downloaded from the reference) and (ii) normalized by ensuring zero-mean and unit variance for each component.
The UCI datasets usually do not provide a validation set explicitly. Since our principal aim here is the comparison of the proposed EOCC with the results presented in Ref. \cite{Juszczak20091859}, we need to consider the same training/test split scheme. As a consequence, in the case of EOCC-2 each validation set, $\mathcal{S}_{vs}$, is generated by applying a suitable zero-mean Gaussian noise to a randomly selected 10\% of the target/non-target data in $\mathcal{S}_{ts}$.
Nonetheless, as a demonstration of reliability of the method, we report also the results achieved defining a proper validation set by taking 10\% of the target and non-target patterns from $\mathcal{S}_{ts}$ (in the following tables, this is indicated as ``EOCC-2\_10\%'').
Since these datasets contain patterns described by features, the dissimilarity measure that we use is the weighted and normalized Euclidean metric, with weights $\underline{\mathbf{w}}\in[0, 1]^u$, where $u$ is the number of features characterizing the dataset at hand.
For settings of the herein considered reference systems we refer the reader to Ref. \cite{Juszczak20091859}.
Tab. \ref{tab:uci_ds_results} and \ref{tab:uci_ds_results2} show the average AUC results for EOCC-1, EOCC-2, and EOCC-2\_10\%, respectively on the considered low- and high-dimensional UCI datasets.
Results of Tab. \ref{tab:uci_ds_results} show very competitive performances of all EOCC variants w.r.t. the others; in four out of seven seven datasets (i.e., BW, D, E, and I) we achieve the highest AUC, considering either the non-normalized and the normalized dataset instances. Data normalization, usually, does not affect the results, with the only exception for the L dataset, in which EOCC performances degrades significantly; it is worth noting that this is observed also for the competitors.
Results of Tab. \ref{tab:uci_ds_results2} still denote good EOCC performances, although we score the best AUC in two datasets only (S in the non-normalized case, and in SP normalized setting). Results on the other datasets are in general comparable, although we observe performance degradation for all three systems on C; however, results are never statistically worse than all competitors.
Nonetheless, it is worth pointing out that, although EOCC relies on an entropy estimator (\ref{eq:rentropy_mst}) that in turn is (indirectly) based on the data PDF, we do not observe any severe performance breakdown when processing high-dimensional data, as it is usually observed with PDF-based methods (such as Parzen).
This follows from the fact that we initially perform a DSR, whose dimensionality is given by the size of the representation set. Therefore, since we used $\mathcal{R}=\mathcal{S}_{tr}$ in the experiments, one should expect to see some performance degradation as the training data size grows (for instance, when considering AB, BW, D, BC-D, and CO). However, on the contrary we achieve good results on those datasets, demonstrating the effectiveness of the combined dissimilarity and graph based approach adopted in EOCC.
Standard deviations are in general very low, denoting highly stable and reliable results.
\begin{table*}[th!]\scriptsize
\begin{center}
\caption{UCI datasets considered in this study.}
\label{tab:uci_ds}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
\textbf{UCI Dataset} & \textbf{Acronym} & \textbf{Target class} & \textbf{\# Target} & \textbf{\# Non-target} & \textbf{\# Params} \\
\hline
\multicolumn{6}{|c|}{\textbf{\textit{Low-dimensional}}} \\
\hline
Abalone & AB & 1 & 1407 & 2770 & 10 \\
Biomed & BI & normal & 127 & 67 & 5 \\
Breast Wisconsin & BW & benign & 458 & 241 & 9 \\
Diabetes (prima indians) & D & present & 500 & 268 & 8 \\
Ecoli & E & pp & 52 & 284 & 7 \\
Iris & I & Iris-setosa & 50 & 100 & 4 \\
Liver & L & healthy & 200 & 145 & 6 \\
\hline
\multicolumn{6}{|c|}{\textbf{\textit{High-dimensional}}} \\
\hline
Arrhythmia & AR & normal & 237 & 183 & 278 \\
Breast cancer wisconsin (diagnostic) & BC-D & B & 357 & 212 & 30 \\
Breast cancer wisconsin (prognostic) & BC-P & N & 151 & 47 & 33 \\
Colon & C & normal & 22 & 40 & 1908 \\
Concordia & CO & 2 & 400 & 3600 & 1024 \\
Sonar & S & M & 111 & 97 & 60 \\
Spectf & SP & 0 & 95 & 254 & 44 \\
\hline
\end{tabular}
\end{center}
\end{table*}
\begin{table*}[thp!]\scriptsize
\begin{center}
\caption{Test set results on low-dimensional datasets. Results show the average AUC with standard deviations and significance test (t-test). Statistically significant results are in bold; not available is denoted with ``-''.}
\label{tab:uci_ds_results}
\begin{tabular}{|c|c|c|c|c|c|c|c|}
\hline
\textbf{System/Dataset} & \textbf{AB} & \textbf{BI} & \textbf{BW} & \textbf{D} & \textbf{E} & \textbf{I} & \textbf{L} \\
\hline
\multicolumn{8}{|c|}{\textbf{\textit{Non normalized data}}} \\
\hline
\rowcolor{lgray}EOCC-1 & 0.685(0.013) & 0.847(0.002) & \textbf{0.990(0.003)} & 0.607(0.046) & 0.953(0.002) & \textbf{1.000(0.000)} & 0.461(0.006) \\
\rowcolor{lgray}EOCC-2 & 0.831(0.001) & 0.864(0.003) & 0.989(0.001) & \textbf{0.717(0.005)} & \textbf{0.957(0.003)} & \textbf{1.000(0.000)} & 0.536(0.021) \\
\rowcolor{lgray}EOCC-2\_10\% & 0.819(0.002) & 0.868(0.007) & 0.929(0.061) & 0.677(0.019) & 0.954(0.003) & \textbf{1.000(0.000)} & 0.493(0.021) \\
\hline
Gauss & 0.861(0.002) & 0.900(0.004) & 0.823(0.002) & 0.705(0.003) & 0.929(0.003) & \textbf{1.000(0.000)} & 0.586(0.005) \\
MoG & 0.853(0.005) & 0.912(0.009) & 0.785(1.003) & 0.674(0.003) & 0.920(0.004) & \textbf{1.000(0.000)} & 0.607(0.006) \\
Na\"{\i}ve Parzen & 0.859(0.004) & \textbf{0.931(0.002)} & 0.965(0.004) & 0.679(0.003) & 0.930(0.008) & \textbf{1.000(0.000)} & \textbf{0.614(0.002)} \\
Parzen & 0.863(0.001) & 0.900(0.011) & 0.723(0.005) & 0.676(0.004) & 0.922(0.004) & \textbf{1.000(0.000)} & 0.590(0.003) \\
\textit{k}-Means & 0.792(0.011) & 0.878(0.012) & 0.846(0.035) & 0.659(0.007) & 0.891(1.006) & \textbf{1.000(0.000)} & 0.578(1.000) \\
1-NN & 0.865(0.001) & 0.891(0.008) & 0.694(0.006) & 0.667(0.007) & 0.902(0.009) & \textbf{1.000(0.000)} & 0.590(0.009) \\
\textit{k}-NN & 0.865(0.001) & 0.891(0.008) & 0.694(0.006) & 0.667(0.007) & 0.902(0.009) & \textbf{1.000(0.000)} & 0.590(0.009) \\
Auto-encoder & 0.826(0.003) & 0.856(0.022) & 0.384(0.009) & 0.598(1.008) & 0.878(1.000) & \textbf{1.000(0.000)} & 0.564(0.009) \\
PCA & 0.802(0.001) & 0.897(0.005) & 0.303(0.010) & 0.587(0.002) & 0.669(0.011) & 0.973(0.008) & 0.549(0.005) \\
SOM & 0.814(0.003) & 0.887(0.008) & 0.790(0.023) & 0.692(0.007) & 0.890(0.011) & \textbf{1.000(0.000)} & 0.596(0.007) \\
MST\_CD & \textbf{0.875(0.001)} & 0.898(0.010) & 0.765(0.018) & 0.669(0.007) & 0.897(0.009) & \textbf{1.000(0.000)} & 0.580(0.009) \\
\textit{k}-Centres & 0.760(0.008) & 0.878(0.024) & 0.715(0.124) & 0.606(0.016) & 0.863(0.012) & \textbf{1.000(0.000)} & 0.537(0.041) \\
SVDD & 0.806(0.001) & 0.220(0.003) & 0.700(0.006) & 0.577(0.098) & 0.894(0.008) & \textbf{1.000(0.000)} & 0.470(0.014) \\
MPM & 0.594(0.001) & 0.792(0.057) & 0.694(0.006) & 0.656(0.007) & 0.802(0.005) & \textbf{1.000(0.000)} & 0.587(0.009) \\
LPDD & 0.697(0.001) & 0.865(0.026) & 0.800(0.005) & 0.668(0.007) & 0.896(0.005) & \textbf{1.000(0.000)} & 0.564(0.026) \\
CHAMELEON & 0.706(0.004) & 0.727(0.019) & 0.669(0.008) & 0.651(0.010) & 0.758(0.016) & \textbf{1.000(0.000)} & 0.580(0.009) \\
\hline
\multicolumn{8}{|c|}{\textbf{\textit{Unit variance normalization}}} \\
\hline
\rowcolor{lgray}EOCC-1 & 0.693(0.005) & 0.867(0.005) & 0.853(0.020) & 0.670(0.024) & 0.928(0.011) & \textbf{1.000(0.000)} & 0.396(0.016) \\
\rowcolor{lgray}EOCC-2 & 0.831(0.001) & 0.878(0.006) & \textbf{0.995(0.001)} & \textbf{0.751(0.012)} & \textbf{0.957(0.004)} & \textbf{1.000(0.000)} & 0.460(0.026) \\
\rowcolor{lgray}EOCC-2\_10\% & 0.841(0.002) & 0.862(0.016) & \textbf{0.995(0.002)} & 0.709(0.023) & 0.954(0.007) & \textbf{1.000(0.000)} & 0.452(0.026) \\
\hline
Gauss & 0.862(0.000) & 0.899(0.005) & 0.985(0.001) & 0.721(0.003) & 0.929(0.003) & \textbf{1.000(0.000)} & 0.509(0.005) \\
MoG & 0.860(0.003) & 0.911(0.008) & 0.984(0.002) & 0.738(0.003) & 0.929(0.003) & \textbf{1.000(0.000)} & 0.494(0.006) \\
Na\"{\i}ve Parzen & 0.859(0.000) & \textbf{0.931(0.002)} & 0.987(0.001) & 0.678(0.003) & 0.930(0.008) & \textbf{1.000(0.000)} & 0.484(0.008) \\
Parzen & \textbf{0.877(0.001)} & 0.915(0.009) & 0.991(0.001) & 0.756(0.002) & 0.929(0.005) & \textbf{1.000(0.000)} & 0.469(0.008) \\
\textit{k}-Means & 0.801(0.003) & 0.902(0.009) & 0.984(0.001) & 0.712(0.010) & 0.878(0.015) & \textbf{1.000(0.000)} & 0.469(0.014) \\
1-NN & 0.862(0.001) & 0.914(0.012) & 0.991(0.001) & 0.721(0.002) & 0.906(0.008) & \textbf{1.000(0.000)} & 0.511(0.007) \\
\textit{k}-NN & 0.862(0.001) & 0.914(0.012) & 0.991(0.001) & 0.721(0.002) & 0.906(0.008) & \textbf{1.000(0.000)} & 0.511(0.007) \\
Auto-encoder & 0.836(0.000) & 0.890(0.013) & 0.960(0.002) & 0.658(0.005) & 0.888(0.023) & \textbf{1.000(0.000)} & \textbf{0.608(0.008)} \\
PCA & 0.826(0.001) & 0.776(0.031) & 0.920(0.004) & 0.640(0.006) & 0.655(0.013) & 0.920(0.008) & \textbf{0.608(0.008)} \\
SOM & 0.838(0.003) & 0.908(0.006) & 0.990(0.002) & 0.709(0.009) & 0.898(0.004) & \textbf{1.000(0.000)} & 0.487(0.017) \\
MST\_CD & - & 0.914(0.012) & 0.992(0.001) & 0.715(0.003) & 0.899(0.009) & \textbf{1.000(0.000)} & - \\
\textit{k}-Centres & 0.767(0.017) & 0.906(0.015) & 0.984(0.002) & 0.678(0.009) & 0.870(0.023) & \textbf{1.000(0.000)} & 0.483(0.006) \\
SVDD & 0.791(0.002) & 0.915(0.009) & 0.988(0.001) & 0.732(0.005) & 0.922(0.010) & \textbf{1.000(0.000)} & 0.490(0.010) \\
MPM & 0.735(0.002) & 0.909(0.010) & 0.991(0.001) & 0.729(0.003) & 0.922(0.007) & \textbf{1.000(0.000)} & 0.521(0.011) \\
LPDD & 0.751(0.002) & 0.889(0.008) & 0.989(0.001) & 0.634(0.005) & 0.947(0.004) & \textbf{1.000(0.000)} & 0.506(0.005) \\
CHAMELEON & - & - & - & - & - & - & - \\
\hline
\end{tabular}
\end{center}
\end{table*}
\begin{table*}[thp!]\scriptsize
\begin{center}
\caption{Same as Tab. \ref{tab:uci_ds_results}, but considering the high-dimensional UCI datasets.}
\label{tab:uci_ds_results2}
\begin{tabular}{|c|c|c|c|c|c|c|c|}
\hline
\textbf{System/Dataset} & \textbf{AR} & \textbf{BC-D} & \textbf{BC-P} & \textbf{C} & \textbf{CO} & \textbf{S} & \textbf{SP} \\
\hline
\multicolumn{8}{|c|}{\textbf{\textit{Non normalized data}}} \\
\hline
\rowcolor{lgray}EOCC-1 & 0.683(0.007) & 0.933(0.000) & 0.554(0.002) & 0.631(0.008) & 0.550(0.016) & 0.439(0.016) & 0.727(0.014) \\
\rowcolor{lgray}EOCC-2 & 0.775(0.016) & 0.938(0.001) & 0.585(0.021) & 0.654(0.014) & 0.783(0.032) & \textbf{0.998(0.000)} & 0.917(0.000) \\
\rowcolor{lgray}EOCC-2\_10\% & 0.707(0.019) & \textbf{0.946(0.005)} & 0.503(0.036) & 0.654(0.014) & 0.783(0.032) & 0.945(0.022) & 0.907(0.000) \\
\hline
Gauss & 0.606(0.006) & - & 0.591(0.009) & 0.704(0.011) & 0.803(0.017) & 0.680(0.031) & 0.833(0.033) \\
MoG & 0.577(0.166) & - & 0.511(0.017) & 0.500(0.000) & 0.500(0.011) & 0.704(0.035) & 0.776(0.031) \\
Na\"{\i}ve Parzen & 0.774(0.007) & - & 0.535(0.015) & 0.700(0.015) & 0.846(0.007) & 0.532(0.039) & 0.902(0.037) \\
Parzen & 0.577(0.166) & - & 0.586(0.029) & 0.364(0.224) & 0.502(0.022) & 0.805(0.031) & 0.879(0.027) \\
\textit{k}-Means & 0.766(0.006) & - & 0.536(0.021) & 0.668(0.031) & 0.862(0.025) & 0.698(0.037) & 0.923(0.017) \\
1-NN & 0.760(0.008) & - & 0.595(0.025) & 0.713(0.033) & 0.901(0.008) & 0.763(0.043) & 0.926(0.029) \\
\textit{k}-NN & 0.760(0.008) & - & 0.595(0.025) & 0.713(0.033) & 0.901(0.009) & 0.696(0.048) & 0.923(0.015) \\
Auto-encoder & 0.522(0.021) & - & 0.548(0.037) & 0.500(0.000) & 0.512(0.015) & 0.596(0.065) & 0.817(0.062) \\
PCA & \textbf{0.807(0.010)} & - & 0.574(0.018) & 0.707(0.016) & 0.824(0.004) & 0.696(0.033) & 0.901(0.030) \\
SOM & 0.772(0.007) & - & 0.523(0.030) & 0.682(0.026) & 0.887(0.020) & 0.801(0.034) & 0.975(0.021) \\
MST\_CD & 0.796(0.006) & - & \textbf{0.611(0.026)} & \textbf{0.733(0.030)} & \textbf{0.911(0.001)} & 0.811(0.031) & \textbf{0.981(0.026)} \\
\textit{k}-Centres & 0.767(0.016) & - & 0.584(0.055) & 0.684(0.029) & 0.815(0.036) & 0.668(0.041) & 0.909(0.016) \\
SVDD & 0.581(0.164) & - & 0.498(0.242) & 0.364(0.224) & 0.121(0.011) & 0.761(0.032) & \textbf{0.978(0.033)} \\
MPM & 0.771(0.005) & - & 0.053(0.001) & 0.500(0.000) & 0.901(0.006) & 0.785(0.030) & \textbf{0.980(0.074)} \\
LPDD & 0.577(0.166) & - & 0.539(0.183) & 0.418(0.200) & 0.864(0.004) & 0.636(0.027) & 0.934(0.033) \\
CHAMELEON & 0.760(0.008) & - & - & 0.391(0.051) & 0.807(0.004) & 0.778(0.010) & 0.944(0.007) \\
\hline
\multicolumn{8}{|c|}{\textbf{\textit{Unit variance normalization}}} \\
\hline
\rowcolor{lgray}EOCC-1 & 0.657(0.035) & 0.843(0.026) & 0.492(0.021) & 0.452(0.010) & 0.550(0.016) & 0.291(0.026) & 0.664(0.014) \\
\rowcolor{lgray}EOCC-2 & 0.745(0.005) & \textbf{0.941(0.018)} & 0.490(0.040) & 0.506(0.009) & 0.783(0.032) & 0.349(0.019) & \textbf{0.959(0.002)} \\
\rowcolor{lgray}EOCC-2\_10\% & 0.707(0.019) & 0.922(0.010) & 0.503(0.016) & 0.506(0.009) & 0.783(0.032) & 0.357(0.027) & 0.927(0.017) \\
\hline
Gauss & 0.768(0.004) & - & 0.508(0.008) & 0.713(0.029) & 0.858(0.000) & 0.657(0.008) & 0.934(0.008) \\
MoG & 0.761(0.004) & - & 0.526(0.016) & - & - & 0.643(0.015) & 0.948(0.008) \\
Na\"{\i}ve Parzen & 0.774(0.007) & - & 0.538(0.022) & 0.700(0.015) & 0.846(0.000) & 0.569(0.012) & 0.892(0.009) \\
Parzen & 0.773(0.005) & - & 0.522(0.017) & 0.364(0.224) & 0.000(0.000) & 0.695(0.008) & 0.958(0.011) \\
\textit{k}-Means & \textbf{0.787(0.006)} & - & 0.520(0.020) & 0.716(0.040) & 0.872(0.005) & 0.625(0.016) & 0.867(0.010) \\
1-NN & 0.776(0.005) & - & 0.517(0.014) & \textbf{0.743(0.012)} & \textbf{0.888(0.000)} & 0.698(0.006) & \textbf{0.959(0.011)} \\
\textit{k}-NN & 0.776(0.005) & - & 0.517(0.014) & \textbf{0.743(0.012)} & \textbf{0.888(0.000)} & 0.698(0.006) & \textbf{0.959(0.011)} \\
Auto-encoder & - & - & 0.520(0.011) & - & - & 0.594(0.014) & 0.850(0.001) \\
PCA & 0.776(0.004) & - & \textbf{0.557(0.011)} & 0.707(0.019) & 0.853(0.000) & 0.608(0.009) & 0.807(0.020) \\
SOM & \textbf{0.787(0.008)} & - & 0.511(0.021) & 0.729(0.019) & 0.878(0.001) & 0.711(0.012) & 0.860(0.003) \\
MST\_CD & 0.778(0.005) & - & 0.527(0.018) & 0.735(0.022) & \textbf{0.888(0.000)} & \textbf{0.715(0.006)} & 0.957(0.011) \\
\textit{k}-Centres & 0.778(0.011) & - & 0.528(0.020) & 0.732(0.018) & 0.849(0.013) & 0.622(0.013) & 0.817(0.013) \\
SVDD & 0.527(0.094) & - & 0.517(0.017) & 0.364(0.224) & 0.000(0.000) & 0.705(0.054) & 0.897(0.032) \\
MPM & 0.771(0.005) & - & 0.518(0.018) & 0.000(0.000) & 0.324(0.000) & 0.696(0.008) & 0.901(0.028) \\
LPDD & 0.783(0.006) & - & 0.531(0.017) & 0.368(0.224) & 0.741(0.013) & 0.644(0.005) & 0.956(0.009) \\
CHAMELEON & - & - & - & - & - & - & - \\
\hline
\end{tabular}
\end{center}
\end{table*}
\subsection{Results on IAM Graph Datasets}
\label{sec:exp_iam}
The IAM repository is a variegated database containing many different datasets of labeled graphs \cite{riesen+bunke2008}.
We consider here six datasets, namely: AIDS, GREC, Letter-Low, Letter-High, Mutagenicity, and Protein. Tab. \ref{tab:iam_ds} shows the relevant information regarding the data. IAM datasets already contain a suitable validation set. Therefore, to adapt the considered dataset to our setting, we just move all non-target patterns of the training set into $\mathcal{S}_{ts}$.
We process the input labeled graphs by means of the graph edit distance algorithm known as TWEC \cite{odse,gm_survey}.
TWEC is a fast (quadratic) heuristic solution to the graph edit distance problem \cite{gm_survey}, which solves the assignment problem of the vertices by a greedy strategy. TWEC is characterized by three parameters ranging in $[0, 1]$, controlling the importance of insertion, deletion, and substitution edit operations.
In Tab. \ref{tab:details_results_iam} we show the obtained results. To our knowledge, there are no results available for comparison considering the one-class classification setting over the IAM datasets. However, we report those results to demonstrate the wide and straightforward applicability of EOCC and for future experimental comparisons.
The herein reported results confirm that EOCC-2 is more effective than EOCC-1, especially when considering harder datasets, such as M and P.
Notably, the P dataset is known to be very hard (see results in \cite{odse,odse2_ijcnn_2013} for the multi-class case); in Tab. \ref{tab:details_results_iam} the obtained AUC denotes nearly a randomized classifier.
However, the accuracy is fairly high ($\simeq 0.9$, with a precision of 1 and recall of $\simeq 0.21$). This proves that when considering the HDF, in this case the system correctly rejects all non-target patterns, while it rejects also some target instance.
In general, the gap between EOCC-1 and EOCC-2 is reduced for both AUC and accuracy.
The number of synthesized DRs is lower for EOCC-1. This aspect magnifies the potential of EOCC-1. In fact, in different experiments the results of EOCC-1 and EOCC-2 are comparable, although EOCC-1 solves the problem in a much lower computing time and with fewer DRs, i.e., with less resources.
This is an important aspect to be evaluated on the basis of the specific application at hand.
Standard deviations are acceptable, confirming the stability of the system.
\begin{table}[thp!]\scriptsize
\begin{center}
\caption{The considered IAM datasets. Full details in \cite{riesen+bunke2008}.}
\label{tab:iam_ds}
\begin{tabular}{|c|c|c|c|c|c|}
\hline
\textbf{IAM Dataset} & \textbf{Acronym} & \textbf{Target class} & \textbf{\# Target} & \textbf{\# Non-target} \\
\hline
AIDS & A & a & 200 & 1600 \\
\hline
GREC & G & 1 & 50 & 1033 \\
\hline
Letter-Low & L-L & A & 150 & 1400 \\
\hline
Letter-High & L-H & A & 150 & 1400 \\
\hline
Mutagenicity & M & mutagen & 2401 & 1713 \\
\hline
Protein & P & 1 & 99 & 332 \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{table*}[thp!]\scriptsize
\begin{center}
\caption{Test set results on IAM datasets. For each dataset, the first row shows results of EOCC-1, the second one those of EOCC-2.}
\label{tab:details_results_iam}
\begin{tabular}{|c|c|c|c|}
\hline
\textbf{Dataset} & \textbf{AUC} & \textbf{Accuracy} & \textbf{\# DRs} \\
\hline
\multirow{2}{*}{A} & 0.977(0.012) & 0.942(0.002) & 2.000(0.000) \\
& 0.974(0.014) & 0.945(0.012) & 3.400(2.607) \\
\hline
\multirow{2}{*}{G} & 0.993(0.006) & 0.969(0.008) & 4.000(0.707) \\
& 1.000(0.000) & 0.973(0.015) & 8.000(0.000) \\
\hline
\multirow{2}{*}{L-L} & 1.000(0.000) & 0.988(0.002) & 2.400(0.547) \\
& 1.000(0.000) & 0.990(0.001) & 3.800(0.447) \\
\hline
\multirow{2}{*}{L-H} & 0.905(0.116) & 0.840(0.144) & 2.000(0.000) \\
& 0.967(0.004) & 0.966(0.002) & 29.400(2.302) \\
\hline
\multirow{2}{*}{M} & 0.501(0.132) & 0.493(0.223) & 3.000(1.000) \\
& 0.682(0.212) & 0.622(0.201) & 123.000(1.050) \\
\hline
\multirow{2}{*}{P} & 0.388(0.028) & 0.158(0.036) & 2.000(0.000) \\
& 0.554(0.025) & 0.921(0.012) & 26.000(1.732) \\
\hline
\end{tabular}
\end{center}
\end{table*}
\section{Conclusions and Future Directions}
\label{sec:conclusions}
In this paper, we have proposed and evaluated a novel one-class classification system, called EOCC.
The classifier has been designed by making use of an interplay of different techniques. The dissimilarity representation is exploited to make the system general-purpose. Graph-based techniques are employed for estimating information-theoretic quantities (i.e., the entropy) and for deriving the model of the classifier. The decision regions forming the model are obtained by exploiting the concept of modularity of a graph partition.
The decision regions are hence defined as clusters of vertices, which are further equipped with suitable membership functions.
This allows us to provide both hard (i.e., Boolean) and soft decisions about the recognition of test patterns.
We have validated the system over two types of benchmarks: (i) different feature-based UCI datasets and (ii) six IAM datasets of labeled graphs.
Overall, the comparisons made over the UCI datasets demonstrate the validity of the approach with respect to several state-of-the-art one-class classification systems taken from the literature.
Results on the IAM datasets prove the versatility and the effectiveness of the system in processing labeled graphs (a less conventional pattern type).
The one-class classification setting is very useful in all those situations where only patterns of interest are known. Such patterns are termed target instances.
Our solution is applicable to virtually any context, being based on the dissimilarity representation. This aspect is of particular interest, since it allows the user to model patterns according to their more suitable representation for the application at hand.
Future directions include usual improvements and variants of the herein discussed system (i.e., by changing suitable components but considering the same overall design). For instance, we used a complete Euclidean graph representation to estimate the entropy of the data mapped into the dissimilarity space. A more sparse representation could become handy when processing large volume of data.
Therefore, in the future we will evaluate entropic spanning graphs based on the so-called \textit{k}-NN graphs.
Other global optimization techniques are of course of interest, as well as other membership function models for the decision regions (e.g., by exploiting non-symmetric membership functions).
Particular attention will be devoted to the issue of \textit{interpretability} of the system.
Usually, researchers focus on improving performances of pattern recognition systems from the pure technical viewpoint, such as by paying attention on the generalization capability and the computing speed. However, when facing applications of true scientific interest, such systems should satisfy the requirement of producing results and using inference rules/functions that are easily understandable by humans.
This fact would allow field experts to easily gather useful insights about the underlying problem.
Our system is suitable for this mission, since it is conceived by exploiting fuzzy sets based techniques.
Therefore, an important future goal is the evaluation of the system from this specific viewpoint.
\bibliographystyle{abbrvnat}
|
\section{Introduction}
The idea of ``Natural SUSY" has become very popular in the last times, especially as a framework that justifies that e.g. the stops should be light (much lighter than the other squarks), say $m_{\tilde t}\stackrel{<}{{}_\sim} 600$~GeV. This is an attractive scenario since it gives theoretical support to searches for light stops and other particles at the LHC, a hot subject from the theoretical and the experimental points of view.
In a few words, the idea is to lie in a region of the minimal supersymmetric Standard Model (MSSM) parameter-space where the electroweak breaking is not fine-tuned (or not too much fine-tuned). This is reasonable since, as it is usually argued, the main phenomenological virtue of supersymmetry (SUSY) is precisely to avoid the huge fine-tuning associated to the hierarchy problem.
Then, the main argument is in brief the following: ``Stops produce the main radiative contributions to the Higgs potential, in particular to the Higgs mass-parameter $m^2$. To avoid fine-tunings these contributions should be reasonably small, not much larger than $m^2$ itself. Since they are proportional to the stop masses, the latter cannot be too large".
Other supersymmetric particles, like gluinos, are also constrained by the same reason; in particular the gluino mass is bounded from above due to its important contribution to the running of the stop masses, which implies a significant 2-loop contribution to $m^2$. In addition, Higgsinos should be light, as their masses are controlled by the $\mu-$parameter, which contributes to $m^2$. These statements sound reasonable and have been often used to quantify the ``naturalness" upper bounds on stop masses, gluino masses, etc.
Apart from the theoretical arguments, the experiments at the LHC, ATLAS and CMS, performed a large number of SUSY searches, covering a significant part of the parameter space~\cite{Aad:2014pda,Aad:2014wea,Aad:2014bva,ATLAS-CONF-2014-006,Chatrchyan:2013iqa,Chatrchyan:2013xna,Chatrchyan:2014lfa,CMS-PAS-SUS-13-020}. From the point of view of Natural SUSY models, the most interesting bounds are those for stops, gluinos and Higgsinos. The lower limits from direct production of stops reach as far as 650~GeV~\cite{Aad:2014kra}, but they are sensitive to the stop details, in particular the mass difference between the stop and the lightest supersymmetric particle (LSP). Much lighter stops are still allowed by the current experimental bounds once certain conditions on their decays are fulfilled~\cite{Rolbiecki:2013fia,Kim:2014eva,Curtin:2014zua}. Concerning the gluino, the current experimental bounds have a strong dependence on the masses of the light squarks. Assuming that the stops are the only squarks
lighter than the gluino (as suggested by the very ``Natural SUSY" rationale), the latter decays through a chain $\tilde{g} \to t\bar{t} \tilde{\chi}^0_1$, and the lower limit reaches $m_{\tilde g}\stackrel{<}{{}_\sim}$ 1.4~TeV~\cite{ATLAS-CONF-2013-061}. Again, with some additional assumptions on the decay chains this limit can be somewhat relaxed. Finally, the $\mu$ parameter is the least constrained at the LHC. Because of the low electroweak production cross-section and the large model dependence, it is entirely possible to have Higgsinos just above the current LEP limit, $\mu \gtrsim 95$~GeV~\cite{Heister:2002mn,Abbiendi:2002vz}. On the other hand, the LEP limit is rather model independent, even if the Higgsino is the LSP with three almost mass degenerate states around 100~GeV. All these bounds are relevant to establish the present degree of fine-tuning in different SUSY scenarios.
Another important experimental ingredient in connection with Natural SUSY is the physics of the Higgs boson \cite{Aad:2012tfa,Chatrchyan:2012ufa}. In particular, the Higgs mass plays a prominent role in naturalness arguments. According to the most recent analyses, $m_h = 125.36 \pm 0.41 $~GeV~\cite{Aad:2014aba} (ATLAS) and $m_h = 125.03 \pm 0.30$~GeV~\cite{CMS:2014ega} (CMS). It is interesting to note that the current measurements are already more accurate than the theoretical predictions, which have a $\sim 2-3$~GeV
uncertainty~\cite{Feng:2013tvd, Hahn:2013ria, Harlander:2008ju}.
Furthermore, all observed Higgs properties are remarkably close to the SM predictions~\cite{ATLAS-CONF-2014-009}, which, within the SUSY context, points to the decoupling limit~\cite{Dobado:2000pw}.
In this paper, we revisit the arguments leading to the previous Natural SUSY scenario, showing that some of them are weak or incomplete. In sect.~\ref{sect:rev}, we review the ``standard" Natural SUSY scenario, pointing out some weaknesses in the usual evaluation of its electroweak fine-tuning, i.e. the tuning to get the correct electroweak scale.
We also address the existence of two potential extra fine-tunings that cannot be ignored in the discussion, namely the tuning to get $m_h=m_h^{\rm exp}$ when stops are too light and the tuning to get a large $\tan\beta$.
In sect.~\ref{sect:EWFT}, we discuss the electroweak fine-tuning of the MSSM, showing its statistical meaning and its generic expression for any theoretical framework. We give in the appendix
tables and plots that allow to easily evaluate the fine-tuning for any theoretical model within the framework of the MSSM, at any value of the high-energy scale.
In sect.~\ref{sect:NatBounds}, we describe our method to rigorously extract bounds on the initial (high-energy) parameters and on the supersymmetric spectrum, from the fine-tuning conditions.
In sect.~\ref{sect:unconstrained}, we apply this method to obtain the numerical values of the various naturalness bounds for the unconstrained MSSM, defined at arbitrary high-energy scale, in a systematic way.
In sec.~\ref{sect:impactextraFT}, we evaluate the impact of the potential extra fine-tunings mentioned above, discussing also the correlation between soft terms that the experimental Higgs mass imposes in the MSSM and its consequences for the electroweak fine-tuning.
Finally, in sect.~\ref{sect:Summary} we present the summary and conclusions of the paper, outlining the main characteristics of Natural SUSY and their level of robustness against changes in the theoretical framework or the high-energy scale at which the soft parameters appear.
\section{The Natural SUSY scenario. A critical review}\label{sect:rev}
\subsection{The ``standard" Natural SUSY}
Naturalness arguments have been used since long ago \cite{Barbieri:1987fn} to constrain from above supersymmetric
masses\footnote{ For a partial list of references on naturalness in SUSY, see \cite{deCarlos:1993yy, Anderson:1994dz, Ciafaloni:1996zh, Bhattacharyya:1996dw, Chankowski:1997zh, Barbieri:1998uv, Kane:1998im, Giusti:1998gz, BasteroGil:1999gu, Chan:1997bi, Chacko:2005ra, Choi:2005hd, Nomura:2005qg, Kitano:2005wc, Nomura:2005rj, Lebedev:2005ge, Allanach:2006jc, Perelstein:2007nx, Allanach:2007qk, Cabrera:2008tj, Cassel:2009ps, Kobayashi:2009rn, Lodone:2010kt, Asano:2010ut, Cassel:2011tg, Kitano:2006gv, Giudice:2006sn, Barbieri:2009ev, Feng:1999mn, Feng:1999zg, Romanino:1999ut, Horton:2009ed} (before LHC ) and
\cite{ Strumia:2011dv, Akula:2011jx,Antusch:2012gv,
Martin:2013aha, Evans:2013jna, Hardy:2013ywa, Baer:2013bba, Arvanitaki:2013yja, Kaminska:2013mya,
Baer:2014ica, Bae:2014fsa, Chakraborti:2014gea, Mustafayev:2014lqa, Fowlie:2014xha, Craig:2013cxa,Antoniadis:2014eta, Kowalska:2014hza, Delgado:2014vha}.
}.
%
Already in the LHC era, they were re-visited in ref.~\cite{Papucci:2011wy} to formulate the so-called Natural SUSY scenario. For the purpose of later discussion, we summarize in this subsection the argument of ref.~\cite{Papucci:2011wy}, which have been invoked in many papers.
Assuming that the extra (supersymmetric) Higgs states are heavy enough, the Higgs potential can be written in the Standard Model (SM) way
\begin{eqnarray}
\label{SMpot}
V= m^2|H|^2+\lambda|H|^4 \ ,
\end{eqnarray}
where the SM-like Higgs doublet, $H$, is a linear combination of the two supersymmetric Higgs doublets, $H\sim \sin\beta H_u + \cos \beta H_d$. Then, the absence of fine-tuning can be expressed as the requirement of not-too-large contributions to the Higss mass parameter, $m^2$. Since the physical Higgs mass is $m_h^2 = 2|m^2|$, a sound measure of the fine-tuning is\footnote{This measure produces similar results to the somewhat standard parametrization of the fine-tuning, see eq.(\ref{BG}) below.} \cite{Kitano:2006gv}
\begin{equation}
\label{FT1}
\tilde \Delta=\left|\frac{\delta m^2}{m^2}\right|=\frac{2\delta m^2 }{m_h^2} \ .
\end{equation}
For large $\tan\beta$, the value of $m^2$ is given by $m^2= |\mu|^2 + m_{H_u}^2$, so one immediately notes that both $\mu$ and $m_{H_u}$ should be not-too-large in order to avoid fine-tuning (as has been well-known since many years ago). For the $\mu-$parameter this implies
\begin{equation}
\mu \stackrel{<}{{}_\sim} 200\, {\rm GeV}\left(\frac{m_{h}}{120\, {\rm GeV}}\right)\left(\frac{\tilde \Delta^{-1}}{20\%}\right)^{-1/2} \ .
\end{equation}
This sets a constraint on Higgsino masses. Constraints for other particles come from the radiative corrections to $m_{H_u}^2$. The most important contribution comes from the stops. Following ref.~\cite{Papucci:2011wy}
\begin{equation}\label{eq:der1}
\delta m_{H_{u}}^{2}|_{\rm stop}=-\frac{3}{8\pi^{2}}y_{t}^{2}\left(m_{Q_{3}}^{2}+ m_{U_{3}}^{2}+|A_{t}|^{2}\right)\log\left(\frac{\Lambda}{\, {\rm TeV}}\right),
\end{equation}
where $\Lambda$ denotes the scale of the transmission of SUSY breaking to the observable sector and the 1-loop leading-log (LL) approximation was used to integrate the renormalization-group equation (RGE). Then, the above soft parameters $m_{Q_{3}}^{2}$, $m_{U_{3}}^{2}$ and $A_{t}$ are to be understood at low-energy, and thus they control the stop spectrum. This sets an upper bound on the stop masses.
In particular one has
\begin{equation}
\label{eq:ft-stop}
\sqrt{ m_{\tilde t_{1}}^{2}+m_{\tilde t_{2}}^{2}}\lesssim 600\, {\rm GeV} \frac{\sin\beta}{(1+x^{2})^{1/2}}\left(\frac{\log\left(\Lambda/\, {\rm TeV}\right)}{3}\right)^{-1/2}\left(\frac{\tilde \Delta^{-1}}{20\%}\right)^{-1/2} \, ,
\end{equation}
where $x=A_{t}/\sqrt{m_{\tilde t_{1}}^{2}+m_{\tilde t_{2}}^{2}}$. Eq.~(\ref{eq:ft-stop}) imposes a bound on the lightest stop. Besides the stops, the most important contribution to $m_{H_u}$ is the gluino one, due to its large 1-loop RG correction to the stop masses. Again, in the 1-loop LL approximation used in ref.~\cite{Papucci:2011wy}, one gets
\begin{equation}\label{eq:gluino}
\delta m_{H_{u}}^{2}|_{\rm gluino} \simeq -\frac{2}{\pi^{2}}y_{t}^{2}\left(\frac{\alpha_{s}}{\pi}\right)|M_{3}|^{2}\log^{2}\left(\frac{\Lambda}{\, {\rm TeV}}\right)\, ,
\end{equation}
where $M_{3}$ is the gluino mass. From the previous equation,
\begin{equation}
M_{3} \lesssim 900\, {\rm GeV} \sin\beta \left(\frac{\log\left(\Lambda/\, {\rm TeV}\right)}{3}\right)^{-1}\left(\frac{m_{h}}{120\, {\rm GeV}}\right)\left(\frac{\tilde \Delta^{-1}}{20\%}\right)^{-1/2} \, .
\end{equation}
Altogether, the summary of the minimal requirements for a natural SUSY spectrum, as given in ref.~\cite{Papucci:2011wy}, is:
\begin{itemize}
\item two stops and one (left-handed) sbottom, both below $500-700~\, {\rm GeV}$.
\item two Higgsinos, \emph{i.e.}, one chargino and two neutralinos below $200-350~\, {\rm GeV}$. In the absence of other [lighter] chargino/neutralinos, their spectrum is quasi-degenerate.
\item a not too heavy gluino, below $900~\, {\rm GeV}-1.5~\, {\rm TeV}$.
\end{itemize}
In the next subsections we point out the weak points of the above arguments that support the `standard" Natural SUSY scenario. Part of those points have been addressed in the literature after ref.~\cite{Papucci:2011wy} (see ref.~\cite{Feng:2013pwa} for a recent and sound presentation of the naturalness issue in SUSY and references therein.)
\subsection{The dependence on the initial parameters}
\label{sec:initialcond}
The one-loop LL approximation used to write eqs.(\ref{eq:ft-stop}, \ref{eq:gluino}), from which the naturalness bounds were obtained, is too simplistic in two different aspects.
First, it is not accurate enough since the top Yukawa-coupling, $y_t$, and the strong coupling, $\alpha_s$, are large and vary a lot along the RG running. As a result, the soft masses evolve greatly and cannot be considered as constant, even as a rough estimate. This effect can be incorporated by integrating numerically the RGE, which corresponds to summing the leading-logs at all orders \cite{Ford:1992mv, Bando:1992np,Bando:1992wy,Einhorn:1982pp}.
Second, and even more important, the physical squark, gluino and electroweakino masses are not initial parameters, but rather a low-energy consequence of the initial parameters at the high-energy scale. This means that one should evaluate the cancellations required among those initial parameters in order to get the correct electroweak scale.
This entails two complications. First, there is not one-to-one correspondence between the initial parameters and the physical quantities, since the former get mixed along their coupled RGEs. Consequently, it is not possible in general to determine individual upper bounds on the physical masses, not even on the initial parameters. Instead, one should expect to obtain contour-surfaces with equal degree of fine-tuning in the parameter-space and, similarly, in the ``space" of the possible supersymmetric spectra. The second complication is that the results depend (sometimes critically) on what one considers as initial parameters.
The most dramatic example of the last statements is the dependence of $m_{H_u}^2$ on the stop masses in the constrained MSSM (CMSSM). In the CMSSM one assumes universality of scalar masses at the GUT scale, $M_X$. This is a perfectly reasonable assumption that takes place in well-motivated theoretical scenarios, such as minimal supergravity. Then one has to evaluate the impact of the initial parameters on $m_{H_u}^2$, and see whether or not the requirement of no-fine-tuning implies necessarily light stops. A most relevant analytic study concerning this issue is the well-known work by Feng et al. \cite{Feng:1999mn}, where they studied the focus
point \cite{Feng:1999mn, Feng:1999zg, Chan:1997bi} region of the CMSSM . In the generic MSSM, the (1-loop) RG evolution of a shift in the initial values of $m_{H_u}^2, m_{U_3}^2, m_{Q_3}^2$ reads
\begin{equation}
\frac{d}{dt} \left[ \begin{array}{c} \delta m_{H_u}^2
\\ \delta m_{U_3}^2 \\ \delta m_{Q_3}^2 \end{array} \right]
= \frac{y_t^2}{8\pi^2} \left[
\begin{array}{ccc}
3 & 3 & 3 \\
2 & 2 & 2 \\
1 & 1 & 1 \end{array} \right]
\left[ \begin{array}{c} \delta m_{H_u}^2
\\ \delta m_{U_3}^2 \\ \delta m_{Q_3}^2 \end{array} \right] \ ,
\end{equation}
where $t\equiv\ln Q$, with $Q$ the renormalization-scale, and $y_t$ is the top Yukawa coupling.
Hence, starting with the CMSSM universal condition at $M_X$: $m_{H_u}^2=m_{U_3}^2=m_{Q_3}^2=m_0^2$, one finds
\begin {equation}
\delta m_{H_u}^2 = \frac{\delta m_0^2}{2} \left\{3\
{\rm exp} \left[ \int_0^t \frac{6y_t^2}{8\pi^2} dt' \right]
- 1 \right\}.
\label{focus}
\end{equation}
Provided $\tan\beta$ is large enough, ${\rm exp}\! \left[ {6\over8\pi^2}
\int_0^t y_t^2 dt' \right] \simeq 1/3$ for the integration between $M_X$ and the electroweak scale, so the value of $m_{H_u}^2$ depends very little (in the CMSSM) on the initial scalar mass, $m_0$. However, the average stop mass is given by (see eq.(\ref{mtbound1}) below)
\begin{eqnarray}
\label{mstops}
\overline m^2_{\tilde t} \simeq 2.97 M_3^2 + 0.50 m_0^2 + \cdots \ ,
\end{eqnarray}
where $M_3$ is the gluino mass at $M_X$.
Therefore, if the stops are heavy {\em because} $m_0$ is large, this does {\em not} imply fine-tuning. This is a clear counter-example to the need of having light stops to ensure naturalness.
From the previous discussion it turns out that the most rigorous way to analyze the fine-tuning is to determine the full dependence of the electroweak scale (and other potentially fine-tuned quantities) on the initial parameters, and then derive the regions of constant fine-tuning in the parameter space.
These regions can be (non-trivially) translated into constant fine-tuning regions in the space of possible physical spectra.
This goal is enormously simplified if one determines in the first place the analytical dependence of low-energy quantities on the high-energy initial parameters, a task which will be carefully addressed in subsection~\ref{sect:fit}.
\subsection{Fine-tunings left aside} \label{sect:extraFT}
In a MSSM scenario, there are two implicit potential fine-tunings that have to be taken into account to evaluate the global degree of fine-tuning. They stem from the need of having a physical Higgs mass consistent with $m_h^{\rm exp}\simeq 125$~GeV and from the requirement of rather large $\tan\beta$. Let us comment on them in order.
\subsubsection*{Fine-tuning to get $m_h^{\rm exp}\simeq 125$~GeV}
As is well known, the tree-level Higgs mass in the MSSM is given by $(m_h^2)_{\rm tree-level}=M_Z^2 \cos^22\beta$, so radiative corrections are needed in order to reconcile it with the experimental value. A simplified expression of such corrections \cite{ Haber:1996fp, Casas:1994us,Carena:1995bx}, useful for the sake of the discussion, is
\begin{equation}\label{eq:der2}
\delta m_{h}^{2}=\frac{3 G_{F}}{\sqrt 2 \pi^{2}}m_{t}^{4}\left(\log\left(\frac{\overline m_{\tilde t}^{2}}{m_{t}^{2}}\right)+\frac{X_{t}^{2}}{\overline m_{\tilde t}^{2}}\left(1-\frac{X_{t}^{2}}{12\overline m_{\tilde t}^{2}}\right)\right)\ +\ \cdots\ ,
\end{equation}
with $\overline m_{\tilde t}$ the average stop mass and $X_{t}=A_{t}-\mu \cot\beta$. The $X_t$-contribution arises from the threshold corrections to the quartic coupling at the stop scale. This correction is maximized for $X_{t}=\sqrt 6 \overline m_{\tilde t}$ ($X_{t}\simeq 2 \overline m_{\tilde t}$ when higher orders are included). Notice that if the threshold correction were not present one would need heavy stops (of about 3~TeV once higher order corrections are included) for large $\tan \beta$ (and much heavier as $\tan\beta$ decreases, see ref. \cite{Cabrera:2011bi,Giudice:2011cg}); which is inconsistent with the requirements of Natural SUSY in its original formulation. However, taking $X_t$ close to the ``maximal" value, it is possible to obtain the correct Higgs mass with rather light stops, even in the $500-700~\, {\rm GeV}$ range; a fact frequently invoked in the literature to reconcile the Higgs mass with Natural SUSY.
On the other side, requiring $X_{t}\sim$ maximal, amounts also to a certain fine-tuning if one needs to lie close to such value with great precision. The precision (and thus the fine-tuning) required depends in turn on the values of $\tan\beta$ and the stop masses. Therefore, when analyzing the naturalness issue one should take into account, besides the fine-tuning associated with the electroweak breaking, the one associated with the precise value required for $X_t$. In subsection~\ref{sect:125} we will discuss the size of this fine-tuning in further detail.
\subsubsection*{Fine-tuning to get large $\tan\beta$}
The value of $\tan\beta \equiv \langle H_u\rangle/\langle H_d\rangle$ is given, at tree level, by
\begin{eqnarray}
\label{B}
\frac{2}{\tan\beta}\simeq \sin 2\beta=\frac{2B\mu}{m_{H_d}^2 + m_{H_u}^2+2\mu^2}=\frac{2B\mu}{m_A^2}\ ,
\end{eqnarray}
where $m_A$ is the mass of the pseudoscalar Higgs state; all the quantities above are understood to be evaluated at the low-scale. Clearly, in order to get large $\tan\beta$ one needs small $B\mu$ at low-energy. However, even starting with vanishing $B$ at $M_X$ one gets a large radiative correction due to the RG running. Consequently, very large values of $\tan\beta$ are very fine-tuned\footnote{The existence of this fine-tuning was first observed in ref. \cite{Nelson:1993vc, Hall:1993gn} and has been discussed, from the Bayesian point of view in ref. \cite{Cabrera:2009dm}.}, as they require a cancellation between the initial value of $B$ and the radiative contributions. On the other hand, moderately large values may be non-fine-tuned, depending on the size of the RG contribution to $B\mu$ and the value of $m_A$. Hence, a complete analysis of the MSSM naturalness has to address this potential source of fine-tuning.
\section{The electroweak fine-tuning of the MSSM}\label{sect:EWFT}
In the MSSM, the vacuum expectation value of the Higgs, $v^2/2 = |\langle H_u\rangle|^2 +|\langle H_d\rangle|^2$, is given, at tree-level, by the minimization relation
\begin{eqnarray}
\label{mu}
-\frac{1}{8}(g^2+g'^2)v^2 = -\frac{M_Z^2}{2}=\mu^2-\frac{m_{H_d}^2 - m_{H_u}^2\tan^2\beta}{\tan^2\beta-1} \ .
\end{eqnarray}
As is well known, the value of $\tan\beta$ must be rather large, so that the tree-level Higgs mass, $(m_h^2)_{\rm tree-level}=M_Z^2 \cos^22\beta$, is as large as possible, $\simeq M_Z^2$; otherwise, the radiative corrections needed to reconcile the Higgs mass with its experimental value, would imply gigantic stop masses \cite{Cabrera:2011bi,Giudice:2011cg} (see subsection~\ref{sect:extraFT} above) and thus an extremely fine-tuned scenario. Notice here that the focus-point regime is not useful to cure such fine-tuning since it only works if $\tan\beta$ is rather large and stop masses are not huge.
Therefore, for Natural SUSY the limit of large $\tan\beta$ is the relevant one. Then, the relation (\ref{mu}) gets simplified
\begin{eqnarray}
\label{min}
-\frac{1}{8}(g^2+g'^2)v^2 = -\frac{M_Z^2}{2}=\mu^2+m_{H_u}^2\ .
\end{eqnarray}
The two terms on the r.h.s have opposite signs and their absolute values are typically much larger than $M_Z^2$, hence the potential fine-tuning associated to the electroweak breaking.
It is well-known that the radiative corrections to the Higgs potential reduce the fine-tuning \cite{deCarlos:1993yy}. This effect can be honestly included taking into account that the effective quartic coupling of the SM-like Higgs runs from its initial value at the SUSY threshold\footnote{A convenient choice of the SUSY-threshold is the average stop mass, since the 1-loop correction to the Higgs potential is dominated by the stop contribution. Hence, choosing $Q_{\rm threshold}\simeq m_{\tilde t}$, the 1-loop correction is minimized and the Higgs potential is well approximated by the tree-level form.}, $\lambda(Q_{\rm threshold})=\frac{1}{8}(g^2+g'^2)$, until its final value at the electroweak scale, $\lambda(Q_{EW})$.
The effect of this running is equivalent to include the radiative contributions to the Higgs quartic coupling in the effective potential, which increase the tree-level Higgs mass, $(m_h^2)_{\rm tree-level}=2\lambda(Q_{\rm threshold})v^2=M_Z^2$, up to the experimental one, $m_h^2=2\lambda(Q_{EW})v^2$. Therefore, replacing $\lambda_{\rm tree-level}$ by the radiatively-corrected quartic coupling is equivalent to replace $M_Z^2 \rightarrow m_h^2$ in eq.(\ref{min}) above, i.e.
\begin{eqnarray}
\label{minh}
-\frac{m_h^2}{2}=\mu^2+m_{H_u}^2\ ,
\end{eqnarray}
which is the expression from which we will evaluate the electroweak fine-tuning in the MSSM. As mentioned above, the radiative corrections slightly alleviate this fine-tuning, since $m_h>M_Z$.
\subsection{The measure of the fine-tuning} \label{sect:FTmeasure}
It is a common practice to quantify the amount of fine-tuning using the parametrization first proposed by Ellis et al. \cite{Ellis:1986yg} and Barbieri and Giudice \cite{Barbieri:1987fn}, which in our case reads
\begin{eqnarray}
\label{BG}
\frac{\partial m_h^2}{\partial \theta_i} = \Delta_{\theta_i}\frac{ m_h^2}{\theta_i}\ , \ \ \ \ \ \Delta\equiv {\rm Max}\ \left|\Delta_{\theta_i}\right|\ ,
\end{eqnarray}
where $\theta_i$ is an independent parameter that defines the model under consideration and $\Delta_{\theta_i}$ is the fine-tuning parameter associated to it. Typically $\theta_{i}$ are the initial (high-energy) values of the soft terms and the $\mu$ parameter. Nevertheless, for specific scenarios of SUSY breaking and transmission to the observable sector, the initial parameters might be particular theoretical parameters that define the scenario and hence determine the soft terms, e.g. a Goldstino angle in scenarios of moduli-dominated SUSY breaking. We will comment further on this issue in subsection~\ref{sect:genericFT}.
It is worth to briefly comment on the statistical meaning of $\Delta_{\theta_i}$. In ref.~\cite{Ciafaloni:1996zh} it was argued that (the maximum of all) $|\Delta_{\theta_i}|$ represents the inverse of the probability of a cancellation among terms of a given size to obtain a result which is $|\Delta_{\theta_i}|$ times smaller. This can be intuitively seen as follows. Expanding $m_h^2(\theta_i)$ around a point in the parameter space that gives the desired cancellation, say $\{\theta_i^0\}$,
up to first order in the parameters, one finds that only a small neighborhood $\delta \theta_i\sim \theta^0_i/\Delta_{\theta_i}$ around this point gives a value of $m_h^2$ smaller or equal to the experimental value \cite{Ciafaloni:1996zh}. Therefore, if one assumes that $\theta_i$ could reasonably have taken any value of the order of magnitude of $\theta_i^0$, then only for a small fraction $\left|{\delta \theta_i}/{\theta^0_i}\right| \sim \Delta_{\theta_i}^{-1}$ of this region one gets $m_h^2\lower.7ex\hbox{$\;\stackrel{\textstyle<}{\sim}\;$} (m_h^{\rm exp})^2$, hence the rough probabilistic meaning of $\Delta_{\theta_i}$.
Note that the value of $\Delta$ can be interpreted as the inverse of the $p$-value to get the correct value of $m_h^2$. If $\theta$ is the parameter that gives the maximum $\Delta$ parameter, then\footnote{Notice that in the particular case when $\theta^0$ minimizes the value of $m_h$, then $\left.\partial m_h/\partial \theta\right|_{\theta=\theta_0}=0$. This lack of sensitivity at first order when $\theta_0$ is close to an stationary point, would seemingly imply no fine-tuning, according to the ``standard criterion". However, from the above discussion, it is clear that in this case the expansion at first order is meaningless; one should start at second order, and then it becomes clear that the fine-tuning is really very high since only when $\theta$ is close to $\theta_0$, one gets $m_h^2\lower.7ex\hbox{$\;\stackrel{\textstyle<}{\sim}\;$} (m_h^{\rm exp})^2$. In other words, the associated $p$-value would be very small.}
\begin{eqnarray}
\label{toy2}
p{\rm -value} \simeq \left|\frac{\delta \theta}{\theta_0}\right| \equiv \Delta^{-1}\ .
\end{eqnarray}
It is noteworthy that for the previous arguments it was implicitly assumed that the possible values of a $\theta_i-$parameter are distributed, with approximately flat probability, in the $[0,\theta^0_i]$ range. In a Bayesian language, the prior on the parameters was assumed to be flat, within the mentioned range. If the assumptions are different (either because the allowed ranges of some parameters are restricted by theoretical consistency or experimental data, or because the priors are not flat), then the probabilistic interpretation has to be consistently modified. These issues become more transparent using a Bayesian approach.
In a Bayesian analysis, the goal is to generate a map of the relative probability of the different regions of the parameter space of the model under consideration (MSSM in our case), using all the available (theoretical and experimental) information. This is the so-called {\em posterior} probability, $p(\theta_i|{\rm data})$, where `data' stands for all the experimental information and $\theta_i$ represent the various parameters of the model. The posterior is given by the Bayes' Theorem
\begin{eqnarray}
\label{Bayes}
p(\theta_i|{\rm data})\ =\ p({\rm data}|\theta_i)\ p(\theta_i)\ \frac{1}{p({\rm data})}\ ,
\end{eqnarray}
where $p({\rm data}|\theta_i)$ is the likelihood (sometimes denoted by ${\cal L}$), i.e. the probability density of observing the given data if nature has chosen to be at the $\{\theta_i\}$ point of the parameter space (this is the quantity used in frequentist approaches); $p(\theta_i)$ is the prior, i.e. the ``theoretical" probability density that we assign a priori to the point in the parameter space; and, finally, $p({\rm data})$ is a normalization factor which plays no role unless one wishes to compare different classes of models.
For the sake of concreteness, let us focus on a particular parameter defining the MSSM, namely the $\mu-$parameter\footnote{Of course, one can take here another parameter and the argument goes the same (actually, in some theoretical scenarios $\mu$ may be not an initial parameter). On the other hand, $\mu$ is a convenient choice since it is the parameter usually solved in terms of $M_Z$ in phenomenological analyses.}.
Now, instead of solving $\mu$ in terms of $M_Z$ and the other supersymmetric parameters using the minimization conditions (as usual), one can (actually should) treat $M_Z^{\rm exp}$, i.e. the electroweak scale, as experimental data on a similar footing with the other observables, entering the total likelihood, ${\cal L}$. Approximating the $M_Z$ likelihood as a Dirac delta,
\begin{eqnarray}
\label{likelihood}
p({\rm data}|M_1, M_2,\cdots, \mu)\ \simeq\ \delta(M_Z-M_Z^{\rm exp})\ {\cal L}_{\rm rest}\ ,
\end{eqnarray}
where ${\cal L}_{\rm rest}$ is the likelihood associated to all the physical observables except $M_Z$, one can marginalize the $\mu-$parameter
\begin{eqnarray}
\label{marg_mu}
&&p(M_1, M_2, \cdots | \ {\rm data} ) = \int d\mu\ p(M_1, M_2, \cdots , \mu |
{\rm data} )
\nonumber\\
&&\hspace{3cm}\propto\ {\cal L}_{\rm rest} \left|\frac{d\mu}{d M_Z}\right|_{\mu_Z}
p(M_1, M_2, \cdots , \mu_Z)\ ,
\end{eqnarray}
where we have used eqs.~(\ref{Bayes}, \ref{likelihood}). Here $\mu_Z$ is the value of $\mu$ that reproduces $M_Z^{\rm exp}$ for the given values of $\{M_1,M_2,\cdots \}$, and $p(M_1, M_2, \cdots, \mu)$ is the prior in the initial parameters (still undefined). Note that the above Jacobian factor in eq.(\ref{marg_mu}) can be written as\footnote{Notice that the dependence of $M_Z$ on $\mu$ is through eq.(\ref{minh}), which determines the Higgs VEV. Thus $\frac{d M_Z^2}{d \mu}\propto \frac{d m_h^2}{d \mu}$.}
\begin{eqnarray}
\label{BG_Bayes}
\left|\frac{d\mu}{d M_Z}\right|_{\mu_Z}\propto
\left|\frac{\mu}{\Delta_\mu}\right|_{\mu_Z},\
\end{eqnarray}
where the constant factors are absorbed in the global normalization factor of eq.(\ref{Bayes}). The important point is that the relative probability density of a point in the MSSM parameter space is multiplied by $\Delta_\mu^{-1}$, which is consistent with the above probabilistic interpretation of $\Delta$ \cite{Cabrera:2008tj, Cabrera:2009dm, Ghilencea:2012qk, Fichet:2012sn}.
Actually, the equivalence is exact if one assumes that the prior in the parameters is factorizable, i.e. $p(M_1, M_2, \cdots , \mu)=p(M_1) p(M_2) \cdots p(\mu)$, {\em and} $p(\mu_Z)\propto 1/\mu_Z$, so that the numerator in the r.h.s of (\ref{BG_Bayes}) is canceled when plugged in eq.(\ref{marg_mu}). This assumption can be realized in two different ways. First, if $\mu$ has a flat prior with range $\sim [0,\mu_Z]$, then the normalization of the $\mu-$prior goes like $\propto 1/\mu_Z$. This is exactly the kind of implicit assumption discussed above. Alternatively, if $\mu$ has a logarithmically flat prior, then $p(\mu)\propto 1/\mu$, with the same result (this is probably the most sensible prior to adopt
since it means that all magnitudes of the
SUSY parameters are equally probable).
In summary, the standard measure of the fine-tuning (\ref{BG}) is reasonable and can be rigorously justified using Bayesian methods. In consequence, we will use it throughout the paper. Nevertheless, it should be kept in mind that the previous Bayesian analysis also provides the implicit assumptions for its validity. If a particular theoretical model does not fulfill them, the standard criterion is inappropriate and should be consistently modified.
\subsection{Generic expression for the fine-tuning} \label{sect:genericFT}
Clearly, in order to use the standard measure of the fine-tuning (\ref{BG}) it is necessary to write the r.h.s. of the minimization equation (\ref{minh}) in terms of the initial parameters. This in turn implies to write the low-energy values of $m_{H_u}^2$ and $\mu$ in terms of the initial, high-energy, soft-terms and $\mu-$term (for specific SUSY constructions, these parameters should themselves be expressed in terms of the genuine initial parameters of the model). Low-energy (LE) and high-energy (HE) parameters are related by the RG equations, which normally have to be integrated numerically. However, it is extremely convenient to express this dependence in an exact, analytical way. Fortunately, this can be straightforwardly done, since the dimensional and analytical consistency dictates the form of the dependence,
\begin{eqnarray}
\label{mHu_gen_fit}
m_{H_u}^2(LE)&=&
c_{M_3^2}M_3^2 +c_{M_2^2}M_2^2 +c_{M_1^2}M_1^2 + c_{A_t^2}A_t^2+c_{A_tM_3}A_tM_3 +c_{M_3M_2}M_3M_2+ \cdots
\nonumber\\
&&\cdots +c_{m_{H_u}^2}m_{H_u}^2 +c_{m_{Q_3}^2} m_{Q_3}^2 +c_{m_{U_3}^2}m_{U_3}^2+\cdots
\\
\mu(LE)&=&c_\mu \mu \ ,
\label{mu_gen_fit}
\end{eqnarray}
where $M_i$ are the $SU(3)\times SU(2)\times U(1)_Y$ gaugino masses, $A_t$ is the top trilinear scalar coupling; and $m_{H_u}, m_{Q_3}, m_{U_3}$ are the masses of the $H_u-$Higgs, the third-generation squark doublet and the stop singlet respectively, all of them understood at the HE scale. The numerical coefficients, $c_{M_3^2}, c_{M_2^2},...$ are obtained by fitting the result of the numerical integration of the RGEs to eqs.(\ref{mHu_gen_fit}, \ref{mu_gen_fit}), a task that we perform carefully in the subsection~\ref{sect:fit}.
The above equations (\ref{mHu_gen_fit}, \ref{mu_gen_fit}) replace the one-loop LL expressions (\ref{eq:der1}, \ref{eq:gluino}) used in the standard Natural-SUSY treatment. If one considers the initial values of the soft parameters and $\mu$ as the independent parameters that define the MSSM, then one can easily extract the associated fine-tuning by applying eq.(\ref{BG}) to (\ref{minh}), and replacing $m_{H_u}^2$ by the expression (\ref{mHu_gen_fit}). Note that the above definition of $\Delta$, eq.(\ref{BG}), is actually not very different from the definition (\ref{FT1}) used in ref.~\cite{Papucci:2011wy}; actually they are identical for the parameters that enter as a single term in the sum of eq.(\ref{mHu_gen_fit}), e.g. $m_{\tilde U_3}^2$. Nevertheless, eq.(\ref{BG}) differs from eq.(\ref{FT1}) when the parameter enters in several terms, e.g. $M_3$.\footnote{Indeed, if eq.(\ref{FT1}) was refined to incorporate the $M_3-$dependent contributions to $m^2$, e.g. through their impact in the stop mixing,
the result would be very similar to that of eq.(\ref{BG}).}
On the other hand, the definition (\ref{BG}), besides being statistically more meaningful, allows to study scenarios where the initial parameters are not soft masses.
From eqs.(\ref{minh}, \ref{mHu_gen_fit}, \ref{mu_gen_fit})) it is easy to derive the $\Delta-$parameters (\ref{BG}) for any MSSM scenario.
A common practice is to consider the (HE) soft terms and the $\mu-$term as the independent parameters, say
\begin{eqnarray}
\label{Thetas}
\Theta_\alpha=\left\{\mu, M_3,M_2,M_1,A_t, m_{H_u}^2, m_{H_d}^2, m_{U_3}^2, m_{Q_3}^2, \cdots\right\},
\end{eqnarray}
which is equivalent to the so-called ``Unconstrained MSSM"\footnote{{ The name ``Unconstrained MSSM" could be a bit misleading in this context, since it would seem to imply that one is not doing any assumptions about the soft terms. But there is in fact an assumption, namely that they are not correlated. Note in particular that although the parameter space of the Unconstrained MSSM includes any MSSM, e.g. the ``Constrained MSSM", the calculation of the fine-tuning for the latter requires to take into account a specific correlation between various soft-terms. Still, we are showing in this section that the results for the Unconstrained MSSM allow to easily evaluate the fine-tuning in any other MSSM scenario.}} .Then one easily computes $\Delta_{\Theta_\alpha}$
\begin{eqnarray}
\label{BGTheta}
\Delta_{\Theta_\alpha}=\frac{\Theta_\alpha}{m_h^2}\frac{\partial m_h^2}{\partial \Theta_\alpha} = -2\frac{\Theta_\alpha}{m_h^2}\frac{\partial m_{H_u}^2}{\partial \Theta_\alpha} \ .
\end{eqnarray}
E.g. $\Delta_{M_3}$ is given by
\begin{eqnarray}
\label{deltaBGM3}
\Delta_{M_3} = -2\frac{M_3}{m_h^2}
\left(2c_{{M_3^2}} M_3+c_{A_tM_3}A_t+c_{M_3M_2}M_2+\cdots
\right)\ .
\end{eqnarray}
The identification $\frac{\partial m_h^2}{\partial \Theta_\alpha}\simeq -2\frac{\partial m_{H_u}^2}{\partial \Theta_\alpha}$ in eq.(\ref{BGTheta}) comes from eq.(\ref{minh}) and thus is valid for all the parameters except $\mu$, for which we simply have
\begin{eqnarray}
\label{BGmu}
\Delta_{\mu}=\frac{\mu}{m_h^2}\frac{\partial m_h^2}{\partial \mu}
=-4 c_\mu^2 \dfrac{\mu^2}{m_h^2} = -4 \left(\dfrac{\mu(LE)}{m_h^2}\right)^2 \ .
\end{eqnarray}
Besides, the term proportional to $m_{H_d}^2$ in eq.(\ref{mu}), which was subsequently neglected, can give relatively important contributions to $\Delta_{m_{H_d}^2}$ if $\tan\beta$ is not too large ($\stackrel{<}{{}_\sim} 10$), namely
%
\begin{eqnarray}
\label{BGmHd}
\Delta_{m_{H_d}^2}\simeq -2\frac{m_{H_d}^2}{m_h^2}\left(c_{m_{H_d}^2} - c'_{m_{H_d}^2}\ {({\tan^2\beta-1})^{-1}\ }\ \right)\ ,
\end{eqnarray}
where $c'_{m_{H_d}^2}\simeq 1$ denotes the $c-$coefficient of $m_{H_d}^2$ in the expression of the LE value of $m_{H_d}^2$ itself, see table \ref{tab:mhiggssq} in the appendix. In any case, the contribution of $m_{H_d}^2$ to the fine-tuning is always marginal.
Note that for any other theoretical scenario, the $\Delta$s associated with the
genuine initial parameters, say $\theta_i$, can be written in terms of $\Delta_{\Theta_\alpha}$
using the chain rule
\begin{eqnarray}
\label{chain}
\Delta_{\theta_i}\equiv\frac{\partial\ln m_h^2}{\partial \ln\theta_i}
=\sum_\alpha\Delta_{\Theta_\alpha} \frac{\partial \ln \Theta_\alpha} {\partial \ln\theta_i}
=
\frac{\theta_i}{m_h^2}
\sum_\alpha\frac{\partial m_h^2}{\partial \Theta_\alpha}
\frac{\partial\Theta_\alpha} {\partial \theta_i}\ .
\end{eqnarray}
Finally, in order to obtain fine-tuning bounds on the parameters of the model we demand $\left|\Delta_{\theta_i}\right| \stackrel{<}{{}_\sim} \Delta^{\rm max}$, where
$\Delta^{\rm max}$ is the maximum amount of fine-tuning one is willing to accept. E.g.
\begin{eqnarray}
\label{Deltamax}
\Delta^{\rm max}=100 \ ,
\end{eqnarray}
represents a fine-tuning of $\sim 1\%$.
\subsection{The fit to the low-energy quantities} \label{sect:fit}
Fits of the kind of eq.(\ref{mHu_gen_fit}) can be found in the literature, see e.g.\cite{Abe:2007kf,Martin:2007gf}.
However, though useful, they should be refined in several ways in order to perform a precise fine-tuning analysis.
The most important improvement is a careful treatment of the various threshold scales.
In particular, the initial MSSM parameters (i.e. the soft terms and the $\mu-$parameter) are defined at a high-energy (HE) scale, which is usually identified as $M_X$, i.e.
the scale at which the gauge couplings unify. Although this is a reasonable assumption, it is convenient to consider the HE scale as an unknown; e.g. in gauge-mediated scenarios it can be in principle any scale.
The low-energy (LE) scale at which one sets the SUSY threshold and the supersymmetric spectrum is computed, is also model-dependent.
A reasonable choice is to take $M_{\rm LE}$ as the averaged stop masses. As discussed above, at this scale the 1-loop corrections to the effective potential are minimized,
so that the potential is well approximated by the tree-level expression; thus eq.(\ref{mHu_gen_fit}) should be understood at this scale.
Nevertheless, in many fits from the literature $M_{\rm LE}$ is identified with $M_Z$. Finally, some parameters are inputs at $M_Z$, e.g. the gauge couplings, while
others, like the soft $B-$parameter (the coefficient of the bilinear Higgs coupling), have to be evaluated in order to reproduce the correct electroweak breaking with the value of $\tan\beta$ chosen.
Similarly, the value of the top Yukawa-coupling has to be settled at high energy in such a way that it reproduces the value of the top mass at the electroweak scale (which is below the LE scale).
All this requires to divide the RG-running into two segments, $[M_{\rm EW},\ M_{\rm LE}]$ and $[M_{\rm LE},\ M_{\rm HE}]$.
Besides this refinement, we have integrated the RG-equations at two-loop order, using {\tt SARAH 4.1.0} \cite{Staub:2013tta}.
The results of the fits for all the LE quantities for $\tan\beta = 10$ and $M_{\rm HE}=M_X$
are given in the appendix, Tables \ref{tab:mhiggssq}, \ref{tab:msquarkthsq}, \ref{tab:msquarkstsq}, \ref{tab:msqslept}, \ref{tab:gauginos}, \ref{tab:trilinears},
and \ref{tab:mu-Bmu} .
The value quoted for each $c-$coefficient has been evaluated at $M_{\rm LE} = 1$~TeV. The dependence of the $c-$coefficients on $M_{\rm LE}$ is logarithmic and can be well approximated by
\begin{eqnarray}
\label{scaledep}
c_i(M_{\rm LE}) \simeq c_i(1\ {\rm~TeV}) + b_i\ln\frac{M_{\rm LE}}{1\ {\rm~TeV}}\ .
\end{eqnarray}
The value of the $b_i$ coefficients is also given in Tables \ref{tab:mhiggssq}--\ref{tab:mu-Bmu} (for $M_{\rm HE}=M_X$). Certainly, the value of $M_{\rm LE}\sim \overline{m_{\tilde t}}$ is itself a (complicated) function of the initial soft parameters.
Nevertheless, it is typically dominated by the (RG) gluino contribution, $M_{\rm LE}\sim \overline{m_{\tilde t}}\sim\sqrt{3}|M_3|$ for $M_{\rm HE}=M_X$. This represents an additional dependence of $m_{H_u}^2$ on $M_3$, which should be taken into account when computing $\Delta_{M_3}$. Actually, this effect diminishes the fine-tuning associated to $M_3$ (which is among the most important ones) because the impact of an increase of $M_3$ in the value of $m_{H_u}^2$ becomes (slightly) compensated by the increase of the LE scale and the consequent
decrease of the $c_{M_3^2}$ coefficient in eq.(\ref{mHu_gen_fit}). We have incorporated this fact in the computations of the fine-tuning.
Let us now turn to the dependence of the fine-tuning on the high-energy scale, $M_{\rm HE}$. The absolute values of all the $c-$coefficients in the fits decrease with $M_{\rm HE}$,
except perhaps the coefficient that multiplies the parameter under consideration (e.g. $c_{m_{H_u}^2}$ in eq.(\ref{mHu_gen_fit})).
In the limit $M_{\rm HE}\rightarrow M_{\rm LE}$ the latter becomes 1, and the others go to zero. Obviously, the fine-tuning decreases as $M_{\rm HE}$ decreases.
The actual dependence of the $c-$coefficients on $M_{\rm HE}$ has to do with the loop-order at which it arises.
If it does at one-loop, the dependence is logarithmic-like, e.g. for $c_{M_2^2}$ in eq.(\ref{mHu_gen_fit}); if it does at two-loop,
the dependence goes like $\sim (\log M_{\rm HE})^2$, e.g. for $c_{M_3^2}$. These dependences are shown in Figs. \ref{fig:mHusq}, \ref{fig:mQ3sq}, \ref{fig:mU3sq}, \ref{fig:Mgauginos} and \ref{fig:Bmu}.
In summary, with the help of Tables \ref{tab:mhiggssq}--\ref{tab:mu-Bmu} and Figs \ref{fig:mHusq}--\ref{fig:Bmu} it is straightforward to evaluate the fine-tuning parameters of any MSSM scenario.
\section{The naturalness bounds}\label{sect:NatBounds}
\subsection{Bounds on the initial (high-energy) parameters} \label{sect:HEbounds}
Let us explore further the size and structure of the fine-tuning, and the corresponding bounds on the initial parameters, in the unconstrained MSSM, i.e. taking as initial parameters the HE values of the soft terms and the $\mu$-term: $\Theta_\alpha=\left\{\mu, M_3,M_2,M_1,A_t, m_{H_u}^2, m_{H_d}^2, m_{U_3}^2, m_{Q_3}^2, \cdots\right\}$.
This is interesting by itself, and, as discussed above, it can be considered as the first step to compute the fine-tuning in any theoretical scenario. For any of those parameters we demand
\begin{eqnarray}
\label{FTbound}
\left|\Delta_{\Theta_\alpha}\right| \stackrel{<}{{}_\sim} \Delta^{\rm max}\ ,
\end{eqnarray}
where $\Delta_{\Theta_\alpha}$ are given by eq.(\ref{BGTheta}).
Now, for the parameters that appear just once in eqs.(\ref{mHu_gen_fit}, \ref{mu_gen_fit}) the corresponding naturalness bound (\ref{FTbound}) is trivial and has the form of an upper limit on the parameter size. For dimensional reasons this is exactly the case for dimension-two parameters in mass units, e.g. for the squared stop masses
\begin{eqnarray}
\label{BGmQ3}
\left|\Delta_{m_{Q_3}^2}\right|= \left|-2\frac{m_{Q_3}^2}{m_h^2}\ c_{m_{Q_3}^2}\right|\stackrel{<}{{}_\sim} \Delta^{\rm max} \ \ \longrightarrow\ \ \
m_{Q_3}^2\stackrel{<}{{}_\sim} 1.36\ \Delta^{\rm max}\ m_h^2
\end{eqnarray}
\begin{eqnarray}
\label{BGmU3}
\left|\Delta_{m_{U_3}^2}\right|= \left|-2\frac{m_{U_3}^2}{m_h^2}\ c_{m_{U_3}^2}\right|\stackrel{<}{{}_\sim} \Delta^{\rm max} \ \ \longrightarrow\ \ \
m_{U_3}^2\stackrel{<}{{}_\sim} 1.72\ \Delta^{\rm max}\ m_h^2 \ ,
\end{eqnarray}
where we have plugged , $c_{m_{Q_3}}=-0.367$, $c_{m_{U_3}}=-0.29$, which correspond to $M_{\rm HE}=M_X$ and $M_{\rm LE}= 1$~TeV, see Table~\ref{tab:mhiggssq}.
For $\Delta^{\rm max}=100$, we get
$m_{Q_3}\stackrel{<}{{}_\sim} 1.46$~TeV, $m_{U_3}\stackrel{<}{{}_\sim} 1.64$~TeV, substantially higher than the usual quoted bounds \cite{Feng:2013pwa}. This is mainly due to the refined RG analysis and the use of the radiatively upgraded expression eq.(\ref{minh}),
rather than eq.(\ref{min}), to evaluate the fine-tuning. We stress that these are the bounds on the high-energy soft masses, the bounds on the physical masses will be worked out in subsection~\ref{sect:boundsspectra}.
The naturalness bounds for the other (HE) dimension-two parameters ($m_{D_{3}}^2$, $m_{Q_{1,2}}^2$, $m_{U_{1,2}}^2$, $m_{D_{1,2}}^2$, $m_{L_3}^2$, ...) have a form similar to eqs.(\ref{BGmQ3}, \ref{BGmU3}) and are also higher than usually quoted. Due to its peculiar RGE, this is also the case of the $\mu-$parameter, see eq.(\ref{BGmu}).
On the other hand, for dimension-one parameters (except $\mu$) the naturalness bounds (\ref{FTbound}) appear mixed. In particular, this is the case for the bounds associated to $M_3, M_2, A_t$. From eqs.(\ref{BGTheta}) and (\ref{mHu_gen_fit})
\begin{eqnarray}
\label{BGM3}
\left|\Delta_{M_3}\right|&=&\frac{1}{m_h^2}\left|6.41M_3^2 -0.57A_tM_3+0.27M_3M_2\right| \stackrel{<}{{}_\sim} \Delta^{\rm max}
\\
\left|\Delta_{M_2}\right|&=&\frac{1}{m_h^2}\left|-0.81M_2^2 -0.14A_tM_2+0.27M_3M_2\right| \stackrel{<}{{}_\sim} \Delta^{\rm max}
\label{BGM2}\\
\left|\Delta_{A_t}\right|&=&\frac{1}{m_h^2}\left|0.44A_t^2 -0.57A_tM_3-0.14A_tM_2\right| \stackrel{<}{{}_\sim} \Delta^{\rm max} \ ,
\label{BGAt}
\end{eqnarray}
where, again, we have plugged the values of the $c-$coefficients corresponding to $M_{\rm HE}=M_X$ and $M_{\rm LE}= 1$~TeV. Other parameters, like $M_1, A_b$, get also mixed with them in the bounds, but their coefficients are much smaller, so we have neglected them. We show in
figure~\ref{fig:paralelogram} the region in the $\{M_2, M_3, A_t\}$ space that fulfills the inequalities for $\Delta^{\rm max} = 100$.
The figure is close to a prism. Their faces are given by the following approximate solution to eqs.(\ref{BGM3}--\ref{BGAt})
\begin{eqnarray}
\label{M3max}
M_3^{\rm max}\ \simeq\ \pm\ m_h \sqrt{\frac{\Delta^{\rm max}}{6.41}} + \frac{1}{12.82}(0.57A_t-0.27M_2)
\end{eqnarray}
\begin{eqnarray}
\label{M2max}
M_2^{\rm max}\ \simeq\ \pm\ m_h \sqrt{\frac{\Delta^{\rm max}}{0.81}} + \frac{1}{1.62}(0.27M_3-0.14A_t)
\end{eqnarray}
\begin{eqnarray}
\label{Atmax}
A_t^{\rm max}\ \simeq\ \pm\ m_h \sqrt{\frac{\Delta^{\rm max}}{0.44}} + \frac{1}{0.88}(0.57M_3+0.14M_2)\ ,
\end{eqnarray}
where the superscript ``max" denotes the, positive and negative, values of the parameter that saturate inequalities (\ref{BGM3}--\ref{BGAt}).
Thus eqs.(\ref{M3max}--\ref{Atmax}) represent the naturalness bounds to $M_3, M_2, A_t$. Each individual bound depends on the values of the other parameters due to the presence of the mixed terms. Depending on the relative signs of the soft terms, the bounds can be larger or smaller than those obtained when neglecting the mixed terms. However, the presence of the latter
stretches each individual {\em absolute} upper bound in a non-negligible way, by doing an appropriate choice of the other soft terms (compatible with their own fine-tuning condition).
\begin{figure}[ht]
\centering
\includegraphics[width=.5\linewidth]{figs/paralelogram_v1}
\caption{
Region in the $\{M_2, M_3, A_t\}$ space that fulfills eq. eqs.(\ref{BGM3}--\ref{BGAt}) for $\Delta^{\rm max}=100$ (axes units:~TeVs). For other values, a
$\sqrt {\Delta^{\rm max}/100 }$ scaling factor has to be applied. }
\label{fig:paralelogram}
\end{figure}
A generic, approximate, expression for the absolute upper bound on a dimension-one parameter, i.e. ${\cal M}_i$ $( {\cal M}_i = M_3, M_2, M_1, A_t, A_b ... )$ can be obtained by replacing the other dimension-one parameters, ${\cal M}_{j\neq i}$, by the values that saturate their zeroth-order fine-tuning bounds, $ \pm {\cal M}_j^{\rm max}\simeq m_h\sqrt {\Delta^{\rm max}/4|c_{{\cal M}_j^2}| }$, with the appropriate sign; namely
\begin{equation}
|{\cal M}_i | < \frac{m_h}{2} \sqrt{ \frac{\Delta^{\rm max}}{ | c_{{\cal M}_i^2 } |}}
\left(
1 + \sum_{j \neq i} \frac{1}{4} \frac { | c_{{\cal M}_i {\cal M}_j} |} { \sqrt{ | c_{{\cal M}_i^2 } c_{{\cal M}_j^2 } |} }
\right) \ .
\label{IneqGen}
\end{equation}
In practice, in order to obtain the absolute upper bounds on $M_3, M_2, A_t$ we have ignored the presence of additional parameters ($M_1, A_b, \cdots$) in (\ref{IneqGen}). Its inclusion would stretch even further the absolute bounds, but quite slightly and artificially since this would imply a certain conspiracy between soft parameters.
As a matter of fact, even playing just with the three parameters which show a sizeable correlation, i.e. $\{M_3, M_2, A_t\}$, implies a certain degree of conspiracy to get the maximum value quoted in (\ref{IneqGen}).
This means that the bound (\ref{IneqGen}) is conservative. A more restrictive and rigorous bound can be obtained by demanding that the addition in quadrature of the $\Delta_i$ parameters never exceeds the reference value, $\Delta^{\rm max}$. In any case, since the fine-tuning conditions of $M_3, M_2, A_t$ are correlated, as shown in eqs.~(\ref{BGM3}--\ref{BGAt}), the most meaningful approach is to determine the regions of the parameter space simultaneously consistent with all the fine-tuning conditions. This will be done in detail in section \ref{sect:Hmass} below.
The numerical modification of eqs.(\ref{BGM3}--\ref{Atmax}) for different values of $M_{\rm LE}$, $M_{\rm HE}$ can be straightforwardly obtained from Table \ref{tab:mhiggssq} and Figure \ref{fig:mHusq}.
Choosing $\Delta^{\rm max}=100$, eqs.(\ref{M3max}--\ref{Atmax}) give $|M_3| \stackrel{<}{{}_\sim} 610$~GeV, $|M_2| \stackrel{<}{{}_\sim} 1630$~GeV, $|A_t| \stackrel{<}{{}_\sim} 2430$~GeV. The limit on $M_3$ is similar to the one found by Feng \cite{Feng:2013pwa}, although this is in part a coincidence. In ref.~\cite{Feng:2013pwa} it was chosen $M_3^2$, rather than $M_3$, as an independent parameter; which reduces the associated $\Delta_{M_3}$ by a factor of 2. So, their bound on $M_3$ was increased (quite artificially in our opinion) by $\sqrt{2}$. On the other hand, in ref.~\cite{Feng:2013pwa} the RG running was not done in two steps, but simply running all the way from $M_X$ till $M_Z$.
Furthermore, they did not consider the mixed terms of eq.(\ref{BGM3}). And finally they used eq.(\ref{min}) instead of eq.(\ref{minh}) to evaluate the fine-tuning. It turns out that, all together, these three approximations increase the estimate of the fine tuning, thus decreasing the upper bound on $M_3$ by a factor which happens to be $\sim 1/\sqrt{2}$.
Actually, for the particular case of the $M_3-$parameter this is not the end of the story. As discussed in subsection~3.2, the $c_{M_3^2}$ coefficient has a dependence on $M_{\rm LE}$ approximately given by eq.(\ref{scaledep}). Since $M_{\rm LE}\simeq \overline{m_{\tilde t}}$ and typically $\overline{m}^2_{\tilde t}\simeq \frac{1}{2}(c_{M_3^2}^{(Q_3)}+c_{M_3^2}^{(U_3)})^2M_3^2$, where $c_{M_3^2}^{(Q_3)}$, $c_{M_3^2}^{(U_3)}$ are the coefficients of $M_3^2$ in the LE expression of $m^2_{Q_3}$, $m^2_{U_3}$ (given in table \ref{tab:msquarkthsq} and Figs.~\ref{fig:mQ3sq}, \ref{fig:mU3sq} for any HE scale), one has an additional contribution to the computation of $\Delta_{M_3}$ in eq.(\ref{BGTheta}).
The corresponding correction to $M_3^{\rm max}$ can be estimated by expanding the new inequality around the previous value of $M_3^{\rm max}$. We find
\begin{eqnarray}
\label{deltaM3}
\delta M_3^{\rm max} \simeq \frac{1}{2}\frac{b_{M_3}}{|c_{M_3^2}|}
\left( \frac{ \sqrt{\frac{1}{2}\left(c_{M_3^2}^{(Q_3)}+c_{M_3^2}^{(U_3)}\right)}\ M_3^{\rm max}}{1\ {\rm~TeV}}
-\frac12\right)\ M_3^{\rm max} \ ,
\end{eqnarray}
where we have neglected subdominant terms\footnote{Note that this correction is applicable as long as $M_{\rm HE}$ is large ($\stackrel{>}{{}_\sim} 10^{10}$~GeV); otherwise, it is quite small, the stop mass is not determined anymore by $M_3$.}.
For $M_{\rm HE}=M_X$ and $M_{\rm LE}= 1\ {\rm~TeV} $ one has $c_{M_3^2}\simeq -1.6$, $c_{M_3^2}^{(Q_3)}+c_{M_3^2}^{(U_3)}\simeq 6$, so the previous correction becomes
\begin{eqnarray}
\label{deltaM3_MX}
\delta M_3^{\rm max} \simeq \frac{2.06 M_3^{\rm max}-0.6 {\rm~TeV}}{10 {\rm~TeV}}\ M_3^{\rm max} \ .
\end{eqnarray}
This increases further $M_3^{\rm max}$ from 610~GeV to $\sim$ 660~GeV, i.e. $m_{\tilde g}\stackrel{<}{{}_\sim} 1440$~GeV which is about the present experimental lower limit on the gluino mass. Recall that this bound has been obtained assuming $\Delta^{\rm max}=100$, thus we conclude that the unconstrained MSSM is fine-tuned at about 1\%. We emphasize that these results have been obtained in the framework of the "unconstrained MSSM", so that $M_3,M_2,A_t$ are treated as independent, non-theoretically-correlated, parameters; and under the assumption $M_{\rm HE}=M_X$.
\subsection{Correlations between the soft terms}
Using the chain rule (\ref{chain}) one can easily evaluate the fine-tuning bounds when the initial soft terms are related in any way determined by the theoretical framework chosen. For instance, it is reasonable to assume that the soft masses at HE come from the same source, and therefore they are related, even if they are not equal. E.g. suppose that at HE
\begin{eqnarray}
\label{relations1}
\left\{m_{H_u}^2, m_{Q_3}^2, m_{U_3}^2\right\} = \left\{a_{H_u}, a_{Q_3}, a_{U_3}\right\} m_0^2 \ .
\end{eqnarray}
Then, plugging eq.(\ref{mHu_gen_fit}) into eq.(\ref{chain}) one immediately derives the fine-tuning condition for $m_0^2$
\begin{eqnarray}
\label{Deltam0}
\left|\Delta_{m_0^2}\right| = \left| -2 \frac{m_0^2}{m_h^2}\left(c_{m_{H_u}^2}a_{H_u} + c_{m_{Q_3}^2}a_{Q_3} + c_{m_{U_3}^2}a_{U_3}
\right)\right|\ \stackrel{<}{{}_\sim}\ \Delta^{\rm max}\ ,
\end{eqnarray}
which entails an upper bound on $m_0^2$, and hence on the stop masses at high energy. E.g.
\begin{eqnarray}
\label{mtRbound}
m_{U_3}^2\stackrel{<}{{}_\sim} \frac{1}{2}\left|\frac{\Delta^{\rm max}}{-0.29+0.631\frac{a_{H_u}}{a_{U_3}}-0.367 \frac{a_{Q_3}}{a_{U_3}}}\right| m_h^2\ ,
\end{eqnarray}
where we have used the $c-$coefficients corresponding to $M_{\rm LE}=1$~TeV, $M_{\rm HE}=M_X$ (Table~\ref{tab:mhiggssq}). This bound can be compared with the bound for the unconstrained MSSM, eq.(\ref{BGmU3}).
Depending on the relative values between the $a${\scriptsize s} , the bound on
$m_{U_3}^2$ gets increased (the usual case) or decreased. For the universal case, $a_{H_u}= a_{Q_3}= a_{U_3}$, one gets
$m_{U_3}\stackrel{<}{{}_\sim} \sqrt{\Delta^{\rm max}}\ 550\ {\rm~GeV}$, which allows for quite heavy stops with very little fine-tuning.
The same game can be played with the gaugino masses and the trilinear couplings. E.g. suppose that
\begin{eqnarray}
\label{relations2}
\left\{M_1,M_2,M_3, A_t\right\} = \left\{a_1, a_2, a_3, a_t\right\} M_{1/2} \ .
\end{eqnarray}
Then, the fine-tuning condition for $M_{1/2}$ reads $\left|\Delta_{M_{1/2}}\right|\stackrel{<}{{}_\sim}\ \Delta^{\rm max}$, with
\begin{eqnarray}
\label{DeltaM12}
\Delta_{M_{1/2}}= -4 \frac{M_{1/2}^2}{m_h^2}\left(c_{M_3^2}a_{3}^2 + c_{M_2^2}a_{2}^2 +c_{A_t^2}a_{t}^2+ c_{M_3M_2}a_3a_2+ c_{M_3A_t}a_3a_t+ c_{M_2A_t}a_2a_t
\right)\ .
\end{eqnarray}
E.g. the bound on $M_3$ becomes
\begin{eqnarray}
\label{m1/2bound}
M_{3}^2\ \stackrel{<}{{}_\sim}\ \frac{a_3^2}{4} \left|\frac{\Delta^{\rm max}}{1.6a_3^2-0.203a_2^2+ 0.109 a_t^2+
0.134a_3a_2 -0.285 a_3a_t -0.068a_2a_t}\right| m_h^2 \ ,
\end{eqnarray}
where, once more, we have used the $c-$coefficients corresponding to $M_{\rm LE}=1$~TeV, $M_{\rm HE}=M_X$.
For the universal case, $a_3=a_2=a_t$, the bound on $M_3$ becomes similar to that of the unconstrained MSSM. However, for other combinations the bound can be much larger. E.g. for $\frac{a_2}{a_3} = 3.16, -2.50$ and $a_t=0$ the
denominator would cancel.\footnote{See \cite{Antusch:2012gv, Martin:2013aha,Kaminska:2013mya} for some studies about non-universal gaugino masses and fine-tuning.} This represents a different kind of focus-point, in this case for gauginos.
Other correlations between the soft parameters and the appearance of alternative focus-point regimes can be explored in a similar way starting at any HE scale,
by using the tables and figures of the appendix. See refs.~\cite{Kowalska:2014hza, Delgado:2014vha} for recent work on this subject.
\subsection{Bounds on the supersymmetric spectrum} \label{sect:boundsspectra}
So far, in this section we have explained in detail how to extract the naturalness limits on the initial (HE) soft terms and $\mu-$term in generic MSSM scenarios. The next step is to translate those bounds into limits on the physical supersymmetric spectrum. Therefore, one has to go back from the high-energy scale to the low-energy one, using the RG equations. Once more, this can be immediately done using the analytical expressions discussed in subsection~\ref{sect:fit} and the appendix for any value of the HE and the LE scales.
Unfortunately, there is no a one-to-one correspondence between the physical masses, and the soft-parameters and $\mu-$term at high-energy. The only approximate exception are the gaugino and Higgsino masses. Namely, from Tables \ref{tab:gauginos}, \ref{tab:mu-Bmu}
\begin{eqnarray}
\label{Mgauginoshiggsinos}
&&M_{\tilde g} \simeq M_3(M_{\rm LE})\simeq 2.22 M_3
\nonumber\\
&&M_{\tilde W} \simeq M_2(M_{\rm LE})\simeq 0.81 M_2
\nonumber\\
&&M_{\tilde B} \simeq M_1(M_{\rm LE})\simeq 0.43 M_1
\nonumber\\
&&M_{\tilde H} \simeq \mu(M_{\rm LE})\simeq 1.002 \mu\ ,
\end{eqnarray}
where the above numbers correspond to $M_{\rm LE}=1$~TeV, $M_{\rm HE}=M_X$.
Of course, these are not yet the physical masses, except, approximately, for the gluino.
For a more precise calculation of the physical (pole) gluino mass, we must incorporate radiative corrections which depend on the
size of the squark masses and that can be rather significant for more than one squark generation with $m_{\tilde q} \gg M_3$ \cite{Martin:1993yx}.
The other gauginos and the Higgsinos get mixed in the chargino and neutralino mass matrices.
However, since we are considering upper limits on these masses, the mixing entries in those matrices are subdominant and do not appreciably affect the bounds.
On the other hand, as discussed in subsection~\ref{sect:HEbounds}, the naturalness limits on (the HE values of)$M_3$, $M_2$ are more involved than for other parameters, since the respective fine-tuning inequalities get mixed with each other and with $A_t$. Using the ($M_{\rm LE}=1$~TeV, $M_{\rm HE}=M_X$) limits on $M_3, M_2, M_1, \mu$ obtained for the unconstrained MSSM (see sects.~\ref{sect:HEbounds} and \ref{sect:unconstrained}) one gets
$M_{\tilde g} \stackrel{<}{{}_\sim} 1440 $~GeV,
$M_{\tilde W} \stackrel{<}{{}_\sim} 1300 $~GeV,
$M_{\tilde B} \stackrel{<}{{}_\sim} 3370 $~GeV and
$M_{\tilde H} \stackrel{<}{{}_\sim} 627 $~GeV.
On the contrary, the physical masses of the sparticles, $m_{\tilde t_1}^2$, $m_{\tilde t_2}^2$, $m_{Q_{1,2}}^2$ $m_{U_{1,2}}^2$, $m_{D_{1,2}}^2$, $m_{H^\pm}^2$, etc., are non-trivial combinations of the various initial soft terms and products of them. The case of the stops is particularly important, since it is a common assumption that Natural SUSY demands light stops. E.g. using $M_{\rm LE}=1$~TeV, $M_{\rm HE}=M_X$, we see from table~\ref{tab:mhiggssq} that the values of $m_{Q_3}^2$, $m_{U_3}^2$ at LE are given by:
\begin{eqnarray}
\label{mtLR_fit}
\hspace{-0.8cm} m_{Q_3}^2(M_{\rm LE})&=& 3.191M_3^2 +0.333M_2^2 +0.871 m_{\tilde Q_3}^2 -0.095m_{\tilde U_3}^2 -0.118m_{H_u}^2+0.072 A_tM_3+\cdots
\nonumber\\
\hspace{-0.8cm} m_{U_3}^2(M_{\rm LE})&=& 2.754M_3^2 -0.151 M_2^2 -0.192 m_{\tilde Q_3}^2 +0.706 m_{\tilde U_3}^2 -0.189 m_{H_u}^2 +0.159 A_tM_3+\cdots
\end{eqnarray}
These are not yet the physical stop masses. One has to take into account the top contribution, $m_t^2$, and the off-diagonal entries in the stop mass matrix, $\sim m_t X_t$ where $X_t = A_t+\mu \cot \beta\simeq A_t$. Finally, one has to extract the mass eigenvalues, $ m_{\tilde t_1}^2$ and $ m_{\tilde t_2}^2$. A representative, and easier to calculate, quantity is the average stop mass,
\begin{eqnarray}
\label{mtbound1}
\overline m_{\tilde t}^2&\equiv& \frac{1}{2}(m_{\tilde t_1}^2 +m_{\tilde t_2}^2)=\frac{1}{2}(m_{Q_3}^2(M_{\rm LE})+m_{U_3}^2(M_{\rm LE}))+m_t^2
\nonumber\\
&\simeq& (2.972 M_3^2+0.339 m_{Q_3}^2+0.305m_{U_3}^2 + 0.091M_2^2 - 0.154m_{H_u}^2 \cdots) +m_t^2 \ .
\end{eqnarray}
The average stop mass is also an important quantity to evaluate the threshold correction to the Higgs mass, and thus it plays an important role in the evaluation of the potential fine-tuning associated to it, see eq.~(\ref{eq:der2}) and subsection~\ref{sect:125}. Setting $M_3$, $m_{Q_3}$, $m_{U_3}$ and $M_2$ at their upper bounds (and neglecting additional terms in the parenthesis of eq.(\ref{mtbound1})) one obtains an upper bound for $\overline m_{\tilde t}$, namely $\overline m_{\tilde t}\stackrel{<}{{}_\sim} 1.7 \ {\rm~TeV}$. However, this is somehow too optimistic since it requires that all these HE parameters are simultaneously at their upper bounds, which is unlikely. A way to deal with this problem is to slightly modify the fine-tuning measure (\ref{BG}), in a (more restrictive) fashion, which counts all the contributions to the fine-tuning. Namely, instead using $\Delta\equiv {\rm Max}\ \left|\Delta_{\theta_i}\right|$, one defines $\Delta\equiv \{\sum_i \left|\Delta_{\theta_i}\right|^2\}^{1/2}$, which, as has been argued \cite{Casas:2005ev}, it is a more meaningful quantity. If the fine-tuning is dominated by one of the HE parameters (which is the usual case) both definitions are equivalent, but if there are several parameters contributing substantially to the fine-tuning, the second definition is more sensible (and restrictive). Then, it is easy to show that the maximum value of $\overline m_{\tilde t}^2$ subject to the condition $\Delta\leq \Delta_{\rm max} $, with $\Delta$ defined in this modified way, is
\begin{equation}
\overline m_{\tilde t}^2=\left[2.972 ^2 (M_3^{\rm max})^4 +
0.339^2 (m_{Q_3}^{\rm max})^4+ 0.305^2 (m_{U_3}^{\rm max})^4+0.091^2(M_2^{\rm max})^4+\cdots
\right] ^{1/2}+ m_t^2\ .
\nonumber
\end{equation}
Using just the dominant terms appearing explicitly above, we get (for $M_{\rm HE}=M_X$) $\overline m_{\tilde t}\leq 1320$ GeV.
From these results it is clear that for the unconstrained MSSM, with $M_{\rm HE}=M_X$, the naturalness bound on the gluino mass is much more important for LHC detection than the one on the stop masses.
Next, we show the numerical values of the various naturalness bounds in a systematic way.
\section{Results for the unconstrained MSSM} \label{sect:unconstrained}
The unconstrained MSSM, where the soft-terms and $\mu-$term at the HE scale are taken as the independent parameters, has been already considered in the previous subsections as a guide to discuss the various naturalness bounds. However, we have so far restricted ourselves to the case $M_{\rm LE}=1$~TeV, $M_{\rm HE}=M_X$. It is interesting to show the limits, both on the initial parameters and on the supersymmetric spectrum, for other choices of $M_{\rm HE}$.
Following the procedure explained in subsections~\ref{sect:HEbounds} and \ref{sect:boundsspectra},
we have computed the fine-tuning constraints for three representative values of $M_{\rm HE}$, namely $M_{\rm HE}= 2\times10^{16}$~GeV, $10^{10}$~GeV and $10^4$~GeV, keeping $M_{\rm LE}=1$~TeV. Using the plots shown in the appendix the reader can evaluate the bounds for any other choice of $M_{\rm HE}$.
The absolute upper bounds on the most relevant HE parameters, obtained from eq.(\ref{IneqGen}), with the additional correction (\ref{deltaM3}) for $M_3$, are shown in table \ref{tab:SoftMax}. Similarly, the corresponding bounds on supersymmetric masses at low energy, evaluated as in subsection~\ref{sect:boundsspectra}, are shown in table \ref{tab:PhysMax}. All the bounds have been obtained by setting $\Delta^{\rm max}=100$, they simply scale as $\sqrt{\Delta^{\rm max}/100}$.
\begin{table}[hbt]
\centering
{\small
\begin{tabular}{| l | c | c | c |}
\hline
~&$M_{\rm M_{\rm HE}}=2\times10^{16}$ &$M_{\rm M_{\rm HE}}=10^{10}$& $M_{\rm M_{\rm HE}}=10^{4}$ \\
\hline
$M_3^{\rm max}(M_{\rm HE})$ & 660 & 1 162 & 5 376 \\
$M_2^{\rm max}(M_{\rm HE})$ & 1 646 & 1 750 & 3 500 \\
$M_1^{\rm max}(M_{\rm HE})$ & 8 002 & 6 100 & 11 048 \\
$A_t^{\rm max}(M_{\rm HE})$ & 2 504 & 2 227& 3 094\\
$m_{H_u}^{\rm max}(M_{\rm HE})$ & 1 038 & 1 046 & 913 \\
$m_{H_d}^{\rm max}(M_{\rm HE})$ & 6 945 & 14 472 & 9 784 \\
$\mu^{\rm max}(M_{\rm HE})$ & 624 & 640 & 630 \\
$m_{Q_3}^{\rm max}(M_{\rm HE})$ & 1 458 & 1 687 & 3 527 \\
$m_{U_3}^{\rm max}(M_{\rm HE})$ & 1 640 & 1 828 & 3 710 \\
$m_{D_3}^{\rm max}(M_{\rm HE})$ & 5 682 & 7 812 & 20 277 \\
$m_{Q_{1,2}}^{\rm max}(M_{\rm HE})$ & 5 601 & 7 693 & 19 288 \\
$m_{U_{1,2}}^{\rm max}(M_{\rm HE})$ & 3 818 & 5 254 & 13 975 \\
$m_{D_{1,2}}^{\rm max}(M_{\rm HE})$ & 5 613 & 7 722 & 19 764 \\
$m_{L_{1,2,3}}^{\rm max}(M_{\rm HE})$ & 5 557 & 7 664 & 20 278 \\
$m_{E_{1,2,3}}^{\rm max}(M_{\rm HE})$ & 5 524 & 7 607 & 20 278\\
\hline
\end{tabular}
}
\caption{Upper bounds on some of the initial (HE) soft terms and $\mu-$term for three different values of $M_{\rm HE}$, in the unconstrained MSSM scenario. All quantities are given in~GeV units.}
\label{tab:SoftMax}
\end{table}
\begin{table}[hbt]
\centering
{\small
\begin{tabular}{| l | c | c | c |}
\hline
~&$M_{\rm HE}=2\times10^{16}$ &$M_{\rm HE}=10^{10}$& $M_{\rm HE}=10^{4}$ \\
\hline
$M_{\tilde g}^{\rm max}$ &1 440 & 1 890 & 5 860 \\
$M_{\tilde W}^{\rm max}$ & 1 303 &1 550 & 3 435\\
$M_{\tilde B}^{\rm max}$ & 3 368 & 4 237 & 10 565 \\
$M_{\tilde H}^{\rm max}$ & 627 & 627 & 627 \\
$\overline m_{\tilde t}^{\rm max}$ & 1 320 & 1 590 & 3 190 \\
$ m_{H^0}^{\rm max}$ & 7 252 & 14 510 & 9 900 \\
\hline
\end{tabular}
}
\caption{Upper bounds on some of the physical masses for three different values of $M_{\rm HE}$, in the unconstrained MSSM scenario. All quantities are given in~GeV units.
}
\label{tab:PhysMax}
\end{table}
\noindent
From the previous tables we can notice some generic facts.
\begin{itemize}
\item
Taking into account the present and future LHC limits, the upper bound on the gluino mass is typically the most stringent one, being at the reach of the LHC (for $\Delta^{\rm max}=100$), unless the high-energy scale is rather low. On the other hand, the gluino bound is the most sensitive to the value of $M_{\rm HE}$, since it is a two-loop effect. For $M_{\rm HE}\simeq10^7$~GeV, it is as already beyond the future LHC limit ($\sim 2.5$~TeV, see e.g. \cite{LHC-talk}) and it increases rapidly as $M_{\rm HE}$ approaches the electroweak scale.
\item
The upper bound on the wino mass, $M_{\tilde W}$, is similar to the gluino one. Note here that (unless $M_{\rm HE}$ is quite small) the weight of $M_2^2$ in the value of $m_{H_u}^2(M_{\rm LE})$ is certainly smaller than that of $M_3^2$; but this effect is compensated, when computing the physical masses, by the large increase of $M_3$ when running from $M_{\rm HE}$ to $M_{\rm LE}$ (see Figs.~1, 4). On the other hand, the bound on $M_{\tilde W}$ is much less restrictive than the one on $M_{\tilde g}$, given the LHC discovery potential. The upper bound on the bino, as expected, is quite mild and always beyond the reach of the next LHC run. This is just a consequence of the little impact that $M_1$ has on $m^2_{H_u}(LE)$.
\item
Concerning electroweakinos, the most relevant upper bounds are those on Higgsinos, $M_{\tilde H}$. Not only they are the strongest ones (hopefully at the reach of the LHC for $\Delta^{\rm max}=100$), but also, by far, the most stable of all bounds. This is a consequence of the fact that $\mu$ runs proportional to itself, so their fine-tuning parameter is insensitive to the HE scale, see eq.(\ref{BGmu}). Apart from that, the running of $\mu$ is very little. It is worth-mentioning that for $\Delta^{\rm max}=100$ the upper bound on $M_{\tilde H}$ is not far from $M_{\tilde H}\simeq 1$~TeV, which is the value required if dark matter is made of Higgsinos \cite{ArkaniHamed:2006mb, Cahill-Rowley:2014boa}.
\item
The upper bounds on stops are not as stringent as the gluino one unless $M_{\rm HE}$ is pretty close to the electroweak scale, in which case none of them is relevant. In general, it is not justified to say that Natural SUSY prefers light stops, close to the LHC limits. Actually, for $\Delta^{\rm max}=100$ the upper bounds on stops are beyond the LHC reach \cite{LHC-talk}. Taking lighter stops does not really improve the fine-tuning since there are other contributions to it which are dominant, in particular the gluino one.
\item
Given the present LHC limits, the contribution of the gluino to the $m_{H_u}^2$ is bigger than that of stops, then it is not useful to have light stops. This conclusion is reinforced when other aspects are considered, see subsection~\ref{sect:125} below. Unless $M_{\rm HE}$ is very small, the gluino mass sets the level of EW fine-tuning of the unconstrained MSSM, which is ${\cal O}(1\%)$.
If $M_{\rm LE}$ becomes close to the electroweak scale, the supersymmetric fine-tuning becomes much less severe. This fact is strengthened by the fact that additional soft dimension-4 Higgs operators may start to become relevant, increasing the tree-level Higgs mass and thus decreasing further the fine-tuning. These aspects were noted in
ref.~\cite{Brignole:2003cm,Casas:2003jx, Dine:2007xi}.
\item
Concerning the squarks of the first two generations and all the generations of sleptons, their bounds are, as expected, far beyond the reach of the LHC; the reason being that their contribution to $m_{H_u}^2$ is very small.
\item
Lastly, we can see the large upper bounds on $m_{H_d}^2$. When $M_{\rm LE}$ is very large, its contribution to the fine-tuning is very small. However, for low values of $M_{\rm LE}$, the term proportional to $m^2_{H_d}(\tan^2\beta-1)^{-1}$ (neglected for simplicity in expr.(\ref{minh})) actually becomes the dominant one, causing a larger impact of $m_{H_d}^2$ on the EW fine-tuning and, as consequence, decreasing the respective upper bound. This can be seen from table (\ref{tab:SoftMax}), where the bound on $m_{H_d}$ gets lower for $M_{\rm LE}=10^4$~GeV. Being the largest bounds as compared to the ones of $\mu$ and $m^2_{H_u}$, the term $m^2_{H_d}$ dominates the bounds on the masses of the heavy Higgses (see table \ref{tab:PhysMax}).
\end{itemize}
\section{Impact of other potential fine-tunings of the MSSM} \label{sect:impactextraFT}
\subsection{Fine-tuning to get $m_h^{\rm exp}\simeq 125$~GeV} \label{sect:125}
From the results of the previous section it is clear that, concerning naturalness, little is gained by going to light stops, say $< 800$~GeV. Actually, such light stops could entail, as already mentioned in section~\ref{sect:extraFT}, an additional fine-tuning since the condition $m_h^{\rm exp}\simeq 125$~GeV may require the threshold contribution to the Higgs mass to be maximal with high accuracy.
The relevant equation is
\begin{eqnarray}
\label{mh}
m_h^2=(m_h^2)_{\rm tree-level} \ +\ \delta_{\rm rad}m_h^2\ +\ \delta_{\rm thr}m_{h}^{2}\ ,
\end{eqnarray}
where $\delta_{\rm rad}m_h^2$ ($\delta_{\rm thr}m_{h}^{2}$) is the radiative (threshold) contribution to $m_h^2$, approximately given by the $X_t-$independent (dependent) part of eq.(\ref{eq:der2}). We recall that for moderately large $\tan\beta$ one can approximate $X_{t}=A_{t}(M_{\rm LE})-\mu \cot\beta\simeq A_t(M_{\rm LE})$. Figure~\ref{fig:higgs_mass} shows the dependence of $m_h$ vs $A_t(M_{\rm LE})$ for different values of the (LE) soft stop-masses, taken as degenerate for simplicity, $m_{Q_3} = m_{U_3} = 500, 1000, 2000$~GeV. If the stops are light, $\sim 500$~GeV, the
correct value of the Higgs boson mass, $m_h = 125 \pm 2$~GeV (the uncertainty is mainly due to the theoretical calculation), requires $A_t(M_{\rm LE})$ to be precisely fine-tuned\footnote{Note that in this case the ``standard criterion" to evaluate the fine-tuning, i.e. $\Delta=\partial \log m_h/\partial\log A_t$ is not applicable (indeed, one would conclude from it that there is no fine-tuning at all), since $A_t$ is close to an stationary point, see footnote 5.} at $\pm 1000$~GeV. On the other hand, if the stop masses are $\sim 1000$ or $2000$~GeV, a broad range of values is allowed, $A_t(M_{\rm LE}) = \pm (2000 \pm 1000)$~GeV, which entails no fine-tuning.
\begin{figure}[ht]
\centering
\includegraphics[width=0.75\linewidth]{figs/Higgs-mass2}
\caption{The Higgs boson mass, $m_h$, as a function of the third generation squark masses, $m_{Q_3} = m_{U_3}$. The black-solid line is for $m_{Q_3} = 500$~GeV, blue-dashed for $m_{Q_3} = 1000$~GeV, and green-dotted for $m_{Q_3} = 2000$~GeV. The red horizontal lines denote $m_h = 125 \pm 2$~GeV band. The Higgs boson mass has been calculated using
\texttt{FeynHiggs~2.10.1}~\cite{Hahn:2013ria, Frank:2006yh,
Degrassi:2002fi, Heinemeyer:1998np, Heinemeyer:1998yj}. \label{fig:higgs_mass}}
\end{figure}
We emphasize that this potential fine-tuning is independent of the one required to obtain the correct electroweak scale, which has been analyzed in the previous section. Therefore, if both fine-tunings are present we should combine them, i.e.\ multiply the two small probabilities of getting both the correct electroweak scale and the correct Higgs mass. This requires to quantify the fine-tuning associated to the Higgs mass in a fashion which has similar statistical meaning as the measure used for the electroweak fine-tuning. Taking into account the discussion of subsection~\ref{sect:FTmeasure}, we adopt here a fine-tuning measure that is also consistent with an interpretation in terms of $p$-value.
In particular, if the stops are light, the fine-tuning is well reflected by the $p-$value of getting $m_h$ as large as $m_h^{\rm exp}$ or larger
\begin{eqnarray}
\label{pvaluemh}
p-{\rm value} = \int_{m_h\geq m_h^{\rm exp}}\ d m_h\ {\cal P} (m_h) \ .
\end{eqnarray}
Here $ {\cal P} (m_h)$ is the probability of a Higgs mass value, given by
\begin{eqnarray}
\label{Pmh}
{\cal P} (m_h)=\left| \frac{dX_t}{dm_h}\right| {\cal P}(X_t(m_h)) \ ,
\end{eqnarray}
where ${\cal P} (X_t)$ is the probability distribution of $X_t-$values.\footnote{For many values of $m_h$, there are four $X_t$ solutions, so ${\cal P} (m_h)$ is the sum of four terms, corresponding to those solutions.} The final step is to assume a shape for ${\cal P} (X_t)$. Note here that $X_t\simeq A_t(M_{\rm LE})$ is a low-energy quantity, so it is not much sense to adopt a prior for it. Strictly speaking, the prior should be assumed for the initial, high-energy parameters that determine the value of $A_t(M_{\rm LE})$, (i.e.\ $A_t, M_3, M_2$), in a similar fashion as the one followed to establish the electroweak fine-tuning in the previous sections. Nevertheless, it is clear from figure~\ref{fig:higgs_mass} that, roughly speaking, for $\overline m_{\tilde t} \gtrsim 1000$~GeV and any sensible theoretical scenario for the soft terms, the $p-$value will be $\sim 20\%$ or larger, which means that there is not really a fine-tuning associated to $m_h\simeq 125$~GeV. Living in this range, the only important
fine-tuning is
the one associated to the electroweak scale. On the other hand, if stops are very light, both fine-tunings should be simultaneously considered. Then, one should multiply the $\Delta_{\rm electroweak}$ parameter by the inverse of the above $p$-value, which necessarily leads to a per-mil (or even more severe) global fine-tuning. So, interestingly, if the average stop mass is light, say $\stackrel{<}{{}_\sim} 800$~GeV, the situation is typically more fine-tuned than for heavier stops,
$\sim {\cal O}(1\ {\rm~TeV})$.
\subsection{The Higgs mass and the parameter space selected by naturalness} \label{sect:Hmass}
\vspace{0.2cm}
On the other hand, even if there is no fine-tuning to get the experimental Higgs mass, the requirement $m_h=m_h^{\rm exp}$ implies a balance between $\delta_{\rm rad}m_h^2$ and $\delta_{\rm thr}m_{h}^{2}$ in eq.(\ref{mh}), which in turn implies a correlation between the initial parameters, especially $M_3$ (the main responsible for the size of the stop masses) and $A_t$. This correlation has non-trivial consequences for the electroweak fine-tuning.
To see this, consider $\Delta_{M_3}$, which is usually the most significant fine-tuning parameter. As discussed in subsection~\ref{sect:HEbounds}, $\Delta_{M_3}$ is a function, not only of $M_3$, but also of $M_2$ and $A_t$. E.g. for $M_{\rm HE}=M_X, M_{\rm LE}= 1$~TeV, $\Delta_{M_3}$ is given by eq.(\ref{deltaBGM3}), where one can note that it will get partially suppressed as long as $M_3$ and $A_t$ are of the same sign. Therefore, fixing $M_3 > 0$ one would expect the lowest electroweak fine-tuning for $A_t > 0$. On the other side, it is evident from table~\ref{tab:trilinears} that the RG running pushes such $A_t$ towards rather low and possibly negative values. However, low values of $A_t$ at LE are in conflict with the measured Higgs boson mass, as can be seen in figure~\ref{fig:higgs_mass}. This will result in a tension between low fine tuning of the electroweak scale and the Higgs mass.
This situation is depicted in figure~\ref{fig:mh_finetuning} where we show the contours of constant Higgs boson mass (black) and fine tuning (red), together in the (high-energy) $M_3$--$A_t$ plane, for different choices of $M_\mathrm{HE}$. For simplicity we have chosen $M_2=M_3$ and $m_{Q_3}^2=m_{U_3}^2=0$ at HE. Note here that, unless the HE stop masses are very large, their LE values are essentially determined by $M_3$ (unless $M_\mathrm{HE}$ is small), so the results of the figures are quite general.
The fine-tuning shown corresponds to the largest $\Delta$ among the parameters. Usually it is given by $\Delta_{M_3}$, especially when there is a significant amount of running, although for large $|A_t|$ it may be given by $\Delta_{A_t}$ (then the red lines get horizontal in the plots).
As expected from the above discussion, when the fine-tuning is dominated by $\Delta_{M_3}$, it tends to be lower for $A_t > 0$;
however, $m_h \sim 125$~GeV prefers $A_t < 0$. For $M_\mathrm{HE} = 2\cdot 10^{16}$~GeV (upper-left panel of figure~\ref{fig:mh_finetuning}), the Higgs boson mass requires $A_t\sim -2000$~GeV, resulting in a large fine tuning, $\Delta \sim 250$. Moreover, $M_3$ is required to be larger than $\sim 750$~GeV which implies that the gluino mass should be (at least) slightly above current exclusion limits. Of course larger values of $M_3$ result in a more severe fine-tuning, as is clear from the figure. The tension between different low energy requirements is clearly visible in the upper-right panel, $M_\mathrm{HE} = 10^{10}$~GeV, where the correct Higgs mass is obtained for $A_t \sim -1500$~GeV with $\Delta\sim 100$ or even smaller, which corresponds to $M_3\sim 900$ GeV and, again, a physical gluino mass just above the current exclusion limits. Once more, higher values of the gluino mass imply higher fine-tuning, but the increase is not as dramatic as for $M_\mathrm{HE} = 2\cdot 10^{16}$~GeV. On the other hand, for positive $A_t$ a much higher value is required, $A_t \sim 3000$~GeV,
which results in a significant increase in fine tuning due to $A_t$, namely $\Delta_{A_t} \sim 300$.
Only for a very low choice of the high-energy scale, $M_\mathrm{HE} = 10^4$~GeV, the positive $A_t$ is preferred. In this case the fine-tuning gets substantially smaller, $\Delta \lesssim 50$.
The result is rather independent of $M_3$ which only enters at 2-loops in the Higgs mass and has a very limited impact on other SUSY parameters due to RGE running.
We can therefore conclude that, unless the scale of SUSY breaking transmission is quite low, the least fine-tuned scenarios (i.e. the most ``natural'' ones) generically demand negative $A_t$, a requirement driven by the measured Higgs mass. The corresponding fine-tuning is ${\cal O}(100)$, with gluinos only slightly heavier than the current limits, promising interesting discovery prospects at the second run of the LHC with increased center-of-mass energy.
\begin{figure}[ht]
\centering
\includegraphics[width=0.45\linewidth]{figs/mhe_2e16}
\includegraphics[width=0.45\linewidth]{figs/mhe_e10}\\
\includegraphics[width=0.45\linewidth]{figs/mhe_e4}
\caption{Contours of constant Higgs boson mass (black, contours for $m_h = 120,123,125,127$~GeV) and fine-tuning (red), eq.~\eqref{BGTheta}, in the $M_3$--$A_t$ plane. We have chosen $M_2=M_3$ and $m_{Q_3}^2=m_{U_3}^2=0$ at HE. From left to right to bottom, $M_\mathrm{HE} = 2\cdot 10^{16}, 10^{10}, 10^4$~GeV. The unphysical region with tachyonic stops is shaded in gray.
\label{fig:mh_finetuning}}
\end{figure}
\subsection{Fine-tuning to get large $\tan\beta$}\label{tanbeta}
As pointed out in subsection~2.3, a large value of $\tan\beta$ generically requires a small value of $B\mu$ at low energy, which requires a cancellation between the initial value and the radiative contribution from the RG-running. Here, we quantify this fine-tuning and discuss its consequences.
From eq.(\ref{B}), we can write, for $\tan\beta\gg 1$,
\begin{eqnarray}
\label{tB}
\tan\beta\simeq \frac{m_{H_d}^2 + m_{H_u}^2+2\mu^2}{B\mu}=\frac{m_A^2}{B\mu}\ ,
\end{eqnarray}
where $m_A$ is the mass of the pseudoscalar Higgs and all the quantities are understood at the low-energy (LE) scale. As discussed in subsection~\ref{sect:extraFT}, the fine-tuning to get large $\tan\beta$ can be reasonable quantified using the standard criterion. Namely, for any initial parameter of the theory, $\theta$, we define the associated fine-tuning, $\Delta_\theta^{(\tan\beta)}$
\begin{eqnarray}
\label{DeltatB1}
\Delta_\theta^{(\tan\beta)}=\frac{\theta}{\tan\beta}\frac{d \tan\beta}{d \theta}=\frac{\theta}{m_A^2}\left[\frac{d m_A^2}{d\theta}-\tan\beta\ \frac{d (B\mu)}{d \theta}\right]\ ,
\end{eqnarray}
where we have used eq.(\ref{tB}). For large $\tan\beta$, $\Delta_\theta^{(\tan\beta)}$ is normally dominated by the second term within brackets in (\ref{DeltatB1})
\begin{eqnarray}
\label{DeltatB}
\left|\Delta_\theta^{(\tan\beta)}\right|\simeq \tan\beta\left|\frac{\theta}{m_A^2}\frac{d (B\mu)}{d \theta}\right|\ .
\end{eqnarray}
The next step is to express the LE value of $B\mu$ in terms of the initial (HE) parameters. E.g. assuming $M_{\rm HE}=M_X$, $M_{\rm LE}=1$~TeV, from table~\ref{tab:mu-Bmu}
\begin{eqnarray}
\label{Bmulow}
B\mu (LE)\simeq B\mu + 0.46 M_3\mu -0.35M_2\mu-0.34 A_t \mu-0.03M_1\mu+\cdots ,
\end{eqnarray}
where the quantities on the r.h.s. are at the HE scale. Then, the corresponding fine-tuning $\Delta$s for the relevant parameters\footnote{Note that $\Delta_\mu^{(\tan\beta)}\simeq 1$.}, $B, M_3,M_2,A_t$, read
\begin{eqnarray}
\label{DeltatBs}
\left|\Delta_{\{B, M_3,M_2,A_t\}}^{(\tan\beta)}\right|\simeq \tan\beta\left| \frac{\mu}{m_A^2}\{B, \ 0.46M_3,\ 0.35M_2, \ 0.34 A_t
\}\right| \ ,
\end{eqnarray}
where we recall that r.h.s. parameters at the HE-scale. Going to particular models, one clearly expects some of the $\{\mu B, \ \mu M_3,\ \mu M_2, \ \mu A_t
\}$ quantities to be of the order of $m_A^2$. Indeed, the HE value of $B$ could be zero, but $M_3,M_2$ cannot.
This means that a certain fine-tuning, $\Delta^{(\tan\beta)}\stackrel{>}{{}_\sim} 5-10$ occurs if $\tan\beta\stackrel{>}{{}_\sim} 15-30$. Since this fine-tuning has a different nature from the electroweak one (discussed in detail in the previous sections), and given the probabilistic meaning of the fine-tuning parameters, this implies that the two $\Delta$s have to be multiplied, $\Delta=\Delta^{({\rm EW})}\Delta_\theta^{(\tan\beta)}$, which generically results in an exaggerated fine-tuning ($> 500-1000$). Notice that these conclusions are alleviated if the HE scale is smaller, since the numerical coefficients in (\ref{Bmulow}) decrease.
On the other hand, for $\Delta^{(\tan\beta)}\stackrel{<}{{}_\sim} 5$ there is no really fine-cancellation to get the value of $\tan\beta$ and one can ignore the $\Delta^{(\tan\beta)}$ fine-tuning factor.
The conclusion is that very large $\tan\beta$, say $\tan\beta\stackrel{>}{{}_\sim} 15-30$, implies a high fine-tuning price, unless the special characteristics of the model lead to a small r.h.s. in (\ref{DeltatB}), e.g. if $m_A^2$ is abnormally large.
Let us conclude this section pointing out that for $\tan\beta\stackrel{>}{{}_\sim} 30$ the impact of the bottom and tau Yukawa couplings in the RGEs become non-negligible, so the previous numerical values would be modified, but the general conclusion would be the same.
\section{Summary and Conclusions: the most robust predictions of a Natural SUSY scenario} \label{sect:Summary}
The idea of ``Natural SUSY", understood as a supersymmetric (MSSM) scenario where the fine-tuning is as mild as possible, is a reasonable guide to explore supersymmetric phenomenology, since, as usually argued, the main phenomenological virtue of SUSY is precisely to avoid the huge fine-tuning associated to the hierarchy problem. Much work has been done in the literature to quantify the fine-tuning of a generic MSSM and to extract the features of Natural SUSY. However, these analyses often ignore relevant aspects, such as the ``mixing" of the fine-tuning conditions or the presence of other potential fine-tunings.
In this paper, we have addressed the supersymmetric fine-tuning in a comprehensive way, including the discussion of the measure of the fine-tuning and its probabilistic meaning, the mixing of the fine-tuning conditions, the method to extract fine-tuning bounds on the initial parameters and the low energy supersymmetric spectrum, as well as the role played by extra potential fine-tunings. We have given tables and plots that allow to easily evaluate the fine-tuning and the corresponding naturalness bounds for any theoretical model defined at any high-energy (HE) scale. Finally, we have analyzed in detail the complete fine-tuning bounds for the unconstrained MSSM, defined at any HE scale, including the impact that the experimental Higgs mass imposes on the soft terms.
From the results of the previous sections, we summarize below the most important implications of fine-tuning in the MSSM; or, in other words, the characteristics of a Natural-SUSY scenario.
\begin{enumerate}
\item
For the evaluation of the fine-tuning it is crucial to define: i) the initial (independent) parameters of the theoretical setup, ii) the high-energy (HE) scale at which they are defined and iii) the criterion to quantify the fine-tuning.
We have seen that the `standard' fine-tuning criterion (\ref{BG}) normally has a sound statistical meaning, though one should be careful about the implicit assumptions of the prior for the initial parameters hold (if not, the standard criterion has to be consistently modified). Besides, we have provided tables and plots (see the appendix) that allow to straightforwardly evaluate the fine-tuning for any theoretical setup at any HE-scale.
\item
Concerning the electroweak fine-tuning of the MSSM (i.e. the one required to get the correct electroweak scale), the most robust result is by far that Higgsinos should be rather light, certainly below 700~GeV for $\Delta<100$, i.e. to avoid a fine-tuning stronger than 1\% (all the bounds on masses scale as $\sqrt{\Delta^{\rm max}}$). This result is enormously stable against changes in the HE scale since the $\mu-$parameter runs proportional to itself (besides running very little from HE to LE). The only way it could be substantially relaxed would be that the $\mu-$parameter were theoretically related to the soft masses in such a way that there occurred a cancellation at LE between $\mu^2$ and $m_{H_u}^2$ (see eq.(\ref{minh})). This is difficult to conceive and, certainly, it is not realized in the known theoretical SUSY frameworks. Incidentally, this upper bound is not far from $M_{\tilde H}\simeq 1$~TeV, which is the value required if dark matter is made of Higgsinos.
\item
The most stringent naturalness upper bound, from the phenomenological point of view, is the one on the gluino mass. If $M_{\rm HE} \simeq M_X$ one gets $M_{\tilde g}\stackrel{<}{{}_\sim} 1.5$~TeV for $\Delta^{\rm max}=100$, i.e. just around the corner at the LHC. In other words, the gluino mass typically sets the level of the electroweak fine-tuning of the MSSM, which at present is ${\cal O}(1\%)$.
However, this limit is not as robust as the one on Higgsinos. First, it presents a strong dependence on the HE-scale (due to the two-loop dependence of the electroweak scale on the gluino mass). Actually, for $M_{\rm HE}\simlt10^7$~GeV and $\Delta^{\rm max}=100$ the upper bound on $M_{\tilde g}$ (about 2.7~TeV) goes beyond the present LHC reach. In addition, it could be relaxed if the initial soft parameters (e.g. the gaugino masses) are theoretically related in a favorable way.
\item
The upper limit on the wino mass, $M_{\tilde W}$, is slightly smaller than the gluino one, but less relevant for LHC phenomenology. It also has a similar degree of robustness, though it is less dependent on $M_{\rm HE}$. The upper bound on the bino mass, $M_{\tilde B}$ is weaker and beyond the LHC reach.
\item
A remarkable conclusion is that light stop masses are {\em not} really a generic requirement of Natural SUSY. Actually, stops could be well beyond the LHC limits without driving the electroweak fine-tuning of the MSSM beyond 1\%. Even more, in some scenarios, like universal scalar masses with $M_{\rm HE}=M_X$, stops above 1.5~TeV are consistent with a quite mild fine-tuning of $\sim$ 10\%. Hence, the upper bounds on stops are neither stringent nor stable under changes of the theoretical scenario.
In contrast, as mentioned above, the gluino mass is required to be light with much more generality, although its impact on the fine-tuning depends crucially on the size of $M_{\rm HE}$ (it is maximum for $M_{\rm HE}=M_X$). Consequently, the electroweak scale is typically fine-tuned at 1\% in most cases, and having light stops does not help, since the electroweak fine-tuning stems from a single cancellation between terms, essentially between the ones proportional to $M_3^2$ and $\mu^2$ in eq.(\ref{minh}).
\item
In addition to the conventional fine-tuning to get the correct electroweak scale, there are two potential extra fine-tunings, namely the tuning of the threshold correction to get $m_h=m_h^{\rm exp}$ when stops are too light, and the tuning of $B\mu$ (at low energy) to get a large $\tan\beta$. It is convenient to avoid these additional fine-tunings, otherwise they have to be combined with (i.e. multiplied by) the electroweak fine-tuning, normally resulting in a gigantic global fine-tuning. Typically, this requires a not-too-light average stop mass, i.e. $\overline m_{\tilde t}\stackrel{>}{{}_\sim} 800$~GeV; and not-too-large $\tan\beta$, i.e. $\tan\beta\stackrel{<}{{}_\sim} 15-30$. The precise conditions to avoid these tunings are discussed in sect.~\ref{sect:impactextraFT}. Note that a small average stop mass is disfavored, but the mass of the lightest stop could be light or very light.
\item
Unless the high-energy scale is quite low, the less fine-tuned scenarios generically demand negative $A_t$, a requirement driven by the measured Higgs mass. The corresponding fine-tuning is ${\cal O}(100)$, with gluinos only slightly heavier than the current limits, which offers interesting prospects for the second run of the LHC.
\item
Lastly, the fine-tuning bounds on all the sleptons, the first two generations of squarks and the heavy Higgs states, are, as expected, far beyond the reach of LHC. This is a consequence of the little effect these parameters have on the value of $m_{H_u}^2$ at low energy.
\end{enumerate}
\section*{Acknowledgements}
The authors want to thank L. Calibbi, D. Cerdeno, S. Heinemeyer and L. Ibanez for very useful discussions.
This work has been partially supported by the MICINN, Spain, under contract FPA2010-17747, FPA2013-44773-P, Consolider-Ingenio CPAN CSD2007-00042, as well as MULTIDARK CSD2009-00064.
We also thank the Spanish MINECO {\em Centro de excelencia Severo Ochoa Program} under
grant SEV-2012-0249.
B.Z. is supported by the IISN, an ULB-ARC grant and by the Belgian Federal Science Policy
through the Interuniversity Attraction Pole P7/37, S.R. by Campus of Excellence UAM+CSIC and K.R. by Spanish Research Council (CSIC) within the JAE-Doc program.
\newpage
|
\section{Introduction}
An important way to get deeper insights into the power of various quantum resources and operations is to explore the power of various quantum variations of the basic models of classical automata. Of a special interest is to do that for various quantum variations of the classical automata, especially for those models that use very limited amounts of quantum resources: states, correlations, operations and measurements. This paper aims to contribute to such a line of research.
Number of (basis) states used is a natural complexity measure for (quantum) finite automata. The size of a (quantum) finite automaton is defined as the number of (basis) states of the (Hilbert) space on which the automaton will operate.
In case of a hybrid, that is quantum/classical finite automata, it is natural to consider both complexity measures~--~the number of classical states and also the number of quantum (basis) states.
Quantum finite automata were introduced by Kondacs and Watrous \cite{Kon97} and also by Moore and Crutchfields \cite{Moo97}, and since that time they were intensively explored \cite{Amb02,Bro99,Qiu11,Yak11}. State complexity and succinctness results are an important
research area of the classical finite automata theory, see \cite{Yu95}, with a
variety of applications. Once quantum versions of classical finite automata were introduced and explored, it started to be of large interest to find out, also through succinctness results, a relation between the power of classical and
quantum finite automata models. This has turned out to be an area of surprising
outcomes that again indicated that the relations between
the classical and the corresponding quantum finite automata models are intriguing.
In the past twenty years, state complexity of several variants of quantum finite automata were deeply and broadly studied \cite{Amb98,AmbNay02,Amb09,AmYa11,Ber05,BMP06,Fre09,Gru00,GQZ14,GQZ14b,Le06,Mer01,Mer02,Qiu14,Yak10,ZhgQiu112,Zhg12,Zhg13}.
State succinctness results were proved for some special languages and promise problems and for several automata models. The methods used to prove those results are various and often ad hoc.
It is therefore natural to try to find out whether there are quite general methods to get state succinctness results for quantum finite automata.
The answer is yes. We will show, in this paper, that state succinctness results can be derived in a nice way out of query complexity results. Here is the basic idea: State complexity is deeply related to communication complexity \cite{KusNis97b}.
Buhrman et al. proved that various communication complexity results can be derived out of query complexity results \cite{Buh98}. If a communication protocol is simple enough, then we can use quantum finite automata to implement it. By using this line of thought, state succinctness results can be derived.
Quantum query complexity is the quantum generalization of the model of decision tree complexity. In this model, an algorithm to compute a Boolean function $f:\{0,1\}^n\to \{0,1\}$ is charged for ``queries" to the input bits, while any intermediate computation is considered as free (see \cite{BdW02}).
Communication complexity was introduced by Yao \cite{Yao79} in 1979.
In the setting of two parties, Alice is given $x\in\{0,1\}^n$, Bob is given $y\in\{0,1\}^n$ and their task is to communicate in order to determine the value of some Boolean function $f:\{0,1\}^n\times\{0,1\}^n\to\{0,1\}$, while exchanging as small number of bits as possible. In this model, local computation
is considered to be free, but communication is considered to be expensive and has to be minimized.
Moreover, for computation, Alice and Bob can use all power available.
There are usually three types of communication complexities considered according to the models of protocols used by Alice and Bob: deterministic, probabilistic and quantum.
Query complexity and communication complexity are related to each other. By using a simulation technique that transforms quantum query algorithms
to quantum communication protocols, Buhrman et al. \cite{Buh98,Buh09} obtained new quantum communication protocols and showed the first impressively (exponential) gap between quantum and classical communication complexity. In the reverse direction, Buhrman et al. showed that how to use lower
bounds for quantum communication protocols to derive lower bounds
for quantum query algorithms.
State complexity of finite automata and communication complexity are also related to each other.
We can use communication complexity results to prove lower bounds on state complexity \cite{Hro01,Kla00,KusNis97b}. On the other hand, if the communication protocols are easy enough, then they can be simulated by finite automata and obtain new state complexity results (upper bounds) for finite automata.
Therefore, we can build connections from query complexity to state complexity. This could be a potential framework to get state succinctness results for quantum finite automata comparing to classical finite automata. We will demonstrate for several cases in this paper, that how to use quantum query complexity results to derive state succinctness results of finite automata.
We first consider the promise problem (partial function) studied in \cite{MJM11}. Namely, the problem
\begin{equation}
\mbox{DJ}'(x)=
\left\{
\begin{array}{ll}
1 & \hbox{if } W(x)\in\{0, 1, n-1, n\}\\
0 & \hbox{if }W(x)=\frac{n}{2},
\end{array}
\right.
\end{equation}
where $W(x)$ is the Hamming weight of $x$.
Montanaro et al. \cite{MJM11} gave a quantum query algorithm for $\mbox{DJ}'$ with 2 queries. However, their proof is quite complicated. Motivated by the method from \cite{AISJ13}, we give a simpler quantum query algorithm with 2 queries for $\mbox{DJ}'$.
Based on this simple query algorithm, we design a quantum communication protocol for the following promise problem
\begin{equation}
\mbox{EQ}'(x,y)=\left\{\begin{array}{ll}
1 &\ \mbox{if}\ H(x,y)\in\{0, 1, n-1, n \} \\
0 &\ \mbox{if}\ H(x,y)=\frac{n}{2},
\end{array}
\right.
\end{equation}
where $H(x,y)$ is the Hamming distance between bit strings $x$ and $y$. We further prove that the exact quantum communication complexity of $\mbox{EQ}'$ is ${\bf O}(\log n)$ while the deterministic communication complexity is ${\bf \Omega}( n)$.
Finally, we consider the promise problem $A(n)=(A_{yes}(n),A_{no}(n))$, where $A_{yes}(n)=\{x\#y\#\#x\#y\,|\,H(x,y)\in\{0, 1, n-1, n \}, \ x,y\in\{0,1\}^n\}$ and $ A_{no}(n)=\{x\#y\#\#x\#y\,|\,H(x,y)=\frac{n}{2},\ x,y\in\{0,1\}^n \}.$
We will prove that the promise problem $A(n)$ can be solved exactly by a one-way finite automata with quantum and classical state (1QCFA) with ${\bf O}(n^2)$ quantum basis states and ${\bf O}(n^3)$ classical states, whereas the sizes of the corresponding one-way deterministic finite automata (1DFA) are $2^{{\bf \Omega}(n)}$.
The paper is structured as follows. In Section 2 basic concepts and notations are introduced and models involved are
described in some details. A new quantum query algorithm is given for $\mbox{DJ}'$ in Section~3. Communication complexity of $\mbox{EQ}'$ is explored in Section~4. State complexity results for the promise problem $A(n)$ are showed in Section~5.
\section{Preliminaries}
In this section, we recall some basic definitions about query complexity, communication complexity and quantum finite automata. Concerning basic concepts and notations of quantum information processing and finite automata, we refer the reader to \cite{Gru97,Gru99,Gru00,Qiu12}.
\subsection{Exact query complexity}
Exact quantum query complexity for partial functions was dealt with in \cite{BH97,DJ92,GQZ14} and for total functions in \cite{Amb13,AISJ13,AGZ14,BBC+98,MJM11}.
Concerning more basic concepts and notations concerning query complexity, we refer the reader to \cite{BdW02}.
An exact classical (deterministic) query algorithm for computing a Boolean function $f:\{0,1\}^n\to \{0,1\}$ can be described
by a decision tree. A decision tree $T$ is a rooted binary tree where each internal vertex
has exactly two children, each internal vertex is labeled with a variable $x_i$ and each leaf is labeled with a value 0 or 1. $T$ computes a Boolean function $f$ as follows: The start is at the root. If this is a leaf then stop. Otherwise, query the variable $x_i$ that labels the root. If $x_i=0$, then recursively evaluate the left subtree, if $x_i=1$ then recursively evaluate the right subtree. The output of the tree is the value of the leaf that is reached at the end of this
process.
The depth
of $T$ is the maximal length of a path from the root to a leaf (i.e. the worst-case number of queries
used on any input). The {\em exact classical query complexity} (deterministic
query complexity, decision tree complexity) is the minimal depth over all decision trees computing $f$.
Let $f:\{0,1\}^n\to \{0,1\}$ be a Boolean function and $x = x_1x_2\cdots x_n$ be an input
bit string. An exact quantum query algorithm for $f$
works in a Hilbert space with some fixed number of basis states. It starts in a
fixed starting state, then performs on it a sequence of transformations
$U_1$, $Q$, $U_2$, $Q$, \ldots, $U_t$, $Q$, $U_{t+1}$.
Unitary transformations $U_i$ do not depend on
the input bits, while $Q$, called the query transformation, does,
in the following way. Each of the basis states corresponds to either one or none
of the input bits. If the basis state $|\psi\rangle$ corresponds to the $i$-th
input bit, then $Q|\psi\rangle=(-1)^{x_i}|\psi\rangle$. If it does not correspond to any
input bit, then $Q$ leaves it unchanged: $Q|\psi\rangle=|\psi\rangle$. Finally, the algorithm performs a measurement in the standard basis.
Depending on the result of the measurement, the algorithm outputs either 0 or 1
which must be equal to $f(x)$. The {\em exact quantum query complexity} is the minimum number of queries made by any quantum algorithm computing $f$.
\subsection{Communication complexity}
We recall here only very basic concepts and notations of communication complexity,
and we refer the reader to \cite{KusNis97b} for
more details. We will deal with the situation that there are two
communicating parties and with very simple tasks of
computing two inputs Boolean functions for the case one input is known to
one party and the other input to the other party.
We will completely ignore computational resources needed by
parties and focus solely on the amount of communication need to be
exchanged between both parties in order to compute the value of a given Boolean function.
More technically, let $X, Y$ be finite subsets of $\{0,1\}^n$. We will consider two-input functions $f: X\times Y\rightarrow \{0,1\}$ and two communicating parties. Alice is given $x\in X$ and Bob is given $y\in Y$. They want to compute $f(x,y)$. If $f$ is defined only on a proper subset of $X\times Y$, $f$ is said to be a partial function or a promise problem.
The computation of $f(x,y)$ will be done using a communication protocol. During the
execution of the protocol, the two parties alternate roles in sending messages. Each of these
messages is a bit-string. The protocol, whose steps are based on the communication so far, specifies also for each step whether the communication terminates
(in which case it also specifies what is the output). If the communication is
not to terminate, the protocol specifies what kind of message the sender (Alice or Bob) should send next, as
a function of its input and communication so far.
A deterministic communication protocol ${\cal P}$ computes
a (partial) function $f$, if for every (promised) input pair $(x,y)\in X\times Y$ the protocol terminates with the
value $f(x,y)$ as its output.
In a probabilistic protocol, Alice and Bob may also flip coins during the protocol execution and proceed according to its output and the protocol can also have an erroneous output with a small probability.
In a quantum protocol, Alice and Bob may use quantum resources to produce the output or (qu)bits for communication.
Let ${\cal P}(x,y)$ denote the output of the protocol ${\cal P}$.
For an exact protocol, that always outputs the correct answer, $Pr({\cal P}(x,y)=f(x,y))=1$.
The communication complexity of a protocol ${\cal P}$ is the
worst case number of (qu)bits exchanged. The communication complexity of $f$ is, with which respect to the communication mode used, the complexity of an
optimal protocol for $f$.
We will use $D(f)$ to denote the {\em deterministic communication complexity} and $Q_E(f)$ to denote the {\em exact quantum communication complexity}.
\subsection{One-way finite automata with quantum and classical states}
In this subsection we recall the definition of 1QCFA.
{\em Two-way finite automata with quantum and classical states} were introduced by Ambainis and Watrous \cite{Amb02} and then explored in \cite{Yak10,ZhgQiu11,ZhgQiu112,Zhg12,Zhg13,Zhg13a}.
1QCFA are one-way versions of 2QCFA, which were introduced by Zheng et al. \cite{ZhgQiu112}.
Informally, a 1QCFA can be seen as a 1DFA which has access to a quantum
memory of a constant size (dimension), upon which it performs
quantum transformations and measurements. Given a finite set of quantum basis states $Q$, we denote by $\mathcal{H}(Q)$
the Hilbert space spanned by $Q$. Let
$\mathcal{U}(\mathcal{H}(Q))$ and $\mathcal{O}(\mathcal{H}(Q))$
denote the sets of unitary operators and projective measurements over $\mathcal{H}(Q)$, respectively.
\begin{definition}
A {\it one-way finite automaton with quantum and classical states} $\mathcal{A}$ is specified by a 10-tuple
\begin{equation}
\mathcal{A}=(Q,S,\Sigma,\Theta,\Delta,\delta,|q_{0}\rangle,s_{0},S_{acc},S_{rej})
\end{equation}
where:
\begin{enumerate}
\item $Q$ is a finite set of orthonormal quantum basis states.
\item $S$ is a finite set of classical states.
\item $\Sigma$ is a finite alphabet of input symbols and let
$\Sigma'=\Sigma\cup \{\cent ,\$\}$, where $\cent$ will be used as the left end-marker and $\$$ as the right end-marker.
\item $|q_0\rangle\in Q$ is the initial quantum state.
\item $s_0$ is the initial classical state.
\item $S_{acc}\subset S$ and $S_{rej}\subset S$, where $S_{acc}\cap S_{rej}=\emptyset$, are sets of
the classical accepting and rejecting states, respectively.
\item $\Theta$ is a quantum transition function
\begin{equation}
\Theta: S\setminus(S_{acc}\cup S_{rej})\times \Sigma'\to \mathcal{U}(\mathcal{H}(Q)),
\end{equation}
assigning to each pair $(s,\gamma)$ a
unitary transformation.
\item $\Delta$ is a mapping
\begin{equation}
\Delta:S\times \Sigma'\rightarrow
\mathcal{O}(\mathcal{H}(Q)),
\end{equation}
where each $\Delta(s,\gamma)$ corresponds to a projective measurement
{\em (}a projective measurement will be taken each time a unitary
transformation is applied; if we do not need a measurement,
we denote that $\Delta(s,\gamma)=I$, and we assume the result of the measurement to be a fixed $c${\em )}.
\item $\delta$ is a special transition function of classical states.
Let the results set of the measurement be $\mathcal{C}=\{c_{1},c_{2},\dots$,
$c_{s}\}$, then
\begin{equation}
\delta:S\times \Sigma' \times \mathcal{C}\rightarrow
S,
\end{equation}
where $\delta(s,\gamma)(c_{i})=s'$ means
that if a tape symbol $\gamma\in \Sigma'$ is being
scanned and the projective measurement result is $c_{i}$, then the state $s$ is changed to $s'$.
\end{enumerate}
\end{definition}
Given an input $w=\sigma_1\cdots\sigma_l$, the word on the tape will be
$w=\cent w\$$ (for convenience, we denote $\sigma_0=\cent $ and $\sigma_{l+1}=\$$).
Now, we define the behavior of 1QCFA $\mathcal{A}$ on the input word $w$.
The computation starts in the classical state $s_0$ and the quantum state $|q_0\rangle$, then the
transformations associated with symbols in the word $\sigma_0\sigma_1\cdots, \sigma_{l+1}$ are applied in succession.
The transformation associated
with a state $s\in S$ and a symbol $\sigma\in\Sigma'$ consists of three steps:
\begin{enumerate}
\item Firstly, $\Theta(s,\sigma)$ is applied to the current quantum state
$|\phi\rangle$, yielding the new state
$|\phi'\rangle=\Theta(s,\sigma)|\phi\rangle$.
\item Secondly, the observable $\Delta(s,\sigma)=\mathcal{O}$ is measured on
$|\phi'\rangle$. The set of possible results is $\mathcal{C}=\{c_1, \cdots, c_s\}$.
According to quantum mechanics principles, such a
measurement yields the classical outcome $c_k$ with probability
$p_k=||P(c_k)|\phi'\rangle||^2$, and the quantum state of $\mathcal{A}$
collapses to $P(c_k)|\phi'\rangle/\sqrt{p_k}$.
\item Thirdly, the current classical state $s$ will be changed to $\delta(s,\sigma)(c_k) =s'.$
\end{enumerate}
An input word $w$ is assumed to be accepted (rejected) if and only if the classical state after scanning $\sigma_{l+1}$
is an accepting (rejecting) state. We assume that $\delta$ is well defined so that 1QCFA $\mathcal{A}$ always accepts or
rejects at the end of the computation.
Language acceptance is a special case of so called promise problem solving.
A {\em promise problem} is a pair $A = (A_{yes}, A_{no})$, where $A_{yes}$, $A_{no}\subset \Sigma^*$
are disjoint sets. Languages may be viewed as promise problems that obey the additional constraint
$A_{yes}\cup A_{no}=\Sigma^*$.
A promise problem $A = (A_{yes}, A_{no})$ is solved exactly by a finite automaton ${\cal A}$ if
\begin{enumerate}
\item[1.] $\forall w\in A_{yes}$, $Pr[{\cal A}\ \mbox{accepts}\ w]=1$, and
\item[2.] $\forall w\in A_{no}$, $Pr[{\cal A}\ \mbox{rejects}\ w]=1$.
\end{enumerate}
\section{An exact quantum query algorithm for $\mbox{DJ}'(x)$}
Montanaro et al. \cite{MJM11} gave a quantum algorithm for $\mbox{DJ}'$ with 2 queries. However, their proof is complicated.
Motivated by the method from \cite{AISJ13},
we give a simpler algorithm with 2 queries for $\mbox{DJ}'$ as follow:
We use basis states $|0,0\rangle$, $|i,0\rangle$, $|i,j\rangle$ and $|k\rangle$ with $0\leq i<j\leq n$ and $1\leq k\leq n-2$.
A basis state $|i,j\rangle$ corresponds to an input bit $x_i$ for $1\leq i\leq n$; a basis state $|k\rangle$ corresponds to an input bit $y_k$ for $1\leq k\leq n-2$ ($y_k$ is some certain bit $x_i$) and the other basis states do not correspond to any input bit.
\begin{enumerate}
\item
The algorithm ${\cal A}$ begins in the state $|0,0\rangle$ and then a unitary mapping $U_1$ is applied on it:
\begin{equation}
U_1|0,0\rangle=\sum_{i=1}^n\frac{1}{\sqrt{n}}|i,0\rangle.
\end{equation}
\item
${\cal A}$ then performs the query:
\begin{equation}
\sum_{i=1}^n\frac{1}{\sqrt{n}}|i,0\rangle\to \sum_{i=1}^n\frac{1}{\sqrt{n}}(-1)^{x_i}|i,0\rangle.
\end{equation}
\item ${\cal A}$ performs a unitary mapping $U_2$ to the current state such that
\begin{equation}
U_2|i,0\rangle=\sum_{j>i\geq 1}\frac{1}{\sqrt{n}}|i,j\rangle-\sum_{1\leq j<i}\frac{1}{\sqrt{n}}|j,i\rangle+\frac{1}{\sqrt{n}}|0,0\rangle
\end{equation}
and the resulting quantum state will be
\begin{equation}
U_2\sum_{i=1}^n\frac{1}{\sqrt{n}}(-1)^{x_i}|i,0\rangle=\frac{1}{n}\sum_{i=1}^n(-1)^{x_i}|0,0\rangle+\frac{1}{n}\sum_{1\leq i<j}((-1)^{x_i}-(-1)^{x_j})|i,j\rangle.
\end{equation}
\item ${\cal A}$ measures the resulting state in the standard basis. If the outcome is $|0,0\rangle$, then $\sum_{i=1}^n(-1)^{x_i}\neq 0 $ and $\mbox{DJ}'(x)=1$. Otherwise, suppose that we get the state $|i,j\rangle$, then we have $x_i\neq x_j$. Let $y=x\setminus \{x_i,x_j\}$, we have $W(y)\in\{0,n-2,\frac{n-2}{2}\}$. If $W(y)=\frac{n-2}{2}$, then $\mbox{DJ}'(x)=0$. If $W(y)\in \{0,n-2\} $, then $\mbox{DJ}'(x)=1$.
The remaining question is exactly the Deutsch-Jozsa promise problem \cite{DJ92} and we can get the answer with 1 query as follows:
we use the subalgorithm ${\cal B}$ to solve the remaining promise problem using $n-2$ quantum basis states $|1\rangle,\ldots,|n-2\rangle$ that will work as follows:
\begin{enumerate}
\item ${\cal B}$ begins in the state $|1\rangle$ and performs on it a unitary transformation $U_3$ such that
\begin{equation}
U_3|1\rangle=\sum_{k=1}^{n-2} \frac{1}{\sqrt{n-2}}|k\rangle.
\end{equation}
\item ${\cal B}$ performs a query $Q$:
\begin{equation}
\sum_{k=1}^{n-2} \frac{1}{\sqrt{n-2}}|k\rangle\to \sum_{k=1}^{n-2} \frac{1}{\sqrt{n-2}}(-1)^{y_k}|k\rangle
\end{equation}
\item ${\cal B}$ performs a unitary transformation $U_4=U_3^{-1}$ and
\begin{equation}
U_3^{-1} \sum_{k=1}^{n-2} \frac{1}{\sqrt{n-2}}(-1)^{y_k}|k\rangle=\frac{1}{n-2}\sum_{k=1}^{n-2}(-1)^{y_k}|1\rangle+\sum_{k=2}^{n-2}\beta_k|k\rangle,
\end{equation}
where $\beta_k$ are amplitudes that we do not need to be specified exactly.
\item ${\cal B}$ measures the resulting state in the standard basis and outputs 1 if the measurement outcome is $|1\rangle$ and 0 otherwise.
\end{enumerate}
According to \cite{AISJ13}, the unitary mapping $U_2$ exists. The rest of the proof is easy to verify. Obviously, the algorithm ${\cal A}$ uses 2 queries.
\end{enumerate}
\section{Communication complexity of $\mbox{EQ}'(x,y)$}
In this section, we will prove that $Q_E(\mbox{EQ}')$ is ${\bf O}(\log n)$ while $D(\mbox{EQ}')$ is ${\bf \Omega}(n)$.
\begin{theorem}\label{Th-1}
$Q_E(\mbox{EQ}')\in {\bf O}(\log n)$.
\end{theorem}
\begin{proof}
Assume that Alice is given an input $x=x_1,\cdots,x_n$ and Bob an input $y=y_1,\cdots,y_n$. The following quantum communication protocol ${\cal P}$ computes $\mbox{EQ}'$ using ${\bf O}(n^2)$ quantum basis states $|0,0\rangle$, $|i,0\rangle$, $|i,j\rangle$ and $|k\rangle$ with $0\leq i<j\leq n$ and $1\leq k\leq n-2$
as follows:
\begin{enumerate}
\item Alice begins with the quantum state $|\psi_0\rangle=|0,0\rangle$ and performs on it the unitary map $U_1$. The quantum state is changed to
\begin{equation}
|\psi_1\rangle=U_1|0,0\rangle=\frac{1}{\sqrt{n}}\sum_{i=1}^n|i,0\rangle.
\end{equation}
\item Alice applies the unitary map $U_x$ to the current state such that $U_x|i,0\rangle=(-1)^{x_i}|i,0\rangle$ for $i>0$ and the quantum state is changed to
\begin{equation}
|\psi_2\rangle=U_x|\psi_1\rangle=\frac{1}{\sqrt{n}}\sum_{i=1}^n(-1)^{x_i}|i,0\rangle.
\end{equation}
\item Alice then sends her current quantum state $|\psi_2\rangle$ to Bob.
\item Bob applies the unitary map $U_y$ to the state that he has received such that $U_y|i,0\rangle=(-1)^{y_i}|i,0\rangle$ for $i>0$ and the quantum state is changed to
\begin{equation}
|\psi_3\rangle=U_y|\psi_2\rangle U_x=\frac{1}{\sqrt{n}}\sum_{i=1}^n(-1)^{x_i+y_i}|i,0\rangle.
\end{equation}
\item Bob applies the unitary map $U_2$ to his quantum state and the quantum state is changed to
\begin{equation}
|\psi_4\rangle=U_2|\psi_3\rangle=\frac{1}{n}\sum_{i=1}^n(-1)^{x_i+y_i}|0,0\rangle+\frac{1}{n}\sum_{1\leq i<j}((-1)^{x_i+y_i}-(-1)^{x_j+y_j})|i,j\rangle.
\end{equation}
\item Bob measures the resulting state in the standard basis and outputs 1 if the measurement outcome is $|0,0\rangle$. Otherwise, suppose that the outcome is $|i,j\rangle$. Bob sends $i$ and $j$ to Alice using classical bits.
\item After Alice receives $i$ and $j$, let $x'=x_1\ldots x_{i-1}x_{i+1}\ldots x_{j-1}x_{j+1}\ldots x_n$. (In convenience, we write $x'=x'_1\ldots x'_{n-2}$).
Alice applies $U_{3}$ to the basis state $|1\rangle$ such that the quantum state is changed to
\begin{equation}
|\psi_5\rangle=U_3|1\rangle=\frac{1}{\sqrt{n-2}}\sum_{k=1}^{n-2} |k\rangle.
\end{equation}
\item Alice then applies $U_{x'}$ to the current state such that $U_{x'}|k\rangle=(-1)^{x'_k}|k\rangle$ for $k>0$ and the quantum state is changed to
\begin{equation}
|\psi_6\rangle= \frac{1}{\sqrt{n-2}}\sum_{k=1}^{n-2}(-1)^{x_k'}|k\rangle.
\end{equation}
\item Alice sends her current quantum state $|\psi_6\rangle$ to Bob.
\item Bob applies the unitary map $U_{y'}$ to the state that he has received such that $U_{y'}|k\rangle=(-1)^{y'_k}|k\rangle$ for $k>0$, where $y'=y_1\ldots y_{i-1}y_{i+1}\ldots y_{j-1}y_{j+1}\ldots y_n$. (In convenience, we write $y'=y'_1\ldots y'_{n-2}$). The quantum state is changed to
\begin{equation}
|\psi_7\rangle=\frac{1}{\sqrt{n-2}}\sum_{k=1}^{n-2} (-1)^{x_k'+y_k'}|k\rangle.
\end{equation}
\item Bob performs a unitary transformation $U_4=U_3^{-1}$ to the current state and the quantum state is changed to
\begin{equation}
|\psi_8\rangle=U_4 |\psi_7\rangle=\frac{1}{n-2}\sum_{k=1}^{n-2} (-1)^{x_k'+y_k'}|1\rangle+\sum_{k=2}^{n-2}\beta_k|k\rangle,
\end{equation}
where $\beta_k$ are amplitudes that we do not need to be specified exactly.
\item Bob measures the resulting state in standard basis and outputs 1 if the measurement outcome is $|1\rangle$ and outputs 0 otherwise.
\end{enumerate}
The unitary transformations $U_1$, $U_2$, $U_3$ and $U_4$ are the same ones as defined in Section~3.
If $H(x,y)\in\{0,n\}$, then the quantum state in Step~5 is
\begin{equation}
|\psi_4\rangle=\frac{1}{n}\sum_{i=1}^n(-1)^{x_i+y_i}|0,0\rangle=\pm|0,0\rangle.
\end{equation}
Bob will get the quantum state $|0,0\rangle$ after the measurement in Step~6 and output 1 as the result of $\mbox{EQ}'(x,y)$.
If $H(x,y)\in\{1,n-1\}$, then there are two cases:
\begin{description}
\item[a)] If the measurement outcome in Step~6 is $|0,0\rangle$ and Bob outputs 1 as the result of $\mbox{EQ}'(x,y)$.
\item[b)] If the measurement outcome in Step~6 is $|i,j\rangle$, then $H(x', y')\in\{0,n-2\}$ and the quantum state in Step~11 is
\begin{equation}
|\psi_8\rangle=\frac{1}{n-2}\sum_{k=1}^{n-2} (-1)^{x_k'+y_k'}|1\rangle=\pm |1\rangle
\end{equation}
Bob will get the quantum state $|1\rangle$ after the measurement in Step~12 and output 1 as the result of $\mbox{EQ}'(x,y)$.
\end{description}
If $H(x,y)=\frac{n}{2}$, then Bob will output 0 as the result of $\mbox{EQ}'(x,y)$ in Step~12.
In Step~3 Alice sends $\lceil\log (n^2)\rceil$ qubits, in Step~6 Bob sends $2\lceil\log (n)\rceil$ bits and in Step~9 Alice sends $\lceil\log (n-2)\rceil$ qubits. Since we can use qubits to send bits, it is clear that this protocol uses only ${\bf O}(\log n)$ qubits for communication.
\end{proof}
The proof for deterministic communication lower bound is similar to the ones in \cite{BdW02,Buh09}. In order to obtain an exponential quantum speed-up in \cite{Buh09}, $\frac{n}{2}$ must be an even integer in the distributed Deutsch-Jozsa promise problem (see \cite{GQZ14} for argument). However, $\frac{n}{2}$ can be arbitrary integer in the promise problem $\mbox{EQ}'$ in this paper.
We use so called ``rectangles" lower bound method \cite{KusNis97b} to prove the result.
A {\em rectangle} in $X\times Y$ is a subset $R\subseteq X\times Y$ such that $R=A\times B$ for some $A\subseteq X$ and $ B\subseteq Y$. A rectangle $R=A\times B$ is called
$1(0)$-rectangle of a function $f:X\times Y\to \{0,1\}$
if for every $x\in A$ and $y\in B$ the value of $f(x,y)$ is 1 (0). Moreover, $C^i(f)$ is defined as the minimum number of $i$-rectangles that partition the space of $i$-inputs (such inputs $x$ and $y$ that $f(x,y)=i$) of $f$.
\begin{lemma}\label{lm-d-lowbound} {\em \cite{KusNis97b}}
$D(f)\geq \max\{\log{ C^1(f)},\log{ C^0(f)}\}$.
\end{lemma}
\begin{remark}
For a partial
function $f: X\times Y\to \{0,1\}$ with domain $\mathcal{ D}$, a rectangle $R=A\times B$ is called
$1(0)$-rectangle if the value of $f(x,y)$ is 1(0) for every $(x,y)\in \mathcal{ D}\cap (A\times B)$ -- we do not care about values for $(x,y)\not\in \mathcal{ D}$.
The above lemma still holds for promise problems (that is for partial functions).
\end{remark}
\begin{theorem}\label{Th-lowerbound}
{\em $D(\mbox{EQ}')\in {\bf \Omega}( n)$}.
\end{theorem}
\begin{proof}
Let ${\cal P}$ be a deterministic protocol for $\mbox{EQ}'$. There are two cases:
{\bf Case 1:} $\frac{n}{2}$ is even. We consider the set $E=\{(x,x),(x,\overline{x}) \,|\, W(x)=\frac{n}{2}\}$. For every $(x,y)\in E$, we have ${\cal P}(x,y)=1$. Suppose there is a 1-monochromatic rectangle $R=A\times B\subseteq \{0,1\}^n\times\{0,1\}^n$ such that ${\cal P}(x,y)=1$ for every promise pair $(x,y)\in R$. Let $S=R\cap E$. For $x,y\in\{0,1\}^n$, let us denote $|x\wedge y|=\sum_{i=1}^n x_i\wedge y_i$.
We now prove that for any distinct $(x,x'),(y,y')\in S$, $|x\wedge y|\neq \frac{n}{4}$.
According to the assumption that $(y,y')\in S\subset E$, we have $y'=y$ or $y'=\overline{y}$. If $|x\wedge y|=\frac{n}{4}$, then $H(x,y)=2(\frac{n}{2}-\frac{n}{4})=\frac{n}{2}=H(x,\overline{y})$ and ${\cal P}(x,y')=0$. Since $(x,x')\in R$ and $(y,y')\in R$, we have $(x,y')\in R$ and ${\cal P}(x,y')=0$, which is a contradiction.
According to Corollary 1.2 from \cite{Fr87}, we have $|S|\leq 1.99^n$. Therefore, the minimum number of 1-monochromatic rectangles that partition the space of inputs is
\begin{equation}
C^1(\mbox{EQ}')\geq\frac{|E|}{|S|}\geq \frac{2{n\choose n/2}}{(1.99)^n}>\frac{2^{n+1}/n}{(1.99)^n}.
\end{equation}
The deterministic communication complexity is then
\begin{equation}
D(\mbox{EQ}')\geq \log{C^1(\mbox{EQ}')}>\log{\frac{2^{n+1}/n}{(1.99)^n}}>0.0073n.
\end{equation}
{\bf Case 2:} $\frac{n}{2}$ is odd. We assume that $n=4k+2$. We consider the set $E=\{(x,x') \,|\, W(x)=\frac{n}{2}\}$, where $x_n'=x_n=1$ and $x_i'=1-x_i$ for $i<n$. For every $(x,x')\in E$, we have $H(x,x')=n-1$ and ${\cal P}(x,x')=1$. Suppose there is a 1-monochromatic rectangle $R=A\times B\subseteq \{0,1\}^n\times\{0,1\}^n$ such that ${\cal P}(x,y)=1$ for every promise pair $(x,y)\in R$. Let $S=R\cap E$. We now prove that for any distinct $(x,x'),(y,y')\in S$, $\sum_{i=1}^{n-1}|x_i\wedge y_i|\neq k$, that is $|x\wedge y|\neq k+1$.
If $|x\wedge y|= k+1$, without a lost of generality, let $
x=\overbrace{1\cdots 1}^{k}\overbrace{ 1\cdots 1}^{k}\overbrace{ 0\cdots 0}^{k}\overbrace{ 0\cdots 0}^{k+1}1,$ and $y=\overbrace{1\cdots 1}^{k}\overbrace{ 0\cdots 0}^{k}\overbrace{ 1\cdots 1}^{k}\overbrace{ 0\cdots 0}^{k+1}1.
$
We have $H(x,y')=k+k+1=\frac{n}{2}$ and ${\cal P}(x,y')=0$. Since $(x,x')\in R$ and $(y,y')\in R$, we have $(x,y')\in R$ and ${\cal P}(x,y')=0$, which is a contradiction.
According to Corollary 1.2 from \cite{Fr87}, we have $|S|\leq 1.99^n$. Therefore, the minimum number of 1-monochromatic rectangles that partition the space of inputs is
\begin{equation}
C^1(\mbox{EQ}')\geq\frac{|E|}{|S|}\geq \frac{{n-1\choose n/2-1}}{(1.99)^{n-1}}>\frac{2^{n-1}/(n-1)}{(1.99)^{n-1}}.
\end{equation}
The deterministic communication complexity is then
\begin{equation}
D(\mbox{EQ}')\geq \log{C^1(\mbox{EQ}')}>\log{\frac{2^{n-1}/(n-1)}{(1.99)^{n-1}}}>0.0073n.
\end{equation}
Therefore the theorem has been proved.
\end{proof}
\section{State succinctness results}
Now we are ready to derive the state succinctness result.
\begin{theorem}
The promise problem $A(n)$ can be solved exactly by a 1QCFA ${\cal A}(n)$ with ${\bf O}(n^2)$ quantum basis states and ${\bf O}(n^3)$ classical states, whereas the sizes of the corresponding 1DFA are $2^{{\bf \Omega}(n)}$.
\end{theorem}
\begin{proof}
Let $x=x_1\cdots x_n$ and $y=y_1\cdots y_n$ be in $\{0,1\}^n$. The input word on the tape will be $w=\cent x\#y\#\#x\#y\$$. Let us consider a 1QCFA ${\cal A}(n)$ with ${\bf O}(n^2)$ quantum basis states
$\{|0,0\rangle,|i,0\rangle,|i,j\rangle,|k\rangle:0\leq i<j\leq n,\ 1\leq k\leq n-2 \}$ and ${\bf O}(n^3)$ classical states $\{S_{ijp}: 0\leq i,j, p\leq n+1\}$ (some of the states may be not used in the automaton actions).
${\cal A}(n)$ starts in the initial quantum state $|0,0\rangle$ and the initial classical state $S_{000}$. We use classical states $S_{ijp}$ ($1\leq p\leq n+1$) to point out the positions of the tape head that will provide some information for quantum transformations. If the classical state of ${\cal A}(n)$ will be $S_{ijp}$ ($1\leq p\leq n$) that will mean that the next scanned symbol of the tape head is the $p$-th symbol of $x$($y$) and $S_{ijn+1}$ means that the next scanned symbol of the tape head is $\#$($\$$).
The behavior of ${\cal A}(n)$ is composed of two parts. The first part is the behavior of ${\cal A}(n)$ when reading the prefix of the input, namely $\cent x\#y\#$. In this part, ${\cal A}(n)$ uses quantum basis state $\{|0,0\rangle,|i,0\rangle,|i,j\rangle:0\leq i<j\leq n \}$ and classical states $S_{00p}$ ($0\leq p\leq n+1$) to simulate Steps 1 to 6 in the proof of Theorem \ref{Th-1}.
After the measurement at the end of the first part, if the outcome is $|0,0\rangle$, then the input is accepted. Otherwise, suppose that the outcome is $|i,j\rangle$, the classical state will be changed to $S_{ij0}$ ($1\leq i<j\leq n$, which means that $H(x_ix_j,y_iy_j)=1$ and the input bits $x_i,x_j,y_i,y_j$ will be skipped during the second part of the behavior of ${\cal A}(n)$.
The second part is the behavior of the automation when reading the second part of the input $\#x\#y\$$. In this part, ${\cal A}(n)$ uses quantum basis states
$\{|k\rangle:\ 1\leq k\leq n-2 \}$ and classical states $S_{ijp}$ ($0\leq p\leq n+1$) to simulate Steps 7 to 12 in the proof of Theorem \ref{Th-1}.
The automaton proceeds as follows:
\begin{enumerate}
\item ${\cal A}(n)$ reads the left end-marker $\cent $, performs $U_1$ on the initial quantum state $|0,0\rangle$, changes its classical state to $\delta(S_{000},\ \cent )=S_{001}$, and moves the tape head one cell to the right.
\item Until the currently scanned symbol $\sigma$ is not $\#$, ${\cal A}(n)$ does the following:
\begin{enumerate}
\item Applies $\Theta(S_{00p},\sigma)=U_{p,\sigma}$ to the current quantum state.
\item Changes the classical state $S_{00p}$ to $S_{00p+1}$ and moves the tape head one cell to the right.
\end{enumerate}
\item ${\cal A}(n)$ changes the classical state $S_{00p+1}$ to $S_{001}$ and moves the tape head one cell to the right.
\item Until the currently scanned symbol $\sigma$ is not $\#$, ${\cal A}(n)$ does the following:
\begin{enumerate}
\item Applies $\Theta(S_{00p},\sigma)=U_{p,\sigma}$ to the current quantum state.
\item Changes the classical state $S_{00p}$ to $S_{00p+1}$ and moves the tape head one cell to the right.
\end{enumerate}
\item When $\#$ is reached, ${\cal A}(n)$ performs $U_{2}$ on the current quantum state.
\item
${\cal A}(n)$ measures the current quantum state in the standard basis. If the outcome is $|0,0\rangle$, ${\cal A}(n)$ accepts the input; otherwise, suppose that the outcome is $|i,j\rangle$, ${\cal A}(n)$ changes the classical state to $S_{ij0}$, moves the tape head one cell to the right.
\item ${\cal A}(n)$ reads $\#$, applies $\Theta(S_{ij0},\#)=U_3U_{ij}$ to the current quantum state, changes its classical state to $S_{ij1}$, and moves the tape head one cell to the right.
\item Until the currently scanned symbol $\sigma$ is not $\#$, ${\cal A}(n)$ does the following:
\begin{enumerate}
\item Applies $\Theta(S_{ijp},\sigma)=U_{ijp,\sigma}$ to the current quantum state.
\item Changes the classical state $S_{ijp}$ to $S_{ijp+1}$ and moves the tape head one cell to the right.
\end{enumerate}
\item ${\cal A}(n)$ changes the classical state $S_{ijp+1}$ to $S_{ij1}$ and moves the tape head one cell to the right.
\item While the currently scanned symbol $\sigma$ is not the right end-marker $\$$, ${\cal A}(n)$ does the following:
\begin{enumerate}
\item Applies $\Theta(S_{ijp},\sigma)=U_{ijp,\sigma}$ to the current quantum state.
\item Changes the classical state $S_{ijp}$ to $S_{ijp+1}$ and moves the tape head one cell to the right.
\end{enumerate}
\item When the right end-marker is reached, ${\cal A}(n)$ performs $U_{4}$ on the current quantum state.
\item
${\cal A}(n)$ measures the current quantum state in the standard basis. If the outcome is $|1\rangle$, ${\cal A}(n)$ accepts the input; otherwise, rejects the input.
\end{enumerate}
where unitary transformations $U_1$, $U_2$, $U_3$, $U_4$ are the ones defined in the proof of Theorem \ref{Th-1} and
\begin{eqnarray*}
U_{p,\sigma}|i,j\rangle&=&(-1)^{\sigma}|i,j\rangle \ \mbox{if}\ i=p;\\
U_{p,\sigma}|i,j\rangle&=& |i,j\rangle \ \mbox{if}\ i\neq p;\\
U_{ij}|i,j\rangle&=& |1\rangle;\\
U_{ijp,\sigma}|k\rangle&=&(-1)^{\sigma}|k\rangle \ \mbox{if}\ p<i\ \mbox{and}\ k=p; \\
U_{ijp,\sigma}|k\rangle&=&|k\rangle \ \mbox{if}\ p<i\ \mbox{and}\ k\neq p; \\
U_{ijp,\sigma}|k\rangle&=&|k\rangle \ \mbox{if}\ p=i; \\
U_{ijp,\sigma}|k\rangle&=&(-1)^{\sigma}|k\rangle \ \mbox{if}\ i<p<j\ \mbox{and}\ k=p-1; \\
U_{ijp,\sigma}|k\rangle&=&|k\rangle \ \mbox{if}\ i<p<j\ \mbox{and}\ k\neq p-1; \\
U_{ijp,\sigma}|k\rangle&=&|k\rangle \ \mbox{if}\ p=j; \\
U_{ijp,\sigma}|k\rangle&=&(-1)^{\sigma}|k\rangle \ \mbox{if}\ p>j\ \mbox{and}\ k= p-2; \\
U_{ijp,\sigma}|k\rangle&=&|k\rangle \ \mbox{if}\ p>j\ \mbox{and}\ k\neq p-2.
\end{eqnarray*}
It is easy to verify that unitary transformation $U_{p,\sigma}$, $U_{ij}$ and $U_{ijp,\sigma}$ exist.
The rest of the proof is analogues to the proof in Theorem \ref{Th-1}.
According to Theorem \ref{Th-lowerbound} , $D(\mbox{EQ}')\in {\bf \Omega}( n)$. Therefore, it is easy to see that the sizes of the corresponding 1DFA for $A(n)$ are $2^{{\bf \Omega}(n)}$ \cite{KusNis97b}.
\end{proof}
\section*{Acknowledgements}
The first author would like to thank Jozef Gruska for having a most constructive influence on him, teaching him how to think and write in a clear way, and for continuous support during his postdoctoral research with him.
|
\section{\label{I} Introduction}
Soon after the discovery of the quantum statistics by Fermi \cite{Fermi1926} and Dirac \cite{Dirac1926}, which incorporates Pauli's exclusion principle \cite{Pauli1925}, the ideal Fermi gas (IFG) has been extensively used to describe, approximately, many physical phenomena in a wide range of values of the particles density, from cosmological scales to nuclear ones.
At the end of the 20th century, the experimental realization of quantum degeneration of a trapped Fermi gas of $^{40}$K atoms \cite{DemarcoSc} raised the interest on the theoretical study of the thermodynamical and dynamical properties of the Fermi gas in the ideal approximation \cite{Vignolo2000,Gleisber2000,Tran2001,Akdeniz2002,Vignolo2003,Anghel2003,Tran2003,vanZyl2003,Mueller2004,Anghel2005,Song2006}.
More recently, the IFG has been used in the context of quantum information and the entanglement entropy of it has been obtained in Ref. \cite{GioevPRL2006}, while in Ref. \cite{CalabreseEPL2012} exact relations between the Renyi entanglement entropies and the particle number fluctuations in a system of noninteracting fermions have been derived.
On the other hand, the thermodynamics of Fermi gases at extreme conditions of density and/or temperature are of great interest to understand processes in white dwarf stars \cite{StockNature1989}, or the properties of the gluon-quark plasma \cite{SatzNature1986}, which is thought to occurred, some microseconds after the Bing-Bang at the early stage of the Universe.
In the noninteracting regime, analysis of the the relativistic effects on the thermodynamics of the IFG are limited to the consideration of the energy spectrum of a relativistic single-particle \cite{HorePRA75,Dunning-DaviesJPhysA1981,HowardJPhysA2004,ElzeJphysG1980}, generally disregarding particle-antiparticle pair production predicted by quantum field theory. This contrasts to the thermodynamics of the relativistic Bose gas which has been thoroughly studied considering pair production \cite{FrotaPRA1989,GretherPRL2007,SuJPA2008,HaberPRL81}. One of the goals of the present paper is to fill out this gap.
In the same noninteracting regime, P.-H. Chavanis \cite{ChavanisPRE2004,ChavanisPRD2007} discusses the effects of the spatial dimensionality in the balance between quantum pressure due to degeneracy and gravitational collapse due to self-gravitation in white dwarf stars. In his analysis, the author shows that the collapse or evaporation of the star is unavoidable in dimensions larger than four, giving a special character to systems of spatial dimension $d\le3$ through an anthropic principle. The properties of the electron gas in the star are rather well approximated by those in the limit of complete degeneration, {\it i.e.} by those at zero temperature, due to the disparate difference between the system temperature and the Fermi temperature, $T/T_{F}\simeq10^{-3}-10^{-2}.$
Moreover, the effects of low spatial-dimensionality on the non-relativistic IFG at finite temperatures are exhibited in the form of an \emph{unusual} temperature dependence of the chemical potential $\mu(T)$ at constant volume \cite[and reference therein]{GretherEPJD2003}. These effects are markedly shown in an IFG trapped in an impenetrable, one dimensional box potential, for which at low temperatures $\mu(T)$ starts rising quadratically with $T$ above the Fermi energy instead of decreasing from it, as does in the three-dimensional case. Eventually, at larger temperatures, $\mu(T)$ turns to its usual monotonic decreasing behavior at a characteristic temperature that can be as large as twice the Fermi temperature. This unusual behavior has been already related with a possible phase transition with the maximum of the chemical potential pointing out a sort of critical point \cite{sevilla}.
In this paper we focus our study on the thermodynamic properties of the IFG, where the effects of quantum degeneracy, relativity and spatial dimensionality, are all combined. Though, particular attention is paid to the temperature dependence of the chemical potential, which has motivated several discussion of its importance on different levels and contexts \cite{cook_ajp95,baierlein,job2006,MuganEJP2009,ShegelskiSSC86,LandsbergSST87,ShegelskiAJP2004,KaplanJStatPhys2006,SevillaEJP2012}, our main results focus on the thermodynamic susceptibilities or response functions, namely the specific heat at constant volume $C_{V}$ and the isothermal compressibility $\kappa_{T},$ for which there is a great interest at conditions of extreme densities and/or temperatures.
Interestingly, our calculations reveal the appearance of a transition in the temperature dependence of $C_{V}$ and $\kappa_{T}$ due to pair production, that occurs at a few tenths of the Fermi temperature. This drastic qualitative change in behavior can be plausibly considered as a phase transition, from a phase at which the compressibility diminishes with temperature, as in standard matter (defining the normal phase), to another at which matter becomes arbitrarily compressible.
The paper is organized as follows. In section \ref{II} we present the system of our study and the chemical potential is calculated from the principle of charge conservation. In section \ref{III} the isothermal compressibility and the heat capacity at constant volume are calculated. Finally, conclusion and final remarks are given in \ref{IV}.
\section{\label{II} Finite temperature: the effects of pair production}
The system under consideration corresponds to a $d$-dimensional gas of non-interacting fermions at finite temperature and chemical potential. We consider the pair production process in thermal equilibrium with a bath of neutral spinless bosons. This corresponds to an over- simplification of the actual physical situation which is under current investigations for interacting systems in Quantum Chromodynamics \cite{Nagata2012} and in the study of collective phenomena \cite{BlaizotPRD2014}.
At zero temperature the system consists of, without loss of generality, $N_{0}$ of spin-$\frac{1}{2}$ fermions (antifermions may be equally chosen instead), of rest mass $m$ in a volume $V_{d}$, and we consider the exact relativistic energy spectrum, given by
\begin{equation}
E_{k}=\sqrt{c^{2}\hbar ^{2}k^{2}+m^{2}c^{4}}, \label{Energia-Rel}
\end{equation}
where $\hbar k$ is the momentum of the particle and $c$ is the speed of light. For simplicity we assume the spin balanced case in which the number of fermions in each projection $s=\uparrow,\downarrow$ of the spin are equal and spin dependent interactions are neglected.
We introduce the ratio of the rest mass to the \emph{Fermi mass}, $\widetilde{m}=m/m_{F},$ as the parameter that tunes the system from the non-relativistic limit, $\widetilde{m}\gg1$, $E_{k}\simeq mc^{2}+\hbar^{2}k^{2}/2m$, to the ultrarelativistic one $\widetilde{m}\ll1,$ $E_{k}\simeq\hbar c k.$ $m_{F}\equiv\hbar k_{F}/c$ and the Fermi wavevector $k_{F}$ is defined through the Fermi energy $E_{F}\equiv\sqrt{c^{2}\hbar ^{2}k_{F}^{2}+m^{2}c^{4}},$ that gives the energy of the higher occupied state at zero temperature which depends on the zero temperature density of particles in the system $n_{0}=N_{0}/V_{d}$.
In $d$ dimensions the Fermi mass has the following explicit dependence on $n_{0}$
\begin{equation}
m_{F}=2\hbar\pi^{1/2}\left[2\,\Gamma(d/2+1)\right]^{1/d}\, \vert n_{0}\vert^{1/d}/c.
\end{equation}
These relations make clear why, for high dense systems, the ultrarelativistic limit corresponds to $\widetilde{m}\ll1.$
In the non-relativistc limit $E_{F}\simeq mc^{2}+E_{F}^{NR}$, with $E_{F}^{NR}=\hbar ^{2}k_{F}^{2}/2m$ is the well known non-relativistic Fermi energy. In the opposite limit $\widetilde{m}\ll 1,$ we have $E_{F}=E_{F}^{UR}+\widetilde{m}^{2}/2+\ldots$ with $E_{F}^{UR}=m_{F}c^{2}$.
According to Quantum Field Theory the relativistic effects of pair production are expected to be important at temperatures of the order of $mc^{2}/k_{B}$ \cite{HaberPRL81,Huang} and the equilibrium state of the mixture of particles-antiparticles is taken into account by the condition $\mu=-\bar{\mu}$ \cite{Huang}, which is straightforwardly obtained by the thermodynamical equilibrium condition on the Helmholtz free energy $F(T,V_{d},N,\overline{N}).$ $N$ and $\overline{N}$ are the number of particle and antiparticles in the system at temperature $T$ and volume $V_{d}$, respectively. Unless otherwise indicated, we denote with an overbar, those quantities related to antiparticles.
The thermodynamic properties are obtained from the grand partition function
\begin{equation}\label{PartitionF}
\Xi(T,V_{d},\mu)\equiv\hbox{Tr}\left\{\exp \left({-\beta \left[\boldsymbol{H}-\mu ( \boldsymbol{N}-\overline{\boldsymbol{N}}) \right] }\right)\right\},
\end{equation}
where $\beta=(k_{B}T)^{-1}$ and $\hbox{Tr}$ denotes the trace over all the states $|n_{\boldsymbol{k}_{1},s}n_{\boldsymbol{k}_{2},s}\ldots\rangle \otimes |\overline{n}_{\boldsymbol{k}_{1},s}\overline{n}_{\boldsymbol{k}_{2},s}\ldots\rangle$ in Fock space, where $\boldsymbol{k}_{i}$ denotes the $d$-dimensional wavevector and $s$ the two projections of spin. $\boldsymbol{H}=\sum_{\boldsymbol{k},s}E_{k}\left(\boldsymbol{n}_{\boldsymbol{k},s}+\overline{\boldsymbol{n}}_{\boldsymbol{k},s}\right),$ $\boldsymbol{N}=\sum_{\boldsymbol{k},s}\boldsymbol{n}_{\boldsymbol{k},s}$ and $\overline{\boldsymbol{N}}=\sum_{\boldsymbol{k},s}\overline{\boldsymbol{n}}_{\boldsymbol{k},s}$ denote the Hamiltonian, the total number of particles and a anti-particles operators respectively, in terms of the number
operators $\boldsymbol{n}_{\boldsymbol{k},s}=\boldsymbol{a}^{\dagger}_{\boldsymbol{k},s}\boldsymbol{a}_{\boldsymbol{k},s}$, $\overline{\boldsymbol{n}}_{\boldsymbol{k},s}=\overline{\boldsymbol{a}}^{\dagger}_{\boldsymbol{k},s}\overline{\boldsymbol{a}}_{\boldsymbol{k},s}$, with eigenvalues $n_{k,s}$, $\overline{n}_{k,s}$, where $\boldsymbol{a}^{\dagger}_{\boldsymbol{k},s}$ ($\overline{\boldsymbol{a}}^{\dagger}_{\boldsymbol{k},s}$) and $\boldsymbol{a}_{\boldsymbol{k},s}$ ($\overline{\boldsymbol{a}}_{\boldsymbol{k},s}$) are the creation and annihilation operators of particles (antiparticles) respectively that satisfy the relations of anti-commutation $\left\{\boldsymbol{a}_{\boldsymbol{k}^{\prime},s^{\prime}},\boldsymbol{a}^{\dagger}_{\boldsymbol{k},s}\right\}=\delta_{\boldsymbol{k},\boldsymbol{k}^{\prime}}\delta_{s,s^{\prime}}$, $\left\{\boldsymbol{a}_{\boldsymbol{k}^{\prime},s^{\prime}}^{\dagger},\boldsymbol{a}^{\dagger}_{\boldsymbol{k},s}\right\}=\left\{\boldsymbol{a}_{\boldsymbol{k}^{\prime},s^{\prime}},\
\boldsymbol{a}_{\boldsymbol{k},s}\right\}=0$. The grand canonical partition function results
\begin{equation}
\Xi(T,V_{d},\mu)= \prod_{\boldsymbol{k},s}\left(1+ze^{-\beta E_{k}}\right)\left(1+\overline{z}e^{-\beta E_{k}}\right)
\end{equation}
with $z=e^{\beta\mu},$ $\overline{z}=z^{-1},$ the fugacity of particles and antiparticle respectively. From this, we have that
\begin{equation}
\ln\Xi(T,V_{d},\mu)=\sum_{\boldsymbol{k},s}\left[\ln\left(1+ze^{-\beta E_{k}}\right)+\ln\left(1+z^{-1}e^{-\beta E_{k}}\right)\right].
\end{equation}
The net number of particles in the system at $T$ y $V_{d}$ is given by
\begin{eqnarray}
N-\overline{N}&=& \left[z\frac{\partial\ln\Xi}{\partial z}\right]_{T,V_{d}}\notag \\
&\equiv &\sum_{\boldsymbol{k},s}\left[\langle n_{E_{k}}\rangle-\langle\bar{n}_{E_{k}}\rangle\right],
\end{eqnarray}
where $\langle{n}_{E_{k}}\rangle=\lbrace\exp\left[\beta(E_{k}-\mu)\right]+1\rbrace^{-1}$ and $\langle\bar{n}_{E_k}\rangle=\lbrace\exp \left[ \beta (E_{k}+\mu )\right] +1\rbrace^{-1}$ give, respectively, the average number of fermions and anti-fermions in the energy state $E_{k}.$
This equation relates the chemical potential of the system with the initial density of particles $n_{0}$ with $N_{0}=(N-\overline{N})$, in the limit of the continuum we have
\begin{equation}\label{NoEq}
n_{0}=R_{d}\int_{0}^{\infty}dk\, k^{d-1}\left[\langle n_{E_{k}}\rangle-\langle\bar{n}_{E_{k}}\rangle\right],
\end{equation}
where the $R_{d}\equiv4\pi^{d/2}/[(2\pi)^{d}\Gamma(d/2)]$ is a constant that depends only on $d.$ Expression \eqref{NoEq} can be written in terms of hyperbolic functions as
\begin{equation}
n_{0}=R_{d}\int _0^{\infty }dk\, k^{d-1}\frac{\sinh\beta\mu}{\cosh\beta E_k+\cosh\beta \mu }
\end{equation}
and simplifies to
\begin{equation}\label{NumEqUR}
n_{0}=-\frac{R_{d}\Gamma(d)}{(\beta\hbar c)^{d}}[\Li{d}{-z}-\Li{d}{-z^{-1}}]
\end{equation}
in the ultrarelativistic limit and to
\begin{equation}
n_{0}=\frac{R_{d}\Gamma(d/2)}{2(\beta\hbar^{2}/2m)^{d/2}}\left[-\Li{d/2}{-z^{NR}}\right]
\end{equation}
in the non-relativistic one, with $z^{NR}\equiv e^{\beta\mu^{NR}}$ the non-relativistic fugacity and $\mu^{NR}\equiv \mu-mc^{2}$. In last expressions $-\Li{\sigma}{-z}\equiv[1/\Gamma(\sigma)]\int_{0}^{\infty}dx\, x^{\sigma-1}/[e^{x}z^{-1}+1]$ is the polylogarithm function, which has the series representation $-\sum_{l=1}^{\infty}(-z)^{l}/l^{\sigma}$ for $\vert z\vert<1.$
It is worthwhile pointing out that expression \eqref{NumEqUR} can be written as a polynomial of degree $d$ in $(\beta\mu)$ for odd dimension\cite{ElzeJphysG1980}, namely
\begin{equation}
n_{0}=\frac{R_{d}}{d}\left(\frac{\mu}{\hbar c}\right)^{d}\left[1+\sum_{j=0}^{d-2}\frac{d!}{j!}[1+(-1)^{j+d}](1-2^{j+1-d})\zeta(d-j)(\beta\mu)^{j-d}\right]
\end{equation}
where $\zeta(z)$ denotes the usual Riemman zeta function.
\subsection{The chemical potential}
Before discussing the temperature dependence of $\mu$ in the regime of interest, we comment in passing that at low enough temperatures, when pair production is negligible, application of the commonly used Sommerfeld expansion \cite{Ashcroft} to Eq. (\ref{NoEq}), gives for the chemical potential
\begin{equation}\label{Sommerfeld}
\frac{\mu(T)}{E_F}=1-\frac{\pi^2}{6}\left(\frac{T}{T_F}\right)^{2}\left[1+(d-2)(1+\widetilde{m}^{2})\right],
\end{equation}
which depends explicitly on $\widetilde{m}.$ From this expression, a simple analysis shows that a \emph{non-monotonic} dependence on $T$ is possible whenever the dimensionality of the system is strictly smaller than $2-\left(1+\widetilde{m}^{2}\right)^{-1}$. This result generalizes the one reported in \cite{GretherEPJD2003} by incorporating the finite rest mass effects, and reduces to the inequalities $d<1$ and $d<2$ in the ultrarelativistic and non-relativistic limits respectively. These two extreme values of dimension corresponds to those for which the IFG is thermodynamically equivalent (in that the specific heat has the same temperature dependence) to the ideal Bose gas.
In Fig.\ref{fig:ChemPot}(a) $\mu(T)$ without pair production is shown (dashed lines) for dimensions 1/2, 1 ,2, 3 and 4 for $\widetilde{m}=1$. The non-monotonous behavior expected for $d<2$, is exhibited as a local maximum for dimensions 1/2 (first dashed line from the far right) and 1 (second dashed line from the far right). These maxima survive only for $d<1$ in the limit $\widetilde{m}\rightarrow0.$ For $\widetilde{m}=100$ Fig.\ref{fig:ChemPot}(b), the chemical potential apparently does not reveal a maximum value expected for $d<2,$ this is because we have chosen to scale $\mu$ with the exact $E_{F},$ however by shiftting by $mc^{2}$ and changing the scaling factor to $E_{F}^{NR}$ we recover the non-monotonic behavior of the non-relativistic IFG $\mu^{NR}=\mu-mc^{2}$ (see Fig.$1$ in Ref.\cite{GretherEPJD2003}).
Without pair production and in the high temperature regime $\mu$ is given by
\begin{equation*}
-k_{B}T\, \ln\left[\frac{V_{d}}{N\lambda^{d}}\, 2\left(\frac{2\pi k_{B}T}{mc^{2}}\right)^{(d-1)/2} K_{(d+1)/2}\left(\frac{mc^{2}}{k_{B}T}\right)\right],
\end{equation*}
where $\lambda=h/mc$ is the Compton wavelength and $K_{\nu}(z)$ denotes the Bessel function of the second kind of order $\nu.$ Last expression corresponds to the classical result for which the chemical potential is negative and decreases monotonically with temperature (see dashed lines in Fig. \ref{fig:ChemPot}). In addition, the same expression is also obtained for $N$ spin-less relativistic bosons of mass $m$ in the same limit \cite{FrotaPRA1989}. This trivial relationship between the Bose and Fermi gas is simply established by the loss of quantum degeneracy due to thermal fluctuations.
\paragraph{Effects of pair production.}
By solving Eq. (\ref{NoEq}) at constant volume, we show that the combined effects of pair production and system dimensionality are conspicuous on the temperature dependence of $\mu(V_{d},T)$ as is shown in Fig. \ref{fig:ChemPot} (solid lines with symbols).
\begin{figure}[h]
\centering
\includegraphics[width=0.7\columnwidth]{ChemPot_ExactRel_PartAntiPart_ratio_1}
\includegraphics[width=0.7\columnwidth]{ChemPot_ExactRel_PartAntiPart_ratio_100}
\caption{(Color online) Dimensionless chemical potential as function of the dimensionless temperature $T/T_{F}$, for different dimensions: 1/2 (circles), 1 (squares), 2 (diamonds), 3 (up-triangles) and 4 (down-triangles). Panel (a), (b), corresponds to $\widetilde{m}=\,1,$ 100, respectively. To exhibit the effects of pair production, the curves of $\mu(T,V_{d})$ for the case when only particles are present in the system are also shown (dashed lines). The solid-red triangle in panel (a) gives the value $1/\sqrt{2}$ which corresponds to the asymptotic value given by expression \eqref{muasymptotic} for $d=1$.}
\label{fig:ChemPot}
\end{figure}
In the high temperature regime, the chemical potential has three distinct asymptotic limits: i) it goes to zero if $d>1$; ii) it goes to the constant value $E_{F}\left[1+\widetilde{m}^{2}\right]^{-1/2}$ if $d=1$ and iii) diverge sub-linearly as a power law for $0<d<1.$ These behaviors are accounted for by the expression
\begin{equation}\label{muasymptotic}
\mu(T)\sim E_{F}\left(\frac{T}{T_{F}}\right)^{1-d}\Phi(\widetilde{m}^{2}+1,d),
\end{equation}
which is approximately obtained from Eq. (\ref{NoEq}), with $\Phi(\xi,d)$ a temperature-independent quantity defined through the expression
\begin{equation}\label{Phi}
\left[\Phi(\xi,d)\right]^{-1}=d\int_{0}^{\infty}dx\, x^{d-1}\left[1+\cosh\left(\frac{x^{2}}{\xi}\right)^{1/2}\right]^{-1}.
\end{equation}
In Table \ref{table} explicit functional forms for $\left[\Phi(\xi,d)\right]^{-1}$ are given for $d=4,\,3,\,2$ and 1.
\begin{table}[h]
\caption{Explicit functional forms for $\left[\Phi(\xi,d)\right]^{-1}$ which appears in eq. \eqref{muasymptotic}.}
\begin{tabular}{|c|c|c|c|c|}
\hline
& $d=4$ & $d=3$ & $d=2$ & $d=1$ \\ \hline
$\Phi(\xi,d)^{-1}$ & $36\, \xi^2\, \zeta(3)$ & $\pi^{2}\, \xi^{3/2}$ & $2\, \xi\,\ln 4$ & $\xi^{1/2}$\\ \hline
\end{tabular}
\label{table}
\end{table}
In the ultrarelativistic limit, the explicit dependence on temperature can be obtained for odd dimensions, namely $\mu/E_{F}=1$ for dimension one and
\begin{equation}
\frac{\mu}{E_{F}}=\left[\frac{1}{2}+\sqrt{\left(\frac{\pi}{\sqrt{3}}\frac{T}{T_{F}}\right)^{6}+\frac{1}{4}}\right]^{1/3}+\left[\frac{1}{2}-\sqrt{\left(\frac{\pi}{\sqrt{3}}\frac{T}{T_{F}}\right)^{6}+\frac{1}{4}}\right]^{1/3}
\end{equation}
for the three-dimensional case \cite{ElzeJphysG1980}.
In Fig.\ref{fig:ChemPot} the effects of pair production on $\mu(T)$ are shown (solid-lines with different symbols which denote different values of the system dimensionality), for the mass ratio $\widetilde{m}=1$ (panel (a)) and $\widetilde{m}=100$ (panel (b)). In both cases, the solid-red line with squares which corresponds to $d=1$, marks the division from the two different behaviors i) and iii).
The behavior exhibited for $d<1$ is puzzling. Though, thermodynamics at these dimensions would seem out of place, the limit $d\rightarrow0$ has been analyzed in Ref. \cite{LeePRE1996} for the non-relativistic IFG, giving a physically consistent interpretation on the meaning of the large values of the chemical potential as $d\rightarrow0$ \cite{ChavezPhysicaE2011}. The effects of pair production makes the chemical potential to grow monotonically for all temperature if $d<1$ and $\widetilde{m}\lesssim2$, since the larger the temperature the larger the number of particles in the system, thus $\langle n_{E_{k}}\rangle>1/2$ for all $T,$ instead of diminishing as happens when the chemical potential decreases monotonically. This considerations make clear why the system at high temperature behaves quite differently from the classical gas counterpart for which $\langle n_{E_{k}}\rangle\ll1$. This explanation agrees with the one given in Ref. \cite{LeePRE1996} for the non-relativistic IFG in diminishing dimensions, where it is argued that the high values of $\mu$ are a manifestation of the Pauli exclusion principle. For larger masses, $\widetilde{m}\gtrsim4$, this behavior is changed as can be noticed in Fig.\ref{fig:ChemPot}(b) for $\widetilde{m}=100$, where the chemical potential exhibits a different non-monotonic behavior, it goes from a decreasing behavior to an increasing one.
For $d>1$, the effects of the original number of particles are outweighed by the large rate of pair production, thus tending to the limit $N\approx\overline{N}$. This is so due to the dependence on $T$ of $\mu,$ which goes to zero as $T^{1-d}$, this implies $z\rightarrow\overline{z}$.
The particular dependence of $\mu$ on $T,$ for different dimensions and values of $\widetilde{m},$ leads to different particle-antiparticle pair production rate as is exhibited in Fig.\ref{fig:ratio}, where the ratio of the antiparticles number to the particles number, $\overline{N}/N$, is shown as function of temperature for $\widetilde{m}=1$. In the inset, the effects of disparate masses, namely $\widetilde{m}=0.01,\, 1,\, 100,$ are shown for $d=3.$
\begin{figure}[ht!]
\centering
\includegraphics[width=0.7\columnwidth]{NbarNratio_1a}
\caption[]{(Color online) Ratio of anti-particles number to the particle number as function of $T/T_{F}$ for dimension 1/2 (circles), 1 (squares), 2 (diamonds), 3 (up-triangles) and 4 (down-triangles). Inset correspond to the three-dimensional case for the mass ratio values $\widetilde{m}=0.01$ (dashed-dotted line), $1$ (continuous line) and $100$ (dashed line).}
\label{fig:ratio}
\end{figure}
\section{\label{III} Thermodynamic susceptibilities}
It has been suggested \cite{sevilla}, on the grounds of the energy-entropy argument that the non-monotonic behavior of the chemical potential can be related to a phase transition. Indeed, for temperatures below the temperature at which $\mu$ acquires a maximum value, the change of the system's free energy is dominated by the change in the internal energy, while for temperatures above such temperature, the free energy change is dominated by entropic differences. Possible phase transitions are pointed out by the thermodynamical susceptibilities, in particular by the isothermal compressibility $\kappa_{T},$ which has been directly measured for a neutral interacting-fermionic gas and revealed a clear signature of the superfluid transition \cite{KuScience2012}.
The susceptibilities also play an important role in equilibrium transformations, such as the cooling by adiabatic compression or by an isocoric transformation of a gas. In such cases, the constant volume specific heat $C_{V}$ and the isothermal compressibility $\kappa_{T}$ are of particular importance.
Regarding the $C_{V},$ it is known that the non-relativistic IFG shows a monotonic non-decreasing behavior as function of $T$ for dimensions $d\ge2$ and a ``hump'' is developed for $d<2$ \cite{GretherEPJD2003}. The hump is related to the non-monotonic behavior of $\mu(T)$ and means that, for low dimensional systems, the IFG dissipate thermal fluctuations more effectively in the temperature region where $\mu>E_{F}$. The isothermal compressibility also exhibits a ``hump'' for $d<2$ with a maximum at a characteristic temperature $T_{\kappa}$ \cite{SevillaPina}. For $T>T_{\kappa}$ the system compres\-sibility diminish, vanishing as the temperature goes to infinity just like the ideal classical gas. Unexpectedly, below $T_{\kappa}$, the compressibility of the gas rises with $T$ above its $\kappa_{0}$ value, the gas turns to be more compresible than the $T=0$ state. This qualitative change in the behavior of the IFG in low dimensions has been suggested to be related to a phase transition \cite{sevilla, SevillaPina}. In addition, a thermodynamic ``equivalence'' between the ideal Bose and Fermi gases in $d=2$ has been analyzed \cite{MayPR64,LeePRE97} and extended to a more general energy-momentum dispersion relation \cite{PathriaPRE98}. Such equivalence is understood as the fact that both gases have the same temperature dependence of their respective specific heat at constant volume.
Now we turn to analyze the effects of pair production on $\kappa_{T}$ and $C_{V}$ of the relativistic IFG, A quantity of interest that is relevant in the study of the system thermodynamic fluctuations corresponds to $\langle{n}_{E_{k}}\rangle\left(1-\langle{n}_{E_{k}}\rangle\right),$ denoted with $\Lambda_{E_{k}},$ which gives account of the variance of the occupation number of the energy-state $E_{k}.$ The dependence on $\beta$ and $\mu$ has not been made explicit for the economy of writing. In figure \ref{lambda_fluctuations} we present $\Lambda_{E}$ and $\overline{\Lambda}_{E}$ for the three-dimensional case as function of $E,$ for $\widetilde{m}=1$ and for temperatures at which: a) pair production is negligible $T/T_{F}=0.1$ (circles); b) pair production starts rising $T/T_{F}=0.3$ (triangles); and c) $T/T_{F}=0.7$ where antiparticles almost equals the particles number (squares).
\begin{figure}
\includegraphics[width=0.7\columnwidth]{lambda}
\caption{(Color online) $\Lambda_{E},$ $\overline{\Lambda}_{E}$ (as defined in text) \emph{vs} the normalized energy $E/E_{F}$ are shown for the three-dimensional relativistic IFG and for different temperatures, namely $T/T_{F}=0.1$ at which pair production is negligible (circles); $T/T_{F}=0.3$ when pair production starts rising (triangles); and $T/T_{F}=0.7$ where antiparticles almost equals the particles number (squares).}
\label{lambda_fluctuations}
\end{figure}
\subsection{The isothermal compressibility $\kappa_{T}$}
The isothermal compressibility is worth of analysis since is directly related to the number fluctuations of the system and such quantity can be used to characterize many situations of the IFG, as entanglement of the system \cite{CalabreseEPL2012} for instance.
At finite temperature, $\kappa_{T}$ can be computed from the expression $(1/n_{0}^{2})\left(\partial n_{0}/\partial\mu\right)_{T}$ which results into
\begin{equation}
\kappa_{T}=\frac{R_{d}}{n_{0}^{2}k_{B}T}\int_{0}^{\infty}dk\, k^{d-1}\left[\Lambda_{E_{k}}+\overline{\Lambda}_{E_{k}}\right].
\end{equation}
In the ultrarelativistic limit last expression simplifies, in terms of polylogarithms, to
\begin{equation}\label{CompressUR}
\kappa_{T}=\frac{R_{d}\Gamma(d)}{n_{0}^{2}(\hbar c)^{d}}(k_{B}T)^{d-1}\left[-\hbox{Li}_{d-1}(-z)-\hbox{Li}_{d-1}(-z^{-1})\right]
\end{equation}
and to
\begin{equation}
\kappa_{T}=-\frac{R_{d}\Gamma(d/2)}{n_{0}^{2}\left(\hbar^{2}/2m\right)^{d/2-1}}(k_{B}T)^{d/2-1}\hbox{Li}_{d/2-1}(-z^{NR})
\end{equation}
in the non-relativistic one.
Expression \eqref{CompressUR} can be written as an even polynomial of order $d-1$ in $\beta\mu$ since for odd dimension, the term within square parenthesis can be written as
\begin{equation}\label{kapPolyn}
(-1)^{(d-1)/2}\frac{(2\pi)^{d-1}}{(d-1)!}\sum_{l=0}^{(d-1)/2}(-1)^{l}\eta_{l,d}\frac{(\beta\mu)^{2l}}{(2\pi)^{2l}}
\end{equation}
where the coefficients $\eta_{l,n}$ are given in terms of the Bernoulli numbers $B_{j}.$ Some of the coefficients are: $\eta_{(d-1)/2,d}=B_{0},$ \ldots, $\eta_{1,d}=\binom{d-1}{d-3}\sum_{j=0}^{d-3}\binom{2+j}{2}\frac{B_{j}}{2^{j}}$, $\eta_{0,d}=\sum_{j=0}^{d-1}\binom{d-1}{d-j-1}\frac{B_{j}}{2^{d-j-1}}.$ Explicit expressions for expression \eqref{kapPolyn} are given in Table \ref{Tab2} for dimensions 1, 2, and 3.
\begin{table}[h]
\caption{Some explicit even polynomial in $\beta\mu$ calculated from expression \eqref{kapPolyn}.}
\begin{tabular}{|c|c|}
\hline
dimension $d$ & $-\hbox{Li}_{d-1}(-z)-\hbox{Li}_{d-1}(-z^{-1})$ \\
\hline
1 & 1 \\
3 & $\zeta(2)+\dfrac{1}{2}(\beta\mu)^{2}$\\
5 & $\dfrac{7}{4}\zeta(4)+\dfrac{1}{2}\zeta(2)(\beta\mu)^{2}+\dfrac{1}{4!}(\beta\mu)^{4}$\\
\hline
\end{tabular}
\label{Tab2}
\end{table}
In Fig. \ref{fig:compressibility}, $\kappa_{T}$ is shown as function of temperature for $\widetilde{m}=1$ [relativistic case, panel (a)] and $\widetilde{m}=100$ [non-relativistic case, panel (b)] and $d=1/2,\,1,\,2,\,3,$ and $4$. The ultrarelativistic limit $\widetilde{m}\ll1$, has been omitted since analytical expression have been obtained.
In the low temperature regime, the compressibility rises and eventually starts diminishing with temperature exhibiting a maximum at $T_{\kappa}$. This behavior is determined by $\widetilde{m}$ and $d$. A calculation based on the observation that the product $\langle n_{E}\rangle(1-\langle n_{E}\rangle)$ is different from zero only in a narrow interval of energies around $\mu,$ gives up to second order terms in $T/T_{F}$
\begin{equation}
\kappa_{T}\simeq\frac{d \mu\left(\mu^{2}-m^{2}c^{4}\right)^{d/2-1}}{n_{0}(m_{F}c^{2})^{d}}\left(1+\frac{\pi^{2}}{6}\left(k_{B}T\right)^{2}(d-2)\left(\mu^{2}-m^{2}c^{4}\right)^{-2}\left[3\left(\mu^{2}-m^{2}c^{4}\right)+(d-4)\mu^{2}\right]\right)
\end{equation}
and by using Eq. (\ref{Sommerfeld}) we have that
\begin{equation}
\frac{\kappa_{T}}{\kappa_{0}}\simeq1-\frac{\pi^{2}}{6}\left(\frac{T}{T_{F}}\right)^{2}\left[1-2(d-2)(1+\widetilde{m}^{2})-(d-2)(d-4)(1+\widetilde{m}^{2})^{2}\right],
\end{equation}
which shows a nonmonotonic dependence with temperature whenever
\begin{equation}\label{conditiondim2}
d<\frac{2+3\widetilde{m}^{2}-\left(1+\widetilde{m}^{4}\right)^{1/2}}{(1+\widetilde{m}^{2})}.
\end{equation}
This raising of the compressibility with temperature is an abnormal feature that would have important effects on some thermodynamical transformations in low dimensional systems at low temperatures \cite{SayginJApplPhys2001,SismanApplEn2001}.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.7\columnwidth]{Compressibilty_ExactRel_Part_Antipart_ratio_1}
\includegraphics[width=0.7\columnwidth]{Compressibilty_ExactRel_Part_Antipart_ratio_100}
\caption{(Color online) Isothermal compressibility $\kappa_{T}$ normalized with its value a $T=0$ as function of the dimensionless temperature $T/T_{F}$ and dimension 1/2 (circles), 1 (squares), 2 (diamonds), 3 (up-triangles) and 4 (down-triangles) for the mass ratio $\widetilde{m}=1$ [panel (a)] and $\widetilde{m}=100$ [panel (b)]. Thin-dashed lines correspond to the cases for which pair production is neglected.}
\label{fig:compressibility}
\end{figure}
In the limit $\widetilde{m}\rightarrow0$ such abnormal behavior is presented for systems in dimensions smaller than 1, as can be checked from the expression
\begin{equation}
\frac{\kappa_{T}}{\kappa_{0}}=1+\frac{\pi^{2}}{6}(d-1)(d-2)\left(\frac{T}{T_{F}}\right)^{2}
\end{equation}
or directly from \eqref{conditiondim2}.
As can be checked straightforwardly from condition \eqref{CompresibilityUR} $\kappa_{T}$ becomes temperature independent for $d=1$ getting the value $\kappa_{0}^{UR}=\left(\pi\hbar c\, n_{0}^{2}\right)^{-1}$ [see Eq. (\ref{CompresibilityUR}) in the Appendix]. Note that for the case $d=2$ $\kappa_{T}$ is proportional to $\mu,$ which turns to be a monotonic decreasing function of $T$ for any value of $\widetilde{m}$.
For $\widetilde{m}\gg1,$ condition \eqref{conditiondim2} turns into $d<2$ which corresponds to the case analyzed by Sevilla and Pi\~na \cite{SevillaPina}.
As temperature is increased, $\kappa_{T}$ suffers a striking change for $d>1$ at the temperature when pair creation starts to be important. Instead of diminishing to zero as occurs if pair creation is neglected (dashed lines in Fig.\ref{fig:compressibility}), it starts to grow with temperature. This behavior is set on when the number of antiparticles is of the order of particles and is marked by a local minimum $T_{\kappa}^{*}$ in the range of tenths of the Fermi temperature. At higher temperatures, $\kappa_{T}$ grows with $T$ asymptotically as
\begin{equation*}
\kappa_{0}\left(1+\widetilde{m}^{2}\right)^{d/2-1}\, (d-1)!\, \zeta(d-1)\,2\left(1-2^{2-d}\right)\left(\frac{T}{T_{F}}\right)^{d-1},
\end{equation*}
with $\zeta(x)$ the Riemann zeta function.
In contrast, $\kappa_{T}/\kappa_{0}$ tends asymptotically to the cons\-tant $(1+\widetilde{m})^{-1/2}$ for $d=1$, while it goes to zero for $d<1$ as can be seen from Fig.\ref{fig:compressibility} for $d=1/2$. In this latter case, though the system behave qualitatively as standard matter, the effects due to pair creation can be noted quantitatively from the departure to the case when no pair creation is considered (thin-dashed lines).
The change of the temperature dependence of $\kappa_{T}$ that occurs at $T_{\kappa}^{*}$ for $d>1$, marks a drastic change in the qualitative behavior of the system driven by pair creation, suggesting a possible transition from a normal phase, to a phase in which the system becomes arbitrarily compressible with temperature.
For the three-dimensional case we have that the minimum of $\kappa_{T}$ occurs at approximately at $T_{\kappa}^{*}=0.34732$ for $\widetilde{m}=0.01$, $T_{\kappa}^{*}=0.32018$ for $\widetilde{m}=1$, and $T_{\kappa}^{*}=0.07465$ for $\widetilde{m}=100$.
At particle densities of the order of the nuclear matter 0.122 fm$^{-3}$, the corresponding Fermi energy is approximately $480.618$ MeV in the limit $\widetilde{m}\ll1$. With these values we estimate $k_{B}T_{\kappa}^{*}\simeq 166.928$ MeV for $\widetilde{m}=0.01$. This value is of the order of the expected crossover temperature to the quark-gluon plasma \cite{Martinez2013}, which from QCD calculations is expected to be 173$\pm15$ MeV for massless
quarks \cite{Karsch2002}, while QCD lattice calculations with non-zero quark masses ($m_{u}=2.3\pm0.5$ MeV, $m_{d}=4.8^{+0.7}_{-0.3}$ MeV \cite{BeringerPRD2012}) give crossover temperatures between 150 and 200 MeV \cite{Martinez2013}. Though for temperatures above $T_{\kappa}^{*}$ the ultrarelativistic ideal gas provides a good description of the quark-gluon plasma, for $T<T_{\kappa}^{*}$ the strong interactions become relevant in the thermodynamics of the system, interactions that our oversimplified approach neglects.
\subsection{The specific heat at constant volume $C_{V}$}
The specific heat at constant volume is expressed in terms of the $\Lambda_{E_{k}}$'s as
\begin{widetext}
\begin{equation}\label{SpecificHeat}
C_{V}=\frac{R_{d}V_{d}}{k_{B}T^{2}}\int_{0}^{\infty}dk\, k^{d-1} E_{k}^{2}\left[\Lambda_{E_{k}}+\overline{\Lambda}_{E_{k}}\right]-
\frac{V_{d}}{\kappa_{T}T}\left[\frac{R_{d}}{n_{0}k_{B}T}\int_{0}^{\infty}dk\, k^{d-1}E_{k}\left[\Lambda_{E_{k}}+\overline{\Lambda}_{E_{k}}\right]\right]^{2}.
\end{equation}
\end{widetext}
As shown in Fig. \ref{fig:SpecificHeat}, the low temperature behavior is given by the well known linear dependence, with the pre\-factor $d\pi^{2}(1+\widetilde{m}^{2})/3$ which comes only from the Fermi-Dirac statistics of the particles and the dimensionality of the system.
In the same range of temperatures where a local minimum in $\kappa_{T}$ is found, the specific heat changes its linear dependence characteristic of the low temperature regime, to the temperature dependence $T^{d}$ as is shown in Fig. \ref{fig:SpecificHeat}. For $\widetilde{m}\ll1$ (not shown in Fig. \ref{fig:SpecificHeat}) the transition is smooth and becomes more marked as the mass ratio increases, as is contrasted in panels (a) and (b), where a plateau appears before the power-law growth. This behavior differs from the case for which only particles are considered (thin-dashed lines) which reaches the Dulong-Pettit limit $C_{V}/dNk_{B}=1$.
\begin{figure}[h]
\centering
\includegraphics[width=0.7\columnwidth]{SpecificHeat_ExactRel_Part_AntiPart_ratio_1}
\includegraphics[width=0.7\columnwidth]{SpecificHeat_ExactRel_Part_AntiPart_ratio_100}
\caption{(Color online) Dimensionless specific heat at constant volume as function of the dimensionless temperature $T/T_{F}$ for dimension 1/2 (circles), 1 (squares), 2 (diamonds), 3 (up-triangles) and 4 (down-triangles) and mass ratio $\widetilde{m}=1$ [panel (a)], $\widetilde{m}=100$ [panel (b)]. Thin-dashed lines correspond to the cases for which pair production is neglected.}
\label{fig:SpecificHeat}
\end{figure}
The exact result $C_{V}/N_{0}k_{B}=\left(\pi^{2}/3\right)(T/T_{F})$ is found in the limit $\widetilde{m}\rightarrow0$ for $d=1$.
In Ref. \cite{BlasPRE1999} the authors considered the relativistic Bose and Fermi gases, at low temperatures, they rightly neglected the antiparticles, and concluded that in two dimensions both gases are thermodynamically inequivalent, in contrast to the non-relativistic case in which they do, however it seems they missed that both gases are thermodynamically equivalent in one dimension in the ultrarelativistic limit. In fact, it is known that the condition for the equivalence between the two quantum gases consists of the constancy of the single-particle density of states $g(E)$. In the exact relativistic case, is the finite rest mass of the particle what avoids such possibility, since there is no value of $d$ which makes the density of states $g(E)=E(E^{2}-m^{2}c^{4})^{d/2-1}\theta(E-mc^{2})$, $\theta(x)$ being the Heaviside step function, to be a constant as occurs in the non-relativistic and ultrarelativistic cases, where $g(E)\propto E^{d/2-1}$ and $\propto E^{d-1}$, respectively.
\section{\label{IV}Conclusions and final remarks}
We have studied the effects of the system dimensionality and quantum-relativity on the thermodynamics of an ideal Fermi gas.
The temperature dependence of the chemical potential is determined by the system dimensionality and by the particles rest mass. We recovered the unusual low temperature dependence of $\mu(T)$ for $d<2$ \cite{GretherEPJD2003} in the non-relativistic limit $m\gg m_{F}$. For arbitrary values of the rest mass, the nonmonotonic behavior of $\mu$ in the low temperature regime appears if $d<2-(1+\widetilde{m}^{2})^{-1}$, which includes the ultrarelativistic case for $\widetilde{m}\ll1$.
Singularly, for dimensions smaller than one, $\mu$ increases monotonically with $T.$ This peculiar behavior occurs since for low dimensional systems, the creation of particle-antiparticle pairs occurs at a so low rate that the initial number of fermions dominates the thermodynamic behavior of the system. This argument is supported from the temperature dependence of $\kappa_{T}$ which vanishes as $T\rightarrow\infty,$ just as in the case when only particles are considered (dashed lines in Fig.\ref{fig:ChemPot}). The temperature dependence of $\mu$ for high temperatures described in Fig.\ref{fig:ChemPot}(a) is also observed in the relativistic Bose gas with pair production \cite{FrotaPRA1989} for $d>1$, with the remarkable difference that for the Bose gas, $\vert\mu_{B}\vert\le mc^{2}$, and therefore the chemical potential vanishes as $T\rightarrow\infty$ even for $0<d<1$. Except in the case $d=1$ for which we have $\mu_{B}=0$ for all $T$ where $\mu_{B}$ is the chemical potential of the Bose gas.
The effects of pair production are exhibited in the thermodynamical susceptibilities as a change in their temperature dependence that appears at some tenths of the Fermi temperature (as shown in Figs. \ref{fig:compressibility} and \ref{fig:SpecificHeat}) corresponding to the temperature range at which the pair production becomes significantly important. Both susceptibilities start growing without limit as a power law of $T$ after this crossover. The temperature that points out the crossover $T_{\kappa}^{*}$ could be determined from the apparent local minimum exhibited in the isothermal compressibility.
Interestingly for three dimensions and at particle densities of the size of the nuclear densities, the crossover occurs at approximately 167 MeV for a rest mass close to 4.86 MeV. This temperature is close to the quark-gluon plasma critical temperature $\lambda_{QCD}$ expected to occur for light quark masses. Above such a temperature the strong interaction among quarks can be neglected leading to the ideal situation described in this paper.
Since our calculations exhibits the exact thermal behavior of the relativistic IFG, evidently can describe systems beyond the standard complete-degeneracy approximation ($T=0$), generally used in situations where a disparate difference between the system's and Fermi temperature exists, as in the case of white dwarf stars, where $T/T_{F}\simeq10^{-2}$. Since the effects of dimensionality on the thermodynamical susceptibilities at low temperatures are not negligible, it would be desirable to extend the analysis presented in Ref. \cite{ChavanisPRD2007} to incorporate finite temperature calculations, particularly for $d<3$, where the susceptibilities exhibit an anomalous behavior. By doing so, an study of the response of white dwarf stars to thermal and pressure fluctuations could plausibly establish a lower bound on the anthropic dimension of the Universe. Thus, in addition with the upper bound ($d<4$) for the anthropic dimension of the Universe given by Chavanis \cite{ChavanisPRD2007}, it would be plausibly justified the three spatial dimensions of the observed Universe.
\section{Acknowledgments}
The authors acknowledge financial support from DGAPA-UNAM grant PAPIIT-IN111070.
\section{Appendix: The zero temperature relativistic IFG}
The zero point energy per particle, $u_{0}=U_{0}/N_{0},$
can be written in terms of the Gaussian or ordinary hypergeometric function $_{2}F_{1}(a_{1},a_{2};b_{1};z)$ \cite{Abramowitz} as
\begin{align}
u_{0}=& mc^{2}\, _{2}F_{1}\left[-1/2,d/2;1+d/2;-\widetilde{m}^{-2}\right].
\end{align}
that reduces to elementary functions for integer values of $d$.
For $\widetilde{m}\gg1$ we have
\begin{equation}
u_{0}=mc^{2}+\frac{d/2}{1+d/2}E_{F}^{NR}-\frac{d/2}{2+d/2}\frac{\left(E_{F}^{NR}\right)^{2}}{mc^{2}}+\ldots
\end{equation}
In the opposite limit, $\widetilde{m}\ll1,$ we can write
\begin{equation}
u_{0}=\frac{d}{d+1}E_{F}^{UR}\left[1+\frac{1}{2}\frac{d+1}{d-1}\widetilde{m}^{2}-\frac{1}{8}\frac{d+1}{d-3}\widetilde{m}^{4}+\ldots\right]
\end{equation}
for $d\neq1,\, 3,\, 5\ldots$.
For $d=3$ and 1 we have, respectively
\begin{align}
u_{0}&=\frac{3}{4}\, E_{F}^{UR}(1+\widetilde{m}^{2})^{1/2}\left[1+\frac{\widetilde{m}^{2}}{2}\left(1-\frac{\widetilde{m}^{2}\arcsinh(\widetilde{m}^{-1})}{(1+\widetilde{m}^{2})^{1/2}}\right)\right]\\
u_{0}&=\frac{1}{2}\, E_{F}^{UR}\left[(1+\widetilde{m}^{2})^{1/2}+\widetilde{m}^{2}\arcsinh(\widetilde{m}^{-1})\right],
\end{align}
In the $\widetilde{m}\ll1$ limit, last expressions can approximated by
\begin{align}
u_{0}&=\frac{3}{4}\, E_{F}^{UR}\left[1+\widetilde{m}^{2}+\frac{1}{8}\widetilde{m}^{4}\ln\left(\widetilde{m}\right)+\ldots\right]\\
u_{0}&=\frac{1}{2}\, E_{F}^{UR}\left[1-\widetilde{m}^{2}\ln\widetilde{m}+\frac{1}{8}\widetilde{m}^{4}+\ldots\right].
\end{align}
For the zero point pressure $P_{0}$ we have
\begin{align}
P_{0}/n_{0}=& m_{F}c^{2}\, \sqrt{1+\widetilde{m}^{2}} - u_{0}
\end{align}
which for $\widetilde{m}\gg1$ last expression reduces to
\begin{equation}
P_{0}/n_{0}=\frac{2}{d+2}E_{F}^{NR}+\frac{d/2}{2+d/2}\frac{\left(E_{F}^{NR}\right)^{2}}{mc^{2}}+\ldots
\end{equation}
where the first term corresponds to the well known non-relativistic case. In the opposite limit
\begin{equation}
P_{0}/n_{0}=\frac{1}{d+1}E_{F}^{UR}-\frac{1}{2(d-1)}E_{F}^{UR}\widetilde{m}^{2} +\frac{d}{8(d-1)}E_{F}^{UR}\widetilde{m}^{4}\ldots
\end{equation}
the first term corresponds to the well known result in the ultrarelativistic case, the next terms are valid always that $d$ is not an odd integer.
For $d=3,1$ we have respectively
\begin{align}
P_{0}/n_{0}=&\frac{1}{4}E_{F}^{UR}-\frac{1}{4}E_{F}^{UR}\widetilde{m}^{2} +\frac{3}{2^{5}}E_{F}^{UR}\widetilde{m}^{4}\ln\widetilde{m}\ldots\\
P_{0}/n_{0}=&\frac{1}{2}E_{F}^{UR}+\frac{1}{2}E_{F}^{UR}\widetilde{m}^{2}\left[1-\ln\widetilde{m}\right]-\frac{1}{2^{4}}E_{F}^{UR}\widetilde{m}^{4}+\ldots
\end{align}
The inverse of the isothermal compressibility $\kappa_{T}=-\left(1/V_{d}\right)\left(\partial V_{d}/\partial P_{0}\right)_{T}$ is given by
\begin{equation}
\kappa_{0}^{-1}=\frac{n_{0}}{d}\frac{m_{F}c^{2}}{\sqrt{1+\widetilde{m}^{2}}}
\end{equation}
in the limit of $\widetilde{m}\gg1$ we have that
\begin{equation}
\kappa_{0} \simeq \kappa_{0}^{NR}\left[1+\frac{1}{2\widetilde{m}^{2}}\right]
\end{equation}
where $\kappa_{0}^{NR}=d/[(d+2)P_{0}]$ is the NR isothermal compressibility, which reduces to the well known result $\kappa_{0}=(3/5)P_{0}^{-1}$ for $d=3$, and
\begin{equation}\label{CompresibilityUR}
\kappa_{0} \simeq \kappa_{0}^{UR}\left[1+\frac{1}{2}\widetilde{m}^{2}\right]
\end{equation}
in the $\widetilde{m}\ll1$ case, where $\kappa_{0}^{UR}=d/[(d+1)P_{0}]$. These results show that the gas is more compressible than in their respective limits $\widetilde{m}\rightarrow\infty$ and $\widetilde{m}=0$.
|
\section{Introduction}
The recent results from {\sc Bicep2}\ \cite{Ade:2014xna}, hinting at a
detection of primordial $B$-mode power in the Cosmic Microwave
Background (CMB) polarisation, place the inflationary paradigm on much
firmer footing. This result, in combination with the {\sc Planck}\ total
intensity measurement \cite{Ade:2013ktc}, imply that primordial
perturbations are generated from an almost de-Sitter like phase of
expansion early in the Universe's history before the standard big bang
scenario.
At first glance there is potential tension between the polarisation
measurements made by {\sc Bicep2}\ and {\sc Planck}'s total intensity
measurements. {\sc Planck}'s power spectrum is lower than the best-fit
$\Lambda$CDM models at multipoles $\ell \lesssim 40$ and {\sc Bicep2}'s high
$B$-mode measurement exacerbates this since tensor modes also
contribute to the total intensity. The tension is indicated by the
difference in the $r\sim 0.2$ value implied by {\sc Bicep2}'s measurements
and the 95\% limit of $r < 0.1$ implied by the {\sc Planck}\ data for
$\Lambda$CDM models. Many authors have pointed out how the tension can
be alleviated by going beyond the primordial power-law, $\Lambda$CDM
paradigm by allowing running of the spectral indices, enhanced
neutrino contributions (see for examples \cite{Czerny:2014wua,Contaldi:2014zua,Gong:2014cqa,Zhang:2014dxk}) or
more exotic scenarios \cite{Anchordoqui:2014dpa}. However the simplest explanation,
that also fits the data best, is one where there is a slight change in
acceleration trajectory during the inflationary phase when the largest
modes were exiting the horizon. This was shown by \cite{Contaldi:2014zua} where
a specific model was used to generate a slightly faster rolling
trajectory at early times. The effect of such a ``slow-to-slow-roll''
transition is to result in a slightly suppressed primordial, scalar
power spectrum that fits the {\sc Planck}\ data despite the large tensor
contribution required by {\sc Bicep2}. In \cite{Contaldi:2014rna} the author analyses
generalised accelerating, or inflating, trajectories that fit the
combination of {\sc Bicep2}\ and {\sc Planck}\ data and conclude that the
suppression is required at a significant level and the best-fit
trajectories are all of the form where the acceleration has a slight
enhancement at early times.
An alternative explanation is that the $B$-mode power observed by
{\sc Bicep2}\ is not due to foregrounds and is not primordial. This
possibility has been discussed by various authors
\cite{Mortonson:2014bja,Flauger:2014qra} who point out that more measurements on the
frequency dependence of the signal are required to definitively state
whether we have detected the signature of primordial tensor
modes. These measurements will be provided in part by the {\sc Planck}\
polarisation analysis and {\sc Bicep2}'s cross-correlation with further KECK
data \cite{Ade:2014xna}.
If the {\sc Bicep2}\ result stands the test of time then the signal we point
out in the analysis below is expected to be present if the simplest
models of inflation driven by a single, slow-rolling scalar field are
the explanation behind the measurements. In this case a measurement of
tensor mode amplitude, or $r$, is a direct measurement of the
background acceleration since $r \sim 16\epsilon$ and the tension
between {\sc Bicep2}\ polarisation and {\sc Planck}\ total intensity measurements
implies a change in the acceleration at early times. In turn, the
change in acceleration enhances the non-Gaussianity on scales that
were exiting the horizon while the acceleration was changing.
In this {\sl paper} we construct a simple toy-model inspired by the best
fitting trajectories found in \cite{Contaldi:2014rna} and calculate its
bispectrum numerically. At small scales, as one would expect, the
non-Gaussianity is small $\mathcal{O}(10^{-2})$ \cite{Maldacena:2002vr,Creminelli:2004yq} but at large scales,
where the scalar power spectrum is suppressed, the non-Gaussianity can
be significantly larger, $\mathcal{O}(10^{-1})$. The results are compared
against the slow-roll approximation in the equilateral configuration
and the squeezed limit consistency relation. Whilst at small scales
there is exceptional agreement with the slow-roll approximation, at
large scales the results can deviate by up to 10\%.
This {\sl paper} is organised as follows. We outline the calculation of the
scalar and tensor power spectra in Section \ref{power_spec} and
summarise the calculation of the bispectrum in Section
\ref{bispectrum}. Our results are presented in Section \ref{results}
and we discuss their implications in Section \ref{conclusion}.
\section{Computation of the scalar power spectrum}\label{power_spec}
The calculation is best performed in a gauge where all the scalar
perturbations are absorbed into the metric such that $g_{ij} =
a^{2}\,(t)e^{2\zeta(t, \mathbf{x})}\delta_{ij}$ and the inflaton
perturbation $\delta\phi(t, \mathbf{x}) = 0$. The primordial power
spectrum is then simply given by:
\begin{equation}\label{PowerSpectrum}
\langle\zeta^{}_{k_{1}}\zeta^\star_{k_{2}}\rangle = (2\pi)^{3}\delta^{(3)}(\mathbf{k}_{1}+\mathbf{k}_{2})P_\zeta(k_1)\,,
\end{equation}
where $\mathbf{k}$ is the Fourier wavevector and $k\equiv |\mathbf{k}|$. The mode $\zeta_{k}(t)$ satisfies the Mukhanov-Sasaki equation \cite{Mukhanov:1985rz,Sasaki:1986hm}
\begin{equation}\label{Mukh}
\frac{\mathrm{d}^{2}\zeta_{k}}{\mathrm{d}N^{2}} + (3 + \epsilon - 2\eta)\frac{\mathrm{d}\zeta_{k}}{\mathrm{d}N} + \frac{k^{2}}{a^{2}H^{2}}\zeta_{k} = 0\,.
\end{equation}
In the above $N$ is the number of $e-$folds which increases with time or alternatively
\begin{equation}
H = \frac{\dot{a}}{a} = \frac{dN}{dt}\,,
\end{equation}
and $\epsilon$ and $\eta$ are the usual slow-roll variables defined by
\begin{equation}
\epsilon = -\frac{\dot{H}}{H^{2}},\,\,\,\,\,\,\,\, \eta = \epsilon - \frac{1}{2H}\frac{d\ln\epsilon}{dt}\,.
\end{equation}
Outside the horizon $\zeta_{k}$ quickly goes to a constant and the
power spectrum is then related to the freeze-out value of $\zeta_k$ on
scales $k\ll aH$
\begin{equation}
P_\zeta(k) = \left|\zeta_{k\ll aH}\right|^{2}\,.
\end{equation}
The initial conditions for the solutions to (\ref{Mukh}) can be set
when the mode is much smaller than the horizon $k\gg aH$ and takes on
the Bunch-Davies form \cite{Bunch:1978yq}
\begin{equation}\label{Initial_zeta}
\zeta_{k} \to \frac{1}{M_{\rm pl}}\,\frac{e^{-ik\tau}}{2a\sqrt{k\epsilon}}\,,
\end{equation}
where $\tau$ is conformal time defined by $\mathrm{d}N/\mathrm{d}\tau =
aH$.
An identical calculation can be performed for the tensor power
spectrum $P_{h}(k) = \left|h_{k\ll aH}\right|^{2}$ with $h_{k}$
satisfying the following differential
equation
\begin{equation}\label{tensors}
\frac{\mathrm{d}^{2}h_{k}}{\mathrm{d}N^{2}} + (3 -
\epsilon)\frac{\mathrm{d}h_{k}}{\mathrm{d}N} +
\frac{k^{2}}{a^{2}H^{2}}h_{k} = 0\,,
\end{equation}
with initial condition
\begin{equation}
h_{k} \to \frac{1}{M_{\rm pl}}\,\frac{e^{-ik\tau}}{a\sqrt{2k}}\,,
\end{equation}
in the limit where $k \gg aH$. Solving for $P_{\zeta}(k)$ and $P_{h}(k)$ numerically we can calculate $n_{s}, r$ and $n_{t}$ directly from their definitions:
\begin{eqnarray}\label{nsr_numeric}
n_{s}(k_\star) & = & 1 + \left.\frac{\mathrm{d}\ln \left[k^{3}P_{\zeta}(k)\right]}{\mathrm{d}\ln k}\right|_{k = k_\star}\,\\
r(k_\star) & = & 8\,\frac{P_h(k_\star)}{P_{\zeta}(k_\star)}\,\nonumber\\\nonumber
n_{t}(k_\star) & = & \left.\frac{\mathrm{d}\ln \left[k^{3}P_{h}(k)\right]}{\mathrm{d}\ln k}\right|_{k = k_\star}\,\\\nonumber
\end{eqnarray}
The factor of 8 comes from how the tensor perturbations are normalised in the second order action.
The above procedure outlines the general calculation of the primordial
power spectrum from inflation. In this work we are interested in
specifying a background model favoured by the recent {\sc Bicep2}\ + {\sc Planck}\
data. In particular we choose a function for $\epsilon$, then $\eta$
and $H$ are easily obtained by its derivative and integral
respectively.
Instead of a direct function of time or $N$ though we specify
$\epsilon(x)$ where $x = \ln (k'/k_{\rm min})$. $k'$ is the mode
crossing the horizon at $e-$foldings $N$ ($k' = aH$) and $k_{\rm min} \sim
10^{-5} (\text{Mpc})^{-1}$ is the largest scale observable today. In
addition to being proportional to $r$ this condition allows one to
easily specify how the background should evolve in our observational
window. For concreteness we require $\epsilon$ to be relatively large,
but still satisfying the slow-roll limit,
at large scales and then to flatten out into another slow-roll regime
with a smaller value. To this end we adopt a simple toy-model for
$\epsilon$ as a function of x
\begin{equation}\label{toy_model}
\epsilon = \left\{\epsilon_1 \tanh\left[(x - x_0)\right] + \epsilon_2\right\}\left(1 + m x\right)\,,
\end{equation}
where the coefficients $\epsilon_1$, $\epsilon_2$, $m$, and
$x_0$ are chosen to give a final power spectrum with the required
suppression and position ($\sim 26\%$ and 1.5$\times 10^{-3}$
Mpc$^{-1}$ respectively \cite{Contaldi:2014zua}) and $n_s\sim 0.96$ on small
scales. Fig.~\ref{fig:background} shows $\epsilon$ and $\eta$ as a
function of $N$ for this toy-model and the resulting power spectra are
shown in Fig.~\ref{fig:ps}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=8.5cm,trim=0cm 0cm 0cm 0cm,clip]{fig}
\caption{ Background functions $\epsilon$ (red, solid) and $\eta$ (blue, dashed) of our toy-model plotted as a function of $e-$folds $N$. The grey vertical line indicates roughly the time when the first observable mode crosses the horizon.}
\label{fig:background}
\end{center}
\end{figure}
\begin{figure*}[t]
\begin{center}
\begin{tabular}{cc}
\makebox[8.5cm][c]{
\includegraphics[width=8.5cm,trim=0cm 0cm 0cm
0cm,clip,angle=0]{fig1}}&
\makebox[8.5cm][c]{\includegraphics[width=8.5cm,trim=0cm 0cm 0cm
0cm,clip,angle=0]{fig2}}
\end{tabular}
\caption{Left: Scalar (red, solid) and tensor (blue, dashed)
dimensionless power-spectra. The tensors have been multiplied by
a factor of 25 for comparison. Right: $r$ (red, solid) and
$n_{s} - 1$ as functions of $k$. The parameters in the toy-model
were chosen to give a good match to the {\sc Planck}\ and
{\sc Bicep2}\ data.}
\label{fig:ps}
\end{center}
\end{figure*}
\section{Computation of the bispectrum}\label{bispectrum}
The largest contribution to primordial non-Gaussianity will come from the bispectrum of the curvature perturbation
\begin{equation}\label{3rd}
\langle\zeta_{k_{1}}\zeta_{k_{2}}\zeta_{k_{3}}\rangle = (2\pi)^{3}\delta^{(3)}(\mathbf{k}_{1}+\mathbf{k}_{2}+\mathbf{k}_{3})B(k_{1}, k_{2}, k_{3})\,.
\end{equation}
The quantity that is often quoted in observational constraints is the dimensionless, reduced bispectrum
\begin{eqnarray}\label{fnl_def}
f_{\mathrm{NL}}(k_{1}, k_{2}, k_{3}) &=& \frac{5}{6}\,B(k_{1}, k_{2},
k_{3}) / \left(|\zeta_{k_{1}}|^{2}|\zeta_{k_{2}}|^{2}+\right.\nonumber\\
&&\left.|\zeta_{k_{1}}|^{2}|\zeta_{k_{3}}|^{2}+|\zeta_{k_{2}}|^{2}|\zeta_{k_{3}}|^{2}\right)\,,
\end{eqnarray}
The analytical calculation is much simpler if we consider the
equilateral configuration $f_{\text{NL}}(k,k,k)$ however this is not a
directly observed quantity as the estimator requires $B(k_{1},
k_{2}, k_{3})$ to be factorizable \cite{Creminelli:2005hu}. This is not true for the general
case, which we are considering. However the overall amplitude of the
reduced bispectrum gives a good indication of the size of the expected
observable $f_{\mathrm{NL}}$.
All theories of inflation will produce a non-zero bispectrum. This is
simply because gravity coupled to a scalar field is a non-linear
theory and will contain interaction terms for the primordial curvature
perturbation $\zeta(t,\textbf{x})$. These interaction terms will
source the bispectrum with the largest contributors coming from
tree-level diagrams associated with the cubic interaction terms. The
bispectrum can then be calculated using the ``in-in'' formalism
\cite{Adshead:2009cb,Maldacena:2002vr,Seery:2005gb}, which to tree
level becomes
\begin{equation}\label{in_in}
\langle\zeta^{3}(t)\rangle = -i\int_{-\infty}^{t}\mathrm{d}t'\langle\left[\zeta^{3}(t), H_{\text{int}}(t')\right]\rangle\,,
\end{equation}
where $H_{\text{int}}$ is the interaction Hamiltonian associated with the following third order action
\begin{eqnarray}\label{action_final}
S_{3} &=& \!\!\int d^4x\, a^{3}\epsilon \left[\left(2\eta - \epsilon\right)\zeta \dot{\zeta}^{2} + \frac{1}{a^{2}}\epsilon\zeta(\partial\zeta)^{2}\right.\nonumber\\
& &\!\!\!\!\left. - (\epsilon - \eta)\zeta^{2}\partial^{2}\zeta - 2\epsilon\left(1 - \frac{\epsilon}{4}\right)\dot{\zeta}\partial_{i}\zeta\partial_{i}\partial^{-2}\dot{\zeta}\right.\nonumber\\
& & \!\!\!\!\left. + \frac{\epsilon^{2}}{4}\partial^{2}\zeta\partial_{i}\partial^{-2}\dot{\zeta}\partial_{i}\partial^{-2}\dot{\zeta}\right]\,,
\end{eqnarray}
The numerical calculation of the bispectrum is technically challenging
and is described in more detail in \cite{Horner:2013sea}. Briefly, for the
equilateral configuration it requires the calculation of the
following integral
\begin{equation}\label{fnl_prelim}
f_{\mathrm{NL}} = \frac{1}{3|\zeta|^{4}}\times{\cal I}\left[\zeta^{*3}\int_{N_{0}}^{N_{1}} dN\, (f_{1}\zeta^{3} + f_{2}\zeta\zeta^{\prime 2})\right]\,,
\end{equation}
where $\zeta = \zeta_{k}$, $\zeta^{\prime} =
\mathrm{d}\zeta/\mathrm{d}N$, and $\cal I$ represents the imaginary
part. The background functions $f_{i}$ are given by
\begin{eqnarray}
\!\!\!\!\!f_{1} & = & \frac{5k^{2}a\epsilon}{H}(2\eta - 3\epsilon)\,,\nonumber\\
\!\!\!\!\!f_{2} & = & -5Ha^{3}\epsilon\left(4\eta - \frac{3}{4} \epsilon^{2}\right)\,.
\end{eqnarray}
The times $N_{0}$ and $N_{1}$ correspond to when the mode is
sufficiently sub- and super-horizon respectively. For calculating the
shape dependence we restrict ourselves to the case of isosceles
triangles so we parametrise our modes in the following
way. $|\textbf{k}_{1}| = |\textbf{k}_{2}| = k, |\textbf{k}_{3}| =
\beta k $. This covers most configurations of interest ($\beta = 0$ is
squeezed, $\beta = 1$ is equilateral, $\beta = 2$ is folded) and is
simple to interpret.
\begin{figure*}[t]
\begin{center}
\begin{tabular}{cc}
\makebox[8.5cm][c]{
\includegraphics[width=8.5cm,trim=0cm 0cm 0cm
0cm,clip,angle=0]{fig3}}&
\makebox[8.5cm][c]{\includegraphics[width=8.5cm,trim=0cm 0cm 0cm
0cm,clip,angle=0]{fig3b}}
\end{tabular}
\caption{$f_{\rm NL}$ as a function of $k$ for equilateral (left) and
squeezed (right) configurations. The blue (dashed) curves
represents the numerical calculation. The red curves represent
the slow roll approximation (\ref{SRapprox}) (left) and the consistency
condition $5/12 (n_{s} - 1)$ (right). It is not
possible to calculate the exact squeezed configuration
numerically so a configuration with $\beta = 0.1$ was used to
approximate the squeezed limit. }
\label{fig:fnl1}
\end{center}
\end{figure*}
\section{Results}\label{results}
For the toy-model given in (\ref{toy_model}) the
non-Gaussianity amplitude is plotted in Fig.~\ref{fig:fnl1}. For
comparison, as well as a consistency check, we plot the full-numerical
calculation (blue-dashed) as well as the the slow-roll approximation
(red-solid) which, in the equilateral limit, is given by \cite{Maldacena:2002vr}
\begin{equation}\label{SRapprox}
f_{\mathrm{NL}}(k) = \frac{5}{12}\left(n_{s}(k) - 1 + \frac{5}{6}n_{t}(k)\right)\,.
\end{equation}
In applying this formula we used the exact values of $n_{s}$ and
$n_{t}$ given by equations \ref{nsr_numeric}. As can be seen from
Fig.~\ref{fig:fnl1}, if values close to $r\sim 0.2$ are confirmed from
polarisation measurements, the non-Gaussianity on large scales are
likely to be an order of magnitude larger than expected. This is
simply because $r \propto \epsilon$ but on smaller scales $\epsilon$
is constrained to be lower by the total intensity measurements. The
only way to reconcile the two regimes is by having $\epsilon$ change
to a lower value at later times and this results in an enhancement of
non-Gaussianity being generated as the value is changing.
Fig.~\ref{fig:fnl1} also shows that, even with strong scale
dependence, there is remarkable agreement between the full numerical
results and the Maldacena formula, with deviations only occurring at
the largest scales. Fig.~\ref{fig:fnl2} shows the complete scale and
shape dependence of $f_{\rm NL}$.
\begin{figure}[t]
\begin{center}
\includegraphics[width=8.5cm,trim=0cm 0cm 0cm 0cm,clip]{fnl}
\caption{ $f_{\rm NL}$ as a function of scale $k$ and shape $\beta$. There
is a mild peak in the equilateral limit, $\beta = 1$. For all
shapes the non-Gaussianity peaks around the scales corresponding
to the size of the horizon at the time when the background acceleration is
changing.}
\label{fig:fnl2}
\end{center}
\end{figure}
\section{Discussion}\label{conclusion}
Models of inflation that contain a feature causing the background
acceleration to change can reconcile {\sc Planck}\ and {\sc Bicep2}\ observations of
the CMB total intensity and polarisation power spectra. We have shown
that these models result in enhanced non-Gaussianity at scales
corresponding to the size of the horizon at the time when the
acceleration is changing. The level of non-Gaussianity at these scales
is an order of magnitude larger than what is expected in the
standard case with no feature and is strongly scale dependent.
Whilst the effect was illustrated using a simple toy-model of the
background evolution $H(t), \epsilon(t)$, etc, we expect the
non-Gaussian enhancement to be present in any model where the
acceleration changes relatively quickly in order to fit the {\sc Planck}\ and
{\sc Bicep2}\ combination. The exact form of non-Gaussianity will obviously
be model dependent.
It is not clear that this level of non-Gaussianity will be observable
since it corresponds to scales $\ell \sim 2\to 80$ where there may not be a
sufficient number of CMB modes on the sky to ever constrain $f_{\rm NL}$ to
${\cal O}(10^{-1})$. However cross-correlation with other surveys of
large scale structure may help to constrain non-Gaussianity on these
scales. In particular it may be possible to detect any anomalous
correlation of modes induced by the non-Gaussianity.
The biggest question at this time however is whether or not the
claimed detection of primordial tensor modes by {\sc Bicep2}\ is
correct. This will be addressed in the near future as the polarisation
signal is observed at more frequencies at the same signal-to-noise
levels reached by the {\sc Bicep2}\ experiment.
\begin{acknowledgments}
We thank Marco Peloso for useful discussions. JSH is supported by a
STFC studentship. CRC and JSH acknowledge the hospitality of the
Perimeter Institute for Theoretical Physics and the Canadian
Institute for Theoretical Astrophysics where some of this work was
carried out.
\end{acknowledgments}
|
\section{\bf Introduction}\par
Now that we have entered the era of precision QCD,
by which we mean
predictions for QCD processes at the total precision tag of $1\%$ or better,
it is paramount to have rigorous baselines with respect to which to compare
theoretical results both against one another and against the new LHC precision data as well as for expectations for the future FCC~\cite{fcc} device.
For example, we have argued in Refs.~\cite{radcor2013,1305-0023,herwiri,qced} that exact, amplitude-based
resummation allows one to have better than 1\% theoretical precision
as a realistic goal in such comparisons as those needed in determining the detailed properties of the newly discovered BEH~\cite{EBH} boson~\cite{atlas-cms-2012}, so that one can indeed
distinguish new physics(NP) from higher order SM processes and can distinguish
different models of new physics from one another as well. One of the
ingredients in exact amplitude-based resummation is the respective set of hard gluon residuals which determine the order of exactness in
the respective QCD predictions.
These residuals obtain from exact fixed-order QCD perturbation theory and thus
any attempt to determine their precision tag necessarily entails determining the respective precision tag of the corresponding fixed-order results.
Unfortunately, in the current state of the art, even though we have for a process such as single $Z/\gamma^*$ production at LHC(FCC) even the NNLO exact result~\cite{nnlo-gZ},
when one tries to compare the predicted $p_T \;(\text{or}\; \phi^*_\eta)$\footnote{ Here, $\phi^*_\eta$ is a new $p_T$-related variable~\cite{phietastr} used in
some of the comparisons with data and it is defined as follows: $\phi_\eta^*=\tan(\frac{1}{2}(\pi-\Delta\phi))\sin\theta^* \cong \left|\sum \frac{{p_i}_T \sin\phi_i}{Q}\right| +{\cal O}(\frac{{{p_i}_T}^2}{Q^2})$, where $\Delta\phi=\phi_1-\phi_2$ is the azimuthal angle
between the two leptons which have transverse momenta $\vec{p_i}_T,\; i=1,2,$
and $\theta^*$ is the scattering angle of the dilepton system relative to the beam direction when one boosts to the frame along the beam direction such that the leptons are back to back.} spectrum with the LHC and FNAL data, one sees the type of divergence shown in Figs.~\ref{atlas-hass-1} and ~\ref{yin-1} taken from Refs.~\cite{atlas-hassani,yin}\footnote{Note that in Fig.~\ref{yin-1} the comparisons with RESBOS~\cite{resbos1,resbos2,resbos3}, which realizes the ``CSS'' resummation in Ref.~\cite{css}, show that in the regime where fixed-order result takes over from the resummed terms we see the prediction overshoot the data -- see also the discussion below from Ref.~\cite{rick}.}.
\begin{figure}[h]
\begin{center}
\setlength{\unitlength}{0.1mm}
\begin{picture}(1600, 930)
\put( 810, 920){\makebox(0,0)[cb]{\bf\Large $Z/\gamma^*$ transverse momentum $\left(d\sigma/d\phi^*_\eta(\ell\ell)\right)$} }
\put( 40, -20){\makebox(0,0)[lb]{\includegraphics[width=150mm]{fig18-atlas-v1a.pdf}}}
\end{picture}
\end{center}
\caption{\baselineskip=11pt Comparisons of some theoretical predictions with the ATLAS $Z/\gamma^*$ $\phi^*_\eta$ spectrum in single $Z/\gamma^*$ production with decay to lepton pairs as given in Ref.~\cite{atlas-hassani}. Here Banfi {\em et al.} refers also to a resummed calculation of the ``CSS'' type~\cite{css}, so that it has the same physical precision limitations as RESBOS~\cite{resbos1,resbos2,resbos3} as discussed in Ref.~\cite{1305-0023} -- see the second reference in Refs.~\cite{phietastr}. }
\label{atlas-hass-1}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=100mm]{yin-1.pdf}
\end{center}
\caption{\baselineskip=11pt Comparisons of theoretical predictions with the CDF $Z/\gamma^*$ $p_T$ spectrum in single $Z/\gamma^*$ production with decay to lepton pairs as given in Ref.~\cite{yin}.}
\label{yin-1}
\end{figure}
The ``soft'' limit of the prediction has no reasonable relation to that in the data and even at $p_T \cong 20 GeV$ the so-called exact NNLO result is just useless. Obviously, it will make no sense to talk about the precision tag of such results! Indeed, one of the consequences of the discrepancy in Figs.~\ref{atlas-hass-1},~\ref{yin-1}
is that, until one understands how to fix it, one cannot even be sure of the normalization one gets when one integrates over the theoretical prediction itself.
Indeed, in the precision theory developed and implemented~\cite{yfs-jw} for the LEP physics program, it was in fact true that ``fixing'' such discrepancies changed the normalization.
\par
More precisely, in Refs.~\cite{1305-0023,herwiri}, we have shown
that the current
realization of the exact amplitude-based resummation
approach to precision LHC physics as effected by the
implementation of the IR-improved DGLAP-CS~\cite{dglap,cs} theory~\cite{irdglap1,irdglap2} via HERWIRI1.031~\cite{herwiri}
in the HERWIG6.5~\cite{herwig} environment improves the agreement between
LHC data on single $Z/\gamma^*$ production in comparison to
the un-improved predictions. This prepares the stage naturally for
setting baselines for the respective theoretical precision tags especially
when we focus on the NLO exact, matrix element matched parton shower MC precision issues involved in comparing the predictions of MC@NLO/HERWIRI1.031 and MC@NLO/HERWIG6.5 in the MC@NLO~\cite{mcatnlo} methodology. Here, we define
MC@NLO/A to be the NLO exact, matrix element matched parton shower MC realization of MC A in the MC@NLO methodology. When we try to address these issues,
we are faced with determining the precision of the respective NLO exact matrix element prediction. This latter issue then brings us to the NLO
version of the type of behavior discussed above for Figs.~\ref{atlas-hass-1},~\ref{yin-1}, which is illustrated in Fig.~\ref{fig-rick}
\begin{figure}[h]
\begin{center}
\includegraphics[width=80mm]{rick-2.pdf}
\end{center}
\caption{\baselineskip=11pt The ratio of the u-quark probability distribution
defined in the Drell-Yan process to that defined from the ${\rm F}_2$ structure function defined in deep inelastic lepton-nucleon scattering as discussed in Ref.~\cite{rick} at $Q= 10$ GeV, where the solid (dashed) curve corresponds to including (excluding) the total cross section in the attendant Drell-Yan distribution.}
\label{fig-rick}
\end{figure}
and discussed at length in Ref.~\cite{rick} where in its eq.(5.5.30) it is shown that
\begin{equation}
\frac{G^{DY}_{p\rightarrow q}(x,Q^2)}{G_{p\rightarrow q}(x,Q^2)}\operatornamewithlimits{\rightarrow}_{x\rightarrow 1}1+\frac{2\alpha_s(Q^2)}{3\pi}\ln^2(1-x)
\label{rickeq1}
\end{equation}
where $G^{DY}_{p\rightarrow q}$ ($G_{p\rightarrow q}$) is the respective Drell-Yan(DIS) structure function~\cite{rick} in a standard type of notation.
No observable data, at LHC or the new FCC,
can have this behavior and it calls into question what a precision tag could even mean here?
\par
With an eye toward ``taming'' this what we see in Fig.~\ref{fig-rick} and (\ref{rickeq1}), we revisit our master formula from Refs.~\cite{qced} for our $QED\otimes QCD$ exact resummation theory
\begin{eqnarray}
&d\bar\sigma_{\rm res} = e^{\rm SUM_{IR}(QCED)}
\sum_{{n,m}=0}^\infty\frac{1}{n!m!}\int\prod_{j_1=1}^n\frac{d^3k_{j_1}}{k_{j_1}} \cr
&\prod_{j_2=1}^m\frac{d^3{k'}_{j_2}}{{k'}_{j_2}}
\int\frac{d^4y}{(2\pi)^4}e^{iy\cdot(p_1+q_1-p_2-q_2-\sum k_{j_1}-\sum {k'}_{j_2})+
D_{\rm QCED}} \cr
&\tilde{\bar\beta}_{n,m}(k_1,\ldots,k_n;k'_1,\ldots,k'_m)\frac{d^3p_2}{p_2^{\,0}}\frac{d^3q_2}{q_2^{\,0}},
\label{subp15b}
\end{eqnarray}\noindent
where $d\bar\sigma_{\rm res}$ is either the reduced cross section
$d\hat\sigma_{\rm res}$ or the differential rate associated to a
DGLAP-CS~\cite{dglap,cs} kernel involved in the evolution of PDF's and
where the {\em new} (YFS-style~\cite{yfs-jw,yfs}) {\em non-Abelian} residuals
$\tilde{\bar\beta}_{n,m}(k_1,\ldots,k_n;k'_1,\ldots,k'_m)$ have $n$ hard gluons and $m$ hard photons and we show the final state with two hard final
partons with momenta $p_2,\; q_2$ specified for a generic $2f$ final state for
definiteness. The infrared functions ${\rm SUM_{IR}(QCED)},\; D_{\rm QCED}\; $ are
defined in Refs.~\cite{qced,irdglap1,irdglap2} as follows:
\begin{eqnarray}
{\rm SUM_{IR}(QCED)}=2\alpha_s\Re B^{nls}_{QCED}+2\alpha_s{\tilde B}^{nls}_{QCED}\cr
D_{\rm QCED}=\int \frac{d^3k}{k^0}\left(e^{-iky}-\theta(K_{max}-k^0)\right){\tilde S}^{nls}_{QCED}
\label{irfns}
\end{eqnarray}
where the dummy parameter $K_{max}$ is such that nothing depends on it.
We have introduced
\begin{eqnarray}
B^{nls}_{QCED} \equiv B^{nls}_{QCD}+\frac{\alpha}{\alpha_s}B^{nls}_{QED},\cr
{\tilde B}^{nls}_{QCED}\equiv {\tilde B}^{nls}_{QCD}+\frac{\alpha}{\alpha_s}{\tilde B}^{nls}_{QED}, \cr
{\tilde S}^{nls}_{QCED}\equiv {\tilde S}^{nls}_{QCD}+{\tilde S}^{nls}_{QED}.
\end{eqnarray}
The DGLAP-CS synthesization of the infrared functions is denoted
by the superscript $nls$ as explained in Refs.~\cite{dglpsyn,qced,irdglap1,irdglap2} while the infrared functions
$B_A,\; {\tilde B}_A,\; {\tilde S}_A, \; A=QCD,\; QED,$ are given
in Refs.~\cite{yfs-jw,yfs,qced,irdglap1,irdglap2}.
The exactness of the
simultaneous resummation of QED and QCD large IR effects that we show here
cannot be emphasized too much.
\par
In the interest of pedagogy, we note that,
in the language of Ref.~\cite{gatheral},
the exponent ${\rm SUM_{IR}(QCED)}$ sums up to the infinite order the maximal leading IR singular terms in the cross section for soft emission
below a dummy parameter $K_{\text{max}}$ and the exponent
$D_{\rm QCED}$ does the same for the regime above $K_{\text{max}}$ so that
(\ref{subp15b}) is independent of $K_{\text{max}}$\footnote{If we want to include more of the
maximal exponentiating terms from the formalism of Ref.~\cite{gatheral} in
the two exponents ${\rm SUM_{IR}(QCED)},\;D_{\rm QCED}$, we may do so with a consequent change in the attendant residuals $\tilde{\bar\beta}_{n,m}$.}.
Exactness order by order in perturbation theory in both
$\alpha$ and $\alpha_s$ in the presence of these resummed terms, as explained
in Refs.~\cite{qced,irdglap1,irdglap2}, is maintained by iterative
computation of the residuals $\tilde{\bar\beta}_{n,m}$ to match the attendant
exact results to all orders in $\alpha$ and $\alpha_s$. In particular, in
our formulation in (\ref{subp15b})
{\it the entire soft gluon phase space is included in the representation -- no part of it
is dropped}. As it is shown in Refs.~\cite{qced},
the new non-Abelian residuals $\tilde{\bar\beta}_{m,n}$
carry a realization
of rigorous shower/ME matching via their shower subtracted analogs:
in (\ref{subp15b}) we make the replacements
\begin{equation}
\tilde{\bar\beta}_{n,m}\rightarrow \hat{\tilde{\bar\beta}}_{n,m}
\end{equation}
where the $\hat{\tilde{\bar\beta}}_{n,m}$ have had all effects in the showers
associated to the attendant PDF's $\{F_j\}$ removed from them. Here
we have in mind the standard formula for the
fully differential representation of a hard LHC(FCC) scattering process:
\begin{equation}
d\sigma =\sum_{i,j}\int dx_1dx_2F_i(x_1)F_j(x_2)d\hat\sigma_{\text{res}}(x_1x_2s),
\label{bscfrla}
\end{equation}
where
$d\hat\sigma_{\text{res}}$ is given in (\ref{subp15b})
and thus is consistent~\cite{radcor2013,1305-0023,herwiri,qced}
with our achieving a total precision tag of 1\% or better for the total
theoretical precision of (\ref{bscfrla}).
\par
For completeness, we also recall the connection between our constructs in the master formula (\ref{subp15b}) and the constructs in the MC@NLO methodology:
We may represent the MC@NLO differential cross section
via~\cite{mcatnlo}
\begin{equation}
\begin{split}
d\sigma_{MC@NLO}&=\left[B+V+\int(R_{MC}-C)d\Phi_R\right]d\Phi_B[\Delta_{MC}(0)+\int(R_{MC}/B)\Delta_{MC}(k_T)d\Phi_R]\\
&\qquad\qquad +(R-R_{MC})\Delta_{MC}(k_T)d\Phi_Bd\Phi_R
\label{mcatnlo1}
\end{split}
\end{equation}
where $B$ is Born distribution, $V$ is the regularized virtual contribution,
$C$ is the corresponding counter-term required at exact NLO, $R$ is the respective
exact real emission distribution for exact NLO, $R_{MC}=R_{MC}(P_{AB})$ is the parton shower real emission distribution
so that the Sudakov form factor is
$$\Delta_{MC}(p_T)=e^{[-\int d\Phi_R \frac{R_{MC}(\Phi_B,\Phi_R)}{B}\theta(k_T(\Phi_B,\Phi_R)-p_T)]},$$
where as usual it describes the respective no-emission probability.
The respective Born and real emission differential phase spaces are denoted by $d\Phi_A, \; A=B,\; R$.
We find it very important {\em still}
to emphasize that the representation of the differential distribution
for MC@NLO in (\ref{mcatnlo1}) illustrates the compensation
between real and virtual divergent soft effects discussed in the
Appendices of Refs.~\cite{irdglap1,irdglap2} in establishing the validity of
(\ref{subp15b}) for QCD. More specifically,
from comparison with (\ref{subp15b}) restricted to its QCD aspect we get the identifications, accurate to ${\cal O}(\alpha_s)$,
\begin{equation}
\begin{split}
\frac{1}{2}\hat{\tilde{\bar\beta}}_{0,0}&= \bar{B}+(\bar{B}/\Delta_{MC}(0))\int(R_{MC}/B)\Delta_{MC}(k_T)d\Phi_R\\
\frac{1}{2}\hat{\tilde{\bar\beta}}_{1,0}&= R-R_{MC}-B\tilde{S}_{QCD}
\label{eq-mcnlo}
\end{split}
\end{equation}
where we defined~\cite{mcatnlo} $$\bar{B}=B(1-2\alpha_s\Re{B_{QCD}})+V+\int(R_{MC}-C)d\Phi_R$$ and we understand
that the DGLAP-CS kernels in $R_{MC}$ are to be taken as the IR-improved ones
as we derived in Refs.~\cite{irdglap1,irdglap2}.
Although we have suppressed the superscript $nls$
for simplicity of notation, to avoid double counting of effects the QCD virtual and real infrared functions
$B_{QCD}$ and $\tilde{S}_{QCD}$ are understood to be DGLAP-CS synthesized as explained in Refs.~\cite{qced,irdglap1,irdglap2}. Most importantly, in view of
(\ref{eq-mcnlo}), we observe that
the way to the extension of frameworks such as MC@NLO to exact higher
orders in $\{\alpha_s,\;\alpha\}$ is open via our $\hat{\tilde{\bar\beta}}_{n,m}$
and will be taken up elsewhere~\cite{elswh}.
\par
We see from the relationship between the hard gluon residuals and the
exact NLO corrections that a serious study of the theoretical precision
of (\ref{bscfrla}) when it uses the results of (\ref{subp15b}), such as it
is done in Refs.~\cite{herwiri},
necessarily involves the studying of the theoretical
precision of these exact NLO results and if we have the behavior in
(\ref{rickeq1}) we do have to ask what would such a study mean in relation
to LHC (or FCC) data? To address this question, we proceed as follows.\par
We recall the well-known representation of the exact NLO
differential cross section for the Drell-Yan process (we focus on
the $\gamma^*$ part of the $Z/\gamma^*$ exchange for simplicity of
presentation, as adding in the effect of the
$Z$ is straightforward and does not affect the analysis here in any
essential way; similarly, we treat the simple case of one flavor with
unit charge following Ref.~\cite{guido-mart,humpvn} for the same reason --
inserting the proper charges and sums is trivial)
\begin{equation}
\begin{split}
\frac{d\sigma^{DY}}{dQ^2}&=\frac{4\pi\alpha^2}{9sQ^2}\int_0^1\frac{dx_1}{x_1}\int_0^1\frac{dx_2}{x_2}\big\{\left[q^{(1)}(x_1)\bar{q}^{(2)}(x_2)+(1\leftrightarrow 2)\right]\big[\delta(1-z_{12})\\
&\quad + \alpha_s(t)\theta(1-z_{12})(\frac{1}{2\pi}P_{qq}(z_{12})(2t)+f^{DY}_q(z_{12}))\big]\\
&\quad +\left[(q^{(1)}(x_1)+\bar{q}^{(1)}(x_1))G^{(2)}(x_2)+(1\leftrightarrow 2)\right] \\
&\quad \times [\alpha_s(t)\theta(1-z_{12})(\frac{1}{2\pi}P_{qG}(z_{12})t+f^{DY}_G(z_{12}))]\big\}
\end{split}
\label{guido-eq2}
\end{equation}
where $z_{12}=\tau/(x_1x_2), \; \tau=Q^2/s$ in the usual conventions~\cite{guido-mart,rick,humpvn}, the labels $1$ and $2$ refer to the two respective incoming protons
and we follow the generic notation of Refs. ~\cite{rick,guido-mart} here.
The unimproved DGLAP-CS~\cite{dglap,cs} kernels in (\ref{guido-eq2}) are
well-known as
\begin{align}
P_{qq}(z)&= C_F \left[\frac{1+z^2}{(1-z)_+} + \frac{3}{2}\delta(1-z)\right],\nonumber\\
P_{qG}(z)&=\frac{1}{2}(z^2+(1-z)^2),
\label{guido-eq3}
\end{align}
where we define $t=\ln(Q^2/\mu^2)$ following Refs.\cite{guido-mart,rick} so that
$\mu$ is the 't Hooft~\cite{thft-mass} unity of mass.
The scheme dependent hard correction terms are given as follows~\cite{humpvn,guido-mart} if one uses massless quarks and gluons and dimensional regularization, for example:
\begin{equation}
\begin{split}
\alpha_s f^{DY}_G(z)&=\frac{\alpha_s}{2\pi}\frac{1}{2}[(z^2+(1-z)^2)\ln\frac{(1-z)^2}{z}-\frac{3}{2}z^2+z+\frac{3}{2}+2P_{qG}(z)\zeta]\\
\alpha_s f^{DY}_q(z)&=C_F\frac{\alpha_s}{2\pi}\Big[4(1+z^2)\left(\frac{\ln(1-z)}{1-z}\right)_+ -2\frac{1+z^2}{1-z}\ln{z}\\
& \quad+\left(\frac{2\pi^2}{3}-8\right)\delta(1-z)+\frac{2}{C_F}P_{qq}(z)\zeta\Big]
\end{split}
\label{guido-eq4}
\end{equation}
where we define~\cite{humpvn} $\zeta=-\frac{1}{\epsilon}+C_E-\ln{4\pi}$ for
$\epsilon=2-n/2$ when $n$ is the dimension of space-time. $C_E$ is Euler-Mascheroni constant.
In the $\overline{\text{MS}}$ scheme, the terms proportional to $\zeta$ are removed
by mass factorization, which also replaces $\mu$ by $\Lambda$ in $t$ following Ref.~\cite{rick}. This leaves the +-functions in the hard corrections and it is the divergent behavior of these distributions as $z\rightarrow 1$ that produces the attendant unphysical results referenced above. How can we fix this?
\par
We imbed the calculation of the hard correction terms into the master formula
(\ref{subp15b}) restricted to its QCD aspect. This gives the following resummed
version of (\ref{guido-eq2}):
\begin{equation}
\begin{split}
\frac{d\sigma^{DY}_{res}}{dQ^2}&=\frac{4\pi\alpha^2}{9sQ^2}\int_0^1\frac{dx_1}{x_1}\int_0^1\frac{dx_2}{x_2}\big\{\left[q^{(1)}(x_1)\bar{q}^{(2)}(x_2)+(1\leftrightarrow 2)\right]2\gamma_qF_{YFS}(2\gamma_q)(1-z_{12})^{2\gamma_q-1}e^{\delta_q}\\
&\quad \times \theta(1-z_{12})\big[ 1+\gamma_q -7C_F\frac{\alpha_s}{2\pi}+ (1-z_{12})(-1+\frac{1-z_{12}}{2})\\
&\quad +2\gamma_q(-\frac{1-z_{12}}{2}-\frac{z_{12}^2}{4}\ln{z_{12}})\\
&\quad + \alpha_s(t)\frac{(1-z_{12})}{2\gamma_q}f^{DY}_q(z_{12})\big]\\
&\quad +\left[(q^{(1)}(x_1)+\bar{q}^{(1)}(x_1))G^{(2)}(x_2)+(1\leftrightarrow 2)\right] \\
&\quad \times \gamma_G F_{YFS}(\gamma_G)e^{\frac{\delta_G}{2}}[\alpha_s(t)\theta(1-z_{12})\big(\frac{t}{2\pi\gamma_G}(\frac{1}{2}(z_{12}^2(1-z_{12})^{\gamma_G}+(1-z_{12})^2z_{12}^{\gamma_G}))\\
&\quad +f^{DY'}_G(z_{12})/\gamma_G\big)]\big\}
\end{split}
\label{guido-eq5}
\end{equation}
where we have introduced here
\begin{equation}
\begin{split}
\alpha_s f^{DY'}_G(z)&=\frac{\alpha_s}{2\pi}\frac{1}{2}[(z^2(1-z)^{\gamma_G}+(1-z)^2z^{\gamma_G})\ln\frac{(1-z)^2}{z}-\frac{3}{2}z^2(1-z)^{\gamma_G}+z(1-z)^{\gamma_G}\\
&\quad +\frac{3}{4}((1-z)^{\gamma_G}+z^{\gamma_G})],\\
\end{split}
\label{guido-eq6}
\end{equation}
and the following exponents and YFS infrared function, $F_{\mathrm{YFS}}$, already needed for the IR-improvement of DGLAP-CS theory in Refs.~\cite{irdglap1,irdglap2}:
\begin{align}
\gamma_q &= C_F\frac{\alpha_s}{\pi}t=\frac{4C_F}{\beta_0}, \qquad \qquad
\delta_q =\frac{\gamma_q}{2}+\frac{\alpha_sC_F}{\pi}(\frac{\pi^2}{3}-\frac{1}{2}),\nonumber\\
\gamma_G &= C_G\frac{\alpha_s}{\pi}t=\frac{4C_G}{\beta_0}, \qquad \qquad
\delta_G =\frac{\gamma_G}{2}+\frac{\alpha_sC_G}{\pi}(\frac{\pi^2}{3}-\frac{1}{2}),\nonumber\\
F_{\mathrm{YFS}}(\gamma)&=\frac{e^{-{C_E}\gamma}}{\Gamma(1+\gamma)}.
\label{resfn1}
\end{align}
We define $\beta_0=11-\frac{2}{3}n_f$ for $n_f$ active flavors
in a standard way and $\Gamma(w)$ is Euler's gamma function of the complex variable $w$.
Note that we have mass factorized in (\ref{guido-eq5}) and (\ref{guido-eq6})
as indicated above.
It can be seen immediately that the regime at $z_{12}\rightarrow 1$ is now under control in (\ref{guido-eq5}) so that we will no longer have the unphysical
behavior discussed above. This is the main result of this paper.\par
Specifically, instead of the result in (\ref{rickeq1}), we now get the
behavior such that the $\ln^2(1-x)$ on the RHS of (\ref{rickeq1}) is replaced by
$$\frac{2(1-x)^{\gamma_q}\ln(1-x)}{\gamma_q} - \frac{2(1-x)^{\gamma_q}}{\gamma_q^2},$$
and this vanishes for $x\rightarrow 1$. What our result
means that the hard correction now has the possibility to be compared
{\em exclusively} to the data in a rigorously meaningful way. We take up such matters elsewhere.~\cite{elswh}.
\par
We stress that the parton shower/ME matching formulas in MC@NLO (shown above) and in POWHEG~\cite{powheg} do not remove the IR divergence which we just tamed, as the latter retains the NLO correction with its bad IR limit in the soft regime for $z_{12} \rightarrow 1$ and the former replaces the bad IR behavior of the NLO correction in the soft $z_{12} \rightarrow 1$ limit with that of the parton shower real emission at the same order and it is well known that the respective unimproved parton shower real emission is infrared divergent for $z_{12} \rightarrow 1$ and requires an ad hoc IR cut-off $k_0$-parameter, as we have discussed in Ref.~\cite{herwiri}. {\em No such parameter is needed in our new approach.}\par
To sum up, we have introduced a new approach to hard corrections in perturbative QCD that will allow us to establish the same type of semi-analytical baselines for QCD that we had in Refs.~\cite{yfs-jw} for the higher order
corrections in the
Standard Model EW theory.
We look forward to its exploitation in precision LHC and FCC physics scenarios.
In closing, we
thank Prof. Ignatios Antoniadis for the support and kind
hospitality of the CERN TH Unit while part of this work was completed.\par
|
\section{Introduction}
The Anfinsen's paradigm states that the native structure of a protein is encoded in its primary structure \cite{Tanuichi}.
This paradigm comes from in vitro unfolding/refolding experiments. While certainly Anfinsen’s paradigm is valid in principle, it is severely hampered in the actual cell environment.
As a matter of fact, many protein species require the intervention of a variety of ``chaperone'' proteins in order to undergo a correct folding process, ending up into the correct soluble native structure.
Protein folding competes with intermolecular aggregation that is driven by very similar forces but ends up into a very different final state: instead of producing soluble and independent molecules, aggregation provokes the formation of insoluble bodies made by many molecules \cite{Agostini2012237}.
The balance between ``within-molecule-interaction'' and ``between-molecules-interactions'' shifts the equilibrium towards correct folding (prevalence of within-molecule-interactions, soluble systems) or aggregation (prevalence of between-molecules-interactions, insoluble systems).
This balance can be so subtle that in some cases a single point mutation \cite{doi:10.1021/ja1116233} is able to provoke a dramatic shift of this equilibrium. It is important to stress that, despite the presence of chaperones, a certain level of aggregation is present in the cells for any protein system, moreover the cell chemico-physical microenvironment (pH, crowding, etc.) has a crucial importance in determining the folding/aggregation propensity of different protein systems.
The existence of several diseases provoked by misfolding and consequent aggregation of protein molecules \cite{doi:10.1146/annurev.biochem.75.101304.123901} gives the elucidation of folding/aggregation balance a great applicative importance.
In their 2009 paper appeared on PNAS, Niwa and colleagues \cite{niwa2009} produced the most precise and pure data base upon which test different explanatory models of folding/aggregation determinants.
At odds with other studies, the authors generated a fair test bench in which the complete ORF library (Open Reading Frames; they correspond to the structural genes of a given genome, i.e., the set of the sequences that are translated into protein molecules) of E. coli was translated in a pure, cell-free and thus chaperone-free, system under the same conditions.
This allows to get data on around 70\% of the E. coli proteins and can be considered as the best experimental framework on which test the intrinsic propensity of a given sequence to go toward the aggregation or correct folding pole.
This is a very crucial point that deserves further explanation. In general, solubility cannot be considered as an intrinsic property of the solute molecule, given that it depends from the solvent and chemico-physical microenvironment features, like temperature, presence of other molecules, pH, ionic strength etc.
The Niwa et al. data set, in which microenvironment is kept invariant, allows to concentrate solely on solute (protein molecule) properties, so making it possible to devise a meaningful pattern recognition strategy.
As a matter of fact, any protein whose 3D structure is known is amenable to be solubilized in some way, but this fact does not imply that can be solubilized by the Niwa et al. minimalistic recipe.
Strictly speaking, \citet{niwa2009} observed the relative solubility of different protein species; only as a second step not correctly folded molecules precipitate and aggregate. In a recent paper, \citet{Agostini2012237} clearly demonstrated that the solubility degree in the Niwa et al. data base negatively correlates with the aggregation propensity, when the aggregation is estimated from the folded state.
This implies that we can safely consider the solubility as a measure of the relative stability of the folded and aggregated states \cite{Agostini2012237}.
The authors observed a bi-modal distribution as for the relative solubility of the 3173 protein species assayed \cite{niwa2009} consistent with the presence of two classes of molecules: proteins not necessitating any chaperone to reach the correct stable folding and proteins that need chaperone intervention in order to avoid aggregation.
This bi-modal character is far to be perfect and there are large areas in between the two peaks confirming the fuzzy boundaries between the two folding behaviors.
The authors describe some general properties showing statistically significant trends between the two groups of soluble and insoluble proteins but they were not able to predict folding/aggregation propensity on a single molecule scale using both sequence and final 3D structure based algorithms.
Thus the authors conclude that proteins more prone to aggregate are bigger than more soluble ones, have a lower isoelectric point, and a higher content of negatively charged residues \cite{niwa2009}.
On the structural side, all-$\beta$ class was more frequent in the low-solubility population like some specific SCOP folds, such as c-94 or c-67. However, the authors were not able to go beyond a population-based statistical description.
On the other hand, \citet{Agostini2012237} obtained a good prediction of actual relative solubility of proteins by a Support Vector Machine (SVM) based approach as applied to chemico-physical representations of sequences, hence giving a proof of concept of the possibility to express solubility/aggregation balance in terms of chemico-physical properties of amino acid residues.
Instead of having as main goal the one of obtaining the best correlation between continuous solubility and chemico-physical description of sequences, in this paper we compared alternative representations of the E. coli proteins as for their ability to discriminate between spontaneous foldable and chaperone-needing proteins.
We analyzed the Niwa et al. dataset by considering pattern recognition methods operating on both sequence and graph based pattern representations.
Departing from the more conventional vector-based representation of patterns implies the adoption data-driven inference mechanisms that able to operate in the so-called non-geometric input spaces \cite{si_asoc_grc}. On the other hand, such types of representations offer also more interpretable information that could be exploited to infer knowledge from the data-driven process itself.
Experimental results show the possibility to reach good accuracy in the prediction of folding/aggregation propensity of single protein systems.
This constitutes a neat progress with respect to the statistical evidence produced by the authors \cite{niwa2009}, which allows us to sketch some general hypotheses on folding process based on some features of the adopted modeling and computational approaches.
Comparing the biophysical/biochemical premises of successful vs. unsuccessful predictions, we were able to single out some important general properties of protein architectural and dynamical features.
Notably, (i) we suggest a threshold around 250 residues discriminating ``easily foldable'' from ``hardly foldable'' molecules consistent with other independent experiments, and (ii) we highlight the relevance of contact graph spectra for folding behavior discrimination and characterization of the E. coli solubility data.
The contact graph spectral description, which we have found to be suitable for this specific problem, can be generalized to other network classification problems and considering different structure-function relationships.
The remainder of the paper is structured as follows. In Section \ref{sec:material} we first present the considered datasets (Section \ref{sec:dataset}) corresponding to alternative protein molecule representations, which we elaborated starting from the original data of \citet{niwa2009}. Then we introduce the methodological basis underlying the particular classification systems used for processing such data (Section \ref{sec:class}).
In Section \ref{sec:complexity}, we describe the theoretical frameworks that we used to experimentally evaluate the structural complexity of the graphs representing the E. coli proteins.
In Section \ref{sec:results}, we present and discuss the experimental results, which include: (i) a statistical analysis demonstrating the consistency and soundness of our results, (ii) the detailed test set classification accuracy percentages obtained on the various datasets, and finally (iii) a preliminary analysis focusing on the study of the structural complexity of the developed protein graph representations.
Finally, Section \ref{sec:conclusions} concludes the paper pointing at future directions.
\section{Materials and methods}
\label{sec:material}
\subsection{Elaboration of datasets}
\label{sec:dataset}
Niwa and colleagues \cite{niwa2009} demonstrated a bi-modal distribution of solubility with a class of largely insoluble and a class of very soluble proteins.
In Figure \ref{fig:solubility_density}, we report the plot showing the protein solubility density according to the normalized solubility degree; each protein solubility value has been normalized by considering the maximum solubility degree in the dataset.
This unbalanced bi-modal character of the solubility makes difficult the selection of a universally valid threshold partitioning proteins into soluble and not soluble hard categories.
In fact, considering $[0, 0.3]$ and $[0.7, 1]$ as the two intervals characterizing the insoluble and soluble proteins, the original dataset would split into 1631 insoluble and 180 soluble proteins, respectively.
The original data is provided in terms of a sequence-based representation of the proteins, which basically includes only the symbolic information of the residues (i.e., a character that identifies an amino acid).
\begin{figure}[ht!]
\centering
\includegraphics[viewport=0 0 341 243,scale=1,keepaspectratio=true]{./density_all}
\caption{Density of the normalized solubility degree of proteins.}
\label{fig:solubility_density}
\end{figure}
The experiments that we conducted have been organized by considering two different sub-samplings of the original dataset. The first sub-sampling considers a perfectly balanced dataset made of the 100 proteins with the highest solubility degree and the 100 proteins with the lowest solubility degree.
Such a dataset, which is evidently a ``simplification'' of the original problem, has been analyzed in our previous work \cite{grapsec_ijcnn_2013}, where we demonstrated the possibility to achieve a very good classification accuracy.
The second dataset sub-sampling, which we elaborate here in this paper, determines the insoluble and soluble classes by taking the proteins falling in the $[0, 0.3]$ and $[0.7, 1]$ ranges, respectively.
The solubility degree ranges have been considered of the same interval length, by placing such intervals on the extremes of the normalized solubility degree range.
This dataset has been split into a training set and a test set.
The training set contains 180 proteins, 70 of which belong to the soluble class and the remaining 110 to the insoluble class. The test set is considerably larger, since in fact it contains 1631 proteins, of which 1521 are insoluble proteins.
Consequently, this dataset (in the following denoted as DS-1811) is largely unbalanced, especially for what concerns the test set.
The training set, although it is much smaller than the test set, it is conceived to ``cover'' the considered data instance as much as possible with \textit{characterizing} proteins, that is, with proteins that suitably represent the soluble or insoluble prototypical patterns. Each protein in DS-1811 is represented by a sequence of symbols (that is characters) associated with amino acids.
As mentioned before, our first goal is to show the possibility to achieve a good predictive ability on the sole basis of pure sequence symbolic information.
Successively, we project the solution -- in terms of substitution matrix between amino acid residues and sequence matching techniques -- on a chemico-physical space, in which the residues are encoded by the principal component scores derived by a large collection of well-known chemico-physical descriptors \cite{apdbase}.
The sequence dissimilarity measure between proteins is assessed by means of a Levenshtein-type metric.
The Levenshtein distance is a string metric that calculates the distance among two strings as the minimum cost of alignment \cite{Deza.Deza2009EncyclopediaofDistances}.
The alignment procedure is a computation that operates globally by means of the so-called edit operations of insertion, deletion, and substitution, quantifying the cost involved in transforming a string into the other.
The algorithm is usually implemented in the well-known dynamic programming framework, resulting thus in a quadratic computational complexity.
The Levenshtein distance has been originally conceived to process sequences of symbols (i.e. strings). However, it can be generalized to sequences of general objects for which it is possible to define a suitable distance measure in the domain of those objects \cite{t2vsdiss__ifsanafips2013,t2apdiss__ifsanafips2013}.
In the general case, the edit costs used to calculate the overall minimum-cost of alignment among sequences are obtained by evaluating such a distance between the objects of the sequences.
It is worth noting that the Levenshtein distance, at odds with other metrics used in protein sequences comparison, is not limited to homologue comparisons.
Successively, we tried to solve the prediction task on the structural side, by representing the proteins as graphs.
At odds with Niwa et al., we made use of a topological enriched representation in which the vertices of a graph are labeled by the three main chemico-physical components coming from a collection of amino acid chemico-physical features (described in detail in Sec. \ref{sec:cca}).
The prediction-by-structure could only be achieved for proteins whose 3D structure is known and present in the PDB database \cite{pdb}.
As a consequence, we were able to construct a dataset made by the 454 contact graphs that were available from PDB files.
We will refer to such dataset of labeled graphs as DS-G-454.
A contact graph is constructed by considering as vertices the centres of mass of each residue (i.e., their alpha carbon atoms). Edges are generated by connecting residues at a distance in between 4 and 8 \AA{}; this choice was dictated by the need to eliminate trivial contacts due to neighboring residues along the sequence while focusing on the elective contacts characterizing the native structure \cite{doi:10.1021/cr3002356}.
Each graph representing a (folded) protein is enriched by considering suitable attributes for both vertices and edges: such types of graphs are termed labeled (attributed) graphs in the pattern recognition literature \cite{gm_survey,gralg_2012,odse,odse2_ijcnn_2013,foggia2012graph}.
The vertices of the graphs are characterized with the three aforementioned chemico-physical components.
Edges are labeled by using the Euclidean distance among the residues in the 3D space. Such a characterization allows us to include in the graph representation both topological and chemico-physical information regarding proteins.
DS-G-454 is constituted by 377 insoluble and 77 soluble proteins; 27 soluble and 132 insoluble proteins have been randomly considered for the test set, while the remaining proteins are used for training.
It is worth noting that, although even in this data set we have a prevalence of insoluble proteins, the obvious fact that in order to get a reliable 3D structure from X-ray or NMR, the molecule must be solubilized. The point is that here we are not studying ``solubility as such'', which in turn is a devoid of chemico-physical sense concept, but the solubility of the molecule in the Niwa et al. specific experimental conditions, which are different from those adopted (on a case-by-case basis) for structural analysis.
By considering only the proteins in DS-G-454 represented as graphs, we have generated another dataset instance containing the related 454 sequences -- thus it is a subset of DS-1811.
We denoted this dataset DS-454. Such a dataset has been considered with the aim of evaluating the effect of sub-sampling DS-1811 from the classification performance viewpoint. DS-454 has been divided into training and test set by considering the same split percentages used for DS-G-454.
The last dataset that we considered is obtained as a direct elaboration of DS-G-454.
Notably, each graph has been mapped into a sequence of vertex attributes according to the procedure of graph serialization that we explain in the following.
By elaborating the whole dataset of graphs, we obtained another dataset of 454 sequences of three-dimensional, real-valued vectors (i.e., the three aforementioned components). We denoted such a dataset DS-S-454.
A sequence of vertices related to a graph is obtained by using the so-called seriation algorithm discussed in Ref. \cite{livi+delvescovo+rizzi_seriation+gradis,seriation+gradis_lncs_2012}.
A graph $G=(\mathcal{V}, \mathcal{E}), |\mathcal{V}|=n, $ where $\mathcal{V}$ is the set of vertices and $\mathcal{E}$ is the set of edges, can be represented according to its transition matrix, $\mathbf{T}^{n\times n}$, which is given as $\mathbf{T}=\mathbf{D}^{-1}\mathbf{A}$. The matrix $\mathbf{D}$ is diagonal and contains the vertex degrees (i.e., number of incident edges), while $\mathbf{A}$ is the adjacency (i.e., the contact) matrix of $G$.
To include the information about the edges in $\mathbf{T}$, we consider the degree of a vertex to be weighted by the quantity stored in the edge attributes (i.e., the atomic distances). The seriation of the graph is then implemented by analyzing the first eigenvector of $\mathbf{T}$ (i.e., the eigenvector associated to the largest, in magnitude, eigenvalue, i.e., 1 in this case), which corresponds to the stationary distribution, $\pi$, of the Markovian random walk on $G$ \cite{Lovasz1996}.
$\mathbf{T}$ may not be symmetric, hence it may not allow a spectral decomposition. However, $\mathbf{T}$ can be symmetrized by computing $\mathbf{D}^{-1/2}\mathbf{T}\mathbf{D}^{-1/2}$, with which shares the same eigenvectors.
Interestingly, the first eigenvector of the contact matrix is known to provide a consistent description of proteins \cite{PROT:PROT22113}, and so is the principal eigenvector of the transition matrix, although it should be interpreted in the random walk setting.
This further sequence representation of proteins allows us to deal with the recognition problem on the ``less complex domain'' of sequences, while at the same time it includes the topological information of the graph representations as a byproduct of the seriation. Training and test set splits are compatible with those of DS-G-454.
To summarize, in this study we considered four datasets related to the same initial set of proteins, namely DS-1811, DS-454, DS-G-454, and DS-S-454.
DS-1811 and DS-454 contain sequences of characters denoting the usual amino acid identifiers.
DS-G-454 is a dataset of labeled graphs describing the proteins according to their 3D native structure, including also suitable vertex--edge auxiliary information.
Finally, DS-S-454 is a direct elaboration of DS-G-454, containing sequences of three-dimensional, real-valued vectors, which are a characterization of the graphs according to the topological information provided by the corresponding transition matrices.
\subsection{The considered pattern recognition systems}
\label{sec:class}
Each of the aforementioned datasets has been processed with an ad hoc similarity-based classification system. Design of dissimilarity and similarity measures is of utmost importance in pattern recognition systems, since they effectively induce the underlying algebraic structure of the input space \cite{gm_survey}.
Kernel methods are an important class of pattern recognition systems, which are based on the definition of a (positive semi-definite) kernel function, $k(\cdot, \cdot)$, which is used to calculate the input pattern similarity \cite{schoelkopf+smola2002}.
Assuming to deal with $\mathbb{R}^d$ data, it is well-known that distances and similarity among vectors are related by the inner product $\langle\cdot, \cdot\rangle$ construct:
\begin{equation}
d(x, y) = \sqrt{\langle x, x\rangle + \langle y, y\rangle - 2\langle x, y \rangle}.
\end{equation}
Let $K_{ij}=\langle x_i, x_j\rangle$ be the $n\times n$ positive definite similarity matrix (also called Gram matrix), containing the pairwise inner products among the $n$ input patterns. Let $D_{ij}=d(x_i, x_j)$ be the matrix containing the pairwise Euclidean distances. It is possible to obtain $\mathbf{K}$ by computing
\begin{equation}
\mathbf{K} = -\frac{1}{2}\mathbf{C}\mathbf{D}^2\mathbf{C},
\end{equation}
where $\mathbf{C}=\mathbf{I}-\mathbf{1}/n$ is called centering matrix and $\mathbf{D}^2$ contains the squared distances \cite{Duin2012826,si_asoc_grc}.
Since the evaluation of a positive definite kernel function $k(\cdot, \cdot)$ corresponds to the evaluation of an inner
product on an \textit{induced} high-dimensional Hilbert space (Mercer’s theorem), we can determine the similarity matrix as $K_{ij}=k(x_i, x_j)$.
Therefore, if we are able to compute a metric distance among the input patterns, we can obtain the corresponding similarity/kernel matrix, $\mathbf{K}$, regardless the nature of the input domain (the converse is also obviously true).
Since in this paper we deal with sequences of variable length, we compute the input data distance by using the Levenshtein distance.
Depending on the specific sequence type, we use static (i.e., predefined) and ad hoc substitution weights for characterizing the edit operation costs. A kernel function is obtained from the Levenshtein string metric according to the aforementioned mathematical expressions.
Correction techniques \cite{Duin2012826} are not used here to make sure that the obtained kernel is positive definite regardless the definition of $d(\cdot, \cdot)$.
Dataset DS-1811, which contains sequences of characters, has been processed with a \textit{kernelized} C-SVM based classifier \cite{schoelkopf+smola2002}. We use three different weighting schemes to process DS-1811.
The first one uses a static cost scheme: all substitutions have unitary costs.
In the second case we use the substitution costs given by the PAM120 matrix \cite{sub_matrix}.
In the latter setting, we determine the substitution costs as the solution of a problem-dependent, global optimization described in the following.
Let $\mathbf{S}\in[0, 1]^{20\times 20}$ be a non-negative, zero-diagonal, cost matrix which contains at position $S_{ij}$ the cost of substituting the \textit{i}-th amino acid with the \textit{j}-th.
The best-performing $\mathbf{S}$ is determined as the cost matrix that maximizes the classification accuracy achieved on a suitable control dataset.
The global optimization is driven by the performance of the classifier on the dataset instance, and it is implemented here by means of a standard genetic algorithm.
In practice, we search for a custom solution, i.e., a matrix $\mathbf{S}$ that is specifically tuned for the particular problem at hand.
Each substitution matrix is formally encoded as a real-valued vector lying in $[0, 1]^{400}$.
Selection, cross-over, and random mutation operators of the genetic algorithm are applied on such vectors to let the population of candidate solutions (i.e., cost matrix instances) evolve over the iterations.
The stop criterion used to determine the convergence of the genetic algorithm is implemented by checking for a maximum number of iterations (100) and by evaluating if the fitness has not changed during the last 10 iterations.
In DS-G-454, instead, we process input patterns that are labeled graphs. Therefore, we implement the kernel function as the graph coverage kernel \cite{livi2012gc,livi2012_pgm}.
DS-454 is a subset of DS-1811, and therefore the same classification systems described for DS-1811 apply here.
Finally, DS-S-454 is again a dataset of sequences, although the objects in the sequences are three-dimensional real-valued vectors; therefore in this case we use the Euclidean distances among such vectors for the substitution weights of the Levenshtein distance.
\subsection{Global structural complexity measures for the protein graphs}
\label{sec:complexity}
Besides their huge differences in shape, proteins contact maps (graphs) share a common modular pattern \cite{tasdighian2013modules,mixbionets2}. The application of Guimer\`{a} and Amaral network cartography \cite{guimera2005cartography} on a large set of proteins, demonstrated the existence of a common role subdivision among amino acid residues for all protein molecules \cite{tasdighian2013modules}. This suggests a common intrinsic complexity of protein graphs wiring pattern.
The basic invariance of protein graph complexity re-assures us of the possibility of relying on a common scaling and thus on the homogeneity of the considered objects, upon which pattern recognition is applied.
To provide experimental evidence of the conjecture that folded proteins ``look all the same'' from the pure topological viewpoint \cite{doi:10.1021/cr3002356}, we analyzed the complexity of the graphs in DS-G-454. The complexity of a graph topology, which is a well-known and widely discussed concept in the scientific literature \cite{Dehmer201157,havlin2010}, have been measured for the data in DS-G-454 in two complementary ways: by (i) calculating the 2-order R\'{e}nyi entropy \cite{rrnyi1961measures} of the associated Markov chain stationary distribution \cite{norris98Markovchains}, and (ii) computing the recently-proposed graph ambiguity measure \cite{Livi_ga_2013}, which is rooted in the fuzzy sets context.
Since the two methods are conceived in two different mathematical settings, we effectively analyze the graph complexity in terms of two different interpretations of the same concept: the uncertainty. In fact, the first method provides an interpretation of the uncertainty of a graph in terms of randomness (transition probability among the states--vertices), while the second method gives us an evaluation of the ambiguity, which refers instead to the degree of irregularity of the graph's topology.
\subsubsection{Uncertainty of a random walk in the graph}
\label{sec:2order_entropy}
Let $G=(\mathcal{V}, \mathcal{E}), |\mathcal{V}|=n$, be an undirected graph. The Markov chain associated to $G$ is completely described by the graph transition matrix, $\mathbf{T}^{n\times n}$.
The stationary probability distribution (also called equilibrium distribution) of the chain is a probability vector $\pi\in[0, 1]^n, \sum_{i=1}^{n} \pi_i = 1$, such that $\pi\mathbf{T}=\pi$; it can be seen also as the left eigenvector of $\mathbf{T}$ with eigenvalue 1.
The stationary distribution $\pi$ always exists but in general it is not unique; moreover, it depends on the initial distribution and on other properties of the chain \cite{Lovasz1996}. A straightforward and always valid way to calculate a stationary distribution of a graph $G$ is to set the initial distribution of the chain as $\pi$, defined as follows:
\begin{equation}
\pi_i = \frac{\left( \sum_{j=1}^{n} A_{ij} \right)}{2|\mathcal{E}|}.
\end{equation}
Given $\pi$, we computed the $\alpha$-order R\'{e}nyi entropy, with $\alpha=2$, to evaluate the complexity.
The 2-order R\'{e}nyi entropy is computed as:
\begin{equation}
H(\pi) = -\log\left(\sum_{i=1}^{n} \pi_{i}^{2}\right) .
\end{equation}
It is well-known that $H(\pi)$ is upper bounded by $\log(n)$ and therefore it can be normalized in the unit interval. The quantity approaches one as the distribution becomes uniform (i.e., maximally uncertain--complex), while it tends to zero as it converges to the degenerate distribution. A $d$-regular graph, for instance, is associated with a uniform stationary distribution of the states (all states--vertices are equally likely to be visited).
\subsubsection{Ambiguity of the graphs}
\label{sec:ambiguity}
The ambiguity of a graph $G=(\mathcal{V}, \mathcal{E}), |\mathcal{V}|=n$, gives a measure of uncertainty elaborated according to a fuzzy set based interpretation \cite{Livi_ga_2013}.
The ambiguity of the graph $G$ is calculated by embedding the graph into a fuzzy hypercube $\mathcal{I}=[0, 1]^n$, which, in short, encodes the membership values of the vertices.
A graph $G$ is mapped to a type-1 fuzzy set $\mathcal{F}$, defined as
\begin{equation}
\mathcal{F}=\{(v, \mu_{\mathcal{F}}(v)) | \ v\in\mathcal{V}(G), \mu_{\mathcal{F}}(v)\in[0, 1] \},
\end{equation}
by generating the membership function $\mu_{\mathcal{F}}(\cdot)$ of the graph vertices, $\mathcal{V}(G)$.
Such a membership function is constructed by considering the so-called partition $P$ of the graph, which basically is a representation of the vertex set $\mathcal{V}(G)$ as the union of $k$ disjoint subsets $\mathcal{C}_i$ of vertices:
\begin{equation}
P=\bigcup_{i=1}^{k} \mathcal{C}_i.
\end{equation}
The partition $P$ is then fuzzified by computing the t-conorm $\bot$ \cite{pedrycz1998introduction} among the fuzzy sets $\mathcal{F}_i$ associated to each $\mathcal{C}_i$, yielding the resulting fuzzy set $\mathcal{F}$:
\begin{equation}
\label{eq:fuzz}
\mathcal{F}= \displaystyle\bot_{i=1}^{k} \mathcal{F}_i.
\end{equation}
Eq. \ref{eq:fuzz} is synthetically denoted as $\mathcal{F}=\phi^{P}(G)$ in the following.
The membership function $\mu_{\mathcal{F}_{i}}(\cdot)$ describing the fuzzy set $\mathcal{F}_i$ is generated according to the following expression:
\begin{equation}
\mu_{\mathcal{F}_{i}}(v) = \alpha_{\mathcal{C}_{i}}(v) \times \beta_{\mathcal{C}_{i}}(v).
\end{equation}
Notably, $\alpha_{\mathcal{C}_{i}}(v)$ accounts for the degree concentration of vertex $v$ in $\mathcal{C}_i$, while $\beta_{\mathcal{C}_{i}}(v)$ gives the importance of the vertex in $\mathcal{C}_i$ in terms of centrality.
Given the fuzzy set $\mathcal{F}$ representing the uncertainty of the graph $G$, the measure of ambiguity of $G$, denoted as $A(G)$, is obtained by computing (any monotonic non-decreasing transformation of) the fuzzy entropy of $\mathcal{F}$ \cite{pedrycz1998introduction}.
Since there are exponentially-many fuzzy set representations for a single graph $G$, the ambiguity value is calculated as the solution of the following combinatorial optimization problem:
\begin{equation}
A(G) = \min_{P} A(\phi^{P}(G)),
\end{equation}
where $A(G)$ assumes values within the $[0, 1]$ range, approaching one as the graph is maximally ambiguous (i.e., maximally irregular). It has been proved that $A(G)$ is zero when the graph is regular (e.g., complete) \cite{Livi_ga_2013}. Accordingly, $A(G)$ can be used as a global complexity descriptor for the topology of $G$, which in particular characterizes the irregularity of the topology of $G$.
\section{Experimental results and discussion}
\label{sec:results}
In this section, we present the experimental results achieved in this paper. In Section \ref{sec:cca} we first demonstrate the soundness, in terms of statical analysis, of the data used for the classification of the E. coli proteome. In Section \ref{sec:class_results} we show the test set classification accuracy results obtained on the alternative protein representations herein considered.
Finally, in Section \ref{sec:complexity_results} we analyze the complexity of the graphs representing the (available) E. coli proteins.
\subsection{Canonical correlation analysis}
\label{sec:cca}
In order to assess the chemico-physical relevance of the obtained classification performance, we checked the possibility to project the SVM solution on DS-1811 by means of the learned cost matrix $\mathbf{S}$ -- that was based only on literal information of residues -- into a chemico-physical consistent solution.
To reach this goal, a combined principal component/canonical analysis strategy was adopted \cite{cca}, which is described in the following.
The available 243 chemico-physical descriptors \cite{apdbase} were submitted to a principal component analysis (PCA), ending up in a three-component solution, globally explaining 70.2\% of total variation (PC1 = 42.3\%, PC2 = 16.3\%, and PC3 = 11.6\%).
The three extracted components allowed for a straightforward interpretation: PC1 is a hydrophobicity score (positive correlation with hydrophobicity scales, negative correlation with hydrophilicity scales); PC2 is a size/steric hindrance component (very high loadings with volume / size scales); finally, PC3 can be considered as a ``rigidity'' scale (high positive loading with turn propensity, negative loadings with alpha-helix propensity scales).
Those three components allow us to generate a matrix having as rows the 20 amino acid species and as columns the three correspondent ChemPhys1 (Hydrophobicity), ChemPhys2 (Size), and ChemPhys3 (Structural Rigidity) component scores.
An analogous approach was followed on the calculated best-performing cost matrix, $\mathbf{S}$, corresponding to the SVM-based solution obtained by considering the classification performance as feedback.
We computed the three principal components of $\mathbf{S}$, having as rows the different amino acids and as columns the ``cost'' of their transformation into different amino acid species.
In this case, a seven-components solution was obtained, explaining 68.8\% of total variability. In the following, we refer to those components with SVM1--SVM7.
We are now in the condition to look for the existence of a sound correlation structure linking the chemico-physical and substitution cost spaces.
Mathematically, this problem translates into a canonical correlation analysis. The linear combinations of the chemico-physical components and the cost matrix components are then computed, with the objective of maximizing the linear correlation. These best correlated combinations are called canonical variates.
They are extracted by the algorithm in order of correlation magnitude; moreover, they are mutually orthogonal to each other, so describing independent ``transformation rules'' between the two heterogeneous spaces.
The fact that we use as original variables principal component scores, which are each other mutually independent inside their respective spaces, reassures us that any observed correlation will be driven by inter-fields correlations only.
The results of the canonical correlation analysis are reported in Table \ref{tab:table1}.
\begin{table}[tph!]\footnotesize
\caption{Results of canonical correlation analysis.}
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
\rowcolor{lgray} & \textbf{Canonical Correlation (CC)} & \textbf{Adjusted CC} & \textbf{Approx Std Err} & \textbf{Squared CC} \\
\hline
\cellcolor{lgray}\textbf{1} & 0.901527 & 0.856986 & 0.042958 & 0.812751 \\
\hline
\cellcolor{lgray}\textbf{2} & 0.747438 & 0.666824 & 0.101250 & 0.558663 \\
\hline
\cellcolor{lgray}\textbf{3} & 0.468202 & 0.333368 & 0.179125 & 0.219213 \\
\hline
\end{tabular}
\label{tab:table1}
\end{center}
\end{table}
It is possible to note two statistically significant canonical variates: the first one scoring a squared canonical correlation around 0.8 (with $r=0.9$), while the second canonical variate gives $r=0.75$, correspondent to a model accounting for 55\% of explained variance (R-square).
In Table \ref{tab:table2} and Table \ref{tab:table3}, respectively, are reported the correlation among chemico-physical and substitution cost variables with the respective canonical variates; the significant correlations are bolded.
\begin{table}[tph!]
\begin{center}
\caption{Correlations among the chemico-physical variables and their canonical variates.}
\begin{tabular}{|c|c|c|c|}
\hline
\rowcolor{lgray} & \textbf{Variate 1} & \textbf{Variate 2} & \textbf{Variate 3} \\
\hline
\cellcolor{lgray}\textbf{ChemPhys1} & -0.1632 & \textbf{0.9759} & -0.1447 \\
\hline
\cellcolor{lgray}\textbf{ChemPhys2} & \textbf{0.8269} & 0.2153 & 0.5194 \\
\hline
\cellcolor{lgray}\textbf{ChemPhys3} & \textbf{-0.5381} & 0.0349 & 0.8422 \\
\hline
\end{tabular}
\label{tab:table2}
\end{center}
\end{table}
\begin{table}[tph!]
\begin{center}
\caption{Correlations between the SVM variables and their canonical variates.}
\begin{tabular}{|c|c|c|c|}
\hline
\rowcolor{lgray} & \textbf{Variate 1} & \textbf{Variate 2} & \textbf{Variate 3} \\
\hline
\cellcolor{lgray}\textbf{SVM1} & \textbf{0.4465} & -0.0808 & 0.4137 \\
\hline
\cellcolor{lgray}\textbf{SVM2} & 0.3212 & 0.3470 & -0.5716 \\
\hline
\cellcolor{lgray}\textbf{SVM3} & -0.1259 & -0.1012 & 0.3203 \\
\hline
\cellcolor{lgray}\textbf{SVM4} & -0.1851 & -0.0256 & 0.4989 \\
\hline
\cellcolor{lgray}\textbf{SVM5} & \textbf{0.5303} & -0.0620 & 0.2210 \\
\hline
\cellcolor{lgray}\textbf{SVM6} & \textbf{-0.6024} & 0.1849 & 0.0061 \\
\hline
\cellcolor{lgray}\textbf{SVM7} & -0.0566 & \textbf{-0.9078} & -0.3189 \\
\hline
\end{tabular}
\label{tab:table3}
\end{center}
\end{table}
The most relevant (first canonical variate) transfer rule linking chemico-physical and discrimination spaces is mainly driven by amino acid size (ChemPhys1) acting in opposition with ``rigidity'' component (ChemPhys3).
The second canonical variate, instead, is linked to the unique role of hydrophobicity, but this is a minor effect (only the seventh component of the discrimination space scales significantly with the canonical variate).
Cost matrix reports the ``expense'' of a given amino acid substitution: high costs (near 1) imply a significant change in the classification, while low costs (near zero) the relative indifference of the substitution in terms of effects on classification.
The presence of significant canonical correlations between chemico-physical features and amino acids substitution costs implies that the hidden discrimination logic is rooted on justifiable chemico-physical properties.
However, by no means we can derive ``general linear models'' based on single amino acid features, because the same move (e.g., substituting residue A with V) has an opposite effect in terms of ``soluble/not soluble'', depending on the starting point.
Moreover, the different moves have both positive and negative loadings on different SVM components. This means that we cannot derive any global aggregation or folding value for a given move: the effect of each move depends on the starting point and has a different value on different eigenvectors, thus a move can be at a high cost for SVM3 and low cost for SVM4.
The SVM components, corresponding to the hidden rules at the basis of cost matrix $\mathbf{S}$, are organized in terms of variance explained.
This implies that SVM1 is more relevant than SVM2. The resulting cost is a sort of majority vote over the components, and it can by no means be collapsed to simple general rules expressed in context independent chemico-physical terms without further specification.
This explains both the lack of any relevant general correlation between average physical properties and solubility classification, and the extreme sensitivity of solubility/aggregation equilibrium by single point mutations.
In Figure \ref{fig:correlation}, the plot showing the relation between chemico-physical and SVM spaces in terms of canonical variates is reported, keeping in mind that SVM variates refer to a ``derivative'' space.
That is, the entity of the effect exerted by a single amino acid changes on its solubility and not the general soluble/insoluble propensity of amino acid residues that, as we explained, does not exist as such. The plot can thus be intended as a proof of the chemico-physical relevance of the solution found by the proposed algorithm -- i.e., the best-performing matrix $\mathbf{S}$ -- and thus of its scientific soundness.
\begin{figure}[ht!]
\centering
\includegraphics[viewport=0 0 476 346,scale=0.75,keepaspectratio=true]{./Figure2}
\caption{Correlation analysis among the two most important canonical variates.}
\label{fig:correlation}
\end{figure}
\subsection{Classification accuracy results}
\label{sec:class_results}
In their paper, \citet{niwa2009} noted that soluble proteins were considerably smaller than insoluble ones.
This is in line with many results coming from other groups \cite{Waldo200333,PRO:PRO122057,Kubelka200476} and with the well-known dependence of folding process from molecules size (\citet{PRO:PRO122057}).
In order to have a baseline with which contrasting our results in terms of recognition, we developed a discriminator operating on the basis of protein size only, i.e., the number of residues in the protein.
Tests have been performed on DS-1811. The best discriminating threshold was found at 246 residues: a protein containing less than 246 residues is deterministically classified as soluble.
Such a classification scheme commits 301 errors for the insoluble proteins (error rate of 0.198) and 33 for the soluble (error rate of 0.3).
This implies a marked tendency of small proteins to be more soluble than large ones \cite{ramshini2011large}.
The fact that the best discrimination threshold corresponds to 246 residues length is consistent with the results outlined by \citet{PROT:PROT21179} that, on the sole basis of between-residues contact density as a function of protein length, discovered a maximum preferred size for protein domains located between 180 and 320 residues length with a peak at 274.
In the analysis reported in \citet{PROT:PROT21179}, no solubility data were specifically addressed; they were solely based on the number of protein internal contacts. Moreover, the results were demonstrated to be very general and independent of the analyzed dataset.
The fact that the present analysis, in a completely independent manner, re-discovers the same threshold is of extreme importance, especially if we consider that the density of within-molecule contacts exponentially decreases at increasing length until approximately 250-300 residues length, and that from this threshold onward the density of contacts remains approximately constant.
The higher density of within-molecule contacts corresponds to a preference for within-molecule interactions (and consequently toward the soluble pole). This gives a simple, albeit very raw, rationalization to the observed convergence toward the same threshold of contact density scaling and soluble/insoluble threshold.
On a more general perspective, this result tells us that the consideration of topological features (contact graphs) of protein 3D structure can be a useful perspective to be adopted for the discrimination of soluble/insoluble proteins.
Let us now move to the results obtained by means of the various SVM-based classifiers.
The results reported in Table \ref{tab:testset_results_ecoli} have been obtained by setting $C=2$ for C-SVM; such a setting has been obtained by preliminary tests.
Overall, the comparison with respect to the aforementioned baseline discrimination approach is in favour of C-SVM.
Results on DS-1811 (sequence-based analysis) show satisfactory error rates for the two classes of insoluble and soluble proteins, regardless the weighting scheme adopted for the Levenshtein distance. Note that the row of this table indicating ``EVOLUTIVE'' as the substitution matrix is the solution that adopts the aforementioned best-performing substitution cost matrix, $\mathbf{S}$.
Results on DS-G-454, i.e., the dataset of labeled graphs, show considerably worse results. In fact, the error rate for the soluble class approaches 0.8, denoting overall an unreliable classification.
Error rates on DS-454 denote a considerable drop of performances for the recognition of the soluble class with respect to DS-1811. Since the dataset of DS-454 is composed only of those sequences for which we have the corresponding graph representations in DS-G-454, we deduce that such a subset of proteins induces a more complex problem than the one on DS-1811 from the pure class discrimination viewpoint; this is also substantiated by the reduced number of available patterns.
On the other hand, results for DS-S-454 are interesting, considering both per-class and global error rates.
We stress that this dataset has been obtained by the seriation of the graph representations in DS-G-454, which is performed by using the information of the first eigenvector of the transition matrix.
Therefore, such an approach could be considered as a graph-based pattern recognition technique, even if the classification is performed by dealing with sequences of vectors.
To further strengthen the results obtained on DS-S-454, we repeated the experiment by randomizing the class labels of the training set patterns; test set is left unaltered. As expected, test set results are much worse than the ones presented in Tab. \ref{tab:testset_results_ecoli}.
In fact, the global error rate drops to 0.5471, with 75 and 12 errors for the (randomized) insoluble and soluble classes, respectively.
The important differences of results achieved on DS-G-454 and DS-S-454, by considering the adopted computational systems, are not easy to justify from a pure algorithmic viewpoint.
That is, the precise reason why operating directly on the labeled graph space yields worse results are not easy to rationalize, since both approaches (i.e., DS-G-454 and DS-S-454) operate basically on the same structural and chemico-physical information, although arranged in two different settings (i.e., respectively graph and sequence).
However, as recently pointed out \cite{PROT:PROT22113}, a subset of the eigenvectors of the contact (adjacency) matrix provides sound descriptors of the protein structure, showing good correlation with hydrophobicity.
It is worth to stress that, since the classification approach used for DS-S-454 effectively makes explicit use of the eigenvalues/eigenvectors of the transition matrix (which is an elaboration of the contact/adjacency matrix) to define the order of the graph vertices in the sequence, such a technique is well-justified from the biological viewpoint, taking also into account chemico-physical information derived by the previously discussed statistical analysis.
\begin{table}[tph!]\footnotesize
\begin{center}
\caption{Test set classification accuracy results achieved on the considered datasets.}
\begin{tabular}{|c|p{1.95cm}|p{2.4cm}|p{2.3cm}|p{2.5cm}|p{1.7cm}|}
\hline
\rowcolor{lgray}\textbf{DS} & \textbf{Subst. Matrix} & \textbf{\# ERR. INS. / \# INS.} & \textbf{\# ERR. SOL. / \# SOL.} & \textbf{ERR. RATE INS. -- ERR. RATE SOL.} & \textbf{GLOBAL ERR. RATE} \\
\hline
\multirow{3}{*}{DS-1811}
& -- & 232/1521 & 33/110 & 0.1525 -- 0.3000 & 0.1624 \\
\cline{2-6}
& PAM120 & 246/1521 & 28/110 & 0.1617 -- 0.2545 & 0.1679 \\
\cline{2-6}
& EVOLUTIVE & 304/1521 & 22/110 & 0.1998 -- 0.2000 & 0.1998 \\
\hline
DS-G-454 & -- & 12/132 & 21/27 & 0.0909 -- 0.7777 & 0.2075 \\
\hline
DS-454 & -- & 4/132 & 20/27 & 0.0303 -- 0.7407 & 0.1509 \\
\hline
DS-S-454 & -- & 10/132 & 11/27 & 0.0757 -- 0.4074 & 0.1320 \\
\hline
\end{tabular}
\label{tab:testset_results_ecoli}
\end{center}
\end{table}
\subsection{Evaluation of protein graph structural complexity}
\label{sec:complexity_results}
We calculated the normalized 2-order R\'{e}nyi entropy (as described in Section \ref{sec:2order_entropy}) for every graph in DS-G-454. As expected, the graphs have almost identical entropy (i.e., complexity), regardless the soluble/insoluble original classification. The insoluble proteins have average entropy of 0.9784, with standard deviation 0.0057; the soluble ones have average entropy 0.9726 and standard deviation 0.0084.
As for the 2-order R\'{e}nyi entropy, we calculated also the graph ambiguity (Section \ref{sec:ambiguity}) of the same graphs in DS-G-454. Analogously, the ambiguity does not seem to provide an effective discriminating feature among soluble and insoluble proteins (although it gives a slightly better average discrimination than the 2-order R\'{e}nyi entropy).
In fact, the insoluble proteins have average ambiguity of 0.1745, with standard deviation 0.0221; the soluble ones have average ambiguity of 0.1980 with standard deviation 0.0313.
According to herein considered interpretations of the structural complexity of a graph, such results points toward the fact that proteins have a common global topological architecture.
It is worth mentioning that we performed additional experiments by representing each graph in the DS-G-454 dataset as a feature vector consisting of two real-valued features: the calculated values of entropy and ambiguity.
The related data distribution is shown in Figure \ref{fig:complex_distr}. We considered a C-SVM classifier operating with a ``regular'' Gaussian kernel on such two-dimensional vectors, with the aim of formally quantifying the class discrimination performance.
The obtained results (full details are not shown here) demonstrate the inability of discriminating soluble and insoluble proteins when represented according to only those two structural complexity descriptors, achieving an error rate of 100\% for the soluble class on the test set (training and test sets are constructed by considering several instances with the same split percentages of DS-G-454).
Nonetheless, such results cannot be considered as definitive. In future studies we will evaluate additional structural descriptors, enriching this type of vector-based representation of protein graphs with further, more complex characteristics.
\begin{figure}[ht!]
\centering
\includegraphics[viewport=0 0 349 243,scale=0.8,keepaspectratio=true]{./entropy_ambiguity}
\caption{Data distribution of the 454 graphs represented as two-dimensional vectors containing the corresponding calculated entropy and ambiguity values.}
\label{fig:complex_distr}
\end{figure}
\section{Conclusions and future directions}
\label{sec:conclusions}
The aim of this study was the derivation of useful insights on the nature of protein folding/aggregation by following a data-driven paradigm.
We exploited different representations of proteins, all based on a consistent and pure experimental data base \cite{niwa2009}.
The use of sequence and graph based representations of patterns implied the use of pattern recognition methods capable of dealing with non-geometric data.
The problem of aggregation propensity recognition was demonstrated to be solvable at the single molecule scale starting from pure sequence data.
Moreover, the scoring of a statistically significant canonical correlation among computed substitution cost matrix and classical chemico-physical properties of amino acids, assures us of the existence of a still controversial and largely context-dependent ``physical code'' of protein folding \cite{dill2012protein}, which we can interpret as a sort of ``transfer equation'' expressed in chemico-physical terms linking sequence and aggregation propensity spaces.
This transfer equation was demonstrated to arise from the combinations of different mutually orthogonal components rooted on chemico-physical properties of amino acid residues. This is strictly dependent on contextual information consistently with the presence of a very delicate balance between folding and aggregation modes, which could be dramatically shifted even by single mutation events.
Additionally, we also reconfirmed the results in Ref. \cite{Agostini2012237} by using a more parsimonious model, that is, with three chemico-physical descriptive variables instead of 28.
The interplay of different forces in folding (van der Waals force here represented by amino acids size component, hydrophobic interaction, and backbone angle preference here represented by structural rigidity component), confirms the general picture given by \citet{dill2012protein}.
The dramatic effect of protein size upon folding/aggregation balance was demonstrated to be mediated by contact density consistently with the recognized importance of contact order in folding rate (\citet{PRO:PRO122057}) and, more in general, by the relative prevalence of ``within-molecule'' vs. ``between-molecules'' contacts.
It is worth noting that the 246 residues length we demonstrated to be the best threshold length for the soluble/insoluble, baseline discrimination fits very well with the limit where an obliged multi-domain organization starts, as described in \citet{PROT:PROT21179}.
This result constitutes a link between size and topological considerations.
The striking differences among the test set classification accuracy results achieved on DS-S-454, DS-454, and DS-G-454 deserve a further analysis. The DS-454 dataset, which is constituted by the proteins for which the 3D structural information is available as PDB, constitutes a not-so-good sample, as evident by the high rate of error for the soluble class (0.74 vs. 0.20 achieved on DS-1811).
On the base of the results achieved with the adopted technique operating directly on the labeled graphs space, we can conclude that this specific computational approach was not sufficiently effective.
In fact, the system was tested for its discrimination power on the DS-G-454 dataset, obtaining classification accuracy results that are slightly worse that the ones on DS-454; the recognition rate of the soluble class is weak, with an error rate equal to 0.77.
However, such a conclusion regarding the results on DS-G-454 cannot be considered as definitive. In fact, we adopted a SVM-based classification system operating in the labeled graph space by means of a graph kernel function.
However, in the graph-based pattern recognition context there is a plethora of techniques that could be used for such a purpose \cite{gm_survey,odse,Hancock2012833,gralg_2012}.
On the other hand, when the same graph topological information was expressed in another form by ordering the amino acid residues of each protein according to the spectral decomposition of their transition matrix, the discrimination power reached a global error rate of 0.13 (the best general performance achieved in this paper). Moreover, the error rate for the soluble class went down to 0.40, definitely better than the 0.74 and 0.77 achieved on DS-454 and DS-G-454, respectively.
This result tells us that we designed a very interesting type of protein representation (DS-S-454), which allowed us to achieve a reasonable classification accuracy, much higher than the one based on sequence of amino acids identifiers only (i.e., DS-454).
\citet{PROT:PROT22113} predicted such efficiency of multilevel representation of protein topology obtained by the projection of topological features on protein sequence.
The Effective Connectivity (EC) profile \cite{PROT:PROT20240} the authors proposed is related to our approach, being based on the eigenvalues of the proteins contact maps.
Bastolla et al. explicitly state ``...the EC profile allows to define a natural measure of modularity that correlates with the number of domains composing the protein...'' and ``...structurally similar proteins have very similar EC profiles, so that the similarity between aligned EC profiles can be used as a structural similarity measure...''.
In their work, the authors complement the pure topological EC with hydrophobicity information showing a marginal but statistically significant improvement in the ``reverse folding problem'' of predicting sequence order by the EC profile.
Here we showed compatible results with the Bastolla et al. proposal, inserting a more refined chemico-physical description of residues (the first three chemico-physical components obtained by the PCA) and imposing an evolution-motivated substitution metrics, obtaining a very good discrimination power as for the hard task of aggregation/folding discrimination.
At odds with Bastolla et al., we adopted a transition matrix based graph representation instead of contact (i.e., adjacency) matrix.
This choice was motivated by the goal of obtaining a global structural complexity measure characterizing the protein 3D structures (based on the stationary distribution of the associated Markov chain). This fact was based on the hypothesis that the more complex the protein architecture is, the higher the difficulty of reaching a stable folding state.
To this end, we also reported preliminary experimental results that showed the high similarity of protein topology, regardless their solubility degree and number of residues.
Nonetheless, more effort will be devoted in future studies to this specific aspect of analyzing the protein graph complexity, for the purpose of discrimination and prediction of the solubility degree.
There are plenty of results (for a review see \citet{doi:10.1021/cr3002356}) pointing to the fact that protein contact maps play, in the realm of proteins, the same role played by structural formula in organic chemistry.
The apparently dramatic loss of information consequent to shift from the whole-rank structural information (the 3D coordinates of each single atom) to the yes/no information of effective contacts between alpha carbons, ends up in the focusing on the actually relevant part of information analogously to what happens with chemical graphs of small organic molecules, where the simple consideration of structural formula was demonstrated to be much more efficient than more sophisticated approaches for predicting their biological activity \cite{bender2005discussion}.
Here in this paper we gave another proof of the central role played in protein science by the use of topological information combined with chemico-physical descriptions of residues.
In our opinion, these results points toward the realization of a consistent and multilevel formalization of protein molecules integrating heterogeneous information.
On a more general ground, we demonstrated the possibility to derive valuable information on the nature of a given investigated phenomenon, operating on the different discrimination ability of pattern recognition techniques as applied to alternative representations of the same objects.
This fact opens the way to the consideration of pattern recognition methods as measurement devices, not so different from a microscope or a spectrophotometer, so closing the gap between computational and content-oriented classical scientific approaches.
In addition, the combination of the the information coming from both wiring architecture and features of specific vertices into a synthetic descriptive frame can be important as a general quantification scheme for any classification problem involving networked systems.
\bibliographystyle{abbrvnat}
|
\section{Introduction}
Diffraction is centuries-old understanding of the single-beam-in multi-beam-out phenomenon in optics, acoustics, quantum mechanics, and all wave equation governed scattering processes, when the middle media has some property periodically varying in space. Naturally we should know that as a simple case transmission of the electron through a sinusoidal-height quantum potential barrier or sinusoidal-depth well demonstrates diffraction effect. Although in vast numerical treatments such as the plane-wave expansion method for metamaterials, this problem is just a building block within, we think it is important to particularly treat the quantum grating effect by developing a general scattering method and provide the detailed physical picture \emph{ab initio} from the Schr\"odinger equation, which to our knowledge is not covered in literature. Also, the sinusoidal-height quantum potential barrier itself can be the model of a real material and a real transport device based on it. The recent widely-focused material, helimagnet (HM), lends a very appropriate platform.
The HM is a kind of magnetic state\cite{Ref13} with its spin spiraling in two or three dimensions characterized by a single spiral wavevector ${\bf{Q}}$, which is different from conventional spatially uniform ferromagnet and antiferromagnet.
When a single electron passes through the HM structure, spin-dependent diffraction occurs. Recently the transport properties of the HM-embedded devices are targeted from different view angles in literature. Manchon \emph{et al.}\cite{Ref42} and us\cite{Ref43} observed the spin-dependent diffraction effect in the transmission at the ferromagnet/HM interface and through the thin-layer HM junction, respectively. Some functional devices were proposed based on the HM such as the persistent spin currents\cite{Ref19}, spin-field-effect transistor\cite{Ref17}, tunneling anisotropic magnetoresistance\cite{Ref18}, and spin resonance\cite{Ref13}. Conductance characteristic of the HM spin configuration and spiral period was found in the Fano resonance spectrum of a quasi-one-dimensional WG containing a thin conducting HM layer as a donor impurity\cite{Ref52}. Ac gate potential driven Quantum pumping behavior was also investigated recently\cite{ZhuPLA2014}. Although some works were done to investigate the HM-related transport properties, there lacks an overall description of the diffraction scenario.
As far as a good transport approach can go, beyond the conductance, we also considered the diffraction governed shot noise properties. As a consequence of the
quantization of charge and defined by quantum contribution in the current fluctuations, shot noise is useful to obtain information
on a system which is not available through conductance
measurements\cite{BlanterPR2000}. Two of the most significant shot noise experiments are carrier charge confirmations of the Cooper pair\cite{CooperShotNoise} and Laughlin quasiparticle\cite{LaughlinShotNoise}. In most cases, the properties of quantum correlation are reflected in the Fano factor $F$, which is defined by proportion of the real shot noise $S$ to Poisson noise $2eI$ ($I$ is the average current), the latter of which corresponds to single quasiparticle transmission without correlation. Therefore, some levels of the Fano factor have typical physical meaning. $F=1$ characterizes Poisson noise. Besides the ideal case, the Fano factor approaches $1$ when the transmission is extremely small corresponding to uncorrelated transport and closed channel in ballistic tunneling. $F=0$ characterizes full correlation and maximal quantum coherence. In real conductors, the Fano factor approaches $0$ when the transmission reaches $1$ corresponding to open channels in ballistic transport. In some cases with strong electron-electron interaction involved\cite{EnhacedShotNoise}, the shot noise can be enhanced beyond $1$. $F=1/2$ characterizes the effect of Pauli exclusion and $F=1/3$ characterizes diffusive transport when open and closed channels distributes in disorder such as diffusive metals\cite{DiffusiveMetal} and graphene\cite{Tworzydlo}. With understanding of the physics underlying different Fano factor levels, we could suppose that diffraction enhances transmission, different diffraction channels have strong coherence, and the shot noise should be thus suppressed. Our theory would confirm this supposition in detail.
Also, the transport properties are governed by diffraction. For the HM spiral wave vector $q$ larger than two times the electron Fermi wave vector $2 k_F$, both diffraction beams degrade into evanescent surface modes and do not contribute to the transmission, in which case the transmission is identical to that of a plain barrier. As a result, sharp change in the conductance, shot noise, and Fano factor occurs at $q=2 k_F$, lending a potential transport measurement of $q$.
\section{Spin Spiral Monolayer Toy Model}
\subsection{Theoretic Formulism}
Our model is sketched in Fig. 1 (a). An HM interlayer is put between two semiinfinite free regions extending in the $x$-$y$ plane. For electrons, those free regions can be normal metal leads. Transport direction is along the $z$ coordinate. The HM spin varies in space with the vector field
\begin{equation}
{{\bf{n}}_r} = \left[ {\sin \left( {qx} \right),0,\cos \left( {qx} \right)} \right].
\label{SpinVectorFiled}
\end{equation}
$q$ is the spin wave vector and the helix is two dimensional modulating in the $x$ direction. The spin exchange between the free electron and the HM magnetization giving rise to a space-dependent Zeeman term in the Hamiltonian, which is
\begin{equation}
H = - \frac{{{\hbar ^2}}}{{2{m^*}}}{\nabla ^2} + \left( {J{{\bf{n}}_r} \cdot {\bf{\sigma }} + {V_0}} \right)\delta \left( z \right).
\label{HMHamiltonian}
\end{equation}
Here, $m^*$ is the HM electron effective mass. $J$ refers to
space and momentum averages of the exchange coupling strength. $\bf{\sigma }$ is the Pauli matrix. We assume an ultrathin HM layer located at the $z=0$ plane, so its effect in the Hamiltonian can be approximated by a $\delta$-function. $V_0$ is the electrostatic potential of the HM. For insulating HM, $V_0 >0$ is a barrier potential; for conducting HM, $V_0 <0$ is a well potential. We consider the former case.
Scattered by the HM interlayer, the electron wave function with incidence, reflection, and transmission beams in the two free regions can be written as
\begin{equation}
\left\{ \begin{array}{l}
{\psi _I (x,z)} = \sum\limits_{n = - \infty ,\sigma }^{ + \infty } {\left( {A_{n\sigma }^i{e^{i{k_{xn}}x}}{e^{i{k_{zn}}z}}{\chi _\sigma } + A_{n\sigma }^o{e^{i{k_{xn}}x}}{e^{ - i{k_{zn}}z}}{\chi _\sigma }} \right)} ,\begin{array}{*{20}{c}}
{}&{z < 0,}
\end{array}\\
{\psi _{II} (x,z) } = \sum\limits_{n = - \infty ,\sigma }^{ + \infty } {\left( {B_{n\sigma }^i{e^{i{k_{xn}}x}}{e^{ - i{k_{zn}}z}}{\chi _\sigma } + B_{n\sigma }^o{e^{i{k_{xn}}x}}{e^{i{k_{zn}}z}}{\chi _\sigma }} \right)} ,\begin{array}{*{20}{c}}
{}&{z > 0,}
\end{array}
\end{array} \right.
\label{WeveFunctions}
\end{equation}
where $k_{xn}=k_x + nq$ and ${k_{zn}} = \sqrt {k_F^2 - k_y^2 - k_{xn}^2} $ with the Fermi wave vector ${k_F} = {{\sqrt {2{m_e}{E_F}} } \mathord{\left/
{\vphantom {{\sqrt {2{m_e}{E_F}} } \hbar }} \right.
\kern-\nulldelimiterspace} \hbar }$, $E_F$ the electron Fermi energy and $m_e$ the free electron mass. $A_{n\sigma }^i$ and $B_{n\sigma }^i$ are the probability amplitudes of the incoming waves from the lower and upper leads, respectively, while $A_{n\sigma }^o$ and $B_{n\sigma }^o$ are those of the outgoing waves. Diffraction occurs in the $x$ direction. Translation symmetry protects the plane wave component in the $y$-direction $e^{ik_y y}$ unchanged during transmission.
By continuity equation at the HM interface
\begin{equation}
{\psi _I}\left( {x,{0^ - }} \right) = {\psi _{II}}\left( {x,{0^ + }} \right),
\end{equation}
and
\begin{equation}
\frac{{{\hbar ^2}}}{{2{m_e}}}{\left. {\frac{{\partial {\psi _I}}}{{\partial z}}} \right|_{z = {0^ - }}} + \left( {{V_0} + { J \bf{w}}} \right){\psi _I}\left( {x,{0^ - }} \right) = \frac{{{\hbar ^2}}}{{2{m_e}}}{\left. {\frac{{\partial {\psi _{II}}}}{{\partial z}}} \right|_{z = {0^ + }}},
\end{equation}
with
\begin{equation}
{\bf{w}} = \left[ {\begin{array}{*{20}{c}}
{\cos \left( {qx} \right)}&{\sin \left( {qx} \right)}\\
{\sin \left( {qx} \right)}&{ - \cos \left( {qx} \right)}
\end{array}} \right],
\end{equation}
the scattering matrix relation
\begin{equation}
\left( {\begin{array}{*{20}{c}}
{A_{n\sigma }^o}\\
{B_{n\sigma }^o}
\end{array}} \right) =\sum _m { \left( {\begin{array}{*{20}{c}}
{r_{nm}^{\sigma \tau }} & {{t'}_{nm}^{\sigma \tau }}\\
{t_{nm}^{\sigma \tau }} & {{r'}_{nm}^{\sigma \tau }}
\end{array}} \right)\left( {\begin{array}{*{20}{c}}
{A_{m\tau }^i}\\
{B_{m\tau }^i}
\end{array}} \right)}=\sum _m {{{\bf{\hat s}}_{\sigma \tau }}\left( {{{\bf{k}}_n},{{\bf{k}}_m}} \right)\left( {\begin{array}{*{20}{c}}
{A_{m\tau }^i}\\
{B_{m\tau }^i}
\end{array}} \right)}
\end{equation}
can be obtained. $r_{nm}^{\sigma \tau }$ and $t_{nm}^{\sigma \tau }$ represent respectively the reflection and transmission amplitudes from the spin-$\tau$, $m$th-order diffraction channel to the spin-$\sigma$, $n$th-order diffraction channel. ${r'}_{nm}^{\sigma \tau }$ and ${t'}_{nm}^{\sigma \tau }$ are the corresponding backward amplitudes. ${\bf{k}}_n=(k_{xn},k_y,k_{zn})$ is the three-dimensional wavevector labeling diffraction channels. Considering the real current flux, the scattering matrix
\begin{equation}
{{\bf{\hat S}}_{\sigma \tau }}\left( {{{\bf{k}}_n},{{\bf{k}}_m}} \right) = \sqrt {\frac{{{\mathop{\rm Re}\nolimits} \left( {{k_{zn}}} \right)}}{{{\mathop{\rm Re}\nolimits} \left( {{k_{zm}}} \right)}}} {{\bf{\hat s}}_{\sigma \tau }}\left( {{{\bf{k}}_n},{{\bf{k}}_m}} \right) .
\label{DefiningBigScatteringMatrix}
\end{equation}
With the scattering matrix ${{\bf{\hat S}}_{\sigma \tau }}\left( {{k_n},{k_m}} \right)$, the conductance and shot noise at low temperatures can be calculated as follows\cite{BlanterPR2000}.
\begin{equation}
G = \frac{e^2}{{{h}}}\int_0^{2\pi } {\int_0^{{\pi \mathord{\left/
{\vphantom {\pi 2}} \right.
\kern-\nulldelimiterspace} 2}} {\sum\limits_{\sigma ,\tau ,n} {{{\left| {{\bf{\hat S}}_{\sigma \tau }^{RL}\left( {{k_n},{k_0}} \right)} \right|}^2}} k_F^2\sin {\theta _{in}}d{\theta _{in}}d{\phi _{in}}} } .
\label{G}
\end{equation}
\begin{equation}
S = \frac{{2{e^3}}}{h}\int_0^{2\pi } {\int_0^{{\pi \mathord{\left/
{\vphantom {\pi 2}} \right.
\kern-\nulldelimiterspace} 2}} {\sum\limits_{\sigma ,\tau ,n} {\left\{ {{{\left| {{\bf{\hat S}}_{\sigma \tau }^{RL}\left( {{k_n},{k_0}} \right)} \right|}^2}\left[ {1 - {{\left| {{\bf{\hat S}}_{\sigma \tau }^{RL}\left( {{k_n},{k_0}} \right)} \right|}^2}} \right]} \right\}} k_F^2\sin {\theta _{in}}d{\theta _{in}}d{\phi _{in}}} } .
\label{S}
\end{equation}
Here, $\theta _{in}$ and $\phi _{in}$ are the incident polar and azimuthal angles, respectively. The Fano factor can be defined by $F=S/(2eG)$.
\subsection{Numerical Results and Interpretations}
Numerical results of the diffracted transmission for $q=0.5 k_F$ are shown in Fig. 1. $T_{0,\pm 1}$ is defined as $\sum\limits_{\sigma ,\tau } {{{\left| {{\bf{\hat S}}_{\sigma \tau }^{RL}\left( {{k_{0, \pm 1}},{k_0}} \right)} \right|}^2}} $. The zero-order transmission $T_0$ has spherical symmetry in the incident angle space, namely, in the electron wave vector space. It approaches maximum at normal incidence, which is natural for standard barrier tunneling as the wave vector in the propagating direction and hence the current flux approaches maximum. The $1$ and $-1$ order transmission is at minimum in normal incidence due to the $x$-$z$ plane symmetry of the HM spiral. When the incident angle increases, $T_1$ and $T_{-1}$ increases, approaches maximum and abruptly disappears into evanescent modes for the incident azimuthal angle $\phi _{in}=0$ and polar angle $\theta _{in} =\pi /6$ and $\theta _{in} =- \pi /6$, respectively. The angle can be analytically obtained by the equation of ${k_{z, \pm 1}} = \sqrt {k_F^2 - {{\left( {{k_F}\sin {\theta _{in}}\sin {\phi _{in}}} \right)}^2} - {{\left( {{k_F}\sin {\theta _{in}}\cos {\phi _{in}} \pm 0.5{k_F}} \right)}^2}} =0$. To the other side of the incident sphere, $T_1$ and $T_{-1}$ disappears at the glazing angle.
To investigate the conductance and shot noise properties of the HM tunnel junction, angle averaged quantities of Eqs. (\ref{G}) and (\ref{S}) are shown in Fig. 2. From panel (a), it can be seen that the conductance monotonously increases with the Fermi wave vector $k_F$. It weakly depends on the HM spiral wave vector $q$ and is not visible in the figure. The conductance increase in $k_F$ is due to larger current flux and more contributing channels. The weak dependence on $q$ is due to diffusion of the diffraction effect by angle average. The angle-averaged shot noise as functions of $q$ and $k_F$ is shown in panel (b). Prominent diffraction effect can be seen in the noise spectrum. As a result of diffraction, different diffracted channels interact and give rise to the multiple peaks in the shot noise. The pink and red dotted lines correspond to $q= k_F$ and $q=2 k_F$, respectively. For $q$ larger than $2 k_F$, all diffracted waves degrade into evanescent modes and variation of the shot noise became smooth without abrupt rises and falls. For $q$ larger than $k_F$ and smaller than $2 k_F$, $T_1$ disappears for all positive incident angles and $T_{-1}$ disappears for all negative incident angles. For $q$ smaller than $k_F$, diffracted transmission from part of the incident semi-sphere becomes evanescent and does not contribute to the transport as can be seen in Fig. 1. At the two sides of the pink dotted line in Fig. 2 (b) when $q$ is close to $k_F$, strength of the two diffracted waves $T_1$ and $T_{-1}$ matches each other, giving rise to maximal channel-coherence. Therefore, peaks of the shot noise are dramatically enhanced at the two side of the pink dotted line labeling $q= k_F$. When the absolute value of $k_F$ is small relative to $q$, almost no diffraction channel exists and the shot noise is extremely small. There is an abrupt increase in the shot noise when diffracted channels begin to contribute to the transport.
The relative strength of the shot noise in comparison with the Poisson noise is measured by the Fano factor. In Fig. 2 (c) and (d), we show numerical results of the Fano factor. From panel (c), it can be seen that the absolute value of the Fano factor is smaller than $1/3$ throughout the considered parameter space. As summarized in the Introduction, this regime is between the complete ballistic tunneling regime of $F=0$ and the diffusive tunneling regime of $F=1/3$. Our considered HM tunnel junction is a low $\delta$-barrier with the spiral modulating in the spin space. The $\delta$-barrier strength $V_0=50$ meV$ \cdot $\AA and the HM exchange coupling strength $J=20$ meV$ \cdot $\AA. Therefore, even without the spiral modulation, the transmission is very large and close unity. The spiral modulation trifurcated the incident electron plane wave into the $0$ and $\pm 1$ order diffracted waves. The main $0$-order transmission still governs with $T_0$ approaching $1$ and $T_{\pm 1}$ three to four orders smaller than it. These transmission properties can be clearly seen in Fig. 1. As an effect, the system approximates an open conductor and the Fano factor is very small. However, coherence between different diffraction channels significantly influences the shot noise. There are two prominent phenomena. One is that the shot noise is dramatically suppressed relative to the poisson value. The other is that the shot noise demonstrates rise-and-fall variations. Therefore, the Fano factor is smaller for larger $k_F$ for stronger transmission and more contributing channels. Also oscillations can be seen in the Fano factor specified in the panel (d). The Fano factor oscillates as a function of $q$ for $q<2 k_F$. The oscillation is small in comparison with its decrease as a function of $k_F$ and is not prominently seen in the panel (c). When $q> 2 k_F$ diffraction disappears, the shot noise properties resemble a plain barrier. For small Fermi energies, there is a dividing line in the shot noise between the diffraction affected transport for $q< 2 k_F$ and the ordinary barrier scattering governed transport for $q>2 k_F$, which is already seen in the panel (b). In the panel (d) the black and red star symbols label the dividing position of $q=2 k_F$ for different Fermi energies. The shot noise properties also lend a potential detection of the HM spiral period and spin wave vector.
\section{Investigation of the Multiferroic Helimagnet $\rm{TbMnO_3}$ }
Our theoretical treatment introduced in the previous section can be generalized into arbitrary spin spiral structures with finite thickness. In this section, the multiferroic HM\cite{YamasakiPRL2007} $\rm{TbMnO_3}$ is considered by taking into account the combined effect of its helimagnetism and electric polarization. The helimagnetic structure and electric potential profile are sketched in Fig. 3.
\subsection{Theoretic Formulism}
The spiral magnetic structure can be described as
\begin{equation}
{{\bf{M}}_i} = {{\bf{m}}_b}\cos \left( {2\pi {{\bf{Q}}_m} \cdot {{\bf{R}}_i}} \right) + {{\bf{m}}_c}\sin \left( {2\pi {{\bf{Q}}_m} \cdot {{\bf{R}}_i}} \right).
\label{HelimagnetismTbMnO3}
\end{equation}
It was reported\cite{YamasakiPRL2007} for $\rm{TbMnO_3}$ that ${\bf{m}}_b \approx (0,3.9,0)\mu _B$ and ${\bf{m}}_c \approx (0,0,2.8)\mu _B$ at 15 K. We define ${\bf{m}}_b = (0,m_b,0)\mu _B$ and ${\bf{m}}_c = (0,0,m_c)\mu _B$ to consider their variation with temperature. ${\bf{R}}_i$ extends in the three-dimensional crystal lattice. ${\bf{R}}_i$ and ${\bf{Q}}_m$ can be expressed relatively as ${{\bf{Q}}_m} = \left( {0, \pm 2\pi q/b, 2 \pi/c } \right)$ and ${{\bf{R}}_i} = \left( {{n_a} a,{n_b}b,{n_c}\frac{c}{2}} \right),$ with $n_a$, $n_b$, and $n_c$ arbitrary integers. From experimental results\cite{YamasakiPRL2007}, $q \approx 0.27$. It could be seen from Eq. (\ref{HelimagnetismTbMnO3}) that the magnetic moments are reversed every other $c/2$ layer (see Fig. 3).
Every atom layer in the transport direction ($c/z$) can be treated as a $\delta$-barrier. To consider a finite-thickness HM, we use a transfer matrix technique on multiple $\delta$-barriers. Magnetic spiral variation in the $b/y$-direction can be approximated to be continuous. Hamiltonian in each atomic layer is Eq. (\ref{HMHamiltonian}), with the exchange coupling terms in neighboring atom layers
\begin{equation}
J{{\bf{n}}_r} \cdot {\bf{\sigma }} = J\left[ {\begin{array}{*{20}{c}}
{{ \tilde m_c}\sin \left( {qy} \right)} & { - i{\tilde m_b}\cos \left( {qy} \right)} \\
{i{\tilde m_b}\cos \left( {qy} \right)} & { - {\tilde m_c}\sin \left( {qy} \right)} \\
\end{array}} \right],
\end{equation}
\begin{equation}
J{{\bf{n}}_r} \cdot {\bf{\sigma }} = J\left[ {\begin{array}{*{20}{c}}
{ - {\tilde m_c}\sin \left( {qy} \right)} & {i{\tilde m_b}\cos \left( {qy} \right)} \\
{ - i{\tilde m_b}\cos \left( {qy} \right)} & {{\tilde m_c}\sin \left( {qy} \right)} \\
\end{array}} \right],
\end{equation}
respectively. $\tilde m_b =m_b/ \sqrt {m_b^2+m_c^2}$ and $\tilde m_c =m_c/ \sqrt {m_b^2+m_c^2}$ are altered to normalize ${\bf{n}}_r$.
Ferroelectric polarization and helimagnetism coexist and are related in the multiferroic HM $\rm{TbMnO_3}$. Dielectric constant is reported\cite{Ref18} to be $\varepsilon = 30$. As a result of the screening charges, the potential difference in the $c/z$-direction between the two interfaces of the HM could be approximated to be $\Delta U=P_c d/(\varepsilon_0 \varepsilon)$, with $P_c$ $c$-component of the electric polarization (it has only $c$-component) and $d$ thickness of the HM. Experimental data of $P_c$ from Ref. \onlinecite{YamasakiPRL2007} are used.
To consider scattering through an arbitrary atom layer approximated into a $\delta$-barrier, the flux-normalized wave functions at the two interfaces could be written as
\begin{equation}
\left\{ {\begin{array}{*{20}{c}}
{{\psi _I (y,z)} = \sum\limits_{n,\sigma } {\left( {\frac{{A_{n\sigma }^i}}{{\sqrt {{k_n}} }}{e^{i{q_n}y + i{k_n}z}}{\chi _\sigma } + \frac{{A_{n\sigma }^o}}{{\sqrt {{k_n}} }}{e^{i{q_n}y - i{k_n}z}}{\chi _\sigma }} \right)} ,} \hfill & {z < 0,} \hfill \\
{{\psi _{II} (y,z)} = \sum\limits_{n,\sigma } {\left( {\frac{{B_{n\sigma }^i}}{{\sqrt {{k_n}} }}{e^{i{q_n}y - i{k_n}z}}{\chi _\sigma } + \frac{{B_{n\sigma }^o}}{{\sqrt {{k_n}} }}{e^{i{q_n}y + i{k_n}z}}{\chi _\sigma }} \right)} ,} \hfill & {z > 0.} \hfill \\
\end{array}} \right.
\label{DiffractedWaveFunctions}
\end{equation}
Since the real position of a $\delta$-barrier does not affect its transmission, here we assume it locate at $z=0$. In Eq. (\ref{DiffractedWaveFunctions}), summation is over all diffraction and spin channels, ${q_n} = {k_y} + n\tilde q$, and ${k_n} = \sqrt {k_F^2 - k_x^2 - q_n^2} $ with $k_F$ the Fermi wave vector and $\tilde q =2 \pi q/b$. $k_x$ is conserved in transmission due to translational invariance in the $a/x$ crystal direction. $\chi _{\sigma}$ are arbitrary eigenspinors and we set them to be ${\chi _ \uparrow } = {\left( {\begin{array}{*{20}{c}}
1 & 0 \\
\end{array}} \right)^{\rm{T}}}$ and ${\chi _ \downarrow } = {\left( {\begin{array}{*{20}{c}}
0 & 1 \\
\end{array}} \right)^{\rm{T}}}$ for algebra simplicity. $A_{n\sigma }^i$($B_{n\sigma }^i$) and $A_{n\sigma }^o$($B_{n\sigma }^o$) are the incident and outgoing probability amplitudes at the $z<0$($z>0$) interfaces, respectively. A cutoff of ${n_{\max }} =8 > 2{k_F}/\tilde q$ secures accuracy in our numerical treatment.
By continuity at the $z=0$ HM interface
\begin{equation}
{\psi _I}\left( {y,{0^ - }} \right) = {\psi _{II}}\left( {y,{0^ + }} \right),
\end{equation}
and
\begin{equation}
\frac{{{\hbar ^2}}}{{2{m_e}}}{\left. {\frac{{\partial {\psi _I}}}{{\partial z}}} \right|_{z = {0^ - }}} + \left( {{V_0} + J{{\bf{n}}_r} \cdot {\bf{\sigma }}} \right){\psi _I}\left( {y,{0^ - }} \right) = \frac{{{\hbar ^2}}}{{2{m_e}}}{\left. {\frac{{\partial {\psi _{II}}}}{{\partial z}}} \right|_{z = {0^ + }}},
\end{equation}
we could obtain the following matrix element equations
\begin{equation}
A_{n\sigma }^i + A_{n\sigma }^o = B_{n\sigma }^i + B_{n\sigma }^o,
\label{ContinuityOne}
\end{equation}
and
\begin{equation}
\begin{array}{l}
i\sqrt {{k_n}} \left( {B_{n\sigma }^o - B_{n\sigma }^i + A_{n\sigma }^o - A_{n\sigma }^i} \right) \\
= \sum\limits_{m,\sigma '} {\left[ {\left( {\frac{{{V_0}{\delta _{nm}}{\delta _{\sigma \sigma '}}}}{{\sqrt {{k_n}} }} + \frac{{{J_{ - 1\sigma \sigma '}}{\delta _{n - 1,m}}}}{{\sqrt {{k_{n - 1}}} }} + \frac{{{J_{ + 1\sigma \sigma '}}{\delta _{n + 1,m}}}}{{\sqrt {{k_{n + 1}}} }}} \right)\left( {B_{m\sigma '}^i + B_{m\sigma '}^o} \right)} \right]} , \\
\end{array}
\label{ContinuityTwo}
\end{equation}
with matrices operating on the spin space
\begin{equation}
{J_{ - 1}} = J\left( {\begin{array}{*{20}{c}}
{ - \frac{{{{\tilde m}_c}}}{{2i}}} & {\frac{{i{{\tilde m}_b}}}{2}} \\
{ - \frac{{i{{\tilde m}_b}}}{2}} & {\frac{{{{\tilde m}_c}}}{{2i}}} \\
\end{array}} \right),\begin{array}{*{20}{c}}
{} & {{J_{ + 1}} = J\left( {\begin{array}{*{20}{c}}
{\frac{{{{\tilde m}_c}}}{{2i}}} & {\frac{{i{{\tilde m}_b}}}{2}} \\
{ - \frac{{i{{\tilde m}_b}}}{2}} & { - \frac{{{{\tilde m}_c}}}{{2i}}} \\
\end{array}} \right).} \\
\end{array}
\label{JPL1}
\end{equation}
With Eqs. (\ref{ContinuityOne}) and (\ref{ContinuityTwo}), we could obtain the transfer matrix connecting the probability amplitudes on the two interfaces of the HM $\delta$-barrier as
\begin{equation}
\left( {\begin{array}{*{20}{c}}
{{B^i}} \\
{{B^o}} \\
\end{array}} \right) = \left( {\begin{array}{*{20}{c}}
{{M_{ii}}} & {{M_{io}}} \\
{{M_{oi}}} & {{M_{oo}}} \\
\end{array}} \right)\left( {\begin{array}{*{20}{c}}
{{A^i}} \\
{{A^o}} \\
\end{array}} \right),
\label{DefiningSingleLayerTransferMatrix}
\end{equation}
in which $A^{i/o}$ and $B^{i/o}$ are matrices made up of the precious elements and a Kronecker product between the diffraction and spin spaces.
The total transfer matrix of multiple $\delta$-barrier follows as
\begin{equation}
{M_f} = \prod\limits_{i = 1}^N {{M_i}} = {M_N} \cdot \ldots \cdot {M_2} \cdot {M_1},
\end{equation}
with
\begin{equation}
{M_i} = \left( {\begin{array}{*{20}{c}}
{{M_{ii}}} & {{M_{io}}} \\
{{M_{oi}}} & {{M_{oo}}} \\
\end{array}} \right)
\end{equation}
for odd $i$'s. $M_i$ with even $i$'s can be obtained by changing $J$ into $-J$ of Eq. (\ref{JPL1}). Secured by the flux normalization of the eigen-spinor wave functions, the scattering matrix ${\bf{\hat S}}$ defined in Eq. (\ref{DefiningBigScatteringMatrix}) can be derived from the transfer matrix $M_f$ as
\begin{equation}
{\bf{\hat S}} = \left( {\begin{array}{*{20}{c}}
{{M_{aa}}} & {{M_{ab}}} \\
{{M_{ab}}} & {{M_{bb}}} \\
\end{array}} \right),
\end{equation}
with
\begin{equation}
\left\{ \begin{array}{l}
{M_{aa}} = - {\left( {{M_{fio}}} \right)^{ - 1}}{M_{fii}}, \\
{M_{ab}} = {\left( {{M_{fio}}} \right)^{ - 1}}, \\
{M_{ba}} = {M_{foi}} - {M_{foo}}{\left( {{M_{fio}}} \right)^{ - 1}}{M_{fii}}, \\
{M_{bb}} = {M_{foo}}{\left( {{M_{fio}}} \right)^{ - 1}}. \\
\end{array} \right.
\end{equation}
Here, relation (\ref{DefiningSingleLayerTransferMatrix}) also holds for $M_f$ connecting probability amplitudes of the beginning and ending interfaces of the total $N$-layer scattering. The total transmission probability can be defined as
$\sum\limits_{\sigma ,\tau ,n} {{{\left| {\hat S_{\sigma \tau }^{RL}\left( {{k_n},{k_0}} \right)} \right|}^2}} $.
To consider the helimagnetism variation in temperature, the conductance, bias current, and noise can be expressed as follows,
\begin{equation}
G = \frac{{{e^2}}}{h}\int_{ - \infty }^{ + \infty } {\int_0^{2\pi } {\int_0^{{\pi \mathord{\left/
{\vphantom {\pi 2}} \right.
\kern-\nulldelimiterspace} 2}} {\sum\limits_{\sigma ,\tau ,n} {{{\left| {{\bf{\hat S}}_{\sigma \tau }^{RL}\left( {{k_n},{k_0}} \right)} \right|}^2}} \frac{{df\left( E \right)}}{{dE}}k_F^2\sin {\theta _{in}}d{\theta _{in}}d{\phi _{in}}} } dE} ,
\end{equation}
\begin{equation}
I = \frac{e}{h}\int_{ - \infty }^{ + \infty } {\int_0^{2\pi } {\int_0^{{\pi \mathord{\left/
{\vphantom {\pi 2}} \right.
\kern-\nulldelimiterspace} 2}} {\sum\limits_{\sigma ,\tau ,n} {{{\left| {{\bf{\hat S}}_{\sigma \tau }^{RL}\left( {{k_n},{k_0}} \right)} \right|}^2}} \left[ {f\left( {E - eV} \right) - f\left( E \right)} \right]k_F^2\sin {\theta _{in}}d{\theta _{in}}d{\phi _{in}}} } dE} ,
\end{equation}
\begin{equation}
S = \frac{{{e^2}}}{{2\pi \hbar }}\sum\limits_{\gamma \delta } {\sum\limits_{mn} {\sum\limits_{{\sigma _1}{\sigma _2}} {\int {dEd{\theta _{in}}d{\phi _{in}}A_{\gamma \delta {\sigma _1}{\sigma _2}}^{mn}\left( {L;E} \right)A_{\delta \gamma {\sigma _2}{\sigma _1}}^{nm}\left( {L;E} \right)} } } } ,
\end{equation}
with
\begin{equation}
A_{\gamma \delta {\sigma _1}{\sigma _2}}^{mn}\left( {L;E} \right) = {\delta _{mn}}{\delta _{{\sigma _1}{\sigma _2}}}{\delta _{L\gamma }}{\delta _{L\delta }} - \sum\limits_{k,{\sigma _3}} {s_{\gamma L;km;{\sigma _3}{\sigma _1}}^\dag \left( E \right){s_{L\delta ;nk;{\sigma _2}{\sigma _3}}}\left( E \right)} .
\end{equation}
The the equilibrium, or Nyquist--Johnson noise is $S_t=4 k_B T G$.
Therefore the Fano factor is defined as $F=(S-S_t)/(2eI)$. To prominently see the quantum transport effect, we define the relative conductance $G_r$ as
\begin{equation}
{G_r} = \frac{G}{{\int_{ - \infty }^{ + \infty } {\frac{{df\left( E \right)}}{{dE}}dE} }}.
\label{RelativeConductance}
\end{equation}
\subsection{Numerical Results and Interpretations}
To obtain results comparable with the experiment, we used real parameters directly simulated from Ref. \onlinecite{YamasakiPRL2007}. The simulated ferroelectric polarization $P_c$ and spiral ellipticity $m_c / m_b$ as a function of the temperature are shown in Fig. 4. From Ref. \onlinecite{YamasakiPRL2007}, it can be seen that $m_b$ linearly varies with the temperature. $m_c$ is nonzero for temperatures below the HM transition temperature $T_C =27 $ K and has a power law relation to the temperature. By fitting to the experimental data, we assume
\begin{equation}
m_b=-0.14 T+6.07,
\end{equation}
and
\begin{equation}
{m_c} = \left\{ {\begin{array}{*{20}{c}}
{\frac{{{{\left( {27 - T} \right)}^{{1 \mathord{\left/
{\vphantom {1 2}} \right.
\kern-\nulldelimiterspace} 2}}}}}{{{m_b}}},} \hfill & {T < {T_C}} \hfill \\
{0,} \hfill & {{T_C} < T < {T_N},} \hfill \\
\end{array}} \right.
\end{equation}
and
\begin{equation}
P_c = A \sin (2\pi qb) m_b m_c,
\end{equation}
with the temperature $T$ measured in degree kelvin and $T_N = 42$ K is the long-range-spin-order-emerging temperature. $A=40$ $\mu $F is an experimentally determined constant. The ferroelectric polarization generates screening charges on the interfaces between the HM and the metal electrodes, which results in an electric field and electric voltage difference. We divide the voltage increase (decrease) into ultrathin steps with one layer corresponding to a real atomic layer illustrated in Fig. 3 (b) and the voltage change between neighboring atomic layers $\Delta U=P_c d/(\varepsilon_0 \varepsilon)$.
Numerical results of the relative conductance defined in Eq. (\ref{RelativeConductance}) are shown in Fig. 4 (c). By defining the relative conductance, the thermal effect is minimized and the quantum effect is manifested. It can be seen from the figure that variation of the $G_r$ follows the pattern of that of the HM spiral ellipticity. For temperatures above $T_C$ and below $T_N$, $\rm {TbMnO_3}$ is ferromagnetic with its magnetic polarization in the $b/y$ direction. For temperatures below $T_C$, the conductance is suppressed by the HM diffraction. Although at some incident angles the transmission is enhanced by the HM diffraction, it is more strongly suppressed at others. As their combined effect, the angularly-averaged conductance is smaller in the HM phase than in the ferromagnetic phase. Also affected by the ferroelectric polarization, the conductance is larger for decreasing electric potential than increasing electric potential.
It should be noted that we consider an HM-layer with the thickness $d=3.2$ nm, which consists of thirty-one semi-atomic layer in the $c/z$-direction. The direction of the HM spin-field is opposite between two neighboring semi-atomic layers, which is illustrated in Fig. 3. The travelling beam is diffracted once more during the scattering by each semi-atomic layer. Our numerical results demonstrates that the effect of the diffraction over diffraction is simple and not divergent or chaotic. This property at least to some extent justifies our theory and numerical techniques. Also the physics underlying the conductance shown in Fig. 4 and the total transmission shown in Fig. 5 is clear and meaningful.
Properties of the conductance can be illustrated by the total transmission probabilities shown in Fig. 5. The transmission demonstrates diffraction features, i.e., maximums and minimums occur at certain incident angles. Variation of the transmission probabilities as a function of the incident azimuthal angle $\phi _{in}$ differs dramatically for different polar angles $\theta _{in}$. For $\theta _{in} =0$, the transmission is constant for all $\phi _{in}$ due to the symmetry of the HM configuration. The transmission is also symmetric between $\phi _{in} < \pi$ and $\phi _{in} > \pi$ for all $\theta _{in}$. The transmission is largest in normal incidence of $\theta _{in}=0$ and smallest in nearly grazing incidence of $\theta _{in} =1.5$ radian for all incident azimuthal angles $\phi _{in}$. As a result of the diffraction by the HM spiral, maximums and minimums occur. The angularly averaged conductance is suppressed when the total contribution from wavy angular distribution of the transmission is smaller than a constant transmission.
Numerical results of the electrical current, shot noise, and Fano factor at zero temperature are shown in Fig. 6. It is natural that both the electrical current and the shot noise increase with the bias voltage because of more energy channels contributing to the transport. The current and shot noise for different HM spiral helicity characterized by $P_c >0$ and $P_c <0$ are nearly the same. However, the Fano factor decreases with the bias voltage and the two HM helical states are distinctly separated in the Fano factor. The Fano factor indicates the coherent correlation between transport channels. The Fano factor is larger for uncorrelated single-channel transport and smaller for correlated multi-channel transport varying in the range between 0 and 1 for ideal quantum coherent tunneling\cite{BlanterPR2000}. The quantum correlation between different diffraction orders during transport further suppresses the shot noise in addition to the Pauli exclusion and interference in the orbital and spin degrees of freedom compared with the Poisson noise. The Fano factor decreases with the bias voltage because of the strengthening of the correlation between different orbital channels when larger $V$ opens up a wider tunneling window in energy. The difference in the Fano factor between the $P_c>0$ and $P_c<0$ states is a direct effect of the ferroelectricity. The total transmission probability for $P_c>0$ is larger than that for $P_c<0$, which is shown in Fig. 5. As a result, interference is stronger for $P_c>0$ than $P_c<0$, giving rise to a smaller Fano factor.
Numerical results of the electrical current, noise, and the Fano factor as a function of the temperature at a fixed bias voltage of 35 mV are shown in Fig. 7. As an effect of the finite temperature, the noise is made up of the shot noise and the thermal noise. Variation of the finite-bias-voltage electrical current $I$ as a function of the temperature follows a similar pattern to the conductance $G_r$. The properties of the electrical current $I$ are governed by the quantum effect of the scattering by the HM-barrier. When the temperature is decreased, the HM spiral ellipticity changes and the electrical current changes with it. For temperatures above $T_C$ and below $T_N$, $\rm {TbMnO_3}$ is ferromagnetic polarized in the $b/y$ direction. At this temperature range, the ferroelectric polarization is absent and no diffraction occurs. Therefore the current does not change with the temperature and the two cases of $P_c>0$ and $P_c<0$ are not separated. For temperatures below $T_C$, the current is suppressed by the HM diffraction and the two cases of $P_c>0$ and $P_c<0$ are differentiated solely by the electric effect.
In Fig. 7 (b), numerical results of the noise are provided. It can be seen that in the finite-temperature case the thermal effect in the noise is strong and over the quantum effect. The thermal effect could be partially suppressed by removing the thermal noise $S_t$ in defining the Fano factor. The Fano factor in Fig. 7 (c) is much larger than 1 due to contributions made by the Fermi distribution function numerically. The difference between the two cases of $P_c>0$ and $P_c<0$ in the noise and Fano factor is a combined effect of the ferroelectric polarization and the diffraction by the HM-barrier.
\section{Conclusions}
In conclusion, we theoretically investigated the conductance and shot noise properties in the HM tunnel junction. We separately considered a single-HM-layer toy model and a multi-HM-layer real model based on the experimental data of the $\rm{TbMnO_3}$ multiferroic HM material. With the developed scattering matrix scheme, the general procedure and formulas to calculate the diffracted transmission probabilities, the conductance, and the shot noise are given. Numerical results of the single-HM-layer toy model show that at a certain incident angle, one of the diffracted waves $T_1$ or $T_{-1}$ disappears into an evanescent mode and that transmission coefficient abruptly falls into zero giving rise to prominent rise-and-fall oscillations in the angle-averaged shot noise. Two dividing lines of $q = k_F$ and $q= 2 k_F$ characterize the regimes that one or both of the diffracted waves complete disappear into evanescent modes for all incident angles. When $q> 2 k_F$, the shot noise properties resemble that of a plain $\delta$-barrier. At the two sides of $q= k_F$, strength of the two diffracted channel matches each other giving rise to strong oscillation in the shot noise. Due to correlation among different diffracted channels, the shot noise are additionally suppressed relative to the Poisson value. The diffraction-affected transport properties are prominently demonstrated in the shot noise and Fano factor. Numerical results of the multi-HM-layer real model of $\rm{TbMnO_3}$ show that clockwise and counterclockwise spin helix is distinctly separated in the conductance and shot noise spectrum. Variation of the electrical conductance and the finite-bias current as a function of the temperature resembles the variation pattern of the HM spiral ellipticity. The shot noise is further suppressed by the quantum correlation between different diffraction orders during transport in addition to the Pauli exclusion and interference in the orbital and spin degrees of freedom compared with the Poisson noise. The current and shot noise properties are a combined result of helimagnetism and electric polarization.
\section{Acknowledgements}
This project was supported by the National Natural Science
Foundation of China (No. 11004063) and the Fundamental Research
Funds for the Central Universities, SCUT (No. 2014ZG0044).
\clearpage
|
\section{Introduction}
An important emerging scientific discipline is to quantify the evolution of low-order statistics of dynamical systems in the presence of uncertainties due to errors in modeling, numerical approximations, parameters, and initial and boundary conditions \cite{smith:13}. While many uncertainty quantification (UQ) methods have been proposed, they rely on some knowledge about the underlying dynamics, at least at the coarse grained levels. One class of popular UQ methods is low-order truncation modeling based on projection of the dynamics onto some basis \cite{smith:13,mk:10}, such as the principal component analysis \cite{jolliffe:86,kutz:13}, the truncated polynomial chaos basis \cite{najm:09,mk:10}, the dynamically orthogonal basis \cite{sl:09}, the Nonlinear Laplacian Spectral Analysis (NLSA) basis \cite{gm:12}, and recently, the ``ROMQG" which stands for Reduced Order Modified Quasilinear Gaussian" method \cite{sm:13} which was carefully designed for predicting statistics of turbulent systems.
This paper considers UQ problems in which one has no explicit parametric form for the underlying dynamics. Given only a time series of the underlying dynamics, our goal is to devise UQ methods based on nonparametric modeling, applying ideas from diffusion maps \cite{BN,diffusion,VB}. While the classical Cauchy problem for solving the backward Kolmogorov equation is to find the semigroup solutions of this PDE, the diffusion maps technique can be interpreted as the inverse of this Cauchy problem. Namely, given a realization of the associated stochastic process, we construct a stochastic matrix (also known as Markov matrix), whose generator is a discrete approximation to the backward Kolmogorov operator. Numerically, the stochastic matrix is constructed by evaluating an appropriate kernel function on all pairs of data points and carefully re-normalizing the resulting matrix to account for sampling bias in the data set and to insure the Markov property. In this paper, we apply the recently developed variable bandwidth diffusion kernel \cite{VB}, which generalizes the original result of diffusion maps introduced in \cite{diffusion} to non-compact manifolds. Our choice is motivated by the fact that this variable bandwidth kernel is more robust and it significantly improves the operator estimation in areas of sparse sampling, e.g. on the tail of a distribution \cite{VB}. The UQ methods discussed in this paper are only applicable for a class of nonlinear dynamical systems which can be described by stochastically forced gradient flows, since this is the class of problems for which one can learn the (backward) Kolmogorov operator from the diffusion maps algorithm \cite{diffusion,VB}. In the conclusion we discuss the possibility of applying the UQ framework developed here to more general systems.
The representation of the probabilistic forecasting problem in diffusion coordinates was first noted in \cite{gradFlow1,gradFlow2}, however this forecasting approach was never demonstrated. In fact, such forecasting would have been difficult for the non-compact examples which we will consider, because they used fixed bandwidth kernels in \cite{gradFlow1,gradFlow2} which would not recover the true Kolmogorov operator, as shown in \cite{VB}. In fact, the goal of \cite{gradFlow1,gradFlow2} was not to solve the forecasting problem, but instead to a low-dimensional representations of complex dynamical systems. We should note that a `locally-scaled' version of the approach in \cite{gradFlow1,gradFlow2} was taken in \cite{gradFlow3}, which used an ad-hoc variable bandwidth function. While the ad-hoc bandwidth of \cite{gradFlow3} is valid for finding low-dimensional coordinates, it would not be useful for the forecasting problem since, as shown in \cite{VB}, the bandwidth function would affect the estimated operator. The previous work in \cite{gradFlow1,gradFlow2,gradFlow3} all focused on using diffusion coordinates as a nonlinear map to obtain a low-dimensional representation of the observed dynamics. In this paper we take a significantly different perspective, namely, we treat the diffusion coordinates as a basis for smooth functions and we represent probability densities in this basis for the purposes of UQ.
We will consider three UQ problems, namely the problems of forecasting, filtering \cite{mh:12}, and response \cite{mag:05}. While most stochastic projection methods chose the projection basis coordinate \cite{smith:13,mk:10} depending on the nature of the problems (such as the geometry of the state space or the distribution of the parameters), here we consider solving these UQ problems in the diffusion coordinates. For diffusion processes on nonlinear manifolds, this data-driven basis is the most natural choice since these eigenfunctions implicitly describe both the dynamics and geometry of the underlying dynamical system. For example, in this coordinate basis, the forecasting problem reduces to solving a diagonal infinite-dimensional systems of linear ODEs, as originally noted in \cite{gradFlow1}. For continuous-time nonlinear filtering problems, the solutions are characterized by the unnormalized conditional distribution which solves the Zakai equation \cite{bc:09}. In diffusion coordinates, we shall see that this nonlinear filtering problem reduces to solving a system of infinite-dimensional multiplicatively forced linear stochastic differential equations. Finally, we consider the response problem studied in \cite{mag:05}, which applied the fluctuation-dissipation theorem to estimate the statistical linear response of perturbed systems given the data set of the unperturbed system and the functional form of the external perturbation. Finding the response in our framework requires the diffusion coordinates of the perturbed system. By assuming the perturbation is given by a known change in the potential function, we show that this perturbed basis can be constructed using an appropriate kernel, evaluated on unperturbed data. In this basis, the corresponding nonlinear response problem reduces to solving an infinite dimensional linear system of ODEs. In each case described above, the nonparametric UQ algorithms consist of solving truncated systems of ODEs and SDEs with finitely many diffusion coordinates.
This paper will be organized as follows: In Section~\ref{DM}, we briefly review the diffusion maps and several computational manipulations for obtaining the appropriate nonparametric probabilistic models by projecting the operators to a basis of the data set itself. In Section~\ref{UQ}, we formulate two UQ problems (prediction and filtering) in the diffusion coordinates. In Section~\ref{sec4}, we will formulate the third UQ problem (nonlinear response) in the diffusion coordinates. In Section~\ref{sec5}, we verify our numerical methods on linear and nonlinear examples. We conclude the paper with a short summary and discussion in Section~\ref{sec6}. We accompany this paper with several movies in the electronic supplementary materials, depicting the evolution of the estimated time-dependent distributions.
\section{Diffusion Kernels as Nonparametric Models}\label{DM}
Consider a state variable $x \in \mathcal{M}$ evolving on a $d$-dimensional manifold $\mathcal{M} \subset \mathbb{R}^n$ according to the stochastic gradient flow,
\begin{align}\label{SDE} dx = -\nabla U(x)dt + \sqrt{2D}dW_t, \end{align}
where $U(x)$ denotes the potential of the vector field at $x\in\cal{M}$ and $D$ is a positive scalar that characterizes the amplitude of a $d-$dimensional white noise $dW_t$ on the manifold $\cal M$. Given data $\{x_i\}_{i=1}^N$ sampled independently from the invariant measure of \eqref{SDE}, $p_{{\rm eq}}(x) \propto \exp(-U/D)$, our goal is to approximate the generator, \begin{align}\label{BKop} \mathcal{L} = D\Delta - \nabla U \cdot \nabla , \end{align}
with a stochastic matrix constructed from the data set. In order to estimate $p_{{\rm eq}}$ from the data set, we will assume that \eqref{SDE} is ergodic, and for estimating the coefficient $D$ we will require the system to be wide-sense stationary as well. The Markov matrix that we will construct is a nonparametric model in the sense that the structure of the manifold $\mathcal{M}$ and the form of the potential function are not assumed to be known. In \eqref{BKop}, $\Delta$ is the Laplacian (with negative eigenvalues), and $\nabla$ is the gradient operator on $\cal M$ with respect to the Riemannian metric inherited from the ambient space $\mathbb{R}^n$.
The diffusion maps algorithm \cite{diffusion} generates a stochastic matrix by evaluating a fixed bandwidth isotropic kernel function on all pairs of data points and then renormalizing the matrix and extracting the generator. When the data set lies on a compact manifold $\cal M$, the resulting matrix converges to $D^{-1}\mathcal{L}$ in the limit of large data for any smooth potential $U$. Recently, this result was extended by both authors to non-compact manifolds by using a variable bandwidth kernel \cite{VB}. They showed that the variable bandwidth kernels produce significantly improved estimation in areas of sparse sampling (e.g. such as rare events which occur on the tail of a distribution) and are less dependent on the choice of bandwidth. As mentioned in the introduction, we will consider the variable bandwidth kernels \cite{VB} as the key ingredient in constructing the nonparametric models in this paper.
In Section \ref{findingBK} below, we briefly review the methodology in \cite{VB} for approximating the generator $D^{-1}\mathcal{L}$. Since our goal is to approximate $\mathcal{L}$, then we need to determine $D$ from the data. In Section~\ref{findingD} will provide one method to determine $D$ using the correlation time of the data and we also mention several other methods to determine $D$.
\subsection{Approximating the Generator of Stochastic Gradient Flows}\label{findingBK}
The key to our approach is the intuition that continuous notions such as functions and operators have discrete representation in the basis of the data set itself. Given a data set $\{x_i\}_{i=1}^N\subset \mathcal{M}\subset \mathbb{R}^n$, sampled independently from the invariant measure $p_{{\rm eq}}(x)$, a function $f$ is represented by a vector $\vec f = (f(x_1),f(x_2),...,f(x_N))^\top$. Similarly, an integral operator $G_{\epsilon} f(x) \equiv \int_{\mathcal{M}} K(x,y)f(y)p_{{\rm eq}}(y)dV(y)$, where $dV(y)$ denotes the volume form on $\cal M$, is represented by a matrix-vector multiplication between the $N\times N$ matrix, $K_{ij} = K(x_i,x_j)$ and the $N$-dimensional vector, $\vec f$. With these definitions, the matrix product $K\vec f$ yields a vector of length $N$ with $i$-th component,
\begin{align}\frac{1}{N}\left(K \vec f \right)_i = \frac{1}{N}\sum_{j=1}^N K_{ij}\vec f_j = \frac{1}{N}\sum_{j=1}^N K(x_i,x_j)f(x_j) \stackrel{N\rightarrow\infty}{\longrightarrow} G_\epsilon f(x_i), \nonumber \end{align}
where the limit follows from interpreting the summation as a Monte-Carlo integral. Thus, the matrix $K$ takes functions defined on $\{x_i\}$ to functions defined on $\{x_i\}$, so in this sense we think of $K$ as an operator written in the basis of delta functions $\{\delta_{x_i}\}$ on the data set. Notice the crucial fact that the data is not uniformly distributed on $\cal M$, so the operator is biased by the sampling measure $p_{{\rm eq}}(y)$. This same bias applies to inner products, if $f,g$ are functions on $\cal M$ and $\vec f, \vec g$ are their representations as vectors evaluated at $\{x_i\}$, then the dot product has the following interpretation,
\begin{align} \frac{1}{N}\vec f \cdot \vec g = \frac{1}{N} \sum_{i=1}^N f(x_i)g(x_i) \stackrel{N\rightarrow\infty}{\longrightarrow} \int_{\mathcal{M}}f(y)g(y)p_{{\rm eq}}(y)dV(y) \equiv \langle f,g\rangle_{L^2(\mathcal{M},p_{{\rm eq}})}. \nonumber \end{align}
The previous formula shows that inner products weighted by the sampling measure $p_{{\rm eq}}$ will be easy to compute in this framework.
With the above intuition in mind, we construct a matrix $L_{\epsilon}$ that will converge to the generator $D^{-1}\mathcal{L}$, where $\cal L$ is defined in \eqref{BKop}, in the limit as $N\to\infty$ in the sense that,
\[\lim_{N\to\infty} \sum_{j=1}^N (L_{\epsilon})_{ij}\vec f_j = D^{-1}\mathcal{L}f(x_i) + \mathcal{O}(\epsilon).\]
The theory developed in \cite{VB} shows that such an $L_\epsilon$ can be constructed with a variable bandwidth kernel,
\begin{align}
K_{\epsilon}(x,y) &= \exp\left\{\frac{-\|x-y\|^2}{4\epsilon \rho(x)\rho(y)}\right\},\label{VBkernel}
\end{align}
where the bandwidth function, $\rho$, is chosen to be inversely proportional to the sampling density, $\rho \approx p_{{\rm eq}}^{-1/2}$. This choice enables us to control the error bounds in the area of sparse sample (see \cite{VB} for the detailed error estimates). Using this kernel requires us to first estimate the sampling density $p_{{\rm eq}}$ in order to define the bandwidth function $\rho$. There are many methods of estimating $p_{{\rm eq}}$, such as kernel density estimation methods \cite{RosenblattFBK,ParzenFBK,ScottVBK,ScottVBK2}. In this paper, we estimate $p_{{\rm eq}}$ using a variable bandwidth kernel method based on the distance to the nearest neighbors (see Section~4 of \cite{VB} for details).
Given the kernel in \eqref{VBkernel}, we apply the following normalization (first introduced in \cite{diffusion}) to remove the sampling bias,
\begin{align}
K_{\epsilon,\alpha}(x_i,x_j) &= \frac{K_{\epsilon}(x_i,x_j)}{q_{\epsilon}(x_i)^{\alpha}q_{\epsilon}(x_j)^{\alpha}} \mbox{, \quad where}\quad q_{\epsilon}(x_i) = \sum_{j=1}^N \frac{K_{\epsilon}(x_i,x_j)}{\rho(x_i)^d}. \nonumber
\end{align}
Note that $q_{\epsilon}$ is a kernel density estimate of the sampling distribution $p_{{\rm eq}}$ based on the kernel $K_{\epsilon}$ and this is the estimate which will be used in all our numerical examples below. Throughout this paper we will use the normalization with $\alpha=-d/4$ as suggested by the theory of \cite{VB}, where $d$ is the intrinsic dimension of the manifold $\mathcal{M}$. We note that $d$ and $\epsilon$ can be estimated from data using the method described in \cite{VB,BGH14}. The next step is to construct a Markov matrix from $K_{\epsilon,\alpha}$ by defining,
\begin{align}
\hat K_{\epsilon,\alpha}(x_i,x_j) = \frac{K_{\epsilon,\alpha}(x_i,x_j)}{q_{\epsilon,\alpha}(x_i)}\mbox{, \quad where}\quad q_{\epsilon,\alpha}(x_i) = \sum_{j=1}^N K_{\epsilon,\alpha}(x_i,x_j). \nonumber
\end{align}
Finally, the discrete approximation to the continuous generator $D^{-1}\cal L$ is given by,
\begin{align}
L_{\epsilon}(x_i,x_j) &= \frac{\hat K_{\epsilon,\alpha}(x_i,x_j)-\delta_{ij}}{\epsilon\rho(x_i)^2}.\label{approxgenerator}
\end{align}
By computing the eigenvectors $\vec \varphi_i$ and eigenvalues $\lambda_i$ of $L_{\epsilon}$, for $i=0,1,\ldots$, we are approximating the eigenfunctions and eigenvalues of $D^{-1}\mathcal{L}$, respectively. Since $D$ is a scalar constant, obviously, the eigenvectors $\vec \varphi_i$ and eigenvalues $D\lambda_i$ are discrete approximations of the eigenfunctions and eigenvalues of the continuous generator $\cal L$, respectively. We should note that this trivial argument does not hold when $D$ is a general non-diagonal matrix describing anisotropic diffusion on $\cal M$ with respect to the Riemannian metric inherited from the ambient space $\mathbb{R}^n$. In that case, we suspect that one must use a different kernel function, such as the Local Kernels defined in \cite{bs:14} which we discuss in the conclusion.
In addition to the generator $\mathcal{L}$, we can also approximate the solution semi-group $e^{\epsilon\mathcal{L}}$ by the matrix,
\[ F_{\epsilon} \equiv I_{N\times N} + \epsilon DL_{\epsilon}, \]
where $F_{\epsilon}$ has the same eigenvectors as $L_{\epsilon}$ but with eigenvalues $\xi_i = 1+\epsilon D\lambda_i$ where $\lambda_i$ are the eigenvalues of $L_{\epsilon}$.
For arbitrary $t$ we can approximate the eigenvalues of the solution semi-group $e^{t\mathcal{L}}$ by $\xi_i^{t/\epsilon}$ which converges to $e^{D\lambda_i t}$ as $\epsilon\rightarrow 0$.
In order to insure that the eigenvectors $\vec \varphi_i$ (which approximate the eigenfunctions $\varphi_i$) form an orthonormal basis with respect to the inner product $\langle \cdot,\cdot\rangle _{L^2(\mathcal{M},p_{{\rm eq}})}$, we note that,
\[ \delta_{ij} = \langle \varphi_i,\varphi_j \rangle_{L^2(\mathcal{M},p_{{\rm eq}})} = \int_{\mathcal{M}} \varphi_i(x)\varphi_j(x)p_{{\rm eq}}(x)dV(x) = \lim_{N\to\infty} \frac{1}{N} \sum_{l=1}^N (\vec \varphi_i)_l (\vec \varphi_j)_l = \lim_{N\to\infty} \frac{ \vec \varphi_i^\top \vec \varphi_j}{N}. \]
Since $\{\vec \varphi_i\}_{i=0}^\infty$ are eigenvectors, they are already orthogonal, thus we only need to renomalize $\vec \varphi_i$ so that $\vec \varphi_i^\top \vec\varphi_i = N$. To do this we simply replace $\vec \varphi_i$ with $\frac{\sqrt{N}}{||\vec\varphi_i||}\vec\varphi_i$. To simplify the notation below, we define $\langle \cdot,\cdot\rangle _{p_{{\rm eq}}}\equiv\langle \cdot,\cdot\rangle _{L^2(\mathcal{M},p_{{\rm eq}})}$.
\subsection{Determining the Diffusion Coefficient}\label{findingD}
The algorithm described in Section \ref{findingBK} approximates the generator $D^{-1}\mathcal{L}$, where $\cal L$ is defined in \eqref{BKop}. Since our aim is to obtain the generator $\cal L$, then we must approximate the scalar diffusion coefficient $D$. Several nonparametric methods for estimating $D$ have been proposed. For example, \cite{cve:06} considered an optimization problem based on a variational formulation which matches information from the discrete approximation of a conditional Markov chain to the generators of the forward and backward Kolmogorov operators. Another method employed in \cite{mg:12} used the one-lag correlations to approximate the diffusion coefficient. The problem with this estimate is that when the true dynamics are not described by a gradient flow, we would like to use a gradient flow to approximate the dynamics on long time scales (see Section \ref{sec52} for an example). If we estimate $D$ using the one-lag correlation as in \cite{mg:12}, we will only match the short-time correlation of the dynamics, but we are interested in the long-time correlation which is better captured by the correlation time.
In this section, we introduce a method to determine $D$ using the correlation time of a one-dimensional observable $S(x(t))$, which results from applying the observation function $S:\mathcal{M}\to\mathbb{R}$ to the multidimensional time series, $x(t)$. By computing the correlation time of this observable for the estimated dynamical system, $D^{-1}\mathcal{L}$, and comparing to the empirical correlation time estimated from the training data set, we will be able to extract the intrinsic parameter $D$. This approach generalizes the Mean Stochastic Model (MSM) of \cite{mgy:10,mh:12}. Intuitively, in the previous section we used the invariant measure to implicitly determine the potential function (through the kernel based generator), and in this section we use the correlation time to fit the stochastic forcing constant $D$. The correlation time is defined by,
\begin{align}
T_c = \int_0^\infty C(\tau)C(0)^{-1}d\tau,\nonumber
\end{align}
where $C(\tau) \equiv \langle S(x(t+\tau))S(x(t)) \rangle$ is the correlation function of the one-dimensional time series, $S(x(t))$. The correlation function can be determined from the data by averaging over $t$, or by the inverse Fourier transform of the power spectrum, which follows from the Wiener-Khinchin formula, $C(\tau) = \mathcal{F}^{-1}\left(\|\mathcal{F}(S(x(t)))\|^2\right)$. For small data sets we found the Wiener-Khinchin approach to be more robust and this is the approach used in the examples in Section \ref{sec5}. We note that the observation function $S$ can be any centered functional on the data set as long as the observed process is stationary and is not identically zero (which guarantees that the correlation time is well defined). Note that in the case of anisotropic diffusion, one would need to compute the entire correlation matrix $\mathbb{E}[x(t-\tau)x(t)^\top]$ as a function of $\tau$. Since we assume the diffusion is isotropic, we will only need a single correlation statistic, and the correlation time of the time series $S(x(t))$ will be sufficient.
Once $T_c$ is estimated from the data, we need to find the value of $D$ which makes \eqref{SDE} have correlation time $T_c$. In order to do this, we will show that $T_c$ can be estimated from the eigenvalues and eigenfunctions of $D^{-1}\mathcal{L}$. Let $e^{\tau\mathcal{L}}$ be the semigroup solution for the backward equation. Following \cite{mag:05} we have,
\[ C(\tau) = \langle S(x(t+\tau)) S(x(t)) \rangle = \int_{\mathcal{M}} \big(e^{\tau\mathcal{L}}(S)(x)\big) S(x) p_{{\rm eq}}(x)dV(x). \]
Writing $S = \sum_i \langle S,\varphi_i\rangle_{p_{{\rm eq}}}\varphi_i$ in the eigenbasis of $\mathcal{L}$, we have,
\begin{align}\label{correlationfunction} C(\tau) = \int_{\cal M} \sum_i e^{D\lambda_i\tau}\langle S,\varphi_i\rangle_{p_{{\rm eq}}}^{\top} \varphi_i(x) S(x)p_{{\rm eq}}(x) dV(x) = \sum_i e^{D\lambda_i\tau} \langle S,\varphi_i \rangle _{p_{{\rm eq}}}^2. \end{align}
Note that since $\lambda_0=0$ and $\varphi_0=1$, $\langle S,\varphi_0\rangle _{p_{{\rm eq}}} = \int_{\mathcal{M}} S(x)p_{{\rm eq}}(x)dV(x) = \mathbb{E}_{p_{{\rm eq}}}[S(x)]$ so we require $S(x)$ to be centered (otherwise the integral of the first term diverges). Noting that $\lambda_i < 0$ for $i>0$, we can compute the correlation time analytically as,
\[ T_c = \int_0^{\infty} C(\tau)C(0)^{-1}d\tau = - \frac{\sum_{i\geq 1} (D\lambda_i)^{-1}\langle S,\varphi_i \rangle _{p_{{\rm eq}}}^2}{\sum_{j\geq 1}\langle S,\varphi_j \rangle _{p_{{\rm eq}}}^2}, \]
and therefore,
\begin{align}\label{Dformula} D = - \frac{1}{T_c}\frac{\sum_{i\geq 1} \lambda_i^{-1}\langle S,\varphi_i \rangle _{p_{{\rm eq}}}^2}{\sum_{j\geq 1}\langle S,\varphi_j \rangle _{p_{{\rm eq}}}^2}. \end{align}
This gives us an equation for the diffusion coefficient $D$ which matches any given correlation time $T_c$. Using the empirical correlation time allows us to find the diffusion coefficient of the system defined by the data. We should note that the correlation function, $C(\tau)$, and hence the correlation time $T_c$, are not intrinsic to the manifold $\mathcal{M}$ due to the dependence on the observation function, $S$. However, the intrinsic diffusion constant $D$ can be recovered from the ratio \eqref{Dformula} since both formulas for $T_c$ are based the same function $S$.
For one-dimensional linear problems, such as the Ornstein-Uhlenbeck process, \eqref{Dformula} reduces to the mean stochastic model (MSM) introduced in \cite{mgy:10,mh:12}. To see this, let the potential function be $U(x) = -\alpha x^2/2\in\mathbb{R}$, so that the gradient flow system in \eqref{SDE} is a one-dimensional Ornstein-Uhlenbeck (OU) process and the eigenfunctions of $\mathcal{L}$ are the Hermite polynomials $\varphi_i(x) = H_i(x)$. All the inner products $\langle S,\varphi_i\rangle _{p_{{\rm eq}}} = \langle x,H_i(x) \rangle _{p_{{\rm eq}}} = \langle H_1(x),H_i(x)\rangle _{p_{{\rm eq}}}$ are zero except for $i=1$, in which case the inner product is $1$ and the associated eigenvalue is $\lambda_1 = \alpha/D$ so that \eqref{Dformula} becomes $\alpha = -1/T_c$ which is the MSM formula. For nonlinear problems, the correlation function \eqref{correlationfunction} for the gradient flow system is more accurate than that of the MSM fit as we will show in Section \ref{sec5}.
To numerically estimate \eqref{Dformula}, we first approximate,
\[ \left<S,\varphi_i\right>_{p_{{\rm eq}}} \approx \frac{1}{N}\sum_{l=1}^N S(x_l)\varphi_i(x_l) = \frac{1}{N}S(x)^\top \vec{\varphi}_i, \]
where $S(x)$ is an $N\times 1$ matrix with $l-$th row given by, $S(x_l)$. We then approximate the diffusion coefficient as,
\[ D \approx -\frac{1}{T_c} \frac{\sum_{i=1}^M \lambda_i^{-1} (S(x)^\top \vec\varphi_i)^2}{\sum_{i=1}^M (S(x)^\top \vec\varphi_i)^2}, \]
which estimates \eqref{Dformula} with summations over $M$ modes. We obtained the best empirical results with the observation function $S(x(t)) \equiv \sum_{j=1}^n (x(t))_j - \mathbb{E}[(x(t))_j]$, which results from summing the (centered) coordinates of the data point $x(t)$. Of course, this observable is vulnerable to certain pathological examples, such as a two dimensional time series with $x_2 = -x_1$. A more robust choice for $S$ would be the norm of the multi-dimensional time series $x(t)$, however we found better results with the summation.
To verify the formula for estimating $D$, we simulate the Ornstein-Uhlenbeck process with true $D=1$ to produce a time series $\{x_i\}_{i=1}^{T}$ with $T=1000000$ and discrete time step $\Delta t=0.01$. We estimate the correlation function as a simple average of $\frac{1}{T-j}\sum_{i=1}^{T-j} S(x_{i+j})S(x_i)$ for lags $j$ starting at $j=1$ and increasing until the first value of $j$ for which the average was negative. Using the trapezoid rule to estimate the integral over all the values of the shift, $j$, we estimate the correlation time to be $T_c \approx 1.0238$. We then subsample every $50$-th data point to produce a time series of length $20000$, the samples of which are approximately independent samples of the invariant measure. We apply the diffusion maps algorithm with the variable bandwidth kernel to the subsampled data to estimate the operator $D^{-1}\mathcal{L}$ as well as the first $M=500$ eigenvalues and eigenfunctions. Approximating the formula \eqref{Dformula} as described above, we found $D \approx 1.0073$. In order to verify that this result for $D$ depends only on the intrinsic geometry, we then map the time series $\{x_i\}$ into $\mathbb{R}^3$ with the isometric embedding,
\[ x_i \mapsto \mathcal{F}(x_i) = (\sin(2x),\cos(2x),x)^\top /\sqrt{5}. \]
One can easily check that $\mathcal{F}$ is an isometry since $D_{x}\mathcal{F}^\top D_x\mathcal{F} \equiv 1$. We then repeat the above procedure for the time series $S \circ \mathcal{F}(x_i)$ and found $T_c \approx 0.7571$ and $D \approx 1.0186$. Notice that the estimates for $D$ are very similar, whereas the estimates for $T_c$ are different. The isometric change in the geometry, $\mathcal{F}$, has decreased the correlation time for the observable $S \circ \mathcal{F}$, however the intrinsic variable $D$ is not effected. This is because $D$ is approximated in \eqref{Dformula} as a ratio between two different methods of estimating the correlation time, and the isometry has the same effect on both estimates of the correlation time.
\section{Forecasting and Filtering in Diffusion Coordinates}\label{UQ}
In this section we formulate two UQ problems, namely forecasting and filtering, in diffusion coordinates. These two problems involve pushing a probability measure forward in time, and therefore they involve the Fokker-Planck operator $\mathcal{L}^*$.
By projecting these infinite dimensional systems onto the eigenfunctions of the Fokker-Planck operator, we find infinite dimensional linear systems which govern the evolution on the projected coordinates. We can then approximate the solution by truncating the linear system to describe finitely many coordinates.
For gradient flows, it is easy to determine the eigenvalues and eigenfunctions of the Fokker-Planck operator, $\mathcal{L}^*$, from the eigenvalues and eigenfunctions of the generator $\cal L$. To see this, note that $p_{{\rm eq}}\propto e^{-U/D}$, and,
\begin{align}\label{conj} \frac{1}{p_{{\rm eq}}}\mathcal{L}^*(fp_{{\rm eq}}) &= \frac{1}{p} \textup{div}\left(D \nabla(fp_{{\rm eq}}) + fp_{{\rm eq}} \nabla U \right) = \frac{1}{p_{{\rm eq}}}\textup{div}\left(Dp_{{\rm eq}}\nabla f +D f\nabla p_{{\rm eq}} + fp_{{\rm eq}}\nabla U \right) \nonumber \\
&= \frac{1}{p_{{\rm eq}}}\textup{div}(Dp_{{\rm eq}}\nabla f) = D\Delta f + D\nabla f \cdot \frac{\nabla p_{{\rm eq}}}{p_{{\rm eq}}} = D\Delta f - \nabla f \cdot \nabla U = \mathcal{L}f,
\end{align}
since $D\nabla p_{{\rm eq}} = - p_{{\rm eq}}\nabla U $. Also, it follows from \eqref{conj} that $\mathcal{L}^*f = p_{{\rm eq}}\mathcal{L}(f/p_{{\rm eq}})$. The two operators have the same eigenvalues and it is easy to show that the semi-group solutions are related by $e^{t\mathcal{L}^*} f= p_{{\rm eq}} e^{t\mathcal{L}}(f/p_{{\rm eq}})$ and the eigenfunctions $\psi_i$ of $\mathcal{L}^*$ and $e^{t\mathcal{L}^*}$ are given by $\psi_i = p_{{\rm eq}} \varphi_i$ where $\varphi_i$ are the eigenfunctions of $\mathcal{L}$ and $e^{t\mathcal{L}}$. From the orthonormality of $\varphi_i$, it is easy to deduce that,
\begin{align}\langle \psi_i,\psi_j \rangle_{1/p_{{\rm eq}}} =\langle \psi_i,\varphi_j \rangle = \langle \varphi_i,\varphi_j \rangle_{p_{{\rm eq}}} =\delta_{ij}.\label{innerproductrule}
\end{align}
Of course, since these eigenfunctions are all represented by $N$-dimensional eigenvectors (which represent evaluation on the data set $\{x_j\}_{j=1}^N$) the actual computation of any inner product will always be realized by rewriting it with respect to $p_{{\rm eq}}$, following the rule in \eqref{innerproductrule}.
\subsection{Nonlinear forecasting}
The \emph{forecasting} problem is, given an arbitrary initial distribution $p_0(x)$ for a state variable $x$, find the density $p(x,t)$ at any future time $t$. This will be the most straightforward problem to solve in diffusion coordinates since $p(x,t)$ solves the Fokker-Planck equation,
\begin{align}\label{forecasteq} \frac{\partial p}{\partial t} = \mathcal{L}^*p, \quad\quad p(x,0) =p_0(x).
\end{align}
Let $\psi_i$ be the eigenfunctions of $\cal L^*$ with eigenvalues $D\lambda_i$, and write $p$ in this basis as, $p(x,t) = \sum_i c_i(t) \psi_i(x)$ where $c_i = \langle p,\psi_i\rangle _{1/p_{{\rm eq}}}$. We define the vector $\vec c = \vec c(t)$ of eigencoordinates and the diagonal matrix $\Lambda_{ii} = \lambda_i$, and writing \eqref{forecasteq} in these coordinates we have,
\begin{align}\label{forecastprojection} \frac{d\vec c}{dt} = D\Lambda \vec c, \quad\quad\vec c(0) = \langle p_0,\psi_i \rangle_{1/p_{{\rm eq}}}.
\end{align}
In order to solve the linear system in \eqref{forecastprojection}, we first need to find the initial conditions $c_i(0)$ by projecting the initial density $p(x,0)=p_0(x)$ onto $\psi_i$. Since densities $p_0(x)$ and $p_{{\rm eq}}(x)$ are given on the training data set $x_i$, we define the discrete densities $(\vec p_0)_l = p_0(x_l)$ and $(\vec p_{{\rm eq}})_l = p_{{\rm eq}}(x_l)$.
Following the inner product rule in \eqref{innerproductrule}, we write the projection of the initial condition as,
\begin{align}\label{coeffest} c_i(0) = \left<p_0,\psi_i\right>_{1/p_{{\rm eq}}} = \left<p_0/p_{{\rm eq}},\varphi_i \right>_{p_{{\rm eq}}} \approx \frac{1}{N} \sum_{j=1}^N \frac{p_0(x_j)}{p_{{\rm eq}}(x_j)}\varphi_i(x_j), \end{align}
since this converges to the true value as $N\to\infty$. With this initial condition, each component in \eqref{forecastprojection} can be solved analytically as $c_i(t) = e^{tD\lambda_i}c_i(0)$. Numerically, we will approximate the solutions with finitely many modes, $i=0,1,\ldots,M$. In Section \ref{QoI} we show how to use these solutions to reconstruct the density $p(x,t)$ or to estimate a quantity of interest at any time $t$.
\subsection{Nonlinear filtering}\label{nonlinearfilter}
The \emph{filtering} problem is, given a time series of noisy observations $z$ of the state variable $x$, find the posterior density $P(x,t | z(s), s\leq t)$ given all the observations $z(s)$ up to time $t$. For an observation function $h:\mathcal{M}\rightarrow\mathbb{R}^m$, we consider continuous-time observations of the form,
\begin{align}\label{obseq} dz = h(x) \, dt + \sqrt{R} \, dW_t, \end{align}
where $dW_t$ are i.i.d Gaussian noise in $\mathbb{R}^m$.
With these observations, the evolution of the unnormalized posterior density, $p(x,t)$, defined as follows,
\[
P(x,t | z(s), s\leq t)= \frac{p(x,t)}{\int_{\cal M} p(x,t)dV(x)},
\]
is given by the Zakai equation \cite{bc:09},
\begin{align}\label{filtereq} dp = \mathcal{L}^*p dt + ph^\top R^{-1} dz. \end{align}
Writing $p(x,t) = \sum_i c_i(t)\psi_i(x)$ the evolution becomes,
\[ \sum_i dc_i \psi_i = \sum_i D\lambda_i c_i \psi_i dt + \sum_i c_i \psi_i h^\top R^{-1} dz, \]
and projecting both sides of this equation onto $\psi_j$ we find,
\begin{align}\label{infinitesde} dc_j = D\lambda_j c_j dt + \sum_i c_i \langle \psi_i h^\top,\psi_j\rangle _{1/p_{{\rm eq}}} R^{-1} dz, \end{align}
which is an infinite-dimensional system of stochastic differential equations with multiplicative stochastic forcing.
To numerically estimate the solution of \eqref{infinitesde}, we truncate this infinite-dimensional system of SDEs by solving only a system of $(M+1)-$dimensional SDEs for $\vec{c}=(c_0,c_1,\ldots,c_M)^{\top}$. This strategy is simply a Galerkin approximation for the Zakai equation for which we choose the diffusion coordinates as a basis rather than the Gaussian series as proposed in \cite{ar:97} or the Hermite polynomial basis as proposed in \cite{fsx:13}. We should note that for scalar OU processes, the diffusion coordinates are exactly the Hermite polynomials since these polynomials are the eigenfunctions of the backward Kolmogorov operator for the OU process.
Defining $H_{ji} = \left<\psi_i h^\top,\varphi_j \right>$ as a $1\times m$ dimensional vector for each pair $(i,j)$; where the $k$-th component is given by $H_{ji}^k = \left<\psi_i h_k,\varphi_j \right>$, we can write the truncated $(M+1)-$dimensional system of SDEs in compact form as follows,
\begin{align}\label{filterprojection} d\vec c = D\Lambda \vec c dt + (H\vec c) R^{-1} dz. \end{align}
To solve this system of SDEs, we first project the initial condition $p_0(x)$ to obtain $c_i(0)$ as in \eqref{coeffest} for $i=0,1,\ldots,M$.
We then numerically solve the system of SDEs in \eqref{filterprojection} using the splitting-up method (see for example \cite{fsx:13}). Explicitly, given $\vec c(t_{i-1})$ we compute $\vec c(t_i)$ by first using the solution to the deterministic part,
\[ \vec c_0(t_i) = \exp\left(D\Lambda \Delta t \right)\vec c(t_{i-1}), \]
and then the solution to the stochastic part,
\[ \vec c(t_i) = \exp\left(H R^{-1}dz(t_i) - \frac{1}{2}H^2 R^{-1}\vec 1 \Delta t \right)\vec c_0(t_i). \]
Here, the exponent term consists of an $M\times M$ matrix given by
\[ HR^{-1}dz(t_i) - \frac{1}{2}H^2 R^{-1}\vec 1 \Delta t = \sum_k H^k (R^{-1}dz(t_i))_k - \frac{1}{2}(H^k)^2 (R^{-1}\vec 1)_k \Delta t. \]
Notice that unlike the forecasting problem which had an analytic solution for any time $t$, this procedure must be iterated for each observation $dz(t_i)$.
\subsection{Quantity of Interest}\label{QoI}
In the uncertainty quantification problems described above, we reduce the PDE and SPDE problems which describe the evolution of density $p(x,t)$ into systems of ODEs in \eqref{forecastprojection} and SDEs in \eqref{filterprojection}, respectively, for the coefficients $c_i(t)$. Using the coefficients $c_i(t)$ we can approximate the density function as follows,
\begin{align}\label{reconstruct} p(x,t) \approx p_M(x,t) =\sum_{i=0}^{M} c_i(t)\psi_i(x) =\sum_{i=0}^{M} c_i(t)\varphi_i(x)p_{{\rm eq}}(x) , \end{align}
where the coefficients $\{c_i(t)\}_{i=0}^M$ are from the solution of the $(M+1)-$dimensional systems of ODEs or SDEs.
One complication in some of the algorithms above is that the evolution of the coefficients may not preserve the normalization of the corresponding density. In fact, for the nonlinear filtering problems in \eqref{filtereq}, the solutions are unnormalized densities. Fortunately we can use the eigenfunctions $\varphi_i$ to estimate the normalization factor of an unnormalized density from its diffusion coordinates. Assume that $p_M$ in \eqref{reconstruct} is an unnormalized density with diffusion coordinates $c_i$ obtained from the solution of the $(M+1)-$dimensional systems of ODEs or SDEs. We can then estimate the normalization constant as,
\[ Z = \int_{\mathcal{M}} p_M(x,t)dV(x) = \sum_{i=0}^M c_i \int_{\mathcal{M}} \varphi_i(x)p_{{\rm eq}}(x)dV(x) = \sum_{i=0}^M c_i \langle 1,\varphi_i \rangle_{p_{{\rm eq}}}. \]
where we estimate inner product, $\langle 1, \varphi_i \rangle_{p_{{\rm eq}}} \approx \frac{1}{N}\sum_{j=1}^N \varphi_i(x_j)$. Subsequently, we normalize our density by replacing $p_M$ with $p_M/Z$. In the remainder of this paper, we will refer to $p_M$ as the normalized density without ambiguity. We also note that by reconstructing the density from the coefficients we are implicitly projecting the density into the subspace of Hilbert space spanned by the first $M+1$ eigenfunctions. This has a smoothing effect which improves the clarity of figures and we always show reconstructed densities in the figures and videos.
In many cases, we are not interested in the entire density $p$; instead we are often interested in the expectation of an observable $A:\mathcal{M}\to\mathbb{R}$ with respect to $p$, namely $\mathbb{E}_{p(x,t)}[A(x)]$. A common example of functionals of interest are the moments $A(x) = x^m$ for $m\in\mathbb{N}$. Although $A$ is a functional on $\mathcal{M}$, $A$ can be given as a functional on the ambient space $\mathbb{R}^n$ because we only evaluate $A$ on the data set, which lies on the manifold $\cal M$. Using the formula \eqref{reconstruct} for the density estimate at time $t$, we can approximate the expectation of the functional $A$ at time $t$ as,
\begin{align}\label{functionalexpectation} \mathbb{E}_{p(x,t)}[A(x)] \approx \langle A,p_M \rangle = \sum_{i=0}^{M} c_i(t) \langle A,\varphi_i\rangle _{p_{{\rm eq}}} = \sum_{i=0}^M c_i(t) a_i = \vec c(t)^\top \vec a.
\end{align}
The formula \eqref{functionalexpectation} shows that we can find the expectation of a functional $A$ by writing $A$ in diffusion coordinates as $a_i = \langle A,\varphi_i\rangle _{p_{{\rm eq}}}$. In these coordinates, the expectation of $A$ is simply the inner product of the diffusion coefficients of $A$ and $p(x,t)$. Finally, if the quantity of interest is given by a functional $A$ evaluated on the training data set as $(\vec A)_l = A(x_l)$, we approximate the inner product $a_i = \left<A,\varphi_i\right>_{p_{{\rm eq}}} \approx \frac{1}{N} \vec A^\top \vec \varphi_i$.
\section{Nonlinear Response in Diffusion Coordinates}\label{sec4}
The response problem considered in \cite{mag:05} is to determine the change in the expectation of a functional $A(x)$ as function of time, after perturbing the dynamics of a system at equilibrium. Letting $p^\delta(x,t)$ be the density evolved according to the perturbed system, we define the response of a functional $A$ as the expected difference,
\begin{align}
\delta\mathbb{E}[A(x)](t) = \mathbb{E}_{p^\delta}[A(x)](t) - \mathbb{E}_{p_{{\rm eq}}}[A(x)],\quad t\geq 0, \label{response}
\end{align}
We assume the unperturbed system has the form,
\begin{align}
dx = F(x)\,dt + \sqrt{2D}dW.\label{unperturbed}
\end{align}
with equilibrium density $p_{{\rm eq}}$ and Fokker-Planck operator $\cal L^*$, so that $\mathcal{L}^*p_{{\rm eq}} = 0$. We assume the perturbed system has the form,
\begin{align}
dx = (F(x)+\delta F(x,t))\,dt + \sqrt{2D}dW,\label{perturbed}
\end{align}
with associated Fokker-Planck operator $\tilde{\mathcal{L}}^*$ so that $p^\delta$ solves the Fokker-Planck equation,
\begin{align}\label{responseeq} \frac{\partial p^\delta}{\partial t} = \tilde{\mathcal{L}}^*p^\delta = \mathcal{L}^*p^\delta + \delta\mathcal{L}^*p^\delta, \end{align}
where $\delta\mathcal{L}^*p=-\textup{div}(\delta Fp)$ denotes the Liouville operator corresponding to the external forcing with functional form $\delta F$.
This reveals that the response problem is closely related to the forecasting problem. Given the Fokker-Planck operator $\tilde{\mathcal{L}}^*$ for the perturbed system, one must solve \eqref{responseeq} with initial condition chosen to be the equilibrium distribution of the unperturbed system, $p^\delta(x,0)=p_{{\rm eq}}(x)$. Since computational cost becomes an issue, especially for higher dimensional problems, one typically approximates $p^\delta$ by solving ensemble solutions of \eqref{perturbed} with initial ensemble sampled from $p_{{\rm eq}}$, and in fact we will use this approach as a benchmark in Section \ref{dwresponsesection}. However, as in \cite{mag:05}, we are interested in cases where the Fokker-Planck operators $\mathcal{L}^*$ and $\tilde{\mathcal{L}}^*$ and also the functional form of $F$ are all unknown. We will assume that we only have access to data from the unperturbed system and the functional form of the external perturbation $\delta F$. Of course, since the techniques of Section \ref{DM} are restricted to gradient flow systems, we will assume that both the unperturbed and perturbed systems are well approximated by stochastically forced gradient flows.
For gradient flow systems, we have $F(x)=-\nabla U(x)$ and $\delta F(x)=-\nabla \delta U(x)$, where the perturbation is a time-independent potential. Given a data set from $\{x_i\}_{i=1}^N$ sampled from $p_{{\rm eq}}$ and the functional form of $\delta U$, it is tempting to proceed as in Section~\ref{UQ} by writing $p^\delta(x,t) = \sum_i c_i(t) \psi_i(x)$ and projecting \eqref{responseeq} to eigenfunctions of $\mathcal{L}^*$. The main issue with this projection is that it is difficult to computationally attain the external perturbed forcing term,
\[\langle\delta \mathcal{L}^*\psi_i,\psi_j\rangle_{1/p_{{\rm eq}}} = -\langle \textup{div}(\delta F\psi_i),\psi_j\rangle_{1/p_{{\rm eq}}}=\langle \delta F, \psi_i\nabla\varphi_j)\rangle = \langle \delta U, \textup{div}(\varphi_i\nabla\varphi_j)\rangle_{p_{{\rm eq}}}.\]
Even if $\delta F$ or $\delta U$ are known, we have no provably convergent technique to compute the gradient or divergence operators on the unknown Riemannian manifold. In fact, the technique discussed in Section \ref{DM} does not provide a framework to represent vector fields on manifolds.
As a remedy, we propose to project \eqref{responseeq} onto the eigenfunctions of $\tilde{\mathcal{L}}^*$ which can be accessed from the functional form of $\delta U$ and the data set $\{x_i\}$. This is an application of a result in \cite{VB}, which showed that one can approximate the generator of any gradient flow system with a (partially) known potential even if the data is sampled from a different gradient flow system. In particular, it was shown (see eq.(17) in \cite{VB}) that for general choice of bandwidth function $\rho$ in \eqref{VBkernel} and any $f\in L^2(\mathcal{M},p_{{\rm eq}})\cap\mathcal{C}^3(\mathcal{M})$, the matrix $L_\epsilon$ constructed in \eqref{approxgenerator} converges to a weighted Laplacian in the following sense,
\begin{align}
L_\epsilon f = \Delta f+2\left(1+\frac{d}{4}\right)\nabla f\cdot \frac{\nabla p_{{\rm eq}}}{p_{{\rm eq}}} + (d+2) \nabla f\cdot \frac{\nabla\rho}{\rho}+\mathcal{O}(\epsilon).\label{Lgeneral}
\end{align}
If we choose the following bandwidth function,
\begin{align}
\rho = p_{{\rm eq}}^{-1/2} e^{-\frac{\delta U}{D(d+2)}},\label{modbandwidth}
\end{align}
where $p_{{\rm eq}}$ is estimated from the data set as before, substituting this into \eqref{Lgeneral}, we obtain,
\begin{align}
L_\epsilon f &= \Delta f+\left(2+\frac{d}{2}\right)\nabla f\cdot \frac{\nabla p_{{\rm eq}}}{p_{{\rm eq}}} - \nabla f\cdot \frac{\nabla\delta U}{D} -\left(\frac{d}{2}+1\right)\nabla f\cdot \frac{\nablap_{{\rm eq}}}{p_{{\rm eq}}} + \mathcal{O}(\epsilon) \nonumber\\ &= \Delta f +\nabla f\cdot \frac{\nabla p_{{\rm eq}}}{p_{{\rm eq}}} - \nabla f\cdot \frac{\nabla\delta U}{D} + \mathcal{O}(\epsilon)\nonumber\\ &= \Delta f - \nabla f\cdot \frac{(\nabla U +\nabla\delta U)}{D} + \mathcal{O}(\epsilon)\nonumber\\ &= D^{-1}\tilde{\mathcal{L}}f+ \mathcal{O}(\epsilon).\label{newL}
\end{align}
With this result, we construct a stochastic matrix $L_\epsilon$ following exactly the procedure described in Section~\ref{findingBK}, except with the bandwidth function in \eqref{modbandwidth}.
Let $\tilde{\lambda}_i$ and $\tilde\varphi_i$ be the eigenvalues and eigenfunctions of $L_\epsilon$, which is a discrete approximation to the continuous operator $D^{-1}\tilde{\mathcal{L}}$ as shown in \eqref{newL}.
Following the same argument as in Section~\ref{UQ}, the corresponding Fokker-Planck operator, $\tilde{\mathcal{L}}^*$, has eigenvalues $D\tilde{\lambda}_i$ with eigenfunctions, $\tilde\psi_i = \tilde\varphi_i \tilde{p}_\textup{eq}$, where $\tilde{p}_\textup{eq} \proptop_{{\rm eq}} e^{-\delta U/D} $ is the equilibrium measure of the perturbed system. Now letting $p^\delta(x,t) = \sum_i \tilde{c}_i(t) \tilde{\psi}_i(x)$ and projecting \eqref{responseeq} with initial condition, $p^\delta(x,0)=p_{{\rm eq}}(x)$, on the perturbed coordinate basis, the evolution of the perturbed density becomes a linear forecasting problem,
\begin{align}\label{responseprojection}
\frac{d\tilde{c}_i}{dt} = D\tilde\lambda_i \tilde c_i, \quad\quad
\tilde c_i(0) = \langle p_{{\rm eq}}, \tilde\psi_i \rangle_{1/\tilde{p}_\text{eq}}.
\end{align}
Therefore, the response in \eqref{response} formula can be approximated by,
\begin{align}\label{responsefunctionalexpectation}
\delta\mathbb{E}[A](t) \approx \sum_{i=1}^M (\tilde{c}_i(t)-\tilde{c}_i(0)) \langle A, \tilde\varphi_i \rangle_{\tilde{p}_\text{eq}},
\quad t\geq0\end{align}
utilizing \eqref{functionalexpectation} on the perturbed diffusion coordinates. Notice that the zeroth mode is excluded in the summation in \eqref{responsefunctionalexpectation} since $\tilde{\lambda}_0=0$ and therefore, $\tilde{c}_0(t)=\tilde{c}_0(0)$, for any $t$.
A complication in numerically implementing the above technique is that the eigenfunctions $\tilde \varphi_i$ are orthonormal with respect to the inner product $\langle\cdot,\cdot\rangle_{\tilde p_{\rm eq}}$ weighted by the perturbed equilibrium density, $\tilde p_{\rm eq}$, but we can only estimate inner products weighted with respect to the unperturbed equilibrium density, $p_{{\rm eq}}$. This requires us to rewrite all of the inner products with respect to $p_{{\rm eq}}$. In particular, we can rewrite the initial coefficients in \eqref{responseprojection} as,
\[ \tilde c_i(0) = \langle p_{{\rm eq}},\tilde \psi_i \rangle_{1/\tilde p_{\rm eq}} = \left<p_{{\rm eq}},\tilde \varphi_i \right> = \langle 1,\tilde\varphi_i \rangle_{p_{{\rm eq}}} \approx \frac{1}{N} \sum_{j=1}^N \tilde\varphi_i(x_j).\]
With these initial conditions, we have explicit solutions for the linear problem in \eqref{responseprojection} given by $\tilde c_i(t)= e^{D\tilde\lambda_it}\tilde c_i(0)$. To compute the response formula in \eqref{responsefunctionalexpectation}, we evaluate the inner product terms in \eqref{responsefunctionalexpectation} as follows,
\[\langle A, \tilde\varphi_i\rangle_{\tilde{p}_\text{eq}} = \langle A, \frac{\tilde{p}_\text{eq}}{p_{{\rm eq}}} \tilde\varphi_i \rangle_{p_{{\rm eq}}} \approx \frac{1}{N}\sum_{j=1}^N A(x_j) \frac{\tilde{p}_\text{eq}(x_j)}{p_{{\rm eq}}(x_j)} \tilde\varphi_i(x_j) = \frac{1}{NZ}\sum_{j=1}^N A(x_j) e^{-\delta U(x_j)/D}\tilde\varphi_i(x_j),\]
using the fact that the equilibrium density of the perturbed system is given by, $\tilde p_{\rm eq} = \frac{1}{Z} p_{{\rm eq}} e^{-\delta U/D}$, where $Z$ is the normalization factor for the density $\tilde p_{\rm eq}$ obtained as discussed in Section~\ref{QoI}.
\section{Numerical Results}\label{sec5}
For a state variable $x$ evolving according to \eqref{SDE}, we have shown how to estimate the density function $p(x,t)$, which solves either the forecasting, filtering, or response problem. We have also shown how to find the expectation of any quantity of interest, $A(x)$, with respect to this measure. In this section, we first validate the numerical algorithms developed in Sections~\ref{UQ} and \ref{sec4} on a simple linear gradient flow system. We then test our algorithms on a chaotically forced gradient flow system which is known to be well approximated as a stochastically forced gradient flow. In each example we will use only $M=50$ eigenvectors in the non-parametric model.
\subsection{Linear Example: Ornstein-Uhlenbeck processes}\label{sec51}
We first validate our algorithms on the Ornstein-Uhlenbeck processes on $\mathcal{M}=\mathbb{R}$, given by,
\begin{align}\label{OU} dx = -x\, dt + \sqrt{2}\, dW_t, \end{align}
where $dW_t$ denotes standard Gaussian white noise. Our technique assumes that we do not know the model in \eqref{OU}, but instead we are only given sample solutions of \eqref{OU} as a training data set. In our numerical experiment, we generate a training data set consisting of $N=10000$ data points $x_l = x(l\Delta t)$ by numerically integrating \eqref{OU} with the Euler-Maruyama method and sub-sample at every $\Delta t=0.2$ time units. We intentionally sub-sample with a large $\Delta t$ so that the data set is approximately independently sampled from the invariant measure. We numerically validate our methods on this simple system \eqref{OU} since it is possible to find an analytic expression for the evolution of the moments (see Appendix \ref{OUanalysis}) for the forecasting and response problems. For the filtering problem, we compare the results to the Kalman-Bucy solutions which are optimal in a least square sense \cite{kalman:61}.
\subsubsection{Forecasting}\label{forecastalgorithm}
\begin{figure}[h]
\centering
\includegraphics[width=0.45\textwidth]{fig1a.pdf}\includegraphics[width=0.45\textwidth]{fig1b.pdf}
\caption{\label{forecastingfig} Forecasting the evolution of a Gamma distributed initial condition in the Ornstein-Uhlenbeck system. Left: Evolution of the density. Right: Evolution of the moments.}
\end{figure}
To validate our forecasting algorithm, we take the initial condition to be a Gamma distribution with density $p_0(x) = 4 (x+1/2) e^{-2(x+1/2)}$ on $[-1/2,\infty)$. For this distribution the first four moments are, $\bar x = \mathbb{E}_{p_0}[x] = 1/2$, $\mathbb{E}_{p_0}[(x-\bar x)^2] = 1/2$, $\mathbb{E}_{p_0}[(x-\bar x)^3] = 1/2$, and $\mathbb{E}_{p_0}[(x-\bar x)^4] = 3/2$. We assume that the initial distribution is known, and we simply evaluate $p_0$ on the training data set to find $(\vec p_0)_l = p_0(x_l)$. The results of the forecasting algorithm are shown in Figure \ref{forecastingfig} where we reconstruct the forecast density at various times as it approaches the invariant measure. Notice the oscillation in the reconstructed $p_0$ near the non-smooth part is a Gibbs phenomenon-like behavior. We also show the evolution of the first four centered moments which are evaluated as linear combinations of the uncentered moments $A(x) = x^r$, $r=1,\ldots, 4$. Notice that the initial kurtosis of the Gamma distribution is difficult to estimate since the data on the tail is very sparse (there are only 3 sample points for $x>3$ in the training data set). The Monte-Carlo integral converges slowly since the Gamma distribution decays slower than the Gaussian distribution.
\subsubsection{Filtering}\label{filteralgorithm}
To verify our filtering algorithm, we consider the observation \eqref{obseq} with $h(x) = x$ so that,
\[ dz = x\, dt + \sqrt{R}\,dW_t. \]
The filtering problem for \eqref{OU} with this observation has optimal solutions given by the Kalman-Bucy equations \cite{kalman:61}, which we approximate for a finite observation time step $\Delta t$ by the discrete time Kalman filter. Since discrete time observations are typically given in the form,
\[ \mathcal{Z}(t_i) = x(t_i) + \sqrt{R_o}\,\omega(t_i) \]
where $\omega(t_i)$ are independent random samples from $\mathcal{N}(0,1)$, we will approximate $dz$ for a finite time step $\Delta t = 0.01$ as,
\[ dz(t_i) \approx x(t_i)\Delta t + \sqrt{\Delta t R_o}\, dW_{t_i} = \mathcal{Z}(t_i) \Delta t . \]
So in the numerical algorithm of Section \ref{nonlinearfilter} we use $dz(t_i) = \mathcal{Z}(t_i)\Delta t$ and we set $R = \Delta t R_o$.
\begin{figure}[h]
\centering
\includegraphics[width=0.45\textwidth]{fig2a.pdf} \includegraphics[width=0.45\textwidth]{fig2b.pdf}
\caption{\label{filterfig} Nonlinear filter mean estimates as functions of time, compared to the truth, Kalman filter solutions, and observations (left). The posterior covariance estimates as functions of time (right). The nonlinear filter recovers the Kalman filter solutions in this linear example. The posterior evolution is shown in {\tt fig2.mov}.}
\end{figure}
In our numerical simulation, we take $R_o=1$ and assume an initial distribution $p_0(x)=p_{{\rm eq}}(x)$, which is the Gaussian invariant measure of \eqref{OU}. The results are shown in Figure \ref{filterfig}, where we show that we recover the same statistical solutions (mean and covariance) as the discrete Kalman filter. We also include a video, {\tt fig2.mov}, comparing the evolution of the reconstructed posterior density with the Kalman filter posterior.
\subsubsection{Response}\label{responsealgorithm}
To validate the response algorithm of Section \ref{sec4}, we perturb the Ornstein-Uhlenbeck system in \eqref{OU}, which has potential $U(x) = x^2/2$, with a potential $\delta U(x) = \frac{a}{2}(x-b/a)^2$. The vector field for the potential $U+\delta U$ becomes $-\nabla(U+\delta U) = -x-a x + b$ so that the perturbed system is again an Ornstein-Uhlenbeck system with different mean and damping parameters. This fact allows us to easily compute the analytic response (see Appendix \ref{OUanalysis}) which we compare to the numerical estimate using the above algorithm in Figure \ref{responsefig}. In this example we choose $a=-0.1$ and $b=0.03$, since the damping is decreased by this perturbation the variance increases in response while the effect of $b$ is a shift in the mean. Notice the accuracy of the response estimates for the mean and variance at all times.
\begin{figure}[h]
\centering
\includegraphics[width=0.5\textwidth]{fig3.pdf}
\caption{\label{responsefig} Nonlinear response of the first two moments of the Ornstein-Uhlenbeck system to a perturbation of the potential function. The analytic response is derived in Appendix \ref{OUanalysis} and the DM response is our nonparametric estimate.}
\end{figure}
\subsection{Nonlinear Example: Chaotically driven double well potential}\label{sec52}
In this section we test our nonparametric UQ methods on the following nonlinear system,
\begin{align}
\label{doublewellfull}
\begin{split}
\dot x &= x-x^3 + \frac{\gamma}{\epsilon} y_2,\\
\dot y_1 &=\frac{10}{\epsilon^{2}}(y_2 - y_1), \\
\dot y_2 &=\frac{1}{\epsilon^{2}}(28 y_1 - y_2 - y_1 y_3),\\
\dot y_3 &=\frac{1}{\epsilon^{2}}(y_1 y_2 - \frac{8}{3}y_3).
\end{split}
\end{align}
which was discussed in \cite{gks:04} as an example of a stochastic homogenization problem and used in \cite{mg:12} as a test model for a reduced filtering problem. The fast variables $(y_1,y_2,y_3)$ solve a chaotic Lorenz-63 system, and $x$ is a scalar ODE driven by this chaotic oscillator with characteristic time $\epsilon^2$. In \cite{gks:04,mg:12} it was shown that for $\gamma, \epsilon$ sufficiently small the dynamics of the variable $x$ are well approximated by the reduced stochastic model.
\begin{align}\label{doublewellreduced}
dX = X(1-X^2) \, dt + \sigma \, dW_t,
\end{align}
where $\sigma$ is a diffusion constant that can be numerically estimated by a linear regression based algorithm or by matching the correlation time of $\gamma y_2$. This reduced model is a stochastically forced gradient flow system, where the potential function $U(X) = X^2/2 - X^4/4$ is a standard double well potential. While this suggests that our technique should be successful, we emphasize that our method makes no use of either \eqref{doublewellfull} or \eqref{doublewellreduced}, instead it relies on a training data set to represent the reduced model for the variable $x$, with no assumed parametric form. Following \cite{mg:12}, we set $\gamma = 4/90$ and $\epsilon=\sqrt{0.1}$. We note that \cite{mg:12} suggested $\sigma^2 =0.113$ for the homogenization and $\sigma^2 = 0.126$ for the filtering problem.
\begin{figure}[h]
\centering
\includegraphics[width=0.45\textwidth]{fig4a.pdf}\includegraphics[width=0.45\textwidth]{fig4b.pdf}
\caption{\label{statistics} Left: Invariant measure estimated by the kernel. Right: Tuning the stochastic forcing using the correlation time.}
\end{figure}
In order to apply our method to this problem, we generated a training data set containing 200000 data points $\{x_j = x(t_j)\}$ with time spacing $\Delta t = t_{j+1}-t_j = .1$ by solving \eqref{doublewellfull} using a fourth order Runge-Kutta scheme with integration time step $\delta t = 0.002$. We applied the variable bandwidth diffusion kernel as described in Section \ref{findingBK} to the data set $\{x_{10j}\}_{j=1}^{20000}$. We then estimate the diffusion constant $D=\sigma^2/2 \approx 0.0535$ as in Section \ref{findingD}, using the entire data set $\{x_j\}$ to find the correlation time and the empirical correlation function. In Figure \ref{statistics} we compare the invariant measure and correlation function of the nonparametric diffusion model to the empirical estimates of these statistics taken from training data set. We also show the statistics of the mean stochastic model (MSM) of \cite{mgy:10,mh:12}, which is a linear (and Gaussian) model constructed by matching the correlation time and the equilibrium density variance. From Figure~\ref{statistics}, one can see that the density and the autocorrelation function from the diffusion map fitting are very accurate.
\subsubsection{Forecasting}\label{doublewellforecast}
We first test our forecasting algorithm for an initial condition given by a normal distribution centered at $0.5$ with variance $0.01$. Unlike the Ornstein-Uhlenbeck system in Section \ref{sec51}, we do not have an analytical expressions for the evolution the moments. For diagnostic purposes, we will compare our forecast to a Monte-Carlo simulation. We initialize an ensemble of $100000$ points $(x,y_1,y_2,y_3)^\top \in \mathbb{R}^4$ with $x$ chosen randomly according to the initial density and triples $(y_1,y_2,y_3)^\top$ randomly selected from the training data set (so that they lie on the attractor of the chaotic system). This ensemble is then evolved according to the system \eqref{doublewellfull} with the same Runge-Kutta integrator used to generate the training data. For any time $t$ we generate the Monte-Carlo density using a histogram of the ensemble at time $t$, and we call the moments of the ensemble the Monte-Carlo moments. These Monte-Carlo estimates make use of the full true model and are computationally more intensive than our model free method, and we will consider them the `truth' for the purposes of comparison.
Next, we estimate the evolution of the density and the first four centered moments using the technique of Section \ref{UQ}. For this initial condition, the density moves to the right and centers on the right well of the potential on the first 10 model time units. Over the next 1000 model time units, small amounts of density slowly migrate across the barrier between the two potential wells until the system is at equilibrium. In Figure \ref{doublewellprediction} we compare the moment estimates of the nonparametric method to the Monte-Carlo moments. We also show snapshots of the evolution of the density as reconstructed from the diffusion coordinates at times $t = 0, 10, \infty$. The evolution of the reconstructed density is compared to the Monte-Carlo simulation in {\tt fig5a.mov} up to time $t=10$ and {\tt fig5b.mov} up to time $t=2000$.
\begin{figure}[h]
\centering
\includegraphics[width=0.425\textwidth]{fig5a.pdf}\includegraphics[width=0.45\textwidth]{fig5b.pdf}
\caption{\label{doublewellprediction} Left: Evolution of the density is summarized in snapshots for $t=0,10,\infty$ and the potential function is shown. Right: Non-parametric estimates for the evolution of the moments is compared to the Monte-Carlo simulation. The reconstructed density evolution is shown in {\tt fig5a.mov} and {\tt fig5b.mov}. }
\end{figure}
\subsubsection{Filtering}
We now test our nonlinear filtering algorithm on the chaotically driven double well potential using the observation \eqref{obseq} with $h(x) = x$ for several values of observation noise variance, $R$, ranging from $10^{-3}$ to $1$. We choose a time discretization with $\Delta t = 1$ between observations, which is relatively long compared to $\Delta t=0.1$ considered in \cite{mg:12}. For each noise level $R$, we generate 10000 discrete observations using the full system \eqref{doublewellfull}, integrated as in Section \ref{doublewellforecast} and observed with the discretization of \eqref{obseq}. We then apply the filtering algorithm from Section \ref{nonlinearfilter}, which we refer to as the nonlinear filter in Figure \ref{doublewellfilter} where we show the mean estimate as the quantity of interest. For comparison we also use the standard ensemble Kalman filter (EnKF) for the reduced model \eqref{doublewellreduced} using $\sigma^2 = .126$ as suggested in \cite{mg:12,gks:04} with ensemble size 50 without additional variance inflation.
In Figure \ref{doublewellfilter} we see that the nonlinear filter tracks the transition between the potential wells more accurately compared to EnKF for large observation noise, $R=0.5$ (see left panel). In the large noise regime, the posterior estimate is very far from Gaussian. Conversely, for very small observation noise and also for very short discretization time, the performance of nonlinear filter and the EnKF are similar, their mean-squared errors are close to the observation noise variance (see the right panel of Figure~\ref{doublewellfilter}). In the small noise regime, the posterior density is close to Gaussian since it is very close to the observation likelihood. The evolution of the reconstructed non-Gaussian posterior density is compared to the Gaussian posterior density of the EnKF in {\tt fig6.mov}.
\begin{figure}[h]
\centering
\includegraphics[width=0.45\textwidth]{fig6a.pdf}\includegraphics[width=0.45\textwidth]{fig6b.pdf}
\caption{\label{doublewellfilter} Left: Evolution of the posterior mean during a transition from the left potential well to the right potential well. Right: Mean squared error between filter estimates and the true state as a function of the observation noise covariance. The evolution of the posterior densities is shown in {\tt fig6.mov}.}
\end{figure}
Next, we consider a nonlinear observation with $h(x) = |x|$, which is a pathological observation since it is impossible to determine from this observation which potential well the truth lies in. As shown in Figure \ref{doublewellnonlinearobs}, the nonlinear filter respects this symmetry with the posterior mean always very close to zero. In contrast, the EnKF outperforms the nonlinear filter when it happens to be on the right side of the well, however it frequently does not track the transitions leading to a higher overall mean squared error. The evolution of the posterior densities are shown in {\tt fig7a.mov} which shows the posterior density of the nonlinear filter to be a symmetric bi-modal distribution. We also consider another nonlinear observation with $h(x) = (x-0.05)^2$, which is nearly pathological, however, since the symmetry is not perfect, repeated observations allow the filter to determine which well the true state is in. As shown in Figure \ref{doublewellnonlinearobs} and in {\tt fig7b.mov}, the posterior of the nonlinear filter becomes increasingly bi-modal when observations occur near $0.05$. However, with repeated observations which are far from $0.05$ the posterior density accumulates information on the true location and the incorrect mode is progressively damped away.
\begin{figure}[h]
\centering
\includegraphics[width=0.45\textwidth]{fig7a.pdf}\includegraphics[width=0.45\textwidth]{fig7b.pdf}
\caption{\label{doublewellnonlinearobs} Left: Evolution of the posterior mean for $h(x) = |x|$. Right: Evolution of the posterior mean for $h(x) = (x-0.05)^2$. Both filtering experiments assimilate observations at every $\Delta t=1$ model time unit and noise variance $R=0.05$. The evolution of the posterior densities are shown in {\tt fig7a.mov} and {\tt fig7b.mov} respectively.}
\end{figure}
\subsubsection{Response}\label{dwresponsesection}
In this section we apply the method of Section \ref{sec4} to quantify the response of the system \eqref{doublewellfull} to perturbing the potential $U(x) = x^2/2-x^4/4$ with
\[ \delta U(x) = -\exp(-100(x-0.5)^2)/10.\]
This introduces a third potential well centered at $x=0.5$ as shown in Figure \ref{doublewellresponse} and changes the vector field for the variable $x$ in \eqref{doublewellfull} which becomes,
\begin{align}\label{perturbeddoublewell}
\dot x &= x-x^3 - 20(x-0.5)\exp(-100(x-0.5)^2) + (\gamma/\epsilon) y_2.
\end{align}
We first create a benchmark for comparison by estimating the response through a Monte-Carlo simulation. Since the system is assumed to be initially at equilibrium for the unperturbed system \eqref{doublewellfull}, we initialize an ensemble of 100000 members by simply taking the final ensemble from the long simulation in Section \ref{doublewellforecast}. We first use a stochastic Monte-Carlo simulation by replacing $(\gamma/\epsilon)y_2$ in \eqref{perturbeddoublewell} with the stochastic term $\sigma dW_t$ with $\sigma^2 = 0.113$ as suggest in \cite{mg:12}. This yields a reduced model as suggested in \cite{mg:12},
\begin{align}\label{reducedperturbeddoublewell}
dx &= \left(x-x^3 - 20(x-0.5)\exp(-100(x-0.5)^2)\right)\, dt + \sigma \, dW_t
\end{align}
We apply the Euler-Maruyama scheme with integration time step $\delta t = 0.002$ to evolve this ensemble using \eqref{reducedperturbeddoublewell}. At any fixed time $t$ we can then use the centered moments of the ensemble to produce a Monte-Carlo estimate of the response. We also tested the reduced model using our estimate of $D$ to define $\sigma^2 = 2D= 0.107$ and the results were similar the those shown below.
\begin{figure}[h]
\centering
\includegraphics[width=0.425\textwidth]{fig8a.pdf}
\includegraphics[width=0.45\textwidth]{fig8b.pdf}
\caption{\label{doublewellresponse} Left: Evolution of the density and the perturbed potential function. Right: Evolution of the moments compared to those of the stochastic Monte-Carlo simulation of the reduced model with $\sigma^2=0.113$.}
\end{figure}
We then apply the algorithm in Section \ref{sec4} to estimate the response. This technique uses the same training data as Section \ref{doublewellforecast} along with the perturbation $\delta U$ given above to estimate the response without any parametric model. In the first 10 time units, density migrates quickly from the potential well centered at $x=1$ into the new potential well centered at $x=0.5$. Over the next 1000 time units, density migrates from the potential well centered at $x=-1$ into the wells at $x=0.5$ and $x=1$, which are now more stable due to the perturbation. In Figure \ref{doublewellresponse} we summarize the evolution of the density and compare the response estimates of the nonparametric technique to the Monte-Carlo estimates of \eqref{reducedperturbeddoublewell}.
\begin{figure}[h]
\centering
\includegraphics[width=0.45\textwidth]{fig9a.pdf}
\includegraphics[width=0.45\textwidth]{fig9b.pdf}
\caption{\label{doublewellresponsefull} Left: Evolution of the moments compared to the full model with $\epsilon=\sqrt{0.1}$. Right: Evolution of the moments compared to the full model with $\epsilon=\sqrt{0.005}$. The evolution of the densities is shown in {\tt fig9.mov} for $\epsilon = \sqrt{0.1}$.}
\end{figure}
Finally, we compare the non-parametric response estimate to a Monte-Carlo simulation of the full model \eqref{doublewellfull} with perturbation \eqref{perturbeddoublewell}. We initialize an ensemble of 100000 members as in the stochastic simulation, and then integrate the perturbed system using a fourth order Runge-Kutta scheme with $\delta t = 0.002$. As shown in Figure \ref{doublewellresponsefull}, for $\epsilon = \sqrt{0.1}$ this procedure only matches the evolution of the stochastically forced gradient flow for a short time. The evolution of the densities up to time $t=10$ is shown for $\epsilon=\sqrt{0.1}$ in {\tt fig9.mov}. By repeating the Monte-Carlo simulation with $\epsilon = \sqrt{0.005}$ and decreasing the integration time step to $\delta t = 0.0005$ we find better agreement for longer time, however the system still does not behave as a gradient flow for very long times as shown in Figure \ref{doublewellresponsefull}.
\section{Summary and Discussion}\label{sec6}
The techniques developed in this paper solve uncertainty quantification problems using only a training data set without any parametric model. The key tool is the diffusion maps algorithm developed in \cite{diffusion,VB}, which uses the training data to build a non-parametric model for any gradient flow system with isotropic homogeneous stochastic forcing. The model is represented implicitly through a kernel matrix and which converges to the generator of the dynamical system in the limit of large data \cite{VB}. Using properties of gradient flow systems, we can use the eigenvectors of this kernel matrix to approximate the eigenfunctions of the corresponding Fokker-Planck operator. Subsequently, we solved three uncertainty quantification problems, forecasting, filtering, and response estimation, by projecting the associated PDE or SPDE onto this basis of eigenfunctions. In this basis the system becomes a linear ODE or SDE which we can then truncate onto a finite number of modes and solve analytically or approximate numerically.
A key advantage of this technique is that all the densities and eigenfunctions are represented by their values on the training data set, which alleviates the need for a grid. This is important because it is difficult to define a grid on an unknown manifold $\mathcal{M}\subset\mathbb{R}^n$. While our numerical examples focused on one-dimensional systems on $\mathcal{M}=\mathbb{R}$, the theory of \cite{diffusion,VB} allows the gradient flow system to evolve on any $d$-dimensional manifold of $\mathcal{M}\subset\mathbb{R}^n$. It is important to note that these techniques provably solve the full nonlinear uncertainty quantification problem, but only in the limit of infinite training data and an infinite number of eigenvectors. In practice, the amount of data and number of eigenvectors required will depend strongly on the intrinsic dimensionality of the underlying manifold. For high-dimensional systems the amount of data required will quickly make these techniques computationally infeasible even if such large data sets are available.
The techniques are also limited by the assumption of a gradient flow system with stochastic forcing which is isotropic and constant over the entire state space. Concurrent research in \cite{bs:14} suggests that it may be possible to overcome these limitations using different kernels based on non-Euclidean norms which vary over the state space. Overcoming the restriction to gradient flow systems may also be possible using kernels which are not centered as shown in \cite{bs:14}. However, one limitation of the results of \cite{bs:14} is the need to estimate the drift and diffusion coefficients that are used to define the appropriate un-centered kernel. Finding provably convergent algorithms for estimating these coefficients can be difficult, particularly on non-compact domains as shown in \cite{VB}. Instead of explicitly estimating the drift and diffusion coefficients, an alternative approach has been introduced in \cite{BGH14} which implicitly recovers these coefficients. The method of \cite{BGH14} was applied to the problem of forecasting turbulent Fourier modes in \cite{bh:15physd} with small noisy training data sets, whereas here we considered large data sets to verify our approach in the limit of large data. Crucially, the approach of \cite{BGH14} is still based on representing an initial density in the diffusion coordinates used here, and we believe that the UQ methods developed here can be extended to the more general context of \cite{BGH14,bh:15physd}. Regardless of these developments, the fundamental limitation for nonparametric modeling will likely remain the requirement that the manifold be low-dimensional.
\bibliographystyle{siam.bst}
|
\section*{References}}
\makeatletter
\def\ps@pprintTitle{%
\let\@oddhead\@empty
\let\@evenhead\@empty
\def\@oddfoot{\reset@font\hfil\thepage\hfil}
\let\@evenfoot\@oddfoot
}
\makeatother
\begin{document}
\title{Chip-firing games on Eulerian digraphs and {\cl{NP}}-hardness of computing the rank of a divisor on a graph}
\author[anal]{Viktor Kiss}
\ead{<EMAIL>}
\author[cs]{Lilla T\'othm\'er\'esz}
\ead{<EMAIL>}
\address[anal]{Department of Analysis, E\"otv\"os Lor\'and University, P\'azm\'any P\'eter s\'et\'any 1/C, Budapest H-1117, Hungary}
\address[cs]{Department of Computer Science, E\"otv\"os Lor\'and University, P\'azm\'any P\'eter s\'et\'any 1/C, Budapest H-1117, Hungary}
\begin{abstract}
Baker and Norine introduced a graph-theoretic analogue of the Riemann-Roch theory.
A central notion in this theory is the rank of a divisor.
In this paper we prove that computing the rank of a divisor on a graph is \cl{NP}-hard, even for simple graphs.
The determination of the rank of a divisor can be translated to a question about a
chip-firing game on the same underlying graph.
We prove the \cl{NP}-hardness of this question by relating chip-firing on directed and undirected graphs.
\end{abstract}
\begin{keyword}
chip-firing game \sep Riemann-Roch theory \sep computational complexity \sep Eulerian graph
\MSC[2010] 05C57 \sep 05C45 \sep 14H55
\end{keyword}
\maketitle
\section{Introduction}
The Riemann-Roch theory for graphs was introduced by Baker and Norine in 2007 as the discrete analogue of the Riemann-Roch theory for Riemann surfaces \cite{BN-Riem-Roch}.
They defined the notions \emph{divisor, linear equivalence} and \emph{rank} also in this combinatorial setting, and showed that the analogue of basic theorems as for example the Riemann-Roch theorem, remains true.
Theorems like Baker's specialization lemma \cite{specializ_lemma}
establish a connection between the rank of a divisor on a graph and on a curve, which enables a rich interaction of the discrete and continuous theories.
A central notion in the Riemann-Roch theory is the rank of a divisor. The question whether the rank can be computed in polynomial time has been posed in several papers \cite{hladky,Manjunath,pot_theory}, originally attributed to H.~Lenstra.
Let us say a few words about previous work concerning the computation of the rank. Hladk\'y, Kr\'al' and Norine \cite{hladky} gave a finite algorithm for computing the rank of a divisor on a metric graph. Manjunath \cite{Manjunath} gave an algorithm for computing the rank of a divisor on a graph (possibly with multiple edges), that runs in polynomial time if the number of vertices of the graph is a constant. It can be decided in polynomial time, whether the rank of a divisor on a graph is at least $c$, where $c$ is a constant \cite{pot_theory}. Computing the rank of a divisor on a complete graph can be done in polynomial time \cite{cori}. For divisors of degree greater than $2g-2$ (where $g$ is the genus of the graph), the rank can be computed in polynomial time \cite{Manjunath}. On the other hand, there is a generalized model in which deciding whether the rank of a divisor is at least zero is already \cl{NP}-hard \cite{lattice}.
Our main goal in this paper is to show that
computing the rank of a divisor on a graph is \cl{NP}-hard, even for simple graphs. This result implies also the \cl{NP}-hardness of computing the rank of a divisor on a tropical curve by \cite[Theorem 1.6]{Luo}. We also show that deciding whether the rank of a divisor on a graph is at most $k$ is in \cl{NP}.
Our method is the following: We translate the question of
computing the rank of a divisor to a question about the chip-firing
game of Bj\"orner, Lov\'asz and Shor
using the duality between these frameworks discovered by Baker
and Norine \cite{BN-Riem-Roch}. We get that the following question is computationally
equivalent to the determination of the rank: Given an initial
chip-distribution on an (undirected) graph $G$, what is the minimum number of
extra chips we need to put on this distribution to make the game non-terminating.
We first prove the \cl{NP}-hardness of computing the minimum number of chips that enables a non-terminating game on a simple Eulerian digraph by showing that it equals to the number of arcs in a minimum cardinality feedback arc set. This result is mentioned in a note added in proof of \cite{BL92}, where only the larger or equal part is proved. Recently, Perrot and Pham \cite{Perrot} solved an analogous question in the abelian sandpile model, which is a closely related variant of the chip-firing game. Our result follows by applying their method to the chip-firing game.
Then we show that the second question (concerning chip-firing games on directed graphs) can be reduced to the first one (concerning undirected graphs).
In order to do so, to any Eulerian digraph and initial chip-distribution, we assign an undirected graph with a chip-distribution
such that in the short run, chip-firing on the undirected graph imitates chip-firing on the digraph.
\section{Preliminaries}
\subsection{Basic notations}
Throughout this paper, \emph{graph} means a connected undirected graph that can have multiple edges but no loops. A graph is \emph{simple} if it does not have multiple edges. A graph is usually denoted by $G$.
The vertex set and the edge set of a graph $G$ are denoted by $V(G)$ and $E(G)$, respectively.
The degree of a vertex $v$ is denoted by $d(v)$, the multiplicity of the edge $(u,v)$ by $d(u,v)$. The \emph{Laplacian matrix} of a graph $G$ means the following matrix $L$:
$$L(i,j) = \left\{\begin{array}{cl} -d(v_i) & \text{if } i=j \\
d(v_i, v_j) & \text{if } i\neq j.
\end{array} \right.$$
\emph{Digraph} means a (weakly) connected directed
graph that can have multiple edges but no loops.
We usually denote a digraph by $D$. The vertex set and edge
set are denoted by $V(D)$ and $E(D)$, respectively.
For a vertex $v$ the indegree and the outdegree of $v$ are
denoted by $d^-(v)$ and $d^+(v)$, respectively.
A digraph $D$ is \emph{Eulerian} if $d^+(v) = d^-(v)$
for each vertex $v \in V(D)$.
The \emph{head} of the directed edge $(u, v)\in E(D)$
is $v$, and the \emph{tail} of the edge is $u$.
The multiplicity of the directed edge $(u,v)$ is
denoted by $\overrightarrow{d}(u, v)$.
A digraph is \emph{simple} if $\overrightarrow{d}(u, v) \le 1$ for each pair of different
vertices $u, v \in V(D)$.
The Laplacian matrix of a digraph $D$ means the
following matrix $L$:
$$L(i,j) = \left\{\begin{array}{cl} -d^+(v_i) & \text{if } i=j \\
\overrightarrow{d}(v_j, v_i) & \text{if } i\neq j.
\end{array} \right.$$
An important notion concerning digraphs is the feedback arc set. It also plays a crucial role in this paper.
\begin{definition}
A \emph{feedback arc set} of a digraph $D$ is a set of edges
$F \subseteq E(D)$ such that the digraph $D'=(V(D), E(D) \setminus F)$
is acyclic. We denote
$$
{\rm minfas}(D) = \min\{|F| : F \subseteq E(D) \text{ is a feedback
arc set}\}.
$$
\end{definition}
Let $G$ be a graph. An \emph{orientation} of $G$ is a directed
graph $D$ obtained from $G$ by directing each edge.
We identify the vertices of $G$ with the corresponding vertices
of $D$. We denote the
indegree and the outdegree of a vertex $v \in V(G)$ in the
orientation $D$ by $d^-_D(v)$ and $d^+_D(v)$,
respectively.
For a graph $G$ let us denote by $\mathbf{0}_G$
the vector with each
coordinate equal to $0$, and by $\mathbf{1}_G$
the vector with each
coordinate equal to $1$,
where the coordinates are
indexed by the vertices of $G$. For a vertex $v$ of $G$ we
denote the characteristic vector of $v$ by $\mathbf{1}_v$.
We use the same notations for digraphs.
\subsection{Riemann-Roch theory on graphs}
In this section we give some basic definitions of the Riemann-Roch theory on graphs.
The basic objects are called \emph{divisors}.
For a graph $G$, ${\rm Div}(G)$ is the free abelian group on the set of vertices of $G$. An element $f\in {\rm Div}(G)$ is called \emph{divisor}. We either think of a divisor $f\in {\rm Div}(G)$ as a function $f:V(G)\to \mathbb{Z}$, or as a vector $f \in \mathbb{Z}^{|V(G)|}$, where the coordinates are indexed by the vertices of the graph.
The \emph{degree} of a divisor is the following: $${\rm deg}(f)=\sum_{v\in V(G)} f(v).$$
The following equivalence relation on ${\rm Div}(G)$ is called \emph{linear equivalence}:
For $f, g \in {\rm Div}(G)$, $f\sim g$ if there exists a $z\in \mathbb{Z}^{|V(G)|}$ such that $g = f + Lz$.
A divisor $f\in{\rm Div}(G)$ is \emph{effective}, if $f(v)\geq 0$
for each $v\in V(G)$.
\begin{definition}[The rank of a divisor]
For a divisor $f\in {\rm Div}(G)$, the \emph{rank} of $f$ is
\begin{equation*}
\begin{split}
{\rm rank}(f) = \min\{{\rm deg}(g) - 1 : \text{$g \in {\rm Div}(G)$, $g$ is effective, } \\
\text{$\nexists h \in {\rm Div}(G)$ such that $h \sim f-g$ and $h$ is effective}\}.
\end{split}
\end{equation*}
When we wish to emphasize the underlying graph, we write ${\rm rank}_G(f)$ instead of ${\rm rank}(f)$.
\end{definition}
\subsection{Chip-firing}
It was noted already by Baker and Norine \cite{BN-Riem-Roch}, that there is a duality between divisors on graphs and the objects of the chip-firing game, as defined by Bj\"orner, Lov\'asz and Shor \cite{BLS91}. Using this duality we can translate some questions in divisor theory to questions in chip-firing. We would like to use the latter language in the article, so let us include here a short introduction to chip-firing.
\begin{remark}
Often the term ``chip-firing game'' is used also in the setting of Baker and Norine, but for clarity, we only use this term for the game of \cite{BLS91}.
\end{remark}
The theory of chip-firing games was developed both for
graphs (see \cite{BLS91}) and digraphs (see \cite{BL92}).
Although the notions of the Riemann-Roch theory on graphs are in duality with
notions concerning the chip-firing game on undirected graphs, later on,
we also need chip-firing on digraphs.
The basic idea is that on each vertex of a graph (or digraph),
there is a certain amount of chips. If a vertex has at least as many chips as its degree
(in the directed case: as its outdegree), then it can be fired.
In the undirected case, this means, that the vertex passes
a chip to its neighbors along each edge incident to it, and so the number of chips on itself decreases by its degree.
In the directed case, the fired vertex passes a chip along each outgoing edge.
In fact, if we think of an undirected graph as a special digraph where we replace
each edge by a pair of oppositely directed edges, then the two definitions coincide for these graphs.
Now we give the exact definitions.
Let $H$ be an undirected graph or a digraph.
A \emph{chip-distribution}, or \emph{distribution}, is a function
$x:V(H) \to \mathbb{Z}^+\cup \{0\}$. We sometimes say: vertex
$v$ has $x(v)$ chips. We use the notation $|x|$ for the number of
chips in the distribution $x$, i.e., $|x| = \sum_{v \in V(H)} x(v)$.
We denote the set of chip-distributions on $H$ by ${\rm Chip}(H)$.
\emph{Firing} a vertex $v$ means taking the
new chip-distribution $x + L\mathbf{1}_v$ instead of $x$.
Note that the Laplacian matrix $L$ is different in the
undirected and in the directed case, and that in both cases $|x+L\mathbf{1}_v|=|x|$
so a firing preserves the number of chips.
A vertex $v \in V(H)$ is \emph{active}
(with respect to $x$) if after firing it, it still has
a nonnegative number of chips, i.e., in the undirected case
if $x(v)\geq d(v)$, while in the directed case if
$x(v) \ge d^+(v)$.
The firing of a vertex $v \in V(H)$ is \emph{legal},
if $v$ was active before the firing.
A \emph{legal game} is a sequence of distributions in which
every distribution is obtained from the previous one by a legal
firing.
A game terminates if there is no active vertex with respect to the
last distribution.
The following theorem was proved by Bj\"orner, Lov\'asz and Shor
for undirected graphs and by Bj\"orner and Lov\'asz for digraphs.
(Originally the theorem for the undirected case was proved for
simple graphs, but the proof works also for graphs with multiple edges.)
\begin{thm}[{\cite[Theorem 1.1]{BLS91}}, {\cite[Theorem 1.1]{BL92}}] \label{thm::vegesseg_kommutativ}
Let $H$ be a graph or a digraph and let $x \in {\rm Chip}(H)$
be a distribution. Then starting from $x$, either every legal game can be continued indefinitely, or every legal game terminates after the same number of moves with the same final distribution. Moreover, the number of times a given node is fired is the same in every legal game.
\end{thm}
Let us call a chip-distribution $x\in {\rm Chip}(H)$ \emph{terminating}, if every legal chip-firing game played starting from $x$ terminates, and call it \emph{non-terminating}, if every legal chip-firing game played starting from $x$ can be continued indefinitely. According to Theorem \ref{thm::vegesseg_kommutativ}, a chip-distribution is either terminating or non-terminating.
It can easily be seen by the pigeonhole principle that
if for a graph $G$ a distribution $x \in {\rm Chip}(G)$ has $|x| > 2|E(G)| - |V(G)|$
then $x$ is non-terminating (see \cite{BLS91}). And similarly,
for a digraph $D$ if a distribution $x \in {\rm Chip}(D)$
has $|x| > |E(D)| - |V(D)|$ then $x$ is non-terminating (see \cite{BL92}).
From this it follows that the following quantity, which measures how far a
given distribution is from being non-terminating, is well defined.
For a distribution $x \in {\rm Chip}(H)$, let
$$
{\rm dist}(x) = \min\{|y| : y \in {\rm Chip}(H), x + y \text{ is non-terminating}\}.
$$
We say that ${\rm dist}(x)$ is the
\emph{distance} of $x$ from non-terminating distributions.
Note that ${\rm dist}(\mathbf{0}_H)$ is exactly the minimum number of chips in a non-terminating distribution on the graph/digraph $H$.
\subsection{Chip-firing and the Riemann-Roch theory}
Now we describe the duality between divisors on graphs and chip-distri\-bu\-ti\-ons discovered by Baker and Norine \cite{BN-Riem-Roch}.
Let $G$ be a graph and let $K^+ = K^+_G$ be the chip-distribution with
$K^+(v)=d(v) - 1$ for each $v \in V(G)$.
For a divisor $f\in {\rm Div}(G)$ with $f(v)\leq d(v)-1$ for each $v\in V(G)$, we have $(K^+ - f)(v)\geq 0$ for each $v\in V(G)$, therefore $K^+ - f\in {\rm Chip}(G)$.
We call $K^+ - f$ the \emph{dual pair} of $f$. Note that each chip-distribution is a dual pair of some divisor.
\begin{prop}[{\cite[Corollary 5.4]{BN-Riem-Roch}}]
\label{prop::dual_pair}
For a divisor $f\in {\rm Div}(G)$ with $f(v)\leq d(v)-1$ for each $v\in V(G)$, there exists an effective divisor equivalent to $f$ if and only if $K^+ - f$ is a terminating distribution.
\end{prop}
\begin{remark}
We could have defined the chip-firing game for not necessarily nonnegative distributions as well with the same rules (only active vertices can fire). In this case Theorem \ref{thm::vegesseg_kommutativ} would still hold, and we could have a dual pair for any divisor, but this is not necessary for our purposes.
\end{remark}
The following is a straightforward consequence of
Proposition \ref{prop::dual_pair}.
\begin{prop}
\label{prop::rank_of_dual}
Let $f \in {\rm Div}(G)$ be a divisor with $f(v)\leq d(v)-1$ for each $v\in V(G)$,
and let
$x \in {\rm Chip}(G)$ be its dual pair. Then ${\rm rank}(f) = {\rm dist}(x) - 1$.
\end{prop}
\subsection{{\cl{NP}}-hardness results}
In Section \ref{sec::dist_digraph}, based on recent results of Perrot
and Pham \cite{Perrot}, we prove the following.
\begin{thm} \label{thm::rang_eulerben_nehez}
Given a digraph $D$, computing ${\rm dist}(\mathbf{0}_D)$ is \cl{NP}-hard,
even for simple Eulerian digraphs.
\end{thm}
Using this result we prove the main theorem of this article.
\begin{thm} \label{thm::chip_rang_NP-teljes}
For a distribution $x\in {\rm Chip}(G)$ on a graph $G$, computing ${\rm dist}(x)$ is \cl{NP}-hard.
\end{thm}
The proof can be found in Section \ref{sec::rank_chip_dist_NP_hard}.
As a corollary of the theorem and Proposition \ref{prop::rank_of_dual}, we get the following.
\begin{coroll}
For a divisor $f\in {\rm Div}(G)$ on a graph $G$, computing ${\rm rank}(f)$ is \cl{NP}-hard, even for a divisor $f$ with $f(v)\leq d(v)-1$ for every $v \in V(G)$.
\end{coroll}
In \cite{hladky}, Hladk\'y, Kr\'al' and Norine prove the following statement:
\begin{prop}[{\cite[Corollary 22.]{hladky}}] \label{prop::hladky_elosztos}
Let $f$ be a divisor on a graph $G$.
Let $G'$ be the simple graph obtained from $G$ by subdividing
each edge of $G$ by an inner point and
let $f'$ be the divisor on $G'$ that agrees with $f$ on
the vertices of $G$ and has value $0$ on new points.
Then ${\rm rank}_G(f)={\rm rank}_{G'}(f')$.
\end{prop}
By dualizing this statement, we get the following: For a distribution $x\in{\rm Chip}(G)$,
if we get $x'\in {\rm Chip}(G')$ from $x$ so that we put $d(v) - 1 - 0 = 1$
chip on each new vertex, and on the vertices of $G$, $x'$ agrees with $x$,
then
${\rm dist}_G(x) = {\rm rank}_G(K^+_G - x) + 1 = {\rm rank}_{G'}(K^+_{G'} - x') + 1 =
{\rm dist}_{G'}(x')$.
When proving Theorem \ref{thm::chip_rang_NP-teljes}, we show a
somewhat stronger statement: By Remark
\ref{rem::dist_NP_erosites}, computing ${\rm dist}$ is \cl{NP}-hard even
for graphs with $|E(G)|\leq 9|V(G)|^5$. For such a $G$,
$|V(G')|\leq |V(G)|+|E(G)|\leq 10|V(G)|^5$. Hence $G'$ and $x'$ can be computed in polynomial time for such a graph $G$ and $x\in {\rm Chip}(G)$, giving the following corollary.
\begin{coroll}
For a distribution $x\in {\rm Chip}(G)$ on a simple graph $G$, computing ${\rm dist}(x)$ is \cl{NP}-hard.
\end{coroll}
Using Proposition \ref{prop::rank_of_dual} again,
we have the following.
\begin{coroll}
For a divisor $f\in {\rm Div}(G)$ on a simple graph $G$, computing ${\rm rank}(f)$ is \cl{NP}-hard.
\end{coroll}
Using a result of \cite{Luo}, we get that computing the rank of divisors is also \cl{NP}-hard for so called tropical curves.
Informally, a metric graph is a graph, where each edge has a positive length, and we consider our graph to be a metric space (the inner points of the edges are also points of this metric space). Tropical curve is more general in that we also allow some edges incident with vertices of degree one to have infinite length.
A divisor on a tropical curve is an integer-valued function on the curve with only finitely many nonzero values. The notions of the degree of a divisor, linear equivalence, effective divisor and the rank can be defined as well, see \cite{hladky}.
A metric graph $\Gamma$ corresponds to the graph $G$, if $\Gamma$ is obtained from $G$ by assigning some positive length to each edge.
\begin{thm}[{\cite[Theorem 1.6]{Luo}}] \label{thm::metrikus_rang_rang}
Let $f$ be a divisor on a graph $G$, and $\Gamma$ be a metric graph corresponding to $G$. Then ${\rm rank}_G(f)={\rm rank}_{\Gamma}(f)$.
\end{thm}
As a metric graph is a special tropical curve, we get the following corollary:
\begin{coroll}
For a tropical curve $\Gamma$, $f\in {\rm Div}(\Gamma)$,
computing ${\rm rank}(f)$ is \cl{NP}-hard.
\end{coroll}
From the positive side, we show the following:
\begin{prop}
Deciding whether for a given divisor $f$ on a graph $G$, and integer $k$, ${\rm rank}(f)\leq k$ is in \cl{NP}.
\end{prop}
\begin{proof}
For an input $(f,k)$ with ${\rm rank}(f)\leq k$, our witness is the divisor $g\geq 0$ such that ${\rm rank}(f-g)=-1$, and ${\rm deg}(g)\leq k+1$ (such a $g$ exists because ${\rm rank}(f)\leq k$).
First, we need to check that $g$ can be given so that it has size polynomial in the size of $(f,k)$. As ${\rm deg}(g)\leq k+1$, and $g\geq 0$, we have $g(v)\leq k+1$ for each vertex $v$. Therefore, the size of $g$ is at most $O(|V(G)|\cdot \log k)$.
On the other hand, it can be checked in polynomial time if ${\rm rank}(f-g)=-1$ \cite{pot_theory}, and also whether ${\rm deg}(g)\leq k+1$.
\end{proof}
By applying Proposition \ref{prop::rank_of_dual}, deciding whether for a given chip-distribution $x$, and integer $k$, ${\rm dist}(x)\leq k$ is also in \cl{NP}.
\section{Minimal non-terminating distributions on Eulerian digraphs}
\label{sec::dist_digraph}
In this section we prove
Theorem \ref{thm::rang_eulerben_nehez}, i.e.,
that computing the minimum number of chips in a non-terminating distribution is \cl{NP}-hard for a simple Eulerian digraph $D$.
We use the method of Perrot and Pham. In the paper \cite{Perrot},
they prove the \cl{NP}-hardness of an analogous question in the abelian sandpile model,
which is a closely related variant of the chip-firing game.
Using the ideas of \cite{Perrot}, we first give a formula for the minimum
number of chips in a non-terminating distribution on an Eulerian digraph.
As a motivation, let us have a look at the analogous question on undirected graphs, which was solved by Bj\"orner, Lov\'asz and Shor.
\begin{thm}[{\cite[Theorem 2.3]{BLS91}}]
\label{thm::also_korlat}
Let $G$ be a graph. Then ${\rm dist}(\mathbf{0}_G)=|E(G)|$.
\end{thm}
We sketch the proof as a motivation for the directed
case.
\begin{proof}
First we prove the following useful lemma.
\begin{lemma}[\cite{BLS91}]
\label{lem::chip_dist_over_acyclic}
Let $D$ be an acyclic orientation of $G$ and let
$x\in {\rm Chip}(G)$ be a distribution with $x(v) \ge d^-_D(v)$
for each $v \in V(G)$. Then $x$ is non-terminating.
\end{lemma}
\begin{proof}
Since the orientation is acyclic, there is a sink, i.e., a
vertex $v_0 \in V(G)$ with $d(v_0) = d^-_D(v_0) \leq x(v_0)$.
Hence $v_0$ is active with respect to $x$. Fire $v_0$ and
denote the resulting distribution by $x'$.
Reverse the direction of the edges incident to $v_0$ and
denote the resulting directed graph by $D'$. It is easy to
see that $D'$ is acyclic and $d^-_{D'}(v) \le x'(v)$ for each
$v \in V(G)$. Hence we can repeat the above argument. This
shows that the distribution $x$ is indeed non-terminating.
\end{proof}
Now taking an acyclic orientation $D$ of $G$ and setting
$x(v) = d^-_D(v)$ for each $v \in V(G)$ we have a distribution
with $|x| = |E(G)|$ that is non-terminating from the lemma.
This shows that ${\rm dist}(\mathbf{0}_G)\leq |E(G)|$.
For proving ${\rm dist}(\mathbf{0}_G)\geq |E(G)|$, take a non-terminating distribution
$x \in {\rm Chip}(G)$. It is enough to show that $|x|\geq |E(G)|$.
Since in a non-terminating game every vertex is fired infinitely often
(see \cite[Lemma 2.1]{BLS91}), after finitely many firings,
every vertex of $G$ has been fired at least once.
Let $x'$ be the distribution at such a moment.
Then $|x| = |x'|$.
Let $D$ be the orientation of $G$ that we get by
directing each edge toward the vertex whose last firing occurred
earlier. It is straightforward to check that
$x'(v) \ge d^-_D(v)$ for each $v \in V(G)$.
This fact implies that $|x| = |x'| \ge |E(G)|$,
completing the proof.
\end{proof}
Now let us consider Eulerian digraphs.
\begin{thm}
\label{thm::dist = minfas}
Let $D$ be an Eulerian digraph.
Then ${\rm dist}(\mathbf{0}_D) = {\rm minfas}(D)$.
\end{thm}
This theorem is already stated in a note added in
proof of \cite{BL92}, but there only the direction
${\rm dist}(\mathbf{0}_D) \ge {\rm minfas}(D)$ is proved.
We give a proof following ideas of Perrot and Pham \cite{Perrot}.
The idea of the proof can be
thought of as the generalization of the idea of the proof of
Theorem \ref{thm::also_korlat}.
For proving Theorem \ref{thm::dist = minfas} we need a classical result about chip-firing on an Eulerian digraph:
\begin{prop}[{\cite[Lemma 2.1]{BL92}}] \label{prop::euler_vegt_mind_vegt_sokszor_lo}
On an Eulerian digraph $D$ if a chip-distribution is non-terminating then in any legal game every vertex is fired infinitely often.
\end{prop}
\begin{proof}[Proof of Theorem \ref{thm::dist = minfas}]
The key lemma is the following observation of Perrot and Pham; they proved it for the recurrent configurations of the abelian sandpile model, but the two models are very closely related.
\begin{lemma}[{\cite{Perrot}}] \label{lemma::euler_ir_felett_vegtelen}
Let $F\subseteq E(D)$ be a minimum cardinality feedback arc set. Denote by $d^+_F(v)$ and $d^-_F(v)$ the outdegree and indegree of a vertex $v$ in the digraph $D_F=(V(D), F)$.
Then a distribution $x\in {\rm Chip}(D)$ satisfying
\begin{equation}
\label{eq::feedback_felett}
x(v)\geq d^-_F(v) \text{ for every $v\in V(D)$}
\end{equation}
is non-terminating.
\end{lemma}
\begin{proof}
First we prove that if $F$ is a minimum cardinality
feedback arc set then there exists a vertex $v \in V(D)$ such
that among the edges incident to $v$, $F$ contains exactly
the in-edges of $v$.
Let $A = E(D) \setminus F$. From the definition of feedback
arc set, $D_A = (V(D), A)$ is an acyclic graph.
Therefore, it has a source $v_0$.
We claim that no out-edge of $v_0$ is in $F$.
Indeed, if some out-edges of $v_0$ would be in $F$, removing
them from $F$ would mean adding some out-edges to the source $v_0$ in
$D_A$, which cannot create a cycle. So we could get a smaller
feedback arc set.
The fact that $v_0$ is a source in $D_A$ means that all the
in-edges of $v_0$ are in $F$. Hence from the edges
incident to $v_0$, $F$ contains exactly the in-edges of
$v_0$.
Now take such a vertex $v_0$.
From \eqref{eq::feedback_felett}, the choice of $v_0$ and
the fact that $D$ is Eulerian, we have that
$x(v_0)\geq d^-_F(v_0)=d^-(v_0)=d^+(v_0)$,
therefore $v_0$ is active with respect to $x$.
Fire $v_0$. Let $x'$ be the resulting distribution.
We show, that we can modify the feedback arc set $F$, such
that for the new feedback arc set $F'$ we have
$x'(v)\geq d^-_{F'}(v)$ for every $v\in V(D)$.
Let $F'$ be the set of arcs obtained from $F$ by removing the
in-edges of $v_0$ and adding the out-edges of $v_0$
(see Figure \ref{fig::feedback}).
Then $D_{A'} = (V(D), A')$ with $A'=E(D)\setminus F'$ is
acyclic, since the new edges are all incident to a sink in
$D_{A'}$.
Moreover, since $D$ is Eulerian, and from the choice of $v_0$, we have $|F'|=|F|$, hence
$F'$ is also a feedback arc set of minimum cardinality.
It is straightforward to check that indeed
$x'(v)\geq d^-_{F'}(v)$ for every $v \in V(D)$.
So we are again in the starting situation, which shows that
$x$ is indeed non-terminating.
\end{proof}
\begin{figure}[ht]
\begin{center}
\begin{tikzpicture}[->,>=stealth',auto,scale=2.3,
thick,every node/.style={circle,draw,font=\sffamily\small}]
\node[label=left:0] (1) at (0, 1) {};
\node[label=left:0] (2) at (0, -1) {};
\node[label=below:0] (3) at (-1, 0) {};
\node[label=below:2] (4) at (1, 0) {};
\path[every node/.style={font=\sffamily\small}]
(1) edge [dashed] node {} (4)
(3) edge node {} (1)
(2) edge node {} (3)
(4) edge node {} (2)
(4) edge [bend left=20] node {} (3)
(3) edge [dashed,bend left=20] node {} (4);
\end{tikzpicture}
\hspace{0.2cm}
\begin{tikzpicture}[->,>=stealth',auto,scale=2.3,
thick,every node/.style={circle,draw,font=\sffamily\small}]
\node[label=left:0] (1) at (0, 1) {};
\node[label=left:1] (2) at (0, -1) {};
\node[label=below:1] (3) at (-1, 0) {};
\node[label=below:0] (4) at (1, 0) {};
\path[every node/.style={font=\sffamily\small}]
(1) edge node {} (4)
(3) edge node {} (1)
(2) edge node {} (3)
(4) edge [dashed] node {} (2)
(4) edge [dashed,bend left=20] node {} (3)
(3) edge [bend left=20] node {} (4);
\end{tikzpicture}
\caption{An example for simultaneously firing a vertex and changing the feedback arc set. The arcs of the feedback arc sets are drawn by dashed lines.}
\label{fig::feedback}
\end{center}
\end{figure}
Now take a feedback arc set $F$ of minimum cardinality,
and let $x(v)=d^-_F(v)$ for every $v\in V(D)$.
Then $|x|=|F|={\rm minfas}(D)$, and from the lemma, $x$ is non-terminating.
This proves that ${\rm dist}(\mathbf{0}_D)\leq {\rm minfas}(D)$.
The direction ${\rm dist}(\mathbf{0}_D)\geq {\rm minfas}(D)$ is shown in the note
added in proof of \cite{BL92} for a general digraph,
however, as we need implications of its idea, we also include this part of the proof.
Take a non-terminating
distribution $x$. It is enough to prove that $|x|\geq {\rm minfas}(D)$.
Let us play a chip-firing game with initial distribution $x$.
Proposition \ref{prop::euler_vegt_mind_vegt_sokszor_lo} says that after finitely many steps, every vertex
has fired. Play until such a moment, and let the distribution at that moment be $x'$.
Let $A$ be the following set of edges:
$$
A=\{(u,v) \in E(D): \textrm{ the last firing of $u$ preceeds the last firing of $v$}\}.
$$
As every vertex has fired, $A$ is well defined.
Let $v_1,v_2, \dots v_{|V(D)|}$ be the ordering of the vertices by the time of their last firing.
Then $v_1,v_2, \dots v_{|V(D)|}$ is a topological order of $D_A=(V(D),A)$, so $D_A$ is acyclic, hence $F=E(D)\setminus A$ is a feedback arc set.
We show that $x'(v)\geq d^-_F(v)$ for every $v\in V(D)$.
For $1 \le i \le |V(D)|$ the vertex $v_i$ has $d^-_F(v_i)=\sum_{j>i}\overrightarrow{d}(v_j,v_i)$.
After its last firing, $v_i$ had a nonnegative number of chips.
Since then, it kept all chips it received. And as $v_{i+1}, \dots, v_{|V(D)|}$
all fired since the last firing of $v_i$, it received at least $\sum_{j>i}\overrightarrow{d}(v_j,v_i)=d^-_F(v_i)$ chips. So indeed, we have $x'(v_i)\geq d^-_F(v_i)$.
Therefore $|x|=|x'| \ge |F|\geq {\rm minfas}(D)$.
\end{proof}
Note that in the above setting, starting from $x'$, then firing the vertices in the order $v_1,v_2,\dots, v_{|V(D)|}$ (once each) is a legal game. Indeed, we proved that $x'(v_i)\geq d^-_F(v_i)=\sum_{j>i}\overrightarrow{d}(v_j,v_i)$. After firing $v_1, \dots v_{i-1}$, the vertex $v_i$ receives $\sum_{j<i}\overrightarrow{d}(v_j,v_i)$ more chips, so it indeed becomes active
($d^+(v_i)=d^-(v_i)=\sum_{j\neq i}\overrightarrow{d}(v_j,v_i))$ as we did not allow loops).
We need this observation in the next section, so we state it as a proposition:
\begin{prop} \label{prop::euler_ha mar mindenki lott}
In a chip-firing game on an Eulerian digraph $D$, if at some moment every vertex has already fired then there is an order of the vertices in which they can be legally fired once each, starting from that moment.\qed
\end{prop}
It is worth noting that on an Eulerian digraph if starting from an initial
distribution $x$ we fired each vertex exactly once, then we get back to distribution $x$:
each vertex $v$ gave and received $d^-(v)=d^+(v)$ chips.
Finally, we prove Theorem \ref{thm::rang_eulerben_nehez}.
\begin{proof}[Proof of Theorem \ref{thm::rang_eulerben_nehez}]
Perrot and Pham proved that computing ${\rm minfas}(D)$ for a
simple Eulerian digraph $D$ is \cl{NP}-hard \cite[Theorem 2]{Perrot},
by reducing it to the \cl{NP}-hardness of computing ${\rm minfas}(D)$
for general digraphs.
From this, and from Theorem \ref{thm::dist = minfas}, the
statement follows.
\end{proof}
\section{The distance from non-terminating distributions is {\cl{NP}}-hard on graphs}
\label{sec::rank_chip_dist_NP_hard}
In this section we prove Theorem \ref{thm::chip_rang_NP-teljes}, the main theorem of this article.
In our proof of the \cl{NP}-hardness, we rely on the fact
that a terminating chip-firing game on an Eulerian digraph $D$
terminates after at most $2|V(D)|^2 |E(D)| \Delta(D)$ steps
(see \cite[Corollary 4.9]{BL92}), where $\Delta(D)$ denotes
the maximum of all the indegrees and the outdegrees of $D$, i.e.,
$\Delta(D) = \max_{v \in V(D)}\max\{d^-(v), d^+(v)\}$.
With this in mind, we define the following transformation:
\begin{definition}
Let $\varphi$ be the following transformation, assigning an undirected graph $G=\varphi(D)$ to any digraph $D$:
Split each directed edge by an inner point, and substitute the tail segment by $M=8|V(D)|^2 |E(D)| \Delta(D)$ parallel edges. Then forget the orientations.
We maintain the effect of the transformation by a bijective function $\psi: (V(D)\cup E(D)) \to V(\varphi(D))$:
For a vertex $v\in V(D)$ let $\psi(v)$ be the corresponding vertex of $\varphi(D)$. For an edge $e \in E(D)$, let $\psi(e)$ be the vertex with which we have split $e$.
\end{definition}
Then the degrees in $\varphi(D)$ are the following:
\begin{equation*}
d(v) = \left\{\begin{array}{cl} d^+\left(\psi^{-1}(v)\right)\cdot M + d^-(\psi^{-1}(v)) & \text{if } \psi^{-1}(v) \in V(D) \\
M + 1 & \text{if } \psi^{-1}(v) \in E(D).
\end{array} \right.
\end{equation*}
\begin{figure}[ht]
\begin{center}
\begin{tikzpicture}[->,>=stealth',auto,scale=2.5,
thick,every node/.style={circle,draw,font=\sffamily\small}]
\node (1) at (0, 1) {\small $v_1$};
\node (2) at (0, -1) {\small $v_2$};
\node (3) at (-1, 0) {\small $v_3$};
\node (4) at (1, 0) {\small $v_4$};
\path[every node/.style={font=\sffamily\small}]
(1) edge node {$e_1$} (4)
(3) edge node {$e_2$} (1)
(2) edge node {$e_3$} (3)
(4) edge node {$e_4$} (2)
(4) edge [bend left=20] node {$e_5$} (3)
(3) edge [bend left=20] node {$e_6$} (4);
\end{tikzpicture}
\begin{tikzpicture}[auto,scale=2.5,
thick,every node/.style={circle,draw,font=\sffamily\small}]
\node (1) at (0, 1) {\tiny $\psi(v_1)$};
\node (2) at (0, -1) {\tiny $\psi(v_2)$};
\node (3) at (-1, 0) {\tiny $\psi(v_3)$};
\node (4) at (1, 0) {\tiny $\psi(v_4)$};
\node (41) at (0.5, 0.5) {\tiny $\psi(e_1)$};
\node (13) at (-0.5, 0.5) {\tiny $\psi(e_2)$};
\node (32) at (-0.5, -0.5) {\tiny $\psi(e_3)$};
\node (24) at (0.5, -0.5) {\tiny $\psi(e_4)$};
\node (34) at (0, -0.25) {\tiny $\psi(e_5)$};
\node (43) at (0, 0.25) {\tiny $\psi(e_6)$};
\path[every node/.style={font=\sffamily\small}]
(4) edge (41)
(41) edge [bend right=10] (1)
(41) edge [bend left=10] (1)
(41) edge (1)
(1) edge (13)
(13) edge [bend right=10] (3)
(13) edge [bend left=10] (3)
(13) edge (3)
(3) edge (32)
(32) edge [bend right=10] (2)
(32) edge [bend left=10] (2)
(32) edge (2)
(2) edge (24)
(24) edge [bend right=10] (4)
(24) edge [bend left=10] (4)
(24) edge (4)
(3) edge (34)
(34) edge [bend right=10] (4)
(34) edge [bend left=10] (4)
(34) edge (4)
(4) edge (43)
(43) edge [bend right=10] (3)
(43) edge [bend left=10] (3)
(43) edge (3);
\end{tikzpicture}
\caption{A schematic picture for a digraph $D$ and the corresponding $\varphi(D)$. In the reality the multiple edges should be 1536-fold.}
\end{center}
\end{figure}
Let us define a certain chip-distribution on the graph $\varphi(D)$:
\begin{definition}[base-distribution]
Let $base_D \in {\rm Chip}(\varphi(D))$ on a vertex $v \in V(\varphi(D))$ be the following:
\begin{equation*}
\begin{split}
base_D(v) = \left\{\begin{array}{cl} d^+\left(\psi^{-1}(v)\right)\cdot M & \text{if } \psi^{-1}(v) \in V(D) \\
M/2 & \text{if } \psi^{-1}(v) \in E(D).
\end{array} \right.
\end{split}
\end{equation*}
\end{definition}
The key lemma in our proof of Theorem \ref{thm::chip_rang_NP-teljes} is the following:
\begin{lemma}
For an Eulerian digraph $D$, ${\rm dist}_D(\mathbf{0}_D) = {\rm dist}_{\varphi(D)}(base_D)$.
\end{lemma}
\begin{proof}
Let $G = \varphi(D)$.
First we show that ${\rm dist}_D(\mathbf{0}_D) \ge {\rm dist}_{G}(base_D)$.
Let $x \in {\rm Chip}(D)$ be a non-terminating chip-distribution such that $|x|$ is minimal.
We can assume that there is an order
of the vertices of $D$ such that from initial distribution
$x$ we can fire the vertices in that order (once each).
Otherwise, from Proposition \ref{prop::euler_vegt_mind_vegt_sokszor_lo} we can play a
chip-firing game from $x$ until each vertex has fired. Denoting
the distribution at that moment by $x'$, from Proposition
\ref{prop::euler_ha mar mindenki lott} for $x'$ there is such an order. As firing
does not change the number of chips in the game, $|x'|$ is still minimal, so we can
substitute $x$ with $x'$.
Let $y\in{\rm Chip}(G)$ be the distribution ``$x + base_D$'', i.e., for a vertex
$v \in V(D)$ let $y(\psi(v)) = x(v) + base_D(\psi(v))$ and for an
edge $e \in E(D)$ let $y(\psi(e)) = base_D(\psi(e))$.
Since $y(w) \ge base_D(w)$ for each $w \in V(G)$ and
$|y - base_D| = |x| = {\rm dist}_D(\mathbf{0}_D)$,
it is enough to show that $y$ is non-terminating.
For that, it is enough to show that we can fire each vertex of $G$
exactly once in some order. Then each vertex $w \in V(G)$ gives and receives $d(w)$ chips,
so we get back to the distribution $y$ and can repeat this period indefinitely.
To get such an order of the vertices of $G$, we will play the chip-firing game simultaneously on $D$ and $G$.
To firing a vertex $v$ in $D$, let the corresponding firings in
$G$ be: Fire $\psi(v)$, then fire $\psi(e)$ for every out-edge
$e$ of $v$ (in some order).
\begin{claim}
If a sequence of firings of length $k \leq M/2$ on
$D$ with initial distribution $x$ is legal then the sequence of the
corresponding firings on $G$ with
initial distribution $y$ is also legal. Moreover, if we
denote the resulting distribution on $D$ by $\tilde{x}$ and on $G$ by
$\tilde{y}$ then
\begin{equation}
\label{eq::vertex}
\tilde{y}(\psi(v)) = \tilde{x}(v) + d^+(v) \cdot M
\text{ for each $v \in V(D)$}
\end{equation}
and
\begin{equation}
\label{eq::edge}
M/2 - k \le \tilde{y}(\psi(e)) \le M/2 + k \text{ for each $e \in E(D)$.}
\end{equation}
\end{claim}
\begin{proof}
We show this by induction on $k$. For $k = 0$ this is trivial.
Take a sequence of firings of length $k \leq M/2$ and assume that the
claim holds for $k - 1$. Denote the distribution on $D$ after
the first $k - 1$ firings by $x'$ and the corresponding
distribution on $G$ by $y'$. Assume that the vertex $v$ is the
last to be fired on $D$. Hence $v$ is active with respect
to $x'$. Denote the distribution after firing $v$ by
$x''$. Vertex $\psi(v)$ is active with respect to $y'$,
since using \eqref{eq::vertex} of the induction hypothesis, the fact that $v$
is active with respect to $x'$ and that $D$ is Eulerian, we get that
$y'(\psi(v)) = x'(v) + d^+(v) \cdot M \ge
d^+(v) + d^+(v) \cdot M =
d^-(v) + d^+(v) \cdot M = d(\psi(v))$.
Fire $\psi(v)$. Now for each out-edge $e$ of $v$ the vertex
$\psi(e)$ is active, since using \eqref{eq::edge} of the induction hypothesis,
it has at least
$M + y'(\psi(e)) \ge M + M/2 - (k - 1) \geq M + 1 = d(\psi(e))$ chips.
Fire these vertices in an arbitrary order.
(Firing one leaves the others active.)
Denote by $y''$ the resulting distribution.
It is easy to check that the distributions $x''$ and $y''$ satisfy
conditions \eqref{eq::vertex} and \eqref{eq::edge}.
\end{proof}
We have chosen the distribution $x$ such that we can fire the vertices
of $D$ in some order (once each) with initial distribution $x$.
This is a legal sequence of firings of length $|V(D)|<M/2$.
According to the previous
claim, the sequence of the corresponding firings on $G$ is also legal.
Moreover, on $G$ we also fire each
vertex exactly once. This finishes the proof of the direction
${\rm dist}_D(\mathbf{0}_D) \ge {\rm dist}_G(base_D)$.
Now we prove that ${\rm dist}_D(\mathbf{0}_D) \le {\rm dist}_G(base_D)$.
For this, let $y \in {\rm Chip}(G)$ be a minimal non-terminating
chip-distribution with $base_D(w) \le y(w)$ for each $w \in V(G)$.
Let $x(v) = y(\psi(v)) - base_D(\psi(v))$ on each $v\in V(D)$.
It is enough to show that $x$ is non-terminating.
First note that ${\rm dist}_D(\mathbf{0}_D) \le |E(D)| - |V(D)| + 1$,
since having a chip-distribu\-ti\-on
with at least $|E(D)|-|V(D)|+1$ chips, at every stage of the game
at least one of the vertices has the sufficient
number of chips to fire.
Consequently, using also the first part of the lemma, we have
that $|y - base_D| = {\rm dist}_G(base_D) \le {\rm dist}_D(\mathbf{0}_D) \le
|E(D)|-|V(D)| + 1 \le \frac{1}{8}M$.
Now we play the game on $G$ and $D$ simultaneously
from initial distributions $y$ and $x$, respectively,
in the following way. Let a step be the following:
Choose a vertex $v \in V(D)$ for which $\psi(v)$ can fire.
On $G$ fire $\psi(v)$, then for every out-edge $e$ of $v$,
fire $\psi(e)$. On $D$ fire $v$.
We show that
for $\frac{3}{8}M \ge
2|V(D)|^2 |E(D)| \Delta(D) + 1$ steps we
can play this legally on both graphs.
Note first that for an edge $e$ of $D$, the change of the
number of chips on $\psi(e)$ is at most one after each step.
Hence at the beginning of a step a vertex of $G$ of
the form $\psi(e)$ can have at most
$M/2 + |y - base_D| + \frac{3}{8}M \leq
M/2 + \frac{1}{8}M + \frac{3}{8}M < M + 1 = d(\psi(e))$ chips, so
it cannot be fired. It also follows from this that on every such
vertex the number of chips is positive, since it is at least
$M/2 - \frac{3}{8}M > 0$.
But $y$ is a non-terminating distribution, hence at the beginning of a
step we can find an active vertex, which therefore must be of the form $\psi(v)$ with $v \in V(D)$.
After firing $\psi(v)$, $\psi(e)$ becomes active
for every out-edge $e$ of $v$, since $\psi(e)$
had a positive number of chips at the beginning of the step,
and received $M$ chips. Hence on $G$ we can play in the desired way for $\frac{3}{8}M$ steps.
For the initial distributions, we have
$y(\psi(v)) = d^+(v) \cdot M + x(v)$ for each $v\in V(D)$,
so a vertex $v \in V(D)$ is active with respect to $x$
if and only if $\psi(v)$ is active with respect to $y$.
Let $x'$ be the distribution on $D$ and $y'$ the distribution
on $G$ at the end of an arbitrary (but at most $\frac{3}{8}M^\text{th}$) step.
Then it can be shown by induction
that $y'(\psi(v)) = d^+(v) \cdot M + x'(v)$
for each $v \in V(D)$. So in each step we have that
a vertex $v \in V(D)$ is active if and only if $\psi(v)$ is active.
Hence for $\frac{3}{8}M$ steps, the corresponding game on $D$ is also legal.
This means that there is a chip-firing game of length at least
$\frac{3}{8}M \ge 2|V(D)|^2 |E(D)| \Delta(D) + 1$ on $D$ with initial distribution $x$, which by
\cite[Corollary 4.9]{BL92} implies that the distribution $x$ is non-terminating.
This finishes the proof.
\end{proof}
For a general digraph, the construction of the proof imitates the following game: If a vertex $v$ fires, each of its out-neighbors $u$ receives $\overrightarrow{d}(vu)$ chips, but the number of chips on $v$ decreases by the in-degree of $v$. This modification of the chip-firing game has been studied by Asadi and Backman \cite{asadi}.
\begin{proof}[Proof of Theorem \ref{thm::chip_rang_NP-teljes}]
The theorem follows from Theorem \ref{thm::rang_eulerben_nehez}
and the previous lemma.
\end{proof}
\begin{remark} \label{rem::dist_NP_erosites}
For a simple Eulerian digraph $D$, one has $$|E(\varphi(D))|\leq |E(D)|\cdot 9|V(D)|^3|E(D)|\leq 9|V(\varphi(D))|^5,$$ therefore the computation of ${\rm dist}$ is \cl{NP}-hard even for graphs with $|E(G)|\leq 9|V(G)|^5$.
\end{remark}
\section{Polynomial time computability in a special case}
In this section we consider undirected graphs, and observe that for
chip-distributions that are in a sense ``small'', computing the
distance from non-terminating distributions
can be done in polynomial time. Moreover, for these distributions,
the distance from non-terminating distributions only depends on the number of edges of the graph and the
number of chips in the distribution.
The corollaries of this observation for the case of divisors
give a special case of the Riemann-Roch theorem.
Recall that Theorem \ref{thm::also_korlat} stated that
${\rm dist}(\mathbf{0}_G)=|E(G)|$ for any
undirected graph $G$. We would like to generalize this statement for
``small enough'' distributions.
We say that a distribution $x \in {\rm Chip}(G)$ is \emph{under an
acyclic orientation}, if there exists an acyclic orientation $D$ of $G$ such that
$x(v) \le d^-_D(v)$ for each $v \in V(G)$.
\begin{prop} \label{prop::acikl_alatt}
Let $G$ be a graph and let $x \in {\rm Chip}(G)$ be a
distribution. If $x$ is
under an acyclic orientation then ${\rm dist}(x) = |E(G)| - |x|$.
\end{prop}
\begin{proof}
From Theorem \ref{thm::also_korlat}, a
non-terminating distribution has at least $|E(G)|$ chips, therefore
${\rm dist}(x) \ge |E(G)| - |x|$.
For the other direction, let $D$ be an acyclic orientation of $G$
with $x(v) \le d^-_D(v)$ for each $v \in V(G)$. Let $y$ be the
distribution on $G$ corresponding to the indegrees of the
orientation, i.e., $y(v) = d^-_D(v)$ for each $v \in V(G)$.
Then, using Lemma \ref{lem::chip_dist_over_acyclic}, $y$ is
non-terminating, moreover $|y| = |E(G)|$ and $y(v) \ge x(v)$ for
each $v \in V(G)$.
Hence ${\rm dist}(x) \le |y - x| = |E(G)| - |x|$.
This completes the proof of the proposition.
\end{proof}
\begin{remark}
It can also be decided in polynomial time whether a distribution
$x \in {\rm Chip}(G)$ is under an acyclic orientation.
A greedy algorithm solves the problem.
\end{remark}
From the previous proposition, using the duality between
chip-distributi\-ons and divisors, we get a special case of the
Riemann-Roch theorem for graphs.
Let us denote by $K$ the canonical divisor on a graph $G$, that is,
$K(v)=d(v)-2$ for each vertex $v\in V(G)$.
\begin{thm}[Riemann-Roch for graphs, \cite{BN-Riem-Roch}]
Let $G$ be a graph, and let $f$ be a divisor on $G$. Then
$${\rm rank}(f)-{\rm rank}(K-f)={\rm deg}(f)-|E(G)|+|V(G)|.$$
\end{thm}
Now, from Proposition \ref{prop::rank_of_dual} and Proposition
\ref{prop::acikl_alatt} we have for $f=K^+ - x$ that
$${\rm rank}(f) = {\rm dist}(x) - 1 = |E(G)| - |x| - 1 =
{\rm deg}(f)- |E(G)| + |V(G)| -1,$$
if $x$ is under an acyclic orientation.
We claim that in this case, ${\rm rank}(K-f)=-1$.
Indeed, $K-f=K-(K^+ - x)= x - \mathbf{1}$, so the dual of $K-f$ is
$K^+ - x + \mathbf{1}$. The distribution $x$ is under an acyclic orientation, let $D$ be an orientation witnessing this, i.e.,
$x(v) \le d^-_D(v)$ for each vertex $v \in V(G)$.
Then $(K^+ - x + \mathbf{1})(v) = d(v) - x(v) \ge d^+_D(v)$
for each vertex $v \in V(G)$, hence we can use Lemma \ref{lem::chip_dist_over_acyclic} for $K^+ - x + \mathbf{1}$ and
the directed graph obtained from $D$ by reversing every edge.
It follows that $K^+ - x + \mathbf{1}$ is non-terminating,
hence for its dual, ${\rm rank}(K-f)=-1$ by Proposition \ref{prop::rank_of_dual}.
Therefore, we have ${\rm rank}(f) - {\rm rank}(K-f)= {\rm deg}(f)- |E(G)| + |V(G)|$,
showing the Riemann-Roch theorem in this special case.
\section*{Acknowledgement}
Research was supported by the MTA-ELTE Egerv\'ary Research Group and by the
Hungarian Scientific Research Fund - OTKA, K109240 (Lilla T\'oth\-m\'e\-r\'esz), 104178 and 105645 (Viktor Kiss).
We would like to thank B\'alint Hujter for introducing us to this topic, Erika B\'erczi-Kov\'acs and Krist\'of B\'erczi for suggesting us to try to reduce the feedback arc set problem to the computation of the rank. We would also like to thank Tam\'as Kir\'aly, Zolt\'an Kir\'aly, M\'arton Elekes and the anonymous referees for their useful comments about the manuscript.
\bibliographystyle{abbrv}
|
\section{Introduction}
G\"odel's \cite{Goedel(58)} so-called dialectica interpretation reduces the consistency of Peano arithmetic to the consistency of the quantifier-free calculus of functionals $T$. In order to extend G\"odel's interpretation to full classical analysis ${\rm PA}^\omega + {\rm CA}$, Spector \cite{Spector(62)} made use of the fact that ${\rm PA}^\omega + {\rm CA}$ can be embedded, via the negative translation, into ${\rm HA}^\omega + {\rm AC}_\mathbb{N} + {\rm DNS}$. Here ${\rm PA}^\omega$ and ${\rm HA}^\omega$ denote Peano and Heyting arithmetic, respectively, formulated in the language of finite types, and
\eqleft{{\rm CA} \; \colon \; \exists f^{\mathbb{N} \to \mathbb{B}} \forall n^\mathbb{N} (f(n) \leftrightarrow A(n))}
is \emph{full comprehension},
\eqleft{{\rm AC}_\mathbb{N} \; \colon \; \forall n^\mathbb{N} \exists x^X A(n, x) \rightarrow \exists f \forall n A(n, f n)}
is \emph{countable choice}, and
\eqleft{{\rm DNS} \; \colon \; \forall n^\mathbb{N} \neg \neg B(n) \rightarrow \neg \neg \forall n B(n),}
is the \emph{double negation shift}, with $A(n)$ and $A(n, x)$ standing for arbitrary formulas, and $B(n) \equiv \exists x \neg A(n, x)$. Since ${\rm HA}^\omega + {\rm AC}_\mathbb{N}$, excluding the double negation shift, has a straightforward (modified) realizability interpretation \cite{Troelstra(73)}, as well as a dialectica interpretation \cite{Avigad(98),Goedel(58)}, the remaining challenge is to give a computational interpretation to ${\rm DNS}$.
A computational interpretation of ${\rm DNS}$ was first given by Spector \cite{Spector(62)}, via the dialectica interpretation. Spector devised a form of recursion on well-founded trees, nowadays known as \emph{Spector bar recursion}, and showed that the dialectica interpretation of ${\rm DNS}$ can be witnessed by such kind of recursion. A computational interpretation of ${\rm DNS}$ via realizability only came recently, first in \cite{Berardi(98)}, via a non-standard form of realizability, and then in \cite{BO(02A),BO(05)}, via Kreisel's modified realizability. The realizability interpretation of ${\rm DNS}$ makes use of a new form of bar recursion, termed \emph{modified bar recursion}.
It has been shown in \cite{BO(05)} that Spector's bar recursion is definable in system $T$ extended with modified bar recursion, but not conversely, since Spector's bar recursion is S1-S9 computable in the model of total continuous functionals, but modified bar recursion is not.
In the present paper we revisit these functional interpretations of classical analysis from the perspective of the newly developed theory of selection functions \cite{Escardo(2008),EO(2010A),EO(2010B),EO(2009),EO(2011A)}. \emph{Selection functionals} are functionals of type $(X \to R) \to X$, for arbitrary finite types $X, R$. We think of mappings $p \colon X \to R$ as generalised predicates, and of functionals $\varepsilon \colon (X \to R) \to X$ as witnessing, when possible, the ``non-emptiness" of any given such predicate. For instance, if $R = \mathbb{B}$ is the set of booleans, Hilbert's $\varepsilon$-constant can be viewed as a selection function. Just as $\varepsilon$-terms in Hilbert's calculus can be used to define the existential quantifier, so can any selection function $\varepsilon \colon (X \to R) \to X$ be used to define a \emph{generalised quantifier} $\phi \colon (X \to R) \to R$ as
\[ \phi(p) \stackrel{R}{=} p(\varepsilon(p)). \]
Moreover, just like the usual quantifiers $\exists^X$ and $\forall^Y$ can be nested to produce a quantifier on the product space $X \times Y$, so can generalised quantifiers and selection functions. We prefer to think about the nesting of selection functions (and quantifiers) as a \emph{product operation}, since it transforms selection functions over spaces $X$ and $Y$ into a new selection function on the product space $X \times Y$ (cf. \cite{EO(2009)}).
In this article we define two different iterations of the binary product of selection functions, one which we call \emph{implicitly controlled} and the other which we call \emph{explicitly controlled} (cf. also \cite{EO(2011A)}) and show that:
\begin{itemize}
\item Modified bar recursion is $T$-equivalent to the implicitly controlled product of selection functions.
\item Spector's bar recursion is $T$-equivalent to the explicitly controlled product of selection functions.
\item The two different products can be used to interpret ${\rm DNS}$ directly via modified realizability and the dialectica interpretation, respectively.
\item The implicitly controlled product of selection functions is strictly stronger than the explicitly controlled one.
\item Apparently stronger iterations of the dependent products are in fact $T$-equivalent to the iterations of the simple products.
\end{itemize}
\section{Preliminaries}
Before we present our main results, let us first define the formal systems used, and give an introduction to our recent work on selection functions.
\subsection{Heyting arithmetic ${\rm HA}^\omega$ and system $T$}
\label{bar-ind}
In this section we define the formal systems used to prove the inter-definability results. These include Heyting arithmetic in all finite types and extensions including bar induction and a continuity principle.
\begin{definition}[Finite types] The set of all \emph{finite types} $\mathcal{T}$ are defined inductively as
\begin{itemize}
\item $\mathbb{B}$ (booleans) and $\mathbb{N}$ (integers) are in $\mathcal{T}$ \\[-3mm]
\item If $X$ and $Y$ are in $\mathcal{T}$ then $X \times Y$ (product) and $X \to Y$ (functions) are in $\mathcal{T}$ \\[-3mm]
\item If $X$ is in $\mathcal{T}$ then $X^*$ (finite sequence) is in $\mathcal{T}$.
\end{itemize}
We will also make informal use of the following type construction: Given a sequence of types $(X_i)_{i \in \mathbb{N}}$ we also consider $\Pi_{i \in \mathbb{N}} X_i$ and $\Pi_{i < n} X_i$ as types. The main purpose of this is to make the constructions more readable, since we can keep track of the positions which are being changed. A formal extension of system $T$ with such type construction has been considered by Tait \cite{tait(1965)}, hence we also hope that our presentation below will extend smoothly to a more general setting\footnote{If the reader prefers, however, she can assume that in a sequence of types $(X_i)_{i \leq \mathbb{N}}$ all $X_i$ are equal $X$, replacing infinite sequence types $\Pi_{i \in \mathbb{N}} X_i$ with $\mathbb{N} \to X$, and finite sequence types $\Pi_{i < n} X_i$ with $X^*$.}, although in this paper we focus on the standard version of system $T$.
\end{definition}
We use $X, Y, Z$ for variables ranging over the elements of $\mathcal{T}$. We often write $\Pi_i X_i$ for $\Pi_{i \in \mathbb{N}} X_i$, and also $\Pi_{i \geq k} X_i$ for $\Pi_i X_{i+k}$.
Let ${\rm HA}^\omega$ be usual Heyting arithmetic in all finite types with a fully extensional treatment of equality, as in the system E-${\rm HA}^\omega$ of \cite{Troelstra(73)}. Its quantifier-free fragment is the usual G\"odel's system $T$, also extended with sequence types. G\"odel's primitive recursion for each sequence of types $(X_i)_{i \in \mathbb{N}} \in \mathcal{T}$ is given by
\[
\begin{array}{lcl}
{\sf R} f g 0 & \stackrel{X_0}{=} & g \\[2mm]
{\sf R} f g (n+1) & \stackrel{X_{n+1}}{=} & f n ({\sf R} f g n)
\end{array}
\]
where ${\sf R}$ has finite type $\Pi_n (X_n \to X_{n+1}) \to X_0 \to \Pi_i X_i$. We also assume that we have a constant ${\bf 0}^X$ of each finite type $X$, and the usual constructors and destructors such as $\pair{t^X, s^Y} \colon X \times Y$ and $\pi_i(\pair{s_0^{X_0}, s_1^{X_1}}) = s_i$, where $i = \{0,1\}$, for instance. For the newly introduced sequence types we have that if $t \colon \Pi_i X_i$ then $t i \colon X_i$; and if $t \colon X_i$ then $\lambda i . t \colon \Pi_i X_i$. If $s \colon \Pi_{i < n} X_i$, we write $s_i \colon X_i$ for the $i$-th element of the sequence, for $i < n$. If $s \colon \Pi_{i < n} (X_i \times Y_i)$ is a sequence of pairs, we write $s^0 \colon \Pi_{i < n} X_i$ and $s^1 \colon \Pi_{i < n} Y_i$ for the projection of the sequence on the first and second coordinates, respectively. If $\alpha$ has type $\Pi_{i \in \mathbb{N}} X_i$ we use the following abbreviations
\[
\begin{array}{lcl}
\alpha^n & \equiv & \lambda i . \alpha(i + n), \quad \mbox{(the $n$-left shift of $\alpha$, hence $\alpha^n \colon \Pi_i X_{i + n}$)} \\[2mm]
q^n(\alpha) & \equiv & q(\alpha^n), \quad \mbox{(so $q^n \colon \Pi_i X_i \to R$ if $q \colon \Pi_i X_{i + n} \to R$)} \\[2mm]
\alpha[k,n] & \equiv & \pair{\alpha(k), \ldots, \alpha(n)}, \quad \mbox{(finite segment from position $k$ to $n$)} \\[2mm]
\initSeg{\alpha}{n} & \equiv & \alpha[0, n-1], \quad \mbox{(initial segment of $\alpha$ of length $n$)} \\[2mm]
\initSegZ{\alpha}{n} & \equiv & \pair{\alpha(0), \ldots, \alpha(n-1), {\bf 0}, {\bf 0}, \ldots}, \; \mbox{(infinite extension of $\initSeg{\alpha}{n}$ with ${\bf 0}$'s)}
\end{array}
\]
where in the last case the type of ${\bf 0}$ at the $i$-th coordinate is the same type of $\alpha(i)$.
We use $*$ for all forms of \emph{concatenation}. For instance, if $x$ has type $X_n$ and $s$ has type $\Pi_{i<n} X_i$ then $s * x$ is the concatenation of~$s$ with~$x$, which has type $\Pi_{i<n+1} X_i$. Similarly, if $x$ has type $X_0$ and $\alpha$ has type $\Pi_i X_{i+1}$ then $x * \alpha$ has type $\Pi_i X_i$. Given a functional $q \colon \Pi_i X_i \to R$ and a finite sequence $s \colon \Pi_{i < n} X_i$ we write $q_s \colon \Pi_{i \geq n} X_i \to R$ for the function $\lambda \alpha . q(s * \alpha)$. When $s = \langle x \rangle$ we write $q_s$ as simply $q_x$.
Given a finite sequence $s$ and an infinite sequence $\alpha$ let us write $s \,@\, \alpha$ for the ``overwriting" of $s$ on $\alpha$, i.e. $(s \,@\, \alpha)(i)$ equals $s_i$ if $i < |s|$ and equals $\alpha(i + |s|)$ otherwise.
In the following we shall assume that certain types are
\emph{discrete}. Semantically, in the model of total continuous
functionals, discreteness means that singletons are open or that all
points are isolated. Syntactically, the following grammar produces discrete types in that
model (along with compact types) \cite{Escardo(2008)}.
\begin{definition}[Discrete and compact types] Define the two subsets of $\mathcal{T}$ inductively as follows:
\[
\begin{array}{lcl}
{\sf compact} & ::= & \mathbb{B} \;|\; {\sf compact} \times {\sf compact} \;|\; {\sf discrete} \to {\sf compact} \\[2mm]
{\sf discrete} & ::= & \mathbb{B} \;|\; \mathbb{N} \;|\; {\sf discrete} \times {\sf discrete} \;|\; {\sf discrete}^* \;|\; {\sf compact} \to {\sf discrete}.
\end{array}
\]
\end{definition}
In this paper we work with a model independent notion of definability.
Formally, given a term $t$ in system~$T$, we view an equation $F(x) = t(F, x)$ as \emph{defining} or \emph{specifying} a functional~$F$.
We do not worry whether such an equation has a solution in any particular model of ${\rm HA}^\omega$, or whether it is unique, when it has a solution.
\begin{definition} We say that a functional $G$ is $T$-definable from a functional $F$ (written $G \leq_T F$) over a theory $\mathcal{S}$ if there exists a term $s$ in system $T$ such that $s(F)$ satisfies the defining equation of $G$ provably in $\mathcal{S}$. We say that $F$ and $G$ are $T$-equivalent over $\mathcal{S}$, written $F =_T G$, if $G \geq_T F$ and $F \geq_T G$.
\end{definition}
When stating in a theorem or proposition that $G$ is $T$-definable in $F$, we will explicitly write after the theorem/proposition number the theory $\mathcal{S}$ that is needed for the verification. In a few cases this theory will be an extension of ${\rm HA}^\omega$ with some the following three principles: \emph{Spector's condition}
\eqleft{{\rm SPEC} \; \colon \; \forall \omega^{\Pi_i X_i \to \mathbb{N}} \forall \alpha^{\Pi_i X_i} \exists n (\omega(\initSegZ{\alpha}{n}) < n),}
the \emph{axiom of continuity}
\eqleft{{\rm CONT} \; \colon \; \forall q^{\Pi_i X_i \to R} \forall \alpha \exists n \forall \beta (\initSeg{\alpha}{n} \stackrel{\Pi_{i < n} X_i}{=} \initSeg{\beta}{n} \to q(\alpha) \stackrel{R}{=} q(\beta))}
with $R$ discrete, and the scheme of \emph{relativised bar induction}
\eqleft{{\rm BI} \; \colon \;
\left\{
\begin{array}{c}
S(\langle\,\rangle) \\
\wedge \\
\forall \alpha \!\in\! S \, \exists n P(\initSeg{\alpha}{n}) \\
\wedge \\
\forall s \in S (\forall x [S(s * x)\to P(s * x)] \to P(s))
\end{array}
\right\} \to P(\langle\,\rangle),
}
where $S(s)$ and $P(s)$ are arbitrary predicates in the language of ${\rm HA}^\omega$, and $\alpha\in S$ and $s\in S$ are shorthands for $\forall n S(\initSeg{\alpha}{n})$ and $S(s)$ respectively.
We note that ${\rm SPEC}$ follows from ${\rm CONT}$, but it also holds in the model of strongly majorizable functionals \cite{Bezem(85)}.
\newcommand{\mathcal{A}}{\mathcal{A}}
\subsection{Selection functions and generalised quantifiers}
\label{binary:product}
In \cite{EO(2009),EO(2011A)} we have studied the properties of functionals having the type $(X \to R) \to R$, and called these \emph{generalised quantifiers}. When $R = \mathbb{B}$ we have that $(X \to \mathbb{B}) \to \mathbb{B}$ is the type of the usual logical quantifiers $\forall, \exists$. We also showed that some generalised quantifiers $\phi \colon (X \to R) \to R$ are \emph{attainable}, in the sense that for some \emph{selection function} $\varepsilon \colon (X \to R) \to X$, we have
\[ \phi p = p(\varepsilon p) \]
for all (generalised) predicates $p$. In the case when $\phi$ is the usual existential quantifier, for instance, $\varepsilon$ corresponds to Hilbert's epsilon term.
Since the types $(X \to R) \to R$ and $(X \to R) \to X$ will be used quite often, we abbreviate them as $K_R X$ and $J_R X$, respectively. Moreover, when $R$ is fixed, we often simply write $K X$ and $J X$, omitting the subscript $R$. In \cite{EO(2009)} we also defined products of quantifiers and selection functions.
\begin{definition}[Product of selection functions and quantifiers] \label{main-simple} Given generalised quantifiers $\phi \colon K X$ and $\psi \colon K Y$, define the product quantifier $(\phi \ttimesTensor \psi) \colon K (X \times Y)$ as
\[ (\phi \ttimesTensor \psi)(p^{X \times Y \to R}) \stackrel{R}{=} \phi(\lambda x^X . \psi (\lambda y^Y . p(x, y))). \]
Also, given selection functions $\varepsilon \colon J X$ and $\delta \colon J Y$, define the product selection function $(\varepsilon \ttimesTensor \delta) \colon J(X \times Y)$ as
\[ (\varepsilon \ttimesTensor \delta)(p^{X \times Y \to R}) \stackrel{X \times Y}{=} (a, b(a)) \]
where
\eqleft{
\begin{array}{lcl}
a & \stackrel{X}{=} & \varepsilon(\lambda x^X. p(x, b(x))) \\[2mm]
b(x^X) & \stackrel{Y}{=} & \delta(\lambda y^Y . p(x, y)).
\end{array}
}
\end{definition}
One of the results we obtained is that the product of attainable quantifiers is also attainable. This follows from the fact that the product of quantifiers corresponds to the product of selection functions, as made precise in the following lemma.
\begin{lemma}[\cite{EO(2009)}, Lemma 3.1.2] \label{basic} Let $R$ be fixed. Given a selection function $\varepsilon : J X$, define a quantifier $\selEmb{\varepsilon} \colon K X$ as
\eqleft{\selEmb{\varepsilon} p = p(\varepsilon p).}
Then for $\varepsilon \colon J X$ and $\delta \colon J Y$ we have $\selEmb{\varepsilon \ttimesTensor \delta} = \selEmb \varepsilon \ttimesTensor \selEmb{\delta}.$
\end{lemma}
Given a finite sequence of selection functions or quantifiers, the two binary products defined above can be iterated so as to give rise to finite products of selection functions and quantifiers. We have shown that such a construction also appears in game theory (backward induction), algorithms (backtracking), and proof theory (interpretation of the infinite pigeon-hole principle) -- see \cite{EO(2009)} for details.
In the following (Sections \ref{conditional} and \ref{implicit}) we
will describe two possible ways of iterating the binary product of
selection function an infinite, or unbounded, number of times.
\section{Explicitly Controlled Product}
\label{conditional}
The finite product of selection functions of Definition \ref{main-simple} can be infinitely iterated in two ways. The first, which we define in this section is via an \emph{explicitly controlled} iteration, which we will show to correspond to Spector's bar recursion. In the following section we also define an \emph{implicitly controlled} iteration, which we will show to correspond to modified bar recursion.
\begin{definition}[${\sf eps}$] \label{eps-def} Let $\varepsilon \colon \Pi_k J X_k$ be a sequence of selection functions. Define their \emph{explicitly controlled infinite product} as
\[
{\sf eps}_n^l(\varepsilon)(q) \stackrel{\Pi_i X_{i + n}}{=}
\left\{
\begin{array}{ll}
{\bf 0} & {\rm if} \; l(q({\bf 0})) < n \\[2mm]
(\varepsilon_n \ttimesTensor {\sf eps}_{n+1}^l(\varepsilon))(q) \quad & {\rm otherwise},
\end{array}
\right.
\]
where $q \colon \Pi_i X_{i+n} \to R$ and $l \colon R \to \mathbb{N}$. We call $l$ the \emph{length function} since it controls the length of the recursive path. Unfolding the definition of $\ttimesTensor$ we can write the defining equation of ${\sf eps}$ as
\begin{equation} \label{eps-eq-def} \tag{{\sf eps}}
{\sf eps}_n^l(\varepsilon)(q) \stackrel{\Pi_i X_{i + n}}{=}
\left\{
\begin{array}{ll}
{\bf 0} & {\rm if} \; l(q({\bf 0})) < n \\[2mm]
c * {\sf eps}_{n+1}^l(\varepsilon)(q_c) \quad & {\rm otherwise},
\end{array}
\right.
\end{equation}
where $c = \varepsilon_n(\lambda x . \selEmb{{\sf eps}_{n+1}^l(\varepsilon)}(q_x))$.
\end{definition}
The next lemma (essentially Lemma 1 of \cite{Spector(62)}) states one of the most crucial properties of this product of selection functions.
\begin{lemma}[${\rm HA}^\omega + (\ref{eps-eq-def})$] \label{spector-main-lemma} Let $\alpha = {\sf eps}_n^{l}(\varepsilon)(q)$. Then, for all $i \colon \mathbb{N}$
\[ \alpha = \initSeg{\alpha}{i} * {\sf eps}_{n + i}^l(\varepsilon)(q_{\initSeg{\alpha}{i}}). \]
\end{lemma}
\begin{proof} By induction on $i$. If $i = 0$ this follows by the definition of $\alpha$. Assume this holds for $i$, we wish to show it also holds for $i+1$. Consider two cases. \\[2mm]
If $l(q_{\initSeg{\alpha}{i}}({\bf 0})) = l(q(\initSegZ{\alpha}{i})) < n + i$ then
\begin{itemize}
\item[$(i)$] ${\sf eps}_{n + i}^l(\varepsilon)(q_{\initSeg{\alpha}{i}}) = {\bf 0}$
\end{itemize}
and hence
\begin{itemize}
\item[$(ii)$] $\alpha \stackrel{\textup{(IH)}}{=} \initSeg{\alpha}{i} * {\sf eps}_{n+i}^l(\varepsilon)(q_{\initSeg{\alpha}{i}}) \stackrel{(i)}{=} \initSegZ{\alpha}{i} = \initSegZ{\alpha}{i+1}$.
\end{itemize}
Therefore,
\begin{itemize}
\item[$(iii)$] $l(q(\initSegZ{\alpha}{i+1})) \stackrel{(ii)}{=} l(q(\initSegZ{\alpha}{i})) < n + i < n + i + 1$.
\end{itemize}
Hence, by $(iii)$ we have
\begin{itemize}
\item[$(iv)$] ${\sf eps}_{n + i + 1}^l(\varepsilon)(q_{\initSeg{\alpha}{i+1}}) = {\bf 0}$.
\end{itemize}
So
\eqleft{\alpha \stackrel{(ii)}{=} \initSegZ{\alpha}{i+1} \stackrel{(iv)}{=} \initSeg{\alpha}{i+1} * {\sf eps}_{n + i + 1}^l(\varepsilon)(q_{\initSeg{\alpha}{i+1}}).}
On the other hand, if $l(q_{\initSeg{\alpha}{i}}({\bf 0})) = l(q(\initSegZ{\alpha}{i})) \geq n + i$, then
\eqleft{\alpha \stackrel{\textup{(IH)}}{=} \initSeg{\alpha}{i} * {\sf eps}_{n+i}^l(\varepsilon)(q_{\initSeg{\alpha}{i}}) = \initSeg{\alpha}{i} * c * {\sf eps}_{n+i+1}^l(\varepsilon)(q_{\initSeg{\alpha}{i} * c}),}
so that $\alpha(i) = c$. Hence $\alpha = \initSeg{\alpha}{i+1} * {\sf eps}_{n+i+1}^l(\varepsilon)(q_{\initSeg{\alpha}{i + 1}})$.
\end{proof}
An immediate consequence of the lemma above is that it allows us to calculate the $i$-th element of the infinite sequence ${\sf eps}_n^l(\varepsilon)(q)$ (see also Theorem \ref{main-spec} for another important consequence).
\begin{corollary}[${\rm HA}^\omega + (\ref{eps-eq-def})$] \label{unwinding-cps} For all $n$ and $i$
\[ {\sf eps}_n^l(\varepsilon)(q)(i) \stackrel{X_{n + i}}{=}
\left\{
\begin{array}{ll}
{\bf 0} & {\rm if} \; l(q_t({\bf 0})) < n + i \\[2mm]
\varepsilon_{n + i}(\lambda x . \selEmb{{\sf eps}_{n + i + 1}^l(\varepsilon)}(q_{t*x})) \;\; & {\rm otherwise},
\end{array}
\right.
\]
where $t = \initSeg{{\sf eps}_n^l(\varepsilon)(q)}{i}$.
\end{corollary}
\begin{proof} Let $\alpha = {\sf eps}_n^l(\varepsilon)(q)$ so that $t = \initSeg{{\sf eps}_n^l(\varepsilon)(q)}{i} = \initSeg{\alpha}{i}$. By Lemma \ref{spector-main-lemma} we have that $\alpha(i) = {\sf eps}_{n+i}^l(\varepsilon)(q_t)(0)$. Hence, by ($\ref{eps-eq-def}$) we have the desired result. \end{proof}
The fact that ${\sf eps}$ exists in the model of total continuous functionals, and is in fact uniquely characterized by its defining equation, can be seen as follows. First, note that the ${\sf eps}_n^l(\varepsilon)(q)$ is an infinite sequence, say $\alpha \colon \Pi_i X_{i + n}$. Intuitively, at each recursive call the functional $q$ gets information about one more element of its input sequence. Assuming continuity we will have that $l \circ q \colon \Pi_i X_{i + n} \to \mathbb{N}$ will eventually always return a fixed value, no matter what the rest of the input sequence is. This means that as $n$ increases we will eventually have $l(q({\bf 0})) < n$. It is perhaps surprising that such a functional also exists in the model of strongly majorizable functionals~\cite{Bezem(85)}, which contains discontinuous functionals! Following the construction of Bezem \cite{Bezem(85)} one can prove this directly, but this result will also follow from our result that ${\sf eps}$ is $T$-definable from Spector's bar recursion (Section \ref{sec-spector}).
We also define the corresponding explicitly controlled product of \emph{quantifiers} as follows:
\begin{definition}[${\sf epq}$] \label{def-epq} Let $\phi \colon \Pi_k K X_k$ be a sequence of quantifiers. Their \emph{explicitly controlled infinite product} is defined as
\[ {\sf epq}_n^l(\phi)(q) \stackrel{R}{=}
\left\{
\begin{array}{ll}
q({\bf 0}) & {\rm if} \; l(q({\bf 0})) < n \\[2mm]
(\phi_n \ttimesTensor {\sf epq}_{n+1}^l(\phi))(q) & {\rm otherwise},
\end{array}
\right.
\]
where $q \colon \Pi_i X_{i + n} \to R$ and $l \colon R \to \mathbb{N}$. Unfolding the definition of the binary product of quantifiers we have
\begin{equation} \label{def-epq-equation} \tag{{\sf epq}}
{\sf epq}_n^l(\phi)(q) \stackrel{R}{=}
\left\{
\begin{array}{ll}
q({\bf 0}) & {\rm if} \; l(q({\bf 0})) < n \\[2mm]
\phi_n(\lambda x^{X_n} . {\sf epq}_{n+1}^l(\phi)(q_x)) & {\rm otherwise}.
\end{array}
\right.
\end{equation}
\end{definition}
Howard (proof attributed to Kreisel) shows in Lemma 3C of \cite{Howard(1968)} that assuming Spector's bar recursion one can prove Spector's stopping condition ${\rm SPEC}$. It is easy to see that the form of bar recursion used by Howard is also an instance of ${\sf epq}$ and hence we obtain:
\begin{lemma} \label{howard-kreisel} ${\rm HA}^\omega + (\ref{def-epq-equation}) \vdash {\rm SPEC}$.
\end{lemma}
We now show that ${\sf epq}$ and ${\sf eps}$ are $T$-equivalent. That ${\sf epq}$ is $T$-definable in ${\sf eps}$ has been recently shown in \cite{Oliva(2012A)}. Hence, it remains to show that ${\sf epq}$ defines ${\sf eps}$. The proof makes use of the fact that each selection function $\varepsilon$ defines a quantifier, as $\phi(p) = p(\varepsilon(p))$ (cf. Lemma \ref{basic}). In order to define ${\sf eps}$ for the types $(X_i, R)$ we shall use ${\sf epq}$ for the types $(X_i, R')$ with $R' = \Pi_i X_i$.
\begin{lemma}[${\rm HA}^\omega + {\rm BI} + {\rm SPEC}$] \label{epq-eps-lemma} Let $R' = \Pi_i X_i$. Given $\varepsilon_i \colon J_R X_i$ and $q \colon \Pi_i X_i \to R$ define
\begin{equation} \label{epq-eps-lemma-eq}
\phi_i^{\varepsilon, q}(p^{X_i \to R'}) \stackrel{R'}{=} p(\varepsilon_i(\lambda x^{X_i} . q(p(x)))).
\end{equation}
Defined also the sequence of functions $f^n = \lambda \alpha . \cZr{\Pi_{i < n} X_i} * \alpha$. Then (with $q \colon \Pi_{i \geq n} X_i \to R$ so $q^n \colon \Pi_i X_i \to R$)
\[ {\sf epq}_{n+1}^{l \circ q^n}(\phi^{\varepsilon, q^n})((f^n)_x) = ({\bf 0} * x^{X_n}) \,@\, {\sf epq}_{n + 1}^{l \circ (q_x)^{n+1}}(\phi^{\varepsilon, (q_x)^{n+1}})(f^{n+1}). \]
\end{lemma}
\begin{proof} By bar induction ${\rm BI}$ and the axiom ${\rm SPEC}$. We take $S(s) = {\sf true}$ and
\[ \underbrace{{\sf epq}_{n+|s|+1}^{l \circ q^n}(\phi^{\varepsilon, q^n})((f^n)_{x * s}) = ({\bf 0} * x) \,@\, {\sf epq}_{n+|s|+1}^{l \circ (q_x)^{n+1}}(\phi^{\varepsilon, (q_x)^{n+1}})((f^{n+1})_s)}_{P(s)} \]
where $s \colon \Pi_{n < i \leq n + |s|} X_i$. \\[1mm]
($i$) $\forall \alpha \exists j \, P(\initSeg{\alpha}{j})$. By ${\rm SPEC}$, for any $\alpha \colon \Pi_{i > n} X_i$ there is a point $j$ such that
\[
\begin{array}{lcl}
(l \circ q^n)(\cZr{\Pi_{i < n} X_i} * x * \initSeg{\alpha}{j} * {\bf 0}) & = & l(q(x * \initSeg{\alpha}{j} * {\bf 0})) \\[1mm]
& < & n + j + 1.
\end{array}
\]
For such $j$ and $s = \initSeg{\alpha}{j}$ it is easy to see that $P(s)$ holds as both sides of $P(s)$ are equal to ${\bf 0} * x * s * {\bf 0}$. \\[1mm]
($ii$) $\forall s (\forall y P(s * y) \to P(s))$. Let $s$ be such that $\forall y P(s * y)$; we show $P(s)$. We can assume that
\[ (l \circ q^n \circ (f^n)_{x * s})({\bf 0}) = (l \circ (q_x)^{n+1} \circ (f^{n+1})_s)({\bf 0}) \geq n + |s| + 1, \]
as otherwise the proof can be carried out as in case $(i)$ above. Hence, we calculate
\[
\begin{array}{l}
{\sf epq}_{n+|s|+1}^{l \circ q^n}(\phi^{\varepsilon, q^n})((f^n)_{x * s}) \\[1.5mm]
\quad \;\;\; \stackrel{(\ref{def-epq-equation})}{=} \phi^{\varepsilon, q^n}_{n+|s|+1}(\lambda y . {\sf epq}_{n+|s|+2}^{l \circ q^n}(\phi^{\varepsilon, q^n})((f^n)_{x * s * y}) \\[1.5mm]
\quad \quad \,\stackrel{(\ref{epq-eps-lemma-eq})}{=} {\sf epq}_{n+|s|+2}^{l \circ q^n}(\phi^{\varepsilon, q^n})((f^n)_{x * s * c}) \\[1.5mm]
\quad \quad \,\stackrel{\textup{(IH)}}{=} ({\bf 0} * x) \,@\, {\sf epq}_{n+|s|+2}^{l \circ (q_x)^{n+1}}(\phi^{\varepsilon, (q_x)^{n+1}})((f^{n+1})_{s * c}) \\[1.5mm]
\quad \quad\; \stackrel{(*)}{=} ({\bf 0} * x) \,@\, {\sf epq}_{n+|s|+2}^{l \circ (q_x)^{n+1}}(\phi^{\varepsilon, (q_x)^{n+1}})((f^{n+1})_{s * \tilde{c}}) \\[1.5mm]
\quad \quad\; \stackrel{(\ref{epq-eps-lemma-eq})}{=} ({\bf 0} * x) \,@\, \phi_{n+|s|+1}^{\varepsilon, (q_x)^{n+1}}(\lambda y . {\sf epq}_{n+|s|+2}^{l \circ (q_x)^{n+1}}(\phi^{\varepsilon, (q_x)^{n+1}})((f^{n+1})_{s * y})) \\[1.5mm]
\quad \quad\, \stackrel{(\ref{def-epq-equation})}{=} ({\bf 0} * x) \,@\, {\sf epq}_{n+|s|+1}^{l \circ (q_x)^{n+1}}(\phi^{\varepsilon, (q_x)^{n+1}})((f^{n+1})_s)
\end{array}
\]
where
\begin{itemize}
\item[] $c = \varepsilon_{n+|s|+1}(\lambda y . q^n({\sf epq}_{n+|s|+2}^{l \circ q^n}(\phi^{\varepsilon, q^n})((f^n)_{x * s * y})))$
\item[] $\tilde{c} = \varepsilon_{n+|s|+1}(\lambda y . (q_x)^{n+1}({\sf epq}_{n+|s|+2}^{l \circ (q_x)^{n+1}}(\phi^{\varepsilon, (q_x)^{n+1}})((f^{n+1})_{s * y})))$
\end{itemize}
so that $(*) \; c = \tilde{c}$ follows directly from the induction hypothesis $\forall y P(s * y)$.
\end{proof}
We are now ready to show that ${\sf eps}$ is $T$-definable from ${\sf epq}$. The proof presented here is essentially the same as Spector's proof that his restricted form of bar recursion ${\sf SBR}$ follows from the general form ${\sf BR}$ (cf. Section \ref{sec-spector}).
\begin{theorem}[${\rm HA}^\omega + {\rm BI}$] \label{cps-sbr} ${\sf epq} \geq_T {\sf eps}$.
\end{theorem}
\begin{proof} Let $\phi^{\varepsilon, q}_n$ and $f^n$ be as defined in Lemma \ref{epq-eps-lemma}. We claim that ${\sf eps}$ can be defined from ${\sf epq}$ as
\begin{itemize}
\item[$(i)$] ${\sf eps}_n^l(\varepsilon)(q) \stackrel{\Pi_i X_{i+n}}{=} ({\sf epq}_n^{l \circ q^n}(\phi^{\varepsilon, q^n})(f^n))^n$.
\end{itemize}
We consider two cases. \\[1mm]
If $l(q({\bf 0}^{\Pi_{i \geq n} X_i})) < n$ then we also have $l(q^n({\bf 0}^{\Pi_i X_i})) < n$. Therefore
\eqleft{{\sf eps}_n^l(\varepsilon)(q)
\,\stackrel{(i)}{=}\, ({\sf epq}_n^{l \circ q^n}(\phi^{\varepsilon, q^n})(f^n))^n
\,\stackrel{(\ref{def-epq-equation})}{=}\, (f^n({\bf 0}))^n
\,=\, \cZr{\Pi_i X_{i + n}}.
}
On the other hand, if $l(q({\bf 0}^{\Pi_{i \geq n} X_i})) \geq n$ then $l(q^n({\bf 0}^{\Pi_i X_i})) \geq n$ and hence:
\eqleft{
\begin{array}{lcl}
{\sf eps}_n^l(\varepsilon)(q)
& \stackrel{(i)}{=} & ({\sf epq}_n^{l \circ q^n}(\phi^{\varepsilon, q^n})(f^n))^n \\[1mm]
& \stackrel{(\ref{def-epq-equation})}{=} & (\phi_n^{\varepsilon, q^n}(\lambda x^{X_n} . {\sf epq}_{n+1}^{l \circ q^n}(\phi^{\varepsilon, q^n})((f^n)_x)))^n \\[1.5mm]
& \stackrel{(\ref{epq-eps-lemma-eq})}{=} & ({\sf epq}_{n+1}^{l \circ q^n}(\phi^{\varepsilon, q^n})((f^n)_c))^n \\[1.5mm]
& \stackrel{\textup{L}\ref{epq-eps-lemma}}{=} & (({\bf 0} * c) \,@\, {\sf epq}_{n + 1}^{l \circ (q_c)^{n+1}}(\phi^{\varepsilon, (q_c)^{n+1}})(f^{n+1}))^n \\[1.5mm]
& = & c * ({\sf epq}_{n+1}^{l \circ (q_c)^{n+1}}(\phi^{\varepsilon, (q_c)^{n+1}})(f^{n+1}))^{n+1} \\[1mm]
& \stackrel{(i)}{=} & c * {\sf eps}_{n+1}^l(\varepsilon)(q_c)
\end{array}
}
where
\eqleft{
\begin{array}{lcl}
c
& = & \varepsilon_n(\lambda x^{X_n} . q^n({\sf epq}_{n+1}^{l \circ q^n}(\phi^{\varepsilon, q^n})((f^n)_x))) \\[1mm]
& \stackrel{\textup{L}\ref{epq-eps-lemma}}{=} & \varepsilon_n(\lambda x^{X_n} . q^n(({\bf 0} * x) \,@\, {\sf epq}_{n+1}^{l \circ (q_x)^{n+1}}(\phi^{\varepsilon, (q_x)^{n+1}})(f^{n+1}))) \\[1.5mm]
& = & \varepsilon_n(\lambda x^{X_n} . (q_x)^{n+1}({\sf epq}_{n+1}^{l \circ (q_x)^{n+1}}(\phi^{\varepsilon, (q_x)^{n+1}})(f^{n+1}))) \\[1.5mm]
& = & \varepsilon_n(\lambda x^{X_n} . q_x({\sf epq}_{n+1}^{l \circ (q_x)^{n+1}}(\phi^{\varepsilon, (q_x)^{n+1}})(f^{n+1}))^{n+1}) \\[1.5mm]
& \stackrel{(i)}{=} & \varepsilon_n(\lambda x^{X_n} . q_x({\sf eps}_{n+1}^{l}(\varepsilon)(q_x))).
\end{array}
}
\end{proof}
\subsection{Dialectica interpretation of classical analysis}
\label{dialectica}
In order to find witnesses for the dialectica interpretation of ${\rm DNS}$, and hence full classical analysis, Spector arrived at the following system of equations
\begin{equation} \label{spector-eq}
\begin{array}{lcl}
n & \stackrel{\mathbb{N}}{=} & \omega \alpha, \\[2mm]
\alpha(n) & \stackrel{X}{=} & \varepsilon_n(p), \\[2mm]
p(\alpha(n)) & \stackrel{R}{=} & q \alpha,
\end{array}
\end{equation}
where $\varepsilon_n \colon J_R X$ and $q \colon (\mathbb{N} \to X) \to R$ and $\omega \colon (\mathbb{N} \to X) \to \mathbb{N}$ are given and $n \colon \mathbb{N}$ and $\alpha \colon \mathbb{N} \to X$ and $p \colon X \to R$ are the unknowns.
We now show how ${\sf eps}$ can be used to solve Spector's equations. We first solve a slightly different set of equations, and as a corollary we obtain a solution to Spector's original one.
\begin{theorem}[${\rm HA}^\omega + (\ref{eps-eq-def})$] \label{main-spec} Let $q \colon \Pi_i X_i \to R$ and $l \colon R \to \mathbb{N}$ and $\varepsilon \colon \Pi_i J_R X_i$ be given. Define
\eqleft{
\begin{array}{lcl}
\alpha & = & {\sf eps}_0^{l}(\varepsilon)(q) \\[2mm]
p_n(x) & = & \selEmb{{\sf eps}^l_{n+1}(\varepsilon)}(q_{\initSeg{\alpha}{n} * x}).
\end{array}
}
For $n \leq l(q(\alpha))$ we have
\[
\begin{array}{lcl}
\alpha(n) & \stackrel{X_n}{=} & \varepsilon_n(p_n) \\[2mm]
p_n(\alpha(n)) & \stackrel{R}{=} & q \alpha.
\end{array}
\]
\end{theorem}
\begin{proof} This is essentially Spector's proof (cf. lemma 11.5 of \cite{Kohlenbach(2008)}). Assume $n \leq l(q(\alpha))$. We first argue that $n \leq l(q(\initSegZ{\alpha}{n}))$. Otherwise, assuming $n > l(q(\initSegZ{\alpha}{n})) = l(q_{\initSeg{\alpha}{n}}({\bf 0}))$ we would have, by Lemma \ref{spector-main-lemma}, that $\alpha = \initSegZ{\alpha}{n}$. And hence, by extensionality, $n > l(q_{\initSeg{\alpha}{n}}({\bf 0})) = l(q(\alpha)) \geq n$, which is a contradiction. \\[2mm]
Hence, assuming $n \leq l(q(\alpha))$ we have $n \leq l(q(\initSegZ{\alpha}{n}))$ and hence
\eqleft{\alpha(n) \,\stackrel{\textup{C}\ref{unwinding-cps}}{=}\, \varepsilon_n(\lambda x . \selEmb{{\sf eps}_{n + 1}^l(\varepsilon)}(q_{\initSeg{\alpha}{n} * x})) \,=\, \varepsilon_n(p_n).}
For the second equality, we have
\eqleft{
\begin{array}{lcl}
p_n(\alpha(n)) & = & \selEmb{{\sf eps}_{n+1}^l(\varepsilon)}(q_{\initSeg{\alpha}{n+1}}) \\[2mm]
& = & q_{\initSeg{\alpha}{n+1}}({\sf eps}_{n+1}^l(\varepsilon)(q_{\initSeg{\alpha}{n+1}})) \\[2mm]
& = & q(\initSeg{\alpha}{n+1} * {\sf eps}_{n+1}^l(\varepsilon)(q_{\initSeg{\alpha}{n+1}})) \\[1mm]
& \stackrel{\textup{L}\ref{spector-main-lemma}}{=} & q(\alpha).
\end{array}
}
\end{proof}
\begin{corollary} \label{spector-solution} For any given $q \colon X^\mathbb{N} \to R$ and $\omega \colon X^\mathbb{N} \to \mathbb{N}$ and sequence of selection functions $\varepsilon_n \colon J_R X$ (of common type $J_R X$) there are $\alpha \colon \mathbb{N} \to X$ and $p \colon X \to R$ satisfying the system of equations (\ref{spector-eq}).
\end{corollary}
\begin{proof} Let $R' = R \times \mathbb{N}$, and let $\pi_0 \colon R \times \mathbb{N} \to R$ and $\pi_1 \colon R \times \mathbb{N} \to \mathbb{N}$ denote the first and second projections. Define
\eqleft{
\begin{array}{lcl}
q'(\alpha) & \stackrel{R'}{=} & \pair{q(\alpha), \omega(\alpha)} \\[2mm]
\varepsilon_n'(p^{X \to R'}) & \stackrel{X}{=} & \varepsilon_n(\lambda x^{X} . \pi_0(p(x))).
\end{array}
}
so $q' \colon (\mathbb{N} \to X) \to R'$ and $\varepsilon_n' \colon J_{R'} X$. Let
\eqleft{
\begin{array}{lcl}
\alpha & \stackrel{X^\mathbb{N}}{=} & {\sf eps}_0^{\pi_1}(\varepsilon')(q') \\[2mm]
p_n'(x^X) & \stackrel{R'}{=} & \selEmb{{\sf eps}_{n+1}^{\pi_1}(\varepsilon')}(q_{\initSeg{\alpha}{n} * x}').
\end{array}
}
Assume $n \leq \omega(\alpha) = \pi_1(q'(\alpha))$. By Theorem \ref{main-spec} we have
\[
\begin{array}{lcl}
\alpha(n) & \stackrel{X}{=} & \varepsilon_n'(p_n') \\[2mm]
p_n'(\alpha(n)) & \stackrel{R'}{=} & q' \alpha.
\end{array}
\]
Finally, let $n = \omega(\alpha)$ and $p(x) = \pi_0(p_n'(x))$. Then it is easy to check that $\alpha$ and $p$ satisfy the desired equation, e.g. $\alpha(n) = \varepsilon_n'(p_n') = \varepsilon_n(\pi_0 \circ p_n') = \varepsilon_n(p)$.
\end{proof}
\subsection{Dependent variants of ${\sf eps}$ and ${\sf epq}$}
\label{EPS-eps}
In the dialectica interpretation of ${\rm DNS}$ given above (Section \ref{dialectica}), the selection functions $\varepsilon_n$ do not depend on the history of choices already made. Thus, it was sufficient to use an iteration of the \emph{simple} product of selection functions. Nevertheless, Spector bar recursion and modified bar recursion are normally formulated in the most general form, where selection functions at point $n$ have access to the values $X_i$ for $i < n$.
In the same vein, in previous papers \cite{EO(2010A),EO(2009)} we have considered generalisations of the product of selection functions, where a selection function (or a quantifier) at stage $n$ can have access to the previously computed values. We called these the \emph{dependent} product of selection functions and quantifiers.
\begin{definition}[Dependent product of selection functions and quantifiers] \label{main-dependent} Given a quantifier $\phi \colon K X$ and a family of quantifiers $\psi \colon X \to K Y$, define the dependent product quantifier $(\phi \ttimesTensor_d \psi) \colon K (X \times Y)$ as
\[ (\phi \ttimesTensor_d \psi)(p^{X \times Y \to R}) \stackrel{R}{=} \phi(\lambda x^X . \psi (x, \lambda y^Y . p(x, y))). \]
Also, given a selection function $\varepsilon \colon J X$ and a family of selection functions $\delta \colon X \to J Y$, define the dependent product selection function $(\varepsilon \ttimesTensor_d \delta) \colon J(X \times Y)$ as
\[ (\varepsilon \ttimesTensor_d \delta)(p^{X \times Y \to R}) \stackrel{X \times Y}{=} (a, b(a)) \]
where
\eqleft{
\begin{array}{lcl}
a & \stackrel{X}{=} & \varepsilon(\lambda x^X. p(x, b(x))) \\[2mm]
b(x) & \stackrel{Y}{=} & \delta(x, \lambda y^Y . p(x, y)).
\end{array}
}
\end{definition}
As done for the simple product of selection functions and quantifiers, we can also iterate the dependent products as follows:
\begin{definition}[$\EPQ$ and $\EPS$] \label{definition-EPS} Given a family of quantifiers \[ \phi \colon \Pi_k (\Pi_{i < k} X_i \to K X_k), \] define their \emph{dependent explicitly controlled product} (denoted $\EPQ$) as
\[
\EPQ_s^l(\phi)(q) \stackrel{R}{=}
\left\{
\begin{array}{ll}
q({\bf 0}) & {\rm if} \;l(q({\bf 0})) < |s| \\[2mm]
(\phi_s \ttimesTensor_d (\lambda x^{X_{|s|}} . \EPQ_{s * x}^l(\phi)))(q) & {\rm otherwise}.
\end{array}
\right.
\]
Unpacking the definition of the binary dependent product $\ttimesTensor_d$ this is equivalent to
\begin{equation} \label{EPQ-def} \tag{\EPQ}
\EPQ_s^l(\phi)(q) \stackrel{R}{=}
\left\{
\begin{array}{ll}
q({\bf 0}) & {\rm if} \;l(q({\bf 0})) < |s| \\[2mm]
\phi_s(\lambda x^{X_{|s|}} . \EPQ_{s * x}^l(\phi)(q_x)) & {\rm otherwise}.
\end{array}
\right.
\end{equation}
Moreover, given a family of selection functions \[ \varepsilon \colon \Pi_k (\Pi_{i < k} X_i \to J X_k), \] define their \emph{dependent explicitly controlled product} (denoted $\EPS$) as
\[
\EPS_s^l(\varepsilon)(q) \stackrel{\Pi_i X_{|s|+i}}{=}
\left\{
\begin{array}{ll}
{\bf 0} & {\rm if} \;l(q({\bf 0})) < |s| \\[2mm]
(\varepsilon_s \ttimesTensor_d (\lambda x^{X_{|s|}} . \EPS_{s * x}^l(\varepsilon)))(q) & {\rm otherwise}.
\end{array}
\right.
\]
Similarly, unfolding the definition of $\ttimesTensor_d$ the defining equation for $\EPS$ is equivalent to
\begin{equation} \label{EPS-def} \tag{\EPS}
\EPS_s^l(\varepsilon)(q) \stackrel{\Pi_i X_{|s|+i}}{=}
\left\{
\begin{array}{ll}
{\bf 0} & {\rm if} \;l(q({\bf 0})) < |s| \\[2mm]
c * \EPS_{s * c}^l(\varepsilon)(q_c) & {\rm otherwise}
\end{array}
\right.
\end{equation}
where $c = \varepsilon_s(\lambda x . \selEmb{\EPS_{s * x}^l(\varepsilon)}(q_x))$.
\end{definition}
Clearly ${\sf eps}$ is $T$-definable from $\EPS$. We now show that in fact ${\sf eps}$ and $\EPS$ are $T$-equivalent. In Theorem \ref{EPS-from-eps} we will make use of the following construction. Given $\alpha \colon \Pi_{i \geq n}(\Pi_{k < i} X_k \to X_i)$ and $s \colon \Pi_{k<n} X_k$ define $\alpha^s \colon \Pi_{i \geq n} X_i$ by course-of-values as
\begin{equation} \label{s-construction}
\alpha^s(i) \stackrel{X_{n+i}}{=} \alpha(i)(s * \initSeg{\alpha^s}{i}).
\end{equation}
Clearly, given a finite sequence $t \colon \Pi_{i \in [n, m]}(\Pi_{k < i} X_k \to X_i)$ we can perform the same construction to obtain a $t^s \colon \Pi_{i \in [n, m]} X_i$.
\begin{lemma}[${\rm HA}^\omega$] \label{s-constr-lemma} $(d * \alpha)^s = d(s) * (\alpha)^{s * d(s)}$, where $d \colon \Pi_{k < n} X_k \to X_n$.
\end{lemma}
\begin{proof} Straightforward.
\end{proof}
Finally, given $q \colon \Pi_{i \geq n} X_i \to R$ define $q^s \colon \Pi_{i \geq n}(\Pi_{k < i} X_k \to X_i) \to R$ as
\eqleft{
\begin{array}{lcl}
q^s(\alpha) & \stackrel{R}{=} & q(\alpha^s).
\end{array}
}
\begin{theorem}[${\rm HA}^\omega$] \label{EPS-from-eps} ${\sf eps} \geq_T \EPS$.
\end{theorem}
\begin{proof} To define $\EPS$ of type $(X_k, R)$ we use ${\sf eps}$ of type $(\Pi_{i < k} X_i \to X_k, R)$. Given selection functions $\varepsilon_s \colon J X_{|s|}$ define $\tilde{\varepsilon}_k \colon J(\Pi_{i<k} X_i \to X_k)$ as
\begin{itemize}
\item[$(i)$] $\tilde{\varepsilon}_k(P^{(\Pi_{i<k} X_i \to X_k) \to R}) \stackrel{\Pi_{i<k} X_i \to X_k}{=} \lambda s^{\Pi_{i<k} X_i} . \varepsilon_s(\lambda y^{X_k}. P(\lambda t. y))$.
\end{itemize}
Note that the infinite (simple) product of the selection functions $\tilde{\varepsilon}_k$ has type
\eqleft{{\sf eps}_n^l(\tilde{\varepsilon}) \; \colon \; J(\Pi_{i \geq n}(\Pi_{k < i} X_k \to X_i))}
where $l \colon R \to \mathbb{N}$. We claim that $\EPS$ can be defined from ${\sf eps}$ as
\begin{itemize}
\item[$(ii)$] $\EPS_s^l(\varepsilon)(q^{\Pi_i X_{i + |s|} \to R}) \stackrel{\Pi_i X_{i + |s|}}{=} ({\sf eps}_{|s|}^l(\tilde{\varepsilon})(q^s))^s$,
\end{itemize}
where $s \colon \Pi_{k<|s|} X_k$. Let us show that $\EPS$ as defined above satisfies the defining equation ($\ref{EPS-def}$).
Consider two cases: \\[1mm]
If $l(q({\bf 0})) < |s|$ then $l(q^s({\bf 0})) = l(q({\bf 0})) < |s|$. Hence, by the definition of $(\cdot)^s$
\eqleft{\EPS_s^l(\varepsilon)(q) \stackrel{(ii)}{=} ({\sf eps}_{|s|}^l(\tilde{\varepsilon})(q^s))^s \stackrel{(\ref{eps-eq-def})}{=} ({\bf 0})^s = {\bf 0}.}
On the other hand, if $l(q({\bf 0})) \geq |s|$ then $l(q^s({\bf 0})) = l(q({\bf 0})) \geq |s|$. Hence
\eqleft{
\begin{array}{lcl}
\EPS_s^l(\varepsilon)(q)
& \stackrel{(ii)}{=} & ({\sf eps}_{|s|}^l(\tilde{\varepsilon})(q^s))^s \\[1mm]
& \stackrel{(\ref{eps-eq-def})}{=} & (d * {\sf eps}_{|s |+1}^l(\tilde{\varepsilon})((q^s)_d))^s \\[1mm]
& \stackrel{\textup{L}\ref{s-constr-lemma}}{=} & d(s) * ({\sf eps}_{|s|+1}^l(\tilde{\varepsilon})((q^s)_d))^{s * d(s)} \\[1mm]
& \stackrel{\textup{L}\ref{s-constr-lemma}}{=} & d(s) * ({\sf eps}_{|s|+1}^l(\tilde{\varepsilon})((q_{d(s)})^{s * d(s)}))^{s * d(s)} \\[1mm]
& \stackrel{(*)}{=} & c * ({\sf eps}_{|s*c|}^l(\tilde{\varepsilon})((q_c)^{s * c}))^{s * c} \\[1mm]
& \stackrel{(ii)}{=} & c * \EPS_{s * c}^l(\varepsilon)(q_c),
\end{array}
}
where $d = \tilde{\varepsilon}_{|s|}(\lambda f . (q^s)_f({\sf eps}_{|s|+1}^l(\tilde{\varepsilon})((q^s)_f)))$ and $c = \varepsilon_s(\lambda x . q_x(\EPS_{s * x}^l(\varepsilon)(q_x))$ so that $(*)$
\eqleft{
\begin{array}{lcl}
d(s) & = & \tilde{\varepsilon}_{|s|}(\lambda f . (q^s)_f({\sf eps}_{|s|+1}^l(\tilde{\varepsilon})((q^s)_f)))(s) \\[1mm]
& \stackrel{(i)}{=} & \varepsilon_s(\lambda x . (q^s)_{\lambda t . x}({\sf eps}_{|s| + 1}^l(\tilde{\varepsilon})((q^s)_{\lambda t . x}))) \\[1mm]
& \stackrel{\textup{L}\ref{s-constr-lemma}}{=} & \varepsilon_s(\lambda x . (q_x)^{s * x}({\sf eps}_{|s| + 1}^l(\tilde{\varepsilon})((q_x)^{s * x}))) \\[2mm]
& = & \varepsilon_s(\lambda x . q_x({\sf eps}_{|s * x|}^l(\tilde{\varepsilon})((q_x)^{s * x}))^{s * x}) \\[1mm]
& \stackrel{(ii)}{=} & \varepsilon_s(\lambda x . q_x(\EPS_{s * x}^l(\varepsilon)(q_x)) \\[2mm]
& = & c.
\end{array}
}
\end{proof}
\begin{remark} Note that a similar construction does not work in the case of quantifiers, since there is no $\lambda$-term (in the pure simply typed $\lambda$-calculus) of type $(X \to K Y) \to K(X \to Y)$, for arbitrary $X$ and $Y$. Nevertheless, it will follow from our results that the explicitly controlled iteration of the simple product of quantifiers ${\sf epq}$ is $T$-equivalent to the explicitly controlled iteration of the dependent product of quantifiers $\EPQ$ (cf. summary of results in Figure \ref{table}).
\end{remark}
\subsection{Relation to Spector's bar recursion}
\label{sec-spector}
As we have shown in Theorem \ref{main-spec}, which is essentially Spector's solution, the explicitly controlled product of selection functions ${\sf eps}$ can also be used to give a computational interpretation of classical analysis. When presenting his solution in \cite{Spector(62)}, Spector first formulates a general ``construction by bar recursion" as
\begin{equation} \label{BR-def} \tag{{\sf BR}}
{\sf BR}_s^\omega(\phi)(q) \stackrel{R}{=}
\left\{
\begin{array}{ll}
q_s({\bf 0}) & {\rm if} \; \omega_s({\bf 0}) < |s| \\[2mm]
\phi_s(\lambda x^{X_{|s|}} . {\sf BR}_{s*x}^\omega(\phi)(q)) \quad & {\rm otherwise},
\end{array}
\right.
\end{equation}
where $\phi_s \colon K_R X_{|s|}$, $q \colon \Pi_i X_i \to R$ and $\omega \colon \Pi_i X_i \to \mathbb{N}$. This is usually referred to as \emph{Spector's bar
recursion}, but we argue that this is misleading. We show that ${\sf BR}$ is
closely related to the product of \emph{quantifiers} $\EPQ$, whereas
the special case of this used by Spector is equivalent to the
(dependent) product of \emph{selection functions} $\EPS$, which we
have shown to be equivalent to ${\sf eps}$ (Section \ref{EPS-eps}).
\begin{remark} In fact, Spector's definition seems slightly more general than ${\sf BR}$ as defined here, since in Spector's definition $q$ might also depend on the length of the sequence $s$. As we show in Lemma \ref{spec-search}, however, it is possible to reconstruct $|s|$ from the sequence $s * {\bf 0}$ if $s$ is the point where Spector's condition first happens.
\end{remark}
\begin{theorem}[${\rm HA}^\omega + {\rm BI}$] \label{BR-EPQ-def} ${\sf BR} \geq_T \EPQ$.
\end{theorem}
\begin{proof} In order to define $\EPQ$ of type $(X_i, R)$ we use ${\sf BR}$ of the same type $(X_i, R$). ${\sf BR}$ and $\EPQ$ have very similar definitions, except that in ${\sf BR}$ the stopping condition is given directly on the current sequence $s * {\bf 0}$, whereas in $\EPQ$ a ``length" function $l \colon R \to \mathbb{N}$ is used so that the stopping condition involves the composition $l \circ q$. Hence, in order to define $\EPQ$ from ${\sf BR}$ it is essentially enough to take $\omega = l \circ q$, taking care of the fact that the types of $q$ in $\EPQ$ and ${\sf BR}$ are slightly different as $q$ in $\EPQ$ takes a ``shorter" input sequence starting at point $|s|$. We show that $\EPQ$ defined as
\begin{itemize}
\item[$(i)$] $\EPQ_s^{l}(\phi)(q) = {\sf BR}_s^{l \circ q^{|s|}}(\phi)(q^{|s|})$
\end{itemize}
satisfies the equation $(\ref{EPQ-def})$. Consider two cases. \\[1mm]
If $(l \circ q^{|s|})_s({\bf 0}) = l(q({\bf 0})) < |s|$ then
\eqleft{
\EPQ_s^{l}(\phi)(q) \stackrel{(i)}{=} {\sf BR}_s^{l \circ q^{|s|}}(\phi)(q^{|s|}) \stackrel{(\ref{BR-def})}{=} (q^{|s|})_s({\bf 0}) = q({\bf 0}).
}
On the other hand, if $(l \circ q^{|s|})_s({\bf 0}) = l(q({\bf 0})) \geq |s|$ then
\eqleft{
\begin{array}{lcl}
\EPQ_s^l(\phi)(q)
& \stackrel{(i)}{=} & {\sf BR}_s^{l \circ q^{|s|}}(\phi)(q^{|s|}) \\[1.5mm]
& \stackrel{(\ref{BR-def})}{=} & \phi_s(\lambda x^{X_{|s|}} . {\sf BR}_{s*x}^{l \circ q^{|s|}}(\phi)(q^{|s|})) \\[1mm]
& \stackrel{(*)}{=} & \phi_s(\lambda x^{X_{|s|}} . {\sf BR}_{s * x}^{l \circ (q_x)^{|s*x|}}(\phi)((q_x)^{|s*x|}) \\[2mm]
& = & \phi_s(\lambda x^{X_{|s|}} . \EPQ_{s * x}^l(\phi)(q_x))
\end{array}
}
where
\begin{itemize}
\item[$(*)$] ${\sf BR}_{s*x}^{l \circ q^{|s|}}(\phi)(q^{|s|}) = {\sf BR}_{s * x}^{l \circ (q_x)^{|s * x|}}(\phi)((q_x)^{|s*x|})$
\end{itemize}
can, as in Lemma~\ref{epq-eps-lemma}, be proved by bar induction ${\rm BI}$ and axiom ${\rm SPEC}$, since
\eqleft{q^{|s|}(s * x * \alpha) = (q_x)^{|s * x|}(s * x * \alpha).}
Finally, recall that ${\rm HA}^\omega + (\ref{BR-def}) \vdash {\rm SPEC}$ (Lemma 3C of \cite{Howard(1968)}).
\end{proof}
Spector, however, explicitly says that only a \emph{restricted form} of ${\sf BR}$ is used for the dialectica interpretation of (the negative translation of) countable choice. It is this restricted form that we shall from now on call \emph{Spector's bar recursion}:
\begin{definition}[Spector's bar recursion] Spector's bar recursion \cite{Spector(62)} is the
recursion schema
\[
{\sf SBR}_s^\omega(\varepsilon) \stackrel{\Pi_i X_i}{=}
s
\,@\,
\left\{
\begin{array}{ll}
{\bf 0} & {\rm if} \; \omega_s({\bf 0}) < |s| \\[2mm]
{\sf SBR}_{s * c}^\omega(\varepsilon) & {\rm otherwise},
\end{array}
\right.
\]
where $c \stackrel{X_{|s|}}{=} \varepsilon_{s}(\lambda x^{X_{|s|}} . {\sf SBR}_{s * x}^\omega(\varepsilon))$, and where $\varepsilon_s \colon J_{\Pi_i X_i} X_{|s|}$ and $\omega \colon \Pi_i X_i \to \mathbb{N}$.
\end{definition}
We now show that Spector's bar recursion is $T$-definable from the explicitly controlled product of selection functions $\EPS$. It will follow from other results that they are in fact $T$-equivalent (see Figure \ref{table}).
\begin{theorem}[${\rm HA}^\omega$] \label{main-eps-sbr} $\EPS \geq_T {\sf SBR}$.
\end{theorem}
\begin{proof} To define ${\sf SBR}$ of type $(X_i)$ we use $\EPS$ of type $(X_i, (\Pi_i X_i) \times \mathbb{N})$. $\EPS$ and ${\sf SBR}$ have very similar definitions, except that $\EPS$ has an extra argument $q \colon \Pi_{i \geq |s|} X_i \to R$. We can obtain ${\sf SBR}$ from $\EPS$ by simply taking $q(\alpha)$ to be the identity function plus the stopping value $\omega(\alpha)$. So, the length function $l \colon R \to \mathbb{N}$ can be taken to be the second projection. The details are as follows: Let $R = (\Pi_i X_i) \times \mathbb{N}$. Given $\omega \colon \Pi_i X_i \to \mathbb{N}$ and $\varepsilon_s \colon J_{\Pi_i X_i} X_{|s|}$, define
\eqleft{
\begin{array}{lcl}
l(r^R) & \stackrel{\mathbb{N}}{=} & \pi_1(r) \\[1mm]
q^\omega(\alpha^{\Pi_i X_i}) & \stackrel{R}{=} & \pair{\alpha, \omega(\alpha)} \\[1mm]
\tilde \varepsilon_s (p^{X_{|s|} \to R}) & \stackrel{X_{|s|}}{=} & \varepsilon_s(\pi_0 \circ p).
\end{array}
}
Define
\begin{itemize}
\item[$(i)$] ${\sf SBR}_s^\omega(\varepsilon) \stackrel{\Pi_i X_i}{=} s * \EPS_s^l(\tilde \varepsilon)((q^\omega)_s)$.
\end{itemize}
If $\omega_s({\bf 0}) < |s|$ then $l((q^\omega)_s({\bf 0})) = \omega_s({\bf 0}) < |s|$ and hence
\eqleft{
\begin{array}{lcl}
{\sf SBR}_s^\omega(\varepsilon)
& \stackrel{(i)}{=} & s * \EPS_s^l(\tilde \varepsilon)((q^\omega)_s) \\[0.5mm]
& \stackrel{(\ref{EPS-def})}{=} & s * {\bf 0} \\[1.5mm]
& = & s \,@\, {\bf 0}.
\end{array}
}
On the other hand, if $\omega_s({\bf 0}) \geq |s|$ then also $l((q^\omega)_s({\bf 0})) \geq |s|$ and we have
\eqleft{
\begin{array}{lcl}
{\sf SBR}_s^\omega(\varepsilon)
& \stackrel{(i)}{=} & s * \EPS_s^l(\tilde \varepsilon)((q^\omega)_s) \\[0.5mm]
& \stackrel{(\ref{EPS-def})}{=} & s * c * \EPS_{s * c}^l(\tilde \varepsilon)((q^\omega)_{s * c}) \\[1mm]
& \stackrel{(*)}{=} & s * d * \EPS_{s * d}^l(\tilde \varepsilon)((q^\omega)_{s * d}) \\[1mm]
& \stackrel{(i)}{=} & {\sf SBR}_{s*d}^\omega(\varepsilon)
\end{array}
}
where $c = \tilde \varepsilon_s(\lambda x . (q^\omega)_{s * x}(\EPS_{s * x}^l(\tilde \varepsilon)((q^\omega)_{s * x})))$ and $d = \varepsilon_s(\lambda x . {\sf SBR}_{s * x}^\omega(\varepsilon))$ so that $(*)$
\eqleft{
\begin{array}{lcl}
c & = & \tilde \varepsilon_s(\lambda x . (q^\omega)_{s * x}(\EPS_{s * x}^l(\tilde \varepsilon)((q^\omega)_{s * x}))) \\[1mm]
& \stackrel{\textup{def} \; q^\omega,\, \tilde \varepsilon_s}{=} & \varepsilon_s(\lambda x . s * x * \EPS_{s * x}^l(\tilde \varepsilon)((q^\omega)_{s * x})) \\[1mm]
& \stackrel{(i)}{=} & \varepsilon_s(\lambda x . {\sf SBR}_{s * x}^\omega(\varepsilon)) \;=\; d.
\end{array}
}
\end{proof}
\section{Implicitly Controlled Product}
\label{implicit}
We have seen in Section \ref{conditional} above that the explicitly
controlled iterated product of selection functions is sufficient to
witness the dialectica interpretation of the double negation shift (and hence, classical countable
choice). In this section we show that when interpreting this same
principle via modified realizability, one seems to need an
\emph{unrestricted} or, as we we shall call it, \emph{implicitly
controlled} infinite product of selection functions.
\begin{definition}[${\sf ips}$] The \emph{implicitly controlled product} of a sequence of selection functions $\varepsilon \colon \Pi_k J X_k$ is defined as
\[ {\sf ips}_n(\varepsilon) \stackrel{J (\Pi_i X_{i + n})}{=} \varepsilon_n \ttimesTensor {\sf ips}_{n+1}(\varepsilon). \]
Unfolding the definition of $\ttimesTensor$, this is the same as
\begin{equation} \label{ips-def-eq} \tag{{\sf ips}}
{\sf ips}_n(\varepsilon)(q) \stackrel{\Pi_i X_{i + n}}{=} \underbrace{\varepsilon_n(\lambda x . q_x({\sf ips}_{n+1}(\varepsilon)(q_x)))}_{c} * \,{\sf ips}_{n+1}(\varepsilon)(q_c).
\end{equation}
\end{definition}
We call the above infinite product \emph{implicitly controlled} because under the assumption of continuity for $q \colon \Pi_i X_{i + n} \to R$, for discrete $R$, the bar recursive calls
eventually terminate.
\begin{remark} As shown in section 5.6. of \cite{EO(2009)}, an implicitly controlled product of quantifiers ${\sf ipq}$
\[
{\sf ipq}_n(\phi) = \phi_n \ttimesTensor {\sf ipq}_{n+1}(\phi)
\]
does not exist. It is enough to consider the case when $R = X_i = \mathbb{N}$. Let $\phi_n(p) = 1 + p(0)$ and $q$ be any function. Assuming the equation above, it follows by induction that for all $n$
\[
{\sf ipq}_0(\phi)(q) = n + {\sf ipq}_n(\phi)(q_{0^n}),
\]
where $0^n = \langle 0, 0, \ldots, 0 \rangle$, with $n$ zeros; which implies ${\sf ipq}_0(\phi)(q) \geq n$, for all $n$.
\end{remark}
\subsection{Realizability interpretation of classical analysis}
\label{realizability}
We now describe how ${\sf ips}$ can be used to interpret the double negation shift (and hence countable choice) via modified realizability. As discussed in the introduction, a computational interpretation of full classical analysis can be reduced to an interpretation of the double negation shift ${\rm DNS}$. Given that the formula $A(n)$ (in ${\rm DNS}$) can be assumed to be of the form $\exists x \neg B(n, x)$, ${\rm DNS}$ is equivalent to
\[
\forall n ((A(n) \to \perp) \to A(n)) \to (\forall n A(n) \to \perp) \to \forall n A(n).
\]
That is because, for $A(n) \equiv \exists x \neg B(n, x)$, we have both $\perp \, \to A(n)$ and $\perp \, \to \forall n A(n)$ in \emph{minimal logic}. Moreover, since the negative translation brings us into minimal logic, falsity $\perp$ can be replaced by an arbitrary $\Sigma^0_1$-formula~$R$. This is known as the (refined) $A$-translation \cite{Berger(95)}, and is useful to analyse proofs of $\Pi^0_2$ theorems in analysis. Recall that we are using the abbreviation
\eqleft{J_R A = (A \to R) \to A.}
The resulting principle we obtain is what we shall call the $J$-shift
\eqleft{J\textup{-}{\sf shift} \quad \colon \quad \forall n J_R A(n) \to J_R \forall n A(n).}
${\rm DNS}$ is then the particular case of the $K$-${\sf shift}$
\eqleft{K\textup{-}{\sf shift} \quad \colon \quad \forall n K_R A(n) \to K_R \forall n A(n),}
when $R = \perp$; considering the other type construction
\eqleft{K_R A = (A \to R) \to R.}
One advantage of moving to the $J\textup{-}{\sf shift}$ is that $A(n)$ now can be taken to be an arbitrary formula, not necessarily of the form $\exists x \neg B(n, x)$. Hence the principle $J\textup{-}{\sf shift}$ is more general than ${\rm DNS}$. We analyse the logical strength of the principle $J\textup{-}{\sf shift}$ in more detail in \cite{EO(2010B)}, where a proof translation based on the construction $J_R A$ is also defined.
Our proof of the following theorem is very similar to that of \cite[Theorem 3]{BO(02A)}.
We assume continuity and relativised bar induction as formulated in Section \ref{bar-ind}.
\begin{theorem}[${\rm HA}^\omega + {\rm BI} + {\rm CONT}$] \label{ips-jshift} ${\sf ips}_0$ modified realizes $J\textup{-}{\sf shift}$.
\end{theorem}
\begin{proof} Given a term $t$ and a formula $A$ we write ``$t \,{\sf mr}\, A$" for ``$t$ modified realizes $A$" (see \cite{Troelstra(73)} for definition). Assume that
\eqleft{
\begin{array}{lcl}
\varepsilon_n & \,{\sf mr}\, & (A(n) \to R) \to A(n), \\[2mm]
q & \,{\sf mr}\, & \forall n A(n) \to R.
\end{array} }
Let
\eqleft{
\begin{array}{lcl}
P(s) & \equiv & s * {\sf ips}_{|s|}(\varepsilon)(q_s) \,\,{\sf mr}\,\, \forall n A(n) \\[2mm]
S(s) & \equiv & \forall n \!<\! |s| \, (s_n \,{\sf mr}\, A(n)).
\end{array}
}
We show $P(\langle \, \rangle)$ by bar induction relativised to the predicate $S$. Let us write $\alpha \in S$ as an abbreviation for $\forall n (\initSeg{\alpha}{n} \in S)$. The first assumption of ${\rm BI}$ (i.e. $S(\pair{\,})$) is vacuously true. We now prove the other two assumptions. \\[2mm]
($i$) $\forall \alpha \!\in\! S \, \exists k \, P(\initSeg{\alpha}{k})$. Given $\alpha \in S$ let $k$ be a point of continuity of $q$ on $\alpha$. Let $r := q \alpha$. By ${\rm CONT}$ we have $q(\initSeg{\alpha}{k} * \beta) = r$, for all $\beta$. By the assumptions on $\alpha$ and $q$ we have that $r \,{\sf mr}\, R$. We must show that for all $n$
\[ (\initSeg{\alpha}{k} * {\sf ips}_k(\varepsilon)(\lambda \beta . r))(n) \,\,{\sf mr}\,\, A(n), \]
If $n < k$ this follows directly from the assumption $\alpha \in S$. In case $n \geq k$ we must show $\varepsilon_n(\lambda x . c) \, \,{\sf mr}\, \, A(n)$, which follows from the assumptions on $\varepsilon_n$ and $r$. \\[2mm]
($ii$) $\forall s \!\in\! S (\forall x [S(s * x) \to P(s * x)] \to P(s))$. Assume $\forall x [S(s * x) \to P(s * x)]$ with $s \!\in\! S$. We must prove $P(s)$, i.e.
\[ s * {\sf ips}_{|s|}(\varepsilon)(q_s) \,\,{\sf mr}\,\, \forall n A(n). \]
Unfolding the definition of ${\sf ips}_{|s|}$ (cf. $(\ref{ips-def-eq})$) this is equivalent to
\[ \underbrace{s * c * {\sf ips}_{|s| + 1}(\varepsilon)(q_{s * c}) \,\,{\sf mr}\,\, \forall n A(n)}_{P(s * c)}. \]
where $c = \varepsilon_{|s|}(\lambda x. q_{s * x}({\sf ips}_{|s| + 1}(\varepsilon)(q_{s * x})))$. Since $s \in S$, by the bar induction hypothesis it is enough to show that $c \,{\sf mr}\, A(|s|)$, i.e.
\[ \varepsilon_{|s|}(\lambda x. q_{s * x}({\sf ips}_{|s| + 1}(\varepsilon)(q_{s * x}))) \,{\sf mr}\, A(|s|) \]
so that also $s * c \in S$. By the assumption on $\varepsilon_{|s|}$, the above follows from
\[ \lambda x. q_{s * x}({\sf ips}_{|s| + 1}(\varepsilon)(q_{s * x})) \,{\sf mr}\, A(|s|) \to R. \]
Finally, by the assumption on $q$ the above is a consequence of
\[ \underbrace{s * x * {\sf ips}_{|s| + 1}(\varepsilon)(q_{s * x}) \,\,{\sf mr}\,\, \forall n A(n)}_{P(s * x)}, \]
for $x \,{\sf mr}\, A(|s|)$, which follows from the (bar) induction hypothesis. \end{proof}
\subsection{Dependent variant of ${\sf ips}$}
\label{IPS-ips}
Consider also the implicitly controlled dependent product of selection functions.
\begin{definition}[$\IPS$] \label{definition-IPS} Let $\varepsilon \colon \Pi_k (\Pi_{i < k} X_i \to J X_k)$. Define the \emph{dependent implicitly controlled product of selection functions} (denoted $\IPS$) as
\begin{equation} \label{IPS-def} \tag{\IPS}
\IPS_s(\varepsilon) \stackrel{J (\Pi_i X_{|s|+i})}{=} \varepsilon_s \ttimesTensor_d (\lambda x^{X_{|s|}} . \IPS_{s * x}(\varepsilon)).
\end{equation}
\end{definition}
Again (similar to Section \ref{EPS-eps}), it is clear that $\IPS$ is a generalisation of ${\sf ips}$. We now show that the proof that $\IPS$ is $T$-definable from ${\sf ips}$ can be easily adapted to show that also ${\sf ips}$ is $T$-equivalent to its dependent variant $\IPS$. In fact, in the case of ${\sf ips}$ and $\IPS$ the proof is slightly simpler since we do not have to worry about the stopping condition and the length function $l$.
\begin{theorem}[${\rm HA}^\omega$] \label{IPS-from-ips} ${\sf ips} \geq_T \IPS$.
\end{theorem}
\begin{proof} Let $\tilde{\varepsilon}_k$ be as defined in Theorem \ref{EPS-from-eps}. Note that the infinite (simple) product of selection functions applied to $\tilde{\varepsilon}$ has type
\eqleft{{\sf ips}_n(\tilde{\varepsilon}) \;\; \colon \;\; J(\Pi_{i \geq n}(\Pi_{k < i} X_k \to X_i)).}
We claim that $\IPS$ can then be defined from ${\sf ips}$ as
\begin{itemize}
\item[$(i)$] $\IPS_s(\varepsilon)(q^{\Pi_j X_{j + |s|} \to R}) \stackrel{\Pi_j X_{j + |s|}}{=} ({\sf ips}_{|s|}(\tilde{\varepsilon})(q^s))^s$
\end{itemize}
where $s \colon \Pi_{k<|s|} X_k$ and $(\cdot)^s$ is as defined in (\ref{s-construction}). We have
\[
\begin{array}{lcl}
\IPS_s(\varepsilon)(q)
& \stackrel{(i)}{=} & ({\sf ips}_{|s|}(\tilde{\varepsilon})(q^{s}))^{s} \\[1mm]
& \stackrel{(\ref{ips-def-eq})}{=} & (d * {\sf ips}_{|s|+1}(\tilde{\varepsilon})((q^{s})_d))^{s} \\[1mm]
& \stackrel{\textup{L}\ref{s-constr-lemma}}{=} & d(s) * ({\sf ips}_{|s|+1}(\tilde{\varepsilon})((q^{s})_d))^{s * d(s)} \\[2mm]
& \stackrel{\textup{L}\ref{s-constr-lemma}}{=} & d(s) * ({\sf ips}_{|s|+1}(\tilde{\varepsilon})((q_{d(s)})^{s * d(s)}))^{s * d(s)} \\[2mm]
& \stackrel{(*)}{=} & c * ({\sf ips}_{|s * c|}(\tilde{\varepsilon})((q_c)^{s * c}))^{s * c} \\[2mm]
& \stackrel{(i)}{=} & c * \IPS_{s * c}(\varepsilon)(q_c)
\end{array}
\]
where, as in Theorem \ref{EPS-from-eps}, we can show that $(*) \; d(s) = c$ for
\eqleft{
\begin{array}{lcl}
d & = & \tilde{\varepsilon}_{|s|}(\lambda f . (q^s)_f({\sf ips}_{|s|+1}(\tilde{\varepsilon})((q^s)_f))) \\[2mm]
c & = & \varepsilon_s(\lambda x . (q_x)(\IPS_{s * x}(\varepsilon)(q_x)).
\end{array}
}
\end{proof}
\subsection{Relation to modified bar recursion}
\label{sec-mbr}
The proof that ${\sf ips}$ interprets full classical analysis, via modified realizability, is very similar to the one given in \cite{BO(02A),BO(05)} that \emph{modified bar recursion} ${\sf MBR}$ interprets full classical analysis. In this section we show how ${\sf MBR}$ corresponds directly to the infinite iteration of a different form of binary product of selection functions. We also show (cf. Section \ref{mbr-def-ips}) that this different product when iterated leads to a form of bar recursion (${\sf MBR}$) which is nevertheless $T$-equivalent to $\IPS$.
\begin{definition} \label{skewed-prod} Given a function $\varepsilon \in (X \to R) \to X \times Y$ and a selection function $\delta \in J Y$ define a selection function $\varepsilon \,\tilde{\ttimesTensor}\, \delta \in J(X \times Y)$ as
\[
(\varepsilon \,\tilde{\ttimesTensor}\, \delta)(p) \stackrel{X \times Y}{=} \varepsilon(\lambda x. p(x, b(x)))
\]
where $b(x) \stackrel{Y}{=} \delta(\lambda y . p(x, y))$. We shall also consider a dependent version $\,\tilde{\ttimesTensor}_d\,$ of the product where $\delta \colon X \to J Y$ and $b(x) = \delta(x, \lambda y . p(x, y))$.
\end{definition}
The above construction shows how a mapping of type $(X \to R) \to X \times Y$ can be extended to a selection function on the product space, given a selection function on $Y$. We shall use this with $X = X_n$ and $Y = \Pi_i X_{i + n + 1}$, so that we obtain a selection function in $J(\Pi_i X_{i+n})$.
\begin{definition}[${\sf mbr}$] Let $\varepsilon_n \colon (X_n \to R) \to \Pi_i X_{i + n}$ and $\varepsilon = (\varepsilon_n)_{n \in \mathbb{N}}$. Define the \emph{iterated skewed product} $\isp$ as
\[ {\sf mbr}_n(\varepsilon) \stackrel{J (\Pi_i X_{i + n})}{=} \varepsilon_n \,\tilde{\ttimesTensor}\, {\sf mbr}_{n+1}(\varepsilon).\]
Unfolding the definition of $\,\tilde{\ttimesTensor}\,$ we have
\begin{equation} \label{mbr-def-eq} \tag{{\sf mbr}}
{\sf mbr}_n(\varepsilon)(q) \stackrel{\Pi_i X_{i + n}}{=} \varepsilon_n(\lambda x . q_x( {\sf mbr}_{n+1}(\varepsilon)(q_x) ).
\end{equation}
Define also the \emph{dependent iterated skewed product} ${\sf MBR}$
\[
{\sf MBR}_s(\varepsilon) \stackrel{J (\Pi_i X_{i + |s|})}{=} \varepsilon_s \,\tilde{\ttimesTensor}_d\, (\lambda x . {\sf MBR}_{s * x}(\varepsilon)),
\]
where in this case $\varepsilon_s \colon (X_{|s|} \to R) \to \Pi_i X_{i + |s|}$. Again, unfolding the definition of $\,\tilde{\ttimesTensor}\,$ we have
\begin{equation} \label{MBR-def-eq-unfold} \tag{{\sf MBR}}
{\sf MBR}_s(\varepsilon)(q) \stackrel{\Pi_i X_{i + |s|}}{=} \varepsilon_s(\lambda x . q_x( {\sf MBR}_{s * x}(\varepsilon)(q_x)).
\end{equation}
We name this $\isp$ and ${\sf MBR}$ because we will show this is essentially \emph{modified bar recursion} as defined in \cite{BO(02A),BO(05)}.
\end{definition}
We think of $\varepsilon$ as a sequence of \emph{skewed selection functions}. The idea is that sometimes a witness for $X_k$ is automatically a witness for all types $X_i$ for $i \geq k$. In such cases, a selection function $\varepsilon_n \colon (X_n \to R) \to X_n$ gives rise to a skewed selection function $\varepsilon_n \colon (X_n \to R) \to \Pi_{i \geq n} X_i$, so that the more intricate product of selection functions (Definition \ref{main-simple}) can be replaced by the simpler product given in Definition \ref{skewed-prod}.
Similarly to $\IPS$ and $\EPS$ (Sections \ref{EPS-eps} and \ref{IPS-ips}), we now show that the \emph{simple} iterated skewed product ${\sf mbr}$ is $T$-equivalent to its \emph{dependent} variant ${\sf MBR}$. Given a sequence of types $X_i$ let us define the new sequence
\[
Y_j \equiv \mathbb{B} \times (\Pi_{i < j} X_i \to \Pi_k X_{k + j}).
\]
The intuition for the construction below is the same as the one used to show that ${\sf eps}$ $T$-defines $\EPS$ (Theorem \ref{EPS-from-eps}), except that here we need an extra boolean flag as the whole result of the skewed selection function will be returned on the first position of the output. The flag is used so that functions querying such sequences can know which are proper values and which are dummy values. Let us first define the construction $(\cdot)^{[s]} \colon \Pi_i Y_{i + |s|} \to \Pi_i X_{i + |s|}$ that given $\alpha \colon \Pi_i Y_{i + |s|}$ is defined as
\eqleft{
\alpha^{[s]}(i) \stackrel{X_{i + |s|}}{=}
\begin{cases}
g(s * \initSeg{\alpha^{[s]}}{n})(i-n) & \text{if $\exists n \leq i (\alpha(n) = \pair{{\sf tt}, g})$} \\[2mm]
{\bf 0} & \text{if $\forall n \leq i (\alpha(n) = \pair{{\sf ff}, \ldots})$}
\end{cases}
}
where in the first case $n$ is the greatest $n \leq i$ such that $\alpha(n)$ is of the form $\pair{{\sf tt}, g}$. Finally, for $q \colon \Pi_i X_{i + |s|} \to R$ we define
\eqleft{q^{[s]}(\alpha) \stackrel{R}{=} q(\alpha^{[s]})}
so $q^{[s]} \colon \Pi_i Y_{i + |s|} \to R$; and for $x \colon X_j$ define $\hat x \colon Y_j$ as
\eqleft{\hat x = \pair{{\sf tt}, \lambda s^{\Pi_{i < j} X_i} . \pair{x^{X_j}, {\bf 0}^{X_{j+1}}, {\bf 0}^{X_{j+2}}, \ldots}}.}
\begin{lemma}[${\rm HA}^\omega$] \label{isp-ISP-lemma} $(q^{[s]})_{\hat x} = (q_x)^{[s * x]}$.
\end{lemma}
\begin{proof} By course-of-values induction on $i$ we have $(\hat x * \alpha)^{[s]}(i) = (x * \alpha^{[s * x]})(i)$. Hence $(\hat x * \alpha)^{[s]} = x * \alpha^{[s * x]}$ and
\[ (q^{[s]})_{\hat x}(\alpha) = q((\hat x * \alpha)^{[s]}) = q_x(\alpha^{[s * x]}) = (q_x)^{[s * x]}(\alpha). \]
\end{proof}
\begin{theorem}[${\rm HA}^\omega$] $\isp \geq_T {\sf MBR}$.
\end{theorem}
\begin{proof} In order to define ${\sf MBR}$ of type $(X_i, R)$ we use ${\sf mbr}$ of type $(Y_j, R)$. For $\varepsilon_s \colon (X_j \to R) \to \Pi_k X_{k + j}$, where $s \colon \Pi_{i < j} X_i$, define $\tilde \varepsilon_j \colon (Y_j \to R) \to \Pi_k Y_{k + j}$ as
\begin{itemize}
\item[$(i)$]
$\tilde \varepsilon_j(P^{Y_j \to R})(k) \stackrel{Y_{k + j}}{=}
\begin{cases}
\pair{{\sf tt}, \lambda t^{\Pi_{i < j} X_i} . \varepsilon_t(\lambda x^{X_{j}}. P(\hat x))} & \text{if $k = 0$,} \\[2mm]
\pair{{\sf ff}, {\bf 0}^{\Pi_{i < j + k} X_i \to \Pi_i X_{i + j + k}}} & \text{if $k > 0$}.
\end{cases}$
\end{itemize}
By the definition of $(\cdot)^{[s]}$ and definition $(i)$ it is easy to check that
\begin{itemize}
\item[$(ii)$] $(\tilde \varepsilon_{|s|}(P))^{[s]} = \varepsilon_s(\lambda x . P(\hat x))$.
\end{itemize}
We claim that ${\sf MBR}$ can be defined from $\isp$ as
\begin{itemize}
\item[$(iii)$] ${\sf MBR}_s(\varepsilon)(q^{\Pi_i X_{i + |s|} \to R}) \stackrel{\Pi_i X_{i + |s|}}{=} (\isp_{|s|}(\tilde \varepsilon)(q^{[s]}))^{[s]}$.
\end{itemize}
We have
\eqleft{
\begin{array}{lcl}
{\sf MBR}_s(\varepsilon)(q)
& \stackrel{(iii)}{=} & (\isp_{|s|}(\tilde \varepsilon)(q^{[s]}))^{[s]} \\[1mm]
& \stackrel{(\ref{mbr-def-eq})}{=} & \big( \tilde \varepsilon_{|s|}(\lambda f^{Y_{|s|}} . (q^{[s]})_f (\isp_{|s| + 1}(\tilde \varepsilon)((q^{[s]})_f))) \big)^{[s]} \\[1mm]
& \stackrel{(ii)}{=} & \varepsilon_s(\lambda x^{X_{|s|}} . (q^{[s]})_{\hat x} (\isp_{|s| + 1}(\tilde \varepsilon)((q^{[s]})_{\hat x}))) \\[1mm]
& \stackrel{\textup{L}\ref{isp-ISP-lemma}}{=} & \varepsilon_s (\lambda x^{X_{|s|}}. q_x(\big(\isp_{|s * x|}(\tilde \varepsilon)((q_x)^{[s * x]})\big)^{[s * x]})) \\[1mm]
& \stackrel{(iii)}{=} & \varepsilon_s (\lambda x^{X_{|s|}}. q_x({\sf MBR}_{s * x}(\varepsilon)(q_x))).
\end{array}
}
\end{proof}
We now show that a slight generalisation of modified bar recursion \cite{BO(02A),BO(05)} is $T$-equivalent to the iterated product of skewed selection functions. Define ${\sf MBR}'$ as
\begin{equation} \label{MBR-var-eq} \tag{{\sf MBR}'}
{\sf MBR}'_s(\varepsilon)(q) \stackrel{\Pi_i X_i}{=} s * \varepsilon_s(\lambda x^{X_{|s|}} . q({\sf MBR}'_{s * x}(\varepsilon)(q)))
\end{equation}
where $q \colon \Pi_i X_i \to R$ and $\varepsilon_s \colon (X_{|s|} \to R) \to \Pi_i X_{|s|+i}$. ${\sf MBR}'$ is a generalisation of modified bar recursion (as defined in \cite{BO(02A)}, cf. lemma 2) to sequence types. If all $X_i = X$ we have precisely the definition given in \cite{BO(02A),BO(05)}.
\begin{theorem}[${\rm HA}^\omega + {\rm BI} + {\rm CONT}$] \label{mbr} ${\sf MBR} =_T {\sf MBR}'$.
\end{theorem}
\begin{proof} For one direction, let $q \colon \Pi_i X_i \to R$ and $s \colon \Pi_{i < n} X_i$ and define
\begin{itemize}
\item[$(i)$] ${\sf MBR}'_s(\varepsilon)(q) \stackrel{\Pi_i X_i}{=} s * {\sf MBR}_s(\varepsilon)(q_s)$.
\end{itemize}
Unfolding definitions we have
\eqleft{
\begin{array}{lcl}
{\sf MBR}'_s(\varepsilon)(q) & \stackrel{(i)}{=} & s * {\sf MBR}_s(\varepsilon)(q_s) \\[1mm]
& \stackrel{(\ref{MBR-def-eq-unfold})}{=} & s * \varepsilon_s(\lambda x^{X_{|s|}} . q_{s*x}({\sf MBR}_{s*x}(\varepsilon)(q_{s*x}))) \\[2mm]
& = & s * \varepsilon_s(\lambda x^{X_{|s|}} . q(s * x * {\sf MBR}_{s*x}(\varepsilon)(q_{s*x}))) \\[1mm]
& \stackrel{(i)}{=} & s * \varepsilon_s(\lambda x^{X_{|s|}} . q({\sf MBR}'_{s*x}(\varepsilon)(q))).
\end{array}
}
For the other direction, let $q \colon \Pi_i X_{i + |s|} \to R$. Define
\begin{itemize}
\item[$(ii)$] ${\sf MBR}_s(\varepsilon)(q) \stackrel{\Pi_i X_{i + |s|}}{=} {\sf MBR}'_{\pair{\,}}(\lambda t . \varepsilon_{s * t})(q)$.
\end{itemize}
We then have
\eqleft{
\begin{array}{lcl}
{\sf MBR}_s(\varepsilon)(q) & \stackrel{(ii)}{=} & {\sf MBR}'_{\pair{\,}}(\lambda t . \varepsilon_{s * t})(q) \\[1mm]
& \stackrel{(\ref{MBR-var-eq})}{=} & \varepsilon_s(\lambda x^{X_{|s|}} . q({\sf MBR}'_{x}(\lambda t . \varepsilon_{s * t})(q))) \\[1mm]
& \stackrel{(*)}{=} & \varepsilon_s(\lambda x^{X_{|s|}} . q_x({\sf MBR}'_{\pair{\,}}(\lambda t . \varepsilon_{s * x * t})(q_x))) \\[1mm]
& \stackrel{(ii)}{=} & \varepsilon_s(\lambda x^{X_{|s|}} . q_x({\sf MBR}_{s*x}(\varepsilon)(q_x))),
\end{array}
}
where $(*) \; {\sf MBR}'_{x * r}(\lambda t . \varepsilon_{s * t})(q) = x * {\sf MBR}'_{r}(\lambda t . \varepsilon_{s * x * t})(q_x)$ can be proven by bar induction on the sequence $r$, assuming continuity of $q$ (cf. Lemma \ref{epq-eps-lemma}).
\end{proof}
\begin{corollary} Gandy's functional $\Gamma$ is $T$-equivalent to ${\sf MBR}$ with $X_i = \mathbb{N}$ for all $i \in \mathbb{N}$.
\end{corollary}
\begin{proof} It has been shown in \cite{BO(05)} that the $\Gamma$ functional is $T$-equivalent to ${\sf MBR}'$ of lowest type. It remains to observe that the equivalence of Theorem \ref{mbr} respects the types. \end{proof}
\begin{question} It should be mentioned that in \cite{Berardi(98)} yet another form of bar recursion is used for the interpretation of the double negation shift (although they also use modified bar recursion when dealing with dependent choice). We refer to this different bar recursion as the ${\sf bbc}$ functional. Thomas Powell \cite{Powell(2014A)} has recently shown that ${\sf bbc}$ is $T$-equivalent to $\IPS$ (see also \cite{Berger04}).
\end{question}
\section{Further Inter-definability Results}
\label{further-def}
In this section we prove three further inter-definability results, namely ${\sf ips} \geq {\sf mbr}$, ${\sf MBR} \geq \IPS$ and $\IPS \geq \EPQ$.
\begin{theorem}[${\rm HA}^\omega$] ${\sf ips} \geq_T {\sf mbr}$.
\end{theorem}
\begin{proof} Given a type $X$ let us denote by $X'$ the type $\mathbb{B} \times X$. In order to define ${\sf mbr}$ of type $(X_i, R)$ we use ${\sf ips}$ of type $(\Pi_j X_{i + j}', R)$. The main idea for the construction is to turn a skewed selection function into a proper selection function as follows. Given $\varepsilon_i \colon (X_i \to R) \to \Pi_j X_{i+j}$ we define $\tilde{\varepsilon}_i \colon J (\Pi_j X_{i+j}')$ as
\begin{itemize}
\item[$(i)$] $\tilde{\varepsilon}_i(f^{\Pi_j X_{i+j}' \to R}) \stackrel{\Pi_j X_{i+j}'}{=} \lambda j . \pair{{\sf ff}, \varepsilon_i(\lambda x^{X_i} . f(\hat x))(j)}$,
\end{itemize}
where
\eqleft{
\hat x(j) =
\left\{
\begin{array}{ll}
\pair{{\sf tt}, x^{X_i}} & {\rm if} \; j = 0 \\[2mm]
\pair{{\sf tt}, {\bf 0}^{X_{i+j}}} & {\rm if} \; j > 0.
\end{array}
\right.
}
Intuitively, the booleans $\mathbb{B} = \{ {\sf tt}, {\sf ff} \}$ are used to distinguish between values returned by $\varepsilon_i$ and those values $\hat x$ passed into a recursive call. \\[1mm]
Given $\alpha \colon \Pi_k (\Pi_j X'_{j + i + k})$ we define $\tilde \alpha \colon \Pi_j X_{j + i}$ as
\eqleft{\tilde{\alpha}(j) \stackrel{X_{j + i}}{=}
\left\{
\begin{array}{ll}
(\alpha(j)(0))_1 & {\rm if} \; \forall k \!<\! j \, ((\alpha(k)(0))_0 = {\sf tt} \\[2mm]
(\alpha(k)(j-k))_1 & {\rm if} \; \exists k \!<\! j \, ((\alpha(k)(0))_0 = {\sf ff},
\end{array}
\right.
}
where $k = \mu k \!<\! j \, (\alpha(k)(0))_0 = {\sf ff}$. The construction $\tilde \alpha$ receives as input a matrix $\alpha \colon \Pi_i \Pi_{j \geq i} X_j'$ and produces a sequence $\Pi_j X_j$ as follows: As long as the value $\alpha(j)$ is some $\hat x$ (boolean flag will be ${\sf tt}$) we filter out the $x$; once we reach a value returned by an $\varepsilon_k$ (boolean will be ${\sf ff}$) then we return the whole sequence returned by the skewed selection function $\varepsilon_k$. Hence, given a $q \colon \Pi_j X_{i+j} \to R$ we define $\tilde q \colon \Pi_k (\Pi_j X'_{i+k+j}) \to R$ as $\tilde q(\alpha) = q(\tilde \alpha)$ where, Clearly
\begin{itemize}
\item[$(ii)$] $\tilde{q}_{\hat x}(\beta) = \widetilde{(q_x)}(\beta)$.
\end{itemize}
We claim that $\isp$ can be defined as
\begin{itemize}
\item[$(iii)$] $\isp_i(\varepsilon)(q) \stackrel{\Pi_j X_{i+j}}{=} (\big({\sf ips}_i(\tilde{\varepsilon})(\tilde{q})\big)^{\Pi_k \Pi_j X'_{i+k+j}}(0))^1$
\end{itemize}
where $\varepsilon_i \colon (X_i \to R) \to \Pi_j X_{i+j}$ and $q \colon \Pi_j X_{i+j} \to R$. Recall that given a sequence $\beta \colon \Pi_{i < n} (X_i \times Y_i)$ we write $\beta^1 \colon \Pi_{i < n} Y_i$ for the projection of the sequence on the second coordinates. We have
\eqleft{
\begin{array}{lcl}
\isp_i(\varepsilon)(q)
& \stackrel{(iii)}{=} & ({\sf ips}_i(\tilde{\varepsilon})(\tilde{q})(0))^1 \\[1mm]
& \stackrel{(\ref{ips-def-eq})}{=} & (\tilde{\varepsilon}_i(\lambda \alpha^{\Pi_j X'_{i+j}} . \tilde{q}_{\alpha}({\sf ips}_{i+1}(\tilde{\varepsilon})(\tilde{q}_{\alpha}))))^1 \\[1mm]
& \stackrel{(i)}{=} & \varepsilon_i(\lambda x^{X_i} . \tilde{q}_{\hat x}({\sf ips}_{i+1}(\tilde{\varepsilon})(\tilde{q}_{\hat x}))) \\[1mm]
& \stackrel{(ii)}{=} & \varepsilon_i(\lambda x^{X_i} . \widetilde{(q_x)}({\sf ips}_{i+1}(\tilde{\varepsilon})(\widetilde{(q_x)}))) \\[1mm]
& \stackrel{(iv)}{=} & \varepsilon_i(\lambda x^{X_i} . q_x(({\sf ips}_{i+1}(\tilde{\varepsilon})(\widetilde{(q_x)})(0))^1)) \\[1mm]
& \stackrel{(iii)}{=} & \varepsilon_i(\lambda x^{X_i} . q_x({\sf mbr}_{i+1}(\varepsilon)(q_x)))
\end{array}
}
using that
\begin{itemize}
\item[$(iv)$] $\tilde \beta = ({\sf ips}_{i+1}(\tilde{\varepsilon})(\widetilde{(q_x)})(0))^1$, for $\beta = {\sf ips}_{i+1}(\tilde{\varepsilon})(\widetilde{(q_x)})$.
\end{itemize}
\end{proof}
\subsection{${\sf MBR} \geq \IPS$}
\label{mbr-def-ips}
We now show that the implicitly controlled dependent product of selection functions $\IPS$ is $T$-definable from (and hence $T$-equivalent to) modified bar recursion ${\sf MBR}$. Since in our proof we need to work with infinite sequences of finite sequences (of arbitrary length), the use of the infinite product type $\Pi_i X_i$ in here is unhelpful. The problem is that keeping track of indices would imply introducing the $\Sigma$ type to record the length of each finite sequence. Although this can be done, it would require much more of dependent type theory than we have assumed so far. Hence, for this section only we work with selection functions of a fixed type $(X \to R) \to X$. Similarly, skewed selection functions will have type $(Y \to R) \to Y^\mathbb{N}$.
Let $X^+$ denote non-empty finite sequences of elements of type $X$. We make use of the following two mappings $G \colon X \to X^+$ and $F \colon (X^+)^\mathbb{N} \to X^\mathbb{N}$ where
\eqleft{
\begin{array}{lcl}
G(x) & = & \langle x \rangle \\[1mm]
F(\alpha) & = & \mbox{concatenation of non-empty finite sequences $\alpha(i)$'s.}
\end{array}
}
For the definition of $F$ it is important that we are considering non-empty sequences, as otherwise such concatenation operation would not be defined in general. Consider two variants $G^* \colon X^* \to (X^+)^*$ and $F^* \colon (X^+)^* \to X^*$, where $G^*$ is the function $G$ applied pointwise to a given finite sequence, and $F^*$ is the concatenation of a finite sequence of non-empty finite sequences.
\begin{lemma} \label{mbr-ips-lemmaA} The following can be easily verified:
\begin{itemize}
\item[($i$)] $F(\lambda i . G(v_i))(i) = v_i$ and $F^*(G^*(s)) = s$, where $v_i \colon X$ and $s \colon X^*$. \\[-2mm]
\item[($ii$)] $F^*(s * t) = F^*(s) * F^*(t)$, where $s, t \colon (X^+)^*$. \\[-2mm]
\item[($iii$)] $F^*(G^*(s) * t) = s * F^*(t)$, where $s \colon X^*$ and $t \colon (X^+)^*$.
\end{itemize}
\end{lemma}
Given selection functions $\varepsilon_s \colon J_R X$ define, by course-of-values, skewed selection functions of type
\eqleft{\nu_r \colon (X^+ \to R) \to (X^+)^\mathbb{N},}
where $r \colon (X^+)^*$, as
\eqleft{\nu_r(P^{X^+ \to R})(i) \stackrel{X^+}{=} G(\varepsilon_{F^*(r * t^i)}(\lambda x^X . P(\langle (F^* t^i) * x \rangle)))}
with $t^i \stackrel{(X^+)^*}{=} \initSeg{\nu_r(P^{X^+ \to R})}{i}$.
\begin{lemma} \label{mbr-ips-lemmaB} If $F^*(r) = F^*(r')$ then $\nu_{r}(P)(i) = \nu_{r'}(P)(i)$.
\end{lemma}
\begin{proof} Directly from Lemma \ref{mbr-ips-lemmaA} $(ii)$ since $F^*(r * t^i) = F^*(r) * F^*(t^i)$.
\end{proof}
Now, given a functional $q \colon X^\mathbb{N} \to R$ define $\tilde q \colon (X^+)^\mathbb{N} \to R$ as
\eqleft{\tilde q (\alpha) \stackrel{R}{=} q(F \alpha).}
Again, it is easy to see that:
\begin{lemma} \label{mbr-ips-lemmaC} $(\tilde q)_{\langle s \rangle} \stackrel{(X^+)^\mathbb{N}}{=} \widetilde{(q_s)}$, where $s \colon X^+$.
\end{lemma}
\begin{proof} $(\tilde q)_{\langle s \rangle}(\alpha) = q(F(\langle s \rangle * \alpha)) = q(F^*(\langle s \rangle) * F(\alpha))) = q_s(F \alpha) = \widetilde{(q_s)}(\alpha).$
\end{proof}
\begin{lemma}[${\rm HA}^\omega + {\rm BI} + {\rm CONT}$] \label{mbr-ips-lemmaD} If $F^*(s) = F^*(s')$ then
\[ {\sf MBR}_{s}(\nu)(\tilde q) = {\sf MBR}_{s'}(\nu)(\tilde q). \]
\end{lemma}
\begin{proof} Define the predicate
\eqleft{P(t^{(X^+)^*}) \,\equiv\, F^*(s * t) = F^*(s' * t) \to {\sf MBR}_{s * t}(\nu)(\tilde{q}_t) = {\sf MBR}_{s' * t}(\nu)(\tilde{q}_t).}
We show $P(\langle \, \rangle)$ by bar induction ${\rm BI}$ (assuming ${\rm CONT}$). \\[1mm]
$(i)$ $\forall \alpha \exists k P(\initSeg{\alpha}{k})$. Given $\alpha$, by ${\rm CONT}$ let $k$ be such that $\tilde{q}_{\initSeg{\alpha}{k}}$ is a constant function, say $\tilde{q}_{\initSeg{\alpha}{k}}(\beta) = r$, for all $\beta$. Assuming $F^*(s * \initSeg{\alpha}{k}) = F^*(s' * \initSeg{\alpha}{k})$
\[
\begin{array}{lcl}
{\sf MBR}_{s * \initSeg{\alpha}{k}}(\nu)(\tilde{q}_{\initSeg{\alpha}{k}})
& = & \nu_{s * \initSeg{\alpha}{k}}(\lambda y . \tilde{q}_{\initSeg{\alpha}{k} * y}({\sf MBR}_{s * \initSeg{\alpha}{k} * y}(\nu)(\tilde{q}_{\initSeg{\alpha}{k} * y}))) \\[2mm]
& = & \nu_{s * \initSeg{\alpha}{k}}(\lambda y . r) \\[1mm]
& \stackrel{\textup{L}\ref{mbr-ips-lemmaB}}{=} & \nu_{s' * \initSeg{\alpha}{k}}(\lambda y . r) \\[2mm]
& = & \nu_{s' * \initSeg{\alpha}{k}}(\lambda y . \tilde{q}_{\initSeg{\alpha}{k} * y}({\sf MBR}_{s' * \initSeg{\alpha}{k} * y}(\nu)(\tilde{q}_{\initSeg{\alpha}{k} * y}))) \\[2mm]
& = & {\sf MBR}_{s' * \initSeg{\alpha}{k}}(\nu)(\tilde{q}_{\initSeg{\alpha}{k}}).
\end{array}
\]
$(ii)$ $\forall t (\forall x P(t * x) \to P(t))$. Let $t$ be such that $\forall x P(t * x)$. Assuming $F^*(s * t) = F^*(s' * t)$, and noting that this implies $F^*(s * t * y) = F^*(s' * t * y)$, we have
\eqleft{
\begin{array}{lcl}
{\sf MBR}_{s * t}(\nu)(\tilde{q}_t)
& = & \nu_{s * t}(\lambda y . \tilde{q}_{t * y}({\sf MBR}_{s * t * y}(\nu)(\tilde{q}_{t * y}))) \\[1mm]
& \stackrel{(\textup{IH})}{=} & \nu_{s * t}(\lambda y . \tilde{q}_{t * y}({\sf MBR}_{s' * t * y}(\nu)(\tilde{q}_{t * y}))) \\[1mm]
& \stackrel{\textup{L}\ref{mbr-ips-lemmaB}}{=} & \nu_{s' * t}(\lambda y . \tilde{q}_{t * y}({\sf MBR}_{s' * t * y}(\nu)(\tilde{q}_{t * y}))) \\[2mm]
& = & {\sf MBR}_{s' * t}(\nu)(\tilde{q}_t).
\end{array}
}
\end{proof}
We can now show that $\IPS$ of type $(X, R)$ is $T$-definable from ${\sf MBR}$ of type $(X^+, R)$.
\begin{theorem}[${\rm HA}^\omega + {\rm BI} + {\rm CONT}$] \label{mbr-cbr} ${\sf MBR} \geq_T \IPS$.
\end{theorem}
\begin{proof} Define $\IPS$ from ${\sf MBR}$ as
\eqleft{\IPS_s(\varepsilon)(q) \stackrel{X^\mathbb{N}}{=} F({\sf MBR}_{G^*(s)}(\nu)(\tilde q))}
where $\nu$ (and $\tilde q$) is defined from $\varepsilon$ ($q$, respectively) as above. We show that $\IPS$ as defined above satisfies its defining equation.
Let
\begin{itemize}
\item $t^i = \initSeg{\nu_{G^*s}(\lambda y . \tilde{q}_y({\sf MBR}_{(G^*s) * y}(\nu)(\tilde{q}_y)))}{i}$
\item $r^i = \initSeg{\IPS_s(\varepsilon)(q)}{i}$.
\end{itemize}
We first show that $(\dagger)~F^*(t^i) = r^i$. By course-of-values assume $F^*(t^j) = r^j$ for $j < i$, then
\[
\begin{array}{lcl}
F^*(t^i)(j)
& \stackrel{X^+}{=} & F^*(\initSeg{\nu_{G^*s}(\lambda y . \tilde{q}_y({\sf MBR}_{(G^*s) * y}(\nu)(\tilde{q}_y)))}{i})(j) \\[1mm]
& \stackrel{\textup{L}\ref{mbr-ips-lemmaA}(i)}{=} & \varepsilon_{F^*((G^* s) * t^j)}(\lambda x . \tilde{q}_{\langle (F^* t^j) * x \rangle}({\sf MBR}_{(G^* s) * \langle (F^* t^j) * x \rangle}(\nu)(\tilde{q}_{\langle (F^* t^j) * x \rangle}))) \\[1mm]
& \stackrel{\textup{L}\ref{mbr-ips-lemmaA}(iii)}{=} & \varepsilon_{s * F^*(t^j)}(\lambda x . \tilde{q}_{\langle (F^* t^j) * x \rangle}({\sf MBR}_{(G^* s) * \langle (F^* t^j) * x \rangle}(\nu)(\tilde{q}_{\langle (F^* t^j) * x \rangle}))) \\[1mm]
& \stackrel{\textup{L}\ref{mbr-ips-lemmaC}}{=} & \varepsilon_{s * F^*(t^j)}(\lambda x . \widetilde{q_{(F^* t^j) * x}}({\sf MBR}_{(G^* s) * \langle (F^* t^j) * x \rangle}(\nu)(\widetilde{q_{(F^* t^j) * x}}))) \\[1mm]
& \stackrel{\textup{(IH)}}{=} & \varepsilon_{s * r^j}(\lambda x . \widetilde{q_{r^j * x}}({\sf MBR}_{(G^* s) * \langle r^j * x \rangle}(\nu)(\widetilde{q_{r^j * x}}))) \\[1mm]
& \stackrel{\textup{L}\ref{mbr-ips-lemmaD}}{=} & \varepsilon_{s * r^j}(\lambda x . \widetilde{q_{r^j * x}}({\sf MBR}_{G^*( s * r^j * x)}(\nu)(\widetilde{q_{r^j * x}}))) \\[2mm]
& = & \varepsilon_{s * r^j}(\lambda x . q_{r^j * x}(F({\sf MBR}_{G^*(s * r^j * x)}(\nu)(\widetilde{q_{r^j * x}})))) \\[2mm]
& = & \varepsilon_{s * r^j}(\lambda x . q_{r^j * x}(\IPS_{s * r^j * x}(\varepsilon)(q_{r^j * x}))) \\[2mm]
& = & (r^i)(j).
\end{array}
\]
We then have
\[
\begin{array}{lcl}
\IPS_s(\varepsilon)(q)(i)
& \stackrel{X}{=} & F({\sf MBR}_{G^*(s)}(\nu)(\tilde q))(i) \\[2mm]
& = & F(\nu_{G^* s} (\lambda y^{X^+} . \tilde{q}_y({\sf MBR}_{(G^*s) * y}(\nu)(\tilde{q}_y))))(i) \\[1mm]
& \stackrel{\textup{L}\ref{mbr-ips-lemmaA}(i)}{=} & \varepsilon_{F^*((G^* s) * t^i)} (\lambda x . \tilde{q}_{\langle (F^* t^i) * x \rangle}({\sf MBR}_{(G^* s) * \langle (F^* t^i) * x \rangle}(\nu)(\tilde{q}_{\langle (F^* t^i) * x \rangle})))) \\[1mm]
& \stackrel{\textup{L}\ref{mbr-ips-lemmaA}(ii)}{=} & \varepsilon_{s * F^*(t^i)} (\lambda x . \tilde{q}_{\langle (F^* t^i) * x \rangle}({\sf MBR}_{(G^* s) * \langle (F^* t^i) * x \rangle}(\nu)(\tilde{q}_{\langle (F^* t^i) * x \rangle})))) \\[1mm]
& \stackrel{(\dagger)}{=} & \varepsilon_{s * r^i} (\lambda x . \tilde{q}_{\langle r^i * x \rangle}({\sf MBR}_{(G^* s) * \langle r^i * x \rangle}(\nu)(\tilde{q}_{\langle r^i * x \rangle})))) \\[1mm]
& \stackrel{\textup{L}\ref{mbr-ips-lemmaC}}{=} & \varepsilon_{s * r^i} (\lambda x . \widetilde{q_{r^i * x}}({\sf MBR}_{(G^* s) * \langle r^i * x \rangle}(\nu)(\widetilde{q_{r^i * x}})))) \\[2mm]
& = & \varepsilon_{s * r^i} (\lambda x . q_{r^i * x}(F({\sf MBR}_{(G^* s) * \langle r^i * x \rangle}(\nu)(\widetilde{q_{r^i * x}})))) \\[1mm]
& \stackrel{\textup{L}\ref{mbr-ips-lemmaD}}{=} & \varepsilon_{s * r^i} (\lambda x . q_{r^i * x}(F({\sf MBR}_{G^* (s * r^i * x)}(\nu)(\widetilde{q_{r^i * x}})))) \\[2mm]
& = & \varepsilon_{s * r^i} (\lambda x . q_{r^i * x}(\IPS_{s * r^i * x}(\varepsilon)(q_{r^i * x}))).
\end{array}
\]
\end{proof}
\subsection{$\IPS \geq \EPQ$}
\label{ips-def-epq}
It has been shown in \cite{BO(05)} that ${\sf BR}$ is $T$-definable from modified bar recursion. Here we simplify that construction and use it to show that $\EPQ$ is $T$-definable from $\IPS$. Moreover, we make explicit the assumption ${\rm SPEC}$ which is used in \cite{BO(05)}. First we prove that (the totalisation of) \emph{Spector's search functional} is definable in G\"odel's system $T$.
\begin{lemma}[${\rm HA}^\omega$] \label{spec-search} The totalisation of \emph{Spector's search functional}
\eqleft{\mu_{{\sf sc}}(\omega)(\alpha) = {\sf least} \, n (\omega(\initSegZ{\alpha}{n}) < n)}
is $T$-definable. More precisely, there exists a term $\chi$ in G\"odel's system $T$ such that the following is provable in ${\rm HA}^\omega$
\eqleft{\exists n (\omega(\initSegZ{\alpha}{n}) < n) \to (\omega(\initSegZ{\alpha}{\chi \omega \alpha}) < \chi \omega \alpha \wedge \forall i < \chi \omega \alpha (\omega(\initSegZ{\alpha}{i}) \geq i)).}
In particular,
\eqleft{{\rm HA}^\omega + {\rm SPEC} \vdash (\omega(\initSegZ{\alpha}{n}) < n) \wedge \forall i < n (\omega(\initSegZ{\alpha}{i}) \geq i)}
where $n = \chi \omega \alpha$.
\end{lemma}
\begin{proof} We show how the unbounded search in $\mu_{{\sf sc}}$ can be turned into a bounded search. Abbreviate $A_n(\omega, \alpha) = (\omega(\initSegZ{\alpha}{n}) < n)$. Consider the following construction, given $\alpha \colon \Pi_i X_i$ define $\alpha^\omega \colon \Pi_i X_i$ as
\eqleft{\alpha^\omega(i) =
\left\{
\begin{array}{ll}
{\bf 0}^{X_i} & {\rm if} \; \exists k \!\leq\! i+1 \, A_k(\omega, \alpha) \\[2mm]
\alpha(i) \quad & {\rm otherwise}.
\end{array}
\right.
}
Assume $\exists n (\omega(\initSegZ{\alpha}{n}) < n)$. Let $n$ be the least number such that $A_n(\omega, \alpha)$ holds. Then it is easy to see that $\alpha^\omega = \initSegZ{\alpha}{n-1}$. Because $n$ is least, we must have that $\omega(\alpha^\omega) \geq n-1$, and hence $n \leq \omega(\alpha^\omega) + 1$. Therefore, $\omega(\alpha^\omega) + 1$ serves as an upper bound on the search $\mu_{\sf sc}$, i.e. $\chi \omega \alpha = \mu n \leq \omega(\alpha^\omega) + 1 \; (\omega(\initSegZ{\alpha}{n}) < n)$. \end{proof}
The construction above shows that Spector's search functional can be made total in system $T$, so that whenever it is well-defined for inputs $\omega$ and $\alpha$ the term $\chi$ indeed computes the correct value.
\begin{theorem} \label{IPS-EPQ-def} $\EPQ$ is $T$-definable from $\IPS$ over ${\rm HA}^\omega + {\rm SPEC}$. However, $\EPQ$ is not $T$-definable from $\IPS$, even over ${\rm HA}^\omega + {\rm BI} + {\rm CONT}$.
\end{theorem}
\begin{proof} First, note that combining the results above we have the equivalences $\IPS =_T {\sf MBR}$ and $\EPQ =_T {\sf BR}$, over ${\rm HA}^\omega + {\rm BI} + {\rm CONT}$. Hence, that $\IPS$ is not $T$-definable from $\EPQ$, even over ${\rm HA}^\omega + {\rm BI} + {\rm CONT}$, follows from fact that ${\sf MBR}$ is not S1-S9 computable in the model of total continuous functions while ${\sf BR}$ is (see \cite{BO(05)}). \\[1mm]
In order to show $\IPS \geq_T \EPQ$ we use the search operator $\chi$ of the above lemma. Define
\eqleft{\chi^{+k}(\omega)(\alpha) = \mu i \leq \chi(\lambda \beta . \omega(\beta) - k)(\alpha) \; (\omega(\initSeg{\alpha}{i}) < i + k) ,}
where $\omega(\beta) - k$ is the cut-off subtraction. By Lemma \ref{spec-search} we have that $n = \chi^{+k}(\omega)(\alpha)$ is the least such that $\omega(\initSeg{\alpha}{n}) - k < n$. But since $\omega(\initSeg{\alpha}{i}) - k < i$ implies $\omega(\initSeg{\alpha}{i}) < i + k$, we have that, provably in ${\rm HA}^\omega + {\rm SPEC}$,
\begin{itemize}
\item[$(i)$] $\omega(\initSeg{\alpha}{n}) < n + k$ and $\forall i < n (\omega(\initSeg{\alpha}{i}) \geq i + k)$, for $n = \chi^{+k}(\omega)(\alpha)$.
\end{itemize}
Let $\psi_s \colon K_R X_{|s|}$ be a given family of quantifiers. We first turn each quantifier $\psi_s \colon K_R X_{|s|}$, where $s \colon \Pi_{i < |s|} X_i$, into a a selection function $\tilde{\psi}_t$ of type $J_R (X_{|t|} \uplus R)$ as\footnote{We are here making use of the sum type $X \uplus Y$, which can be implemented as $\mathbb{B} \times X \times Y$, since we assume all types are inhabited, with ${\sf inj}_X \colon X \to X \uplus Y$ and ${\sf inj}_Y \colon Y \to X \uplus Y$ the standard injections.}
\begin{itemize}
\item[$(ii)$] $\tilde{\psi}_t(F^{(X_{|t|} \uplus R) \to R}) \stackrel{X_{|t|} \uplus R}{=} {\sf inj}_R (\psi_{\check{t}}(\lambda x^{X_{|t|}} . F({\sf inj}_{X_{|t|}} x)))$
\end{itemize}
where $t \colon \Pi_{i < |t|} (X_i \uplus R)$, and $\check{(\cdot)} \colon \Pi_{i < n} (X_i \uplus R) \to \Pi_{i < n} X_i$ is defined as
\eqleft{(\check{s})_i \stackrel{X_i}{=}
\left\{
\begin{array}{ll}
x_i & {\rm if} \; s_i = {\sf inj}_{X_i}(x_i) \\[2mm]
{\bf 0}^{X_i} \quad & {\rm otherwise}.
\end{array}
\right.
}
We will also make use of the dual operation $\tilde{(\cdot)} \colon \Pi_{i < n} X_i \to \Pi_{i < n} (X_i \uplus R)$ that maps ${\sf inj}_X(\cdot)$ pointwise on a given sequence. Clearly we have
\begin{itemize}
\item[$(iii)$] $\check{\tilde{s}} = s$ and $\tilde s * {\sf inj}_{X_{|s|}}(x) = \widetilde{s*x}$, for $s \colon \Pi_{i < n} X_i$.
\end{itemize}
Note that both construction $\check{(\cdot)}$ and $\tilde{(\cdot)}$ can similarly defined on infinite sequences as well. Hence, given $s \colon \Pi_{i < k} X_i$ and $q \colon \Pi_i X_{i + k} \to R$ and $l \colon R \to \mathbb{N}$ let us define the function $q^{l, s} \colon \Pi_{i \geq k} (X_i \uplus R) \to R$ as
\begin{itemize}
\item[$(iv)$] $q^{l, s}(\alpha^{\Pi_i(X_{i + k} \uplus R)}) \stackrel{R}{=}
\left\{
\begin{array}{ll}
q(\initSeg{\check{\alpha}}{n} * {\bf 0}) & {\rm if} \; \forall i \!<\! n \, (\alpha(i) \in X_{i + k}) \\[2mm]
a & {\rm if} \; \exists i \!<\! n \, (\alpha(i) \in R),
\end{array}
\right.$
\end{itemize}
where $n = \chi^{+|s|}(l \circ q)(\check{\alpha})$ and $\alpha(\mu i \!<\! n \, (\alpha(i) \in R)) = {\sf inj}_R(a)$. Intuitively, when $q^{l, s}$ reads an input sequence $\alpha \colon \Pi_i (X_{i + |s|} \uplus R)$ it finds the first point $n$ where $l(q(\initSegZ{\check{\alpha}}{n})) < n + |s|$. If all values in $\alpha$ up to that point are $X_i$ values it means this $\alpha$ was generated by a sequence of bar recursive calls until the stopping condition was reached, and hence we must apply the outcome function $q$ to the sequence up to that point. Otherwise, it means that the bar recursive calls have already reached the leaves of the bar recursion (i.e. the stopping conditions) and we are now backtracking and calculating the values of intermediate notes, i.e. computations of the $R$-values. In which case the first such value is then returned. We claim that $\EPQ$ defined as
\begin{itemize}
\item[$(v)$] $\EPQ_s^{l}(\psi)(q^{\Pi_i X_{i + |s|} \to R}) \stackrel{R}{=} q^{l, s}(\IPS_{\tilde s}(\tilde{\psi})(q^{l, s}))$
\end{itemize}
satisfies equation $(\ref{EPQ-def})$. Consider two cases. \\[1mm]
If $l(q({\bf 0})) < |s|$ then, by $(i)$, $n = \chi^{+|s|}(l \circ q)(\beta) = 0$, for any $\beta$. Hence, by $(iv)$, we have that $q^{l, s}(\beta) = q({\bf 0})$, again for any $\beta$. Therefore,
\eqleft{
\begin{array}{lcl}
\EPQ_s^{l}(\psi)(q)
& \stackrel{(v)}{=} & q^{l, s}(\IPS_{\tilde s}(\tilde{\psi})(q^{l, s})) \\[2mm]
& = & q({\bf 0}).
\end{array}
}
On the other hand, if $l(q({\bf 0})) \geq |s|$ then, again by $(i)$, $n = \chi^{+|s|}(l \circ q)(\beta) > 0$, for any $\beta$. This implies both
\begin{itemize}
\item[($vi$)] $q^{l, s}(c * \beta) = r$, for $c = {\sf inj}_R(r)$ and arbitrary $\beta$, and \\[-2mm]
\item[($vii$)] $q^{l, s}(d * \beta) = (q_x)^{l, s*x}(\beta)$, for $d = {\sf inj}_{X_{|s|}}(x)$ and arbitrary $\beta$.
\end{itemize}
Hence
\eqleft{
\begin{array}{lcl}
\EPQ_s^{l}(\psi)(q)
& \stackrel{(v)}{=} & q^{l, s}(\IPS_{\tilde s}(\tilde{\psi})(q^{l, s})) \\[1mm]
& \stackrel{(\ref{IPS-def})}{=} & q^{l, s}(c * \IPS_{\tilde s * c}(\tilde{\psi})((q^{l, s})_c)) \\[1mm]
& \stackrel{(ii), (vi)}{=} & \psi_{\check{\tilde{s}}}(\lambda x . (q^{l, s})_{{\sf inj}_{X_{|s|}}(x)}(\IPS_{\tilde s * {\sf inj}_{X_{|s|}}(x)}(\tilde \psi)((q^{l, s})_{{\sf inj}_{X_{|s|}(x)}}))) \\[1mm]
& \stackrel{(iii)}{=} & \psi_s(\lambda x . (q^{l, s})_{{\sf inj}_{X_{|s|}}(x)}(\IPS_{\widetilde{s*x}}(\tilde \psi)((q^{l, s})_{{\sf inj}_{X_{|s|}(x)}}))) \\[1mm]
& \stackrel{(vii)}{=} & \psi_s(\lambda x . (q_x)^{l, s*x}(\IPS_{\widetilde{s*x}}(\tilde \psi)((q_x)^{l, s*x}))) \\[1mm]
& \stackrel{(v)}{=} & \psi_s(\lambda x . \EPQ_{s*x}^{l}(\psi)(q_x)) \\[2mm]
\end{array}
}
where $c = \tilde{\psi}_{\tilde s}(\lambda x . \selEmb{\IPS_{\tilde s * x}(\tilde{\psi})}((q^{l, s})_x))$.
\end{proof}
\begin{remark} As shown in \cite{Howard(1968)} (cf. also Lemma \ref{howard-kreisel}), if one extends system $T$ with Spector's bar recursion, one can actually prove ${\rm SPEC}$. Hence, the result above says that in all models of system $T$ where $\EPQ$ could exist, it indeed does whenever $\IPS$ also exists. We leave it as an open question whether $\IPS$ already defines $\EPQ$ without assuming ${\rm SPEC}$.
\end{remark}
\begin{figure}
\includegraphics[width=11.5cm]{diagram.pdf}
\caption{Diagram of inter-definability results}
\label{table}
\end{figure}
\section{Summary of Results}
Figure \ref{table} gives a diagrammatic picture of the results presented above. We use a full-line-arrow to represent that the inter-definability holds over ${\rm HA}^\omega$, whereas a dotted-line-arrow indicates that extra assumptions are needed. We have used extra assumptions in four cases. In Theorems \ref{cps-sbr} and \ref{BR-EPQ-def} we made use of bar induction ${\rm BI}$; in Theorem \ref{IPS-EPQ-def} we use ${\rm SPEC}$; and in Theorem \ref{mbr-cbr} we seem to need both bar induction ${\rm BI}$ and the axiom of continuity ${\rm CONT}$. It is an interesting open question whether any of these four results can be shown in ${\rm HA}^\omega$ alone, or under weaker assumptions.
Given that ${\rm CONT}$ implies ${\rm SPEC}$, our results show that over the theory ${\rm HA}^\omega + {\rm BI} + {\rm CONT}$ the different forms of bar recursion considered here fall into two distinct equivalence classes with respect to $T$-definability.
\medskip
\noindent {\bf Acknowledgements}. The authors would like to thank Ulrich Berger, Thomas Powell and in particular the anonymous referee for suggesting numerous improvements and spotting inaccuracies in earlier versions of the paper. The second author also acknowledges support of The Royal Society under grant 516002.K501/RH/kk.
\bibliographystyle{asl}
|
\section{Introduction}
\vspace*{-5pt}
Since the start-up of the LHC, the ATLAS, ALICE, CMS and LHCb detectors have collected
data at higher energies, at higher transverse momenta, with better precision and
with more exclusivity towards direct production compared to that which was achieved before.
Unfortunately, all this seems insufficient to clear up the complexity of the quarkonium-production mechanism.
In this context, a growing attention was directed towards the study of Associated-Quarkonium Production (AQP) channels. In a number of cases,
they are expected to put specific constraints on certain parameters of the theoretical approaches proposed
to describe quarkonium production.
One of the reasons for the complications in describing these elementary reactions is certainly connected
to the large expected impact of perturbative corrections in $\alpha_s$. It is for instance well known
that $\alpha^4_s$ and $\alpha^5_s$ corrections to the colour-singlet
mechanism (CSM)~\cite{CSM_hadron} have a significant impact on the $P_T$ dependence of the $J/\psi$ and $\Upsilon$ cross sections observed in
high-energy hadron collisions~\cite{Campbell:2007ws,Artoisenet:2007xi,Gong:2008sn,Artoisenet:2008fc}
as well as on their polarisation~\cite{Gong:2008sn,Lansberg:2010vq,Li:2008ym,Lansberg:2009db}.
NLO corrections to the colour-octet mechanism (COM) cannot also be overlooked since
they significantly affect polarisation predictions and the extraction~\cite{Butenschoen:2012px} of the non-perturbative octet matrix
elements (also referred to as LDMEs). All this renders the phenomenological analyses rather complex to interpret owing to the
large uncertainties in the current theoretical predictions. On the contrary,
when one focuses on the $P_T$-integrated yields, both for the charmonia and the bottomonia,
the CSM contributions agree relatively well with the existing data at colliders
energies~\cite{Brodsky:2009cf,Lansberg:2010cn}.
The introduction of new observables, such as AQP channels, may thus be of great help.
For some, the impact of QCD corrections is expected to be smaller, thus with smaller uncertainties
on the renormalisation scale for instance. Others are expected to be specifically discrimant towards the separation
of CO and CS contributions, being particularly sensitive to either of these contributions. Finally, some
AQP reactions have similar properties as the inclusive-production reactions
and simply provide further constraints for the fits of NRQCD LDMEs.
\vspace*{-5pt}
\section{Quarkonium plus Quarkonium}
\vspace*{-5pt}
We start with the first AQP ever measured, {\it i.e.}~that of a pair of $J/\psi$'s.
It has been measured for the first time at large $x_F$
by the CERN-NA3 collaboration~\cite{Badier:1982ae,Badier:1985ri} in the eighties.
The rates were higher than
expected and they seemed to be only explainable by the coalescence of
a double intrinsic-charm pair in the proton projectile~\cite{Vogt:1995tf}.
Recently, LHCb has also studied this process at the LHC~\cite{Aaij:2011yc}.
It is important to realise that the cross section measured by LHCb covers a totally different region than NA3,
which a priori should be accounted for by the conventional pQCD approaches, {\it e.g.}~by the
CSM. As a matter of fact, the $P_T$-integrated rate seen by LHCb is in very good agreement
with the recent theoretical expectations from the CSM at LO~\cite{Qiao:2009kg,Berezhnoy:2011xy}.
As far as the $P_T$-integrated yields are concerned, it is reasonable to say that the CSM predictions are as
satisfactory for double $J/\psi$ production as for single $J/\psi$ production.
Last year, we evaluated the leading-$P_T$ contribution at NLO, dubbed as NLO$^\star$, for $J/\psi$
pair production along with that of $J/\psi + \eta_c$ at LO~\cite{Lansberg:2013qka} using the
automated matrix element and event generator {\sc HELAC-Onia}~\cite{Shao:2012iz}. Our partial NLO$^\star$ evaluation
has been confirmed by a full NLO evaluation by Sun and Chao~\cite{Sun:2014gca}. We have found that the NLO$^\star$ is indeed
enhanced w.r.t. LO for increasing $P_T$. It is already 8 times larger for $P_T \gtrsim$ 5~GeV and nearly 400
times larger for $P_T \gtrsim$ 30~GeV. This is compatible with a relative suppression scaling as $P_T^{-2}$ between
the LO and NLO topologies, which justifies the use of the NLO$^\star$ approximation (see~\cite{Artoisenet:2008fc}).
As regards the $P_T$ spectrum for $J/\psi + \eta_c$ at LO, it lies exactly in between these cases.
\begin{figure}[hbt!]
\begin{center}
\subfigure{\includegraphics[width=7.5cm,clip=true]{dphi-norm245-smearing-0512c.pdf}}
\subfigure{\includegraphics[width=7.5cm,clip=true]{dphi-norm5245-smearing-0512c.pdf}}
\caption{Azimuthal distribution of quarkonium pairs for different initial $k_T$ of the colliding partons with (left) and
without (right) a $P_T$ cut on one of the quarkonium.
\label{fig:deltaphi}}\vspace*{-10pt}
\end{center}
\end{figure}
In addition to the yields and their kinematical dependences, we have also looked at the azimuthal correlations
between the quarkonia. At NLO$^{(\star)}$, the presence of a gluon in the final state of the leading topologies
naturally creates an imbalance between the quarkonia, which are then not necessarily created back to back ($\Delta \phi = \pi$). The presence of initial $k_T$ in the colliding partons is also another source of imbalance.
If $\langle k_T \rangle$ is as high as 2 GeV, the azimuthal distribution is completely flattened
(see \cf{fig:deltaphi} (left)); in other words,
the quarkonia are completely decorrelated in azimuth. If one imposes a $P_T$ cut on one of the quarkonia, they
remain produced back to back, except for $J/\psi-J/\psi$ at NLO (see \cf{fig:deltaphi} (right)). Indeed, there is a possibility that the $J/\psi$ pair be
produced recoiling on a gluon. In such a case, the quarkonia are ``near'' to each other ($\Delta \phi \simeq 0$).
These results, in any case, show that one has to be careful when trying to separate out the contribution of Single-Parton Scatterings (SPS)
--expected to be {\it anti}-correlated-- from that of Double-Parton Scatterings
--expected to be uncorrelated-- uniquely from the analysis of the azimuthal distributions. Our findings
confirm the discussion in~\cite{Kom:2011bd}.
Finally, let us add that double $J/\psi$ production has recently been studied by D0 at Fermilab~\cite{Abazov:2014qba}
and by CMS at the LHC~\cite{Khachatryan:2014iia}. These studies respectively brought in new information about the
rapidity and the $P_T$ difference dependence. A possible next step might be to look at the polarisation as we did at NLO$^{(\star)}$
in~\cite{Lansberg:2013qka}.
\vspace*{-5pt}
\section{Quarkonium plus heavy quarks: $J/\psi + c$ and $\Upsilon + \hbox{non-prompt } J/\psi$}
\vspace*{-5pt}
In addition to quarkonium-pair production, LHCb has recently investigated the production of $J/\psi$ with charm hadrons~\cite{Aaij:2012dz}.
Depending on the scheme considered for the number of flavours, this process can come from partonic sub-processes such as $gc \to J/\psi c$~\cite{Brodsky:2009cf}
or $gg \to J/\psi c \bar c$~\cite{Artoisenet:2007xi}. A comparison with the
LHCb data~\cite{Aaij:2012dz} is not straighforward since one has to take into account in the kinematics the charm-quark fragmentation.
In addition, the LHCb data are only for $P_T^C$ of the charm hadrons above 3 GeV. In $2\to 2$ processes, this automatically excludes $P_T^{J/\psi} \leq 3$~GeV.
QCD corrections may therefore be important for these kinematical configurations. As for now, these reactions have only been evaluated at LO and
the current theoretical uncertainties from the scales and from the charm-quark mass are unfortunately large.
In spite of this, the theory-data comparison hints at an important DPS contribution. Yet, since the
$P_T$ spectra of the $J/\psi$'s produced with a charmed hadron and of those without tend to differ, SPS contributions may be sizeable.
Overall, much still has to be learnt from these interesting pieces of data.
\begin{wrapfigure}{r}{0.375\textwidth}
\vspace{-20pt}
\begin{center}
\subfigure[Leading-$P_T$ CSM and COM graphs]{\includegraphics[width=5cm,clip=true]{gg-Upsilon-bb-CSM-COM-crop.pdf}}
\subfigure[CSM crosss-section at the LHC]{\includegraphics[width=5.3cm,clip=true]{dsdpt_upsi_bb-14-080714-crop.pdf}}
\end{center}
\vspace{-20pt}
\caption{Hadroproduction of $\Upsilon + b\bar b $}\label{fig:upsilon_plus_b}
\vspace{-10pt}
\end{wrapfigure}
Along the same lines, bottomonia can also be produced with a beauty hadron. Since these are usually studied via their
weak decay into a $J/\psi$, final states such as $\Upsilon + \hbox{non-prompt } J/\psi$ are also worth investigating. As for
$J/\psi\,+\,$charm, one can look at the event topology, in particular, at how ``near'' (or ``far'') the heavy-flavoured particle is (w.r.t. the quarkonium).
This can indeed be an interesting way to pin down the dominant production (SPS) mechanism\footnote{In the case of DPS production,
one naturally expects the absence of any kind of correlations.}. If one assumes that such a reaction is initated by gluon fusion, the
leading-$P_T$ processes for, respectively the CSM and COM contributions, are as depicted on \cf{fig:upsilon_plus_b} (a). One clearly sees that the
COM contributions can only produce two beauty hadrons ({\it i.e.}~$J/\psi$'s) back-to-back to the $\Upsilon$. At large enough $P^\Upsilon_T$, one would not expect at all
beauty hadrons ``near'' the $\Upsilon$. This is precisely the opposite of what one expects from the CSM, where one beauty hadron would be ``near'' and
one ``away''. The cross section for the LHC kinematics is shown on \cf{fig:upsilon_plus_b} (b). Since about 1\% of $b$ quarks
eventually end up decaying into a $J/\psi$, one can reasonably expect up to 400 $\Upsilon + \hbox{non-prompt } J/\psi$ events at $P^\Upsilon_T\simeq 10$~GeV
with an integrated luminosity of 20 fb$^{-1}$ at the LHC.
\vspace*{-5pt}
\section{Quarkonium plus a back-to-back isolated photon}
\vspace*{-5pt}
The production of a quarkonium associated with an isolated photon has been discussed
in the literature at many instances since the early 90's. The yields have been evaluated at NLO in~\cite{Li:2008ym}
and a partial NNLO (NNLO$^\star$) evaluation has been performed~\cite{Lansberg:2009db}. This
observable is an interesting complement to inclusive measurements. For instance,
it may put stringent constraints on the CO LDMEs\footnote{An updated NLO analysis
taking into account CO channels recently showed that some of the NLO fits
of CO LDMEs~\cite{Butenschoen:2012px} provided unphysical predictions --to be precise, negative yields-- for $J/\psi+\gamma$ production~\cite{Li:2014ava}.}. Yet,
the expected rates are necessarily lower than for inclusive quarkonium production and the theoretical
uncertainties are not necessarily smaller.
In this context, it is worth noting that the requirement for back-to-back quarkonium--photon production selects out
an interesting part of the phase space where (i) neither the QCD corrections, (ii) nor the CO contributions, are kinematically
enhanced, and where (iii) an extension of collinear factorisation -- the TMD factorisation-- can
fully be applied, providing a rigourous set of tools to study the transverse dynamics of the gluon content of the proton.
This is what we discussed in~\cite{Dunnen:2014eta}.
In particular, we computed the expected rates at the LHC, which are summarised in \ct{tab:xsec-onium-gamma}.
On the one hand, we can see that, at the LHC, this observable is essentially from gluon fusion,
making it an excellent gluon probe. On the other hand,
it is most likely a pure CS yield in the $\Upsilon$ case whereas, in the $J/\psi$ case, the predicted
CS yield is above the CO ones where the cross section for accessible events
($J/\psi-\gamma$ invariant mass pair between 20 and 25 GeV) is the highest. For the sake of
TMD factorisation applicability,
it may also be worth isolating the $J/\psi$ in order to prevent any contamination from CO transitions.
\begin{wraptable}{r}{0.67\textwidth}
\vspace{-30pt}
\begin{center}
\includegraphics[width=10cm,clip=true]{TabJPsiGamma-crop.pdf}
\end{center}
\vspace{-10pt}
\caption{ Cross section for back-to-back quarkonium--photon production at $\sqrt{s}=$7~TeV for
gluon fusion and quark-annhilation and for CS and CO transitions for different
pair invariant masses, $Q$, and for $|Y|<0.5$ \& $|\cos\theta|<0.45$ (see \protect\cite{Dunnen:2014eta}).}\label{tab:xsec-onium-gamma}
\vspace{-10pt}
\end{wraptable}
Our conclusion is that such an observable can already be studied at the LHC with
data on tapes (about 20 fb$^{-1}$ of $pp$ collisions). In particular, one should be
able to extract -- for the first time -- the transverse momentum dependence of the gluon content in the proton and
to tell whether the distribution of linearly polarised gluon in an unpolarised proton is nonzero.
It has to be noted that such an observable is also sensitive to gluons at lower energies, such as at the proposed
fixed-target experiment using the LHC beams~\cite{Brodsky:2012vg} (AFTER@LHC). With 20 fb$^{-1}$ of $pp$ data at $\sqrt{s}=115$~GeV, it
should be possible to probe the gluon TMDs up to $x\simeq 0.5$~\cite{transversity} with this process at $Q\simeq 10$ GeV.
\vspace*{-5pt}
\section{Quarkonium plus $W$/$Z$ bosons}
\vspace*{-5pt}
The production of a quarkonium associated with a vector boson has also been the object of a number
of studies. Previous theoretical analyses of $J/\psi+W$ production at hadron colliders~\cite{Barger:1995vx,Kniehl:2002wd,Li:2010hc}
focused on the leading contributions in $\alpha$ or $\alpha_s$, only arising from COM. For a long time, it was believed that
{\it ``$\psi+W$ offers a clean test of the colour-octet contributions''}~\cite{Barger:1995vx}
and that {\it ``If the $J/\psi+W$ production is really detected, it would be a solid basis for testing the color-octet
mechanism of the NRQCD''}~\cite{Li:2010hc}, the latter statement being made following a
NLO study in $\alpha_s$.
In~\cite{Lansberg:2013wva}, we have shown that the CSM contributions to direct $J/\psi+W^\pm$
are not small at all as compared to that from CO transitions. These CSM contributions are due to
two sub-processes: a) the fusion of a gluon and a strange quark which turns into a charm quark by the emission of the $W$,
followed by the fragmentation of the charm quark into a $J/\psi+c$ pair; b) the quark $q$ and antiquark $\bar q'$
annihilation into an off-shell photon, $\gamma^\star$, and a $W$, followed by
the fluctuation of the $\gamma^\star$ into a $J/\psi$. The former process appears at $\alpha_s^3 \alpha$ and the latter at $\alpha^3$ compared to $\alpha_s^2 \alpha$ for the COM process. The CSM contributions were earlier disregarded since
formally appearing at higher orders in $\alpha$ or $\alpha_s$ despite the suppression of the CO in the $v$ expansion of NRQCD.
At the Tevatron, we found that the COM contribution was significantly larger than that of the CSM via $sg$ fusion,
but of similar size as the CSM contribution via $\gamma^\star$. At LHC energies, our results was that the three contributions were of the same order.
As a result, the total CSM cross section is about twice as large as the COM one at 7 TeV, probably a little more at
14 TeV and at large $P_T$ --because of the most favourable running of $\alpha$ w.r.t $\alpha_s$ for increasing scales.
Yet, at LHC energies, based on the ATLAS results~\cite{Aad:2014rua}, it seems that there is a significant DPS contribution.
We have also studied the hadroproduction of $J/\psi+Z$ and $\Upsilon+Z$ at NLO in~\cite{Gong:2012ah}. We found
out that the quarkonium polarisation is not affected by the QCD corrections, whereas these significantly affect the
yield for increasing $P_T$.
\vspace*{-5pt}
\section{Conclusion}
\vspace*{-5pt}
A number of associated-quarkonium-production channels have been theoretically
studied since the early nineties. Thanks to the large luminosities recently collected at both
Fermilab and the LHC, a number of experimental studies have been --and are being-- carried out. There is no doubt that the confrontation
between these theory predictions and these measurements will bring in very soon new information which will be very useful in order to solve the
quarkonium-production puzzles. They may also play a key role in the understanding of DPS physics.
|
\section{Introduction}
Power supply is considered a crucial aspect affecting the performance of wireless communication systems especially for portable nodes \cite{f1,lu2014dynamic}. Traditionally, portable/mobile wireless nodes are battery-based, i.e., operate with energy supply from a battery. These batteries have limited storage capacity and frequently need to be recharged or replaced \cite{lu2014dynamic}. In the last few years, radio frequency (RF) energy harvesting technology has been developed. Such technology is able to supply energy to wireless nodes through the conversion of RF energy to direct current (DC) energy. The authors of \cite{mikeka2011design} report RF-to-DC conversion efficiencies above 0.4\% at -40 dBm, above 18.2\% at -20dBm and over 50\% at -5 dBm RF signals. There are expectations on the improvement of both the sensitivity of energy harvesting circuits and the conversion efficiency in the near future \cite{lu2014dynamic}. Consequently, the use of RF energy harvesting technology has been gaining increasing world-wide interest. In addition, compared with other forms of energy harvesting (e.g., solar, wind and acoustic noise), RF energy harvesting provides relatively predictable energy supply. The amount of RF harvested energy depends on the wavelength of the harvested RF signal and the channel and distance between an RF energy source and the harvesting device \cite{f3}.
Several works have considered nodes with energy harvesting capability, e.g., \cite{hoang2009opportunistic,sharma2010optimal,ho2010optimal,yang2010transmission,yang2010optimal,tutuncuoglu2010optimum}. Various energy sources, such as light, vibration or heat are available to be harvested \cite{survey}. In \cite{hoang2009opportunistic}, Hoang \emph{et al.} studied the optimal policies for a cognitive node. The problems of throughput maximization and mean delay minimization for single-node communication were considered in \cite{sharma2010optimal}. Using a finite horizon setup, the authors of \cite{ho2010optimal} addressed the energy allocation for the maximization of the throughput.
Energy harvested from ambient natural energy resources, i.e., wind and solar energy, has got wide attention \cite{pappas,wimob,ourletter,sultan,wcmpaper,gc2013}. The statistics of energy arrivals are assumed to follow certain random processes. For instance, the authors of \cite{pappas,wimob,ourletter,wcmpaper,gc2013} assume Bernoulli energy arrivals. However, the authors of \cite{krikidis2012stability} assume Poisson arrivals. The authors of \cite{le2008efficient,f3,f7,f8,f9} consider energy harvesting from RF transmissions. In particular, it is assumed that nodes can harvest energy from the transmissions of nearby nodes.
Recently, the availability of free RF energy has increased due to the advent of wireless communication and broadcasting systems \cite{le2008efficient}. Radio waves are ubiquitous in our daily lives in
form of signal transmissions from TV, radio, wireless local area networks (LANs) and mobile phones.
Powering a cognitive radio network through RF energy harvesting can be efficient in terms of spectrum usage and energy limits for wireless networking \cite{lu2014dynamic,f7,f8}. In cognitive radio networks with RF energy harvesting capability, secondary users (SUs) harvest energy from RF signals induced from the nearby RF sources, e.g., primary users (PUs), cellular base stations and other ambient RF sources. Such RF signals can be converted into DC electricity. The converted energy can be stored in an energy storage and used to operate the
devices and transmit data. The SUs spend that energy for their own data transmission. Recent literature on RF-powered cognitive radio networks mainly focuses on investigating throughput maximization under
various constraints. The authors of \cite{f9} consider an RF energy harvesting-enabled cognitive radio sensor networks. The total consumed energy should not be greater than the total harvested energy. This condition represents an energy casuality constraint. An optimal mode selection policy is proposed to balance the immediate throughput and harvested RF energy in transmitting and harvesting modes, respectively. The authors in \cite{f8} assume that mobile nodes in a cognitive radio network opportunistically either harvest energy from RF transmissions of nearby devices in a primary network, or transmit data if the secondary nodes are not in the interference range of any PU. The secondary network throughput is maximized via obtaining the optimal transmit power and density of the secondary transmitting terminals under an outage-probability constraint.
In this work, we consider a mixture between RF energy harvesting and energy harvested from natural resources in the surrounding environment. We assume a simple configuration composed of one energy harvesting SU and a PU. The SU collects energy from nature and converts energy from the PU's RF transmissions. We assume fading channels and propose a power allocation policy at the PU. We investigate the impact of the proposed power allocation policy on the throughput of both primary and secondary links. Results reveal that the throughput of both PU and SU is enhanced under the proposed power allocation and RF energy harvesting techniques.
The rest of this paper is organized as follows. The system model under consideration is described in Section \ref{system_model}. Energy harvesting employed at the SU is studied and analyzed in Section \ref{energy_harvesting}. Section \ref{proposed_policy} presents the proposed transmission policy adopted by the SU along with its throughput analysis. Numerical results are presented in Section \ref{numerical_results}. Finally, concluding remarks are drawn in Section \ref{conclusion}.
\section{System Model}\label{system_model}
Fig. \ref{Fig1} shows the system model under consideration.
We consider a simple cognitive network composed of one PU ($\rm p$) plugged to a reliable power supply and one SU ($\rm s$) equipped with energy harvesting capability.
The PU transmits its packets to a primary destination ($\rm pd$), while the intended destination for SU packet transmissions is $\rm sd$.
Each user has its own infinite-length data queue (buffer) to store the incoming data traffic. We denote the primary and the secondary data queues as $Q_{\rm p}$ and $Q_{\rm s}$, respectively. The SU maintains another queue denoted by $Q_{\rm e}$ which stores fixed-length energy packets. An energy packet is assumed to contain ${\rm e}$ energy units. $Q_{\rm e}$ has a limited capacity of $E_{\max}$ packets. Thus, it contains $E_{\max} {\rm e}$ energy units at maximum. We assume fixed-length data packets of size $\beta$ bits.
Time is slotted and the transmission of a packet takes exactly one time slot of duration $T$ seconds.
The arrivals at the primary queue, $Q_{\rm p}$, are assumed to be independent and identically distributed across time slots following a Bernoulli process with mean $\lambda_{\rm p}\in [0,1]$ packets/slot. We assume that the secondary data queue is saturated with packets. This means that the SU is backlogged with data. The SU harvests energy from the primary RF transmissions and from the environment (nature), e.g., solar, wind, etc. The arrivals to $Q_{\rm e}$ due to environmental energy harvesting are assumed to be distributed according to a Poisson process with rate $\lambda_{\rm e}$ energy packets/slot. For secondary data transmissions, we assume that multiple energy packets may be used for a single data packet transmission. The PU has the priority to transmit if $Q_{\rm p}$ is non-empty, whereas the SU transmits a packet from $Q_{\rm s}$ if $Q_{\rm e}$ maintains at least $\mathcal{G}$ energy packets and the PU is inactive. Whenever it transmits, the PU utilizes the whole time slot duration for its data transmission. However, the SU performs spectrum sensing for a duration of $\tau < T$ seconds to detect the PU's activity. We do not consider erroneous sensing outcomes. For similar assumption of perfect channel sensing, the reader is referred to \cite{krikidis2009protocol,krikidis2010stability,wcmpaper}.
The channel between every transmitter-receiver pair exhibits frequency-flat Rayleigh block fading, i.e., the channel coefficient remains constant for one time slot and changes independently from a slot to another. We denote the gain of the channel between transmitter ${\rm i}$ and receiver ${\rm \ell}$ at the $t$th time slot by $h_{\rm i\ell}(t)$\footnote{We refer by the channel gain to the absolute squared value of the channel fading coefficient.}. According to the Rayleigh fading assumption, $h_{\rm i\ell}(t)$ is exponentially distributed with mean $\rm \sigma_{i\ell}$. All links are statistically independent.
\begin{figure}[t]
\begin{center}
\includegraphics[width=1.\columnwidth]{RF_system_model.eps}
\caption{Cognitive radio network under consideration. $\mathcal{A}_{\rm p}$ is the number of primary data packet arrivals in a given time slot. $\mathcal{H}_{\rm RF}$ and $\mathcal{H}_{\rm N}$ are the number of energy packets harvested from RF and natural resources, respectively.}\label{Fig1}
\end{center}
\vspace{-5mm}
\end{figure}
The PU adopts a power allocation scheme in which it selects its transmission power level each time slot based on the channel realization between its transmitter and the destination. For the link $\rm p \rightarrow pd$, we assume perfect channel state information at the transmitter side (CSIT).
The transmit power is assumed to be continuous over the set $P=[0,P_\mathcal{M}]$, where $P_\mathcal{M}$ is the maximum transmit power constraint per time slot and is dependent on the application and the used transceiver. If the primary direct link is in a deep fade such that the maximum power, $P_\mathcal{M}$, cannot overcome the link outage, the PU remains idle during the whole time slot to avoid wasting its energy without adding further contribution to its throughput. An outage occurs on the link $\rm p \rightarrow pd$ when the transmission rate exceeds the link capacity. The probability that an outage occurs on that link in the $t$th time slot is given by
\begin{equation}\label{outage_p}
\mathbb{P}_{\rm ppd}(t)=\Pr \left \{\mathcal{R}_{\rm p}> \log_2\left(1+\frac{P_{\rm p}(t) h_{\rm ppd}(t)}{\mathcal{N}_\circ W}\right)\right \}
\end{equation}
where $\Pr\{\cdot\}$ denotes the probability of the event $\{\cdot\}$, $\mathcal{R}_{\rm p}=\beta/T/W$ bits/second/Hz is the targeted primary spectral efficiency, $P_{\rm p}(t)$ is the average transmit power in Watts used by the PU in time slot $t$, ${\mathcal{N}_\circ}$ is the additive white Gaussian noise (AWGN) power spectral density in Watts/Hz and $W$ is the channel bandwidth in Hz. It is worth noting from (\ref{outage_p}) that the minimum power required to guarantee a successful primary transmission in time slot $t$, i.e., no outage occurs on the link $\rm p \rightarrow pd$, is given by
\begin{equation}\label{PU_power}
P^*_{\rm p}(t)=\frac{\mathcal{N}_\circ W (2^{\mathcal{R}_{\rm p}}-1)}{h_{\rm ppd}(t)}.
\end{equation}
Hereafter, we omit the temporal index $t$ for simplicity. Nevertheless, it is implicitly understood that power allocation at the PU is performed on a slot-by-slot basis.
The PU sets its transmission power level according to (\ref{PU_power}) and $P_{\mathcal{M}}$. If $P_{\rm p}^*>P_{\mathcal{M}}$, the PU remains silent. It is worth noting that as the instantaneous channel gain, $h_{\rm ppd}$, decreases, the power level used by the PU increases. This potentially increases the amount of RF energy harvested by the SU. However, as $h_{\rm ppd}$ decreases, this also increases the probability of the PU being idle which in turn reduces its throughput and might result in increasing the availability of time slots in which the SU can transmit. Therefore, this point arises a fundamental tradeoff between the throughput of both the PU and SU which is studied in details in the context of the paper. Interestingly, results reveal that the throughput of both users could be enhanced simultaneously.
\section{Energy harvesting}\label{energy_harvesting}
In this section, we study and analyze the arrivals of energy packets to $Q_{\rm e}$ as a result of harvesting energy from primary RF transmissions as well as from natural resources. We lay down the basis of the probabilistic energy arrivals model based on which the analysis in the rest of the paper is performed.
\subsection{RF energy harvesting}
The one and only source of RF energy harvesting considered in this paper is the transmission of PU's packets. Therefore, to analyze the impact of these transmissions on the amount of harvested energy by the SU, we start with studying the behavior of the PU. The SU accesses the channel when the PU is inactive. This occurs in two cases. 1) $Q_{\rm p}$ is empty; or 2) $Q_{\rm p}$ is non-empty but $P_{\rm p}^{*}>P_{\mathcal{M}}$. Thus, the probability of the PU being inactive is given by
\begin{equation}
\begin{split}
\Pi\!&=\!\Pr\{Q_{\rm p}=0\}\!+\!\Pr\{Q_{\rm p}\!\ne\!0\} \Pr\{P_{\rm p}^*>P_{\mathcal{M}}\}\!
\\& =\!1\!-\!\Pr\{Q_{\rm p}\ne0\} (1-\Pr\{P_{\rm p}^*\!>\!P_{\mathcal{M}}\}).
\end{split}
\end{equation}
We first compute $\Pr\{P_{\rm p}^*>P_{\mathcal{M}}\}$. It is given by
\begin{equation}
\Pr\{P_{\rm p}^*>P_{\mathcal{M}}\}= \Pr \left \{\frac{\mathcal{N}_\circ W(2^{\mathcal{R}_{\rm p}}-1)}{h_{\rm ppd}}>P_{\mathcal{M}} \right \}.
\end{equation}
Since $h_{\rm ppd}$ is exponentially distributed, we have
\begin{equation}
\Pr\{P_{\rm p}^*>P_{\mathcal{M}}\}= 1-\exp\Big(-\frac{\mathcal{N}_\circ W(2^{\mathcal{R}_{\rm p}}-1)}{P_{\mathcal{M}}\sigma_{\rm ppd}}\Big).
\end{equation}
The mean service rate of the primary queue, $\mu_{p}$, is equal to the probability that the link $\rm p \rightarrow pd$ is not in outage. Therefore, $\mu_{p}$ is given by
\begin{equation}\label{mu_p}
\begin{split}
&\mu_{\rm p}=\Pr\{P_{\rm p}^* \leq P_{\mathcal{M}}\}=\exp\Big(-\frac{\mathcal{N}_\circ W (2^{\mathcal{R}_{\rm p}}-1)}{P_{\mathcal{M}}\sigma_{\rm ppd}}\Big).
\end{split}
\end{equation}
If $Q_{\rm p}$ is stable, i.e., $\lambda_{\rm p}<\mu_{\rm p}$, the probability of the queue being empty is given by
\begin{equation}
\Pr\{Q_{\rm p}=0\}=1-\frac{\lambda_{\rm p}}{\mu_{\rm p}}.
\end{equation}
However, if $Q_{\rm p}$ is unstable, i.e., $\lambda_{\rm p}\ge\mu_{\rm p}$, the queue is always saturated with data packets and hence, the probability of the queue being empty is $0$.
Combining both cases, the probability of the primary queue being empty is given by
\begin{equation}
\Pr\{Q_{\rm p}=0\}=1-\min \left \{ \frac{\lambda_{\rm p}}{\mu_{\rm p}},1 \right \}.
\end{equation}
Thus, the probability of the PU being inactive is given by
\begin{equation}\label{Pi}
\Pi=1-\min \left \{ \frac{\lambda_{\rm p}}{\mu_{\rm p}},1 \right \} \mu_{\rm p}
\end{equation}
where $\mu_{\rm p}$ is given by (\ref{mu_p}).
Consequently, the throughput of the PU is given by
\begin{equation}\label{PU_thrpt}
\overline{\Pi}=\min \left \{ \frac{\lambda_{\rm p}}{\mu_{\rm p}},1 \right \} \mu_{\rm p}.
\end{equation}
where the notation $\overline{\mathcal{X}}=1-\mathcal{X}$ is used throughout the paper.
The received power at the SU due to the primary transmission, when the PU is {\it active}, is given by
\begin{equation}\label{rxed_p_su}
P_{\rm p}^* h_{\rm ps} = \frac{\mathcal{N}_\circ W (2^{\mathcal{R}_{\rm p}}-1)}{h_{\rm ppd}} h_{\rm ps}
\end{equation}
with $P_{\rm p}^*\le P_{\mathcal{M}}$. Received signals from primary transmissions at the SU is converted into energy packets with an RF-to-DC conversion efficiency $\eta \leq 1$. From (\ref{rxed_p_su}), it is obvious that the amount of energy harvested by the SU depends on the quality of the links between the PU and SU and the PU and its destination, i.e., $\rm p \rightarrow s$ and $\rm p \rightarrow pd$. Another interesting observation is the dependence of the harvested energy on the burstiness of the primary source. Specifically, more data packets at the PU increases the possibility of transmission and hence, increases the possibility of harvesting energy at the SU. However, this comes at the expense of lowering the possibility of finding available time slots in which the SU can transmit.
Since one energy packet contains ${\rm e}$ energy units, the number of energy packets harvested by the SU when the PU is {\it active} in a time slot is given by
\begin{equation}
\mathcal{H}_{\rm RF}=\Big\lfloor \frac{\eta \frac{\mathcal{N}_\circ W (2^{\mathcal{R}_{\rm p}}-1)}{h_{\rm ppd}} h_{\rm ps}T}{{\rm e}}\Big\rfloor
\end{equation}
where $\lfloor\cdot\rfloor$ is the largest integer not greater than the argument.
Therefore, when the PU transmits in a given time slot, the probability that the SU harvests $n$ energy packets is given by
\begin{equation}\label{P(A_e=n)}
\Pr \lbrace \mathcal{H}_{\rm RF}\!=\!n \rbrace\!=\! \Pr \! \left \lbrace \!\! n \leq \frac{\eta \frac{\mathcal{N}_\circ W(2^{\mathcal{R}_{\rm p}}-1)}{h_{\rm ppd}} h_{\rm ps}T}{{\rm e}}
< n+1 \!\! \right \rbrace
\end{equation}
with $h_{\rm ppd} \geq a=\frac{\mathcal{N}_\circ W(2^{\mathcal{R}_{\rm p}}-1)}{P_{\mathcal{M}}}$. This condition on $h_{\rm ppd}$ originates from the power allocation policy employed at the PU. From (\ref{P(A_e=n)}), $\Pr \lbrace \mathcal{H}_{\rm RF}\!=\!n \rbrace$ can be written as
\begin{equation}\label{Z}
\Pr \left \{ \frac{n\rm e}{\eta \mathcal{N}_\circ W(2^{\mathcal{R}_{\rm p}}-1)T} \leq
\frac{h_{\rm ps}}{h_{\rm ppd}}
< \frac{(n+1){\rm e}}{\eta \mathcal{N}_\circ W(2^{\mathcal{R}_{\rm p}}-1)T}\right \}.
\end{equation}
Let $\alpha=\frac{{\rm e}}{\eta {\mathcal{N}_\circ W(2^{\mathcal{R}_{\rm p}}-1)} T}$. We solve (\ref{Z}) to obtain a closed-form expression for the probability of harvesting $n$ packets through primary RF transmissions. We get
\begin{equation}
\Pr \lbrace \mathcal{H}_{\rm RF}\!=\!n \rbrace = F((n+1)\alpha)-F(n\alpha)
\end{equation}
where the function $F(\cdot)$ is derived in Appendix A.
\subsection{Energy harvesting from natural resources}
The energy packets harvested from nature are assumed to follow a Poisson process with rate $\lambda_{\rm e}$ \cite{krikidis2012stability}. Thus, the probability of harvesting $k$ energy packets in a given time slot is given by
\begin{equation}
\Pr\{\mathcal{H}_{\rm N}=k\}=\frac{(\lambda_{\rm e} T)^k \exp(-\lambda_{\rm e} T)}{k!}.
\end{equation}
Note that the number of harvested packets from any of the sources cannot be negative; hence, $P^{\prime}_{n}=P^{\prime \prime}_{n}=0$ for all $n\in\{-\infty,\dots,-2,-1\}$.
\subsection{Combining nature and RF energy harvesting}
When the PU is inactive, the energy packet arrivals at $Q_{\rm e}$ comes from natural resources only. Thus, the probability of harvesting $n$ packets of energy from nature is given by
\begin{equation}
P^{\prime}_{n}= \Pr \{\mathcal{H}_{\rm N}=n\}.
\label{dfgg2}
\end{equation}
However, when the PU is active, the probability of harvesting $n$ packets of energy from both the PU's transmission and natural resources in an arbitrary time slot is given by
\begin{equation}
P^{\prime \prime}_{n}=\sum_{k=0}^{\infty}\Bigg(\begin{array}{cc}
n \\
k
\end{array}\Bigg)
\Pr \{\mathcal{H}_{\rm N}=k\} \Pr \{\mathcal{H}_{\rm RF}=n-k\}.
\label{dfgg}
\end{equation}
where $\Big(\begin{array}{cc}
n \\
k
\end{array}\Big)$ denotes $n$ choose $k$. Note that the joint probabilities in (\ref{dfgg}) are the multiplication of the marginal probabilities due to the independence of the two energy sources, i.e., RF and natural energy sources.
The energy queue evolves as follows:
\begin{equation}
Q^{t+1}_{\rm e}=(Q^{t}_{\rm e}- \mathcal{D}^t_{\rm e})^+ +\mathcal{H}^{t}_{\rm RF}+\mathcal{H}^{t}_{\rm N}
\end{equation}
where $(\cdot)^+$ returns the maximum between the argument and $0$, $\mathcal{D}^t_{\rm e}$ is the number of departures from the energy queue in time slot $t$, $\mathcal{H}^t_{\rm RF}$ is the number of harvested energy packets from the primary transmission in time slot $t$, and $\mathcal{H}^t_{\rm N}$ is the number of energy harvested from nature in time slot $t$.
\section{Proposed Transmission Policy and SU throughput analysis}\label{proposed_policy}
Whenever the SU transmits a data packet, it uses $\mathcal{G}~\in~\{1,2,3,\dots,E_{\max}\}$ energy packets. This means that once that number of packets is available at the energy queue and at the same time the PU is inactive, they are dissipated. Clearly, as the number of energy packets invested in a secondary data packet transmission increases, i.e., as $\mathcal{G}$ increases, the probability that $\rm sd$ successfully decodes that packet increases. However, increasing $\mathcal{G}$ depletes $Q_{\rm e}$ faster. Therefore, we need to choose the optimal value of $\mathcal{G}$ that maximizes the SU throughput.
Let $\chi_m$ denote the probability of the energy queue being in state $m\in\{0,1,\dots,E_{\max}\}$.
Therefore, it can be easily shown that the mean service rate of the energy queue is given by
\begin{equation}
\begin{split}
\mu_{\rm e}&= \mathcal{G} \Pi \sum_{j=\mathcal{G}}^{E_{\max}} \chi_j
\end{split}
\end{equation}
where $\Pi$ is given by (\ref{Pi}).
The SU transmits with fixed energy $\mathcal{G} {\rm e}$ energy units. Each time slot, it senses the channel for $\tau$ seconds and transmits only if the channel is sensed to be free. Thus, its transmission spectral efficiency is $\mathcal{R}_{\rm s}=b/T_{\rm s}/W$ bits/second/Hz, where $T_{\rm s}=T-\tau$. The probability of success of a secondary data packet transmission, i.e., the link $\rm s \rightarrow sd$ is not in outage, is given by
\begin{equation}\label{P_ssd}
\overline{\mathbb{P}_{\rm ssd,\mathcal{G}}}= \exp\Big(-\mathcal{N}_\circ W(T-\tau)\frac{(2^{\mathcal{R}_{\rm s}}-1)}{\mathcal{G} {\rm e}\sigma_{\rm ssd}}\Big).
\end{equation}
As indicated earlier, $\overline{\mathbb{P}_{\rm ssd,\mathcal{G}}}$ is monotonically increasing with the transmit energy, $\mathcal{G} {\rm e}$.
A packet from the secondary data queue is served when the PU is inactive, the secondary link is not in outage and the secondary energy queue maintains $\mathcal{G}$ energy packets. The throughput of the SU is then given by
\begin{equation}\label{SU_thrpt}
\begin{split}
\mu_{\rm s}&=\Pi \overline{\mathbb{P}_{{\rm ssd},\mathcal{G}}} \sum_{j=\mathcal{G}}^{E_{\max}} \chi_j.
\end{split}
\end{equation}
The maximum SU throughput is obtained via solving the following optimization problem:
\begin{equation}
\begin{split}
& \underset{\mathcal{G}\in \{1,2,\dots,E_{\max}\}}{\max.} \,\,\,\,\,\,\ \mu_{\rm s}=\Pi \overline{\mathbb{P}_{{\rm ssd},\mathcal{G}}} \sum_{j=\mathcal{G}}^{E_{\max}} \chi_j.
\end{split}
\end{equation}
$\Pi$ and $\overline{\mathbb{P}_{{\rm ssd},\mathcal{G}}}$ have already been determined by (\ref{Pi}) and (\ref{P_ssd}), respectively. Thus, it now remains to compute the steady state distribution of the energy queue, i.e., $\{\chi_j\}_{j=\mathcal{G}}^{\infty}$, to completely characterize the SU throughput for a given value of $\mathcal{G}$, given by (\ref{SU_thrpt}). Towards this objective, we model the evolution of the energy queue with a Markov chain. The probability of transition from state $j$ to state $k$, denoted by $P_{j \rightarrow k}$, is
\begin{itemize}
\item For $k < E_{\max}$,
\begin{enumerate}
\item at $j<\mathcal{G}$,
\begin{enumerate}
\item if $k<j$,
\begin{equation}
P_{j \rightarrow k}=0
\end{equation}
\item if $k \geq j$,
\begin{equation}
P_{j \rightarrow k}=\Pi P^{\prime}_{k-j} +
\overline{\Pi}P^{\prime \prime}_{k-j}
\end{equation}
\end{enumerate}
\item at $j \geq \mathcal{G}$,
\begin{enumerate}
\item if $k<j-\mathcal{G}$,
\begin{equation}
P_{j \rightarrow k}=0
\end{equation}
\item if $k \geq j-\mathcal{G}$,
\begin{equation}
P_{j \rightarrow k}=\Pi P^{\prime}_{k-(j-\mathcal{G})} +
\overline{\Pi}P^{\prime \prime}_{k-j}.
\end{equation}
\end{enumerate}
\end{enumerate}
\item For $k=E_{\max}$,
\begin{enumerate}
\item at $j<\mathcal{G}$,
\begin{align}
P_{j \rightarrow k}&=\Pi \left[\displaystyle \sum_{n=k-j}^{\infty} P_{n}^{\prime}\right] +
\overline{\Pi} \left[\displaystyle \sum_{n=k-j}^{\infty} P_{n}^{\prime \prime}\right] \notag \\
&= \Pi \left[ 1-\displaystyle \sum_{n=0}^{k-j-1} P_{n}^{\prime} \right] +
\overline{\Pi} \left[1- \displaystyle \sum_{n=0}^{k-j-1} P_{n}^{\prime \prime} \right]
\end{align}
\item at $j \geq \mathcal{G}$,
\begin{align}
P_{j \rightarrow k}&=\Pi \left[\displaystyle \sum_{n=k-(j-\mathcal{G})}^{\infty} P_{n}^{\prime}\right] +
\overline{\Pi} \left[ \displaystyle \sum_{n=k-j}^{\infty} P_{n}^{\prime \prime} \right] \notag \\
&= \Pi \left[ 1-\displaystyle \sum_{n=0}^{k-(j-\mathcal{G})-1} P_{n}^{\prime} \right] +
\overline{\Pi} \left[1- \displaystyle \sum_{n=0}^{k-j-1} P_{n}^{\prime \prime} \right].
\end{align}
\end{enumerate}
\end{itemize}
\begin{figure}[t]
\begin{center}
\includegraphics[width=1\columnwidth , height=0.7\columnwidth]{RF_eta_Emax.eps}
\caption{SU throughput versus $\lambda_{p}$ for different values of $\eta$ and $E_{\max}$.} \label{Fig2}
\end{center}
\vspace{-5mm}
\end{figure}
The $(E_{\max}+1) \times (E_{\max}+1)$ transition probability matrix, denoted by $\Omega$, is constructed according to the above description. Then, the $1 \times E_{\max}$ steady state distribution vector $\boldsymbol{\chi}=[\chi_0,\chi_1,\dots,\chi_{E_{\max}}]$ is obtained through solving
\begin{equation}
\boldsymbol{\chi}=\boldsymbol{\chi} \Omega.
\end{equation}
\section{Numerical Results}\label{numerical_results}
In this section, the system performance is evaluated in terms of the SU throughput under the proposed transmission policy. Specifically, we investigate the effect of the following factors on the achievable throughput for the secondary link $\rm s \rightarrow sd$;
\begin{enumerate}
\item We study the impact of the bursty arrivals at the PU's queue on the SU throughput.
\item We analyze the dependence of both primary and secondary throughput on the quality of the direct link between the PU and its destination, i.e., we focus on the role of the power allocation policy employed at the PU.
\item We show how each energy harvesting technique employed at the SU influences its throughput, i.e., we investigate the effect of harvesting energy from natural resources and PU's RF transmissions individually. Then, we combine both energy harvesting techniques to enhance the secondary throughput.
\item The roles of the maximum capacity of the energy queue, $E_{\max}$, energy packets' arrival rate at $Q_{\rm e}$, $\lambda_{\rm e}$, and RF-to-DC conversion efficiency, $\eta$, are highlighted.
\end{enumerate}
We show our results at a fixed packet-length of $\beta=10^3$ bits. Time slot duration $T$ is set to $1$ second, while the sensing duration $\tau=0.1$ second. We note that $10 \%$ of the time slot is used by the SU to perform sensing which to some extent justifies the assumption of perfect sensing.
Channel bandwidth $W=1$ KHz. The noise power spectral density $\mathcal{N}_{\circ}=10^{-6}$ Watts/Hz. We assume that an energy packet contains $10^{-3}$ joules, i.e., ${\rm e}=10^{-3}$ joules. The maximum power constraint imposed at the PU $P_\mathcal{M}=10$ dBm unless otherwise stated. The variances of both links $\rm p \rightarrow s$ and $\rm s \rightarrow sd$ are set to unity, i.e., $\rm \sigma_{ps}=\sigma_{ssd}=1$. However, we consider $\rm \sigma_{ppd}=0.5$.
\begin{figure}[t]
\begin{center}
\includegraphics[width=1\columnwidth , height=0.7\columnwidth]{PU_SU_thrpt_vs_sigma_pDp.eps}
\caption{Primary and secondary throughput versus $\sigma_{\rm ppd}$.} \label{Fig3}
\end{center}
\vspace{-5mm}
\end{figure}
In Figs. \ref{Fig2} and \ref{Fig3}, we demonstrate the effect of RF energy harvesting on the system performance. We consider a scenario in which the SU harvests energy from the PU's RF transmissions only, i.e., $\lambda_{\rm e}=0$. We study the effect of the packet arrival rate at $Q_{\rm p}$, $\lambda_{\rm p}$, on the maximum achievable SU throughput. Towards this objective, we plot in Fig. \ref{Fig2} the maximum SU throughput versus $\lambda_{\rm p}$. Interestingly, we observe that the SU throughput is initially enhanced as $\lambda_{\rm p}$ increases. However, this behavior is reversed as $\lambda_{\rm p}$ continues to increase until the secondary throughput saturates as $\lambda_{\rm p} \in [\mu_{\rm p},1]$. What happens beyond the scene goes as follows. Initially, as $\lambda_{\rm p}$ increases, the PU transmits more frequently and hence, the amount of energy harvested by the SU due to primary RF transmissions increases. Thus, the SU possesses enough energy for its prospective transmissions which enhances its throughput. However, as $\lambda_{\rm p}$ continues to increase, the SU suffers from the lack of available time slots in which it can transmit, i.e., slots in which the PU is inactive. This becomes the dominant factor reducing the throughput of the SU even though it has enough amount of energy. As $\lambda_{\rm p}$ exceeds $\mu_{\rm p}$, the PU's activity no longer depends on $\lambda_{\rm p}$ since its queue becomes backlogged. From (\ref{PU_thrpt}), we know that $\overline{\Pi}=\mu_{\rm p}$ which is a constant number independent of $\lambda_{\rm p}$. Therefore, the effect of $\lambda_{\rm p}$ on the SU throughput in the interval $\lambda_{\rm p} \in
[\mu_{\rm p},1]$ is neutralized and the secondary throughput saturates as shown in Fig. \ref{Fig2}.
Furthermore, we depict the impact of RF-to-DC conversion efficiency on the SU throughput via plotting the results for different values of $\eta$ in Fig. \ref{Fig2}. Obviously, the SU throughput is enhanced as $\eta$ increases. Moreover, we notice from Fig. \ref{Fig2} that increasing the capacity of the SU's battery, i.e., $E_{\max}$, enhances its throughput. These results are intuitive and match our expectations.
\begin{figure}
\begin{center}
\includegraphics[width=1\columnwidth , height=0.7\columnwidth]{RF_N_RFN.eps}
\caption{SU throughput versus $\lambda_{\rm p}$ for different energy harvesting techniques.} \label{Fig4}
\end{center}
\vspace{-5mm}
\end{figure}
We proceed next with studying the power allocation policy employed at the PU and how it affects the throughput of both primary and secondary links. As explained in Section \ref{system_model}, the PU's transmit power depends on the quality of the link between itself and its destination, i.e., $\rm \sigma_{ppd}$. In Fig. \ref{Fig3}, we remain focusing on the case of RF energy harvesting via setting $\lambda_{\rm e}=0$. We plot the throughput of both the PU and SU versus $\rm \sigma_{ppd}$ for $\eta=0.6$, $\lambda_{\rm p}=0.4$, $E_{\max}=10$ energy packets and $P_{\mathcal{M}}=1.76$ dBm. As expected, the PU's throughput is a monotonically increasing function of $\rm \sigma_{ppd}$. From (\ref{PU_thrpt}), we note that $\mu_{\rm p}$ increases as $\rm \sigma_{ppd}$ increases. Thus, when $\mu_{\rm p}>\lambda_{\rm p}$, the PU throughput saturates at $\lambda_{\rm p}=0.4$. On the other hand, the SU throughput is enhanced initially as $\rm \sigma_{ppd}$ increases. This is attributed to the increase in the frequency of primary RF transmissions and hence, the amount of energy packets harvested by the SU increases resulting in enhancing its throughput. However, as $\rm \sigma_{ppd}$ continues to increase, the behavior of the SU throughput is reversed. This originates from the adaptive power allocation policy employed at the PU. As the quality of the link $\rm p \rightarrow pd$ is enhanced, the PU can guarantee successful packet transmissions at lower average power levels. Thus, as the PU reduces its transmission power, the amount of energy harvested by the SU is reduced yielding the reduction in its throughput.
In Fig. \ref{Fig4}, we show the resulting SU throughput for different energy harvesting scenarios at $E_{\max}=6$ energy packets. We consider each energy harvesting technique independently, i.e., natural resources and RF energy harvesting, then we combine both of them. The case of harvesting energy from natural resources only corresponds to $\eta=0$. Since the amount of energy harvested from natural resources is independent of the PU's activity, the energy packets arrivals at $Q_{\rm e}$ does not change as $\lambda_{\rm p}$ varies. Therefore, the SU throughput changes with $\lambda_{\rm p}$ only due to the variation of the availability of time slots in which the PU is inactive. Consequently, SU throughput decreases as $\lambda_{\rm p}$ increases which is evident from Fig. \ref{Fig4}. On the other hand, Fig. \ref{Fig4} shows the SU performance under RF energy harvesting only, i.e., $\lambda_{\rm e}=0$, that has already been explained in the comments on Fig. \ref{Fig2}. Finally, from Fig. \ref{Fig4}, we can notice the performance gain originating from combining both energy harvesting techniques.
In Fig. \ref{Fig5}, we investigate the dependence of the SU throughput on energy packets' arrival rate at $Q_{\rm e}$, $\lambda_{\rm e}$, and the maximum capacity of $Q_{\rm e}$, $E_{\max}$, at $\eta=0.2$. We note that the SU throughput is enhanced as both $\lambda_{\rm e}$ in energy packets/slot and $E_{\max}$ in packets increase. This is attributed to the increase in the availability of energy at the SU required to support its packets transmission.
\begin{figure}
\begin{center}
\includegraphics[width=1\columnwidth , height=0.7\columnwidth]{RF_N_Le_Emax.eps}
\caption{SU throughput versus $\rm \lambda_{p}$ for different values of $\lambda_{\rm e}$ and $E_{\max}$.}\label{Fig5}
\end{center}
\vspace{-5mm}
\end{figure}
\section{Conclusion}\label{conclusion}
In this work, we have investigated the maximum achievable throughput for an SU sharing the spectrum with a PU. The SU has a limited-capacity battery and is equipped with energy harvesting capability. We assume that the SU harvests energy from primary RF transmissions as well as through natural resources. The obtained results reveal the promises of employing RF energy harvesting. We have showed that the SU throughput can be enhanced as the arrival rate to the primary queue increases. This is attributed to the increase in the amount of energy that the SU harvests from primary RF transmissions. Thus, we conclude that a win-win scenario can be approached in which the performance of both users is enhanced simultaneously.
\section*{Appendix A}
For simplicity, we define new variables with which we proceed in the derivation. Let $X=h_{\rm ps}$, $Y=h_{\rm ppd}$ and $Z=X/Y$. The first step towards solving (\ref{Z}) requires the computation of $F(z)=\Pr\{Z \leq z , Y \geq a\}$ for an arbitrary $z \geq 0$. After some manipulation, we have
\begin{equation}\label{17}
F(z)=\int_{a}^{\infty} \int_{0}^{zy} P_{X,Y}(x,y) dx dy
\end{equation}
where $P_{X,Y}(\cdot,\cdot)$ denotes the joint probability density function (PDF) of $X$ and $Y$. Since $X$ and $Y$ are independent random variables, their joint PDF is given by the product of their marginal distributions, i.e., $P_{X,Y}(x,y)=P_{X}(x)P_{Y}(y)$. Following the Rayleigh fading assumption, we know that $P_{X}(x)=\lambda_{x}\exp(-x \lambda_{x})$ and $P_{Y}(y)=\lambda_{y}\exp(-y \lambda_{y})$, where $\lambda_{x}$ and $\lambda_{y}$ are given by $1/\sigma_{\rm ps}$ and $1/\sigma_{\rm ppd}$, respectively. Substituting in (\ref{17}) and solving the double integral, we get
\begin{equation}
\begin{split}
F(z)\!&=\!\int_{a}^{\infty} \int_0^{zy} \lambda_x \exp(-\lambda_x x) \lambda_y \exp(-\lambda_y y) dx dy\\& \!=\! \int_a^\infty \lambda_x \lambda_y \exp(-\lambda_y y) \int_{0}^{zy}\! \exp(-\lambda_x x) dx\\& \!=\!\lambda_x \lambda_y \int_a^ \infty \! \! \exp(- \lambda_y y) \Big[ \frac{\exp(-\lambda_x x)}{-\lambda_x}\Big]_0^{zy} dy\\& \!=\!\lambda_y \int_a^ \infty\exp(- \lambda_y y) \Big[ {1-\exp(-\lambda_x zy)}\Big]_0^{zy} dy \\& \!=\!\lambda_y \int_a ^\infty\!\!\exp(-\lambda_y y)\!-\!\exp(-y [\lambda_y\!+\!\lambda_x z]) dy\\& \!=\! \Bigg[\!-\exp(-\lambda_y y)\!+\!\frac{\lambda_y}{\lambda_y\!+\!\lambda_x z}\exp(-y [\lambda_y\!+\!\lambda_x z])\Bigg]_a^\infty\\&=
\exp(-\lambda_y a)\Bigg[\!1\!-\!\frac{\lambda_y}{\lambda_y\!+\!\lambda_x z}\exp(-a \!\lambda_x z)\Bigg].
\end{split}
\end{equation}
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\noindent Spontaneous parametric down-conversion (SPDC), spontaneous four-wave mixing (SFWM), and frequency conversion (FC) are three of the most common nonlinear
processes used in quantum optics. In the first two a nonlinear medium is used to convert photons from a pump laser into pairs of quantum correlated photons \cite{christ13}. In the third, the frequency of a photon is increased or decreased after it interacts with a strong pump field inside a nonlinear medium. Pair generation is often modeled in a straightforward manner using the first order of a Dyson series, or the first term in a Taylor-like series that results if consequences of time-ordering are ignored \cite{grice97}. But when the nonlinear interaction becomes sufficiently strong a more involved theoretical treatment is required. In particular, the Taylor-like series is suspect beyond a few orders because the interaction Hamiltonian does not commute with itself at different times.
In previous work \cite{aggie10,aggie11} the Dyson series has been used to investigate parametric down-conversion, and it was shown that when the initial state of the down-converted modes is vacuum (SPDC), the Taylor-like and Dyson series give identical results up to second order. In a subsequent study, Christ et al. \cite{christ13n} used the Heisenberg picture to investigate time ordering issues in both SPDC and FC. They noticed that when the pump is approximated as undepleted and treated classically, the Hamiltonian is quadratic in the bosonic operators of the down-converted fields, and thus the equations of motion of the operators are necessarily \textit{linear.} This implies that the outgoing fields are related to the incoming fields by a linear Bogoliubov transformation \cite{Braunstein05}. They used this observation to build an \textit{ansatz} solution, and develop a numerical algorithm that permits the calculation of the linear transformation relating output to input. Another consequence of the linear Bogoliubov relation between output and input is that the state of the down-converted photons in SPDC cannot be more complicated than a two-mode squeezed vacuum. This is at odds with a result of Branczyk et al. \cite{aggie11}, who found that the third order correction to the SPDC state contains a six photon energy correlated state. As we shall show, this discrepancy is
related to the fact that the Dyson series does not retain the Gaussian preserving nature of the quadratic Hamiltonian, or equivalently the linear Bogoliubov nature of the transformation relating input and output fields.
In this paper we employ the Magnus Expansion (ME) perturbation series, which is able to deal properly with time ordering at each level, and thus is unitary to all orders. \ Yet, unlike the Dyson series, it respects the Gaussian preserving nature of the process for the class of Hamiltonians that are quadratic in bosonic operators, giving rise to a linear Bogoliubov transformation to whatever order it is taken. Thus we feel it should be the preferred strategy for the description of photon generation and conversion in nonlinear quantum optics.
In Section II we introduce the ME and mention some of its important properties, deriving a useful simplification of its terms when dealing with a broad class of Hamiltonians used in quantum optics; we also show why the ME automatically respects the Gaussian nature of usual nonlinear quantum optical processes. In Section \ref{sec:pdc} we apply the Magnus series generated by the ME to SPDC and SFWM Hamiltonians. While the ME has been used in the past in quantum nonlinear optics (see e.g. Yang et al. \cite{Yang08} and references cited therein), usually only the first term in the expansion has been kept. We calculate the second and third order terms, and also introduce diagrams akin to Feynman diagrams that allow for a simple tabulation of the higher order processes in the ME. Using these diagrams we easily demonstrate why it is necessary to go to third order in the perturbation expansion to see a correction to the SPDC and SFWM states. In section \ref{sec:fc} we extend these ideas to FC processes. We show that the expansion formulas derived for FC are completely analogous to those derived for SPDC and SFWM. In section \ref{sec:dis} we show how the ME can be taken to a form more amenable to direct calculation and develop a perturbation series for the joint spectral amplitude (JSA) of the down converted photons in SPDC and SFWM. In section \ref{sec:broad} we show that for perfect phase matching the second and third order Magnus terms vanish. Finally, conclusions and remarks are presented in section \ref{sec:end}.
\section{Properties of the Magnus expansion}\label{sec:magnus}
\noindent In quantum mechanics it is often necessary to solve the time dependent Schr\"{o}dinger equation for the time evolution operator $\mathcal{U}(t,t_0)$,
\begin{eqnarray}\label{schro}
i\hbar \frac{d}{dt} \mathcal{U} (t,t_0)=H_I(t) \mathcal{U} (t,t_0),
\end{eqnarray}
where $H_I(t)$ is the interaction picture Hamiltonian.
One common way to approximate the solution of Eq. (\ref{schro}) is to use the Dyson series in which the evolution operator is expanded as a power series
\begin{eqnarray}\label{dysondef}
\mathcal{U}(t_1,t_0)&=& \mathbb{I}+T_1+T_2+\ldots\\
&=&\mathbb{I}+(-i)\int_{t_0}^{t_1} dt' \frac{H_I(t')}{\hbar}\nonumber\\
&&+(-i)^2\int_{t_0}^{t_1} dt' \int_{t_0}^{t'} dt'' \frac{H_I(t')H_I(t'')}{\hbar^2}+\ldots \nonumber .
\end{eqnarray}
Often after time-ordering some of the terms appearing, depicted as Feynman diagrams, can be grouped into classes and summed. Here instead we use the Magnus expansion \cite{magnus54,blanes09} in which a series expansion is developed as the argument of an exponential,
\begin{eqnarray}\label{magnusdef}
\mathcal{U}(t_1,t_0) =\exp\big(&&\Omega_1(t_1,t_0)\\
&&+\Omega_2(t_1,t_0)+\Omega_3(t_1,t_0)+\ldots\big), \nonumber
\end{eqnarray}
where unless otherwise indicated we take $t_0 \to -\infty$ and $t_1 \to \infty $.
In the following we calculate the terms in such a series for a rather broad class of Hamiltonians relevant to nonlinear quantum optics. We write write the time dependence of the interaction Hamiltonian as
\begin{eqnarray}\label{hamiltonian}
H_I(t)=\hbar \int d \boldsymbol{\omega} \left( h(\boldsymbol{ \omega}) e^{i \Delta t}+ h^\dagger(\boldsymbol{ \omega}) e^{-i \Delta t} \right),
\end{eqnarray}
where we use the shorthand notation
\begin{eqnarray}
h(\boldsymbol{ \omega})&=&h(\omega_1,\omega_2,\ldots,\omega_n),\\
d\boldsymbol{ \omega}&=&d \omega_1 d \omega_2 \ldots d\omega_n,\nonumber\\
\Delta&=& \Delta(\omega_1,\omega_2,\ldots,\omega_n).\nonumber
\end{eqnarray}
For the moment we leave the integration limits of the frequencies $\boldsymbol{\omega}$ unspecified; in the next sections we will specify them for particular cases.
\indent For the first order Magnus term one simply finds
\begin{eqnarray}\label{Omega1}
\Omega_1&=&-\frac{i}{\hbar} \int dt H_I(t)\\
&=&-2 \pi i \int d \boldsymbol{\omega} \left(h(\boldsymbol{ \omega}) +\text{h.c.} \right) \delta(\Delta).
\end{eqnarray}
In the last equation we have adopted the convention that whenever the lower (or upper) limit of an integral over time is not specified it is meant to be understood as $-\infty$ (or $+\infty$). The term \ref{Omega1}, exponentiated, corresponds to the sum of the Taylor-like series for $\mathcal{U}$ that could be constructed by time-ordering the series (\ref{dysondef}) and ignoring the fact that the Hamiltonians $H_I(t)$ at different times $t$ need not commute. In fact, if $[H_I(t),H_I(t')]=0$ for all $t$ and $t'$, then $\exp \Omega_1$ constitutes the \emph{exact} solution of the problem, since all the higher order Magnus terms vanish. Hence we refer to terms in the Magnus expansion expansion beyond $\Omega_1$ as ``time ordering corrections''.
In general no simplifications can be made to the higher order Magnus terms if no further assumptions are made about $H_I(t)$. We make the following:
\begin{eqnarray}\label{comm}
[ h(\boldsymbol{ \omega}), h(\boldsymbol{ \omega}')]=[ h^\dagger(\boldsymbol{ \omega}), h^\dagger(\boldsymbol{ \omega}')]=0.
\end{eqnarray}
which is indeed valid for the Hamiltonians usually employed to describe both type I and II parametric down conversion, four wave mixing, and frequency conversion, either if the pump is treated classically or if it is treated quantum mechanically.
Using property (\ref{comm}) and the fact that
\begin{eqnarray}\label{timeorder2}
\int dt \int ^{t} dt' e^{i \Delta t-i \Delta'
t'}&=&2 \pi^2 \delta(\Delta) \delta(\Delta')\\
&&+2 \pi i \delta(\Delta-\Delta') \frac{\mathcal{P}}{ \Delta} , \nonumber
\end{eqnarray}
with $\mathcal{P}$ indicating a principal value, the second order Magnus term, generally given by
\begin{eqnarray}
\Omega_2=\frac{(-i)^2}{2 \hbar^2}\int dt \int ^{t} dt' \left[H_I(t),H_I(t') \right],
\end{eqnarray}
can be simplified to
\begin{eqnarray}\label{magnus2}
\Omega_2=-2 \pi i \fint d \boldsymbol{\omega} d \boldsymbol{\omega}'
\left[ h(\boldsymbol{\omega}), h^\dagger(\boldsymbol{\omega}') \right] \frac{\delta(\Delta-\Delta')}{\Delta}.
\end{eqnarray}
In the last equation we have used the principal value integral symbol $\fint$. This is necessary because the result (\ref{timeorder2}) is only valid under a principal value integration sign. Only the second term of (\ref{timeorder2}) contributes to the second order term. This is to be expected since the first term is just
\begin{eqnarray}\label{to2}
2 \pi^2 \delta(\Delta)\delta(\Delta') = \frac{1}{2} \int dt \int dt' e^{i \Delta t-i \Delta' t'},
\end{eqnarray}
which is the expression that would be obtained from (\ref{timeorder2}) ignoring the time ordering; note the difference in limits between Eq. (\ref{to2}) and Eq. (\ref{timeorder2}).
Now let us look at the third order Magnus term, which in general
\begin{align}
&\Omega_3=\frac{(-i)^3}{6 \hbar^3}\int dt \int ^{t} dt' \int ^{t'} dt'' \times \\
&\left( \left[H_I(t),\left[H_I(t'),H_I(t'') \right] \right]+\left[\left[H_I(t),H_I(t') \right],H_I(t'') \right] \right). \nonumber
\end{align}
Here the relevant time integrals take the form
\begin{eqnarray}\label{triple}
\int &dt& \int ^{t} dt' \int ^{t'} dt'' e^{i \mu t+i \nu t'+i \xi t''}\\
&=&2 \pi ^3 \delta \left(\mu \right) \delta \left(\nu\right) \delta \left(\xi\right)\nonumber\\
&&-2 \pi ^2 i \delta \left(\nu+\xi\right) \delta \left(\mu+\nu+\xi\right)\frac{\mathcal{P}}{\xi}\nonumber\\
&&+2 \pi ^2 i \delta \left(\mu+\nu\right) \delta \left(\mu+\nu+\xi\right)\frac{\mathcal{P}}{\mu}\nonumber\\
&&-2 \pi \delta \left(\mu+\nu+\xi\right)
\frac{\mathcal{P}}{\mu+\nu} \frac{\mathcal{P}}{\mu}.\nonumber
\end{eqnarray}
Because of the condition (\ref{comm}) the only term that contributes to $\Omega_3$ in (\ref{triple}) is the last one, and then upon relabeling one obtains:
\begin{eqnarray}\label{mag3}
\Omega_3=-2 \pi i \fint && d \boldsymbol{\omega} d \boldsymbol{\omega}' d \boldsymbol{\omega}'' \left([h(\boldsymbol{\omega}),[h(\boldsymbol{\omega}'), h^\dagger(\boldsymbol{\omega}'')]]+\text{h.c.} \right)\nonumber\\
&& \times \frac{ \delta \left(\Delta+\Delta'-\Delta''\right)}{3} \left(\frac{1}{\Delta \Delta''}+\frac{1}{\Delta \Delta'} \right).
\end{eqnarray}
We emphasize that the simple form obtained for $\Omega_3$ critically depends on assumption (\ref{comm}). If this did not hold, then the second and third terms in (\ref{triple}) would also contribute, leading to a much more complicated expression. However, even if (\ref{comm}) were not to hold the first term of (\ref{triple}) would never contribute to $\Omega_3$, which is related to the fact that it is proportional to terms that would arise if time order were ignored, and their effects have all been included in $\Omega_1$.
Finally, let us note that the Magnus expansion is more robust in preserving qualitative features of the exact solution for $\mathcal{U}$ than the Dyson expansion, particularly for problems in nonlinear quantum optics. Generally, of course, the Magnus expansion truncated at any level will always give rise to a unitary evolution operator if the Hamiltonian is Hermitian, as is assumed when writing (\ref{hamiltonian}), while a truncated Dyson expansion will not. More specifically, the Magnus expansion truncated at any level will always give rise, in the Heisenberg picture, to a linear Bogoliubov transformation of the bosonic operators if the Hamiltonian is quadratic in them. Since such quadratic bosonic Hamiltonians (QBH) often appear in nonlinear quantum optics, and since for such a Hamiltonian the exact solution for the Heisenberg picture evolution of the bosonic operators indeed takes the form of a linear Bogoliubov transformation \cite{Braunstein05}, this is a very desirable feature of the Magnus approximation. Here is the proof that any Magnus approximation displays this feature: From the general commutator identity
\begin{eqnarray}
[A B, C D]&=&[A,C]BD+A[B,C]D\\
&&+C[A,D]B+CA[B,D]. \nonumber
\end{eqnarray}
we see that if $A,B,C,D$ are all bosonic operators then the commutators on the right hand side are all $c-$numbers, and in fact if they do not vanish they are delta functions, either Kronecker or Dirac. Since all the terms $\Omega_1, \Omega_2$ etc., in the Magnus series involve commutators, it is clear that for a QBH all the terms $\Omega_i$ will be quadratic functions of the bosonic operators, and the $M^{\text{th}}$ order approximation to the Magnus expansion
\begin{eqnarray}\label{orderm}
\mathcal{U}_M(t_f,t_0)&=&\exp(\Omega_1(t_f,t_0)+\Omega_2(t_f,t_0)+\nonumber\\
&&\ldots+\Omega_M(t_f,t_0))\nonumber\\
&=&\exp(\mathcal{O}_M(t_f,t_0)),
\end{eqnarray}
where we reinstate general initial and final times, $t_0$ and $t_f$ respectively, is characterized by an operator $\mathcal{O}(t_f,t_0)$ that is anti-Hermitian and a quadratic function of the boson operators. Now in the Heisenberg picture and initial operator $A(t_0)$ evolves to a later operator $A(t)$ according to
\begin{eqnarray}
A(t_f)&=&\mathcal{U}^\dagger(t_f,t_0) A(t_0) \mathcal{U}(t_f,t_0).
\end{eqnarray}
Using the approximation given by (\ref{orderm}) and the Baker-Campbell-Hausdorff (BCH) formula
\begin{eqnarray}\label{BCH}
e^{X}Y e^{-X} &=&Y+\left[X,Y\right]+\frac{1}{2!}[X,[X,Y]]+\cdots ,
\end{eqnarray}
we find
\begin{eqnarray}\label{transinout}
A(t_f)&=&A(t_0)+[A(t_0),\mathcal{O}_M(t_f,t_0)]\\
&&+\frac{1}{2!}[[A(t_0),\mathcal{O}_M(t_f,t_0)],\mathcal{O}_M(t_f,t_0)]+\ldots \nonumber.
\end{eqnarray}
Using the general commutator identity
\begin{eqnarray}
[A,BC]=[A,B]C+B[A,C],
\end{eqnarray}
we see that if $A(t_0)$ is a bosonic operator the commutators on the right hand side of (\ref{transinout}) will all be bosonic operators, and thus the Heisenberg bosonic operators at $t_f$ are indeed linear functions of the Heisenberg bosonic operators at $t_0$. This completes the proof. In the Schr\"{o}dinger picture, the result implies that any Magnus approximation for evolution under a QBH will share the property of the exact solution that the general Gaussian nature of an initial state will be preserved; that is, the exact evolution constitutes a unitary Gaussian channel, and any Magnus approximation for the evolution constitutes a unitary Gaussian channel as well\cite{eisert05}.
\section{Time ordering in Spontaneous Parametric Down Conversion and Four Wave Mixing}\label{sec:pdc}
\noindent In type II SPDC an incoming photon at frequency $\omega_p$ is transformed into two photons of different polarizations at frequencies $\omega_a$ and $\omega_b$ which are roughly half the frequency of the incoming photon. Within the usual rotating wave and undepleted pump approximations, the interaction Hamiltonian governing this process in an effectively one-dimensional structure with all relevant modes propagating in the same direction, such as a channel or ridge waveguide, is (see Appendix \ref{chi2app} for a derivation)
\begin{eqnarray}\label{pdc}
H_I(t)=-\hbar \varepsilon \int && d\omega_a d\omega_b d\omega_p e^{i \Delta t} \Phi(\omega_a,\omega_b,\omega_p) \nonumber\\
&&\times \alpha(\omega_p) a^\dagger(\omega_a) b^\dagger (\omega_b )+\text{h.c.},
\end{eqnarray}
\begin{eqnarray}
\Delta=\omega_a+\omega_b-\omega_p.
\end{eqnarray}
where $a(\omega_a)$ and $b(\omega_b)$ are photon destruction operators for modes associated with the wave vectors $k_a(\omega_a)$ and $ k_b(\omega_b)$, constructed to satisfy the commutation relations
\begin{eqnarray}\label{commspdc}
[a(\omega),a^\dagger(\omega')]&&=[b(\omega),b^\dagger(\omega')]=\delta(\omega-\omega'),
\end{eqnarray}
and all other commutators between destruction and creation operators
vanishing; $\alpha \left( \omega _{p}\right) $ is a classical function
proportional to the electric field amplitude of the undepleted pump mode
with wave vector $k(\omega _{p})$. The assumption (\ref{commspdc}) is well justified
because for type II SPDC the down converted photons appear in orthogonal
polarizations, and their frequency support is far away from the frequency
support of the pump mode characterized by $\alpha \left( \omega _{p}\right) $. Because of this the limits of integration for the frequencies of the modes will range from 0 to $\infty$. Nevertheless, since the bandwidth of the photons is much smaller than their central frequencies these integrals can be safely extended in their lower limit to $-\infty$.
The phase matching function $\Phi$ incorporates the spatial dependence of the modes and the nonlinearity profile; It takes the form
\begin{eqnarray}\label{pm1}
\Phi(\omega_a,\omega_b,\omega_p) = \text{sinc} \left(\frac{\Delta k L}{2} \right),\\
\Delta k =k_a(\omega_a)+k_b(\omega_b)-k_p(\omega_p),
\end{eqnarray}
when the nonlinearity is assumed to be uniform in a region of space from $-L/2$ to $L/2$, and in Eq. (\ref{pdc}) $\varepsilon$ is a dimensionless constant incorporating parameters such as the strength of the nonlinearity profile and the effective cross sectional area of the crystal.
There will be corrections to the integrand of (\ref{pdc}) depending in a benign way on the frequencies \cite{grice97}, and for simplicity we neglect them here.\\
In SFWM a pair of photons from a pump mode at frequency $\omega _{p}$ are
converted into a pair of photons of different frequencies $\omega _{p}\pm
\Delta \omega _{p}$. The description of this process is inherently more
complicated than that of SPDC, but we show in Appendix C that if the pump
field is treated classically and undepleted, then the Hamiltonian governing the process is formally identical to (\ref{pdc}) if the effects of group velocity dispersion on the propagation of the pump pulse can be neglected. \ In that limit we find
\begin{eqnarray}
H_{I}(t)=-\hbar \varepsilon \int && d\omega _{a}d\omega _{b}d\omega _{+}e^{i\tilde{\Delta}t}\tilde{\Phi}(\omega _{a},\omega _{b},\omega _{+}) \nonumber\\
&&\times \tilde{\alpha}(\omega _{+})a^{\dagger }(\omega _{a})b^{\dag }(\omega _{b})+\text{h.c.},
\label{HISFWM}
\end{eqnarray}
where here $\tilde{\Delta}=\omega _{a}+\omega _{b}-\omega _{+}$, and
\begin{eqnarray}
\tilde{\Phi}(\omega _{a},\omega _{b},\omega _{+}) =\text{sinc}\left( \frac{\Delta k L}{2}\right) , \\
\Delta k = k_{a}(\omega _{a})+k_{b}(\omega _{b})-2k_{p}\left( \frac{\omega_{+}}{2}\right) ,
\end{eqnarray}
and now $\tilde{\alpha}(\omega _{+})$ is proportional to the square of the
amplitude of the electric field of the pump. Again we have assumed that
the generated photons at $\omega _{a}$ and $\omega _{b}$ are characterized
by distinct polarizations and/or frequency ranges -- the latter, for
example, might arise due to phase matching constraints \cite{moreno12} -- and thus the only nonvanishing commutation relations
are given by (\ref{commspdc}).
While the approach introduced in this paper can be applied to more
complicated scenarios for SFWM, we restrict ourselves here to situations
where (\ref{HISFWM}) is applicable; then the analysis of SPDC and SFWM is
essentially the same. \ The interaction Hamiltonian $H_{I}(t)$ takes the
form (\ref{hamiltonian}) with
\begin{eqnarray}\label{Fdef}
h(\omega _{a},\omega _{b},\omega _{p}) &=&F(\omega _{a},\omega _{b},\omega_{p})a^{\dagger }(\omega _{a})b^{\dagger }(\omega _{b}), \\
F(\omega _{a},\omega _{b},\omega _{p}) &=&-\varepsilon \alpha (\omega_{p})\Phi (\omega _{a},\omega _{b},\omega _{p}),
\end{eqnarray}%
for SPDC; for SFWM we simply replace $\omega _{p}$ by $\omega _{+}$, $\alpha
(\omega _{p})$ by $\tilde{\alpha}(\omega _{+})$, and $\Phi (\omega
_{a},\omega _{b},\omega _{c})$ by $\tilde{\Phi}(\omega _{a},\omega
_{b},\omega _{+})$. \ So although in the calculations below we explicitly
use the notation for SPDC, the generalization to SFWM is trivial.
With the notation introduced in the paragraphs above, the first order Magnus term is
\begin{eqnarray}
\Omega_1=-2 \pi i \int d\omega_a d\omega_b \big( && J_1(\omega_a,\omega_b)a^\dagger(\omega_a) b^\dagger (\omega_b)\\
&&+ \text{h.c.} \big), \nonumber
\end{eqnarray}
with
\begin{eqnarray}\label{J1pdc}
J_1(\omega_a,\omega_b)&=&F(\omega_a,\omega_b,\omega_a+\omega_b).
\end{eqnarray}
This process can be represented by the diagram (a) in Fig. \ref{fdiag} in which a photon of frequency of $\omega_p$ (or two photons of frequency $\omega_p$ for SFWM) is converted to two photons of frequency $\omega_a$ and $\omega_b$; we associate it with the term $ J_1\left(\omega _a,\omega _b\right) a^\dagger (\omega _a)b^\dagger (\omega _b)$. To have a Hermitian Hamiltonian the reverse process needs to be possible and this is associated with $ \bar{J}_1\left(\omega _a,\omega _b\right) a (\omega _a) b (\omega _b)$ where $\bar{x}$ is the complex conjugate of $x$.
\begin{figure}
\centering
\includegraphics[width=0.40\textwidth]{magnusdiag.pdf}
\caption{\label{fdiag} Diagrams representing the first, second and third order Magnus terms for SPDC. Dashed lines are used to represent pump photons, full lines are used to represent lower energy down converted ones. (a) depicts a photon of frequency $\omega_p$ being converted to two photons of frequencies $\omega_a$ and $\omega_b$. (b) depicts the second order Magnus term in which one of the photons from a down-conversion event is, with the help of a low energy photon previously present, up converted to a pump photon. Finally, (c) depicts the third order Magnus term in which a pair of photons from two previous down-conversion events is converted to a pump photon.
}
\end{figure}
We note that with the approximation of keeping only the first term in the Magnus expansion, as well as the assumption that its effect is small, we have
\begin{eqnarray}
\ket{\psi(\infty)}&=&\mathcal{U} \ket{\psi(-\infty)}\approx e^{\Omega_1}\ket{\psi(-\infty)}\nonumber \\
&=&(\mathbb{I}+\Omega_1) \ket{\psi(-\infty)} =\mathbb{I}+T_1 \ket{\psi(-\infty)},
\end{eqnarray}
where $T_1$ is the first term of the Dyson series (\ref{dysondef}), this is, as shown in the last equation, identical to the first order Magnus term $\Omega_1$.
For the second order Magnus term we need the following:
\begin{eqnarray}
\left[ h(\boldsymbol{\omega}), h^\dagger (\boldsymbol{\omega}') \right]=&-&\delta(\omega _b'-\omega _b)F\left(\omega _a,\omega _b,\omega _p\right) \times \\
&&\bar{F}\left(\omega _a',\omega _b',\omega _p'\right) a ^{\dagger } (\omega _a)a(\omega _a')\nonumber\\
&-&\delta(\omega _a'-\omega _a)F\left(\omega _a,\omega _b,\omega _p\right) \times \nonumber\\
&&\bar{F}\left(\omega _a',\omega _b',\omega _p'\right) b ^{\dagger }(\omega _b)b(\omega _b')\nonumber.
\end{eqnarray}
After some relabeling the second order Magnus term is
\begin{eqnarray}\label{fun2}
\Omega_2=-2 \pi i \Big( \int && d\omega_a d\omega_a' G_2^a(\omega_a ,\omega_a') a ^{\dagger}\left(\omega_a\right)a\left(\omega_a'\right) \\
&&+\int d\omega_b d\omega_b' G_2^b(\omega_b ,\omega_b') b^{\dagger}\left(\omega_b\right) b\left(\omega_b'\right)\Big),\nonumber
\end{eqnarray}
with
\begin{eqnarray}\label{g2}
G_2^a(\omega_a&,&\omega_a') =
\int d\omega_{d} \fint \frac{d\omega_p}{\omega_p} \times \\
&& F\left(\omega_a,\omega _d,\omega _a+\omega _p+ \omega _d \right) \bar{F}\left(\omega _a',\omega _d,\omega _a'+\omega _p+\omega _d\right), \nonumber
\end{eqnarray}
\begin{eqnarray}\label{h2}
G_2^b(\omega_b&,&\omega_b')= \int d\omega_c \fint \frac{d\omega_p}{\omega_p} \times\\
&& F\left(\omega _c,\omega _b,\omega _b+\omega_c+\omega _p\right) \bar{F}\left(\omega _c,\omega _b',\omega _b'+\omega _p+\omega_c\right). \nonumber
\end{eqnarray}
The last expression shows that the second order term corresponds to a frequency conversion process in which a photon of frequency $\omega_a'$ ($\omega_b'$) is converted to one of frequency $\omega_a$ ($\omega_b$) with probability amplitude $G_2^a (G_2^b)$. If one recognizes that $F$ is associated with down conversion and $\bar{F}$ with up conversion, then Eq. \ref{g2} (\ref{h2}) can be read as follows: first a photon at the pump energy (or two photons for the case of SFWM) is down converted to two photons of energies $\omega_a$ ($\omega_b$) and $\omega_d (\omega_c)$. Later the photon of energy $\omega_d (\omega_c)$ interacts with a photon of energy $\omega_a'$ ($\omega_b'$) to be up converted to another photon of energy close to the pump energy (or a pair of photons for SFWM), which effectively gives a frequency conversion process from $\omega_a'$($\omega_b'$) to $\omega_a$ ($\omega_b$). This is sketched in diagram (b) in Fig. \ref{fdiag}.
For the third order Magnus term we need to calculate the following:
\begin{eqnarray}\label{mag3p}
[h(\boldsymbol{\omega})&,&[h(\boldsymbol{\omega}'), h^\dagger(\boldsymbol{\omega}'')]]=\\
&& F\left(\omega _a,\omega _b,\omega _p\right) F\left(\omega _a',\omega _b',\omega _p'\right) \bar{F}\left(\omega _a'',\omega _b'',\omega _p''\right)\times \nonumber\\
&& \Big(\delta \left(\omega _a''-\omega _a'\right) \delta \left(\omega _b''-\omega _b\right)
b^{\dagger }\left(\omega _b'\right)a^{\dagger }\left(\omega _a\right)\nonumber\\
&&
+\delta
\left(\omega _a''-\omega _a\right) \delta \left(\omega _b''-\omega _b'\right) b^{\dagger }\left(\omega
_b\right)a^{\dagger }\left(\omega _a'\right)\Big). \nonumber
\end{eqnarray}
Using (\ref{mag3}), after some rearranging of the dummy integration variables we obtain
\begin{eqnarray}
\Omega_3&&=-2 \pi i \int d\omega_a d\omega_b a^\dagger(\omega_a) b^\dagger(\omega_b) \\
&&\fint d\omega_c d\omega_d d\omega_p d\omega_q \frac{ \bar{F}\left(\omega _d,\omega _c,-\omega _a-\omega _b+\omega _p+\omega _q\right)}{3}\nonumber\\
&&\times F\left(\omega _a,\omega _c,\omega _p\right) F\left(\omega _d,\omega _b,\omega _q\right) \Big\{\nonumber\\
&&\frac{2}{\left(\omega _a+\omega _c-\omega _p\right) \left(\omega _b+\omega _d-\omega _q\right)}\nonumber\\
&&+ \frac{1}{\left(\omega _a+\omega _c-\omega _p\right) \left(\omega _a+\omega _b+\omega _c+\omega_d-\omega _p-\omega _q\right)} \nonumber\\
&& +\frac{1}{\left(\omega _b+\omega _d-\omega _q\right) \left(\omega _a+\omega _b+\omega _c+\omega_d-\omega _p-\omega _q\right)} \nonumber
\Big\}\nonumber.
\end{eqnarray}
Switching $\omega_p \to \omega_p+\omega_a+\omega_c$ and $\omega_q \to \omega_q+\omega_b+\omega_d$ the innermost integral becomes
\begin{eqnarray}\label{bpv}
&&\int d\omega_c d\omega_d \fint d\omega_p d\omega_q \frac{ \bar{F}\left(\omega _d,\omega _c,\omega _c+\omega _d+\omega_p+\omega _q\right)}{3} \nonumber\\
&&\times F\left(\omega _a,\omega _c,\omega _a+\omega _c+\omega _p\right) F\left(\omega _d,\omega _b,\omega _b+\omega _d+\omega _q\right)\nonumber\\
&&\times \Big\{\frac{2}{\omega _p \omega _q}+\frac{1}{\omega_p+\omega_q}\left(\frac{1}{\omega_p}+\frac{1}{\omega_q} \right) \Big\}.
\end{eqnarray}
At this point we use the identity \cite{barnett02}
\begin{eqnarray}
\frac{\mathcal{P}}{\omega_p+\omega_q}\left(\frac{\mathcal{P}}{\omega_p}+ \frac{\mathcal{P}}{\omega_q} \right)= \frac{\mathcal{P}}{\omega_p} \frac{\mathcal{P}}{\omega_q}+\pi^2 \delta(\omega_p) \delta(\omega_q),
\end{eqnarray}
which is valid under the assumption that the test function over which the principal values are being calculated is well behaved. We can then write
\begin{eqnarray}
\Omega_3=-2 \pi i \int d\omega_a d\omega_b \big(&&J_3(\omega_a,\omega_b) a^\dagger(\omega_a)b^\dagger(\omega_b)\\
&&+\text{h.c.}\big), \nonumber
\end{eqnarray}
with
\begin{eqnarray}\label{J_3}
&& J_3(\omega_a,\omega_b)=\int d\omega_c d\omega_d \Big\{ \frac{\pi^2}{3}
\bar{F}\left(\omega _c,\omega _d,\omega _c+\omega _d\right) \times \\
&& F\left(\omega _a,\omega _d,\omega _a+\omega _d\right) F\left(\omega _c,\omega _b,\omega _b+\omega_c\right) \nonumber\\
&&+\fint \frac{d \omega_p}{\omega_p} \frac{d \omega_q}{\omega_q} \bar{F}\left(\omega _c,\omega _d,\omega _c+\omega _d+\omega _p+\omega _q\right) \times \nonumber\\
&& F\left(\omega _a,\omega _d,\omega _a+\omega _d+\omega _p\right) F\left(\omega _c,\omega _b,\omega _b+\omega_c+\omega _q\right) \Big\}. \nonumber
\end{eqnarray}
As before, a diagram representing the third order processes can be drawn. This corresponds to two down conversion processes in which one of the converted photons from each process is later used for up-conversion, as is seen in diagram (c) of Fig. \ref{fdiag}.
Note that the treatment for type I SPDC follows easily from what was done above. One needs only to replace $b$ by $a$ to characterize the fact that
both down converted photons are generated with the same polarization and in the same frequency range; the quantity $F$ in (\ref{Fdef}) will be symmetric in the first two arguments, $F(x,y,z)=F(y,x,z)$, and the expression for $\varepsilon $ will differ by a factor of two (see Appendix B).
Finally, let us mention that we have developed a numerical library \cite{github} that uses the \verb|cubature| software package \cite{cubature} to efficiently calculate the different Magnus terms for SPDC and SFWM within the approximations introduced in this paper.
\section{Time ordering in Frequency Conversion}\label{sec:fc}
\noindent Frequency conversion (FC) is another nonlinear process in which three fields interact. In this section we explicitly consider FC employing a $\chi_2$ nonlinearity, but the extension to treat FC by a $\chi_3$ nonlinearity follows in the same way the description of photon generation by SFWM followed from the description of photon generation by SPDC/ For a $\chi_2$ nonlinearity, FC occurs when a pump photon with frequency $\omega_p$ fuses with a photon with frequency $\omega_a$ to create a higher energy photon with frequency $\omega_b=\omega_a+\omega_p$. The Hamiltonian governing this process is derived in a completely analogous manner to the way (\ref{pdc}) is derived in Appendix \ref{chi2app}. The Hamiltonian governing the process is
\begin{eqnarray}\label{fc}
H_I(t)=-\hbar \varepsilon \int && d\omega_a d\omega_b d\omega_p e^{i \Delta t} \Phi(\omega_a,\omega_b,\omega_p) \nonumber\\
&&\times \alpha(\omega_p) a(\omega_a) b^\dagger (\omega_b )+\text{h.c.},
\end{eqnarray}
where all the variables have the same meaning as in Eq. (\ref{pdc}) except that now the phase matching function for uniform non-linearity in the region $-L/2$ to $L/2$ is given by
\begin{eqnarray}\label{pm2}
\Phi(\omega_a,\omega_b,\omega_p) = \text{sinc}\left(\frac{\Delta k L}{2} \right),\\
\Delta k =k_b(\omega_b)-k_a(\omega_a)-k_p(\omega_p),
\end{eqnarray}
and
\begin{eqnarray}
\Delta=\omega_b-\omega_a-\omega_p.
\end{eqnarray}
The interaction Hamiltonian (\ref{fc}) can be rewritten as
\begin{eqnarray}\label{notationfc}
H_I(t)&=& \int d\omega_a d\omega_b d\omega_p \left( e^{i \Delta t} h(\omega_a,\omega_b,\omega_p)+ \text{h.c.} \right),
\end{eqnarray}
with
\begin{eqnarray}
h(\omega_a,\omega_b,\omega_p)=a (\omega _a)b^\dagger (\omega _b) F\left(\omega _a,\omega _b,\omega _p\right),
\end{eqnarray}
\begin{eqnarray}\label{f2}
F(\omega_a,\omega_b,\omega_p)=-\varepsilon \alpha(\omega_p) \Phi(\omega_a,\omega_b,\omega_p).
\end{eqnarray}
In this case the first Magnus term is
\begin{eqnarray}\label{Omega1fc}
\Omega_1=- 2 \pi i\int d\omega_a d\omega_b \big(&& J_1(\omega_a,\omega_b)a(\omega_a) b^\dagger (\omega_b)\\
&&+ \text{h.c.} \big). \nonumber
\end{eqnarray}
with
\begin{eqnarray}\label{magnus1fc}
J_1(\omega_a,\omega_b)&=&F(\omega_a,\omega_b,\omega_b-\omega_a).
\end{eqnarray}
For the higher order terms we find:
\begin{eqnarray}\label{fun2fc}
\Omega_2=-2 \pi i \Big( \int && d\omega_a d\omega_a' G_2^a(\omega_a ,\omega_a') a ^{\dagger}\left(\omega_a\right)a\left(\omega_a'\right) \\
&&+\int d\omega_b d\omega_b' G_2^b(\omega_b ,\omega_b') b^{\dagger}\left(\omega_b\right) b\left(\omega_b'\right)\Big)\nonumber.
\end{eqnarray}
with
\begin{eqnarray}\label{g2fcn}
G_2^a(\omega_a&,&\omega_a') =
-\int d\omega_{q} \fint \frac{d\omega_p}{\omega_p} \times \\
&& F\left(\omega_a',\omega _q,\omega _q-\omega _a'+\omega _p \right) \bar{F}\left(\omega _a,\omega _q,\omega _q-\omega _a+\omega _p\right), \nonumber\\
G_2^b(\omega_b&,&\omega_b')= \int d\omega_q \fint \frac{d\omega_p}{\omega_p} \times\\
&& F\left(\omega _q,\omega _b, \omega _b-\omega_q+\omega _p\right) \bar{F}\left(\omega _q,\omega _b',\omega _b'+\omega _p-\omega_q\right), \nonumber
\end{eqnarray}
and
\begin{eqnarray}
\Omega_3=-2 \pi i \int d\omega_a d\omega_b \left(J_3(\omega_a,\omega_b) a (\omega_a)b^\dagger(\omega_b)+\text{h.c.}\right),
\end{eqnarray}
with
\begin{eqnarray}
&& J_3(\omega_a,\omega_b)=\int d\omega_c d\omega_d \Big\{ \frac{-\pi^2}{3}
\bar{F}\left(\omega _c,\omega _d,\omega _d-\omega _c\right) \times \\
&& F\left(\omega _a,\omega _d,\omega _d-\omega _a\right) F\left(\omega _c,\omega _b,\omega _b-\omega_c\right) \nonumber\\
&&+\fint \frac{d \omega_p}{\omega_p} \frac{d \omega_q}{\omega_q} \bar{F}\left(\omega _c,\omega _d,\omega _d-\omega _c+\omega _p-\omega _q\right) \times \nonumber\\
&& F\left(\omega _a,\omega _d,\omega _d-\omega _a-\omega _q\right) F\left(\omega _c,\omega _b,\omega _b-\omega_c+\omega _p\right) \Big\}. \nonumber
\end{eqnarray}
In the last equation it has been assumed again that the function $F$ is well behaved.
Note the very elegant connection the Magnus series provides between FC and SPDC. To obtain the hamiltonian of FC from that of SPDC it suffices to make the substitutions $a(\omega_a) \rightarrow a^\dagger(\omega_a), a^\dagger(\omega_a) \rightarrow a(\omega_a)$ and $\omega_a \rightarrow -\omega_a$. The corresponding Magnus terms of FC are obtained by performing exactly the same substitutions in the SPDC terms, and rearranging the sign of the variables in the principal value integrals over $\omega_q$ and $\omega_p$.
\section{Disentangling the Magnus expansion}\label{sec:dis}
\noindent The calculations for photon generation used the fact that the commutator of two two-mode squeezing Hamiltonians gives two single-mode frequency conversion terms,
\begin{eqnarray}
&&\left[ a^{\dagger }\left(\omega_p\right) b^{\dagger }\left(\omega _q\right)+\text{h.c.} \ , a^{\dagger }\left(\omega _r\right) b^{\dagger }\left(\omega _s\right)+\text{h.c.} \right]=\\
&& \delta(\omega _q-\omega _s) a^{\dagger }\left(\omega _r\right)a\left(\omega _p\right)+\delta(\omega _p-\omega _r) b^{\dagger }\left(\omega_s\right)b\left(\omega _q\right)-\text{h.c.} \nonumber,
\end{eqnarray}
and the fact that the commutator of a single-mode frequency conversion term with a two-mode squeezing term is again a two-mode squeezing term,
\begin{eqnarray}
&&\left[ a^{\dagger }\left(\omega _s\right) a\left(\omega_r\right)+\text{h.c.} \ , a^{\dagger } \left(\omega _p\right) b^{\dagger }\left(\omega_q\right)+\text{h.c.} \right]=\\
&&\delta(\omega _s-\omega _p) b^{\dagger} \left(\omega _q\right) a^{\dagger }\left(\omega _r\right)+ \delta(\omega _r-\omega_p) b^{\dagger }\left(\omega _q\right)a^{\dagger } \left(\omega _s\right)-\text{h.c.}\nonumber
\end{eqnarray}
Likewise for FC they used the fact that the commutator of two two-mode frequency conversion terms gives single-mode frequency conversion terms,
\begin{eqnarray}
&&\left[ a\left(\omega_p\right) b^{\dagger }\left(\omega _q\right)+\text{h.c.} \ , a\left(\omega _r\right) b^{\dagger }\left(\omega _s\right)+\text{h.c.} \right]=\\
&& \delta(\omega _q-\omega _s) a^{\dagger }\left(\omega _p\right)a\left(\omega _r\right)+\delta(\omega _p-\omega _r) b^{\dagger }\left(\omega_q\right)b\left(\omega _s\right)-\text{h.c.}, \nonumber
\end{eqnarray}
and that the commutator of a two-mode frequency conversion term and a single-mode frequency conversion term gives again a two-mode frequency conversion term,
\begin{eqnarray}
&&\left[ a^{\dagger }\left(\omega _s\right) a\left(\omega_r\right)+\text{h.c.} , a^{\dagger } \left(\omega _p\right) b\left(\omega_q\right)+\text{h.c.} \right]=\\
&&\delta(\omega _s-\omega _p) a^{\dagger}\left(\omega _r\right) b \left(\omega _q\right) + \delta(\omega _r-\omega_p) a^{\dagger } \left(\omega _s\right) b\left(\omega _q\right)-\text{h.c.}\nonumber
\end{eqnarray}
These results can be used to inductively show that the unitary that connects states at $t_0 \to -\infty $ and $t \to \infty$ has to be of the form:
\begin{eqnarray}\label{magnusgen}
\mathcal{U}(t&,&t_0)=\exp\Big(-2 \pi i \int d \omega_1 d\omega_2 \\
&& \Big\{\tilde G^a(\omega_1,\omega_2)a^\dagger(\omega_1) a(\omega_2) +\tilde G^b(\omega_1,\omega_2)b^\dagger(\omega_1) b(\omega_2) \nonumber\\
&&+ (\tilde J(\omega_1,\omega_2)a^\dagger(\omega_1) b^\dagger(\omega_2) +\text{h.c.}) \Big\} \Big) \nonumber,
\end{eqnarray}
with
\begin{eqnarray}
\tilde J=\sum_i J_{2i+1}; \quad \tilde G^a=\sum_i G^a_{2i}; \quad \tilde G^b=\sum_i G^b_{2i},
\end{eqnarray}
for SPDC and SFWM. For FC one only need to change
\begin{eqnarray}\label{change}
a^\dagger(\omega_1) b^\dagger(\omega_2) \Rightarrow a^\dagger(\omega_1) b(\omega_2),
\end{eqnarray}
in Eq. (\ref{magnusgen}).
The result (\ref{magnusgen}) confirms in detail the proof given at the end of section II. Eq. (\ref{magnusgen}) shows the great utility of the Magnus expansion. It provides us with the general form of the solution. One only needs to calculate the higher corrections to the coefficients that appear in front of the squeezing and frequency conversion terms to get an increasingly better approximation to the solution.
Diagrams similar to those presented in Fig. \ref{fdiag} can be drawn for the higher order terms. We emphasize that they are used here just as a convenient way of representing the terms in the Magnus expansion; they are \emph{not} Feynman diagrams, of course, and we have not provided any rule to actually do calculations using them. Yet they can help in understanding the physics. In particular, given that the amplitudes they represent will always be multiplied by only pairs of bosonic operators, no matter what the order of the diagram there can be only two free legs at the low down-converted frequencies, which correspond precisely to the frequencies of the bosonic operators that they multiply. The diagrams in Fig. \ref{fdiag} are precisely the three simplest diagrams satisfying this selection rule.
While the form (\ref{magnusdef}) of the unitary operator given by the Magnus expansion identifies the physics of the evolution, particularly as illustrated in the example of (\ref{magnusgen}), it cannot be directly used to calculate the evolution of an initial state, since the terms associated with the functions $J_{2i+1}$ do not commute with the terms associated with the functions $G_{2i}^a$ and $G_{2i}^b$. In this section we will provide a simple strategy to disentangle the Magnus expansion.We will be able to refactorize the unitary (\ref{magnusgen}) as a product of two unitaries. For the SPDC and SFWM Hamiltonians (\ref{pdc},\ref{HISFWM}) these unitaries will correspond to a pure squeezing Hamiltonian and pure single mode frequency conversion
\begin{eqnarray}\label{fact}
\mathcal{U}&=&\mathcal{U}_{\text{sq}} \mathcal{U}_{\text{fc}},\\
\mathcal{U}_{\text{sq}}&=&e^{-2\pi i \int d \omega_a d \omega_b J(\omega_a,\omega_b) a^\dagger(\omega_a) b^\dagger(\omega_b)+\text{h.c.}},\\
\mathcal{U}_{\text{fc}}&=&e^{-2\pi i \int d \omega_a d \omega_a' G(\omega_a,\omega_a') a^\dagger(\omega_a) a(\omega_a')} \times\\
&&e^{-2\pi i \int d \omega_b d \omega_b' G(\omega_b,\omega_b') b^\dagger(\omega_b) b(\omega_b)}\nonumber.
\end{eqnarray}
The case of FC is completely analogous and one only need to follow prescription (\ref{change}).
The factorization (\ref{fact}) has two important features. The first one is that if time ordering is irrelevant then $J=J_1$ as given in Eq. (\ref{J1pdc}) and $G^a=G^b=0$. The second is that when acted on the vacuum the SPDC or SFWM state is simply
\begin{eqnarray}
\ket{\psi}&=&\mathcal{U} \ket{\text{vac}}=\mathcal{U}_{\text{sq}}\mathcal{U}_{\text{fc}}\ket{\text{vac}}=\mathcal{U}_{\text{sq}}\ket{\text{vac}}\\
&=&e^{-2\pi i \int d \omega_a d \omega_b J(\omega_a,\omega_b) a^\dagger(\omega_a) b^\dagger(\omega_b)+\text{h.c.}} \ket{\text{vac}} \nonumber
\end{eqnarray}
that is $J$ is proportional to the joint spectral amplitude (JSA) of the down converted photons.
To obtain the factorization (\ref{fact}) we will use the BCH formula (\ref{BCH}) and the Zassenhaus formula \cite{casas12},
\begin{eqnarray}
e^{X+Y}&=& e^{X} e^{Y} e^{-\frac{1}{2} [X,Y]} e^{\frac{1}{6}(2[Y,[X,Y]]+ [X,[X,Y]] ) \cdots} .
\end{eqnarray}
The lowest order factorization that can be obtained by just going to second order in the Magnus expansion is to write
\begin{eqnarray}
\mathcal{U}=e^{\Omega_1+\Omega_2} \approx e^{\Omega_1}e^{\Omega_2}=\mathcal{U}_{\text{sq}} \mathcal{U}_{\text{fc}}.
\end{eqnarray}
When applied to vacuum one simply obtains
\begin{eqnarray}
\ket{\psi_2}=\mathcal{U} \ket{\text{vac}}=e^{\Omega_1}\ket{\text{vac}},
\end{eqnarray}
that is, up to second order the state is not modified by time ordering (since $\Omega_2\ket{\text{vac}}=0$), reminiscent of what happens when the Dyson series is used for this same problem \cite{aggie11}. Nevertheless, in the Magnus expansion the meaning of the null contribution of the second order in SPDC and SFWM is transparent: this second order term requires at least one down-converted photon to be present and hence it will vanish when applied to the vacuum.\\
\noindent Thus it is clear that to see time-ordering corrections to the expression for the state of the generated photons it is necessary to go to at least to third order. In this case one finds
\begin{eqnarray}\label{spdc3}
e^{\Omega_1+\Omega_2+\Omega_3}&\approx& e^{\Omega_1+\Omega_3} e^{\Omega_2} e^{-\frac{1}{2}[\Omega_1+\Omega_3,\Omega_2]}\\
&\approx& e^{\Omega_1+\Omega_3} e^{\Omega_2} e^{-\frac{1}{2}[\Omega_1,\Omega_2]}\nonumber \\
&=& e^{\Omega_1+\Omega_3} e^{\Omega_2} e^{-\frac{1}{2}[\Omega_1,\Omega_2]}e^{-\Omega_2}e^{\Omega_2}\nonumber \\
&\approx&e^{\Omega_1+\Omega_3} e^{-\frac{1}{2}e^{\Omega_2}[\Omega_1,\Omega_2]e^{-\Omega_2}}e^{\Omega_2}\nonumber \\
&\approx&e^{\Omega_1+\Omega_3} e^{-\frac{1}{2}([\Omega_1,\Omega_2]+[\Omega_2,[\Omega_1,\Omega_2]])}e^{\Omega_2}\nonumber \\
&\approx&e^{\Omega_1+\Omega_3} e^{-\frac{1}{2}[\Omega_1,\Omega_2]}e^{\Omega_2}\nonumber \\
&\approx&e^{\Omega_1+\Omega_3-\frac{1}{2}[\Omega_1,\Omega_2]}e^{\Omega_2}. \nonumber
\end{eqnarray}
Here we ignore any fourth order term, resolved the identity as $\mathbb{I}=e^{-\Omega_2}e^{\Omega_2}$, used the BCH and Zassenhaus identities and the following relation $e^W e^V e^{-W}\equiv e^{e^W V e^{-W}} $.
From the last derivation is seen that to calculate the first non-vanishing correction to the JSA we not only need $\Omega_3$ but also $[\Omega_1,\Omega_2]$. This extra third order correction is easy to understand: The diagram corresponding to $\Omega_2$ needs an incoming photon, which can only be provided by a previous down conversion process event included in $\Omega_1$. This process will be of course of third order since it includes two down conversion events and one up conversion event. With $\Omega_1$ a two mode squeezing operator and $\Omega_2$ containing two single mode frequency converters, their commutator is another two mode squeezing operator:
\begin{eqnarray}
\frac{[\Omega_1, \Omega_2 ]}{2}=-2 \pi && i^2 \int d \omega_a d \omega_b \times \\
&&(K_3(\omega_a,\omega_b) a^\dagger(\omega_a) b^\dagger(\omega_b) +\text{h.c.}) \nonumber,
\end{eqnarray}
with
\begin{eqnarray}
K_3(\omega_a,\omega_b)&=& \pi \int d \omega \Big(G_2^a\left(\omega _a,\omega \right) J_1\left(\omega ,\omega _b\right)\\
&&+J_1\left(\omega _a,\omega \right) G_2^b\left(\omega _b,\omega \right) \Big)\nonumber.
\end{eqnarray}
Eq. (\ref{spdc3}) provides a compact formula accounting for the effects of time ordering, and leads to an identification of the form of the joint spectral amplitude, which to this level of approximation is proportional to
\begin{eqnarray}
J(\omega_a,\omega_b)&=&J_1(\omega_a,\omega_b)\\
&&+\left(J_3(\omega_a,\omega_b)-i K_3(\omega_a,\omega_b)\right) \nonumber,
\end{eqnarray}
where the first contribution $J_1(\omega_a,\omega_b)$ neglects effects of time-ordering, while the additional terms take corrections into account to all orders lower than the fifth (this last fact we shall prove shortly). This illustrates that as the intensity of the pump field increases and the corrections become larger not only are more squeezed photons generated in the same modes, but their spectral properties change. In particular, if the functions characterizing the pump pulse and the phase-matching functions are real, then as the pump intensity is increased the phase structure of the joint spectral amplitude evolves from trivial to nontrivial.
From the general properties of the Magnus expansion, \emph{i.e.} the fact that the $n^{\text{th}}$ order Magnus terms is made of $n-1$ commutators, we know that all even Magnus terms will be pairs of single mode frequency conversion generators, whereas all odd terms will be two mode squeezing operators. This fact can be used to show that the commutator of two Magnus terms $[\Omega_i,\Omega_j]$ (which is an $i+j$ order term) will be a frequency conversion term if $i+j$ is even and a squeezing term if $i+j$ is odd. With these two facts and after some algebra it can be shown, that the extension of (\ref{spdc3}) to fifth order is
\begin{eqnarray}
\mathcal{U}= \exp(\Omega_1+\Omega_2+\Omega_3+\Omega_4+\Omega_5)= \mathcal{U}_{\text{sq}} \mathcal{U}_{\text{fc}},
\end{eqnarray}
where the squeezing and frequency conversion unitaries $\mathcal{U}_{\text{sq}}$, $\mathcal{U}_{\text{fc}}$ are
\begin{eqnarray}
\mathcal{U}_{\text{sq}}&=&\exp(X),\\
X&=&\Omega_1+\Big(\Omega_3-\frac{1}{2}[\Omega_1,\Omega_2]\Big)+\Big(\Omega_5-\frac{1}{2}[\Omega_1,\Omega_4] \\
&&-\frac{1}{2}[\Omega_3,\Omega_2]+\frac{5}{6}[\Omega_2,[\Omega_2,\Omega_1]]+\frac{1}{2}[\Omega_1,[\Omega_1,\Omega_3]]\Big) \nonumber,\\
\mathcal{U}_{\text{fc}}&=&\exp(Y),\\
Y&=&\Omega_2+\Big(\Omega_4+\frac{1}{2}[\Omega_1,[\Omega_1,\Omega_2]]\Big).
\end{eqnarray}
\indent We now consider difference frequency generation, where besides an injected pump pulse that is treated classically we also seed the modes at $\omega_a$ and/or $\omega_b$, with injected pulses described by coherent states. Here the output state can be written as
\begin{eqnarray}
\ket{\psi_3}&=&e^{\Omega_1+\Omega_3-\frac{1}{2}[\Omega_1,\Omega_2]}e^{\Omega_2} \times\\
&&e^{i \sum_{c={a,b}}\left(\int d \omega_c f_c(\omega_c) c^{\dagger}(\omega_c)+\text{h.c.}\right)}\ket{\text{vac}}\nonumber\\
&=& \left(e^{\Omega_1+\Omega_3-\frac{1}{2}[\Omega_1,\Omega_2]} \right)\times\nonumber\\
&&\left(e^{\Omega_2}e^{i \sum_{c={a,b}}\left(\int d \omega_c f_c(\omega_c) c^{\dagger}(\omega_c)+\text{h.c.}\right)}e^{-\Omega_2}\right)\ket{\text{vac}}.\nonumber
\end{eqnarray}
Keeping only terms up to third order, we find
\begin{eqnarray}\label{manexp}
&&e^{\Omega_2}e^{i \sum_{c={a,b}}\left(\int d \omega_c f_c(\omega_c) c^{\dagger}(\omega_c)+\text{h.c.}\right)}e^{-\Omega_2}=\\
&&e^{i \sum_{c={a,b}}\left(\int d \omega_c f_c(\omega_c) e^{\Omega_2}c^{\dagger}(\omega_c)e^{-\Omega_2}+\text{h.c.}\right)}=\nonumber\\
&&e^{i \sum_{c={a,b}}\left(\int d \omega_c f_c(\omega_c) \left(c^{\dagger}(\omega_c)+[\Omega_2,c^{\dagger}(\omega_c)] \right)+\text{h.c.}\right)}\nonumber.
\end{eqnarray}
The commutator inside the exponential can be easily calculated to be
\begin{eqnarray}
[\Omega_2,a^\dagger(\omega_a')]&=&-2 \pi i \int d \omega_a a^{\dagger}(\omega_a) G_2^a(\omega_a,\omega_a'),
\end{eqnarray}
\begin{eqnarray}
[\Omega_2,b^\dagger(\omega_b')]&=&-2 \pi i \int d \omega_b b^{\dagger}(\omega_b) G_2^b(\omega_b,\omega_b').
\end{eqnarray}
This shows that the first effect due to time ordering corrections in stimulated processes is to modify the shape of the injected seed pulses.
Note that although the discussion in this section was phrased in terms of the SPDC and SFWM hamiltonians all its conclusions are directly applicable to the case of FC.
\section{Time ordering corrections in broadly phase-matched Processes}\label{sec:broad}
\noindent Here we consider time ordering corrections in the special limit that the phase matching function is independent of is variables,
\begin{eqnarray}\label{broad}
\Phi(\omega_a,\omega_b,\omega_p) \propto 1,
\end{eqnarray}
in Eqs. (\ref{pm1}) and (\ref{pm2}). This implies that the function $F$ in (\ref{Fdef}) and (\ref{f2}) is a function only of its last argument, $F(x,y,z)=F(z)$.
We show below that if this holds the second and third order Magnus terms vanish. We conjecture that in this limit \emph{all} Magnus terms $\Omega_i$ beyond $\Omega_1$ vanish for PDC and FC; in the special case of FC in which the initial state is a single photon this has been shown to hold\cite{donohue14}. \\
\indent To proceed, we look at $G_2^a(\omega_a, \omega_a')$ as a typical term. For PDC, in the limit (\ref{broad}) it becomes
\begin{eqnarray}
G_2^a(\omega_a&,&\omega_a') =
\int d\omega_{q} \fint \frac{d\omega_p}{\omega_p} \times \\
&& F\left(\omega_a,\omega _q,\omega _a+\omega _p+ \omega _q \right) \bar{F}\left(\omega _a',\omega _q,\omega _a'+\omega _p+\omega _q\right) \nonumber\\
&&= \int d\omega_{q} \fint \frac{d\omega_p}{\omega_p} F\left(\omega _a+\omega _p+ \omega _q \right) \bar{F}\left(\omega _a'+\omega _p+\omega _q\right). \nonumber
\end{eqnarray}
In the last equation we are free to move the origin of the integral in $\omega_q$ by $\omega_p$ to get:
\begin{eqnarray}
G_2^a = \int d\omega_{q} F\left(\omega _a+ \omega _q \right) \bar{F}\left(\omega _a'+\omega _q\right) \fint \frac{d\omega_p}{\omega_p},
\end{eqnarray}
using the identity
\begin{eqnarray}\label{id1}
\fint \frac{dx}{ x+y}= 0,
\end{eqnarray}
the innermost integral is identically zero.
The same kind of argument shows that $G_2^b$ vanishes. Finally, to show that the third order term is also zero we proceed along similar lines. Using expression (\ref{bpv}), with $F(x,y,z)=F(z)$ we obtain:
\begin{eqnarray}
&&\int d\omega_c d\omega_d \fint d\omega_p d\omega_q \frac{
\bar{F}\left(\omega _c+\omega _d+\omega_p+\omega _q\right)
}{3} \nonumber\\
&&\times F\left(\omega _a+\omega _c+\omega _p\right) F\left(\omega _b+\omega _d+\omega _q\right)\nonumber\\
&&\times \Big\{\frac{2}{\omega _p \omega _q}+\frac{1}{\omega_p+\omega_q}\left(\frac{1}{\omega_p}+\frac{1}{\omega_q} \right) \Big\}.
\end{eqnarray}
and as before we shift the integration axes, in this case, $\omega_c \rightarrow \omega_c-\omega_q$, $\omega_d \rightarrow \omega_d-\omega_p$ to get:
\begin{eqnarray}
\int d\omega_c d\omega_d && \bar{F}\left(\omega _c+\omega _d\right) F\left(\omega _a+\omega _d\right) F\left(\omega _b+\omega_c\right) \times \\
&& \fint d \omega_p d \omega_q \left\{ \frac{2}{\omega_p \omega_q}+\frac{1}{\omega_p+\omega_q}\left(\frac{1}{\omega_p}+\frac{1}{\omega_q} \right) \right\}= 0. \nonumber
\end{eqnarray}
\section{Conclusions}\label{sec:end}
\noindent In this paper we have used the Magnus expansion to construct solutions to the time dependent Schr\"{o}dinger equation describing three nonlinear processes governed by Hamiltonians that do not commute at different times: spontaneous parametric down-conversion, spontaneous four wave mixing, and frequency conversion. In the derivation of the Magnus terms for these three processes we found a rather simple form for the second and third order Magnus terms for a broad class of Hamiltonians that satisfy Eq. (\ref{comm}). The Magnus expansion solution that we found for the time evolution operator was shown to be Gaussian preserving; equivalently, when the unitary operator is used to transform input and output bosonic operators the transformation is Bogoliubov linear. With the expansion we developed, it is easy to see why the second order effect due to the time ordering in SPDC vanishes: Essentially the second order process needs to be stimulated by a previous photon. We also showed how to calculate corrections to the joint spectral amplitude of the photons in SPDC and SFWM, and how even when the phase matching function and pump functions are real the JSA acquires a non-trivial phase structure due to time ordering; this structure can be investigated via tomographic methods\cite{sipe14}. We also explored how the effect of time ordering acts in stimulated PDC; in particular, we looked at the case when the lower energy photons are prepared in coherent states and showed that the first correction due to time ordering is to modify their spectral structure. Finally, the Magnus expansion also allowed us to argue that time ordering is irrelevant if the phase matching function is infinitely broad.
\begin{acknowledgments}
N.Q. gratefully acknowledges insightful conversations with J.M. Donohue and A.M. Bra{\'n}czyk and funding from NSERC Canada.
\end{acknowledgments}
|
\subsection{Please Capitalize the First Letter of Each Notional Word in Subsection Title}
\section{Introduction}
\label{sect:intro}
Over the years, a wide range of oscillatory phenomena have been observed in various regions of the Sun. Oscillations in the solar atmosphere have been studied since 1960s (Leighton, Noyes \& Simon \cite{Leighton1962}). These studies have improved our understanding of the internal structure of the Sun as~well~as the dynamic structure of sunspots. Intensity, velocity and magnetic field observations of the Sun in various spectral lines have been used in studying these oscillations (Staude \cite{Staude1999}).
The studies related to the photosphere emphasize on understanding the internal structure of the Sun through acoustic waves, while the studies with the chromospheric lines focus on understanding the propagation of waves to the higher atmosphere, their interaction with the magnetic fields there, and consequently understanding the problem of coronal heating.
The important findings of these studies in magnetic media are: (i) the presence of the photospheric five minute oscillations, and its absorption in the regions of strong photospheric magnetic fields (Braun~et~al. \cite{Braun1987}, \cite{Braun1988}, \cite{Braun1992}, \cite{Braun1993}; Bogdan~et~al. \cite{Bogdan1993}; Hindman \& Brown \cite{Hindman1998}; Kumar~et~al. \cite{Kumar2000}; Jain \& Haber \cite{Jain2002}; Venkatakrishnan~et~al. \cite{PVK2002}), (ii) enhanced oscillations in chromospheric umbra in the three minute band (Bhatnagar \& Tanaka \cite{Bhatnagar1972}; Lites \cite{Lites1986}; Kentischer \& Mattig \cite{Kenti1995}; Nagashima~et~al. \cite{Nagashima2007}), and (iii) running penumbral waves (Zirin \& Stein \cite{Zirin1972}; Giovanelli \cite{Giovanelli1972}, \cite{Giovanelli1974}; Maltby \cite{Maltby1975}; Christopoulou~et~al. \cite{Christopoulou1999}, \cite{Christopoulou2000}, \cite{Christopoulou2001}; Bloomfield~et~al. \cite{Bloom2007}). Some of the review articles (Bogdan \cite{Bogdan2000}, Solanki \cite{Solanki2003}, Bogdan \& Judge \cite{Bogdan2006}) explain the work done on the nature of sunspot oscillations and related problems. In this context, simultaneous time-series observations in various spectral lines that sample the sunspot atmosphere at different heights using high resolution instruments such as the Solar Optical Telescope (SOT) (Tsuneta~et~al. \cite{HinodeSOT2008}) on board {\it Hinode} (Kosugi~et~al. \cite{Hinode2007}) and the Helioseismic and Magnetic Imager (Schou~et~al. \cite{HMI2012}) on board {\it Solar Dynamics Observatory} ({\it SDO}; Pesnell~et~al. \cite{SDO2011}) can be useful in studying these oscillatory processes and their contribution to the dynamics in the solar atmosphere.
The high-resolution and multi-wavelength capability of {\it Hinode} provide several important opportunities to local helioseismologists. These allow to understand and confirm many physical processes in the sub-surface layers of the Sun and also in its atmosphere (Sekii \cite{sekii2009}; Kosovichev \cite{koso2012}), some of which are as follows. Nagashima~et~al. (\cite{Nagashima2007}) studied intensity oscillations in a sunspot and showed that G~band power is suppressed in sunspot umbra, while Ca~{\sc ii}~H observations revealed high-frequency oscillations with a peak at 6~mHz. Kosovichev \& Sekii (\cite{koso2007}) studied the flare-induced high-frequency chromospheric oscillations in a sunspot. Sekii~et~al. (\cite{sekii2007}) confirmed that the supergranulation is a shallow phenomenon. Similarly, Mitra-Kravev~et~al. (\cite{Mitra2008}) examined the phase difference between oscillations of the photosphere and chromosphere. Zhao~et~al. (\cite{Zhao2010}) obtained travel-time measurements for short distances without phase-speed filtering and confirmed the sound-speed results, which were obtained using data from Michelson Doppler Imager (MDI) ( Scherrer~et~al. \cite{MDI1995}) on board the {\it Solar and Heliospheric Observatory} ({\it SoHO}; Domingo~et~al. \cite{Domingo95}) with the phase-speed filtering.
We study intensity fluctuations in different regions of an active region and the correlation between the different parameters of the photospheric magnetic fields and the intensity oscillatory power at different heights in the solar atmosphere in different frequency bands using {\it Hinode}/SOT data. Earlier investigations carried out by Mathew (\cite{Mathew2008}) using Dopplergrams from {\it SoHO}/MDI and potential field computations from {\it SoHO}/MDI line-of-sight magnetograms revealed that the umbra-penumbra boundary showed enhanced absorption of power, where the transverse potential field was strongest. Gosain~et~al. (\cite{Gosain2011}) confirmed the aforementioned result by relating the power obtained from {\it SoHO}/MDI Dopplergrams with a vector magnetogram obtained from {\it Hinode}. Using high-resolution observations of G~band and Ca~{\sc ii}~H obtained from {\it Hinode}, Nagashima~et~al. (\cite{Nagashima2007}) studied the power in spatial scales corresponding to umbral flashes and they observed a node-like structure at the center of umbra with suppressed power in Ca~{\sc ii}~H power maps at all frequencies. To that end, we employ high temporal and spatial resolution observations from {\it Hinode} to study the nature of sunspot oscillations and its relation to the photospheric magnetic field parameters.
\section{The observational data}
\label{sec:obs}
\begin{figure}
\centering
\includegraphics[width=1.0\textwidth]{ms1631fig1.eps}
\caption{Intensity filtergrams of the active region NOAA 10953 in: (a) G~band and (b) Ca II H line with a field-of-view of 112 arc-sec$^2$. The magnetic field strength map is shown in (c).}
\label{fig1}
\end{figure}
We have used a 3 hr 30 min sequence of G band and Ca II H filtergrams of the active region NOAA 10953 recorded by the Broad-band Filter Imager (BFI) of the {\it Hinode}/SOT to study the intensity oscillations in the active region. The filtergrams were acquired on 2007 May 1 during 14:31-17:57 UT and have a spatial sampling of $0\farcs11$ pixel$^{-1}$ and a cadence of one minute. The active region was located at S10W05 on the solar disk. The field-of-view (FOV) of the filtergrams is 112 arc-sec$^2$. G~band filtergrams were acquired nearly 3~s later to Ca~{\sc ii}~H filtergrams. In-addition to the broad band images, near simultaneous spectro-polarimetric observations of the active region from SOT/SP (Ichimoto~et~al. \cite{HinodeSP2008}) have been used in our analysis. SOT/SP records the four Stokes spectra of the Fe line pair at 630 nm with a spectral sampling of 21.5 m\AA~ and an exposure time of 4.8 s at each slit position. We have used level-2 maps comprising magnetic field strength, inclination, and azimuth which were obtained by inverting the observed Stokes profiles employing the MERLIN\footnote{Level-2 data of the active region was made available by the Community Spectro-polarimetric Analysis Center(CSAC) at HAO.} code. The active region was scanned in the fast mode with a step width of 0\farcs29 and a sampling of 0\farcs32 along the slit. These maps were interpolated to a spatial scale of 0\farcs32~pixel$^{-1}$ in both the directions. Consequently, the images obtained with BFI were also re-scaled to 0\farcs32~pixel$^{-1}$ resolution to match that of the SOT/SP maps.
\section{Analysis and Results}
The images were corrected for flat-fielding, dark current and bad pixels using standard Solarsoft routines. Although, the correlation tracker is employed to take care of global motion of the region of interest, we performed a rigid alignment of the sunspot as a function of time. This was done using a FFT based 2D cross-correlation routine, updating the reference frame at every 10$^{\textrm{\tiny{th}}}$ frame to account for the evolution of the sunspot. Figure~\ref{fig1} shows the snapshot of the active region in G~band and Ca~{\sc ii}~H along the map of magnetic field strength. Figure~2 shows time averaged G~band and Ca~{\sc ii}~H images deduced over the period 14:31-17:57 UT on 2007 May 1. The average images were normalized with their exposure times to obtain the counts in same scale in both the images.
A two point backward difference filter (Garc{\'{\i}}a \& {Ballot}, \cite{Garcia2PBF}) was applied to obtain the first difference of the time series and the filtered data were normalized by the mean intensity in the two running frames as shown in the equation (1). First difference enhances the oscillatory signals above the background variations and the normalization by the mean intensity causes a smooth transition between the umbra and the penumbra (Nagashima~et~al. \cite{Nagashima2007}).
\begin{eqnarray}
\hat{I_k} = 2\left(I_k-I_{k-1}\right)/\left(I_k+I_{k-1}\right)
\end{eqnarray}
Where, $\hat{I_k}$ and $I_k$ are normalized intensity and intensity of the $k^{th}$ image of the sequence, respectively.
Further, we computed the Power Spectral Density (PSD) from the mean normalized differential intensity fluctuations at each pixel and generated 3D power maps with frequency along the z-direction. PSD in each pixel is corrected for $\omega^2$ to remove the effect of the time-derivative (Nagashima~et~al. \cite{Nagashima2007}). The variation of intensity oscillatory power in different regions of the active region in G~band and Ca~{\sc ii}~H observations are studied in the following sections.
\begin{figure}[b]
\centering
\includegraphics[width=0.85\textwidth]{ms1631fig2.eps}
\caption{Average intensity maps of the active region in G~band (left) and Ca~{\sc ii}~H line (right). The contours correspond to umbra, umbra-penumbra boundary, penumbra and quiet Sun. The bright regions outside the sunspot are considered as plage while the regions that are neither bright in G band and Ca~{\sc ii}~H are assumed to be quiet Sun.}
\label{fig3}
\end{figure}
\subsection{Power distribution in different regions of the active region and quiet Sun}
\label{sec:3}
In order to investigate the power distribution in different regions of the active region, the FOV was divided into the following regions: umbra, umbra-penumbra boundary (UPB), penumbra, plage, and quiet Sun. The iso-intensity contours overlaid on the time averaged G-band and Ca~{\sc ii}~H images in Figure 2 were determined from the peaks in the intensity distribution. The contours enclosing the umbra, UPB, and penumbra were obtained from the time averaged G-band image while the same for the plage region was determined from the time averaged Ca image; the bright regions outside the sunspot are considered as plage while the regions that are neither bright in G band and Ca~{\sc ii}~H are assumed to be quiet Sun. The fraction of pixels corresponding to umbra, UPB, and penumbra are 0.022, 0.030, 0.22, respectively. Plage region and quiet region contain 0.626 and 0.091 fraction of pixels, respectively. The average power in each of these regions as a function of frequency is shown in Figures~\ref{layerwise}~and~\ref{Rwise}. Figure~\ref{layerwise} allows us to compare the power in the different regions described above, separately in G~band and Ca~{\sc ii}~H. On the other hand, Figure~\ref{Rwise} shows a comparison between the powers in G~band and Ca~{\sc ii}~H in each of the regions.
\begin{figure}[t]
\centering
\includegraphics[width=0.425\textwidth]{ms1631fig3a.eps}
\includegraphics[width=0.425\textwidth]{ms1631fig3b.eps}
\caption{Power distribution of G~band intensity (left) and Ca~{\sc ii}~H intensity (right) in different regions of the FOV. Power in umbra (solid line with diamond symbols), umbra-penumbra boundary (dotted line), penumbra (dashed line), plage (dot-dashed) and quiet Sun (solid line) are shown here.}
\label{layerwise}
\end{figure}
\begin{figure}[b]
\centering
\includegraphics[width=0.275\textwidth]{ms1631fig4a.eps}
\hspace{1mm}
\includegraphics[width=0.275\textwidth]{ms1631fig4b.eps}
\hspace{0.5mm}
\includegraphics[width=0.275\textwidth]{ms1631fig4c.eps}
\includegraphics[width=0.275\textwidth]{ms1631fig4d.eps}
\hspace{0.5mm}
\includegraphics[width=0.275\textwidth]{ms1631fig4e.eps}
\hspace{1mm}
\includegraphics[width=0.275\textwidth]{ms1631fig4f.eps}
\caption{Power distribution of G~band intensity (solid line) and Ca~{\sc ii}~H intensity (dotted line) in the regions: umbra (top-left), umbra-penumbra boundary (top-middle), penumbra (top-right), plage (bottom-left), quiet Sun (bottom-middle) and sunspot (bottom-right).}
\label{Rwise}
\end{figure}
It is observed that intensity oscillatory power in G~band decreases from the quiet Sun to umbra (i.e., in the order quiet Sun, plage, penumbra, UPB, and umbra) at all frequencies. On the other hand, power in Ca~{\sc ii}~H shows such a trend from quiet Sun to penumbra only, with the exception of slight enhancement of power in 3-4 mHz frequency range (c.f., Figure ~\ref{layerwise}). It is also observed that overall power is lower in G~band as compared to Ca~{\sc ii}~H with a cross-over seen at 3~mHz which shifts to 0~mHz as we move from the quiet Sun to the umbra of the sunspot (c.f. Figure~\ref{Rwise}). Thus, Ca~{\sc ii}~H oscillations are seen to be richer in high-frequency power in the magnetized environment, which is in good agreement with the results of Nagashima~et~al. (\cite{Nagashima2007}). We also observe the presence of 5-minute oscillations in quiet Sun, plage and penumbral regions.
\subsection{Intensity power maps as function of magnetic field strength and frequency}
In order to examine the relation between oscillatory power in G~band and Ca~{\sc ii}~H with respect to the photospheric magnetic field strength ($B$) and frequency, we constructed power maps as function of magnetic field strength and frequency as indicated in Figure~\ref{magpow}. These maps were derived by averaging the power in all pixels having similar magnetic field strengths over a 30~G interval at each frequency ranging from 0.1 to 8.3~mHz. We observe the following from these power maps.
\begin{enumerate}
\item In general, Ca~{\sc ii}~H shows larger power in comparison to that of G~band for any magnetic field strength and frequencies above 1~mHz.
\item Power in G~band decreases with increase in magnetic field strength at all frequencies.
\item Ca~{\sc ii}~H shows larger power in stronger magnetic fields ($|B|$~\textgreater 2200~G) at all frequencies and less power in the intermediate field regime (100~G~\textless~$|B|$~\textless~2000 G) at frequencies above 5~mHz. Weaker fields ($|B|$~\textless~100~G) show larger power in the frequency range 0-6~mHz in comparison to the same in the intermediate field strength regime.
\item Power map of Ca~{\sc ii}~H is more structured than that of G~band. It shows the signature of 5-minute oscillations for the magnetic field strengths between 1200-2000 G. These fields are mostly located in the penumbral region where the inclination is in the range of 90-130 degrees. In the higher magnetic field regions (mostly, umbra) it also shows oscillatory behavior.
\end{enumerate}
\begin{figure}[h]
\centering
\includegraphics[width=0.40\textwidth]{ms1631fig5a.eps}
\hspace{5mm}
\includegraphics[width=0.40\textwidth]{ms1631fig5b.eps}
\caption{Power maps from the intensity variations in the active region NOAA 10953 in G~band (left) and Ca~{\sc ii}~H (right) as function of magnetic field strength and frequency. The maps are shown in logarithmic scale as indicated by the color bar.}
\label{magpow}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.575\textwidth]{ms1631fig6a.eps}
\includegraphics[width=0.375\textwidth]{ms1631fig6b.eps}
\caption{Left: Enclosed curves overlaid on the intensity filtergram of the active region to study the radial and azimuthal variation of the magnetic field parameters and oscillations in various regions of the sunspot. Right: The number of pixels enclosed by each annular-like region are shown in the plot. The regions are numbered from center of umbra to sunspot boundary.}
\label{contou}
\end{figure}
\subsection{Relation between magnetic field parameters and oscillatory power in the sunspot}
\label{sec3}
The radial variation of the observed magnetic field parameters is studied using the enclosed curves shown in Figure~\ref{contou}. The umbra-penumbra boundary and penumbra-quiet Sun boundary obtained from the time-averaged G band image (c.f., Fig. 2), were used to construct 19 equidistant annular-like regions in the azimuthal direction which yielded the curves shown in Figure~\ref{contou}. The curves inside the umbra were hand drawn, which are separated by $\approx$1$^{\prime\prime}$. The other curves between umbra-penumbra boundary and penumbra are spaced apart by 0\farcs6. The regions 1-3 and 8-20 correspond to umbra and penumbra, respectively, while the regions 4-7 represent the umbra-penumbra boundary. The number of pixels enclosed by each region is illustrated in Figure~\ref{contou}.
\begin{figure}[t]
\centering
\includegraphics[width=0.30\textwidth]{ms1631fig7a.eps}
\hspace{1mm}
\includegraphics[width=0.30\textwidth]{ms1631fig7b.eps}
\hspace{1mm}
\includegraphics[width=0.30\textwidth]{ms1631fig7c.eps}
\caption{Radial variation of power in G~band and Ca {\sc ii}~H are shown in left and right panels, respectively. Corresponding variation of magnetic field parameters ($B$, $\gamma$, $B_l$ and $B_t$) are shown in the middle panels. The power is averaged over frequency regimes: 0-8~mHz, 0-2~mHz, 2-5~mHz and 5-8~mHz. The vertical bars show $\pm1\sigma$ errors in the estimated values. The regions 1-3, 4-7 and 8-20 correspond to umbra, umbra-penumbra boundary and penumbra, respectively. The regions 21-23 correspond to sunspot-quiet Sun boundary.}
\label{SPvar}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.425\textwidth]{ms1631fig8a.eps}
\includegraphics[width=0.425\textwidth]{ms1631fig8b.eps}
\includegraphics[width=0.425\textwidth]{ms1631fig8c.eps}
\includegraphics[width=0.425\textwidth]{ms1631fig8d.eps}
\caption{Plots show correlation between the photospheric magnetic field parameters and intensity oscillatory power of umbra (top-panels) and penumbra (bottom panels) in G~band (left panels) and Ca~{\sc ii}~H (right panels) in different frequency bands. Correlation between $B$ vs. power (solid line), $B_t$ vs. power (dashed line), and $B_l$ vs. power (dotted) are shown here.}
\label{Corr}
\end{figure}
The azimuthally averaged radial profiles of various magnetic field parameters are shown in Figure~\ref{SPvar}. It is observed that magnetic field strength ($B$), inclination ($\gamma$), and line-of-sight ($B_l$) components of magnetic field show smooth variation with the radial distance. Transverse component of magnetic field ($B_t$) increases from center of the umbra to the inner-mid penumbra and, thereafter, this trend reverses. The vertical bars correspond to $\pm1\sigma$ errors in the estimated values. In general, the power in umbra-penumbra boundary (regions 4-7) and sunspot-quiet Sun boundary (regions 21-23) shows relatively larger spread in G~band. Whereas in Ca~{\sc ii}~H, the power in umbra and umbra-penumbra boundary shows larger spread.
The radial variation of power in Ca~{\sc ii}~H is dominant in the umbra in comparison to the other regions of the sunspot in the frequency regimes 2-5~mHz and 5-8~mHz. On the other hand, in 0-2~mHz band umbra-penumbra boundary shows enhancement of power both in Ca~{\sc ii}~H and G~band. Umbra-penumbra boundary in Ca~{\sc ii}~H shows reduction of power in 2-5~mHz and 5-8~mHz bands, while it shows enhancement of power in G~band for the above frequency ranges. This reduction/enhancement of power at umbra-penumbra boundary occurs around the peak value of the transverse magnetic field. The inclination of magnetic field at this location is about $130^\circ$. On the other hand, power estimated from the photospheric Dopplergrams obtained by MDI/{\it SoHO} showed enhanced {\it p-}mode absorption near umbra-penumbra boundary, where the inclination angle is nearly $135^\circ$ (Mathew~\cite{Mathew2008}; Gosain~et~al.~\cite{Gosain2011}).
In order to understand the relation between magnetic field parameters and intensity oscillatory power of the sunspot, we estimated the correlation-coefficients between radial variation of power and the magnetic field parameters averaged over each region. Here, it is observed that none of the magnetic field parameters ($B$, $B_l$, and $B_t$) show any correlation with the power in the sunspot. A Similar analysis by Gosain~et~al.~(\cite{Gosain2011}) did not reveal any strong association of the magnetic field parameters with Doppler power in a sunspot. However, when we derive the correlation-coefficients separately for the umbra and penumbra, we find distinct relations between the magnetic field parameters and the PSD which are shown in Figure~\ref{Corr}. When we consider sunspot as a whole, the distinct behaviors in the umbra and penumbra gets mixed-up and thus there is no correlation seen in our analysis.
Intensity oscillatory power in Ca~{\sc ii}~H penumbra shows anti-correlation with $B$, $|B_l|$ and $B_t$ at all the frequencies except in the frequency range of 3.0-3.5~mHz, i.e. in the 5-minute regime. G~band intensity power in penumbra also shows anti-correlation with magnetic field parameters. Similarly, the umbral power in Ca~{\sc ii}~H shows correlation with $B$, and $|B_l|$ above 1.5~mHz and opposite relation with $B_t$. Whereas, G~band umbral power shows completely opposite trend; i.e. it does show anti-correlation with $B$ and $|B_l|$, while correlation with $B_t$ at all the frequencies.
\section{Summary and Discussion}
We have analyzed high-resolution G~band and Ca~{\sc ii}~H line filtergrams of the active region NOAA 10953 obtained by {\it Hinode}/SOT along with near simultaneous spectro-polarimetric observations of this active region from {\it Hinode} SOT/SP to study the relation between various magnetic parameters and the intensity oscillatory power in the photosphere and chromosphere. Our chief findings are as follows:
\begin{enumerate}
\item Photospheric power maps derived from G~band time series reveal more power in the quiet Sun (weaker fields) in comparison to the sunspot. This is in agreement with the previous reports that stronger magnetic fields absorb more power in the 5-minute band (Braun~et~al. \cite{Braun1992}; Hindman \& Brown \cite{Hindman1998}; Kumar~et~al. \cite{Kumar2000}; Venkatakrishnan et~al. \cite[and references therein]{PVK2002}). We, however, do not observe a reversal of this trend at higher frequencies in magnetic concentrations as shown by Venkatakrishnan et~al. (\cite{PVK2002}) in oscillatory power derived from photospheric Dopplergrams. Our analysis showing the absence of photospheric power at high frequencies in strong magnetic field regions is consistent with the results of Jain \& Haber (\cite{Jain2002}) who reported that only power spectra derived from Dopplergrams exhibit the above trend and not from intensity filtergrams. We emphasize here that the well known 5-minute oscillations are not dominantly seen in G~band power maps, whereas the 5-minute oscillations have been distinctly observed by Jain \& Haber (\cite{Jain2002}) in the continuum intensity power of the data obtained from {\it SoHO}/MDI.
To confirm this, we have analyzed Dopplergrams and continuum intensity images obtained by the Helioseismic and Magnetic Imager (HMI) instrument onboard {\it Solar Dynamics Observatory} ({\it SDO}) on 2013 April 11. We estimated the PSD of the quiet Sun near the disk center for both the data sets. The Dopplergrams exhibit dominant power in 5-minute regime and similar nature of power is also observed in the continuum intensity images. However, the power in intensity is weaker in comparison to the Doppler power in the 5-minute regime. These results obtained with {\it SDO}/HMI is in agreement with the work done by Jain \& Haber (\cite{Jain2002}) using the data obtained from {\it SoHO}/MDI. We conjecture that the reason behind the observed lower power of 5-minute oscillations in G~band data relative to that in continuum intensity power from {\it SoHO}/MDI and {\it SDO}/HMI could be due to difference in their formation heights in the solar atmosphere.
\item It is well known that in the chromosphere, the umbra shows an enhancement in power at frequencies above 5~mHz (Bhatnagar \& Tanaka \cite{Bhatnagar1972}; Braun~et~al. \cite{Braun1992}; Brown~et~al. \cite{Brown1992}; Kentischer \& Mattig \cite{Kenti1995}; Lites \cite{Lites1986}). We also observe such a behavior in our analysis of Ca~{\sc ii}~H power (c.f., Figure~3 and 5). The umbral 3-minute chromospheric oscillations are suggested to emanate directly from the photosphere through linear wave propagation (Centeno et~al. \cite{Centeno2006}). Spectropolarimetric investigations by Centeno et~al. (\cite{Centeno2006}) using simultaneous observations taken in photosphere and chromosphere have provided observational evidence for the upward propagation of slow magneto-acoustic waves from the photosphere to the chromosphere inside the umbra of a sunspot. The phase spectra derived by Centeno et~al. (\cite{Centeno2006}) yield a value of the atmospheric cut-off frequency around 4~mHz and shows evidence for the upward propagation of higher frequency oscillations. Similarly, the presence of 5-minute oscillations in the chromospheric quiet Sun, and penumbral regions is attributed to the inclined magnetic field lines along which the photospheric 5-minute oscillations propagate to the higher atmosphere (Jefferies~et~al.~\cite{Jeff2006}; McIntosh \& Jefferies~\cite{McI2006}).
\item We have used near simultaneous {\it Hinode} SOT/SP observations of an active region, close to disk-center, to study the relationship between the magnetic-field parameters and the intensity oscillatory power in G~band and Ca~{\sc ii}~H. The azimuthally averaged radial profiles of field strength and inclination show a smooth decrease from the centre of the umbra to the periphery of the sunspot (c.f., Figure~\ref{SPvar}). However, the power does not exhibit a similar behavior. Umbra in Ca~{\sc ii}~H and G~band shows reduction of power in 0-2~mHz band. While this behavior is illustrated at all other frequencies in G~band, the same shows enhancement in 5-8~mHz band in Ca~{\sc ii}~H. The umbra-penumbra boundary shows enhancement of G~band power at all frequencies, where the transverse magnetic field is the highest.
\item The correlation analysis to the umbra and penumbra, separately, shows correlation between power and magnetic field parameters. In order to examine the influence of magnetic field on oscillatory power, we determined the correlation as a function of frequency bins. We observe that the photospheric magnetic parameters (except the transverse component) are correlated with umbral power in Ca~{\sc ii}~H and anti-correlated with umbral power in G~band at all frequencies, while the transverse magnetic field exhibits the opposite result. However, the analysis that includes only the penumbra shows that G~band power in the penumbra is anti-correlated with magnetic field parameters at all frequencies. Ca~{\sc ii}~H power in penumbra shows strong correlation with photospheric magnetic field in the frequency band 3.0-3.5~mHz, while at every other frequencies they are anti-correlated.
The observed correlation between chromospheric penumbral power and photospheric magnetic fields in the 5-minute band could be the result of the magnetic inclination becoming large enough to allow photospheric 5-minute power to tunnel through higher acoustic cut-off frequency as demonstrated by Bloomfield et~al. (\cite{Bloom2007}) using simultaneous spectro-polarimetric observations of photosphere and chromosphere. As the umbral 3-minute chromospheric oscillations are inferred as field-aligned propagating slow magneto-acoustic waves (Centeno et~al. \cite{Centeno2006}), the observed correlation between photospheric magnetic fields and the chromospheric umbral power could also be due to the transportation of the photospheric power to the chromosphere through the magnetic field lines.
\end{enumerate}
\begin{acknowledgements}
{\it Hinode} is a Japanese mission developed and launched by ISAS/JAXA, collaborating with NAOJ as a domestic partner, NASA and STFC (UK) as international partners. Scientific operation of the {\it Hinode} mission is conducted by the {\it Hinode} science team organized at ISAS/JAXA. This team mainly consists of scientists from institutes in the partner countries. Support for the post-launch operation is provided by JAXA and NAOJ (Japan), STFC (U.K.), NASA (U.S.A.), ESA, and NSC (Norway). We also thank Community Spectro-polarimetric Analysis Center team at High-Altitude Observatory for providing the vector magnetogram (level2) data. Thanks to B.~Ravindra for his valuable suggestions related to this work. Rohan E. Louis is grateful for the financial assistance from the German Science Foundation (DFG) under grant DE 787/3-1. We thank the anonymous referee for constructive comments and suggestion that improved the presentation of the manuscript.
\end{acknowledgements}
|
\section{Introduction}
\label{sec:intro}
\begin{figure*}
\begin{center}
\begin{subfigure}[b]{0.4\textwidth}
\includegraphics[trim=2.5cm 2.5cm 2.5cm 2.5cm, clip=true, width=\textwidth]{system1_2.pdf}
\caption*{Observers (front) view}
\label{fig:system1}
\end{subfigure} \quad
\begin{subfigure}[b]{0.4\textwidth}
\includegraphics[trim=2.5cm 2.5cm 2.5cm 2.5cm, clip=true, width=\textwidth]{system2_2.pdf}
\caption*{Side view}
\label{fig:system2}
\end{subfigure}
\end{center}
\hspace{1.5cm}
\begin{subfigure}[b]{0.4\textwidth}
\includegraphics[trim=3.9cm 1.5cm 3.8cm 3.5cm, clip=true, width=\textwidth]{system3-eps-converted-to.pdf}
\caption*{Top view}
\label{fig:system3}
\end{subfigure}
\caption{\footnotesize Configuration of the HAT-P-7 system to scale using values for HAT-P-7b (full black circle) from \citet[][]{2013ApJ...774L..19V}. Top Left: the observers view of the system from Earth, with the angular momentum vectors of the planetary orbit, $\bar{n}_p$ (normal to the orbital plane), and stellar spin, $\bar{n}_{\star}$, given as red arrows. The projected angle, $\lambda$, is indicated by the shaded magenta region and is found as the angle between $\bar{n}_p$ and $\bar{n}_{\star}$ when these are projected onto the plane of the sky (dashed lines). This angle is obtained from RM measurements and in this panel we have used $\lambda=155\pm 37^{\circ}$ following \citet{2012ApJ...757...18A} (the uncertainty on $\lambda$ is not included in the figure). The stellar inclination, $i_{\star}$, which is the parameter measured from asteroseismology, is set to $15^{\circ}$, and is given by the direct angle between the line of sight (midpoint of star) and $\bar{n}_{\star}$. The inclination of the planetary orbit, $i_p$, is set to $83^{\circ}$ following \citet[][]{2013ApJ...774L..19V}. The true angle, $\psi$, is the direct angle between $\bar{n}_p$ and $\bar{n}_{\star}$. Top Right: side view of the system, with the observers view-point from the right, indicated by the ``line of sight''. To properly show $i_{\star}$ (shaded green) and $i_p$ (shaded red), and not their projected values, we have set $\lambda=180^{\circ}$ such that both $\bar{n}_p$ and $\bar{n}_{\star}$ lie in the same plane as the line of sight. For illustrative purposes we have in addition decreased $i_p$ to $63^{\circ}$. When adopting this configuration $\psi$ (shaded blue) is given by the sum of $i_{\star}$ and $i_p$.
Bottom Left: Top view of the system, with the observers view-point from the bottom, indicated by the ``line of sight''. }
\label{fig:system}
\end{figure*}
Asteroseismology can provide detailed information about stellar parameters such as mass, radius, and age \citep[see][and references therein]{2013ARA&A..51..353C}. Furthermore, an estimate for the stellar inclination can be obtained for solar-like oscillators \citep[][]{2003ApJ...589.1009G}, which in turn is needed in asserting the obliquity of planet hosting systems.
The obliquity of planetary systems, $\psi$, the angle between the stellar spin-axis and the angular momentum vector of the planetary orbit, is an important parameter for a better understanding of how these systems form and evolve \citep[see, e.\,g.,][]{2008ApJ...678..498N,2010ApJ...718L.145W,2010A&A...524A..25T,2011ApJ...729..138M,2012ApJ...758L...6R}.
The obliquity is especially interesting for systems that appear to have retrograde orbits from measurements of the sky-projected obliquity, $\lambda$. The reason is that, while the orbit might indeed be retrograde, it is not known whether it is closer to polar than equatorial. This distinction makes a great difference for theories dealing with planetary system formation and evolution as they must be able to account for such a configuration \citep[see, e.\,g.,][]{2008ApJ...686..580C,2008ApJ...686..621F,2010ApJ...725.1995M,2011MNRAS.412.2790L,2012ApJ...757...18A}.
Stellar obliquities are difficult to measure as stars are unresolved by modern telescopes. Therefore no spacial information can be obtained\footnote{With the exception of stars with spacial interferometric constraints.}. However during planetary transits parts of the stellar surface are covered breaking the degeneracy.
Such information on the obliquity can, e.\,g., be obtained from studies of the anomalous effect in the radial velocity (RV) curve known as the \emph{Rossiter-McLaughlin} (RM) effect \citep[][see \citet{2009ApJ...696.1230F} for an overview]{1924ApJ....60...15R, 1924ApJ....60...22M}.
Unfortunately, only $\lambda$ can be obtained from RM measurements (see \fref{fig:system}).
Other means of obtaining $\lambda$ are, e.\,g., from spot-crossing anomalies observed during planetary transits \citep[see, e.\,g.,][]{2011ApJ...733..127S,2013ApJ...775...54S,2011ApJS..197...14D}, Doppler tomography \citep[see, e.\,g.,][]{2012A&A...543L...5G}, or from the effects of gravity darkening \citep[see, e.\,g.,][]{2011ApJS..197...10B,2014ApJ...786..131A}.
The true obliquity can only be unequivocally determined if the stellar angle of inclination, $i_{\star}$\footnote{We define $i_{\star}$ as the angle between the stellar spin axis and the observers line of sight, thus going from $i_{\star}=0^{\circ}$ for a pole-on view to $i_{\star}=90^{\circ}$ for an equator-on view.}, can be measured, and combined with $\lambda$ and the inclination of the planetary orbital plane, $i_p$.
The orbital inclination can be estimated with relative ease from analysis of the photometric light curve if the planet happens to transit its host star.
A measure of the stellar inclination angle can be obtained from combining $v\sin i_{\star}$ from spectroscopy with the stellar rotation period from modulations of the light curve from stellar spots \citep[see][for recent uses of this method for planetary systems]{2012ApJ...756...66H,2014ApJ...783....9H}. These estimates can be quite uncertain due to the difficulty of calibrating the spectroscopic $v\sin i_{\star}$, disentangling the rotational signal from other broadening effects, and the need for an estimate of the stellar radius $R_{\star}$.
A more direct method for obtaining $i_\star$ is that of asteroseismology, where the stellar inclination can be estimated by analysing solar-like acoustic ($p$-mode) oscillations \citep[][]{2003ApJ...589.1009G}.
Another great advantage of using asteroseismology is that a detailed stellar model can be obtained with well determined parameters that are needed in simulations of planetary systems dynamics.
The high photometric quality that enabled the \emph{Kepler}{} mission \citep[see][]{2010Sci...327..977B, 2010ApJ...713L..79K} to detect the transits of extrasolar planets, and thus allows for a determination of $i_p$, also make the data ideal for asteroseismic analysis \citep{2010PASP..122..131G}.
Analysis of the obliquity using the asteroseismic method have to date only been performed in the systems Kepler-50 and 65 by \citet[][]{2013ApJ...766..101C}, Kepler-56 by \citet[][]{2013Sci...342..331H}, Kepler-410 by \citet[][]{2014ApJ...782...14V}, and 16 Cygni by Davies et al. (submitted). However, for these systems the obliquity could only be assessed in a statistical sense, since the projected angle, $\lambda$, was unavailable.
Our aim in this paper is to use asteroseismology to determine precise stellar parameters of HAT-P-7 and to determine the obliquity of the HAT-P-7\footnote{We would like to emphasize the efforts made by O. Benomar and his collaborators for their work on HAT-P-7. This system turned out to be studied simultaneously and independently by our respective teams.} system with a F6V \citep[][]{2013MNRAS.433.2097F} type star and a close-in ${\sim}1.78\, M_{\rm J}$ planet (HAT-P-7b) in a ${\sim}2.2$ day orbit \citep{2008ApJ...680.1450P}.
However, from the very onset it is clear that this is a challenging task as HAT-P-7 is a late-F-type star. This spectral type is notorious for having short lifetimes of the $p$-mode oscillations and consequently very broad (in frequency) oscillation modes, which highly obscures the potentially small effects imposed by rotation.
A fortuitous feature of the system is that not only $i_{\star}$ and $i_p$ can be estimated from \emph{Kepler}{} data, but the RM effect has been studied independently by \citet{2009ApJ...703L..99W}, \citet{2009PASJ...61L..35N}, and \citet[][]{2012ApJ...757...18A} using HIRES and/or HDS data\footnote{HIRES@Keck-I: High Resolution Spectrograph \citep[][]{1994SPIE.2198..362V};\\ HDS@Subaru: High Dispersion Spectrograph \citep[][]{2002PASJ...54..855N}.}. Values for $\lambda$ and $v\sin i_{\star}$ from these studies are given in Table~\ref{tab:orbit}. Despite disagreement on the actual values, all three works agree that the system is misaligned, that the planetary orbit might be retrograde, and that the very low value measured for $v\sin i_{\star}$ suggests a low $i_{\star}$, which for a transiting planet implies a near-polar orbit of the planet
An interesting aspect of the system with regard to the obliquity is that a third body (a M5.5V dwarf known as HAT-P-7B) is found to be associated with the system \citep[][]{2010PASJ...62..779N,2012PASJ...64L...7N,2013MNRAS.428..182B,2013MNRAS.433.2097F}, and that a fourth associated body, likely more massive than Jupiter, is speculated based on an unexplained RV excess \citep{2009ApJ...703L..99W,2012PASJ...64L...7N}.
The system is thus a prime candidate for obliquity studies and theories dealing with planetary system formation and evolution.
The paper is structured as follows: In \sref{sec:data} we describe the data used in our analysis. Section~\ref{sec:inc} deals with the model we use for obtaining the stellar inclination angle. In \sref{sec:res} we present our analysis and results, including stellar modelling using two different codes (\sref{sec:modelling}), and the results obtained for the stellar inclination and rotation (\sref{sec:splitinc}).
Our results for the obliquity is the topic of \sref{sec:obliquity}. In \sref{sec:gyro} we compare our result on the stellar rotation rate to gyrochronology, and \sref{sec:activity} is concerned with possible activity signatures. Finally we discuss our results in \sref{sec:dis} and make some concluding remarks in \sref{sec:con}.
\begin{figure*}
\centering
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{power_spectrum3-eps-converted-to.pdf}
\label{fig:Power_spectrum}
\end{subfigure} \quad
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{Bgfit3-eps-converted-to.pdf}
\label{fig:Background_fit}
\end{subfigure}
\caption{\footnotesize Left: power spectrum of \hat{} (black) with the fitted model over-plotted (red; cf. \eqref{eq:limitspec}). Right: power spectrum of \hat{}\ (black) over-plotted with the optimum fit to the background (red; cf. \eqref{eq:bg}). The light-grey part up to $\rm 100\, \mu Hz$ was not included in the fit. The fit includes, besides the Gaussian envelope from $p$-modes centred around $\nu_{\rm max}\approx1115\,\rm \mu Hz$, a granulation component (\textcolor[rgb]{0, 0.5, 0}{\textbf{P}$\rm _G$}; green) and a white/shot noise (\textcolor{black}{\textbf{P$\rm _S$}}; black) level. The dashed red line shows the background fit without the Gaussian envelope.}
\label{fig:Powerspecs_full}
\end{figure*}
\section{Data}
\label{sec:data}
We extracted short cadence (SC; $\Delta t = 58.8\,\rm s$) \emph{simple aperture photometry} (SAP) data from \emph{target pixel files} (TPFs) using the procedure of Steven Bloemen (private communication), which starts from the original \emph{Kepler}{} mask and adds or removes pixels to the aperture based on the amount of signal in each pixel. The outcome of this is a new mask that most often is slightly bigger than the original one.
The data span quarters from Q$0$ to Q$17$ (${\sim}1470$ days), with a duty cycle of ${\sim}90.4\%$. This constitutes the full amount of data available from \emph{Kepler}{} in its normal mode of operation.
Data were downloaded from \emph{The Mikulski Archive for Space Telescopes} (MAST) and corrected using the procedure described in Handberg \& Lund (submitted). Briefly, two median-filtered versions of the time series are computed with different filter windows. A weighted combination of the two are then used together with a filtered version of the planetary phase-curve to correct the time series for both instrumental features and the planetary signal.
For the asteroseismic analysis we used filter windows of $1$ ($\tau_{\rm long}$) and $0.028$ ($\tau_{\rm short}$) days, while windows of $15$ ($\tau_{\rm long}$) and $5$ ($\tau_{\rm short}$) days were adopted in preparing the time series used in \sref{sec:peak} to search for a low-frequency imprint of rotational modulation. We refer the reader to Handberg \& Lund (submitted) for further details on the filter and for a view of the corrected time series for HAT-P-7.
For the transit parameters needed in both the correction of the time series (orbital period) and for the estimation of the obliquity (inclination of orbital plane) we used the results of \citet[][]{2013ApJ...774L..19V}.
The power spectrum was calculated using a weighted sine-wave fitting method \citep[see, e.\,g.,][]{1992PhDT.......208K,1995A&A...301..123F}, normalized according to the amplitude-scaled version of Parseval's theorem \citep[see][]{1992PASP..104..413K}, in which a sine wave of peak amplitude, A, will have a corresponding peak in the power spectrum of $\rm A^2$.
\section{Fitting the power spectrum}
\label{sec:inc}
The harmonic eigenmodes of acoustic solar-like oscillations are characterised by their degree, $l$, which gives the number of nodal lines on the stellar surface, and their radial order, $n$, giving the number of radial nodes. In addition, an eigenmode is characterised by its azimuthal order, $m$, of which there are $2l +1$. Only in the case of broken spherical symmetry, e.\,g.\ by rotation, will the degeneracy between different $m$-values be lifted. The removal of this degeneracy makes it possible to measure the stellar rotation rate and inclination angle.
\subsection{Modelling the power spectrum}
\label{sec:psmodel}
We model the power spectral density of the oscillations with a series of standard Lorentzian functions. The limit spectrum (noise free) to be fit to the power spectrum can be expressed as follows:
\begin{equation}
\mathcal{P}(\nu_j ; \boldsymbol \Theta) = \sum_{n=n_{a}}^{n_{b}}\sum_{l=0}^{2}\sum_{m=-l}^{l} \frac{\mathcal{E}_{l m}(i_{\star})S_{nl}}{1+\frac{4}{\Gamma_{nl}^2}(\nu_j-\nu_{nl m})^2 } + B(\nu_j)\, .
\label{eq:limitspec}
\end{equation}
Here $n_a$ and $n_b$ represent the first and last mode orders included from the power spectrum, respectively, while $\nu_{nl m}$ is the mode frequency including the effect of rotation, $B(\nu_j)$ describes the contribution from the stellar noise background at frequency $\nu_j$, $S_{nl}$ is the overall height of the multiplet, \ie, the maximum power spectral density, and $\Gamma_{nl}$ is the mode linewidth. A geometrical modulation of the relative visibility between components of a split multiplet is given by $\mathcal{E}_{l m}(i_{\star})$. The fitted parameters are denoted by $\boldsymbol \Theta$. This fit can be seen in the left panel of \fref{fig:Powerspecs_full}.
The background signal $B(\nu)$ is described by a series of power laws \citep{1985ESASP.235..199H}, each of which relate to a specific physical phenomenon. The power laws included in this work describe the signals from granulation and faculae. The specific functional from for the background signal is that suggested by \citet{KarPhD}:
\begin{align}
B(\nu) = \sum_{i=1}^2{\frac{4\sigma_i^2 \tau_i}{1 + (2\pi \nu \tau_i)^2 + (2\pi \nu \tau_i)^4}} + B_0 \, .
\label{eq:bg}
\end{align}
In this equation, $\sigma_i$ and $\tau_i$ gives, respectively, the flux rms variation in time and the characteristic time scales of the different phenomena. The constant $B_0$ is a measure of the photon shot-noise. The fit to the background can be seen in the right panel of \fref{fig:Powerspecs_full}.
\subsection{Optimisation procedure}
\label{sec:Optimisation_procedure}
The fit of \eqref{eq:limitspec} to the power spectrum is optimised in a Bayesian manner using the Markov Chain Monte Carlo (MCMC) sampler \texttt{emcee} (\citet[][]{2013PASP..125..306F}.
Frequencies for all modes are free parameters in the fit.
Heights and widths are only free parameters for radial ($l=0$) modes. For a non-radial mode the width is found from a linear interpolation between the two nearest radial modes. The same goes for the heights where the linearly interpolated value for the nearest radial modes is scaled via the visibility parameter (assumed constant as a function of frequency). In this work we keep the relative visibilities as free parameters.
With this set-up we have the following set of free parameters in the fitting: $\boldsymbol\Theta~=~\{\nu_{nl},\, i_{\star},\, \nu_s,\, S_{n,0},\, \tilde{V}_{l=1}^2,\, \tilde{V}_{l=2}^2,\, \Gamma_{n,0}\}$.
We refer the reader to Appendix~\ref{app:modelling_power} for details on the fitted model and the adopted optimisation.
\section{Analysis and Results}
\label{sec:res}
\subsection{Overall results from peak-bagging}
\label{sec:res_overall}
For estimating mode frequencies for the modelling we fit \eqref{eq:limitspec} to the frequency range $\rm 600 - 1650\, \mu Hz$, which is the range where we could visually identify modes.
All estimated frequencies are given in Table~\ref{tab:mcmc}. Here we report the median of the marginalized posterior (kernel) probability distributions (PPDs) for the respective modes, while the uncertainties were obtained from the $68\%$ \emph{highest probability density} (HPD) credible region. For the frequency uncertainties used in the stellar modelling (see \sref{sec:modelling}) we adopted the mean value of the (potentially asymmetric) uncertainties from the HPD credible region.
For results on mode linewidths and visibilities we refer the reader to Appendix~\ref{app:lwvis}.
For the estimation of the inclination and splitting parameters we fit \eqref{eq:limitspec} to a smaller range in the power spectrum including only the frequencies in the range $\rm 780 - 1400\, \mu Hz$. The selection of this interval was based on the estimates for the mode linewidths from the large fit and to get modes of relatively high signal-to-noise ratio (SNR). For the splitting we used a flat prior from $\rm -8\, \text{to}\,8\, \mu Hz$, and then used the absolute value of the splitting in \eqref{eq:limitspec}. For the inclination we used a flat prior from $-90\, \text{to}\,180^{\circ}$, and then folded values onto the interval from $0\, \text{to}\,90^{\circ}$. The symmetry of the priors was chosen to avoid potential boundary effects in the MCMC sampling.
We note, that a component of \eqref{eq:limitspec} that can cause problems in the fitting is that of the noise-background. Initially we did not fix the background in the fit of \eqref{eq:limitspec}, but rather set Gaussian priors on the background parameters from the posteriors of a background-only fit. However, given that the background is very poorly constrained in the relatively small part of the power spectrum occupied by the oscillation modes, we found that not even the Gaussian priors could constrain the background, and especially the granulation time scale, $\tau_g$, wandered to lower values. These lower values for $\tau_g$ were found to correlate with a large range of high values for the splitting at a particular value for the inclination. Due to this apparent degeneracy we chose to fix the background in the fits.
\subsection{Modelling of \hat{}}
\label{sec:modelling}
\begin{table*}
\begin{center}
\begin{threeparttable}
\caption{Selected properties of our best-fit models compared to the first asteroseismic modelling by \citet{2010ApJ...713L.164C} (CD10). The GARSTEC model (\sref{sec:GARSTEC}) constitutes our preferred values.}
\begin{tabular}{lllll}
\toprule \\ [-0.3cm]
& & \multicolumn{2}{c}{This Work} \\
\cmidrule(r){3-4}
Parameter & & GARSTEC (preferred) &MESA & CD10\tnote{a} \\[0.1cm]
\midrule
$M_\star\,\rm [M_\sun]$ & & $1.51^{+0.04}_{-0.05}$ &$1.63 \pm 0.09$ & $1.52$ \\ [0.15cm]
$R_\star\,\rm [R_\sun]$ & & $2.00^{+0.01}_{-0.02}$ &$2.04 \pm 0.04$ & $1.992$ \\ [0.15cm]
$L_\star\,\rm [L_\sun]$ & & $5.91^{+0.31}_{-0.33}$ &$6.2 \pm 0.5$ & $5.81$ \\ [0.15cm]
Age [Gyr] & & $2.07^{+0.28}_{-0.23}$ &$1.9 \pm 0.4$ & $1.875$ \\ [0.15cm]
$\log g\,\rm [cm\, s^{-2}]$ & & $4.01^{+0.01}_{-0.01}$ &$4.03 \pm 0.01$ & $4.021$ \\ [0.15cm]
$T_\mathrm{eff} \,\rm [K]$ & & $6366^{+78}_{-80}$ &$6374 \pm 130$ & $6355$ \\ [0.15cm]
$[\mathrm{Fe}/\mathrm{H}]$ [dex] & & $+0.28^{+0.11}_{-0.11}$ &$+0.36 \pm 0.07$ & \\ [0.15cm]
$Y_{\rm ini}$ & & $0.288^{+0.009}_{-0.008}$ &$0.27 \pm 0.03$ & $0.2901$ \\ [0.15cm]
$X_{\rm ini}$ & & $0.685^{+0.013}_{-0.015}$ &$0.70 \pm 0.03$ & $0.6809$ \\ [0.15cm]
$M_{\rm core}\,\rm [M_\star]$\tnote{b} & & $0.077^{+0.007}_{-0.008}$ & $0.076$ & \\ [0.15cm]
$\alpha$ & & $1.791$ (fixed) &$1.88 \pm 0.16$ & $2.00$ (fixed) \\ [0.15cm]
$f_\mathrm{ov}$ & & $0.016$ (fixed) &$0.003 \pm 0.002$ & \\ [0.15cm]
\bottomrule
\end{tabular}
\label{tab:mod_results}
\begin{tablenotes}
\tiny
\item [a] Values from the model with the smallest $\chi^2$, with convective overshoot over $0.1$ pressure scale heights included (model No.~2 in their Table~2). No uncertainties are reported by the authors.
\item [b] Mass of the convective core from the position of the Schwarzschild boundary. The overshooting region extends beyond this point.
\end{tablenotes}
\end{threeparttable}
\end{center}
\end{table*}
\begin{figure}
\includegraphics[width=\columnwidth]{echelle5.pdf}
\caption{\footnotesize \'{E}chelle diagram for \hat{} using $\rm \Delta\nu = 59.22\,\mu Hz$. The grey scale range from white at low power to black at high power. Circles give the extracted frequencies with corresponding uncertainties, triangles connected by red lines give model frequencies from GARSTEC, while squares connected by green lines give model frequencies from MESA. The degree of each ridge is indicated in the top part of the plot. The radial order of the $l=0$ modes is indicated by the numbers in the right side of the plot.}
\label{fig:model_compare}
\end{figure}
A detailed modelling of HAT-P-7 was first made by \citet{2010ApJ...713L.164C} based on asteroseismic measurements of the solar-like $p$-mode oscillations in the star. This work was based on the SC Q0-Q1 data from the \emph{Kepler}\ satellite. A fit of $33$ individual mode frequencies, with values obtained from peaks of a smoothed power spectrum, was made to models computed using ASTEC \citep{2008Ap&SS.316...13C}, with adiabatic pulsation frequencies calculated using ADIPLS \citep[Aarhus adiabatic oscillation package,][]{2008Ap&SS.316..113C}.
Here we use the updated set of $50$ frequencies from our peak-bagging to model \hat{}, and use two different codes and modelling schemes to asses the robustness of our results. We note that this approach does not safeguard against other potential sources of systematics that can arise from, for instance, input physics not covered by the two codes or from assumptions made on certain quantities in the modelling. The effect of such systematics will be studied elsewhere (Silva~Aguirre et al. (in prep.)).
The results from the modelling can be found in Table~\ref{tab:mod_results}. The \'{e}chelle diagram \citep[][]{1983SoPh...82...55G} of \hat{} is given in \fref{fig:model_compare} after correcting for the background, and overlaid are both peak-bagged and modelled frequencies. In the construction of the \'{e}chelle diagram we have on the ordinate plotted the mid frequency of the respective $\Delta\nu$-length segments. Here $\Delta\nu$ denotes the so-called \emph{large-separation}, computed as the frequency difference between consecutive radial orders of a given degree. To obtain a better representation of the ridges for illustrative purposes a constant value was added to the frequencies before taking the modulo, thus allowing a shift of the ridges on the abscissa \citep[see, e.\,g.,][]{2011arXiv1107.1723B}. The $\Delta\nu$ from \citet[][]{2013ApJ...767..127H} was used.
\subsubsection{MESA model}
\label{sec:MESA}
As a first approach \hat{} was modelled using MESA \citep[Modules for Experiments in Stellar Astrophysics,][]{ref:mesa11,ref:mesa13} and ADIPLS \citep[][]{2008Ap&SS.316..113C}. For the modelling we used the value of the effective temperature, $T_\mathrm{eff}\rm = 6350 \pm 126 \, K$, the heavy-element abundance, $[\mathrm{Fe}/\mathrm{H}] = +0.26 \pm 0.15$ \citep[both from][]{2014ApJS..211....2H}, and the frequencies given in Table~\ref{tab:mcmc}, except for the lowest $l = 0$ mode which gave consistently extremely poor agreement with the models and was subsequently excluded.
For the input physics we chose to neglect diffusion and settling following \citet{2010ApJ...713L.164C}. We used the 2005 update of the OPAL EOS \citep{ref:opal_1996,ref:opal_2002} and the NACRE nuclear reaction rates \citep{ref:nacre_reac_rates} with the updated $^{14}\mathrm{N}(p,\gamma)^{15}\mathrm{O}$ reaction rate by \citet{ref:14N} and the updated $^{12}\mathrm{C}(\alpha,\gamma)^{16}\mathrm{O}$ reaction rate by \citet{ref:12C}. Furthermore we used the OPAL opacities \citep{ref:opal_opacity} assuming the solar chemical composition given by \citet{ref:AGSS09}, supplemented by the \citet{ref:lowT_2005} opacities at low temperatures. We used the mixing-length theory of convection as formulated by \citet{ref:mixinglength}. Finally, we chose to use the "simple photosphere" option in MESA for the atmospheric boundary condition, which constitutes a grey atmosphere with the optical depth, $\tau_{\rm s}$, to the base of the atmosphere of $2/3$ \citep[see][their Equation~(3)]{ref:mesa11}.
In the matching of the model frequencies to the observed frequencies, we corrected the model frequencies for near-surface effects \citep{ref:surfaceterm} using the prescription by \citet{ref:brandao} with $b = 4.90$ and a reference frequency of $\rm \nu_0=1100 \, \mu Hz$. The best-fitting model was found via a $\chi^2$-minimisation \citep[see, e.\,g.,][]{2010ApJ...713L.164C,ref:brandao,2013ApJ...763...49D}. The total $\chi^2$ was found as the weighted average of a normalised frequency and a spectroscopic component with weights given as $\chi^2 = \tfrac{2}{3}\chi_\nu^2 + \frac{1}{3}\chi_\mathrm{spectro}^2$ \citep[following][]{ref:mesa13} and the normalisation given by the number of values entering the $\chi^2$ component. The component $\chi_\mathrm{spectro}$ included the match of model and spectroscopic values for $T_\mathrm{eff}$ and $[\mathrm{Fe}/\mathrm{H}]$. The uncertainty of a given model parameter was found as the likelihood-weighted standard deviation of total $\chi^2$ values for all the computed models. Since we have chosen to use the model parameters for the best-fitting model instead of the likelihood-weighted mean values, we have added the difference between the mean and the best-fitting values in quadrature to the uncertainties.
The parameters of the best-fitting model and their uncertainties are listed in Table~\ref{tab:mod_results}.
\subsubsection{GARSTEC model}
\label{sec:GARSTEC}
\begin{figure}
\includegraphics[width=\columnwidth]{ratios2-eps-converted-to.pdf}
\caption{\footnotesize Ratios $\rm r_{02}$ and $\rm r_{010}$ as a function of frequency. The lines show the ratios obtained for the best-fit GARSTEC (black solid) and MESA (dashed blue) models. Note that the relatively small uncertainties on frequencies (horizontal error bars) renders them undiscernible on the given scale.}
\label{fig:ratio}
\end{figure}
In the second approach we have used two grids of stellar models computed with the Garching Stellar Evolution Code \citep[GARSTEC, see][]{Weiss:2008jy}.
The input physics is similar to the description given for the MESA model with the difference that we used the mixing-length theory of convection of \citet{2013sse..book.....K}, and \citet{Grevesse:1998cy} solar abundances. Also, the updated $^{12}\mathrm{C}(\alpha,\gamma)^{16}\mathrm{O}$ reaction rate was not used.
In one grid we included the effects of core overshooting using an exponential decay of the convective velocities with an efficiency of $f_\mathrm{ov}=0.016$ \citep{Magic:2010iz}.
The grids cover a mass range between $0.7$ and $1.8\,M_{\sun}$ in steps of $0.01\,M_{\sun}$ and initial compositions of $-0.65<[\rm Fe/H]<+0.50$ in steps of 0.05~dex. These were determined using a galactic chemical evolution law of $\Delta Y/\Delta Z=1.4$ \citep[see, e.\,g.,][]{2006AJ....132.2326B} anchored to the Big Bang nucleosynthesis value of $Y_p=0.248$ \citep{Steigman:2010gz}. For hundreds of models along each evolutionary track, we computed theoretical oscillations frequencies using ADIPLS \citep[][]{2008Ap&SS.316..113C}. These allowed us to construct dense grids of models covering the spectroscopic and asteroseismic parameter space of our target.
To determine the stellar parameters, we used the Bayesian approach described in Silva~Aguirre et al. (in prep.). Briefly, we assumed a flat prior in $\rm [Fe/H]$ and age including only a strict cut on the latter at $15\rm\, Gyr$, and we used a standard Salpeter IMF \citep[][]{1955ApJ...121..161S}. We computed the likelihood of the observed values given a set of model parameters assuming Gaussian distributed errors. In our case, the observables included were the spectroscopic $T_{\rm eff}$ and $\rm [Fe/H]$, and the frequency ratios defined as \citep[][]{2003A&A...411..215R}
\begin{align}
r_{01}(n) &= \frac{d_{01}(n)}{\Delta\nu_1(n)}, \quad r_{10}(n)=\frac{d_{10}(n)}{\Delta\nu_0(n+1)}\\
r_{02}(n) &=\frac{\nu_{n,0}-\nu_{n-1,2}}{\Delta\nu_1(n)}\, .
\end{align}
Here the $d_{01}$ and $d_{10}$ are the smooth 5-point small frequency separations given as
\begin{align}
d_{01}(n) &= \tfrac{1}{8} \left( \nu_{n-1,0} - 4\nu_{n-1,1} + 6\nu_{n,0} -4\nu_{n,1} +\nu_{n+1,0} \right)\\
d_{10}(n) &= -\tfrac{1}{8} \left( \nu_{n-1,1} - 4\nu_{n,0} + 6\nu_{n,1} -4\nu_{n+1,0} +\nu_{n+1,1} \right)\,.
\end{align}
We refer to \citet[][]{SilvaAguirre:2013in} for further details.
In \fref{fig:ratio} we show the ratios obtained from the peak-bagging as a function of frequency along with ratios from the best-fit model. Here we also show the corresponding ratios obtained from the best-fit MESA model. Note that the ratios were not fitted in the MESA modelling. The construction of the ratios introduces correlations as a function of frequency that need to be taken into account when calculating the likelihood.
This is done by calculating the $\chi^2$ entering the likelihood function as
\begin{equation}
\chi^2 = \frac{1}{N}\left(\boldsymbol{x}_{\rm obs} - \boldsymbol{x}_{\rm model} \right)^T \boldsymbol{\rm C}^{-1} \left(\boldsymbol{x}_{\rm obs} - \boldsymbol{x}_{\rm model} \right)\, ,
\end{equation}
where $\boldsymbol{\rm C}$ is the covariance matrix and the vectors $\boldsymbol{x}$ (of length $N$) give the observed and modelled values.
As an illustration of $\boldsymbol{\rm C}$ we show in \fref{fig:hinton} the Hinton diagram for the $\rm r_{010}$ ratios, which provides a qualitative view of the correlation matrix for these ratios. To construct ratios we use the PPDs from the individual frequencies. From these distributions we obtain our central values and uncertainties and use the Pearson standard correlation coefficient\footnote{Using the Python package \texttt{Pandas}. } to compute the correlation matrix.
We found that the grid including core overshooting provided the best results.
The final set of parameters from the modelling are given in Table~\ref{tab:mod_results}. These were obtained from the median of the posterior probability distribution function and the 16 and 84 percent values. For the sake of comparison with the MESA model, the frequencies of the best-fit model are shown in \fref{fig:model_compare} after applying the \citet[][]{2008ApJ...683L.175K} method to correct for near-surface effects. Note, that the surface correction is not needed for this modelling approach as frequency ratios, not strongly affected by the surface layers \citep[][]{2003A&A...411..215R, 2011A&A...529A..63S}, are used rather than the actual frequencies. Consequently, any apparent misfit there might be between observed and surface corrected model frequencies in the \'{e}chelle diagram cannot simply be interpreted as being caused by a poor model, but could just as well be due to a poor surface correction. In the current case the fit, qualitatively speaking, reproduce the observations satisfactorily.
\begin{figure}
\centering
\includegraphics[scale=0.45, trim=0cm 1.2cm 0cm 0cm, clip=true,]{hinton_R010_cov-eps-converted-to.pdf}
\caption{\footnotesize Hinton diagram for the correlation matrix of $\rm r_{010}$ ratios. The first part of the subscript denotes if the ratio if of type $"01"$ or $"10"$, while the second gives the radial order of the central frequency. White (black) squares indicate positive (negative) covariances between ratios in question, and the size give the relative size of the correlation (one along the diagonal). }
\label{fig:hinton}%
\end{figure}
\subsubsection{Model comparison}
\label{sec:model_compare}
An improvement in our analysis compared to other works using \emph{Kepler}{} data is that now the full available dataset can be utilised. Furthermore, we find that the approach adopted in this work for the extraction of mode frequencies, \ie{} using an MCMC peak-bagging scheme, should provide more reliable estimates and uncertainties for the frequencies compared to the approach of assigning frequencies from peaks in a smoothed power spectrum.
In Table~\ref{tab:mod_results} we compare our results to the original asteroseismic results from \citet{2010ApJ...713L.164C}.
It is noteworthy that the GARSTEC model agrees quite well with these values, where only data from Q$0$-Q$1$ was used.
One contributing factor to the relatively small changes in model results from the addition of significantly more data is the F-type character of the star, where the frequency precision is limited by the large mode linewidths (see Appendix~\ref{app:lwvis}).
Our MESA modelling results in a slightly more massive star than those obtained by GARSTEC and \citet{2010ApJ...713L.164C}, but still agree within uncertainties despite the different fitting techniques.
The asteroseismic modelling by \citet{2012AN....333.1088V} resulted in a less massive ($M_{\star}=1.36\, M_{\sun}$) and less metal-rich ($[\mathrm{Fe}/\mathrm{H}]=0.13$) star than the current modelling efforts. This is likely because the authors explored a limited metallicity range and no spectroscopic constraints were included in the optimisation.
Our stellar parameters agree with those derived by \citet{2008ApJ...680.1450P} from a combined analysis of stellar isochrones, photometry, and spectroscopy. With our asteroseismic analysis we reduce the uncertainties on mass, radius, and age to $4.2\%$, $1.1\%$, and $11.2\%$ respectively from the original estimates of $6.4\%$, $13.9\%$, and $45.5\%$ given in \citet{2008ApJ...680.1450P}. In this comparison asymmetric uncertainties were added in quadrature.
Since the GARSTEC values are computed using frequency ratios that are insensitive to near-surface effects, this set of stellar parameters will be adopted for the remainder of the paper. Furthermore, this Bayesian approach offers a direct estimation of the parameter uncertainties from the posterior distributions from a dense grid of models.
Finally, we note that the values of mass, radius, and age obtained from AME \citep{ref:ame}; $M/M_\sun = 1.55 \pm 0.05$, $R/R_\sun = 1.99 \pm 0.02$, $\mathrm{age} = 1.9 \pm 0.5 \, \rm Gyr$, are consistent with the results from this work.%
\subsubsection{Updated planetary parameters}
\label{sec:planet}
With our new estimates for the stellar parameters we can update the mass and radius for HAT-P-7b. The mass is found from \citep[see, e.\,g.,][]{2010arXiv1001.2010W}
\begin{equation}
\frac{M_p}{(M_p + M_{\star})^{2/3}} = \frac{K_{\star}\sqrt{1-e^2}}{\sin i_p} \left( \frac{P_{\rm orb}}{2\pi G}\right)^{1/3}\, ,
\end{equation}
where $M_p$ is the planet mass, $e$ is the eccentricity, $P_{\rm orb}$ is the orbital period, and $K_{\star}$ is the stellar reflex velocity. As input we use $i_p$ and $P_{\rm orb}$ from \citet[][]{2013ApJ...774L..19V}, $e$ is set to zero following \citet[][]{2012MNRAS.422.3151H}, and $K_{\star}=212.2\pm 3.2\, \rm m\, s^{-1}$ from combining estimates of $K_{\star}$ by \citet{2009ApJ...703L..99W} and \citet{2009PASJ...61L..35N}. Solving for the planet mass then gives\footnote{Using $M_J = 1.899\,10^{30}\,\rm g$ and $R_J=7.1492\, 10^9\,\rm cm$.} $M_p=1.80\pm 0.05\, M_J$. For the planetary radius we get $R_p=1.51\pm 0.02\, R_J$ when using $R_p/R_{\star}$ from \citet[][]{2013ApJ...774L..19V}. Considering the discussion in \citet[][]{2013ApJ...774L..19V} on the size of systematic effects on transit depth measurements, we adopted a one per cent uncertainty on $R_p/R_{\star}$. This gives a planetary density of $\rho_p=0.65\pm 0.03\, \rm g\, cm^{-3}$, which is just within the uncertainty of the estimate from \citet{2008ApJ...680.1450P} of $\rho_p=0.876^{+0.17}_{-0.24}\, \rm g\, cm^{-3}$. We note that the stellar parameters estimated from asteroseismology offers a greatly improved precision on the planetary parameters.
\subsection{Splitting and inclination}
\label{sec:splitinc}
In \fref{fig:bestfit} we present the results for the splitting and inclination of HAT-P-7 as the PPDs for these parameters, along with the 2D correlation map. These results were obtained from the small fit described above. Indicated in \fref{fig:bestfit} are the $68\%$ and $95\%$ HPD credible regions.
The distributions for inclination and splitting for HAT-P-7 are unfortunately not simple and Gaussian, but rather cover an extended region of parameter space. Considering this we will refrain from using measures of central tendency, such as the median or mode of the distributions, but rather report values from the $68\%$ HPD credible regions. The overall behaviour seen in most of the correlation map correspond well to what might be expected given the relatively large linewidths: $\rm\Gamma = 4.8 - 8.4\, \mu Hz$ for the range fitted (see \fref{fig:linewidth}). At low inclinations, the central $m=0$ component dominates the relative heights of azimuthal components of a rotationally split multiplet, while at high inclinations the sectoral ($m=\pm l$) components dominate. For $l=2$ the tesseral ($0<|m|<l$) components dominate at an inclination of ${\sim}50^{\circ}$. Thereby, at low inclinations the splitting is generally less constrained as the $m\neq0$ components are small and harder to fit. Furthermore, a linewidth exceeding the splitting effectively hides these small components up to a splitting of half the linewidth (\ie, up to $\rm{\sim}4\, \mu Hz$). At higher inclinations the splitting is more easily discernible as the visible azimuthal components are the ones furthest apart, thus the linewidth becomes less of a problem.
The reader will notice a high density at $i_{\star} \sim 10^{\circ}$ and $\nu_s > 5\, \rm \mu Hz$. This should not be interpreted as a good solution to the fit of the power spectrum but is rather the result of walkers trying to escape the parameter space, with the consequence that the walkers pile up at the boundary of the prior on the splitting. A contributing factor could also be that the small separation between $l=0$ and $l=2$ modes is of the order ${\sim}5\, \rm \mu Hz$, whereby $l=2$ azimuthal components would be within the $l=0$ mode profile. We see no correlations between these high splitting values and any of the other parameters of the fit and so these parameters are unaffected by this feature. These walkers will, however, contribute to the shape of the splitting and inclination PPDs and therefore we chose to exclude walkers with $\nu_s > 5\, \rm \mu Hz$ in the marginalization of the $i_{\star}$ and $\nu_s$ PPDs.
This is the same effect we observed when the background was kept free in the fit, but now much decreased. It is, however, not clear to us what is special about the inclination of $i_{\star} \sim 10^{\circ}$ in facilitating this escape of walkers.
From the $68\%$ HPD credible regions of the inclination and splitting PPDs we find $i_{\star}<36.5^{\circ}$ and $\nu_s < 0.87 \, \rm \mu Hz$ (see Table~\ref{tab:astro}). The value for the inclination is in overall agreement with the notions made by others based on the low value for $v\sin i_{\star}$. In addition, agreement is found with the statistically derived estimate by \citet[][]{2010ApJ...719..602S} of $i_{\star} = 6.7 - 10^{\circ}$ based on a simple $P_{\rm rot}\propto t^{1/2}$ model \citep[][]{1967ApJ...148..217W,1972ApJ...171..565S} for the evolution of the stellar rotation period as a function of mass and age.
The splitting, on the other hand, is lower than expected for an F6-type star (see below).
However, it is clear from the correlation map that at low inclinations, values of the splitting up to $\rm{\sim}4\, \mu Hz$ is allowed.
\begin{figure*}
\centering
\includegraphics[width=0.8\textwidth]{corr_inc_splitting3.pdf}
\caption{\footnotesize Top: marginalized posterior probability distribution (PPD) for the stellar inclination $i_{\star}$. Here we have folded the full distribution ($-90$ to $180^{\circ}$) onto the the range from $0$ to $90^{\circ}$. Bottom right: PPD for the rotational splitting. Bottom left: correlation map between the inclination and the rotational splitting.
The $68\%$ credible regions (highest posterior density credible regions) are indicated by the dark gray part, while the light gray indicates the additional part covered in a $95\%$ credible region. Included are also lines of constant $v\sin i_{\star}$, computed using the radius estimate from our analysis and with $v\sin i_{\star}$ values from the literature (see Table~\ref{tab:orbit}), arranged in the legend in order of increasing $v\sin i_{\star}$ value (top to bottom). The dashed black horizontal line indicates the value of $\nu_s = 5\, \rm \mu Hz$ above which walkers were excluded from the PPDs. The grey dash-dot lines give lines of constant $v\sin i_{\star}$ from $8$ to $48\,\rm km\, s^{-1}$ in steps of $8\,\rm km\, s^{-1}$.}
\label{fig:bestfit}
\end{figure*}
\begin{table}
\begin{center}
\begin{threeparttable}
\caption{Values related to the stellar inclination and rotation. All values are estimated from the $68\%$ HPD credible region of their corresponding parameter distributions.}
\begin{tabular}{ll}
\toprule \\ [-0.3cm]
Parameter & $68\%$ HPD limit \\[0.1cm]
\midrule
$i_{\star}\, [^{\circ}]$ & \hspace{0.5cm}$<36.5$ \\ [0.2cm]
$\rm \nu_s\, [\mu Hz]$ & \hspace{0.5cm}$<0.87$ \\ [0.2cm]
$P_{\rm rot}\,\rm [days]$ & \hspace{0.5cm}$>13.23$ \\ [0.2cm]
$v\sin i_{\star}\, [\rm km\, s^{-1}]$ & \hspace{0.5cm}$< 2.21$ \\ [0.2cm]
$v_{\rm surf}\, [\rm km\, s^{-1}]$ & \hspace{0.5cm}$<7.66$ \\ [0.2cm]
\bottomrule
\end{tabular}
\label{tab:astro}
\end{threeparttable}
\end{center}
\end{table}
With our estimate for the stellar radius (see \sref{sec:modelling}) we may convert the estimated ranges for the splitting and inclination to a measure of $v\sin i_{\star}$ as \citep[][]{2013ApJ...766..101C}
\begin{equation}
v\sin i_{\star} = 2\pi R \nu_s \sin i_{\star}\, .
\label{eq:vsini}
\end{equation}
Using the PPDs obtained for $\nu_s$ and $i_{\star}$, while assuming a distribution for the radius as $R/R_{\odot}\sim \mathcal{N}(2.00, 0.02)$, we find value of $ v\sin i_{\star} < 2.21$ km s$^{-1}$ ($68\%$ HPD credible region). In \fref{fig:vsini} we have plotted the full distribution for $ v\sin i_{\star}$ and indicated the values obtained from spectral and RM analysis (see Table~\ref{tab:orbit}). We find that the value of \citet{2009PASJ...61L..35N} agree within uncertainties with our results. \citet{2009ApJ...703L..99W} found the largest value for $ v\sin i_{\star}$. Their higher $ v\sin i_{\star}$ value results from the lower orbital inclination value (see Table~\ref{tab:orbit}) they used for modelling the RM effect. For HAT-P-7 a larger impact parameter (lower inclination) requires a larger $ v\sin i_{\star}$. If one would repeat the analysis with the inclination derived from the \emph{Kepler}{} light curve, instead of the ground
based data available to them, then this disagreement in $ v\sin i_{\star}$ would vanish.
Having $ v\sin i_{\star}$ and $i_{\star}$ measured and assuming solid body rotation we can now calculate the true rotation speed of HAT-P-7, $v<7.66 \,\rm km\, s^{-1}$ ($68\%$ HPD credible region).
This upper limit estimate agrees with the values from \citet[][]{2013A&A...557L..10N} from the rotation periods of \emph{Kepler}{} stars of an approximate spectral type between F4 and 6.
\begin{figure}[h]
\includegraphics[width=\columnwidth]{vsini.pdf}
\caption{\footnotesize Distribution for $ v\sin i_{\star}$ constructed from the PPDs obtained for $\nu_s$ and $i_{\star}$ and assuming $R/R_{\odot}\sim \mathcal{N}(2.00, 0.02)$ (see \eqref{eq:vsini}). The dark and light grey regions of the distribution corresponds to the $68\%$ and $95\%$ credible regions, respectively. Literature values for $v\sin i_{\star}$ obtained from spectral and RM analysis (see Table~\ref{tab:orbit}) are given by vertical lines where the shaded regions above the distribution give the corresponding uncertainties.}
\label{fig:vsini}
\end{figure}
\section{Obliquity}
\label{sec:obliquity}
\begin{table*}
\begin{center}
\begin{threeparttable}
\caption{Literature Values for Parameters Related to the \hat{} System.}
\begin{tabular}{lcccc}
\toprule \\ [-0.3cm]
Source & $v\sin i_{\star}$ [km s$^{-1}$] & $\lambda\, [^{\circ}]$ & $i_{p}\, [^{\circ}]$ & $\psi$ (this work) \\[0.1cm]
\midrule
\citet{2008ApJ...680.1450P} & 3.8 $\pm$ 0.5 & & 85.7$^{+3.5}_{-3.1}$ & \\ [0.2cm]
\citet{2009ApJ...703L..99W} & 4.9$^{+1.2}_{-0.9}$ & 182.5 $\pm$ 9.4 & 80.8$^{+2.8}_{-1.2}$ & $83^{\circ}<\psi<119^{\circ}$\\[0.2cm]
\citet{2009PASJ...61L..35N} & 2.3$^{+0.6}_{-0.5}$ & -132.6$^{+10.5}_{-16.3}$\tnote{a} & 85.7$^{+3.5}_{-3.1}$ & $83^{\circ}<\psi<106^{\circ}$\tnote{b}\\[0.2cm]
\citet{2012ApJ...757...18A} & 2.7 $\pm$ 0.4 & 155 $\pm$ 37\tnote{c} & & $83^{\circ}<\psi<111^{\circ}$\\[0.2cm]
\citet{2012ApJ...757..161T} & 4.2$\pm$ 0.5 & & & \\[0.2cm]
\citet{2013ApJ...774L..19V} & & & 83.151$^{+0.030}_{-0.033}$ & \\[0.2cm]
This work & $< 2.21$ & & & \\[0.1cm]
\bottomrule
\end{tabular}
\label{tab:orbit}
\begin{tablenotes}
\tiny
\item [a] The value for the projected obliquity is equivalent to $\lambda=227.4^{+10.5\circ}_{-16.3}$.
\item [b] We used an uncertainty on $\lambda$ of $\pm 16.3^{\circ}$.
\item [c] From a fit to RM data \citet{2012ApJ...757...18A} obtained an uncertainty on $\lambda$ of $\pm14^{\circ}$. The uncertainty adopted here was ascribed by \citet{2012ApJ...757...18A} as the standard deviation of the three independent measurements of $\lambda$.
\end{tablenotes}
\end{threeparttable}
\end{center}
\end{table*}
With our estimate for the stellar inclination, $i_{\star}$, $\lambda$, and the planetary orbital inclination, $i_p$, we are now able to calculate the system obliquity, $\psi$, from \citep[][]{2005ApJ...631.1215W}
\begin{equation}
\cos \psi = \sin i_{\star} \cos \lambda \sin i_p + \cos i_{\star} \cos i_p \, .
\end{equation}
In \fref{fig:ob} we show the distribution for $\psi$ when using the obtained distribution for $i_{\star}$ (see \fref{fig:bestfit}), while adopting $i_p$ from \citet[][]{2013ApJ...774L..19V}, and $\lambda$ from \citet{2012ApJ...757...18A}. From the $68\%$ HPD credible region we obtain $83^{\circ}<\psi<111^{\circ}$, consistent with a polar orbit. The corresponding results from using $\lambda$ from \citet{2009ApJ...703L..99W} and \citet{2009PASJ...61L..35N} are given in Table~\ref{tab:orbit}.
We refer to \citet{2012ApJ...757...18A} for a discussion on the different values for $\lambda$ and the possible reasons for their disagreement \citep[see also][]{2011ApJ...738...50A}. Here we note that regardless which $\lambda$-value we use we find a polar orbit for HAT-P-7 b.
When an assessment of the obliquity is made without knowing the stellar inclination, a flat distribution in $\cos i_{\star}$ is generally assumed for the stellar orientation, as this results in an isotropic distribution for the stellar inclination. From such a distribution it is \emph{a priori} much more likely to observe a random star in an equator-on configuration. In \fref{fig:ob} we show the distribution in $\psi$ from adopting this isotropic distribution in $i_{\star}$. In this way \citet{2009ApJ...703L..99W} estimated $\psi>86.3^{\circ}$ with $99.73\%$ confidence, while \citet{2009PASJ...61L..35N} found $\psi>90^{\circ}$ with $99.70\%$ confidence \citep[both used $i_p$ from][]{2008ApJ...680.1450P}. From this approach a retrograde orbit is thus strongly suggested, but the orbit has a higher probability of being more equatorial than polar.
With our asteroseismic estimate for the inclination we can substantiate these statistical results as we find $\psi>90^{\circ}$ with $68\%$ credibility using $\lambda$ from \citet{2012ApJ...757...18A}, and now a near polar orbit is the most likely configuration for the system.
\begin{figure}[h]
\includegraphics[width=\columnwidth]{obliquity-eps-converted-to.pdf}
\caption{\footnotesize Distribution of the true angle $\psi$ using the distribution for $i_{\star}$ from the peak-bagging, and with the assumption of Normal distributions for the planetary inclination $i_p$ and projected angle $\lambda$. For $i_p$ the value from \citet{2013ApJ...774L..19V} was adopted, while $\lambda$ was taken from \citet{2012ApJ...757...18A}. The dark and light grey regions of the distribution corresponds to the $68\%$ and $95\%$ credible regions, respectively. The dashed curve gives the distribution if one instead assumes an isotropic distribution for $i_{\star}$, \ie, flat in $\cos i_{\star}$. Indicated is also which values of $\psi$ that corresponds to a retrograde or prograde orbit of HAT-P-7b. }
\label{fig:ob}
\end{figure}
\section{Comparison with gyrochronology}
\label{sec:gyro}
Our result for the limits on the stellar rotation rate (see Table~\ref{tab:astro}) can be compared to empirically calibrated gyrochronology relations.
Here we use the form described by \citet[][]{2007ApJ...669.1167B} given as
\begin{equation}
P(B-V, t) = t^n \times a\left[(B-V)_0 - c \right]^b\, ,
\label{eq:gyro}
\end{equation}
where $t$ is the stellar age in Myr, while $a$, $b$, and $n$ are empirically determined coefficients that vary depending on the calibration set used \citep[see, e.\,g.,][]{2014ApJ...780..159E}.
To compare with this relation, we first need an estimate for the de-reddened colour $(B-V)_0$. The procedure used to get $(B-V)_0$ is described in Appendix~\ref{app:BV}.
Using the estimate $(B-V)_0=0.495\pm 0.022$ together with the age determined in \sref{sec:modelling} of $t=\rm 2.07\pm 0.36 \, Gyr$ (see Table~\ref{tab:mod_results}; asymmetric uncertainties were added in quadrature), we can estimate the rotation period from the relation in \eqref{eq:gyro}. In \fref{fig:gyro} we show two versions of this relation, namely those by \citet[][]{2007ApJ...669.1167B} (B07), and \citet[][]{2009ApJ...695..679M} \citep[M09; see also][]{2010ApJ...721..675B,2011ApJ...733..115M}. From these we get periods of $9.9\pm 1.8$ (B07) and $5.0\pm 4.0$ days (M09).
The upper value from the B07 relation is comparable to our lower limit period estimate of about $13$ days (see Table~\ref{tab:astro}), so the agreement is not very convincing. On the other hand, the splittings at low inclinations match well this range of rotation periods as the level of ${\sim}4\, \rm \mu Hz$ corresponds to a rotation period of ${\sim}2.9$ days (see \fref{fig:bestfit}).
We note that a $P_{\rm rot}\propto t^{1/2}$ law might provide a poor description of the rotational evolution for HAT-P-7, as its $(B-V)_0$ puts it in close proximity to the so-called \emph{Kraft break} \citep[][]{1967ApJ...150..551K} where loss of angular momentum via a stellar wind is inhibited by the lack of a sufficiently deep convection zone. If HAT-P-7 is on the low $(B-V)_0$ side of this break the rotation rate becomes a strong function of the initial conditions \citep[see, e.\,g.,][]{2013ApJ...776...67V}, and a gyrochronology scaling is not applicable.
The close-in hot-jupiter HAT-P-7b has potentially had an impact on the rotation rate of its host star. While a detailed dynamical analysis of this system is beyond the scope of this paper, the synchronisation of rotation and alignment clearly has not been reached yet. The time scale for circularisation has been reached though, with an upper limit on the eccentricity of $e<0.038$ given by \citet[][]{2012MNRAS.422.3151H}. While it is difficult to assess the past interaction between the planet and the star, we note that with the configuration we estimate for the system, \ie, close to a polar orbit of the planet, the rotational synchronisation time will likely be very long as the angular momentum vectors are close to being orthogonal, thereby decreasing the tidal interaction.
For discussions on the interaction between hot jupiters and their host stars we refer the reader to, e.\,g., \citet[][]{1981A&A....99..126H}, \citet[][]{2009ApJ...702.1413M}, \citet[][]{2010ApJ...723L..64C}, \citet[][]{2010ApJ...725.1995M}, \citet[][]{2011IAUS..276..238M}, and \citet[][]{2014ApJ...786..102V}.
\begin{figure}
\includegraphics[width=\columnwidth]{gyro2.pdf}
\caption{\footnotesize Gyrochronology relations from \citet[][]{2009ApJ...695..679M} (M09), \citet[][]{2007ApJ...669.1167B} (B07). Shaded region around each relation (with the same colour) represent the standard error of the relationship between $(B-V)_0$ and rotation period from propagating the uncertainties reported for the coefficients entering \eqref{eq:gyro} together with the uncertainty in our age estimate. The vertical line and shaded region give the value of $(B-V)_0=0.495\pm0.022$ of \hat{}, while darker shaded regions around the relations vertically bound the corresponding uncertainty in the period from the uncertainty in $(B-V)_0$, age, and the gyrochronology relations. Horizontal lines indicates the limits on the $68\%$ (dashed) and $95\%$ (dotted) credible regions on the stellar rotation period from $p$-mode splittings. The upward pointing arrows indicate that these rotation periods are lower limits from the respective HPD credible regions.}
\label{fig:gyro}
\end{figure}
\section{Activity signatures}
\label{sec:activity}
Activity of stars is linked to the interaction between rotation, convection, and magnetic fields. The interaction can cause magnetic features, such as dark spots and bright faculae, to appear in the photosphere of the stars \citep[see, e.\,g.,][for a review]{lrsp-2005-8}, and plage in the chromosphere. If a star is rotating, such dark or bright regions will induce a temporal modulation of the integrated stellar flux.
The photospheric features can also cause departures from radiative equilibrium in the stellar chromosphere in plage, and induce emission in the cores of specific spectral lines. For example is emission in cores of Ca \textsc{ii} H\&K lines an often used indicator for magnetic activity \citep[][]{2001A&A...376.1080K,2004ApJS..152..261W,2011A&A...532A..81F,2012A&A...543A.146F}, as it is believed to reflect the amount of non-thermal chromospheric heating above faculae \citep[see, e.\,g.,][for a review]{2008LRSP....5....2H}. For close-in hot-jupiter systems there is furthermore the possibility for a direct magnetic interaction between the planet and the star, where magnetic reconnections, similar to those seen in flares, can cause a heating of the chromosphere \citep[see, e.\,g.,][]{2005ApJ...622.1075S}.
\subsection{Chromospheric activity}
To test for signatures of chromospheric activity we searched the Ca \textsc{ii} H\&K lines in the HIRES spectra from \citet{2009ApJ...703L..99W}, but found no signs of emission or high activity levels. As we see $p$-mode oscillations in HAT-P-7, it must have an outer convection zone, and the $(B-V)_0$ colour derived in \sref{sec:gyro} places it above the limit from \citet[][]{1991ApJ...380..200S} for the onset of activity. The absence of an emission signal could point towards an intrinsically low surface activity caused by a low rotation rate.
Another effect that would influence the signal is if \hat{} has a low inclination angle. On the Sun, faculae are primarily located in the active latitude bands between around $5$ and $40^{\circ}$ latitude. As faculae and plage are believed to have the same magnetic driver the Ca \textsc{ii} H\&K emission from plage will thus decrease with decreasing inclination from the decreasing projected area of the active regions \citep[][]{2001A&A...376.1080K,2004A&A...413..657F,2007MNRAS.377...17C}.
\subsection{Low frequency region}
\label{sec:peak}
We searched for a signal imparted by a temporal flux modulation at low frequencies in the power spectrum of \hat{} \citep[see, e.\,g.,][]{2011A&A...534A...6C,2013A&A...557L..10N}. Examples of such activity signals in F-type \emph{Kepler}{} stars can for instance be seen in \citet[][]{2014A&A...562A.124M}.
The signal from the stellar rotation can also be estimated by, e.\,g., the autocorrelation function (ACF) of the time series, as shown for instance by \citet[][]{2013MNRAS.432.1203M}.
In \fref{fig:accps} we show the ACF of the corrected time series together with the low-frequency end of the power spectrum (in units of period for convenience).
The ACF shows no clear sign of modulation. To test the hypothesis that a signal in the time series has died out at lag $k$, we use the large-lag standard error \citep[][]{anderson1976time}, indicated in \fref{fig:accps}. From this it is clear that the signal seen in the ACF is not significant.
However, the low-amplitude hump at around ${\sim}9$ days in the ACF does seem to align with an increase in power in the power spectrum.
We investigated the presence of a modulation further via the Morlet wavelet transform \citep[][]{1998BAMS...79...61T} of the time series \citep[see also][]{2013A&A...550A..32M}, this is shown in \fref{fig:wave}. No clear periodicity is seen, and the signal, giving rise to the ${\sim}9$ day hump in the ACF and the power spectrum, seems to be very intermittent.
Finally we checked the magnetic proxy from \citet[][]{2014ApJ...783..123C}, but also here we did not find indications of a signal.
The fact that a stronger or more localised signal is missing for \hat{} might again be linked to a low stellar inclination. Indeed, if magnetic features such as stellar spots reside primarily near the equator of the star, while we view it close to pole-on, a strong modulation in the light curve would not be expected.
We also note a collection of peaks in the power spectrum around a frequency of $\rm 6.7\pm0.4\, \mu Hz$ (corresponding to a period of ${\sim}41.6$ hours).
The presence of this power excess is quite robust and is consistently seen when looking at random segments of the total time series and in the wavelet spectrum. We therefore rule out that it originates from random noise. First, we checked that there are no known artefacts at this frequency \citep[][]{kepdatachar}. To see if the signal might originate from another star in the vicinity of \hat{} ($K_p = 10.463$) we located all \emph{Kepler}{} targets with available data within a radius of $3\arcmin$ ($12$ were found). Of these, the star KIC~10666727 ($K_p=13.166$) was found to have strong signatures of spot modulations in its time series, and broadened power excess peaks in its power spectrum around $\rm {\sim}2.2\, \mu Hz$, $\rm {\sim}4.6\, \mu Hz$, and most interestingly in the region $\rm 6.7-7.2\, \mu Hz$. The angular separation between this target and HAT-P-7 is around $155\arcsec$, which puts the two stars close enough that direct PRF (pixel response function) contamination could occur between them \citep{ref:contamination}. However, according to \citet{ref:contamination}, for stars to contaminate over such a large separation they have to be bright.
We do not see any significant power excess in HAT-P-7 at other frequencies where KIC~10666727 shows even stronger excess than at $\rm 6.7\pm0.4\, \mu Hz$. Therefore we find it possible, but unlikely, that KIC~10666727 is contaminating the light curve of HAT-P-7.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{acc_ps_lowfilter-eps-converted-to.pdf}
\caption{\footnotesize Top: autocorrelation function (ACF) of the \hat{}\ time series (black). The dashed horizontal line gives the expectation value for random independent and identically distributed values, while the grey curves give the large-lag $95\%$ confidence levels. Bottom: low-frequency end of the power spectrum in units of period.}
\label{fig:accps}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=2\columnwidth]{wave2-eps-converted-to.pdf}
\caption{\footnotesize Morlet wavelet power spectrum as a function of time for \hat{}, where the time series was binned by 31 points. The colour goes from red at low power to blue at high power - the colours are on a logarithmic scale. The cross-hatched regions at high and low times indicate the \emph{cone of influence}, where edge effects become important \citep[see][]{1998BAMS...79...61T}. No clear signatures from rotation are apparent.}
\label{fig:wave}
\end{figure*}
\section{Discussion}
\label{sec:dis}
With an asteroseismic modelling we have provided a precise and detailed model for \hat{}, with parameters that can be used in tests of theories on the formation and evolution of planetary systems. These, together with the obliquity, are especially important to constrain when attempting to explain a system such as HAT-P-7 which likely has a close-to polar orbit of its hot-jupiter planet.
Our estimate of the obliquity of the \hat{} system supports the hypothesis described by \citet[][]{2010ApJ...718L.145W} stating that hot-jupiters are born with a large range of obliquities. Planet-planet scatterings \citep[][]{2008ApJ...686..580C}, and the effect of Kozai cycles and tidal friction \citep[][]{2007ApJ...669.1298F} could for example create large initial obliquities. For cool dwarfs with deep convection zones, tidal dissipation would operate efficiently and result in aligned systems. Hot stars fail to align because they lack such a deep convection zone for most of their life on the main-sequence, if not all together. This hypothesis was put forth based on the observed trend of $\lambda$ against $T_{\rm eff}$ where a broad distribution is seen in $\lambda$ for $T_{\rm eff}>6250\, \rm K$, while predominantly low values for $\lambda$ is seen below this temperature. This result was corroborated by \citet{2012ApJ...757...18A} using a larger sample of measurements. These authors also found that the systems with respect to $\lambda$ could be sorted in relative tidal timescales, which in turn depends on, e.\,g., the planet to star mass ratio, and $a/R_{\star}$ ($a$ being the semimajor axis). While \hat{} falls on the hot side of the dividing temperature between the two regimes, and tidal interactions are expected to be weak, it is worth noting that the temperature is quite close to this dividing line. Also, $a/R_{\star}$ is small which increases tidal interaction \citep{2012ApJ...757...18A}. The relatively slow rotation found suggests that some magnetic braking has and is taking place. The lack of alignment is likely linked to the combined effects of a low mass of the convection zone (making tides ineffective), the close-to orthogonal alignment of the angular momentum vectors, and the relatively young age of the system. Indeed, the star might have started out with a more rapid rotation, where magnetic braking from the developing convection zone has been more effective in slowing down the star than has the tidal interaction in realigning the system.
With regard to the obliquity of the system, the fact that a third, and possibly a fourth, body is found to be associated \citep{2009ApJ...703L..99W,2012PASJ...64L...7N}, could support the scenario of few-body dynamical interaction early in the life of the system, which resulted in the high obliquity of HAT-P-7b.
\section{Conclusions}
\label{sec:con}
Using asteroseismology we have estimated the stellar inclination and provided new and precise stellar parameters for \hat{} based on the extracted mode frequencies.
Mode frequencies were extracted using a Bayesian MCMC approach to peak-bag the frequency power spectrum from corrected \emph{Kepler}{} data, where we utilised the full SC dataset from Q0-Q17 available from the \emph{Kepler}{} satellite.
From this, information is obtained about stellar, planetary, and system parameters (age, mass, radius, composition, luminosity, mass of convective core) that are important ingredients in, e.\,g., dynamical simulations used to test theories for the evolution of planetary systems.
We have found the star of the \hat{} system to have a low inclination, $i_{\star}<36.5^{\circ}$ with $68\%$ credibility, meaning that the star is seen close to pole-on. Combining this with estimates for the planetary orbital inclination and the projected obliquity from RM measurements, we find that the close-in hot-jupiter planet is in a high obliquity and likely retrograde orbit (the retrograde solutions account for ${\sim}68\%$ of the PDF in \fref{fig:ob}). Our estimate for the stellar rotation matches empirical findings for stars of the same spectral type, and using the age of the system from our modelling of the oscillation frequencies in combination with an improved estimate for the colour of the star (see Appendix~\ref{app:BV}), yields that estimates from gyrochronology are not in conflict with our results. While the lack of signatures from activity are by no means proof of a low inclination for the star, they are not incompatible with a low inclination.
To our knowledge this analysis is the first wherein asteroseismology has been able to provide an estimate for $i_{\star}$ that, together with $\lambda$ from RM measurements and $i_p$ from analysis of the transit profile, has allowed for a near complete description of the system geometry.
For theories attempting to explain the formation and evolution of planetary systems, the \hat{} system is highly interesting, as any such theory must be able to offer an explanation for the system geometry. The analysis presented in this paper has shed more light on the obliquity of the system, and not just the projected obliquity which is normally used in obliquity studies. The result on the obliquity is in agreement with assumptions based on the high value of $\lambda$ and the low value for $v\sin i_{\star}$. Furthermore, the high value for $\psi$ corroborates the theory where the degree of alignment is connected to the tidal evolution of the system \citep[see, e.\,g.,][]{2010ApJ...718L.145W,2012ApJ...757...18A,2014ApJ...786..102V}.
An aspect of the \hat{} system that could be investigated to constrain the obliquity even further is the apparent asymmetry of the transit light curve, seen for instance in the phase curve presented in \citet[][]{2013ApJ...772...51E} and \citet[][]{2013ApJ...774L..19V}. Such an asymmetry is also seen in, e.\,g., the KOI-13 system, and it was found by \citet[][]{2011ApJ...736L...4S} and \citet[][]{2011ApJS..197...10B} to be in good agreement with the predictions by \citet[][]{2009ApJ...705..683B} for a planet crossing over the gravity brightened polar region of its rapidly rotating host star on a high obliquity orbit.
\begin{acknowledgements}
The authors wish to thank the entire \emph{Kepler}{} team, without whom these results would not be possible.
We are thankful to O. Benomar and his collaborators for their approach to the contemporaneousness of our respective studies.
We would like to thank Eric Stempels and Heidi Korhonen for useful discussions in the early phases of the project.
Thanks to Joshua N. Winn and John A. Johnson for making their Keck spectra available to us.
Funding for the Stellar Astrophysics Centre (SAC) is provided by The Danish National Research Foundation. The research is supported by the ASTERISK project (ASTERoseismic Investigations with SONG and \emph{Kepler}{}) funded by the European Research Council (Grant agreement no.: 267864).
M.L. would like to thank the asteroseismology group at SIfA for their hospitality during a research stay where some of this work was carried out.
M.L. also wishes to thank Niels Bohr Fondet for financial support for the research stay at SIfA.
T.L.C., G.R.D., W.J.C., and R.H. acknowledge the support of the UK Science and Technology Facilities Council (STFC).
C.K. acknowledge the support of the Villum Foundation.
M.B.N. acknowledges research funding by Deutsche Forschungsgemeinschaft (DFG) under grant SFB 963/1 ``Astrophysical flow instabilities and turbulence'' (Project A18).
The research leading to these results has received funding from the European Community's Seventh Framework Programme (FP7/2007-2013) under grant agreement no. 269194.
This research has made use of the following web resources: the SIMBAD database (simbad.u-strasbg.fr), operated at CDS, Strasbourg, France; NASAs Astrophysics Data System Bibliographic Services (adswww.harvard.edu); arxiv.org, maintained and operated by the Cornell University Library.
\end{acknowledgements}
|
\section{Introduction}
Motivated by \cite{brto_eicb}, the authors were constructing a pair of
2-adjunctions but they realized that \cite{cljs_klem} discovered them already,
in a very general case, nevertheless due to the fact of different motivations,
the examples they apply such 2-adjunctions to, are other. Among the examples given in this article, there are two that are of high importance. The first one is the equivalence between the
monoidal structures for the category of Eilenberg-Moore algebras with the
colax monad structures, and the dual case for Kleisli categories, cf
\cite{moie_motc} and \cite{zama_fotm}. The second example is the equivalence
between the right strong monads and left actions of the underlying category over the Kleisli one, cf \cite{mepa_pame}.\\
We give the structure of the article along with the list of compiled examples.\\
In Section 2, we give the formal 2-adjunction corresponding to the Kleisli situation. In Section 3, we give the formal 2-adjunction corresponding to the Eilenberg-Moore case.\\
In Section 4, we apply the case where the 2-categroy is ${}_{2}Cat$ to the
2-adjunction of EM. In Section 5, we proved the theorem of I. Moerdijk on the equivalence for the monoidal structures induced on the category of algebras and colax monads.\\
In Section 6, we use the Kleisli 2-adjunction for the case ${}_{2}Cat$. In
Section 7, we apply this 2-adjunction to induced a monoidal structure on the Kleisli category and relate it with lax monads.\\
In Section 8, we apply the adjunction to the most known case of liftings of functors and commutative diagrams for the forgetful functor, check \cite{bofr_haca_II} and \cite{tami_psdl}, just to mention a couple of references.\\
In Section 9, we relate actions of the category $\mathcal{C}$ over its Kleisli category $\mathcal{C}_{\scriptscriptstyle{F}}$ with strong monads.\\
In Section 10, we finalize with left and right functor algebras for a monad and relate this to certain liftings and extensions, respectively, for the underlying functors, cf. \cite{duej_kaee}.\\
We give some remarks on notation. Suppose that we had an adjunction of the
form $\mathfrak{L}\dashv\mathfrak{R}$, then the unit and counit for this
adjunction will be denoted as $\eta^{\mathfrak{R}\mathfrak{L}}$ and
$\varepsilon^{\mathfrak{L}\mathfrak{R}}$, respectively. We are completely
aware that this notation is very cumbersome, nonetheless, it is clear and in
the case of the proliferation of several adjunctions it has the property of
avoiding the need for extra greek letters to denote the new units and
counits. As the article develops, the reader will see the advantage in using this notation.\\
We will be working with monoidal categories denoted as $(\mathcal{C}, \otimes, I, a, l, r)$ and also as $(\mathcal{C}, \otimes, I)$ , as a contraction, that leaves understood the natural constrain transformations. We will be working with the constant functor $\delta_{I}:\mathbf{1}\longrightarrow \mathcal{C}$, on $I$, where $\mathbf{1}$ is the category with only one object $0$ and only one arrow $1_{0}$. That is to say, $\delta_{I}(0) = I$.\\
On the other hand, it is known that a category with binary products and a terminal object has a canonical (cartesian) monoidal structure. This is the case for the category $Cat$, of small categories. The natural constraint transformations, taken on components, are functors, for example, for $\mathcal{C}, \mathcal{D}, \mathcal{E}$, $a_{\mathcal{C},\mathcal{D},\mathcal{E}}:(\mathcal{C}\times\mathcal{D})\times\mathcal{E}\longrightarrow \mathcal{C}\times(\mathcal{D}\times\mathcal{E})$ is the obvious functor. In order to compact the notation, we will agree that in the case that the component be the object $\mathcal{C}, \mathcal{C}, \mathcal{C}$, the asociativity functor will be denoted simply as $a_{\scriptscriptstyle{\mathcal{C}}}$. In turn, the respective constraint functors will be denoted as $l_{\scriptscriptstyle{\mathcal{C}}}$ and $r_{\scriptscriptstyle{\mathcal{C}}}$. \\
The horizontal composition in a general 2-category $\mathcal{A}$ will be denoted as $\cdot$ or by juxtaposition, this notation will be used indistinctively. The vertical composition on 2-cells will be given the symbol $\circ$.
\section{Formal Kleisli 2-Adjunction}
Consider a 2-category such that $\mathcal{A}^{op}$ admits the construction of algebras. Due to this property of the 2-category $\mathcal{A}^{op}$, we will be able to construct a 2-adjunction of the form
\begin{equation}\label{1405261543}
\xymatrix@C=1.6cm{
\mathbf{Mnd}(\mathcal{A}^{op}) \ar@<-3pt>[r]_{\Psi_{\scriptscriptstyle{K}}} & \mathbf{Adj}_{\scriptscriptstyle{R}}(\mathcal{A}^{op})\ar@<-3pt>[l]_{\Phi_{\scriptscriptstyle{K}}}
}
\end{equation}\\
If we describe the adjunction over $\mathcal{A}$ and not on the opposite one
then the 2-category $\mathbf{Mnd}(\mathcal{A}^{op})$ will be isomorphic to
$\mathbf{Mnd}^{\scriptscriptstyle{\bullet}}(\mathcal{A})$ and the 2-category
$\mathbf{Adj}_{\scriptscriptstyle{R}}(\mathcal{A}^{op})$ will be isomorphic to $\mathbf{Adj}_{\scriptscriptstyle{L}}(\mathcal{A})$. Note that in \cite{stro_fotm}, the category $\mathcal{A}^{op}$ is denoted as $\mathcal{A}^{\ast}$. \\
The description of the 2-category $\mathbf{Mnd}^{\scriptscriptstyle{\bullet}}(\mathcal{A})$ is given as follows.
\begin{enumerate}
\item [1.-] The 0-cells are monads in $\mathcal{A}$, \emph{i.e.} $(A, f,\mu^{\scriptscriptstyle{f}}, \eta^{\scriptscriptstyle{f}})$. The short notation $(A,f)$ will be used for such a monad.
\item [2.-] The 1-cells, which we call indistinctively as morphisms of monads, are pairs of the form $(m, \pi):(A, f)\longrightarrow (B, h)$; where $m:A\longrightarrow B$ is a 1-cell in $\mathcal{A}$ and $\pi:mf\longrightarrow hm$ is a 2-cell in $\mathcal{A}$ such that the following diagrams commute
\begin{equation*}
\begin{array}{ccc}
\xymatrix@C=1.5cm@R=1.5cm{
mff \ar[r]^-{\pi f}\ar[d]_{m\mu^{\scriptscriptstyle{f}}} & hmf\ar[r]^-{h \pi } &
hhm\ar[d]^{\mu^{\scriptscriptstyle{h}} m}\\
mf\ar[rr]_{\pi} & & hm
} &
\xy<1cm,0cm>
\POS (0,-20.5) *+{,},
\endxy
&
\xy<1cm,0cm>
\POS (0, -6)
\xymatrix@C=1cm@R=0.8cm{
& m\ar[ld]_{m\eta^{f}}\ar[rd]^{\eta^{h} m}& \\
mf\ar[rr]_{\pi} & & hm
}
\endxy
\end{array}
\end{equation*}
\item [3.-] The 2-cells, which we call indistinctively as transformations of monads, have the form $\vartheta:(m, \pi)\longrightarrow (n, \tau):(A, f)\longrightarrow (B, h)$, such that $\vartheta:m\longrightarrow n:A\longrightarrow B$ is a 2-cell in $\mathcal{A}$ and the following diagram commutes
\begin{equation*}\label{1405261627}
\xy<1cm,0cm>
\POS (0, 0) *+{mf} = "c1",
\POS (20, 0) *+{hm} = "c2",
\POS (0, -20) *+{nf} = "d1",
\POS (20, -20) *+{hn} = "d2",
\POS (18, -10) = "a",
\POS (10, -18) = "b",
\POS "c1" \ar^{\pi} "c2",
\POS "d1" \ar_{\tau} "d2",
\POS "c1" \ar_{\vartheta f} "d1",
\POS "c2" \ar^{h \vartheta } "d2",
\endxy
\end{equation*}
This 2-cell is displayed as follows
\begin{equation*}
\xy<1cm,0cm>
\POS (0, 0) *+{(A, f)} = "c1",
\POS (32, 0) *+{(B, h)} = "c2",
\POS (16,0) *+{\vartheta},
\POS (10, 3.5) = "a",
\POS (10, -3.5) = "b",
\POS "c1" \ar@/^1.5pc/^{(m,\: \pi)} "c2",
\POS "c1" \ar@/_1.5pc/_{(n,\:\tau)} "c2",
\POS "a" \ar "b",
\endxy
\end{equation*}
\end{enumerate}
The structure of the 2-category $\mathbf{Adj}_{\scriptscriptstyle{L}}(\mathcal{A})$ is given as follows
\begin{enumerate}
\item [1.-] The $0$-cells are made of adjunctions
\begin{equation*}\label{1407051035}
\begin{array}{cc}
\xymatrix@C=1.6cm{
A \ar@<-3pt>[r]_{l} & B\ar@<-3pt>[l]_{r}
} &\!\!\! .
\end{array}
\end{equation*}
\item [2.-] The $1$-cells are pairs of the form $(j, k)$ such that the second diagram is the 2-cell mate to the first one
\begin{equation*}\label{1407051036}
\begin{array}{ccc}
\xy<1cm,0cm>
\POS (0, 0) *+{A} = "c1",
\POS (20, 0) *+{\overline{A}} = "c2",
\POS (0, -20) *+{B} = "d1",
\POS (20, -20) *+{\overline{B}} = "d2",
\POS (18, -10) = "a",
\POS (10, -18) = "b",
\POS "c1" \ar^{j} "c2",
\POS "c1" \ar_{l} "d1",
\POS "c2" \ar^{\overline{l}} "d2",
\POS "d1" \ar_{k} "d2",
\endxy &
\xy<1cm,0cm>
\POS (0,-22) *+{,},
\endxy
&
\xy<1cm,0cm>
\POS (0, 0) *+{A} = "c1",
\POS (20, 0) *+{\overline{A}} = "c2",
\POS (0, -20) *+{B} = "d1",
\POS (20, -20) *+{\overline{B}} = "d2",
\POS (10, -2) = "b",
\POS (18, -10) = "a",
\POS "c1" \ar^{j} "c2",
\POS "d1" \ar^{r} "c1",
\POS "d2" \ar_{\overline{r}} "c2",
\POS "d1" \ar_{k} "d2",
\POS "b" \ar@/_.5pc/_{\rho}"a",
\endxy
\end{array}
\end{equation*}
The mate is described by
\begin{equation}\label{1407052237}
\rho = \overline{r}k\varepsilon\circ\overline{\eta}jr
\end{equation}
This morphism can be represented as
\begin{equation*}\label{1407052238}
\xy<1cm,0cm>
\POS (0, 0) *+{A} = "c1",
\POS (20, 0) *+{\overline{A}} = "c2",
\POS (0, -20) *+{B} = "d1",
\POS (20, -20) *+{\overline{B}} = "d2",
\POS "c1" \ar^{j} "c2",
\POS "c1" \ar@<-2pt>_{l} "d1",
\POS "d1" \ar@<-2pt>_{r} "c1",
\POS "c2" \ar@<-2pt>_{\overline{l}} "d2",
\POS "d2" \ar@<-2pt>_{\overline{r}} "c2",
\POS "d1" \ar_{k} "d2",
\POS (10, -2) = "b",
\POS (16, -9) = "a",
\POS "b" \ar@/_.4pc/_{\rho}"a",
\endxy
\end{equation*}
\noindent and denoted as $(j,k,\rho):l\dashv r\longrightarrow \overline{l}\dashv \overline{r}$.
\item [3.-] The $2$-cells are made of a pair of 2-cells in $\mathcal{A}$, $(\alpha, \beta)$ as in
\begin{equation*}\label{1405141814}
\xy<1cm,0cm>
\POS (0, 0) *+{A} = "c1",
\POS (25, 0) *+{\overline{A}} = "c2",
\POS (0, -25) *+{B} = "d1",
\POS (25, -25) *+{\overline{B}} = "d2",
\POS (12.5, 0) *+{\alpha} = "e1",
\POS (12.5, -25) *+{\beta} = "e2",
\POS (7, 3) = "a",
\POS (7, -2.5) = "b",
\POS (7,-22) = "ap",
\POS (7, -27.5) = "bp",
\POS "c1" \ar@/^1.2pc/^{j} "c2",
\POS "c1" \ar@/_1.2pc/_{j'} "c2",
\POS "c1" \ar@<-2pt>_{l} "d1",
\POS "d1" \ar@<-2pt>_{r} "c1",
\POS "c2" \ar@<-2pt>_{\overline{l}} "d2",
\POS "d2" \ar@<-2pt>_{\overline{r}} "c2",
\POS "d1" \ar@/^1.2pc/^{k} "d2",
\POS "d1" \ar@/_1.2pc/_{k'} "d2",
\POS "a" \ar "b",
\POS "ap" \ar "bp",
\endxy
\end{equation*}
\noindent such that they fulfill one of the following equivalent conditions
\begin{enumerate}
\item [(i)] $\overline{l}\alpha = \beta l$,
\item [(ii)] $\rho'\circ\alpha r = \overline{r}\beta\circ\rho$.\\
\end{enumerate}
\newtheorem{1407060859}{Remark}[section]
\begin{1407060859}
Note that the previous conditions can be seen as commutative surface diagrams.
\end{1407060859}
This $2$-cell can be displayed as follows
\begin{equation*}
\xy<1cm,0cm>
\POS (0, 0) *+{l\dashv r} = "c1",
\POS (32, 0) *+{\overline{l}\! \dashv \! \overline{r}} = "c2",
\POS (16,0) *+{\scriptstyle{(}\scriptstyle{\alpha},\scriptstyle{\beta})},
\POS (10, 3.5) = "a",
\POS (10, -3.5) = "b",
\POS "c1" \ar@/^1.5pc/^{(j, k, \rho)} "c2",
\POS "c1" \ar@/_1.5pc/_{(j', k',\rho')} "c2",
\POS "a" \ar "b",
\endxy
\end{equation*}
The $n$-cell structure described arrange itself to form a $2$-category.
\end{enumerate}
Before going into the details on the construction of the 2-functor $\Psi_{\scriptscriptstyle{K}}$, we develop some
calculations. These calculations are dual to those made in
\cite{stro_fotm}. Note that we are going to be switching between the
2-categories $\mathcal{A}^{op}$ and $\mathbf{Mnd}(\mathcal{A}^{op})$ to $\mathcal{A}$
and $\mathbf{Mnd}^{\scriptscriptstyle{\bullet}}(\mathcal{A})$, respectively. \\
Since the 2-category $\mathcal{A}^{op}$ admits the construction of algebras,
the functor $\mathnormal{Inc}_{\mathcal{A}^{op}}:\mathcal{A}^{op}\longrightarrow \mathbf{Mnd}(\mathcal{A}^{op})$ admits a right adjoint, denoted as
$\mathnormal{Alg}_{\mathcal{A}^{op}}:\mathbf{Mnd}(\mathcal{A}^{op})\longrightarrow
\mathcal{A}^{op}$. These 2-functors are going to be short denoted as
$\mathnormal{I}^{\scriptscriptstyle{\bullet}}$ and $\mathnormal{A}^{\scriptscriptstyle{\bullet}}$
respectively. \\
The corresponding counit, on the component $(A, f^{op})$, is $\varepsilon^{I\! A^{\scriptscriptstyle{\bullet}}}(A, f^{op}):\mathnormal{Inc}_{\mathcal{A}^{op}}\mathnormal{Alg}_{\mathcal{A}^{op}}(A, f^{op})\longrightarrow (A, f^{op})$. If we define $\mathnormal{Alg}_{\mathcal{A}^{op}}(A, f^{op}) = A_{\scriptscriptstyle{f}}$ then $\varepsilon^{I\! A^{\scriptscriptstyle{\bullet}}}(A,f^{op}) = (g_{\scriptscriptstyle{f}}, \iota_{\scriptscriptstyle{f}}): (A, f)\longrightarrow (A_{\scriptscriptstyle{f}},1_{A_{\scriptscriptstyle{f}}})$. This last 1-cell belongs to $\mathbf{Mnd}^{\scriptscriptstyle{\bullet}}(\mathcal{A})$, where $g_{\scriptscriptstyle{f}}:A\longrightarrow A_{\scriptscriptstyle{f}}$ and $\iota_{\scriptscriptstyle{f}}:g_{\scriptscriptstyle{f}}v_{\scriptscriptstyle{f}}g_{\scriptscriptstyle{f}}\longrightarrow g_{\scriptscriptstyle{f}}$.\\
Following \cite{stro_fotm}, for any monad $(A, f^{op})$ in $\mathbf{Mnd}(\mathcal{A}^{op})$, there exists an adjunction in $\mathcal{A}$,
\begin{equation*}\label{1406142138}
\begin{array}{cc}
\xymatrix@C=1.6cm{
A \ar@<-3pt>[r]_{g_{\scriptscriptstyle{f}}} & A_{\scriptscriptstyle{f}}\ar@<-3pt>[l]_{v_{\scriptscriptstyle{f}}}
} &\!\!\! ,
\end{array}
\end{equation*}\\
\noindent such that it generates the monad $(A, f)$, with unit $\eta^{\scriptscriptstyle{f}}$ and counit $\varepsilon^{\phantom{.}\! gv_{\!\scriptscriptstyle{f}}}$. It can be checked that $\iota_{\scriptscriptstyle{f}} = \varepsilon^{\phantom{.}\! gv_{\!\scriptscriptstyle{f}}}g_{\scriptscriptstyle{f}}$.\\
Suppose that there is a morphism of monads $(m^{op}, \pi):(B, h^{op})\longrightarrow (A, f^{op})$ in $\mathbf{Mnd}(\mathcal{A}^{op})$, \emph{i.e.} $(m,\pi):(A,f)\longrightarrow (B,h)$ in $\mathbf{Mnd}^{\scriptscriptstyle{\bullet}}(\mathcal{A})$. Take the following composition of morphisms of monads $(g_{\scriptscriptstyle{h}},\iota_{\scriptscriptstyle{h}})\cdot(m,\pi)= (g_{\scriptscriptstyle{h}}m, \iota_{\scriptscriptstyle{h}}m\circ g_{\scriptscriptstyle{h}}\pi):(A,f)\longrightarrow (B_{\scriptscriptstyle{h}}, 1_{B_{\scriptscriptstyle{h}}})$.\\
Since the counit is universal from $\mathnormal{Inc}_{\mathcal{A}^{op}}$ to $(A,f^{op})$, there exists a 1-cell $m_{\pi}:A_{\scriptscriptstyle{f}}\longrightarrow B_{\scriptscriptstyle{h}}$, in $\mathcal{A}$, such that the following diagram commute\\
\begin{equation*}\label{1406021518}
\xy<1cm,0cm>
\POS (25,15) *+{(A, f)} = "a12"
\POS (0,0) *+{(A_{\scriptscriptstyle{f}}, 1_{A_{\scriptscriptstyle{f}}})} = "a21",
\POS (50,0) *+{(B_{\scriptscriptstyle{h}}, 1_{B_{\scriptscriptstyle{h}}})} = "a23"
\POS "a12" \ar_{(g_{\scriptscriptstyle{f}},\iota_{\scriptscriptstyle{f}})} "a21",
\POS "a12" \ar^{\phantom{12}(g_{\scriptscriptstyle{h}}m, \phantom{.}\!\iota_{\scriptscriptstyle{h}}m\phantom{.}\!\circ\phantom{.}\! g_{\scriptscriptstyle{h}}\pi)} "a23",
\POS "a21" \ar_{(m_{\pi}, 1_{m_{\pi}})} "a23",
\endxy
\end{equation*}\\
In particular, $g_{\scriptscriptstyle{h}}m = m_{\pi}g_{\scriptscriptstyle{f}}$ and $\iota_{\scriptscriptstyle{h}}m\circ g_{\scriptscriptstyle{h}}\pi = m_{\pi}\iota_{\scriptscriptstyle{f}}$. Note that the associated mate to the first equality is $\rho_{\pi} = v_{\scriptscriptstyle{h}}m_{\pi}\varepsilon^{g v_{\scriptscriptstyle{h}}}\circ\eta^{\scriptscriptstyle{h}}m v_{\scriptscriptstyle{f}}$ and that $\rho_{\pi} g_{\scriptscriptstyle{f}} = \pi$.\\
Consider a 2-cell of monads $\vartheta: (m,\pi)\longrightarrow (n,\tau):(A,f)\longrightarrow (B,h)$ in $\mathbf{Mnd}^{\scriptscriptstyle{\bullet}}(\mathcal{A})$. Because of the construction of algebras for $\mathcal{A}^{op}$, the 2-adjunction $\mathnormal{Alg}_{\mathcal{A}^{op}}\dashv \mathnormal{Inc}_{\mathcal{A}^{op}}$ provides an isomorphism of categories, for $(A,f^{op})$ in $\mathbf{Mnd}(\mathcal{A}^{op})$ and $B$ in $\mathcal{A}^{op}$, of
the form
\begin{equation*}\label{1406191447}
Hom_{\mathbf{Mnd}(\mathcal{A}^{op})}\big((A,f^{op}), \mathnormal{Inc}_{\mathcal{A}^{op}}(B)\big)
\cong Hom_{\mathcal{A}^{op}}\big(\mathnormal{Alg}_{\mathcal{A}^{op}}(A,f^{op}), B\big)
\end{equation*}\\
\noindent this translates, in the non-opposite case, into the following assigment
\begin{equation}\label{1406191526}
\begin{array}{ccc}
\xy<1cm,0cm>
\POS (0,0) *+{A_{\scriptscriptstyle{f}}} = "a11",
\POS (25,0) *+{B} = "a12",
\POS (9, 3.5) = "a",
\POS (9,-3.5) = "b"
\POS "a11" \ar@/^1.3pc/^{a} "a12",
\POS "a11" \ar@/_1.3pc/_{b} "a12",
\POS "a" \ar^{\phantom{1}\alpha} "b",
\endxy & \longmapsto &
\xy<1cm,0cm>
\POS (0,0) *+{(A,f)} = "a11",
\POS (30,0) *+{(B, 1_{B})} = "a13",
\POS (9, 3.5) = "a",
\POS (9,-3.5) = "b",
\POS (15, -0.4) *+{\scriptstyle{\alpha\phantom{.}\! g_{\!\scriptscriptstyle{f}}}} = "a12",
\POS "a11" \ar@/^1.5pc/^{(ag_{\!\scriptscriptstyle{f}},\phantom{.}\! a\iota_{\!\scriptscriptstyle{f}})} "a13",
\POS "a11" \ar@/_1.5pc/_{(b\phantom{.}\! g_{\!\scriptscriptstyle{f}}, \phantom{.}\! b\phantom{.}\!\iota_{\!\scriptscriptstyle{f}})} "a13",
\POS "a" \ar "b",
\endxy
\end{array}
\end{equation}\\
\noindent On the other hand, we have an equality of
2-cells\\
\begin{equation*}\label{1406191527}
\begin{array}{ccc}
\xy<1cm,0cm>
\POS (0,0) *+{(A,f)} = "a11",
\POS (30,0) *+{\phantom{123}(B_{\scriptscriptstyle{h}}, 1_{B_{\scriptscriptstyle{h}}})} = "a13",
\POS (10, 3.5) = "a",
\POS (10,-3.5) = "b",
\POS (15, 0.5) *+{\scriptstyle{g_{\scriptscriptstyle{h}}\vartheta}} = "a12",
\POS "a11" \ar@/^1.5pc/^{(g_{\scriptscriptstyle{h}}m, \phantom{.}\! \iota_{\scriptscriptstyle{h}}m\circ g_{\scriptscriptstyle{h}}\pi)} "a13",
\POS "a11" \ar@/_1.5pc/_{(g_{\scriptscriptstyle{h}}n, \phantom{.}\! \iota_{\scriptscriptstyle{h}} n\circ g_{\scriptscriptstyle{h}}\tau)} "a13",
\POS "a" \ar "b",
\endxy & = &
\xy<1cm,0cm>
\POS (0,0) *+{(A, f)} = "a11",
\POS (30,0) *+{\phantom{123}(B_{\scriptscriptstyle{h}}, 1_{B_{\scriptscriptstyle{h}}})} = "a13",
\POS (11, 3.5) = "a",
\POS (11,-3.5) = "b",
\POS (16, 0.5) *+{\scriptstyle{g_{\scriptscriptstyle{h}}\vartheta}} = "a12",
\POS "a11" \ar@/^1.5pc/^{(m_{\pi}g_{\!\scriptscriptstyle{f}} , \phantom{.}\! m_{\pi}\iota_{\!\scriptscriptstyle{f}})} "a13",
\POS "a11" \ar@/_1.5pc/_{(n_{\tau}g_{\!\scriptscriptstyle{f}} , \phantom{.}\! n_{\tau}\iota_{\!\scriptscriptstyle{f}})} "a13",
\POS "a" \ar "b",
\endxy
\end{array}
\end{equation*}\\
Therefore, to the 2-cell $g_{\scriptscriptstyle{h}}\vartheta$ there corresponds, through the
asignment \eqref{1406191526}, a 2-cell $\beta_{\vartheta} =
\mathnormal{Alg}_{\mathcal{A}^{op}}(g_{\scriptscriptstyle{h}}\vartheta)\cdot
\eta^{IA^{\scriptscriptstyle{\bullet}}}(B_{\scriptscriptstyle{h}})$, such that $g_{\scriptscriptstyle{h}}\vartheta =
\beta_{\vartheta}g_{\scriptscriptstyle{f}}$, where $\vartheta:m_{\pi}\rightarrow n_{\tau}$. We
change, at this point, the notation as $\beta_{\vartheta} = \widetilde{\vartheta}$.\\
Without any further ado, we provide the description of the 2-functor $\Psi_{\scriptscriptstyle{K}}$ \\
\begin{enumerate}
\item [1.-] For the monad $(A, f, \mu^{\scriptscriptstyle{f}}, \eta^{\scriptscriptstyle{f}})$ in $\mathbf{Mnd}^{\scriptscriptstyle{\bullet}}(\mathcal{A})$, $\Psi_{\scriptscriptstyle{K}}(A, f) = g_{\scriptscriptstyle{f}}\dashv v_{\scriptscriptstyle{f}}$.
\item [2.-] For the morphism $(m, \pi):(A,f)\longrightarrow (B, h)$, $\Psi_{\scriptscriptstyle{K}}(m, \pi) = (m, m_{\pi}, \rho_{\pi})$
\item [3.-] For the transformation $\vartheta: (m,\pi)\longrightarrow (n,\tau):(A,f)\longrightarrow(B,g)$, $\Psi_{\scriptscriptstyle{K}}(\vartheta) =(\vartheta, \widetilde{\vartheta})$, where $\widetilde{\vartheta}$ is given as above.
\end{enumerate}
\noindent The description of the functor $\Phi_{\scriptscriptstyle{K}}$ is given as follows
\begin{enumerate}
\item [1.-] For the adjunction $l\dashv r$, $\Phi_{\scriptscriptstyle{K}}(l\dashv r) = (A, rl)$
\item [2.-] For the morphism of adjunctions $(j,k,\rho):(l\dashv r)\longrightarrow(\overline{l}\dashv \overline{r})$, $\Phi_{\scriptscriptstyle{K}}(j,k,\rho) = (j,\pi_{\rho})$. Where $\pi_{\rho} = \rho\phantom{.}\! l$.
\item [3.-] For the transformation of adjunctions $(\alpha, \beta):(j,k,\rho)\longrightarrow (j',k',\rho'):l\dashv r\longrightarrow \overline{l}\dashv\overline{r}$, $\Phi_{\scriptscriptstyle{K}}(\alpha, \beta) = \vartheta_{(\alpha,\beta)} = \alpha$.
\end{enumerate}
Yet again, following \cite{stro_fotm}, it can be shown that for the adjunction $l\dashv r$, there exists a \emph{dual comparison 1-cell} $k_{rl}:A_{rl}\longrightarrow B$, such that $l = k_{rl}g_{rl}$, $v_{rl} = rk_{rl}$ and $\varepsilon^{rl} l = k_{rl}\iota_{rl}$.\\
The unit of the 2-adjunction in \eqref{1405261543}, $\eta^{\Phi\Psi_{\scriptscriptstyle{K}}}:1_{\mathbf{Mnd}^{\scriptscriptstyle{\bullet}}(\mathcal{A})}\longrightarrow\Phi_{\scriptscriptstyle{K}}\Psi_{\scriptscriptstyle{K}}$ is defined, in the component $(A, f)$, as follows
\begin{equation*}\label{1406281141}
\eta^{\Phi\Psi_{\scriptscriptstyle{K}}}(A, f) := (1_{A}, 1_{\scriptscriptstyle{f}}):(A, f)\longrightarrow (A,f)
\quad in \quad \mathbf{Mnd}^{\scriptscriptstyle{\bullet}}(\mathcal{A})
\end{equation*}\\
In turn, the counit $\varepsilon^{\Psi\Phi_{\scriptscriptstyle{K}}}:\Psi_{\scriptscriptstyle{K}}\Phi_{\scriptscriptstyle{K}}\longrightarrow 1_{\mathbf{Adj}_{l}(\mathcal{A})}$ is defined, in the component $l\dashv r$, as follows
\begin{equation*}\label{1406282009}
\varepsilon^{\Psi\Phi_{\scriptscriptstyle{K}}}(l\dashv r) := (1_{A}, k_{rl}, 1_{v_{rl}}):g_{rl}\dashv v_{rl}\longrightarrow l\dashv r
\quad in \quad \mathbf{Adj}_{\scriptscriptstyle{L}}(\mathcal{A})
\end{equation*}\\
\newtheorem{1406031238}{Theorem}[section]
\begin{1406031238}
There exists a 2-adjunction $\Psi_{\scriptscriptstyle{K}}\dashv \Phi_{\scriptscriptstyle{K}}$.
\end{1406031238}
\noindent \emph{Proof}:\\
We prove only one of the triangular identities, \emph{i.e.} $\Phi_{\scriptscriptstyle{K}}\phantom{.}\!\varepsilon^{\Psi\Phi_{\scriptscriptstyle{K}}}\circ\eta^{\Phi\Psi_{\scriptscriptstyle{K}}}\Phi_{\scriptscriptstyle{K}} = 1_{\Phi_{\scriptscriptstyle{K}}}$.\\
\begin{eqnarray*}
\big(\Phi_{\scriptscriptstyle{K}}\phantom{.}\!\varepsilon^{\Psi\Phi_{\scriptscriptstyle{K}}}\circ\eta^{\Phi\Psi_{\scriptscriptstyle{K}}}\Phi_{\scriptscriptstyle{K}}\big)(l\dashv r) &=& \Phi_{\scriptscriptstyle{K}}\phantom{.}\!\varepsilon^{\Psi\Phi_{\scriptscriptstyle{K}}}(l\dashv r)\cdot\eta^{\Phi\Psi_{\scriptscriptstyle{K}}}\Phi_{\scriptscriptstyle{K}}(l\dashv r) \\
&=& \Phi_{\scriptscriptstyle{K}}(1_{A}, k_{rl}, 1_{v_{rl}}) \cdot \eta^{\Phi\Psi_{\scriptscriptstyle{K}}}(A, rl)\\
&=& (1_{A}, 1_{v_{rl}}g_{rl})\cdot (1_{A}, 1_{rl}) = (1_{A}, 1_{rl}) = 1_{(A, rl)}\\
&=& 1_{\Phi_{\scriptscriptstyle{K}}(l\dashv\phantom{.}\! r)} = 1_{\Phi_{\scriptscriptstyle{K}}}(l\dashv r).
\end{eqnarray*}
$\square$
\section{Formal Eilenberg-Moore 2-Adjunction}
Consider a $2$-category $\mathcal{A}$ which admits the construction of algebras. With this property of $\mathcal{A}$, we will construct a 2-adjunction of the form
\begin{equation*}\label{1405151703}
\begin{array}{cc}
\xymatrix@C=1.6cm{
\mathbf{Adj}_{\scriptscriptstyle{R}}(\mathcal{A}) \ar@<-3pt>[r]_{\Phi_{\scriptscriptstyle{E}}} & \mathbf{Mnd}(\mathcal{A})\ar@<-3pt>[l]_{\Psi_{\!\scriptscriptstyle{E}}}
} &\!\!\! ,
\end{array}
\end{equation*}\\
\noindent The 2-category $\mathbf{Adj}_{\scriptscriptstyle{R}}(\mathcal{A})$ is described as follows
\begin{enumerate}
\item [1.-] The $0$-cells are made of adjunctions
\begin{equation*}\label{1404151713}
\begin{array}{cc}
\xymatrix@C=1.6cm{
A \ar@<-3pt>[r]_{l} & B\ar@<-3pt>[l]_{r}
} &\!\!\! .
\end{array}
\end{equation*}
\item [2.-] The $1$-cells are pairs, of 1-cells in $\mathcal{A}$, $(j, k)$ such that the first diagram is the 2-cell mate to the second one
\begin{equation*}\label{1405141753}
\begin{array}{ccc}
\xy<1cm,0cm>
\POS (0, 0) *+{A} = "c1",
\POS (20, 0) *+{\overline{A}} = "c2",
\POS (0, -20) *+{B} = "d1",
\POS (20, -20) *+{\overline{B}} = "d2",
\POS (18, -10) = "a",
\POS (10, -18) = "b",
\POS "c1" \ar^{j} "c2",
\POS "c1" \ar_{l} "d1",
\POS "c2" \ar^{\overline{l}} "d2",
\POS "d1" \ar_{k} "d2",
\POS "a" \ar@/_.5pc/_{\lambda}"b",
\endxy &
\xy<1cm,0cm>
\POS (0,-22) *+{,},
\endxy
&
\xy<1cm,0cm>
\POS (0, 0) *+{A} = "c1",
\POS (20, 0) *+{\overline{A}} = "c2",
\POS (0, -20) *+{B} = "d1",
\POS (20, -20) *+{\overline{B}} = "d2",
\POS "c1" \ar^{j} "c2",
\POS "d1" \ar^{r} "c1",
\POS "d2" \ar_{\overline{r}} "c2",
\POS "d1" \ar_{k} "d2",
\endxy
\end{array}
\end{equation*}
The mate is described by
\begin{equation}\label{1405151810}
\lambda = \overline{\varepsilon}kl\circ \overline{l}j\eta
\end{equation}
This morphism can be represented as
\begin{equation*}\label{1403151811}
\xy<1cm,0cm>
\POS (0, 0) *+{A} = "c1",
\POS (20, 0) *+{\overline{A}} = "c2",
\POS (0, -20) *+{B} = "d1",
\POS (20, -20) *+{\overline{B}} = "d2",
\POS (17, -13) = "a",
\POS (11, -19) = "b",
\POS (17,-13) = "ap",
\POS (11, -19) = "bp",
\POS "c1" \ar^{j} "c2",
\POS "c1" \ar@<-2pt>_{l} "d1",
\POS "d1" \ar@<-2pt>_{r} "c1",
\POS "c2" \ar@<-2pt>_{\overline{l}} "d2",
\POS "d2" \ar@<-2pt>_{\overline{r}} "c2",
\POS "d1" \ar_{k} "d2",
\POS "ap" \ar@/_.25pc/_{\lambda}"bp",
\endxy
\end{equation*}
\noindent and denoted as $(j,k,\lambda):l\dashv r\longrightarrow \overline{l}\dashv \overline{r}$.
\item [3.-] The $2$-cells are made of a pair of 2-cells in $\mathcal{A}$, $(\alpha, \beta)$ as in
\begin{equation*}\label{1407190919}
\xy<1cm,0cm>
\POS (0, 0) *+{A} = "c1",
\POS (25, 0) *+{\overline{A}} = "c2",
\POS (0, -25) *+{B} = "d1",
\POS (25, -25) *+{\overline{B}} = "d2",
\POS (12.5, 0) *+{\alpha} = "e1",
\POS (12.5, -25) *+{\beta} = "e2",
\POS (7, 3) = "a",
\POS (7, -2.5) = "b",
\POS (7,-22) = "ap",
\POS (7, -27.5) = "bp",
\POS "c1" \ar@/^1.2pc/^{j} "c2",
\POS "c1" \ar@/_1.2pc/_{j'} "c2",
\POS "c1" \ar@<-2pt>_{l} "d1",
\POS "d1" \ar@<-2pt>_{r} "c1",
\POS "c2" \ar@<-2pt>_{\overline{l}} "d2",
\POS "d2" \ar@<-2pt>_{\overline{r}} "c2",
\POS "d1" \ar@/^1.2pc/^{k} "d2",
\POS "d1" \ar@/_1.2pc/_{k'} "d2",
\POS "a" \ar "b",
\POS "ap" \ar "bp",
\endxy
\end{equation*}
\noindent such that they fulfill one of the following equivalent conditions
\begin{enumerate}
\item [(i)] $\lambda'\circ \overline{l} \alpha = \beta l \circ\lambda$,
\item [(ii)] $\alpha r = \overline{r} \beta$.
\end{enumerate}
\newtheorem{1407152212}{Remark}[section]
\begin{1407152212}
Note that the previous conditions can be seen as commutative surface diagrams.
\end{1407152212}
This $2$-cell can be displayed as follows
\begin{equation*}
\xy<1cm,0cm>
\POS (0, 0) *+{l\dashv r} = "c1",
\POS (32, 0) *+{\overline{l}\! \dashv \! \overline{r}} = "c2",
\POS (16,0) *+{\scriptstyle{(}\scriptstyle{\alpha},\scriptstyle{\beta})},
\POS (10, 3.5) = "a",
\POS (10, -3.5) = "b",
\POS "c1" \ar@/^1.5pc/^{(j, k, \lambda)} "c2",
\POS "c1" \ar@/_1.5pc/_{(j',\phantom{.}\! k',\phantom{.}\!\lambda')} "c2",
\POS "a" \ar "b",
\endxy
\end{equation*}
The described $n$-cell structure arrange itself to form a $2$-category.
\end{enumerate}
\noindent The 2-category $\mathbf{Mnd}(\mathcal{A})$ is formed as follows
\begin{enumerate}
\item [1.-] The 0-cells are monads in $\mathcal{A}$, $(A,f,\mu^{\scriptscriptstyle{f}},\eta^{\scriptscriptstyle{f}})$. The short notation $(A, f)$ will be used for such a monad.
\item [2.-] The 1-cells are \emph{morphims of monads} $(p,\varphi):(A, f)\longrightarrow (B, h)$ which consist of a 1-cell $p:A\longrightarrow B$ and a 2-cell $\varphi:hp\longrightarrow pf$ such that the following diagrams commutes
\begin{equation*}\label{1405151917}
\begin{array}{ccc}
\xymatrix@C=1.5cm@R=1.5cm{
hhp \ar[r]^-{h\varphi }\ar[d]_{\mu^{h}\! p} & hpf\ar[r]^-{\varphi f} & pff\ar[d]^{p\mu^{f}}\\
hp\ar[rr]_{\varphi} & & pf
} &
\xy<1cm,0cm>
\POS (0,-20.5) *+{,},
\endxy
&
\xy<1cm,0cm>
\POS (0, -7)
\xymatrix@C=1cm@R=0.8cm{
& p\ar[ld]_{\eta^{h}p}\ar[rd]^{p\eta^{f}}& \\
hp\ar[rr]_{\varphi} & & pf
}
\endxy
\end{array}
\end{equation*}
\item [3.-] The 2-cells or \emph{transformations of monads} $\theta:(p,\varphi)\longrightarrow(q,\psi):(A,f)\longrightarrow(B,h)$, consists of a 2-cell $\theta:p\longrightarrow q$ in $\mathcal{A}$ and fulfills the commutativity of the following diagram
\begin{equation*}\label{1405151923}
\xy<1cm,0cm>
\POS (0, 0) *+{hp} = "c1",
\POS (20, 0) *+{pf} = "c2",
\POS (0, -20) *+{hq} = "d1",
\POS (20, -20) *+{qf} = "d2",
\POS (18, -10) = "a",
\POS (10, -18) = "b",
\POS "c1" \ar^{\varphi} "c2",
\POS "d1" \ar_{\psi} "d2",
\POS "c1" \ar_{h\theta } "d1",
\POS "c2" \ar^{\theta f} "d2",
\endxy
\end{equation*}
This 2-cell is displayed as follows
\begin{equation*}
\xy<1cm,0cm>
\POS (0, 0) *+{(A, f)} = "c1",
\POS (32, 0) *+{B, h)} = "c2",
\POS (16,0) *+{\theta},
\POS (10, 3.5) = "a",
\POS (10, -3.5) = "b",
\POS "c1" \ar@/^1.5pc/^{(p,\: \varphi)} "c2",
\POS "c1" \ar@/_1.5pc/_{(q,\:\psi)} "c2",
\POS "a" \ar "b",
\endxy
\end{equation*}
\end{enumerate}
The description of $2$-functor $\Phi_{\scriptscriptstyle{E}}$ is given as follows
\begin{enumerate}
\item [1.-] On $0$-cells, $\Phi_{\scriptscriptstyle{E}}(l\dashv r) = (A, rl, r\varepsilon l, \eta)$, \emph{i.e.} the induced monad by an adjunction.
\item [2.-] On $1$-cells, $(j,k,\lambda):(A, rl)\longrightarrow (\overline{A}, \overline{r}\overline{l})$, $\Phi_{\scriptscriptstyle{E}}(j,k,\lambda) = (j, \overline{r}\lambda): (A, rl) \longrightarrow (\overline{A}, \overline{r}\overline{l})$.
\item [3.-] On $2$-cells, $(\alpha,\beta):(j,k,\lambda)\longrightarrow(j',k',\lambda')$, $\Phi_{\scriptscriptstyle{E}}(\alpha,\beta)= \alpha: (j, \overline{r}\lambda)\longrightarrow (j', \overline{r}\lambda')$.
\end{enumerate}
\noindent Before the description of the 2-functor $\Psi_{\scriptscriptstyle{E}}$, we realize some calculations.\\
Since the 2-category $\mathcal{A}$ admits the construction of algebras, the 2-functor $\mathnormal{Inc}_{\mathcal{A}}:\mathcal{A}\longrightarrow \mathbf{Mnd}(\mathcal{A})$ admits a right adjoint, denoted as $\mathnormal{Alg}_{\mathcal{A}}:\mathbf{Mnd}(\mathcal{A})\longrightarrow\mathcal{A}$.\\
The corresponding counit, on the component $(A, f)$, is $\varepsilon^{IA}(A,f):\mathnormal{Inc}_{\mathcal{A}}\mathnormal{Alg}_{\mathcal{A}}(A, f)\longrightarrow (A, f)$. If we define $\mathnormal{Alg}_{\mathcal{A}}(A,f) = A^{\scriptscriptstyle{f}}$ then $\varepsilon^{IA}(A,f) :=(u^{\scriptscriptstyle{f}},\chi^{\scriptscriptstyle{f}}):(A^{\scriptscriptstyle{f}}, 1_{A^{\scriptscriptstyle{f}}})\longrightarrow (A,f)$, where $u^{\scriptscriptstyle{f}}:A^{\scriptscriptstyle{f}}\longrightarrow A$ and $\chi^{\scriptscriptstyle{f}}:u^{\scriptscriptstyle{f}}d^{\scriptscriptstyle{f}}u^{\scriptscriptstyle{f}}\longrightarrow u^{\scriptscriptstyle{f}}$.\\
In Theorem 2, at \cite{stro_fotm}, the author proved that if $\mathcal{A}$ admits the construction of algebras then for any monad $(A, f)$ in $\mathbf{Mnd}(\mathcal{A})$, there exists an adjunction in $\mathcal{A}$
\begin{equation*}\label{1405151952}
\begin{array}{cc}
\xymatrix@C=1.6cm{
A \ar@<-3pt>[r]_{d^{\scriptscriptstyle{f}}} & A^{\scriptscriptstyle{f}}\ar@<-3pt>[l]_{u^{\scriptscriptstyle{f}}}
} &\!\!\! ,
\end{array}
\end{equation*}
\noindent such that it generates the monad $(A, f)$, with unit $\eta^{\scriptscriptstyle{f}}$ and counit $\varepsilon^{d\phantom{.}\! u^{\scriptscriptstyle{f}}}$. It can be checked that $\chi^{\scriptscriptstyle{f}} = u^{\scriptscriptstyle{f}}\varepsilon^{d\phantom{.}\! u^{\scriptscriptstyle{f}}}$.\\
Suppose there is a morphism of monads $(p, \varphi):(A, f)\longrightarrow (B,h)$. Take the composition of morphisms of monads $(p,\varphi)\cdot(u^{\scriptscriptstyle{f}},\chi^{\scriptscriptstyle{f}})= (pu^{\scriptscriptstyle{f}}, p\chi^{\scriptscriptstyle{f}}\circ\varphi u^{\scriptscriptstyle{f}}):\mathnormal{Inc}_{\mathcal{A}}(A^{\scriptscriptstyle{f}}) = (A^{\scriptscriptstyle{f}},1_{A^{\scriptscriptstyle{f}}})\longrightarrow (B, h)$.\\
The previous counit, $\varepsilon^{I\! A}$, is universal from the functor $\mathnormal{Inc}_{\mathcal{A}}$, in particular, for the 1-cell
$(pu^{\scriptscriptstyle{f}}, p\chi^{\scriptscriptstyle{f}}\circ\varphi u^{\scriptscriptstyle{f}}): \mathnormal{Inc}_{\mathcal{A}}(A^{\scriptscriptstyle{f}}) \longrightarrow (A, f)$ exists a unique 1-cell in $\mathcal{A}$ of the form $p^{\varphi}:A^{\scriptscriptstyle{f}}\longrightarrow \mathnormal{Alg}_{\mathcal{A}}(B, h) = B^{\scriptscriptstyle{h}}$ such that the following diagram commutes
\begin{equation*}\label{1405161557}
\xy<1cm,0cm>
\POS (0,15) *+{\mathnormal{Inc}_{\mathcal{A}}(A^{\scriptscriptstyle{f}})} = "a11",
\POS (50,15) *+{\mathnormal{Inc}_{\mathcal{A}} (B^{\scriptscriptstyle{h}})} = "a13"
\POS (25,0) *+{(A, f)} = "a22"
\POS "a11" \ar^{\mathnormal{Inc}_{\mathcal{A}}(\phantom{.}\! p^{\varphi})} "a13",
\POS "a11" \ar_{(pu^{\scriptscriptstyle{f}}\!, \phantom{.}\! p\chi^{\scriptscriptstyle{f}}\circ\phantom{.}\! \varphi u^{\scriptscriptstyle{f}})\phantom{12}} "a22",
\POS "a13" \ar^{(u^{\scriptscriptstyle{h}},\chi^{\scriptscriptstyle{h}})} "a22",
\endxy
\end{equation*}\\
In particular, $pu^{\scriptscriptstyle{f}} = u^{\scriptscriptstyle{h}}p^{\varphi}$ and $p\chi^{\scriptscriptstyle{f}}\circ\phantom{.}\! \varphi u^{\scriptscriptstyle{f}} = \chi^{\scriptscriptstyle{h}}p^{\varphi}$. Observe that the associated mate, to the first equality, is $\lambda = \varepsilon^{\scriptscriptstyle{h}}p^{\varphi}d^{\scriptscriptstyle{f}}\circ d^{\scriptscriptstyle{h}}p\phantom{.}\!\eta^{\scriptscriptstyle{f}}$ and that $u^{\scriptscriptstyle{h}}\lambda = \varphi$.\\
Consider a 2-cell of monads, $\theta:(p, \varphi)\longrightarrow (q,\psi):(A, f)\longrightarrow (B, h)$. Because of the construction of algebras for $\mathcal{A}$, the 2-adjunction provides an isomorphism of categories, for $A$ in $\mathcal{A}$ and $(X, f)$ in $\mathbf{Mnd}(\mathcal{A})$,
\begin{equation*}\label{1405251634}
Hom_{\mathcal{A}}(A, \mathnormal{Alg}_{\mathcal{A}}(X,f))\cong Hom_{\mathbf{Mnd}(\mathcal{A})}(\mathnormal{Inc}_{\mathcal{A}}(A), (X,f))
\end{equation*}\\
\noindent given by the following assignment
\begin{equation}\label{1405251638}
\begin{array}{ccc}
\xy<1cm,0cm>
\POS (0,0) *+{A} = "a11",
\POS (25,0) *+{X^{\scriptscriptstyle{f}}} = "a12",
\POS (9, 3.5) = "a",
\POS (9,-3.5) = "b"
\POS "a11" \ar@/^1.4pc/^{a} "a12",
\POS "a11" \ar@/_1.4pc/_{b} "a12",
\POS "a" \ar^{\phantom{1}\alpha} "b",
\endxy & \longmapsto &
\xy<1cm,0cm>
\POS (0,0) *+{(A, 1_{A})} = "a11",
\POS (30,0) *+{(X, f)} = "a13",
\POS (9, 3.5) = "a",
\POS (9,-3.5) = "b",
\POS (15, 0.5) *+{\scriptstyle{u^{\scriptscriptstyle{f}}\alpha}} = "a12",
\POS "a11" \ar@/^1.5pc/^{(u^{\scriptscriptstyle{f}}a,\phantom{.}\! \chi^{\scriptscriptstyle{f}}a)} "a13",
\POS "a11" \ar@/_1.5pc/_{(u^{\scriptscriptstyle{f}}b, \phantom{.}\! \chi^{\scriptscriptstyle{f}}b)} "a13",
\POS "a" \ar "b",
\endxy
\end{array}
\end{equation}\\
\noindent cf. \cite{stro_fotm}. On the other hand, we have an equality of 2-cells\\
\begin{equation*}\label{1405251719}
\begin{array}{ccc}
\xy<1cm,0cm>
\POS (0,0) *+{(A^{\scriptscriptstyle{f}}, 1_{A^{\scriptscriptstyle{f}}})} = "a11",
\POS (30,0) *+{(B, h)} = "a13",
\POS (10, 3.5) = "a",
\POS (10,-3.5) = "b",
\POS (15, 0.5) *+{\scriptstyle{\theta u^{\scriptscriptstyle{f}}}} = "a12",
\POS "a11" \ar@/^1.5pc/^{(pu^{\scriptscriptstyle{f}}, \phantom{.}\! p\chi^{\scriptscriptstyle{f}}\circ\phantom{.}\!\varphi u^{\scriptscriptstyle{f}})} "a13",
\POS "a11" \ar@/_1.5pc/_{(qu^{\scriptscriptstyle{f}}, \phantom{.}\! q\chi^{\scriptscriptstyle{f}}\circ\phantom{.}\!\psi u^{\scriptscriptstyle{f}})} "a13",
\POS "a" \ar "b",
\endxy & = &
\xy<1cm,0cm>
\POS (0,0) *+{(A^{\scriptscriptstyle{f}}, 1_{A^{\scriptscriptstyle{f}}})} = "a11",
\POS (30,0) *+{(B, h)} = "a13",
\POS (11, 3.5) = "a",
\POS (11,-3.5) = "b",
\POS (16, 0.5) *+{\scriptstyle{\theta u^{\scriptscriptstyle{f}}}} = "a12",
\POS "a11" \ar@/^1.5pc/^{(u^{\scriptscriptstyle{h}}p^{\varphi}, \phantom{.}\! \chi^{\scriptscriptstyle{h}}p^{\varphi})} "a13",
\POS "a11" \ar@/_1.5pc/_{(u^{\scriptscriptstyle{h}}q^{\psi}, \phantom{.}\! \chi^{\scriptscriptstyle{h}}q^{\psi})} "a13",
\POS "a" \ar "b",
\endxy
\end{array}
\end{equation*}\\
Therefore, to the 2-cell $\theta u^{\scriptscriptstyle{f}}$ there corresponds, through the
assignment \eqref{1405251638}, a 2-cell $\mathnormal{Alg}_{\mathcal{A}}(\theta
u^{\scriptscriptstyle{f}})\eta^{AI}(A^{\scriptscriptstyle{f}}):= \beta^{\theta}$, where
$\beta^{\theta}:p^{\varphi}\longrightarrow q^{\psi}$ and such that
$u^{\scriptscriptstyle{h}}\beta^{\theta} = \theta u^{\scriptscriptstyle{f}}$. We change the notation as follows
$\beta^{\theta} \phantom{.}\!\separ\phantom{.}\!\separ = \phantom{.}\!\separ\phantom{.}\! \phantom{.}\!\separ\widehat{\theta}$.\\
With these calculations at hand, we define the 2-functor $\Psi_{\scriptscriptstyle{E}}$.
\begin{enumerate}
\item [1.-] On 0-cells, $(A, f)$ , $\Psi_{\scriptscriptstyle{E}}(A, f) = d^{\scriptscriptstyle{f}}\dashv u^{\scriptscriptstyle{f}}$.
\item [2.-] On 1-cells, $(p, \varphi):(A, f)\longrightarrow (B, h)$, $\Psi_{\scriptscriptstyle{E}}(p, \varphi) = (p, p^{\varphi}): d^{\scriptscriptstyle{f}}\dashv u^{\scriptscriptstyle{f}}\longrightarrow d^{\scriptscriptstyle{h}}\dashv u^{\scriptscriptstyle{h}}$.
\item [3.-] On 2-cells, $\theta:(p, \varphi)\longrightarrow (q,\psi):(A,
f)\longrightarrow (B, h)$, $\Psi_{\scriptscriptstyle{E}}(\theta) = (\theta, \widehat{\theta}):(p, p^{\varphi})\longrightarrow (q, q^{\psi}): d^{\scriptscriptstyle{f}}\dashv u^{\scriptscriptstyle{f}}\longrightarrow d^{\scriptscriptstyle{h}}\dashv u^{\scriptscriptstyle{h}}$.
\end{enumerate}
The unit and the counit for this 2-adjunction are given as follows. The component of the unit, at $l\dashv r$, is
$\eta^{\Psi\Phi_{\!\scriptscriptstyle{E}}}(l\dashv r):l\dashv r\longrightarrow\Psi_{\scriptscriptstyle{E}}\Phi_{\scriptscriptstyle{E}}(l\dashv r)$, where $\Psi_{\scriptscriptstyle{E}}\Phi_{\scriptscriptstyle{E}}(l\dashv r) = d^{\phantom{.}\! rl}\dashv u^{rl}$. In \cite{stro_fotm}, Theorem 3, the author proved the existence of a \emph{comparison} 1-cell $k^{rl}:B\longrightarrow A^{rl}$, such that $u^{rl}k^{rl} = r$ and $d^{rl} = k^{rl}l$. Therefore, we can make the following definition $\eta^{\Psi\Phi_{\!\scriptscriptstyle{E}}}(l\dashv r) = (1_{A}, k^{rl}, 1_{d^{rl}}): l\dashv r \longrightarrow d^{\phantom{.}\! rl}\dashv u^{rl}$.\\
In turn, the component of the counit, at $(A,f)$, is $\varepsilon^{\Phi\Psi_{\scriptscriptstyle{E}}}(A, f):\Phi_{\scriptscriptstyle{E}}\Psi_{\scriptscriptstyle{E}}(A, f)\longrightarrow (A,f)$, where $\Phi_{\scriptscriptstyle{E}}\Psi_{\scriptscriptstyle{E}}(A, f) = (A, f)$. In this case, the counit is defined as $\varepsilon^{\Phi\Psi_{\scriptscriptstyle{E}}}(A, f) = (1_{A}, 1_{\scriptscriptstyle{f}}):(A, f)\longrightarrow (A, f)$.
\newtheorem{1405251557}{Theorem}[section]
\begin{1405251557}
There exists a 2-adjunction $\Phi_{\scriptscriptstyle{E}}\dashv \Psi_{\scriptscriptstyle{E}}$.
\end{1405251557}
\noindent \emph{Proof}:\\
We prove only one of the triangular identities and the other one is left to the reader. Using the definition of the unit and counit for this 2-adjunction, the triangular identity $\varepsilon^{\Phi\Psi_{\!\scriptscriptstyle{E}}}\Phi_{\!\scriptscriptstyle{E}}\circ\Phi_{\!\scriptscriptstyle{E}}\eta^{\Psi\Phi_{\!\scriptscriptstyle{E}}} = 1_{\Phi_{\scriptscriptstyle{E}}}$ is proved as indicated.
\begin{eqnarray*}
(\varepsilon^{\Phi\Psi_{\!\scriptscriptstyle{E}}}\Phi_{\!\scriptscriptstyle{E}}\circ\Phi_{\!\scriptscriptstyle{E}}\eta^{\Psi\Phi_{\!\scriptscriptstyle{E}}})
(l\dashv r) &=& \varepsilon^{\Phi\Psi_{\!\scriptscriptstyle{E}}}\Phi_{\!\scriptscriptstyle{E}}(l\dashv r)\cdot\Phi_{\!\scriptscriptstyle{E}}\eta^{\Psi\Phi_{\!\scriptscriptstyle{E}}}
(l\dashv r) \\
&=& \varepsilon^{\Phi\Psi_{\!\scriptscriptstyle{E}}}(A,rl)\cdot \Phi_{\!\scriptscriptstyle{E}}(1_{A}, k^{rl},
1_{d^{rl}}) = (1_{A}, 1_{rl})\cdot(1_{A}, u^{rl}1_{d^{rl}})\\
&=& (1_{A}, 1_{rl}) = 1_{(A, rl)} = 1_{\Phi_{\scriptscriptstyle{E}}(l\dashv r)} = 1_{\Phi_{\scriptscriptstyle{E}}}(l\dashv r).
\end{eqnarray*}
$\square$
\section{Eilenberg-Moore 2-Adjunction}
In this section, we apply the results of the Section 3 to the 2-category
$_{2}Cat$, the 2-category of small categories and functors, because the
2-category ${}_{2}Cat$ admits the construction of algebras. The 2-adjunction given in that section gives a usual adjunction,
\begin{equation*}\label{1310282200}
\begin{array}{cc}
\xymatrix@C=1.6cm{
\mathbf{Adj}_{\scriptscriptstyle{R}}(_{2}Cat)\ar@<-3pt>[r]_-{\Phi_{\scriptscriptstyle{E}}} & \mathbf{Mnd}\ar@<-3pt>[l]_-{\Psi_{\!\scriptscriptstyle{E}}}(_{2}Cat)
} &\!\!\! ,
\end{array}
\end{equation*}
Since the complete description, for a general $\mathcal{A}$, has been given above, we only give some remarks on the derived properties for this particular 2-category. \\
The description of the $2$-functor $\Psi_{\scriptscriptstyle{E}}$, for this particular 2-category, is given by the following entries
\begin{enumerate}
\item [1.-] On $0$-cells, $\Psi_{\scriptscriptstyle{E}}(\mathcal{C}, F) = D^{\scriptscriptstyle{F}}\dashv U^{\scriptscriptstyle{F}}$, \emph{i.e.} the Eilenberg-Moore adjunction.
\item [2.-] On $1$-cells, $(P,\varphi) : (\mathcal{C}, F)\longrightarrow (\mathcal{D}, H)$, $\Psi_{\scriptscriptstyle{E}}(P, \varphi) = (P, P^{\phantom{.}\!\varphi}, \lambda^{\varphi})$. The action of the functor $P^{\phantom{.}\!\varphi}:\mathcal{C}^{\scriptscriptstyle{F}}\longrightarrow \mathcal{D}^{\scriptscriptstyle{H}}$ is the following
\begin{enumerate}
\item [(i)] On objects, $(M, \chi_{\scriptscriptstyle{M}})$ in $\mathcal{C}^{\scriptscriptstyle{F}}$, $P^{\phantom{.}\!\varphi}(M, \chi_{\scriptscriptstyle{M}}) = (PM, P\chi_{\scriptscriptstyle{M}}\cdot \varphi_{M})$.
\item [(ii)] On morphisms, $p$, $P^{\phantom{.}\!\varphi}(p) = Pp$.
\item [(iii)] The natural transformation $\lambda^{\varphi}$ is the mate of the identity $U^{\scriptscriptstyle{H}}P^{\varphi} = P U^{\scriptscriptstyle{F}}$. Using \eqref{1405151810}, we get the component of $\lambda^{\varphi}$ at $A$, in $\mathcal{C}$,
\begin{eqnarray*}
\lambda^{\varphi}A &=& \big(\varepsilon^{\scriptscriptstyle{D}\!\scriptscriptstyle{U}^{\!\scriptscriptstyle{H}}}P^{\varphi}D^{\scriptscriptstyle{F}}\circ D^{\scriptscriptstyle{H}}\!P\eta^{\scriptscriptstyle{U}\!\scriptscriptstyle{D}^{\!\scriptscriptstyle{F}}}\big)(A)\\
&=& P\mu^{\scriptscriptstyle{F}}A\cdot \varphi FA\cdot HP\eta^{\scriptscriptstyle{F}}A
\end{eqnarray*}
\end{enumerate}
\item [3.-] On $2$-cells, $\theta:(P,\varphi)\longrightarrow(Q,\psi)$, we have
\begin{equation*}\label{1403152054}
\Psi_{\scriptscriptstyle{E}}(\theta) = (\alpha^{\theta}, \beta^{\theta}) = (\theta, \hat{\theta})
\end{equation*}\\
The induced natural tranformation $\hat{\theta}:P^{\varphi}\longrightarrow Q^{\psi}:\mathcal{C}^{\scriptscriptstyle{F}}\longrightarrow \mathcal{D}^{\scriptscriptstyle{H}}$ is defined through its components as
\begin{equation*}\label{1403152055}
\hat{\theta}(M,\chi_{\scriptscriptstyle{M}}) = \theta M
\end{equation*}
\noindent It is clear that this definition is equivalent to the condition $ \theta\phantom{.}\! U^{\scriptscriptstyle{F}}=U^{\scriptscriptstyle{H}}\hat{\theta} $. \\
\end{enumerate}
Since we have a 2-adjunction, the following isomorphism of categories takes place, natural for all $L\dashv R$ and $(\mathcal{X}, H)$:
\begin{equation}\label{1403191902}
Hom_{\mathbf{Adj}_{\scriptscriptstyle{R}}({}_{2}Cat)}(L\dashv R, \Psi_{\scriptscriptstyle{E}}(\mathcal{X}, H)) \cong Hom_{\mathbf{Mnd}({}_{2}Cat)}(\Phi_{\scriptscriptstyle{E}}(L\dashv R), (\mathcal{X}, H))
\end{equation}
\section{Monoidal Liftings (Eilenberg-Moore Type)}
\subsection{Colax Monads}
In this section, we give the definition of a colax monad.\\
\newtheorem{1403161259}{Definition}[section]
\begin{1403161259}
A colax monad $\big((F, \xi, \gamma), \mu^{\scriptscriptstyle{F}}, \eta^{\scriptscriptstyle{F}}\big)$ over the monoidal category $(\mathcal{C}, \otimes, I)$ consists of the following
\begin{enumerate}
\item $(F,\mu^{\scriptscriptstyle{F}}, \eta^{\scriptscriptstyle{F}})$ is a monad on $\mathcal{C}$.
\item $(F,\xi,\gamma):(\mathcal{C}, \otimes, I)\longrightarrow (\mathcal{C}, \otimes, I)$ is a colax monoidal functor. That is to say, the natural transformations $\xi:F\cdot\otimes\longrightarrow\otimes\cdot (F\times F) $ and $\gamma:F\cdot\delta_{I}\longrightarrow \delta_{I}$ fulfills the commutativity on the following diagrams
\begin{equation}\label{1403181443}
\xymatrix@C=1.8cm@R=1.4cm{
F((A\otimes B)\otimes C)\ar[r]^{\xi_{A\otimes B,C}}\ar[d]_{Fa_{A,B,C}} & F(A\otimes B)\otimes FC\ar[r]^{\xi_{A,B}\otimes\: FC} & (FA\otimes FB)\otimes FC\ar[d]^{\alpha_{FA,FB,FC}}\\
F(A\otimes (B\otimes C))\ar[r]_{\xi_{A, B\otimes C}} & FA\otimes F(B\otimes C)\ar[r]_{FA\:\otimes\:\xi_{B,C}} & FA\otimes (FB\otimes FC)
}
\end{equation}
\begin{equation}\label{1403191813}
\begin{array}{ccc}
\xymatrix@C=1.2cm@R=1.2cm{
F(I\otimes A)\ar[r]^-{\xi_{I, A}}\ar@/_0.8pc/[rd]_{Fl_{A}} & FI \otimes FA\ar[r]^-{\gamma\phantom{.}\!\otimes FA} & I\otimes FA\ar@/^0.8pc/[ld]^{l_{FA}} \\
& FA &
} &
&
\xymatrix@C=1.2cm@R=1.2cm{
FA\otimes I\ar@/_0.8pc/[rd]_{r_{FA}} & FA\otimes FI\ar[l]_-{FA\phantom{.}\! \otimes \gamma} & F(A\otimes I)\ar[l]_-{\xi_{A,I}}\ar@/^0.8pc/[ld]^{Fr_{A}}\\
& FA &
}
\end{array}
\end{equation}
\item $\mu^{\scriptscriptstyle{F}}:(F,\xi,\gamma)\cdot(F,\xi,\gamma)\longrightarrow
(F,\xi,\gamma)$ and $\eta^{\scriptscriptstyle{F}}:(1_{\mathcal{C}}, 1_{\otimes},
1_{\delta_{I}})\longrightarrow (F,\xi,\gamma)$ are colax natural
transformations, \emph{i.e.} apart from the fact that they are natural
transformations, they fulfill additionally the following commutative diagrams
\begin{equation}\label{1403191825}
\begin{array}{ccc}
\xymatrix@C=1.5cm@R=1.5cm{
FF\otimes \ar[r]^-{F\xi}\ar[d]_{\mu^{\scriptscriptstyle{F}}\otimes} & F\otimes (F\times F)\ar[r]^-{\xi (F\times F) } & \otimes (FF\times FF)\ar[d]^{\otimes(\mu^{\scriptscriptstyle{F}}\times \mu^{\scriptscriptstyle{F}})}\\
F\otimes\ar[rr]_{\xi} & & \otimes (F\times F)
} &
\xy<1cm,0cm>
\POS (0,-20.5) *+{,},
\endxy
&
\xymatrix@C=1.15cm@R=1.5cm{
FF\delta_{I}\ar[d]_{\mu^{\scriptscriptstyle{F}}\delta_{I}}\ar[r]^-{F\gamma} & F\delta_{I}\ar[r]^-{\gamma} & \delta_{I}\\
F\delta_{I}\ar@/_1pc/[rru]_{\gamma} & &
}
\end{array}
\end{equation}
\noindent \&
\begin{equation}\label{1403191842}
\begin{array}{ccc}
\xy
\POS (-22, 0) *+{\otimes} = "a",
\POS (22, 0) *+{\otimes} = "d",
\POS (-22, -20) *+{F\otimes} = "c",
\POS (22, -20) *+{\otimes(F\times F)} = "b",
\POS "a" \ar^{1_{\otimes}} "d",
\POS "d" \ar^{\otimes(\eta^{\scriptscriptstyle{F}}\times\eta^{\scriptscriptstyle{F}})} "b",
\POS "a" \ar_{\eta^{\scriptscriptstyle{F}}\otimes} "c",
\POS "c" \ar_{\xi} "b",
\endxy
&
\xy<1cm,0cm>
\POS (0,-20.5) *+{,},
\endxy
&
\xy<1cm,0cm>
\POS (0, -4)
\xymatrix@C=1.2cm@R=1cm{
\delta_{I}\ar[d]_{\eta^{\scriptscriptstyle{F}}\delta_{I}} \ar[r]^{1_{\delta_{I}}} & \delta_{I}\\
F\delta_{I}\ar@/_0.7pc/[ru]_{\gamma} &
}
\endxy
\end{array}
\end{equation}
\end{enumerate}
\end{1403161259}
Since the natural transformation $\gamma$ has only one component, at $0$, then this natural transformation and its component will be denoted indistinctly as $\gamma$.\\
Using the isomorphism \eqref{1403191902}, the following bijection can be obtained, cf. \cite{moie_motc}
\newtheorem{1403192127}{Theorem}[section]
\begin{1403192127}
There is bijective correspondance between the following structures
\begin{enumerate}
\item [1.-] Colax monads $\big((F, \xi, \gamma), \mu^{\scriptscriptstyle{F}}, \eta^{\scriptscriptstyle{F}}\big)$, for the monoidal structure $(\mathcal{C}, \otimes, I, a, l, r)$.
\item [2.-] Morphisms and natural transformations of monads of the form
\begin{eqnarray*}
(\otimes, \xi)&:& (\mathcal{C}\times\mathcal{C}, F\times F)\longrightarrow (\mathcal{C}, F),\\
(\delta_{I}, \gamma)&:&(\mathbf{1}\:, 1_{\mathbf{1}}\:)\longrightarrow (\mathcal{C},F)\\
a&:&(\otimes\cdot(\otimes\times \mathcal{C}),\otimes(\xi\times F)\circ\xi(\otimes\times\mathcal{C}))\longrightarrow(\otimes\cdot (\mathcal{C}\times\otimes)\cdot a_{\mathcal{C}}, \otimes(F\times\xi)a_{\mathcal{C}}\circ\xi(\mathcal{C}\times\otimes)a_{\mathcal{C}})\\
&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad:((\mathcal{C}\times\mathcal{C})\times\mathcal{C}, (F\times F)\times F)\longrightarrow (\mathcal{C}, F), \\
l&:& (\otimes\cdot(\delta_{I}\times \mathcal{C})\cdot l_{\mathcal{C}}^{-1}, \otimes(\gamma\times F)\phantom{.}\! l^{-1}_{\mathcal{C}}\circ
\phantom{.}\!\xi(\delta_{I}\times\mathcal{C})\phantom{.}\! l^{-1}_{\mathcal{C}})\longrightarrow (1_{\mathcal{C}}, 1_{F}):(\mathcal{C}, F)\longrightarrow (\mathcal{C}, F),\\
r&:&(\otimes\cdot( \mathcal{C}\times\delta_{I})\cdot r_{\mathcal{C}}^{-1},
\otimes(F\times\gamma)\phantom{.}\! r^{-1}_{\mathcal{C}}\circ
\xi(\mathcal{C}\times\delta_{I})\phantom{.}\! r^{-1}_{\mathcal{C}})\longrightarrow (1_{\mathcal{C}}, 1_{F}):(\mathcal{C}, F)\longrightarrow (\mathcal{C}, F).\\
\end{eqnarray*}
\item [3.-] Monoidal structures for the Eilenberg-Moore category, $(\mathcal{C}^{\scriptscriptstyle{F}}, \widehat{\otimes} , \hat{I}, \hat{a}, \hat{l}, \hat{r})$ such that the following diagram of arrows and surfaces commutes
\begin{equation}\label{1303192133}
\begin{array}{cc}
\xy<1cm,0cm>
\POS (0, 10) *+{(a)},
\POS (0, 0) *+{\mathcal{C}\times\mathcal{C}} = "c1",
\POS (25, 0) *+{\mathcal{C}} = "c2",
\POS (0, -20) *+{\mathcal{C}^{\scriptscriptstyle{F}}\times\mathcal{C}^{\scriptscriptstyle{F}}} = "d1",
\POS (25, -20) *+{\mathcal{C}^{\scriptscriptstyle{F}}} = "d2",
\POS (18, -10) = "a",
\POS (10, -18) = "b",
\POS "c1" \ar^-{\otimes} "c2",
\POS "d1" \ar^{U^{\scriptscriptstyle{F}}\times U^{\scriptscriptstyle{F}}} "c1",
\POS "d2" \ar_{U^{\scriptscriptstyle{F}}} "c2",
\POS "d1" \ar_-{\widehat{\otimes}} "d2",
\endxy &
\xy<1cm,0cm>
\POS (0, 10) *+{(b)},
\POS (0,0) *+{\mathbf{1}} = "c1",
\POS (25,0) *+{\mathcal{C}} = "c2",
\POS (25,-20) *+{\mathcal{C}^{\scriptscriptstyle{F}}} = "d2",
\POS (0,-20) *+{\mathbf{1}^{1_{\mathbf{1}}}} = "d1",
\POS "c1" \ar^-{\delta_{I}} "c2",
\POS "d1" \ar^{U^{1_{\mathbf{1}}}} "c1",
\POS "d2" \ar_{U^{\scriptscriptstyle{F}}} "c2",
\POS "d1" \ar_-{\delta_{\hat{I}}} "d2",
\endxy
\end{array}
\end{equation}
\begin{equation*}\label{1404271827}
\begin{array}{ccc}
\xy<1cm,0cm>
\POS (0, 0) *+{\mathcal{C}^{3}} = "c1",
\POS (7, 0) *+{\phantom{(\mathcal{C})}}= "c3",
\POS (30, 0) *+{\mathcal{C}} = "c2",
\POS (11, 4) = "a",
\POS (11, -4) = "b",
\POS (11,-36) = "c",
\POS (11,-44) = "d",
\POS (0,-40) *+{(\mathcal{C}^{\scriptscriptstyle{F}})^{3}} = "d1",
\POS (30,-40) *+{\mathcal{C}^{\scriptscriptstyle{F}}} = "d2",
\POS "c1" \ar@/^1.5pc/^{\otimes\cdot (\otimes\times \mathcal{C})} "c2",
\POS "c1" \ar@/_1.5pc/_{\otimes\cdot (\mathcal{C}\times\otimes)\cdot a_{\mathcal{C}}} "c2",
\POS "a" \ar^{\phantom{11}a} "b",
\POS "d1" \ar@/^1.5pc/^{\widehat{\otimes}\cdot (\widehat{\otimes}\times \mathcal{C}^{\scriptscriptstyle{F}})} "d2",
\POS "d1" \ar@/_1.5pc/_{\widehat{\otimes}\cdot (\mathcal{C}^{\scriptscriptstyle{F}}\times\widehat{\otimes})\cdot a_{\mathcal{C}^{\scriptscriptstyle{F}}}} "d2",
\POS "c" \ar^{\phantom{11}\widehat{a}} "d",
\POS "d1" \ar^{(U^{\scriptscriptstyle{F}})^{3}} "c1",
\POS "d2" \ar_{U^{\scriptscriptstyle{F}}} "c2",
\endxy &
\xy<1cm,0cm>
\POS (0, 0) *+{\mathcal{C}} = "c1",
\POS (7, 0) *+{\phantom{(\mathcal{C})}}= "c3",
\POS (30, 0) *+{\mathcal{C}} = "c2",
\POS (11, 4) = "a",
\POS (11, -4) = "b",
\POS (11,-36) = "c",
\POS (11,-44) = "d",
\POS (0,-40) *+{\mathcal{C}^{\scriptscriptstyle{F}}} = "d1",
\POS (30,-40) *+{\mathcal{C}^{\scriptscriptstyle{F}}} = "d2",
\POS "c1" \ar@/^1.5pc/^{\otimes\cdot (\delta_{I}\times \mathcal{C})\cdot l^{-1}_{\mathcal{C}}} "c2",
\POS "c1" \ar@/_1.5pc/_{1_{\mathcal{C}}} "c2",
\POS "a" \ar^{\phantom{12}l} "b",
\POS "d1" \ar@/^1.5pc/^{\widehat{\otimes}\cdot (\delta_{\hat{I}}\times
\mathcal{C}^{\scriptscriptstyle{F}})\cdot l^{-1}_{\mathcal{C}^{\scriptscriptstyle{F}}}} "d2",
\POS "d1" \ar@/_1.5pc/_{1_{\mathcal{C}^{\scriptscriptstyle{F}}}} "d2",
\POS "c" \ar^{\phantom{12}\widehat{l}} "d",
\POS "d1" \ar^{U^{\scriptscriptstyle{F}}} "c1",
\POS "d2" \ar_{U^{\scriptscriptstyle{F}}} "c2",
\endxy
&
\xy<1cm,0cm>
\POS (0, 0) *+{\mathcal{C}} = "c1",
\POS (7, 0) *+{\phantom{(\mathcal{C})}}= "c3",
\POS (30, 0) *+{\mathcal{C}} = "c2",
\POS (11, 4) = "a",
\POS (11, -4) = "b",
\POS (11,-36) = "c",
\POS (11,-44) = "d",
\POS (0,-40) *+{\mathcal{C}^{\scriptscriptstyle{F}}} = "d1",
\POS (30,-40) *+{\mathcal{C}^{\scriptscriptstyle{F}}} = "d2",
\POS "c1" \ar@/^1.5pc/^{\otimes\cdot (\mathcal{C}\times\delta_{I})\cdot r^{-1}_{\mathcal{C}}} "c2",
\POS "c1" \ar@/_1.5pc/_{1_{\mathcal{C}}} "c2",
\POS "a" \ar^{\phantom{12}r} "b",
\POS "d1" \ar@/^1.5pc/^{\widehat{\otimes}\cdot
(\mathcal{C}^{\scriptscriptstyle{F}}\times\delta_{\hat{I}})\cdot r^{-1}_{\mathcal{C}^{\scriptscriptstyle{F}}}} "d2",
\POS "d1" \ar@/_1.5pc/_{1_{\mathcal{C}^{\scriptscriptstyle{F}}}} "d2",
\POS "c" \ar^{\phantom{12}\widehat{r}} "d",
\POS "d1" \ar^{U^{\scriptscriptstyle{F}}} "c1",
\POS "d2" \ar_{U^{\scriptscriptstyle{F}}} "c2",
\endxy
\end{array}
\end{equation*}
\end{enumerate}
\end{1403192127}
\noindent\emph{Proof}:\\
$1\Rightarrow 2$)\\
Consider a colax monad $\big((F, \xi, \gamma), \mu^{\scriptscriptstyle{F}}, \eta^{\scriptscriptstyle{F}}\big)$, for the monoidal structure $(\mathcal{C}, \otimes, I)$. In particular, the multiplication and the unit of the monad are colax natural transformations and the first diagrams in \eqref{1403191825} and \eqref{1403191842} commute. Therefore, we have a monad morphism $(\otimes, \xi):(\mathcal{C}\times\mathcal{C} , F\times F)\longrightarrow (\mathcal{C}, F)$.\\
Likewise, the commutativity of the second diagrams in \eqref{1403191825} and \eqref{1403191842} implies that $(\delta_{I}, \gamma):(\mathbf{1}\:, 1_{\mathbf{1}}\:)\longrightarrow (\mathcal{C},F)$ is a morphism of monads. Note that the requirement $(\delta_{I}, \gamma)$ is a monad morphism is equivalent to the statement $(I, \gamma)$ is an Eilenberg-Moore algebra.\\
Since $(\otimes, \xi)$ is a morphism of monads then the following morphisms are also morphisms of monads $\big(\otimes\cdot(\otimes\times\mathcal{C}), \otimes(\xi\times F)\circ\xi(\otimes\times\mathcal{C})\big)$ and $(\otimes\cdot(\mathcal{C}\times\otimes)\cdot a_{\mathcal{C}}, \otimes(F\times\xi)a_{\mathcal{C}}\circ \xi(\mathcal{C}\times\otimes)a_{\mathcal{C}})$ from $((\mathcal{C}\times \mathcal{C})\times \mathcal{C}, (F\times F)\times F)$ to $(\mathcal{C}, F)$ and due to the commutativity of the diagram \eqref{1403181443}, the following is a $2$-cell in $\mathbf{Mnd}$
\begin{equation*}
\xy<1cm,0cm>
\POS (0, 0) *+{((\mathcal{C}\times\mathcal{C})\times\mathcal{C}, (F\times F)\times F)} = "c1",
\POS (7, 0) *+{\phantom{(\mathcal{C})}}= "c3",
\POS (50, 0) *+{(\mathcal{C}, F)} = "c2",
\POS (27, 5) = "a",
\POS (27, -5) = "b",
\POS "c3" \ar@/^2pc/^{(\otimes\cdot (\otimes\times \mathcal{C}),\:\otimes(\xi\times F)\circ\xi(\otimes\times\mathcal{C}))} "c2",
\POS "c3" \ar@/_2pc/_{(\otimes\cdot (\mathcal{C}\times\otimes)\cdot a_{\mathcal{C}},\:\otimes(F\times\xi)a_{\mathcal{C}}\circ\xi(\mathcal{C}\times\otimes)a_{\mathcal{C}})} "c2",
\POS "a" \ar^{\phantom{1}a} "b",
\endxy
\end{equation*}\\
Likewise, because $(\otimes, \xi)$ and $(\delta_{I}, \gamma)$ are monad morphisms, $(\otimes\cdot (\delta_{I}\times \mathcal{C})\cdot l^{-1}_{\mathcal{C}}, \otimes(\gamma\times F)\phantom{.}\! l^{-1}_{\mathcal{C}}\circ \phantom{.}\!\xi(\delta_{I}\times\mathcal{C})\phantom{.}\! l^{-1}_{\mathcal{C}})$ is also a monad morphism. Using the commutativity of the first diagram in \eqref{1403191813}, we can consider the monad $2$-cell
\begin{equation*}
\xy<1cm,0cm>
\POS (0, 0) *+{(\mathcal{C}, F)} = "c1",
\POS (7, 0) *+{\phantom{(\mathcal{C})}}= "c3",
\POS (55, 0) *+{(\mathcal{C}, F)} = "c2",
\POS (19, 5) = "a",
\POS (19, -5) = "b",
\POS "c1" \ar@/^2pc/^{(\otimes\cdot (\delta_{I}\times \mathcal{C})\phantom{.}\!
l^{-1}_{\mathcal{C}},\:\otimes\!\phantom{.}(\gamma\times
F)\!\phantom{.}l^{-1}_{\mathcal{C}}\circ\phantom{.}\! \xi (\delta_{I}\times \mathcal{C})l^{-1}_{\mathcal{C}})} "c2",
\POS "c1" \ar@/_2pc/_{(1_{\mathcal{C}}, 1_{\scriptscriptstyle{F}})} "c2",
\POS "a" \ar^{\phantom{1234}l} "b",
\endxy
\end{equation*}\\
In a similar way, the following is a monad transformation, $r:(\otimes\cdot(\mathcal{C}\times\delta_{I})\cdot r_{\mathcal{C}}^{-1}, \otimes(F\times\gamma)\phantom{.}\! r^{-1}_{\mathcal{C}}\circ\xi(\mathcal{C}\times\delta_{I})\phantom{.}\! r^{-1}_{\mathcal{C}})\longrightarrow (1_{\mathcal{C}}, 1_{F}):(\mathcal{C}, F)\longrightarrow (\mathcal{C}, F)$.\\
$2\Rightarrow 1$)\\
Note that the aforementioned claims can be reverted.\\
2 $\Rightarrow$ 3)\\
Take the monad morphism $(\otimes, \xi): (\mathcal{C}\times\mathcal{C},F\times F)\longrightarrow (\mathcal{C}, F)$. In order to use the isomorphism \eqref{1403191902}, we make $L\dashv R = D^{\scriptscriptstyle{F}}\times D^{\scriptscriptstyle{F}} \dashv U^{\scriptscriptstyle{F}}\times U^{\scriptscriptstyle{F}}$ and $(\mathcal{X}, H,\mu^{\scriptscriptstyle{H}},\eta^{\scriptscriptstyle{H}}) = (\mathcal{C}, F, \mu^{\scriptscriptstyle{F}}, \eta^{\scriptscriptstyle{F}})$. Therefore, to this monad morphism corresponds a morphism of adjunctions of the form $(\otimes, \otimes^{\xi}): D^{\scriptscriptstyle{F}}\times D^{\scriptscriptstyle{F}} \dashv U^{\scriptscriptstyle{F}}\times U^{\scriptscriptstyle{F}} \longrightarrow D^{\scriptscriptstyle{F}}\dashv U^{\scriptscriptstyle{F}}$ such that a diagram like (\ref{1303192133}a) commutes. According to the definition of $\Psi^{\scriptscriptstyle{E}}$, the functor $\otimes^{\xi}$ acts as follows
\begin{equation*}\label{1403201159}
\otimes^{\xi}\big((M,\chi_{\scriptscriptstyle{M}}),(N,\chi_{\scriptscriptstyle{N}})\big) = \big(\otimes (M, N), \otimes(\chi_{\scriptscriptstyle{M}}, \chi_{\scriptscriptstyle{N}})\cdot \xi_{\scriptscriptstyle{M},\scriptscriptstyle{N}}\big)
\end{equation*}\\
The previous action is defined at the beginning of the proof of Theorem 7.1, \cite{moie_motc}.
\begin{equation*}\label{1403201301}
\otimes^{\xi}(p,q) = \otimes(p,q)
\end{equation*}\\
\noindent We change the notation from $\otimes^{\xi}$ to $\widehat{\otimes}$.\\
If in the isomorphism \eqref{1403191902}, we make $L\dashv R = 1_{\mathbf{1}}\dashv 1_{\mathbf{1}}$ and $(\mathcal{X}, H,\mu^{\scriptscriptstyle{H}},\eta^{\scriptscriptstyle{H}}) = (\mathcal{C}, F, \mu^{\scriptscriptstyle{F}}, \eta^{\scriptscriptstyle{F}})$. The monad morphism $(\delta_{I}, \gamma)$ has an associated morphism of adjunctions of the form $(\delta_{I}, \delta_{I}^{\phantom{I}\gamma}):(1_{\mathbf{1}}\dashv 1_{\mathbf{1}})\longrightarrow D^{\scriptscriptstyle{F}}\dashv U^{\scriptscriptstyle{F}}$ such that a diagram like (\ref{1303192133}b) commutes. According to the definition of $\Psi^{\scriptscriptstyle{E}}$, the functor $\delta_{I}^{\phantom{I}\gamma}$ acts as follows
\begin{equation*}\label{1403201409}
\delta_{I}^{\phantom{I}\gamma}(0, 1_{0}) = (\delta_{I}(0), \delta_{I}(1_{0})\cdot \gamma) = (I, \gamma).
\end{equation*}\\
\noindent On morphisms,
\begin{equation*}\label{1403201416}
\delta_{I}^{\phantom{I}\gamma}(1_{0}) = \delta_{I}(1_{0}) = 1_{I} = 1_{(I,\gamma)}.
\end{equation*}\\
\noindent If we make the following definition $\hat{I} = (I, \gamma)$, then $\delta_{I}^{\phantom{I}\gamma} := \delta_{\hat{I}}$. The algebra $(I, \gamma)$ is the unit of the monoidal structure on $\mathcal{C}^{\scriptscriptstyle{F}}$. \\
Suppose that we have a natural transformation of the form $a:(\otimes\cdot(\otimes\times \mathcal{C}),\otimes(\xi\times F)\circ\xi(\otimes\times\mathcal{C}))\longrightarrow(\otimes\cdot (\mathcal{C}\times\otimes)\cdot a_{\mathcal{C}}, \otimes(F\times\xi)a_{\mathcal{C}}\phantom{.}\!\circ\phantom{.}\!\xi(\mathcal{C}\times\otimes)a_{\mathcal{C}}):((\mathcal{C}\times\mathcal{C})\times\mathcal{C}, (F\times F)\times F)\longrightarrow (\mathcal{C}, F)$ then we can make $L\dashv R = (D^{\scriptscriptstyle{F}}\times D^{\scriptscriptstyle{F}})\times D^{\scriptscriptstyle{F}}\dashv (U^{\scriptscriptstyle{F}}\times U^{\scriptscriptstyle{F}})\times U^{\scriptscriptstyle{F}}$ and $(\mathcal{X}, H,\mu^{\scriptscriptstyle{H}},\eta^{\scriptscriptstyle{H}}) = (\mathcal{C}, F, \mu^{\scriptscriptstyle{F}}, \eta^{\scriptscriptstyle{F}})$. Therefore, to the previous $2$-cell of monads, we can associate a $2$-cell of adjunctions of the form
\begin{equation*}
\xy<1cm,0cm>
\POS (0, 0) *+{\mathcal{C}^{3}} = "c1",
\POS (7, 0) *+{\phantom{(\mathcal{C})}}= "c3",
\POS (40, 0) *+{\mathcal{C}} = "c2",
\POS (14, 4) = "a",
\POS (14, -4) = "b",
\POS (14,-36) = "c",
\POS (14,-44) = "d",
\POS (0,-40) *+{(\mathcal{C}^{\scriptscriptstyle{F}})^{3}} = "d1",
\POS (40,-40) *+{\mathcal{C}^{\scriptscriptstyle{F}}} = "d2",
\POS "c1" \ar@/^1.7pc/^{\otimes\cdot (\otimes\times \mathcal{C})} "c2",
\POS "c1" \ar@/_1.7pc/_{\otimes\cdot (\mathcal{C}\times\otimes)\cdot a_{\mathcal{C}}} "c2",
\POS "a" \ar^{\phantom{12}a} "b",
\POS "d1" \ar@/^1.7pc/^{[\otimes\cdot (\otimes\times \mathcal{C})]^{\phantom{.}\!{}^{\cdot}\!\xi^{2}}} "d2",
\POS "d1" \ar@/_1.7pc/_{[\otimes\cdot (\mathcal{C}\times\otimes)\cdot a_{\mathcal{C}}]^{{}_{\cdot}\xi^{2}}} "d2",
\POS "c" \ar^{\phantom{12}\beta^{a}} "d",
\POS "c1" \ar@<-4pt>_{(D^{\scriptscriptstyle{F}})^{3}} "d1",
\POS "d1" \ar@<-1pt>_{(U^{\scriptscriptstyle{F}})^{3}} "c1",
\POS "c2" \ar@<-2pt>_{D^{\scriptscriptstyle{F}}} "d2",
\POS "d2" \ar@<-3pt>_{U^{\scriptscriptstyle{F}}} "c2",
\endxy
\end{equation*}\\
In order to reduce expressions, we used and will be using the following notation
\begin{eqnarray*}
{}^{\cdot}\xi^{2} &:=& \otimes(\xi\times F)\circ\xi(\otimes\times\mathcal{C}), \\
{}_{\cdot}\xi^{2} &:=& \otimes(F\times\xi)a_{\mathcal{C}}\circ\xi(\mathcal{C}\times\otimes)a_{\mathcal{C}}, \\
(\cdot)^{3} &:=& (\cdot\times\cdot)\times\cdot
\end{eqnarray*}
It can be prove that $[\otimes\cdot (\otimes\times \mathcal{C})]^{\phantom{.}\!{}^{\cdot}\!\xi^{2}} = \widehat{\otimes}\cdot (\widehat{\otimes}\times \mathcal{C}^{\scriptscriptstyle{F}})$. On objects and morphisms
\begin{eqnarray*}
&&\!\!\!\!\!\!\!\!\![\otimes\cdot (\otimes\times
\mathcal{C})]^{\phantom{.}\!{}^{\cdot}\!\xi^{2}}\big(\big((M,\chi_{\scriptscriptstyle{M}}),(N,\chi_{\scriptscriptstyle{N}})\big),
(M',\chi_{\scriptscriptstyle{M}'})\big) \\
&=& \big\lbrack(\otimes\cdot(\otimes\times \mathcal{C}))\big( (M,N), M'\big),(\otimes\cdot (\otimes\times \mathcal{C}))\big((\chi_{\scriptscriptstyle{M}},\chi_{\scriptscriptstyle{N}}), \chi_{\scriptscriptstyle{M}'}\big)\cdot (\xi_{M,N}\otimes FM') \cdot \xi_{M\otimes N, M'} \big\rbrack\\
&=& \big\lbrack (M\otimes N)\otimes M', \big((\chi_{\scriptscriptstyle{M}}\otimes\chi_{\scriptscriptstyle{N}})\otimes\chi_{\scriptscriptstyle{M}'}\big)\cdot \big(\xi_{M,N}\otimes FM'\big) \cdot \xi_{M\otimes N, M'}\big\rbrack\\
&=& \big\lbrack (M\otimes N)\otimes M', \big(\lbrack(\chi_{\scriptscriptstyle{M}}\otimes\chi_{\scriptscriptstyle{N}})\cdot \xi_{M,N}\rbrack\otimes\chi_{\scriptscriptstyle{M}'}\big) \cdot \xi_{M\otimes N, M'}\big\rbrack\\
&=& \widehat{\otimes}\big((M\otimes N,(\chi_{\scriptscriptstyle{M}}\otimes\chi_{\scriptscriptstyle{N}})\cdot\xi_{\scriptscriptstyle{M},\scriptscriptstyle{N}}),(M',\chi_{\scriptscriptstyle{M}'})\big)\\
&=& \widehat{\otimes}\cdot(\widehat{\otimes}\times \mathcal{C}^{\scriptscriptstyle{F}})\big(\big((M,\chi_{\scriptscriptstyle{M}}),(N,\chi_{\scriptscriptstyle{N}})\big), (M',\chi_{\scriptscriptstyle{M}'})\big)
\end{eqnarray*}
\begin{eqnarray*}
[\otimes\cdot (\otimes\times \mathcal{C})]^{\phantom{.}\!{}^{\cdot}\!\xi^{2}} \big((p,q), p'\big) &=& \otimes\cdot (\otimes\times \mathcal{C}) ((p,q),p')\\
& =& (p\otimes q)\otimes p'
\end{eqnarray*}
In the same way, we can check that $[\otimes\cdot(\mathcal{C}\times\otimes)\cdot a_{\mathcal{C}}]^{{}_{\cdot}\xi^{2}} = \widehat{\otimes}\cdot (\mathcal{C}^{\scriptscriptstyle{F}}\times\widehat{\otimes})\cdot a_{\mathcal{C}^{\scriptscriptstyle{F}}}$. We change the notation $\beta^{a}$ for $\widehat{a}$ and we get a natural transformation $\widehat{a}:\widehat{\otimes}(\widehat{\otimes}\times\mathcal{C}^{\scriptscriptstyle{F}})\longrightarrow \widehat{\otimes}(\mathcal{C}^{\scriptscriptstyle{F}}\times\widehat{\otimes})\cdot a_{\mathcal{C}^{\scriptscriptstyle{F}}}:(\mathcal{C}^{\scriptscriptstyle{F}}\times\mathcal{C}^{\scriptscriptstyle{F}})\times\mathcal{C}^{\scriptscriptstyle{F}}\longrightarrow \mathcal{C}^{\scriptscriptstyle{F}}$. Using the definition of the functor $\Psi_{\scriptscriptstyle{E}}$ on the $2$-cell $a$, we get the component at $\big(\big((M,\chi_{\scriptscriptstyle{M}}),(N,\chi_{\scriptscriptstyle{N}})\big), (M',\chi_{\scriptscriptstyle{M}'})\big) $
\begin{equation*}\label{1403221055}
\widehat{a}\big(\big((M,\chi_{\scriptscriptstyle{M}}),(N,\chi_{\scriptscriptstyle{N}})\big), (M',\chi_{\scriptscriptstyle{M}'})\big)= a(M,N,M')
\end{equation*}\\
Suppose we have a $2$-cell in $\mathbf{Mnd}$ of the form $l:(\otimes\cdot(\delta_{I}\times \mathcal{C})\cdot l_{\mathcal{C}}^{-1}, \otimes(\gamma\times F)\phantom{.}\! l^{-1}_{\mathcal{C}}\circ \phantom{.}\!\xi(\delta_{I}\times\mathcal{C})\phantom{.}\! l^{-1}_{\mathcal{C}})\longrightarrow (1_{\mathcal{C}}, 1_{F}):(\mathcal{C}, F)\longrightarrow (\mathcal{C}, F)$. If in the isomorphism \eqref{1403191902}, we make $L\dashv R = D^{\scriptscriptstyle{F}} \dashv U^{\scriptscriptstyle{F}}$ and $(\mathcal{X},H,\mu^{\scriptscriptstyle{H}},\eta^{\scriptscriptstyle{H}}) = (\mathcal{C}, F, \mu^{\scriptscriptstyle{F}}, \eta^{\scriptscriptstyle{F}})$, it can be obtained a $2$-cell in the $2$-category $\mathbf{Adj}_{\scriptscriptstyle{R}}({}_{2}Cat)$ of the form $(l,\beta^{l}):(\otimes\cdot (\delta_{I}\times \mathcal{C})\cdot l^{-1}_{\mathcal{C}}, 1_{\mathcal{C}})\longrightarrow (\lbrack \otimes\cdot (\delta_{I}\times \mathcal{C})\cdot l^{-1}_{\mathcal{C}}\rbrack^{\gamma\circ\xi}, \lbrack 1_{\mathcal{C}}\rbrack^{1_{\scriptscriptstyle{F}}}):D^{\scriptscriptstyle{F}}\dashv U^{\scriptscriptstyle{F}}\longrightarrow D^{\scriptscriptstyle{F}}\dashv U^{\scriptscriptstyle{F}}$. Where we used the notation $\gamma\circ\xi = \otimes\!\phantom{.}(\gamma\times F)\!\phantom{.}l^{-1}_{\mathcal{C}}\circ\xi (\delta_{I}\times \mathcal{C})\phantom{.}\! l^{-1}_{\mathcal{C}}$. We change the notation from $\beta^{l}$ to $\hat{l}$.\\
In the same way as before, it can be proved that $\lbrack \otimes\cdot (\delta_{I}\times \mathcal{C})\cdot l^{-1}_{\mathcal{C}}\rbrack^{\gamma\circ\xi} = \widehat{\otimes}\phantom{.}\! (\delta_{\hat{I}}\times \mathcal{C}^{\scriptscriptstyle{F}})\phantom{.}\! l^{-1}_{\mathcal{C}^{\scriptscriptstyle{F}}} $ and $\lbrack 1_{\mathcal{C}}\rbrack^{1_{\scriptscriptstyle{F}}} = 1_{\mathcal{C}^{\scriptscriptstyle{F}}}$. Therefore, we obtain a natural transformation $\hat{l}:\widehat{\otimes}\phantom{.}\!(\delta_{\hat{I}}\times \mathcal{C}^{\scriptscriptstyle{F}})\phantom{.}\! l^{-1}_{\mathcal{C}^{\scriptscriptstyle{F}}} \longrightarrow 1_{\mathcal{C}^{\scriptscriptstyle{F}}}:\mathcal{C}^{\scriptscriptstyle{F}}\longrightarrow \mathcal{C}^{\scriptscriptstyle{F}} $. Using the definition of the $2$-functor $\Psi_{E}$ on the $2$-cell $l$, the component of the natural transformation $\hat{l}$ on $(M, \chi_{\scriptscriptstyle{M}})$ is
\begin{equation*}\label{1405071231}
\hat{l}(M,\chi_{\scriptscriptstyle{M}}) = l_{\scriptscriptstyle{M}}
\end{equation*}\\
Similarly to the $2$-cell $r:(\otimes\cdot(\mathcal{C}\times\delta_{I})\cdot r^{-1}_{\mathcal{C}}, \:\otimes\!\phantom{.}(F\times \gamma)\!\phantom{.}r^{-1}_{\mathcal{C}}\circ\xi\phantom{.}\! (\mathcal{C}\phantom{.}\!\times \delta_{I})\phantom{.}\! r^{-1}_{\mathcal{C}})\longrightarrow(1_{\mathcal{C}}, 1_{\scriptscriptstyle{F}}):(\mathcal{C}, F)\longrightarrow(\mathcal{C}, F)$ there corresponds a natural transformation $\hat{r}:\widehat{\otimes} \phantom{.}\!(\mathcal{C}^{\scriptscriptstyle{F}}\times \delta_{\hat{I}})\phantom{.}\! r^{-1}_{\mathcal{C}^{\scriptscriptstyle{F}}}\longrightarrow 1_{\mathcal{C}^{\scriptscriptstyle{F}}}: \mathcal{C}^{\scriptscriptstyle{F}}\longrightarrow \mathcal{C}^{\scriptscriptstyle{F}}$. The component of this natural transformation, at $(M,\chi_{\scriptscriptstyle{M}})$, is
\begin{equation}\label{1405071326}
\hat{r}(M,\chi_{\scriptscriptstyle{M}}) = r_{\scriptscriptstyle{M}}
\end{equation}\\
Since the natural transformations $a, l$ and $r$ fulfill the coherence conditions for a monoidal struture and $U^{\scriptscriptstyle{F}}$ is faithfull then $\hat{a}, \hat{l}$ and $\hat{r}$ fulfill the pentagon and the triangle coherence conditions. Therefore, $(\mathcal{C}^{\scriptscriptstyle{F}}, \widehat{\otimes}, \hat{I}, \hat{a},\hat{l}, \hat{r})$ is a monoidal structure over $\mathcal{C}^{\scriptscriptstyle{F}}$.\\
3 $\Rightarrow$ 2)\\
Note that the aforementioned statements can be reverted. For example, take the
morphism of adjunctions
$(a,\widehat{a}):(\otimes\cdot(\otimes\times\mathcal{C}),
\widehat{\otimes}\cdot(\widehat{\otimes}\times\mathcal{C}^{\scriptscriptstyle{F}}))\longrightarrow
(\otimes\cdot(\mathcal{C}\times\otimes)\cdot a_{\mathcal{C}},
\widehat{\otimes}\cdot(\mathcal{C}^{\scriptscriptstyle{F}}\times\widehat{\otimes})\cdot
a_{\mathcal{C}^{\scriptscriptstyle{F}}}):(U^{\scriptscriptstyle{F}}\times U^{\scriptscriptstyle{F}})\times U^{\scriptscriptstyle{F}}\dashv (D^{\scriptscriptstyle{F}}\times
D^{\scriptscriptstyle{F}})\times D^{\scriptscriptstyle{F}}\longrightarrow U^{\scriptscriptstyle{F}}\dashv D^{\scriptscriptstyle{F}}$. The image of this
2-cell, under $\Phi_{\scriptscriptstyle{E}}$, is $a:(\otimes, \varphi_{\otimes})\cdot(\otimes\times\mathcal{C}, \varphi_{\otimes}\times F)\longrightarrow(\otimes,\varphi_{\otimes})\cdot(\mathcal{C}\times\otimes, F\times\varphi_{\otimes})\cdot(a_{\mathcal{C}}, 1_{F\times(F\times F)\cdot a_{\mathcal{C}}}):((\mathcal{C}\times\mathcal{C})\times\mathcal{C}, (F\times F)\times F)\longrightarrow (\mathcal{C}, F)$, \emph{i.e.}
\begin{equation*}
a:(\otimes\cdot(\otimes\times\mathcal{C}), \otimes(\varphi_{\otimes}\times
F)\circ\varphi_{\otimes}(\otimes\times\mathcal{C}))\longrightarrow
(\otimes\cdot(\mathcal{C}\times\otimes)\cdot a_{\mathcal{C}},
\otimes(F\times\varphi_{\otimes})\cdot a_{\mathcal{C}}\circ
\varphi_{\otimes}\cdot(\mathcal{C}\times\otimes)\cdot
a_{\mathcal{C}}):(\mathcal{C}^{3}, F^{3})\longrightarrow (\mathcal{C}, F)
\end{equation*}
Everytime we used the isomorphism \eqref{1403191902}, the monad $(\mathcal{C}, F, \mu^{\scriptscriptstyle{F}}, \eta^{\scriptscriptstyle{F}})$ was always
taken fixed, therefore the implication 2 $\Rightarrow$ 3 is natural in the monad $(\mathcal{C}, F, \mu^{\scriptscriptstyle{F}}, \eta^{\scriptscriptstyle{F}})$.
$\square$\\
The authors did not check for the naturality of the implication $1\Rightarrow
2$, but the reader can do it.
\section{Kleisli 2-Adjunction}
Based on either \cite{brto_eicb} or \cite{cljs_klem}, the following 2-adjunction takes place
\begin{equation*}\label{1310282201}
\begin{array}{cc}
\xymatrix@C=1.6cm{
\mathbf{Mnd}^{\scriptscriptstyle{\bullet}}({}_{2}Cat) \ar@<-3pt>[r]_-{\Psi_{\scriptscriptstyle{K}}} & \mathbf{Adj}_{\scriptscriptstyle{L}}({}_{2}Cat)\ar@<-3pt>[l]_-{\Phi_{\!\scriptscriptstyle{K}}}
} &\!\!\! ,
\end{array}
\end{equation*}
\noindent which can also be deduced from the general 2-adjunction given by \eqref{1405261543}. In this sense, we provide only a few remarks on the structure for the several objects that build this 2-adjunction. \\
The description of $2$-functor, $\Psi_{\scriptscriptstyle{K}}$, is given completely in order to provide the necessary notation. The structure of such 2-functor goes as follows\\
\begin{enumerate}
\item [1.-] On $0$-cells, $\Psi_{\scriptscriptstyle{K}}(\mathcal{C}, F) = G_{\scriptscriptstyle{F}}\dashv V_{\scriptscriptstyle{F}}$, \emph{i.e.} the Kleisli adjunction.
\item [2.-] On $1$-cells, $(P,\pi) : (\mathcal{C}, F)\longrightarrow (\mathcal{D}, H)$, $\Psi_{\scriptscriptstyle{K}}(P, \pi) = (P, P_{\pi}, \rho_{\pi})$. In the definition of the functor $P_{\pi}:\mathcal{C}_{\scriptscriptstyle{F}}\longrightarrow \mathcal{D}_{\scriptscriptstyle{H}}$, we use the notation $(\cdot)^{\sharp}$ given for a morphism in $\mathcal{C}_{\scriptscriptstyle{F}}$ and $(\cdot)^{\flat}$ for a morphism in $\mathcal{D}_{\scriptscriptstyle{H}}$. This notation is used in \cite{masa_cawm_2ned} and \cite{tami_psdl}.
\begin{enumerate}
\item [(i)] On objects, $X$ in $\mathcal{C}_{\scriptscriptstyle{F}}$, $P_{\pi}X= PX$.
\item [(ii)] On morphisms, $x^{\sharp}:X\longrightarrow Y$ in $\mathcal{C}_{\scriptscriptstyle{F}}$, $P_{\pi}x^{\sharp}= (\pi C_{x^{\sharp}}\cdot Px)^{\flat}$, where $C_{x^{\sharp}}$ is the notation for the codomain of the morphism $x^{\sharp}$ as in $C_{\scriptscriptstyle{F}}$, which in this case is $Y$.
\item [(iii)] In order to define $\rho_{\pi}$ we have to prove that the following equality of functors takes place, $G_{\scriptscriptstyle{H}}P = P_{\pi}G_{\scriptscriptstyle{F}}$.
On objects and morphisms $f:A\longrightarrow B$ in $\mathcal{C}$,
\begin{eqnarray*}
G_{\scriptscriptstyle{H}}PA &=& PA = P_{\pi}A = P_{\pi}G_{\scriptscriptstyle{F}}A ,\\
G_{\scriptscriptstyle{H}}Pf &=& (HPf\cdot \eta^{\scriptscriptstyle{H}}PA)^{\flat} = (HPf\cdot \pi A\cdot P\eta^{\scriptscriptstyle{F}}A)^{\flat} \\
&=& (\pi B\cdot PFf\cdot P\eta^{\scriptscriptstyle{F}}A)^{\flat} = P_{\pi}(Ff\cdot \eta^{\scriptscriptstyle{F}}A)^{\sharp} = P_{\pi}G_{\scriptscriptstyle{F}}f
\end{eqnarray*}
\noindent where the second equality takes place because of the unitality condition on $\pi$ and the third one is due to the naturality on $\pi$. Using \eqref{1407052237}, we get the mate for this identity
\begin{equation*}\label{1310291824}
\rho_{\pi} = V_{\scriptscriptstyle{H}}P_{\pi}\varepsilon^{DU^{\scriptscriptstyle{F}}}\scriptstyle{\circ}\textstyle{} \:\eta^{\scriptscriptstyle{H}}PV_{\scriptscriptstyle{F}},
\end{equation*}
\noindent whose component, at $X$ in $\mathcal{C}_{\scriptscriptstyle{F}}$, is $\rho_{\pi}X = \mu^{\scriptscriptstyle{H}}PX\cdot H\pi X\cdot \eta^{\scriptscriptstyle{H}}PFX = \pi X$.
\end{enumerate}
\item [3.-] On $2$-cells, $\vartheta:(P,\pi)\longrightarrow(Q,\tau)$, we have
\begin{equation*}\label{1311011235}
\Psi_{\scriptscriptstyle{K}}(\vartheta) = (\alpha_{\vartheta}, \beta_{\vartheta})
\end{equation*}
\noindent where $\alpha_{\vartheta}:= \vartheta$ and we rename $\beta_{\vartheta}$ as $\tilde{\vartheta}$. The induced natural tranformation $\tilde{\vartheta}:P_{\pi}\longrightarrow Q_{\tau}:\mathcal{C}_{\scriptscriptstyle{F}}\longrightarrow \mathcal{D}_{\scriptscriptstyle{H}}$ is defined through its components as
\begin{equation}\label{1311011822}
\tilde{\vartheta}X = (\eta^{\scriptscriptstyle{H}}QX\cdot \vartheta X)^{\flat}
\end{equation}
\noindent It is clear that this definition is equivalent to the condition $G_{\scriptscriptstyle{H}}\phantom{.}\!\vartheta=\tilde{\vartheta}\phantom{.}\! G_{\scriptscriptstyle{F}}$. \\
\end{enumerate}
Since we have a 2-adjunction, the following isomorphism of categories takes place, natural in $(\mathcal{X}, H)$ and $L\dashv R$
\begin{equation}\label{1311042204}
Hom_{\mathbf{Mnd}^{\scriptscriptstyle{\bullet}}({}_{2}Cat)}((\mathcal{X}, H), \Phi_{\scriptscriptstyle{K}}(L\dashv R))\cong Hom_{\mathbf{Adj}_{\scriptscriptstyle{L}}({}_{2}Cat)}(\Psi_{\scriptscriptstyle{K}}(\mathcal{X}, H),L\dashv R)
\end{equation}
\section{Monoidal Extensions (Kleisli Type)}
\subsection{Lax Monads}
Dual to colax monads, we give the definition of a \emph{lax monad}.\\
\newtheorem{1403061850}{Definition}[section]
\begin{1403061850}
A lax monad $((F, \zeta, \omega), \mu^{\scriptscriptstyle{F}}, \eta^{\scriptscriptstyle{F}})$ over a monoidal category $(\mathcal{C}, \otimes, I, a, l, r)$ consists of the following
\begin{enumerate}
\item $(F,\mu^{\scriptscriptstyle{F}}, \eta^{\scriptscriptstyle{F}})$ is a monad on $\mathcal{C}$.
\item $(F,\zeta,\omega):(\mathcal{C}, \otimes, I)\longrightarrow (\mathcal{C}, \otimes, I)$ is a lax monoidal functor. This means that the natural transformations $\zeta:\otimes\cdot (F\times F)\longrightarrow F\cdot\otimes$ and $\omega:\delta_{I}\longrightarrow F\cdot\delta_{I}$, fulfills the commutativity on the following diagrams
\begin{equation}\label{1403061941}
\xymatrix@C=1.8cm@R=1.4cm{
(FA\otimes FB)\otimes FC\ar[r]^{\zeta_{A,B}\otimes FC}\ar[d]_{a_{FA,FB,FC}} & F(A\otimes B)\otimes FC\ar[r]^{\zeta_{A\otimes B, C}} & F((A\otimes B)\otimes C)\ar[d]^{Fa_{A,B,C}} \\
FA\otimes (FB\otimes FC)\ar[r]_{FA\otimes \zeta_{B,C}} & FA\otimes F(B\otimes C) \ar[r]_{\zeta_{A,B\otimes C}} & F(A\otimes (B\otimes C))
}
\end{equation}
\begin{equation}\label{1403131301}
\begin{array}{ccc}
\xymatrix@C=1.2cm@R=1.2cm{
I\otimes FA\ar[r]^-{\omega\phantom{.}\!\otimes\phantom{.}\! FA}\ar@/_0.8pc/[rd]_{l_{FA}} & FI \otimes FA\ar[r]^-{\zeta_{I, A}} & F(I\otimes A)\ar@/^0.8pc/[ld]^{Fl_{A}} \\
& FA &
} &
&
\xymatrix@C=1.2cm@R=1.2cm{
F(A\otimes I)\ar@/_0.8pc/[rd]_{Fr_{A}} & FA\otimes FI\ar[l]_-{\zeta_{A,I}} & FA\otimes I\ar[l]_-{FA\phantom{.}\!\otimes\phantom{.}\! \omega}\ar@/^0.8pc/[ld]^{r_{FA}}\\
& FA &
}
\end{array}
\end{equation}
\item $\mu^{\scriptscriptstyle{F}}:(F,\zeta,\omega)\cdot(F,\zeta,\omega)\longrightarrow (F,\zeta,\omega)$ and $\eta^{\scriptscriptstyle{F}}:(1_{\mathcal{C}}, 1_{\otimes}, 1_{\delta_{I}})\longrightarrow (F,\zeta,\omega)$ are lax natural transformations, the adjective lax adds, to the naturality, the following commutative diagrams
\begin{equation}\label{1403131208}
\begin{array}{ccc}
\xymatrix@C=1.5cm@R=1.5cm{
\otimes (FF\times FF) \ar[r]^-{\zeta (F\times F)}\ar[d]_{\otimes(\mu^{\scriptscriptstyle{F}}\times \mu^{\scriptscriptstyle{F}})} & F\otimes (F\times F)\ar[r]^-{F\zeta } & FF\otimes \ar[d]^{\mu^{\scriptscriptstyle{F}}\otimes}\\
\otimes (F\times F)\ar[rr]_{\zeta} & & F\otimes
} &
\xy<1cm,0cm>
\POS (0,-20.5) *+{,},
\endxy
&
\xymatrix@C=1.15cm@R=1.5cm{
\delta_{I} \ar[r]^-{\omega}\ar@/_1pc/[rrd]_{\omega} & F\delta_{I}\ar[r]^-{F\omega} & FF\delta_{I}\ar[d]^{\mu^{\scriptscriptstyle{F}}\delta_{I}}\\
& & F\delta_{I}
}
\end{array}
\end{equation}
\noindent \&
\begin{equation}\label{1403131209}
\begin{array}{ccc}
\xy
\POS (-22, 0) *+{\otimes} = "a",
\POS (22, 0) *+{\otimes} = "d",
\POS (22, -20) *+{F\otimes} = "c",
\POS (-22, -20) *+{\otimes(F\times F)} = "b",
\POS "a" \ar^{1_{\otimes}} "d",
\POS "d" \ar^{\eta^{\scriptscriptstyle{F}}\otimes} "c",
\POS "a" \ar_{\otimes(\eta^{\scriptscriptstyle{F}}\times\eta^{\scriptscriptstyle{F}})} "b",
\POS "b" \ar_{\zeta} "c",
\endxy
&
\xy<1cm,0cm>
\POS (0,-20.5) *+{,},
\endxy
&
\xy<1cm,0cm>
\POS (0, -4)
\xymatrix@C=1.2cm@R=1cm{
\delta_{I}\ar@/_0.7pc/[rd]_{\omega}\ar[r]^{1_{\delta_{I}}}& \delta_{I}\ar[d]^{\eta^{\scriptscriptstyle{F}}\delta_{I}}\\
& F\delta_{I}
}
\endxy
\end{array}
\end{equation}
\end{enumerate}
\end{1403061850}
\newtheorem{1403131236}[1403061850]{Note}
\begin{1403131236}
Necessarily $\omega(0)= \eta^{\scriptscriptstyle{F}}\! I$.
\end{1403131236}
Since the natural transformation $\omega$ has only one component then this
natural transformation and its component will be denoted indistinctly with
$\omega$, for example $\omega = \omega(0) = \eta^{\scriptscriptstyle{F}}\! I$.\\
We are going to make use of the isomorphism \eqref{1311042204}. The result we want to obtain using this isomorphism is the following.
\newtheorem{1311042045}[1403061850]{Theorem}
\begin{1311042045}
There is a bijective correspondance between the following structures
\begin{enumerate}
\item [1.-] Colax monads $((F, \zeta, \omega),\mu^{\scriptscriptstyle{F}}, \eta^{\scriptscriptstyle{F}})$, for the
monoidal structure $(\mathcal{C}, \otimes, I, a, l, r)$.
\item [2.-] Morphims and transformations of monads of the form
\begin{eqnarray*}
(\otimes, \zeta)&:& (\mathcal{C}\times\mathcal{C}, F\times F)\longrightarrow (\mathcal{C}, F),\\
(\delta_{I}, \omega)&:&(\mathbf{1}\:, 1_{\mathbf{1}}\:)\longrightarrow (\mathcal{C},F)\\
a&:&(\otimes\cdot(\otimes\times
\mathcal{C}),\zeta(\otimes\times\mathcal{C})\circ\otimes(\zeta\times
F))\longrightarrow(\otimes\cdot (\mathcal{C}\times\otimes)\cdot a_{\mathcal{C}}, \zeta(\mathcal{C}\times\otimes)a_{\mathcal{C}}\circ\otimes(F\times\zeta)a_{\mathcal{C}})\\
&&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad:((\mathcal{C}\times\mathcal{C})\times\mathcal{C}, (F\times F)\times F)\longrightarrow (\mathcal{C}, F), \\
l&:& (\otimes\cdot(\delta_{I}\times \mathcal{C})\cdot l_{\mathcal{C}}^{-1},\zeta\phantom{.}\!(\delta_{I}\times\mathcal{C})\phantom{.}\! l^{-1}_{\mathcal{C}}\circ
\otimes(\omega\times F)\phantom{.}\! l^{-1}_{\mathcal{C}} )\longrightarrow (1_{\mathcal{C}}, 1_{F}):(\mathcal{C}, F)\longrightarrow (\mathcal{C}, F),\\
r&:&(\otimes\cdot(\mathcal{C}\times\delta_{I})\cdot r_{\mathcal{C}}^{-1},
\zeta\phantom{.}\!(\mathcal{C}\times\delta_{I})\phantom{.}\! r^{-1}_{\mathcal{C}}\circ\otimes(F\times\omega)\phantom{.}\! r^{-1}_{\mathcal{C}}
)\longrightarrow (1_{\mathcal{C}}, 1_{F}):(\mathcal{C}, F)\longrightarrow (\mathcal{C}, F).\\
\end{eqnarray*}
\item [3.-] Monoidal structures for the Kleisli category $(\mathcal{C}_{\scriptscriptstyle{F}}, \widetilde{\otimes} , \tilde{I})$ such that the following diagrams of arrows and surfaces commute
\begin{equation}\label{1405081433}
\begin{array}{cc}
\xy<1cm,0cm>
\POS (0, 10) *+{(a)},
\POS (0, 0) *+{\mathcal{C}\times\mathcal{C}} = "c1",
\POS (25, 0) *+{\mathcal{C}} = "c2",
\POS (0, -20) *+{\mathcal{C}_{\scriptscriptstyle{F}}\times\mathcal{C}_{\scriptscriptstyle{F}}} = "d1",
\POS (25, -20) *+{\mathcal{C}_{\scriptscriptstyle{F}}} = "d2",
\POS (18, -10) = "a",
\POS (10, -18) = "b",
\POS "c1" \ar^-{\otimes} "c2",
\POS "c1" \ar_{G_{\scriptscriptstyle{F}}\times G_{\scriptscriptstyle{F}}} "d1",
\POS "c2" \ar^{G_{\scriptscriptstyle{F}}} "d2",
\POS "d1" \ar_-{\widetilde{\otimes}} "d2",
\endxy &
\xy<1cm,0cm>
\POS (0, 10) *+{(b)},
\POS (0,0) *+{\mathbf{1}} = "c1",
\POS (25,0) *+{\mathcal{C}} = "c2",
\POS (25,-20) *+{\mathcal{C}^{\scriptscriptstyle{F}}} = "d2",
\POS (0,-20) *+{\mathbf{1}_{1_{\mathbf{1}}}} = "d1",
\POS "c1" \ar^-{\delta_{I}} "c2",
\POS "c1" \ar_{G_{1_{\mathbf{1}}}} "d1",
\POS "c2" \ar^{G^{\scriptscriptstyle{F}}} "d2",
\POS "d1" \ar_-{\delta_{\tilde{I}}} "d2",
\endxy
\end{array}
\end{equation}
\begin{equation*}\label{1405072237}
\begin{array}{ccc}
\xy<1cm,0cm>
\POS (0, 0) *+{\mathcal{C}^{3}} = "c1",
\POS (7, 0) *+{\phantom{(\mathcal{C})}}= "c3",
\POS (30, 0) *+{\mathcal{C}} = "c2",
\POS (11, 4) = "a",
\POS (11, -4) = "b",
\POS (11,-36) = "c",
\POS (11,-44) = "d",
\POS (0,-40) *+{(\mathcal{C}_{\scriptscriptstyle{F}})^{3}} = "d1",
\POS (30,-40) *+{\mathcal{C}_{\scriptscriptstyle{F}}} = "d2",
\POS "c1" \ar@/^1.5pc/^{\otimes\cdot (\otimes\times \mathcal{C})} "c2",
\POS "c1" \ar@/_1.5pc/_{\otimes\cdot (\mathcal{C}\times\otimes)\cdot a_{\mathcal{C}}} "c2",
\POS "a" \ar^{\phantom{11}a} "b",
\POS "d1" \ar@/^1.5pc/^{\widetilde{\otimes}\cdot (\widetilde{\otimes}\times \mathcal{C}_{\scriptscriptstyle{F}})} "d2",
\POS "d1" \ar@/_1.5pc/_{\widetilde{\otimes}\cdot (\mathcal{C}_{\scriptscriptstyle{F}}\times\tilde{\otimes})\cdot a_{\mathcal{C}_{\scriptscriptstyle{F}}}} "d2",
\POS "c" \ar^{\phantom{11}\tilde{a}} "d",
\POS "c1" \ar_{(G_{\scriptscriptstyle{F}})^{3}} "d1",
\POS "c2" \ar^{G_{\scriptscriptstyle{F}}} "d2",
\endxy &
\xy<1cm,0cm>
\POS (0, 0) *+{\mathcal{C}} = "c1",
\POS (7, 0) *+{\phantom{(\mathcal{C})}}= "c3",
\POS (30, 0) *+{\mathcal{C}} = "c2",
\POS (11, 4) = "a",
\POS (11, -4) = "b",
\POS (11,-36) = "c",
\POS (11,-44) = "d",
\POS (0,-40) *+{\mathcal{C}_{\scriptscriptstyle{F}}} = "d1",
\POS (30,-40) *+{\mathcal{C}_{\scriptscriptstyle{F}}} = "d2",
\POS "c1" \ar@/^1.5pc/^{\otimes\cdot (\delta_{I}\times \mathcal{C})\cdot l^{-1}_{\mathcal{C}}} "c2",
\POS "c1" \ar@/_1.5pc/_{1_{\mathcal{C}}} "c2",
\POS "a" \ar^{\phantom{12}l} "b",
\POS "d1" \ar@/^1.5pc/^{\widetilde{\otimes}\cdot (\delta_{\tilde{I}}\times
\mathcal{C}_{\scriptscriptstyle{F}})\cdot l^{-1}_{\mathcal{C}_{\scriptscriptstyle{F}}}} "d2",
\POS "d1" \ar@/_1.5pc/_{1_{\mathcal{C}_{\scriptscriptstyle{F}}}} "d2",
\POS "c" \ar^{\phantom{12}\tilde{l}} "d",
\POS "c1" \ar_{G_{\scriptscriptstyle{F}}} "d1",
\POS "c2" \ar^{G_{\scriptscriptstyle{F}}} "d2",
\endxy
&
\xy<1cm,0cm>
\POS (0, 0) *+{\mathcal{C}} = "c1",
\POS (7, 0) *+{\phantom{(\mathcal{C})}}= "c3",
\POS (30, 0) *+{\mathcal{C}} = "c2",
\POS (11, 4) = "a",
\POS (11, -4) = "b",
\POS (11,-36) = "c",
\POS (11,-44) = "d",
\POS (0,-40) *+{\mathcal{C}_{\scriptscriptstyle{F}}} = "d1",
\POS (30,-40) *+{\mathcal{C}_{\scriptscriptstyle{F}}} = "d2",
\POS "c1" \ar@/^1.5pc/^{\otimes\cdot (\mathcal{C}\times\delta_{I})\cdot r^{-1}_{\mathcal{C}}} "c2",
\POS "c1" \ar@/_1.5pc/_{1_{\mathcal{C}}} "c2",
\POS "a" \ar^{\phantom{12}r} "b",
\POS "d1" \ar@/^1.5pc/^{\widetilde{\otimes}\cdot
(\mathcal{C}_{\scriptscriptstyle{F}}\times\delta_{\tilde{I}})\cdot r^{-1}_{\mathcal{C}_{\scriptscriptstyle{F}}}} "d2",
\POS "d1" \ar@/_1.5pc/_{1_{\mathcal{C}_{\scriptscriptstyle{F}}}} "d2",
\POS "c" \ar^{\phantom{12}\tilde{r}} "d",
\POS "c1" \ar_{G_{\scriptscriptstyle{F}}} "d1",
\POS "c2" \ar^{G_{\scriptscriptstyle{F}}} "d2",
\endxy
\end{array}
\end{equation*}
\end{enumerate}
\end{1311042045}
\emph{Proof}:\\
$1\Rightarrow 2$)\\
Consider a lax monad $((F, \zeta, \omega),\mu^{\scriptscriptstyle{F}}, \eta^{\scriptscriptstyle{F}})$ for the
monoidal category $(\mathcal{C}, \otimes, I)$. In particular, $\mu^{\scriptscriptstyle{F}}$ and
$\eta^{\scriptscriptstyle{F}}$ are natural lax monoidal transformations. Therefore, the
commutativity of the first diagram in \eqref{1403131208} and the first one in
\eqref{1403131209} is equivalent to the condition that the following be a
monad morphism $(\otimes, \zeta):(\mathcal{C}\times\mathcal{C} , F\times F)\longrightarrow (\mathcal{C},
F)$.\\
The commutativity condition on the second diagrams in \eqref{1403131208} and
\eqref{1403131209} is equivalent to the condition for the following be a monad
morphism $(\delta_{I},\omega): (\mathbf{1}\:,
1_{\mathbf{1}}\:)\longrightarrow (\mathcal{C},F) $. \\
Since $(\otimes, \zeta)$ is a morphism of monads so are $(\otimes\cdot
(\otimes\times \mathcal{C}),\:\zeta(\otimes\times
\mathcal{C})\circ\otimes(\zeta\times F))$ and $(\otimes\cdot
(\mathcal{C}\times\otimes)\cdot
a_{\mathcal{C}},\:\zeta(\mathcal{C}\times\otimes)a_{\mathcal{C}}\circ\otimes(F\times\zeta)a_{\mathcal{C}})$. Yet
again, since
$((F, \zeta), \mu^{\scriptscriptstyle{F}}, \eta^{\scriptscriptstyle{F}})$ is a lax monad over the monoidal category
$(\mathcal{C},\otimes, I, a, l, r)$, then a commutative diagram like
\eqref{1403061941} takes place. Therefore the following is a 2-cell in $\mathbf{Mnd}^{\scriptscriptstyle{\bullet}}({}_{2}Cat)$.
\begin{equation*}
\xy<1cm,0cm>
\POS (0, 0) *+{((\mathcal{C}\times\mathcal{C})\times\mathcal{C}, (F\times F)\times F)} = "c1",
\POS (7, 0) *+{\phantom{(\mathcal{C})}}= "c3",
\POS (50, 0) *+{(\mathcal{C}, F)} = "c2",
\POS (27, 5) = "a",
\POS (27, -5) = "b",
\POS "c3" \ar@/^2pc/^{(\otimes\cdot (\otimes\times \mathcal{C}),\:\zeta(\otimes\times \mathcal{C})\circ\otimes(\zeta\times F))} "c2",
\POS "c3" \ar@/_2pc/_{(\otimes\cdot (\mathcal{C}\times\otimes)\cdot a_{\mathcal{C}},\:\zeta(\mathcal{C}\times\otimes)a_{\mathcal{C}}\circ\otimes(F\times\zeta)a_{\mathcal{C}})} "c2",
\POS "a" \ar^{\phantom{1}a} "b",
\endxy
\end{equation*}\\
Since $(\otimes, \zeta)$ and $(\delta_{I}, \omega)$ are monad morphisms so is $(\otimes\cdot(\delta_{I}\times \mathcal{C})\cdot l_{\mathcal{C}}^{-1},\zeta\phantom{.}\!(\delta_{I}\times\mathcal{C})\phantom{.}\! l^{-1}_{\mathcal{C}}\circ
\otimes\phantom{.}\!(\omega\times F)\phantom{.}\! l^{-1}_{\mathcal{C}})$ and taking into
account the commutativity of the diagram (\ref{1403131301}a), we can state
that the following is a 2-cell in $\mathbf{Mnd}^{\scriptscriptstyle{\bullet}}({}_{2}Cat)$
\begin{equation*}
\xy<1cm,0cm>
\POS (0, 0) *+{(\mathcal{C}, F)} = "c1",
\POS (40, 0) *+{(\mathcal{C}, F)} = "c2",
\POS (13, 3.5) = "a",
\POS (13, -3.5) = "b",
\POS "c1" \ar@/^1.5pc/^{(\otimes\cdot (\delta_{I}\times\mathcal{C})\cdot l^{-1}_{\mathcal{C}}, \:\zeta\phantom{.}\!(\delta_{I}\times\mathcal{C})\phantom{.}\! l^{-1}_{\mathcal{C}}\circ\phantom{.}\!
\otimes\phantom{.}\!(\omega\times F)\phantom{.}\! l^{-1}_{\mathcal{C}})} "c2",
\POS "c1" \ar@/_1.5pc/_{(1_{\mathcal{C}},1_{F})} "c2",
\POS "a" \ar^{\phantom{111}l} "b",
\endxy
\end{equation*}
In the very same way, the following is a $2$-cell of monads,
$r:(\otimes\cdot(\mathcal{C}\times\delta_{I})\cdot
r^{-1}_{\mathcal{C}},\zeta(\mathcal{C}\times\delta_{I})\phantom{.}\!
r^{-1}_{\mathcal{C}}\circ\otimes(F\times\omega)\phantom{.}\! r^{-1}_{\mathcal{C}}
)$\\
2 $\Rightarrow$ 1) The previous assertions can be reverted.\\
2 $\Rightarrow$ 3)\\
Suppose we have a monad morphism $(\otimes, \zeta)$. Use the isomorphism \eqref{1311042204}, with $(\mathcal{D}, H,\mu^{\scriptscriptstyle{H}},\eta^{\scriptscriptstyle{H}}) = (\mathcal{C}\times\mathcal{C}, F\times F, \mu^{\scriptscriptstyle{F}}\times \mu^{\scriptscriptstyle{F}}, \eta^{\scriptscriptstyle{F}}\times \eta^{\scriptscriptstyle{F}})$ and $L\dashv R = G_{\scriptscriptstyle{F}}\dashv V_{\scriptscriptstyle{F}}$ to get an associated morphism of adjunctions $(\otimes, \otimes_{\zeta}): G_{\scriptscriptstyle{F}}\times G_{\scriptscriptstyle{F}}\dashv V_{\scriptscriptstyle{F}}\times V_{\scriptscriptstyle{F}}\longrightarrow G_{\scriptscriptstyle{F}}\dashv V_{\scriptscriptstyle{F}}$, such that a diagram like (\ref{1405081433}a) commutes. According to the definition of $\Psi_{\scriptscriptstyle{K}}$, the functor $\otimes_{\zeta}$ acts as follows. On objects,
\begin{equation*}\label{1403131226}
\otimes_{\zeta}(X,Y) = \otimes(X, Y) = X\otimes Y,
\end{equation*}
\noindent and on morphisms,
\begin{equation*}\label{1403131227}
\otimes_{\zeta}(x^{\sharp},y^{\sharp}) = (\zeta_{C_{x^{\sharp}},C_{y^{\sharp}}}\cdot (x\otimes y))^{\sharp}
\end{equation*}\\
\noindent where $C_{x^{\sharp}}$ is codomain of the morphism $x^{\sharp}$ for example. We rename $\otimes_{\zeta}$ as $\widetilde{\otimes}$.\\
For the monad morphism, $(\delta_{I},\omega):(\mathbf{1},1_{\mathbf{1}})\longrightarrow (\mathcal{C}, F)$, use the mentioned
isomorphism with $(\mathcal{D}, H,\mu^{\scriptscriptstyle{H}},\eta^{\scriptscriptstyle{H}}) = (\mathbf{1}\:,1_{\mathbf{1}}, 1_{1_{\mathbf{1}}}, 1_{1_{\mathbf{1}}})$,
\emph{i.e.} the trivial monad on the category $\mathbf{1}$, and $L\dashv R = G_{\scriptscriptstyle{F}}\dashv V_{\scriptscriptstyle{F}}$. Therefore, there exists
an adjunction morphism $(\delta_{I}, \lbrack\delta_{I}\rbrack_{\omega}): G_{1_{\mathbf{1}}}\dashv V_{1_{\mathbf{1}}}\longrightarrow G_{\scriptscriptstyle{F}}\dashv V_{\scriptscriptstyle{F}}$. According to the 2-functor $\Psi_{\scriptscriptstyle{K}}$, the functor $\lbrack\delta_{I}\rbrack_{\omega}:\mathbf{1}\longrightarrow \mathcal{C}_{\scriptscriptstyle{F}}$, acts in the following way
\begin{equation*}\label{1403131233}
\lbrack\delta_{I}\rbrack_{\omega}(0) = \delta_{I}(0) = I
\end{equation*}
\begin{equation*}\label{1403131234}
\lbrack\delta_{I}\rbrack_{\omega}(1_{0}) = (\omega(0)\cdot \delta_{I}(1_{0}))^{\sharp} = (\eta^{\scriptscriptstyle{F}}I)^{\sharp}
\end{equation*}\\
That is to say $\lbrack\delta_{I}\rbrack_{\omega} = \delta_{\tilde{I}}:\mathbf{1}_{1_{\mathbf{1}}}\longrightarrow \mathcal{C}_{\scriptscriptstyle{F}}$, where $\tilde{I} = I$.\\
Suppose that we have the following $2$-cell in $\mathbf{Mnd}$,\\
\begin{equation*}
\xy<1cm,0cm>
\POS (0, 0) *+{((\mathcal{C}\times\mathcal{C})\times\mathcal{C}, (F\times F)\times F)} = "c1",
\POS (7, 0) *+{\phantom{(\mathcal{C})}}= "c3",
\POS (50, 0) *+{(\mathcal{C}, F)} = "c2",
\POS (27, 5) = "a",
\POS (27, -5) = "b",
\POS "c3" \ar@/^2pc/^{(\otimes\cdot (\otimes\times \mathcal{C}),\:\zeta(\otimes\times \mathcal{C})\circ\otimes(\zeta\times F))} "c2",
\POS "c3" \ar@/_2pc/_{(\otimes\cdot (\mathcal{C}\times\otimes)\cdot
a_{\mathcal{C}},
\:\zeta(\mathcal{C}\times\otimes)a_{\mathcal{C}}\circ\otimes(F\times\zeta) a_{\mathcal{C}})} "c2",
\POS "a" \ar^{\phantom{1}a} "b",
\endxy
\end{equation*}\\
In order to continue with the calculations, we use the following notation, for
the sake of simplification\\
\begin{eqnarray*}
{}^{\cdot}\zeta^{2} &:=& \zeta(\otimes\times \mathcal{C})\circ\otimes(\zeta\times F),\\
{}_{\cdot}\zeta^{2} &:=& \zeta(\mathcal{C}\times\otimes)a_{\mathcal{C}}\circ\otimes(F\times\zeta)a_{\mathcal{C}}, \\
(\cdot)^{3} &:=& (\cdot\times\cdot)\times\cdot\:.
\end{eqnarray*}
According to the isomorphism of categories given by \eqref{1311042204}, to the previous $2$-cell in $\mathbf{Mnd}^{\scriptscriptstyle{\bullet}}({}_{2}Cat)$ corresponds a $2$-cell, $(\alpha_{a}, \beta_{a})$ in $\mathbf{Adj}_{\scriptscriptstyle{L}}({}_{2}Cat)$, where $\alpha_{a} = a$ and we rename $\beta_{a}= \tilde{a}$ and such that
\begin{equation*}
\xy<1cm,0cm>
\POS (0, 0) *+{\mathcal{C}^{3}} = "c1",
\POS (7, 0) *+{\phantom{(\mathcal{C})}}= "c3",
\POS (40, 0) *+{\mathcal{C}} = "c2",
\POS (14, 4) = "a",
\POS (14, -4) = "b",
\POS (14,-36) = "c",
\POS (14,-44) = "d",
\POS (0,-40) *+{(\mathcal{C}_{\scriptscriptstyle{F}})^{3}} = "d1",
\POS (40,-40) *+{\mathcal{C}_{\scriptscriptstyle{F}}} = "d2",
\POS "c1" \ar@/^1.7pc/^{\otimes\cdot (\otimes\times \mathcal{C})} "c2",
\POS "c1" \ar@/_1.7pc/_{\otimes\cdot (\mathcal{C}\times\otimes)\cdot a_{\mathcal{C}}} "c2",
\POS "a" \ar^{\phantom{12}a} "b",
\POS "d1" \ar@/^1.7pc/^{[\otimes\cdot (\otimes\times \mathcal{C})]_{{}^{\cdot}\!\zeta^{2}}} "d2",
\POS "d1" \ar@/_1.7pc/_{[\otimes\cdot (\mathcal{C}\times\otimes)\cdot a_{\mathcal{C}}]_{{}_{\cdot}\zeta^{2}}} "d2",
\POS "c" \ar^{\phantom{12}\widetilde{a}} "d",
\POS "c1" \ar@<-4pt>_{(G_{\scriptscriptstyle{F}})^{3}} "d1",
\POS "d1" \ar@<-1pt>_{(V_{\scriptscriptstyle{F}})^{3}} "c1",
\POS "c2" \ar@<-2pt>_{G_{\scriptscriptstyle{F}}} "d2",
\POS "d2" \ar@<-3pt>_{V_{\scriptscriptstyle{F}}} "c2",
\endxy
\end{equation*}\\
It can be show that
\begin{eqnarray*}
\lbrack\otimes\cdot (\otimes\times \mathcal{C})\rbrack_{{}^{\cdot}\zeta^{2}} &=& \widetilde{\otimes}\cdot(\widetilde{\otimes}\times \mathcal{C}_{\scriptscriptstyle{F}})\\
\lbrack\otimes\cdot (\mathcal{C}\times\otimes)\cdot a_{\mathcal{C}} \rbrack_{{}_{\cdot}\zeta^{2}} &=& \widetilde{\otimes}\cdot (\mathcal{C}_{\scriptscriptstyle{F}}\times\widetilde{\otimes})\cdot a_{\mathcal{C}_{\scriptscriptstyle{F}}}
\end{eqnarray*}\\
Therefore, we have a natural transformation $\widetilde{a}:\widetilde{\otimes}\cdot(\widetilde{\otimes}\times\mathcal{C}_{\scriptscriptstyle{F}})\longrightarrow \widetilde{\otimes}\cdot(\mathcal{C}_{\scriptscriptstyle{F}}\times\widetilde{\otimes})\cdot a_{\mathcal{C}_{\scriptscriptstyle{F}}}$ that will
be part of a monoidal structure on
$\mathcal{C}_{\scriptscriptstyle{F}}$. According to the 2-functor $\Psi_{\scriptscriptstyle{K}}$, the component of $\widetilde{a}$ at $((X, Y), Z)$ is\\
\begin{equation*}\label{1405091636}
\widetilde{a}_{X,Y,Z} = (\eta^{\scriptscriptstyle{F}}(X\otimes (Y\otimes Z))\cdot a_{X,Y,Z})^{\sharp}
\end{equation*}\\
Suppose that we have a 2-cell in $\mathbf{Mnd}^{\scriptscriptstyle{\bullet}}({}_{2}Cat)$ of the form $l:\big(\otimes\cdot(\delta_{I}\times\mathcal{C})\cdot l^{-1}_{\mathcal{C}}, \zeta(\delta_{I}\times\mathcal{C})\phantom{.}\! l^{-1}_{\mathcal{C}}\circ\otimes\phantom{.}\!(\omega\times F)\phantom{.}\! l^{-1}_{\mathcal{C}}\big)\longrightarrow (1_{\mathcal{C}}, 1_{\scriptscriptstyle{F}}):(\mathcal{C}, F)\longrightarrow (\mathcal{C}, F)$. \\
Therefore, we obtain a natural transformation $\tilde{l}:\widetilde{\otimes}\cdot (\delta_{\tilde{I}}\times\mathcal{C}_{\scriptscriptstyle{F}})\cdot
l^{-1}_{\mathcal{C}_{\scriptscriptstyle{F}}}\longrightarrow 1_{\mathcal{C}_{\scriptscriptstyle{F}}}$. Using the
definition of the functor $\Psi_{\scriptscriptstyle{K}}$ on the 2-cell $l$, the component of $\tilde{l}$, on the object $X$ in $\mathcal{C}_{\scriptscriptstyle{F}}$, is \\
\begin{equation}\label{1405091630}
\tilde{l}X = (\eta^{\scriptscriptstyle{F}}X\cdot lX)^{\sharp}
\end{equation}\\
Similarly, for the monad morphism
$r:(\otimes\cdot(\mathcal{C}\times\delta_{I})\cdot r^{-1}_{\mathcal{C}}, \zeta
(\mathcal{C}\times\delta_{I})r^{-1}_{\mathcal{C}}\circ\otimes(F\times\omega)r^{-1}_{\mathcal{C}})\longrightarrow
(1_{\mathcal{C}}, 1_{\scriptscriptstyle{F}}):(\mathcal{C}, F)\longrightarrow(\mathcal{C}, F)$, we
obtain a natural transformation $\tilde{r}:
\otimes_{\zeta}\cdot(\mathcal{C}_{\scriptscriptstyle{F}}\times\delta_{\tilde{I}})\cdot
r^{-1}_{\mathcal{C}_{\scriptscriptstyle{F}}}\longrightarrow 1_{\mathcal{C}_{\scriptscriptstyle{F}}}:
\mathcal{C}_{\scriptscriptstyle{F}}\longrightarrow \mathcal{C}_{\scriptscriptstyle{F}}$.\\
The proof of the coherence conditions are left to the reader.\\
In summary, $(\mathcal{C}_{\scriptscriptstyle{F}}, \widetilde{\otimes}, \tilde{I}, \tilde{a}, \tilde{l}, \tilde{r})$ has a monoidal structure on $\mathcal{C}_{\scriptscriptstyle{F}}$.\\
3 $\Rightarrow$ 2)\\
Using the isomorphism, given by \eqref{1311042204}, we get the return of the
proof. For example, the image, under $\Phi_{\scriptscriptstyle{K}}$, for the 2-cell of adjunctions $(a,\tilde{a}):
(\otimes\cdot(\otimes\times\mathcal{C}),
\otimes\cdot(\mathcal{C}\times\otimes)\cdot a_{\mathcal{C}})\longrightarrow
(\widetilde{\otimes}\cdot(\widetilde{\otimes}\times\mathcal{C}_{\scriptscriptstyle{F}}),\widetilde{\otimes}\cdot(\mathcal{C}_{\scriptscriptstyle{F}}\times\widetilde{\otimes})\cdot
a_{\mathcal{C}_{\scriptscriptstyle{F}}}): (G_{\scriptscriptstyle{F}}\times G_{\scriptscriptstyle{F}})\times G_{\scriptscriptstyle{F}}\dashv (V_{\scriptscriptstyle{F}}\times
V_{\scriptscriptstyle{F}})\times V_{\scriptscriptstyle{F}}$ is
\begin{eqnarray*}
\Phi_{\scriptscriptstyle{K}}((a,\tilde{a})) &=& a:(\otimes,
\pi_{\otimes})(\otimes\times\mathcal{C},\pi_{\otimes\times\mathcal{C}})\longrightarrow
(\otimes, \pi_{\otimes})(\mathcal{C}\times\otimes,
\pi_{\mathcal{C}\times\otimes})(a_{\mathcal{C}},
\pi_{a_{\mathcal{C}}}):(\mathcal{C}^{3}, F^{3})\longrightarrow (\mathcal{C},
F)\\
&=& a:(\otimes,
\pi_{\otimes})(\otimes\times\mathcal{C},\pi_{\otimes}\times
F)\longrightarrow (\otimes,
\pi_{\otimes})(\mathcal{C}\times\otimes,F\times\pi_{\otimes})(a_{\mathcal{C}},
1_{\scriptscriptstyle{F}\times(\scriptscriptstyle{F}\times\scriptscriptstyle{F})\cdot a_{\mathcal{C}}}):\\
&& (\mathcal{C}^{3},
F^{3})\longrightarrow (\mathcal{C}, F)\\
&=&
a:\big(\otimes\cdot(\otimes\times\mathcal{C}), \pi_{\otimes}(\otimes\times\mathcal{C})\circ\otimes(\pi_{\otimes}\times
F)\big)\longrightarrow\\
&& \big(\otimes\cdot(\mathcal{C}\times\mathcal{C})\cdot a_{\mathcal{C}},\pi_{\otimes}(\mathcal{C}\times\otimes)a_{\mathcal{C}}\circ
\otimes(F\times\pi_{\otimes})a_{\mathcal{C}}\big):(\mathcal{C}^{3},
F^{3})\longrightarrow (\mathcal{C}, F)
\end{eqnarray*}\\
Note that $a_{\mathcal{C}}$ is a morphism of adjunctions. $\square$
\section{Liftings to the Eilenberg-Moore algebras \& Extensions to the Kleisli Categories}
This is probably the more explored section of all this article, a few examples
of the detailed proofs for the following statements are found in
\cite{bofr_haca_II} and \cite{tami_psdl}. In this section, we treated these statements only as direct
consequences of the isomorphisms of categories given by \eqref{1403191902} and \eqref{1311042204}.
\newtheorem{1405102218}{Theorem}[section]
\begin{1405102218}
There is a bijective correspondence, natural in $(\mathcal{C}, F, \mu^{\scriptscriptstyle{F}}, \eta^{\scriptscriptstyle{F}})$ and $(\mathcal{D}, H, \mu^{\scriptscriptstyle{H}}, \eta^{\scriptscriptstyle{H}})$, between the following structures
\begin{enumerate}
\item [1.-] Liftings to the Eilenberg-Moore algebras, for the functor
$P:\mathcal{C}\longrightarrow \mathcal{D}$. That is to say, the following diagram commutes
\begin{equation*}
\xy<1cm,0cm>
\POS (0, 0) *+{\mathcal{C}} = "a21",
\POS (20, 0) *+{\mathcal{D}}= "a22",
\POS (0, 20) *+{\mathcal{C}^{\scriptscriptstyle{F}}} = "a11",
\POS (20,20) *+{\mathcal{D}^{\scriptscriptstyle{H}}} = "a12",
\POS "a11" \ar^{Q} "a12",
\POS "a21" \ar_{P} "a22",
\POS "a11" \ar_{U^{\scriptscriptstyle{F}}} "a21",
\POS "a12" \ar^{U^{\scriptscriptstyle{H}}} "a22",
\endxy
\end{equation*}
\item [2.-] Morphisms of monads $(P,\varphi) : (\mathcal{C}, F)\longrightarrow (\mathcal{D}, H)$. That is to say, a natural transformation $\varphi:HP\longrightarrow PF$, such that the following diagrams commute
\begin{equation*}
\begin{array}{cc}
\xy<1cm,0cm>
\POS (0,22) *+{HHP} = "a11",
\POS (30,22) *+{HPF} = "a12",
\POS (60,22) *+{PFF} = "a13",
\POS (0,0) *+{HP} = "a21",
\POS (60,0) *+{PF} = "a23",
\POS "a11" \ar^{H\varphi} "a12",
\POS "a12" \ar^{\varphi F} "a13",
\POS "a21" \ar_{\varphi} "a23",
\POS "a11" \ar_{\mu^{\scriptscriptstyle{H}}\! P} "a21",
\POS "a13" \ar^{P\mu^{\scriptscriptstyle{F}}} "a23"
\endxy &
\xy<1cm,0cm>
\POS (20,15) *+{P} = "a12",
\POS (0,0) *+{HP} = "a21",
\POS (40,0) *+{PF} = "a23",
\POS "a12" \ar_{\eta^{\scriptscriptstyle{H}}\! P} "a21",
\POS "a12" \ar^{P\eta^{\scriptscriptstyle{F}}} "a23",
\POS "a21" \ar_{\varphi} "a23",
\endxy
\end{array}
\end{equation*}\\
\end{enumerate}
\end{1405102218}
\newtheorem{1405102219}[1405102218]{Theorem}
\begin{1405102219}
There exists a bijective correspondence, natural in $(\mathcal{C}, F, \mu^{\scriptscriptstyle{F}}, \eta^{\scriptscriptstyle{F}})$ and $(\mathcal{D}, H, \mu^{\scriptscriptstyle{H}}, \eta^{\scriptscriptstyle{H}})$, between the following structures
\begin{enumerate}
\item [1.-] Extensions to the Kleisli categories, for the functor $P:\mathcal{C}\longrightarrow \mathcal{D}$. That is to say, the following diagram commutes
\begin{equation*}
\xy<1cm,0cm>
\POS (0, 0) *+{\mathcal{C}_{\scriptscriptstyle{F}}} = "a21",
\POS (20, 0) *+{\mathcal{D}_{\scriptscriptstyle{H}}}= "a22",
\POS (0, 20) *+{\mathcal{C}} = "a11",
\POS (20,20) *+{\mathcal{D}} = "a12",
\POS "a11" \ar^{P} "a12",
\POS "a21" \ar_{Q} "a22",
\POS "a11" \ar_{G_{\scriptscriptstyle{F}}} "a21",
\POS "a12" \ar^{G_{\scriptscriptstyle{H}}} "a22",
\endxy
\end{equation*}
\item [2.-] Morphisms of monads $(P, \varphi):(\mathcal{C}, F)\longrightarrow (\mathcal{D}, H)$. That is to say, a natural transformation $\varphi:PF\longrightarrow HP$
\begin{equation*}
\begin{array}{cc}
\xy<1cm,0cm>
\POS (0,22) *+{PFF} = "a11",
\POS (30,22) *+{HPF} = "a12",
\POS (60,22) *+{HHP} = "a13",
\POS (0,0) *+{PF} = "a21",
\POS (60,0) *+{HP} = "a23",
\POS "a11" \ar^{\varphi F} "a12",
\POS "a12" \ar^{H\varphi } "a13",
\POS "a21" \ar_{\varphi} "a23",
\POS "a11" \ar_{P\mu^{\scriptscriptstyle{F}}} "a21",
\POS "a13" \ar^{\mu^{\scriptscriptstyle{H}}P} "a23"
\endxy &
\xy<1cm,0cm>
\POS (20,15) *+{P} = "a12",
\POS (0,0) *+{PF} = "a21",
\POS (40,0) *+{HP} = "a23",
\POS "a12" \ar_{P\eta^{\scriptscriptstyle{F}}} "a21",
\POS "a12" \ar^{\eta^{\scriptscriptstyle{H}}\! P} "a23",
\POS "a21" \ar_{\varphi} "a23",
\endxy
\end{array}
\end{equation*}\\
\end{enumerate}
\end{1405102219}
\section{Actions on the Kleisli Category}
\subsection{Categorical Actions}
In this section we give the definition of a \emph{categorical action}.
\newtheorem{1407221804}{Definition}[section]
\begin{1407221804}
Let $(\mathcal{C}, \otimes, I)$ be a monoidal category. A left
$\mathcal{C}$-action on the category $\mathcal{B}$ is a functor
$\boxtimes:\mathcal{C}\times\mathcal{B}\longrightarrow \mathcal{B}$ together with natural
transformations
$\nu:\boxtimes(\otimes\times\mathcal{B})\longrightarrow\boxtimes(\mathcal{C}\times\otimes)\alpha_{\ast}:
(\mathcal{C}\times\mathcal{C})\times \mathcal{B}\longrightarrow \mathcal{B}$
and $j:\boxtimes(\delta_{I}\times\mathcal{B})l^{-1}_{\scriptscriptstyle{\mathcal{C}}}\longrightarrow
1_{\mathcal{B}}:\mathcal{B}\longrightarrow \mathcal{B}$ such that they fulfill the following commutative diagrams, for
objects $C,C', C''$ in $\mathcal{C}$ and $B$ in $\mathcal{B}$,
\begin{equation*}
\xy<1cm,0cm>
\POS (0, 40) *+{\lbrack (C\otimes C')\otimes C''\rbrack\boxtimes B} = "a11",
\POS (70, 40) *+{(C\otimes C')\boxtimes(C''\boxtimes B)} = "a12",
\POS (0,20) *+{\lbrack C\otimes (C'\otimes C'')\rbrack\boxtimes B} = "a21",
\POS (0,0) *+{C\boxtimes\lbrack (C'\otimes C'')\boxtimes B\rbrack} = "a31",
\POS (70,0) *+{C\boxtimes\lbrack C'\boxtimes (C''\boxtimes B)\rbrack} = "a32",
\POS "a11" \ar^{\nu_{C\otimes C', C'', B}} "a12",
\POS "a12" \ar^{\nu_{C, C', C''\boxtimes B}} "a32",
\POS "a11" \ar_{\alpha_{C, C', C''}\boxtimes B} "a21",
\POS "a21" \ar_{\nu_{C, C'\otimes C'', B}} "a31",
\POS "a31" \ar_{C\boxtimes\nu_{C', C'', B}} "a32",
\endxy
\end{equation*}
\noindent and
\begin{equation*}
\begin{array}{ccc}
\xy<1cm,0cm>
\POS (0,20) *+{(C\otimes I)\boxtimes B} = "a11",
\POS (40,20)*+{C\boxtimes(I\boxtimes B)} = "a13",
\POS (20, 0)*+{C\boxtimes B} = "a22",
\POS "a11" \ar^{\nu_{C, I, B}} "a13",
\POS "a13" \ar^{C\phantom{.}\!\boxtimes j_{B}} "a22",
\POS "a11" \ar_{r_{C}\boxtimes B} "a22",
\endxy & &
\xy<1cm,0cm>
\POS (0,20) *+{(I\otimes C)\boxtimes B} = "a11",
\POS (40,20)*+{I\boxtimes(C\boxtimes B)} = "a13",
\POS (20, 0)*+{C\boxtimes B} = "a22",
\POS "a11" \ar^{\nu_{I, C, B}} "a13",
\POS "a13" \ar^{j_{C\phantom{.}\!\boxtimes B}} "a22",
\POS "a11" \ar_{l_{C}\boxtimes B} "a22",
\endxy
\end{array}
\end{equation*}
\end{1407221804}
\subsection{Strong Monads}
In this section we give the definition of a \emph{strong monad}.
\newtheorem{1405191543}{Definition}[section]
\begin{1405191543}
A \emph{right strong monad} $((F,\sigma^{r}), \mu^{\scriptscriptstyle{F}}, \eta^{\scriptscriptstyle{F}})$, on the monoidal category $(\mathcal{C}, \otimes, I)$, is a usual monad $(F, \mu^{\scriptscriptstyle{F}}, \eta^{\scriptscriptstyle{F}})$, on $\mathcal{C}$, with a natural transformation $\sigma^{r}:A \otimes FB\longrightarrow F(A\otimes B)$ such that the following diagrams commute
\begin{equation}\label{1405191712}
\begin{array}{cc}
\xy<1cm,0cm>
\POS (0,20) *+{A\otimes FFB} = "a11",
\POS (35,20) *+{F(A\otimes FB)} = "a12",
\POS (70,20) *+{FF(A\otimes B)} = "a13"
\POS (0,0) *+{A\otimes FB} = "a21",
\POS (70,0) *+{F(A\otimes B)} = "a23",
\POS (-12,25) *+{(a)},
\POS "a11" \ar^{\sigma^{r}_{A,FB}} "a12",
\POS "a12" \ar^{F\sigma^{r}_{\! A,B}} "a13",
\POS "a13" \ar^{\mu^{F}(A\otimes B)} "a23"
\POS "a11" \ar_{A\otimes\mu^{\scriptscriptstyle{F}}\! B} "a21",
\POS "a21" \ar_{\sigma^{r}} "a23",
\endxy &
\xy<1cm,0cm>
\POS (20,20) *+{A\otimes B} = "a12",
\POS (0,0) *+{A\otimes FB} = "a21",
\POS (40,0) *+{F(A\otimes B)} = "a23",
\POS (-5,25) *+{(b)},
\POS "a12" \ar^{\eta^{\scriptscriptstyle{F}}(A\otimes B)} "a23",
\POS "a12" \ar_{A\otimes\eta^{\scriptscriptstyle{F}}\! B} "a21",
\POS "a21" \ar_{\sigma^{r}} "a23",
\endxy
\end{array}
\end{equation}
\noindent and
\begin{equation}\label{1405191723}
\begin{array}{cc}
\xy<1cm,0cm>
\POS (0,20) *+{(A\otimes B)\otimes FC} = "a11",
\POS (90,20) *+{F((A\otimes B)\otimes C)} = "a13",
\POS (0,0) *+{A\otimes (B\otimes FC)} = "a21",
\POS (45,0) *+{A\otimes F(B\otimes C)} = "a22",
\POS (90,0) *+{F(A\otimes (B\otimes C))} = "a23"
\POS (0,30) *+{(a)},
\POS "a11" \ar^{\sigma^{r}_{A\otimes B, C}} "a13",
\POS "a13" \ar^{Fa_{A,B,C}} "a23",
\POS "a11" \ar_{a_{A,B,FC}} "a21",
\POS "a21" \ar_{A\otimes\sigma^{r}_{B, C}} "a22",
\POS "a22" \ar_{\sigma^{r}_{A,B\otimes C}} "a23",
\endxy &
\xy<1cm,0cm>
\POS (0, 20) *+{I\otimes FA} = "a11",
\POS (40,20) *+{F(I\otimes A)} = "a13",
\POS (20,0) *+{FA} = "a22",
\POS (0,30) *+{(b)},
\POS "a11" \ar^{\sigma^{r}_{I, A}} "a13",
\POS "a13" \ar^{Fl_{A}} "a22",
\POS "a11" \ar_{l_{FA}} "a22",
\endxy
\end{array}
\end{equation}
\end{1405191543}
\newtheorem{1405201903}[1405191543]{Definition}
\begin{1405201903}
A \emph{left strong monad} $((F,\sigma^{r}), \mu^{\scriptscriptstyle{F}}, \eta^{\scriptscriptstyle{F}})$ on a monoidal category $(\mathcal{C}, \otimes, I)$, is a usual monad $(F, \mu^{\scriptscriptstyle{F}},\eta^{\scriptscriptstyle{F}})$ on $\mathcal{C}$, together with a natural transformation $\sigma^{l}_{A,B}:FA\otimes B \longrightarrow F(A\otimes B)$ such that fulfills the commutativity of dual diagrams like \eqref{1405191712} and \eqref{1405191723}.
\end{1405201903}
In \cite{mepa_pame}, the author relates actions of a category $\mathcal{C}$
over its Kleisli category and strong monads. In this section, with the use of
the 2-adjunction of Kleisli, we rediscover this relation between certain actions on the Kleisli categories for a given monad $(\mathcal{C}, F, \mu^{\scriptscriptstyle{F}}, \eta^{\scriptscriptstyle{F}})$ with strong monads.
\newtheorem{1505191350}[1405191543]{Theorem}
\begin{1505191350}
There exists a bijection between the following structures
\begin{enumerate}
\item [1.-] Right strong monads $((F,\sigma^{r}),\mu^{\scriptscriptstyle{F}},\eta^{\scriptscriptstyle{F}})$ on the
monoidal category $(\mathcal{C}, \otimes, I, a, r, l)$.
\item [2.-] Morphisms and transformations of monads of the form
\begin{eqnarray*}
(\otimes, \sigma^{r}) &:& (\mathcal{C}\times\mathcal{C},
\mathcal{C}\times F)\longrightarrow (\mathcal{C}, F)\\
a &:& (\otimes\cdot(\otimes\times\mathcal{C}),
\sigma^{r}(\otimes\times\mathcal{C}))\longrightarrow
(\otimes\cdot(\mathcal{C}\times\otimes)\cdot a_{\mathcal{C}},
\sigma^{r}(\mathcal{C}\times\otimes)a_{\mathcal{C}}\circ
\otimes(\mathcal{C}\times\sigma^{r})a_{\mathcal{C}})\\
&:& ((\mathcal{C}\times\mathcal{C})\times\mathcal{C},
(\mathcal{C}\times\mathcal{C})\times F)\longrightarrow (\mathcal{C}, F)\\
l&:& (\otimes\cdot(\delta_{I}\times\mathcal{C})\cdot l^{-1}_{\mathcal{C}},
\phantom{.}\!\sigma^{r}(\delta_{I}\times\mathcal{C})\phantom{.}\!
l^{-1}_{\mathcal{C}})\longrightarrow (1_{\mathcal{C}}, 1_{\scriptscriptstyle{F}}):(\mathcal{C},
F)\longrightarrow (\mathcal{C}, F)
\end{eqnarray*}
\item [3.-] Left actions on the Kleisli category, $\mathcal{C}_{\scriptscriptstyle{F}}$, $\boxtimes:\mathcal{C}\times\mathcal{C}_{\scriptscriptstyle{F}}\longrightarrow \mathcal{C}_{\scriptscriptstyle{F}}$ such that the following diagrams of morphisms and surfaces commute
\begin{equation}\label{1405222031}
\xy<1cm,0cm>
\POS (0,20) *+{\mathcal{C}\times\mathcal{C}} = "a11",
\POS (20,20) *+{\mathcal{C}} = "a12",
\POS (0,0) *+{\mathcal{C}\times\mathcal{C}_{\scriptscriptstyle{F}}} = "a21",
\POS (20,0) *+{\mathcal{C}_{\scriptscriptstyle{F}}} = "a22",
\POS "a11" \ar^-{\otimes} "a12",
\POS "a12" \ar^{G_{\scriptscriptstyle{F}}} "a22",
\POS "a11" \ar_{\mathcal{C}\times G_{\scriptscriptstyle{F}}} "a21",
\POS "a21" \ar_-{\boxtimes} "a22",
\endxy
\end{equation}
\begin{equation}\label{1405202049}
\begin{array}{ccc}
\xy<1cm,0cm>
\POS (0,15) *+{(a)}
\POS (0, 0) *+{\mathcal{C}^{2}\times\mathcal{C}} = "c1",
\POS (7, 0) *+{\phantom{(\mathcal{C})}}= "c3",
\POS (30, 0) *+{\mathcal{C}} = "c2",
\POS (11, 4) = "a",
\POS (11, -4) = "b",
\POS (11,-36) = "c",
\POS (11,-44) = "d",
\POS (0,-40) *+{\mathcal{C}^{2}\times\mathcal{C}_{\scriptscriptstyle{F}}} = "d1",
\POS (30,-40) *+{\mathcal{C}_{\scriptscriptstyle{F}}} = "d2",
\POS "c1" \ar@/^1.5pc/^{\otimes\cdot (\otimes\times \mathcal{C})} "c2",
\POS "c1" \ar@/_1.5pc/_{\otimes\cdot (\mathcal{C}\times\otimes)\cdot a_{\mathcal{C}}} "c2",
\POS "a" \ar^{\phantom{11}a} "b",
\POS "d1" \ar@/^1.5pc/^{\boxtimes\cdot (\otimes\times \mathcal{C}_{\scriptscriptstyle{F}})} "d2",
\POS "d1" \ar@/_1.5pc/_{\boxtimes\cdot (\mathcal{C}\times\boxtimes)\cdot a_{\mathcal{C}_{\ast}}} "d2",
\POS "c" \ar^{\phantom{11}\tilde{a}} "d",
\POS "c1" \ar_{\mathcal{C}^{2}\times G_{\scriptscriptstyle{F}}} "d1",
\POS "c2" \ar^{G_{\scriptscriptstyle{F}}} "d2",
\endxy &
\xy<1cm,0cm>
\POS (0, 0) *+{\phantom{123}}
\endxy
&
\xy<1cm,0cm>
\POS (0,15) *+{(b)}
\POS (0, 0) *+{\mathcal{C}} = "c1",
\POS (7, 0) *+{\phantom{(\mathcal{C})}}= "c3",
\POS (30, 0) *+{\mathcal{C}} = "c2",
\POS (11, 4) = "a",
\POS (11, -4) = "b",
\POS (11,-36) = "c",
\POS (11,-44) = "d",
\POS (0,-40) *+{\mathcal{C}_{\scriptscriptstyle{F}}} = "d1",
\POS (30,-40) *+{\mathcal{C}_{\scriptscriptstyle{F}}} = "d2",
\POS "c1" \ar@/^1.5pc/^{\otimes\cdot (\delta_{I}\times\mathcal{C})\cdot l^{-1}_{\mathcal{C}}} "c2",
\POS "c1" \ar@/_1.5pc/_{1_{\mathcal{C}}} "c2",
\POS "a" \ar^{\phantom{12}l} "b",
\POS "d1"
\ar@/^1.5pc/^{\boxtimes\cdot(\delta_{I}\times\mathcal{C}_{\scriptscriptstyle{F}})\cdot l^{-1}_{\mathcal{C}_{\scriptscriptstyle{F}}}} "d2",
\POS "d1" \ar@/_1.5pc/_{1_{\mathcal{C}_{\scriptscriptstyle{F}}}} "d2",
\POS "c" \ar^{\phantom{12}\tilde{l}} "d",
\POS "c1" \ar_{G_{\scriptscriptstyle{F}}} "d1",
\POS "c2" \ar^{G_{\scriptscriptstyle{F}}} "d2",
\endxy
\end{array}
\end{equation}
\end{enumerate}
\end{1505191350}
\emph{Proof}:\\
1 $\Rightarrow$ 2)\\
Take a right strong monad $(F, \sigma^{r})$ over the monoidal category
$(\mathcal{C}, \otimes, I)$. Due to this fact,
the following is a morphism of monads $(\otimes, \sigma^{r}):(\mathcal{C}\times\mathcal{C}, \mathcal{C}\times F)\longrightarrow (\mathcal{C}, F)$, and so are the following, $(\otimes\cdot(\otimes\times\mathcal{C}),\sigma^{r}(\otimes\times\mathcal{C}))$ and $(\phantom{.}\!\otimes\cdot(\mathcal{C}\times\otimes)\cdot a_{\mathcal{C}},
\sigma^{r}(\mathcal{C}\times\otimes)a_{\mathcal{C}}\circ\otimes(\mathcal{C}\times\sigma^{r})a_{\mathcal{C}}\phantom{.}\!)$.\\
The commutativity of the diagram (\ref{1405191723}a) implies that the
following is a transformation of monads
\begin{equation*}
\xy<1cm,0cm>
\POS (0, 0) *+{((\mathcal{C}\times\mathcal{C})\times\mathcal{C}, (\mathcal{C}\times \mathcal{C})\times F)} = "c1",
\POS (7, 0) *+{\phantom{(\mathcal{C})}}= "c3",
\POS (50, 0) *+{(\mathcal{C}, F)} = "c2",
\POS (27, 5) = "a",
\POS (27, -5) = "b",
\POS "c3" \ar@/^2pc/^{(\otimes\cdot (\otimes\times \mathcal{C}),\: \sigma^{r}(\otimes\times\mathcal{C}))} "c2",
\POS "c3" \ar@/_2pc/_{(\otimes\cdot (\mathcal{C}\times\otimes)\cdot a_{\mathcal{C}},\:\sigma^{r}(\mathcal{C}\times\otimes)a_{\mathcal{C}}\circ\otimes(\mathcal{C}\times\sigma^{r})a_{\mathcal{C}})} "c2",
\POS "a" \ar^{\phantom{1}a} "b",
\endxy
\end{equation*}\\
By the same reason as before, the following is a morphism of monads
$(\otimes\cdot(\delta_{I}\times\mathcal{C})\cdot l^{-1}_{\mathcal{C}},
\phantom{.}\!\sigma^{r}(\delta_{I}\times\mathcal{C})\phantom{.}\!
l^{-1}_{\mathcal{C}})$. Furthermore, due to the commutativity of (\ref{1405191723}b), we have a 2-cell of monads $l:(\otimes\cdot(\delta_{I}\times\mathcal{C})\cdot l^{-1}_{\mathcal{C}},
\phantom{.}\!\sigma^{r}(\delta_{I}\times\mathcal{C})\phantom{.}\!
l^{-1}_{\mathcal{C}})\longrightarrow (1_{\mathcal{C}}, 1_{\scriptscriptstyle{F}}):(\mathcal{C},
F)\longrightarrow (\mathcal{C}, F)$ as required.\\
The return of the implication is inmediate. For example, a monad transformation $a$
as indicated implies that a diagram like (\ref{1405191723}a) commutes.\\
2 $\Rightarrow$ 3)\\
In the isomorphism \eqref{1311042204}, make $(\mathcal{X}, H) =
(\mathcal{C}\times\mathcal{C}, \mathcal{C}\times F)$ and $L\dashv R =
G_{\scriptscriptstyle{F}}\dashv V_{\scriptscriptstyle{F}}$. Therefore, exists a bijection between morphisms of
monads of the form $(\otimes,\varphi)$ and morphisms of adjunctions $(\otimes,
\otimes_{\varphi})$, where the corresponding induced functor is denoted as $\boxtimes = \otimes_{\varphi}$. This pair of functors make a diagram like \eqref{1405222031} commute. In particular, the action of the second functor, on morphisms, is $\boxtimes (f, x^{\sharp}) = (\varphi_{A', Y}\cdot(f\otimes x))^{\sharp}$, where $x^{\sharp}:X\longrightarrow Y$ and $f:A\longrightarrow A'$.\\
Yet again, use the isomorphism \eqref{1311042204} with $(\mathcal{X}, H) = ((\mathcal{C}\times\mathcal{C})\times\mathcal{C},
(\mathcal{C}\times\mathcal{C})\times F)$ and $L\dashv R = G_{\scriptscriptstyle{F}}\dashv V_{\scriptscriptstyle{F}}$. Therefore, there exists a bijection between the tranformations of monads of the form $a$ and transformations of adjunctions $(a, \tilde{a})$. Where the
second natural tranformation has the form $\tilde{a}: \lbrack \otimes\cdot(\otimes\times\mathcal{C})\rbrack_{\varphi(\boxtimes\times\mathcal{C})}\longrightarrow\lbrack \otimes\cdot(\mathcal{C}\times\otimes)\cdot a_{\mathcal{C}})\rbrack_{\varphi^{2}}:(\mathcal{C}\times\mathcal{C})\times\mathcal{C}_{\scriptscriptstyle{F}}\longrightarrow \mathcal{C}_{\scriptscriptstyle{F}}$. Note that we use the following short notation $\varphi^{2} = \varphi(\mathcal{C}\times\otimes)\phantom{.}\! a_{\mathcal{C}}\circ\otimes\phantom{.}\!(\mathcal{C}\times\varphi)\phantom{.}\! a_{\mathcal{C}}$.\\
We prove only that $\lbrack \otimes\cdot(\otimes\times\mathcal{C})\rbrack_{\varphi(\boxtimes\times\mathcal{C})}= \boxtimes(\otimes\times\mathcal{C}_{\scriptscriptstyle{F}})$. Since for objects there is nothing to prove, let $((f,g),x^{\sharp}):((A,B), X)\longrightarrow ((A',B'), Y)$ be a morphim in $(\mathcal{C}\times\mathcal{C})\times\mathcal{C}_{\scriptscriptstyle{F}}$, therefore
\begin{eqnarray*}
\lbrack
\otimes\cdot(\otimes\times\mathcal{C})\rbrack_{\varphi(\boxtimes\times\mathcal{C})}((f,g),x^{\sharp})
&=& \big( \varphi(\otimes\times\mathcal{C})((A',B'),
Y)\cdot \otimes(\otimes\times\mathcal{C})((f, g), x)\big)^{\sharp}\\
&=& \big( \varphi(\otimes\times\mathcal{C})((A',B'),
Y)\cdot (f\otimes g)\otimes x\big)^{\sharp} = \boxtimes(f\otimes g,
x^{\sharp})\\
&=& \boxtimes (\otimes\times\mathcal{C}_{\scriptscriptstyle{F}})((f, g), x^{\sharp})
\end{eqnarray*}\\
At this moment, we change the notation to $\widetilde{a} = \nu$. Therefore,
the refered natural transformation can be written as $\nu:\boxtimes (\otimes\times\mathcal{C}_{\scriptscriptstyle{F}})\longrightarrow
\boxtimes(\mathcal{C}\times\boxtimes)a_{\mathcal{C}_{\ast}}:(\mathcal{C}\times\mathcal{C})\times\mathcal{C}_{\scriptscriptstyle{F}}\longrightarrow \mathcal{C}_{\scriptscriptstyle{F}}$,
according to the requirement. Note that the notation $a_{\mathcal{C}_{\ast}}$
stands for the object $a_{\mathcal{C},\mathcal{C},\mathcal{C}_{\scriptscriptstyle{F}}}$. The
component of the natural transformation $\nu$, on the object $((A,
A'), X)$, is \\
\begin{equation*}
\nu_{A, A', X} = \big(\eta^{\scriptscriptstyle{F}}(A\otimes(A'\otimes X))\cdot a_{A, A', X}\big)^{\sharp}
\end{equation*}
\noindent according to the equation \eqref{1311011822}.\\
The same procedure can be applied to the natural transformation $l$, in order
to get a natural transformation $j:\boxtimes(\delta_{I}\times\mathcal{C}_{\scriptscriptstyle{F}})l^{-1}_{\scriptscriptstyle{\mathcal{C}}_{\scriptscriptstyle{F}}}\longrightarrow
1_{\scriptscriptstyle{\mathcal{C}}_{\scriptscriptstyle{F}}}:\mathcal{C}_{\scriptscriptstyle{F}}\longrightarrow \mathcal{C}_{\scriptscriptstyle{F}}$, whose
component, on the object $X$, is $j_{X} = (\eta^{\scriptscriptstyle{F}}X\cdot l_{X})^{\sharp}$.\\
The reader is invited to realize the remain calculations in order to prove that the
structure $(\mathcal{C}_{\scriptscriptstyle{F}}, \boxtimes, \nu, j)$ is that of an left
$\mathcal{C}$-action on $\mathcal{C}_{\scriptscriptstyle{F}}$.\\
3 $\Rightarrow$ 2)\\
Since we have the isomorphism, the return is already given, nonetheless we comment some
part of the procedure.\\
Under the isomorphism, a commutative diagram like \eqref{1405222031} give rise
to a morphism of monads of the form $(\otimes, \varphi^{\boxtimes}):
(\mathcal{C}\times\mathcal{C}, \mathcal{C}\times F)\longrightarrow
(\mathcal{C}, F)$ where the commutative diagrams that fulfill this morphism
are \eqref{1405191712} and \eqref{1405191723}.\\
If we have a commutative surface like (\ref{1405202049}a), we obtain, through
the isomorphism, a transformation of monads of the form $a: (\otimes\cdot(\otimes\times\mathcal{C}),
\varphi^{\boxtimes}(\otimes\times\mathcal{C}))\longrightarrow
(\otimes\cdot(\mathcal{C}\times\otimes)\cdot a_{\mathcal{C}},
\varphi^{\boxtimes}(\mathcal{C}\times\otimes)a_{\mathcal{C}}\circ
\otimes(\mathcal{C}\times\varphi^{\boxtimes})a_{\mathcal{C}})$.\\
$\square$
We state the dual theorem\\
\newtheorem{1505241921}[1405191543]{Theorem}
\begin{1505241921}
There exists a bijection between the following structures
\begin{enumerate}
\item [1.-] Left strong monads $((F,\sigma^{l}),\mu^{\scriptscriptstyle{F}},\eta^{\scriptscriptstyle{F}})$ on the
monoidal category $(\mathcal{C}, \otimes, I, a, r, l)$.
\item [2.-] Morphisms and transformations of monads of the form
\begin{eqnarray*}
(\otimes, \varphi) &:& (\mathcal{C}\times\mathcal{C},
F\times\mathcal{C})\longrightarrow (\mathcal{C}, F)\\
a &:& (\otimes\cdot(\otimes\times\mathcal{C}),
\varphi(\otimes\times\mathcal{C})\circ\otimes(\varphi\times\mathcal{C}))\longrightarrow
(\otimes\cdot(\mathcal{C}\times\otimes)\cdot a_{\mathcal{C}},
\varphi(\mathcal{C}\times\otimes)a_{\mathcal{C}})\\
&:& ((\mathcal{C}\times\mathcal{C})\times\mathcal{C},
(F\times\mathcal{C})\times \mathcal{C})\longrightarrow (\mathcal{C}, F)\\
r&:& (\otimes\cdot(\mathcal{C}\times\delta_{I})\cdot r^{-1}_{\mathcal{C}},
\phantom{.}\!\varphi(\mathcal{C}\times\delta_{I})\phantom{.}\!
r^{-1}_{\mathcal{C}})\longrightarrow (1_{\mathcal{C}}, 1_{\scriptscriptstyle{F}}):(\mathcal{C},
F)\longrightarrow (\mathcal{C}, F)
\end{eqnarray*}
\item [3.-] Right actions on the Kleisli category, $\mathcal{C}_{\scriptscriptstyle{F}}$,
$\boxtimes:\mathcal{C}_{\scriptscriptstyle{F}}\times\mathcal{C}\longrightarrow \mathcal{C}_{\scriptscriptstyle{F}}$
such that the following diagrams of morphisms and surfaces commute
\begin{equation*}\label{1405241922}
\xy<1cm,0cm>
\POS (0,20) *+{\mathcal{C}\times\mathcal{C}} = "a11",
\POS (20,20) *+{\mathcal{C}} = "a12",
\POS (0,0) *+{\mathcal{C}_{\scriptscriptstyle{F}}\times\mathcal{C}} = "a21",
\POS (20,0) *+{\mathcal{C}_{\scriptscriptstyle{F}}} = "a22",
\POS "a11" \ar^-{\otimes} "a12",
\POS "a12" \ar^{G_{\scriptscriptstyle{F}}} "a22",
\POS "a11" \ar_{G_{\scriptscriptstyle{F}}\times\mathcal{C}} "a21",
\POS "a21" \ar_-{\boxtimes} "a22",
\endxy
\end{equation*}
\begin{equation*}\label{1405241923}
\begin{array}{ccc}
\xy<1cm,0cm>
\POS (0,15) *+{(a)}
\POS (0, 0) *+{\mathcal{C}^{2}\times\mathcal{C}} = "c1",
\POS (7, 0) *+{\phantom{(\mathcal{C})}}= "c3",
\POS (30, 0) *+{\mathcal{C}} = "c2",
\POS (11, 4) = "a",
\POS (11, -4) = "b",
\POS (14,-36) = "c",
\POS (14,-44) = "d",
\POS (0,-40) *+{(\mathcal{C}_{\scriptscriptstyle{F}}\times\mathcal{C})\times\mathcal{C}} = "d1",
\POS (30,-40) *+{\mathcal{C}_{\scriptscriptstyle{F}}} = "d2",
\POS "c1" \ar@/^1.5pc/^{\otimes\cdot (\otimes\times \mathcal{C})} "c2",
\POS "c1" \ar@/_1.5pc/_{\otimes\cdot (\mathcal{C}\times\otimes)\cdot a_{\mathcal{C}}} "c2",
\POS "a" \ar^{\phantom{11}a} "b",
\POS "d1" \ar@/^1.5pc/^{\boxtimes\cdot (\boxtimes\times \mathcal{C})} "d2",
\POS "d1" \ar@/_1.5pc/_{\boxtimes\cdot (\mathcal{C}_{\scriptscriptstyle{F}}\times\otimes)\cdot a_{\mathcal{C}_{\ast}}} "d2",
\POS "c" \ar^{\phantom{11}\tilde{a}} "d",
\POS "c1" \ar_{(G_{\scriptscriptstyle{F}}\times\mathcal{C})\times\mathcal{C}} "d1",
\POS "c2" \ar^{G_{\scriptscriptstyle{F}}} "d2",
\endxy &
\xy<1cm,0cm>
\POS (0, 0) *+{\phantom{123}}
\endxy
&
\xy<1cm,0cm>
\POS (0,15) *+{(b)}
\POS (0, 0) *+{\mathcal{C}} = "c1",
\POS (7, 0) *+{\phantom{(\mathcal{C})}}= "c3",
\POS (30, 0) *+{\mathcal{C}} = "c2",
\POS (11, 4) = "a",
\POS (11, -4) = "b",
\POS (11,-36) = "c",
\POS (11,-44) = "d",
\POS (0,-40) *+{\mathcal{C}_{\scriptscriptstyle{F}}} = "d1",
\POS (30,-40) *+{\mathcal{C}_{\scriptscriptstyle{F}}} = "d2",
\POS "c1" \ar@/^1.5pc/^{\otimes\cdot (\mathcal{C}\times\delta_{I})\cdot r^{-1}_{\mathcal{C}}} "c2",
\POS "c1" \ar@/_1.5pc/_{1_{\mathcal{C}}} "c2",
\POS "a" \ar^{\phantom{12}r} "b",
\POS "d1"
\ar@/^1.5pc/^{\boxtimes\cdot(\mathcal{C}_{\scriptscriptstyle{F}}\times\delta_{I})\cdot r^{-1}_{\mathcal{C}_{\scriptscriptstyle{F}}}} "d2",
\POS "d1" \ar@/_1.5pc/_{1_{\mathcal{C}_{\scriptscriptstyle{F}}}} "d2",
\POS "c" \ar^{\phantom{12}\tilde{r}} "d",
\POS "c1" \ar_{G_{\scriptscriptstyle{F}}} "d1",
\POS "c2" \ar^{G_{\scriptscriptstyle{F}}} "d2",
\endxy
\end{array}
\end{equation*}
\end{enumerate}
\end{1505241921}
We left to the reader the writing of dual statements, \emph{i.e.} the ones that
corresponds to the Eilenberg-Moore category, where the direction of the
natural transformations are inverted, for example $\widehat{\sigma}^{r}_{A,B}:
F(A\otimes B)\longrightarrow A\otimes FB$.
\section{Functor Algebras}
Check Proposition II.1.1 in \cite{duej_kaee} and \cite{mewi_nobh} for this section. we define the category of $H$-left functor algebras for a given monad $(\mathcal{D}, H, \mu^{\scriptscriptstyle{H}}, \eta^{\scriptscriptstyle{H}})$.
\newtheorem{1407022006}{Definition}[section]
\begin{1407022006}
The category of left $H$-functor algebras, for the pair $(\mathcal{C},\mathcal{D})$, denoted as ${}_{\scriptscriptstyle{H}}\mathcal{F}$ or ${}_{\scriptscriptstyle{H}}\mathcal{M}$ is defined as follows. The objects are given by $(J,\lambda_{\scriptscriptstyle{J}})$, where $J:\mathcal{C}\longrightarrow\mathcal{D}$ is a functor and $\lambda_{\scriptscriptstyle{J}}: HJ\longrightarrow J$ is a natural transformation such that the following diagrams commute
\begin{equation}\label{1407022029}
\begin{array}{cc}
\xy<1cm,0cm>
\POS (0,20) *+{HHJ} = "a11",
\POS (27,20) *+{HJ} = "a12",
\POS (0,0) *+{HJ} = "a21",
\POS (27,0) *+{J} = "a22",
\POS "a11" \arrow^{\mu^{\scriptscriptstyle{H}}J} "a12",
\POS "a12" \arrow^{\lambda_{\scriptscriptstyle{J}}} "a22",
\POS "a11" \arrow_{H\lambda_{\scriptscriptstyle{J}}} "a21",
\POS "a21" \arrow_{\lambda_{\scriptscriptstyle{J}}} "a22",
\endxy &
\xy<1cm,0cm>
\POS (0,20) *+{J} = "a11",
\POS (20,20) *+{HJ} = "a12",
\POS (20,0) *+{J} = "a22",
\POS "a11" \arrow^{\eta^{\scriptscriptstyle{H}}J} "a12",
\POS "a12" \arrow^{\lambda_{\scriptscriptstyle{J}}} "a22",
\POS "a11" \arrow_{1_{\scriptscriptstyle{J}}} "a22",
\endxy
\end{array}
\end{equation}
A \emph{morphism of functor algebras}, is a natural transformation, $\theta:(J,\lambda_{\scriptscriptstyle{J}})\longrightarrow (K,\lambda_{\scriptscriptstyle{K}})$,
$\theta:J\longrightarrow K$ such that the following diagram commute
\begin{equation}\label{1407022032}
\xy<1cm,0cm>
\POS (0,20) *+{HJ} = "a11",
\POS (25,20) *+{HK} = "a12",
\POS (0,0) *+{J} = "a21",
\POS (25,0) *+{K} = "a22",
\POS "a11" \arrow^{H\theta} "a12",
\POS "a12" \arrow^{\lambda_{\scriptscriptstyle{K}}} "a22",
\POS "a11" \arrow_{\lambda_{\scriptscriptstyle{J}}} "a21",
\POS "a21" \arrow_{\theta} "a22",
\endxy
\end{equation}
\end{1407022006}
We realize that the diagrams given by \eqref{1407022029}, for a left
$H$-functor algebra, account for a monad morphism of the form
$(J,\lambda_{\scriptscriptstyle{J}}):(\mathcal{C}, 1_{\mathcal{C}})\longrightarrow (\mathcal{D},
H)$. In the same way, the commutative diagram for a morphism of left
$H$-functor algebras, as in \eqref{1407022032}, account for a monad
transformation $\theta:(J,\lambda_{\scriptscriptstyle{J}})\longrightarrow(K,
\lambda_{\scriptscriptstyle{K}}):(\mathcal{C}, 1_{\mathcal{C}})\longrightarrow(\mathcal{D}, H)$.\\
Using the isomorphism for the Eilenberg-Moore 2-adjunction, given by \eqref{1403191902}, the category ${}_{\scriptscriptstyle{H}}\mathcal{F}$ is isomorphic to the following category, named possibly as \emph{category of liftings to} $\mathcal{D}^{\scriptscriptstyle{H}}$, for the pair $(\mathcal{C},\mathcal{D})$. The objects of such category are functor pairs $(J, \hat{J}\phantom{.}\!)$ such that they complete to an adjunction morphism, in $\mathbf{Adj}_{\scriptscriptstyle{R}}({}_{2}Cat)$, of the form $(J, \hat{J}\phantom{.}\!): 1_{\mathcal{C}}\dashv 1_{\mathcal{C}}\longrightarrow D^{\scriptscriptstyle{H}}\dashv U^{\scriptscriptstyle{H}}$. That is to say, the following diagram commutes
\begin{equation*}
\xy<1cm,0cm>
\POS (0, 0) *+{\mathcal{C}} = "c1",
\POS (20, 0) *+{\mathcal{D}} = "c2",
\POS (0, -20) *+{\mathcal{C}} = "d1",
\POS (20, -20) *+{\mathcal{D}^{\scriptscriptstyle{H}}} = "d2",
\POS "c1" \ar^{J} "c2",
\POS "d1" \ar^{1_{\mathcal{C}}} "c1",
\POS "d2" \ar_{U^{\scriptscriptstyle{H}}} "c2",
\POS "d1" \ar_{\hat{J}} "d2",
\endxy
\end{equation*}
\emph{i.e.}
\begin{equation*}
\xy<1cm,0cm>
\POS (0, 0) *+{\mathcal{C}} = "c1",
\POS (20, 0) *+{\mathcal{D}} = "c2",
\POS (20, -18) *+{\mathcal{D}^{\scriptscriptstyle{H}}} = "d2",
\POS "c1" \ar^{J} "c2",
\POS "d2" \ar_{U^{\scriptscriptstyle{H}}} "c2",
\POS "c1" \ar_{\hat{J}} "d2",
\endxy
\end{equation*}
The morphisms of such category are the usual morphisms of adjunctions $(\alpha, \beta): (J, \hat{J}\phantom{.}\!)\longrightarrow (K,\hat{K}):1_{\mathcal{C}}\dashv 1_{\mathcal{C}}\longrightarrow D^{\scriptscriptstyle{H}}\dashv U^{\scriptscriptstyle{H}}$. We proved then the following theorem.
\newtheorem{1406031326}{Theorem}[section]
\begin{1406031326}
There exists an isomorphism, natural on $\mathcal{C}$ and $(\mathcal{D}, H)$, between the following categories
\begin{enumerate}
\item [1.-] The category of left $H$-functor algebras ${}_{\scriptscriptstyle{H}}\mathcal{F}$.
\item [2.-] The category of liftings to $\mathcal{D}^{\scriptscriptstyle{H}}$, for the pair $(\mathcal{C},\mathcal{D})$.
\end{enumerate}
\end{1406031326}
\noindent $\square$
Dually, we have the category of right $H$-functor algebras, for the monad $(\mathcal{D}, H, \mu^{\scriptscriptstyle{H}}, \eta^{\scriptscriptstyle{H}})$, denoted as $\mathcal{F}_{H}$ or $\mathcal{M}_{\scriptscriptstyle{H}}$. The objects are pairs $(J, \rho_{\scriptscriptstyle{J}})$, where the natural transformation $\rho_{\scriptscriptstyle{J}}:JH\longrightarrow J$ is such that it fulfills diagrams dual to those in \eqref{1407022029}. In the same (dual) way as before, this category is the same as the category $Hom_{\mathbf{Mnd}^{\scriptscriptstyle{\bullet}}({}_{2}Cat)}((\mathcal{D}, H), (\mathcal{C}, 1_{\mathcal{C}}))$. Therefore using the isomorphism \eqref{1311042204}, the previous category is isomorphic to the category named as \emph{extensions from} $\mathcal{D}_{\scriptscriptstyle{H}}$, for the pair $(\mathcal{D},\mathcal{C})$. The objects of this category are pairs of functors $(J, \tilde{J}\phantom{.}\!)$ such that thery complete to an adjunction morphism $(J,\tilde{J}\phantom{.}\!):G_{\scriptscriptstyle{H}}\dashv V_{\scriptscriptstyle{H}}\longrightarrow 1_{\mathcal{C}}\dashv 1_{\mathcal{C}}$ in $\mathbf{Adj}_{\scriptscriptstyle{L}}({}_{2}Cat)$. In particular, the following diagram commutes
\begin{equation*}
\xy<1cm,0cm>
\POS (0, 0) *+{\mathcal{D}} = "c1",
\POS (20, 0) *+{\mathcal{C}} = "c2",
\POS (0, -18) *+{\mathcal{D}_{\scriptscriptstyle{H}}} = "d2",
\POS "c1" \ar^{J} "c2",
\POS "d2" \ar_{\tilde{J}} "c2",
\POS "c1" \ar_{G_{\scriptscriptstyle{H}}} "d2",
\endxy
\end{equation*}
\noindent We also proved the following theorem
\newtheorem{1406032115}[1406031326]{Theorem}
\begin{1406032115}
There exists an isomorphism, natural on $(\mathcal{D}, H)$ and $\mathcal{C}$, between the following categories
\begin{enumerate}
\item [1.-] The category of right $H$-functor algebras $\mathcal{F}_{\scriptscriptstyle{H}}$.
\item [2.-] The category of extensions from $\mathcal{D}_{\scriptscriptstyle{H}}$, for the pair $(\mathcal{D},\mathcal{C})$.
\end{enumerate}
\end{1406032115}
\noindent $\square$
\section{Conclusions and Future Work}
This review has the objective to show how several situations for the theory of
monads are connected in a very simple way, through a 2-adjunction. Any person who has
tought a course on monads would agree that this structure, of a 2-adjunction,
can be used as an educational purpose in the sense that a simple structure can
account for several situations and which can spare the, otherwise cumbersome,
details of the proofs. \\
For future work we have a few recommendations. The reader may find interesting
to extent the part of strong monads and actions over the Kleisli categories to
strong symmetrical monads and use the actions for the Eilenberg-Moore case. It
would be interesting also to contextualize the case of the monoidal liftings
and monoidal extensions according to the formal theory of monoidal monads, and the
\emph{standar} objects, given in \cite{zama_fotm}.\\
The reader may want to find more situations in the monad theory that can use the isomorphism provided by this pair of 2-adjunctions, the authors will certainly pursue this issue.
\newpage
|
\section{Effects of Interaction on the wave function in the perpendicular dimensions}
Fig. 5 shows the numerical evaluation of $\left|\psi_{k_{1\perp},k_{2\perp}}\right|^{2}$, the modulus squared of the two-excitation wave function in the dimensions perpendicular to the separation. While the second order of the interaction changes the Gaussian cross section from circular to elliptical by entangling the two excitations, the first order does not have any effect in this dimension.
\begin{figure}[h]\label{SiPerpendicular}
\scalebox{0.5}{\includegraphics*{perpendicular.pdf}}
\caption{Numerical results for $\left|\psi_{k_{1\perp},k_{2\perp}}\right|^{2}$. There is no momentum displacement in this dimension. The parameters are the same as for Fig. 2.}
\end{figure}
\section{Sensitivity of Fidelity to positioning errors in Swapping}
Fig. 6 shows that the swapping protocol is more sensitive to positioning errors for the collective excitations in the parallel dimension than in the perpendicular dimension.
\begin{figure}\label{errorinswapping}
\scalebox{0.35}{\includegraphics*{errorinswapping.pdf}}
\caption{Average fidelity reduction as a function of positioning error for the swapping protocol in (a) parallel and (b) perpendicular dimension.
The parameters are the same as for Fig.~2.}
\end{figure}
\section{``Frozen Collision''}
Fig. 7 shows the redistribution of the momentum vectors of the collective excitations due to the interaction. One sees that collective excitations that are created by the storage of co-propagating photons will yield diverging photons upon retrieval.
\begin{figure}[h!]\label{AngularDestribution}
\scalebox{0.3}{\includegraphics{AngularDistribution.pdf}}
\caption{Angular distribution of the momenta of the collective excitations before (black for both) and after (red and blue) interaction. The parameters are the same as for Fig. 2.}
\end{figure}
\section{Gate Efficiency: thermal motion and uniform loss}
\subsection{Thermal motion}
The efficiency factor due to the thermal motion of the atoms is given by $\eta_{_{th}}=\frac{1}{(1+(\frac{t}{\xi})^{2})^{2}}exp[\frac{-t^{2}/\tau^{2}}{(1+(t/\xi)^{2})}]$
\cite{key-1}, where $\tau=\frac{\Lambda}{2\pi v}$ is the dephasing
time scale, which is determined by the wave length $\Lambda$ of the collective excitations and the thermal speed $v$, and $\xi=\frac{w}{v}$ is the time scale on which an atom traverses the width $w$ of a collective excitation.
\subsection{Uniform loss}
Photon loss that is uniform over the four rails has no effect on the conditional fidelity and can be quantified independently in terms of the efficiency. Since only the interacting rails ($\left|1_{C}\right\rangle ,\left|1_{T}\right\rangle $) are excited
to Rydberg levels, they experience an extra loss due to the life time of the Rydberg level. Furthermore, the shorter wavelength of the Rydberg excitations in the interacting rails creates a stronger loss due to the atomic thermal motion compared to the non-interacting rails. Finally, the extra process of storage and retrieval during the swapping of the interacting rails causes an extra loss of efficiency in these rails. The loss can be made uniform by adding a controlled external source of loss
on the non-interacting rails ($\left|0_{C}\right\rangle ,\left|0_{T}\right\rangle $).
|
\section{Introduction}
Around a collapsing star, gravity takes an ultra-strong intensity.
Hence corresponding matter density and space-time curvature diverge
to infinity. It is predicted from singularity theorems in general
relativity [2] that end of stellar collapse leads to a visible naked
or invisible covered singularity. Property of visibility of
singularities are determined by the causal structure of collapsing
process, satisfying the gravitational field equations. Internal
dynamics of the collapse can be determine type of singularities.
They will be naked and hence visible from view of external observer
if collapse process delays the formation of the event horizon
[3,4,5]. Penrose represented `cosmic
censorship hypotheses` initially [6], which explains properties of the final singularity of a
gravitational collapse. In this conjecture, space-time singularities created from gravitational collapse usually should be covered
by hyper-surface called as event horizon. Hence
it will be invisible from view of outside observers of collapsed
object namely black hole [7]. However there is still a doubt for
reality of censorship hypotheses and has been
recognized as one of the most open problems in classical general
theory of relativity. The censorship hypotheses may to be still valid in quantum gravity approach (see [8] and references therein).
Black holes and naked
singularities have different properties observationally. They have different characteristics
in the gravitational lensing [9,10] and also many exact solutions of Einstein`s gravitational field equations are obtained which
admit naked singularities
[11-18].\\
In the present paper we use
vacuum sector of the Brans Dicke gravity [19] in two dimensional analogue [20] and obtain non-static solutions of the
field equations. We apply post Newtonian perturbation series expansion method
to solve the dynamical field equation because there is not obtained exact solutions.
Stress tensor of our solutions treats as anisotropic spherically symmetric perfect
fluid. Equation of state takes dark and visible regular matter regime during the collapse
process under particular conditions on the values of the parameters of the solutions. Apparent and event horizon of the collapse process is
derived and in static regime of the solutions apparent horizon will be inside of event horizon for $\omega>0$ (see fig. 1) \\
In second part of the paper we study corresponding de
Broglie-Bohm quantum wave functional of the collapsing fluid sphere.
Probability amplitude of particles ensemble and their radial
velocities is calculated from incident current density of particles
ensemble. Also total mass of the collapsing fluid is calculated
restricting to choose $\omega<-1;\omega>\frac{2}{3}$ (see fig.2).
Diagram of the incident current density of particles ensemble is
given in fig.3 comparable with evaluated quantum potential on the
horizons. This current density describes radially accelerating
particles of fluid sphere which come crosses apparent and then event
horizon in the presence of quantum potential effect. Organization
of the article is as follows.\\
In section 2 we present dilatonic Brans Dicke gravity in 2 dimensions [1]. Varying the action,
we obtain dynamical equations of the two dimensional induced metric, the Brans Dicke scalar field and the dilation field.
Dilaton field comes from dimensional reduction of the Brans Dicke action from four to two dimensions.
General form of the anisotropic spherically symmetric Brans Dicke perfect fluid stress
tensor is discuss in terms of the fluid density, radial and tangential pressures respectively.
In section 3 we
solve dynamical field equations and obtain non-static dynamical
solutions of the system in terms of time and radial coordinates. Our
solutions are obtained by applying post Newtonian approach because
the non-linear dynamical field equations have not exact solutions.
In this section mass function of the collapsing fluid is also
derived and discuss with respect to different values of the
Brans-Dicke parameter. In section 4 we use result of the paper [1]
and obtained shape of de Broglie quantum wave functional of our
collapsing Brans Dicke fluid sphere. In this section we calculate
current density of particles ensemble moving on the apparent and
event horizons compare with values of quantum potential of the
system for different values of the Brans-Dicke parameter. Section 5
allocates to summery and conclusion.
\section{2d Dilatonic Brans-Dicke gravity} We take the Brans-Dicke scalar tensor gravity
theory [18] such as follows.
\begin{equation}
I[g_{\mu\nu},\phi]=\frac{1}{16\pi}\int d^4x\sqrt{g}\{\phi R^{[4]}-\frac{\omega}{\phi}g^{\mu\nu}\partial_{\mu}\phi\partial_{\nu}\phi
\}\end{equation}
where $g$ is absolute value of determinant of the metric
$g_{\mu\nu}$ with Lorentzian signature $(-,+,+,+).$ $\omega$ is
dimensionless Brans-Dicke parameter and its present limits based on
time-delay experiments [20-23] requires
$\omega\geq4\times10^4.$ When $\omega\to\infty$, then the theory
leads to the Einstein`s general relativity [24]
and its negative values come from higher dimensions of curved space time (see [25] and references therein).\\
Using general form of spherically symmetric curved space times
metric such that
\begin{equation}
d s^2=g_{a b}dx^{a}dx^{b}+\psi^{2}(x^a)
\left(d\theta^{2}+\sin^{2}\theta d\varphi^{2}\right),\label{1.2}
\end{equation}
and eliminating
divergence-less terms, the action (2.1) become [1]:
\begin{equation} I[\phi,\psi,g_{ab}]=\frac{1}{2}\int
d^2x\sqrt{g}\{\phi+\frac{1}{2}\phi\psi^2 R+\phi
g^{ab}\partial_a\psi\partial_b\psi$$$$+2\psi
g^{ab}\partial_a\psi\partial_b\phi-\frac{\omega\psi^2}{2\phi}g^{ab}\partial_a\phi\partial_b\phi\}\end{equation}
where $\psi^2$ is 2-sphere conformal factor and $R $ is Ricci scalar
of 2-dimensional metric $g_{ab}$ defined by 2-dimensional
coordinates system $x^{a}$ with $a \equiv 1,2.$ Varying the action
(2.3), with respect to the fields $g^{ab},$ $\psi$ and $\phi$ the
corresponding field equations are obtained respectively as
\begin{equation}
-2\frac{\nabla_a\nabla_b\psi}{\psi}+g_{ab}[2\frac{\Box\psi}{\psi}+\frac{\partial_c\psi\partial^c\psi}{\psi^2}-\frac{1}{\psi^2}]=T_{ab}[\phi],\end{equation}
\begin{equation}\frac{\Box\psi}{\psi}-\frac{R}{2}=T^\varphi_\varphi[\phi]=T^\theta_{\theta}[\phi]\end{equation}
and \begin{equation}
\omega\frac{\Box\phi}{\phi}+\frac{1}{\psi^2}+\frac{R}{2}-\frac{\omega}{2}\frac{\partial_c\phi\partial^c\phi}{\phi^2}
+\frac{2\omega\partial_c\psi\partial^c\phi}{\phi\psi}-\frac{2\Box\psi}{\psi}-\frac{\partial_c\psi\partial^c\psi}{\psi^2}=0
\end{equation}
where $\Box=\partial_a(\sqrt{g}g^{ab}\partial_b)/\sqrt{g},$
$\nabla_a\nabla_b=\partial_a\partial_b-\Gamma^c_{ab}\partial_c,$
$T_{ab}$ and
$T^\theta_{\theta}=T^\varphi_\varphi$ is 2d and angular parts of the
the dilaton-Brans-Dicke field stress tensor defined by respectively
as
\begin{equation}
T_{ab}[\phi]=\frac{\omega}{\phi^2}\partial_a\phi\partial_b\phi+\frac{\nabla_a\nabla_b\phi}{\phi}-g_{ab}
[\frac{\Box\phi}{\phi}+\frac{2\partial_c\psi\partial^c\phi}{\psi\phi}+\frac{\omega\partial_c\phi\partial^c\phi}{2\phi^2}],\end{equation}
and \begin{equation}
T^\varphi_\varphi[\phi]=T^{\theta}_{\theta}[\phi]
=-\frac{1}{\psi\phi}\{\partial_c\phi\partial^c\psi+\frac{\omega}{2}\frac{\psi}{\phi}\partial_c\phi\partial^c\phi\}.
\end{equation}
We consider here, space time source fluids (the Brans-Dicke and dilaton scalar fields)
for which the energy momentum tensor can be decompose relative to the fluid flow (velocity field) $U^{\mu}
$ in four dimension $\mu,\nu=t,r,\theta, \varphi$ as [26,27,28]
\begin{equation} T_{\mu\nu}=(\varrho+p_t)U_\mu U_\nu+p_tg_{\mu\nu}+(p_r-p_t)S_{\mu}S_{\nu}\end{equation}
where \begin{equation}\varrho=T_{\mu\nu}U^{\mu}U^{\nu}\end{equation}
is the energy density relative to $U^{\mu}$,
\begin{equation}p_r=\frac{1}{3}(T_{\mu\nu}h^{\mu\nu}) \end{equation} is the isotropic radial pressure relative to $U^\mu$,
and where $h_{\mu\nu}=U_{\mu}U_{\nu}+g_{\mu\nu} $. The fluid
velocity vector $U^{\mu}$ is time-like $g_{\mu\nu}U^{\mu}U^{\nu}=-1$
and $S^{\mu}$ is space-like $S^{\mu}S_{\mu}=3$, on the direction of
anisotropy, with $S^{\mu}U_{\mu}=0.$ $S^{\mu}$ is called the
momentum density (the energy flux) relative to the fluid flow
$U^{\mu}$. The transverse or tangential pressure $p_t$ is derived
from (2.9) and (2.11) as
\begin{equation}p_t=\frac{1}{2}\left(h^{\mu\nu}-\frac{1}{3}S^{\mu}S^{\nu}\right)T_{\mu\nu}.\end{equation}
When $p_r=p_t$ the stress tensor (2.9) leads to an isotropic perfect
fluid stress tensor. \\
It is convenient to introduce 2 dimensional
space time metric in $t,r$ coordinates with line element
\begin{equation}ds^2=-e^{2A(t,r)}dt^2+2e^{2B(t,r)}dtdr+e^{2C(t,r)}dr^2.\end{equation}
With the above background metric and from point of view of free
falling comoving observers, $U^\mu U_{\mu}=-1,$ $S^{\mu}S_{\mu}=3$
and $U_{\mu}S^{\mu}=0$ yields
\begin{equation} U^{\mu}(t,r)=(e^{-A(t,r)},0,0,0)\end{equation}
and \begin{equation}
S^\mu(t,r)=(0,\sqrt{3}e^{-C(t,r)},0,0).\end{equation} With these
definitions, (2.10), (2.11) and (2.12) leads to the following
relations respectively. \begin{equation}
\varrho(t,r)=T_{tt}e^{-2A}=-T^t_t,\end{equation}
\begin{equation}p_r(t,r)=\frac{1}{3}\{2T^{tr}e^{2B}+2T^{\theta}_{\theta}+T^{rr}e^{2C}\}\end{equation}
and
\begin{equation} p_t(t,r)=e^{2B}T^{tr}+T^{\theta}_{\theta}
\end{equation}
where $p_r, p_t$ is pressure along radial and tangential direction
respectively. We can also define magnitude of the anisotropic stress
tensor $S(t,r)$ and isotropic part of pressure $p(t,r)$ of the fluid
as
\begin{equation}S(t,r)=p_r-p_t=\frac{p-p_t}{3},~~~p(t,r)=3p_r(t,r)-2p_t(t.r).
\end{equation} In the next section, we seek
dynamical solutions of the metric equations defined by (2.4), (2.5)
and (2.6).
\section{ Dynamical solutions}
Applying the background metric equation (2.13) the dynamical field
equations (2.4), (2.5) and (2.6) do not take simple form and will
not have exact solutions. Hence we use post Newtonian approach to
solve the field equations such as follows. Assuming perturbation
series expansion of the background metric (2.13) as
\begin{equation} e^{2A(t,r)}=1+\varepsilon
x(t,r)+O(\varepsilon^2),~~~e^{2B(t,r)}=1+\varepsilon
y(t,r)+O(\varepsilon^2),$$$$e^{2C(t,r)}=1+\varepsilon
z(t,r)+O(\varepsilon^2)
\end{equation} in which $|\varepsilon|<<1$ is assumed to be order parameter of the
perturbation\footnote{ In section 4 we obtain a good physical mining
for $\varepsilon,$ namely it become proportional to the quantum
potential value evaluated on the event horizon of our dilatoinc
Brans-Dicke collapsing fluid sphere and it will be also independent
of different values of the Brans Dicke parameter $\omega.$ } we
will have up to terms in order $\varepsilon^2$\begin{equation}
g^{tt}(t,r)\approx-\frac{1}{2}\{1+\varepsilon[z-x-2y]/2\}
,~~~~g^{tr}(t,r)\approx\frac{1}{2}\{1-\varepsilon[x+z]/2 \}$$$$
g^{rr}(t,r)\approx\frac{1}{2}\{1+\varepsilon[x-z-2y]/2\},~~~Ricci\approx\frac{\varepsilon}{8}[\partial^2_rx-4\partial^2_tz+8\partial_t\partial_ry],
\end{equation}
\begin{equation}
\Gamma^t_{tt}\approx\frac{\varepsilon}{4}[\partial_tx+2\partial_ty+\partial_rx],~~
~\Gamma^t_{tr}=\Gamma^t_{rt}\approx\frac{\varepsilon}{4}[\partial_rx+\partial_tz]
\end{equation}
\begin{equation}
\Gamma^t_{rr}\approx\frac{\varepsilon}{4}[\partial_rz+\partial_tz-2\partial_ry],~
~~\Gamma^r_{tt}\approx\frac{\varepsilon}{4}[\partial_rx+2\partial_ty-\partial_tx]
\end{equation}
\begin{equation}\Gamma^r_{tr}=\Gamma^r_{rt}\approx\frac{\varepsilon}{4}[\partial_tz-\partial_rx],~~
~\Gamma^r_{rr}\approx\frac{\varepsilon}{4}[2\partial_ry+\partial_rz-\partial_tz].\end{equation}
With these approximations and definitions \begin{equation}
\psi(t,r)=\sqrt{G}e^{\frac{1}{2}+\varepsilon
u(t,r)},~~~\phi(t,r)=\frac{e^{\varepsilon v(t,r)}}{G}\end{equation}
the field equations defined by (2.4), (2.5) and (2.6), up to terms
of order in $\varepsilon^2,$ leads to the following linear partial
differential equations.
\begin{equation}
\partial^2_\tau u+2\partial_\tau\partial_\rho(u+v)+\partial^2_\rho(u+v)-x+2u\approx0,\end{equation}
\begin{equation}\partial^2_\tau(u+v)-\partial_\tau\partial_\rho v-\partial^2_\rho (u+v)-2u+y\approx0,
\end{equation}
\begin{equation}\partial^2_\tau(u+v)-2\partial_\tau\partial_\rho(u+v)+\partial^2_\rho u-2u+z\approx0,
\end{equation}
\begin{equation}
\partial^2_\tau(u-z/2)+\partial_\tau\partial_\rho(y-2u)-\partial^2_\rho(u-x/8)\approx0
\end{equation}
and
\begin{equation}
\partial^2_\tau(u-z/4-\omega v/2)+\partial_\tau\partial_\rho(\omega v-2u+y/2)+\partial^2_\rho(\omega v/2-u+x/16)-2u\approx0\end{equation}
where we defined dimensionless coordinates
\begin{equation}\tau=\frac{t}{\sqrt{G}},~~~~\rho=\frac{r}{\sqrt{G}}.
\end{equation}
One can obtain solutions of the above equations as \begin{equation}
u(\tau,\rho)=u_0e^{a\tau+b\rho},~~~v(\tau,\rho)=v_0e^{a\tau+b\rho},~~~x(\tau,\rho)=x_0e^{a\tau+b\rho},\end{equation}
and
\begin{equation}y(\tau,\rho)=y_0e^{a\tau+b\rho},~~~~z(\tau,\rho)=z_0e^{a\tau+b\rho}\end{equation}
in which the constants $u_0, v_0, x_0, y_0, z_0, \omega, a, b$
satisfies the following
relations.
\begin{equation}\omega(a,b)=\frac{(a^2-2ab-b^2-4)(16a^3b-8a^2b^2-10ab^3-b^4-4a^4)}{(b^2+2ab-a^2)(16a^3b-5a^2b^2-10ab^3-b^4-4a^4+6b^2)}
\end{equation}
\begin{equation}\frac{u_0}{v_0}=\frac{16a^3b-8a^2b^2-10ab^3-b^4-4a^4}{-16a^3b+5a^2b^2+10ab^3+b^4+4a^4-6b^2}
\end{equation}
\begin{equation}\frac{z_0}{v_0}=$$$$\frac{32(a^3b-ab^3)-2b^4-8a^4-22a^3b^3+10ab^5+b^6+7a^4b^2+8a^2b^4-10a^2b^2}{-16a^3b+5a^2b^2+10ab^3+b^4+4a^4-6b^2}
\end{equation}
\begin{equation}\frac{x_0}{v_0}=\frac{16ba^2(a^3-ab^2-b)-4a^2(b^4+a^4)+32ab(a^2-b^2)-8(a^4+b^4+a^4b^2)}{-16a^3b+5a^2b^2+10ab^3+b^4+4a^4-6b^2}
\end{equation}
\begin{equation}\frac{y_0}{v_0}=$$$$\frac{32a^3b-26ab^3-8(a^4+b^4)+6a^2b^2(2b^2-3a^2)+ab(b^4+4a^4)+10a^2b^2(ab-1)}{-16a^3b+5a^2b^2+10ab^3+b^4+4a^4-6b^2}.
\end{equation}
Applying (2.7), (2.8), (2.19), (3.1), (3.2), (3.3), (3.4), (3.5),
(3.6), (3.13) and (3.14), up to terms of order in $\varepsilon^2$
(2.16), (2.17), (2.18) and (2.19) become
\begin{equation}G\varrho(t,r)\approx\varepsilon(2ab+b^2)v(t,r),
\end{equation}
\begin{equation}
Gp_r(t,r)\approx\varepsilon\left(\frac{5a^2-10ab-3b^2}{12}\right)v(t,r),\end{equation}
\begin{equation}
Gp_t(t,r)\approx\varepsilon\left(\frac{a^2-4ab-b^2}{4}\right)v(t,r),\end{equation}
\begin{equation} GS(t,r)\approx\varepsilon
a\left(\frac{a+b}{6}\right)v(t,r)\end{equation} and
\begin{equation}Gp(t,r)\approx\varepsilon\left(\frac{3a^2-2ab-b^2}{4}\right)v(t,r).
\end{equation} State equation of the fluid become
\begin{equation}\gamma(a,b)=\frac{p}{\varrho}=\frac{1}{4}\left(\frac{3a^2-2ab-b^2}{2ab+b^2}\right).
\end{equation} Applying
(3.1), (3.13) and (3.14) the metric equation (2.13) become
\begin{equation} Gds^2\approx-(1+\varepsilon x_0 e^{a \tau+b\rho})d\tau^2+
2(1+\varepsilon y_0 e^{a\tau+b\rho})d\tau d\rho+(1+\varepsilon
z_0e^{a\tau+b\rho}) d\rho^2.\end{equation} Static regime of our
solution is obtained by setting $a=0$ in which barotropic index of
state equation (3.25) leads to $\gamma(0,b)=-0.25$ describing a dark
matter dominant of the gravitational system and (3.15) reduces to
the condition
$b=-\sqrt{\frac{6\omega-4}{\omega+1}},\omega>\frac{2}{3}$ where we
do not consider its positive sign root because of asymptotically
flat condition of the metric solution (3.26) at large scales
$\rho>>1.$ In the latter case (3.16), (3.17), (3.18) and (3.19)
become
\begin{equation} \frac{u_0}{v_0}=\frac{3\omega-2}{5},~~~
\frac{x_0}{v_0}=\frac{y_0}{v_0}=\frac{8(3\omega-2)}{5},~~~\frac{z_0}{v_0}=\frac{2(2-3\omega)(2\omega-3)}{5(\omega+1)}.\end{equation}
Position of event horizon is obtained from the equation $g_{tt}(\tau
, \rho)=0$ as
\begin{equation} a\tau+b\rho=-\ln|\varepsilon x_0|\end{equation}
which in static regime become
\begin{equation}\rho_{EH}=\sqrt{\frac{\omega+1}{6\omega-4}}\ln\left|\frac{4\varepsilon v_0}{5}(6\omega-4)\right|.\end{equation} Surface area of apparent horizon is
$4\pi\psi^2$ and its position is obtained from the equation
$g^{ab}\partial_a\psi\partial_b\psi=0$ as
\begin{equation} a\tau+b\rho=\ln\left|\frac{2(b^2+2ab-a^2)}{\varepsilon[(a^2+b^2)z_0+2(b^2+ab-a^2)y_0-(a-b)^2x_0]}\right|\end{equation}
which in static regime become \begin{equation}
\rho_{AH}=\sqrt{\frac{\omega+1}{6\omega-4}}\ln\left|\frac{4\varepsilon
v_0}{5}\frac{(3\omega-2)(\omega+5)}{(\omega+1)}\right|.\end{equation}
Eliminating $v_0$ between (3.29) and (3.31) we obtain relative
distance of apparent and event horizons positions as
\begin{equation}\rho_{EA}(\omega)=\rho_{EH}-\rho_{AH}=\sqrt{\frac{\omega+1}{6\omega-4}}\ln\left|10\left(\frac{\omega+1}{\omega+5}\right)\right|.
\end{equation} In general relativistic limits $\omega\to+\infty$ the above relation leads to $\rho_{EA}=0.94$ (see figure
1).\\
Let us see how our solutions work inside a spherically symmetric
relativistic star with total mass $M$ and radius $R.$ Take an
observer located at some distance $r$ from the star center. The
gravitational field there depends only on the mass function $m(r)$
up to the radius $r$ which is derived from the density function as
[29]
\begin{equation} m(r)=\int_0^r4\pi r^2\varrho(r)dr.\end{equation}
Therefore, one may use the metric of outside of star as
Schwarzschild like
with $g_{tt}^{outside}=-\left(1-\frac{2m(r)}{r}\right),$ because coefficients of inside
and outside metric of the star are equal on the surface of star. One
can use the junction conditions to match both inside and outside
metric of the star and obtain exact form of external Schwarzschild
like metric described in terms of external coordinated system.
However we will not calculate it here and our study is restricted to
obtain corresponding mass function of
collapsing fluid sphere only in what follows. \\
Applying (3.13), (3.20), (3.33)
and setting $a=0,$ $r=\sqrt{G}\rho,$
$b=-\sqrt{\frac{6\omega-4}{\omega+1}},$ one infer
\begin{equation} m(\rho)=4\pi\sqrt{G}\varepsilon v_0 b^2\int_0^\rho
\rho^2 e^{b\rho} d\rho\end{equation} leading to
\begin{equation} m^*(\rho)=\sqrt{\frac{\omega+1}{6\omega-4}}
\bigg\{1-\frac{1}{2}\bigg[\left(\frac{6\omega-4}{\omega+1}\right)^2-\left(\frac{6\omega-4}{\omega+1}\right)+2\bigg]
e^{-\sqrt{\frac{6\omega-4}{\omega+1}}\rho}\bigg\}\end{equation}
where we defined dimensionless mass function as
$m^*(\rho)=\frac{m(\rho)}{8\pi\sqrt{G}v_0\varepsilon}.$ Observer
which is located at infinity
$`\rho_{\infty}\sqrt{\frac{6\omega-4}{\omega+1}}>>1`$ with respect
to center of collapsing sphere evaluates its total mass such as
follows.
\begin{equation}M^*(\omega)=m^*(\infty)=\sqrt{\frac{\omega+1}{6\omega-4}}\end{equation}
which its diagram is plotted against $\omega$ in figure 2.
Evidently, the above total mass is eliminated with particular value
$\omega=-1$ and it takes an infinite value for $\omega=\frac{2}{3}.$
In the next section we study de-Broglie-Bohm approach of quantum
gravity of our collapsing fluid model.
\section{de Broglie-Bohm quantum gravity perspective} Using
the transformations\begin{equation}
\phi\psi^2=\left(\frac{3+2\omega}{2+\omega}\right)\frac{\mathfrak{R}}{4},~~~\psi=\sqrt{G}\mathfrak{R}^{\left(\frac{1+\omega}{3+2\omega}\right)}
e^{i\mathfrak{D}\frac{\sqrt{4+2\omega}}{3+2\omega}},
\end{equation} one can show that the action (2.3) is transformed to the well known action of Klein-Gordon
scalar field $\Psi(\psi,\phi)=\sqrt{\mathfrak{R}}e^{i\mathfrak{D}}$
moving on the particular conformal frame metric \begin{equation}
\tilde{g}_{ab}=\Omega^2g_{ab},~~~
~\Box\ln\Omega=-\frac{1}{\psi^2}\end{equation} in which covariant
conservation of current density
$J_a=\sqrt{\mathfrak{R}}\partial_a\mathfrak{D}$ of the particles
ensemble with radial velocity $V_a=\partial_a \mathfrak{D},$ is
satisfied [1]. Usually
$\Psi(\psi,\phi)=\sqrt{\mathfrak{R}}e^{i\mathfrak{D}}$ is called as
de Broglie pilot wave of the gravitational action (2.3) where
$\sqrt{\mathfrak{R}}$ and $\mathfrak{D}$ are its probability
amplitude and phase functionals respectively.\\
Applying (3.6) and (3.13) the equation (4.1) leads to the following
relations.
\begin{equation} \mathfrak{R}(\tau, \rho)=\left(\frac{8+2\omega}{3+2\omega}\right)\exp\{1+\varepsilon (u_0+v_0)e^{a\tau+b\rho}\}\end{equation}
and
\begin{equation}\mathfrak{D}(\tau,\rho)
=\frac{1}{\sqrt{(4+2\omega)(3+2\omega)}}\bigg\{\frac{1}{2}-(1+\omega)\ln\left(\frac{8+4\omega}{3+2\omega}
\right)$$$$+\varepsilon[(2+\omega)u_0-(1+\omega)v_0]e^{a\tau+b\rho}\bigg\}.
\end{equation}
Setting $a=0$ one can obtain static regime of the de Broglie wave
solution of the system given by (4.3) and (4.4) as
\begin{equation} \mathfrak{R}(\rho)=\left(\frac{8+2\omega}{3+2\omega}\right)
\exp\{1+\varepsilon (3v_0/5)(\omega+1)e^{-\sqrt{\frac{6\omega-4}{\omega+1}}\rho}\}\end{equation}
and \begin{equation}\mathfrak{D}(\rho)=\frac{1}{\sqrt{(4+2\omega)(3+2\omega)}}\bigg\{\frac{1}{2}-(1+\omega)\ln\left(\frac{8+4\omega}{3+2\omega}
\right)$$$$+\varepsilon(v_0/5)(3\omega^2-\omega-9)e^{-\sqrt{\frac{6\omega-4}{\omega+1}}\rho}\bigg\}
\end{equation}
in which we used (3.27). One of useful quantities which can be
consider here is evaluation of incident current density of particles
ensemble with the Plank mass $m_p=\frac{1}{\sqrt{G}}$ (in units
$\hbar=c=1$ ) on particular space-like hypersurface
$\rho=\rho_H=constant$ (for instance: the event and apparent
horizons). Applying the well known formula
$J_{incident}=\frac{i\hbar}{2m_p}\{\Psi\partial_{\rho}\Psi^*-\Psi^*\partial_\rho\Psi\}$
where $\Psi=\sqrt{\mathfrak{R}}e^{i\mathfrak{D}}$ we obtain
\begin{equation}
J_{incident}=\sqrt{\mathfrak{R}}\partial_{\rho}\mathfrak{D}_{|_{\rho=\rho_H}}\end{equation}
where $\partial_{\rho}\mathfrak{D}_{|_{\rho=\rho_H}}$ describes
velocity of particles ensemble on hyper-surface of the horizons
(event or apparent). Applying (4.5) and (4.6) the current density
(4.7) become
\begin{equation}
J_{incident}(\rho)=\left(\frac{9-\omega-3\omega^2}{3+2\omega}\right)
\sqrt{\frac{(4+\omega)(3\omega-2)}{(\omega+1)(\omega+2)}}\left(\frac{3\sqrt{2}v_0}{5}\right)\times$$$$
\exp\bigg\{\frac{1}{2}-\sqrt{\frac{6\omega-4}{1+\omega}}\rho+\varepsilon\left(\frac{3v_0}{10}\right)
(1+\omega)e^{-\sqrt{\frac{6\omega-4}{1+\omega}}\rho}
\bigg\}\end{equation} which at infinity $\rho\to+\infty$ leads to
$J_{incident}(+\infty)=0.$
Applying (3.29) and (3.31) we will have
\begin{equation}J_{incident}(\rho_{EH})=\left(\frac{9-\omega-3\omega^2}{3+2\omega}\right)
\sqrt{\frac{(4+\omega)}{(\omega+1)(\omega+2)(3\omega-2)}}
\left(\frac{3\sqrt{2}}{8\varepsilon}\right)e^{\frac{5}{8}\left(\frac{3\omega-1}{3\omega-2}\right)}
\bigg\}\end{equation} and
\begin{equation}J_{incident}(\rho_{AH})=\left(\frac{9-\omega-3\omega^2}{3+2\omega}\right)\sqrt{\frac{(4+\omega)(1+\omega)}{(3\omega-2)(\omega+2)}}
\left(\frac{3\sqrt{2}}{4\varepsilon}\right)\times$$$$
\exp\bigg\{\frac{15\omega^2+58\omega-37}{8(3\omega-2)(\omega+5)}\bigg\}\end{equation}
$V_a=\partial_a\mathfrak{D},$ given in the relation (4.7) is radial velocity of particles ensemble
which shows collapsing trajectories in direction orthogonal to wave
front $\mathfrak{D}=constant$ located on the horizons. \\
Applying (3.26) and (4.5) one can derive dimension-less quantum
potential $Q=-\frac{\Box\sqrt{\mathfrak{D}}}{\sqrt{\mathfrak{D}}}$
effecting on the particles ensemble in static regime of the system
in terms of dimensionless radial coordinate as
\begin{equation} Q(\rho)=-\varepsilon^2\left(\frac{3v_0}{5}\right)(3\omega-2)e^{-\sqrt{\frac{6\omega-4}{\omega+1}}\rho}\end{equation} which by applying
(3.29) and (3.31) reduces to the following values
\begin{equation}Q(\rho_{EH})=-\frac{3}{8}\varepsilon,~~~~Q(\rho_{AH})=-\frac{3}{4}\left(\frac{1+\omega}{5+\omega}\right)\varepsilon.
\end{equation} Obviously our results and deductions will be
corrected by regarding backreaction corrections of de Broglie-Bohm quantum potential effects
making quantum trajectories of the particles ensemble on the background metric solution (3.26), which is not
considered here and it will be considered as future work.
\section{Concluding remarks}
Two dimensional analogue of vacuum sector of the Brans Dicke scalar
tensor gravity model is used to solve its dynamical equations. Our
classical solutions describe anisotropic perfect fluid collapsing
sphere which treats as a dark matter gravitational source in its
static regime. The fluid can be treat as regular visible matter in
its dynamical non-static regime. Total mass of the fluid is obtained
in terms of the Brans Dicke parameter $\omega$ leading to conditions
$\omega\leq-1$ and $\omega>\frac{2}{3}.$ In general relativistic
approach in which $\omega\to+\infty$ dimensionless normalized total
mass of the fluid leads to $M^*=\frac{1}{\sqrt{6}}.$ It is
eliminated with $\omega=-1$ and takes infinite value with
$\omega=\frac{2}{3}.$ de Broglie-Bohm quantum gravity approaches of
the model is also studied. Its de Broglie pilot wave is obtained
from which incident current density of particles ensemble is
evaluated on both event and apparent horizons describing the thermal
`Hawking radiation` in statistical mechanics point of view.
Mathematical calculations in both of classical and quantum regime
predicts that the spherical symmetric stellar collapse with
dilaton-Brans-Dicke scalar source will be reach to black hole
structure finally where cosmic censorship hypothesis maintained
valid still and causal singularity $r=0$ is covered by the black
hole event horizon.
\vskip 0.5 cm
{\bf References}
\begin{description}
\item[1.] H. Ghaffarnejad, Int. J.
Theor. Phys. 53, 2616 (2014), gr-qc/1309.6157 v2 (2014).
\item[2.] R. M. Wald, \textit{General Relativity}, (The University of
Chicago Press, Chicago, 1984).
\item [3.] P. S. Joshi, \textit{Gravitational Collapse and Space-time
Singularities}, (Cambridge University Press, 2007).
\item[4.]T. P. Singh, J. Astrophys. Astr., 20, 221, (1999).
\item[5.]A. H. Ziaie, Kh. Atazadeh and Y. Tavakoli, Class. Quantum Grav. 27, 075016 (2010); Erratum-ibid.27, 209801
(2010); gr-qc/1003.1725v3.
\item[6.] R. Penrose, Rev. Nuovo Cimento,1, 252 (1969).
\item[7.] A. Ori and T. Piran, Phys. Rev. D42, 1068 (1990).
\item[8.] H. Ghaffarnejad, H. Neyad and M. A. Mojahedi, Astrophys.
Space Sci, 346, 497 (2013), arXiv:1305.6914 [physics.gen-ph].
\item[9.] K. S. Virbhadra and C. R. Keeton, Phys. Rev. D77, 124014
(2008), gr-qc/0710.2333.
\item[10.] K. S. Virbhadra and G. F. R. Ellis, Phys. Rev. D65, 103004
(2002).
\item[11.]F. C. Mena, B. C. Nolan and R. Tavakol, Phys. Rev. D70,
084030, (2004), gr-qc/0405041.
\item[12.]J. M. Martin-Garcia and C. Gundlach, Phys. Rev. D68, 024011
(2003).
\item[13.] K. D. Patil, Phys. Rev. D67, 024017 (2003).
\item[14.]P. S. Joshi, Phys. Rev. D51, 6778 (1995).
\item[15.]J. Lemos, Phys. Rev. Lett.68, 1447 (1992).
\item[16.] P. S. Joshi and I. H. Dwivedi, Commun. Math.Phys. 146,
333 (1992).
\item[17.] S. G. Ghosh and N. Dadhich, Gen. Relativ. Gravit35, 359
(2003).
\item[18.] P. Szekeres and V. Lyer, Phys. Rev. D47, 4362 (1993).
\item[19.] C. Brans and R. Dicke, Phys. Rev. 124, 925 (1961).
\item[20.] C. M. Will, \textit{Theory And Experiment In Gravitational
Physics}, Cambridge University press (1993); revised version:
gr-qc/9811036.
\item[21.] E. Gaztanaga and J. A. Lobo, Astrophys. J., 548, 47
(2001).
\item[22.] R. D. Reasenberg et al, Astrophys. J., 234, 925 (1961).
\item[23.] C. M. Will, Living Rev. Rel. 9 (2006);
http://WWW.livingreviews.org/lrr-2006-3.
\item[24.] L. J. Garay and J. G. Bellido, gr-qc/9209015 (1992).
\item[25.] Y. M. Chi, Class Quantum Gravi, 14, 10, 2963 (1997).
\item[26.] H. Culetu, gr-qc/1101.2980 (2011)
\item[27.] H. Culetu, hep-th/0711.0062 (2007).
\item [28.] T. Koivisto and D. F. Motta, astro-ph/ 0801.3676 (2008).
\item[29.] S. Weinberg, \textit{Gravitation and Cosmology}, John
Wiley and Sons, (1972).
\end{description}
\begin{figure}
\hspace{0cm} \includegraphics[width=13.5cm]{figure_1.eps}
\hspace{0cm} \caption{\label{mutbeta.figs.} Diagram shows relative
distance positions of event horizon with respect to apparent horizon
$\rho_{EA}=\rho_{EH}-\rho_{AH}$ against the Brans-Dicke parameter
with solid line. Also values of quantum potential
$\frac{Q}{\varepsilon}$ is plotted against $\omega$ with dash
(dash-dot) lines on the apparent (event) horizons. It is given here
to compare with $\rho_{EA}.$}
\end{figure}
\begin{figure}
\hspace{0cm} \includegraphics[width=13.5cm]{figure_2.eps}
\hspace{0cm} \caption{\label{mutbeta.figs.} Diagram of normalized
total mass $M^*=\frac{M}{8\pi \sqrt{G}v_0\varepsilon}$ given by
(3.36) is plotted against $\omega$ with dot-line. It is comparable
with diagrams of quantum potential (given on the apparent horizon
(dash line) and event horizon (solid line)) plotted against
$\omega.$ }
\end{figure}
\begin{figure}
\hspace{0cm} \includegraphics[width=13.5cm]{figure_3.eps}
\hspace{0cm} \caption{\label{mutbeta.figs.}Diagram of incident
current density of particles ensemble $\varepsilon J_{incident}$
(see Eqs. (4.9) and (4.10)) on the event (apparent) horizon
hyper-surface with solid (dot) line. Also quantum potential values
$\frac{Q}{\varepsilon}$ (see Eqs. (4.12)) is plotted against
$\omega$ on the apparent (event) horizons with dash (dash-dot) lines
. This is given to compare with diagram of the current density. }
\end{figure}
\end{document}
|
\section{Introduction}
\label{sec:introduction}
Integer linear programming problems are widely described in the
combinatorial optimisation literature, and include many well-known and
important applications. Typical problems of this type include lot
sizing, scheduling, facility location, vehicle routing, and more; see
for example \cite{IntCombOptim,lscopt}. The problem consists of
optimising a linear function subject to a set of linear constraints,
in the presence of integer and, possibly, continuous variables. If
the subset of continuous variables is empty, the problem is called
\emph{pure integer} (IP). In the more general case, where there are
also continuous variables, the problem is usually called \emph{mixed
integer} (MIP).
The general formulation of a mixed integer linear program is
\begin{equation}
\label{eq:mip}
\max_{x,y}\{c x + h y : A x + G y \leq b,
x \in {\mathchoice {\hbox{$\sf\textstyle Z\kern-0.4em Z$}_+^n, y \in \bbbr_+^p\}
\end{equation}
where ${\mathchoice {\hbox{$\sf\textstyle Z\kern-0.4em Z$}_+^n$ is the set of nonnegative integral $n$-dimensional
vectors and $\bbbr_+^p$ is the set of nonnegative $p$-dimensional
vectors. $A$ and $G$ are $m \times n$ and $m \times p$ matrices,
respectively, where $m$ is the number of constraints. The integer
variables are $x$, and the continuous variables are
$y$.
\subsection{The evolutionary structure}
The main idea for the conception of the algorithm described here is
that if the integer variables of a MIP are fixed, what remains to
solve is a standard LP problem; this can be done exactly and
efficiently, for example by the simplex algorithm or by interior point
methods. We are therefore able to make the integer variables evolve
through an evolutionary algorithm (EA); after they are fixed by the
EA, we can determine the continuous variables in function of them.
\subsection{Branch-and-bound.}
The most well known algorithm for solving MIPs is branch-and-bound
(B\&B) (for a detailed description see, for
example,~\cite{IntCombOptim}). This algorithm starts with a
continuous relaxation of the MIP, and proceeds with a systematic
division of the domain of the relaxed problem, until the optimal
solution is found. There are two main advantages of the B\&B
algorithm. The first and most important is that its solution is
optimal (or there are no feasible solutions); the other is that some
nodes of the B\&B exploration graph can be pruned, and therefore the
algorithm's speed and memory requirement improved. These are two
important reasons to dissuade the application of an EA for the same
purpose: EAs cannot prove that the solution found is optimal, and in
what concerns convergence the best that can be proved is that, for
elitist EAs, we obtain a sequence of evaluations which converges to
the optimal objective value as the number of generations tends to
infinity.
We believe that it is nevertheless worthy to try to use an EA for this
type of problems, because of two other important reasons. The first
is that it is easy to incorporate in the EA a problem-specific local
search method, possibly working on primal solutions, taking advantage
of the problem structure; this could provide a speedup of one order of
magnitude. The second reason is that in some cases B\&B fails to find
a good feasible solution, sufficient for most practical applications,
in a reasonable computational time. It can be hoped that an EA does
better than B\&B in these cases.
\subsection{Benchmark problems}
Instances of integer linear problems correspond to specifications of
the data: the matrices $A$ and $G$, and the vectors $b$, $c$ and $h$
in equation~\ref{eq:mip}. The most commonly used representation of
instances of these problems is through \emph{MPS files}. The format
of these files has the advantage of being standard, and hence readable
by most of the solvers; the disadvantage being that it can not provide
information concerning the specific characteristics of the problem.
We have tested the EA with a subset of the benchmark problems that are
available in the \emph{MIPLIB}~\cite{miplib}. These problems range
from the moderately easy to the very difficult, for the solution
techniques available nowadays.
\section{The evolutionary operators}
\label{sec:evol-op}
Evolutionary algorithms function by maintaining a set of solutions,
generally called a \emph{population}, and making these solutions
evolve through operations that mimic the natural evolution:
reproduction, and selection of the fittest. Some of these operators
where customised for the concrete type of problems that we are
dealing with; we focus on each of them in the following
sections.
\subsection{Representation of the solutions}
\label{sec:representation}
The part of the solution that is determined by the EA is the subset of
integer variables, $x$ in equation~\ref{eq:mip}. Integer variables
are fixed by the EA, leading to an LP with only the continuous
variables, $y$, free; these are determined afterwards by solving a
linear problem.
We use the term \emph{individual} to mean a solution of the original
mixed-integer problem, and the term \emph{genome} to mean the subset
of integer variables of that solution. The solution corresponding to
a particular individual is represented in the EA by its genome, an
$n$-dimensional vector $\bar{x}^{E\hspace{-.6mm}A} = (\bar{x}^{E\hspace{-.6mm}A}_1 \ldots \bar{x}^{E\hspace{-.6mm}A}_n)$; we call each
$\bar{x}^{E\hspace{-.6mm}A}_k$ a \textit{chromosome}.
An individual $i$ kept in the algorithm's population is hence
represented by the vector of integer variables $\bar{x}^{{E\hspace{-.6mm}A}^i}$, and the
corresponding vector of continuous variables $\yea^i$ is determined by
an LP solver, at the time of its evaluation.
\subsection{Evaluation of individuals}
\label{sec:objective}
The solutions that are kept by the algorithm---or, in other words, the
individuals that compose the population---may be feasible or not. For
the algorithm to function appropriately it has to be able to deal with
both feasible and infeasible individuals coexisting in the population.
In the process of evaluation of an individual, we first formulate an
LP by fixing all the variables of the MIP at the values of the
individual's genome:
\begin{equation}
\label{eq:lp-conv}
z = \max_{y}\{c \bar{x}^{E\hspace{-.6mm}A} + h y : G y \leq b - A \bar{x}^{E\hspace{-.6mm}A}, y \in \bbbr_+^p\}
\end{equation}
We are now able to solve this (purely continuous) linear problem using
a standard algorithm, like the simplex.
\subsubsection{Feasible solutions}
If problem~\ref{eq:lp-conv} is feasible, the evaluation (fitness)
attributed to the corresponding individual is the objective value $z$,
and the individual is labelled feasible. We denote this fitness by
$\zea$, a data structure consisting of the objective value and a
flag stating that the solution is feasible.
\subsubsection{Infeasible solutions}
If problem~\ref{eq:lp-conv} is infeasible, we formulate another LP for
the minimisation of the infeasibilities. This is accomplished by
setting up artificial variables and minimising their sum (a procedure
that is identical to the phase I of the simplex algorithm):
\begin{equation}
\label{eq:lp-infeas}
\zeta = \min_{s}\{\sum_{k=1}^{m} s_k : G y \leq b - A \bar{x}^{E\hspace{-.6mm}A} + s,
y \in \bbbr_+^p \;,\; s \in \bbbr_+^m\}
\end{equation}
where $m$ is the number of constraints.
The evaluation attributed to such an individual is the value~$\zeta$
of the optimal objective of the LP of equation~\ref{eq:lp-infeas}, and
the individual is labelled infeasible. The fitness data structure
$\zea$ consists of the value~$\zeta$ and an infeasibility
flag.
\subsubsection{Comparison and selection of individuals}
\label{sec:comparison}
For the selection of individuals, we have to provide a way for
comparing them, independently of the corresponding solutions being
feasible or not. What we propose is to rank the solutions, so that:
feasible solutions are always better than infeasible ones, feasible
solutions are ranked among them according to the objective of the MIP
problem, and infeasible solutions are ranked among them according to
the sum of infeasibilities (i.e., according to a measure of their
distance from the feasible set). For this purpose, we define an
operator to compare two individuals. We say that $\bar{z}^{{E\hspace{-.6mm}A}^i} \succ \bar{z}^{{E\hspace{-.6mm}A}^j}$
($i$ \emph{is better} than $j$) iff:
\begin{itemize}
\item $i$ is feasible and $j$ is infeasible;
\item $i$ and $j$ are feasible, and $z^i > z^j$ ($i$ has a better
objective);
\item $i$ and $j$ are infeasible, and $\zeta^i < \zeta^j$
($i$ is closer to the feasible region than~$j$).
\end{itemize}
As there is the possibility that both feasible and infeasible
individuals coexist in the population, their fitness cannot be
attributed as in common EAs, based only on the value of an objective
function. Therefore, selection of an individual has to be (directly
or indirectly) based on its ranking in the population, which can be
determined through the comparison operator defined above (see also
section~\ref{sec:selection}).
\subsection{Initialisation}
\label{sec:initialisation}
The population that it used at the beginning an evolutionary process
is usually determined randomly, in such a way that the initial
diversity is very large. In the case of MIP, it is appealing to bias
the initial solutions, so that they are distributed in regions of the
search space that are likely to be more interesting. A way to provide
this bias, inspired in an algorithm provided in~\cite{gunluk96}, is to
firstly solve the LP relaxation of the problem, and then round the
solutions obtained to one of the closest integers. The probabilities
for rounding up or down each of the variables are given by the
distance from the fractional solution to its closest integer points.
If we denote the solution of the LP relaxation by $x^{L\hspace{-.6mm}P} = (x^{L\hspace{-.6mm}P}_1
\ldots x^{L\hspace{-.6mm}P}_n)$, each element of the initial population will be
determined as follows. For all the chromosomes $k \in \{1, \ldots,
n\}$, the corresponding variable $\bar{x}^{E\hspace{-.6mm}A}_k$ is rounded down with
probability
$$P(\bar{x}^{E\hspace{-.6mm}A}_k = \lfloor x^{L\hspace{-.6mm}P}_k \rfloor) = x^{L\hspace{-.6mm}P}_k - \lfloor x^{L\hspace{-.6mm}P}_k
\rfloor$$
or rounded up with probability $1 - P(\bar{x}^{E\hspace{-.6mm}A}_k = \lfloor
x^{L\hspace{-.6mm}P}_k \rfloor)$.
\subsection{The genetic operators}
\label{sec:gen-operators}
The generation of a new individual from two parents is composed of
three steps: recombination (meiosis and crossover), possibly followed
by mutation, followed by local search. Each of the genetic operators
is controlled by two parameters: probability of occurrence and
intensity of the operation.
We use the following notation: $\nu^p$, $\chi^p$, $\mu^p$, are the
probabilities of meiosis, crossover, and mutation, respectively;
$\nu^s$, $\chi^s$, $\mu^s$ are their respective intensities. The
distribution of the perturbations added by mutation is $\delta(s) = 1
- \rand^{s^2}$, were $s$ is the intensity and $\rand$ is a random
number uniformly distributed in $[0,1]$. The value of $\delta(s)$ is
scaled, so that it covers the whole region between the value $\bar{x}^{E\hspace{-.6mm}A}_k$
and its bounds.
The process of reproduction for creating a new genome $\bar{x}^{E\hspace{-.6mm}A}$ from two
parents $\bar{x}^{{E\hspace{-.6mm}A}^f}$ and $\bar{x}^{{E\hspace{-.6mm}A}^m}$ is presented in figure~\ref{fig:genetic}.
\begin{figure}[htbp!]
\begin{center}
\begin{minipage}[tl]{115mm}
\begin{program}
\item[select parents ($\bar{x}^{{E\hspace{-.6mm}A}^f}$, $\bar{x}^{{E\hspace{-.6mm}A}^m}$)] \comment{\ }
\item[procedure Reproduce($\bar{x}^{{E\hspace{-.6mm}A}^f}$, $\bar{x}^{{E\hspace{-.6mm}A}^m}$)] \comment{\ }
\item[if $\rand<\nu^p$] \comment{Do the meiosis with probability $\nu^p$}
\begin{program}
\item[for $k=1$ to $n$ do ] \comment{\ }
\begin{program}
\item[set $p := \rand (n-k+1) (1-\nu^s)$] \comment{Determine the
size of the ``path'' to select from one of the parents
(inversely proportional to the intensity of meiosis)}
\item[if $\rand < 1/2$] \comment{With 50\% probability, copy from
the father ($\bar{x}^{{E\hspace{-.6mm}A}^f}$)}
\begin{program}
\item[for $l=1$ to $p$ do] \comment{\ }
\begin{program}
\item[if $\rand<\chi^p$] \comment{With some probability do crossover,}
\begin{program}
\item[set $\bar{x}^{E\hspace{-.6mm}A}_k := \bar{x}^{{E\hspace{-.6mm}A}^f}_k + (\bar{x}^{{E\hspace{-.6mm}A}^m}_k - \bar{x}^{{E\hspace{-.6mm}A}^f}_k)\; \chi^s \rand$]
\comment{with intensity $\chi^s$}
\end{program}
\item[else] \comment{\ }
\begin{program}
\item[set $\bar{x}^{E\hspace{-.6mm}A}_k := \bar{x}^{{E\hspace{-.6mm}A}^f}_k$] \comment{No crossover, exact copy of
$\bar{x}^{{E\hspace{-.6mm}A}^f}_k$}
\end{program}
\end{program}
\item[done] \comment{\ }
\end{program}
\item[else] \comment{With 50\% probability, copy from
the mother ($\bar{x}^{{E\hspace{-.6mm}A}^m}$)}
\begin{program}
\item[. . .] \comment{(swap the roles of $\bar{x}^{{E\hspace{-.6mm}A}^f}$ and $\bar{x}^{{E\hspace{-.6mm}A}^m}$)}
\end{program}
\item[end if] \comment{\ }
\end{program}
\item[done] \comment{\ }
\end{program}
\item[else ] \comment{In this case, no meiosis occurs:}
\begin{program}
\item[set $\bar{x}^{E\hspace{-.6mm}A} := \bar{x}^{{E\hspace{-.6mm}A}^f}$, or $\bar{x}^{E\hspace{-.6mm}A} := \bar{x}^{{E\hspace{-.6mm}A}^m}$] \comment{copy exactly
$\bar{x}^{{E\hspace{-.6mm}A}^f}$ or $\bar{x}^{{E\hspace{-.6mm}A}^m}$, with same probability}
\end{program}
\item[end if] \comment{\ }
\item[for $k=1$ to $n$ do ] \comment{Now, do the mutation: for each
element of $\bar{x}^{E\hspace{-.6mm}A}$,}
\begin{program}
\item[if $\rand < \mu^p$] \comment{with probability $\mu^p$, add
mutation}
\begin{program}
\item[set $\bar{x}^{E\hspace{-.6mm}A}_k := $round($\bar{x}^{E\hspace{-.6mm}A}_k \pm \delta(\mu^s)$)]
\comment{of intensity $\mu^s$, and round to
nearest integer.}
\end{program}
\end{program}
\item[done] \comment{\ }
\item[end procedure] \comment{\ }
\end{program}
\end{minipage}
\caption{Pseudo programming code for the genetic operations.}
\label{fig:genetic}
\end{center}
\end{figure}
The recombination process produces a linear combination of the genomes
of two individuals selected from the population, and is based on two
sub-operations: meiosis and crossover. Given two progenitor genomes,
the meiosis consists of selecting "paths", or sequences of $\bar{x}^{E\hspace{-.6mm}A}_k$'s,
alternately from each of them, to create a new genome. The greater
the meiosis intensity ($\nu^s$), the smaller these paths are likely to
be. Crossover consists of, for each of the chromosomes (indices of
the genome vector), perturbing the value obtained by meiosis in the
direction of its value for the other progenitor. The smaller the
\emph{crossover intensity} parameter is, the closer the produced
chromosome is to that of one of the parents.
The mutation adds a random perturbation to the genome created this
way. For each mutation we randomly choose, with identical
probability, to add or subtract $\delta(s)$ to the value of the
chromosome, where $s$ is the intensity, or magnitude of the mutation.
We then round the value to the closest integer.
Local search tries to improve the newly created individual's
performance by hill climbing in its neighbourhood, as described below.
\subsection{Local search}
\label{sec:local-search}
We propose a rather rough---but general---local search method, for
hill climbing in the integer variables space. This search is
performed whenever a new individual is created. It is based on what
is called \emph{hunt search}, originally conceived for locating values
in an ordered table. The idea is to check for improvements in the
objective when each of the $n$ integer variables $\bar{x}^{E\hspace{-.6mm}A}_k$ is
independently perturbed, with a geometrically increasing step.
The algorithm is the presented in figure~\ref{fig:local-search}. Note
that this local search method is completely problem-independent, and
that its use does not exclude the possibility of using an additional,
problem-specific local search method to speedup the search.
\begin{figure}[htbp!]
\begin{center}
\begin{minipage}[t]{115mm}
\begin{program}
\item[procedure Local\_Search($\bar{x}^{E\hspace{-.6mm}A}$)] \comment{\ }
\item[set $\bar{x}^{{L\hspace{-.6mm}S}} := \bar{x}^{E\hspace{-.6mm}A}$] \comment{Start the local search at the
individual's solution}
\item[for $k=1$ to $n$, do] \comment{For all the integer variables (in a
random order):}
\begin{program}
\item[set $\bar{x}^{{L\hspace{-.6mm}S}}_k := \bar{x}^{E\hspace{-.6mm}A}_k+1$, $\bar{z}^{\mathrm{up}} :=$ LPsolution($\bar{x}^{{L\hspace{-.6mm}S}}$)]
\comment{Evaluate the perturbed solution}
\item[set $\bar{x}^{{L\hspace{-.6mm}S}}_k := \bar{x}^{E\hspace{-.6mm}A}_k-1$, $\bar{z}^{\mathrm{down}} :=$ LPsolution($\bar{x}^{{L\hspace{-.6mm}S}}$)]
\comment{by solving the corresponding LP}
\item[if $\zea \succ \bar{z}^{\mathrm{up}}$ and $\zea \succ \bar{z}^{\mathrm{down}}$] \comment{Solution
is a local optimum with respect to index $k$}
\begin{program}
\item[continue with next $k$] \comment{\ }
\end{program}
\item[if $\bar{z}^{\mathrm{up}} \succ \bar{z}^{\mathrm{down}}$] \comment{Stepping up is better than
stepping down}
\begin{program}
\item[set ${\mathit{step}} := 1$] \comment{\ }
\end{program}
\item[else] \comment{\ }
\begin{program}
\item[set ${\mathit{step}} := -1$] \comment{\ }
\end{program}
\item[while improving $\zea$ do] \comment{\ }
\begin{program}
\item[set $\bar{x}^{{L\hspace{-.6mm}S}}_k := \bar{x}^{E\hspace{-.6mm}A}_k+{\mathit{step}}$, $\bar{z}^{{L\hspace{-.6mm}S}} :=$
LPsolution($\bar{x}^{{L\hspace{-.6mm}S}}$)] \comment{\ }
\item[if $\bar{z}^{{L\hspace{-.6mm}S}} \succ \zea$] \comment{An improvement was
found:}
\begin{program}
\item[set $\bar{x}^{E\hspace{-.6mm}A}_k := \bar{x}^{{L\hspace{-.6mm}S}}_k$, $\zea := \bar{z}^{{L\hspace{-.6mm}S}}$] \comment{update
the solution}
\end{program}
\item[set ${\mathit{step}} := 2 \times {\mathit{step}}$]
\comment{Geometrically increase the step}
\end{program}
\item[done] \comment{\ }
\end{program}
\item[done] \comment{\ }
\item[end procedure]
\end{program}
\end{minipage}
\caption{The local search procedure.}
\label{fig:local-search}
\end{center}
\end{figure}
\section{Niche search}
\label{sec:niche-search}
Niche search is an evolutionary algorithm where the total population
is grouped into separate niches, each of which evolves independently
of the others for some (sub-)generations. The claim is that this way,
as the global evolutionary search pursues, more localised searches are
done inside each of the niches. The algorithm is therefore expected
to keep a good compromise between intensification of the search and
diversification of the population. This method has some similarities
with that described in~\cite{muhlenbein94}, where \emph{competing
subpopulations} play a role similar to that of the niches. An
application of niche search to a specific combinatorial optimisation
problem has been shown in~\cite{pedroso98icec}; here, it is extended
to the general MIP case.
Niches are subject to competition between them. The bad niches (i.e.,
those which have worse populations) tend to extinguish: they are
replaced by new ones, which are formed by elements selected from a
``good'' niche and the extinguishing one. All the parameters that
control the genetic operators described in
section~\ref{sec:gen-operators} (mutation intensity and probability,
etc.), together with a selectivity factor, are assigned exogenously
and randomly to each newly created niche. (The selectivity determines
how good an individual must be in relation to the average of the niche
in order to have a favoured probability of being selected for
reproducing.)
\subsection{Niche search core algorithm}
\label{sec:algorithm}
We summarise now the main steps of functioning for the niche search
algorithm. This is the kernel algorithm, which drives the population
operations making use of the solution representation and genetic
operators described in the preceding sections. Niche search is
characterised by evolution in two layers: in the higher layer, there
is the evolution of niches, subject to competition between them. Each
iteration of this process is called a \emph{niche generation}, or
simply a generation. In the lower layer, the individuals that
compose each niche evolve inside it, competing with other individuals
of the niche. Each iteration of this lower layer process is called an
\emph{individual's generation}, or a subgeneration.
The code describing the evolution of the set of niches, in what we
call a niche generation, is presented in figure~\ref{fig:niche-search}.
\begin{figure}[htbp!]
\begin{center}
\begin{minipage}[tl]{115mm}
\begin{program}
\item[set t := 0] \comment{Start with an initial time.}
\item[niches(t) = CreateNiches(t)] \comment{Create desired number
of niches for the run.}
\item[InitParameters(niches(t))] \comment{Randomly
initialise the parameters that characterise each niche:
crossover probability and intensity, mutation probability and
intensity, etc.}
\item[InitialisePopulation(niches(t))] \comment{Randomly
initialise the pop.\ of each niche.}
\item[Evaluate(niches(t))] \comment{Evaluate the fitness of all the
niches in the initial population. For evaluating a niche, we
used the fitness of its best element (other strategies
are also possible).}
\item[iterate] \comment{Start evolution.}
\begin{program}
\item[Breed(niches(t))] \comment{Create a new generation of
individuals in each of the niches, through the lower layer
evolution process described below.}
\item[Evaluate(niches(t))] \comment{Evaluate the new niches.}
\item[set weak(t) := SelectWeak(niches(t))] \comment{Select
niches that will extinguish.}
\item[set strong(t) := SelectStrong(niches(t))] \comment{Select
niches that will be used for generating new niches.}
\item[set newniches(t) := Recombination(weak(t),strong(t))]
\comment{Create a new niche for replacing each of the
extinguishing ones. The recombination strategy used is to
create a population formed of the union of the weak niche with
a strong one. Then, replace the individuals of the weak niche
by a selection of the best individuals from that population.}
\item[InitParameters(newniches(t))] \comment{Assign random
parameters to new niches.}
\item[Evaluate(newniches(t))] \comment{Evaluate the new niches.}
\item[Extinguish(weak(t), niches(t))] \comment{Remove weak
niches from the population}
\item[Insert(newniches(t), niches(t))] \comment{and include the
newly created ones.}
\item[set niches(t+1) := niches(t)] \comment{ }
\item[set t := t + 1] \comment{Increase the time counter.}
\end{program}
\item[until Terminated()] \comment{Termination criteria: number of
generations completed.}
\item[display solution] \comment{Solution is the best individual
found.}
\end{program}
\end{minipage}
\caption{Niche search: evolution of niches.}
\label{fig:niche-search}
\end{center}
\end{figure}
We now turn to the evolution of the individuals inside each of the
niches. Pseudo-programming code describing how individuals breed at
each generation of the inside-niche evolution (i.e., describing what a
\emph{subgeneration} is) is presented in
figure~\ref{fig:inside-niche}. Note that this process is repeated
for each of the niches, at each niche generation.
\begin{figure}[htbp!]
\begin{center}
\begin{minipage}[tl]{115mm}
\begin{program}
\item[Procedure Breed(niches(t))] \comment{ }
\item[for all niche in niches(t) do] \comment{(\textup{\textsf{t}}
is the \emph{niche generation} counter).}
\begin{program}
\item[set g := 0] \comment{Initialise the ``subgeneration''
counter.}
\item[set population(g) := niche] \comment{Set the reference
population: (only) the elements of the niche that is now
breeding.}
\item[iterate] \comment{Start evolution.}
\begin{program}
\item[for all element in offspring(g) do] \comment{ }
\begin{program}
\item[$p_1$ = Selection(population(g))] \comment{Select
parents for reproduction (in our imple-}
\item[$p_2$ = Selection(population(g))]
\comment{mentation through roulette wheel
selection).}
\item[set element := Reproduce($p_1, p_2$)] \comment{Create
the offspring using the}
\end{program}
\item[done] \comment{ operators described in
(section\protect~\ref{sec:local-search}).}
\item[Evaluate(offspring(g))] \comment{Evaluate the objective of
all the individuals in the niche's population. Scale to
obtain the fitnesses (section\protect~\ref{sec:selection}).}
\item[set population(g+1) := offspring(g)]
\comment{Future population is the offspring.}
\item[set g := g + 1] \comment{Increase the
subgeneration counter.}
\end{program}
\item[until Terminated()] \comment{Termination criteria: best
individual has not improved.}
\item[set niche := population(g)]
\comment{Update niche's population. This niche is now ready
to start}
\end{program}
\item[done] \comment{competition with the others.}
\item[end procedure] \comment{ }
\end{program}
\end{minipage}
\caption{Niche search: evolution inside the niches.}
\label{fig:inside-niche}
\end{center}
\end{figure}
\subsection{Selection in each niche: rank-based fitnesses}
\label{sec:selection}
As explained in section~\ref{sec:objective}, the solution process is
divided into two goals: obtaining feasibility and optimisation. This
has motivated the implementation of an order-based fitness attribution
scheme. The selection of the individuals that are able to reproduce
at each generation is based on a fitness value, called
\textit{rank-fitness}, that is proportional to their ranking
according to the comparison operator defined in
section~\ref{sec:comparison}.
In niche search there is a parameter of each niche, called the
\textit{selectivity}, that controls the probability of selection of
each individual in relation to their competitors. If this parameter
is very low, then the probability of selection of the best individuals
is only slightly greater than the probability of selection of the
worst; if it is high, then the best individuals have a much greater
probability of selection, what means that the ``genetic information''
of the worse ones is not likely to propagate to the future
generations.
In a niche with $n$ elements, the best of them is assigned a
rank-fitness of $1$ (i.e., $n/n$), the second-best $(n-1)/n$, up to
the worse, whose rank-fitness is $1/n$. We then elevate this value to
a power, greater or equal to zero---the selectivity parameter of the
niche---to obtain the scaled-fitness of each individual. The
selection is then performed through roulette wheel selection, giving
to each individual a probability of selection proportional to its
scaled-fitness (see, for example,~\cite{goldberg89} for a description
of roulette wheel selection).
\subsection{Elitism}
\label{sec:elitism}
Elitism determines whether the best solution found so far by the
algorithm is kept in the population or not. Elitism generally
intensifies the search in the region of the best solution. As
mentioned before, niche search keeps several groups, or niches,
evolving with some independence. Each of these groups may be elitist
(keeping \emph{its} best element in its population) or not.
Our objectives are two fold: we want the search to be as deep as
possible around good regions, but we do not want to neglect other
possible regions. The strategy that we devised for accomplishing this
is the following. Niches whose best individual is different of the
best individual of other niches are elitist, but when several niches
have an identical best individual (and this occurs frequently), only
one of them is elitist. With this strategy we hope to have an
intensified search on regions with good solutions, and at the same
time enforce a good degree of diversification.
\section{Numerical results}
\label{sec:results}
The instances of MIP problems used as benchmarks are defined in the
MIPLIB~\cite{miplib}. The evolutionary system starts by reading an
MPS file, and stores the information contained there into an internal
representation. The number of variables and constraints, their type
and bounds, and all the matrix information is, hence, determined at
runtime.
Note that the LPs solved by the EA are often much simpler than those
solved by B\&B; as all the integer variables are fixed, its size may
be much smaller (for a large proportion of integer variables).
Therefore, it is not surprising that numerical problems that the LP
solver may show up in B\&B, generally do not arise for LPs formulated
by the EA.
\subsection{Branch-and-bound}
In our implementation we have used a publicly available LP solver
called \emph{lp\_solve}~\cite{lp-solve} for the solution of the linear
programs. This solver also comprises an implementation of the B\&B
algorithm, that was used for producing results to compare with the
evolutionary algorithm.
The B\&B scheme consists on depth-first search, branching on the first
non-integer variable. Results obtained using B\&B on the series of
benchmark problems selected are provided in table~\ref{tab:bb-sols}.
The maximum number of LPs solved in B\&B was limited to 50 million; in
cases where this was exceeded, the best solution found within that
limit is reported.
\begin{table}[htbp!]
\begin{center}
\leavevmode
\begin{tabular}[t]{|l|c|c|l|}
\hline
Problem & Best solution & Number of & Remarks \\
name & found & LPs solved & \\
\hline \hline
bell3a & 878430.316 & 438737 & Optimal \\
bell5 & 8966406.492 & 2159885 & Optimal \\
egout & 562.27 & 55057 & Solution incorrect (rounding problems?) \\
enigma & 0 & 9321 & Optimal \\
flugpl & 1201500 & 2179 & Optimal \\
gt2 & -- & -- & Failed (unknown error) \\
lseu & 1120 & 252075 & Optimal \\
mod008 & 307 & 2848139 & Optimal \\
modglob & 27124594.43 &$>$50000000 & Stopped (excessive CPU time) \\
noswot & -23.0 (infeas.) & 3042 & Failed (numerical instability?)\\
p0033 & 3089 & 7409 & Optimal \\
pk1 & 12 & 704208 & Failed (numerical instability)\\
pp08a & 9880 &$>$50000000 & Stopped (excessive CPU time) \\
pp08acut & 7900 &$>$50000000 & Stopped (excessive CPU time) \\
rgn & 82.2 & 4747 & Optimal \\
stein27 & 18 & 12031 & Optimal \\
stein45 & 30 & 235087 & Optimal \\
vpm1 & 21 & 1685443 & Failed (numerical instability?)\\
\hline
\end{tabular}
\caption{Solutions obtained for branch-and-bound.}
\label{tab:bb-sols}
\end{center}
\end{table}
\subsection{Evolutionary algorithm}
Niche search was used to make 5 niches, each with 5 individuals,
evolve for 250 niche generations. In each of these generations, the
population of each niche would reproduce until no improvements in its
best element were observed. Although tuning up the population and
generation numbers would likely lead to better results, we have made
not attempt to do so, and used the same values for all the problems.
The MIPLIB minimisation problems were converted into maximisations.
In table~\ref{tab:ea-sols} we report the optimal solutions, as stated
in the MIPLIB, and the range of the final solutions determined in an
experiment with 25 independent runs of niche search for each of the
benchmark problems. For more than 50\% of the tests, the optimal
solution could be determined. The EA failed to systematically find a
feasible solution only for the \emph{enigma} problem.
The average number of LPs that were solved until obtaining the
solutions for niche search reported is written on the rightmost
column.
\begin{table}[htbp!]
\begin{center}
\leavevmode
\begin{tabular}[t]{|l|r||r|r|r||r|}
\hline
\mcc{Problem} & \mccr{Optimal} & \multicolumn{4}{|c|}{Niche search
solutions (25 runs)} \\ \cline{3-6}
\mcc{name} & \mccr{solution} & \mcc{Worst} & \mcc{Mean} &
\mccr{Best}& Avg.\#LPs \\ \hline\hline
bell3a & -878430.32 & -1502340 & -929125.2 & -881935 & 312455 \\
bell5 & -8966406.49& -9342570 & -9121121.2 & -9030450 & 597709 \\
egout & -568.101 & -575.983 & -568.73156 & -568.101 & 48843\\
enigma & -0.0 & -24 \small{\textit{(infeas.)}}&
-15.2 \small{\textit{(infeas.)}}&
-7\small{ \textit{(infeas.)}} & 252582 \\
flugpl & -1201500 & -1240500 & -1209300 & -1201500 & 37478\\
gt2 & -21166.0 & -42006 & -32375.84 & -22342 & 1249782\\
lseu & -1120 & -1542 & -1270.12 & -1120 & 232348 \\
mod008 & -307 & -349 & -317.04 & -307 & 154303\\
modglob & -20740508 & -20740508& -20740508 &-20740508 & 99478\\
noswot & +43 & 29 & 39.56 & 41 & 401825 \\
p0033 & -3089 & -3188 & -3097.64 & -3089 & 25785\\
pk1 & -11 & -29 & -23.6 & -19 & 93031 \\
pp08a & -7350 & -7390 & -7358.8 & -7350 & 45780\\
pp08acut & -7350 & -7350 & -7350 & -7350 & 45582 \\
rgn & -82.1999 & -82.2 & -82.2 & -82.2 & 8050\\
stein27 & -18 & -18 & -18 & -18 & 286 \\
stein45 & -30 & -31 & -30.04 & -30 & 43108\\
vpm1 & -20 & -20 & -20 & -20 & 7397 \\
\hline
\end{tabular}
\caption{Optimal solutions of the benchmark problems reported in
MIPLIB, solutions obtained in an experiment with 25 independent
runs of niche search, and average number of LPs solved for
obtaining them.}
\label{tab:ea-sols}
\end{center}
\end{table}
\begin{table}[hbtp!]
\begin{center}
\leavevmode
\begin{tabular}[t]{|l||r|r||r|r||c|}
\hline
Problem name & $r^f/R$ & $E[n^f]$ & $r^o/R$ & $E[n^o]$ & Best algorithm \\
\hline \hline
bell3a & 100\% & 2053 & 0\% &$>$18246645 & B\&B\\
bell5 & 100\% & 33748 & 0\% &$>$18024642 & B\&B\\
egout & 100\% & 423 & 92\% & 133764 & EA \\
enigma & 0\% & $>$11876637& 0\% &$>>$11876637& B\&B\\
flugpl & 100\% & 29048 & 80\% & 91004 & B\&B\\
gt2 & 100\% & 6383 & 0\% &$>$37665907 & EA \\
lseu & 100\% & 1985 & 4\% & 10269416 & B\&B\\
mod008 & 100\% & 17 & 48\% & 2557585 & EA? \\
modglob & 100\% & 3 & 100\% & 99478 & EA \\
noswot & 100\% & 33627 & 0\% &$>$34335094 & EA \\
p0033 & 100\% & 8350 & 80\% & 93571 & B\&B\\
pk1 & 100\% & 3 & 0\% &$>$6259152 & EA \\
pp08a & 100\% & 49 & 72\% & 177969 & EA \\
pp08acut & 100\% & 33 & 100\% & 45582 & EA \\
rgn & 100\% & 21 & 100\% & 8050 & B\&B\\
stein27 & 100\% & 41 & 100\% & 286 & EA \\
stein45 & 100\% & 61 & 96\% & 54791 & EA \\
vpm1 & 100\% & 123 & 100\% & 7397 & EA \\
\hline
\end{tabular}
\caption{Niche search: number of successes and expected
number of LP solutions for finding a feasible and the optimal
solution, respectively, and performance comparison with B\&B.}
\label{tab:ea-summary}
\end{center}
\vspace{-10mm}
\end{table}
In order to assess the empirical efficiency of the algorithm, we
provide a measure of the expectation of the number of LP solutions
required for finding a feasible and the optimal solution. Let $R$ be
the number of runs per benchmark problem in a given experiment, and
$r^f$ and $r^o$ be the number of runs in which a feasible and the
optimal solution are found, respectively. Let $n^f_i$ be the number
of LP solutions that were required for obtaining a feasible solution
in run $i$, or the total number of LPs solved in that run if no
feasible solution was found. Similarly, let $n^o_i$ be the number of
LP solutions required for reaching optimality, or the total number of
LPs solved in $i$ if no optimal solution was found. Then, the
expected number of LPs for reaching feasibility, based on these $R$
observations, is:
$$E[n^f] = \sum_{i=1}^{R} \frac{n^f_i}{r^f}$$
Equivalently, the expected number of LPs for reaching optimality is
$$E[n^o] = \sum_{i=1}^{R} \frac{n^o_i}{r^o}$$
These values are
reported for each of the benchmark problems in
table~\ref{tab:ea-summary}. On the case of $r^o=0$, the sum of the LP
solutions of the total experiment ($R$ runs) provides a lower bound on
the expectations for optimality. The same for feasibility, when
$r^f=0$. These are the values reported in table~\ref{tab:ea-summary}
for those situations. In this table we also make a comparison of B\&B
and the EA. The judgement is based on the reliability and on the
expected number of LPs required for optimality, for each of the
algorithms.
For some problems (e.g. pp08a) the EA quickly obtained a good
solution, even though B\&B has failed. For other (e.g.\ modglob,
mod008), a feasible solution was easily found at the beginning of the
EA, suggesting its possible use as a method for obtaining a feasible
solution to speedup B\&B. Some benchmarks---especially
\textit{enigma}---were easily solved by B\&B, even though the EA had
problems tackling them.
\begin{figure}[htbp!]
\begin{center}
\epsfig{file=pp08a.ps,width=80mm,angle=-90}
\caption{Typical log of the evolution of the solution with the
number of LPs solved, in this case for the \emph{pp08a}
benchmark. A feasible solution was found at around the 70th
LP solved. Dotted line for infeasible solutions (left side $y$
axis), continuous line for feasible ones (right side $y$ axis).}
\label{fig:graph}
\end{center}
\end{figure}
In figure~\ref{fig:graph} is plotted a log of the evolution of the
population's best solution in a typical run of the EA, for the case of
the \emph{pp08a} problem. The curve at the beginning of the process
corresponds to infeasible solutions; a feasible solution is found in
the middle of the process. In most of the cases, these two phases of
the search can be distinctly observed: first minimising the
infeasibilities and then, when a feasible solution is found,
optimising the objective.
In order to assert the importance of each of the operators used in the
evolutionary system, we executed some experiments for assessing their
efficiency. These experiments consisted on keeping track of which of
the operators were responsible for improvements on the solutions, and
of analysing the behaviour of the algorithm in their absence. They
showed that the three genetic operators, the local search, and the
initialisation procedure, where all necessary for a good performance
of the algorithm.
We also made a series of runs with only one niche, increasing the
number of generations so that the maximum number of LP solutions was
approximately the same as the one used for the results reported in
this section. The solutions obtained provided an empirical
confirmation of the importance of the separation of the population in
niches. With a single niche the algorithm decreased its performance,
both in terms of the number of runs that lead to feasibility and
optimality, and in terms of the number of calls to the LP solver that
were required for obtaining an equivalent final solution.
\section{Conclusion}
\label{sec:conclusion}
In this paper we present an evolutionary algorithm for the solution of
integer linear programs based on the separation of the variables into
the integer part and the continuous part. The integer variables are
fixed by the evolutionary system, and replaced in the original LP,
leading to a pure continuous problem. The optimisation of this LP
determines the continuous variables corresponding to that integer
solution, and the objective value leads to the solution's fitness.
The results obtained for some of the standard benchmark problems were
compared to those obtained by B\&B. The performance of the
evolutionary algorithm is promising. In some of the benchmark tests
it outperformed B\&B, either by requiring less LP solutions to
systematically reach the optimum, or by succeeding in determining a
good feasible---sometimes optimal---solution in cases where B\&B
failed.
The success of the algorithm in finding good feasible solutions with
limited computational resources for most of the benchmark problems
testify its potentialities for real-world, practical applications.
The algorithm proposed does not take into account any particular
structure of the problems (it is based only on the information
contained in MPS files; nothing about the specific kind of problem
dealt with is taken into account). For obtaining more competitive
results, a problem-specific local search, exploiting the particular
structure of the problem, should be additionally implemented.
The discrepancy between the results obtained by the EA and by B\&B
suggests that these algorithms are probably good complements of each
other, and the integration of both approaches in a single tool seems
to be a promising research direction.
An advantage, not yet exploited, of this evolutionary algorithm is
that the models that it can tackle may include non-linearities, as
long as a linear problem can be obtained by fixing some variables.
These nonlinear variables would also be fixed by the evolutionary
structure, at the time of fixing the integer ones, in such a way that
the resulting problem is linear and continuous.
\section{Acknowledgements}
This work was supported by the European Union Science and Technology
Fellowship Programme in Japan.
|
\section{Introduction and statement of the results. \label{intro}}
This paper improves the results of \cite{DisconziViscousFluidsNonlinearity}, where the formulation
of relativistic viscous fluids has been investigated by the
second author. There, following Lichnerowicz
\cite{LichnerowiczBookGR},
a physically well-motivated
relativistic version of the Navier-Stokes equations has been proposed, and
well-posedness, in Gevrey spaces,
of the corresponding Einstein-Navier-Stokes system
established under the assumption that the fluid is incompressible
(in a relativistic sense, see below) and irrotational. The goal
of the present work is to remove the latter hypothesis,
replacing it by a restriction on the initial data that allows
the vorticity to be non-zero.
Finding the correct way of incorporating viscosity into
General Relativity
is a longstanding problem\footnote{It is interesting to notice that
even the correct formulation of the non-relativistic Navier-Stokes
equations on a general Riemannian manifold does not seem to present
itself in a natural and obvious way, see \cite{ChanCzubakNavierStokes,EbinMarsden,TaylorPDE3}.}
\cite{LichnerowiczBookGR, MisnerThorneWheeler, WeinbergCosmology},
one that has recently attracted attention due to its importance
in the study of heavily dense objects (as neutron stars),
and models of the early
universe. See, for instance,
\cite{Cooketal2003,
Disconzi_Kephart_Scherrer_2015,
Dosetal,
DosTsa,
Duez_review,
DuezetallEinsteinNavierStokes,
Herr_axially,
Herr_axially_shear,
Kreiss_et_al,
Lovelace_Duez_et_al,
MaartensDissipative,
Pahwa_FLRW,
Pal_et_al,
Pal_Reula_Rezzolla,
Piattella_et_al,
RezzollaZanottiBookRelHydro,
Saijo,
SorBran,
Weinberg_GR_book}
and references therein.
A thorough and more up-to-date discussion, including details
on the First and Second Order Theories mentioned below,
can be found in \cite{RezzollaZanottiBookRelHydro}.
The main difficulty in formulating a theory of relativistic viscous fluids
seems to stem from
the absence of
a variational formulation for the classical, non-relativistic, Navier-Stokes
equations (see, however, \cite{EyinkActionPrinciple, YasueVariationalNavierStokes} for formalisms that allow
for a variational principle in some more general sense).
Lacking such a formulation, one does not have a canonical
way of determining what the stress-energy tensor $T_{\alpha\beta}$ ought to be in the
context of General Relativity. Different proposals have been made in this
regard, but they have all led to either ill-posed equations, or to
equations that imply the existence of
superluminal signals. This approach, where one couples Einstein's
equations to the Navier-Stokes equations via the introduction of a
suitable stress-energy tensor is known as (relativistic) Standard Irreversible
Thermodynamics or First Order Theories. We remark that it
consists in the traditional approach of coupling
gravity to matter, one that has been successful for almost all other
matter fields \cite{ChoquetBruhatGRBook, HawkingEllisBook}.
The failure of First Order Theories in producing a consistent theory
of relativistic fluids led researches to devise a different approach, known
as (relativistic) Rational
Extended Irreversible Thermodynamics, or Second Order Theories, or yet
Divergence-type formulation of Extended Irreversible
Thermodynamics
\cite{JouetallBook, MuellerRuggeriBook, RezzollaZanottiBookRelHydro}.
In such theories, one extends the space of variables of the model, and
the resulting equations are of hyperbolic character in several important
situations of physical interest, leading to equations that are well-posed,
with disturbances propagating at finite speed.
It is not at all clear, however, that the equations remain hyperbolic under all
physically realistic scenarios. In fact,
Rezzolla and Zanotti conclude their detailed discussion of relativistic viscous
fluids pointing out that ``the construction of a formulation that is cast in a
divergence-type is not, \emph{per se}, sufficient to guarantee hyperbolicity"
\cite{RezzollaZanottiBookRelHydro}.
Furthermore, the plethora of models that comes out
of the extended thermodynamic approach suggests that it entails many
\emph{ad-hoc} features, in sharp contrast to the usually unique way of coupling
gravity to matter via the introduction of the stress-energy tensor of matter
fields (when the latter
is uniquely determined by a variational characterization).
These considerations
suggest that it is worthwhile to take a fresh look at the question
of whether there is a correct stress-energy tensor $T_{\alpha\beta}$ that describes relativistic viscous
fluids, and that can be coupled to gravity in the traditional way, i.e.,
as in the Standard Irreversible Thermodynamics approach (see also the discussion
in section \ref{status}).
This idea is reinforced by the fact that recent numerical advances in the
modeling of rapidly rotating stars with shear viscosity employ
the first order formalism \cite{DuezetallEinsteinNavierStokes}.
Consider the following stress-energy tensor for a viscous fluid:
\begin{align}
T_{\alpha\beta} & = (p + \varrho ) u_\alpha u_\beta - p g_{\alpha\beta}
+ \kappa \pi_{\alpha\beta} \nabla_\mu C^\mu
+ \vartheta\pi_\alpha^\rho \pi_\beta^\mu (\nabla_\rho C_\mu + \nabla_\mu C_\rho),
\label{T_vis}
\end{align}
where
$p$ and $\varrho$ are respectively the
pressure and density of the fluid, $u$ is its four-velocity,
the bulk viscosity $\kappa$ and the shear viscosity $\vartheta$ are
non-positive constants\footnote{The coefficients of bulk and shear viscosity have a definite sign, the
choice of which depends on conventions. Sometimes $T_{\alpha \beta}$
is written with a shear term $-\vartheta\pi_\alpha^\rho \pi_\beta^\mu (\nabla_\rho C_\mu + \nabla_\mu C_\rho)$, $\vartheta >0$, which corresponds to having $\vartheta < 0$ in our formulation.
While such sign differences are important when one explicitly
computes the values of physical observables, for the results here presented all that matters
is that $\vartheta \neq 0$. In physically realistic models, it is also the case that $\vartheta$ is
not a constant, but a smooth function of the thermodynamic variables. Our result
can be extended to this case with minor changes in the proof, provided that
$\vartheta$ never vanishes, but we do not
include this here for brevity.},
$g$ is a Lorentzian metric\footnote{Our convention for the metric is $(+---)$.}
and
$\pi_{\alpha\beta} = g_{\alpha\beta} - u_\alpha u_\beta$.
$p$ and $\varrho$ are related by an equation known as equation of state,
the choice of which depends on the nature of fluid. $C$ is known as
the dynamic velocity
(also called the current of the fluid), and it is related to $u$ by
\begin{gather}
C_\alpha = F u_\alpha,
\label{dyn_vel}
\end{gather}
where $F$ is the so-called index of the fluid. It is defined as
\begin{equation}\label{indexF}
F = 1 + \epsilon + \frac{p}{r},
\end{equation}
where $\epsilon\geq0$ is the specific internal energy and $r\geq 0$ is the rest mass density \cite{FriRenCauchy}.
The density $\varrho$ is related to the internal energy and the rest mass by
\[
\varrho=r(1+\epsilon),
\]
so that
\begin{equation}\label{iF}
rF=\varrho + p.
\end{equation}
The index of the fluid, $F$, and the dynamic velocity, $C$, have been introduced by
Lichnerowicz in his study of relativistic inviscid fluids
\cite{LichnerowiczBookGR, Lich_fluid_1, Lich_fluid_2, Lichnerowicz_MHD_book}.
Lichnerowicz was also the first one to write down the stress-energy tensor
(\ref{T_vis}) \cite{LichnerowiczBookGR}, except that it contained an extra term of
the form
$\vartheta \pi_{\alpha\beta} u^\mu \partial_\mu F$. This extra term was pointed out by
Lichnerowicz himself and later by Pichon \cite{PichonViscous}, to lead to an
indetermination in the computation of the pressure. Pichon proposed
subtracting this term, which leads to (\ref{T_vis}).
See \cite{LichnerowiczBookGR,PichonViscous} for more background on \eqref{T_vis}.
The reader should notice that
(\ref{T_vis}) reduces to the stress-energy tensor of an ideal, i.e., inviscid,
fluid when $\kappa = \vartheta = 0$. Indeed, this is just one of several natural
requirements that one would impose when looking for an appropriate
definition of a stress-energy tensor for a relativistic fluid with viscosity,
see \cite{DisconziViscousFluidsNonlinearity}. We point out that Choquet-Bruhat has also proposed
a stress-energy tensor similar to (\ref{T_vis}) \cite{ChoquetBruhatGRBook}. Her proposal
does not include the projection terms $\pi_{\alpha\beta}$, and the viscous
terms are, therefore, linear in the velocity. We remark that yet another proposal
for a viscous relativistic stress-energy tensor appears in
\cite{TempleViscous}.
Next, recall the first law of thermodynamics\footnote{See, for instance, \cite{DisconziRemarksEinsteinEuler,FriRenCauchy} for a review of the thermodynamic properties of
relativistic fluids.}
\begin{equation}\label{flt}
\theta ds=d\epsilon+pdv
\end{equation}
where $\theta$ is the absolute temperature, $s$ the specific entropy, and $v$ the specific volume. We have $v=\frac 1r$ \cite{FriRenCauchy},
so by \eqref{indexF}, \eqref{flt} can be written as
\begin{equation}\label{flt2}
\theta ds=dF-\frac 1r dp.
\end{equation}
A fluid with stress-energy tensor (\ref{T_vis}) is said to be incompressible if
\begin{equation}\label{incompressible_def}
\nabla_\mu C^\mu = 0.
\end{equation}
\begin{remark}
Here we follow the literature (e.g., \cite{ChoquetBruhatGRBook, Lichnerowicz_MHD_book})
and call incompressible a fluid satisfying (\ref{incompressible_def}). We stress, however, that this terminology
is based more on an analogy with Newtonian physics (where incompressible fluids are characterized
by vanishing divergence) than on actual physical properties of fluids, in that (\ref{incompressible_def})
does not imply incompressibility in the exact sense of the word.
Pseudo-incompressible would probably be a better terminology, but it is not clear if adopting a different terminology
than what is used in the literature would not cause more confusion.
\label{remark_incompressibility}
\end{remark}
Then, by \eqref{iF}, $T_{\alpha\beta}$
becomes
\begin{align}
T_{\alpha\beta} = rF u_\alpha u_\beta - p g_{\alpha\beta}
+ \vartheta\pi_\alpha^\rho \pi_\beta^\mu (\nabla_\rho C_\mu + \nabla_\mu C_\rho).
\label{T_vis_i}
\end{align}
Moreover, because $\pi^\mu_\beta u_\mu=0$, we can rewrite \eqref{T_vis_i} as
\begin{equation}\label{Tab}
T_{\alpha\beta} = rF u_\alpha u_\beta - p g_{\alpha\beta}
+ F\vartheta\pi_\alpha^\rho \pi_\beta^\mu (\nabla_\rho u_\mu + \nabla_\mu u_\rho).
\end{equation}
Finally, we define the vorticity tensor by
\begin{gather}
\Omega_{\alpha\beta } = \nabla_\alpha C_\beta - \nabla_\beta C_\alpha
\equiv \partial_\alpha C_\beta - \partial_\beta C_\alpha.
\label{vorticity}
\end{gather}
A fluid is called irrotational if $\Omega = 0$. Notice that $\Omega$ is anti-symmetric,
so it has only six independent components.
We follow the standard approach of assuming that only two
of the thermodynamic quantities are independent with the question of which ones left as a matter of choice. The other quantities are then
determined by the first law of thermodynamics and an equation of state. The equation of state
depends on the nature of the fluid, and physically, the relations between the thermodynamic quantities
should be invertible\footnote{Upon making such assumptions, we are restricting ourselves to fluids in a single phase and
ruling out the possibility of phase transitions.}. Here we shall assume that
$r$ and $s$ are independent and postulate an equation of state
of the form
\begin{gather}
\varrho = \mathscr{P}(r,s).
\label{eq_state_r_s}
\end{gather}
It follows that $p = p(r,s)$,
$\theta=\theta(r,s)$, $\epsilon=\epsilon(r,s)$,
and $F=F(r,s)$ are known if $r$ and $s$ are. We note that later on it will be more convenient to treat $s$ and $F$
as independent variables. Then the equation
of state will be given by $r = r(F,s)$.
On physical grounds, one has that $F > 0$.
This allows to restrict to positive values when treating $F$ as
an independent variable.
In this situation, the following
condition will be assumed to hold:
\begin{gather}
\frac{\partial r}{\partial F} \geq \frac{r}{F},
\label{sound_speed_condition}
\end{gather}
in particular $\frac{\partial r}{\partial F} > 0$ if
$r > 0$.
Condition (\ref{sound_speed_condition}) expresses the statement that sound waves in an ideal fluid
travel at most at the speed of light. This condition has to be satisfied if we want to recover the stress-energy
tensor of an ideal fluid when $\kappa = \vartheta = 0$ \cite{Lichnerowicz_MHD_book}
We suppose that the equation of state is such that the temperature
satisfies
\begin{align}
\begin{split}
\theta(r,s) > 0 \text{ if } r > 0, s \geq 0, \\
\theta(F,s) > 0 \text{ if } s \geq 0, F > 0,
\end{split}
\label{temperature}
\end{align}
expressing the positivity of the temperature regardless of the choice
of independent variables.
The full system of equations derived from coupling Einstein's equations
to (\ref{T_vis_i}) is rather complicated, see
\cite{DisconziViscousFluidsNonlinearity}. Thus we consider, besides incompressibility, one further
simplifying assumption, namely, we investigate the sub-class of solutions
for which the vorticity evolves according to
\begin{gather}
\mathcal L_C \Omega_{\alpha \beta}
+ q u^\mu \nabla_\mu \nabla_\alpha C_\beta - q
u^\mu \nabla_\mu \nabla_\beta C_\alpha +
\partial_\alpha(\theta F) \partial_\beta s - \partial_\beta(\theta F) \partial_\alpha s
= \mathcal F_\vartheta.
\label{evolution_vorticity}
\end{gather}
Here, $\mathcal L_C$ denotes the Lie derivative in the direction of $C$,
$\mathcal F_\vartheta$ is a smooth function
of $\Omega$, $g$ and its derivatives up to second
order, $C$ and $u$ and their derivatives up to first order.
$q$ is a constant, and $\mathcal F_\vartheta$ and $q$
may also depend on the parameter $\vartheta$. $\mathcal F_\vartheta$
and $q$ dictate, to a certain extent, which quantities are considered
relevant in some particular model, and, therefore, are chosen according
to the problem one wishes to study (see section \ref{vorticity_in_rel}).
We discuss the restrictions imposed by
(\ref{evolution_vorticity}) in section \ref{restriction_vorticity}.
It should be noticed that one must have $q=0$ and that $\mathcal F_\vartheta$
cannot be chosen freely if $\vartheta = 0$ (although $\vartheta = 0$ will not be treated here).
The starting point is the following system of equations: Einstein's equations coupled to
(\ref{T_vis}) and supplemented by (\ref{incompressible_def}) and
(\ref{evolution_vorticity}), reading
\begin{subnumcases}{\label{original}}
R_{\alpha\beta} - \frac{1}{2}R g_{\alpha\beta} + \Lambda g_{\alpha\beta}= \mathscr{K} T_{\alpha\beta},
\label{original_Einstein} \\
\nabla^\alpha T_{\alpha\beta} = 0,
\label{original_conservation} \\
\nabla_\alpha( r u^\alpha ) = 0,
\label{original_mass_conservation} \\
\nabla_\mu C^\mu = 0,
\label{original_incompressible} \\
\mathcal L_C \Omega_{\alpha\beta} +q u^\mu \nabla_\mu \nabla_\alpha C_\beta - q
u^\mu \nabla_\mu \nabla_\beta C_\alpha = \widetilde{\mathcal F}_\vartheta,
\label{original_vorticity}
\\
u^\alpha u_\alpha = 1.
\label{original_normalization}
\end{subnumcases}
where $R_{\alpha\beta}$ and $R$ are the Ricci and scalar curvature
of the metric $g$, $\mathscr{K}$ is a coupling constant, and $\Lambda$ is
the cosmological constant\footnote{Our results do hold irrespective
of the value of $\Lambda$.}. We recall that (\ref{original_conservation}) is in fact a consequence
of (\ref{original_Einstein}) in view of the Bianchi identities, but it is customary to list it along with the other equations.
In the sequel we set $\mathscr{K}$ to $1$. The equation
(\ref{original_mass_conservation}) says
the mass is locally conserved along the flow lines, and (\ref{original_normalization}) is the standard
normalization condition on the velocity of a relativistic fluid.
In general, without
(\ref{original_mass_conservation}), the motion of the fluid is underdetermined.
Equation (\ref{original_vorticity}) is simply (\ref{evolution_vorticity}),
with $\widetilde{\mathcal F}_\vartheta$ defined in the obvious fashion.
A useful consequence of \eqref{original_normalization} often used in computations is
\begin{equation}\label{useful}
u^\alpha \nabla_\beta u_\alpha=0.
\end{equation}
The unknowns are the metric $g$, the fluid velocity $u$, the specific entropy $s$, and the rest mass density $r$, where $s$ and $r$ are non-negative
real valued functions. We suppose that we are also given a
smooth function $\mathscr{P}: \mathbb R_+ \times \mathbb R_+ \rightarrow \mathbb R$ that gives the
equation of state (\ref{eq_state_r_s}), with the other thermodynamic quantities then given as functions of $s$ and $r$ as discussed above.
\begin{definition}
System (\ref{original}) with $T_{\alpha\beta}$
given by (\ref{T_vis}) will be called the
incompressible Einstein-Navier-Stokes
system.
\end{definition}
\begin{remark}
Here we recall once more that the terminology ``incompressible fluid" is a bit misleading, see remark \ref{remark_incompressibility}.
\end{remark}
\noindent \textbf{Assumption.}
We shall assume for the rest of the text that $\vartheta \neq 0$. \\
An initial data set for the Einstein-Navier-Stokes system consists of the following:
\begin{itemize}
\item a three-dimensional manifold $\Sigma$,
\item a Riemannian metric $g_0$ (with our conventions this metric is negative definite)
\item a symmetric two-tensor $\kappa$,
\item a real valued non-negative function $s_0$,
\item a real valued non-negative function $r_0$,
\item a vector field $v$.
\end{itemize}
The last five quantities are defined on $\Sigma$. As it is well-known, these data must satisfy the constraint equations. In a coordinate system with $\partial_0$ transversal and
$\partial_i$, $i=1,2,3$, tangent to $\Sigma$ the constraint equations are given by
\begin{gather}
S_{\alpha 0 } = T_{\alpha 0 },
\label{constraints}
\end{gather}
where
$S_{\alpha\beta} = R_{\alpha\beta} - \frac{1}{2}R g_{\alpha\beta} + \Lambda g_{\alpha \beta}$
is the Einstein tensor. In our case, it is not enough that
the initial data satisfies (\ref{constraints}). We also need
the compatibility conditions obtained upon restriction
of (\ref{original_vorticity}) to the initial hypersurface,
since initial data for $\Omega$ and $C$ are derived from
$g_0$, $\kappa$, $s_0$, $r_0$, and $v$ (see \cite{DisconziViscousFluidsNonlinearity}).
By definition, an initial data set
always satisfies the constraints and compatibility
conditions. While the construction of initial data
for the Einstein-Navier-Stokes equations is an important task,
here our primary interest is on the evolution problem, and as
such we shall take the standard approach of assuming the initial data
as given (see the discussion in section \ref{restriction_vorticity}).
We are now ready to state the main result. We refer the
reader to the standard literature in General Relativity for the terminology
employed in Theorem \ref{main_theorem}. We remind the reader of the
definition of Gevrey spaces $\gamma^{m, (\sigma)}$ in Section \ref{main_proof}, referring to
references
\cite{LerayOhyaLinear, LerayOhyaNonlinear,RodinoGevreyBook} for more details.
\begin{theorem}
Let $\mathcal I = (\Sigma, g_0, \kappa,v, s_0, r_0)$ be an initial data set for the
incompressible Einstein-Navier-Stokes system (\ref{original}) ,
with $\Sigma$ compact, $s_0 > 0$, $r_0 > 0$, and an equation of state $\mathscr{P}$
such that (\ref{sound_speed_condition}) and (\ref{temperature})
are satisfied. Let $\mathcal F_\vartheta$ be a given smooth function of
$\Omega$, $g$ and its derivatives up to second
order, $C$ and $u$ and their derivatives up to first order, and assume that
$q > 0$.
Assume that the initial data is in
$\gamma^{(\sigma)}(\Sigma)$ for some $1 \leq \sigma <\frac{24}{23}$.
Then there exist a space-time $(M,g)$
that is a development of $\mathcal I$,
real valued functions $s > 0$
and $r > 0$
defined on $M$, and a vector field $u$,
such that
$g \in \gamma^{3, (\sigma)}(M)$,
$u \in \gamma^{2, (\sigma)}(M)$,
$s \in \gamma^{2, (\sigma)}(M)$,
$r \in \gamma^{2, (\sigma)}(M)$, and
$(g,u, s, r)$ satisfy the incompressible
Einstein-Navier-Stokes system in $M$.
Furthermore, this solution satisfies the geometric uniqueness
and domain of dependence properties, in the following sense.
Let $\mathcal I^\prime = (\Sigma^\prime, g_0^\prime, \kappa^\prime, v^\prime, s_0^\prime, r_0^\prime)$
be another initial data set, also with equation of state $\mathscr{P}$,
with corresponding development $(M^\prime,g^\prime)$ and
solution $(g^\prime,u^\prime, s^\prime, r^\prime)$
of the incompressible Einstein-Navier-Stokes equations
in $M^\prime$.
Assume that there exists a diffeomorphism
between $S \subset \Sigma$ and $S^\prime \subset \Sigma^\prime$ that carries
$\left. \mathcal I\right|_S$ onto $\left. \mathcal I^\prime \right|_{S^\prime}$, where $S$ and $S^\prime$
are, respectively, domains in $\Sigma$ and $\Sigma^\prime$. Then there exists
a diffeomorphism between $D_g(S) \subset M$
and $D_{g^\prime}(S^\prime) \subset M^\prime$ carrying
$(g,u, s, r)$ onto
$(g^\prime,u^\prime, s^\prime, r^\prime)$, where $D_g(S)$ denotes
the future domain of dependence of $S$ in the metric $g$; in particular $D_g(S)$ and
$D_{g^\prime}(S^\prime)$ are isometric.
\label{main_theorem}
\end{theorem}
\begin{remark}
Under the further assumption that the fluid is irrotational,
Theorem (\ref{main_theorem}) was proved by the second author in
\cite{DisconziViscousFluidsNonlinearity}, where a better regularity result than
in theorem \ref{main_theorem}, namely, $\sigma < 2$, was obtained.
\end{remark}
\begin{remark}
The space-time $M$ is diffeomorphic to $\Sigma \times [0,T]$ for some $T > 0$,
and to $\Sigma \times [0,\widetilde{T})$ for some $\widetilde{T} > T$
if we require it to be a maximal Cauchy
development.
\end{remark}
\begin{remark}
The compactness of $\Sigma$
is not absolutely necessary due to the domain of dependence property. However, in the case of a
non-compact $\Sigma$ without asymptotic conditions on the initial data, $M$ may not contain any Cauchy surface other than
$\Sigma$ itself.
\end{remark}
\begin{remark}
The hypotheses $s_0>0$ and $r_0 > 0$ guarantee, by continuity,
the positivity of $s$ and $r$ in the neighborhood of $\Sigma$, as stated in the theorem.
The assumption $s_0 > 0$ could be weakened to $s_0 \geq 0$, but in this case
the non-negativity of $s$ in $M$ would have to be derived from the equations of
motion, a task we avoid for brevity. On the other hand, allowing $r_0$ to vanish would cause severe
difficulties. In fact, the well-posedness of the corresponding problem is largely
open even in the case of an ideal fluid \cite{FriRenCauchy}.
\end{remark}
In the following, we adopt:
\begin{convention}Greek indices run from $0$ to $3$ and Latin indices
from $1$ to $3$.
\end{convention}
\section{Discussion on the hypotheses and the thesis of Theorem \ref{main_theorem}}
In this section we comment on the relevance
of Theorem \ref{main_theorem} for the study of relativistic fluids
with viscosity. We highlight the restrictions imposed by (\ref{evolution_vorticity})
and the regularity of solutions,
make some remarks regarding the physical content of the Theorem, and discuss
how this work fits within the broader context of relativistic viscous fluids,
making some general remarks about (\ref{T_vis}) and its particular case (\ref{T_vis_i})
along the way.
Readers interested solely in the proof of Theorem \ref{main_theorem} may skip this section.
\subsection{The evolution of the vorticity.\label{restriction_vorticity}}
Let us start with the evolution condition imposed on $\Omega$, i.e.,
equation (\ref{evolution_vorticity}) or, equivalently, (\ref{original_vorticity}).
In its full-generality,
the incompressible Einstein-Navier-Stokes system consists
of equations (\ref{original_Einstein})-(\ref{original_incompressible}) and
(\ref{original_normalization}), i.e., (\ref{original}) without
(\ref{original_vorticity}).
As such a system is rather complicated, we have imposed
(\ref{evolution_vorticity}) which, of course, is ultimately
a restriction on the unknowns $(g,u,s,r)$. From
the point of view of
(\ref{original_Einstein})-(\ref{original_incompressible}), equation
(\ref{original_vorticity}) should be understood as a constraint,
in the following sense. An initial data set
yielding a solution to
(\ref{original_Einstein})-(\ref{original_incompressible}) (plus
(\ref{original_normalization}))
will also give a solution to (\ref{original}) only if
further relations among the initial data hold.
Indeed, arguing as in \cite{DisconziViscousFluidsNonlinearity},
one determines, from the original Cauchy data and Einstein's equations,
the values of $\partial^2g$, $\partial s$, $\partial u$, $\Omega$, and $\partial^2 C$
(as well as the corresponding lower order derivatives) on the initial hypersurface $\{ t = 0 \}$,
obtaining
a relation of the form $\partial_0 \Omega_{\alpha\beta} = W_{\alpha \beta}$
on $\{ t = 0 \}$, where
$W_{\alpha \beta}$ is a function of the Cauchy data.
From (\ref{evolution_vorticity}), one also obtains
a relation $\partial_0 \Omega_{\alpha\beta} = Z_{\alpha \beta}$
on $\{ t = 0 \}$, with $Z_{\alpha \beta}$ a function of the Cauchy data.
Therefore the initial data must be such data $W_{\alpha\beta} = Z_{\alpha\beta}$,
and hence Theorem \ref{main_theorem} is ultimately a result under restrictions on the initial data, namely,
the initial data ought to satisfy
compatibility
conditions imposed by (\ref{original_vorticity}), as mentioned
earlier\footnote{We notice that a similar, albeit notably simpler, situation happens to perfect
fluids with zero vorticity. In that case, the equation for the vorticity is given by
$\mathcal L_C \Omega = 0$, where $\mathcal L$ is the Lie derivative. This is not compatible with
the equations derived from the divergence of the stress-energy tensor unless the initial data
is such that $\Omega = 0$ on the $\{ t = 0\}$ slice.}.
How large is the class of initial conditions
satisfying the above restrictions is by no means an unimportant question,
but one that is, at this point, premature,
since we do not even know whether the system (\ref{original_Einstein})-(\ref{original_incompressible}) and
(\ref{original_normalization}) has any solution at all
outside the class of analytic functions.
While this leaves open the question of how general
Theorem \ref{main_theorem} is,
it is consistent with some expected physical applications as
discussed in section \ref{vorticity_in_rel}.
Such restrictions notwithstanding, we make two important remarks.
First, one could, in principle, consider the case of zero vorticity, with the function
$\mathcal F_\vartheta$ chosen so that (\ref{evolution_vorticity}) becomes an identity (notice that our results
do not rely on any specific form of $\mathcal F_\vartheta$, except for the dependence
on the number of derivatives of its arguments). In this case, the constraints
reduce to those that have to be imposed on the initial data when $\Omega = 0$. One could
say then that our theorem reproves the earlier result
\cite{DisconziViscousFluidsNonlinearity}. While
obviously this is not our goal here, it at least shows the set of appropriate
initial data to be non-empty.
The interesting question is whether the set of initial data satisfying the constraints
and compatibility conditions is non-empty
modulo zero vorticity. This is will be addressed in a future work.
Second, in light of how little is known about viscosity in General Relativity
(see section \ref{status}), our conditional result should be view
as evidence that further investigation of Lichnerowicz's proposal (\ref{T_vis})
is a worthwhile line of inquiry. In this regard,
it is illustrative to point out that there are other situations in General
Relativity where the evolution problem
is investigated without a decisive answer to the question of solvability of the constraints,
but this has never stopped the community to make conditional statements regarding
Einstein's equations.
One such situation, for example, is the study of vacuum Einstein's equations with low regularity.
As discussed, for instance, in \cite{HolstMeier},
the ``rough solution theory of the
constraints has in fact lagged behind that of the evolution problem."
For instance, well-posedness of the vacuum evolution problem for
data $(g,\kappa$) in $H^s\times H^{s-1}$, $s > \frac{5}{2}$, had been known
since 1977 \cite{Hughes_Kato_Marsden_quasi_linear_1977}.
However, it was not until 2004 that solutions to the constraint
equations in this regularity class could be constructed \cite{CB_constraints_2004}.
Hence, for 27 years it was not known if the classical result \cite{Hughes_Kato_Marsden_quasi_linear_1977}
was not empty modulo large
values of $s$ (for which \cite{Hughes_Kato_Marsden_quasi_linear_1977} would simply reproduce
earlier known results).
\subsection{Regularity of solutions.}
We work in the Gevrey class, because the equations
we derive form a Leray-Ohya system
(see \cite{LerayOhyaNonlinear}), which, in general, is not well-posed
in Sobolev spaces. Gevrey spaces have become an important tool in analyzing the
equations of Fluid Dynamics, especially when viscosity is present
(see, e.g., \cite{TadmorBesovGevrey, TitiGevreyNavier, TitiGevreyParabolic, TemamGevrey, RodinoGevreyBook} and references therein). Hence, it is sensible
that such spaces might play a role in the case of relativistic
viscous fluids as well.
Furthermore, Gevrey spaces are not completely foreign to the study of Einstein's
equations:
in some relevant circumstances, the equations of ideal magneto-hydrodynamics appear
to
have been shown to be well-posed only in the
Gevrey class \cite{ChoquetBruhatGRBook,FriRenCauchy}\footnote{Although it is very likely that
the formulation of \cite{AnilePennisiMHD} would carry over, with almost no modifications, to the
coupling
with Einstein's equations. A proof of this statement, however, does not
seem to be available in the literature.}. On the other hand,
the overwhelming success of Sobolev space techniques in the investigation of the
Cauchy problem
for Einstein's equations\footnote{The literature on this topic is too vast; see,
e.g., the monographs \cite{ChoquetBruhatGRBook, RingstromCauchyBook}.}
almost demands that we employ Sobolev spaces in the study of the
evolution
problem. Moreover, in order to eventually settle the question of
whether (\ref{T_vis}) can give a physically satisfactory description
of relativistic viscous phenomena, we have to be able to explicitly compute several
physical
observables. For this, one has to solve the equations numerically, which, in turn,
requires that the equations be well-posed in some function space characterized by
a finite number of derivatives.
Unfortunately, currently Gevrey regularity seems to be the best one can do for (\ref{T_vis}),
as the corresponding equations of motion do not seem to be amenable to known Sobolev-type techniques.
We remark, however, that simply establishing causality of the equations of motion is already
a step forward in light of the long history of non-causal theories of relativistic viscous fluids
\cite{RezzollaZanottiBookRelHydro}.
\subsection{The status of viscosity in relativity\label{status} and Lichnerowicz's proposal}
In spite of the restriction on the initial data due to (\ref{evolution_vorticity}),
the severity of which we acknowledge is yet
to be understood, and on the regularity class of solutions,
one should not overlook the
conclusion of Theorem \ref{main_theorem}: it is possible, employing
the traditional Standard Irreversible Thermodynamics, to obtain
a description of relativistic viscous fluids that is well-posed and
does not
exhibit faster than light signals. In this regard, we
remind the reader once more that we are attempting
a new look at this problem through a first order formalism.
Hence, it is all but unreasonable to start off with conditions that
render the problem tractable with current mathematical technology.
The first attempt in this direction \cite{DisconziViscousFluidsNonlinearity} dealt with irrotational
fluids. Here, we considered a less dramatic condition on the vorticity,
namely, (\ref{evolution_vorticity}), which seems to be compatible
with some physical applications (see section \ref{vorticity_in_rel}).
The message conveyed by this is that,
while it is wide open whether a full existence result for the incompressible Einstein-Navier-Stokes may be within reach without restrictions on the vorticity,
one can still prove well-posedness results under interesting scenarios.
Another restriction in our Theorem that one would like to remove is the
incompressibility hypothesis, not only for the sake of mathematical generality,
but also because relativistic systems many times exhibit sound waves that
propagate at sub-luminal speeds. This is the subject of ongoing investigations.
In order to put all of the above in perspective, we give rather brief
overview of what is currently known about viscosity in relativity.
The Mueller-Israel-Stewart (MIS) theory
\cite{MIS-2, MIS-3, MIS-5, MIS-6, MIS-1, MIS-4}
is probably the best accepted theory of relativistic viscous phenomena. It consists of
a systematic application of the ideas of Relativistic Extended Irreversible Thermodynamics
\cite{JouetallBook, MuellerRuggeriBook}.
The linearization about equilibrium states of the MIS theory
has been shown to be causal \cite{Hiscock_Lindblom_stability_1983}.
The non-linear theory, however, is also plagued
with non-causality behavior \cite{Hiscock_Lindblom_pathologies_1988}.
To be fair, such loss of causality is known to happen under extreme physical conditions
unlikely to be met by most realistic systems.
More precisely, in \cite{Hiscock_Lindblom_pathologies_1988}
the authors investigate
the relatively simple case
where only heat conduction is
present, so that the bulk and shear viscosity are zero, and under the assumption of planar symmetry.
Under
these assumptions, it is shown in \cite{Hiscock_Lindblom_pathologies_1988} that the equations
of motion are causal under a restriction on the values of the heat-flux, and non-causal
if such a restriction is violated\footnote{In passing, one should note that the MIS is sometimes referred to
as ``causal dissipative relativistic theory," but strictly speaking that
is, in view of what has been said, a misnomer.}.
In contrast, Theorem \ref{main_theorem}, as well as \cite{DisconziViscousFluidsNonlinearity},
makes no symmetry or near-equilibrium assumption, and treats the full non-linear system,
albeit it assumes stiffness and stringent restrictions on the initial data (or irrotationality
in the case of \cite{DisconziViscousFluidsNonlinearity}). It is important to notice,
however, that from the point
of view of causality, such results treat precisely the most ``dangerous" scenario, i.e., they
include
the term
\begin{gather}
\pi_\alpha^\rho \pi_\beta^\mu
(\nabla_\rho C_\mu + \nabla_\mu C_\rho),
\nonumber
\end{gather}
which leads to multiple characteristic due to the presence
of the projections $\pi_\alpha^\rho \pi_\beta^\mu$. The causality obtained in
\cite{Hiscock_Lindblom_pathologies_1988}, on the other hand, is restricted to the case
when the viscous part of the stress-energy tensor contains only the heat flow; in particular, such
projection terms are absent.
We also point out that, to the best of our knowledge,
the aforementioned causality and well-posedness results of the MIS
theory \cite{Hiscock_Lindblom_stability_1983, Hiscock_Lindblom_pathologies_1988} do not include
coupling to Einstein's equations, i.e., they consider the fluid equations in a fixed background
(except for some very simple situations such as FRW cosmologies \cite{MaartensDissipative}),
whereas Theorem \ref{main_theorem} and \cite{DisconziViscousFluidsNonlinearity} do
treat the full Einstein-fluid system.
Another interesting feature of Theorem \ref{main_theorem} is that it circumvents the
instability results of Hiscock and Lindblom
\cite{Hiscock_Lindblom_instability_1985}.
In fact, formally the equations that we study here
correspond to the case $\kappa = \sigma=0$ in \cite{Hiscock_Lindblom_instability_1985}.
Equations (\ref{incompressible_def}) and (\ref{evolution_vorticity}), however,
further constrain the evolution of perturbations
(compare with equations (41) in \cite{Hiscock_Lindblom_instability_1985}).
On the other hand, if condition (\ref{incompressible_def}) is dropped, then
the term $\nabla_\mu C^\mu$ that contributes
to the viscous part in (\ref{T_vis}) will depend on derivatives of the termodynamic
variables along the flow, a case not covered under the assumptions of
\cite{Hiscock_Lindblom_instability_1985}.
One important question about theories based on (\ref{T_vis}) is whether natural physical requirements
are satisfied. One such requirement is that entropy production be non-negative. It is not difficult
to see that, at least for the case investigated here, namely, when (\ref{T_vis}) reduces to
(\ref{T_vis_i}), this is the case. To see this,
one first uses $u^\beta \nabla^\alpha T_{\alpha\beta} = 0$ and the first law of thermodynamics to derive
\begin{gather}
\theta r u^\alpha \partial_\alpha s = \vartheta F ( \nabla^\mu u^\nu \nabla_\mu u_\nu
+ \nabla^\mu u^\nu \nabla_\nu u_\mu - u^\mu \nabla_\mu u^\alpha u^\nu \nabla_\nu u_\alpha).
\nonumber
\end{gather}
On the other hand, direct computation gives
\begin{gather}
\Sigma^{\alpha\beta} \Sigma_{\alpha \beta} = 2F^2
( \nabla^\mu u^\nu \nabla_\mu u_\nu
+ \nabla^\mu u^\nu \nabla_\nu u_\mu - u^\mu \nabla_\mu u^\alpha u^\nu \nabla_\nu u_\alpha),
\nonumber
\end{gather}
thus
\begin{gather}
\theta r u^\alpha \partial_\alpha s = \frac{\vartheta}{2F} \Sigma^{\alpha\beta} \Sigma_{\alpha \beta} \geq 0,
\nonumber
\end{gather}
since $\Sigma^{\alpha\beta} \Sigma_{\alpha \beta} \leq 0$\footnote{Recall our convention for the metric.} and $\vartheta \leq 0$.
A detailed study of physical consistency of models based on (\ref{T_vis}) appears in
\cite{DisconziKephartScherrerNew}
(see \cite{Disconzi_Kephart_Scherrer_2015}, however, for some further results concerning physical
properties of (\ref{T_vis})).
In summarty, it is fair to say that despite considerable progress since the original work of Eckart
\cite{EckartViscous}, the description of relativistic viscous phenomena still presents
many challenges. These remarks are not intended to claim that Lichnerowicz's proposal is better than
the more studied MIS theory or should be favored over any other theory,
but rather to highlight how little is known about viscosity in General
Relativity, which makes, in our opinion, attempts at different approaches, such as those based
on (\ref{T_vis}), welcome.
\subsection{Vorticity in relativistic fluids and some physical considerations\label{vorticity_in_rel}}
In the case of inviscid fluids, the vorticity obeys \cite{Lich_fluid_1, Lichnerowicz_MHD_book}
\begin{gather}
\mathcal L_C \Omega_{\alpha \beta} +
\partial_\alpha(\theta F) \partial_\beta s - \partial_\beta(\theta F) \partial_\alpha s
= 0.
\label{vorticity_ideal}
\end{gather}
This equation, sometimes called the Lichnerowicz, or Carter-Lichnerowicz
equation, plays an important role in the study of inviscid relativistic fluids,
and generalizations have also been
employed in formulations with viscosity
\cite{AndCom, ChrisMalik, Dosetal, DosTsa,
GourgIntro, Herr_axially_shear, Lewis, SorBran, TsagasBarrow}.
Equation (\ref{evolution_vorticity}) reduces to (\ref{vorticity_ideal})
when $\mathcal F_\vartheta = 0 = q$; thus, in particular, we see that
in physically relevant models, $q$ and $\mathcal F_\vartheta$ vanish when viscosity
is absent. Many of the modifications of (\ref{vorticity_ideal})
that include viscosity tend to occur in the context of very specific models,
where the equations are simple as compared to the ones here investigated
(for instance, perturbations of an FRW model). These can generally
be accommodated by (\ref{evolution_vorticity}) with a suitable choice of
$\mathcal F_\vartheta$; see, for example,
\cite{AndCom, ChrisMalik, Dosetal, DosTsa, Herr_axially, Herr_axially_shear, SorBran} and references therein. In more general terms, (\ref{evolution_vorticity})
seems natural when one considers applications with small viscosity,
in that the evolution of $\Omega$ should, to a certain degree,
resemble that of an ideal fluid. We also point out that
(\ref{evolution_vorticity}) is consistent with standard cosmology (with no
viscosity), in that, in such scenarios, $\Omega$ decays with the Hubble expansion,
being, as a consequence, ignored in many circumstances\footnote{Although
vorticity may play an important role in early stages of the Universe.
See, for example,
\cite{ChrisMalik} and references therein.}. Hence, one may, again,
suspect that in these cases $\Omega$ will be governed by an equation
similar to that of ideal fluids, since (\ref{vorticity_ideal})
enjoys the property of preserving zero vorticity.
Our choice (\ref{evolution_vorticity})
is a compromise between the previous considerations and an
algebraic form for which properties of hyperbolic polynomials, necessary for our techniques, hold true.
See section \ref{main_proof}.
\section{A new system of equations.\label{new_system_section}}
Here we derive a different system of equations, whose existence of solutions
implies Theorem \ref{main_theorem}. Thus, for the rest of this section, we assume we have a
sufficiently differentiable solution to (\ref{original}). In particular,
in light of (\ref{dyn_vel}) and (\ref{original_normalization}),
one has
\begin{gather}
F = \sqrt{ C^\mu C_\mu},
\nonumber
\end{gather}
so that $F$ can be viewed as a function of $g$ and $C$.
\begin{convention} Unless stated otherwise, we shall assume from now
on to be working in harmonic (or wave) coordinates.
\end{convention}
\begin{notation}
Below, $B$,
with indices attached when necessary, is used to denote expressions, where the
maximum number of the derivatives of the variables $g$, $s$, $u$, $\Omega$, $F$,
and $C$ is
indicated in the arguments. For instance, $B(\partial g, \partial^2 s)$ indicates an
expression depending on at most first derivatives of $g$ and second derivatives of $s$. The expression represented by $B$ can vary from equation to equation.
\end{notation}
\subsection{Equation for $g$}
By taking the trace of \eqref{original_Einstein} we get
\[
T:=g^{\alpha\beta}T_{\alpha\beta}=-R + 4\Lambda,
\]
so we can rewrite Einstein's equation as
\begin{equation}\label{ens1}
R_{\alpha\beta}=T_{\alpha\beta}-\frac 12 Tg_{\alpha\beta} + \Lambda g_{\alpha\beta}.
\end{equation}
Next, recall that in harmonic coordinates, the Ricci curvature can be written as
\begin{equation}\label{ens2}
R_{\alpha\beta}=\frac 12 g^{\mu\nu}\partial_{\mu\nu}g_{\alpha\beta}+B_{\alpha\beta}(\partial g).
\end{equation}
From \eqref{Tab} we also have
\begin{align}
T_{\alpha\beta}&= rF u_\alpha u_\beta - p g_{\alpha\beta}
+F\vartheta\pi_\alpha^\rho \pi_\beta^\mu (\nabla_\rho u_\mu + \nabla_\mu u_\rho)\nonumber\\
&=B_{\alpha\beta}(\partial g, s, u, C)+ \vartheta\sqrt{C^\nu C_{\nu}}\pi_\alpha^\rho \pi_\beta^\mu (\partial_\rho u_\mu + \partial_\mu u_\rho)\label{ens3}.
\end{align}
Hence
\begin{equation}\label{ens4}
T=rF-4p+\vartheta F\pi^{\rho\mu}(\nabla_\rho u_{\mu}+\nabla_{\mu}u_\rho)=\vartheta\sqrt{C^\nu C_{\nu}}\pi^{\rho\mu} (\partial_\rho u_\mu + \partial_\mu u_\rho)+B(\partial g, s, u, C).
\end{equation}
Inserting \eqref{ens2}, \eqref{ens3}, and \eqref{ens4} into \eqref{ens1} we obtain
the following equation for $g$
\begin{equation}\label{ens5}
g^{\mu\nu}\partial_{\mu\nu}g_{\alpha\beta}-\vartheta\sqrt{C^\nu C_{\nu}}\left(2\pi_\alpha^\rho \pi_\beta^\mu (\partial_\rho u_\mu + \partial_\mu u_\rho) -\pi^{\rho\mu} (\partial_\rho u_\mu + \partial_\mu u_\rho)g_{\alpha\beta}\right)+B_{\alpha\beta}(\partial g, s, u, C) =0.
\end{equation}
\subsection{Equation for $s$}
From \eqref{original_conservation}, \eqref{original_mass_conservation}, \eqref{flt2} by considering $u^\beta\nabla^\alpha T_{\alpha\beta}$ with $T_{\alpha\beta}$ as in \eqref{Tab}, we obtain
\[
r\theta u^\alpha \partial_\alpha s-\vartheta F(\nabla_\mu u_\nu +\nabla_\nu u_\mu)\pi^{\alpha \mu}\nabla_\alpha u^\nu=0,
\]
where we also used \eqref{useful} and that $u^\beta\pi^\mu_\beta=0$.
To obtain the desired quasi-linear structure we apply $u^\sigma\nabla_\sigma$ to the equation. This results in
\begin{equation}\label{s1}
\begin{split}
u^\sigma u^\alpha \partial_{\alpha\sigma} s &- \vartheta\frac{\sqrt{C^\rho C_\rho}}{\theta r} \pi^{\alpha\mu}u^\sigma \partial_\alpha u^\nu (\partial_{\mu \sigma}u_\nu+\partial_{\sigma\nu}u_\mu)- \vartheta\frac{\sqrt{C^\rho C_\rho}}{\theta r} (\partial_{\mu}u_\nu+\partial_{\nu}u_\mu)\pi^{\alpha\mu}u^\sigma \partial_{\alpha\sigma} u^\nu\\
& + B(\partial^2 g, \partial s, \partial C, \partial u)=0.
\end{split}
\end{equation}
We note that in the derivation of $B$ in \eqref{s1} at first one obtains derivatives in $\theta$ and $r$, which get replaced by $\partial s$, and $\partial F$. Then in view of the comment at the beginning of this section, $\partial F$ gets replaced by $\partial C$ and $\partial g$.
\subsection{Equation for $u$.}
The derivation of the equation for $u$ is long, requiring several
calculations. We shall break them into short claims in order to
facilitate the reading.
Since $F >0$, inspired by \cite{LichnerowiczBookGR} we can define the conformal metric
\begin{gather}
\overline{g} = F^2 g,
\nonumber
\end{gather}
and denote by $\overline{\nabla}$ covariant differentiation in the $\overline{g}$-metric.
We also let
\begin{gather}
\overline{C}^\alpha = F^{-1} u^\alpha,
\label{C_bar_inv}
\end{gather}
i.e., $\overline{C}^\alpha$ is $C_\alpha$ with index raised in the $\overline{g}$ metric, so
that
\begin{gather}
\overline{C}^\alpha C_\alpha = 1.
\label{unit_C_bar}
\end{gather}
It also follows that
\begin{gather}
\Omega_{\alpha\beta} = \overline{\nabla}_\alpha C_\beta - \overline{\nabla}_\beta C_\alpha.
\label{vorticity_nabla_bar}
\end{gather}
If $v$ is a one-form, a direct calculation gives
\begin{gather}
\overline{\nabla}_\alpha v_\beta =
\nabla_\alpha v_\beta - K_\alpha v_\beta - K_\beta v_\alpha + K^\rho v_\rho g_{\alpha\beta},
\label{relation_cov_der}
\end{gather}
where $K_\alpha = \partial_\alpha \log F = \frac{\partial_\alpha F}{F}$.
The following standard identities will also be needed,
\begin{gather}
\nabla_\alpha \nabla_\beta v^\lambda - \nabla_\beta \nabla_\alpha v^\lambda
= R_{\alpha\beta \hspace{0.2cm} \gamma}^{\hspace{0.35cm} \lambda} v^\gamma ,
\nonumber
\end{gather}
from which it follows
\begin{gather}
\nabla_\alpha \nabla_\beta v^\alpha - \nabla_\beta \nabla_\alpha v^\alpha
= R_{\beta\gamma} v^\gamma.
\label{contracted_Ricci_id}
\end{gather}
To derive the equation for $u$ we need to compute the divergence of $T_{\alpha\beta}$.
For this it will be convenient to set
\begin{gather}
\Sigma_{\alpha\beta} =
\pi_\alpha^\mu \pi_\beta^\nu (\nabla_\mu C_\nu + \nabla_\nu C_\mu ),
\label{shear_C}
\end{gather}
which can be written as
\begin{gather}
\Sigma_{\alpha\beta} =
F \pi_\alpha^\mu \pi_\beta^\nu (\nabla_\mu u_\nu + \nabla_\nu u_\mu ),
\label{shear_u}
\end{gather}
since $u^\beta\pi^\mu_\beta=0$. $\Sigma_{\alpha\beta}$ is sometimes called the shear tensor.
\begin{claim}
Let
\begin{gather}
\overline{\Sigma}_{\alpha\beta} = \overline{\nabla}_\alpha C_\beta +
\overline{\nabla}_\beta C_\alpha - \overline{C}^\lambda( \overline{\nabla}_\lambda C_\alpha C_\beta +
\overline{\nabla}_\lambda C_\beta C_\alpha ).
\nonumber
\end{gather}
Then the following two identities hold:
\begin{gather}
\overline{\Sigma}_{\alpha\beta} = \Sigma_{\alpha\beta} + 2 \pi_{\alpha\beta} u^\rho \partial_\rho F,
\nonumber
\end{gather}
and
\begin{gather}
\overline{\Sigma}_{\alpha\beta} = 2 \overline{\nabla}_\beta C_\alpha + \Theta_{\alpha\beta},
\nonumber
\end{gather}
where
\begin{gather}
\Theta_{\alpha\beta} = \Omega_{\alpha\beta} - u^\lambda(\Omega_{\lambda \alpha} u_\beta + \Omega_{\lambda \beta} u_\alpha ).
\nonumber
\end{gather}
\label{claim_1}
\end{claim}
\begin{proof}
Using (\ref{relation_cov_der}), (\ref{dyn_vel}), (\ref{C_bar_inv}), a long
but not difficult calculation gives
\begin{gather}
\overline{\Sigma}_{\alpha\beta} =
F(\nabla_\alpha u_\beta + \nabla_\beta u_\alpha - u_\beta u^\nu \nabla_\nu u_\alpha
- u_\alpha u^\mu \nabla_\mu
u_\beta) + 2 \pi_{\alpha\beta} u^\rho \partial_\rho F.
\nonumber
\end{gather}
But in light of (\ref{useful}) we have
\begin{gather}
\pi_\alpha^\mu \pi_\beta^\nu(\nabla_\mu u_\nu + \nabla_\nu u_\mu)
=
\nabla_\alpha u_\beta + \nabla_\beta u_\alpha - u_\beta u^\nu \nabla_\nu u_\alpha
- u_\alpha u^\mu \nabla_\mu
u_\beta ,
\nonumber
\end{gather}
so that (\ref{shear_u}) gives
\begin{gather}
\overline{\Sigma}_{\alpha\beta} = \Sigma_{\alpha\beta} + 2 \pi_{\alpha\beta} u^\rho \partial_\rho F.
\nonumber
\end{gather}
For the second identity, use (\ref{vorticity_nabla_bar}) to get
\begin{align}
\begin{split}
\overline{\Sigma}_{\alpha\beta} & = \overline{\nabla}_\alpha C_\beta +
\overline{\nabla}_\beta C_\alpha - \overline{C}^\lambda( \overline{\nabla}_\lambda C_\alpha C_\beta +
\overline{\nabla}_\lambda C_\beta C_\alpha ) \\
& = \Omega_{\alpha \beta }+
2 \overline{\nabla}_\beta C_\alpha - \overline{C}^\lambda( \Omega_{\lambda\alpha} C_\beta +
\Omega_{\lambda\beta} C_\alpha ) - \overline{C}^\lambda( \overline{\nabla}_\alpha C_\lambda C_\beta +
\overline{\nabla}_\beta C_\lambda C_\alpha ).
\end{split}
\nonumber
\end{align}
The result now follows by noticing that
(\ref{unit_C_bar}) gives $\overline{C}^\lambda \overline{\nabla}_\alpha C_\lambda = 0
= \overline{C}^\lambda \overline{\nabla}_\beta C_\lambda$ and using (\ref{dyn_vel}).
\nonumber
\end{proof}
\begin{claim}
We have
\begin{gather}
\pi^{\gamma\beta} \nabla_\alpha \overline{\Sigma}^\alpha_{\hspace{0.2cm} \beta } =
- 2 \pi^{\gamma\beta} K^\alpha \Omega_{\alpha\beta} + \pi^{\gamma\beta} \nabla_\alpha \Theta^\alpha_{\hspace{0.2cm} \beta}.
\nonumber
\end{gather}
\label{claim_2}
\end{claim}
\begin{proof}
Using claim \ref{claim_1}, (\ref{relation_cov_der}), (\ref{contracted_Ricci_id}),
and (\ref{original_incompressible}),
one gets
\begin{align}
\begin{split}
\nabla_\alpha \overline{\Sigma}^\alpha_{\hspace{0.2cm} \beta } & =
2 \nabla_\alpha \nabla_\beta C^\alpha - 2 \nabla_\alpha K^\alpha C_\beta - 2 K^\alpha \Omega_{\alpha \beta }
+ \nabla_\alpha \Theta^\alpha_{\hspace{0.2cm} \beta} \\
& = 2 R_{\alpha\beta}C^\alpha - 2 \nabla_\alpha K^\alpha C_\beta - 2 K^\alpha \Omega_{\alpha \beta }
+ \nabla_\alpha \Theta^\alpha_{\hspace{0.2cm} \beta}.
\end{split}
\nonumber
\end{align}
Now, from (\ref{ens1}), the fact that $\pi_{\alpha\beta} C^\alpha = F \pi_{\alpha\beta}u^\alpha = 0$,
and the form of $T_{\alpha\beta}$, it follows that
\begin{gather}
R_{\alpha\beta}C^\alpha = (rF - p -\frac{1}{2}T)C_\beta,
\nonumber
\end{gather}
so that
\begin{align}
\begin{split}
\nabla_\alpha \overline{\Sigma}^\alpha_{\hspace{0.2cm} \beta } & =
(2rF - 2p - T - 2 \nabla_\alpha K^\alpha )C_\beta - 2 K^\alpha \Omega_{\alpha \beta }
+ \nabla_\alpha \Theta^\alpha_{\hspace{0.2cm} \beta},
\end{split}
\nonumber
\end{align}
from which the claim follows upon contracting with $\pi^{\gamma\beta} $ and using again
$\pi_{\alpha\beta} C^\alpha = 0$
\end{proof}
\begin{claim} We have
\begin{gather}
2 u^\alpha \pi^{\gamma\rho}\nabla_\alpha \partial_\rho F =
- 2 \pi^{\gamma\beta} K^\alpha \Omega_{\alpha\beta} + \pi^{\gamma\beta}\nabla_\alpha\Theta^\alpha_{\hspace{0.2cm}\beta}
+ B^\gamma(\partial g, \partial s, \partial F, \partial u).
\nonumber
\end{gather}
\label{claim_3}
\end{claim}
\begin{proof}
Combining the first identity of claim \ref{claim_1} with claim \ref{claim_2},
\begin{gather}
\pi^{\gamma\beta} \nabla_\alpha \Sigma^\alpha_{\hspace{0.2cm} \beta } =
-2u^\rho \pi^{\gamma\alpha}\nabla_\alpha \partial_\rho F
- 2 \pi^{\gamma\beta} K^\alpha \Omega_{\alpha\beta} + \pi^{\gamma\beta} \nabla_\alpha \Theta^\alpha_{\hspace{0.2cm} \beta}+ B^\gamma(\partial g, \partial F, \partial u).
\nonumber
\end{gather}
Writing (\ref{T_vis_i}) as $T_{\alpha\beta} = \widehat{T}_{\alpha\beta} + \vartheta \Sigma_{\alpha\beta}$,
noticing that $ \pi^{\gamma\beta} \nabla_\alpha \widehat{T}_{\alpha\beta} = B^\gamma(\partial g,
\partial F, \partial u)$, and invoking (\ref{original_conservation}),
we have
\begin{gather}
\pi^{\gamma\beta} \nabla_\alpha \Sigma^\alpha_{\hspace{0.2cm} \beta } =
B^\gamma(\partial g, \partial s,
\partial F, \partial u),
\label{projection_div_shear_lower_order}
\end{gather}
since $\vartheta > 0$. The claim follows from these last two equalities.
\end{proof}
\begin{claim} We have
\begin{gather}
u^\alpha \nabla_\alpha u_\beta = \pi^\alpha_\beta \frac{\partial_\alpha F}{F} + \frac{1}{F} u^\alpha
\Omega_{\alpha\beta}.
\nonumber
\end{gather}
\label{claim_4}
\end{claim}
\begin{proof}
From (\ref{relation_cov_der}) and (\ref{dyn_vel}),
\begin{gather}
\overline{\nabla}_\alpha C_\beta = F \nabla_\alpha u_\beta - \partial_\beta F u_\alpha+
u^\rho \partial_\rho F g_{\alpha\beta},
\nonumber
\end{gather}
which upon contraction gives
\begin{gather}
\overline{C}^\alpha \overline{\nabla}_\alpha C_\beta = u^\alpha \nabla_\alpha u_\beta
- \pi^\alpha_\beta \frac{\partial_\alpha F}{F}.
\nonumber
\end{gather}
Contracting (\ref{vorticity_nabla_bar}) with $\overline{C}^\alpha$, using
(\ref{unit_C_bar}) and the above equality, one obtains the result, after rewriting
$C$ in terms of $F$ and $u$.
\end{proof}
\begin{claim} We have
\begin{align}
\begin{split}
2 \pi^{\gamma\beta} \nabla^\alpha \Sigma_{\alpha\beta} & =
2 F \nabla^\alpha \nabla_\alpha u^\gamma
- \pi^{\gamma\beta}\nabla_\alpha \Theta^\alpha_{\hspace{0.2cm} \beta} - 2u^\alpha u^\rho \nabla_\alpha \Omega_\rho^{\hspace{0.2cm} \gamma}
\\
& + 2 \pi^{\alpha\mu} \pi^{\gamma \nu} \nabla_\alpha \nabla_\nu C_\mu+
B^\gamma(\partial^2 g, \partial s, \partial F, \partial u, \Omega, \partial C).
\end{split}
\nonumber
\end{align}
\label{claim_5}
\end{claim}
\begin{proof}
Use (\ref{dyn_vel}) in the first term on the right hand side of (\ref{shear_C}) to
write it as
\begin{gather}
\Sigma_{\alpha\beta} = F \pi_\alpha^\mu \pi_\beta^\nu \nabla_\mu u_\nu +
\pi_\alpha^\mu \pi_\beta^\nu \nabla_\nu C_\mu ,
\nonumber
\end{gather}
where $\pi_\beta^\nu u_\nu = 0$ has been employed. Applying
$\pi^{\gamma\beta} \nabla^\alpha$, we get
\begin{gather}
\pi^{\gamma\beta} \nabla^\alpha \Sigma_{\alpha\beta}
= F(\nabla_\alpha \nabla^\alpha u^\gamma - u^\alpha u^\mu \nabla_\alpha \nabla_\mu u^\gamma)
+ \pi^{\alpha\mu} \pi^{\gamma \nu} \nabla_\alpha \nabla_\nu C_\mu
+ B^\gamma (\partial g, \partial F, \partial u, \partial C),
\label{claim_5_exp_1}
\end{gather}
where we have used that (\ref{useful}) implies
\begin{gather}
u^\nu \nabla^\alpha \nabla_\alpha u_\nu
= \nabla^\alpha(u^\nu \nabla_\alpha u_\nu) - \nabla^\alpha u^\nu \nabla_\alpha u_\nu
= B^\gamma(\partial g, \partial u),
\nonumber
\end{gather}
and
\begin{gather}
u^\nu \nabla_\alpha \nabla_\mu u_\nu
= \nabla_\alpha(u^\nu \nabla_\mu u_\nu) - \nabla_\alpha u^\nu \nabla_\mu u_\nu
= B^\gamma(\partial g, \partial u).
\nonumber
\end{gather}
Commuting $u^\mu$ and $\nabla_\alpha$
one obtains, in light of claim \ref{claim_4},
\begin{gather}
F u^\alpha u^\mu \nabla_\alpha \nabla_\mu u^\gamma
= u^\alpha\pi^{\gamma\rho} \nabla_\alpha \partial_\rho F + u^\rho u^\alpha \nabla_\alpha \Omega_\rho^{\hspace{0.2cm} \gamma}
+ B^\gamma(\partial g, \partial F, \partial u, \Omega),
\nonumber
\end{gather}
so that (\ref{claim_5_exp_1}) becomes
\begin{align}
\begin{split}
\pi^{\gamma\beta} \nabla^\alpha \Sigma_{\alpha\beta}
& = F\nabla_\alpha \nabla^\alpha u^\gamma -
u^\alpha\pi^{\gamma\rho} \nabla_\alpha \partial_\rho F - u^\rho u^\alpha \nabla_\alpha \Omega_\rho^{\hspace{0.2cm} \gamma}
\\
& + \pi^{\alpha\mu} \pi^{\gamma \nu} \nabla_\alpha \nabla_\nu C_\mu
+ B^\gamma (\partial g, \partial F, \partial u, \partial C, \Omega).
\end{split}
\nonumber
\end{align}
The result now follows by using claim \ref{claim_3} to eliminate
$u^\alpha\pi^{\gamma\rho} \nabla_\alpha \partial_\rho F$ from the above expression,
after noticing that $K^\alpha \Omega_{\alpha\beta}$ can be absorbed into
$B^\gamma$.
\end{proof}
In light of (\ref{projection_div_shear_lower_order}) and using the definition
of $\Theta$,
claim \ref{claim_5} gives the desired equation for $u$, namely,
\begin{align}
\begin{split}
g^{\mu \nu } \partial_{\mu \nu} u_\gamma &
- \frac{1}{2 \sqrt{C^\rho C_\rho} }
g^{\mu \nu} \partial_\nu \Omega_{\mu \gamma}
- \frac{1}{2 \sqrt{C^\rho C_\rho} }
u^\mu u^\nu \partial_\mu\Omega_{\nu \gamma}
+
\frac{1}{2 \sqrt{C^\rho C_\rho} }
u_\gamma u^\mu g^{\nu \beta} \partial_\beta \Omega_{ \nu \mu}
\\
&
+
\frac{1}{ \sqrt{C^\rho C_\rho} } \pi^{\alpha\mu} \pi^\nu_\gamma \partial_{\alpha\nu} C_\mu +
B_\gamma(\partial^2 g, \partial s, \partial u, \Omega, \partial C) = 0,
\end{split}
\label{u_equation}
\end{align}
where we used $F>0$, and subsequently (\ref{dyn_vel}) to
eliminate
the $F$ dependence.
\subsection{Equations for $\Omega$.}
Recalling that
\begin{gather}
\mathcal L_C \Omega_{\alpha \beta}
= C^\mu \nabla_\mu \Omega_{\alpha\beta} + \nabla_\alpha C^\mu \Omega_{\mu \beta}
+ \nabla_\beta C^\mu \Omega_{\alpha \mu},
\nonumber
\end{gather}
we see that (\ref{evolution_vorticity}) has the form
\begin{gather}
C^\mu \partial_\mu \Omega_{\alpha \beta} + q u^\mu \partial_{\mu\alpha} C_\beta
-qu^\mu \partial_{\mu \beta} C_\alpha + B_{\alpha\beta}(\partial^2 g, \partial s, \partial u, \Omega, \partial C) = 0.
\label{Omega_equation}
\end{gather}
\subsection{Equations for $C$.}
In order to close the system, we need to specify equations of motion for
$C$. Since all the equations from (\ref{original}) have already
been employed above in the derivation of equations for $g$, $s$, $u$,
and $\Omega$,
one suspects that equations for $C$ should be determined by
some extra conditions, not explicitly present in (\ref{original}).
However, in order to do so without changing the content of the original
problem, one must choose equations that are necessarily satisfied
by any solution to (\ref{original}). Thus, convenient identities,
that follow from standard tensor calculus and our basic definitions,
will be employed.
Using the definition of the Hodge-Laplacian gives
\begin{gather}
\Delta C = d \delta C + \delta d C = \delta \Omega,
\nonumber
\end{gather}
where $\delta C = - \nabla^\mu C_\mu = 0$ (see (\ref{original_incompressible})) and
the definition of $\Omega$, (\ref{vorticity}), have been used. On the other hand,
recalling
\begin{gather}
(\Delta C)_\alpha = - \nabla^\mu \nabla_\mu C_\alpha + R_{\mu\alpha}C^\mu,
\nonumber
\end{gather}
we obtain the following equation for $C$:
\begin{gather}
g^{\mu\nu}\partial_{\mu\nu} C_\alpha - g^{\mu\nu} \partial_\mu \Omega_{\nu \alpha}
+ B_\alpha(\partial^2 g, \Omega, \partial C) = 0.
\label{C_equation}
\end{gather}
\subsection{The full system.}
The sought new system of equations consists of (\ref{ens5}),
(\ref{s1}), (\ref{u_equation}),
(\ref{Omega_equation}), and (\ref{C_equation}).
These are 25 equations for the 25 unknowns: ten $g_{\alpha\beta}$, one $s$,
four $u_\alpha$, six $\Omega_{\alpha\beta}$, and four $C_\alpha$.
We shall refer to this system as the modified
incompressible Einstein-Navier-Stokes system.
One important aspect of our proof consists in showing that the
modified incompressible Einstein-Navier-Stokes system
forms a Leray-Ohya system \cite{LerayOhyaNonlinear}, which
depends, among other things, on a counting of
derivatives.
For this purpose, it is convenient to write the system
symbolically
as
\begin{align}
\left\{
\begin{matrix}
a_{11}(g) \partial^2 g & + & 0 & + & a_{13}(g,u,C)\partial u & + & 0
\\
0 & + & a_{22}(g,u)\partial^2 s & + & a_{23}(g,s,\partial u, C) \partial^2 u
&+& 0
\\
0 & + & 0 & + & a_{33}(g) \partial^2 u & + & a_{34}(g,u,C)\partial \Omega
\\
0 & + & 0 & + &0 & + & a_{44}(g,C)\partial \Omega
\\
0 & + & 0 & + &0 & + & a_{54}(g)\partial \Omega
\end{matrix}
\nonumber
\right.
\end{align}
\begin{subequations}
\begin{alignat}{4}
& + && \hspace{1.2cm} 0 && = && \hspace{0.25cm} B_g( \partial g, s, u, C),
\label{eq1} \\
&+ && \hspace{1.2cm} 0 && = && \hspace{0.25cm} B_s(\partial^2 g, \partial s, \partial u, \partial C),
\label{eq2} \\
& + && \hspace{0.25cm} a_{35}(g,u,C) \partial^2 C && = &&
\hspace{0.25cm} B_u(\partial^2 g, \partial s, \partial u, \Omega, \partial C),
\label{eq3} \\
& + && \hspace{0.25cm} a_{45}(g,u) \partial^2 C && = &&
\hspace{0.25cm} B_\Omega(\partial^2 g, \partial s, \partial u, \Omega, \partial C),
\label{eq4} \\
& + && \hspace{0.25cm} a_{55}(g) \partial^2 C && = &&
\label{eq5} \hspace{0.25cm} B_C (\partial^2 g, \Omega, \partial C) .
\end{alignat}
\label{new_system_symbolic}
\end{subequations}
We write this more succinctly as
\begin{gather}
A(V,\partial)V = B(V),
\nonumber
\end{gather}
where $V=(g,s,u,\Omega, C)$, $B(V)=(B_g, B_s, B_u, B_\Omega, B_C)$,
and
\begin{align}
A(V,\partial) = \left(
\begin{matrix}
a_{11}(g) \partial^2 & 0 & a_{13}(g,u,C)\partial & 0 & 0
\\
0 & a_{22}(g,u)\partial^2 & a_{23}(g,s,\partial u, C) \partial^2 & 0 & 0 &
\\
0 & 0 & a_{33}(g) \partial^2 & a_{34}(g,u,C)\partial & a_{35}(g,u,C) \partial^2 \\
0 & 0 & 0 & a_{44}(g,C)\partial & a_{45}(g, u) \partial^2 \\
0 & 0 & 0 & a_{54}(g)\partial & a_{55}(g) \partial^2
\end{matrix}
\right)
\label{matrix}
\end{align}
\section{Proof of Theorem \ref{main_theorem}. \label{main_proof}}
The main tool in the proof of theorem \ref{main_theorem} is
the theory of weakly hyperbolic equations in Gevrey spaces developed
by Leray and Ohya
\cite{LerayOhyaLinear,LerayOhyaNonlinear, OhyaLinear}
and extended by Choquet-Bruhat \cite{CB_diagonal} to the type of non-diagonal
systems which will be of interest here. We shall not review these
constructions, except for those aspects that will be necessary
to fix our notation and conventions, referrring the reader to the
above references for the complete discussion.
Some aspects of the Leray-Ohya theory can
also be found (without proofs) in
\cite{ChoquetBruhatGRBook, ChoquetetallAnalysisManifolds, DisconziViscousFluidsNonlinearity}.
For the reader's convenience, we start by recalling the definition of
Gevrey spaces. As our proof is essentially local, with a global (in space)
solution obtained by a gluing argument, it suffices
to give the definition in the case of $\mathbb R^{n+1}$, whose coordinates
we denote by $(x_0, \dots, x_{n})$.
For a number $|X| > 0$, let $X$ be the strip
$0 \leq x^0 \leq |X|. $
The Gevrey space $\gamma^{m, (\sigma)}(X)$ is defined as follows.
Let $S_t = \{ x^0 =t \}$, and
\begin{gather}
|D^k u |_t = c(n,k) \sup_{|\alpha| \leq k} \parallel D^\alpha u \parallel_{L^2(S_t)},
\nonumber
\end{gather}
where
$c(n,k)$ is a normalization constant. Then, for $\sigma \geq 1$, and
$m$ a non-negative integer,
$ u \in \gamma^{m, (\sigma)}(X)$ means that $u \in C^\infty(X)$, and
\begin{gather}
\sup_{|\beta| \leq m, \, \alpha, \, 0 \leq t \leq |X|} \frac{1}{ (1 + |\alpha| )^\sigma }
\left( |D^{\beta + \alpha} u |_t \right)^\frac{1}{ 1 + |\alpha| } < \infty,
\nonumber
\end{gather}
where the sup over $\alpha$ is taken over multi-indices such that $\alpha_0 = 0$.
Analogously one defines such spaces for open sets, product spaces, etc.
Intuitively, $\gamma^{m, (\sigma)}(X)$ can be thought of as a space between
analytic and smooth functions, in the following sense. An analytic function
$u$
on $S_t$ obeys, on each compact set, an inequality of the form
$|D^\alpha u| \leq C^{|\alpha|+1} \alpha!$, for some $C>0$. Gevrey functions
of class $\sigma \geq 1$ are smooth functions obeying
the weaker inequality $|D^\alpha u| \leq C^{|\alpha|+1} (\alpha!)^\sigma$.
Then, $\gamma^{m, (\sigma)}(X)$ consists of those functions whose
derivatives up to order $m$, restricted to each time slice $S_t$,
belong to the Gevrey space of class $\sigma$ --- except that it is convenient
to characterize the Gevrey spaces of $S_t$ with the help of an integral norm,
as done above. Gevrey spaces of functions defined on a hypersurface
$\Sigma \subset X$ (say, on $\{ x^0 = 0 \}$), are defined in an analogous fashion
and denoted by $\gamma^{(\sigma)}(\Sigma)$. These will be the spaces
where initial data is prescribed (notice that there is no supremum over
$\beta$ in this case).
Gevrey spaces are important in particular because
it is possible to establish in them the well-posedness of certain PDEs that
are known not to be well-posed in Sobolev or smooth spaces \cite{MizohataCauchyProblem}.
At the same time, Gevrey spaces allow constructions with compactly supported functions,
an important tool in analysis not possible in the class of analytic
functions.
See \cite{LerayOhyaNonlinear,RodinoGevreyBook} for details on Gevrey spaces and their applications.
Consider a system of $N$ partial differential
equations and $N$ unknowns in $X = \mathbb R^n \times [0,T]$,
and
denote the unknown as
$V=(v^I)$, $I=1,\dots, N$. Suppose that the
system has the following quasi-linear structure: it is possible
to attach to each unknown $v^I$ a non-negative integer $m_I$, and to
each equation a non-negative integer $n_J$, such that the system reads
\begin{gather}
h^J_I(\partial^{m_K - n_J -1} v^K, \partial^{m_I - n_J}) v^I
+ b^J(\partial^{m_K - n_J - 1} v^K) = 0, \, J=1, \dots, N.
\label{general_system}
\end{gather}
The notation here is similar to the
one we used to write system (\ref{new_system_symbolic}), namely,
$h^J_I(\partial^{m_K - n_J -1} v^K, \partial^{m_I - n_J})$
is a homogeneous differential operator of order $m_I - n_J$ (which can
be zero), whose coefficients depend on at most
$m_K - n_J -1$ derivatives of $v^K$, $K=1,\dots N$. The remaining terms,
$b^J(\partial^{m_K - n_J - 1} v^K)$, also depend on at most
$m_K - n_J -1$ derivatives of $v^K$, $K=1,\dots N$.
Recall that the characteristic polynomial of (\ref{general_system})
at $x \in X$ and for a given $V$
is
the polynomial in the co-tangent space $T^*_x X$,
$p(\xi)$, $\xi \in T_x^* X$,
of degree
\begin{gather}
\ell = \sum_{I=1}^N m_I - \sum_{J=1}^N n_J,
\nonumber
\end{gather}
given by the principal part (of order $\ell$)
of the characteristic determinant of the system, $
\det(h^J_I(\xi))$.
Consider the
Cauchy problem for (\ref{general_system}), with Cauchy data given
on $X_0 = \mathbb R^n \times \{ t = 0\}$. Assume that for any $x \in X_0$,
and with $V$ taking the values of the Cauchy data on $X_0$,
the characteristic polynomial $p(\xi)$ is a product of $q$
hyperbolic polynomials of orders $\ell_q$,
\begin{gather}
p(\xi) = p_{1}(\xi) \cdots p_q(\xi).
\nonumber
\end{gather}
Suppose finally that
\begin{gather}
\max_q \ell_q \geq \max_I m_I - \min_J n_J.
\nonumber
\end{gather}
Building on the techniques developed in \cite{LerayOhyaNonlinear}
Choquet-Bruhat proved \cite{CB_diagonal}
under the above conditions,
the Cauchy problem for (\ref{general_system}) has a unique solution
$V$ in the Gevrey space $\gamma^{\operatorname{m},(\sigma)}(X^\prime)$,
where $X^\prime = \mathbb R^n \times [0,T^\prime]$, $T^\prime \leq T$,
for a suitable integer
$\operatorname{m} \geq 0$, and $1 \leq \sigma <
\sigma_0 = \frac{q}{q-1}$ (the case $q=1$,
$\sigma_0 = \infty$, corresponds to solutions in Sobolev spaces). Furthermore,
the solution enjoys the domain of dependence or finite propagation speed
property, with the domain of dependence of a point $x \in X^\prime$
determined by the characteristic cone $\{ p(\xi) = 0 \}$ at $x$. \\
\noindent \emph{Proof of Theorem \ref{main_theorem}.} We shall verify
that the modified incompressible Einstein-Navier-Stokes system
is of the form (\ref{general_system}) and satisfies all the
conditions given in \cite{CB_diagonal} which we summarized
above.
Consider the unknown $V=(g,s,u,\Omega, C) = (v^1, v^2, v^3, v^4, v^5)$
for the system (\ref{new_system_symbolic}). Naturally, it is understood
that each $v^I$ and each equation in (\ref{new_system_symbolic}) represent,
respectively, a set of unknowns and a set of equations, but they can be
grouped together since they are all of the same form. For instance,
for all the ten unknowns $g$, all the equations take the same form
(\ref{ens5}). We also remark that, as it is standard in the study of the
evolution problem in General Relativity, although $V$ and
(\ref{new_system_symbolic}) are defined in a local coordinate patch, we
rely on the aforementioned results \cite{CB_diagonal, LerayOhyaNonlinear},
given for $\mathbb R^n \times [0,T]$ ($n=3$ in our case),
using the finite propagation speed and a standard gluing argument
to construct global in space solutions (see below).
One verifies that (\ref{new_system_symbolic}) has the form (\ref{general_system})
upon choosing
\begin{align}
\begin{array}{ccccc}
m_1 = 3, & m_2 = 2, & m_3 = 2, & m_4 = 1, & m_5 = 2, \\
n_1 = 1, & n_2 = 0, & n_3 = 0, & n_4 = 0, & n_5 = 0,
\end{array}
\label{indices}
\end{align}
where $m_1 = m(v^1) \equiv m(g)$, $m_2 \equiv m(v^2) = m(s)$,
$m_3 = m(v^3) \equiv m(u)$, $m_4 = m(v^4) \equiv m(\Omega)$,
$m_5 = m(v^5) \equiv m(C)$,
$n_1 = n(\text{equation } (\ref{eq1}) )
\equiv n(\text{equation } (\ref{ens5}) )$,
$n_2 = n(\text{equation } (\ref{eq2}) ) \equiv
n(\text{equation } (\ref{s1}) )$,
$n_3 = n(\text{equation } (\ref{eq3}) ) \equiv
n(\text{equation } (\ref{u_equation}) )$,
$n_4 = n(\text{equation } (\ref{eq4}) ) \equiv
n(\text{equation } (\ref{Omega_equation}) )$,
$n_5 = n(\text{equation } (\ref{eq5}) ) \equiv
n(\text{equation } (\ref{C_equation}) )$,
and letting $h^J_I$ be the differential operator whose matrix $(h^J_J)$
is given by (\ref{matrix}). Indeed, we list below
for each equation $J$ in (\ref{new_system_symbolic}), the value of $n_J$,
the highest
derivatives
of the each unknown entering in the coefficients and the right-hand side of the equation, and the difference
$m_I - n_J$:
\begin{gather}
\text{eq. } (\ref{eq1}):
n_1 = 1;
\partial g, s, u, C;
\begin{cases}
m(g) - n_1 \equiv m_1 - n_1 = 2, \\
m(s) - n_1 \equiv m_2 - n_1 = 1, \\
m(u) - n_1 \equiv m_3 - n_1 = 1, \\
m(\Omega) - n_1 \equiv m_4 - n_1 = 0, \\
m(C) - n_1 \equiv m_5 - n_1 = 1,
\end{cases}
\nonumber
\end{gather}
\begin{gather}
\text{eq. } (\ref{eq2}):
n_2 = 0;
\partial^2 g, \partial s, \partial u, \partial C;
\begin{cases}
m(g) - n_2 \equiv m_1 - n_2 = 3, \\
m(s) - n_2 \equiv m_2 - n_2 = 2, \\
m(u) - n_2 \equiv m_3 - n_2 = 2, \\
m(\Omega) - n_2 \equiv m_4 - n_2 = 1, \\
m(C) - n_2 \equiv m_5 - n_2 = 2,
\end{cases}
\nonumber
\end{gather}
\begin{gather}
\text{eq. } (\ref{eq3}):
n_3 = 0;
\partial^2 g, \partial s, \partial u, \Omega, \partial C;
\begin{cases}
m(g) - n_3 \equiv m_1 - n_3 = 3, \\
m(s) - n_3 \equiv m_2 - n_3 = 2, \\
m(u) - n_3 \equiv m_3 - n_3 = 2, \\
m(\Omega) - n_3 \equiv m_4 - n_3 = 1, \\
m(C) - n_3 \equiv m_5 - n_3 = 2,
\end{cases}
\nonumber
\end{gather}
\begin{gather}
\text{eq. } (\ref{eq4}):
n_4 = 0;
\partial^2 g, \partial s, \partial u, \Omega, \partial C;
\begin{cases}
m(g) - n_4 \equiv m_1 - n_4 = 3, \\
m(s) - n_4 \equiv m_2 - n_4 = 2, \\
m(u) - n_4 \equiv m_3 - n_4 = 2, \\
m(\Omega) - n_4 \equiv m_4 - n_4 = 1, \\
m(C) - n_4 \equiv m_5 - n_4 = 2,
\end{cases}
\nonumber
\end{gather}
\begin{gather}
\text{eq. } (\ref{eq5}):
n_5 = 0;
\partial^2 g, \Omega, \partial C;
\begin{cases}
m(g) - n_5 \equiv m_1 - n_5 = 3, \\
m(s) - n_5 \equiv m_2 - n_5 = 2, \\
m(u) - n_5 \equiv m_3 - n_5 = 2, \\
m(\Omega) - n_5 \equiv m_4 - n_5 = 1, \\
m(C) - n_5 \equiv m_5 - n_5 = 2.
\end{cases}
\nonumber
\end{gather}
As described in \cite{CB_diagonal, LerayOhyaNonlinear},
the Cauchy data for a system of the form (\ref{general_system}) consists
of the functions $v^I$, along with their derivatives up to order $m_J - 1$,
on the initial surface. The initial data is also required to satisfy
some compatibility conditions, which essentially come from requiring
that the equations are satisfied on the initial time slice when
they take the values of the initial data. In our case, we have
to further ensure that the initial data for the system (\ref{new_system_symbolic})
is compatible with solutions of the original set of equations,
i.e., (\ref{original}) (written in harmonic coordinates), and with
(\ref{dyn_vel}) and (\ref{vorticity}).
The derivation, out of the original
initial data, of Cauchy data for (\ref{new_system_symbolic}),
such that the conditions of the last paragraph are satisfied,
is done in similar fashion as in \cite{DisconziViscousFluidsNonlinearity}, and
therefore we shall skip the details. In a nutshell, one uses the
equations of motion to derive what the new initial data ought to be.
In fact, such a procedure is commonly used in General Relativity
when solutions to Einstein equations are found via a different
set of equations
\cite{ChoquetBruhatGRBook,
CB_York,
DisconziRemarksEinsteinEuler,
DisconziVamsiEinsteinEuler,
Fischer_Marsden_Einstein_FOSH_1972, Fri1, Fri2, Fri3, Fri4, Fri5, FriRenCauchy, Lichnerowicz_MHD_book}. We remark, for future reference,
that although we are treating $\Omega$, $C$, and $u$ as independent variables,
and hence we do not know yet that (\ref{dyn_vel}) and (\ref{vorticity})
hold true, these relation are satisfied by the initial data by the way
they are derived; see \cite{DisconziViscousFluidsNonlinearity}.
Next, we need to compute the characteristic determinant of the system,
$\det A(V,\xi)$, where $\xi$ is a co-vector and $A(V,\xi)$ the principal symbol in
the
direction of $\xi$. From (\ref{matrix}),
\begin{gather}
\det A(V,\xi) =
\det a_{11}(g, \xi) \det a_{22}(g,u,\xi) \det a_{33}(g,\xi)
\det \widetilde{A}(g,u,C,\xi),
\label{det_product}
\end{gather}
where $a_{ij}(\cdot,\xi)$ and $\widetilde{A}(g,u,C,\xi)$ are,
respectively, the principal symbols of the differential
operators $a_{ij}$ and
\begin{align}
\widetilde{A}(g,u,C,\partial) = \left(
\begin{matrix}
a_{44}(g,C)\partial & a_{45}(g,u) \partial^2 \\
a_{54}(g)\partial & a_{55}(g) \partial^2
\end{matrix}
\right).
\nonumber
\end{align}
From (\ref{ens5}),
(\ref{s1}), (\ref{u_equation}), we find,
\begin{gather}
\det a_{11}(g, \xi) = (\xi^\mu \xi_\mu)^{10},
\label{det_a_11}
\end{gather}
\begin{gather}
\det a_{22}(g, u, \xi) = (u^\mu \xi_\mu)^2,
\label{det_a_22}
\end{gather}
and
\begin{gather}
\det a_{33}(g, \xi) = (\xi^\nu \xi_\nu)^4,
\label{det_a_33}
\end{gather}
where as usual the indices are raised with $g$, i.e.,
$\xi^\mu \xi_\mu = g^{\mu \nu} \xi_\nu \xi_\mu$, and
$u^\mu \xi_\mu = g^{\mu\nu} u_\nu \xi_\mu$. The powers
$10$ and $4$ in (\ref{det_a_11}) and
(\ref{det_a_33}) come, respectively, from the fact that
(\ref{ens5}) corresponds to
ten equations and (\ref{u_equation}) to
four equations, whereas the power $2$
in (\ref{det_a_22}) comes from the double characteristic
$u^\alpha u^\beta \xi_\alpha \xi_\beta$ of $u^\alpha u^\beta \partial_{\alpha\beta} s$
in (\ref{s1}).
The operator $\widetilde{A}$ has a more complicated structure,
which requires us to be more explicit. Recalling that $\Omega$
has six independent components, the components $(\Omega,C)$ in
$V=(g,s,u,\Omega, C)$ are
\begin{gather}
(\Omega_{01}, \Omega_{02}, \Omega_{03}, \Omega_{12}, \Omega_{13}, \Omega_{23},
C_0, C_1, C_2, C_3),
\nonumber
\end{gather}
From (\ref{Omega_equation}) and (\ref{C_equation}), we then see
that $\widetilde{A}$ has the following form
\begin{gather}
\begin{bmatrix}
C^\mu \partial_\mu & 0 & 0 & 0 & 0 & 0 & - q u^\mu \partial_{\mu 1} &
q u^\mu \partial_{\mu 0} & 0 & 0 \\
0 & C^\mu \partial_\mu & 0 & 0 & 0 & 0 & -q u^\mu \partial_{\mu 2} &
0 & q u^\mu \partial_{\mu 0} & 0\\
0 & 0 & C^\mu \partial_\mu & 0 & 0 & 0 & -q u^\mu \partial_{\mu3} &
0 & 0 & q u^\mu \partial_{\mu 0 } \\
0 & 0 & 0 & C^\mu \partial_\mu & 0 & 0 & 0 & -q u^\mu \partial_{\mu 2} &
q u^\mu \partial_{\mu 1} & 0 \\
0 & 0 & 0 & 0 & C^\mu \partial_\mu & 0 & 0 & -q u^\mu \partial_{\mu 3} &
0 & q u^\mu \partial_{\mu1} \\
0 & 0 & 0 & 0 & 0 & C^\mu \partial_\mu & 0 & 0 & -q u^\mu \partial_{\mu 3} &
qu^\mu \partial_{\mu 2} \\
g^{\mu 1} \partial_\mu & g^{\mu 2} \partial_\mu & g^{\mu 3} \partial_\mu &
0 & 0 & 0 & g^{\mu \nu } \partial_{\mu \nu} & 0 & 0 & 0 \\
- g^{\mu 0} \partial_\mu & 0 & 0 & g^{\mu 2} \partial_\mu & g^{\mu 3} \partial_\mu &
0 & 0 & g^{\mu\nu} \partial_{\mu\nu} & 0 & 0 \\
0 & - g^{\mu 0} \partial_\mu & 0 & - g^{\mu 1} \partial_\mu & 0 &
g^{\mu 3} \partial_\mu & 0 & 0 & g^{\mu\nu}\partial_{\mu\nu} & 0 \\
0 & 0 & - g^{\mu 0} \partial_\mu & 0 & - g^{\mu 1} \partial_\mu &
-g^{\mu 2} \partial_\mu & 0 & 0 & 0 & g^{\mu\nu} \partial_{\mu\nu}
\end{bmatrix}
\nonumber
\end{gather}
$\widetilde{A}(g,u,C,\xi)$ is given by
\begin{gather}
\begin{bmatrix}
C^\mu \xi_\mu & 0 & 0 & 0 & 0 & 0 & - q u^\mu \xi_{\mu} \xi_{ 1} &
q u^\mu \xi_{\mu} \xi_{ 0} & 0 & 0 \\
0 & C^\mu \xi_\mu & 0 & 0 & 0 & 0 & - q u^\mu \xi_{\mu} \xi_{ 2} &
0 & q u^\mu \xi_{\mu} \xi_{ 0} & 0 \\
0 & 0 & C^\mu \xi_\mu & 0 & 0 & 0 & - q u^\mu \xi_{\mu} \xi_{ 3} &
0 & 0 & q u^\mu \xi_{\mu } \xi_{ 0 } \\
0 & 0 & 0 & C^\mu \xi_\mu & 0 & 0 & 0 & - q u^\mu \xi_{\mu} \xi_{ 2} &
q u^\mu \xi_{\mu } \xi_{ 1} & 0 \\
0 & 0 & 0 & 0 & C^\mu \xi_\mu & 0 & 0 & - q u^\mu \xi_{\mu} \xi_{ 3} &
0 & q u^\mu \xi_{\mu } \xi_{ 1} \\
0 & 0 & 0 & 0 & 0 & C^\mu \xi_\mu & 0 & 0 & - q u^\mu \xi_{\mu} \xi_{ 3} &
q u^\mu \xi_{\mu} \xi_{ 2} \\
\xi^1 & \xi^2 & \xi^3 &
0 & 0 & 0 & \xi^\mu \xi_\mu & 0 & 0 & 0 \\
- \xi^0 & 0 & 0 & \xi^2 & \xi^3 &
0 & 0 & \xi^\mu \xi_\mu & 0 & 0 \\
0 & - \xi^0 & 0 & - \xi^1 & 0 &
\xi^3 & 0 & 0 & \xi^\mu \xi_\mu & 0 \\
0 & 0 & - \xi^0 & 0 & - \xi^1 &
-\xi^2 & 0 & 0 & 0 & \xi^\mu \xi_\mu
\end{bmatrix}
\nonumber
\end{gather}
The determinant of the above matrix can be computed, yielding, after much algebra,
\begin{gather}
F^3(F+q)^2 (u^\mu \xi_\mu)^6 (\xi^\lambda \xi_\lambda)^2 P(\xi),
\label{det_complicated}
\end{gather}
where
\begin{gather}
P(\xi) = A \xi_0^4 + B \xi_0^2 + C
\nonumber
\end{gather}
and we have used that $g_{00} = 1$ and $g_{0i} = 0$.
The coefficients $A$, $B$, and $C$ are given by
\begin{gather}
A = F+q,
\nonumber
\end{gather}
\begin{gather}
B = 2 F \xi_1 \xi^1 + 2q \xi_1 \xi^1 + 2F \xi_2 \xi^2 + 2q \xi_2 \xi^2 + q \xi_3 \xi^2 + 2F \xi_3 \xi^3
+ q \xi_3 \xi^3,
\nonumber
\end{gather}
and
\begin{gather}
C = F (\xi_1 \xi^1)^2 + q (\xi_1 \xi^1)^2 + 2F \xi_1 \xi^2 \xi_2 \xi^2 + 2q \xi_1 \xi^2 \xi_2 \xi^2
+ q \xi_1 \xi_3 \xi^1 \xi^2 + F (\xi_2 \xi^2)^2 + q (\xi_2 \xi^2)^2
\nonumber \\
+ q \xi_2 \xi_3 (\xi_2)^2 + 2 F \xi_1 \xi_3 \xi^1 \xi^3 + q \xi_1 \xi_3 \xi^1 \xi^3
+ 2F \xi_2 \xi_3 \xi^2 \xi^3 + q \xi_2 \xi_3 \xi^2 \xi^3 + q (\xi_3)^2 \xi^2 \xi^3 + F (\xi_3 \xi^3)^2.
\nonumber
\end{gather}
We investigate the roots of $P(\xi)$. We have
\begin{gather}
(\xi_0)^2 = \frac{-B \pm \sqrt{ B^2 - 4AC}}{2A}.
\nonumber
\end{gather}
We need to verify that the right-hand side is real and non-negative. A long but not difficult
computation reveals that
\begin{gather}
B^2 - 4AC = q^2 \xi_3^2 (\xi^2 - \xi^3)^2,
\nonumber
\end{gather}
assuring reality. For non-negativity, it suffices to show that $-B - \sqrt{ B^2 - 4AC} \geq 0$.
For this, assume first that we are working at a point where $g$ equals the Minkowski metric, so that,
after some more algebra and recalling our sign convention,
\begin{gather}
-B - \sqrt{ B^2 - 4AC} =
2(F+q)( (\xi_1)^2 + (\xi_2)^2 + \frac{(\xi_3)^2}{2} ) + F (\xi_3)^2 + q \xi_3 \xi_2
- \sqrt{ q^2 \xi_3^2 (\xi_2 - \xi_3)^2 }.
\nonumber
\end{gather}
It suffices to analyze the case where the term $q \xi_3 \xi_2$ gives a non-positive contribution.
Thus we can replace $q \xi_3 \xi_2$ by $- q \xi_3 \xi_2$ and assume that $\xi_2 \geq 0$ and
$\xi_3 \geq 0$, in which case the above becomes
\begin{gather}
-B - \sqrt{ B^2 - 4AC} =
2(F+q)( (\xi_1)^2 + (\xi_2)^2 + \frac{(\xi_3)^2}{2} ) + F (\xi_3)^2 - q \xi_3 \xi_2
- q \xi_3 |\xi_2 - \xi_3 |.
\nonumber
\end{gather}
If $\xi_2 \geq \xi_3$, we find that
\begin{align}
\begin{split}
-B - \sqrt{ B^2 - 4AC} & =
2(F+q)( (\xi_1)^2 + (\xi_2)^2 + \frac{(\xi_3)^2}{2} ) + F (\xi_3)^2 - q \xi_3 \xi_2
- q \xi_3 (\xi_2 - \xi_3 ) \\
& = 2(F+q)( (\xi_1)^2 + (\xi_2)^2 + \frac{(\xi_3)^2}{2} ) + (F+q) (\xi_3)^2 - 2 q \xi_2 \xi_3
\\
& \geq 2(F+q)( (\xi_1)^2 + (\xi_2)^2 + \frac{(\xi_3)^2}{2} ) + (F+q) (\xi_3)^2 - 2 q (\xi_2)^2 \\
& = 2 (F+q) (\xi_1)^2 + 2F (\xi_2)^2 + 2 (F+q) (\xi_3)^2 \geq 0.
\end{split}
\label{inequality_Min_1}
\end{align}
If $\xi_2 \leq \xi_3$:
\begin{align}
\begin{split}
-B - \sqrt{ B^2 - 4AC} & =
2(F+q)( (\xi_1)^2 + (\xi_2)^2 + \frac{(\xi_3)^2}{2} ) + F (\xi_3)^2 - q \xi_3 \xi_2
- q \xi_3 (\xi_3 - \xi_2 ) \\
& = 2(F+q)( (\xi_1)^2 + (\xi_2)^2 ) + (F+q) (\xi_3)^2 + F (\xi_3)^2 - q (\xi_3)^2 \\
& = 2 (F+q) ( (\xi_1)^2 + (\xi_2)^2 ) + 2 F (\xi_3)^2 \geq 0.
\end{split}
\label{inequality_Min_2}
\end{align}
Therefore, we conclude that $P(\xi)$ factors as the product of two hyperbolic polynomials of degree two,
$P(\xi) = P_1(\xi) P_2(\xi)$,
at least at a point where the metric equals the Minkowski metric.
Now we consider the general case, i.e., when $g$ does not necessarily equal the Minkowski metric.
Consider the initial hypersurface $\Sigma$ where the Cauchy data is given. We can assume that the coordinate
chart on $\Sigma$ is the neighborhood of a point $p$ such that $g_{\alpha\beta}(p) = \eta_{\alpha\beta}$, where
$\eta_{\alpha\beta}$ is the Minkowski metric. Notice that we can still assure that the same coordinates
are harmonic coordinates, since the latter are determined by prescribing the first derivatives of $g$.
Since (\ref{inequality_Min_1}) and (\ref{inequality_Min_2}) are strict inequalities when $\xi \neq 0$,
we see that $-B - \sqrt{B^2 - 4AC} \geq 0$ for points sufficiently near $p$. Therefore, $P(\xi)$
is the product of two hyperbolic polynomial, as desired (notice that the reality condition previously
verified did not use $g_{\alpha\beta}(p) = \eta_{\alpha\beta}$).
Combining (\ref{det_product}), (\ref{det_a_11}), (\ref{det_a_22}),
(\ref{det_a_33}), (\ref{det_complicated}), and the above,
we conclude that
\begin{gather}
\det A(V,\xi) = F^3 (F+ q)^3 (\xi^\mu \xi_\mu)^{14}
(u^\nu \xi_\nu)^6 (u^\tau \xi_\tau \xi^\rho \xi_\rho )^2 P_1(\xi) P_2(\xi),
\label{product_hyp_pol}
\end{gather}
where it is understood that the above expression is evaluated
at the initial data since,
as mentioned in the beginning of this section, we need to verify
the hyperbolicity conditions of \cite{CB_diagonal} when
the unknown $V$ takes the values of the Cauchy data\footnote{More precisely,
we need to verify the hyperbolicity conditions when $V$ and its derivatives up to an order determined by the compatibility
conditions take the values of the Cauchy data
\cite{CB_diagonal, DisconziViscousFluidsNonlinearity}. Here, however, the
coefficients appearing in the determinant (\ref{product_hyp_pol}) do not
involve derivatives of $V$.}.
In particular, also as already mentioned,
even though we
are treating $C$ and $u$ as independent variables, for the initial data
it holds that $C = F u$, and we used this fact to eliminate $C$ from (\ref{product_hyp_pol}).
It is well-known (see e.g. \cite{Lichnerowicz_MHD_book}) that
the first, second, and third degree polynomials
$u^\tau \xi_\tau$, $\xi^\mu \xi_\mu$, and $u^\nu \xi_\nu \xi^\rho \xi_\rho$,
are hyperbolic as long as $g$ is a Lorentzian metric and $u$ is time-like, conditions that are to be fulfilled when $V$ takes
our initial conditions. Also, $F+q > 0$ because $q > 0$ by hypothesis and
$F \geq 1$ by (\ref{indexF}), which holds for the initial data,
as well as the fact that $\epsilon \geq 0$, $p \geq 0$,
and $r >0$.
$\det A(V,\xi)$ is, therefore, a product of 24 hyperbolic polynomials,
with the highest degree of such polynomials being equal to three.
Thus, in the notation employed at the beginning of this section and
with indices $m_I$ and $n_J$ given by (\ref{indices}), we verify that
\begin{gather}
3 = \max_q \ell_q \geq \max_I m_I - \min_J n_J = 3 - 0 = 3.
\nonumber
\end{gather}
We also have $\sigma_0 = \frac{24}{24-1} = \frac{24}{23}$.
The coefficients of the the differential operator $A(V,\partial)$
depend polynomially on $V$, whereas $B(V)$ is a rational function
of the functions $v^I$. The denominator of the rational expressions
appearing in $B(V)$ are products of $\sqrt{C^\rho C_\rho}$,
$r = r(F,s) \equiv r(\sqrt{C^\rho C_\rho},s)$, and
$\theta = \theta(F,s) \equiv \theta(\sqrt{C^\rho C_\rho},s)$.
Hence, recalling (\ref{temperature}) and that $F>0$,
the denominators in such rational expressions are, as functions of $V$,
uniformly bounded away from zero (recall that $\Sigma$ is compact)
when $V$ takes the Cauchy data.
We have, therefore, verified all the conditions necessary to apply
Choquet-Bruhat's theorem \cite{CB_diagonal}
combined with Leray and Ohya's results \cite{LerayOhyaNonlinear},
obtaining a short-in-time solution $V$
to (\ref{new_system_symbolic})
with $v^I \in \gamma^{m_I,(\sigma)}(\Sigma \times [0,T])$, for $1 \leq \sigma < \sigma_0$
and some $T>0$.
It has to be shown that the solution $V$ to (\ref{new_system_symbolic})
yields a solution to the original set of equations (\ref{original}).
The argument to show this is very similar to the one employed in
\cite{DisconziViscousFluidsNonlinearity, Lich_MHD_paper} (see also \cite{Lichnerowicz_MHD_book}),
thus we just
mention the general idea. Consider the incompressible Einstein-Navier-Stokes
system written in harmonic coordinates. Pichon \cite{PichonViscous} has shown that this
system can be solved for analytic data (his work treated only the case of
an equation of state that does not include entropy, but it is not
difficult to see that his procedure generalizes to the case of interest here).
By the way the Cauchy data for (\ref{new_system_symbolic}) is derived
out of the initial data for (\ref{original}), the analytic
solution to (\ref{original}) will satisfy the system (\ref{new_system_symbolic})
with $C_\alpha = F u_\alpha$ and $\Omega_{\alpha\beta} = \nabla_\alpha C_\beta - \nabla_\beta C_\alpha$.
For the case of initial data in Gevrey spaces, as in Theorem
\ref{main_theorem}, we approximate the initial data
by analytic Cauchy data, obtaining a sequence
$\{ (g_j, u_j, r_j, s_j )\}$
of analytic solutions to (\ref{original}), and a corresponding
sequence $\{V_j\}$
of analytic solutions
to (\ref{new_system_symbolic})
that converges to the solution $V$ obtained above. The estimates
on solutions derived by Leray and Ohya \cite{LerayOhyaNonlinear} assure that
$\{ ( g_j, u_j, s_j, r_j )\}$ also converges to a limit
$\{ ( g, u, s, r)\}$ that satisfies the incompressible Einstein-Navier-Stokes
system and belongs to the desired Gevrey class. It is well-known
that a solution to Einstein's equations in harmonic coordinates yields
a solution to the full system if and only if the constraint equations
are satisfied, which is the case by hypothesis\footnote{Here it is important to remind
the reader of the discussion of section \ref{restriction_vorticity}. In particular,
the solvability of the constraints under the further requirements imposed by (\ref{evolution_vorticity})
is at this point unknown.}. Finally, we
notice that $\pi^{\gamma\beta} \nabla^\alpha T_{\alpha\beta} = 0$ implies
\begin{gather}
0 = u^\alpha u^\gamma \nabla_\alpha u_\gamma = \frac{1}{2} u^\alpha \partial_\alpha (u^\gamma u_\gamma),
\nonumber
\end{gather}
and therefore $u$, being unitary at time zero, remains unitary.
The existence of a domain of dependence also follows from the results
of \cite{LerayOhyaNonlinear}. The domain of dependence of the solution
is given by the intersection of the interior of the cones determined
by the hyperbolic polynomials appearing in the product
(\ref{product_hyp_pol}). All these cones have a common interior,
namely, the interior of the light-cone
$\xi^\mu \xi_\mu = g^{\mu\nu}\xi_\mu \xi_\nu \geq 0$.
With the domain of dependence at hand, a standard gluing argument produces a
solution that is global in space and geometrically unique. This finishes
the proof of Theorem \ref{main_theorem}.
\vskip 0.2cm
\noindent \textbf{Acknowledgments.} We are grateful to the referees for reading the manuscript
in much detail and making several important comments.
\bibliographystyle{plain}
|
\section{Introduction}
\emph{Convex regression}~(CR) problem is
concerned with fitting a convex function to a finite number of observations. In particular, suppose that we are given $N$ observations $\{(\pmb{x}_\ell,\bar{y}_\ell)\}_{\ell=1}^N\subset\mathbb{R}^n\times\mathbb{R}$ such that
\begin{align}
\label{data}
\bar{y}_\ell=f_0(\pmb{x}_\ell)+\varepsilon_\ell,\quad \ell=1,\ldots,N,
\end{align}
where $f_0: \mathbb{R}^n \rightarrow \mathbb{R}$ is convex, $\varepsilon_\ell$ is a random variable with $E[\varepsilon_\ell]=0$ for all $\ell$. The objective is to estimate the convex function $f_0$ from the observed data points.
CR has many applications in various disciplines, such as statistics, economics, operations research, and electrical engineering. M. Mousavi~\cite{mousavi2013shape} employed
CR to estimate the value function under infinite-horizon discounted rewards for Markov chains, which naturally arises in various control problems. In economics field, CR is used for approximating consumers' concave utility functions from empirical data~~\cite{meyer1968consistent}.
Moreover, in queueing network context, for models where the expectation of performance measure is convex in model parameters -se
~\cite{chen2001fundamentals}, using Monte Carlo methods to compute the expectation give rise to CR problem~\cite{lim2012consistency}.
The most well-known method for CR is the least squares~(LS) problem,
\begin{align}
\label{infinite_dim}
\hat{f}_N=\mathop{\rm argmin}_{f\in\mathcal{C}} \sum\limits_{\ell=1}^{N} \big( f(\pmb{x}_\ell ) - \bar{y}_\ell \big)^2,
\end{align}
where $\mathcal{C}:=\{f:\mathbb{R}^n\rightarrow\mathbb{R} \hbox{ such that } f \hbox{ is convex}\}$. This infinite dimensional problem
is equivalent to a finite dimensional quadratic problem~(QP),
\begin{align}
\label{original}
& \min_{y_\ell\in\mathbb{R},~\pmb{\xi}_\ell\in\mathbb{R}^n} \sum\limits_{\ell=1}^{N} \big| y_\ell -\bar{y}_\ell \big|^2\\
\hbox{s.t.} \quad & {y}_{\ell_1} \geq y_{\ell_2} + {\xi_{\ell_2}}^{T} (\pmb{x}_{\ell_1} - \pmb{x}_{\ell_2}) \quad 1\leq \ell_1 \neq \ell_2\leq N. \nonumber
\end{align}
Indeed, let $\{(y_\ell^*,\xi_\ell^*)\}_{\ell=1}^N$ be an optimal solution to \eqref{original}, it is easy to show that when $N \geq n + 1$, $\{y_\ell^*\}_{\ell=1}^N$ is unique, $\hat{f}_N(\pmb{x}_\ell)=y_\ell^*$ and $\xi_\ell^*\in\partial\hat{f}_N(\pmb{x}_\ell)$ for all $\ell$, where $\partial$ denotes the subdifferential. Moreover, $\hat{f}_N\rightarrow f_0$ almost surely is shown in~\cite{lim2012consistency}; and the convergence rate is established in \cite{groeneboom2001estimation} for one-dimensional case, i.e. $n=1$. LS estimator has some significant advantages over many other estimators proposed in the literature for CR. First, LS estimator is a non-parametric regression method as discussed in \cite{Sen11}, which does not require any tuning parameters and avoids the issue of selecting an appropriate estimation structure. On the other hand, as discussed in~\cite{mousavi2013shape}, methods proposed by Hannah and Dunson~\cite{hannah2011approximate,hannah2013multivariate}, are semi-parametric and require adjusting several parameters before fitting a convex function. Second, LS estimator can be computed by solving the QP in \eqref{original}. Therefore, at least in theory, it can be solved very efficiently using interior point methods (IPM). However, a major drawback of LS estimator in practice is that the number of shape constraints in \eqref{original} is $\mathcal{O}(N^2)$. Consequently, the problem quickly becomes massive even for moderate number of observations: the complexity of each factorization step in IPM is $\mathcal{O}(N^3(n+1)^3)$, and the memory requirement of IPM is $\mathcal{O} \big( N^2(n+1)^2 \big)$ assuming Cholesky factors are stored - see \cite{Boyd04,Nocedal2006}.
In this paper, we develop a methodology for parallel computing the LS estimator on huge-scale CR problems. The proposed method carefully manages the memory usage through parallelization, and efficiently solves large-scale instances of \eqref{original}.
Indeed, by regularizing the objective in~\eqref{original}, we ensure the feasibility of primal iterates in the limit, and Lipchitz continuity of gradient of the dual function. These properties lead to the main result, Theorem~\ref{bound}, which provides error bounds on the distance between the LS estimator and the optimal solution to the regularized problem. In the rest of the paper, after examining the dual decomposition for large-scale CR instances, we briefly discuss a first-order augmented Lagrangian method for solving QP subproblems. Finally, we conclude with a number of numerical examples
\section{Methodology}
Assume that $\{\varepsilon_\ell\}_{\ell=1}^N$ is uniformly bounded by some $B_\varepsilon>0$, $f_0:\mathbb{R}^n\rightarrow\mathbb{R}\cup\{+\infty\}$ is a convex function, and $\{\pmb{x}_\ell\}_{\ell=1}^N$ is a set of independent and identically distributed (i.i.d.) random vectors in $\mathbb{R}^n$ having a common continuous distribution supported on the $n$-dimensional hypercube $\mathcal{H}:=[-B_x,B_x]^n\subset \mathbf{ri}\mathop{\bf dom}(f_0)$ for some $B_x>0$, where $\mathbf{ri}$ denotes the relative interior.
Consider \eqref{original} in the following compact form,
\begin{align}
\label{original_compact}
\min_{\pmb{y}\in\mathbb{R}^N,~\pmb{\xi}\in\mathbb{R}^{Nn}} \quad & \tfrac{1}{2} \left\| \pmb{y} - \bar{\pmb{y}} \right\| _2^2\\
\text{s.t. } \qquad & A_1~\pmb{y} + A_2~\pmb{\xi}\geq 0, \nonumber
\end{align}
where $A_1\in\mathbb{R}^{N(N-1)\times N}$ and $A_2\in\mathbb{R}^{N(N-1)\times Nn}$ are the matrices corresponding to the constraints in \eqref{original}. Let
\begin{align}
\label{tikhonov}
(\pmb{y}^*, \pmb{\xi}^* ) := \mathop{\rm argmin}\limits_{\pmb{y},~\pmb{\xi}} \left\{ \tfrac{1}{2} \big\| \pmb{y} \big\| _2^2 + \tfrac{1}{2} \big\| \pmb{\xi} \big\| _2^2:\ (\pmb{y}, \pmb{\xi} ) \in \chi^* \right\},
\end{align}
where $\chi^*$ denotes the set of optimal solutions to \eqref{original_compact}. Let $B_\xi:=\norm{\pmb{\xi}^*}_2$. Moreover, since \eqref{original_compact} is a convex QP, strong duality holds, and an optimal dual solution $\pmb{\theta}^*\in\mathbb{R}^{N(N-1)}$ exists.
Let $B_\theta>0$ such that $\norm{\pmb{\theta}^*}_\infty\leq B_\theta$ for some optimal dual. The complexity result of the proposed method will be provided in terms of constants $B_\xi$ and $B_\theta$.
\subsection{Separability}
To reduce \textit{curse of dimensionality} and develop a first-order parallel algorithm that can solve~\eqref{original_compact} for large $N$, we use dual decomposition to induce separability. To this aim, we partition $N$ observations into $K$ subsets. Let $\{\mathcal{C}_i\}_{i=1}^K$ denote the collection of indices such that $|\mathcal{C}_i|\geq n+1$ for all $i$. To simplify the notation, let $N=K \bar{N}$ for some $\bar{N}>n+1$, $$\mathcal{C}_i:= \big \{(i-1)\bar{N}+1,~(i-1)\bar{N}+2,\ldots,~i\bar{N} \big\}$$ for $1\leq i\leq K$. Throughout the paper,
$\pmb{y}_i\in\mathbb{R}^{\bar{N}}$ and $\pmb{\xi}_i\in\mathbb{R}^{\bar{N}n}$ denote the sub-vectors of $\pmb{y}\in\mathbb{R}^{N}$ and $\pmb{\xi}\in\mathbb{R}^{Nn}$ corresponding to indices in $\mathcal{C}_i$, respectively.
For every ordered pair $(\ell_1,\ell_2)$ such that $1\leq\ell_1\neq\ell_2\leq N$, there corresponds a constraint in \eqref{original} represented by a row in the matrices $A_1$ and $A_2$ of formulation \eqref{original_compact}. By dualizing all the constraints in \eqref{original} corresponding to $\ell_1\neq\ell_2$ such that they belong to different sets in the partition, i.e. $\ell_1\in\mathcal{C}_{i}$, $\ell_2\in\mathcal{C}_{j}$ and $i\neq j$, we form the partial Lagrangian, \vspace{-2mm}
\begin{align*}
\mathcal{L} \left(\pmb{y}, \pmb{\xi}, \pmb{\theta}\right) :=& \frac{1}{2}\sum\limits_{i=1}^K \norm{\pmb{y}_i - \bar{\pmb{y}}_i}_2^2\\
& - \sumsum\limits_{1\leq i \neq j\leq K} {\pmb{\theta}_{ij}}^{\mathsf{T}} \left( A_1^{ij}
\begin{bmatrix} \pmb{y}_i \\ \pmb{y}_j \end{bmatrix} +A_2^{ij} \begin{bmatrix} \pmb{\xi}_i \\ \pmb{\xi}_j \end{bmatrix} \right),
\end{align*}
which will lead to the following partial dual function
\begin{align}
\label{subgradient}
g(\pmb{\theta}):=&\min_{\pmb{y}\in\mathbb{R}^N,~\pmb{\xi}\in\mathbb{R}^{Nn} } \quad \mathcal{L} \left(\pmb{y}, \pmb{\xi}, \pmb{\theta}\right) \\
\text{s.t.} & \quad A_1^{ii}~\pmb{y}_i + A_2^{ii}~\pmb{\xi}_i \geq 0,\ i=1,\ldots,K, \nonumbe
\end{align}
where for each $i\in\{1,\ldots,K\}$, $A_1^{ii}$ and $A_2^{ii}$ are formed by rows of $A_1$ and $A_2$, respectively, corresponding to all $(\ell_1,\ell_2)$ such that $\ell_1\neq\ell_2$ and $\ell_1,\ell_2\in\mathcal{C}_i$; similarly, for each $(i,j)$ such that $1\leq i\neq j\leq K$, ${A_1}^{ij} $ and ${A_2}^{ij} $ contain the rows of $A_1$ and $A_2$, respectively, corresponding to $\{(\ell_1,\ell_2):\ \ell_1\in\mathcal{C}_i,~\ell_2\in\mathcal{C}_j\}$,
and $\pmb{\theta}_{ij}\in\mathbb{R}^{\bar{N}^2}$ denotes the associated dual variables.
$\pmb{\theta}$ denotes the vector formed by vertically concatenating $\pmb{\theta}_{ij}$ for $1\leq i\neq j\leq K$.
Note that partial Lagrangian $\mathcal{L}$ is separable and can be written as $\mathcal{L} \left(\pmb{y}, \pmb{\xi}, \pmb{\theta}\right)=\sum_{i=1}^K \mathcal{L}_i\left(\pmb{y}_i, \pmb{\xi}_i, \pmb{\theta}\right)$ for some very simple quadratic functions $\mathcal{L}_i$.
Thanks to the separability of $\mathcal{L}$, computing the partial dual function $g(\pmb{\theta})$, given in \eqref{subgradient}, is equivalent to solving $K$ quadratic subproblems of the form:
\begin{align}
\label{subgradient_split}
&\min_{\pmb{y}_i\in\mathbb{R}^{\bar{N}},~\pmb{\xi}_i\in\mathbb{R}^{\bar{N}n} } \mathcal{L}_i (\pmb{y}_i, \pmb{\xi}_i, \pmb{\theta}) \\
&\qquad \text{s.t.} \quad A_1^{ii}~\pmb{y}_i + A_2^{ii}~\pmb{\xi}_i \geq 0,\nonumber
\end{align}
for $1\leq i\leq K$.
Given the dual variables $\pmb{\theta}$, since all $K$ subproblems can be computed in parallel,
one can take advantage of the computing power of multi-core processors.
In the rest of the paper, we discuss
how to compute a solution to \eqref{original} via solving the dual problem: $\max\{g(\pmb{\theta}):\ \pmb{\theta}\geq 0\}$.
\subsection{Projected Subgradient Method for Dual}
One of the most well-known methods for solving the dual problem is the projected subgradient method.
Let $\pmb{\theta}^0_{ij}=\pmb{0}$ for all $i$, $j$ such that $i\neq j$. Given the $k$-th dual iterate $\pmb{\theta}^k$, $( \pmb{y}^k, \pmb{\xi}^k )$ denotes an optimal solution to the minimization problem in \eqref{subgradient} when $\pmb{\theta}$ is set to $\pmb{\theta}^k$,
and $\pmb{\theta}_{ii}^*$ denotes an optimal dual associated with constraints $A_1^{ii}~\pmb{y}_i + A_2^{ii}~\pmb{\xi}_i \geq 0$ in \eqref{subgradient}. The next dual iterate $\pmb{\theta}^{k+1}$ is computed as follows
\begin{align}
\pmb{\theta}_{ij}^{k+1} = \prod_{\mathcal{S}_{ij}} \left( \pmb{\theta}_{ij}^{k} - t_k\left( A_1^{ij} \begin{bmatrix} \pmb{y}_i^k \\ \pmb{y}_j^k \end{bmatrix} +A_2^{ij} \begin{bmatrix} \pmb{\xi}_i^k \\ \pmb{\xi}_j^k \end{bmatrix} \right) \right),
\end{align}
where $\Pi_{\mathcal{S}_{ij}}(.)$ denotes the Euclidean projection on to $$\mathcal{S}_{ij} = \Big\{ \pmb{\theta}_{ij} \geq \pmb{0}: {\pmb{\theta}_{ij}}^{\mathsf{T}} A_2^{ij} + \begin{bmatrix} {\pmb{\theta}_{ii}^*}^{\mathsf{T}} A_2^{ii} & {\pmb{\theta}_{jj}^*}^{\mathsf{T}} A_2^{jj} \end{bmatrix} =\pmb{0} \Big\}.$$
Since $\mathcal{L}$ is linear in $\pmb{\xi}$, $\mathop{\bf dom} g$ is non-trivial and is given by the Cartesian product of $\mathcal{S}_{ij}$'s.
The projected subgradient method is guaranteed to converge in function value with a careful selection of step size sequence $\{ t_k \}_{k=1}^{\infty}$, and it requires $\mathcal{O}( 1/ \epsilon^2)$ iterations to obtain an $\epsilon$-optimal solution~-see~\cite{Nesterov04_1B}.
However, due to lack of strong convexity of the objective function in \eqref{original_compact} (not in $\pmb{\xi}$), even if the dual variables converge to an optimal dual solution, the primal feasibility cannot be guaranteed in the limit.
\subsection{Tikhonov Regularization Approach}
In order to ensure feasibility in the limit, which cannot be guaranteed by the subgradient method discussed above, we employ Tikhonov regularization, of which convergence properties were investigated in \cite{engl1989convergence}. Given $\gamma>0$, consider
\begin{align}
\label{regularize}
(\pmb{y}(\gamma),\pmb{\xi}(\gamma))=\mathop{\rm argmin}_{\pmb{y},~\pmb{\xi}} \quad & \tfrac{1}{2} \left\| \pmb{y} - \bar{\pmb{y}} \right\| _2^2 + \tfrac{\gamma}{2} \left\| \pmb{\xi} \right\| _2^2 \\
\text{s.t. } \quad & A_1\pmb{y} + A_2{\pmb{\xi}}\geq 0.\nonumber
\end{align}
\vspace{-5mm}
\noindent As $\gamma$ decreases to zero from above, the minimizer $(\pmb{y}(\gamma), \pmb{\xi}(\gamma))$ converges to $(\pmb{y}^*, \pmb{\xi}^* )$ defined in \eqref{tikhonov}.
\begin{lem}
\label{continuity}
The minimizer of \eqref{regularize}, $ \pmb{y}(\gamma)$ as a function of regularization parameter $\gamma$, is H\"{o}lder continuous on $[0,\infty)$,\vspace{-1mm}
\begin{equation}
\label{holder_cond}
\norm{\pmb{y}(\gamma) - \pmb{y}^*}_2 \leq B_\xi \sqrt{\gamma}.
\end{equation}
\end{lem}
\vspace{2mm}
\proof
Let $\left(\pmb{y}(\gamma),\pmb{\xi}(\gamma)\right)$ be the optimal solution to \eqref{regularize} and $\left(\pmb{y}^*,\pmb{\xi}^*\right)$ be defined as in \eqref{tikhonov}. From the first-order optimality conditions of \eqref{regularize} and \eqref{original_compact}, we have
\begin{equation}
\label{optimality condition1}
\begin{pmatrix} \pmb{y}(\gamma) - \bar{\pmb{y}} \\ \gamma~\pmb{\xi}(\gamma) \end{pmatrix} ^{\mathsf{T}} \begin{pmatrix} \pmb{y}^* - \pmb{y}(\gamma) \\ \pmb{\xi}^* - \pmb{\xi}(\gamma) \end{pmatrix} \geq 0,
\end{equation}
\begin{equation}
\label{optimality condition2}
\begin{pmatrix} \pmb{y}^* - \bar{\pmb{y}} \\ 0 \end{pmatrix} ^{\mathsf{T}} \begin{pmatrix} \pmb{y}(\gamma) - \pmb{y}^* \\ \pmb{\xi}(\gamma) - \pmb{\xi}^* \end{pmatrix} \geq 0.
\end{equation}
Note that both $(\pmb{y}(\gamma),\pmb{\xi}(\gamma))$ and $(\pmb{y}^*,\pmb{\xi}^*)$ are feasible to \eqref{original_compact} and \eqref{regularize}. This implies $\norm{\pmb{\xi}(\gamma)}_2\leq \norm{\pmb{\xi}^*}_2$. Summing up \eqref{optimality condition1} and \eqref{optimality condition2}, and using $B_\xi=\norm{\pmb{\xi}^*}_2$, it follows that
\begin{align*}
\left\| \pmb{y}(\gamma) - \pmb{y}^* \right\|_2^2 \leq \gamma \pmb{\xi}(\gamma)^{\mathsf{T}} \big( \pmb{\xi}^* - \pmb{\xi}(\gamma) \big)\leq\gamma B^2_\xi.
\end{align*}
\vspace{-1mm}
\endproof
Since the objective function in \eqref{regularize} is strongly convex in both $\pmb{y}$ and $\pmb{\xi}$, Danskin's theorem (see~\cite{Bertsekas99}) implies that the Lagrangian dual function of \eqref{regularize} is differentiable; therefore, one can use gradient type methods to solve the corresponding dual problem. Moreover, strong convexity ensures that, one can solve the primal problem by solving the dual problem. Indeed, let $\pmb{\theta}(\gamma)$ be an optimal solution to the dual problem of \eqref{regularize}, we can recover $(\pmb{y}(\gamma),\pmb{\xi}(\gamma))$ by computing the primal minimizers in \eqref{subgradient} when the dual is set to $\pmb{\theta}(\gamma)$. The discussion above shows that achieving primal feasibility is not an issue provided that we can solve the dual of \eqref{regularize}. This motivates the next section, where we briefly state a first-order algorithm that can efficiently solve the dual of \eqref{regularize}.
\subsection{Accelerated Proximal Gradient~(APG) Algorithm}
Let $\rho:\mathbb{R}^d\rightarrow\mathbb{R}$ be a concave function such that $\nabla \rho$ is Lipschitz continuous on $\mathbb{R}^d$ with constant $L$, and $\mathcal{Q}\subset\mathbb{R}^d$ be a compact convex set. The APG algorithm~\cite{Beck09,Tseng08} displayed in Figure~\ref{fig:apg} is based on Nesterov's accelerated gradient method~\cite{Nesterov04_1B,Nesterov05_1J} and solves $\rho^*=\max\{\rho(\eta):\ \eta\in\mathcal{Q}\}$. Corollary~3 in~\cite{Tseng08}, and Theorem 4.4 in~\cite{Beck09} show that for all $\ell\geq 1$ the error bound is given by $$0\leq\rho^*-\rho(\eta_\ell)\leq \frac{2L}{(\ell+1)^2} \norm{\eta_0-\eta^*}_2^2,$$ where $\eta_0$ is the initial APG iterate and $\eta^*\in\mathop{\rm argmin}_{\eta\in\mathcal{Q}} \rho(\eta)$. Hence, using APG one can compute an $\delta$-optimal solution within at most $\mathcal{O}(\sqrt{L/\delta})$ APG iterations.
\vspace{-5mm}
\begin{figure}[thpb]
\begin{framed}
{\small
\textbf{Algorithm APG} \big( $\eta_0$ \big) \\
Iteration 0: Take $ \eta_0^{(1)}=\eta_1^{(2)}=\eta_0, t_1=1 $\\
Iteration $\ell$: ($\ell \geq 1$) Compute
\begin{enumerate}
\item ${\eta_\ell}^{(1)}= \Pi_{\mathcal{Q}}\left( \eta_\ell^{(2)}+\frac{\nabla \rho(\eta_\ell^{(2)})}{L} \right)$
\item $t_{\ell+1}= ( 1+ \sqrt{ 1+ 4t_\ell^2} )/2$
\item $\eta_{\ell+1}^{(2)}= \eta_\ell^{(1)}+ \frac{t_\ell -1}{t_{\ell+1}} \left(\eta_\ell^{(1)}-\eta_{\ell-1}^{(1)}\right)$
\end{enumerate}
}
\end{framed}
\vspace{-2mm}
\caption{Accelerated Proximal Gradient Algorithm}
\label{fig:apg}
\vspace{-3mm}
\end{figure}
In this paper, we will use APG algorithm on a slightly different but equivalent problem to \eqref{regularize}.
Let $A_3$ and $A_4$ denote the matrices formed by vertically concatenating $A_1^{ij}$ and $A_2^{ij}$, respectively, for $1\leq i\neq j\leq K$; and define
\begin{align}
\label{eq:C}
C=\begin{bmatrix} A_3 & A_4 \\ I & 0 \end{bmatrix},
\end{align}
where $I\in\mathbb{R}^{N\times N}$ identity matrix. For notational convenience, let $ \pmb{\eta}^{\mathsf{T}}=\begin{bmatrix} \pmb{y}^{\mathsf{T}} & \pmb{\xi}^{\mathsf{T}} \end{bmatrix}$, and consider
\begin{align}
\label{omega_primal}
\min_{\pmb{\eta}\in Q_1} \tfrac{1}{2} \norm{\pmb{y} - \bar{\pmb{y}}}_2^2 + \tfrac{\gamma}{2} \left\| \pmb{\xi} \right\| _2^2, \quad \text{s.t.} \quad C~\pmb{\eta} \geq 0,
\end{align}
where $Q_1 := \big\{ (\pmb{y},\pmb{\xi}): A_1^{ii}~\pmb{y}_i +A_2^{ii}~\pmb{\xi}_i \geq 0, 1\leq i \leq K\big\}$. Note that \eqref{omega_primal} is different from \eqref{regularize} only in constraints $\pmb{y}\geq \mathbf{0}$. Via possibly shifting all the observations $\{\pmb{y}_\ell\}_{\ell=1}^N$ up by a sufficiently large quantity, we can assume without loss of generality that $\pmb{y}^*(\gamma)\geq \mathbf{0}$ under bounded error assumption, i.e. $|\varepsilon_\ell|\leq B_\varepsilon$ for all $\ell$. Therefore, \eqref{regularize} and \eqref{omega_primal} are indeed equivalent problems. Consider the dual problem of \eqref{omega_primal},
\begin{equation}
\label{lambda_dual}
\max_{\pmb{\theta}} g_\gamma(\pmb{\theta}) \quad \hbox{ s.t. } \quad \pmb{\theta} \in Q_2,
\end{equation}
where $Q_2 := \big\{ \pmb{\theta} : \norm{\pmb{\theta}}_2 \leq B_\theta,~\pmb{\theta} \geq 0 \big\}$, and
\begin{align}
\label{eq:g_gamma}
g_\gamma(\pmb{\theta}) = \min\limits_{(\pmb{y}, \pmb{\xi}) \in Q_1 } \left\{ \tfrac{1}{2} \big\| \pmb{y} - \bar{\pmb{y}} \big\| _2^2 + \tfrac{\gamma}{2} \left\| \pmb{\xi} \right\| _2^2 - \pmb{\theta}^{\mathsf{T}} C\pmb{\eta} \right\}.
\end{align}
Let $\pmb{\eta}(\pmb{\theta})$ be the minimizer in \eqref{eq:g_gamma}. Theorem 7.1 in~\cite{nesterov2005excessive} and Danskin's theorem imply that
\begin{align}
\label{dual gradient}
\nabla g_\gamma(\pmb{\theta}) = - C \pmb{\eta}(\pmb{\theta})
\end{align}
is Lipschitz continuous with constant
\begin{align}
\label{eq:lischitz}
L_g = \tfrac{1}{\gamma}~\sigma_{\max}^2(C).
\end{align}
Parallel~APG algorithm~(P-APG), displayed in Fig.~\ref{fig:papg}, is the customized version of APG algorithm in Fig.~\ref{fig:apg} to solve \eqref{lambda_dual}. Note that at each iteration computation in Step 1) can be done in parallel using $K$ processors, each solving a smaller QP.
\vspace{-3mm}
\begin{figure}[thpb]
\begin{framed}
{\small
\textbf{Algorithm P-APG} \big( $\gamma$ \big) \\
Iteration 0: Take $ \pmb{\theta}_0^{(1)} = \pmb{\theta}_1^{(2)} = \pmb{0}, t_1=1 $\\
Iteration $\ell$: ($\ell \geq 1$) Compute
\begin{enumerate}
\item $\pmb{\eta}_\ell = \mathop{\rm argmin}\limits_{(\pmb{y}, \pmb{\xi}) \in Q_1 }\left\{ \tfrac{1}{2} \big\| \pmb{y} - \bar{\pmb{y}} \big\| _2^2 + \tfrac{\gamma}{2} \left\| \pmb{\xi} \right\| _2^2 - \left(\pmb{\theta}_\ell^{(2)}\right)^{\mathsf{T}} C\pmb{\eta} \right\}$
\item ${\pmb{\theta}_\ell}^{(1)}= \Pi_{Q_2}\left( \pmb{\theta}_\ell^{(2)}-\frac{1}{L_g}C\pmb{\eta}_\ell \right)$
\item $t_{\ell+1}= ( 1+ \sqrt{ 1+ 4t_\ell^2} )/2$
\item $\pmb{\theta}_{\ell+1}^{(2)}= \pmb{\theta}_\ell^{(1)}+ \frac{t_\ell -1}{t_{\ell+1}} \left(\pmb{\theta}_\ell^{(1)}-\pmb{\theta}_{\ell-1}^{(1)}\right)$
\end{enumerate}
}
\end{framed}
\vspace{-2mm}
\caption{ Parallel APG Algorithm}
\label{fig:papg}
\vspace{-4mm}
\end{figure}
Note that the iteration complexity of gradient ascent method on \eqref{lambda_dual} is $\mathcal{O}(L_g/\delta)=\mathcal{O}(B_\theta^2(\gamma\delta)^{-1})$. On the other hand, P-APG in Fig.~\ref{fig:papg} can compute a $\delta$-optimal solution to \eqref{lambda_dual} within $\mathcal{O}(\sqrt{L_g/\delta})$ iterations. More precisely, \eqref{eq:lischitz} implies $\mathcal{O}(B_\theta (\gamma\delta)^{-1/2})$ complexity for P-APG on \eqref{lambda_dual}.
Let $\pmb{\theta}_\delta$ be a $\delta$-optimal solution to \eqref{lambda_dual}, and $(\pmb{y}_\delta, \pmb{\xi}_\delta)$ be the optimal solution to the minimization problem in \eqref{subgradient} when $\pmb{\theta}$ is set to $\pmb{\theta}_\delta$. In Theorem~\ref{bound}, which is the main result of this paper, we establish an error bounds on suboptimality $\norm{\pmb{y}_\delta-\pmb{y}^*}_2$, and on infeasibility $\norm{ (A_1 \pmb{y}_{\delta} + A_2 \pmb{\xi}_{\delta} )_{-} }_2$, where $(\pmb{x})_{-}:=\max\{ -\pmb{x},\pmb{0}\}$.
\begin{thm}
\label{bound}
Let $\left(\pmb{y}(\gamma), \pmb{\xi}(\gamma)\right)$ and $\pmb{\theta}^*$ denote the optimal solutions to \eqref{regularize} and \eqref{lambda_dual}, respectively. Let $\pmb{\theta}_{\delta}$ be a $\delta$-optimal solution to \eqref{lambda_dual}, and $\left(\pmb{y}_\delta, \pmb{\xi}_\delta\right)$ be the minimizer in
\eqref{eq:g_gamma} when $\pmb{\theta}$ is set to $\pmb{\theta}_\delta$.
For all $\delta > 0$, the following bounds hold:
\vspace{-2mm}
\begin{flalign}
\label{bound ineq}
&\norm{ \pmb{y}_{\delta} - \pmb{y}^*}_2 \leq B_\xi \sqrt{\gamma} + \sqrt{\tfrac{2\delta}{\gamma}} \sigma_{\max}(C),\\
&\norm{ (A_1 \pmb{y}_{\delta} + A_2 \pmb{\xi}_{\delta} )_{-} }_2 \leq \sqrt{\tfrac{2 \delta }{\gamma}}\sigma_{\max}(C). \label{bound vio}
\end{flalign}
\vspace{-5mm}
\end{thm}
\proof
Since $g_\gamma$ is Lipschitz continuous with constant $L_g$ given in \eqref{eq:lischitz}, we have
\vspace{-3mm}
\begin{align*}
\left\| \nabla g_\gamma( \pmb{\theta}_1) - \nabla g_\gamma( \pmb{\theta}_2) \right\|_2 \leq \frac{\sigma_{\max}^2(C)}{\gamma} \left\| \pmb{\theta}_1 - \pmb{\theta}_2\right\|_2.
\end{align*}
\vspace{-4mm}
\noindent Moreover, first order optimality conditions for \eqref{lambda_dual} imply
\vspace{-1mm}
\begin{align}
\label{lambda first order}
- \fprod{ \nabla g( \pmb{\theta}^*), \quad \pmb{\theta} - \pmb{\theta}^*} \geq 0, \qquad \forall \pmb{\theta}\in Q_2.
\end{align}
\vspace{-5mm}
\noindent From (2.1.7) in \cite{Nesterov04_1B}, it follows that
\begin{align*}
- \nabla g_\gamma( \pmb{\theta}^*)^{\mathsf{T}} ( \pmb{\theta}_{\delta} - \pmb{\theta}^* ) &+ \frac{\gamma}{2\sigma_{\max}^2(C)} \left\| \nabla g_\gamma( \pmb{\theta}_{\delta}) - \nabla g_\gamma( \pmb{\theta}^*) \right\|_2^2 \\
& \leq - g_\gamma( \pmb{\theta}_{\delta}) + g( \pmb{\theta}^*) \leq \delta.
\end{align*}
Using \eqref{eq:C}, \eqref{dual gradient} and \eqref{lambda first order}, we have
\begin{align}
\left\| \begin{bmatrix} A_3 (\pmb{y}(\gamma) - \pmb{y}_\delta)+A_4(\pmb{\xi}(\gamma) - \pmb{\xi}_\delta) \\ \pmb{y}(\gamma) - \pmb{y}_\delta \end{bmatrix} \right\|_2^2 &\leq \tfrac{2\delta}{\gamma}\sigma_{\max}^2(C). \label{eq:mixed_bound}
\end{align}
Hence, together with \eqref{holder_cond}, it implies \eqref{bound ineq}. Moreover, since $\norm{\pmb{x}-\pmb{y}}_2\geq\norm{(\pmb{x})_{-}-(\pmb{y})_{-}}_2$ for any $\pmb{x}$ and $\pmb{y}$, we also have
\begin{align*}
\norm{(A_3\pmb{y}_\delta+A_4\pmb{\xi}_\delta)_{-}-(A_3\pmb{y}(\gamma)+A_4\pmb{\xi}(\gamma))_{-}}_2 \leq \sqrt{\tfrac{2 \delta }{\gamma}} \sigma_{\max}(C).
\end{align*}
Since $(\pmb{y}(\gamma),\pmb{\xi}(\gamma))$ is feasible to \eqref{regularize}, and $(\pmb{y}_\delta,\pmb{\xi}_\delta)\in Q_1$, above inequality implies \eqref{bound vio}.
\endproof
Next, we prove an important technical lemma that will be used later in Theorem~\ref{thm:xi-bound} to show that $\norm{\pmb{\xi}_\delta-\pmb{\xi}^*}_2$ is small.
\begin{lem}\label{LIC}
Assuming that $ \{\pmb{x}_i \}_{i=1}^N $ are uniformly sampled at random from set $\phi=\{ \pmb{x} \in \mathbb{R}^n : \| \pmb{x}\|_{\infty} \leq B_x \}$, the matrix $A_4$ in \eqref{eq:C} has linearly independent columns (LIC).
\end{lem}
\proof
Remember $A_4\in\mathbb{R}^{N(N-\bar{N})\times Nn}$ denotes the matrix formed by vertically concatenating all $A_2^{ij}$ for $1\leq i\neq j\leq K$. Note that rows of $[A_1^{ij}~A_2^{ij}]$ correspond to constraints $y_{\ell_1} - y_{\ell_2} + \pmb{\xi}_{\ell_2}^{\mathsf{T}} (\pmb{x}_{\ell_2} - \pmb{x}_{\ell_1}) \geq 0$, where $\ell_1 \in \mathcal{C}_i$, $\ell_2 \in \mathcal{C}_{j}$ and $i\neq j$. For the sake of simplifying the discussion below, without loss of generality, we fix $i=2$ and $j=1$, and focus on the structure of $A_2^{21}$.
Let $\hat{A}_2^{21}$ denote the submatrix of $A_2^{21}$ formed by selecting the rows corresponding to $(\ell_1,\ell_2)\in\mathcal{C}_2\times\mathcal{C}_1$ such that $\ell_1=\bar{N}+1$ and $\ell_2 \in \mathcal{C}_{1}$. Hence, we have
\begin{align}
\label{A_4}
\hat{A}_2^{21} = \begin{bmatrix} \mathbf{X} & \mathbf{0}\end{bmatrix}
\end{align}
where $\mathbf{0}$ is the matrix of zeros and $\mathbf{X}\in\mathbb{R}^{\bar{N}\times\bar{N}n}$ such that
\begin{align}
\label{X}
\mathbf{X}= \begin{bmatrix}
& {\pmb{\bar{x}}_1}^{\mathsf{T}} &\pmb{0}^{\mathsf{T}} &\pmb{0}^{\mathsf{T}} & \ldots & \pmb{0}^{\mathsf{T}} \cr
& \pmb{0}^{\mathsf{T}}& {\pmb{\bar{x}}_2}^{\mathsf{T}} &\pmb{0}^{\mathsf{T}} & \ldots & \pmb{0}^{\mathsf{T}} \cr
& \pmb{0}^{\mathsf{T}} & \pmb{0}^{\mathsf{T}} & {\pmb{\bar{x}}_3}^{\mathsf{T}} &\ldots & \pmb{0}^{\mathsf{T}} \cr
& \vdots & \vdots & \vdots &\ddots & \vdots \cr
& \pmb{0}^{\mathsf{T}} & \ldots & \pmb{0}^{\mathsf{T}} &\ldots & {\pmb{\bar{x}}_{\bar{N}}}^{\mathsf{T}} \cr
\end{bmatrix}
\end{align}
and $\pmb{\bar{x}}_\ell := \pmb{x}_{\ell} - \pmb{x}_{\bar{N}+1}$ for $1\leq\ell\leq\bar{N}$.
Fix $1\leq j\leq K$. Note that for each $\ell\in\mathcal{C}_j$, there corresponds $n$ columns in $A_4$; and
the zero structure in \eqref{A_4} implies that each column of $A_4$ corresponding to $\mathcal{C}_j$ is linearly independent with $\bar{N}n$ columns in $A_4$ corresponding to $\mathcal{C}_k$ with probability 1~(w.p.~1) for all $k\neq j$. Moreover, when we focus on \eqref{X}, we also see that any one of the $n$ columns in $A_4$ corresponding to $\bar{\ell}\in\mathcal{C}_j$ is also linearly independent with $n$ columns in $A_4$ corresponding to $\ell\in\mathcal{C}_j$ with probability 1 for all $\ell\neq \bar{\ell}$. Therefore, to show that $A_4$ has linearly independent columns, it is sufficient to show that for any given $1\leq j\leq K$ and $\ell\in\mathcal{C}_j$, the corresponding $n$ columns of $A_4$ are linearly independent w.p.~1.
Let $D\in\mathbb{R}^{N(N-\bar{N})\times n}$ be the submatrix of $A_4\in\mathbb{R}^{N(N-\bar{N})\times Nn}$ corresponding to columns $\bar{\ell}\in\mathcal{C}_j$ for some $1\leq j\leq K$; and $\pmb{d}_{\ell_1 \ell_2}^{\mathsf{T}}$ denote the row of $D$ corresponding to
$(\ell_1, \ell_2)$ such that $\ell_1$ and $\ell_2$ belong to different sets in the partition. Clearly,
\vspace{-1mm}
\begin{equation}
\pmb{d}_{\ell_1 \ell_2}^{\mathsf{T}}=
\left\{
\begin{array}{ll}
\big(\pmb{x}_{\bar{\ell}} - \pmb{x}_{\ell_1} \big)^{\mathsf{T}}, & \hbox{if $\ell_1\not\in\mathcal{C}_j$ and $\ell_2=\bar{\ell}$;} \\
\pmb{0}^{\mathsf{T}}, & \hbox{o.w.}
\end{array}
\right.
\end{equation}
Without loss of generality, we fix $j>1$ and consider $\bar{D} \in \mathbb{R}^{\bar{N} \times n}$ which denotes the submatrix of $D$ corresponding to the rows $\pmb{d}_{\ell_1 \ell_2}^{\mathsf{T}}$ such that $\ell_1 \in \mathcal{C}_1$ and $\ell_2=\bar{\ell}\in \mathcal{C}_j$. The following discussion is true for any $C_k$ such that $k\neq j$, but setting $k=1$ simplifies the notation in $\bar{D}$.
\[
\bar{D} = \begin{pmatrix} \pmb{x}_{\bar{\ell}}^{\mathsf{T}} - \pmb{x}_1^{\mathsf{T}} \\
\vdots \\
\pmb{x}_{\bar{\ell}}^{\mathsf{T}} - \pmb{x}_\ell^{\mathsf{T}}\\
\vdots \\
\pmb{x}_{\bar{\ell}}^{\mathsf{T}} - \pmb{x}_{\bar{N}}^{\mathsf{T}}
\end{pmatrix}
\]
It suffices to show that $\bar{D}$ has LIC.
Since $\bar{N}\geq n+1$ and $\{\pmb{x}_\ell\}_{\ell=1}^N$ is a set of i.i.d. random vectors in $\mathbb{R}^n$ having a common \emph{continuous} distribution, it can be shown that there exists $n$ linearly independent rows of $\bar{D}$ w.p.~1. Thus,
$A_4$ has LIC.
\endproof
\begin{thm}
\label{thm:xi-bound}
There exists $K_1,K_2>0$ such that \vspace{-1mm}
$$\norm{\pmb{\xi}_\delta-\pmb{\xi}^*}_2\leq K_1\sqrt{\gamma}+K_2\sqrt{\tfrac{\delta}{\gamma}}.$$
\end{thm}
\proof
Since $\pmb{y}^*$ is the unique optimal solution to \eqref{original_compact}, \eqref{tikhonov} implies that $\pmb{\xi}^*=\mathop{\rm argmin}\{\norm{\pmb{\xi}}_2:\ A_1\pmb{y}^*+A_2\pmb{\xi}\geq \pmb{0}\}$. Similarly, \eqref{regularize} implies that $\pmb{\xi}(\gamma)=\mathop{\rm argmin}\{\norm{\pmb{\xi}}_2:\ A_1\pmb{y}(\gamma)+A_2\pmb{\xi}\geq \pmb{0}\}$. Hence, for $\pmb{h}(\gamma):=A_1(\pmb{y}^*-\pmb{y}(\gamma))$,
\begin{equation}
\pmb{\xi}(\gamma)=\mathop{\rm argmin}\{\norm{\pmb{\xi}}_2:\ A_1\pmb{y}^*+A_2\pmb{\xi}\geq \pmb{h}(\gamma)\}.
\end{equation}
Sensitivity of projection onto parametric polyhedral sets was studied in~\cite{Yen95_1J}. Using Theorem~2.1 in~\cite{Yen95_1J} and \eqref{holder_cond}, we have
\begin{equation}
\norm{\pmb{\xi}(\gamma)-\pmb{\xi}^*}_2\leq K\norm{\pmb{h}(\gamma)}_2\leq K\sigma_{\max}(A_1)B_\xi\sqrt{\gamma},
\end{equation}
for some $K>0$. Moreover, \eqref{eq:mixed_bound} and Lemma~\ref{LIC} imply that
\begin{equation*}
\norm{\pmb{\xi}_\delta-\pmb{\xi}(\gamma)}_2\leq\frac{\sqrt{\tfrac{2\delta}{\gamma}}\sigma_{\max}(C)+\sigma_{\max}(A_3)\norm{\pmb{y}(\gamma)-\pmb{y}_\delta}_2}{\sigma_{\min}(A_4)}
\end{equation*}
Hence, $\norm{\pmb{\xi}_\delta-\pmb{\xi}(\gamma)}_2\leq\frac{(\sigma_{\max}(A_3)+1)\sigma_{\max}(C)}{\sigma_{\min}(A_4)}\sqrt{\tfrac{2\delta}{\gamma}}$.
\endproof
\subsection{ALCC - An Augmented Lagrangian Method}
\label{sec:alcc}
Now, we first briefly state a first-order algorithm to directly solve \eqref{regularize}. Let $\bar{B}_y$ and $\bar{B}_\xi$ be given such that $\pmb{y}(\gamma)\in\mathcal{Q}_y:=\{\pmb{y}:\norm{\pmb{y}-\bar{\pmb{y}}}_2\leq\bar{B}_y\}$, and $\pmb{\xi}(\gamma)\in\mathcal{Q}_\xi:=\{\pmb{\xi}:\norm{\pmb{\xi}}_2\leq\bar{B}_\xi\}$. Such $\bar{B}_y$ and $\bar{B}_\xi$ can be found easily, if we are given a feasible solution $(\hat{\pmb{y}},\hat{\pmb{\xi}})$, i.e. $A_1\hat{\pmb{y}}+A_2\hat{\pmb{\xi}}\geq\pmb{0}$. Indeed, selecting $\bar{B}_y=\bar{B}$ and $\bar{B}_\xi=B/\sqrt{\gamma}$ works, where $\bar{B}:=\sqrt{\norm{\hat{\pmb{y}}-\bar{\pmb{y}}}_2^2+\gamma\norm{\hat{\pmb{\xi}}}_2^2}$.
ALCC~\cite{Aybat13} computes a solution to \eqref{regularize} by inexactly solving a sequence of subproblems:
\begin{align}
\label{eq:subproblem}
P_k^*&:=\min\{ P_k(\pmb{y}, \pmb{\xi}):\ \pmb{y}\in\mathcal{Q}_y, \pmb{\xi}\in\mathcal{Q}_\xi\},\\
P_k(\pmb{y}, \pmb{\xi})&:= \tfrac{1}{2\mu_k} \left\| \pmb{y} - \bar{\pmb{y}} \right\| _2^2 + \tfrac{\gamma}{2\mu_k}\| \pmb{{\boldsymbol{\xi}}} \|_2^2 + h_k(\pmb{y}, \pmb{\xi}), \nonumber
\end{align}
where $h_k(\pmb{y}, \pmb{\xi}): = \frac{1}{2} \norm{\left( A_1\pmb{y} + A_2 \pmb{\xi} - \pmb{\theta}_k \right)_{-}}_2^2$,
and $\{\pmb{\theta}_k\}$ sequence is defined in Fig.~\ref{fig:alcc}.
For $k \geq 1$, $h_k( \pmb{y}, \pmb{\xi})$ is convex in $\pmb{y}$ and $\pmb{\xi}$ -see Lemma 3.1 in \cite{Aybat13}. Moreover,
\begin{align*}
\nabla_{\pmb{y}} h_k( \pmb{y}, \pmb{\xi}) &= -{A_1}^{\mathsf{T}} \left( A_1\pmb{y} + A_2 \pmb{\xi} - \pmb{\theta}_k \right)_{-}, \\
\nabla_{\pmb{\xi}} h_k( \pmb{y}, \pmb{\xi}) &= -{A_2}^{\mathsf{T}} \left( A_1\pmb{y} + A_2 \pmb{\xi} - \pmb{\theta}_k \right)_{-}.
\end{align*}
In addition, $\nabla_{\pmb{y}} h_k( \pmb{y}, \pmb{\xi})$ is Lipschitz continuous in $\pmb{y}$ for all fixed $\pmb{\xi}$ with constant $\sigma_{\max}^2(A_1)$, and $\nabla_{\pmb{\xi}} h_k( \pmb{y}, \pmb{\xi} )$ is Lipschitz continuous in $\pmb{\xi}$ for all fixed $\pmb{y}$ with constant $\sigma_{\max}^2(A_2)$. Hence, $\nabla_{\pmb{y}} P_k( \pmb{y}, \pmb{\xi})$ is Lipschitz continuous in $\pmb{y}$ for all fixed $\pmb{\xi}$ with constant $L_k^y:=\frac{1}{\mu_k}+\sigma_{\max}^2(A_1)$, and $\nabla_{\pmb{\xi}} P_k( \pmb{y}, \pmb{\xi} )$ is Lipschitz continuous in $\pmb{\xi}$ for all fixed $\pmb{y}$ with constant $L_k^\xi:=\frac{\gamma}{\mu_k}+\sigma_{\max}^2(A_2)$.
For given $c>1$ and $\kappa>0$, it is shown in~\cite{Aybat13} that the ALCC algorithm, displayed in Fig.~\ref{fig:alcc}, can compute an $\epsilon$-optimal and $\epsilon$-feasible solution to \eqref{original_compact} within $\mathcal{O}(\log(\epsilon^{-1}))$ ALCC iterations that require at most $\mathcal{O}(\epsilon^{-1}\log(\epsilon^{-1}))$ MAPG iterations. The bottleneck step at each MAPG iteration is the matrix-vector multiplication with $A_1 \in \mathbb{R}^{N^2-N \times N}$, $A_2 \in \mathbb{R}^{N^2-N \times Nn}$, ${A_1}^{\mathsf{T}}$ and ${A_2}^{\mathsf{T}}$. Due to specific structures of $A_1$ and $A_2$, without forming $A_1$ and $A_2$ explicitly, we can compute $A_1y$ and ${A_1}^{\mathsf{T}} z$ with $\mathcal{O}(N^2 - N)$ complexity for all $y$ and $z$; $A_2 \xi$ and ${A_2}^{\mathsf{T}} \omega$ with $\mathcal{O} \big( n(N^2 - N) \big)$ for all $\xi$ and $\omega$. Indeed, neither $A_1$ nor $A_2$ is stored in the memory, storing only $\{\pmb{x}_\ell\}_{\ell=1}^N$ is sufficient to be able to compute these matrix-vector multiplications.
\begin{figure}[h!]
\begin{framed}
{\small
\textbf{Algorithm ALCC} ( $\pmb{y}_0, \pmb{\xi}_0, \mu_1, \tau_1, \alpha^y_1, \alpha^\xi_1$ ) \\
Iteration 0: Take $\pmb{\theta}^0=\pmb{0}, k=1$\\
Iteration k: ($k \geq 1$)
\begin{enumerate}
\item $L^y_k=\frac{1}{\mu_k}+\sigma_{\max}^2(A_1)$, $L^\xi_k=\frac{\gamma}{\mu_k}+\sigma_{\max}^2(A_2)$
\item $\ell^{\max}_k= 4\sqrt{\frac{L^y_k \bar{B}_y^2+L^\xi_k\bar{B}_\xi^2}{\tau_k}}$
\item $ (\pmb{y}_k, \pmb{\xi}_k) = \hbox{\textbf{MAPG}} \big( P_k, L_k^y, L_k^\xi, \pmb{y}_{k-1}, \pmb{\xi}_{k-1}, \alpha^y_k, \alpha^\xi_k, \ell^{\max}_k \big)$
\item $\pmb{\theta}_{k+1} = \frac{\mu_k}{\mu_{k+1}} \big( A_1\pmb{y}_k + A_2 \pmb{\xi}_k - \pmb{\theta}_k \big)_{-}$
\item $\mu_{k+1}= c~\mu_k,\ \tau_{k+1}=\tau_k/\left(c~k^{1+\kappa}\right)^2$
\item $\alpha^y_{k+1}=\alpha^y_{k}/\left(c~k^{1+\kappa}\right)^2,\ \alpha^\xi_{k+1}=\alpha^\xi_{k}/\left(c~k^{1+\kappa}\right)^2$
\end{enumerate}
}
\end{framed}
\vspace{-2mm}
\caption{Augmented Lagrangian Algorithm ALCC}
\label{fig:alcc}
\vspace{-2mm}
\end{figure}
Note that at each iteration of ALCC in Step 2) MAPG algorithm is called to inexactly solve \eqref{eq:subproblem}. Instead of MAPG, one can also use APG in Fig.~\ref{fig:apg} to inexactly solve \eqref{eq:subproblem}. Within MAPG algorithm, step sizes taken in each block-coordinate are determined by the block Lipschitz constant, i.e. for $y$-coordinate the step size is $1/L^y_k$, while it is $1/L_k^\xi$ for the $\xi$-coordinate. On the other hand, within APG algorithm displayed in Fig.~\ref{fig:apg}, the step sizes taken in each coordinate are equal and determined by the global Lipschitz constant. Thanks to this property of MAPG, we are able to obtain faster convergence in practice in comparison to APG algorithm. When $L_k^y\approx L_k^\xi$, their performance are almost the same; however, when $\max\{L_k^\xi, L_k^y\}/\min\{L_k^\xi, L_k^y\}\gg 1$, since APG uses the global constant L, it takes very tiny steps in one of the block-coordinates.
\begin{figure}[h!]
\begin{framed}
{\small
\textbf{Algorithm MAPG} \big( $P,~L^y,~L^\xi,~\pmb{y}_0,~\pmb{\xi}_0,~\alpha^y,~\alpha^\xi,\ell^{\max}$ \big) \\
Iteration 0: Take $ \pmb{y}_0^{(1)}=\pmb{y}_1^{(2)}=\pmb{y}_0, \pmb{\xi}_0^{(1)}=\pmb{\xi}_1^{(2)}=\pmb{\xi}_0, t_1=1 $\\
Iteration $\ell$: ($\ell \geq 1$)
\begin{enumerate}
\item ${\pmb{y}_\ell}^{(1)}= \Pi_{\mathcal{Q}_y}\left( \pmb{y}_\ell^{(2)}-\frac{1}{L^y}\nabla_{\pmb{y}} P(\pmb{y}_\ell^{(2)}, \pmb{\xi}_\ell^{(2)})\right)$
\item $ {\pmb{\xi}_\ell}^{(1)}= \Pi_{\mathcal{Q}_\xi}\left(\pmb{\xi}_\ell^{(2)}-\frac{1}{L^\xi}\nabla_{\pmb{\xi}} P(\pmb{y}_\ell^{(2)}, \pmb{\xi}_\ell^{(2)})\right)$
\item $t_{\ell+1}= ( 1+ \sqrt{ 1+ 4t_\ell^2} )/2$
\item $\pmb{y}_{\ell+1}^{(2)}= \pmb{y}_\ell^{(1)} + \frac{t_\ell -1}{t_{\ell+1}} \left(\pmb{y}_\ell^{(1)} - \pmb{y}_{\ell-1}^{(1)}\right)$
\item $\pmb{\xi}_{\ell+1}^{(2)} = \pmb{\xi}_\ell^{(1)}+ \frac{t_\ell -1}{t_{\ell+1}}\left(\pmb{\xi}_\ell^{(1)}-\pmb{\xi}_{\ell-1}^{(1)}\right)$
\item \textbf{if} $\norm{\pmb{y}_\ell^{(1)}-\pmb{y}_\ell^{(2)}}_2\leq \alpha^y$ \textbf{and} $\norm{{\pmb{\xi}_\ell}^{(1)}-\pmb{\xi}_\ell^{(2)}}_2\leq \alpha^\xi$
\item \quad \textbf{return} $\big( \pmb{y}_\ell^{(1)}, \pmb{\xi}_\ell^{(1)}\big)$
\item \textbf{else if} $\ell=\ell^{\max}$
\item \quad \textbf{return} $\big( \pmb{y}_\ell^{(1)}, \pmb{\xi}_\ell^{(1)}\big)$
\item\textbf{end if
\end{enumerate}
}
\end{framed}
\vspace{-2mm}
\caption{ Modified Accelerated Proximal Gradient Algorithm}
\vspace{-4mm}
\end{figure}
Convergence and rate result of MAPG follow directly from APG in \cite{Beck09} with the help of following lemma.
\begin{lem}
\label{lem:MAPG}
Let $f: \mathbb{R}^m \rightarrow \mathbb{R}$ be a convex function, such that $\nabla_{x_1}f( \pmb{x}_1, \pmb{x}_2)$ is Lipschitz continuous with respect to $\pmb{x}_1$ with constant $L_1$, and $\nabla_{x_2}f( \pmb{x}_1, \pmb{x}_2)$ is Lipschitz continuous in $\pmb{x}_2$ with constant $L_2$. Then we have
\begin{align*}
f(\pmb{z}_1, &\pmb{z}_2)\leq \: f(\pmb{x}_1, \pmb{x}_2) + L_1 \| \pmb{z}_1 - \pmb{x}_1 \|_2^2 + L_2 \| \pmb{z}_2 - \pmb{x}_2 \|_2^2 \nonumber \\
& + \nabla_{\pmb{x}_1} f(\pmb{x}_1, \pmb{x}_2)^{\mathsf{T}} (\pmb{z}_1 -\pmb{x}_1) + \nabla_{\pmb{x}_2} f(\pmb{x}_1, \pmb{x}_2)^{\mathsf{T}} (\pmb{z}_2 -\pmb{x}_2).
\end{align*}
\end{lem}
\proof
From Lipschitz continuity of $\nabla_{x_1}f(.,x_2)$ for each $x_2$ and $\nabla_{x_2}f(x_1,.)$ for each $x_1$, it follows that
\begin{align}
\label{lipschitz3}
f( \pmb{y}_1, \pmb{x}_2) \leq f( \pmb{x}_1, \pmb{x}_2) &+ \nabla_{\pmb{x}_1} f( \pmb{x}_1, \pmb{x}_2)^{\mathsf{T}} (\pmb{y}_1 -\pmb{x}_1) \nonumber \\
&+ \tfrac{L_1}{2} \| \pmb{y}_1 - \pmb{x}_1 \|_2^2, \\
\label{lipschitz4}
f( \pmb{x}_1, \pmb{y}_2) \leq f( \pmb{x}_1, \pmb{x}_2) &+ \nabla_{\pmb{x}_2} f(\pmb{x}_1, \pmb{x}_2)^{\mathsf{T}} (\pmb{y}_2 -\pmb{x}_2) \nonumber \\
&+ \tfrac{L_2}{2} \| \pmb{y}_2 - \pmb{x}_2 \|_2^2.
\end{align}
Multiplying \eqref{lipschitz3} and \eqref{lipschitz4} with $\frac{1}{2}$, and summing them up, gives us
\begin{align*}
\tfrac{1}{2} &f( \pmb{y}_1, \pmb{x}_2) + \tfrac{1}{2} f( \pmb{x}_1, \pmb{y}_2) \\
\leq \: & f( \pmb{x}_1, \pmb{x}_2) + \tfrac{1}{2} \nabla f(\pmb{x}_1, \pmb{x}_2)^{\mathsf{T}} \begin{pmatrix} \pmb{y}_1 -\pmb{x}_1 \\ \pmb{y}_2 -\pmb{x}_2 \end{pmatrix} \\
&+ \frac{L_1}{4} \| \pmb{y}_1 - \pmb{x}_1 \|_2^2 +\frac{L_2}{4} \| \pmb{y}_2 - \pmb{x}_2 \|_2^2.
\end{align*}
Let $\pmb{z}_1 = (\pmb{x}_1+\pmb{y}_1)/2$ and $\pmb{z}_2 = (\pmb{x}_2+\pmb{y}_2)/2$, and by convexity of $f$, we have
\begin{align*}
f(\pmb{z}_1, \pmb{z}_2) \leq \frac{1}{2} f( \pmb{y}_1, \pmb{x}_2) + \frac{1}{2}& f( \pmb{x}_1, \pmb{y}_2).
\end{align*}
Combining the last two inequality concludes the proof.
\endproof
Let $\{\pmb{y}^{(1)}_\ell,\pmb{{\boldsymbol{\xi}}}^{(1)}_\ell\}_{\ell\in\mathbb{Z}_+}$ be the iterate sequence generated by MAPG algorithm while running on \eqref{eq:subproblem} starting from $(\pmb{y}_{k-1},\pmb{\xi}_{k-1})$. Using Lemma~\ref{lem:MAPG} and adapting the proof of Theorem 4.4 in~\cite{Beck09}, it can be shown that for all $\ell\geq 1$,
\begin{align*}
0&\leq P_k\left(\pmb{y}^{(1)}_\ell,\pmb{{\boldsymbol{\xi}}}^{(1)}_\ell\right) - P_k^* \\
&\leq \frac{4 \left( L_k^y \| \pmb{y}_{k-1} - \pmb{y}_k^*\|_2^2 + L_k^\xi\| \pmb{{\boldsymbol{\xi}}}_{k-1} - \pmb{{\boldsymbol{\xi}}}_k^*\|_2^2 \right)}{(\ell +1)^2},
\end{align*}
where $(\pmb{y}_k^*, \pmb{{\boldsymbol{\xi}}}_k^*)$ is a minimizer of \eqref{eq:subproblem}. Note that we have $\norm{\pmb{y}_{k-1}-\pmb{y}_k^*}_2\leq2\bar{B}_y$ and $\norm{\pmb{\xi}_{k-1}-\pmb{\xi}_k^*}_2\leq2\bar{B}_\xi$. Hence, for all $\ell\geq\ell^{\max}_k$, it is guaranteed that $\left(\pmb{y}^{(1)}_\ell,\pmb{{\boldsymbol{\xi}}}^{(1)}_\ell\right)$ is $\tau_k$-optimal to \eqref{eq:subproblem}.
Note that one can also use ALCC, displayed in Fig.~\ref{fig:alcc}, to compute the primal iterates
$\pmb{\eta}_\ell$ in Step-1 of P-APG in Fig.~\ref{fig:papg} during the $\ell$-th iteration. In particular, at beginning of every P-APG iteration, $\pmb{\eta}_\ell$ can be computed using ALCC to evaluate $\nabla g_\gamma(\pmb{\theta}_\ell^{(2)})$. More importantly, thanks to the separability of regularized \eqref{subgradient_split}, one can do this computation in parallel running ALCC on each one of the $K$ processors.
Let $N=K\bar{N}$ such that $N\geq n+1$. Below we consider the bottleneck memory requirement for solving \eqref{regularize} in 4 cases: \textbf{a)} P-APG with ALCC computing Step-1 in Fig.~\ref{fig:papg}, \textbf{b)} running ALCC \emph{alone} on \eqref{regularize}, \textbf{c)} P-APG with a primal-dual IPM computing Step-1 in Fig.~\ref{fig:papg}, and \textbf{d)} running IPM \emph{alone} on \eqref{regularize}. The bottleneck for case~\textbf{a)} is determined by Step-2 in Fig.~\ref{fig:papg}, due to dual iterates $\pmb{\theta}$ of size ${(K^2-K)\bar{N}^2+K\bar{N}}$. Similarly, for case \textbf{b)} Step-4 in Fig.~\ref{fig:alcc} requires storing $\pmb{\theta}$ of size $K^2\bar{N}^2$. On the contrary, IPM needs to solve a Newton system in each iteration for both cases \textbf{c)} and \textbf{d)}. Assuming Cholesky factorization is stored, one needs to keep $K$ lower triangular matrices in memory of size $\bar{N}(n+1)$-by-$\bar{N}(n+1)$ for case \textbf{c)}, and to keep 1 lower triangular matrix of size $N(n+1)$-by-$N(n+1)$ for case \textbf{c)}.
Above discussion is summarized in Table~\ref{memory}. Note that running IPM within P-APG reduces the memory requirement significantly by a factor of $K$ in comparison to running IPM alone, e.g. if we partition $N$ observations into $K=10$ subsets and each subproblem requires 1GB of memory, then running IPM alone requires roughly 100GB, while IPM within P-APG requires only 10GB in total.
\begin{table}[htdp]
\centering
\caption{Comparison of Memory Usage}
\begin{tabular}{lll}
\toprule
& IPM & ALCC \\
\midrule
Alone & $\mathcal{O}\big( K^2 \bar{N}^2 (n+1)^2\big)$ & $\mathcal{O}\big( K^2\bar{N^2} \big)$ \\
P-APG & $\mathcal{O}\big( K \bar{N}^2 (n+1)^2 + (K^2-K)\bar{N}^2\big)$ & $\mathcal{O}\big( (K^2-K)\bar{N}^2 \big)$ \\
\bottomrule
\end{tabular}
\label{memory}
\end{table}
\vspace{-4mm}
\section{Numerical Study}
\label{numerical}
In this section, we provide a comparison in Matlab among the following methods: Sedumi, ALCC, Mosek, P-APG with Sedumi, P-APG with ALCC, and P-APG with Mosek, on problem~\eqref{regularize} with increasing dimension. The numerical study is mainly aimed to demonstrate how the performance of each method scales with the dimension of the problem.
First, we start with a small size problem: $n=5, N=100$. $\Lambda\in\mathbb{R}^{n\times n}$, $\{ \pmb{x}_i \}_{i=1}^N\subset\mathbb{R}^n$ and $\{ \epsilon_i \} _{i=1}^N\subset\mathbb{R}$ are generated randomly with all the components being i.i.d. with $\mathcal{N}(0,1)$, and $\bar{y}_i$ are generated according to \eqref{data}, where $f_0(\pmb{x})=\frac{1}{2}\pmb{x}^{\mathsf{T}} Q \pmb{x}$, and $Q= \Lambda^{\mathsf{T}} \Lambda$. We
compare the quality of the solutions computed by P-APG and dual gradient ascent (as the dual function $g_\gamma$ in \eqref{eq:g_gamma} is differentiable). In order to compute dual gradient, $\nabla g_\gamma$, one needs to solve $K$ quadratic subproblems. To exploit this parallel structure, we partition the data into two sets, i.e. $K=2$. Within both the dual gradient ascent and P-APG, we called ALCC to compute the dual gradients via solving $K$ QP subproblems. Since we allow violations for the relaxed constraints, ``duality gap" in the paper is defined as $\pmb{\theta}_k^{\mathsf{T}} C \pmb{\eta}_k$ at $k^{th}$ iteration. Fig.~\ref{Gap_All} represents how the duality gap of both methods changes at each iteration. In order to better understand the behavior of P-APG, we report in Fig.~\ref{Gap_APG} the duality gap of P-APG in a smaller scale. Fig.~\ref{Distance_All} reports the infeasibility of iterates, i.e. $\big\| \big( A_1\pmb{y}_k +A_2\pmb{\xi}_k \big)_{-} \big\|_2$. \vspace{-4mm}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.35\textwidth]{Gap_All_1}
\caption{ Duality Gap for P-APG and Dual Gradient Ascent}
\label{Gap_All}
\vspace{-4mm}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.35\textwidth]{Gap_APG_1}
\vspace{-2mm}
\caption{ Duality Gap for P-APG Method}
\label{Gap_APG}
\vspace{-4mm}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.35\textwidth]{Distance_All_1}
\caption{ Distance to Feasible Region for P-APG and Dual Gradient Ascent}
\label{Distance_All}
\vspace{-4mm}
\end{figure}
\begin{table}[t]
\setlength\tabcolsep{2.7pt}
\caption{ Comparison with test function $\exp( \pmb{p}^{\mathsf{T}}\pmb{x}) $ }
\label{tab:exp}
\centering
{\renewcommand{\arraystretch}{0.65}
\begin{tabular}{llccccc}
\toprule
N & Solver & CPU & W.T. & $\frac{1}{2} \big\| \pmb{y} - \bar{\pmb{y}} \big\|_2^2$ & Gap & Infeas. \\
\midrule
\multirow{6}{*}{200}
& Sedumi & 2.69 & 2.69 & 1.16E-05 & 0 & 0 \\
& ALCC & 1.86 & 1.86 & 1.16E-05 & -1.19E-07 & 9.6E-02 \\
& Mosek & 0.74 & 0.74 & 1.16E-05 & 2.68E-08 & 0 \\
& PAPG(Sedumi) & 29.64 & 14.84 & 1.17E-05 & 1.31E-07 & 9.9E-02 \\
& PAPG(ALCC) & 9.93 & 4.98 & 1.19E-05 & 2.93E-08 & 9.9E-02 \\
& PAPG(Mosek) & 3.77 & 1.91 & 1.17E-05 & 9.85E-08 & 9.4E-02 \\
\midrule
\multirow{6}{*}{400} & Sedumi & O.M. & O.M. & O.M. & O.M. & O.M. \\
& ALCC & 14.74 & 14.74 & 6.01E-05 & -1.15E-07 & 9.8E-02 \\
& Mosek & O.M. & O.M. & O.M. & O.M. & O.M. \\
& PAPG(Sedumi) & 120.48 & 30.28 & 6.00E-05 & 4.52E-09 & 9.7E-02 \\
& PAPG(ALCC) & 35.83 & 9.11 & 6.11E-05 & -2.87E-08 & 9.8E-02 \\
& PAPG(Mosek) & 15.73 & 4.12 & 6.00E-05 & 4.97E-09 & 9.7E-02 \\
\midrule
\multirow{6}{*}{800} & Sedumi & O.M. & O.M. & O.M. & O.M. & O.M. \\
& ALCC & 93.57 & 93.57 & 2.02E-04 & -8.50E-08 & 9.9E-02 \\
& Mosek & O.M. & O.M. & O.M. & O.M. & O.M. \\
& PAPG(Sedumi) & 146 & 19 & 2.46E-04 & 2.41E-08 & 9.9E-02 \\
& PAPG(ALCC) & 118.54 & 15.77 & 2.10E-04 & 7.01E-08 & 9.8E-02 \\
& PAPG(Mosek) & 52.43 & 7.52 & 2.05E-04 & 6.49E-08 & 9.7E-02 \\
\midrule
\multirow{6}{*}{1600} & Sedumi & O.M. & O.M. & O.M. & O.M. & O.M. \\
& ALCC & N/A & N/A & N/A & N/A & N/A \\
& Mosek & O.M. & O.M. & O.M. & O.M. & O.M. \\
& PAPG(Sedumi) & N/A & N/A & N/A & N/A & N/A \\
& PAPG(ALCC) & 323.68 & 23.94 & 1.97E-03 & 5.85E-09 & 9.9E-02 \\
& PAPG(Mosek) & 204.56 & 17.14 & 1.85E-03 & -6.72E-10 & 9.9E-02 \\
\bottomrule
\end{tabular} }
\end{table}
\begin{table}[h]
\setlength\tabcolsep{3pt}
\caption{ Comparison with test function $\frac{1}{2}\pmb{x}^{\mathsf{T}}Q\pmb{x}$ }
\label{tab:xqx}
\centering
{\renewcommand{\arraystretch}{0.65}
\begin{tabular}{llccccc}
\toprule
N & Solver & CPU & W.T. & $\frac{1}{2} \big\| \pmb{y} - \bar{\pmb{y}} \big\|_2^2$ & Gap & Infeas. \\
\midrule
\multirow{6}{*}{200}
& Sedumi & 2.88 & 2.88 & 1.25E-04 & 0 & 0 \\
& ALCC & 3.58 & 3.58 & 1.25E-04 & -3.41E-08 & 9.8E-03 \\
& Mosek & 0.65 & 0.65 & 1.25E-04 & 2.11E-08 & 0 \\
& PAPG(Sedumi) & 16.7 & 8.41 & 1.29E-04 & 4.48E-06 & 6.5E-02 \\
& PAPG(ALCC) & 12.5 & 6.3 & 1.25E-04 & 9.49E-08 & 9.8E-02 \\
& PAPG(Mosek) & 5.57 & 2.8 & 1.26E-04 & 1.45E-07 & 8.6E-02 \\
\midrule
\multirow{6}{*}{400} & Sedumi & O.M. & O.M. & O.M. & O.M. & O.M. \\
& ALCC & 33.3 & 33.3 & 1.02E-03 & -6.84E-08 & 1.4E-02 \\
& Mosek & O.M. & O.M. & O.M. & O.M. & O.M. \\
& PAPG(Sedumi) & 164 & 41.2 & 1.02E-03 & -8.62E-08 & 9.5E-02 \\
& PAPG(ALCC) & 63.2 & 15.7 & 1.01E-03 & 1.53E-07 & 9.7E-02 \\
& PAPG(Mosek) & 29.01 & 7.43 & 1.02E-03 & -8.74E-08 & 9.5E-02 \\
\midrule
\multirow{6}{*}{800} & Sedumi & O.M. & O.M. & O.M. & O.M. & O.M. \\
& ALCC & 140 & 140 & 4.03E-02 & -2.94E-07 & 9.9E-02 \\
& Mosek & O.M. & O.M. & O.M. & O.M. & O.M. \\
& PAPG(Sedumi) & 303.33 & 39.87 & 4.04E-03 & -1.95E-07 & 9.9E-02 \\
& PAPG(ALCC) & 206 & 23.9 & 4.04E-03 & -2.03E-07 & 9.9E-02 \\
& PAPG(Mosek) & 100.32 & 14.34 & 4.04E-03 & -1.95E-07 & 9.9E-02 \\
\midrule
\multirow{6}{*}{1600} & Sedumi & O.M. & O.M. & O.M. & O.M. & O.M. \\
& ALCC & N/A & N/A & N/A & N/A & N/A \\
& Mosek & O.M. & O.M. & O.M. & O.M. & O.M. \\
& PAPG(Sedumi) & N/A & N/A & N/A & N/A & N/A \\
& PAPG(ALCC) & 480 & 29.8 & 5.28E-03 & 1.10E-07 & 9.9E-02 \\
& PAPG(Mosek) & 273.93 & 21.47 & 5.23E-03 & 5.59E-08 & 9.8E-02 \\
\bottomrule
\end{tabular} }
\end{table}
A primal-dual iterate $(\pmb{\eta},\pmb{\theta})$ is optimal if the duality gap and infeasibility are both zero. As the feasibility happens in the limit, the duality gap in Fig.~\ref{Gap_APG} can go below the red line, which can be explained by the infeasibility of iterates. Therefore, observing a decrease in duality gap only tells one part of the story; without convergence to feasibility, it is not valuable alone as a measure. As shown in the Fig.~\ref{Gap_All} and Fig.~\ref{Gap_APG}, the duality gap converges quickly to zero for both methods. On the other hand, as shown in Fig.~\ref{Distance_All}, constraint violation for P-APG iterates decreases to 0 much faster than it does for the dual gradient ascent iterates. Hence, P-APG iterate sequence converges to the unique optimal solution considerably faster.
The larger scale problems are carried out on a single node at a research computing cluster. The node is composed of one 16-core processor sharing 32GB. For P-AGPG numerical tests, in each job submitted to the computing cluster, an instance of \eqref{regularize} is solved using P-APG on the node such that each subproblem is computed on a different core. The dimension of variables $n=80$ and the number of observations $N=200, 400, 800, 1600$. We partition the set of observations into $K$ subsets. Each one of them consists of 100 points. So, $K=2, 4, 8, 16$ for $N=200, 400, 800, 1600$, respectively. In all the tables, \textit{N/A} means that the wall clock time exceeded 2 hours for the job, and \textit{O.M.} means the algorithm in focus runs out of memory. Also CPU denotes the CPU run time in \emph{minutes}; W.T. stands for wall-clock time in \emph{minutes}. Since the number of constraints increases at the rate of $\mathcal{O}(N^2)$, as the size of problem increases in $N$, we reported the normalized infeasibility and normalized duality gap, which are $\norm{\big( A_1\pmb{y} +A_2\pmb{\xi} \big)_{-}}_2/\sqrt{N^2-N}$ and $\pmb{\theta}_k^{\mathsf{T}} C \pmb{\eta}_k/(N^2-N)$, respectively. We report numerical results for the following test functions: $f_0(\pmb{x})=\frac{1}{2}\pmb{x}^{\mathsf{T}} Q \pmb{x}$, $f_0(\pmb{x}) = \exp( \pmb{p}^{\mathsf{T}} \pmb{x} )$, where $Q$ is generated as discussed before, and $\pmb{p} \in \mathbb{R}^n$ is generated using uniform distribution.
\begin{table}[h!]
\setlength\tabcolsep{3.7pt}
\caption{ Replications with test function $\exp( \pmb{p}^{\mathsf{T}}\pmb{x}) $ }
\label{tab:exprep}
\centering
{\renewcommand{\arraystretch}{0.55}
\begin{tabular}{lcccccc}
\toprule
Solver & Rep. & CPU & W.T. & $\frac{1}{2} \big\| \pmb{y} - \bar{\pmb{y}} \big\|_2^2$ & Gap & Infeas. \\
\midrule
\multirow{5}{*}{PAPG(ALCC)} & 1 & 118.54 & 15.77 & 2.10E-04 & 7.01E-08 & 9.8E-02 \\
& 2 & 131.93 & 17.56 & 3.90E-04 & 5.52E-09 & 9.9E-02 \\
& 3 & 136.50 & 18.01 & 4.69E-04 & -3.87E-09 & 9.8E-02 \\
& 4 & 126.43 & 16.74 & 2.91E-04 & -2.53E-09 & 9.9E-02 \\
& 5 & 144.31 & 18.98 & 5.35E-04 & -7.62E-08 & 9.8E-02 \\
\midrule
\multirow{5}{*}{PAPG(Mosek)} & 1 & 52.43 & 7.52 & 2.05E-04 & 6.49E-08 & 9.7E-02 \\
& 2 & 57.33 & 8.12 & 3.77E-04 & 9.39E-09 & 9.9E-02 \\
& 3 & 61.53 & 8.64 & 4.54E-04 & -5.30E-09 & 9.9E-02 \\
& 4 & 55.86 & 8.00 & 2.84E-04 & -7.48E-09 & 9.9E-02 \\
& 5 & 65.04 & 9.15 & 5.15E-04 & -7.74E-08 & 9.8E-02 \\
\bottomrule
\end{tabular} }
\end{table}
\begin{table}[h!]
\setlength\tabcolsep{3.7pt}
\caption{ Replications with test function $\frac{1}{2}\pmb{x}^{\mathsf{T}}Q\pmb{x}$ }
\label{tab:xqxrep}
\centering
{\renewcommand{\arraystretch}{0.55}
\begin{tabular}{lcccccc}
\toprule
Solver & Rep. & CPU & W.T. & $\frac{1}{2} \big\| \pmb{y} - \bar{\pmb{y}} \big\|_2^2$ & Gap & Infeas. \\
\midrule
\multirow{5}{*}{PAPG(ALCC)} & 1 & 206.00 & 23.90 & 4.04E-03 & -2.03E-07 & 9.9E-02 \\
& 2 & 213.27 & 27.63 & 1.00E-03 & -9.04E-08 & 9.6E-02 \\
& 3 & 211.18 & 27.37 & 1.11E-03 & -1.09E-07 & 9.9E-02 \\
& 4 & 178.77 & 23.41 & 7.29E-04 & -6.20E-08 & 9.9E-02 \\
& 5 & 200.62 & 26.14 & 1.27E-03 & -1.05E-07 & 9.6E-02 \\
\midrule
\multirow{5}{*}{PAPG(Mosek)} & 1 & 100.32 & 14.34 & 4.04E-03 & -1.95E-05 & 9.9E-02 \\
& 2 & 79.27 & 10.87 & 9.88E-04 & -1.04E-07 & 9.9E-02 \\
& 3 & 83.11 & 11.43 & 1.09E-03 & -1.19E-07 & 9.9E-02 \\
& 4 & 68.90 & 9.66 & 7.16E-04 & -2.30E-08 & 9.9E-02 \\
& 5 & 79.55 & 10.99 & 1.24E-03 & -1.19E-07 & 9.8E-02 \\
\bottomrule
\end{tabular} }
\end{table}
\addtolength{\textheight}{- 12.3 cm}
All the algorithms are terminated either when they compute an iterate with normalized infeasibility and normalized duality gap are less than 1E-01 and 1E-06, respectively, or at the end of 2 hours. The numerical results reported in Table~\ref{tab:exp} and~\ref{tab:xqx} show that P-APG solution is very close to the real optimal solution of \eqref{regularize}. Note that ALCC fails to terminate within in 2 hours when $N=1600$; and interior point methods fail to run anything beyond $N=200$ due $\mathcal{O}(N^2n^2)$ memory requirement. Moreover, in order to test the robustness of P-APG, we solved 5 random instances when $N=800$, of which results are reported in Table~\ref{tab:exprep} and Table~\ref{tab:xqxrep}.
Numerical results show that advantages of P-APG over running IPM or ALCC alone on \eqref{regularize} become more and more evident as the dimension of the problem increases.
\section{Conclusion}
In this paper, we proposed P-APG method to efficiently compute the least squares estimator for large scale convex regression problems. By relaxing constraints partially, we obtained the separability on the corresponding Lagrangian dual problem. Using Tikhonov regularization, we ensured the feasibility of iterates in the limit, and we provided error bounds on 1) the distance between the inexact solution to the regularized problem and the optimal solution to the original problem, 2) the constraint violation of the regularized solution. The comparison in the numerical section demonstrates the efficiency of P-APG method on memory usage compared to IPM. Furthermore, the extended random tests show the stability of P-APG method. Due to limited space, we could not include computational results on real-life data; but they will be made available online at authors' webpage.
\bibliographystyle{unsrt}
|
\section{\setcounter{equation}{0}\oldsection}
\renewcommand\thesection{\arabic{section}}
\renewcommand\theequation{\thesection.\arabic{equation}}
\newtheorem{claim}{\noindent Claim}[section]
\newtheorem{theorem}{\noindent Theorem}[section]
\newtheorem{lemma}{\noindent Lemma}[section]
\newtheorem{proposition}{\noindent Proposition}[section]
\newtheorem{definition}{\noindent Definition}[section]
\newtheorem{remark}{\noindent Remark}[section]
\newtheorem{corollary}{\noindent Corollary}[section]
\newtheorem{example}{\noindent Example}[section]
\title{On global multidimensional supersonic flows with vacuum states at infinity}
\author{Xu Gang$^{1}$, \quad Yin
Huicheng$^{2}$\footnote{Xu Gang was supported
by the National Natural Science Foundation of China (No.11101190)
and Natural Science Fundamental Research Project of Jiangsu Colleges
(No.10KLB110002); Yin Huicheng was
supported by the NSFC (No.~10931007, No.~11025105) and the Priority
Academic Program Development of Jiangsu Higher Education
Institutions.}\vspace{0.5cm}\\
\small 1. Faculty of Science, Jiangsu University, Zhenjiang, Jiangsu
212013, China.\\
\small 2. Department of Mathematics
and IMS, Nanjing University, Nanjing 210093, China.\\
}
\date{}
\maketitle
\centerline {\bf Abstract} \vskip 0.3 true cm
In this paper, we are
concerned with the global existence and stability of a smooth supersonic flow
with vacuum state at infinity in a 3-D infinitely long divergent nozzle.
The flow is described by a 3-D steady potential equation, which is multi-dimensional quasilinear
hyperbolic (but degenerate at infinity) with respect to the supersonic direction,
and whose linearized part admits the form
$\partial_t^2-\displaystyle\frac{1}{(1+t)^{2(\gamma-1)}}(\partial_1^2+\partial_2^2)+\displaystyle\frac{2(\gamma-1)}{1+t}\partial_t$ for $1<\gamma<2$. From the physical point of view,
due to the expansive geometric property of the divergent nozzle and the mass conservation of gas, the
moving gas in the nozzle will gradually
become rarefactive and tends to a vacuum state at infinity, which implies that such a smooth
supersonic flow should be globally stable for small perturbations since there are no strong
resulting compressions in the motion of the flow. We will confirm such a global stability phenomena
by rigorous mathematical proofs and further show that there do not exist
vacuum domains in any finite part of the nozzle.
\vskip 0.3 true cm
{\bf Keywords:} Supersonic flow, divergent nozzle, vacuum,
anisotropic weighted energy estimate, global existence\vskip 0.3 true cm
{\bf Mathematical Subject Classification 2000:} 35L70, 35L65,
35L67, 76N15
\vskip 0.4 true cm
\centerline{\bf $\S 1$. Introduction and main results}
\vskip 0.3 true cm
In this paper, we are concerned with the global existence and stability of
a smooth supersonic polytropic gas with vacuum state at infinity in a 3-D infinitely long divergent nozzle.
The divergent nozzle is described by the domain $\Omega=\{x=(x_1, x_2, x_3)\in\Bbb R^3: x_1^2+x_2^2\le\tan^2\varphi_0 x_3^2,
x_1^2+x_2^2+x_3^2\ge 1, x_3>0\}$
with $\varphi_0\in (0, \displaystyle\frac{\pi}{2})$ (see the Figure 1 below), and the potential function $\Phi$
of irrotational polytropic gas satisfies the
following steady potential equation in $\Omega$:
$$
\displaystyle\sum_{i=1}^3((\partial_i\Phi)^2-c^2(\rho))\partial_i^2\Phi+2 \displaystyle\sum_{1\le i<j\le
3}\partial_i\Phi\partial_j\Phi\partial_{ij}^2\Phi=0,\eqno{(1.1)}
$$
where $\partial_i=\partial_{x_i}$ ($1\le i\le 3$), $c(\rho)=\sqrt{P'(\rho)}$ is the local sound speed, $P(\rho)$ is the pressure,
$\rho$ is the density, and state equation is given by $P(\rho)=\rho^{\gamma}$ with $1<\gamma<2$ (for the air, $\gamma\approx 1.4$).
Moreover, the density $\rho=\rho(\nabla_x\Phi)$ can be determined by the Bernoulli's law:
$$\frac{1}{2}|\nabla_x \Phi|^2+\frac{\gamma}{\gamma-1}\rho^{\gamma-1}=C_0\equiv \f12 q_0^2+\frac{\gamma}{\gamma-1}\rho_0^{\gamma-1},\eqno{(1.2)}$$
where $\nabla_x=(\partial_1, \partial_2, \partial_3)$, and $q_0>c(\rho_0)$ (this means that the flow at the entrance is supersonic
along the radial direction).
Without loss of generality and for convenience,
$C_0=1$ will be always assumed in the whole paper.
\vskip 0.5 true cm
\includegraphics[width=12cm,height=6.5cm]{1.png}
\centerline{\bf Figure 1. Supersonic flow in a 3-D divergent nozzle}
\vskip 0.8 true cm
Denote the divergent nozzle wall by $\Sigma=\{x:
x_1^2+x_2^2=\tan^2\varphi_0 x_3^2, x_1^2+x_2^2+x_3^2\ge 1, x_3>0\}$, then
$\Phi$ satisfies the following fixed boundary condition on $\Sigma$:
$$x_1\partial_1\Phi+x_2\partial_2\Phi-\tan^2\varphi_0x_3\partial_3\Phi=0.\eqno{(1.3)}$$
Due to the divergent geometric property of $\Omega$, it is convenient to work in
the spherical coordinates $(r, \theta, \varphi)$:
$$(x_1, x_2, x_3)=(r\cos\theta\sin\varphi, r\sin\theta\sin\varphi, r\cos\varphi),\eqno{(1.4)}$$
where
$r=\sqrt{x_1^2+x_2^2+x_3^2}$, $0\le\theta\le 2\pi$ and $0\le\varphi\le\varphi_0$.
Under the coordinate transformation (1.4), (1.1) becomes
\begin{align*}
&\big((\partial_r\Phi)^2-c^2(\rho)\big)\partial_r^2\Phi+
\frac{1}{r^2\sin^2\varphi}\bigg(\frac{1}{r^2\sin^2\varphi}(\partial_{\theta}\Phi)^2-c^2(\rho)\bigg)\partial_{\theta}^2\Phi
+\frac{1}{r^2}\bigg(\frac{1}{r^2}(\partial_{\varphi}\Phi)^2-c^2(\rho)\bigg)\partial_{\varphi}^2\Phi\\
&\qquad +\frac{2\partial_r\Phi\partial_{\theta}\Phi}{r^2\sin^2\varphi}\partial_{r\theta}^2\Phi+\frac{2}{r^2}\partial_r\Phi\partial_{\varphi}\Phi\partial_{r\varphi}^2\Phi
+\frac{2\partial_{\theta}\Phi\partial_{\varphi}\Phi}{r^4\sin^2\varphi}\partial_{\theta\varphi}^2\Phi
-\frac{1}{r^3}\bigg(2r^2c^2(\rho)+(\partial_{\varphi}\Phi)^2\\
&\qquad +\frac{1}{\sin^2{\varphi}}(\partial_{\theta}\Phi)^2\bigg)\partial_r\Phi
-\frac{\cot\varphi}{r^4}\bigg(r^2c^2(\rho)+\frac{1}{\sin^2\varphi}(\partial_{\theta}\Phi)^2\bigg)\partial_{\varphi}\Phi=0.\tag{1.5}
\end{align*}
In particular, if the solution $\Phi$ of (1.5) is axially symmetric, namely, $\Phi(r,\theta,\varphi)\equiv \Phi(r,\varphi)$
is independent of the variable $\theta$, then (1.5) becomes
\begin{align*}
&\big((\partial_r\Phi)^2-c^2(\rho)\big)\partial_r^2\Phi
+\frac{1}{r^2}\bigg(\frac{1}{r^2}(\partial_{\varphi}\Phi)^2-c^2(\rho)\bigg)\partial_{\varphi}^2\Phi
+\frac{2}{r^2}\partial_r\Phi\partial_{\varphi}\Phi\partial_{r\varphi}^2\Phi\\
&\qquad -\frac{1}{r^3}\big(2r^2c^2(\rho)+(\partial_{\varphi}\Phi)^2\big)\partial_r\Phi
-\frac{c^2(\rho)}{r^2}cot\varphi\partial_{\varphi}\Phi=0.\tag{1.6}
\end{align*}
Here we point out that some coefficients in (1.5) or (1.6) admit strong singularities near $\varphi=0$.
Consequently, in order to overcome
the difficulties arisen by the singularities
near $\varphi=0$, we require to rewrite (1.5) or (1.6) by introducing some
smooth vector fields tangent to the sphere $\Bbb S^2$ as in [15].
Set
\begin{equation}
\left\{
\begin{aligned}
&Z_1=x_1\partial_2-x_2\partial_1=\partial_{\theta},\\
&Z_2=x_2\partial_3-x_3\partial_2=-\cot\varphi\cos\theta\partial_{\theta}-\sin\theta\partial_{\varphi},\\
&Z_3=x_3\partial_1-x_1\partial_3=-\cot\varphi\sin\theta\partial_{\theta}+\cos\theta\partial_{\varphi}.
\end{aligned}
\right.\tag{1.7}
\end{equation}
Then it follows from a direct computation that (1.5) or (1.6) has such a new form
\begin{align*}
&((\partial_r\Phi)^2-c^2(\rho))\partial_r^2\Phi+\frac{2\partial_r\Phi}{r^2}\sum\limits_{i=1}^3Z_i\Phi\partial_rZ_i\Phi
-\frac{c^2(\rho)}{r^2}\sum\limits_{i=1}^3Z_i^2\Phi
+\frac{1}{r^4}\sum\limits_{i,j=1}^3Z_i\Phi
Z_j\Phi
Z_iZ_j\Phi\\
&\quad +\sum\limits_{i,j=1}^3\frac{C_{ij}(\omega)}{r^3}\partial_r\Phi Z_i\Phi
Z_j\Phi+\sum\limits_{i,j,k=1}^3\frac{C_{ijk}(\omega)}{r^4}Z_i\Phi
Z_j\Phi
Z_k\Phi-\frac{2c^2(\rho)}{r}\partial_r\Phi=0,\tag{1.8}
\end{align*}
where $\omega=\displaystyle \frac{x}{r}$,
$C_{ij}(\omega)=C_{ij}(\displaystyle \frac{x}{r})$ and $C_{ijk}(\omega)=C_{ijk}(\displaystyle\frac{x}{r})$
are smooth functions on their arguments.
Meanwhile, the fixed boundary condition (1.3) can be changed as
$$
x_1Z_3\Phi-x_2Z_2\Phi=0\qquad \text{on}~~\Sigma.\eqno{(1.9)}
$$
Especially, for the axially symmetric solution $\Phi$, the boundary condition on $\Sigma$ is
$$
Z_2\Phi=Z_3\Phi=0\qquad \text{on}~~\Sigma.\eqno{(1.10)}
$$
In addition, we impose the following initial axially symmetric perturbations:
$$\Phi(1,\theta, \varphi)=\varepsilon\Phi_0(\varphi),\qquad
\partial_r\Phi(1,\theta,\varphi)=q_0+\varepsilon\Phi_1(\varphi),\eqno{(1.11)}
$$
where $\varepsilon>0$ is a small parameter, and
$\Phi_i(\varphi)\in C_0^{\infty}[0,\varphi_0)$ ($i=0,1$). In fact, such kinds of initial conditions (1.11)
can be easily realized by small axially symmetric perturbations on the initial
density and velocity of irrotational gas.
\vskip 0.5 true cm
\includegraphics[width=11cm,height=6.0cm]{3D3.png}
\centerline{\bf Figure 2. Perturbed supersonic flow in the 3-D divergent nozzle}
\vskip 0.8 true cm
Let $\Gamma=\{r=r(\theta, \varphi): 0\le\theta\le 2\pi, 0\le\varphi\le\varphi_0\}$ be any $C^1-$smooth
cross section of $\Omega$ (see the Figure 2 above). Denote the positive constant
$m_{\varepsilon}
=2\pi\int_0^{\varphi_0}\rho_0^{\varepsilon}(\varphi)
(q_0+\varepsilon\Phi_1(\varphi))sin\varphi d\varphi$, where the initial density $\rho_0^{\varepsilon}(\varphi)=(\displaystyle\frac{\gamma-1}{\gamma})^{\frac{1}{\gamma-1}}\biggl\{\displaystyle\frac{\gamma}{\gamma-1}\rho_0^{\gamma-1}-\f12
\biggl(2q_0\varepsilon \Phi_1+\varepsilon^2\Phi_1^2+\varepsilon^2(\Phi'_0)^2\biggr)\biggr\}^{\frac{1}{\gamma-1}}$
is determined by the Bernoulli's law (1.2). The main result in our paper is:
{\bf Theorem 1.1.} {\it There exists a constant $\varepsilon_0>0$ depending on $q_0,\rho_0$
and $\gamma$ such that problem (1.8) with (1.10)-(1.11)
possesses a global $C^{\infty}$ supersonic solution $\Phi(x)$ for
$\varepsilon<\varepsilon_0$ and the mass of gas on any smooth cross surface $\Gamma$ is conserved, namely,
$\int_{\Gamma}\rho\nabla_x\Phi\cdot {\overrightarrow {n}}dS
\equiv m_{\varepsilon}$, where $\overrightarrow {n}$ stands for the unit outward normal direction of
$\Gamma$. Moreover, $\rho(x)>0$ and $\displaystyle\lim_{r\to\infty}\rho(x)=0$ hold in the whole $\Omega$.}
{\bf Remark 1.1.} {\it From Theorem 1.1, one easily knows that there do not exist
vacuum domains in any finite part of $\Omega$ for the problem (1.8) together with (1.10)-(1.11).}
{\bf Remark 1.2.} {\it For the small arbitrarily (not axially symmetric) perturbed supersonic flow in $\Omega$,
which is determined by the equation
(1.8) together with (1.9) and the initial data $(\Phi(1,\theta,\varphi),$ $\partial_r\Phi(1,\theta,\varphi))
=(\varepsilon \Phi_0(\theta,\varphi), q_0+\varepsilon\Phi_1(\theta,\varphi))$ with $\Phi_i(\theta,\varphi)\in C_0^{\infty}([0, 2\pi]\times [0, \varphi_0))$ $(i=0,1)$,
we can also solve the global stability problem as in Theorem 1.1 by analogous but much more complicated analysis.
Nevertheless, due to the lengthy formulas and too heavy computations, we do not give out the related details of
proof procedure here.}
{\bf Remark 1.3.} {\it By the same analysis in this paper, Theorem 1.1 can be extended into the curved
2-D or 3-D divergent nozzles with small and arbitrary perturbations of straight boundaries (one can see the
following Figure 3 and Figure 4).}
\vskip 0.3 true cm
\includegraphics[width=11cm,height=5.5cm]{B.png}
\centerline{\bf Figure 3. 2-D global smooth supersonic flow in a curved divergent nozzle}
\vskip 0.8 true cm
\includegraphics[width=11cm,height=5.5cm]{A.png}
\centerline{\bf Figure 4. 3-D global smooth supersonic flow in a curved divergent nozzle}
\vskip 0.8 true cm
{\bf Remark 1.4.} {\it For the de Laval nozzle, which is constructed by
a converging ``entry'' section and a diverging ``exhaust''
section, when the supersonic flow is formed across the sonic curve
in the slowly variable nozzle and the infinite long nozzle walls approach two symmetric lines (see the Figure 5 below),
then our Theorem 1.1 illustrates that the smooth supersonic flow exists globally for the small
perturbed state. On the other hand, if the de Laval nozzle is finitely long and
an appropriately large exit pressure
$p_e$ is given, as stated in Section 147 of [5],
at a certain place in
the diverging part of the nozzle a shock front intervenes and the gas is compressed and slowed down to
subsonic speed (see the Figure 6 below). This phenomenon has been extensively studied, especially the
stability problem of a transonic shock is completely solved for
a general class of 2-D de
Laval nozzles whose divergent parts are small and arbitrary perturbations of divergent angular domains for
the full steady compressible Euler system in [16].}
\vskip 0.5 true cm
\includegraphics[width=13cm,height=5.5cm]{2D2.png}
\centerline{\bf Figure 5. Global continuous transonic flow in an infinite long de Laval nozzle}
\vskip 0.8 true cm
\vskip 0.5 true cm
\includegraphics[width=16cm,height=5.5cm]{S.png}
\centerline{\bf Figure 6. Stability of a transonic shock in a finitely long de Laval nozzle}
\vskip 0.8 true cm
{\bf Remark 1.5.} {\it The nonlinear equation (1.1) in our case is actually a two dimensional
quasilinear degenerate wave equation
if one regards $r$ as the time since the flow is
supersonic in $r$-direction, whose linearized part is like $\partial_t^2-\displaystyle\frac{1}{(1+t)^{2(\gamma-1)}}
(\partial_1^2+\partial_2^2)+\displaystyle\frac{2(\gamma-1)}{1+t}\partial_t$ (one can see Remark 3.1 below in $\S 3$). On the other hand,
if we consider the Cauchy initial data problem of (1.1) which is of a small
perturbation with respect to the uniform constant density $\rho_0$ and velocity $(0,0,q_0)$
\begin{equation}
\left\{
\begin{aligned}
&\displaystyle\sum_{i=1}^3((\partial_i\Phi)^2-c^2(\rho))\partial_i^2\Phi+2 \displaystyle\sum_{1\le i<j\le
3}\partial_i\Phi\partial_j\Phi\partial_{ij}^2\Phi=0, \qquad x_3\ge 0,\\
&\Phi(x)|_{x_3=0}=\varepsilon\Phi_0(x_1,x_2),\quad \partial_3\Phi(x)|_{x_3=0}=q_0+\varepsilon\Phi_1(x_1,x_2),\qquad (x_1, x_2)
\in\Bbb R^2,
\end{aligned}
\right.\tag{1.12}
\end{equation}
where $q_0>c(\rho_0)$, and $\Phi_i(x_1,x_2)\in C_0^{\infty}(\Bbb R^2)$ ($i=0,1$), then by a direct verification,
one sees that (1.12) does not fulfill the ``null-condition'' put
forward in [4] and [14]. Therefore, in terms of the extensive
results of [1-2], [9], [21], [28] and so on, the classical solution to
(1.12) will blow up for finite $x_3$. However, compared this blowup result with our Theorem 1.1,
we obtain the global existence of a smooth solution to (1.12) together with the divergent
nozzle wall condition due to the rarefactive property of
supersonic gas.}
{\bf Remark 1.6.} {\it If the initial density contains vacuum, the local well-posedness results of
compressible Euler system have been extensively studied in [3], [6-7], [12-13], [18], [20], [25-27]
and so on. In the general case, such local classical solution will blow up in finite time as shown in [3], [25] and
the references therein. With respect to the problem in our paper, the vacuum only appears at infinity
and the smooth solution exists globally.}
{\bf Remark 1.7.} {\it If the initial velocity $u_0(x)$ of gas forces particles to spread out, roughly speaking,
$u_0(x)$ is close to a linear field, which means $\displaystyle{\overline\lim}_{
|x|\rightarrow \infty}|u_0(x)|
=\infty$, the authors in [11] and [22] have proved the global existence
of smooth solutions to the Cauchy problem of compressible Euler system. Here we emphasize that
our initial data (1.11) are not the cases posed in [11] or [22]
(for example, one can see Theorem 1 of [11]).}
Let us comment on the proof of Theorem 1.1. Since the local solvability
of problem (1.8) together with (1.10)-(1.11) has been known
as long as the vacuum does not appear, we will use the continuous induction method to
prove Theorem 1.1. To achieve
this objective, we need to establish the global
energy estimates with suitable anisotropic weights for (1.8) with
(1.10)-(1.11), which is degenerate at infinity and admits a linear part as follows:
$\partial_r^2-\displaystyle\frac{1}{(1+r)^{2(\gamma-1)}}(\partial_1^2+\partial_2^2)+\displaystyle\frac{2(\gamma-1)}{1+r}\partial_r$. Based on such estimates,
one then obtains the absence of vacuum
for any finite $x_3>0$ in $\Omega$, the global
existence, stability, and the asymptotic behavior of the
solution. The key ingredients in
the analysis to obtain weighted energy estimates are to look
for an appropriate multiplier and the suitable anisotropic weights, derive available boundary conditions of higher order
derivatives of $\Phi$ on the boundary $\Sigma$ and search for the required weighted Sobolev
interpolations. Finding a suitable multiplier and anisotropic weights are not easy due to the
following reasons: Firstly, to obtain the global existence with no vacuum state requires to
establish the estimates independent of $x_3$ and $\nabla_x^{\alpha}\Phi$ $(0\le |\alpha|\le 4)$
on the boundaries as well as in the
interior of the domain $\Omega$. This leads to strict constraints on
the multiplier and anisotropic weights, as well as makes the computations delicate and
involved. Secondly, as our background solution tends
to vacuum at infinity with different rates for the density and velocity
and their derivatives respectively,
one needs to take some measures to simplify the
coefficients of the nonlinear equation (1.8) so that the procedure to find the
multiplier and anisotropic weights and meanwhile avoid the appearance of vacuum for finite $x_3$ becomes manageable
(one can see more detailed explanations in Remark 2.5 and $\S 5$ below).
Thirdly, the Neumann-type boundary condition (1.10) fulfilled by $\Phi$
arises additional difficulties since there are no enough information
on $\Phi$ itself and its higher order derivatives. Thanks to some delicate analysis
on the radial derivatives and angular derivatives of $\Phi$, which are closely
accompanied by the weighted Sobolev interpolation inequalities in [17], we
finally overcome all these difficulties and obtain a uniform weighted estimate
of $\Phi$ and its higher-order derivatives with no vacuum state for any finite
$x_3>0$ in $\Omega$. This eventually establishes Theorem~1.1.
This paper is organized as follows. In \S2, we derive some basic
estimates on the background solution with vacuum at infinity,
and show some preliminary results regarding the weighted Sobolev interpolation
inequalities. In $\S 3$, we reformulate problem (1.8) together with (1.10)-(1.11) by
decomposing its solution as a sum of the background solution and a
small perturbation $\dot\Phi$ so that its linearization can be studied
in a convenient way. In \S4, we will establish a uniform weighted
energy estimate for the corresponding linear problem, where an
appropriate multiplier is constructed. In $\S 5$, the uniform higher-order weighted estimates of
$\dot\Phi$ are established by rather delicate analysis on the radial derivatives and angular derivatives of
$\dot\Phi$, where the domain composition
techniques are applied in order to obtain the energy estimates of $\dot\Phi$ near $\varphi=0$. In \S6, based on
the results in $\S 5$, we complete the proof of Theorem~1.1 by applying Sobolev's embedding theorem and continuous
induction method.
\vskip 0.5 true cm
\centerline{\bf $\S 2$. Background solutions and some preliminaries}
\vskip 0.5 true cm
In this section, at first we analyze the background solution to
(1.6) with (1.10)-(1.11) when the initial data (1.11) are replaced by
$$(\Phi(1, \theta, \varphi),\quad \partial_r\Phi(1, \theta, \varphi))
=(0, q_0).\eqno{(2.1)}$$
In this case, the density $\rho(x)$ and velocity $u(x)=\nabla_x\Phi(x)$ in $\Omega$
have such forms: $\rho(x)=\hat\rho(r)$, $u(x)=\displaystyle\frac{x}{r}\hat U(r)$. Consequently, the problem (1.6) with (1.10)
and (2.1) is equivalent to
\begin{equation}
\left\{
\begin{aligned}
&(r^2\hat \rho\hat U)'(r)=0,\qquad r\ge 1,\\
&\displaystyle\frac{1}{2}{\hat U}^2(r)+\frac{\gamma}{\gamma-1}{\hat\rho}^{\gamma-1}(r)=1,\qquad r\ge 1,\\
&\hat \rho(1)=\rho_0,\quad \hat U(1)=q_0.
\end{aligned}
\right.\tag{2.2}
\end{equation}
With respect to problem (2.2), we have
{\bf Lemma 2.1.} {\it For $r\ge 1$, (2.2) has a global smooth solution in $\Omega$ which satisfies
\begin{align*}
&\hat \rho(r)=O(r^{-2})>0,\quad c^2(\hat \rho(r))=O(r^{2(1-\gamma)}),\quad \hat U(r)=\sqrt{2}+O(r^{2(1-\gamma)}),\\
&\hat U'(r)=O(r^{1-2\gamma})>0.\tag{2.3}
\end{align*}
Correspondingly, the potential function $\hat\Phi(r)=\int_1^r\hat U(s)ds$.}
{\bf Remark 2.1.} {\it Lemma 2.1 states an interesting physical phenomenon: along the direction of increasing
area, a supersonic flow is expanded and accelerated, meanwhile becomes more and more rarefactive in the divergent
nozzle. This and more physical phenomena on the supersonic or subsonic
flows in divergent or convergent nozzles can be found in Chapter V of [5].}
{\bf Proof.} It follows from the first equation and the initial data in (2.2) that
$$r^2\hat \rho(r)\hat U(r)=\rho_0q_0.\eqno{(2.4)}$$
This, together with the second equation in (2.2), yields
\begin{equation}
\left\{
\begin{aligned}
&\displaystyle\hat \rho'(r)=-\frac{2\rho_0q_0\hat U}{r^3(\hat U^2-c^2(\hat \rho))},\\
&\displaystyle \hat U'(r)=\frac{2\hat Uc^2(\hat \rho)}{r(\hat U^2-c^2(\hat \rho))}.
\end{aligned}
\right.\tag{2.5}
\end{equation}
Thus, $\hat \rho'(r)<0$, $\hat U'(r)>0$ and $(\hat U^2-c^2(\hat \rho))'(r)>0$ hold as long as $\hat U^2-c^2(\hat \rho)>0$
and $\hat U>0$. From this, we can also obtain $(\hat U^2-\hat c^2(\hat \rho))(r)\ge (\hat U^2-c^2(\hat \rho))(1)=q_0^2-c^2(\rho_0)>0$
and $\hat U(r)\ge q_0$. On the other hand, if we set $f_1(\hat \rho, \hat U, r)=r^2\hat \rho \hat U-\rho_0q_0$
and $f_2(\hat \rho, \hat U, r)=\displaystyle\frac{1}{2}\hat U^2+\frac{\gamma}{\gamma-1}\hat \rho^{\gamma-1}-1$, then a direct computation
yields $\displaystyle\frac{\partial(f_1, f_2)}{\partial (\hat \rho, \hat U)}=r^2(\hat U^2-c^2(\hat \rho))\ge q_0^2-c^2(\rho_0)>0$. Thus, $\hat \rho(r)$ and $\hat U(r)$ in (2.2)
exist globally for $r\ge 1$ by implicit function theorem. In addition, (2.3) can be directly obtained by (2.4) and the second
equality in (2.2), and (2.5) respectively. \qquad\qquad \qquad \qquad \qquad \qquad \qquad \qquad
\qquad \qquad \qquad \qquad \quad $\square$
Next, we cite an important weighted Sobolev interpolation inequality in [17], which will be applied to prove some
crucial weighted inequalities listed in Lemma 2.6 below.
{\bf Lemma 2.2. (see [17])} {Suppose $s, \tau, p, \alpha, \beta, q, a$ are real
numbers, and $j\ge0, m>0$ are integers, satisfying
\begin{equation}
\left\{
\begin{aligned}
&\displaystyle p,q\ge1, \frac{j}{m}\le a\le1, s>0,\\
&\displaystyle \frac{1}{s}+\frac{\tau}{n}>0,
\frac{1}{p}+\frac{\alpha}{n}>0, \frac{1}{q}+\frac{\beta}{n}>0,\\
&\displaystyle m-j-\frac{n}{p} \quad \text {is not a nonnegative integer.}
\end{aligned}
\right.\tag{2.6}
\end{equation}
There exists a positive constant $C$ such that the following
inequality holds for all $v\in C_0^{\infty}(\Bbb R^n)$:
$$
\big||x|^{\tau}\nabla_x^jv\big|_{L^s}\le
C\big||x|^{\alpha}\nabla_x^mv\big|_{L^p}^a\big||x|^{\beta}v\big|_{L^q}^{1-a},\eqno{(2.7)}
$$
if and only if the following conditions hold:
$$
\frac{1}{s}+\frac{\tau-j}{n}=a(\frac{1}{p}+\frac{\alpha-m}{n})+(1-a)(\frac{1}{q}+\frac{\beta}{n})\quad\text{with
$\tau\le a\alpha+(1-a)\beta;$}\eqno{(2.8)}
$$
if $\displaystyle\frac{1}{q}+\frac{\beta}{n}=\frac{1}{p}+\frac{\alpha-m}{n}$, then
$$
a(\alpha-m)+(1-a)\beta+j\le\tau; \eqno{(2.9)}
$$
if $\displaystyle a=\frac{j}{m}$, then
$$\tau=a\alpha+(1-a)\beta.\eqno{(2.10)}$$}
{\bf Corollary 2.3.} {\it For the domain $\Omega$ defined in $\S 1.1$, if
$u\in C^m(\bar\Omega)$ and
$$u|_{r\ge T}\equiv 0,\eqno{(2.11)}$$
where $T>1$ is a constant, then we have
(i) (2.7) still holds under the restrictions (2.6)(2.8)-(2.10),
moreover the constant $C$ in the right hand side of (2.7) does not
depend on $T$.
(ii) for $m=2$, $1<\gamma<2$, $\sigma>0$ and $0<\delta<4\gamma$,
\begin{align*}
&|r^{\frac{2\gamma+2\sigma-1}{4}}\nabla_x u|_{L^4(\Omega)}\le C|r^{\frac{2\gamma-1}{2}}\nabla_x^2u|_{L^2(\Omega)}^{\frac{1}{2}}|r^{\sigma}u|_{L^{\infty}(\Omega)}^{\frac{1}{2}},\tag{2.12}\\
&|r^{\frac{4\gamma+1}{4}}\nabla_x u|_{L^4(\Omega)}\le C|r^{\frac{2\gamma+1}{2}}\nabla_x^2u|_{L^2(\Omega)}^{\frac{1}{2}}|r^{\gamma}u|_{L^{\infty}(\Omega)}^{\frac{1}{2}},\tag{2.13}\\
&|r^{\frac{8\gamma-7-\delta}{4}}\nabla_x u|_{L^4(\Omega)}\le
C|r^{\frac{4\gamma-3-\delta}{2}}\nabla_x^2u|_{L^2(\Omega)}^{\frac{1}{2}}|r^{2(\gamma-1)}u|_{L^{\infty}(\Omega)}^{\frac{1}{2}},\tag{2.14}\\
&|r^{\frac{8\gamma-3-\delta}{4}}\nabla_x u|_{L^4(\Omega)}\le C|r^{\frac{4\gamma-1-\delta}{2}}\nabla_x^2u|_{L^2(\Omega)}^{\frac{1}{2}}
|r^{2\gamma-1}u|_{L^{\infty}(\Omega)}^{\frac{1}{2}},\tag{2.15}
\end{align*}
where the generic positive constant $C$ is independent of $T$.}
{\bf Proof.} (i) The proof is completely parallel to that of Lemma 2.2
(one can check the details in [17]),
then we omit it here.
(ii) In (2.6)(2.8)-(2.10) of Lemma 2.2, set $s=4, p=2, q=\infty,
a=\f12$ and $j=1, m=2$, one then concludes that:
(2.12) and (2.13) come from (2.7) and the choices of $\tau=\displaystyle\frac{2\gamma+2\sigma-1}{4}, \alpha=\displaystyle\frac{2\gamma-1}{2}, \beta=\sigma$
and $\tau=\displaystyle\frac{4\gamma+1}{4}, \alpha=\displaystyle\frac{2\gamma+1}{2}, \beta=\gamma$ respectively;
(2.14) and (2.15) are derived from (2.7) by choosing $\tau=\displaystyle\frac{8\gamma-7-\delta}{4}, \alpha=\displaystyle\frac{4\gamma-3-\delta}{2}, \beta=2(\gamma-1)$
and $\tau=\displaystyle\frac{8\gamma-3-\delta}{4}, \alpha=\displaystyle\frac{4\gamma-1-\delta}{2}, \beta=2\gamma-1$ respectively.
\qquad\qquad\qquad\quad $\square$
In order to apply Lemma 2.2 or Corollary 2.3 to derive some weighted Sobolev inequalities in $\Omega$ without
the restriction (2.11), we require
to establish an extension result as follows:
{\bf Lemma 2.4.} {\it Set $D_T=\{(r,\theta,\varphi): 1<r<T, 0\le\theta\le 2\pi, 0\le\varphi<\varphi_0\}$
for $T>1$. If $u(x)\in C^{3}(\bar D_T)$ and $r^{\beta}\partial^{\alpha}_xu\in L^2(D_T)$ $(|\alpha|\le 3)$ with some $\beta\in\Bbb R$,
then there exists an extension $Eu\in C^3(\bar D_{\frac{9}{8}T})$ of $u$ such that $Eu=u$ in $D_T$,
$Eu|_{r\ge \frac{9}{8}T}\equiv 0$ and
$$
|r^{\beta}Eu|_{L^\infty(D_{\frac{9}{8}T})}\le C|r^{\beta}u|_{L^\infty(D_{T})},\quad |r^{\beta}\nabla^{\alpha}_xEu|_{L^2(D_{\frac{9}{8}T})}\le C \displaystyle\sum_{|\nu|\le |\alpha|}|r^{\beta-|\alpha|+|\nu|}\nabla^{\nu}_xu|_{L^2(D_T)},\eqno{(2.16)}
$$
where $C>0$ is independent of $T$.}
{\bf Proof.} In terms of the geometric property of $D_T$, it is convenient to use the spherical coordinate
to work. Denote by $\tilde u(r,\theta,\varphi)=u(r cos\theta sin\varphi , r sin\theta sin\varphi, r cos\varphi)$.
Let $\tilde E$ be an extension operator defined as follows:
\begin{equation*}
(\tilde Eu)(r,\theta,\varphi)=\left\{
\begin{aligned}
&\tilde u(r,\theta,\varphi),\qquad\q\qquad\q\qquad 1\le r\le T,\\
&\displaystyle\sum_{j=1}^4\lambda_j\tilde u(T+j(T-r),\theta,\varphi),\qquad T<r\le \frac{9}{8}T\\
\end{aligned}
\right.
\end{equation*}
where
$\displaystyle\sum_{j=1}^4(-j)^k\lambda_j=1$ for $k=0, 1, 2, 3$.
Noticing that
$$1\le \frac{r}{T+j(T-r)}\le\frac{9}{4}\qquad \text{for $T\le r\le \frac{9}{8}T$ and $0\le j\le 3$},$$
then a direct computation yields
$$|r^{\beta}\tilde Eu|_{L^\infty(D_{\frac{9}{8}T})}\le C|r^{\beta}u|_{L^\infty(D_{T})}
$$
and
\begin{align*}
&|r^{\beta}\nabla_x \tilde Eu|_{L^2(D_{\frac{9}{8}T})}^2\le |r^{\beta}\nabla_x u|_{L^2(D_T)}^2
+|r^{\beta}\nabla_x \tilde Eu|_{L^2(D_{\frac{9}{8}T}\setminus D_T)}^2\\
&\le |r^{\beta}\nabla_x u|_{L^2(D_T)}^2+|r^{\beta}((\partial_r\tilde Eu)^2+\frac{1}{r^2\sin^2\varphi}(\partial_{\theta}\tilde Eu)^2+\frac{1}{r^2}(\partial_{\varphi}\tilde Eu)^2)|_{L^2(D_{\frac{9}{8}T}\setminus D_T)}^2\\
&\le |r^{\beta}\nabla_x u|_{L^2(D_T)}^2+C|r^{\beta}((\partial_ru)^2+\frac{1}{r^2\sin^2\varphi}(\partial_{\theta}u)^2+\frac{1}{r^2}(\partial_{\varphi}u)^2)|_{L^2(D_{T}\setminus D_{\frac{1}{2}T})}^2\\
&\le C|r^{\beta}\nabla_x u|_{L^2(D_T)}^2.
\end{align*}
Analogously, we have for $|\alpha|\le 3$
$$|r^{\beta}\nabla_x^{\alpha}\tilde Eu|_{L^2(D_{\frac{9}{8}T})}\le C |r^{\beta}\nabla_x^{\alpha}u|_{L^2(D_T)}.$$
Choosing a $C^{\infty}-$smooth function $\eta(s)$ with $\eta(s)\equiv 1$
for $s\le 1$ and $\eta(s)\equiv 0$ for $s\ge\displaystyle\f98$ and setting
$$Eu(x)=\eta(\frac{r}{T})\tilde Eu,$$
then $Eu$ satisfies (2.16) and the proof of Lemma 2.4 is completed.\qquad\qquad
\qquad \qquad \quad $\square$
{\bf Remark 2.2.} {\it From Lemma 2.4, we easily know that Corollary 2.3 still holds when the assumption (2.11)
is removed.}
With respect to the $Z-$fileds introduced in (1.7), we have the following properties by direct verifications
as in [15].
{\bf Lemma 2.5.} {\it
\begin{align*}
&(i)\quad [Z_1, Z_2]=Z_3, [Z_2, Z_3]=Z_1, [Z_3,Z_1]=Z_2.\\
&(ii)\quad [Z_i, \partial_r]=0, Z_ir=0.\\
&(iii)\quad \displaystyle\nabla_x f\cdot\nabla_x g=\partial_r f\cdot\partial_r
g+\frac{1}{r^2}\displaystyle\sum_{i=1}^{3}Z_i f\cdot Z_i g\quad \text{for any $C^1$ smooth
functions $f$ and $g$}.\\
&(iv)\quad |Z v|\leq r|\nabla_x v|\quad\text{for any $C^1$ smooth function $v$},
\text{here and below $Z\in\{Z_1, Z_2, Z_3\}$}.\\
&(v)\quad \displaystyle\partial_1=\frac{x_1}{r}\partial_r+\frac{x_2}{r^2}Z_1-\frac{x_3}{r^2}Z_3;
\quad \partial_2=\frac{x_2}{r}\partial_r+\frac{x_3}{r^2}Z_2-\frac{x_1}{r^2}Z_1;\quad
\partial_3=\frac{x_3}{r}\partial_r+\frac{x_1}{r^2}Z_3-\frac{x_2}{r^2}Z_2.
\end{align*}}
{\bf Remark 2.3.} {\it If $u\in C^m(\Bbb R^3)$ with $m\in\Bbb N$, then by Lemma 2.5 we have
$|\nabla_x^mu|\sim |\partial_r^mu|+\displaystyle\frac{|\partial_r^{m-1}Zu|}{r}+\displaystyle\frac{|\partial_r^{m-2}Z^2u|}{r^2}+...+\displaystyle\frac{|Z^mu|}{r^m}$.}
As direct applications of Remark 2.2 and Lemma 2.5, we have the following inequalities
which will be used again and again in $\S 5$ below.
{\bf Lemma 2.6.} {If $1<\gamma<2$, $\sigma\ge\gamma-1$, $0<\delta<4\gamma$, $u(x)\in C^4(\bar D_T)$,
then there exists a generic positive constant $C$ independent of $T$ such that
\begin{align*}
(i)\quad &|r^{\frac{2\gamma+2\sigma-9}{4}}Z^2u|_{L^4(D_T)}\le
C\bigg(\displaystyle\sum_{k=0}^2|r^{\frac{2\gamma-1}{2}-k}\nabla_x^{2-k}(\frac{1}{r}Zu)|_{L^2(D_T)}^{\frac{1}{2}}\bigg)
|r^{\sigma-1}Zu|_{L^{\infty}(D_T)}^{\frac{1}{2}}.
\tag{2.17}\\
(ii)\quad &|r^{\frac{4\gamma-11}{4}}Z^3u|_{L^4(D_T)}\le
C\bigg(\displaystyle\sum_{k=0}^2|r^{\frac{2\gamma+1}{2}-k}\nabla_x^{3-k}(\frac{1}{r}Zu)|_{L^2(D_T)}^{\frac{1}{2}}\bigg)
|r^{\gamma}\nabla_x (\frac{1}{r}Zu)|_{L^{\infty}(D_T)}^{\frac{1}{2}}\\
&\qquad +C\bigg(\displaystyle\sum_{k=0}^2|r^{\frac{2\gamma-1}{2}-k}\nabla_x^{2-k}(\frac{1}{r}Zu)|_{L^2(D_T)}^{\frac{1}{2}}\bigg)
|r^{\sigma-1}Zu|_{L^{\infty}(D_T)}^{\frac{1}{2}}.\tag{2.18}\\
(iii)\quad &|r^{\frac{8\gamma-11-\delta}{4}}\partial_rZ^2u|_{L^4(D_T)}+|r^{\frac{8\gamma-7-\delta}{4}}\partial_r^2Zu|_{L^4(D_T)}\\
&\quad
\le C\bigg(\displaystyle\sum_{k=0}^2|r^{\frac{4\gamma-3-\delta}{2}-k}\nabla_x^{2-k}\partial_rZu|_{L^2(D_T)}^{\frac{1}{2}}\bigg)
|r^{2(\gamma-1)}\partial_rZu|_{L^{\infty}(D_T)}^{\frac{1}{2}}.\tag{2.19}\\
(iv)\quad &|r^{\frac{8\gamma-3-\delta}{4}}\partial_r^3u|_{L^4(D_T)}\le
C\bigg(\displaystyle\sum_{k=0}^2|r^{\frac{4\gamma-1-\delta}{2}-k}\nabla_x^{2-k}\partial_r^2u|_{L^2(D_T)}^{\frac{1}{2}}\bigg)
|r^{2\gamma-1}\partial_r^2u|_{L^{\infty}(D_T)}^{\frac{1}{2}}.\tag{2.20}
\end{align*}}
{\bf Proof.} Let $E$ be the extension operator given in Lemma 2.4, then we have
(i)
\begin{align*}
&|r^{\frac{2\gamma+2\sigma-9}{4}}Z^2u|_{L^4(D_T)}\\
\le & C|r^{\frac{2\gamma+2\sigma-1}{4}}\nabla_x(\frac{1}{r}Zu)|_{L^4(D_T)} \qquad \text{(Applying $\frac{Z^2}{r^2}=\frac{Z}{r}(\frac{Z}{r})$
due to Lemma 2.5 (ii))}\\
\le & C|r^{\frac{2\gamma+2\sigma-1}{4}}\nabla_xE(\frac{1}{r}Zu)|_{L^4(\Omega)}\\
\le & C|r^{\frac{2\gamma-1}{2}}\nabla_x^2E(\frac{1}{r}Zu)|_{L^2(\Omega)}^{\frac{1}{2}}|r^{\sigma}E(\frac{1}{r}Zu)|_{L^{\infty}(\Omega)}^{\frac{1}{2}}
\quad \text{(Applying (2.12) for $E(\frac{1}{r}Zu)$)}\\
\le &
C\bigg(\displaystyle\sum_{k=0}^2|r^{\frac{2\gamma-1}{2}-k}\nabla_x^{2-k}(\frac{1}{r}Zu)|_{L^2(D_T)}^{\frac{1}{2}}\bigg)
|r^{\sigma-1}Zu|_{L^{\infty}(D_T)}^{\frac{1}{2}}. \quad\text{(By Lemma 2.4)}\\
\end{align*}
(ii)
\begin{align*}
&|r^{\frac{4\gamma-11}{4}}Z^3u|_{L^4(D_T)}\\
\le &C|r^{\frac{4\gamma+1}{4}}\nabla_x^2(\frac{1}{r}Zu)|_{L^4(D_T)}+C|r^{\frac{4\gamma-3}{4}}\nabla_x(\frac{1}{r}Zu)|_{L^4(D_T)}\\
\le &C|r^{\frac{4\gamma+1}{4}}\nabla_xE(\nabla_x(\frac{1}{r}Zu))|_{L^4(\Omega)}+ C|r^{\frac{2\gamma+2\sigma-1}{4}}\nabla_x(\frac{1}{r}Zu)|_{L^4(D_T)}
\quad \text{(By $\sigma\ge \gamma-1$)}\\
\le &C|r^{\frac{2\gamma+1}{2}}\nabla_x^2E(\nabla_x(\frac{1}{r}Zu))|_{L^2(\Omega)}^{\frac{1}{2}}|r^{\gamma}E(\nabla_x (\frac{1}{r}Zu))|_{L^{\infty}(\Omega)}^{\frac{1}{2}}\\
&\qquad + C|r^{\frac{2\gamma+2\sigma-1}{4}}\nabla_x(\frac{1}{r}Zu)|_{L^4(D_T)}\qquad\qquad\qquad \text{(Applying (2.13) for $E(\nabla_x(\frac{1}{r}Zu))$)}\\
\le &C\bigg(\displaystyle\sum_{k=0}^2|r^{\frac{2\gamma+1}{2}-k}\nabla_x^{3-k}(\frac{1}{r}Zu)|_{L^2(D_T)}^{\frac{1}{2}}\bigg)
|r^{\gamma}\nabla_x(\frac{1}{r}Zu)|_{L^{\infty}(D_T)}^{\frac{1}{2}}\\
&+C\bigg(\displaystyle\sum_{k=0}^2|r^{\frac{2\gamma-1}{2}-k}\nabla_x^{2-k}(\frac{1}{r}Zu)|_{L^2(D_T)}^{\frac{1}{2}}\bigg)
|r^{\sigma-1}Zu|_{L^{\infty}(D_T)}^{\frac{1}{2}}.\quad\text{(By Lemma 2.4 and (i))}\\
\end{align*}
(iii)
\begin{align*}
&|r^{\frac{8\gamma-11-\delta}{4}}\partial_rZ^2u|_{L^4(D_T)}
+|r^{\frac{8\gamma-7-\delta}{4}}\partial_r^2Zu|_{L^4(D_T)}\\
\le &C |r^{\frac{8\gamma-7-\delta}{4}}\nabla_x(\partial_rZu)|_{L^4(D_T)}\\
\le &C |r^{\frac{8\gamma-7-\delta}{4}}\nabla_xE(\partial_rZu)|_{L^4(\Omega)}\\
\le & C
|r^{\frac{4\gamma-3-\delta}{2}}\nabla_x^2E(\partial_rZu)|_{L^2(\Omega)}^{\frac{1}{2}}|r^{2(\gamma-1)}E(\partial_rZu)|_{L^{\infty}(\Omega)}^{\frac{1}{2}}
\quad \text{(Applying (2.14) for $E(\partial_rZu)$)}\\
\le & C\bigg(\displaystyle\sum_{k=0}^2|r^{\frac{4\gamma-3-\delta}{2}-k}\nabla_x^{2-k}\partial_rZu|_{L^2(D_T)}^{\frac{1}{2}}\bigg)
|r^{2(\gamma-1)}\partial_rZu|_{L^{\infty}(D_T)}^{\frac{1}{2}}.\quad\text{(By Lemma 2.4)}\\
\end{align*}
(iv)
\begin{align*}
&|r^{\frac{8\gamma-3-\delta}{4}}\partial_r^3u|_{L^4(D_T)}\\
\le & C|r^{\frac{8\gamma-3-\delta}{4}}\nabla_x(\partial_r^2u)|_{L^4(D_T)}\\
\le & C|r^{\frac{8\gamma-3-\delta}{4}}\nabla_xE(\partial_r^2u)|_{L^4(\Omega)}\\
\le &
C|r^{\frac{4\gamma-1-\delta}{2}}\nabla_x^2E(\partial_r^2u)|_{L^2(\Omega)}^{\frac{1}{2}}|r^{2\gamma-1}E(\partial_r^2u)|_{L^{\infty}(\Omega)}^{\frac{1}{2}}\quad \text{(Applying (2.15) for $E(\partial_r^2u)$)}\\
\le &
C\bigg(\displaystyle\sum_{k=0}^2|r^{\frac{4\gamma-1-\delta}{2}-k}\nabla_x^{2-k}\partial_r^2u|_{L^2(D_T)}^{\frac{1}{2}}\bigg)
|r^{2\gamma-1}\partial_r^2u|_{L^{\infty}(D_T)}^{\frac{1}{2}}.\quad\text{(By Lemma 2.4)}
\end{align*}
Therefore, we complete the proof of Lemma 2.6.\qquad\qquad\qquad\qquad\qquad\qquad\qquad $\square$
Based on Lemma 2.6, we further have
{\bf Lemma 2.7.} {\it If $1<\gamma<2$, $\sigma=\min\{1, 2(\gamma-1)\}$, $0<\delta<4\gamma$, $u(x)\in C^4(\bar D_T)$, and the following assumptions hold for some constant $M>0$
$$\sum\limits_{0\le
l_1+l_2\le1} r^{l_1}|\partial_r^{l_1}Z^{l_2}\partial_ru|\le M\varepsilon
r^{-2(\gamma-1)},\qquad r^{-1}|Zu|\le M\varepsilon r^{-\sigma},\qquad
r^{-1}|Z^2u|\le
M\varepsilon r^{-(\gamma-1)},\eqno{(2.21)}$$
then
$$
|r^{\frac{2\gamma+2\sigma-9}{4}}Z^2u|_{L^4}\le C(M)\varepsilon^{\frac{1}{2}}\bigg(\ds\sum_{l=0}^2\big(|r^{\frac{4\gamma-7-\delta+2l}{2}}\nabla_x^l\partial_ru|_{L^2}
+|r^{\frac{2\gamma-5+2l}{2}}\nabla_x^l(\frac{1}{r}Zu)|_{L^2}\big)\bigg)^{\frac{1}{2}},\eqno{(2.22)}$$
and
\begin{align*}
&|r^{\frac{4\gamma-11}{4}}Z^3u|_{L^4}
+|r^{\frac{8\gamma-11-\delta}{4}}\partial_rZ^2u|_{L^4}+|r^{\frac{8\gamma-7-\delta}{4}}\partial_r^2Zu|_{L^4}
+|r^{\frac{8\gamma-3-\delta}{4}}\partial_r^3u|_{L^4}\\
&\le C(M)\varepsilon^{\frac{1}{2}}\bigg(\ds\sum_{l=0}^3\big(|r^{\frac{4\gamma-7-\delta+2l}{2}}\nabla_x^l\partial_ru|_{L^2}
+|r^{\frac{2\gamma-5+2l}{2}}\nabla_x^l(\frac{1}{r}Zu)|_{L^2}\big)\bigg)^{\frac{1}{2}},\tag{2.23}
\end{align*}
where $C(M)>0$ is a constant depending on $M$.}
{\bf Remark 2.4.} {\it By $1<\gamma<2$ and $\sigma=\min\{1, 2(\gamma-1)\}$, we can easily conclude
$\sigma\ge\gamma-1$, which means that $\sigma$ satisfies the requirement in Lemma 2.6.}
{\bf Remark 2.5.} {\it (2.21) actually comes from the induction assumptions in Theorem 5.1 on $\dot\Phi$,
where $\dot\Phi$ is the difference between the solution $\Phi$ of (1.8) and the background solution $\hat\Phi$. By (2.21),
we know that $|\partial_ru|\le M\varepsilon r^{-2(\gamma-1)}$ but $|\frac{Zu}{r}|\le M\varepsilon r^{-\sigma}$, and
$|\partial_rZu|\le M\varepsilon r^{-2(\gamma-1)}$ but $|\frac{Z^2u}{r}|\le M\varepsilon r^{-(\gamma-1)}$, which implies that the decay rates
of the radial derivatives and angular derivatives of $u$ are different. Consequently, in order to obtain the
anisotropic energy estimates of $\dot\Phi$ in $\S 5$, we have to pay much attentions on
distinguishing the different roles
of $\partial_r\dot\Phi$ and $Z\dot\Phi$, and this leads to rather involved and delicate analysis.}
{\bf Proof.} In order to prove (2.22)-(2.23), we only verify
$|r^{\frac{2\gamma+2\sigma-9}{4}}Z^2u|_{L^4}$ to satisfy (2.22)
since the terms in the left hand side of (2.23) can be analogously
done.
It follows from Lemma 2.5, the assumptions on $\gamma$ and $\delta$, and a direct computation that
\begin{align*}
&\displaystyle\sum_{k=0}^2|r^{\frac{2\gamma-1}{2}-k}\nabla_x^{2-k}(\frac{1}{r}Zu)|_{L^2(D_T)}\\
&\le C\ds\sum_{l=0}^2\bigg(|r^{\frac{4\gamma-7-\delta+2l}{2}}(\nabla_x^l\partial_ru)|_{L^2(D_T)}
+|r^{\frac{2\gamma-5+2l}{2}}(\nabla_x^l(\frac{1}{r}Zu))|_{L^2(D_T)}\bigg).
\end{align*}
On the other hand, by (2.21) we have
$$|r^{\sigma-1}Zu|_{L^{\infty}(D_T)}\le
M\varepsilon.
$$
Consequently, by Lemma 2.6 (i), we know that (2.22) holds for $|r^{\frac{2\gamma+2\sigma-9}{4}}Z^2u|_{L^4}$,
and then the proof of (2.23) can be completed similarly.\qquad\qquad \qquad \qquad \qquad \qquad \qquad
\qquad $\square$
\vskip 0.5 true cm
\centerline{\bf $\S 3$. Reformulation of the problem (1.8) with (1.10)-(1.11)}
\vskip 0.5 true cm
At first, we state a local solvability result on the problem (1.8) with (1.10)-(1.11).
{\bf Lemma 3.1.} {\it There exists a $T_0>1$ such that the problem
(1.8) with (1.10)-(1.11) possesses a local $C^{\infty}$ solution
$\Phi(r,\varphi)$ in $\Omega_{T_0}=\{(r,\theta,\varphi): 1\le r\le
T_0, 0\le\theta\le 2\pi, 0\le\varphi\le\varphi_0\}$. Moreover,
for any $k\in\Bbb N\cup\{0\}$, there exists a positive constant $C_k$
such that
$$||\Phi(r,\theta,\varphi)-\hat{\Phi}(r)||_{C^k(\Omega_{T_0})}\le C_k\varepsilon,$$
where $\hat\Phi(r)$ is given in Lemma 2.1.}
{\bf Proof.} The quasilinear equation (1.8) is
strictly hyperbolic with respect to the $r-$direction by $\partial_r\Phi>c(\rho)$. Thus, by the standard
Picard iteration as in [19], one can derive that Lemma 3.1 holds.
\qquad \qquad\quad\quad $\square$
Next, we reformulate (1.8) with (1.10)-(1.11).
Let
$\dot{\Phi}=\Phi-\hat{\Phi}$. Then it follows from a direct computation that (1.8) can be reduced to:
$$
\mathcal {L}\dot{\Phi}=\dot
f\quad\text{in $\Omega$},\eqno{(3.1)}$$
where
\begin{equation}
\left\{
\begin{aligned}
&\mathcal {L}\dot\Phi=\partial_r^2\dot\Phi-\displaystyle\frac{P_1(r)}{r^2}\sum\limits_{i=1}^3Z_i^2\dot\Phi+
\displaystyle\frac{P_2(r)}{r}\partial_r\dot\Phi,\\
&\dot f=f_{00}\partial_r^2\dot\Phi+\displaystyle\frac{1}{r^2}\sum\limits_{i,j=1}^3f_{ij}Z_iZ_j\dot\Phi
+\displaystyle\frac{1}{r}\sum\limits_{i=1}^3f_{0i}\partial_rZ_i\dot\Phi+f_0\\
\end{aligned}
\right.
\end{equation}
with
\begin{equation}
\left\{
\begin{aligned}
&P_1(r)=\frac{\displaystyle{c}^2(\hat\rho)}{\displaystyle\hat {U}^2-{c}^2(\hat\rho)},\\
&P_2(r)=\frac{\displaystyle 2}{\displaystyle( \hat {U}^2-{c}^2(\hat\rho))^2}
\big((\gamma-1)\hat{U}^4+{c}^4(\hat\rho)+\hat{U}^2{c}^2(\hat\rho)\big)\\
\end{aligned}
\right.\tag{3.2}
\end{equation}
and
\begin{equation}
\left\{
\begin{aligned}
&f_{00}=-\frac{\displaystyle 1}{\displaystyle \hat{U}^2-
c^2(\hat\rho)}\biggl((\gamma+1)\hat{U}\partial_r\dot\Phi+\displaystyle\frac{\gamma+1}{2}(\partial_r\dot\Phi)^2
+\displaystyle\frac{\gamma-1}{2r^2}\sum\limits_{i=1}^3(Z_i\dot\Phi)^2\biggr),\\
&f_{ii}=\displaystyle\frac{1}{\hat{U}^2-{c}^2(\hat\rho)}\biggl(\frac{\gamma-1}{2}(\partial_r\dot\Phi)^2-(\gamma-1)\hat{U}\partial_r\dot\Phi
+\displaystyle\frac{\gamma-1}{2r^2}\sum\limits_{k=1}^3(Z_k\dot\Phi)^2-\displaystyle\frac{1}{r^2}(Z_i\dot\Phi)^2\biggr),\quad
1\le i\le 3,\\
&f_{ij}=f_{ji}=-\displaystyle\frac{1}{r^2(\hat{U}^2-{c}^2(\hat\rho))}Z_i\dot\Phi
Z_j\dot\Phi,\qquad 1\le i\neq j\le 3,\\
&f_{0i}=-\displaystyle\frac{1}{r(\hat{U}^2-{c}^2(\hat\rho))}(\hat{U}+\partial_r\dot\Phi)Z_i\dot\Phi,\qquad
1\le i\le 3,\\
&f_0=f_0^1+f_0^2,\\
&\quad f_0^1=\displaystyle\frac{1}{\hat{U}^2-{c}^2(\hat\rho)}\biggl\{-\sum\limits_{i,j=1}^3\displaystyle\frac{C_{ij}}{r^3}\partial_r\dot\Phi
Z_i\dot\Phi
Z_j\dot\Phi-\sum\limits_{i,j,k=1}^3\displaystyle\frac{C_{ijk}}{r^4}Z_i\dot\Phi
Z_j\dot\Phi Z_k\dot\Phi\\
&\qquad\qquad
+\displaystyle\frac{2}{r}\bigg(\frac{\gamma-1}{2}(\partial_r\dot\Phi)^3-(\gamma-1)\hat{U}(\partial_r\dot\Phi)^2
+\displaystyle\frac{\gamma-1}{2r^2}\sum\limits_{i=1}^3(Z_i\dot\Phi)^2\partial_r\dot\Phi
+\displaystyle\frac{\gamma-1}{2}\hat{U}(\partial_r\dot\Phi)^2\bigg)\biggr\},\\
&\quad f_0^2=\displaystyle\frac{\hat U}{\hat{U}^2-{c}^2(\hat\rho)}\biggl(-\sum\limits_{i,j=1}^3\displaystyle\frac{C_{ij}}{r^3}Z_i\dot\Phi
Z_j\dot\Phi+\displaystyle\frac{\gamma-1}{r^3}\sum\limits_{i=1}^3(Z_i\dot\Phi)^2\biggr).\\
\end{aligned}
\right.\tag{3.3}
\end{equation}
Here we point out that the terms $f_0^1=
O(\displaystyle\frac{(\partial_r\dot\Phi Z\dot\Phi)^2}{r^3})+O(\displaystyle\frac{(Z\dot\Phi)^3}{r^4})+O(\frac{(\partial_r\dot\Phi)^2}{r})$ and $f_0^2=O(\displaystyle\frac{(Z\dot\Phi)^2}{r^3})$
appeared in $f_0$ will be treated differently
since only such kinds of estimates of $|\partial_r\dot\Phi|\le C\varepsilon r^{-2(\gamma-1)}\to 0$ and $|Z\dot\Phi|\le C\varepsilon r^{1-\sigma}\not\to
0$ as $r\to\infty$ are derived in $\S 5$ (one can the details in Lemma 5.4 and Lemma 5.5 below). In fact,
$f_0^1$ can be easily estimated since it admits better decay rate with respect to large $r$.
On the nozzle wall $\varphi=\varphi_0$, $\dot\Phi$ satisfies
$$Z_2\dot\Phi=Z_3\dot\Phi=0.\eqno{(3.4)}$$
In addition, we have the following initial data of $\dot\Phi$ from (1.11)
$$\dot\Phi(1,\varphi)=\varepsilon\Phi_0(\varphi),\qquad
\partial_r\dot\Phi(1,\varphi)=\varepsilon\Phi_1(\varphi),\eqno{(3.5)}
$$
By using Lemma 2.1 and direct computations, we can obtain the following estimates on the coefficients
of $\mathcal {L}\dot\Phi$ in (3.1):
{\bf Lemma 3.1.}
\begin{align*}
&P_1(r)=O(r^{2(1-\gamma)})>0,\\
&P_2(r)=2(\gamma-1)+O(r^{2(1-\gamma)})>0,\\
&P'_1(r)=-\displaystyle\frac{c^2(\hat\rho(r))}{r(\hat U^2-c^2(\hat\rho(r))}\bigl(2(\gamma-1)\hat U^4+2\hat U^2c^2(\hat\rho(r))\bigr)=O(r^{1-2\gamma})<0,\\
&\displaystyle\frac{P'_1(r)}{P_1}=O(r^{-1})<0,\\
&P'_2(r)=\frac{c^2(\hat{\rho}(r))}{r(\hat{U}^2-c^2(\hat{\rho}(r)))^3}\bigg(\bigl(12(\gamma-1)\hat{U}^4
-8(\gamma-2)\hat{U}^2c^2(\hat{\rho}(r))\bigr)\bigl(\hat{U}^2-c^2(\hat{\rho}(r))\bigr)\\ &\qquad\quad-2(\gamma+1)\hat{U}^2\bigl(2(\gamma-1)\hat{U}^4+2\hat{U}^2c^2(\hat{\rho}(r))+2c^4(\hat{\rho}(r))\bigr)\bigg)=O(r^{1-2\gamma}).
\end{align*}
{\bf Remark 3.1.} {\it From Lemma 3.1, if we take $r$
as the time $t$, then we know that the main part of $\mathcal {L}$
is like the seconder order operator $\partial_t^2-\displaystyle\frac{1}{(1+t)^{2(\gamma-1)}}\Delta+\displaystyle\frac{2(\gamma-1)}{1+t}\partial_t$,
which is strictly hyperbolic but degenerate as
$t\to\infty$. Recently, with respect to the semilinear wave equations with the forms
of $\partial_t^2u-\Delta u+\displaystyle\frac{\mu}{(1+t)^{\alpha}}\partial_tu=f(u)$, where $\mu>0$ and $\alpha>0$ are suitable constants, there
have been extensive
and interesting works on the global existence or blowup results for the different nonlinear function $f(u)$, one can see
[8], [23-24] and the references therein.}
\vskip 0.5 true cm
\centerline{\bf $\S 4$. A first-order weighted energy estimate}
\vskip 0.5 true cm
In this section, we establish a weighted energy estimate of
$\nabla_x\dot\Phi$ for the linear part of (3.1)
together with (3.4)-(3.5),
which will play a fundamental role in our subsequent analysis.
Set $D_T=\{(r,\theta,\varphi): 1<r<T, 0\le\theta\le 2\pi, 0\le\varphi<\varphi_0\}$
for any $T>1$, $B_T=\{(r,\theta,\varphi)\in \Sigma: 1<r<T,
0\le\theta\le2\pi, \varphi=\varphi_0\}$, and $S_T=\bar\Omega\cap\{r=T\}$.
{\bf Theorem 4.1.} {\it Let $\dot\Phi\in C^{2}(\bar D_T)$ satisfy the boundary
condition (3.4) and initial data condition (3.5). Then there exists a multiplier
$\mathcal {M}\dot\Phi=r^{\mu}a(r)\partial_r\dot\Phi$ such that for fixed
constant $\mu=4\gamma-6$ we have
\begin{align*}
&T^{\mu}\int_{S_T}(\partial_r\dot\Phi)^2dS+T^{\mu-2\gamma}\int_{S_T}(Z\dot\Phi)^2dS
+C\int_{D_T}\big(r^{\mu-1-\delta}(\partial_r\dot\Phi)^2+r^{\mu-1-2\gamma}(Z\dot\Phi)^2\big)dx\\
&\le \int_{D_T}\mathcal {L}\dot\Phi\cdot\mathcal {M}\dot\Phi
dx
+C\varepsilon^2,\tag{4.1}
\end{align*}
where $(Z\dot\Phi)^2=\ds\sum_{k=1}^3(Z_k\dot\Phi)^2$, $C>0$ is a generic positive constant, and $\delta>0$ is a fixed constant.}
{\bf Remark 4.1.} {\it Here we emphasize that the choice of $\mu=4\gamma-6$ in (4.1) is very necessary due to the following two reasons:
First, to guarantee the positivity of (4.3) below, one should let $\mu\le 4\gamma-6$; Second, by the Bernoulli's law (1.2),
we have $c^2(\rho)=
c^2(\hat\rho)-\frac{\gamma-1}{2}\big((\partial_r\dot\Phi)^2+2\hat U\partial_r\dot\Phi+\frac{1}{r^2}(Z\dot\Phi)^2\big)$.
Notice that only the estimate of $|\nabla_x\dot\Phi|\le
C\varepsilon r^{-\frac{\mu}{2}-1}$ can be obtained by the analysis in $\S 5-\S 6$. On the other hand, $c^2(\hat\rho(r))\ge Cr^{-2(\gamma-1)}$ and $\hat U=O(1)$ hold
by Lemma 2.1. Therefore, in order to guarantee the absence of vacuum for any finite $r$ in $\Omega$, we require to choose
the constant $\mu$ such that $-\displaystyle\frac{\mu}{2}-1\le
-2(\gamma-1)$, which leads to $\mu\ge 4\gamma-6.$ Combining these two reasons yields $\mu=4\gamma-6$.}
{\bf Remark 4.2.} {\it In Theorem 4.1, it suffices to choose the constant $\delta>0$. However, to derive the higher order energy estimates
of $\dot\Phi$, we require to give more restrictions on $\delta$ (one can see Theorem 5.1 in $\S 5$).}
{\bf Proof.} It follows from the integration by parts and (3.4)-(3.5) that
\begin{align*}
&\int_{D_T}\mathcal {L}\dot\Phi\cdot\mathcal {M}\dot\Phi
dx\\
&=
\int_{S_T}\frac{1}{2}r^{\mu}a(r)(\partial_r\dot\Phi)^2dS+\int_{S_T}\frac{1}{2}r^{\mu-2}
P_1a(r)(Z\dot\Phi)^2dS\\
&\quad -\bigg(\int_{S_1}\frac{1}{2}r^{\mu}a(r)(\partial_r\dot\Phi)^2dS+\int_{S_1}\frac{1}{2}r^{\mu-2}
P_1a(r)(Z\dot\Phi)^2dS\bigg)\\
&\quad +\int_{D_T}\bigg(r^{\mu-1}\big((P_2-\frac{\mu+2}{2})a(r)-\frac{1}{2}ra'(r)\big)(\partial_r\dot\Phi)^2
-\frac{1}{2}r^{\mu-3}\big((\mu P_1+rP'_1)a(r)+ra'(r)P_1\big)
(Z\dot\Phi)^2\bigg)dx\\
&\quad +\int_{B_T}r^{\mu}a(r)P_1(-x_2Z_2\dot\Phi+x_1Z_3\dot\Phi)\partial_r\dot\Phi\\
&\ge \int_{S_T}\frac{1}{2}r^{\mu}a(r)(\partial_r\dot\Phi)^2dS+\int_{S_T}\frac{1}{2}r^{\mu-2}
P_1a(r)(Z\dot\Phi)^2dS\\
&\quad +\int_{D_T}\bigg(r^{\mu-1}\big((P_2-\frac{\mu+2}{2})a(r)-\frac{1}{2}ra'(r)\big)(\partial_r\dot\Phi)^2
-\frac{1}{2}r^{\mu-3}\big((\mu P_1+rP'_1)a(r)+ra'(r)P_1\big)
(Z\dot\Phi)^2\bigg)dx\\
&\quad -C\varepsilon^2\tag{4.2}
\end{align*}
It is noted that
\begin{align*}
&(P_2-\frac{\mu+2}{2})a(r)-\frac{1}{2}ra'(r)\\
&=\frac{1}{2(\hat{U}^2-{c}^2(\hat\rho))}\big((4\gamma-6-\mu)\hat{U}^4+(4+2(\mu+2))\hat{U}^2{c}^2(\hat\rho)
+(2-\mu){c}^4((\hat\rho))\big)a(r)-\frac{1}{2}ra'(r),\tag{4.3}
\end{align*}
then in order to guarantee the positivity of (4.3) for $\mu=4\gamma-6$, we require
$$a(r)>0\quad \text{and $a'(r)<0$}.\eqno{(4.4)}$$
For this end, we choose
$$a(r)=1+r^{-\delta}\quad\text{with $\delta>0$.}$$
In this case, one can arrive at
$$
(P_2-\frac{\mu+2}{2})a(r)-\frac{1}{2}ra'(r)>\frac{1}{2}\delta r^{-\delta}.\eqno{(4.5)}
$$
On the other hand, it follows from a direct computation and the assumption of $1<\gamma<2$ that
\begin{align*}
&-(\mu P_1+r\partial_rP_1)a(r)-ra'(r)P_1\\
&=\frac{c^2(\hat\rho)}{({\hat U}^2-{c}^2(\hat\rho))^3}\biggl((2(\gamma-1)-\mu){\hat U}^4
+(4+2\mu){\hat U}^2{c}^2(\hat\rho)
-\mu{c}^4(\hat\rho)\biggr)a(r)+\frac{\delta {c}^2(\hat\rho)r^{-\delta}}{{\hat U}^2-{c}^2(\hat\rho)}\\
&>\frac{2{c}^2(\hat\rho)}{({\hat U}^2-{c}^2(\hat\rho))^3}(2-\gamma){\hat U}^4\\
&>Cr^{-2(\gamma-1)}.\tag{4.6}
\end{align*}
Thus, substituting (4.5)-(4.6) into (4.2) yields Theorem 4.1. \qquad \qquad \qquad \qquad \qquad
$\square$
\vskip 0.5 true cm
\centerline{\bf $\S 5$. Higher-order weighted energy estimates of $\dot\Phi$}
\vskip 0.5 true cm
In this section, we will derive the higher-order energy
estimates of solution $\dot\Phi$ to (3.1) with (3.4)-(3.5) so that the suitable decay properties of
$\nabla_x\dot\Phi$ can be obtained and the density $\rho(x)>0$ can be also derived in subsequent $\S 6$.
Due to the Neumann boundary condition (3.4),
the asymptotic degeneracy of some coefficients in (3.1), and the different decay rates of $\partial_r\dot\Phi$ and
$\displaystyle\frac{Z\dot\Phi}{r}$, the related derivation procedure will become rather complicated
and technical.
{\bf Theorem 5.1.} {\it Let $\dot\Phi\in C^4(\bar{D}_T)$ be the solution to
(3.1) with (3.4)-(3.5), and further assume
$$
\sum\limits_{0\le l_1+l_2\le 1}
r^{l_1}|\partial_r^{l_1}Z^{l_2}\partial_r\dot\Phi|\le M\varepsilon r^{-2(\gamma-1)},\quad
r^{-1}|Z\dot\Phi|\le M\varepsilon r^{-\sigma},\quad
r^{-1}|Z^2\dot\Phi|\le M\varepsilon r^{-(\gamma-1)},\eqno{(5.1)}
$$
where $M>0$ is a constant, and $\sigma=\min\{1, 2(\gamma-1)\}$. Then for sufficiently small $\varepsilon>0$ and
$0\le k\le 3$, we have
\begin{align*}
&T^{\mu+2k}\int_{S_T}|\nabla_x^k\partial_r\dot\Phi|^2dS
+T^{\mu-2\gamma+2k}\int_{S_T}|\nabla_x^k
(\frac{1}{r}Z\dot\Phi)|^2dS\\
&\quad +\int_{D_T}\biggl(r^{\mu-1-\delta+2k}|\nabla_x^k\partial_r\dot\Phi|^2+r^{\mu+1-2\gamma+2k}
|\nabla_x^k(\frac{1}{r}Z\dot\Phi)|^2\biggr)dx\\
&\le
C\varepsilon^2,\tag{5.2}
\end{align*}
where $\mu=4\gamma-6$, $0<\delta\le\min\{\gamma-1, \sigma-(\gamma-1)\}$, and the domains $D_T, B_T, S_T$ have been defined
in the beginning of $\S 4$.}
In order to prove Theorem 5.1, we will apply the induction method on $k$ in (5.2) to establish
the following estimates respectively:
(i) $\partial_rS^k\dot\Phi$ and $ZS^k\dot\Phi$ with $S=r\partial_r$ and $1\le k\le 3$ (in this case, all the radial derivatives of $\nabla_x\dot\Phi$
up to third order are treated);
(ii) $\partial_rZ\dot\Phi$ and $Z^2\dot\Phi$ (in this case, together with the case $k=1$ in (i),
all the second order derivatives $\nabla_x^2\dot\Phi$ are treated);
(iii) $\partial_rSZ\dot\Phi$, $ZSZ\dot\Phi$, $\partial_rZ^2\dot\Phi$ and $Z^3\dot\Phi$ (in this case, together with the case $k=2$ in (i),
all the third order derivatives $\nabla_x^3\dot\Phi$ are treated);
(iv) $\partial_rS^2Z\dot\Phi$, $ZS^2Z\dot\Phi$, $\partial_rSZ^2\dot\Phi$, $ZSZ^2\dot\Phi$, $\partial_rZ^3\dot\Phi$ and $Z^4\dot\Phi$ (in this case,
together with the case $k=3$ in (i),
all the fourth order derivatives $\nabla_x^4\dot\Phi$ are treated).
These estimates will be given in Lemma 5.2-Lemma 5.5 respectively.
At first, we establish the radial derivative estimates of $\dot\Phi$ under the suitable induction
assumption. Set $S=r\partial_r$, which is tangent to
fixed nozzle wall $\Sigma$, then we have
{\bf Lemma 5.2. (Radial derivative estimates)} {Under the assumptions of Theorem 5.1, if (5.2)
holds for $0\le l\le m-1$ with $1\le m\le 3$,
then
\begin{align*}
&T^{\mu}\int_{S_T}(\partial_r S^m\dot\Phi)^2dS+T^{\mu-2\gamma}\int_{S_T}(ZS^m\dot\Phi)^2dS
+\int_{D_T}\bigg(r^{\mu-1-\delta}(\partial_r S^m\dot\Phi)^2+r^{\mu-1-2\gamma}(Z S^m\dot\Phi)^2\bigg)dx\\
&\le C\varepsilon^2+ C\varepsilon\int_{D_T}
\displaystyle\sum_{l=0}^m\bigg(r^{\mu+1-2\gamma+2l}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2+r^{\mu-1-\delta+2l}(\nabla_x^l\partial_r\dot\Phi)^2\bigg)
dx\\
&\quad +C\varepsilon\ds\sum_{l=0}^m\bigg(T^{\mu+2l}\int_{S_T}(\nabla_x^l\partial_r\dot\Phi)^2dS+T^{\mu-2\gamma+2l}\int_{S_T}
(\nabla_x^l(Z\dot\Phi))^2dS\bigg)\\
&\quad +C\varepsilon\bigg(\int_{D_T}r^{\mu-1-\delta}(\partial_r S^m\dot\Phi)^2+r^{\mu-1-2\gamma}(Z S^m\dot\Phi)^2dx\bigg)^{\f12},\tag{5.3}
\end{align*}
where $0<\delta\le\gamma-1$.
Especially, for $m=0$, the following estimate holds
\begin{align*}
&T^{\mu}\int_{S_T}(\partial_r\dot\Phi)^2dS+T^{\mu-2\gamma}\int_{S_T}(Z\dot\Phi)^2dS
+\int_{D_T}\bigg(r^{\mu-1-\delta}(\partial_r \dot\Phi)^2+r^{\mu-1-2\gamma}(Z \dot\Phi)^2\bigg)dx\\
&\le C\varepsilon^2+ C\varepsilon\int_{D_T}
\bigg(r^{\mu+1-2\gamma}(\frac{1}{r}Z\dot\Phi)^2+r^{\mu-1-\delta}(\partial_r\dot\Phi)^2\bigg)
dx\\
&\quad +C\varepsilon\bigg(T^{\mu}\int_{S_T}(\partial_r\dot\Phi)^2dS+T^{\mu-2\gamma}\int_{S_T}
(Z\dot\Phi)^2dS\bigg).\tag{5.4}
\end{align*}}
{\bf Remark 5.1.} {\it For the case of $m=0$ in (5.4), we do not require any induction assumption.}
{\bf Remark 5.2.} {\it It is noted that the angular derivatives of $\dot\Phi$
are still included in the right hand side of (5.3), which implies that we have not
obtained the complete estimates on the radial derivative estimates of $\dot\Phi$. However,
since the coefficients of angular derivatives of $\dot\Phi$ in (5.3) are small, then together
with the subsequent angular derivative estimates, we can derive (5.2).}
{\bf Proof.} Noticing that on $\Sigma$
$$S^m Z_2\dot\Phi=S^m Z_3\dot\Phi=0.\eqno{(5.5)}
$$
This, together with Theorem 4.1 and (3.5), yields
\begin{align*}
&T^{\mu}\int_{S_T}(\partial_rS^m\dot\Phi)^2dS+T^{\mu-2\gamma}\int_{S_T}(ZS^m\dot\Phi)^2dS
+\int_{D_T}\biggl(r^{\mu-1-\delta}(\partial_rS^m\dot\Phi)^2+r^{\mu-1-2\gamma}(ZS^m\dot\Phi)^2\biggr)dx\\
&\le C\int_{D_T}\mathcal {L}S^m\dot\Phi\cdot\mathcal {M}S^m\dot\Phi
dx+C\varepsilon^2.\tag{5.6}
\end{align*}
Next, we derive an explicit representation of
$\mathcal {L}S^m\dot\Phi$ for the later uses.
By a direct computation, we have
$$
\mathcal {L} S=S\mathcal {L}+2\mathcal {L}+\displaystyle\frac{P'_1}{r}\sum\limits_{i=1}^3Z_i^2+A_1,
\eqno{(5.7)}
$$
where $A_1=-P'_2\partial_r$ is a first order operator.
By induction, for $1\le m\le3$, we further arrive at
$$
\mathcal {L}S^m=S^m\mathcal {L}+mS^{m-1}\bigl(\displaystyle\frac{P'_1}{r}\sum\limits_{i=1}^3Z_i^2\big)
+\sum\limits_{0\le l\le m-1}C_{lm}S^{l}\mathcal{L}\dot\Phi+A_m,\eqno{(5.8)}
$$
where $C_{lm}$ are some suitable constants, $A_m$ stands for a lower order differential operator whose order is less
than $m$. For examples,
\begin{align*}
&A_2=\big(\frac{2P'_1}{r}-r(\frac{P'_1}{r})'\big)\sum\limits_{i=1}^3Z_i^2
+SA_1+A_1S+A_1,\\
&A_3=SA_2+[\frac{P'_1}{r}\sum\limits_{i=1}^3Z_i^2, S^2]
+2S(\frac{P'_1}{r}\sum\limits_{i=1}^3Z_i^2)+A_2+A_1S^2,
\end{align*}
here and below, $[\cdot , \cdot]$ denotes the usual commutator.
For convenient treatments, for $1\le m\le3$, we rewrite (5.8) as
$$\mathcal {L}S^m=S^m\mathcal {L}+B_{1m}+B_{2m}\eqno{(5.9)}$$
with
\begin{align*}
&B_{1m}=\displaystyle\sum_{0\le l\le
m-1}C_{lm} S^{l}\mathcal{L}+ \displaystyle\sum_{0\le l\le
m-2}C_{lm}S^{m-1-l}\big(\frac{rP'_1}{P_1}\big)S^l\big(\frac{P_1}{r^2}\ds\sum_{i=1}^3Z_i^2\big)+A_m,\\
&B_{2m}=\frac{mrP'_1}{P_1}S^{m-1}\big(\frac{P_1}{r^2}\displaystyle\sum_{i=1}^3Z_i^2\big),
\end{align*}
where $B_{2m}\dot\Phi$ contains the $(m+1)-$th order (the highest order) derivatives of $\dot\Phi$,
but $B_{1m}\dot\Phi$ only includes $\nabla_x^{\alpha}\dot\Phi$ with $|\alpha|\le m$ (the lower order derivatives of $\dot\Phi$)
and $\nabla_x^{m+1}\dot\Phi$ with small coefficients.
In addition, from the equation (3.1), for $0\le m\le3$, we have
$$
S^m\mathcal{L}\dot\Phi=I_{1}^m+I_{2}^m+I_{3}^m,\eqno{(5.10)}
$$
where
\begin{align*}
&I_{1}^m=f_{00}\partial_r^2S^m\dot\Phi+\frac{1}{r^2}\sum\limits_{1\le
i,j\le 3}f_{ij}Z_iZ_jS^m\dot\Phi
+\frac{1}{r}\sum\limits_{i=1}^3f_{0i}\partial_rZ_iS^m\dot\Phi,\\
&I_{2}^m=f_{00}[S^m, \partial_r^2]\dot\Phi
+\frac{1}{r}\sum\limits_{i=1}^3f_{0i}[S^m, \partial_rZ_i]\dot\Phi,\\
&I_{3}^m=\sum\limits_{0\le l\le
m}C_{lm}\big\{\sum\limits_{l_1+l_2=
l, l_1\ge1}\tilde C_{l_1l_2}\big(S^{l_1}f_{00}S^{l_2}\partial_r^2\dot\Phi
+S^{l_1}(\frac{1}{r^2}\sum\limits_{1\le i,j\le3}f_{ij})S^{l_2}Z_iZ_j\dot\Phi\\
&\qquad\quad +S^{l_1}(\frac{1}{r}\sum\limits_{i=1}^3f_{0i})S^{l_2}\partial_rZ_i\dot\Phi\big)\big\}+S^mf_0.\\
\end{align*}
Based on the preparations above, we now treat $\int_{D_T}\mathcal {L}S^m\dot\Phi\cdot\mathcal{M}S^m\dot\Phi$
in the right hand side of (5.6).
This procedure is divided into the following
five parts.
\vskip 0.3 true cm
{\bf Part 1. The estimate of $\int_{D_T}I_{1}^m\cdot\mathcal{M}S^m\Phi dx$}
\vskip 0.3 true cm
Notice that we have for $C^1-$smooth functions $g_i$ ($1\le i\le 3$)
$$
\sum\limits_{i=1}^3Z_ig_i
=\partial_1(x_3g_3-x_2g_1)+\partial_2(x_1g_1-x_3g_2)+\partial_3(x_2g_2-x_1g_3)\eqno{(5.11)}
$$
and
$$
\int_{D_T}\sum\limits_{i=1}^3Z_ig_idx=-\int_{B_T}\frac{1}{\sin\varphi_0}(x_2g_2-x_1g_3)dS.\eqno{(5.12)}
$$
In addition, a direct computation yields for $m\le3$
\begin{align*}
I_{1}^m&\cdot \mathcal{M}S^m\dot\Phi\\
=&\partial_r\bigg(\frac{1}{2}r^{\mu}a(r)f_{00}(\partial_rS^m\dot\Phi)^2-r^{\mu-2}a(r)\displaystyle\sum_{1\le i<j\le3}f_{ij}Z_iS^m\dot\Phi Z_jS^m\dot\Phi
-\frac{1}{2}r^{\mu-2}a(r)\displaystyle\sum_{i=1}^3f_{ii}(Z_iS^m\dot\Phi)^2\bigg)\\
&+\displaystyle\sum_{i=1}^3Z_i\bigg(\frac{1}{2}r^{\mu-1}a(r)f_{0i}(\partial_rS^m\dot\Phi)^2+r^{\mu-2}a(r)\partial_rS^m\dot\Phi\ds\sum_{j=1}^3f_{ij}Z_jS^m\dot\Phi\bigg)\\
&-\frac{1}{2}\partial_r(r^\mu a(r)f_{00})(\partial_rS^m\dot\Phi)^2-\frac{1}{2}r^{\mu-1}a(r)(\partial_rS^m\dot\Phi)^2\displaystyle\sum_{i=1}^3Z_if_{0i}
+\displaystyle\sum_{i=1}^3\partial_r(\frac{1}{2}r^{\mu-2}a(r)f_{ii})(Z_iS^m\dot\Phi)^2\\
&+\displaystyle\sum_{1\le i<j\le3}\partial_r(r^{\mu-2}a(r)f_{ij})Z_iS^m\dot\Phi Z_jS^m\dot\Phi.\tag{5.13}
\end{align*}
On the other hand, by the expressions of $f_{ij}, f_{0i}$ and (5.5), a crucial observation yields on $B_T$
\begin{align*}
&x_2\bigg(\frac{1}{2}r^{\mu-1}a(r)f_{02}(\partial_rS^m\dot\Phi)^2
+r^{\mu-2}a(r)\partial_rS^m\dot\Phi\displaystyle\sum_{j=1}^3f_{2j}Z_jS^m\dot\Phi\bigg)\\
&\quad -x_1\bigg(\frac{1}{2}r^{\mu-1}a(r)f_{03}(\partial_rS^m\dot\Phi)^2
+r^{\mu-2}a(r)\partial_rS^m\dot\Phi\displaystyle\sum_{j=1}^3f_{3j}Z_jS^m\dot\Phi\bigg)\\
&=\frac{1}{2}a(r)(\partial_rS^m\dot\Phi)^2(x_2f_{02}-x_1f_{03})
+r^{\mu-2}a(r)\partial_rS^m\dot\Phi\displaystyle\sum_{j=1}^3(x_2f_{2j}-x_1f_{3j})Z_jS^m\dot\Phi\\
&=\frac{1}{\hat{U}^2-{c}^2(\hat\rho)}r^{\mu-2}a(r)\partial_rS^m\dot\Phi\bigg(\frac{\gamma-1}{2}(\partial_r\dot\Phi)^2
-(\gamma-1)\hat{U}\partial_r\dot\Phi+\frac{\gamma-1}{2r^2}\displaystyle\sum_{k=1}^3(Z_k\dot\Phi)^2\bigg)
(x_2Z_2S^m\dot\Phi\\
&\quad -x_1Z_3S^m\dot\Phi)-\frac{1}{\hat{U}^2-{c}^2(\hat\rho)}r^{\mu-4}a(r)\partial_rS^m\dot\Phi\ds\sum_{j=1}^3Z_j\dot\Phi Z_jS^m\dot\Phi(x_2Z_2\dot\Phi-x_1Z_3\dot\Phi)\\
&\quad -\frac{1}{2}r^{\mu-2}a(r)(\hat{U}+\partial_r\dot\Phi)(\partial_rS^m\dot\Phi)^2(x_2Z_2\dot\Phi-x_1Z_3\dot\Phi)\\
&=0.\tag{5.14}
\end{align*}
Thus, by (5.13) together with (5.12) and (5.14), it follows from an integration by parts
and simultaneously notices
the expressions of $f_i$ and the assumption (5.1) that
\begin{align*}
&|\int_{D_T}I_{1}^m\cdot\mathcal{M}S^m\dot\Phi dx|\\
&\le C\varepsilon^2+C\varepsilon\bigg(T^{\mu}\int_{S_T}(\partial_rS^m\dot\Phi)^2dS
+T^{\mu-2\gamma}\int_{S_T}(ZS^m\dot\Phi)^2dS\\
&\quad+\int_{D_T}\big(r^{\mu-1-\delta}(\partial_rS^m\dot\Phi)^2
+r^{\mu-1-2\gamma}(ZS^m\dot\Phi)^2\big)dx\bigg),\tag{5.15}
\end{align*}
here we have used some facts such as
$$|r^{\mu-1}a(r)(\partial_rS^m\dot\Phi)^2\displaystyle\sum_{i=1}^3Z_if_{0i}|\le C\varepsilon r^{\mu-1-(\gamma-1)}(\partial_rS^m\dot\Phi)^2\le C\varepsilon r^{\mu-1-\delta}(\partial_rS^m\dot\Phi)^2\quad \text{for $0<\delta\le \gamma-1$}.$$
\vskip 0.3 true cm
{\bf Part 2. The estimate of $\int_{D_T}I_{2}^m\cdot\mathcal{M}S^m\Phi dx$}
\vskip 0.3 true cm
It follows from the expressions
of $f_i$, Lemma 2.5 (ii) and (5.1) that
$$|I_{2}^m|\le C\varepsilon
\big(r^{-\sigma-1}\ds\sum_{0\le l\le
m-1}|S^l\partial_rZ\dot\Phi|+r^{-2(\gamma-1)}\ds\sum_{0\le l\le
m-1}|S^l\partial_r^2\dot\Phi|\big),$$
which derives that
$$\int_{D_T}|I_{2}^m\cdot\mathcal{M}S^m\dot\Phi|dx
\le C\varepsilon\int_{D_T}\bigg(\displaystyle\sum_{0\le l\le
m-1}r^{\mu-1-\delta}(S^l\partial_rZ\dot\Phi)^2 +\displaystyle\sum_{1\le l\le
m}r^{\mu-1-\delta}(S^l\partial_r\dot\Phi)^2\bigg)dx.\eqno{(5.16)}
$$
\vskip 0.3 true cm
{\bf Part 3. The estimate of $\int_{D_T}I_{3}^m\cdot\mathcal{M}S^m\Phi dx$}
\vskip 0.3 true cm
At first, we treat the case of $\int_{D_T}I_{3}^m\cdot\mathcal{M}S^m\Phi dx$ with $m\le 2$.
For $m\le 2$, as in Part 2 it follows from the expressions
of $f_i$ and the assumption (5.1) that
$$ |I_{3}^m|\le C\varepsilon \bigg(\displaystyle\sum_{0\le
l\le m}\big(r^{-\gamma}|S^l\partial_r\dot\Phi|+r^{-\sigma-2}|S^lZ\dot\Phi|\big)+\displaystyle\sum_{0\le l\le m-1}
r^{-2\gamma}|S^lZ^2\dot\Phi|\bigg),$$
which derives for $m\le 2$
$$
\int_{D_T}|I_{3}^m\cdot\mathcal{M}S^m\dot\Phi|dx
\le C\varepsilon\int_{D_T}
\displaystyle\sum_{l=0}^m\bigg(r^{\mu+1-2\gamma+2l}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2+r^{\mu-1-\delta+2l}(\nabla_x^l\partial_r\dot\Phi)^2\bigg)
dx.\eqno{(5.17)}$$
Next we deal with $\int_{D_T}I_{3}^3\cdot\mathcal{M}S^3\dot\Phi dx$.
It is noted that the most troublesome terms in $I_{3}^3$ are the ones which include the products of third order derivatives
of $\dot\Phi$ since there are no related weighted $L^{\infty}$ estimates in (5.1). For the convenient treatments, we decompose $I_{3}^3$ into $J_1$ and $J_2$
by using $S^2=r\partial_r+r^2\partial_r^2$, where only $J_2$ contains the product terms of third order derivatives
of $\dot\Phi$. Namely,
$$I_{3}^3=J_1+J_2\eqno{(5.18)}$$
with
\begin{align*}
&J_1=\\
&\sum\limits_{0\le l\le
2}C_{l2}\sum\limits_{l_1+l_2=
l, l_1\ge1}\tilde C_{l_1l_2}\bigg(S^{l_1}f_{00}S^{l_2}\partial_r^2\dot\Phi
+S^{l_1}(\frac{1}{r^2}\sum\limits_{1\le i,j\le3}f_{ij})S^{l_2}Z_iZ_j\dot\Phi+S^{l_1}(\frac{1}{r}\sum\limits_{i=1}^3f_{0i})S^{l_2}\partial_rZ_i\dot\Phi\bigg)\\
&\quad +C_{33}\sum\limits_{(l_1,l_2)\not=(2, 1)}\tilde C_{l_1l_2}\bigg(S^{l_1}f_{00}S^{l_2}\partial_r^2\dot\Phi
+S^{l_1}(\frac{1}{r^2}\sum\limits_{1\le i,j\le3}f_{ij})S^{l_2}Z_iZ_j\dot\Phi+S^{l_1}(\frac{1}{r}\sum\limits_{i=1}^3f_{0i})S^{l_2}\partial_rZ_i\dot\Phi\bigg)\\
&\quad +C_{33}\tilde C_{21}\bigg(r\partial_rf_{00}S\partial_r^2\dot\Phi+r\partial_r(\frac{1}{r^2}\ds\sum_{i,j=1}^3f_{ij})SZ_iZ_j\dot\Phi
+r\partial_r(\frac{1}{r}\ds\sum_{i=1}^3f_{0i})S\partial_rZ_i\dot\Phi\bigg)+S^3f_0
\end{align*}
and
$$J_2=C_{33}\tilde C_{21}\bigg(r^2\partial_r^2f_{00}S\partial_r^2\dot\Phi+r^2\partial_r^2(\frac{1}{r^2}\displaystyle\sum_{i,j=1}^3f_{ij})SZ_iZ_j\dot\Phi
+r^2\partial_r^2(\frac{1}{r}\displaystyle\sum_{i=1}^3f_{0i})S\partial_rZ_i\dot\Phi\bigg).$$
By the assumption (5.1) and the expressions of $f_{ij}, f_0$, then a direct computation yields
$$|J_1|\le C\varepsilon \bigg\{\displaystyle\sum_{0\le
l\le 3}\big(r^{-\gamma}|S^l\partial_r\dot\Phi|+r^{-\sigma-2}|S^lZ\dot\Phi|\big)+\displaystyle\sum_{0\le l\le2}r^{-2\gamma}|S^lZ^2\dot\Phi|\biggr\}.\eqno{(5.19)}$$
Next, by the expressions of $f_{ij}$ we continue to decompose $J_2$ as $J_2=J_{21}+J_{22}$ so that only
$J_{22}$ contains the product terms of third order derivatives
of $\dot\Phi$. More concretely,
$$J_2=J_{21}+J_{22}$$
with
\begin{align*}
J_{22}&=-\frac{r^2}{{\hat U}^2-c^2(\hat\rho)}\bigg\{\bigg((\gamma+1)\big(\hat {U}+\partial_r\dot\Phi\big)\partial_r^3\dot\Phi+\frac{\gamma-1}{r^2}
\displaystyle\sum_{i=1}^3Z_i\dot\Phi\partial_r^2Z_i\dot\Phi\bigg)S\partial_r^2\dot\Phi\\
&-\frac{\gamma-1}{r^2}\displaystyle\sum_{i=1}^3\bigg(\partial_r\dot\Phi\partial_r^3\dot\Phi-\hat{U}\partial_r^3\dot\Phi+\frac{1}{r^2}\displaystyle\sum_{k=1}^3Z_k\dot\Phi\partial_r^2Z_k\dot\Phi
-\frac{1}{r^2}Z_i\dot\Phi\partial_r^2Z_i\dot\Phi\bigg)SZ_i^2\dot\Phi\\
&+\frac{1}{r^4}\displaystyle\sum_{1\le i\not=j\le 3}\big(Z_i\dot\Phi\partial_r^2Z_j\dot\Phi+Z_j\dot\Phi\partial_r^2Z_i\dot\Phi\big)SZ_iZ_j\dot\Phi +\frac{1}{r^2}\displaystyle\sum_{i=1}^3\big(\hat{U}\partial_r^2Z_i\dot\Phi+\partial_r^3\dot\Phi Z_i\dot\Phi\\
&+\partial_r\dot\Phi\partial_r^2Z_i\dot\Phi\big)S\partial_rZ_i\dot\Phi\biggr\}
\end{align*}
and
$$|J_{21}|\le C\varepsilon \bigg(r^{-(\gamma-1)}|S\partial_r^2\dot\Phi|+r^{-\sigma-1}|S\partial_rZ\dot\Phi|+r^{-2\gamma}|SZ^2\dot\Phi|\bigg),\eqno{(5.20)}$$
here we point out that (5.20) is derived by direct but tedious computations through applying Lemma 2.1, assumption
(5.1) and the concrete expression of $J_{21}$.
Combining (5.19) and (5.20) together with Lemma 2.5 can yield
$$
\int_{D_T}|(J_1+J_{21})\cdot\mathcal{M}S^3\dot\Phi|dx
\le C\varepsilon\int_{D_T}
\displaystyle\sum_{l=0}^3\bigg(r^{\mu+1-2\gamma+2l}|\nabla_x^l(\frac{1}{r}Z\dot\Phi)|^2
+r^{\mu-1-\delta+2l}|\nabla_x^l\partial_r\dot\Phi|^2\bigg)dx.\eqno{(5.21)}
$$
Finally we treat $\int_{D_T}|J_{22}\cdot\mathcal {M}S^3\dot\Phi|dx$.
To overcome the difficulties induced by the lack of weighted $L^{\infty}$ estimates of
$|\nabla_x^3 \dot\Phi|$ in $J_{22}$, we will use the interpolation inequalities in Corollary 2.3 and Lemma 2.6.
In fact, by (5.1) and the expression of $J_{22}$, it is only enough to deal with the following typical terms
in $\int_{D_T}|J_{22}\cdot\mathcal {M}S^3\dot\Phi|dx$:
\vskip 0.3 true cm
{\bf (A) Estimate of $|r^{\mu+2}\partial_r^3\dot\Phi S\partial_r^2\dot\Phi\partial_rS^3\dot\Phi|_{L^1(D_T)}$}
\begin{align*}
&|r^{\mu+2}\partial_r^3\dot\Phi S\partial_r^2\dot\Phi\partial_rS^3\dot\Phi|_{L^1}=|r^{\mu+3}(\partial_r^3\dot\Phi)^2\partial_rS^3\dot\Phi|_{L^1}\\
=&|r^{\delta-2(\gamma-1)}\cdot (r^{\frac{8\gamma-3-\delta}{4}}\partial_r^3\dot\Phi)^2\cdot r^{\frac{\mu-1-\delta}{2}}\partial_rS^3\dot\Phi|_{L^1}\\
\le & |r^{\frac{8\gamma-3-\delta}{4}}\partial_r^3\dot\Phi|_{L^4}^2|r^{\frac{\mu-1-\delta}{2}}\partial_rS^3\dot\Phi|_{L^2}\\
\le & C\varepsilon\displaystyle\sum_{l=0}^3\bigg(|r^{\frac{\mu-1-\delta+2l}{2}}\nabla_x^l\partial_r\dot\Phi|_{L^2}
+|r^{\frac{\mu+1-2\gamma+2l}{2}}\nabla_x^l(\frac{1}{r}Z\dot\Phi)|_{L^2}\bigg)
|r^{\frac{\mu-1-\delta}{2}}\partial_rS^3\dot\Phi|_{L^2}\\
&\qquad \quad\qquad \quad\qquad \quad\qquad \quad\qquad \quad\text{(Applying Lemma 2.7 for $\dot\Phi$)}\\
\le & C\varepsilon\displaystyle\sum_{l=0}^3\bigg(|r^{\frac{\mu-1-\delta+2l}{2}}\nabla_x^l\partial_r\dot\Phi|_{L^2}^2
+|r^{\frac{\mu+1-2\gamma+2l}{2}}\nabla_x^l(\frac{1}{r}Z\dot\Phi)|_{L^2}^2\bigg).\tag{5.22}
\end{align*}
\vskip 0.3 true cm
{\bf (B) Estimate of $|r^{\mu}Z\dot\Phi\partial_r^2Z\dot\Phi S\partial_r^2\dot\Phi\partial_rS^3\dot\Phi|_{L^1(D_T)}
+|r^{\mu}Z\dot\Phi\partial_r^3\dot\Phi S\partial_rZ\dot\Phi\partial_rS^3\dot\Phi|_{L^1(D_T)}$}
\begin{align*}
&|r^{\mu}Z\dot\Phi\partial_r^2Z\dot\Phi S\partial_r^2\dot\Phi\partial_rS^3\dot\Phi|_{L^1(D_T)}+|r^{\mu}Z\dot\Phi\partial_r^3\dot\Phi S\partial_rZ\dot\Phi\partial_rS^3\dot\Phi|_{L^1(D_T)}\\
=&2|r^{\mu+1}Z\dot\Phi\partial_r^3\dot\Phi\partial_r^2Z\dot\Phi\partial_rS^3\dot\Phi|_{L^1(D_T)}\\
\le& C|r^{\mu+2-\sigma}\partial_r^3\dot\Phi\partial_r^2Z\dot\Phi\partial_rS^3\dot\Phi|_{L^1(D_T)} \qquad\qquad \qquad\text{(By assumption (5.1))}\\
=&|r^{\delta-2(\gamma-1)-\sigma}\cdot r^{\frac{8\gamma-3-\delta}{4}}\partial_r^3\dot\Phi\cdot r^{\frac{8\gamma-7-\delta}{4}}\partial_r^2Z\dot\Phi
\cdot r^{\frac{\mu-1-\delta}{2}}\partial_rS^3\dot\Phi|_{L^1(D_T)}\\
\le& C|r^{\frac{8\gamma-3-\delta}{4}}\partial_r^3\dot\Phi|_{L^4(D_T)}|r^{\frac{8\gamma-7-\delta}{4}}\partial_r^2Z\dot\Phi|_{L^4(D_T)} |r^{\frac{\mu-1-\delta}{2}}\partial_rS^3\dot\Phi|_{L^2(D_T)}\\
\le&C\varepsilon\displaystyle\sum_{l=0}^3\bigg(|r^{\frac{\mu-1-\delta+2l}{2}}\nabla_x^l\partial_r\dot\Phi|_{L^2(D_T)}^2
+|r^{\frac{\mu+1-2\gamma+2l}{2}}\nabla_x^l(\frac{1}{r}Z\dot\Phi)|_{L^2(D_T)}^2\bigg).\quad\text{(Applying Lemma 2.7 for $\dot\Phi$)}\tag{5.23}
\end{align*}
\vskip 0.3 true cm
{\bf (C) Estimate of $|r^{\mu}\partial_r^3\dot\Phi SZ^2\dot\Phi\partial_rS^3\dot\Phi|_{L^1(D_T)}$}
\begin{align*}
&|r^{\mu}\partial_r^3\dot\Phi SZ^2\dot\Phi\partial_rS^3\dot\Phi|_{L^1(D_T)}
=|r^{\mu+1}\partial_r^3\dot\Phi\partial_rZ^2\dot\Phi\partial_rS^3\dot\Phi|_{L^1(D_T)}\\
=&|r^{\delta-2(\gamma-1)}\cdot r^{\frac{8\gamma-3-\delta}{4}}\partial_r^3\dot\Phi\cdot r^{\frac{8\gamma-11-\delta}{4}}\partial_rZ^2\dot\Phi
\cdot r^{\frac{\mu-1-\delta}{2}}\partial_rS^3\dot\Phi|_{L^1(D_T)}\\
\le& C|r^{\frac{8\gamma-3-\delta}{4}}\partial_r^3\dot\Phi|_{L^4(D_T)}|r^{\frac{8\gamma-11-\delta}{4}}\partial_rZ^2\dot\Phi|_{L^4(D_T)} |r^{\frac{\mu-1-\delta}{2}}\partial_rS^3\dot\Phi|_{L^2(D_T)}\\
\le&C\varepsilon\displaystyle\sum_{l=0}^3\bigg(|r^{\frac{\mu-1-\delta+2l}{2}}\nabla_x^l\partial_r\dot\Phi|_{L^2(D_T)}^2
+|r^{\frac{\mu+1-2\gamma+2l}{2}}\nabla_x^l(\frac{1}{r}Z\dot\Phi)|_{L^2(D_T)}^2\bigg).\quad\text{(Applying Lemma 2.7 for $\dot\Phi$)}\tag{5.24}
\end{align*}
\vskip 0.3 true cm
{\bf (D) Estimate of $|r^{\mu-2}Z\dot\Phi \partial_r^2Z\dot\Phi SZ^2\dot\Phi\partial_rS^3\dot\Phi|_{L^1(D_T)}$}
\begin{align*}
&|r^{\mu-2}Z\dot\Phi \partial_r^2Z\dot\Phi SZ^2\dot\Phi\partial_rS^3\dot\Phi|_{L^1(D_T)}\\
\le &C|r^{\mu-\sigma}\partial_r^2Z\dot\Phi\partial_rZ^2\dot\Phi\partial_rS^3\dot\Phi|_{L^1(D_T)}\qquad\qquad\qquad\qquad\text{(By assumption (5.1))}\\
=& C|r^{\delta-2(\gamma-1)-\sigma}\cdot r^{\frac{8\gamma-7-\delta}{4}}\partial_r^2Z\dot\Phi\cdot r^{\frac{8\gamma-11-\delta}{4}}\partial_rZ^2\dot\Phi
\cdot r^{\frac{\mu-1-\delta}{2}}\partial_rS^3\dot\Phi|_{L^1(D_T)}\\
\le& C|r^{\frac{8\gamma-7-\delta}{4}}\partial_r^2Z\dot\Phi|_{L^4(D_T)}|r^{\frac{8\gamma-11-\delta}{4}}\partial_rZ^2\dot\Phi|_{L^4(D_T)} |r^{\frac{\mu-1-\delta}{2}}\partial_rS^3\dot\Phi|_{L^2(D_T)}\\
\le& C\varepsilon\displaystyle\sum_{l=0}^3\bigg(|r^{\frac{\mu-1-\delta+2l}{2}}\nabla_x^l\partial_r\dot\Phi|_{L^2(D_T)}^2
+|r^{\frac{\mu+1-2\gamma+2l}{2}}\nabla_x^l(\frac{1}{r}Z\dot\Phi)|_{L^2(D_T)}^2\bigg).\quad\text{(Applying Lemma 2.7 for $\dot\Phi$)}\tag{5.25}
\end{align*}
\vskip 0.3 true cm
{\bf (E) Estimate of $|r^{\mu}\partial_r^2Z\dot\Phi S\partial_rZ\dot\Phi\partial_rS^3\dot\Phi|_{L^1(D_T)}$}
\begin{align*}
&|r^{\mu}\partial_r^2Z\dot\Phi S\partial_rZ\dot\Phi\partial_rS^3\dot\Phi|_{L^1(D_T)}\\
= &|r^{\mu+1}(\partial_r^2Z\dot\Phi)^2\partial_rS^3\dot\Phi|_{L^1(D_T)}\\
=& |r^{\delta-2(\gamma-1)}\cdot (r^{\frac{8\gamma-7-\delta}{4}}\partial_r^2Z\dot\Phi)^2\cdot r^{\frac{\mu-1-\delta}{2}}\partial_rS^3\dot\Phi|_{L^1(D_T)}\\
\le &C|r^{\frac{8\gamma-7-\delta}{4}}\partial_r^2Z\dot\Phi|_{L^4(D_T)}^2|r^{\frac{\mu-1-\delta}{2}}\partial_rS^3\dot\Phi|_{L^2(D_T)}\\
\le &C\varepsilon\displaystyle\sum_{l=0}^3\bigg(|r^{\frac{\mu-1-\delta+2l}{2}}\nabla_x^l\partial_r\dot\Phi|_{L^2(D_T)}^2
+|r^{\frac{\mu+1-2\gamma+2l}{2}}\nabla_x^l(\frac{1}{r}Z\dot\Phi)|_{L^2(D_T)}^2\bigg).\quad\text{(Applying Lemma 2.7 for $\dot\Phi$)}
\tag{5.26}
\end{align*}
Collecting (5.22)-(5.26), one has
$$\int_{D_T}|J_{22}\cdot\mathcal{M}S^3\dot\Phi|dx
\le C\varepsilon\int_{D_T}
\displaystyle\sum_{l=0}^3\bigg(r^{\mu+1-2\gamma+2l}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2+r^{\mu-1-\delta+2l}(\nabla_x^l\partial_r\dot\Phi)^2\bigg)
dx.\eqno{(5.27)}$$
This, together with (5.17) and (5.21), yields for $m\le 3$
$$\int_{D_T}|I_{3}^m\cdot\mathcal{M}S^m\dot\Phi|dx
\le C\varepsilon\int_{D_T}
\displaystyle\sum_{l=0}^m\bigg(r^{\mu+1-2\gamma+2l}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2+r^{\mu-1-\delta+2l}(\nabla_x^l\partial_r\dot\Phi)^2\bigg)
dx.\eqno{(5.28)}$$
\vskip 0.3 true cm
{\bf Part 4. The estimate of $\int_{D_T}B_{1m}\dot\Phi\cdot\mathcal{M}S^m\Phi dx$}
\vskip 0.3 true cm
At first, from the expressions of $A_k$ with $1\le k\le m$ in (5.8) and Lemma 2.1, we have
\begin{equation}
\left\{
\begin{aligned}
&|A_1\dot\Phi|\le Cr^{1-2\gamma}|\partial_r\dot\Phi|,\\
&|A_k\dot\Phi|\le Cr^{1-2\gamma}\big(\ds\sum_{l=0}^{k-1}|\partial_rS^l\dot\Phi|+\ds\sum_{l=0}^{k-2}|\frac{1}{r}S^lZ^2\dot\Phi|\big),\qquad
k=2,3.
\end{aligned}
\right.\tag{5.29}
\end{equation}
Substituting (5.29) into $B_{1m}\dot\Phi$ and using the expression of $\dot f$, then we have by (5.1)
and a direct computation
\begin{align*}
&|B_{11}\dot\Phi|\le C\varepsilon\bigg(\ds\sum_{0\le l\le 1}(r^{-\gamma}|S^l\partial_r\dot\Phi|+r^{-\sigma-2}|S^lZ\dot\Phi|)+r^{-2\gamma}|Z^2\dot\Phi|\bigg)
+Cr^{1-2\gamma}|\partial_r\dot\Phi|,
\tag{5.30}\\
&|B_{1k}\dot\Phi|\le C\varepsilon\bigg(\ds\sum_{0\le l\le k}(r^{-\gamma}|S^l\partial_r\dot\Phi|+r^{-\sigma-2}|S^lZ\dot\Phi|)
+\ds\sum_{0\le l\le k-1}r^{-2\gamma}|S^lZ^2\dot\Phi|\bigg)\\
&\quad \qquad +Cr^{1-2\gamma}(\ds\sum_{l=0}^{k-1}|\partial_rS^l\dot\Phi|+\ds\sum_{l=0}^{k-2}|\frac{1}{r}S^lZ^2\dot\Phi|),\qquad k=2,3.\tag{5.31}
\end{align*}
Since (5.2) holds for $l\le m-1$, we then have from (5.30)-(5.31)
\begin{align*}
&\int_{D_T}|B_{1m}\dot\Phi\cdot\mathcal{M}S^m\Phi dx|\le
C\varepsilon\int_{D_T}\ds\sum_{l=0}^m\bigg(r^{\mu+1-2\gamma+2l}(\nabla_x(\frac{Z\dot\Phi}{r}))^2+r^{\mu-1-\delta+2l}
(\nabla_x^l\partial_r\dot\Phi)^2\bigg)dx\\
&\qquad +C\varepsilon\bigg(\int_{D_T}r^{\mu-1-\delta}(\partial_rS^m\dot\Phi)^2dx\bigg)^{\f12}.\tag{5.32}
\end{align*}
\vskip 0.3 true cm
{\bf Part 5. The estimate of $\int_{D_T}B_{2m}\dot\Phi\cdot\mathcal{M}S^m\Phi dx$}
\vskip 0.3 true cm
It is noted that $B_{2m}\dot\Phi$ contains the $(m+1)-$th (the highest order) derivatives of $\dot\Phi$
and then $B_{2m}\dot\Phi\cdot\mathcal{M}S^m\Phi$ will contain the term $\nabla_{x}^{\alpha}\dot\Phi\nabla_{x}^{\beta}\dot\Phi$
($|\alpha|=|\beta|=m+1$) which will yield serious troubles in the general case.
However, thanks to $\displaystyle\frac{P_1'(r)}{P_1(r)}<0$
given in Lemma 3.1 and the good form of (3.1), the bad influence of $\nabla_{x}^{\alpha}\dot\Phi\nabla_{x}^{\beta}\dot\Phi$
with $|\alpha|=|\beta|=m+1$ can be eliminated in the related energy
estimates. We now give the details.
Since
$$\partial_r S^m\dot\Phi=\displaystyle\sum_{0\le l\le
m-1}C_{lm}\partial_r S^l\dot\Phi+r S^{m-1}\partial_r^2\dot\Phi,\eqno{(5.33)}$$
then it
follows from (3.1) that
$$\partial_r S^m\dot\Phi=\displaystyle\sum_{0\le l\le
m-1}C_{lm}\partial_r S^l\dot\Phi+r S^{m-1}\big(\frac{1}{r^2}P_1\ds\sum_{i=1}^3Z_i^2\dot\Phi+\mathcal{L}\dot\Phi
-\frac{1}{r}P_2\partial_r\dot\Phi\big).\eqno{(5.34)}$$
A direct computation yields
\begin{align*}
&\int_{D_T} B_{2m}\dot\Phi\cdot \mathcal{M} S^m\dot\Phi dx=\int_{D_T}
\frac{mrP'_1}{P_1}S^{m-1}\big(\frac{P_1}{r^2}\sum\limits_{i=1}^3Z_i^2\dot\Phi\big)\cdot
r^{\mu}a(r)\partial_r S^m\dot\Phi dx\\
= &\int_{D_T}
\frac{mr^{\mu+1}a(r)P'_1}{P_1}\bigg\{\bigg[S^{m-1}\big(\frac{P_1}{r^2}\sum\limits_{i=1}^3Z_i^2\dot\Phi\big)\bigg]^2\\
&\qquad + \bigg(\displaystyle\sum_{0\le l\le
m-1}C_{lm}\partial_r S^l\dot\Phi+r S^{m-1}\big(\mathcal{L}\dot\Phi-\frac{P_2}{r}\partial_r\dot\Phi\big)\bigg)
S^{m-1}\big(\frac{P_1}{r^2}\sum\limits_{i=1}^3Z_i^2\dot\Phi\big)\bigg\}dx\\
&\le \int_{D_T} \frac{mr^{\mu+1}a(r)P'_1}{P_1}\bigg(\displaystyle\sum_{0\le l\le
m-1}C_{lm}\partial_r S^l\dot\Phi+r S^{m-1}\big(\mathcal{L}\dot\Phi-\frac{P_2}{r}\partial_r\dot\Phi\big)\bigg)
S^{m-1}\big(\frac{P_1}{r^2}\sum\limits_{i=1}^3Z_i^2\dot\Phi\big)dx.\\
&\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad
\quad\quad\text{(By $\displaystyle\frac{P_1'(r)}{P_1(r)}<0$)}\tag{5.35}
\end{align*}
Note that $\displaystyle\sum_{0\le l\le
m-1}C_{lm}\partial_r S^l\dot\Phi-r S^{m-1}\big(\frac{P_2}{r}\partial_r\dot\Phi\big)$ only contains at most
$m-$order derivatives of $\dot\Phi$, then we have by (5.2) for $l\le m-1$
$$
\int_{D_T}r^{\mu-1-\delta}\bigg(\displaystyle\sum_{0\le l\le
m-1}C_{lm}\partial_r S^l\dot\Phi-r S^{m-1}\big(\frac{P_2}{r}\partial_r\dot\Phi\big)\bigg)^2dx\le C\varepsilon^2.\eqno{(5.36)}
$$
On the other hand,
we have
$$
|S^{m-1}\mathcal{L}\dot\Phi|\le C\varepsilon \bigg\{\displaystyle\sum_{0\le
l\le m}\big(r^{-\gamma}|S^l\partial_r\dot\Phi|+r^{-\sigma-2}|S^lZ\dot\Phi|\big)
+\displaystyle\sum_{0\le l\le m-1}r^{-2\gamma}|S^lZ^2\dot\Phi|\bigg\}.\eqno{(5.37)}
$$
Therefore, inserting (5.36)-(5.37) into (5.35) yields
\begin{align*}
&\int_{D_T}B_{2m}\dot\Phi\cdot\mathcal{M}S^m\dot\Phi dx\le C\varepsilon^2+
C\varepsilon\int_{D_T}
\displaystyle\sum_{l=0}^m\bigg(r^{\mu+1-2\gamma+2l}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2+r^{\mu-1-\delta+2l}(\nabla_x^l\partial_r\dot\Phi)^2\bigg)
dx\\
&\qquad +C\varepsilon\bigg(\int_{D_T}r^{\mu-1-2\gamma}|S^{m-1}Z^2\dot\Phi|^2dx\bigg)^{\f12}.\tag{5.38}
\end{align*}
Consequently, inserting (5.9)-(5.10), (5.15)-(5.17), (5.28), (5.32)
and (5.38) into (5.6), we can complete the proof of (5.3).
For the case of $m=0$, (5.4) comes directly from Theorem 4.1, (5.10)
and (5.15)-(5.17). \qquad \qquad $\square$
Based on Lemma 5.2 and the ingredients in Lemma 5.2,
we now derive a series of estimates on the higher order derivatives of $\dot\Phi$.
{\bf Lemma 5.3. (Second order angular derivative estimates)} {\it Under the assumptions of Theorem 5.1, then
\begin{align*}
&T^{\mu}\int_{S_T}(\partial_rZ\dot\Phi)^2dS
+T^{\mu-2\gamma}\int_{S_T}(Z^2\dot\Phi)^2dS
+\int_{D_T}\bigg(r^{\mu-1-\delta}(\partial_rZ\dot\Phi)^2
+r^{\mu-1-2\gamma}(Z^2\dot\Phi)^2\bigg)dx\\
&\le C\varepsilon^2+C\varepsilon\bigg(T^{\mu}\int_{S_T}(\partial_rZ\dot\Phi)^2dS+T^{\mu-2\gamma}\int_{S_T}(Z^2\dot\Phi)^2dS\bigg)\\
&\quad +C\varepsilon\bigg(\displaystyle\sum_{l=0}^1\int_{D_T}r^{\mu+1-2\gamma+2l}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2
+r^{\mu-1-\delta+2l}(\nabla_x^l\partial_r\dot\Phi)^2dx\bigg),\tag{5.39}
\end{align*}
where $0<\delta\le\gamma-1$.}
{\bf Remark 5.3.} {\it Lemma 5.3, together with Lemma 5.2 for $m=1$, yields (5.2) in the case of $k=1$.}
{\bf Proof.} Noting $\mathcal {L} Z\dot\Phi=Z\mathcal {L}\dot\Phi=Z\dot f$, then it follows from Theorem 4.1 that
\begin{align*}
&T^{\mu}\int_{S_T}|\partial_rZ\dot\Phi|^2dS+T^{\mu-2\gamma}\int_{S_T}|Z^2\dot\Phi|^2dS
+\int_{D_T}\bigg(r^{\mu-1-\delta}|\partial_rZ\dot\Phi|^2+r^{\mu-1-2\gamma}|Z^2\dot\Phi|^2\bigg)dx\\
&\le C\int_{D_T}Z\dot f\cdot\mathcal{M}Z\dot\Phi
dx+C\varepsilon^2.\tag{5.40}
\end{align*}
In order to estimate the term $\int_{D_T}Z\dot f\cdot\mathcal{M}Z\dot\Phi
dx$ in the right hand side of (5.40), we rewrite $Z\dot f=D_1+D_2$ with
\begin{align*}
D_1&=f_{00}\partial_r^2Z\dot\Phi+\frac{1}{r^2}\sum\limits_{1\le
i,j\le3}f_{ij}Z_iZ_jZ\dot\Phi
+\frac{1}{r}\sum\limits_{i=1}^3f_{0i}\partial_rZ_i(Z\dot\Phi),\\
D_2&=\frac{1}{r^2}\sum\limits_{1\le
i,j\le3}f_{ij}[Z, Z_iZ_j]\dot\Phi
+\frac{1}{r}\sum\limits_{i=1}^3f_{0i}[Z, \partial_rZ_i]\dot\Phi
+Zf_{00} \partial_r^2\dot\Phi\\
&\quad +Z(\frac{1}{r^2}\sum\limits_{1\le i,j\le3}f_{ij}) Z_iZ_j\dot\Phi
+Z(\frac{1}{r}\sum\limits_{i=1}^3f_{0i}) \partial_rZ_i\dot\Phi+Zf_0,
\end{align*}
where $D_1$ contains the third order derivatives of $\dot\Phi$, and $D_2$ is composed by the lower order
(up to second order)
derivative terms of $\dot\Phi$.
In this case, a direct computation yields
\begin{align*}
&D_1\cdot\mathcal{M}Z\dot\Phi\\
&=
\partial_r\bigg(\frac{1}{2}r^{\mu}a(r)f_{00}(\partial_rZ\dot\Phi)^2-r^{\mu-2}a(r)\ds\sum_{1\le i<j\le3}f_{ij}Z_iZ\dot\Phi Z_jZ\dot\Phi
-\frac{1}{2}r^{\mu-2}a(r)\ds\sum_{i=1}^3f_{ii}(Z_iZ\dot\Phi)^2\bigg)\\
&\quad +\ds\sum_{i=1}^3Z_i\bigg(\frac{1}{2}r^{\mu-1}a(r)f_{0i}(\partial_rZ\dot\Phi)^2
+r^{\mu-2}a(r)\partial_rZ\dot\Phi\ds\sum_{j=1}^3f_{ij}Z_jZ\dot\Phi\bigg)\\
&\quad -\frac{1}{2}\partial_r(r^\mu a(r)f_{00})(\partial_rZ\dot\Phi)^2-\frac{1}{2}r^{\mu-1}a(r)(\partial_rZ\dot\Phi)^2\ds\sum_{i=1}^3Z_if_{0i}
+\ds\sum_{i=1}^3\partial_r(\frac{1}{2}r^{\mu-2}a(r)f_{ii})(Z_iZ\dot\Phi)^2\\
&\quad +\ds\sum_{1\le i<j\le3}\partial_r(r^{\mu-2}a(r)f_{ij})Z_iZ\dot\Phi Z_jZ\dot\Phi.\tag{5.41}
\end{align*}
On the other hand, as in (5.14), it follows from the expressions of $f_{ij}, f_{0i}$ and the boundary condition
(3.4) that on $\Sigma$
\begin{align*}
&x_2\bigg(\frac{1}{2}r^{\mu-1}a(r)f_{02}(\partial_rZ\dot\Phi)^2+r^{\mu-2}a(r)\partial_rZ\dot\Phi\ds\sum_{j=1}^3f_{2j}Z_jZ\dot\Phi\bigg)\\
&-x_1\bigg(\frac{1}{2}r^{\mu-1}a(r)f_{03}(\partial_rZ\dot\Phi)^2+r^{\mu-2}a(r)\partial_rZ\dot\Phi\ds\sum_{j=1}^3f_{3j}Z_jZ\dot\Phi\bigg)=0.\tag{5.42}
\end{align*}
Thus, by integration by parts together with the expressions of $f_i$ and (5.1), we have
\begin{align*}
&|\int_{D_T}D_1\cdot\mathcal{M}Z\dot\Phi dx|\le C\varepsilon^2+C\varepsilon\bigg(T^{\mu}\int_{S_T}(\partial_rZ\dot\Phi)^2dS
+T^{\mu-2\gamma}\int_{S_T}(Z^2\dot\Phi)^2dS\bigg)\\
&\qquad\qquad\qquad\qquad +C\varepsilon\bigg(\displaystyle\sum_{l=0}^1\int_{D_T}r^{\mu+1-2\gamma+2l}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2
+r^{\mu-1-\delta+2l}(\nabla_x^l\partial_r\dot\Phi)^2dx\bigg).\tag{5.43}
\end{align*}
In addition, a direct computation yields
$$|D_2|\le C\varepsilon\big(r^{-2(\gamma-1)}|\partial_r^2\dot\Phi|+r^{-2(\gamma-1)}|\partial_rZ\dot\Phi|+r^{-2-\sigma}|Z^2\dot\Phi|\big),\eqno{(5.44)}$$
which implies
$$\int_{D_T}|D_2\cdot\mathcal{M}Z\dot\Phi|dx
\le C\varepsilon\int_{D_T}\bigg(r^{\mu+1-\delta}(\partial_r^2\dot\Phi)^2+r^{\mu-1-\delta}(\partial_rZ\dot\Phi)^2+r^{\mu-1-2\gamma}(Z^2\dot\Phi)^2\bigg)dx.\eqno{(5.45)}
$$
Substituting (5.43) and (5.45) into (5.40) yields (5.39), we then complete the proof of Lemma 5.3.
\qquad\quad $\square$
{\bf Lemma 5.4. (Third order angular derivative estimates)} {\it Under the assumptions of Theorem 5.1,
then
\begin{align*}
&T^{\mu}\int_{S_T}(\partial_rZ^2\dot\Phi)^2dS
+T^{\mu-2\gamma}\int_{S_T}(Z^3\dot\Phi)^2dS
+\int_{D_T}\bigg(r^{\mu-\gamma}(\partial_rZ^2\dot\Phi)^2
+r^{\mu-1-2\gamma}(Z^3\dot\Phi)^2\bigg)dx\\
&\le C\varepsilon^2+C\varepsilon\int_{D_T}
\ds\sum_{l=0}^2\bigg(r^{\mu+1-2\gamma+2l}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2+r^{\mu-1-\delta+2l}(\nabla_x^l\partial_r\dot\Phi)^2\bigg)
dx\\
&\qquad +C\varepsilon\ds\sum_{l=0}^2\bigg(T^{\mu+2l}\int_{S_T}(\nabla_x^l\partial_r\dot\Phi)^2dS
+T^{\mu-2\gamma+2l}\int_{S_T}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2dS\bigg),\tag{5.46}
\end{align*}
where $0<\delta\le\min\{\gamma-1, 2\sigma-2(\gamma-1)\}$ with $\sigma=\min\{1, 2(\gamma-1)\}$.}
{\bf Remark 5.4.} {\it Under the conditions of Lemma 5.4, as in the proof of Lemma 5.3, we have
\begin{align*}
&T^{\mu}\int_{S_T}(\partial_rSZ\dot\Phi)^2dS+T^{\mu-2\gamma}\int_{S_T}(ZSZ\dot\Phi)^2dS\\
&\qquad +\int_{D_T}\bigg(r^{\mu-1-\delta}(\partial_rSZ\dot\Phi)^2+r^{\mu-1-2\gamma}(ZSZ\dot\Phi)^2\bigg)dx\\
&\le C\varepsilon^2+C\varepsilon\int_{D_T}
\ds\sum_{l=0}^2\bigg(r^{\mu+1-2\gamma+2l}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2
+r^{\mu-1-\delta+2l}(\nabla_x^l\partial_r\dot\Phi)^2\bigg)
dx\\
&\qquad +C\varepsilon\ds\sum_{l=0}^2\bigg(T^{\mu+2l}\int_{S_T}(\nabla_x^l\partial_r\dot\Phi)^2dS
+T^{\mu-2\gamma+2l}\int_{S_T}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2dS\bigg)\\
&\qquad +C\varepsilon\bigg(\int_{D_T}\big(r^{\mu-1-\delta}(\partial_rSZ\dot\Phi)^2+r^{\mu-1-2\gamma}(Z^3\dot\Phi)^2\big)dx\bigg)^{\f12}.
\end{align*}
This, together with (5.46) and (5.3) in the case of $m=2$ and Remark 2.3, yields (5.2) for $k=2$ under the
assumption that (5.2)
holds for $k\le1$.}
{\bf Proof.} At first, we establish an analogous inequality
to (5.46) for $(Z_2^2+Z_3^2)\dot\Phi$. Based on this together with
the domain decomposition technique, we can complete the proof of (5.46).
It follows from (1.9) that on $\Sigma$
$$\partial_r^m\partial_{\varphi}\Phi=0.\eqno{(5.47)}$$
Differentiating (1.6) with respect to $\varphi$ and applying (5.47) yield on $\Sigma$
$$\partial_{\varphi}(\sin\varphi\partial_{\varphi}^2\Phi)=0.\eqno{(5.48)}$$
On the other hand, by a direct computation we have
$$\partial_{\varphi}(Z_2^2+Z_3^2)\Phi=(\partial_{\varphi}^3+\cot\partial_{\varphi}^2-\csc^2\varphi\partial_{\varphi})\Phi.$$
This, together with (5.47)-(5.48) and the definition of $\dot\Phi$, yields
$$
Z(Z_2^2+Z_3^2)\dot\Phi=0\qquad \text{on $\Sigma$}.\eqno{(5.49)}$$
Applying Theorem 4.1 to $(Z_2^2+Z_3^2)\dot\Phi$, one has for $\delta\le\gamma-1$
\begin{align*}
&T^{\mu}\int_{S_T}(\partial_r(Z_2^2+Z_3^2)\dot\Phi)^2dS+T^{\mu-2\gamma}\int_{S_T}(Z(Z_2^2+Z_3^2)\dot\Phi)^2dS\\
&\quad +\int_{D_T}\bigg(r^{\mu-\gamma}(\partial_r(Z_2^2+Z_3^2)\dot\Phi)^2+r^{\mu-1-2\gamma}(Z(Z_2^2+Z_3^2)\dot\Phi)^2\bigg)dx\\
&\le \int_{D_T}\mathcal{L}(Z_2^2+Z_3^2)\dot\Phi\cdot\mathcal{M}(Z_2^2+Z_3^2)\dot\Phi
dx +C\varepsilon^2.\tag{5.50}
\end{align*}
By $\mathcal{L}(Z_2^2+Z_3^2)\dot\Phi=(Z_2^2+Z_3^2)\mathcal{L}\dot\Phi=(Z_2^2+Z_3^2)\dot f$ and the expression of $\dot f$,
we have
$$
\mathcal{L}(Z_2^2+Z_3^2)\dot\Phi=K_1+K_2+K_3+K_4\eqno{(5.51)}
$$
with
\begin{align*}
K_1&=f_{00}\partial_r^2(Z_2^2+Z_3^2)\dot\Phi+\frac{1}{r^2}\sum\limits_{1\le
i,j\le3}f_{ij}Z_iZ_j(Z_2^2+Z_3^2)\dot\Phi
+\frac{1}{r}\sum\limits_{i=1}^3f_{0i}\partial_rZ_i((Z_2^2+Z_3^2)\dot\Phi),\\
K_2&=\frac{1}{r^2}\sum\limits_{1\le
i,j\le3}f_{ij}[(Z_2^2+Z_3^2),Z_iZ_j]\dot\Phi
+\frac{1}{r}\sum\limits_{i=1}^3f_{0i}[(Z_2^2+Z_3^2),\partial_rZ_i]\dot\Phi,\\
K_3&=\ds\sum_{k=2}^3\sum\limits_{l=1}^2C_{l}\bigg(Z_k^{l}f_{00}Z_k^{2-l}\partial_r^2\dot\Phi
+Z_k^{l}(\frac{1}{r^2}\sum\limits_{1\le i,j\le3}f_{ij})Z_k^{2-l}Z_iZ_j\dot\Phi
+Z_k^{l}(\frac{1}{r}\sum\limits_{i=1}^3f_{0i})Z_k^{2-l}\partial_rZ_i\dot\Phi\bigg)\\
&\quad +(Z_2^2+Z_3^2)f_0^1,\\
K_4&=(Z_2^2+Z_3^2)f_0^2.
\end{align*}
Next, we start to deal with each term $\int_{D_T}K_i\cdot\mathcal{M}(Z_2^2+Z_3^2)\dot\Phi
dx$ $(1\le i\le 4)$.
\vskip 0.3 true cm
{\bf (A) The estimate on $\int_{D_T}K_1\cdot\mathcal{M}(Z_2^2+Z_3^2)\dot\Phi
dx$}
\vskip 0.3 true cm
Analogous to the treatment on $\int_{D_T}D_1\cdot\mathcal{M}(Z_2^2+Z_3^2)\dot\Phi
dx$ in (5.41) and (5.43), we can obtain
\begin{align*}
&|\int_{D_T}K_1\cdot\mathcal{M}(Z_2^2+Z_3^2)\dot\Phi dx|\\
&\le C\varepsilon^2+C\varepsilon\bigg(T^{\mu}\int_{S_T}(\partial_r(Z_2^2+Z_3^2)\dot\Phi)^2dS+T^{\mu-2\gamma}\int_{S_T}(Z(Z_2^2+Z_3^2)\dot\Phi)^2dS\\
&\quad +\int_{D_T}\big(r^{\mu-1-\delta}(\partial_r(Z_2^2+Z_3^2)\dot\Phi)^2+r^{\mu-1-2\gamma}(Z(Z_2^2+Z_3^2)\dot\Phi)^2\big)dx\bigg).
\tag{5.52}
\end{align*}
\vskip 0.3 true cm
{\bf (B) The estimate on $\int_{D_T}K_2\cdot\mathcal{M}(Z_2^2+Z_3^2)\dot\Phi
dx$}
\vskip 0.3 true cm
By the expressions of $f_{ij}$ and the assumption (5.1), it follows from a direct computation that
$$|K_2|\le C\varepsilon
\big(r^{-2\gamma}|Z^3\dot\Phi|+r^{-1-\sigma}|\partial_rZ^2\dot\Phi|\big)\eqno{(5.53)}$$
and
$$
\int_{D_T}|K_2\cdot\mathcal{M}(Z_2^2+Z_3^2)\dot\Phi|dx
\le C\varepsilon\int_{D_T}
\ds\sum_{l=0}^2\big(r^{\mu+1-2\gamma+2l}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2+r^{\mu-1-\delta+2l}(\nabla_x^l\partial_r\dot\Phi)^2\big)
dx.\eqno{(5.54)}
$$
\vskip 0.3 true cm
{\bf (C) The estimate on $\int_{D_T}K_3\cdot\mathcal{M}(Z_2^2+Z_3^2)\dot\Phi
dx$}
\vskip 0.3 true cm
By the expressions of $f_{ij}$ and $f_0^1$, we know that $K_3$ only
contains such terms: $\partial_r^2Z^l\dot\Phi$ ($l=0,1$), $\partial_rZ^l\dot\Phi$ ($0\le
l\le 2$) and $Z^l\dot\Phi$ ($1\le l\le 3$) with suitably decayed
coefficients. More concretely, by the assumption (5.1), we have
$$|K_3|\le C\varepsilon \big(r^{-2(\gamma-1)}\ds\sum_{0\le l\le1}|\partial_r^2Z^l\dot\Phi|+r^{-\gamma}\ds\sum_{0\le
l\le 2}|\partial_rZ^l\dot\Phi|+r^{-2\gamma}\ds\sum_{1\le l\le
3}|Z^l\dot\Phi|\big),$$
which implies
$$
\int_{D_T}|K_3\cdot\mathcal{M}(Z_2^2+Z_3^2)\dot\Phi|dx
\le C\varepsilon\int_{D_T}
\ds\sum_{l=0}^2\bigg(r^{\mu+1-2\gamma+2l}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2+r^{\mu-1-\delta+2l}(\nabla_x^l\partial_r\dot\Phi)^2\bigg)
dx.\eqno{(5.55)}
$$
\vskip 0.3 true cm
{\bf (D) The estimate on $\int_{D_T}K_4\cdot\mathcal{M}(Z_2^2+Z_3^2)\dot\Phi
dx$}
\vskip 0.3 true cm
By the expression of $f_0^2$, we have
$$K_4=K_4^1+K_4^2\eqno{(5.56)}$$
with
\begin{align*}
&K_4^1=\frac{\hat{U}}{r^3(\hat{U}^2-c^2(\hat\rho))}\ds\sum_{k=2}^3\ds\sum_{i,j=1}^3\bigg(C_{ij}\big(Z_i\dot\Phi Z_k^2Z_j\dot\Phi+Z_k^2Z_i\dot\Phi Z_j\dot\Phi\big)+2Z_kC_{ij}\big(Z_i\dot\Phi Z_kZ_j\dot\Phi+Z_kZ_i\dot\Phi Z_j\dot\Phi\big)\\
&\qquad\qquad+Z_k^2C_{ij}Z_i\dot\Phi Z_j\dot\Phi\bigg)
+\frac{(\gamma-1)\hat{U}}{r^3(\hat{U}^2-c^2(\hat\rho))}\ds\sum_{k=2}^3\ds\sum_{i=1}^3Z_i\dot\Phi Z_k^2Z_i\dot\Phi,\\
&K_4^2=-\ds\sum_{k=2}^3\bigg(\ds\sum_{i,j=1}^3\frac{2\hat{U}}{r^3(\hat{U}^2-{c}^2(\hat\rho))}C_{ij}Z_kZ_i\dot\Phi
Z_kZ_j\dot\Phi+\ds\sum_{i=1}^3\frac{2(\gamma-1)\hat{U}}{r^3(\hat{U}^2-c^2(\hat\rho))}(Z_kZ_i\dot\Phi)^2\bigg),\\
\end{align*}
where $K_4^1=O(\displaystyle\frac{Z\dot\Phi Z^3\dot\Phi}{r^3})+O(\displaystyle\frac{Z\dot\Phi Z^2\dot\Phi}{r^3})$, and $K_4^2=O(\displaystyle\frac{Z_iZ_j\dot\Phi Z_kZ_l\dot\Phi}{r^3})$.
Here we point out that $K_4^1$ can be easily estimated since $\displaystyle\frac{Z\dot\Phi}{r}$ admits a good decay rate in assumption (5.1).
In fact, we have
$$|K_4^1|\le
Cr^{-3}|Z\dot\Phi|\ds\sum_{l=0}^2|Z^lZ\dot\Phi|\le Cr^{-2-\sigma}\ds\sum_{l=0}^2|Z^lZ\dot\Phi|$$
and
$$
\int_{D_T}|K_4^1\cdot\mathcal{M}(Z_2^2+Z_3^2)\dot\Phi|dx
\le C\varepsilon\int_{D_T}
\ds\sum_{l=0}^2\big(r^{\mu+1-2\gamma+2l}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2+r^{\mu-1-\delta+2l}(\nabla_x^l\partial_r\dot\Phi)^2\big)
dx.\eqno{(5.57)}
$$
Next we deal with $\int_{D_T}|K_4^2\cdot\mathcal{M}(Z_2^2+Z_3^2)\dot\Phi|dx$.
By H\"older inequality and Lemma 2.6-Lemma 2.7, we have for $\delta<2\sigma-2(\gamma-1)$
\begin{align*}
&|r^{\mu-3}Z_kZ_i\dot\Phi Z_kZ_j\dot\Phi\partial_r(Z_2^2+Z_3^2)\dot\Phi|_{L_1(D_T)}\\
=& |r^{\frac{\delta+2(\gamma-1)-2\sigma}{2}}\cdot r^{\frac{2\gamma+2\sigma-9}{4}}Z_kZ_i\dot\Phi\cdot r^{\frac{2\gamma+2\sigma-9}{4}}Z_kZ_j\dot\Phi
\cdot r^{\frac{\mu-1-\delta}{2}}\partial_r(Z_2^2+Z_3^2)\dot\Phi|_{L_1(D_T)}\\
\le &C|r^{\frac{2\gamma+2\sigma-9}{4}}Z_kZ_i\dot\Phi|_{L^4(D_T)} |r^{\frac{2\gamma+2\sigma-9}{4}}Z_kZ_j\dot\Phi|_{L^4(D_T)}|r^{\frac{\mu-1-\delta}{2}}\partial_r(Z_2^2+Z_3^2)\dot\Phi|_{L^2(D_T)}\\
\le &C\varepsilon\ds\sum_{l=0}^2\bigg(|r^{\frac{\mu-1-\delta+2l}{2}}\nabla_x^l\partial_r\dot\Phi|_{L^2(D_T)}
+|r^{\frac{\mu+1-2\gamma+2l}{2}}\nabla_x^l(\frac{1}{r}Z\dot\Phi)|_{L^2(D_T)}\bigg)
|r^{\frac{\mu-1-\delta}{2}}\partial_r(Z_2^2+Z_3^2)\dot\Phi|_{L^2(D_T)}\\
\le &C\varepsilon\ds\sum_{l=0}^2\bigg(|r^{\frac{\mu-1-\delta+2l}{2}}\nabla_x^l\partial_r\dot\Phi|_{L^2(D_T)}^2
+|r^{\frac{\mu+1-2\gamma+2l}{2}}\nabla_x^l(\frac{1}{r}Z\dot\Phi)|_{L^2(D_T)}^2\bigg),
\end{align*}
which derives
$$
\int_{D_T}|K_4^2\cdot\mathcal{M}(Z_2^2+Z_3^2)\dot\Phi|dx\le
C\varepsilon\int_{D_T}
\ds\sum_{l=0}^2\bigg(r^{\mu+1-2\gamma+2l}\nabla_x^l(\frac{1}{r}Z\dot\Phi)^2+r^{\mu-1-\delta+2l}(\nabla_x^l\partial_r\dot\Phi)^2\bigg)
dx.\eqno{(5.58)}
$$
Substituting (5.51), (5.52), (5.54)-(5.58) into (5.50) yields
\begin{align*}
&T^{\mu}\int_{S_T}(\partial_r(Z_2^2+Z_3^2)\dot\Phi)^2dS
+T^{\mu-2\gamma}\int_{S_T}(Z(Z_2^2+Z_3^2)\dot\Phi)^2dS\\
&\qquad +\int_{D_T}\bigg(r^{\mu-\gamma}(\partial_r(Z_2^2+Z_3^2)\dot\Phi)^2
+r^{\mu-1-2\gamma}(Z(Z_2^2+Z_3^2)\dot\Phi)^2\bigg)dx\\
&\le C\varepsilon^2+C\varepsilon\int_{D_T}
\ds\sum_{l=0}^2\bigg(r^{\mu+1-2\gamma+2l}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2+r^{\mu-1-\delta+2l}(\nabla_x^l\partial_r\dot\Phi)^2\bigg)
dx\\
&\qquad +C\varepsilon\bigg(T^{\mu}\int_{S_T}(\partial_r(Z_2^2+Z_3^2)\dot\Phi)^2dS
+T^{\mu-2\gamma}\int_{S_T}(Z(Z_2^2+Z_3^2)\dot\Phi)^2dS\bigg).\tag{5.59}
\end{align*}
By (5.59), we know that the estimates of $Z(Z_2^2+Z_3^2)\dot\Phi$ and $\partial_r(Z_2^2+Z_3^2)\dot\Phi$
on $S_T$ or $D_T$ have been established but such kinds of third order angular
derivatives $Z^3\dot\Phi$ and $\partial_rZ^2\dot\Phi$
are not estimated. However, away from $\varphi=0$ in $\Omega$, we can get the estimates on the third order
derivatives $Z^3\dot\Phi$ and $\partial_rZ^2\dot\Phi$. Indeed, for $\varphi>\frac{\varphi_0}{3}$,
by Lemma 2.5 a direct computation yields
\begin{equation}
\left\{
\begin{aligned}
&\ds\sum_{i,j,k=1}^3|Z_iZ_jZ_k\dot\Phi|\le C\big(\ds\sum_{i=1}^3|Z_i\ds\sum_{j=1}^3Z_j^2\dot\Phi|
+\ds\sum_{i,j=1}^3|Z_iZ_j\dot\Phi|+\ds\sum_{i=1}^3|Z_i\dot\Phi|\big),\\
&\ds\sum_{i,j=1}^3|\partial_rZ_iZ_j\dot\Phi|\le C\big(|\partial_r\ds\sum_{j=1}^3Z_j^2\dot\Phi|+\ds\sum_{i=1}^3|\partial_rZ_i\dot\Phi|+|\partial_r\dot\Phi|\big).
\end{aligned}
\right.\tag{5.60}
\end{equation}
This, together with $Z_1^k\dot\Phi\equiv 0$ $(k\in\Bbb N)$, (5.59) and Lemma 5.3, yields
\begin{align*}
&T^{\mu}\int_{S_T\cap\{\varphi\ge\frac{\varphi_0}{3}\}}(\partial_rZ^2\dot\Phi)^2dS
+T^{\mu-2\gamma}\int_{S_T\cap\{\varphi\ge\frac{\varphi_0}{3}\}}(Z^3\dot\Phi)^2dS\\
&\quad +\int_{D_T\cap\{\varphi\ge\frac{\varphi_0}{3}\}}\bigg(r^{\mu-\gamma}(\partial_rZ^2\dot\Phi)^2
+r^{\mu-1-2\gamma}(Z^3\dot\Phi)^2\bigg)dx\\
&\le C\varepsilon^2+C\varepsilon\int_{D_T}
\ds\sum_{l=0}^2\bigg(r^{\mu+1-2\gamma+2l}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2+r^{\mu-1-\delta+2l}(\nabla_x^l\partial_r\dot\Phi)^2\bigg)
dx\\
&\quad +C\varepsilon\bigg(T^{\mu}\int_{S_T}(\partial_r(Z_2^2+Z_3^2)\dot\Phi)^2dS
+T^{\mu-2\gamma}\int_{S_T}(Z(Z_2^2+Z_3^2)\dot\Phi)^2dS\bigg).\tag{5.61}
\end{align*}
To obtain the estimates of $Z^3\dot\Phi$ and $\partial_rZ^2\dot\Phi$ on the domain $\{\varphi\le\displaystyle\frac{\varphi_0}{3}\}$, we will take
a domain decomposition technique. Namely, we choose
a smooth cut-off function $\chi(\varphi)$ as follows
\begin{equation*}
\chi(\varphi)= \left\{
\begin{aligned}
&1,\quad\text{for $0\le\varphi\le\frac{\varphi_0}{3}$},\\
&0,\quad\text{for $\frac{2\varphi_0}{3}\le\varphi\le\varphi_0$},\\
&\text{smooth connection}, \quad \text{for $\frac{\varphi_0}{3}\le\varphi\le\frac{2\varphi_0}{3}$}\\
\end{aligned}
\right.
\end{equation*}
such that $\chi(\varphi)\dot\Phi$ is studied.
Indeed, by Theorem 4.1 we have
\begin{align*}
&T^{\mu}\int_{S_T}\chi(\varphi)(\partial_rZ^2\dot\Phi)^2dS
+T^{\mu-2\gamma}\int_{S_T}\chi(\varphi)(Z^3\dot\Phi)^2dS\\
&\qquad +
\int_{D_T}\chi(\varphi)\bigg(r^{\mu-\gamma}(\partial_rZ^2\dot\Phi)^2
+r^{\mu-1-2\gamma}(Z^3\dot\Phi)^2\bigg)dx\\
&\le C\bigg(\int_{D_T}r^{\mu-1-2\gamma}|\chi'(\varphi)|(Z^3\dot\Phi)^2dx\bigg)^{\frac{1}{2}}
\bigg(\int_{D_T}r^{\mu-\gamma}|\chi'(\varphi)|(\partial_rZ^2\dot\Phi)^2dx\bigg)^{\frac{1}{2}}\\
&\qquad +\int_{D_T}\mathcal{L}Z^2\dot\Phi\cdot\chi(\varphi)\mathcal{M}Z^2\dot\Phi
dx+C\varepsilon^2,\tag{5.62}
\end{align*}
here we point out that $\chi'(\varphi)$ has a compact support away from $\varphi=0$, which implies that
the first term in the right hand side of (5.62) can be estimated as in
(5.61).
On the other hand, by the compact support property of $\chi(\varphi)$ away from $\varphi=0$,
then completely similar to the treatment
on $\int_{D_T}\mathcal{L}(Z_2^2+Z_3^2)\dot\Phi\cdot\mathcal{M}(Z_2^2+Z_3^2)\dot\Phi
dx$ in (5.50), we can arrive at
\begin{align*}
&\int_{D_T}\mathcal{L}Z^2\dot\Phi\cdot\chi(\varphi)\mathcal{M}Z^2\dot\Phi dx\\
\le &C\varepsilon^2+C\varepsilon\int_{D_T}
\ds\sum_{l=0}^2\bigg(r^{\mu+1-2\gamma+2l}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2+r^{\mu-1-\delta+2l}(\nabla_x^l\partial_r\dot\Phi)^2\bigg)
dx\\
&+C\varepsilon\ds\sum_{l=0}^2\bigg(T^{\mu+2l}\int_{S_T}(\nabla_x^l\partial_r\dot\Phi)^2dS
+T^{\mu-2\gamma+2l}\int_{S_T}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2dS\bigg).\tag{5.63}
\end{align*}
Finally, combining (5.61) and (5.62)-(5.63) yields the proof of (5.46).\qquad\qquad $\square$
Finally, we deal with the estimates of $\nabla_x^4\dot\Phi$.
{\bf Lemma 5.5. (Fourth order angular derivative estimates)} {\it Under the assumptions of Theorem 5.1, then
\begin{align*}
&T^{\mu}\int_{S_T}(\partial_rZ^3\dot\Phi)^2dS
+T^{\mu-2\gamma}\int_{S_T}(Z^4\dot\Phi)^2dS
+\int_{D_T}\bigg(r^{\mu-\gamma}(\partial_rZ^3\dot\Phi)^2
+r^{\mu-1-2\gamma}(Z^4\dot\Phi)^2\bigg)dx\\
\le&C\varepsilon^2+C\varepsilon\int_{D_T}
\ds\sum_{l=0}^3\bigg(r^{\mu+1-2\gamma+2l}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2+r^{\mu-1-\delta+2l}(\nabla_x^l\partial_r\dot\Phi)^2\bigg)
dx\\
& +C\varepsilon\ds\sum_{l=0}^3\bigg(T^{\mu+2l}\int_{S_T}(\nabla_x^l\partial_r\dot\Phi)^2dS
+T^{\mu-2\gamma+2l}\int_{S_T}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2dS\bigg),\tag{5.64}
\end{align*}
where $0<\delta\le\min\{\gamma-1, \sigma-(\gamma-1)\}$ with $\sigma=\min\{1, 2(\gamma-1)\}$.}
{\bf Remark 5.5.} {\it Under the conditions of Lemma 5.5, as in the proofs of Lemma 5.3
and Lemma 5.4 respectively, we have
\begin{align*}
&T^{\mu}\int_{S_T}(\partial_rS^2Z\dot\Phi)^2dS+T^{\mu-2\gamma}\int_{S_T}(ZS^2Z\dot\Phi)^2dS\\
&\qquad +\int_{D_T}\bigg(r^{\mu-1-\delta}(\partial_rS^2Z\dot\Phi)^2+r^{\mu-1-2\gamma}(ZS^2Z\dot\Phi)^2\bigg)dx\\
&\le C\varepsilon^2+C\varepsilon\int_{D_T}
\ds\sum_{l=0}^3\bigg(r^{\mu+1-2\gamma+2l}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2
+r^{\mu-1-\delta+2l}(\nabla_x^l\partial_r\dot\Phi)^2\bigg)
dx\\
&\qquad +C\varepsilon\ds\sum_{l=0}^3\bigg(T^{\mu+2l}\int_{S_T}(\nabla_x^l\partial_r\dot\Phi)^2dS
+T^{\mu-2\gamma+2l}\int_{S_T}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2dS\bigg)\\
&\qquad +C\varepsilon\bigg(\int_{D_T}\big(r^{\mu-1-\delta}(\partial_rS^2Z\dot\Phi)^2+r^{\mu-1-2\gamma}(SZ^3\dot\Phi)^2\big)dx\bigg)^{\f12}
\end{align*}
and
\begin{align*}
&T^{\mu}\int_{S_T}(\partial_rSZ^2\dot\Phi)^2dS+T^{\mu-2\gamma}\int_{S_T}(ZSZ^2\dot\Phi)^2dS\\
&\qquad +\int_{D_T}\bigg(r^{\mu-1-\delta}(\partial_rSZ^2\dot\Phi)^2+r^{\mu-1-2\gamma}(ZSZ^2\dot\Phi)^2\bigg)dx\\
&\le C\varepsilon^2+C\varepsilon\int_{D_T}
\ds\sum_{l=0}^3\bigg(r^{\mu+1-2\gamma+2l}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2
+r^{\mu-1-\delta+2l}(\nabla_x^l\partial_r\dot\Phi)^2\bigg)
dx\\
&\qquad +C\varepsilon\ds\sum_{l=0}^3\bigg(T^{\mu+2l}\int_{S_T}(\nabla_x^l\partial_r\dot\Phi)^2dS
+T^{\mu-2\gamma+2l}\int_{S_T}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2dS\bigg)\\
&\qquad +C\varepsilon\bigg(\int_{D_T}\big(r^{\mu-1-\delta}(\partial_rSZ^2\dot\Phi)^2+r^{\mu-1-2\gamma}(Z^4\dot\Phi)^2\big)dx\bigg)^{\f12}
\end{align*}
This, together with (5.64) and (5.3) in the case of $m=3$ and Remark 2.3, yields (5.2) for $k=3$.}
{\bf Proof.} As in Lemma 5.4, at first we establish an analogous inequality
to (5.53) for $Z(Z_2^2+Z_3^2)\dot\Phi$. Based on this together with
the domain decomposition technique, we can complete the proof of (5.64).
Note that
$$
\partial_rZ(Z_2^2+Z_3^2)\dot\Phi=0\qquad \text{on $\varphi=\varphi_0$}.\eqno{(5.65)}
$$
Then applying Theorem 4.1 to $Z(Z_2^2+Z_3^2)\dot\Phi$ yields
\begin{align*}
&T^{\mu}\int_{S_T}(\partial_rZ(Z_2^2+Z_3^2)\dot\Phi)^2dS
+T^{\mu-2\gamma}\int_{S_T}(Z^2(Z_2^2+Z_3^2)\dot\Phi)^2dS\\
&\qquad +\int_{D_T}\bigg(r^{\mu-\gamma}(\partial_rZ(Z_2^2+Z_3^2)\dot\Phi)^2
+r^{\mu-1-2\gamma}(Z^2(Z_2^2+Z_3^2)\dot\Phi)^2\bigg)dx\\
&\le C\int_{D_T}\mathcal{L}Z(Z_2^2+Z_3^2)\dot\Phi\cdot\mathcal{M}Z(Z_2^2+Z_3^2)\dot\Phi
dx +C\varepsilon^2.\tag{5.66}
\end{align*}
By $\mathcal{L}Z(Z_2^2+Z_3^2)\dot\Phi=Z(Z_2^2+Z_3^2)\mathcal{L}\dot\Phi=Z(Z_2^2+Z_3^2)\dot f$
and the expression of $\dot f$,
we have
$$
\mathcal{L}Z(Z_2^2+Z_3^2)\dot\Phi=M_1+M_2+M_3+M_4\eqno{(5.67)}
$$
with
\begin{align*}
M_1&=f_{00}\partial_r^2Z(Z_2^2+Z_3^2)\dot\Phi+\frac{1}{r^2}\ds\sum_{1\le
i,j\le3}f_{ij}Z_iZ_jZ(Z_2^2+Z_3^2)\dot\Phi
+\frac{1}{r}\ds\sum_{i=1}^3f_{0i}\partial_rZ_iZ(Z_2^2+Z_3^2)\dot\Phi,\\
M_2&=\frac{1}{r^2}\ds\sum_{1\le
i,j\le3}f_{ij}[Z(Z_2^2+Z_3^2),Z_iZ_j]\dot\Phi
+\frac{1}{r}\ds\sum_{i=1}^3f_{0i}[Z(Z_2^2+Z_3^2),\partial_rZ_i]\dot\Phi,\\
M_3&=\ds\sum_{k=2}^3\ds\sum_{l=1}^2C_{l}Z\bigg(Z_k^{l}f_{00}Z_k^{2-l}\partial_r^2\dot\Phi
+\ds\sum_{1\le i,j\le3}Z_k^{l}(\frac{1}{r^2}f_{ij})Z_k^{2-l}Z_iZ_j\dot\Phi
+\ds\sum_{i=1}^3Z_k^{l}(\frac{1}{r}f_{0i})Z_k^{2-l}\partial_rZ_i\dot\Phi\bigg)\\
&\qquad +\bigg(Zf_{00}Z_k^2\partial_r^2\dot\Phi
+\ds\sum_{k=2}^3\ds\sum_{1\le i,j\le3}Z(\frac{1}{r^2}f_{ij})Z_k^2Z_iZ_j\dot\Phi
+\ds\sum_{i=1}^3Z(\frac{1}{r}f_{0i})Z_k^2\partial_rZ_i\dot\Phi\bigg)\\
&\qquad +Z(Z_2^2+Z_3^2)f_0^1,\\
M_4&=Z(Z_2^2+Z_3^2)f_0^2.
\end{align*}
For notational simplifications, we rewrite $M_3$ and $M_4$ as follows
\begin{align*}
M_3&=\sum\limits_{l_1+l_2\le
3,l_1\ge1}C_{l_1l_2}\bigg(Z^{l_1}f_{00}Z^{l_2}\partial_r^2\dot\Phi
+\sum\limits_{1\le i,j\le3}Z^{l_1}(\frac{1}{r^2}f_{ij})Z^{l_2}Z_iZ_j\dot\Phi
+\sum\limits_{i=1}^3Z^{l_1}(\frac{1}{r}f_{0i})Z^{l_2}\partial_rZ_i\dot\Phi\bigg)\\
&\quad +Z^3f_0^1,\\
M_4&=Z^3f_0^2.
\end{align*}
Next, we treat each term $\int_{D_T}M_i\cdot\mathcal{M}Z(Z_2^2+Z_3^2)\dot\Phi
dx$ $(1\le i\le 4)$ respectively.
\vskip 0.3 true cm
{\bf (A) The estimate on $\int_{D_T}M_1\cdot\mathcal{M}Z(Z_2^2+Z_3^2)\dot\Phi
dx$}
\vskip 0.3 true cm
As in (5.52), we can obtain
\begin{align*}
&|\int_{D_T}M_1\cdot\mathcal{M}Z(Z_2^2+Z_3^2)\dot\Phi dx|\le C\varepsilon^2+C\varepsilon\bigg(T^{\mu}\int_{S_T}(\partial_rZ(Z_2^2+Z_3^2)\dot\Phi)^2dS\\
&\qquad\quad +T^{\mu-2\gamma}\int_{S_T}(Z^2(Z_2^2+Z_3^2)\dot\Phi)^2dS+\int_{D_T}\big(r^{\mu-1-\delta}(\partial_rZ(Z_2^2+Z_3^2)\dot\Phi)^2\\
&\qquad\quad +r^{\mu-1-2\gamma}(Z^2(Z_2^2+Z_3^2)\dot\Phi)^2\big)dx\bigg).
\tag{5.68}
\end{align*}
\vskip 0.3 true cm
{\bf (B) The estimate on $\int_{D_T}M_2\cdot\mathcal{M}Z(Z_2^2+Z_3^2)\dot\Phi
dx$}
\vskip 0.3 true cm
It follows from the expressions
of $f_{ij}$, (5.1) and a direct computation that
$$
|M_2|\le C\varepsilon
\big(r^{-2\gamma}|Z^4\dot\Phi|+r^{-\sigma-1}|\partial_rZ^3\dot\Phi|\big)$$
and
$$
\int_{D_T}|M_2\cdot\mathcal{M}Z(Z_2^2+Z_3^2)\dot\Phi|dx\le C\varepsilon\int_{D_T}
\ds\sum_{l=0}^3\bigg(r^{\mu+1-2\gamma+2l}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2+r^{\mu-1-\delta+2l}(\nabla_x^l\partial_r\dot\Phi)^2\bigg)
dx.\eqno{(5.69)}
$$
\vskip 0.3 true cm
{\bf (C) The estimate on $\int_{D_T}M_3\cdot\mathcal{M}Z(Z_2^2+Z_3^2)\dot\Phi
dx$}
\vskip 0.3 true cm
Due to the lack of $L^{\infty}$ assumptions on the third order derivatives of $\dot\Phi$
in (5.1), we will decompose $M_3$ as follows
$$M_3=M_3^1+M_3^2,$$
where $M_3^1$ is linear with respect to the third and fourth order derivatives of $\dot\Phi$,
$M_3^2$ contains the products of two third order
derivatives of $\dot\Phi$, which is required to be specially treated and admits the following concrete
expression:
\begin{align*}
M_3^2=&-\frac{C_{12}}{(\hat{U}^2-c^2(\hat\rho))}\bigg\{\bigg((\gamma+1)\hat{U}\partial_rZ^2\dot\Phi
+(\gamma+1)\partial_r\dot\Phi\partial_rZ^2\dot\Phi+\frac{\gamma-1}{r^2}\ds\sum_{i=1}^3Z_i\dot\Phi
Z^2Z_i\dot\Phi\bigg)\partial_r^2Z\dot\Phi\\
&-\frac{\gamma-1}{r^2}\ds\sum_{i=1}^3\bigg(\partial_r\dot\Phi\partial_rZ^2\dot\Phi-\hat{U}\partial_rZ^2\dot\Phi
+\frac{1}{r^2}\ds\sum_{k=1}^3Z_k\dot\Phi Z^2Z_k\dot\Phi-\frac{2}{r^2(\gamma-1)}Z_i\dot\Phi
Z^2Z_i\dot\Phi\bigg)ZZ_i^2\dot\Phi\\
&+\frac{1}{r^4}\ds\sum_{1\le i\neq
j\le3}Z_i\dot\Phi Z^2Z_j\dot\Phi
ZZ_iZ_j\dot\Phi+\frac{1}{r^2}\ds\sum_{i=1}^3\bigg((\hat{U}+\partial_r\dot\Phi)Z^2Z_i\dot\Phi+Z_i\dot\Phi\partial_rZ^2\dot\Phi\bigg)\partial_rZZ_i\dot\Phi)\bigg\}
\end{align*}
By (5.1) and a direct but tedious computation, we can arrive at
$$
|M_3^1|\le C\varepsilon \big(r^{-2(\gamma-1)}\ds\sum_{0\le i \le2}|\partial_r^2Z^i\dot\Phi|+r^{-\delta-1}\ds\sum_{0\le
i\le 3}|\partial_rZ^i\dot\Phi|+r^{-2-\sigma}\ds\sum_{1\le i\le
4}|Z^i\dot\Phi|\big),
$$
which yields
$$
\int_{D_T}|M_3^1\cdot\mathcal{M}Z(Z_2^2+Z_3^2)\dot\Phi|dx\le C\varepsilon\int_{D_T}
\ds\sum_{l=0}^3\bigg(r^{\mu+1-2\gamma+2l}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2+r^{\mu-1-\delta+2l}(\nabla_x^l\partial_r\dot\Phi)^2\bigg)
dx.\eqno{(5.70)}
$$
Next, we focus on the treatment of $\int_{D_T}|M_3^2\cdot\mathcal{M}Z(Z_2^2+Z_3^2)\dot\Phi|dx$.
By the expression of $M_3^2$, one knows that the typical terms in $M_3^2\cdot\mathcal{M}Z(Z_2^2+Z_3^2)\dot\Phi$
are respectively:
$r^{\mu}\partial_rZ^2\dot\Phi\partial_r^2Z\dot\Phi\partial_r Z(Z_2^2+Z_3^2)\dot\Phi$, $r^{\mu-2}Z\dot\Phi Z^3\dot\Phi\partial_r^2Z\dot\Phi\partial_r Z(Z_2^2+Z_3^2)\dot\Phi$,
$r^{\mu-2}\partial_rZ^2\dot\Phi Z^3\dot\Phi\partial_r Z(Z_2^2+Z_3^2)\dot\Phi$,
$r^{\mu-4}Z\dot\Phi (Z^3\dot\Phi)^2\partial_r Z(Z_2^2+Z_3^2)\dot\Phi$,
$r^{\mu-2}Z^3\dot\Phi\partial_rZ^2\dot\Phi\partial_r Z(Z_2^2+Z_3^2)\dot\Phi$ and $r^{\mu-2}Z\dot\Phi (\partial_rZ^2\dot\Phi)^2\partial_r Z(Z_2^2+Z_3^2)\dot\Phi$.
By the interpolation inequalities in Corollary 2.3 and Lemma 2.6, the terms mentioned above can be treated directly.
Indeed, we have
(i)
\begin{align*}
&|r^{\mu}\partial_rZ^2\dot\Phi\partial_r^2Z\dot\Phi\partial_r Z(Z_2^2+Z_3^2)\dot\Phi|_{L^1(D_T)}\\
\le &C|r^{\delta-2(\gamma-1)}\cdot r^{\frac{8\gamma-11-\delta}{4}}\partial_rZ^2\dot\Phi\cdot r^{\frac{8\gamma-7-\delta}{4}}\partial_r^2Z\dot\Phi\cdot r^{\frac{\mu-1-\delta}{2}}\partial_rZ^3\dot\Phi|_{L_1(D_T)}\\
\le &C |r^{\frac{8\gamma-11-\delta}{4}}\partial_r^2Z\dot\Phi|_{L^4(D_T)}|r^{\frac{8\gamma-7-\delta}{4}}\partial_r^2Z\dot\Phi|_{L^4(D_T)}
|r^{\frac{\mu-1-\delta}{2}}\partial_rZ^3\dot\Phi|_{L^2(D_T)}.\qquad \text{(By $\delta\le\gamma-1$)}
\end{align*}
(ii)
\begin{align*}
&|r^{\mu-2}Z\dot\Phi Z^3\dot\Phi\partial_r^2Z\dot\Phi\partial_r Z(Z_2^2+Z_3^2)\dot\Phi|_{L^1(D_T)}\\
\le &|r^{\mu-1-\sigma}Z^3\dot\Phi\partial_r^2Z\dot\Phi\partial_rZ(Z_2^2+Z_3^2)\dot\Phi|_{L_1(D_T)}\qquad \text{(By assumption (5.1))}\\
\le &C|r^{\frac{3\delta-(\gamma-1)-\sigma}{4}}\cdot r^{\frac{4\gamma-11}{4}}Z^3\dot\Phi\cdot r^{\frac{8\gamma-7-\delta}{4}}\partial_r^2Z\dot\Phi\cdot r^{\frac{\mu-1-\delta}{2}}\partial_rZ^3\dot\Phi|_{L_1(D_T)}\\
\le &C |r^{\frac{4\gamma-11}{4}}Z^3\dot\Phi|_{L^4(D_T)}|r^{\frac{8\gamma-7-\delta}{4}}\partial_r^2Z\dot\Phi|_{L^4(D_T)}
|r^{\frac{\mu-1-\delta}{2}}\partial_rZ^3\dot\Phi|_{L^2(D_T)},
\end{align*}
here we have used $0<\delta\le\min(\gamma-1,\sigma-(\gamma-1))$ and $\sigma=\min(1,2(\gamma-1))$, which derives $3\delta-(\gamma-1)-\sigma\le0$.
(iii)
\begin{align*}
&|r^{\mu-2}\partial_rZ^2\dot\Phi Z^3\dot\Phi\partial_rZ(Z_2^2+Z_3^2)\dot\Phi|_{L_1(D_T)}\\
=& |r^{\frac{3\delta-4(\gamma-1)}{4}}\cdot r^{\frac{8\gamma-11-\delta}{4}}\partial_rZ^2\dot\Phi\cdot r^{\frac{4\gamma-11}{4}}Z^3\dot\Phi
\cdot r^{\frac{\mu-1-\delta}{2}}\partial_rZ^3\dot\Phi|_{L^1(D_T)}\\
\le &C |r^{\frac{8\gamma-11-\delta}{4}}\partial_rZ^2\dot\Phi|_{L^4(D_T)} |r^{\frac{4\gamma-11}{4}}Z^3\dot\Phi|_{L^4(D_T)}|r^{\frac{\mu-1-\delta}{2}}\partial_rZ^3\dot\Phi|_{L^2(D_T)}.\qquad \text{(By $\delta\le\gamma-1$)}
\end{align*}
(iv)
\begin{align*}
&|r^{\mu-4}Z\dot\Phi (Z^3\dot\Phi)^2\partial_r Z(Z_2^2+Z_3^2)\dot\Phi|_{L^1(D_T)}\\
\le &|r^{\mu-3-\sigma}(Z^3\dot\Phi)^2\partial_rZ(Z_2^2+Z_3^2)\dot\Phi|_{L_1(D_T)}\qquad\text{(By assumption (5.1))}\\
=&|r^{\frac{\delta-2\sigma}{2}}\cdot (r^{\frac{4\gamma-11}{4}}Z^3\dot\Phi)^2\cdot r^{\frac{\mu-1-\delta}{2}}\partial_rZ^3\dot\Phi|_{L_1(D_T)}\\
\le &C|r^{\frac{4\gamma-11}{4}}Z^3\dot\Phi|_{L^4(D_T)}^2
|r^{\frac{\mu-1-\delta}{2}}\partial_rZ^3\dot\Phi|_{L^2(D_T)}.\qquad \text{(By $\delta\le \gamma-1\le \sigma$)}\\
\end{align*}
(v)
\begin{align*}
&|r^{\mu-2}Z^3\dot\Phi\partial_rZ^2\dot\Phi \partial_rZ(Z_2^2+Z_3^2)\dot\Phi|_{L_1(D_T)}\\
=&|r^{\frac{3\delta-4(\gamma-1)}{4}}\cdot r^{\frac{4\gamma-11}{4}}Z^3\dot\Phi\cdot r^{\frac{8\gamma-11-\delta}{4}}\partial_rZ^2\dot\Phi
\cdot r^{\frac{\mu-1-\delta}{2}}\partial_rZ^3\dot\Phi|_{L_1(D_T)}\\
\le &C|r^{\frac{4\gamma-11}{4}}Z^3\dot\Phi|_{L^4(D_T)}|r^{\frac{8\gamma-11-\delta}{4}}\partial_rZ^2\dot\Phi|_{L^4(D_T)}
|r^{\frac{\mu-1-\delta}{2}}\partial_rZ^3\dot\Phi|_{L^2(D_T)}.\qquad \text{(By $\delta\le\gamma-1$)}\\
\end{align*}
(vi)
\begin{align*}
&|r^{\mu-2}Z\dot\Phi (\partial_rZ^2\dot\Phi)^2\partial_r Z(Z_2^2+Z_3^2)\dot\Phi|_{L^1(D_T)}\\
\le &|r^{\mu-1-\sigma}(\partial_rZ^2\dot\Phi)^2\partial_rZ(Z_2^2+Z_3^2)\dot\Phi|_{L_1(D_T)}\qquad\text{(By assumption (5.1))}\\
=& |r^{\delta-2(\gamma-1)-\sigma}\cdot (r^{\frac{8\gamma-11-\delta}{4}}\partial_rZ^2\dot\Phi)^2\cdot r^{\frac{\mu-1-\delta}{2}}\partial_rZ^3\dot\Phi|_{L_1(D_T)}\\
\le &C |r^{\frac{8\gamma-11-\delta}{4}}\partial_rZ^2\dot\Phi|_{L^4(D_T)}^2 |r^{\frac{\mu-1-\delta}{2}}\partial_rZ^3\dot\Phi|_{L^2(D_T)}.
\qquad \text{(By $\delta\le\gamma-1$)}\\
\end{align*}
Substituting those estimates above into $\int_{D_T}|M_3^2\cdot\mathcal{M}Z(Z_2^2+Z_3^2)\dot\Phi|dx$ and applying
Lemma 2.6-Lemma 2.7 yield
$$\int_{D_T}|M_3^2\cdot\mathcal{M}Z(Z_2^2+Z_3^2)\dot\Phi|dx\le
C\varepsilon\int_{D_T}\ds\sum_{l=0}^3\bigg(r^{\mu-1-\delta+2l}(\nabla_x^l\partial_r\dot\Phi)^2
+r^{\mu+1-2\gamma+2l}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2\bigg)dx.\eqno{(5.71)}$$
By (5.70) and (5.71), we have
$$
\int_{D_T}|M_3\cdot\mathcal{M}Z(Z_2^2+Z_3^2)\dot\Phi|dx\le C\varepsilon\int_{D_T}
\ds\sum_{l=0}^3\bigg(r^{\mu+1-2\gamma+2l}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2+r^{\mu-1-\delta+2l}(\nabla_x^l\partial_r\dot\Phi)^2\bigg)
dx.\eqno{(5.72)}
$$
\vskip 0.3 true cm
{\bf (D) The estimate on $\int_{D_T}M_4\cdot\mathcal{M}Z(Z_2^2+Z_3^2)\dot\Phi
dx$}
\vskip 0.3 true cm
Note that
\begin{align*}
M_4&=-\frac{\hat{U}}{r^3}\ds\sum_{i,j=1}^3C'_{ij}ZZ_i\dot\Phi ZZ_j\dot\Phi
-\frac{\hat{U}}{r^3}\ds\sum_{i,j=1}^3\tilde{C}_{ij}ZZ_i\dot\Phi Z^2Z_j\dot\Phi
+\frac{6(\gamma-1)}{r^3}\hat{U}\ds\sum_{i=1}^3ZZ_i\dot\Phi ZZ_i^2\dot\Phi\\
&\quad +\text{\{left terms\}}\\
&\equiv M_4^1+M_4^2
\end{align*}
and
$$|M_4^2|\le \frac{C}{r^3}|Z\dot\Phi|\ds\sum_{i=1}^4|Z^i\dot\Phi|.\eqno{(5.73)}$$
On the other hand, due to
\begin{align*}
&|r^{\mu-3}(Z^2\dot\Phi)^2\partial_rZ(Z_2^2+Z_3^2)\dot\Phi|_{L_1(D_T)}\\
\le &|r^{\frac{\delta+2(\gamma-1)-2\sigma}{2}}\cdot (r^{\frac{2\gamma+2\sigma-9}{4}}Z^2\dot\Phi)^2\cdot r^{\frac{\mu-1-\delta}{2}}
\partial_rZ_k(Z_2^2+Z_3^2)\dot\Phi|_{L_1(D_T)}\\
\le &C|r^{\frac{2\gamma+2\sigma-9}{4}}Z^2\dot\Phi|_{L^4(D_T)}^2
|r^{\frac{\mu-1-\delta}{2}}\partial_rZ_k(Z_2^2+Z_3^2)\dot\Phi|_{L^2(D_T)},\qquad\text{(By $\delta<2(\sigma-(\gamma-1))$)}\\
&|r^{\mu-3}Z^2\dot\Phi Z^3\dot\Phi\partial_rZ(Z_2^2+Z_3^2)\dot\Phi|_{L_1(D_T)}\\
\le &|r^{\frac{\delta+(\gamma-1)-\sigma}{2}}\cdot r^{\frac{2\gamma+2\sigma-9}{4}}Z^2\dot\Phi\cdot r^{\frac{4\gamma-11}{4}}Z^3\dot\Phi
\cdot r^{\frac{\mu-1-\delta}{2}}\partial_rZ(Z_2^2+Z_3^2)\dot\Phi|_{L_1(D_T)}\\
\le &C|r^{\frac{2\gamma+2\sigma-9}{4}}Z^2\dot\Phi|_{L^4(D_T)} |r^{\frac{4\gamma-11}{4}}Z^3\dot\Phi|_{L^4(D_T)}|r^{\frac{\mu-1-\delta}{2}}\partial_rZ(Z_2^2+Z_3^2)\dot\Phi|_{L^2(D_T)},
\qquad\text{(By $\delta<\sigma-(\gamma-1)$)}
\end{align*}
then together with Lemma 2.6-Lemma 2.7 we can arrive at
$$
\int_{D_T}|M_4^1\cdot\mathcal{M}Z(Z_2^2+Z_3^2)\dot\Phi|dx\le
C\varepsilon\int_{D_T}\ds\sum_{l=0}^3\bigg(r^{\mu-1-\delta+2l}(\nabla_x^l\partial_r\dot\Phi)^2
+r^{\mu+1-2\gamma+2l}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2\bigg)dx.\eqno{(5.74)}$$
Collecting (5.73) and (5.74) yields
$$
\int_{D_T}|M_4\cdot\mathcal{M}Z(Z_2^2+Z_3^2)\dot\Phi|dx\le C\varepsilon\int_{D_T}
\ds\sum_{l=0}^3\bigg(r^{\mu+1-2\gamma+2l}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2+r^{\mu-1-\delta+2l}(\nabla_x^l\partial_r\dot\Phi)^2\bigg)
dx.\eqno{(5.75)}
$$
Substituting (5.67)-(5.69), (5.72) and (5.75) into (5.66) yields
\begin{align*}
&T^{\mu}\int_{S_T}(\partial_rZ(Z_2^2+Z_3^2)\dot\Phi)^2dS
+T^{\mu-2\gamma}\int_{S_T}(Z^2(Z_2^2+Z_3^2)\dot\Phi)^2dS\\
&\qquad +\int_{D_T}\bigg(r^{\mu-1-\delta}(\partial_rZ(Z_2^2+Z_3^2)\dot\Phi)^2
+r^{\mu-1-2\gamma}(Z^2(Z_2^2+Z_3^2)\dot\Phi)^2\bigg)dx\\
&\le C\varepsilon^2+C\varepsilon\int_{D_T}
\ds\sum_{l=0}^3\bigg(r^{\mu+1-2\gamma+2l}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2
+r^{\mu-1-\delta+2l}(\nabla_x^l\partial_r\dot\Phi)^2\bigg)
dx\\
&\qquad +C\varepsilon\bigg(T^{\mu}\int_{S_T}(\partial_rZ(Z_2^2+Z_3^2)\dot\Phi)^2dS
+T^{\mu-2\gamma}\int_{S_T}(Z^2(Z_2^2+Z_3^2)\dot\Phi)^2dS\bigg).\tag{5.76}
\end{align*}
By (5.76), we have obtained the estimates of $Z^2(Z_2^2+Z_3^2)\dot\Phi$ and $\partial_rZ(Z_2^2+Z_3^2)\dot\Phi$
on $S_T$ or $D_T$. However, the related estimates on such third order derivatives $Z^3\dot\Phi$ and $\partial_rZ^2\dot\Phi$
are not obtained. Note that away from $\varphi=0$ in $\Omega$, as in (5.60), we can actually get the estimates of $Z^3\dot\Phi$ and $\partial_rZ^2\dot\Phi$ like (5.76) since
$$
\ds\sum_{i,j,k,l=1}^3|Z_iZ_jZ_kZ_l\dot\Phi|\le C\big(\ds\sum_{i,k=1}^3|Z_iZ_k(Z_2^2+Z_3^2)\dot\Phi|+\ds\sum_{i,j,k=1}^3|Z_iZ_jZ_k\dot\Phi|+\ds\sum_{i=1}^3|Z_iZ_j\dot\Phi|+\ds\sum_{i=1}^3|Z_i\dot\Phi|\big)$$
and
$$\ds\sum_{i,j,k=1}^3|\partial_rZ_iZ_jZ_k\dot\Phi|\le C\big(\ds\sum_{k=1}^3|\partial_rZ_k(Z_2^2+Z_3^2)\dot\Phi|+\ds\sum_{i,j=1}^3|\partial_rZ_iZ_j\dot\Phi|+\ds\sum_{i=1}^3|\partial_rZ_i\dot\Phi|+|\partial_r\dot\Phi|\big)
$$
hold for $\varphi>\frac{\varphi_0}{3}$.
Near $\varphi=0$, as in (5.62), applying Theorem 4.1 yields
\begin{align*}
&T^{\mu}\int_{S_T}\chi(\varphi)(\partial_rZ^3\dot\Phi)^2dS
+T^{\mu-2\gamma}\int_{S_T}\chi(\varphi)(Z^4\dot\Phi)^2dS\\
&\qquad +\int_{D_T}\bigg(r^{\mu-\gamma}\chi(\varphi)(\partial_rZ^3\dot\Phi)^2
+r^{\mu-1-2\gamma}\chi(\varphi)(Z^4\dot\Phi)^2\bigg)dx\\
&\le C\bigg(\int_{D_T}r^{\mu-1-2\gamma}|\chi'(\varphi)|(Z^4\dot\Phi)^2dx\bigg)^{\frac{1}{2}}
\bigg(\int_{D_T}r^{\mu-\gamma}|\chi'(\varphi)|(\partial_rZ^3\dot\Phi)^2dx\bigg)^{\frac{1}{2}}\\
&\qquad +\int_{D_T}\mathcal{L}Z^3\dot\Phi\cdot\chi(\varphi)\mathcal{M}Z^3\dot\Phi
dx+C\varepsilon^2,\tag{5.77}
\end{align*}
where $\chi(\varphi)$ has been defined in (5.62), and
$\chi'(\varphi)$ has a compact support away from $\varphi=0$, which implies that
the first term in the right hand side of (5.77) can be estimated as in
(5.76).
By the compact support property of $\chi(\varphi)$, similar to the treatment
on $\int_{D_T}\mathcal{L}Z(Z_2^2+Z_3^2)\dot\Phi\cdot\mathcal{M}Z(Z_2^2+Z_3^2)\dot\Phi
dx$ in (5.67), we then have
\begin{align*}
&\int_{D_T}\mathcal{L}Z^3\dot\Phi\cdot\chi(\varphi)\mathcal{M}Z^3\dot\Phi dx\\
&\le C\varepsilon^2+C\varepsilon\int_{D_T}
\ds\sum_{l=0}^3\bigg(r^{\mu+1-2\gamma+2l}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2+r^{\mu-1-\delta+2l}(\nabla_x^l\partial_r\dot\Phi)^2\bigg)
dx\\
&\qquad +C\varepsilon\ds\sum_{l=0}^3\bigg(T^{\mu+2l}\int_{S_T}(\nabla_x^l\partial_r\dot\Phi)^2dS
+T^{\mu-2\gamma+2l}\int_{S_T}(\nabla_x^l(\frac{1}{r}Z\dot\Phi))^2dS\bigg).\tag{5.78}
\end{align*}
Substituting (5.78) into (5.77) and combining the obtained estimates for $\varphi>\frac{\varphi_0}{3}$
in $\Omega$, we then complete the proof of Lemma 5.5.\qquad \qquad \qquad \qquad \qquad $\square$
Based on Lemma 5.2-Lemma 5.5 and Remark 5.1-Remark 5.5, we now start to prove Theorem 5.1.
{\bf Proof of Theorem 5.1.}
By (5.4) in Lemma 5.2, we have
$$
T^{\mu}\int_{S_T}(\partial_r \dot\Phi)^2dS+T^{\mu-2\gamma}\int_{S_T}(Z\dot\Phi)^2dS
+\int_{D_T}\bigg(r^{\mu-1-\delta}(\partial_r \dot\Phi)^2+r^{\mu-1-2\gamma}(Z \dot\Phi)^2\bigg)dx\le C\varepsilon^2.\eqno{(5.79)}
$$
By $\delta\le\gamma-1$, Lemma 2.5 and (5.79), we can complete the proof of (5.2) in the case of $k=0$.
Similarly, by Lemma 5.2-Lemma 5.5, Remark 5.1-Remark 5.5, Lemma 2.5 and Remark 2.3, we can arrive at
\begin{align*}
&\ds\sum_{l=1}^m\bigg(T^{\mu+2l}\int_{S_T}|\nabla_x^l\partial_r\dot\Phi|^2dS
+T^{\mu-2\gamma+2l}\int_{S_T}|\nabla_x^l
(\frac{1}{r}Z\dot\Phi)|^2dS\biggr)\\
&\quad +\ds\sum_{l=1}^m\int_{D_T}\biggl(r^{\mu-1-\delta+2l}|\nabla_x^l\partial_r\dot\Phi|^2+r^{\mu+1-2\gamma+2l}
|\nabla_x^l(\frac{1}{r}Z\dot\Phi)|^2\biggr)dx\\
&\le C\varepsilon^2+C\varepsilon\biggl\{\ds\sum_{l=0}^m\bigg(T^{\mu+2l}\int_{S_T}|\nabla_x^l\partial_r\dot\Phi|^2dS
+T^{\mu-2\gamma+2l}\int_{S_T}|\nabla_x^l
(\frac{1}{r}Z\dot\Phi)|^2dS\bigg)\\
&\quad +\ds\sum_{l=0}^m\int_{D_T}\biggl(r^{\mu-1-\delta+2l}|\nabla_x^l\partial_r\dot\Phi|^2+r^{\mu+1-2\gamma+2l}
|\nabla_x^l(\frac{1}{r}Z\dot\Phi)|^2\biggr)dx\bigg\}\\
&\quad+C\varepsilon\biggl(\ds\sum_{l=1}^m\int_{D_T}\big(r^{\mu-1-\delta+2l}|\nabla_x^l\partial_r\dot\Phi|^2+r^{\mu+1-2\gamma+2l}
|\nabla_x^l(\frac{1}{r}Z\dot\Phi)|^2\big)dx\biggr)^{\f12}.\tag{5.80}
\end{align*}
Then for $1\le m\le3$, it follows from (5.79)-(5.80) that
\begin{align*}
&\ds\sum_{l=1}^m\bigg(T^{\mu+2l}\int_{S_T}|\nabla_x^l\partial_r\dot\Phi|^2dS
+T^{\mu-2\gamma+2l}\int_{S_T}|\nabla_x^l
(\frac{1}{r}Z\dot\Phi)|^2dS\biggr)\\
&\quad +\ds\sum_{l=1}^m\int_{D_T}\biggl(r^{\mu-1-\delta+2l}|\nabla_x^l\partial_r\dot\Phi|^2+r^{\mu+1-2\gamma+2l}
|\nabla_x^l(\frac{1}{r}Z\dot\Phi)|^2\biggr)dx\\
&\le C\varepsilon^2+C\varepsilon\biggl(\ds\sum_{l=1}^m\int_{D_T}r^{\mu-1-\delta+2l}|\nabla_x^l\partial_r\dot\Phi|^2+r^{\mu+1-2\gamma+2l}
|\nabla_x^l(\frac{1}{r}Z\dot\Phi)|^2dx\biggr)^{\f12}.\tag{5.81}
\end{align*}
If
$$\ds\sum_{l=1}^m\int_{D_T}r^{\mu-1-\delta+2l}|\nabla_x^l\partial_r\dot\Phi|^2+r^{\mu+1-2\gamma+2l}
|\nabla_x^l(\frac{1}{r}Z\dot\Phi)|^2dx\le C\varepsilon^2,$$
then (5.2) is derived directly;
If
$$\ds\sum_{l=1}^m\int_{D_T}r^{\mu-1-\delta+2l}|\nabla_x^l\partial_r\dot\Phi|^2+r^{\mu+1-2\gamma+2l}
|\nabla_x^l(\frac{1}{r}Z\dot\Phi)|^2dx\ge C\varepsilon^2,$$
then it follows from (5.81) that
\begin{align*}
&\ds\sum_{l=1}^m\int_{D_T}\biggl(r^{\mu-1-\delta+2l}|\nabla_x^l\partial_r\dot\Phi|^2+r^{\mu+1-2\gamma+2l}
|\nabla_x^l(\frac{1}{r}Z\dot\Phi)|^2\biggr)dx\\
\le& C\varepsilon\biggl(\ds\sum_{l=1}^m\int_{D_T}r^{\mu-1-\delta+2l}|\nabla_x^l\partial_r\dot\Phi|^2+r^{\mu+1-2\gamma+2l}
|\nabla_x^l(\frac{1}{r}Z\dot\Phi)|^2dx\biggr)^{\f12},
\end{align*}
which means
$$\ds\sum_{l=1}^m\int_{D_T}\biggl(r^{\mu-1-\delta+2l}|\nabla_x^l\partial_r\dot\Phi|^2+r^{\mu+1-2\gamma+2l}
|\nabla_x^l(\frac{1}{r}Z\dot\Phi)|^2\biggr)dx\le C\varepsilon^2.$$
Substituting this into (5.81) derives (5.2) for $1\le k\le 3$ and further completes the proof of Theorem 5.1.\qquad$\square$
\vskip 0.5 true cm
{\bf $\S 6$. Proof of Theorem 1.1.}
\vskip 0.5 true cm
It follows from Sobolev's embedding theorem (see also [10, Lemma
14]) that, one has for $1\le r\le T$
\begin{equation}
\left\{
\begin{aligned}
&\sum\limits_{0\le l\le 1}|r^l\nabla_x^l(\partial_r\dot\Phi)|^2\le
Cr^{-2}\int_{S_r}\sum\limits_{0\le l\le 3}
|r^l\nabla_x^l(\partial_r\dot\Phi)|^2dS,\\
&\sum\limits_{0\le l\le
1}|r^l\nabla_x^l(\frac{1}{r}Z\dot\varphi)|^2\le
Cr^{-2}\int_{S_r}\sum\limits_{0\le l\le 3}
|r^l\nabla_x^l(\frac{1}{r}Z\dot\varphi)|^2dS.
\end{aligned}
\right.\tag{6.1}
\end{equation}
On the other hand, (5.2) shows that
$$\int_{S_r}\sum\limits_{0\le l\le
3}|r^l\nabla_x^l(\partial_r\dot\Phi)|^2dS\le C\varepsilon^2r^{-\mu},\qquad
\int_{S_r}\sum\limits_{0\le l\le
3}|r^l\nabla_x^l(\frac{1}{r}Z\dot\Phi)|^2dS\le
C\varepsilon^2r^{-\mu+2\gamma-2}.$$
Hence we arrive at
$$\sum\limits_{0\le l\le
1}|r^l\nabla_x^l(\partial_r\dot\Phi)|^2\le C\varepsilon^2 r^{-\mu-2},\qquad
\sum\limits_{0\le l\le 1}|r^l\nabla_x^l(\frac{1}{r}Z\dot\Phi)|^2\le
C\varepsilon^2r^{-\mu+2\gamma-4}.$$
Subsequently, one has
$$\sum\limits_{0\le l\le
1}|r^l\nabla_x^l(\partial_r\dot\Phi)|\le C\varepsilon r^{-\frac{\mu+2}{2}}=C\varepsilon
r^{-2(\gamma-1)},\qquad \sum\limits_{0\le l\le
+1}|r^l\nabla_x^l(\frac{1}{r}Z\dot\Phi)|\le C\varepsilon
r^{-\frac{\mu-2\gamma+4}{2}}=C\varepsilon r^{-(\gamma-1)}.$$
In addition, for $1<\gamma<2$,
$$
|Z\dot\Phi|\le|Z\dot\Phi(1,\varphi)|+\int_1^r|\partial_rZ\dot\Phi(t,\varphi)|dt\le C\varepsilon(1+r^{1-2(\gamma-1)}),
$$
which means
$$
\frac{1}{r}|Z\dot\Phi|\le C\varepsilon(r^{-1}+r^{-2(\gamma-1)})\le C\varepsilon r^{-\sigma}.
$$
In this case, by the Bernoulli's law (1.2),
we have $c^2(\rho)=
c^2(\hat\rho)-\frac{\gamma-1}{2}\big((\partial_r\dot\Phi)^2+2\hat U\partial_r\dot\Phi+\frac{1}{r^2}(Z\dot\Phi)^2\big)$,
which derives $Cr^{2(1-\gamma)}-C\varepsilon(r^{2(1-\gamma)}+r^{-2\sigma})<c^2(\rho)<Cr^{2(1-\gamma)}+C\varepsilon(r^{2(1-\gamma)}+r^{-2\sigma})$
together with Lemma 2.1. Thus, one obtains $c^2(\rho)\sim r^{2(1-\gamma)}>0$ for any $r\ge 1$ and small $\varepsilon$.
Therefore, the proof of Theorem
1.1 is completed by the local existence result in Theorem 3.1 and continuous induction method, where
the $C^{\infty}$ regularity of $\Phi$ comes from the strong continuity principle (see [19]) and the $C^{\infty}$ smoothness of initial data $(\Phi(x)|_{r=1}, \partial_r\Phi(x)|_{r=1})$ and boundary $\Sigma$.\qquad\qquad $\square$
|
\section{Introduction}
Ideal homogeneous incompressible inviscid fluid flows are governed by the Euler equations:
\[
\mathrm{(E)}
\setcounter{equation}{0}
\left\{ \aligned
&\frac{\partial v}{\partial t} +(v\cdot \nabla )v =-\nabla p ,
\quad (x,t)\in {\mathbb R^3}\times (0, \infty) \\
&\textrm{div }\, v =0 , \quad (x,t)\in {\mathbb R^3}\times (0,
\infty)\\
\endaligned
\right.
\]
where $v=(v_1, \cdots, v_n )$, $v_j =v_j (x, t)$, $j=1,\cdots,n$, $n\geq 2$, is the
velocity of the flow, $p=p(x,t)$ is the scalar pressure.
Let $R_j$, $j=1, \cdots, n$, denote the Riesz transforms, given by
$$R_j(f)(x)= C_n\lim_{\varepsilon\to 0} \int_{\mathbb R^n\setminus B_\varepsilon(x)} \frac{(x_j-y_j)f(y)}{|x-y|^{n+1}} dy,\quad C_n=\frac{\Gamma \left(\frac{n+1}{2} \right)}{\pi ^{\frac{n+1}{2}}}.
$$
The pressure in the Euler (and Navier-Stokes) equations is given in terms of the velocity up to addition of a harmonic function by
\begin{equation}\label{pre-vel} p= \sum_{j,k=1}^n R_j R_k (v_jv_k).
\end{equation}
This is easily seen by taking the divergence of (E).
In \cite{c1}(see also \cite{bm}) the following result is obtained.
\begin{thm}
If $(v,p)$ satisfies (\ref{pre-vel}) and $|p|+|v|^2 \in L^1(\mathbb R^n)$, then
\begin{equation}\label{ortho}
\int_{\mathbb R^n} v_jv_k dx=-\delta_{jk}\int_{R^n} p\,dx.
\end{equation}
\end{thm}
In the next section we present a simple proof of this result using the continuity of the Fourier transform of functions belonging to $L^1 (\mathbb R^n)$. In order to see the implications of the above theorem for Beltrami flows, let us
recall that in the stationary case in $\mathbb R^3$, the first equations of (E) can be rewritten as
\begin{equation}
v\times \o=\nabla (p+\frac12 |v|^2),\quad \o=\mathrm{curl}\, v.
\end{equation}
A vector field $v$ is called a Beltrami flow if there exists a function $\lambda=\lambda (x)$ such that
\begin{equation}\label{bel}
\o =\lambda v.
\end{equation}
Therefore, if $v$ is a Beltrami flow, then the pair $(v,p)$ is a solution of the stationary Euler equations if
\begin{equation}\label{pre}
p+ \frac12 |v|^2 =c, \quad \mbox{$c=$ constant}.
\end{equation}
We call such a solution $(v, p)$ a ``Beltrami solution'' of the stationary Euler equations.
We refer to \cite{ep} for a recent interesting result regarding the Beltrami flows.
Recently, Nadirashvili proved a Liouville type property of Beltrami flows (\cite{n}). He showed that a Beltrami solution $(v,p)$ satisfying
either $v\in L^q(\mathbb R^3)$, $2\leq q\leq 3$, or $|v(x)| =o(1/|x|)$ is necessarily trivial, $v=0$.
In the case of finite energy Beltrami flows we have the following immediate consequence of Theorem 1.1:
\begin{thm}
Let $(v,p)$ be a Beltrami solution of the stationary Euler equations with the $\lambda $ given in (\ref{bel}).
If $v\in L^2(\mathbb R^3)$, then $v=0$. The same conclusion holds, for instance, if there exists $q\in [\frac65, \infty]$ such that $v\in L^q(\mathbb R^3)$ and $\lambda \in L^{\frac{6q}{5q-6}} (\mathbb R^3)$ (if $v\in L^{\frac65} (\mathbb R^3)$, then we require $\lambda \in L^\infty (\mathbb R^3)$ ).
\end{thm}
We have also the following result for the cases considered in the paper \cite{n}, for which we present a different, simple proof.
\begin{thm}
Let $v\in L^\infty _{loc} (\Bbb R^3)$ be a Beltrami solution of the stationary Euler equations satisfying
either $v\in L^q (\Bbb R^3)$ for some $q\in [2, 3)$, or that there exists $\varepsilon >0$ such that $ |v(x)|= O(1/|x|^{1+\varepsilon})$ as $|x|\to \infty$.
Then, $v=0$.
\end{thm}
\section{Proof of the Theorems}
\setcounter{equation}{0}
We use the notation for the Fourier transform of $f(x)$
$$\mathcal{F}(f)(\xi)=\widehat{f}(\xi)=\frac{1}{(2\pi)^{\frac{n}{2}}} \int_{\mathbb R^n} f(x)e^{-ix\cdot \xi } dx,
$$
whenever the right hand side is defined.
In terms of the Fourier transform the Riesz transform is defined as
$$ \widehat{R_j (f)}(\xi)=\frac{i \xi_j}{|\xi|}, \quad i=\sqrt{-1}.
$$
\noindent{\bf Proof of Theorem 1.1 } Without loss of generality we may restrict ourselves to the stationary case, $v(x,t)=v(x), p(x,t)=p(x)$. By the Fourier transform one has
\begin{equation}\label{th1} \hat{p}(\xi)=-\sum_{j,k=1}^n \frac{\xi_j\xi_k}{|\xi|^2} \widehat{v_j v_k }(\xi).
\end{equation}
We note that $\widehat{p}(\xi)$ and $\widehat{v_j v_k}(\xi)$, $j, k =1, \cdots, n$, are continuous at $\xi=0$ from the hypothesis, $|p|+|v|^2\in L^1(\mathbb R^n)$.
Let $w$ be a given constant vector with $|w|=1$. We put $\xi= \rho w $ in (\ref{th1}), and pass $\rho \to 0$ to obtain
\begin{equation}\label{th2}
\int_{\mathbb R^n} p\, dx = -\int_{\mathbb R^n} (v\cdot w )^2 dx.
\end{equation}
If we plug $w=\mathbf{e}^j$ in (\ref{th2}), where $\mathbf{e}^j$ is the canonical basis of $\mathbb R^n$ with its components given by $(\mathbf{e}^j)_k =\delta_{jk}$, then we have
$$\int_{\mathbb R^n} p \, dx= -\int_{\Bbb R^n} v_j ^2 dx\quad \forall j=1, \cdots, n.
$$
On the other hand, for $j\neq k$, if we put $w=\frac{\mathbf{e}^j+\mathbf{e}^k}{\sqrt{2}}$ in (\ref{th2}), we obtain
$\int_{\mathbb R^n} v_j v_k dx=0$.
$\square$\\
\ \\
\noindent{\bf Proof of Theorem 1.2 } Since $(v,p)$ is a Beltrami solution of stationary Euler equations, we have $p-c=\frac12 |v|^2:=\tilde{p}$ for some constant $c$.
In the case $v\in L^2 (\mathbb R^3)$ we find that $|v|^2+|\tilde{p}|\in L^1(\mathbb R^3)$, and by Theorem 1.1 we obtain
$$\int_{\mathbb R^3} \tilde{p} \, dx= -\frac{1}{3}\int_{\mathbb R^3} |v|^2 dx=-\frac12\int_{\mathbb R^3} |v|^2 dx,
$$
which implies that $v=0$, and $\tilde{p}=\sum_{j,k=1}^n R_j R_k (v_jv_k)=0$. On the other hand, if $v\in L^{q} (\mathbb R^3)$ and $\lambda \in L^{\frac{6q}{5q-6}} (\mathbb R^3)$ with $\frac65<q\leq \infty$, or $v\in L^{\frac65} (\mathbb R^3)$ and $\lambda \in L^\infty (\mathbb R^3)$, then
we estimate
\begin{eqnarray*}
\|v\|_{L^2}&\leq& C \|\nabla v\|_{L^{\frac{6}{5}}}\leq C \| \o\|_{L^{\frac{6}{5}}}=C \| \lambda v \|_{L^{\frac{6}{5}}}\\
&\leq& C\|\lambda\|_{L^{\frac{6q}{5q-6}}} \|v\|_{L^q} <\infty,
\end{eqnarray*}
and we reduce to the above case of $v\in L^2(\Bbb R^3)$. $\square$\\
\ \\
\noindent{\bf Proof of Theorem 1.3 }
We first observe that our hypothesis implies that
\begin{equation}\label{th13a} \int_{\Bbb R^3} |v|^2 |x| ^{\lambda -2} dx <\infty.
\end{equation}
for some $\lambda \in (1, 2)$.
Indeed, the case $ |v(x)|= O(1/|x|^{1+\varepsilon})$ as $|x|\to \infty$ is obvious, while in the case $v\in L^q (\Bbb R^3)$ for some $q\in [2, 3)$, we have the following estimate,
\begin{eqnarray*}
\int_{\{|x|\geq 1\}} |v|^2|x|^{\lambda -2} dx\leq C \|v\|_{L^q}^2 \left(\int_1 ^\infty r^{2-\frac{q(\lambda -2)}{q-2}} dr\right)^{\frac{q-2}{q}}<\infty
\end{eqnarray*}
for $\lambda$ with $1< \lambda < \frac{6}{q}-1$.
Let us introduce a standard radial cut-off function $\sigma\in
C_0 ^\infty(\Bbb R^N)$ such that
\begin{equation}\label{16}
\sigma(|x|)=\left\{ \aligned
&1 \quad\mbox{if $|x|<1$},\\
&0 \quad\mbox{if $|x|>2$},
\endaligned \right.
\end{equation}
and $0\leq \sigma (x)\leq 1$ for $1<|x|<2$. Then, for each $R
>0$, we define
$
\sigma \left(\frac{|x|}{R}\right):=\sigma_R (|x|)\in C_0 ^\infty (\Bbb R^N).
$
A Beltrami solution $(v,p)$ with $p=-\frac12 |v|^2 +C$ satisfies
\begin{equation}\label{th13b}
\sum_{j,k=1}^3 \int_{\Bbb R^3} v_jv_k \partial_j\partial_k \varphi dx=\frac12 \int_{\Bbb R^3} |v|^2 \Delta \varphi dx \quad \forall \varphi \in C_0 ^2 (\Bbb R^3).
\end{equation}
We choose our test function $ \varphi (x)=\varphi_{\delta, R}(x)= (|x|^{2\lambda } +\delta)^{\frac12} \sigma_R$ for $\delta, R>0$ in (\ref{th13b}), which is an approximation of $\varphi =|x|^\lambda$, and passing
first $\delta \to 0$, and then $R\to \infty$, using the dominated convergence theorem, taking (\ref{th13a}) into account, we obtain easily that
\begin{equation}\label{th13c}
(\lambda -1)\int_{\Bbb R^3} |v|^2 |x|^{\lambda -2} dx=2(\lambda -2) \int_{\Bbb R^3} (v\cdot x)^2 |x|^{\lambda -4} dx.
\end{equation}
The fact that $\lambda \in (1,2)$ in (\ref{th13c}) implies $v=0$. $\square$
$$\mbox{\bf Acknowledgements}$$
DC was partially supported by NRF grants 2006-0093854 and 2009-0083521.
PC was partially supported by NSF DMS grants 1209394 and 1265132.
|
\section{Introduction}
The gonihedric Ising model was originally formulated by Ambartzumian and
Savvidy~\cite{discretization} as a possible lattice discretization of
{an alternative ``linear'' action for} the string worldsheet in bosonic string theory.
{These} early discretizations {using} triangulations were then translated {to
plaquette surfaces generated as the spin cluster boundaries
of classical generalized Ising spin Hamiltonians} by Savvidy and Wegner~\cite{spinssystem}.
For a recent review, see~\cite{Johnston2008}.
The {resulting} gonihedric Ising model is a generalized three-dimensional Ising model, where spins
$\spin{ }$, living on a three-dimensional cubic lattice, interact via
nearest $\nn{i}{j}$, next-to-nearest $\nnn{i}{j}$ and plaquette interactions
$\plaq{i}{j}{k}{l}$ and the weights of the different interactions are fine-tuned
so that the \emph{area} of spin cluster boundaries does not contribute to the
partition function, but {rather} their edges and self-interactions. This leads to the family of Hamiltonians
\begin{equation}
\Ham^\kappa = -2\kappa\nnsum+\frac{\kappa}{2}\nnnsum-\frac{1-\kappa}{2}\plaqsum,
\label{eq:ham:goni}
\end{equation}
where $\kappa$ parametrizes the self-avoidance of the spin cluster boundaries.
The purely plaquette Hamiltonian with $\kappa=0$,
\begin{equation}
\Ham = -\frac{1}{2}\plaqsum,
\label{eq:ham:gonikappa0}
\end{equation}
allows spin cluster boundaries to intersect without energetic penalty. It has
attracted some attention recently, as it displays a strong first-order
transition~\cite{firstorder} and evidence of glass-like behaviour~\cite{glassy}
at low temperatures in spite of the absence of quenched disorder.
The strong first-order nature of the transition for the purely plaquette
Hamiltonian has meant that its has proved difficult to obtain consistent values
for the inverse transition temperature. Only recently, it was understood that
the exponential degeneracy $q = 2^{3L}$~\cite{degeneracy} in the
low-temperature phase, for the periodic system living on a cube with linear lattice
size $L$, severely changes the finite-size corrections and leads to nonstandard
finite-size-scaling~\cite{mueller2014}.
First estimates {of} the inverse transition temperature were given by a
mean-field approximation that yielded $\beta^\infty = 0.325$ and early canonical
Monte Carlo simulations gave $\beta^\infty=0.50(1)$~\cite{Johnston1996}. {Later},
detailed finite-size scaling analysis of Monte Carlo data with {\it fixed}
boundary conditions found $\beta^\infty = 0.54757(63)$~\cite{Baig2004}. More
recently, another value of $0.510(2)$, that is apparently compatible with the
first simulations~\cite{Johnston1996}, was suggested by analysing a dual
representation of the model with periodic boundary
conditions which turns out to be an anisotropic Ashkin-Teller
model~\cite{Johnston2011}. Here, two spins $\sigma, \tau$ live on each vertex
of a cubic lattice, with nearest-neighbour interactions along the $x,y$
and $z$-axes,
\begin{equation}
\Ham^d = -\frac{1}{2}\sum\limits_{\nn{i}{j}_x}\sigma_i\sigma_j -
\frac{1}{2}\sum\limits_{\nn{i}{j}_y}\tau_i\tau_j
-\frac{1}{2}\sum\limits_{\nn{i}{j}_z}\sigma_i\sigma_j\tau_i\tau_j.
\label{eq:ham:dual}
\end{equation}
To resolve these discrepancies, we have conducted multicanonical
Monte Carlo simulations of the original model and its dual
representation. {In the remainder of the paper we present the results of these simulations
and the underlying theory of finite-size scaling at first order phase transitions used in extracting the conclusions.}
{
In section 2, after first discussing ``standard'' first-order finite-size scaling behaviour for models with
periodic boundary conditions, we observe that nonstandard first-order finite-size scaling behaviour will occur when there is an exponentially large
degeneracy of the low-temperature phase, which is the case for both the plaquette Hamiltonian and
its dual. The scaling corrections for various observables are presented in detail in this case.}
{
In section 3 we discuss the simulations themselves, for the plaquette Hamiltonian
with periodic boundary conditions, for the dual model with periodic boundary conditions and for the plaquette Hamiltonian
with fixed boundary conditions.
We determine characteristic
quantities of these systems, such as the specific heat, Binder's energy
parameter, the latent heat and interface tension, as well as the inverse
temperature of the transition. We find for the first time, with the use of the nonstandard scaling relations, a consistent estimate
of the inverse critical temperature from the plaquette Hamiltonian and its dual with periodic boundary conditions and the plaquette Hamiltonian with fixed boundary conditions. The self-consistency of the simulations is also confirmed by extracting various prefactors of scaling corrections both directly from fits and by calculation of energy moments.}
{
Finally, in section 4 we summarize our results for the measured physical quantities and note that the latent heat of the transition in the gonihedric model appears to be boundary condition dependent.}
\section{Finite-Size Scaling for First-Order Phase Transitions}
\label{sec:fss}
In spite of the ubiquity of first-order phase transitions~\cite{ubiquity} it
was only relatively recently that the initial studies of finite-size scaling
for first-order transitions were carried out~\cite{pioneer} and subsequently
pursued further in~\cite{furtherd}. Rigorous results for periodic boundary
conditions were derived in \cite{rigorous-fss,rigorous-fss-potts}. For the
discussion of scaling laws under periodic boundary conditions here, we will
first employ the two-state model of~\cite{twostatemodel} which is capable of
correctly reproducing the prefactors of the leading contributions. To have this
paper reasonably self-contained, we will recall the principles and main results
in the following. In this model we assume the system {spends} some time in
either one of the $q$ ordered phases or in the disordered phase, where
transitions between the phases are instantaneous and fluctuations within the
phases are neglected. Let $W_{\rm o}$ denote the fraction of the total time {spent} in
one of the $q$ ordered phases and $W_{\rm d} = 1 - W_{\rm o}$ {the fraction spent} in the
disordered phase. We associate energies $e_{\rm o}$ and $e_{\rm d}$
with the phases.
Under {these} assumptions, the energy moments are simply given by
$\left<e^n\right> = W_{\rm o}e_{\rm o}^n + (1-W_{\rm o})e_{\rm d}^n$.
The specific heat
$C_V(\beta, L) = -\beta^2\partial e(\beta, L)/\partial\beta$ as an expression
{in terms} of {these} moments becomes
\begin{equation}
C_V(\beta, L) = L^d\beta^2\left(\left< e^2\right> - \left< e \right>^2\right) =
L^d\beta^2 W_{\rm o}(1-W_{\rm o})\Delta e^2,
\label{eq:specheat}
\end{equation}
with $\Delta e = e_{\rm d} - e_{\rm o}$.
Varying $W_{\rm o}$, we find the maximum
\begin{equation}
C^{\rm max} = L^d (\beta\Delta e/2)^2 \approx L^d (\beta^\infty\Delta \hat{e}/2)^2 \label{eq:fss:specheatmax}
\end{equation} for $W_{\rm o} = W_{\rm d} = 0.5$, where the {ordered
and disordered} peaks of the energy probability density have equal weight.
Here, we have assumed that $\beta$, $e_{\rm o}$ and $e_{\rm d}$ deviate from the
infinite-volume limit $\beta^\infty$, {$\hat{e}_{\rm o} = e_{\rm o}(\beta^\infty)$} and $\hat{e}_{\rm d}$,
respectively, only by terms of order {$1/L^d$}.
In close analogy, we can write the Binder parameter in terms of the two-state moments
\begin{eqnarray}
B(\beta, L) &=& 1 - \frac{\langle e^4\rangle}{3\langle e^2 \rangle^2} = 1 - \frac{ W_{\rm o}\hat{e}_{\rm o}^4 + (1-W_{\rm o})\hat{e}_{\rm d}^4}{3\left(W_{\rm o}\hat{e}_{\rm o}^2 + (1-W_{\rm o})\hat{e}_{\rm d}^2\right)^2}\;,\label{eq:binder}
\end{eqnarray}
from which we can calculate the minimum with respect to the weights at $W_{\rm
o} = \hat{e}_{\rm d}^2/(\hat{e}_{\rm o}^2 + \hat{e}_{\rm d}^2)$, such that
$W_{\rm o}/W_{\rm d} = \hat{e}_{\rm d}^2/\hat{e}_{\rm o}^2$ with a value of
\begin{eqnarray}
B^{\rm min}(L) = 1 - \frac{1}{12}\left( \frac{\hat{e}_{\rm o}}{\hat{e}_{\rm d}} +
\frac{\hat{e}_{\rm d}}{\hat{e}_{\rm o}} \right)^2.\label{eq:fss:bindermin}
\end{eqnarray}
The free-energy densities, $f_{\rm o}$ and $f_{\rm d}$, of any one of the
ordered phases or the disordered phase govern their probability
\begin{equation}
p_{\rm o}\propto {\rm e}^{-\beta L^d f_{\rm o}}\mbox{ and } p_{\rm d} \propto
{\rm e}^{-\beta L^d f_{\rm d}},
\end{equation}
and the fraction of time spent in the ordered phases is proportional to $q p_{\rm o}$.
Neglecting {exponentially} small corrections in the linear lattice size \mbox{$L$~\cite{rigorous-fss,rigorous-fss-potts,twostatemodel}},
the ratio of the fraction of time spent in the respective phases is given by
\begin{equation}
W_{\rm o}/W_{\rm d} \simeq q{\rm e}^{- L^d \beta f_{\rm o}}/{\rm e}^{-\beta L^d f_{\rm d}} \; .
\end{equation}
Expanding the logarithm $\ln (W_{\rm o}/W_{\rm d}) = \ln q + L^d\beta(f_{\rm d} - f_{\rm o})$ around the infinite-volume phase-transition temperature $\beta^\infty$
leads to
\begin{equation}
\ln (W_{\rm o}/W_{\rm d}) = \ln q + L^d\Delta \hat{e}(\beta - \beta^\infty) + \dots \, ,
\end{equation}
which after truncation {can be solved} for the inverse temperatures $\beta$.
For $W_{\rm o} = W_{\rm d} = 0.5$ we find the inverse temperature $\beta^{\rm eqw}$, where both peaks of the energy probability density have equal weight, and, to leading order, the location $\beta^{C_V^{\rm max}}$ of the specific heat maximum,
\begin{equation}
\beta^{C_V^{\rm max}}(L) = \beta^{\rm eqw}(L) = \beta^\infty - \frac{\ln q}{\Delta\hat{e}L^{d}} + \dots \, .
\end{equation}
The minimum of Binder's parameter at $W_{\rm o}/W_{\rm d} = \hat{e}_{\rm
d}^2/\hat{e}_{\rm o}^2$ is located at the inverse temperature
\begin{eqnarray}
\beta^{B^{\rm min}}(L) = \beta^\infty - \frac{\ln(q\hat{e}^2_{\rm
o}/\hat{e}^2_{\rm d})}{L^d\Delta \hat{e}} + \dots \, .
\end{eqnarray}
In spite of its simplicity the model captures the
essential features of first-order phase transitions and, importantly for our purposes, correctly predicts the
prefactors of the leading finite-size scaling corrections for a class of models
with a contour representation, such as the $q$-state Potts model, where a
completely rigorous theory of scaling also exists~\cite{rigorous-fss-potts}.
This {rigorous theory} enables the calculation of the coefficients of higher-order terms in a
systematic asymptotic expansion in powers of $1/L^d$~\cite{twostatemodel,lee-kosterlitz}.
In addition, there are further corrections that decay exponentially fast with
growing system size~\cite{exponentialcorr}.
These models for periodic boundary conditions have, up to {exponentially small}
corrections in $L$, a canonical partition function of the form{~\cite{rigorous-fss-potts}}
\begin{equation}
Z\left(\beta, L\right) = q {\rm e}^{-\beta L^d f_o(\beta)} + {\rm e}^{-\beta L^d f_d(\beta)},
\end{equation}
allowing a more rigorous derivation of inverse transition temperatures.
The inverse temperature of equal peak weight then reads{~\cite{twostatemodel}}
\begin{eqnarray}
\beta^{\rm eqw}(L)
&\!\!=&\!\! \beta^\infty\! - \beta^\infty\frac{\ln q}{\Delta\hat{s}L^{d}}
+ \beta^\infty\left( \frac{\ln q}{\Delta\hat{s}L^d}\right)^2
\left(\frac{\Delta\hat{C}}{2\Delta\hat{s}} \right)
+ {\cal O}\left(\frac{(\ln q)^3}{L^{3d}} \right)\!,\label{eq:fss:beta:eqw}
\end{eqnarray}
where $\Delta\hat{s} = \beta^{\infty}\Delta\hat{e}$ is the transition entropy
and $\Delta\hat{C} = \hat{C}_d - \hat{C}_o$. For the location of the
specific-heat maximum and the minimum of the Binder parameter one finds
{~\cite{twostatemodel,lee-kosterlitz}}
\begin{eqnarray}
\beta^{C^{\rm max}_V}(L)
&=& \beta^\infty - \beta^\infty\frac{\ln q}{\Delta\hat{s}L^{d}}
+ \beta^\infty\left( \frac{1}{\Delta\hat{s}L^d}\right)^2
\left(\frac{\Delta\hat{C}}{2\Delta\hat{s}}\left( \left(\ln q\right)^2 -12 \right)
+ 4\right) \nonumber\\
&& +\ {\cal O}\left(\frac{(\ln q)^3}{L^{3d}} \right), \label{eq:fss:beta:specheat}\\
\beta^{B^{\rm min}}(L)
&=& \beta^\infty - \beta^\infty\frac{\ln(q\hat{e}^2_{\rm
o}/\hat{e}^2_{\rm d})}{L^d\Delta \hat{s}} + \frac{a_2}{L^{2d}} +
{\cal O}\left( \frac{(\ln(q\hat{e}^2_{\rm o}/\hat{e}^2_{\rm d}))^3}{L^{3d}}\right),
\label{eq:fss:beta:binder}
\end{eqnarray}
where the {expression} for $a_2$ is {explicitly} known but very
complicated~\cite{twostatemodel}, and will simplify in our special case (see
below).
The leading correction {to} the inverse temperature of equal peak height,
$\beta^{\rm eqh}$, comes {more} heuristically from a double gaussian approximation of
the energy probability density~\cite{twostatemodel},
\begin{equation}
\beta^{\rm eqh}(L) = \beta^\infty - \beta^\infty\frac{\ln (q \hat{C}_{\rm d}/\hat{C}_{\rm o})}{2\Delta\hat{s}L^{3}} + {\cal O}\left(\frac{(\ln (q \hat{C}_{\rm d}/\hat{C}_{\rm o}))^2}{L^{6}}\right).
\end{equation}
Usually, the low-temperature degeneracy $q$ is a constant
{(like in a $q$-state Potts model)} and standard
finite-size scaling behaviour at a first-order transition displays a leading
contribution proportional to the inverse volume $L^{-d}$. However, for the
three-dimensional gonihedric Ising model~\eref{eq:ham:goni}, the degeneracy $q$
is {exponentially} dependent on the linear system size.
By construction, the model shows a highly degenerate ground-state for
all parameters $\kappa$. In the special case of vanishing energetic penalty for
self-intersecting spin cluster boundaries, $\kappa=0$, the degeneracy,
\begin{equation}
q = 2^{3L} = e^{3 L \ln 2},
\label{eq:degeneracy}
\end{equation}
is apparent even for finite temperatures~\cite{degeneracy}.
{
{The} usual $1/L^{3}$ behaviour {is} {therefore} transmuted to {a} $1/L^{2}$ behaviour~\cite{mueller2014}. {Furthermore}, the
{finite-size scaling} corrections {to $\beta^{\rm eqw}(L)$} in \eref{eq:fss:beta:eqw} and in the scaling {law}
\eref{eq:fss:beta:specheat} {for $\beta^{C^{\rm max}_V}(L)$} now coincide up to order ${\cal O}(L^{-4})$,}
\begin{eqnarray}
\beta^{C^{\rm max}_V}(L) \approx \beta^{\rm eqw}(L)
&=& \beta^\infty - \frac{\ln 2^{3L}}{\Delta\hat{e}L^{3}}
+ \frac{\Delta\hat{C}}{2\Delta\hat{e}}
\left( \frac{\ln 2^{3L}}{\Delta\hat{s}L^3}\right)^2
+ {\cal O}\left(\frac{(\ln q)^3}{L^9} \right)\nonumber \\
&=& \beta^\infty - \frac{3\ln 2}{\Delta\hat{e}L^{2}}
+ \frac{\Delta\hat{C}}{2\Delta\hat{e}}
\left( \frac{3\ln 2}{\Delta\hat{s}L^2}\right)^2
+ {\cal O}(L^{-6})\label{eq:fss:beta:eqw2}\label{eq:fss:beta:specheat2}\,.
\end{eqnarray}
The scaling {law} for the peak location of Binder's
parameter~\eref{eq:fss:beta:binder} becomes
\begin{eqnarray}
\beta^{B^{\rm min}}(L)
&\!=&\! \beta^\infty - \frac{\ln(2^{3L}\hat{e}^2_{\rm o}/
\hat{e}^2_{\rm d})}{\Delta \hat{e}L^3} +
\frac{\Delta\hat{C}}{2\Delta\hat{e}}
\left( \frac{\ln 2^{3L}}{\Delta\hat{s}L^3}\right)^2 + \dots \nonumber\\
&\!=&\! \beta^\infty - \frac{3\ln2}{\Delta \hat{e}L^2}
- \frac{\ln(\hat{e}^2_{\rm o}/\hat{e}^2_{\rm d})}{\Delta\hat{e}L^3}
+ \frac{\Delta\hat{C}}{2\Delta\hat{e}}
\left( \frac{3\ln 2}{\Delta\hat{s}L^2}\right)^2\!\! + {\cal O}(L^{-6}),
\label{eq:fss:beta:binder2}
\end{eqnarray}
where we have used the fact that only the contribution to $a_2$ with the highest
power of $\ln q$, $a_2 = (\ln q/\Delta\hat{s})^2 \Delta\hat{C}/2\Delta\hat{e} + \dots$,
{contributes} to the order given.
The leading contribution {to} the finite-size correction is thus also proportional to
$L^{-2}$, and the prefactor of the contribution ${\cal O}(L^{-4})$ becomes the
same {as that} found for the inverse temperatures of {the} equal peak weight and the peak
location of the specific heat. {Note that there is, however, an additional correction term of ${\cal O}(L^{-3})$.}
The leading correction {to} the inverse temperature of equal peak height,
$\beta^{\rm eqh}$, is now also {of} order ${\cal O}\left( L^{-2} \right)$,
\begin{equation}
\beta^{\rm eqh}(L) = \beta^\infty - \frac{3\ln 2}{\Delta\hat{e}L^{2}} - \frac{\ln(\hat{C}_{\rm d}/\hat{C}_{\rm o})}{2\Delta\hat{e}L^{3}} + {\cal O}(L^{-4}).
\label{eq:fss:beta:eqh}
\end{equation}
The extremal values of the specific heat and Binder's parameter change with
the system size according to
\begin{eqnarray}
C^{\rm max}(L)
&\!\!=&\!\!\left( \frac{\Delta\hat{s}}{2}\right)^2\!L^3 +
\frac{\ln q(\Delta{\hat{C}} - \Delta{\hat{s}})}{2} +
\frac{\hat{C}_{\rm d} + \hat{C}_{\rm o}}{2} + \dots\nonumber\\
&\!\!=&\!\! \left( \frac{\Delta\hat{s}}{2}\right)^2\!L^3 +
\frac{3\ln 2(\Delta{\hat{C}}-\Delta{\hat{s}})}{2}L +
\frac{\hat{C}_{\rm d} + \hat{C}_{\rm o}}{2} + {\cal O}(L^{-1}) \label{eq:fss:specheat}
\end{eqnarray}
and
\begin{equation}
B^{\rm min}(L) = 1 - \frac{1}{12}\left( \frac{\hat{e_{\rm o}}}{\hat{e_{\rm d}}} +
\frac{\hat{e_{\rm d}}}{\hat{e_{\rm o}}} \right)^2 + aL^{-2} + {\cal O}(L^{-3}).
\label{eq:fss:binder}
\end{equation}
The {prefactor in the} first correction {for $B^{\rm min}(L)$} reads
\begin{equation}
a = \left( \frac{1}{\beta^\infty} \right)^2 \frac{ 3\ln 2(\hat{e}_{\rm d} + \hat{e}_{\rm o}) (\hat{C}_{\rm o} \hat{e}_{\rm d} - \hat{C}_{\rm d} \hat{e}_{\rm o}) (\hat{e}_{\rm d}^2 + \hat{e}_{\rm o}^2)}{6\hat{e}_{\rm d}^3 \hat{e}_{\rm o}^3},\label{eq:fss:binder:a2}
\end{equation}
which comes from an even more complicated expression {of} ${\cal O}(L^{-3})$
in the general case~\cite{twostatemodel}. {Here} we {have} already {taken} the degeneracy $q=2^{3L}$ into account.
The Taylor series of the energy in the {ordered and disordered} phases, $e_{\rm o/d}$, around $\beta {=}\beta^\infty$ reads
\begin{equation}
e_{\rm o/d}(\beta) = \hat{e}_{\rm o/d} + \frac{\partial e_{\rm o/d}}{\partial \beta}\bigg|_{\beta=\beta^\infty}(\beta - \beta^\infty) + {\cal O}\left(\left( \beta - \beta^\infty \right)^2 \right),
\end{equation}
where the specific heat of the {ordered and disordered} phase enters the leading correction. Calculating the energies at inverse temperature $\beta^{\rm eqw}$, the scaling of the energy {fulfils}
\begin{equation}
e_{\rm o/d}(\beta^{\rm eqw}) = \hat{e}_{\rm o/d} + \hat{C}_{\rm o/d}\left(\frac{1}{\beta^\infty}\right)^2(\beta^{\rm eqw} - \beta^\infty) + {\cal O}\left(\left( \beta^{\rm eqw} - \beta^\infty \right)^2\right).
\end{equation}
The difference $\beta^{\rm eqw} - \beta^\infty$ is known from \eref{eq:fss:beta:eqw2}, therefore the expression
\begin{equation}
e_{\rm o/d}(\beta^{\rm eqw}) = \hat{e}_{\rm o/d} + \left(\frac{1}{\beta^\infty}\right)^2\hat{C}_{\rm o/d} \frac{3\ln(2)}{\Delta\hat{e} L^2} + {\cal O}\left( L^{-4} \right)
\label{eq:fss:energies}
\end{equation}
shows the {finite-size} corrections {to} the energy.
The same change of leading contributions is
apparent in the dual model~\eref{eq:ham:dual}, where a similar low-temperature
phase degeneracy is expected (but, in contrast to the original model, not
proven). These considerations will also apply to other models with periodic
boundary conditions which have a degeneracy that depends {exponentially} on the system size.
For fixed boundary conditions, surface effects play an important
role~\cite{fssfbc}. Here, the inverse transition temperature of the gonihedric
Ising model is shifted by a leading term of order $\mathcal{O}(L^{-1})$ for
finite lattices of linear lattice size $L$. Thus in this case we expect
\begin{equation}
\beta(L) =\beta^\infty - \frac{a_1}{L} + {\mathcal O}(L^{-2})
\end{equation}
for the peak locations of both the specific heat and Binder's parameter.
\section{Simulation Results}
An effective way of combating supercritical slowing down near first-order phase
transitions, where canonical simulations tend to get trapped in either the
disordered or ordered phases, is to use the multicanonical Monte Carlo
algorithm~\cite{muca,mucajanke}. At such first order transitions cooling a system down or
heating it up also makes estimation of the transition temperature inherently
difficult using standard algorithms due to hysteresis effects. In a
multicanonical simulation it is possible to systematically improve guesses of
the energy probabilty distribution before the actual production run by using
recursive estimates~\cite{mucaweights}.
The rare states lying between the disordered and ordered phases in the energy
histogram are then promoted artificially, decreasing the autocorrelation time
and allowing the system to oscillate more rapidly between phases. Canonical
estimators can then be retrieved by weighting the multicanonical data to yield
Boltzmann-distributed energies. Such reweighting techniques are very powerful
when combined with the multicanonical simulations, allowing the calculation of
observables over a broad range of temperatures.
\subsection{Observables}
Standard observables such as the specific heat
\eref{eq:specheat}
and Binder's energy parameter~\eref{eq:binder} were calculated at
different temperatures from our data for both the gonihedric Ising
model~\eref{eq:ham:gonikappa0} and its dual~\eref{eq:ham:dual}. The
positions of their peaks, $\beta^{C_V^{\rm max}}(L)$ and $\beta^{B^{\rm
min}}(L)$ were then determined by the reweighting techniques~\cite{reweighting}.
Other estimates of the inverse critical temperature are given by $\beta^{\rm eqw}$ and
$\beta^{\rm eqh}$, where the disordered and ordered peaks of the energy
probability density $p(e)$ have the same weight or height, respectively. In
practice, we use reweighting techniques to get an estimator of the energy
probability densities at different temperatures. $\beta^{\rm eqw}$ is chosen
systematically to minimize
\begin{equation}
D^{\rm eqw}(\beta) = \left( \sum_{e < e_{\rm min}} p(e, \beta) - \sum_{e \geq
e_{\rm min}} p(e, \beta) \right)^2
\end{equation}
where the energy of the minimum between both peaks, $e_{\rm min}$, is
determined beforehand to distinguish between the different phases.
The location of the minimum, $\beta^{\rm eqw}$, is then used to calculate the
energy moments of the {ordered and disordered} phases,
\begin{eqnarray}
e^k_{\rm o}(L) = \sum_{e < e_{\rm min}} e^k\, p(e, \beta^{\rm eqw}) \Big/ \sum_{e < e_{\rm min}} p(e, \beta^{\rm eqw}),\nonumber\\
e^k_{\rm d}(L) =
\sum_{e \geq e_{\rm min}} e^k\, p(e, \beta^{\rm eqw}) \Big/\sum_{e \geq e_{\rm min}} p(e, \beta^{\rm eqw}),
\label{eq:e-eqw}
\end{eqnarray}
where $e_{\rm o/d}(L) = e^1_{\rm o/d}(L)$ is the energy in the respective phases, and their difference is an estimator of the latent heat $\Delta e(L)=e_{\rm d}(L)-e_{\rm o}(L)$. Also, the second and first moment combine to give the specific heat of the disordered and ordered phases,
\begin{equation}
C_{\rm o/d}(L) = \beta^2L^d\left( e^2_{\rm o/d}(L) - \left(e^1_{\rm o/d}(L)\right)^2\right).
\end{equation}
To find the inverse transition temperature where both phases have equal
height we minimize
\begin{equation}
D^{\rm eqh}(\beta) = \left( \max_{e < e_{\rm min}}\{p(e, \beta)\} - \max_{e
\geq e_{\rm min}}\{p(e, \beta)\} \right)^2,
\end{equation}
{as a function of $\beta$.}
The probability density $p(e, \beta^{\rm eqh})$ itself
at $\beta^{\rm eqh}$ is also of interest since one can make use of it to extract the reduced interface tension
\begin{equation}
\sigma(L) = \frac{1}{2L^2} \ln \left( \frac { \max \{ p(e, \beta^{\rm eqh})
\} } { \min \{ p(e, \beta^{\rm eqh}) \} } \right),
\label{eq:interface-tension}
\end{equation}
for periodic boundary conditions. This characteristic quantity of first-order
phase transitions is almost impossible to extract reliably from canonical
Metropolis simulations, since getting reasonable statistics on the suppressed
states is very hard. Multicanonical simulations, on the other hand, are
perfectly tailored for measurements of such rare events.
\subsection{Original plaquette model with periodic boundary conditions}
The quality of the simulations for the original plaquette
model~\eref{eq:ham:gonikappa0} with periodic boundary conditions {can be} judged
by the integrated autocorrelation time $\tau^{\rm int}$ and the number of
{sweeps}
in Figure~\ref{fig:tauorig}. Here, $\tau^{\rm int}$ has been
determined with a self-consistent cutoff at $6\tau^{\rm int}$ and the error
comes from the known formula for this algorithm~\cite{dataanalysis},
$\epsilon_{\tau^{\rm int}} = \left( \tau^{\rm int} \right)^{3/2}\left(
\frac{24}{n} \right)^{1/2}$, where {$n$ is the number of measurements ($=$ number of sweeps when performing measurements every sweep).}
We would expect that the integrated autocorrelation time with perfect
multicanonical weights should in principle
increase linearly with the volume,
$\tau^{\rm int} \propto L^3$.
\begin{figure}[hbt]
\begin{center}
\includegraphics{tau_meas_orig}
\caption{
The integrated autocorrelation time $\tau^{\rm int}$ in units of sweeps for
the original model with periodic boundary conditions (blue markers; left
axis) over the linear lattice size $L$. The dashed and dotted curves are
fits {with} a power law or an exponential law for lattices sizes $L\geq 16$.
On the right axis, with green markers, the total number
of {sweeps} $n$ {divided by the volume} is shown.}
\label{fig:tauorig}
\end{center}
\end{figure}
Error-weighted nonlinear least-squares fits
{of a power law, $\tau^{\rm int} \propto L^{\alpha}$,
to the actual measured
integrated autocorrelation times}
yield
{much larger} exponents $\alpha \approx 6.40(17)$ that vary a bit around $6.0$ depending on
the lattice sizes included in the fits. Also for those fits with acceptable
$\chi^2 \approx 1$ that only include lattices with size $L\geq 16$, fits to an
exponential growth with $L$ are of comparable quality, {see Figure~\ref{fig:tauorig}}. With least-squares {fits} and no proper model
testing, we cannot distinguish between the two alternatives~\cite{clauset2009}.
{In any case}, we find that the autocorrelation time grows significantly faster than
expected, an effect that may be attributed to free-energy barriers.
Such hidden barriers appear for instance in droplet
condensation~\cite{freeEbarriers}, whose analog {with} the gonihedric Ising
model is, however, still unclear.
For each lattice, we performed a maximum number of
$n^{\rm max}=2^{17}\,L^3=131072\,L^3$ sweeps with an upper bound on the
computer time of around $500$ hours of real time for each lattice size. All
lattices with size $L\leq 20$ finished the desired number of sweeps, the larger
lattices were aborted after $500$ hours and collected {correspondingly} less statistics. The
ratio $n/\tau^{\rm int}$ is a {measure} for the number of effectively
uncorrelated {data}. Although the autocorrelation time increases
dramatically with the system size, the simulation of the largest lattice of
$V=27^3$ spins still effectively flipped more than 250 times between the two
phases during the simulation. This is remarkable, given that rare states were
suppressed by more than 60 orders of magnitude compared to the most probable
states (see the inset of Figure~\ref{fig:fit:orig-pbc}).
From our multicanonical data, we have extracted the resulting quantities of
interest for different lattice sizes and listed them in
Table~\ref{tab:originalresults}, where errors have been extracted
by jackknife analysis using $20$ blocks for each lattice size~\cite{jackknife}.
We find that the inverse temperatures of the specific-heat maximum
$\beta^{C_V^{\rm max}}$ and of equal peak weights $\beta^{\rm eqw}$ {fall nearly}
together for all lattice sizes. This {is accounted for by the fact} that the
higher-order corrections of order ${\cal O}\left(L^{-4}\right)$ in the scaling
{law \eref{eq:fss:beta:eqw2}} for these quantities
collapse {due to} the {exponential} degeneracy of the low-temperature phase
to induce the same prefactor.
\begin{landscape} \begin{table}[htpb] \tablesize \centering
\caption{Resulting quantities for the gonihedric Ising
model~\eref{eq:ham:gonikappa0} with \emph{periodic\/} boundary conditions:
extremal values for the specific heat $C_V^{\rm max}$, Binder's energy
parameter $B^{\rm min}$, with their respective pseudo-critical inverse
temperatures $\beta$, and temperatures where peaks of the energy
probability density have equal heights and weights for finite lattices
with linear length $L$. The finite-lattice interface tension is listed as
$\sigma$, the energy of the ordered and disordered phases and their
difference as $e_{\rm o}, e_{\rm d}$ and $\Delta e$.
The infinite lattice limits are listed as parameters of fits whose
goodness-of-fit value $Q$ is given. Highlighted in light grey are the
datapoints used for fits with only leading order correction (lo).
Additional datapoints used for fitting with subleading corrections (so)
up to and including order ${\cal O}(L^{-4})$ are marked in dark grey.}
\vspace{3ex} \begin{tabular}{l*{10}{r}} \toprule
\multicolumn{1}{c}{$L$} & \multicolumn{1}{c}{$\beta^{C_V^{\rm max}}$} &
\multicolumn{1}{c}{$C_V^{\rm max}$} & \multicolumn{1}{c}{$\beta^{B^{\rm
min}}$} & \multicolumn{1}{c}{$B^{\rm min}$} &
\multicolumn{1}{c}{$\beta^{\rm eqw}$} & \multicolumn{1}{c}{$\beta^{\rm
eqh}$} & \multicolumn{1}{c}{$\sigma$} & \multicolumn{1}{c}{$e_{\rm o}$} &
\multicolumn{1}{c}{$e_{\rm d}$} & \multicolumn{1}{c}{$\Delta e$} \\
\midrule
\input{original_results_full_data.tex}
\bottomrule \end{tabular}
\label{tab:originalresults} \end{table} \end{landscape}
{Least-squares fits of} the data in Table~\ref{tab:originalresults}
according to the laws in section~\ref{sec:fss} have been conducted.
We have left out the smaller lattices systematically for each fit,
until a goodness-of-fit value of at least $Q = 0.5$ was found
for each observable individually.
The same protocol was employed earlier~\cite{mueller2014} for a reduced time
series, where only every $L^3$-th measurement was used. There we
were not challenging the prefactors of higher-order corrections {so}
the reduced time series was sufficient.
We list all the fit parameters obtained for {both} the full time series and the
reduced one in Table~\ref{tab:resultingfits} along with the quality-of-fit
parameters $Q$ and the degrees of freedom left.
The inverse transition temperatures in the thermodynamic limit
are {effectively identical} and do not depend on whether we use the reduced or the
full dataset {or} on the precise final averaging procedure. A graphical
representation of the best fits is given in
Figure~\ref{fig:fit:orig-pbc}.
\begin{figure}[htb]
\begin{center}
\includegraphics[scale=0.98]{orig_pbc_fit}
\caption{
Best fits up to order ${\cal O}(L^{-4})$ obtained for the original model
with periodic boundary conditions (cf. Table~\ref{tab:resultingfits})
using the (finite lattice) peak
locations for the specific heat $C_V^{\rm max}$, Binder's energy
parameter $B^{\rm min}$; or inverse temperatures $\beta^{\rm eqw}$ and
$\beta^{\rm eqh}$, where the two peaks of the energy probability density
are of same weight or have equal height, respectively. The values for
$\beta^{\rm eqw}$ and $\beta^{C^{\rm max}_V}$ are indistinguishable in
the plot. The inset shows the energy probability density $p(e)$ over $e =
E/L^d$ at $\beta^{\rm eqh}$ for lattices with linear length $L\in \left\{
9, 10, \dots, 27 \right\}$. }
\label{fig:fit:orig-pbc}
\end{center}
\end{figure}
\begin{table}[btp]
\tablesize
\centering
\caption{Resulting parameters of the best fits to the extremal points $\beta$
for the specific heat $C_V^{\rm max}$, Binder's energy parameter $B^{\rm
min}$; or inverse temperatures $\beta^{\rm eqw}$ and $\beta^{\rm eqh}$ to
laws of the form $\beta(L) = \beta^\infty + p_2/L^2 + p_3/L^3 + p_4/L^4$.
Parameters $p_i$ not used in the specific fit are marked with --. The
error-weighted average over all four inverse temperatures are listed as
$\beta^{\rm av}$, whereas $\beta^{\rm av}_\text{w/o eqw}$ is the average,
where $\beta^{\rm eqw}$ is left out, because it is strongly correlated {with}
$\beta^{C^{\rm max}_V}$, and would effectively weight this value twice.
}
\vspace{3ex}
\input{fitting_results.tex}
\label{tab:resultingfits}
\end{table}
\clearpage
Since the inverse temperatures $\beta^{\rm eqw}$ and $\beta^{C_V^{\rm max}}$
are obviously strongly correlated, we leave out the former and average
over $\beta^{C_V^{\rm max}}, \beta^{B^{\rm min}}$, and $\beta^{\rm eqh}$,
neglecting cross-correlations~\cite{combinefitsweigel} between those.
Our best estimate {of} the inverse transition temperature is then given by
\begin{eqnarray}
\beta^\infty &= 0.551\,332(8) &\qquad\textrm{original model, periodic bc,}
\end{eqnarray}
which accounts for the higher-order scaling corrections up to
${\cal O}\left(L^{-4}\right)$.
Although the inverse transition temperatures do not change, we employ the full
data set. The reason is that the error on $\beta^{\rm eqh}$ becomes smaller
for the time series that uses the full, correlated data set. This is attributed
to the fact that the observable relies on the statistics in single bins of the
energy histogram, which in total becomes smoother when using more, correlated
measurements. The same argument is valid for the calculation of the interface
tension{~\eqref{eq:interface-tension}, for which the best fit
with corrections of order ${\cal O}(L^{-2})$ yields a value of}
\begin{eqnarray}
\hat\sigma = 0.12037(18) &\qquad\textrm{original model, periodic bc.}
\end{eqnarray}
Moments of the energy in the pure ordered and disordered phases are
also expected to become more accurate using the full data set, since
autocorrelation times in the pure phases are {then} significantly smaller
{than $\tau^{\rm int}$ for the full energy range (see below).
By fitting the scaling law \eref{eq:fss:energies} to these moments, one obtains the latent
heat in the infinite-volume limit,}
\begin{eqnarray}
\Delta\hat{e} = 0.85148(5) &\qquad\textrm{original model, periodic bc.}
\end{eqnarray}
Taking a careful look at the scaling laws in section~\ref{sec:fss}, we find
that the prefactors of the scaling corrections only depend on the moments of
the energy or their differences. We have two methods at hand to test the
self-consistency of our simulations. {Firstly, since} the statistics of the observables
{are} very high, we can retrieve the prefactors of the corrections as
parameters of (nonlinear) least-squares-fits with all corrections up to and
including ${\cal O}(L^{-4})$. Secondly, from multicanonical simulations we get
estimators~\eref{eq:e-eqw} of the energy moments, allowing a direct
computation of those prefactors.
{In addition}, we {carried out} independent canonical simulations for the original model
under periodic boundary conditions for very large lattices. The goal was to get
independent measurements {of} the moments of the energy in the ordered and
disordered phases. Here we prepared the system in the appropriate phase and
performed the simulations at a fixed temperature $\beta = 0.5513$, near the
transition temperature, exploiting {the fact} that {in canonical simulations}, for large lattices, flips between the
two phases are extremely unlikely.
Of course, {this} was only possible after having determined the transition
temperature with high accuracy by the multicanonical simulations.
The quality of the canonical measurements and estimators on the energy and the
specific heat are summarized in Table~\ref{tab:canonicalresults}. The
autocorrelation times within the phases are significantly smaller, because the
system {does} not traverse suppressed, {improbable} states between the phases. The
statistical error has again been retrieved by jackknife analysis. For
lattices with size $L\geq 32$, physical quantities indicate no further
dependence on the lattice size within the error. Therefore we can safely take
the error-weighted averages over energy moments and their differences for those
lattices. The multicanonical and the canonical estimates {of} energetic moments
agree astonishingly well.
With use of the energy moments {from} both simulations, we can challenge the
prefactors {in} {the finite-size scaling laws}
numerically by comparing the numerical values for the fit parameters
to the {theoretically} expected prefactors {in terms of the energy moments.}
The results of this cross-check {are collected} in Table~\ref{tab:consistency}.
\begin{table}[htpb]
\tablesize
\centering
\caption{Quality and resulting quantities of the canonical simulations. In the time series of $n_{\rm meas} = 1048576$ measurements we found autocorrelation times $\tau < 7$ (in units of measurements), leading to approximately $n$ uncorrelated measurements. The energy $\hat{e}_{\rm o/d}$ and the specific heat $\hat{C}_{\rm o/d}$ in the different phases have been measured by preparing two independent systems for each respective phase at inverse temperature $\beta = 0.5513$. We give the error-weighted average over lattices $L\geq32$, where the dependence on the lattice size is smaller than the statistical error. The last line gives infinite-volume limits from the multicanonical simulations for comparison.}
\vspace{3ex}
\begin{tabular}{l*{8}{l}}
\toprule
\multicolumn{1}{c}{$L$} &
\multicolumn{1}{c}{$n$} &
\multicolumn{1}{c}{$\hat{e}_o$} &
\multicolumn{1}{c}{$\hat{e}_d$} &
\multicolumn{1}{c}{$\Delta \hat{e}$} &
\multicolumn{1}{c}{$\hat{C}_o$} &
\multicolumn{1}{c}{$\hat{C}_d$} &
\multicolumn{1}{c}{$\Delta \hat{C}$}
\\
\midrule
16 & 149796 & $-1.468406(29)$ & $-0.61804(13) $ & 0.85036(21) & 0.1645(5) & 0.861(4) & 0.696(6)\\
32 & 174762 & $-1.468362(10)$ & $-0.61741(4) $ & 0.85095(6) & 0.1645(6) & 0.8464(29) & 0.682(5)\\
48 & 149796 & $-1.468360(6) $ & $-0.617382(19)$ & 0.850978(30) & 0.1658(5) & 0.8445(27) & 0.679(5)\\
64 & 174762 & $-1.468367(6) $ & $-0.617401(15)$ & 0.850966(24) & 0.1656(6) & 0.847(4) & 0.681(6)\\
\midrule
\multicolumn{2}{l}{average $(L\geq 32)$} & $-1.468364(4)$ & $-0.617396(11)$ & 0.850968(18) & 0.16534(30) & 0.8458(18) & 0.6805(27) \\
\multicolumn{2}{l}{multicanonical} & $-1.468373(12)$ & $-0.61771(6)$ & 0.85148(5) & 0.16414(15) & 0.8410(12) & 0.6769(17)\\
\bottomrule
\end{tabular}
\label{tab:canonicalresults}
\end{table}
\clearpage
Employing the scaling relation for the specific-heat
maximum~\eref{eq:fss:specheat}, we can calculate $\Delta\hat{e}$ from the fit
parameter of the leading contribution. Using our estimate {of} $\beta^{\infty} =
0.551\,332(8)$, we get $\Delta\hat{e} = 0.85130(7)$, very close to our estimate
$0.850968(18)$ from the moments of the canonical simulations.
The leading correction to the specific-heat ansatz~\eref{eq:fss:specheat} has a
prefactor which computes to {$0.2197(17)$ from the canonical moments}.
The fits find $0.17(6)$, which is compatible, {if} not quite accurate.
The minimum of the
Binder parameter~\eref{eq:fss:binder} for the infinite lattice is
found to be $0.34729(7)$ from the direct fit {of} our multicanonical data which
{agrees} within error bars {with} $0.34723(9)$ from the canonical and $0.347(4)$
from the multicanonical energy moments. The first correction {in}
Eq.~\eref{eq:fss:binder:a2} yields a value of $-9.195(14)$ when inserting the
energy {moments} from the multicanonical simulation. The fits find $-9.12(4)$
which is very close.
\begin{table}
\tablesize
\centering
\caption{Resulting prefactors of the finite-size scaling corrections of the original model, retrieved by fitting the ansatz, compared to direct calculations from estimators for the energy $\hat{e}_{\rm o/d}$ and specific heat $\hat{C}_{\rm o/d}$ of the ordered and disordered phases. In the multicanonical simulations these moments were determined by finite-size scaling of $e_{\rm o/d}(L), c_{\rm o/d}(L)$; and in the canonical case by measuring time series directly at $\beta = 0.5513\simeq\beta^\infty$.}
\vspace{3ex}
\begin{tabular}{*{6}{l}}
\toprule
input &
$\Delta\hat{e}$ &
$\frac{3\ln(2)}{\Delta \hat{e}}$ &
$\frac{2\ln(\hat{e}_{\rm o}/\hat{e}_{\rm d})}{\Delta \hat{e}}$ &
$\frac{\ln(\hat{C}_{\rm d}/\hat{C}_{\rm o})}{2\Delta \hat{e}}$ &
$B^{\rm min}_{L\rightarrow\infty}$\\
\midrule
fit on $C^{\rm max}_{V}(L)$ & 0.85130(7) & -- & -- & -- & --\\
fit on $B^{\rm min}(L)$ & -- & -- & -- & -- & 0.34729(7) \\
fit on $\beta^{C^{\rm max}_{V}}(L)$ & 0.8771(14) & $2.371(4)$ & -- & -- & --\\
fit on $\beta^{B^{\rm min}}(L)$ & 0.871(14) & $2.39(4)$ & 1.6(8) & -- & --\\
fit on $\beta^{\rm eqw}(L)$ & 0.8770(14) & $2.371(4)$ & -- & -- & --\\
fit on $\beta^{\rm eqh}(L)$ & 0.84(4) & $2.47(11)$ & -- & 2.8(2.4) & --\\
fit on $\beta^{C^{\rm max}_{V}} - \beta^{B^{\rm min}}$ & -- & -- & 2.03469(7) & -- & --\\
fit on $\beta^{C^{\rm max}_{V}} - \beta^{\rm eqh}$ & -- & -- & -- & {0.892(14)} & --\\
\midrule
\multicolumn{6}{l}{energy moments from simulations \dots}\\
\multicolumn{1}{r}{\dots multicanonical
& 0.85148(5) & $2.44215(15)$ & 2.03649(27) & 0.9594(10) & 0.347(4) \\
\multicolumn{1}{r}{\dots canonical
& 0.850968(18) & $2.44362(6)$ & 2.03625(6) & 0.9591(16) & 0.34723(9) \\
\bottomrule
\end{tabular} \label{tab:consistency}
\end{table}
The coefficient of the leading correction apparent for all inverse temperatures,
$p_2 = 3\ln(2)/\Delta \hat{e}$, agrees reasonably well for the
fits on $\beta^{B^{\rm min}}$ and $\beta^{\rm eqh}$ {(cf.~Table~\ref{tab:consistency})}. The fits on
$\beta^{\rm eqw}$ and $\beta^{C^{\rm max}_V}$ yield a slope of $2.371({4})$ which
within error bars is slightly off {from the value of} our best estimate of $2.44362(6)$ {from the energy moments}.
The relative error between the two values is very small though, around $3\%$,
which is acceptable given that the leading contribution probably accounts for
the omitted higher-order contributions with different sign and that we {have}
neglected all exponential corrections.
The second leading correction of order ${\cal O}(L^{-3})$ of $\beta^{B^{\rm
min}}$ has a prefactor of the form $2\ln(\hat{e}_{\rm o}/\hat{e}_{\rm d})/\Delta \hat{e}$, which
we expect to have a value of $2.03625(6)$ from the energy moments. The
corrections of fourth order {to} $\beta^{B^{\rm min}}, \beta^{C^{\rm max}_V}$ and
$\beta^{\rm eqw}$ are supposed to be identical from the analytical expansion in
section~\ref{sec:fss}. The fits of the inverse temperatures $\beta^{C^{\rm
max}_V}$ and $\beta^{\rm eqw}$ in Table~\ref{tab:resultingfits} suggest a value
around $17$ for {the ${\cal O}(L^{-4})$ contribution}. For the lattice sizes
accessible to the multicanonical approach, that is of the same absolute order of
magnitude (but different sign) compared to the third-order contribution. Therefore
they should, in principle, compensate each other. This is reflected by the fact
that the second-order contribution $p_2$ of $\beta^{B^{\rm min}}$ is closest to
the expected one. In accordance, fitting the observable to the law
$\beta^{\infty} + p_2/L^2 + p_3/L^3$ gives a prefactor $p_3 = 0.65(12)$ with
the wrong sign compared to the theoretical prediction~\eref{eq:fss:beta:binder2},
compensating the next contribution. We therefore also {looked} at the fit
including the fourth term of order ${\cal O}(L^{-4})$. Not taking the {explicit values}
too {seriously}, we find $p_3 = -1.6(8),\, p_4 = 13(4)$, which reflect qualitatively
the compensation of those contributions for our lattice sizes at hand. Overall,
we must conclude that least-squares fitting {cannot be pushed any further} given our
statistics.
The observation that $\beta^{C^{\rm max}_V}$ and $\beta^{B^{\rm min}}$ have the
same ${\cal O}(L^{-4})$-contribution can be exploited (implicitly also making
use of the cross-correlations) by looking at their difference.
Here, we expect from \eref{eq:fss:beta:specheat2} and
\eref{eq:fss:beta:binder2} a remainder of $2\ln(\hat{e}_{\rm o}/\hat{e}_{\rm d})/L^3\Delta
\hat{e} + {\cal O}(L^{-5})$ for the scaling. In fact, fitting the difference gives a
prefactor $2.03469(7)$, in {excellent} agreement with $2.03625(6)$ {from the formula}, where the
relative error between the two is less than $0.1\%$.
The difference $\beta^{C^{\rm max}_{V}} - \beta^{\rm eqh}$ should give the
third correction {to} $\beta^{\rm eqh}$, {which} reads $p_3 = \ln(\hat{C}_{\rm
d}/\hat{C}_{\rm o})/2\Delta \hat{e}$. The fit yields $0.892(14)$ with $Q=0.98$
and $8$ degrees of freedom left, which differs from {$0.9594(10)$} by about
$7\%$.
Finally, we can also compare the numerical values for the correction to the
energies via~\eref{eq:fss:energies}. For $\hat{e}_{\rm o}$, we find a prefactor
of $1.329(5)$ from the specific heat $\hat{C}_{\rm o}$, compared to the value
of $1.397(5)$, for $\hat{e}_{\rm d}$ a value of $6.80(3)$ compared to
$6.09(4)$, which is roughly $10\%$ off.
The overall consistency of our results is very good, given that we neglected
all exponential corrections. No estimates for the prefactors differ by more than
$10\%$, and the {various}
estimates of the inverse transition temperature are insensitive to the actual
fitting protocol we use. This clearly demonstates that the first
correction terms are properly predicted by the simple two-state model even in
the case of models with an exponential degeneracy of the low-temperature phase.
The earlier canonical Monte Carlo simulations of the original plaquette model
yielded values of $\beta^\infty = 0.50(1)$~\cite{Johnston1996} and more
recently canonical simulations of the dual model~\eref{eq:ham:dual} gave
$\beta^\infty = 0.510(2)$~\cite{Johnston2011}.
Another previous estimate for the infinite-lattice inverse transition
temperature, reported by Baig et al.~\cite{Baig2004} from canonical simulations
using fixed boundary conditions, $\beta^\infty = 0.54757(63)$, is much closer
to the results here.
We therefore measured the inverse transition temperature again using
multicanonical simulations for both the dual model~\eref{eq:ham:dual} under
periodic boundary conditions and the original model~\eref{eq:ham:gonikappa0}
with fixed boundary conditions. We resolve those inconsistencies, as we
show in the following.
\subsection{Dual model with periodic boundary conditions}
For the dual model, we performed a number of $n^{\rm max} = 4\times 10^6$ sweeps
and took measurements every sweep for even lattices up to $L=24$.
The inverse temperatures of the dual model were fitted to laws with the
leading correction of order ${\cal O}(L^{-2})$, which should be well covered by
our data. The best fits on the inverse temperatures are shown in
Figure~\ref{fig:fit:dual}, where we used the data in
Table~\ref{tab:dualresults} that also lists the other quantities of interest.
\begin{figure}[h]
\begin{center}
\includegraphics[scale=0.98]{dual_fit}
\caption{Best
fits obtained for the dual model with periodic boundary conditions using
the (finite lattice) peak locations for the specific heat $C_V^{\rm
max}$, Binder's energy parameter $B^{\rm min}$; or inverse temperatures
$\beta^{\rm eqw}$ and $\beta^{\rm eqh}$, where the two peaks of the
energy probability density are of same weight or have equal height,
respectively.
The inset shows the energy probability density $p(e)$ over $e = E/L^d$ at
$\beta^{\rm eqh}$
{for lattices of linear size $L\in \left\{ 12, 14, \dots, 24 \right\}$.} }
\label{fig:fit:dual}
\end{center}
\end{figure}
\begin{landscape} \begin{table}[htpb] \tablesize \centering
\caption{ Simulation results for the dual model~\eref{eq:ham:dual}: extremal
values for the specific heat $C_V^{\rm max}$, Binder's energy parameter
$B^{\rm min}$, with their respective pseudo-critical inverse temperatures
$\beta$, and temperatures where peaks of the energy probability density
have equal heights and weights for finite lattices with linear length
$L$. The finite-lattice interface tension is listed as $\sigma$, the
energy of the ordered and disordered phases and their difference as
$e_{\rm o}, e_{\rm d}$ and $\Delta e$. Light grey cells mark the values
used for fitting, so that the goodness-of-fit parameter $Q>0.5$. If $Q <
0.5$ for all fits, we took that one with the largest $Q$.}
\vspace{3ex}
\begin{tabular}{l*{10}{r}} \toprule \multicolumn{1}{c}{$L$} &
\multicolumn{1}{c}{$\beta^{C_V^{\rm max}}$} & \multicolumn{1}{c}{$C_V^{\rm
max}$} & \multicolumn{1}{c}{$\beta^{B^{\rm min}}$} &
\multicolumn{1}{c}{$B^{\rm min}$} & \multicolumn{1}{c}{$\beta^{\rm eqw}$} &
\multicolumn{1}{c}{$\beta^{\rm eqh}$} & \multicolumn{1}{c}{$\sigma$} &
\multicolumn{1}{c}{$e_{\rm o}$} & \multicolumn{1}{c}{$e_{\rm d}$} &
\multicolumn{1}{c}{$\Delta e$} \\ \midrule
\input{dual_results_full_data.tex}
\bottomrule
\end{tabular}\label{tab:dualresults}
\end{table} \end{landscape}
Since the inverse temperatures $\beta^{\rm eqw}$ and $\beta^{C_V^{\rm max}}$
are again obviously strongly correlated, we leave out the former and average
over $\beta^{C_V^{\rm max}}, \beta^{B^{\rm min}}$, and $\beta^{\rm eqh}$,
neglecting cross-correlations~\cite{combinefitsweigel} between those.
We then find the error weighted averages,
\begin{eqnarray}
\beta^\infty_{\rm dual} &= 1.313\,28(12) &\qquad\textrm{dual model, periodic bc,}
\end{eqnarray}
for the inverse transition temperatures of the models. The error is taken
as the smallest error of the contributing estimates.
The temperature $\beta^\infty_{\rm dual}$ of the dual model is related to the temperature in
the original model, $\beta^\infty$, by the duality transformation
\begin{equation}
\beta^\infty = -\ln\left( \tanh\left( \frac{\beta^\infty_{\text{dual}}}{2}
\right) \right). \label{eq:duality}
\end{equation}
Applying standard error propagation, we retrieve a value of
\begin{eqnarray}
\beta^\infty &=
0.551\,43(7) \qquad &\textrm{from duality, periodic bc}
\end{eqnarray}
for the original model. This agrees very well with $0.551\,332(8)$, {obtained}
from the direct simulation, considering that higher-order corrections in the
dual model are omitted and additional exponential
{corrections~\cite{rigorous-fss-potts,twostatemodel,exponentialcorr}} in the
finite-size scaling were ignored completely for both models.
We argue that the earlier estimates on the transition temperature were clearly
hampered by strong hysteresis effects. Apart from the locations of the
hysteresis branches being cooling-rate dependent, it is hard to estimate the
transition temperature reliably from their locations. This is illustrated in
Figure~\ref{fig:hysteresis}, where the multicanonical data of the dual model is
located between the two hysteresis branches.
Such effects are very {difficult} to tackle using canonical
Monte Carlo data, as already remarked on by the authors of
Ref.~\cite{Johnston2011}.
{
For the interface tension~\eref{eq:interface-tension} of the dual model we find
$\sigma = 0.1214(13)$. This value agrees very well with the interface tension of
the original model, $\sigma = 0.12037(18)$, which raises interesting questions
about the duality of the model.}
\begin{figure}[htb] \begin{center} \includegraphics[]{hysterese}
\caption{Strong hysteresis effects in the dual gonihedric
Ising model with periodic boundary conditions. The linear lattice size is
$L=14$ for comparison with Fig.~10 of Ref.~\cite{Johnston2011}. One can
see that heating the system up (decreasing the inverse temperature
$\beta$) or cooling it down (increasing $\beta$) can lead to strong
hysteresis effects in the energy. Our multicanonical data lies in between
both branches of the hysteresis curve, but not in the centre, as one may
heuristically assume.}
\label{fig:hysteresis} \end{center} \end{figure}
\subsection{Original plaquette model with fixed boundary conditions}
The remaining open question is the difference between our {inverse transition temperature} compared to the
value {in} the same model under fixed boundary conditions. In the thermodynamic
limit, we expect a system to be independent of its boundary conditions, since
the boundaries grow like a surface, whereas the system size grows with the
volume.
We therefore reinvestigated the model~\eref{eq:ham:gonikappa0} using the multicanonical
algorithm for such fixed boundary conditions.
\begin{landscape} \begin{table}[htpb] \tablesize \centering
\caption{Extremal points of the gonihedric Ising
model~\eref{eq:ham:gonikappa0} for the specific heat $C_V^{\rm
max}$, Binder's energy parameter $B^{\rm min}$, with their respective
pseudo-critical inverse temperatures $\beta$ for finite lattices with
linear length $L$ contained in a box with \emph{fixed} boundary conditions. For
each lattice size, a number of $n_{\rm prod} = 10^6$ measurements were
taken. Errors have been calculated with jackknife error estimation using
$20$ blocks.} \vspace{3ex} \begin{tabular}{l*{10}{r}} \toprule
\multicolumn{1}{c}{$L$} & \multicolumn{1}{c}{$\beta^{C_V^{\rm max}}$} &
\multicolumn{1}{c}{$C_V^{\rm max}$} & \multicolumn{1}{c}{$\beta^{B^{\rm
min}}$} & \multicolumn{1}{c}{$B^{\rm min}$} &
\multicolumn{1}{c}{$\beta^{\rm eqw}$} & \multicolumn{1}{c}{$\beta^{\rm
eqh}$} & \multicolumn{1}{c}{$\sigma$} & \multicolumn{1}{c}{$e_{\rm o}$} &
\multicolumn{1}{c}{$e_{\rm d}$} & \multicolumn{1}{c}{$\Delta e$} \\
\midrule
08 & 0.44699(14) & 5.29(4) & 0.44233(13) & 0.63030(26) & 0.4475(18) & 0.4479(8) & 0(0.002) & $-1.67(5) $ & $-1.29(6) $ & 0.387(7) \\
09 & 0.45794(13) & 7.63(8) & 0.45488(14) & 0.6305(4) & 0.4594(12) & 0.4577(11) & 0(0.003) & $-1.65(4) $ & $-1.28(5) $ & 0.379(10) \\
10 & 0.46715(15) & 11.08(12) & 0.46510(16) & 0.6293(5) & 0.4678(6) & 0.46683(26) & 0(0.003) & $-1.623(21) $ & $-1.230(29) $ & 0.394(9) \\
11 & 0.47465(14) & 15.81(14) & \hllg{0.47323(14)} & 0.6273(4) & 0.4752(5) & 0.47403(25) & 0.0044(8) & $-1.62(4) $ & $-1.21(6) $ & 0.410(20) \\
12 & 0.48097(14) & 22.70(26) & \hllg{0.47995(14)} & 0.6234(6) & 0.48136(19) & 0.4803(4) & 0.0060(13) & $-1.625(9) $ & $-1.192(19) $ & 0.433(11) \\
13 & 0.48629(9) & \hllg{31.56(30)} & \hllg{0.48552(9)} & 0.6197(6) & 0.48653(9) & 0.48566(9) & 0.0072(14) & $-1.6230(7) $ & $-1.1654(25)$ & 0.4575(27) \\
14 & 0.49086(8) & \hllg{43.7(5)} & \hllg{0.49027(9)} & 0.6146(7) & 0.49099(11) & \hllg{0.49020(27)} & 0.0090(7) & $-1.620(8) $ & \hllg{$-1.129(17) $} & 0.490(9) \\
15 & 0.49483(13) & \hllg{56.8(8)} & \hllg{0.49435(13)} & 0.6116(9) & 0.49490(17) & \hllg{0.49429(12)} & 0.0097(11) & $-1.612(14) $ & \hllg{$-1.104(28) $} & \hllg{0.508(14)} \\
16 & \hllg{0.49825(11)} & \hllg{75.2(8)} & \hllg{0.49786(11)} & 0.6059(8) & 0.49830(11) & \hllg{0.49778(10)} & 0.0119(4) & $-1.6108(5) $ & \hllg{$-1.080(4) $} & \hllg{0.531(4)} \\
17 & \hllg{0.50126(7)} & \hllg{92.6(8)} & \hllg{0.50093(7)} & 0.6043(7) & \hllg{0.50129(7)} & \hllg{0.50090(19)} & \hllg{0.0125(20)} & $-1.6046(5) $ & \hllg{$-1.0664(27)$} & \hllg{0.5382(25)} \\
18 & \hllg{0.50410(6)} & \hllg{115.3(11)} & \hllg{0.50383(6)} & \hllg{0.6008(8)} & \hllg{0.50412(6)} & \hllg{0.50381(24)} & \hllg{0.0135(19)} & $-1.59885(24) $ & \hllg{$-1.0487(28)$} & \hllg{0.5501(28)} \\
19 & \hllg{0.50648(8)} & \hllg{141.9(12)} & \hllg{0.50625(8)} & \hllg{0.5969(8)} & \hllg{0.50649(8)} & \hllg{0.50620(14)} & \hllg{0.0147(15)} & $-1.59286(27) $ & \hllg{$-1.0311(24)$} & \hllg{0.5618(24)} \\
20 & \hllg{0.50866(10)} & \hllg{167.3(16)} & \hllg{0.50846(10)} & \hllg{0.5963(10)} & \hllg{0.50867(10)} & \hllg{0.50841(12)} & \hllg{0.0156(13)} & $-1.58807(27) $ & \hllg{$-1.0248(30)$} & \hllg{0.5633(29)} \\
21 & \hllg{0.51063(8)} & \hllg{200.7(19)} & \hllg{0.51046(8)} & \hllg{0.5930(9)} & \hllg{0.51064(8)} & \hllg{0.51039(7)} & \hllg{0.01706(26)} & $-1.5828(4) $ & \hllg{$-1.0106(27)$} & \hllg{0.5722(28)} \\
22 & \hllg{0.51258(6)} & \hllg{235.1(21)} & \hllg{0.51243(6)} & \hllg{0.5914(9)} & \hllg{0.51259(6)} & \hllg{0.51237(5)} & \hllg{0.0173(12)} & $-1.57837(20) $ & \hllg{$-1.0024(26)$} & \hllg{0.5759(26)} \\
23 & \hllg{0.51433(7)} & \hllg{275.6(24)} & \hllg{0.51420(7)} & \hllg{0.5889(10)} & \hllg{0.51433(7)} & \hllg{0.51415(7)} & \hllg{0.01905(26)} & $-1.57402(20) $ & \hllg{$-0.9920(27)$} & \hllg{0.5820(26)} \\
24 & \hllg{0.51586(6)} & \hllg{317(4) } & \hllg{0.51574(6)} & \hllg{0.5879(11)} & \hllg{0.51586(6)} & \hllg{0.51568(6)} & \hllg{0.0193(6)} & $-1.57031(14) $ & \hllg{$-0.9864(28)$} & \hllg{0.5839(28)} \\
25 & \hllg{0.51728(7)} & \hllg{368.6(21)} & \hllg{0.51718(7)} & \hllg{0.5847(7)} & \hllg{0.51729(7)} & \hllg{0.51716(6)} & \hllg{0.0200(10)} & $-1.56641(24) $ & \hllg{$-0.9750(18)$} & \hllg{0.5914(17)} \\
26 & \hllg{0.51853(6)} & \hllg{422.2(28)} & \hllg{0.51843(6)} & \hllg{0.5827(8)} & \hllg{0.51853(6)} & \hllg{0.51837(15)} & \hllg{0.0207(11)} & \hllg{$-1.56259(18) $} & \hllg{$-0.9670(20)$} & \hllg{0.5956(20)} \\
27 & \hllg{0.51985(7)} & \hllg{472(4) } & \hllg{0.51977(7)} & \hllg{0.5830(8)} & \hllg{0.51985(7)} & \hllg{0.51971(6)} & \hllg{0.0210(14)} & \hllg{$-1.55980(25) $} & \hllg{$-0.9656(20)$} & \hllg{0.5942(20)} \\
28 & \hllg{0.52084(6)} & \hllg{543(5) } & \hllg{0.52077(6)} & \hllg{0.5795(12)} & \hllg{0.52084(6)} & \hllg{0.52073(9)} & \hllg{0.0220(4)} & \hllg{$-1.55660(11) $} & \hllg{$-0.9544(28)$} & \hllg{0.6022(28)} \\
29 & \hllg{0.52198(24)} & \hllg{603(12)} & \hllg{0.52191(24)} & \hllg{0.5796(23)} & \hllg{0.52198(24)} & \hllg{0.52190(13)} & \hllg{0.023(5)} & \hllg{$-1.5538(9) $} & \hllg{$-0.953(6) $} & \hllg{0.601(7)} \\
\midrule
\rowcolor{lightgrey} $\infty$ & 0.55119(11) & 0.0327(6)$L^3$ & 0.55146(7) & 0.5444(14) & 0.55119(12) & 0.55152(12) & 0.0281(7) & $-1.4782(27)$ & $-0.790(4)$ & 0.694(4) \\
\rowcolor{lightgrey} $Q$ & 0.53 & 0.53 & 0.56 & 0.59 & 0.52 & 0.53 & 0.99 & 0.49 & 0.73 & 0.50 \\
\bottomrule
\end{tabular}
\label{tab:fixedresults} \end{table} \end{landscape}
For our simulations we enclosed $L^3$ free spins in a larger cube with frozen
outer planes, so the whole system contained $(L+2)^3$ spins. Our results are
listed in Table~\ref{tab:fixedresults}. We performed linear regression on the
peak locations $\beta(L)$ of the specific heat and Binder's parameter according
to the law \begin{equation} \beta^\infty = \beta(L) + \frac{a_1}{L} +
\frac{a_2}{L^2} \end{equation} that was also used by
Baig~et~al.~\cite{Baig2004}, and fitted the inverse temperatures.
The statistical errors of the constant $a_2$ turned out to be of
the same order as the value itself, therefore we set $a_2=0$ for the fits in
Table~\ref{tab:fixedresults}, intentionally neglecting the contribution
$\mathcal{O}(L^{-2})$. The best fits and the energy probability density are
shown in Figure~\ref{fig:fit:orig-fbc} and the weighted average of inverse
transition temperatures is given by:
\begin{eqnarray}
\beta^\infty &= 0.551\,38(5) \qquad & \textrm{original model, fixed bc}.
\end{eqnarray}
\begin{figure}[htb] \begin{center} \includegraphics[]{orig_fbc_fit}
\caption{
Best fits obtained for the gonihedric Ising model using fixed boundary
conditions using the (finite lattice) peak locations for the specific
heat $C^{\rm max}$, Binder's energy parameter $B_{\rm min}$. The inset
shows the energy probability density $p(e)$ over $e = E/L^3$ at
$\beta^{\rm eqh}$ for lattices with a number of
{$L\in \left\{ 11, 12, \dots, 29 \right\}$}
free spins in each dimension.} \label{fig:fit:orig-fbc}
\end{center} \end{figure}
This estimate of the inverse transition temperature is thus in excellent
agreement with the other results obtained here from multicanonical simulations
with periodic boundary conditions for the gonihedric Ising model
$\beta^\infty = 0.551\,332(8)$ and the dual model $\beta^\infty = 0.551\,43(7)$.
Direct comparison to Ref.~\cite{Baig2004} shows that while inverse transition
temperatures are reproduced, the extremal values of observables are not. The
following observations may help to clarify the deviations. The authors of
Ref.~\cite{Baig2004} simulated the system by applying periodic boundary
conditions and fixing one plane parallel to the $xy$-, $yz$- and $zx$-planes
each. If their simulation box consisted of a total number of $\hat{L}^3$ spins,
they simulated $(\hat{L} - 1)^3$ \emph{free} spins. Thus our data with lattices
of linear length $L$ has to be compared to their data {with} $\hat{L} = L + 1$.
Also, their specific heat and magnetic susceptibility $\hat{\chi} = \beta
\hat{L}^{-d} \left( \langle M^2 \rangle - \langle M \rangle^2 \right)$ have to
be multiplied by a factor of $(L+1)^3/L^3$ to be comparable with our
normalization, since these quantities are proportional to the inverse system
volume. Here, $M=\sum_i\sigma_i$ is the total magnetization, which for fixed
boundary conditions is a well defined order parameter.
Binder's energy parameter has no explicit volume-dependence by design, but it
is sensitive to offsets in the energy, which cancel in the specific heat. Our
values of the Binder parameter minima differ significantly
from~\cite{Baig2004}. However, if we shift our measured energies $E$ to get
$\hat{E} = E - 1.5\hat{L}^2 = E - 1.5(L+1)^2$ and calculate Binder's
parameter~\eref{eq:binder} with $\hat{E}$ instead of $E$, our measurements
reproduce {those} of~\cite{Baig2004} very well. The energy $E$ of the system can
be written in terms of the number of plaquettes with an even or odd number of
parallel aligned spins, $n_+$ or
$n_-$, \begin{eqnarray} E(\beta) = - \frac{1}{2}\left( n_+(\beta) - n_-(\beta)
\right). \end{eqnarray} Since we measure the same cumulant values for
shifted energies $\hat{E}$, in Ref.~\cite{Baig2004} an
additional number of $\hat{n}_+ = n_+ + 3\hat{L}^2$ plaquettes contribute to
the system's energy because energetic contributions from the
fixed planes, where all spins are aligned, were included.
\begin{table}[htpb] \tablesize \centering
\caption{Values for comparison with Ref.~\cite{Baig2004}, with the hat
denoting the observables as calculated by Baig~et~al. Linear lattice lengths
are $\hat{L} = L + 1$, with $L$ being the number of free spins in each
direction. The specific heat $\hat{C}_V^{\rm max} = L^3/(L+1)^3 C_V^{\rm
max}$, Binder's parameter $\hat{B}_{\rm min} = 1.0 -
\langle\hat{E}^4\rangle/3\langle\hat{E}^2\rangle^2$ with $\hat{E} = E -
1.5\hat{L}^2$. The inverse temperatures $\beta$ are the same for their data
and ours. Magnetic susceptibilities $\hat{\chi}^{\rm max} = L^3/(L+1)^3 \chi^{\rm max}$ are
listed as well.} \vspace{3ex}
\begin{tabular}{ll*{6}{r}}
\toprule
\multicolumn{1}{c}{$L$} & \multicolumn{1}{c}{$\hat{L}$} &
\multicolumn{1}{c}{$\beta^{\hat{C}_V^{\rm max}}$} &
\multicolumn{1}{c}{$\hat{C}_V^{\rm max}$} &
\multicolumn{1}{c}{$\beta^{\hat{B}^{\rm min}}$} &
\multicolumn{1}{c}{$\hat{B}^{\rm min}$} & \multicolumn{1}{c}{$\beta^{\hat{\chi}^{\rm max}}$} &
\multicolumn{1}{c}{$\hat{\chi}^{\rm max}$}\\ \midrule
9 & 10 & 0.45794(13) & 5.56(6) & 0.45527(14) &
0.63964(29) & 0.45690(14) & 6.86(9) \\ 11 & 12 &
0.47465(14) & 12.18(11) & 0.47338(14) &
0.63555(30) & 0.47438(14) & 16.37(15) \\ 13 & 14 &
0.48629(9) & 25.27(24) & 0.48559(9) & 0.6282(5)
& 0.48621(9) & 36.5(4) \\ 14 & 15 & 0.49086(8) &
35.5(4) & 0.49032(9) & 0.6234(6) & 0.49083(9)
& 52.9(7) \\ 17 & 18 & 0.50126(7) & 78.0(6) &
0.50096(7) & 0.6133(6) & 0.50125(7) & 126.2(1.2)
\\ 19 & 20 & 0.50648(8) & 121.7(1.0) & 0.50626(8)
& 0.6062(7) & 0.50648(8) & 206.9(1.9) \\ \bottomrule
\end{tabular} \label{tab:fixed_compare} \end{table}
For direct comparison, the resulting quantities are listed in
Table~\ref{tab:fixed_compare} after applying all corrections, showing that our
data is then in very good agreement with Ref.~\cite{Baig2004}. For completeness
we include here also our data for the magnetic susceptibility $\hat{\chi}$.
The deviation from the fitting results in Ref.~\cite{Baig2004} simply stems
from the fact that our simulations are performed with the multicanonical
method that allows a finite-size scaling analysis with more and significantly
larger lattice sizes.
{
The interface tension as a function of the linear lattice size~\eref{eq:interface-tension} is extracted and its infinite-volume limit yields a value of $\sigma = 0.0281(7)$ where we allowed corrections of order ${\cal O}(L^{-2})$ in the fits. Note that this value is about four times smaller than that for the same model with periodic boundary conditions.
}
\section{Summary}
We simulated the plaquette gonihedric Ising model and its dual to shed some
light on discrepancies in the recent literature on the reported value(s) of the
first-order phase transition temperature. High-precision results from
multicanonical simulations forced us to review the traditional scaling ansatz
for first-order finite-size corrections by taking the exponential low-temperature
phase degeneracy of the model into account. The leading correction
in such circumstances is then no longer proportional to the inverse volume of
the system, ${\cal O}\left( L^{-3} \right)$, but is rather ${\cal O}\left(
L^{-2} \right)$. With this finite-size scaling ansatz, our simulations with
periodic boundary conditions produced consistent results for both the original
formulation of the model as well as its dual representation. Since our results
also differed {from} earlier simulations {using} fixed boundary conditions, where
the leading corrections are now ${\cal O}\left( L^{-1} \right)$, we carried out
multicanonical simulations of the gonihedric Ising model with these boundary
conditions too. The resulting inverse transition temperature was fully
consistent with the value found using periodic boundary conditions when larger
lattices were included, and hopefully settle once and for all enduring
inconsistencies.
Interestingly, we do find different values for the latent heat for different
boundary conditions, $\Delta\hat{e} = 0.694(4)$ in the case of fixed
boundary conditions compared to $\Delta\hat{e} = 0.850968(18)$ for periodic
boundaries. That the latent heat may be dependent on the boundary conditions has
been observed earlier for the $q$-states Potts model~\cite{fssfbc}.
The main resulting physical
quantities that characterize the first-order phase transition are summarized in
Table~\ref{tab:infinite_system_results} for the different models and boundary
conditions in our simulations.
We find an overall consistent value for the inverse transition temperature of
\begin{equation}
\beta^\infty = 0.551\,334(8)
\end{equation}
and we measure the interface tension of the original model and its dual for the
first time.
{
We find values of $\sigma = 0.12037(18)$
and $\sigma = 0.1214(13)$ for the original and dual model with periodic boundary conditions, respectively. The interface tension of the original model with fixed boundary conditions is found to be much smaller, $\sigma = 0.0281(7)$.}
\begin{table}[tpb]
\tablesize
\centering
\caption{
Overview of resulting quantities of the infinite systems. } \vspace{3ex}
\begin{tabular}{*{7}{l}}
\toprule \multicolumn{1}{c}{model} &
\multicolumn{1}{c}{bc} & \multicolumn{1}{c}{$\beta^\infty$} &
\multicolumn{1}{c}{$\hat{e}_{\rm o}$} & \multicolumn{1}{c}{$\hat{e}_{\rm d}$} & \multicolumn{1}{c}{$\Delta\hat{e}$} &
\multicolumn{1}{c}{$\hat\sigma$}\\
\midrule
original~\eref{eq:ham:gonikappa0} & periodic & 0.551332(8) & $-1.468364(4)$ & $-0.617396(11)$ & 0.850968(18) & {0.12037(18)} \\
dual~\eref{eq:ham:dual} & periodic & 0.55143(7) & $-1.37644(21)$ & $-0.88227(19)$ & 0.49402(26) & 0.1214(13)\\
& & \multicolumn{2}{l}{[$\beta^\infty_{\rm dual} = 1.31328(12) $]}\\
original~\eref{eq:ham:gonikappa0} & fixed & 0.55138(5) & $-1.4782(27) $ &$ -0.790(4) $ & 0.694(4) & 0.0281(7) \\
\bottomrule
\end{tabular} \label{tab:infinite_system_results} \end{table}
{
Any model with an exponentially degenerate low-temperature phase will display the modified scaling at a
first-order phase transition we have delineated for the {three-dimensional} gonihedric model and its dual here.
Apart from higher-dimensional variants of the gonihedric model or its dual, there are numerous other models in which the scenario could be realized.
Examples range from ANNNI models \cite{selke} to topological {``orbital''} models in the context of quantum computing \cite{nussinov} which all share an extensive ground-state degeneracy. Among the orbital models for transition metal compounds, a particularly promising candidate is the three-dimensional classical compass or $t_{2g}$ orbital model where a highly degenerate ground state is well known and the signature of a first-order transition into the disordered phase has recently been found numerically \cite{compass4}.}
{
Other systems, such as the {three-dimensional} Ising antiferromagnet on {an} FCC lattice, have an exponentially
degenerate number of ground states but a small number (6 in the case of the {FCC} Ising antiferromagnet) of true low-temperature phases. Nonetheless, they do possess
an exponentially degenerate number of low-energy excitations so, depending on the nature of the growth of energy barriers with system size, an {\it effective} modified scaling could still be seen at a first-order transition for the lattice sizes accessible in typical simulations. The crossover to the true asymptotic (standard) scaling would then only appear for very large lattices.
Indeed, previous simulations appear to have found non-standard scaling for the first-order transition in
the {three-dimensional} Ising antiferromagnet on {an}
FCC lattice \cite{beath_ryan}.}
\section*{Acknowledgment}
This work was supported by the Deutsche Forschungsgemeinschaft (DFG)
through the Collaborative Research Centre SFB/TRR 102 (project B04) and by
the Deutsch-Franz\"osische Hochschule (DFH-UFA) under Grant No.\ CDFA-02-07.
\section*{References}
\input{bibliography_iop.tex}
\end{document}
|
\section{Introduction}
Rogue waves (RWs) appear in variety of physical systems, such as in oceans \cite{Kharif1,Kharif2,Osborne}, in atmosphere \cite{JPP75}, in optics \cite{Natural,SR2463}, in plasma \cite{EL96,Sharma1}. A fundamental property of such extreme waves is that they appear from nowhere, reach heights up to twice the height of surrounding waves, and disappear without a trace \cite{PLA373}. Significant theoretical progress has been achieved in recent years in order to describe the formation of RWs within the framework in weakly nonlinear evolution equations \cite{Osborne,ZakharovGelash}. In fact, RWs may be modeled in a general context by exact breather solutions of the nonlinear Schr\"odinger equation (NLS) \cite{Physrep52847}, which describe the propagation of stationary and pulsating wave packet dynamics in nonlinear dispersive media. An appropriate model to describe strong localizations in the medium of interest is the family of doubly-localized breather solutions, also referred to as Akhmediev-Peregrine (AP) solutions. The first-order solution is known as the Peregrine breather (PB) solution \cite{JAMSSB}; it is localized in space and in time, since it is the limiting case of time-periodic Kuznetsov-Ma solitons \cite{Kuznetsov,Ma} as well as space-periodic Akhmediev breathers \cite{AEK1985,tmp691089}, when the period of these periodic solutions tends to infinity. Furthermore, the AP amplifies the amplitudes of the carrier by a factor of three and higher. PB dynamics has been validated experimentally in optics \cite{NP6,ol36112}, in water waves \cite{PRL106,PRE86016311} and in plasma \cite{PRL10725}. More recently higher-order breathers have been observed as well \cite{PRX2011015,PRE86056601,Chabchoub4,FrisquetPRX,FrisquetPRA}.
There are indeed several interesting RW pattern solutions which the NLS may admit. Fundamental patterns may consist of a
simple central high peak, surrounded by several gradually decreasing solitonic peaks, an equal height triangular pattern, or a
circular pattern \cite{AkhmedievCascade}. Doubly-localized AP solutions are symmetric with respect to both the spatial and the temporal co-ordinate in theory. However, several non-symmetrical rogue waves under fundamental pattern have been observed in water tank, see for instance
Figs. 4-6 in \cite{PRE86016311}, Figs. 3, 4 and 6 in \cite{PRX2011015}, Figs. 3-9 in \cite{PRE86056601}. The observed non-symmetry of wave profiles in the wave flume is due to higher-dispersive and to the effect of the mean flow, not captured in the NLS, which can be explained by more accurate evolution equation, such as the generalized NLS in optics and the modified NLS (MNLS) in water waves \cite{Dysthe,DystheTrulsen,ChabchoubSupercontinuum}.
Independently from these higher-order effects, it is an interesting task to derive doubly-localized and non-symmetric solutions of the NLS. Physically, non-symmetric structures have different and fundamental distribution of energy and would be of interest for several physical applications. In this work, we derive exact solutions of the Manakov system up to third-order and as a special class corresponding family of non-symmetric doubly localized solutions of the NLS. In addition, we demonstrate the physical accuracy of the derived solutions to describe strong localizations in water waves by generating experimentally the derived second-order solution in a water wave tank. It is shown that the expected amplitude amplification of the background is reached. Moreover, the experimental results are in very good agreement with numerical MNLS simulations.
This paper is organized as follows. In section 2, we introduce the Lax pair of the Manakov system and solve the Lax pair equations, with spectral parameter $\lambda$, from a periodic \textit{seed} solution. In other words, we derive eigenfunctions of the Manakov
system, associated with the spectral parameter $\lambda$. In section 3, we give a determinant representation of the $n$th-order vector RWs of the Manakov system. In particular, exact expressions for the first-, second- and third-order RW solutions of the NLS equation, localized in both time and space are reported. The latter solutions have the property to be non-symmetric with respect to the spatial co-ordinate. Furthermore, the differences between the non-symmetric and corresponding symmetric doubly-localized RWs of the NLS equation are discussed in detail. In section 4, we shall compare experimental data, related to the evolution of the non-symmetric solution, derived in this work, in a water wave tank, with NLS maximal wave profile predictions as well as with numerical MNLS simulations. Finally, we summarize the main results in section 5.
\section{Vector eigenfunctions for the Manakov system}
Recently, RW solutions of the multi-component soliton equations have been derived. These solutions attracted the scientific interest \cite{EPJ,CPL28,PRE86036601,PRL109,PRE87,PRE88052914, N14, PRE87032910,jpsj81033002,pre86066603,pre87012913, pre88062925}. The Manakov system \cite{SPJETP}
is an important integrable two-component NLS system and can be applied to optical as well as hydrodynamic systems \cite{OnoratoOsorneSerio,GramstadTrulsen}.
In particular, if $q_1=\alpha q_2(\alpha \not=0 \in \mathbb{R})$ is a solution of the Manakov system
\begin{equation}\label{CNLS}
\begin{aligned}
{\rm i}q_{1t}+q_{1xx}+2(|q_1|^2+|q_2|^2)q_1=0,\\
{\rm i}q_{2t}+q_{2xx}+2(|q_2|^2+|q_1|^2)q_2=0, \\
\end{aligned}
\end{equation}
then
\begin{equation}\label{transformationnlsandcnls}
q=\sqrt{1+\frac{1}{\alpha^2}}q_1=\sqrt{1+\alpha^2}q_2
\end{equation}
is a solution of the NLS equation
\begin{equation}\label{NLS}
iq_t+q_{xx}+2q|q|^2=0.
\end{equation}
In the present paper, we shall construct the solutions of the Manakov system under the condition $q_1=\alpha q_2$ by the Darboux transformation (DT) from a periodic \textit{seed} solution, which imply several non-symmetrical extreme localizations of the NLS equation.
The DT is a powerful method to generate soliton \cite{AkhmedievBook,pla467100,actaamathsinica3143,Matveevbook,Gubook2,scichinaa121867,JPA40degasperis,JPA42degasperis},
breather \cite{actaamathsinica281713}, RW \cite{PRE80026601,PRE85026607,PRE87052914} solutions of the NLS equation.
The breather solution of the Manakov system \cite{PhyicaD141104} has been constructed by the DT from a periodic \textit{seed}
solution.
A crucial step of the DT method is to find the solutions, i.e. eigenfunctions of the linear partial differential equations,
defined by the Lax pair of the soliton equations. Therefore, we shall first solve the eigenfunctions of the Manakov system in this section in order to provide necessary preparatory work for the DT in next section.
The Manakov system \eqref{CNLS} is produced by the compatibility of the associated Lax pair equations:
\begin{equation}\label{laxpair}
\left\{
\begin{aligned}
\Psi_{x}&=U\Psi=({\rm i}J\lambda+Q)\Psi,\\
\Psi_{t}&=V\Psi=(2{\rm i}J\lambda^2+2Q\lambda+{\rm i}J(Q^2-Q_{x}))\Psi,
\end{aligned}\right.
\end{equation}
where
\begin{equation}
J=\left(\begin{array}{ccc}
-1 &0 &0\\
0 &1 &0\\
0 &0 &1\\
\end{array}\right),\nonumber\\
\quad Q=\left(\begin{array}{ccc}
0 &q_1 &q_2\\
-q_1^{*} &0 &0\\
-q_2^{*} &0 &0\\
\end{array}\right).
\end{equation}
Here, $\Psi$ is a vector eigenfunction associated with the spectral parameter $\lambda$.
\\
To obtain RW solutions of the Manakov system, we start with periodic \textit{seed} solutions,
\begin{equation}\label{seed}
\begin{aligned}
q_1=c_1e^{{\rm i}(a_1x+b_1t)},\quad
q_2=c_2e^{{\rm i}(a_2x+b_2t)},
\end{aligned}
\end{equation}
where $a_1, a_2, b_1, b_2,c_1,c_2 \in \mathbb{R}$. Then, the dispersion relations can be written as
\begin{equation}\label{dispersion}
\begin{aligned}
b_1=2(c_1^2+c_2^2-a_1^2),\quad
b_2=2(c_1^2+c_2^2-a_2^2).
\end{aligned}
\end{equation}
In this work, we consider only a special case for the condition: $a_1=a_2=a$. Then, $b_1=b_2=b$, and
$\alpha=\frac{c_1}{c_2}$. The solutions $q_1$ and $q_2$ are therefore reduced to a solution of the NLS equation
by the straight-forward transformation \eqref{transformationnlsandcnls}. Hence, by n-fold DT, we can generate new solutions
$q_1^{[n]}$ and $q_2^{[n]}$ of the Manakov system from these \textit{seed} solutions, which satisfy the condition $q^{[n]}_1= \alpha q^{[n]}_2$, and then they are automatically reduced to new solutions $q^{[n]}$ of the NLS.
Now, we solve the eigenfunctions of the Lax pair, associated with \textit{seed} solutions \eqref{seed}. Firstly, we construct the following transformation $\Psi=R\Phi$ to map the variable coefficient partial differential equations \eqref{laxpair} to constant coefficient partial differential equations
\begin{equation}\label{constantlaxpaireq}
\Phi_{x}=\Omega\Phi, \ \ \Phi_t=\Lambda\Phi,
\end{equation}
by the following matrix
\begin{equation}\label{mapmattix}
R=\left( \begin{array}{ccc}
e^{\frac{{\rm i}(ax+bt)}{2}} &0 &0\\
0 &e^{\frac{-{\rm i}(ax+bt)}{2}} &0\\
0 &0 &e^{\frac{-{\rm i}(ax+bt)}{2}}\\
\end{array}\right).
\end{equation}
Here,
\begin{equation}\label{xpart}
\Omega=\left(\begin{array}{ccc}
-{\rm i}\lambda-{\rm i}\frac{a}{2} &c_1 &c_2\\
-c_1 &{\rm i}\lambda+{\rm i}\frac{a}{2} &0\\
-c_2 &0 &{\rm i}\lambda+{\rm i}\frac{a}{2} \\
\end{array}\right),
\end{equation}
\begin{equation}\label{tpart}
\Lambda=\left(\begin{array}{ccc}
-2{\rm i}\lambda^2+{\rm i}(c_1^2+c_2^2)-{\rm i}\frac{b}{2} &(2\lambda-a)c_1 &(2\lambda-a)c_2\\
(a-2\lambda)c_1 &2{\rm i}\lambda^2-{\rm i}c_1^2+{\rm i}\frac{b}{2} &-{\rm i}c_1c_2\\
(a-2\lambda)c_2 &-{\rm i}c_1c_2 &2{\rm i}\lambda^2-{\rm i}c_2^2+{\rm i}\frac{b}{2}\\
\end{array}\right),
\end{equation}
$\Phi(x,t,\lambda)=\left(\begin{matrix}
f(x,t,\lambda),&g(x,t,\lambda),&h=(x,t,\lambda)
\end{matrix}\right)^T$, $\left(\cdot\right)^T$ denotes the transposition operator. Secondly, taking into account the dispersion relation \eqref{dispersion} in Eqs. \eqref{constantlaxpaireq}, \eqref{xpart} and \eqref{tpart},
then
\begin{equation}\label{efun}
\begin{aligned}
f=&\sum_{j=1}^{j=3}K_j{\rm e}^{u_jx+(2\lambda-a)u_jt},\\
g=&\sum_{j=1}^{j=3}K_j(\frac{-c_1}{u_j-{\rm i}\lambda-{\rm i}\frac{a}{2}}){\rm e}^{u_jx+(2\lambda-a)u_jt},\\
h=&\sum_{j=1}^{j=3}K_j(\frac{-c_2}{u_j-{\rm i}\lambda-{\rm i}\frac{a}{2}}){\rm e}^{u_jx+(2\lambda-a)u_jt},
\end{aligned}
\end{equation}
are solved by the method of characteristic equation. Here, $K_j(j=1,2,3)$ are three constants,
$u_1$,$u_2$ and $u_3$ are three roots of the following equation
\begin{equation}\label{eigenfun}
\begin{aligned}
&{u}^{3}-\left(\frac{1}{2}\,{\rm i}a+{\rm i}\lambda \right) {u}^{2}+ \left( \frac{1}{4}\,{a}^{2}
+a\lambda+{c_{{2}}}^{2}+{\lambda}^{2}+{c_{{1}}}^{2} \right) u +\Gamma=0,
\end{aligned}
\end{equation}
with
$$\Gamma={\rm i}{
\lambda}^{3}+\frac{3}{2}\,{\rm i}a{\lambda}^{2}
+{\rm i}{c_{{1}}}^{2}\lambda+{\rm i}{c_{{2}}}
^{2}\lambda+\frac{3}{4}\,{\rm i}\lambda\,{a}^{2}+\frac{1}{2}\,{\rm i}{c_{{2}}}^{2}a+\frac{1}{8}\,{\rm i}{a}^{3}+\frac{1}{2}\,{\rm i}{c_{{1}}}^{2}a.$$
Finally, solving Eq. \eqref{eigenfun}, we have
\begin{equation}\label{algebraroot}
\begin{aligned}
&u_1=\frac{\sqrt{-4\lambda^2-4\lambda a-a^2-4c_1^2-4c_2^2}}{2},\\
&u_2=\frac{-\sqrt{-4\lambda^2-4\lambda a-a^2-4c_1^2-4c_2^2}}{2},\\
&u_3={\rm i}(\frac{a}{2}+\lambda).
\end{aligned}
\end{equation}
Thus, eigenfunctions of the Manakov system associated with \textit{seed} Eq.\eqref{seed} are obtained by taking Eq.\eqref{algebraroot}, Eq.\eqref{efun} and Eq.\eqref{mapmattix}
back into $\Psi=R\Phi$.
\\
\section{Rogue wave solutions of the Manakov system and the NLS}
In this section, we discuss the solutions of the Manakov system and the NLS equation.
Based on early work, where RW solutions have been derived \cite{jpsj81033002,pre86066603,pre87012913, pre88062925,PRE87052914,PhyicaD141104,JPA44305203,JMP53063507,cnsns191706,ps2014}, we set $\lambda=\lambda_0+\epsilon^2=-\frac{a}{2}+{\rm i}\sqrt{c_1^2+c_2^2}+\epsilon^2$. Note that $u_1=u_2=0$ and $u_3=-\sqrt{c_1^2+c_2^2}$ if $ \lambda=\lambda_0$. Furthermore, we set $K_1=1$, $K_2=-1$ and $K_3=0$, it is trivial to verify that $\Psi(\lambda_0)=0$. Thus, the Taylor expansion with respect to $\epsilon^2$ of the new solutions $q_1^{[n]}$ and $q_2^{[n]}$ generates the vector RWs of the Manakov system. For convenience, we introduce
$\Phi_j= \Phi(x,t,\lambda)|_{\lambda=\lambda_j}=\left(\begin{matrix}
f_j,&g_j,&h_j
\end{matrix}\right)^T$, associated with $n$ distinct eigenvalues $\lambda_j, j=1, 2, 3, \cdots, n$.
\subsection{First-order rogue wave solutions of Manakov system}
We refer to \cite{PhyicaD141104} for the use of the DT to solve the Manakov system and we shall not repeat this procedure in this work.
As mentioned earlier, we set $K_1=1$, $K_2=-1$, $K_3=0$ and $\lambda=\lambda_0+\epsilon^2$ in $q_1^{[1]}$ and $q_2^{[1]}$, then
first-order Taylor expansion of them gives the first-order RW of the Manakov system.
\begin{theorem}\label{thm_1rw}
Let $\Psi_{1}=R\Phi_1$ be a solution of the Lax pair equations associated with $\lambda_{1}$ and \textit{seed} solution, $\lambda_1=\lambda_0+\epsilon^2$, then $(q_1^{[1]},q_2^{[1]})$ given by the following
determinant representation are the first-order vector RW of the Manakov system:
\begin{equation}\label{f_1_rw}
q_1^{[1]}=\left(c_1+2{\rm i}\frac{|\Omega_{1}'|}{|\Omega_{2}'|}\right){\rm e}^{{\rm i}(ax+bt)},\quad q_2^{[1]}=\left(c_2-2{\rm i}\frac{|\Omega_{3}'|}{|\Omega_{2}'|}\right){\rm e}^{{\rm i}(ax+bt)},
\end{equation}
where
\begin{equation*}
\begin{aligned}
\Omega_{1}=\left(\begin{matrix}
f_{1} &h_{1} &\lambda_1 f_{1} \\
-g_{1}^* &0 &-\lambda_1^*g_{1}^*\\
-h_{1}^* &f_{1}^* &-\lambda_1^*h_{1}^*\\
\end{matrix}\right),\quad
\Omega_{2}=\left(\begin{matrix}
f_{1} &g_{1} &h_{1} \\
-g_{1}^* &f_{1}^* &0\\
-h_{1}^* &0 &f_{1}^*\\
\end{matrix}\right),\quad
\Omega_{3}=\left(\begin{matrix}
f_{1} &g_{1} &\lambda_1 f_{1} \\
-g_{1}^* &f_{1}^* &-\lambda_1^*g_{1}^*\\
-h_{1}^* &0 &-\lambda_1^*h_{1}^*\\
\end{matrix}\right),
\end{aligned}
\end{equation*}
and
\begin{equation}
\begin{aligned}
\Omega_{1}'=&\left(\left.\frac{\partial^{2}}{\partial{\epsilon^{2}}}\right|_{\epsilon=0}\left(\Omega_{1}\right)
_{ij}\left(\lambda_0+\epsilon^2\right)\right)_{3\times3},\\
\Omega_{2}'=&\left(\left.\frac{\partial^{2}}{\partial{\epsilon^{2}}}\right|_{\epsilon=0}\left(\Omega_{2}\right)
_{ij}\left(\lambda_0+\epsilon^2\right)\right)_{3\times3},\\
\Omega_{3}'=&\left(\left.\frac{\partial^{2}}{\partial{\epsilon^{2}}}\right|_{\epsilon=0}\left(\Omega_{3}\right)
_{ij}\left(\lambda_0+\epsilon^2\right)\right)_{3\times3}.
\end{aligned}
\end{equation}
\end{theorem}
Substituting \eqref{efun} into \eqref{f_1_rw} of \textcolor{red}{T}heorem \ref{thm_1rw}, the explicit expressions of the the first-order vector RWs
of the Manakov system are given as
\begin{equation}\label{1_rw}
\begin{aligned}
q_1^{[1]}=e^{{\rm i}(ax+bt)}(c_1+\frac{F_1}{H_1}),\quad
q_2^{[1]}=e^{{\rm i}(ax+bt)}(c_2+\frac{G_1}{H_1}),
\end{aligned}
\end{equation}
where
\begin{equation}\nonumber
\begin{aligned}
&F_1=4c_{{1}}\,\sqrt {{c_{{1}}}^{2}+{c_{{2}}}^{2}}L_1L_2, \ \ \ \ \ \ \ \ \ G_1=4c_{{2}}\,\sqrt {{c_{{1}}}^{2}+{c_{{2}}}^{2}}L_1L_2,\\
&L_1=x-2\,ta+2\,{\rm i}t\sqrt {{c_{{1}}}^{2}+{c_{{2}}}^{2}},\ \ L_2=2\,{\rm i}t \left( {c_{{1}}}^{2}+{c_{{2}}}^{2} \right) + \left( 2\,t
a -\,x\right) \sqrt {{c_{{1}}}^{2}+{c_{{2}}}^{2}}+1,\\
&H_1=8\,{t}^{2} \left( {c_{{1}}}^{2}+{c_{{2}}}^{2} \right) ^{2}+ \left(2\,{x}^{2}-
8\,xta+8\,{t}^{2}{a}^{2}\right) \left( {c_{{
1}}}^{2}+{c_{{2}}}^{2} \right) + \left(4\,ta-2\,x \right) \sqrt {{
c_{{1}}}^{2}+{c_{{2}}}^{2}}+1.\\
\end{aligned}
\end{equation}
It is obvious that ${q_1^{[1]}}/{q_2^{[1]}}={c_1}/{c_2}$, so that $q^{[1]}=\sqrt{1+\frac{c^2_2}{c^2_1}}q_1^{[1]}=\sqrt{1+\frac{c^2_1}{c^2_2}}q_2^{[1]}$
is the first-order RW of the
NLS, which is plotted in Fig. \ref{fig1}.
\begin{figure}[h]
\subfigure[]{\includegraphics[width=6.5cm]{q1.pdf}}
\subfigure[]{\includegraphics[width=6.5cm]{q1_contour.pdf}}
\caption{(a) Evolution of the first-order doubly-localized RW non-symmetric $q^{[1]}$. (b) Contour plot of the solution $q^{[1]}$.}
\label{fig1}
\end{figure}
By choosing $c_1=c_2=\frac{\sqrt{2}}{2}$, the amplitude of the plane wave, i.e. carrier of $q^{[1]}$ is of one. The maximum amplitude of the RW $q^{[1]}$ is located at $(\frac{1}{2},0)$. Compared with the results of \cite{N14} (Eqs. 10(a) and 10(b), Figs. 1, 2 and 3), it can be concluded that $q^{[1]}$ possesses the same dynamics as the doubly-localized Peregrine solution \cite{JAMSSB}, except for the location of the rogue peak.
Note that $q^{[1]}$ is non-symmetric with respect to ($x=0$)-axis, as displayed in Fig. \ref{fig1}. However, after applying a straight-forward transformation $x=x+\frac{1}{2}$, it becomes symmetric.
Moreover, let $\,c_1=\frac{\sqrt{2}}{2}$ and $c_2=\frac{\sqrt{2}}{2}$ in this case, then $\alpha=1$, and
\begin{equation}
q_s^{[1]}=\sqrt{2}q_1^{[1]}=\sqrt{2}q_2^{[1]}={\frac { \left( 3-16\,{t}^{2}+16\,{\rm i}t-4
\,{x}^{2} \right) }{1+16\,{t}^{2}+4\,{x}^{2}}}{{\rm e}^{2\,{\rm i}t}}
\end{equation}
is the Peregrine soliton of the NLS equation \cite{JAMSSB}, as expected, see Fig. \ref{fig2}.
\begin{figure}[h]
\subfigure[]{\includegraphics[width=6.5cm]{qp.pdf}}
\subfigure[]{\includegraphics[width=6.5cm]{qp_contour.pdf}}
\caption{(a) Temporal and spatial evolution of the Peregrine breather. (b) Contour plot of the Peregrine breather.}
\label{fig2}
\end{figure}
This fact means that the asymmetry feature in the first-order RW is non-essential, since it can be explained by a simple linear shift of the spatial variable. We need to study higher-order RWs of the NLS to explore the non-symmetric property of such solutions.
\subsection{Higher-order rogue wave solutions of the Manakov system}
Similar to the case of NLS \cite{scichinaa121867, PRE87052914} and according to the one-fold DT, we obtain the determinant expressions of $n$th-order solution $(q_1^{[n]},\,q_2^{[n]})$ through iteration. With the method of Taylor expansion with respect to $\epsilon^2$ in $(q_1^{[n]},\,q_2^{[n]})$ through $\lambda_j=\lambda_0+\epsilon^2$,
the determinant expressions of the $n$th-order RWs are obtained as well.
\begin{theorem}\label{thm_nrw}
Let $\Psi_{i}=R\Phi_i$ $(i=1,2,\cdots,n)$ be $n$ distinct eigenfunctions associated with $\lambda_{i}(i=1,2,\cdots,n)$,
then, $(q_1^{[n]},\,q_2^{[n]})$ given by the following determinant representations are the $n$th-order RWs of the Manakov system.
\begin{equation}
q_1^{[n]}=(c_1+2{\rm i}\frac{|\Omega_{1}'|}{|\Omega_{2}'|}){\rm e}^{{\rm i}(ax+bt)},\quad q_2^{[n]}=(c_2-2{\rm i}\frac{|\Omega_{3}'|}{|\Omega_{2}'|}){\rm e}^{{\rm i}(ax+bt)},
\end{equation}
where
\begin{equation}
\nonumber
\Omega_{1}=\left(\begin{matrix}
f_{1} &g_{1} &h_{1} &\lambda_1f_{1} &\lambda_1g_{1} &\lambda_1h_{1} &\cdots
&\lambda_1^{n-1}f_{1} &\lambda_1^{n-1}h_{1} &\lambda_1^{n}f_{1}\\
-g_{1}^* &f_{1}^* &0 &-\lambda_1^*g_{1}^* &\lambda_1^*f_{1}^* &0 &\cdots
&-(\lambda_1^*)^{n-1}g_{1}^* &0 &-(\lambda_1^*)^{n}g_{1}^*\\
-h_{1}^* &0 &f_{1}^* &-\lambda_1^*h_{1}^* &0 &\lambda_1^*f_{1}^* &\cdots
&-(\lambda_1^*)^{n-1}h_{1}^* &(\lambda_1^*)^{n-1}f_{1}^* &-(\lambda_1^*)^{n}h_{1}^*\\
\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots\\
f_{n} &g_{n} &h_{n} &\lambda_nf_{n} &\lambda_ng_{n} &\lambda_nh_{n} &\cdots
&(\lambda_n^*)^{n-1}f_{n} &(\lambda_n^*)^{n-1}h_{n} &(\lambda_n^*)^{n}f_{n}\\
-g_{n}^* &f_{n}^* &0 &-\lambda_n^*g_{n}^* &\lambda_n^*f_{n}^* &0 &\cdots
&-(\lambda_n^*)^{n-1}g_{n}^* &0 &-(\lambda_n^*)^{n}g_{n}^*\\
-h_{n}^* &0 &f_{n}^* &-\lambda_n^*h_{n}^* &0 &\lambda_n^*f_{n}^* &\cdots
&-(\lambda_n^*)^{n-1}h_{n}^* &(\lambda_n^*)^{n-1}f_{n}^* &-(\lambda_n^*)^{n}h_{n}^*\\
\end{matrix}\right),\\
\end{equation}
\begin{equation}\nonumber
\begin{aligned}
\Omega_{2}=\left(\begin{matrix}
f_{1} &g_{1} &h_{1} &\lambda_1f_{1} &\lambda_1g_{1} &\lambda_1h_{1} &\cdots
&\lambda_1^{n-1}f_{1} &\lambda_1^{n-1}g_{1} &\lambda_1^{n-1}h_{1}\\
-g_{1}^* &f_{1}^* &0 &-\lambda_1^*g_{1}^* &\lambda_1^*f_{1}^* &0 &\cdots
&-(\lambda_1^*)^{n-1}g_{1}^* &(\lambda_1^*)^{n-1}f_{1}^* &0\\
-h_{1}^* &0 &f_{1}^* &-\lambda_1^*h_{1}^* &0 &\lambda_1^*f_{1}^* &\cdots
&-(\lambda_1^*)^{n-1}h_{1}^* &0 &(\lambda_1^*)^{n-1}f_{1}^*\\
\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots\\
f_{n} &g_{n} &h_{n} &\lambda_nf_{n} &\lambda_ng_{n} &\lambda_nh_{n} &\cdots
&\lambda_n^{n-1}f_{n} &\lambda_n^{n-1}g_{n} &\lambda_n^{n-1}h_{n}\\
-g_{n}^* &f_{n}^* &0 &-\lambda_n^*g_{n}^* &\lambda_n^*f_{n}^* &0 &\cdots
&-(\lambda_n^*)^{n-1}g_{n}^* &(\lambda_n^*)^{n-1}f_{n}^* &0\\
-h_{n}^* &0 &f_{n}^* &-\lambda_n^*h_{n}^* &0 &\lambda_n^*f_{n}^* &\cdots
&-(\lambda_n^*)^{n-1}h_{n}^* &0 &(\lambda_n^*)^{n-1}f_{n}^*\\
\end{matrix}\right),\\
\Omega_{3}=\left(\begin{matrix}
f_{1} &g_{1} &h_{1} &\lambda_1f_{1} &\lambda_1g_{1} &\lambda_1h_{1} &\cdots
&\lambda_1^{n-1}f_{1} &\lambda_1^{n-1}g_{1} &\lambda_1^{n}f_{1}\\
-g_{1}^* &f_{1}^* &0 &-\lambda_1^*g_{1}^* &\lambda_1^*f_{1}^* &0 &\cdots
&-(\lambda_1^*)^{n-1}g_{1}^* &(\lambda_1^*)^{n-1}f_{1}^* &-(\lambda_1^*)^{n}g_{1}^*\\
-h_{1}^* &0 &f_{1}^* &-\lambda_1^*h_{1}^* &0 &\lambda_1^*f_{1}^* &\cdots
&-(\lambda_1^*)^{n-1}h_{1}^* &0 &-(\lambda_1^*)^{n}h_{1}^*\\
\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots\\
f_{n} &g_{n} &h_{n} &\lambda_nf_{n} &\lambda_ng_{n} &\lambda_nh_{n} &\cdots
&(\lambda_n^*)^{n-1}f_{n} &(\lambda_n^*)^{n-1}g_{n} &(\lambda_n^*)^{n}f_{n}\\
-g_{n}^* &f_{n}^* &0 &-\lambda_n^*g_{n}^* &\lambda_n^*f_{n}^* &0 &\cdots
&-(\lambda_n^*)^{n-1}g_{n}^* &(\lambda_n^*)^{n-1}f_{n}^* &-(\lambda_n^*)^{n}g_{n}^*\\
-h_{n}^* &0 &f_{n}^* &-\lambda_n^*h_{n}^* &0 &\lambda_n^*f_{n}^* &\cdots
&-(\lambda_n^*)^{n-1}h_{n}^* &0 &-(\lambda_n^*)^{n}h_{n}^*\\
\end{matrix}\right),
\end{aligned}
\end{equation}
and
\begin{equation}
\begin{aligned}
\Omega_{1}'=&\left(\left.\frac{\partial^{n_i}}{\partial{\epsilon^{n_i}}}\right|_{\epsilon=0}\left(\Omega_{1}\right)
_{ij}\left(\lambda_0+\epsilon^2\right)\right)_{3n\times3n},\\
\Omega_{2}'=&\left(\left.\frac{\partial^{n_i}}{\partial{\epsilon^{n_i}}}\right|_{\epsilon=0}\left(\Omega_{2}\right)
_{ij}\left(\lambda_0+\epsilon^2\right)\right)_{3n\times3n},\\
\Omega_{3}'=&\left(\left.\frac{\partial^{n_i}}{\partial{\epsilon^{n_i}}}\right|_{\epsilon=0}\left(\Omega_{3}\right)
_{ij}\left(\lambda_0+\epsilon^2\right)\right)_{3n\times3n},
\end{aligned}
\end{equation}
$n_i=2[\frac{i-1}{3}]+2,\,[i]$ denotes the floor function of $i$.
\end{theorem}
For convenience, let $a=0,\,c_1=\frac{\sqrt{2}}{2}$ and $c_2=\frac{\sqrt{2}}{2}$ in the following. Then, $\alpha=1$, and $q^{[n]}=\sqrt{2}q_1^{[n]}=\sqrt{2}q_2^{[n]}$ is the $n$th-order RWs of the NLS equation.
Let $n=2$ in \textcolor{red}{T}heorem \ref{thm_nrw}, the second-order vector RWs of the Manakov system are given by
\begin{equation}\label{2_rw}
\begin{aligned}
q_1^{[2]}=q_2^{[2]}=(\frac{\sqrt{2}}{2}+\frac{F_2}{H_2}){\rm e}^{2{\rm i}t},
\end{aligned}
\end{equation}
with
\begin{equation}\nonumber
\begin{aligned}
F_2=&12\,{\rm i}\sqrt {2} \left( 12\,t-32\,{t}^{3}-12\,xt+16\,{x}^{3}t+64\,{t}^{
3}x-8\,{x}^{4}t-64\,{x}^{2}{t}^{3}-128\,{t}^{5}\right.\\
&\left.+48\,{\rm i}{x}^{2}{t}^{2}-4
\,{\rm i}{x}^{3}+6\,{\rm i}{x}^{2}+160\,{\rm i}{t}^{4}-48\,{\rm i}x{t}^{2}+48\,{\rm i}{t}^{2}+2\,{\rm i}{x
}^{4}-3\,{\rm i}x \right),\\
H_2=&1024\,{t}^{6}+768\,{x}^{2}{t}^{4}-768\,{t}^{4}x+1920\,{t}^{4}
+192\,{x}^{4}{t}^{2}-384\,{x}^{3}{t}^{2}+288\,x{t}^{2}\\
&+288\,{t}^{2}+16
\,{x}^{6}-48\,{x}^{5}+72\,{x}^{4}-72\,{x}^{3}+72\,{x}^{2}-36\,x+9.
\end{aligned}
\end{equation}
It is a non-symmetrical doubly-localized RW $q^{[2]}=\sqrt{2}q_1^{[2]}$ of the NLS, and is displayed in Fig. \ref{fig3}.
\begin{figure}[h]
\subfigure[]{\includegraphics[width=6.5cm]{q2.pdf}}
\subfigure[]{\includegraphics[width=6.5cm]{q2_contour.pdf}}
\caption{(a) Evolution of the second-order doubly-localized RW non-symmetric $q^{[2]}$. (b) Contour plot of the solution $q^{[2]}$.}
\label{fig3}
\end{figure}
It is obvious that the structure of this solution is highly non-symmetrical in the $(x,t)$ plane and the highest peak
is not located at $(0,0)$, unlike the profile of the symmetrical doubly-localized AP breathers' fundamental pattern \cite{PRX2011015}. In fact, as can been observed qualitatively in Fig. \ref{fig3} (a), there are four small soliton peaks surrounding one main and highest peak. However, the left two soliton peaks with respect to the ($x=0$)-axis are higher than the two on the right. Besides, the highest peak is not located on the center of the four small peaks as the symmetrical fundamental pattern\cite{pla3732137}. It is located on the left side of the co-ordinate origin center. Specifically, the co-ordinate of the main peak is $(0.372, 0)$ in the $(x,t)$ plane, and the height of the main peak is of 4.695. This value corresponds to the amplitude amplification of this solution, since the carrier amplitude is of one. This also means that the amplitude of this RW is not distributed symmetrically along $x$-axis.
In Fig. \ref{fig3} (a), it can be also seen that there is only one significant trough at the left side of the main peak. This is not the case for the symmetric AP solution, where two identical and significant troughs can be noticed at each side of the main peak along ($t=0$)-axis, see \cite{pla3732137}. In order to depict these troughs, we have displayed the contour line at asymptotic plane in Fig. \ref{fig3} (b). Generally, Fig. \ref{fig3} (a) and (b) show the noticeable and clear differences to the dynamics, described by the symmetrical fundamental pattern of the second-order NLS RW \cite{pla3732137}. The centers of the depicted three circles are given by (0.6,0.41), (0.84, 0), (0.6,-0.41) in the ($x,t$)-plane.
Next, let $n=3$ in {}{T}heorem \ref{thm_nrw}, then the third-order RW of Manakov system is given as
\begin{equation}\label{3_rw}
q_1^{[3]}=q_2^{[3]}=e^{2{\rm i}t}(\frac{\sqrt{2}}{2}+\frac{F_3}{H_3}).
\end{equation}
Here, $F_3$ and $G_3$ {}{are} two polynomials with degree 12 of $x$ and $t$, which are given in Appendix.
Because $\frac{F_3}{H_3}=-\sqrt{2}$ when $x\rightarrow \infty $ and $t\rightarrow \infty$, the asymptotical
height of $|q_1^{[3]}|$ and $|q_2^{[3]}|$ is $\frac{\sqrt{2}}{2}$.
Once again $q^{[3]}=\sqrt{2}q_1^{[3]}=\sqrt{2}q_1^{[3]}$ gives the third-order non-symmetrical RW of the NLS,
which has asymptotical height $1$ when $x\rightarrow \infty $ and $t\rightarrow \infty$.
The profile and the density plot of $q^{[3]}$ are displayed in Fig. \ref{fig4}, which as expected are different from
the symmetrical pattern of the third-order AP \cite{PRE80026601}.
\begin{figure}[h]
\subfigure[]{\includegraphics[width=6.5cm]{q3.pdf}}
\subfigure[]{\includegraphics[width=6.5cm]{q3_contour.pdf}}
\caption{(a) Evolution of the third-order doubly-localized RW non-symmetric $q^{[3]}$. (b) Contour plot of the solution $q^{[3]}$.}
\label{fig4}
\end{figure}
Similar to the case of $q^{[2]}$, it is clearly not symmetric with respect to the spatial co-ordinate, however still the symmetry is conserved with respect to the ($t=0$)-axis. The
co-ordinate of main peak of $|q^{[3]}|$ is indeed not at $(0,\, 0)$. The maximum amplitude occurs at $(0.305,\,0)$ in the ($x,t$)-plane, is of $6.269$. This value is smaller than 7, which characterizes the third-order AP solution \cite{PRE80026601}.
Note also that the latter solution maps six zero amplitude points located on the $x$-axis.
{Since the DT of the Manakov system requires solving a cubic algebraic equation \eqref{eigenfun}, thus, it is possible to derive novel non-trivial solutions for the NLS, when the condition described in Eq. \eqref{transformationnlsandcnls} applies. Indeed, it would be also possible to derive the presented doubly-localized solutions directly from the NLS. This needs further work, which has been started.}
\subsection{Differences between symmetrical and non-symmetrical RWs}
In the section, we analyze the differences between the derived solution to the family of APs up to third-order. Since the non-symmetric solution of Manakov system can generate the solutions of the NLS equation through the simple scaling transformation $q^{[n]}=\sqrt{2} q_1^{[n]}$, when $c_1=c_2=\frac{\sqrt{2}}{2}$ in Eqs. \eqref{seed}, we just discuss the difference between symmetrical
and non-symmetrical solution of the NLS equation. The main characteristic features are summarized in Table 1.
\begin{table}[h]
\centering
\begin{tabular}{|c|c|c|c|c|c|}
\hline
$q_s^{[j]}\&q^{[j]}$ & Symmetric axis &No. zeros & Amplitude &Denom. degree\\\hline
$q^{[1]}$ & $x=\frac{1}{2},\,t=0$ & $2$ &$3$ &$2$ \\
$q_s^{[1]}$ & $x=0,\,t=0$ & $2$ &$3$ &$2$ \\\hline
$q^{[2]}$ & $x=0.372,\,t=0$ & $4$ &$4.695$ &$6$ \\
$q_s^{[2]}$ & $x=0,\,t=0$ & $4$ &$5$ &$6$ \\\hline
$q^{[3]}$ & $x=0.305,\,t=0$ & $6$ &$6.269$ & $12$ \\
$q_s^{[3]}$ & $x=0,\,t=0$ & $6$ &$7$ & $12$ \\
\hline
\end{tabular}\\
\caption{ The difference between symmetric $q_s^{[j]}$ and non-symmetric $q^{[j]}$ RWs of the NLS equation. Here, \textit{No. zeros} denotes the number of zero amplitudes in $q_s^{[j]}\&q^{[j]}$ along $x$-axis,
whereas \textit{Denom. degree} labels the degree of the polynomials of $x$ and $t$
in the denominators of $q_s^{[j]}\&q^{[j]}$. Furthermore, $q_s^{[j]}$ and $q^{[j]}$ represent the symmetric and non-symmetric solution, respectively. The symmetrical fundamental patterns $q_s^{[j]}$ have a fixed asymptotical plane
with a scaled height of $1$, as constructed in \cite{PRE80026601,PRE87052914}. }
\end{table}
As already mentioned, we can conclude that the solutions $q^{[j]}(j=2,3)$ (the non-symmetric solution) are just symmetrical with respect to
the $t$-axis and not symmetrical with respect to the $x$-axis. Furthermore, the amplitude of the non-symmetric solution
is slightly smaller than the corresponding symmetric AP case. Besides, zero amplitudes in $|q^{[j]}(x,0)(j=2,3)|$ are not distributed along $x$-axis as in the case of symmetrical RWs\cite{PRE80026601}. The number of the zero amplitudes of the RWs $q^{[n]}$ is increasing by increasing the order $n$. Therefore, it is reasonable to infer that the non-symmetric feature becomes more and more remarkable and noticeable, thus, such RW solutions deserve further studies.
\section{Numerical validation and laboratory water wave experiments}
This section is dedicated to the evolution of the derived second-order non-symmetrical RW $q^{[2]}$ in dimensional physical units and to laboratory experiments, which have been conducted in order to observe the latter NLS solution. For that purpose the solution has to be written first in dimensional units, in order to satisfy the deep-water NLS \cite{Zakharov}:
\begin{equation}\label{nls}
i\left(\frac{\displaystyle\partial \psi}{\displaystyle \partial
\tau}+\frac{\displaystyle\omega_0}{\displaystyle2k_0}\frac{\displaystyle\partial \psi}{\displaystyle \partial
\xi}\right)-\frac{\displaystyle\omega_0}{\displaystyle
8k_0^2}\frac{\displaystyle\partial^2 \psi}{\displaystyle \partial
\xi^2}-\frac{\displaystyle\omega_0 k_0^2}{\displaystyle
2}\left|\psi\right|^2\psi=0,
\end{equation}
while the wave frequency $\omega_0$ and the wave number $k_0$ are connected through the dispersion relation $\omega_0=\sqrt{gk_0}$, where $g$ labels the gravitational acceleration. Then, the spatio-temporal evolution of the free surface elevation $\eta(\xi,\tau)$ is approximated to first-order in steepness:
\begin{equation}
\eta(\xi,\tau)=\operatorname{Re}\left(\psi(\xi,\tau)\cdot\exp\left[i\left(k_0\xi-\omega_0\tau\right)\right]\right).
\label{se}
\end{equation}
The amplitude of the carrier $a_0$ and its steepness $\varepsilon_0=a_0k_0$ have been chosen to be of 0.003 m and 0.05, respectively. Therefore, the wave frequency is $\omega_0=12.78\textnormal{ rad}\cdot\textnormal{s}^{-1}$. These values have been chosen in order to satisfy the deep-water gravity waves conditions and to avoid wave breaking, which would significantly violate the evolution of the wave field, as predicted and approximated by the NLS.
\subsection{Numerical MNLS simulations}
The first step consists in determining the appropriate initial spatial co-ordinate for the boundary conditions Eq. (\ref{se}), which are needed in order to drive the physical wave maker. Therefore, preliminary numerical simulations based on the modified NLS (MNLS) \cite{Dysthe}, known to describe more accurately the evolution of nonlinear wave packets compared with the NLS since it takes into account higher-order dispersion as well as the mean flow of the wave field, have been performed for the dimensional non-symmetrical second-order $\psi^{[2]}$. Note that as described earlier and differently than the second-order AP solution the maximal amplification of the corresponding field is not at $\xi=0$. Results of the simulations, showing the last stage of envelope propagation over 50 m, before reaching its maximal amplification and for the initial spatial co-ordinate $\xi=-45$ m, are depicted in Fig. \ref{fig5}.
\begin{figure}[h]
\centering
\includegraphics[width=10cm]{MNLS.pdf}
\caption{MNLS simulations of the dimensional second-order non-symmetric solution $\psi^{[2]}$ for the carrier ampliude $a_0=0.003$ m and $\varepsilon=0.05$, while evolving in space with the group velocity, which is $c_g:=\dfrac{\omega_0}{2k_0}=0.38\textnormal{ m}\cdot\textnormal{s}^{-1}$.}
\label{fig5}
\end{figure}
Results in Fig. \ref{fig5} show the evolution of the $\psi^{[2]}$-wave field, which initial perturbation consists of two small envelope modulations, as described by NLS boundary conditions. During this propagation with the group velocity, a strong maximal localization is formed after 50 m and is the result of the nonlinear interaction of the two initially small envelopes.
\subsection{Wave flume experiments}
As next, we will trigger the same evolution dynamics in a laboratory wave flume. Experiments have been conducted in the same facility, as described in detail in \cite{PRX2011015}. The carrier parameters as well as initial conditions have been chosen to be the same as for the numerical simulations of the evolution dynamics. The wave gauge is placed 10 m from the wave maker, which consists of a single flap, controlled by a hydraulic cylinder. The from the gauge measured wave profiles are then reemitted to the flap in order to overcome the length restriction of the wave flume facility and to increase its length, as described in \cite{Chabchoub4}. Therefore, four iteration loops have to be conducted to satisfy a propagation distance of 50 m. The experimental results for the same initial condition as for the numerical simulations are shown in Fig. \ref{fig6}.
\begin{figure}[h]
\centering
\includegraphics[width=10cm]{evolution.pdf}
\caption{Evolution of the second-order non-symmetric solution $\psi^{[2]}$ in a water wave tank for $a_0=0.003$ m and $\varepsilon=0.05$. The five measurements are aligned by the value of the group velocity.}
\label{fig6}
\end{figure}
As for the doubly-localized second-order AP solution \cite{PRX2011015} and as already provided by the MNLS simulations, the characteristic amplification of the wave field is due to the overlapping two envelope modulations. However, the difference with respect to the symmetric AP solution is that this nonlinear envelope interaction is non-symmetric already at the initial stage of the small amplitude modulations, as can be noticed in Fig. \ref{fig6}. This asymmetrical interaction persists for the whole propagation distance along the flume. Indeed, as predicted by the MNLS simulations, the maximal amplitude amplification is reached after 50 m of wave propagation. Next,
we compare the maximal wave profile with the NLS prediction. Fig. \ref{fig7} shows the results.
\begin{figure}[h]
\centering
\includegraphics[width=10cm]{comparison.pdf}
\caption{Comparison of the maximal measured wave profiles with respect to the theoretical NLS prediction to second-order in steepness.}
\label{fig7}
\end{figure}
Clearly, the agreement is remarkable. The observed and theoretical wave profiles, shown in Fig. \ref{fig7} are \textit{almost} identical with a noticeable asymmetry compared to the maximal wave, related to this particular solution. Furthermore, the maximal characteristic amplitude amplification of 4.6 is reached. These experimental observations prove that higher-order nonlinear envelopes, generating non-symmetric RWs, can be described by weakly nonlinear evolution equations and can be observed in nonlinear dispersive media, governed by the NLS. The derived non-symmetric solutions naturally have different physical and hydrodynamic properties in terms of surface wave profiles as well as flow field variations, compared with the symmetric AP solutions. This should be analyzed and investigated in more detail numerically and experimentally in order to quantify these physical differences. Furthermore, the results may motivate similar analytical, numerical and experimental studies in other media, governed by the NLS, such as in optics or in plasma.
\section{Conclusion}
To summarize, we presented the determinant expressions of the $n$th-order RWs of the Manakov system in {}{T}heorem 1 and {T}heorem 2,
and gave the explicit expressions for the first-, second- and third-order RW solutions in a particular case such that $q_1^{[j]}=q_2^{[j]}(j=1,2,3)$. As special case, these three solutions determine non-symmetrical RWs $q^{[j]}(j=1,2,3)$ with respect to the spatial co-ordinate under the fundamental framework of the NLS. Fundamental characteristics of these doubly-localized NLS solutions as well as the differences to the corresponding symmetrical AP case, have been discussed in detail. Furthermore, we confirmed the physical validity of the second-order solution $q^{[2]}$, by performing numerical simulations as well as laboratory experiments. The theoretical, numerical and experimental results are in a very good agreement and confirm the accuracy of weakly nonlinear evolution equations to describe symmetric and non-symmetric strongly localized pattern dynamics. This work may motivate further numerical simulations of the fully nonlinear hydrodynamic evolution equations and further experiments in different nonlinear dispersive media.
\mbox{\vspace{1cm}}
{\bf Acknowledgments} This work is supported by the NSF of China under Grant No.11271210, the K. C. Wong Magna Fund in Ningbo University and the Natural Science Foundation of Ningbo under Grant No. 2011A610179. J. S. H acknowledges sincerely Prof. A. S. Fokas for arranging the visit to Cambridge University in 2012-2014 and for many useful discussions.
{A. C. acknowledges partial support from the Australian Research Council (Discovery Projects No. DP1093349) and support from the Isaac Newton Institute for Mathematical Sciences.}
|
\section*{Acknowledgments}
The computing resources used for our numerical study and the related technical support have been partly provided by the CRESCO/ENEA\-GRID High Performance Computing infrastructure and its staff \cite{cresco}. CRESCO ({\color{red}C}omputational {\color{red} RES}earch centre on {\color{red} CO}mplex systems) is funded by ENEA and by Italian and European research programmes.
\section{Dirichlet integrals}
\subsection{Normalization of ${\cal D}_\alpha$}
It is worthwhile describing a simple technique to calculate polynomial
integrals on the simplex by means of a specific example, namely the
normalization coefficient of the Dirichlet distribution
${\cal D}_\alpha(\bar\phi) = Z_{{\cal D}}(\alpha,s)^{-1}
\phi_1^{\slashed{\alpha}_1}\ldots
\phi_{Q-1}^{\slashed{\alpha}_{Q-1}}\left(s-|\bar\phi|\right)^{\slashed{\alpha}_Q}$. Specifically,
the integral that we aim at calculating is
\begin{equation}
Z_{\cal D}(\alpha,s) = \int_{\bar T_Q(s)}\text{d}
\bar\phi\ \phi_1^{\slashed{\alpha}_1}\ldots
\phi_{Q-1}^{\slashed{\alpha}_{Q-1}}\left(s-|\bar\phi|\right)^{\slashed{\alpha}_Q}\,.
\end{equation}
By introducing a Dirac delta function, this integral can be brought to
the equivalent form
\begin{equation}
Z_{\cal D}(\alpha,s) = \int_0^s\text{d} \phi_1\ldots \int_0^s\text{d} \phi_{Q}
\ \phi^{\slashed{\alpha}}\delta\left(s-|\phi|\right)\,.
\end{equation}
Moreover, owing to the Dirac delta function, all the upper integration
limits can be pushed to infinity without changing the integral. If we
replace the Dirac delta function by its Fourier representation
\begin{equation}
\delta(z) = \frac{1}{2\pi}\int_{-\infty}^{+\infty}\text{d} \lambda \text{e}^{-i\lambda z}\,,
\end{equation}
and rotate $\lambda\to i\lambda$, the integral turns into a complex
one, performed along the imaginary axis, {\it i.e.\ }
\begin{equation}
Z_{\cal D}(\alpha,s) = \frac{1}{2\pi i}\int_{-i\infty}^{+i\infty}\text{d} \lambda\ \text{e}^{\lambda s}\ \left(\prod_{k=1}^{Q} \int_0^{+\infty}\text{d} \phi_k\ \phi_k^{\alpha_k-1} \text{e}^{-\lambda \phi_k}\right)\,,
\end{equation}
The inner integrals are Laplace transforms of monomials. They sum to
\begin{equation}
\int_0^{+\infty}\text{d} \phi\ \phi^{\alpha-1} \text{e}^{-\lambda \phi} = \frac{\Gamma(\alpha)}{\lambda^{\alpha}}\,.
\end{equation}
Hence it follows
\begin{equation}
Z_{\cal D}(\alpha,s) = \dfrac{\prod_{k=1}^{Q}\Gamma(\alpha_k)}{2\pi
i}\int_{-i\infty}^{+i\infty}\text{d} \lambda\ \frac{\text{e}^{\lambda
s}}{\lambda^{|\alpha|}}\ =
\dfrac{\prod_{k=1}^{Q}\Gamma(\alpha_k)}{\Gamma(|\alpha|)}
s^{|\alpha|-1}\,,
\label{eq:wk}
\end{equation}
as a result of the Laplace antitransform of $\lambda^{-|\alpha|}$.
\subsection{Integrals of $V_\alpha$ and $U_\alpha$ on $\bar T_Q(s)$}
The integral of $V_\alpha$ on $\bar T_Q(s)$ follows trivially from
a term--by--term integration of its monomial expansion, namely
\begin{align}
Z_V(\alpha,s) & \equiv \int_{\bar T_Q(s)}\text{d}\bar\phi\,
V_\alpha(\bar\phi) \ = \sum_{\beta\le\alpha}v_{\alpha\beta}(\kappa)
\frac{\prod_{m=1}^{Q-1}\Gamma(\beta_m+1)}{\Gamma(|\beta|+Q)}\,.
\label{eq:Vint}
\end{align}
The integral of $U_\alpha$ on $\bar T_Q(s)$ can be similarly
calculated, provided we first represent it as a monomial sum. To this
aim,
we just need to apply the standard binomial formula
\begin{equation}
\partial_x^{k}[f(x)g(x)] = \sum_{\ell = 0}^k {k\choose
l}[\partial_x^{k-\ell}f(x)][\partial_x^\ell g(x)]\,,
\end{equation}
in sequence to the various factors of eq.~(\ref{eq:Ubasis}). We also
observe that, given $k,m\in\mathds{N}$ with $k\le m$, it holds
\begin{align}
& \partial_x^k x^m = k!{m\choose k}x^{m-k}\,,\\[2.0ex]
& \partial_x^k (s-x)^{m} = (-1)^k k!{m\choose k}(s-x)^{m-k}\,.
\end{align}
Accordingly, we have
\begin{equation}
U_\alpha(\bar\phi) =
\frac{1}{{\cal D}_\kappa(\bar\phi)}\partial_2^{\alpha_2}\ldots\partial_{Q-1}^{\alpha_{Q-1}}
P_1(\bar\phi)\,.
\end{equation}
with
\begin{align}
P_1(\bar\phi) & = \phi_2^{\alpha_2+\kappa_2-1}\ldots
\phi_{Q-1}^{\alpha_{Q-1}+\kappa_{Q-1}-1}\partial_1^{\alpha_1}
\left\{\phi_1^{\alpha_1+\kappa_1-1}(s-|\bar\phi|)^{|\alpha|+\kappa_{Q}-1}\right\}
\nonumber\\[2.0ex]
& =
\alpha_1!\sum_{\eta_1=0}^{\alpha_1}(-1)^{\eta_1}{\alpha_1+\kappa_1-1\choose
\alpha_1-\eta_1}{|\alpha|+\kappa_{Q}-1 \choose
\eta_1}\nonumber\\[2.0ex]
& \cdot\phi_1^{\eta_1+\kappa_1-1}\phi_2^{\alpha_2+\kappa_2-1}\ldots
\phi_{Q-1}^{\alpha_{Q-1}+\kappa_{Q-1}-1}(s-|\bar\phi|)^{|\alpha|-\eta_1+\kappa_{Q}-1}\,.
\end{align}
Upon iterating the above calculation over all derivatives, we arrive at
\begin{align}
U_\alpha(\bar\phi) & =
\left[\prod_{m=1}^{Q-1}\Gamma(\alpha_m+1)\right]\sum_{\eta\le\alpha}(-1)^{|\eta|}\prod_{m=1}^{Q-1}{\alpha_m
+
\kappa_m-1\choose \alpha_m-\beta_m}{|\alpha|-L_m+\kappa_{Q}-1\choose
\eta_m}\nonumber\\[2.0ex]
& \hskip 2.0cm \cdot
\phi_1^{\eta_1}\ldots\phi_{Q-1}^{\eta_{Q-1}}(s-|\phi|)^{|\alpha|-|\eta|}\,,
\end{align}
with $L_1\equiv 0$ and $L_m \equiv \sum_{k=1}^{m-1}\eta_k$ for $m\ge
2$. Therefore, we have
\begin{align}
& Z_U(\alpha,s) \equiv \int_{\bar T_Q(s)}\text{d}\bar\phi\,U_\alpha(\bar\phi) =
\frac{\prod_{m=1}^{Q-1}\Gamma(\alpha_m+1)}{\Gamma(|\alpha|+Q)} \nonumber\\[2.0ex]
& \hskip 1.0cm \cdot
\sum_{\eta\le\alpha}(-1)^{|\eta|}\,\Gamma(|\alpha|-|\eta|+1)
\prod_{m=1}^{Q-1}\Gamma(\eta_m+1){\alpha_m+\kappa_m-1\choose
\alpha_m-\eta_m}{|\alpha|-L_m+\kappa_{Q}-1\choose \eta_m}\,.
\end{align}
\end{appendices}
\section{Introduction}
Multivariate birth--death models have since long captured the interest
of researchers in statistical physics as they represent a natural
mathematical framework to investigate a plethora of interdisciplinary
problems, ranging from opinion diffusion to language emergence, cultural
dissemination and epidemic
spreading~\cite{FortCastLor,Keeling}. Broadly speaking, such
models describe an evolving population of agents, each lying in one of
$Q$ allowed physical states. The system is macroscopically represented by a
state vector $\phi=(\phi_k)_{k=1}^Q$, with $\phi_k$ denoting the fraction of
agents in the $k$--th state. By definition, for
$Q<\infty$ the state vector lives on the $Q$--simplex
\begin{equation}
T_Q(s) = \left\{\phi\in\mathds{R}_+^Q:\quad \sum_{k=1}^Q\phi_k = s\right\}\,.
\label{eq:simplex}
\end{equation}
The meaning of the parameter $s$ will become clear in the sequel,
while for the time being the reader may assume $s=1$. If the
microscopic dynamics of the model is determined by Markovian agent--agent
interactions altering the components of $\phi$, then in the
thermodynamic limit the system is known to obey a Fokker--Planck equation (FPE)
(see for instance ref. \cite[chapt.~7]{Gardiner}),
\begin{align}
& \label{eq:fpe}\partial_t {\cal P}(t,\bar\phi) = -\sum_{k=1}^{Q-1}
\partial_k\left[A_k(t,\bar\phi){\cal P}(t,\bar\phi)\right] +
\frac{1}{2}\sum_{i,k=1}^{Q-1}\partial_i\partial_k\left[B_{ik}(t,\bar\phi)
{\cal P}(t,\bar\phi)\right] \equiv{\cal L}_{\scriptscriptstyle{\rm FP}}\cdot{\cal P}(t,\bar\phi)\,,\\[1.0ex]
& \label{eq:bc} {\cal P}(t,\bar\phi) = 0 \quad \text{if}\quad \phi\notin T_Q(s)\,.
\end{align}
where $\bar\phi=(\phi_k)_{k=1}^{Q-1}$ denotes the essential state vector
(obtained from $\phi$ by conventionally leaving out its $Q$--th
component), ${\cal P}(t,\bar\phi)$ represents the probability density of
$\bar\phi$ at time $t$ and we define $\partial_t \equiv
\partial/\partial t$ and $\partial_k \equiv
\partial/\partial\phi_k$. ${\cal L}_{\scriptscriptstyle{\rm FP}}$ is commonly referred to as
the Fokker--Planck operator. We do not impose any initial condition to
eq.~(\ref{eq:fpe}), as this is not relevant to our aims.
We shall make the assumption -- valid for several birth--death models
-- that the drift coefficients $(A_k)_{k=1}^{Q-1}$ and the diffusion
ones $(B_{ik})_{i,k=1}^{Q-1}$ are time--independent polynomials of the
components of $\bar\phi$. We shall also assume that the stochastic
dynamics of the model has no exit states such as consensus or
no--infected--agents configurations. If this the case, the system is
expected to asymptotically relax to a dynamic equilibrium, with
$\phi$ wandering across $T_Q(s)$ according to the stochastic process
\begin{equation}
\text{d}\phi_\ell(t) = A_\ell(\bar\phi)\text{d} t + \sum_{k=1}^{Q-1}C_{\ell
k}(\bar\phi)\text{d} W_k(t) + \text{d} K_\ell(t)\,,\qquad
\ell=1,\ldots,Q-1\,,
\label{eq:stochproc}
\end{equation}
and eventually distributing according to a limit probability density
${\cal P}(\bar\phi)=\lim_{t\to\infty}{\cal P}(t,\bar\phi)$. Here, the matrix $C$
is related to the diffusion matrix via $B = C \cdot \trans{C}$
($\trans{C}$ is the transposed matrix of $C$), $W(t)=\left(W_k(t)\right)_{k=1}^{Q-1}$ is a
Wiener process describing the stochastic diffusion of the state vector
and $K(t) = \left(K_k(t)\right)_{k=1}^{Q-1}$ is a Skorokhod bounded variation
process~\cite{Skorokhod}, increasing only when $\phi\in\partial
T_Q(s)$ so as to ensure the boundary condition, eq.~(\ref{eq:bc}).
Numerical simulations of eq.~(\ref{eq:stochproc}) can be efficiently
used to make quantitative statements on ${\cal P}(\bar\phi)$, yet they give
little insight on its analytic structure. From this point of view, a
more convenient approach would be to represent the equilibrium
distribution in terms of a properly chosen function basis. A
legitimate possibility is to consider polynomial distributions on the
$Q$--simplex. In regard to this choice, we recall that a Dirichlet
distribution
$\text{Dir}(\gamma)$ of order $Q$ with parameter $\gamma\in\mathds{N}^Q$, has
probability density
\begin{align}
\label{eq:Dirichlet}
& {\cal D}_\gamma(\bar\phi) =
\frac{\Gamma(|\gamma|)}{\prod_{k=1}^{Q}\Gamma(\gamma_k)}s^{1-|\gamma|}\left(
\prod_{k=1}^{Q-1}
\phi_k^{\slashed{\gamma}_k}\right)(s-|\bar\phi|)^{\slashed{\gamma}_Q}\,,\qquad
\bar\phi\in\bar T_{Q}(s)\,,\\[1.0ex]
& \bar T_Q(s) = \{\bar\phi\in\mathds{R}_+^{Q-1}: \ |\bar\phi|\le s\}\,,
\end{align}
with $\slashed{\gamma}_k \equiv \gamma_k-1$ and $|x| \equiv \sum
x_k$. We notice that $\bar T_Q(s)$ is in one--to--one correspondence
with $T_Q(s)$, hence we can equivalently write $\phi\in T_Q(s)$ or
$\bar\phi\in\bar T_Q(s)$. It is crucial for the reader who is
unfamiliar with the mathematics of the simplex to learn how to
calculate Dirichlet integrals, {\it i.e.\ } polynomial integrals on $\bar
T_Q(s)$. As an example, we review in App.~A an elegant way to work out
the normalization constant of eq.~(\ref{eq:Dirichlet}). This is
sufficient to be able to reproduce all calculations presented in the
paper. That being said, an important feature of the Dirichlet
distributions is represented by
\begin{myprop} Dirichlet distributions with positive integer
indices provide a basis of polynomials, that is to say
\vskip -0.5cm
\begin{equation}
{\rm span}\left\{{\cal D}_\gamma(\bar\phi):\ \gamma\in\mathds{N}^Q \text{ and
} \ |\slashed{\gamma}| =
n\right\} = {\rm
span}\left\{\bar\phi^\alpha:\ \alpha\in\mathds{N}_0^{Q-1} \text{ and }
|\alpha|\le n\right\}\,,
\label{eq:span}
\end{equation}
where we make use of the multi--index notation $\bar\phi^\alpha
\equiv \phi_1^{\alpha_1}\cdot\ldots\cdot\phi_{Q-1}^{\alpha_{Q-1}}$.
\end{myprop}
\noindent\emph{Proof.} Given $\gamma\in\mathds{N}^{Q}$ with
$|\slashed{\gamma}| = n$, ${\cal D}_\gamma(\bar\phi)$ is a polynomial with
degree~$n$, hence it can be written as a linear combination of
monomials with degree $\le n$. Conversely, suppose that $\alpha\in
\mathds{N}_0^{Q-1}$ and $|\alpha|=n$. Then $\bar\phi^\alpha \propto
{\cal D}_{\gamma}(\bar\phi)$ with $\gamma =
(\alpha_1+1,\ldots,\alpha_{Q-1}+1,1)$. Finally, if $\alpha\in
\mathds{N}_0^{Q-1}$ and $|\alpha|=m<n$, then we define $\gamma =
(\alpha_1+1,\ldots,\alpha_{Q-1}+1,n-m+1)$ such that
$|\slashed{\gamma}|=n$, and we observe that
\begin{equation}
{\cal D}_{\gamma}(\bar\phi) \propto \bar\phi^{\alpha}(s-|\bar\phi|)^{n-m}
\propto \bar\phi^\alpha + {\cal E}_\alpha(\bar\phi)
\end{equation}
with ${\cal E}_\alpha(\bar\phi)$ being a linear combination of monomials,
each with degree $>m$. Therefore, the proof can be obtained by
backward induction on $m=n-1,n-2,\ldots.$\qed
\noindent Motivated by this observation, we introduce a polynomial
approximation ${\cal P}_n$ to ${\cal P}$ with degree $n$, reading
\begin{align}
\label{eq:polexp}
{\cal P}_n(\bar\phi) & = \sum_{\gamma\in\Omega_n}c_\gamma
{\cal D}_\gamma(\bar\phi)\,,\qquad \Omega_n =
\{\gamma\in\mathds{N}^Q:\ |\slashed{\gamma}|=n\}\,,\\[1.0ex]
\label{eq:polnorm}
\sum_{\gamma\in\Omega_n}c_\gamma & = 1\,.
\end{align}
Owing to Prop.~1, eq.~(\ref{eq:polexp}) is equivalent to a complete
sum over all monomials of degree $\le n$, while eq.~(\ref{eq:polnorm})
is just obtained by imposing that ${\cal P}_n$ is correctly normalized on
$\bar T_Q(s)$. Sometimes, $\Omega_n$ is referred to by mathematicians
as the \emph{bucket space}. The choice of the Dirichlet distributions
as a polynomial basis is favourable for several reasons, as we shall
explain in next sections.
The aim of the present paper is to describe how estimates of the
expansion coefficients $c \equiv \{c_\alpha\}$ can be determined
straightaway from the FPE, with a view to providing a hopefully
helpful analysis tool to practitioners in the physics of complex
systems. To this end, we adapt to eqs.~(\ref{eq:fpe})--(\ref{eq:bc}) a
mathematical technique known as the Ritz--Galerkin (RG) method for
partial differential equations (see for instance~\cite{finitel} for
a technical introduction), which is commonly used by engineers in many
applicative fields, including fluid and solid mechanics,
hydrodynamics, wave propagation, electromagnetism and many
others~\cite{spectral}. Our approach makes use of orthogonal
polynomials on $\bar T_Q(s)$ as test functions and point--like
zero--orthogonal--flux conditions on $\partial \bar T_Q(s)$.
The paper, which is written in a pedagogical style with detailed
calculations, is organized as follows. In sect.~2, we provide a short
compendium of orthogonal polynomials on $\bar T_Q(s)$, while in
sect.~3 we review the basics of the RG method and discuss
how to apply it to the FPE for birth--death models with polynomial
drift and diffusion coefficients. In sects.~4 and~5, we show
applications of the method respectively to the binary voter model with
zealots studied in \cite{Mobilia2007}, for which an exact solution of
the FPE is known, and to its generalization to the multi--state
case. In sect.~6, we discuss how symmetry arguments can help reduce
the computational budget needed to implement the method. We finally
draw our conclusions in sect.~7.
\section{Orthogonal polynomials on the simplex}
Since ${\cal D}_\gamma$ is a polynomial distribution on $\bar
T_Q(s)$ with degree $n$ for $\gamma\in\Omega_n$, it is rather natural
to look for orthogonal polynomial bases on the simplex. The
general theory of multivariate orthogonal polynomials is still an open
research field: it does not belong to the average undergraduate
background of physicists and is not even discussed in many essays in
the mathematical literature. Fortunately, an excellent introduction is
provided in ref.~\cite{XuBook}. We refer the reader to that book for a
comprehensive presentation of classical and recent developments on the
subject, while for the sake of readability and self--consistency of
the paper we review here those aspects which are closely related to
our ends.
\newpage
First of all, a multivariate polynomial $P_\alpha$ on $\bar T_Q(s)$,
indexed by $\alpha\in\mathds{N}_0^{Q-1}$ can be always represented by its
monomial expansion
\begin{equation}
P_\alpha(\bar\phi) = \sum_{\beta\le\alpha} c_\beta \bar\phi^\beta
\equiv \sum_{\beta_1=0}^{\alpha_1}\ldots
\sum_{\beta_{Q-1}=0}^{\alpha_{Q-1}}
c_{\beta_1\ldots\beta_{Q-1}}\phi_1^{\beta_1}\ldots
\phi_{Q-1}^{\beta_{Q-1}}\,.
\end{equation}
The degree of $P_\alpha$ is the maximum degree of its monomials, {\it i.e.\ }
$\deg\{P_\alpha\}=|\alpha|$. Secondly, the orthogonality notion on
$\bar T_Q(s)$ depends on the introduction of a scalar product, which
in turn requires the specification of a measure. The standard choice
-- which we adopt here -- is to weight the Lebesgue measure by a
Dirichlet distribution, {\it i.e.\ } to define
\begin{equation}
\langle f,g\rangle_\kappa = \int_{\bar T_Q(s)}\text{d}
\bar\phi\ f(\bar\phi)g(\bar\phi){\cal D}_\kappa(\bar\phi)\,,\qquad
\kappa\in\mathds{N}^{Q}\,,
\label{eq:defscalprod}
\end{equation}
for sufficiently regular functions $f,g$ on $\bar T_Q(s)$. The
Dirichlet weight is such that $\langle 1,1\rangle_\kappa =
1$. Thirdly, two polynomials $P$ and $Q$ on $\bar T_Q(s)$ are said to
be orthogonal if $\langle P,Q\rangle_\kappa = 0$, while a polynomial
$P$ is called an orthogonal polynomial if it is orthogonal to all
polynomials of lower degree, {\it i.e.\ }
\begin{equation}
\langle P,Q\rangle_\kappa = 0\,, \qquad \forall Q \quad \text{ with
}\quad \deg\{Q\}<\deg\{P\}\,.
\end{equation}
The main difference between orthogonal polynomial bases in one and
several variables, is that the former count just one element per
degree, whereas the latter count many of them. To be precise, it can
be shown as a trivial consequence of Prop.~1 that
\begin{equation}
\dim\left\{\text{orthogonal polynomials } P \text{ on } \bar
T_Q(s):\ \deg\{P\}\le n \right\} = |\Omega_n| = {n+Q-1 \choose n}\,.
\end{equation}
Notice that the \emph{bucket space} expands roughly as
\begin{equation}
|\Omega_n| \approx
\frac{\exp\left\{(Q-1)(H_n-\gamma_{\scriptscriptstyle{\rm
E}})\right\}}{\Gamma(Q)}\,,\qquad H_n = \sum_{k=1}^n
\frac{1}{k}\,,
\end{equation}
with $H_n$ being the $n$--th harmonic number and
$\gamma_{\scriptscriptstyle{\rm E}} = 0.57721...$ the
Euler--Mascheroni constant. For this reason, the RG method becomes
computationally challenging even for models with a moderately large
value of $Q$.
Now, there exist several sets of orthogonal polynomials
with respect to eq.~(\ref{eq:defscalprod}). Along with
ref.~\cite{XuBook}, we focus on two of them, namely
\vskip 0.2cm
{\raggedleft \underline{the monomial basis}}
\begin{equation}
\label{eq:Vbasis}
V_\alpha(\bar\phi) = \sum_{\beta\le\alpha}
(-1)^{|\alpha|+|\beta|}s^{-|\beta|}\prod_{i=1}^{Q-1}{\alpha_i \choose
\beta_i}\dfrac{({\kappa_i})_{\alpha_i}}{({\kappa_i})_{\beta_i}}
\dfrac{(|\kappa|-1)_{|\alpha|+|\beta|}}{(|\kappa|-1)_{2|\alpha|}}\,\bar\phi^\beta
\equiv \sum_{\beta\le\alpha}v_{\alpha\beta}(\kappa)\bar\phi^\beta\,,
\end{equation}
{\raggedleft \underline{the Appel basis}}
\begin{equation}
\label{eq:Ubasis}
U_\alpha(\bar\phi) =
{\cal D}_\kappa(\bar\phi)^{-1}\partial^{|\alpha|}_\alpha
\left[\phi_1^{\alpha_1+\kappa_1-1}\ldots
\phi_{Q-1}^{\alpha_{Q-1}+\kappa_{Q-1}-1}
(s-|\bar\phi|)^{|\alpha|+\kappa_{Q}-1} \right]\,,
\end{equation}
for $\alpha\in\mathds{N}_0^{Q-1}$, with $\partial^{|\alpha|}_\alpha \equiv
\partial^{|\alpha|}/\partial x_1^{\alpha_1}\ldots \partial
x_d^{\alpha_d}$ and with $(x)_n = x(x+1)\ldots (x+n-1)$ denoting the
Pochhammer symbol (also known as the raising factorial). The
polynomials $\{V_\alpha\}$ and $\{U_\alpha\}$ fulfill the following
properties:
\begin{myprop}
For any polynomial $P_\beta$ on $\bar T_Q(s)$, it holds
\begin{equation}
\langle V_\alpha,P_\beta\rangle_\kappa = \langle
U_\alpha,P_\beta,\rangle_\kappa = 0 \quad \text{if} \quad
|\beta| < |\alpha|\,.
\end{equation}
Moreover, the polynomials $\{V_\alpha:\,\alpha\in\mathds{N}_0^{Q-1}\}$ and
$\{U_\alpha:\,\alpha\in\mathds{N}_0^{Q-1}\}$ are biorthogonal, {\it i.e.\ } they
fulfill
\begin{equation}
\langle V_\alpha, U_\beta\rangle_\kappa = f_\alpha\,
\delta_{\alpha\beta}\,,\qquad f_\alpha = s^{|\alpha|}
\dfrac{\left[\prod_{m=1}^{Q-1}
(\kappa_m)_{\alpha_m}\Gamma(\alpha_m+1)\right](\kappa_{Q})_{|\alpha|}}
{(|\kappa|)_{2|\alpha|}}\,.
\end{equation}
\end{myprop}
\noindent\emph{Proof.} The proof is contained in
ref.~\cite[chap.~2]{XuBook}. Here, we only review the argument showing
that $\langle V_\alpha, P_\beta\rangle_\kappa=0$, since we shall need
a formula which is derived along the proof. Owing to Prop. 1, it is
sufficient to prove that $\langle V_\alpha,X_\gamma
\rangle_{\kappa}=0$ for
\begin{equation}
X_\gamma(\bar\phi) =
\,\left[\prod_{m=1}^{Q-1}\phi_m^{\slashed{\gamma}_m}\right]
(s-|\bar\phi|)^{\slashed{\gamma}_{Q}}\,,
\end{equation}
with $\gamma\in\mathds{N}^{Q}$ and $|\slashed{\gamma}|=|\alpha|-1$. Indeed, it holds
\begin{align}
\langle V_\alpha,X_\gamma\rangle_\kappa & = \int_{\bar T_Q(s)}\text{d}\bar\phi\
V_\alpha(\bar\phi) X_\gamma(\bar\phi) {\cal D}_{\kappa}(\bar\phi)\nonumber\\[1.0ex]
& =
\frac{\Gamma(|\kappa|)}{\prod_{k=1}^{Q}\Gamma(\kappa_k)}s^{1-|\kappa|}
\sum_{\beta\le\alpha}(-1)^{|\alpha|+|\beta|}s^{-|\beta|}
\prod_{m=1}^{Q-1}{\alpha_m \choose
\beta_m}\dfrac{({\kappa_m})_{\alpha_m}}
{({\kappa_m})_{\beta_m}}\dfrac{(|\kappa|-1)_{|\alpha|+|\beta|}}
{(|\kappa|-1)_{2|\alpha|}}\nonumber\\[0.0ex]
\phantom{\langle V_\alpha,X_\gamma\rangle_\kappa} & \cdot \int_{\bar
T_Q(s)}\text{d}\bar\phi\ \prod_{\ell=1}^{Q-1}\phi_\ell^{\beta_\ell+
\slashed{\gamma}_\ell+\kappa_\ell-1}
(s-|\bar\phi|)^{\kappa_{Q}+\slashed{\gamma}_{Q}-1}\nonumber\\[1.0ex]
& =
s^{|\slashed{\gamma}|}\frac{\Gamma(|\kappa|)}{\Gamma(|\kappa|+2|\alpha|-1)}
\left[\prod_{m=1}^{Q-1}\frac{\Gamma(\kappa_m+\alpha_m)\Gamma(\slashed{\gamma}_m+1)}
{\Gamma(\kappa_m)}\right]\frac{\Gamma(\kappa_{Q}+\slashed{\gamma}_{Q})}
{\Gamma(\kappa_Q)}\nonumber\\[0.0ex]
& \cdot
\sum_{\beta\le\alpha}\prod_{m=1}^{Q-1}(-1)^{\beta_m}{\alpha_m\choose\beta_m}
{\kappa_m+\slashed{\gamma}_m+\beta_m-1\choose
\slashed{\gamma}_m} \frac{\Gamma(|\kappa| + |\alpha| +
|\beta|-1)} {\Gamma(|\kappa| + |\slashed{\gamma}| + |\beta|)}\,.
\end{align}
If $|\slashed{\gamma}|=|\alpha|-1$, the rightmost ratio of
$\Gamma$--functions simplifies and we are left with a product of
independent sums. From a Chu--Vandermonde formula
\begin{equation}
\sum_{k=0}^n(-1)^k{n\choose k}{a+k\choose m} =
\frac{(-m)_n}{\Gamma(m+1)(1+a)_{n-m}}\,,\qquad
m,n\in\mathds{N}\,,\ a\in\mathds{R}\,,
\end{equation}
it follows
\begin{align}
\langle V_\alpha,X_\gamma\rangle_\kappa & =
s^{|\slashed{\gamma}|}\frac{\Gamma(|\kappa|)}{\Gamma(|\kappa|+2|\alpha|-1)}
\left[\prod_{m=1}^{Q-1}\frac{\Gamma(\kappa_m+\alpha_m)\Gamma(\slashed{\gamma}_m+1)}
{\Gamma(\kappa_m)}\right]\frac{\Gamma(\kappa_{Q}+
\slashed{\gamma}_{Q})}{\Gamma(\kappa_Q)}\nonumber\\[0.0ex]
& \cdot
\prod_{m=1}^{Q-1}\frac{(-\slashed{\gamma}_m)_{\alpha_m}}
{\Gamma(\slashed{\gamma}_m+1)(\kappa_m+\slashed{\gamma}_m)_{\alpha_m-\slashed{\gamma}_m}}\,.
\end{align}
However, we know that $(-m)_n=0$ for $m<n$. Since
$|\slashed{\gamma}|=|\alpha|-1$, there is at least one value of $m$
for which $\slashed{\gamma}_m<\alpha_m$. Therefore we conclude that
$V_\alpha$ is orthogonal to $X_\gamma$. \qed
\begin{myremark}
The polynomials $\{V_\alpha\}$ can be easily coded. Indeed, we use
them for computations. However, it should be noticed that the
numerical evaluation of eq.~(\ref{eq:Vbasis}) can be critical,
specially for $|\alpha|\gg 1$, since $V_\alpha$ adds largely different
ratios of factorials with alternating signs. For this reason,
computations should be performed and crosschecked with different
levels of floating point rounding. Most of the numerical experiments
described in next sections have been done in Maple$\texttrademark$,
which allows to control the numerical precision by the environment
variable {\tt Digits}. We employ the biorthogonal basis $\{U_\alpha\}$
essentially to develop theoretical arguments. \qed
\end{myremark}
\begin{figure}[t!]
\centering
\includegraphics[width=0.6\textwidth]{chi.pdf}
\caption{ \footnotesize Heat map of $\log|\chi_{\alpha\gamma}|$ for
$Q=4$, $n=14$, $\kappa = (2,2,2,2)$ and $s=1$. In this case
$|\Omega_n| = 680$.}
\label{fig:chi}
\end{figure}
\noindent If $|\slashed{\gamma}|\ne |\alpha|-1$, the factorization
property does not hold, hence we are left with the general formula
\begin{empheq}[box=\mybluebox]{align}
\label{eq:scalprod}
\chi_{\alpha\gamma} \equiv \langle V_\alpha,{\cal D}_\gamma\rangle_\kappa &
= s^{1-Q}
\frac{\Gamma(|\gamma|)\Gamma(|\kappa|)}{\Gamma(|\kappa|+2|\alpha|-1)}
\left[\prod_{m=1}^{Q-1}\frac{\Gamma(\kappa_m+\alpha_m)}{\Gamma(\kappa_m)}\right]\frac{\Gamma(\kappa_{Q}+{\gamma}_{Q}-1)}{\Gamma(\kappa_Q)\Gamma(\gamma_Q)}\nonumber\\[0.0ex]
& \hskip -0.4cm \cdot \sum_{\beta\le\alpha}\frac{\Gamma(|\kappa| +
|\alpha| + |\beta|-1)}{\Gamma(|\kappa| + |{\gamma}| +
|\beta|-Q)}\prod_{m=1}^{Q-1}(-1)^{\beta_m}{\alpha_m\choose\beta_m}{\kappa_m+{\gamma}_m+\beta_m-2\choose
{\gamma}_m-1}\,,
\end{empheq}
valid for $|\alpha|\le|\slashed{\gamma}|$. We can extend Remark 1 to
eq.~(\ref{eq:scalprod}) as well. The meaning of the matrix $\chi =
\{\chi_{\alpha\gamma}\}$ becomes clear if we expand ${\cal D}_\gamma$ along
the Appel basis, namely
\begin{equation}
{\cal D}_{\gamma}(\bar\phi) =
\sum_{0\le|\beta|\le|\slashed{\gamma}|}d_{\gamma\beta}\,U_\beta(\bar\phi)\,.
\label{eq:Dirproj}
\end{equation}
By projecting both sides of eq.~(\ref{eq:Dirproj}) onto $V_\alpha$, we
obtain
\begin{align}
\chi_{\alpha\gamma} & = \langle V_\alpha,{\cal D}_{\gamma}\rangle_\kappa =
\sum_{0\le|\beta|\le|\slashed{\gamma}|}d_{\gamma\beta}\, \langle
V_\alpha,U_\beta\rangle_\kappa = \sum_{0\le|\beta|\le
|\slashed{\gamma}|}d_{\gamma\beta}\, f_\beta\, \delta_{\alpha\beta}
= d_{\gamma\alpha}\,f_\alpha\,\,,
\end{align}
whence it follows
\begin{equation}
{\cal D}_\gamma(\bar\phi) = \sum_{0\le|\beta|\le|\slashed{\gamma}|}\frac{\chi_{\beta\gamma}}{f_\beta}\,U_\beta(\bar\phi)\,.
\end{equation}
We thus conclude that, given $n\ge 0$, $\gamma\in\Omega_n$ and
$\alpha\in\mathds{N}_0^{Q-1}$ such that $|\alpha|\le |\slashed{\gamma}|$, the
matrix elements $\chi_{\alpha\gamma}$ are essentially the expansion
coefficients of the (non--orthogonal) basis $\{{\cal D}_\gamma\}$ along
the (orthogonal) Appel basis $\{U_\alpha\}$, {\it i.e.\ } $\chi$ is essentially
a change--of--basis matrix. It is interesting to look at the numerical
values of $\chi_{\alpha\gamma}$ in some specific case. As an example,
in Fig.~1 we show a heat map of $\log|\chi_{\alpha\gamma}|$ for
$Q=4$, $n=14$, $\kappa = (2,2,2,2)$ and $s=1$; here, the index
arrays are sorted in their respective domains according to a reverse
lexicographic ordering (RLO) $\alpha \to i_\text{rlx}(\alpha)$, which
we recall to be defined by
\begin{mydef}
Given $d\ge 1$ and $\alpha,\beta\in\mathds{N}_0^d$, we say that $\alpha
\prec \beta$ if $|\alpha|<|\beta|$ or
$|\alpha|=|\beta|$ and $\exists k\in\{1,\ldots,
d\}:\ \alpha_i=\beta_i$ for $i=1\ldots k-1$, and $\alpha_k<\beta_k$.
\end{mydef}
\noindent An efficient indexing algorithm for this specific ordering
is discussed in \cite{deboor}, to which we refer the reader for
details. We see from the heat map that the coefficients
$\chi_{\alpha\gamma}$ decrease exponentially as $|\alpha|$
increases. This behaviour looks natural if one considers that
$V_\alpha$ is not positive definite on $\bar T_Q(s)$: its zeros are an
algebraic variety, whose structure becomes more and more complex as
$|\alpha|$ increases. The sign of $V_\alpha$ is important since
${\cal D}_\gamma(\bar\phi)\ge 0$ for $\bar\phi\in\bar T_Q(s)$, hence
$\chi_{\alpha\gamma}$ receives contributions of opposite signs from
adjacent domains separated by zeros of $V_\alpha$. For this reason, it
averages progressively to zero as $|\alpha|$ increases. The isolated
white points in the upper part of the plot correspond to values of
$\alpha$ and $\gamma$ for which $\chi_{\alpha\gamma}=0$; they have
been coloured as the lowest non--zero observed value of
$\log|\chi_{\alpha\gamma}|$, just to preserve the colour map.
The reader could feel uncomfortable with the fact that the orthogonal
bases are not normalized on $\bar T_Q(s)$ such as ${\cal D}_\gamma$
is. This is not really a problem as far as we are concerned, since
normalization constants change the rows of the RG
matrix by an irrelevant overall rescaling, as we shall see in the next
section. Moreover, normalizing the orthogonal polynomials
$\{V_\alpha\}$ and $\{U_\alpha\}$ as if they were probability densities
on $\bar T_Q(s)$ is not possible. Indeed, their integrals $Z_V$ and
$Z_U$, which are discussed in App.~A for the sake
of completeness, vanish for specific values of $\alpha$ and~$\kappa$
due to non--positiveness.
\section{Ritz--Galerkin orthogonality}
The Fokker--Planck operator ${\cal L}_{\scriptscriptstyle{\rm FP}}$ is a linear operator: given two
functions ${\cal P}_1$, ${\cal P}_2\in{\cal H}$, with ${\cal H}$ a sufficiently regular
function space on the $Q$--simplex, such that ${\cal L}_{\scriptscriptstyle{\rm FP}}\cdot{\cal P}_k=0$ for
$k=1,2$, then ${\cal L}_{\scriptscriptstyle{\rm FP}}\cdot(a_1{\cal P}_1+a_2{\cal P}_2) = 0$ for any
$a_1,a_2\in\mathds{R}$. In other words, the solutions of the stationary FPE
belong to $\ker{\cal L}_{\scriptscriptstyle{\rm FP}}\subset {\cal H}$. Since in general $\dim\ker{\cal L}_{\scriptscriptstyle{\rm FP}}>1$,
a specific solution can be singled out by imposing a set of additional
conditions. We shall come to this point in a while. For the time
being, we observe that
\begin{equation}
\langle V,{\cal L}_{\scriptscriptstyle{\rm FP}}\cdot {\cal P}\rangle_\kappa = 0 \quad \forall\, V\in
{\cal H}\quad \text{ if } \quad {\cal P}\in\ker{\cal L}_{\scriptscriptstyle{\rm FP}}\,.
\label{eq:completesol}
\end{equation}
Conversely, a function ${\cal P}\in{\cal H}$ fulfilling $\langle V,{\cal L}_{\scriptscriptstyle{\rm FP}}\cdot
{\cal P}\rangle_\kappa = 0\ \ \forall\, V\in {\cal H}$ is called \emph{a weak
solution} of the FPE. The idea underlying the RG approximation
method is to look for a weak solution by enforcing
eq.~(\ref{eq:completesol}) only for $V\in \bar{\cal H}$, with $\bar{\cal H}$ a
properly chosen subset of ${\cal H}$. For instance, for $n\ge 1$, we could
opt for
\begin{equation}
\bar{\cal H}_n = {\rm span}\left\{\bar\phi^\alpha:\ \alpha\in\mathds{N}_0^{Q-1}
\text{ and } |\alpha|\le n\right\}\,.
\end{equation}
Since $\{V_\alpha\}_{|\alpha|\le n}$ is a basis of $\bar{\cal H}_n$, a RG
weak solution ${\cal P}_n$ has to fulfill
\begin{equation}
\langle V_\alpha,{\cal L}_{\scriptscriptstyle{\rm FP}}\cdot {\cal P}_n\rangle_\kappa = 0\,\qquad \forall\,
\alpha:\ |\alpha|\le n\,.
\label{eq:projectFPE}
\end{equation}
As proved in a celebrated theorem by Lax and Milgram
\cite{laxmilgram}, a sufficient condition to make the search of weak
solutions (and therefore of RG weak solutions) a well--posed problem,
is that the following two properties are fulfilled:
\begin{alignat}{4}
& \bullet \ \text{boundedness}& & \quad \rightsquigarrow & & \qquad
\exists C<+\infty:\quad |\langle V,{\cal L}_{\scriptscriptstyle{\rm FP}}\cdot V'\rangle| \le
C\,||V|| ||V'||\,,\quad \forall\,
V,V'\in{\cal H}\,,\\[2.0ex]
& \bullet \ \text{coerciveness}& & \quad \rightsquigarrow & & \qquad
\exists c>0: \quad |\langle V,{\cal L}_{\scriptscriptstyle{\rm FP}}\cdot V \rangle| \ge
c||V||^2\,,\quad \forall\, V\in{\cal H}\,,
\end{alignat}
for some scalar product $\langle\,\cdot\,,\cdot\,\rangle$ on $\bar T_Q(s)$ (not necessarily $\langle\,\cdot\,,\,\cdot\,\rangle_\kappa$), with $||\cdot||^2 = \langle \cdot,\cdot\rangle$ being
the induced norm. Though it is
not difficult to check the boundedness condition for a ${\cal L}_{\scriptscriptstyle{\rm FP}}$ with
polynomial coefficients $(A_k)_{k=1}^{Q-1}$ and $(B_{ik})_{i,k=1}^{Q-1}$ on a compact domain
such as $\bar T_Q(s)$, checking the coerciveness of ${\cal L}_{\scriptscriptstyle{\rm FP}}$ is more
problematic, since this is related to the structure of the eigenvalue
spectrum of $(B_{ik})_{i,k=1}^{Q-1}$. We do not attempt any general proof in the
present paper. Instead, we adopt a heuristic approach where we just
apply the RG method to a given complex model and check out the
outcome. However, if the Lax--Milgram conditions are fulfilled, then
the C\'ea estimate
\begin{equation}
||{\cal P} - {\cal P}_n|| \le \frac{C}{c}\inf_{{\cal Q}_n\in\bar{\cal H}_n}||{\cal P}-{\cal Q}_n||\,
\end{equation}
follows straightaway, stating that the RG solution ${\cal P}_n$ is a {\it
quasi}--best approximation on $\bar{\cal H}_n$ to a truly weak solution
${\cal P}\in{\cal H}$. In addition, the error ${\cal P}-{\cal P}_n$ is weakly orthogonal
to $\bar {\cal H}_n$. That being said, we are ready to show how to adapt
the RG method to birth--death models with polynomial drift and
diffusion coefficients.
\vskip 0.2cm
\noindent {$i$) If we expand ${\cal P}_n$ according to
eq.~(\ref{eq:polexp}) and insert the expansion into
eq.~(\ref{eq:projectFPE}), we obtain
\vskip -0.5cm
\begin{equation}
0 = \sum_{\gamma\in\Omega_n} \langle V_\alpha,{\cal L}_{\scriptscriptstyle{\rm FP}}
\cdot{\cal D}_\gamma\rangle_\kappa \,c_\gamma = \sum_{\gamma\in\Omega_n}
\psi_{\alpha\gamma}c_\gamma\,,\qquad \psi_{\alpha\gamma} \equiv
\langle V_\alpha,{\cal L}_{\scriptscriptstyle{\rm FP}}
\cdot{\cal D}_\gamma\rangle_\kappa\,.
\label{eq:homogRG}
\end{equation}
Accordingly, the stationary FPE turns into a square homogeneous linear
system with coefficient matrix $\psi\in
\mathds{R}^{|\Omega_n|\times|\Omega_n|}$ and unknown vector
$c\in\mathds{R}^{|\Omega_n|}$, both indexed (for instance) via $\alpha \to
i_\text{rlx}(\alpha)$. If the problem is well posed, the eigenvalue
spectrum of $\psi$ must have a certain number of zeros $\nu_0>0$,
depending on $n$, $Q$ and the specific form of $(A_k)_{k=1}^{Q-1}$ and
$(B_{ik})_{i,k=1}^{Q-1}$. The linear system $\psi\cdot c=0$ must be augmented by
imposing that $c$ is normalized according to eq.~(\ref{eq:polnorm})
and by introducing a set of --~say~-- $Q{n}_{\scriptscriptstyle{\rm bc}}$ additional equations to
enforce the boundary conditions, eq.~(\ref{eq:bc}). This leads us to a
larger non--homogeneous linear system $\Psi\cdot c = \eta$, of which
we know at present that
\begin{alignat}{3}
\label{eq:systeq1}
& \Psi\in \mathds{R}^{(|\Omega_n|+1+Q{n}_{\scriptscriptstyle{\rm bc}})\times |\Omega_n|}\,: & &
\qquad\left\{\begin{array}{ll}\Psi_{ij} = \psi_{ij} & \hskip 0.12cm
\text{for } i,j=1,..,|\Omega_n|\,;\\[2.0ex]
\Psi_{(|\Omega_n|+1)j} = 1 & \hskip 0.12cm \text{for }
j=1,\ldots,|\Omega_n|\,;
\end{array}\right.\\[2.0ex]
\label{eq:systeq2}
& \eta\in\mathds{R}^{(|\Omega_n|+1+Q{n}_{\scriptscriptstyle{\rm bc}})}\,: & & \qquad
\left\{\begin{array}{ll}\eta_{i} = 0 & \quad \text{for }
i=1,..,|\Omega_n|\,;\\[2.0ex]
\eta_{|\Omega_n|+1} = 1\,. & \end{array}\right.
\end{alignat}
It should be observed that eq.~(\ref{eq:homogRG}) is left invariant by
any change of normalization of $V_\alpha$. Such a change would just
correspond to rescaling the rows of $\psi$.
\vskip 0.2cm
\noindent $ii$) Let us see how to set up the boundary conditions and
fill in the lowest $Q{n}_{\scriptscriptstyle{\rm bc}}$ rows of $\Psi$ and elements
of~$\eta$. First, we recall that the stationary FPE can be written in
the form of a local conservation law, namely
\begin{equation}
0 = \sum_{k=1}^{Q-1}\partial_k J_k(\bar\phi)\,, \qquad\qquad
J_k(\bar\phi) = -A_k(\bar\phi){\cal P}(\bar\phi) +
\frac{1}{2}\sum_{i=1}^{Q-1}\partial_i[B_{ik}(\bar\phi){\cal P}(\bar\phi)]\,,
\label{eq:locFPE}
\end{equation}
where $J$ is naturally interpreted as a vector probability
current. Integrating both sides of eq.~(\ref{eq:locFPE}) over $\bar
T_Q(s)$ and making use of the divergence theorem yields
\begin{equation}
0 = \int_{\partial \bar T_Q(s)}\text{d} \bar\phi \ \ \hat n(\bar\phi)\cdot
J(\bar\phi)\,,\qquad Q>2\,,
\label{eq:zeroflux}
\end{equation}
with $\hat n(\bar\phi)$ representing the inward pointing unit vector
orthogonal to $\partial \bar T_Q(s)$ at $\bar\phi$. Clearly,
eq.~(\ref{eq:zeroflux}) means that there is no overall probability
flux across $\partial \bar T_Q(s)$ when the system is in
equilibrium. In order for eq.~(\ref{eq:bc}) to hold, the orthogonal
component of the probability current must vanish point-by-point on the
boundary (reflecting boundary conditions) and not just on average, {\it i.e.\ }
the sought solution must fulfill $\hat n(\bar\phi)\cdot J(\bar\phi)=0$
for $\bar \phi\in\partial \bar T_Q(s)$. Unfortunately, this is a
continuous infinite set of conditions, which we however approximate by
a finite subset. To this end, we observe that $\partial \bar T_Q(s)$
is made of $Q$ $(Q-1)$--dimensional hypersurfaces, namely
\begin{equation}
\partial\bar T_Q(s) = \bigcup_{k=1}^Q {\mathfrak H}_k\,, \qquad \left\{\begin{array}{ll}
{\mathfrak H}_k\, = \{\bar\phi\in\bar T_Q(s):\quad \phi_k=0\}\,;\qquad &
k=1,\ldots,Q-1\,,\\[2.0ex]
{\mathfrak H}_Q = \{\bar\phi\in\bar T_Q(s):\quad |\bar\phi|=s\}\,. &
\end{array}\right.
\end{equation}
On each hypersurface ${\mathfrak H}_k$ we consider a regular grid of
zero--orthogonal--flux points $(\bar\phi_{km})_{m=1}^{{n}_{\scriptscriptstyle{\rm bc}}}$ at
which we impose the condition $0= \hat n(\bar\phi_{km})\cdot
J(\bar\phi_{km})$, {\it i.e.\ }
\begin{alignat}{3}
\label{eq:bcorth}
& 0 = J_k(\bar\phi_{km}) \,,& & \qquad k = 1,\ldots,Q-1\, \quad
m=1,\ldots,{n}_{\scriptscriptstyle{\rm bc}}\,,\\[1.0ex]
& 0 = \sum_{k=1}^{Q-1}J_k(\bar\phi_{Qm})\,, & & \qquad
m=1,\ldots,{n}_{\scriptscriptstyle{\rm bc}}\,.
\label{eq:bcskew}
\end{alignat}
An illustrative example corresponding to $Q=3$ and ${n}_{\scriptscriptstyle{\rm bc}}=10$ is shown
in Fig.~\ref{fig:bc}, where the boundary points have been chosen
according to
\begin{equation}
\bar\phi_{1m} = \left(0,s\frac{m}{{n}_{\scriptscriptstyle{\rm bc}}+1}\right)\,,\quad
\bar\phi_{2m} = \left(s\frac{m}{{n}_{\scriptscriptstyle{\rm bc}}+1},0\right)\,,\quad
\bar\phi_{3m} =\left(s\frac{m}{{n}_{\scriptscriptstyle{\rm bc}}+1},s\frac{{n}_{\scriptscriptstyle{\rm bc}}+1-m}{{n}_{\scriptscriptstyle{\rm bc}}+1} \right)\,.
\end{equation}
\begin{figure}[t!]
\centering
\includegraphics[width=0.6\textwidth]{bc.pdf}
\vskip -0.2cm
\caption{ \footnotesize Zero--orthogonal--flux points for $Q=3$ and ${n}_{\scriptscriptstyle{\rm bc}}=10$.}
\label{fig:bc}
\end{figure}
\noindent $iii$) As already observed, the rank $r_\psi$ of $\psi$ is
expected not to be maximal, {\it i.e.\ }
$r_\psi=|\Omega_n|-\nu_0<|\Omega_n|\equiv r_{\psi,\text{max}}$. The
normalization condition eq.~(\ref{eq:polnorm}) adds a linearly
independent row to the system $\psi\cdot c=0$, thus increasing the
rank of the coefficient matrix by one. Each additional boundary
condition adds another linearly independent row and
further increases the rank of the coefficient matrix until this
becomes maximal. From this point on, {\it i.e.\ } for $Q{n}_{\scriptscriptstyle{\rm bc}}>\nu_0-1$, the rank
of the coefficient matrix keeps maximal, while the system becomes
overconstrained and thus inconsistent (to understand this, imagine to
perform a row echelon reduction of the system $\Psi\cdot c = \eta \to
\Psi_\text{red}\cdot c=\eta_\text{red}$; the reduced row echelon form
$\Psi_\text{red}$ has still maximal rank $r_\Psi=|\Omega_n|$; its last
$Q{n}_{\scriptscriptstyle{\rm bc}}-\nu_0+1$ rows are full of zeros, while in general the last
$Q{n}_{\scriptscriptstyle{\rm bc}}-\nu_0+1$ elements of $\eta_\text{red}$ are expected not to
vanish). This is particularly inconvenient, as it compels us to very
carefully choose an exact number $Q{n}_{\scriptscriptstyle{\rm bc}} = \nu_0-1$ of boundary
points. Though reasonable, we have no theoretical argument to prove
that $\nu_0-1 \propto Q$. We follow a different approach: an
alternative is indeed to impose an arbitrary number $Q{n}_{\scriptscriptstyle{\rm bc}}>\nu_0-1$ of
boundary conditions and consider the normal system
\begin{equation}
(\trans{\Psi}\cdot\Psi)\cdot c = \trans{\Psi}\cdot \eta\,.
\label{eq:normsyst}
\end{equation}
in place of the original one (least--squares
problem). Eq.~(\ref{eq:normsyst}) is consistent for any choice of
$Q{n}_{\scriptscriptstyle{\rm bc}}>\nu_0-1$. Indeed, since $\Psi$ has maximal rank,
$\trans{\Psi}\cdot\Psi\in \mathds{R}^{|\Omega_n|\times |\Omega_n|}$ has no
zero eigenvalues, hence it can be inverted. The system is consistent
as $(\trans{\Psi}\cdot\eta)_k = 1$ for all
$k=1,\ldots,|\Omega_n|+1+Q{n}_{\scriptscriptstyle{\rm bc}}$, as a consequence of
eqs.~(\ref{eq:systeq1})--(\ref{eq:systeq2}). Clearly, the original
system and eq.~(\ref{eq:normsyst}) are not equivalent.
Notice that $\Psi$ is in general expected to have a large condition
number (the ratio between its largest and lowest singular value) and
the latter is expected to get larger as $n$ increases. Since the
condition number of $\trans{\Psi}\cdot\Psi$ is the square of the
condition number of $\Psi$, the inversion of $\trans{\Psi}\cdot\Psi$
might be computationally critical. Therefore, an appropriate inversion
algorithm should be used in order to solve eq.~(\ref{eq:normsyst}). We
use the CGNR algorithm in our numerical tests, see
ref.~\cite[chapt.~8]{saad} for details.
The effects of imposing more and more boundary conditions will be
discussed in a specific example in sect.~5. We can say in advance
that the condition number of $\trans{\Psi}\cdot \Psi$ is not
sensitive to ${n}_{\scriptscriptstyle{\rm bc}}$ and that ${\cal P}_n$ rapidly converges as ${n}_{\scriptscriptstyle{\rm bc}}$ increases.
\vskip 0.2cm
\noindent $iv$) We need to discuss how to concretely work out and
compute the matrix elements $\psi_{\alpha\gamma}$. Here, the
assumption that $(A_k)_{k=1}^{Q-1}$ and $(B_{ik})_{i,k=1}^{Q-1}$ are polynomials becomes
practically decisive. Indeed, we observe that $\partial_k^m{\cal D}_\gamma$
is a polynomial on $\bar T_Q(s)$ with $\deg\{\partial_k^m{\cal D}_\gamma\}
= |\slashed{\gamma}|-m$ for $m\le |\slashed{\gamma}|$, while
$\phi_k^m{\cal D}_\gamma$ is a polynomial on $\bar T_Q(s)$ with
$\deg\{\phi_k^m{\cal D}_\gamma\} = |\slashed{\gamma}|+m$. Since
$\{{\cal D}_\gamma\}$ is a polynomial basis, it must be possible to express
both $\partial_k^m{\cal D}_\gamma$ and $\phi_k^m{\cal D}_\gamma$ as linear
combinations of some $\{{\cal D}_{\gamma'}\}$. Now, since ${\cal L}_{\scriptscriptstyle{\rm FP}}$ is a second
order partial differential operator, we never need to differentiate
more than twice. Analogously, since ${\cal L}_{\scriptscriptstyle{\rm FP}}$ is usually derived from a
Master Equation resulting from a detailed balance, $(A_k)_{k=1}^{Q-1}$ and
$(B_{ik})_{i,k=1}^{Q-1}$ are usually not more than quadratic polynomials (this
statement is of course less universal -- as the reader may
understand~-- since transition rates depend on the specific model, but
is often true). Instead of writing a general formula to expand
$\phi_i^p\phi_k^q\partial^r_m\partial^s_n{\cal D}_\gamma(\bar\phi)$ as a
linear combination of Dirichlet distributions, we prefer to report
formulae for specific choices of indices and exponents. To this
aim, we need to introduce some additional notation. We define
\begin{equation}
\gamma_{\ell^\pm} \equiv (\gamma_1,\ldots,\gamma_{\ell-1},\gamma_\ell\pm 1,\gamma_{\ell+1},\ldots,\gamma_Q)\,.
\end{equation}
Similarly, we define $\gamma_{\ell^+m^+}$, $\gamma_{\ell^+m^-}$,
$\gamma_{\ell^{++}}$, {\it etc}. as results of the iterated application of
index--raising operators $\oplus_\ell\cdot \gamma \equiv \gamma_{\ell^+}$
and index--lowering operators $\ominus_\ell \cdot \gamma \equiv
\gamma_{\ell^-}$, somewhat similar to the creation and destruction
operators of the quantum harmonic oscillator. Based on this, reference
formulae read
\vskip 0.3cm
\hrule
\vskip 0.2cm
\begin{equation}
\phi_\ell{\cal D}_\gamma(\bar\phi) = s\frac{\gamma_\ell}{|\gamma|}{\cal D}_{\gamma_{\ell^+}}(\bar\phi)\,,
\label{eq:firstalg}
\end{equation}
\begin{equation}
\partial_\ell{\cal D}_\gamma(\bar\phi) = s^{-1}(|\gamma|-1)[\theta_{\gamma_\ell,2}{\cal D}_{\gamma_{\ell^-}}(\bar\phi) - \theta_{\gamma_Q,2}{\cal D}_{\gamma_{Q^-}}(\bar\phi)]\,,
\label{eq:delDgamma}
\end{equation}
\begin{equation}
\phi_\ell\partial_\ell {\cal D}_\gamma(\bar\phi) = \theta_{\gamma_\ell,2}(\gamma_\ell-1){\cal D}_\gamma(\bar\phi) - \theta_{\gamma_Q,2}\gamma_\ell {\cal D}_{\gamma_{\ell^+Q^-}}(\bar\phi)\,,
\label{eq:phidelDgamma}
\end{equation}
\begin{equation}
\phi^2_\ell\partial_\ell {\cal D}_\gamma(\bar\phi) =\frac{s}{|\gamma|}\left\{\theta_{\gamma_\ell,2}\gamma_\ell(\gamma_\ell-1){\cal D}_{\gamma_{\ell^+}}(\bar\phi) - \theta_{\gamma_Q,2}\gamma_\ell(\gamma_\ell+1) {\cal D}_{\gamma_{\ell^{++}Q^-}}(\bar\phi)\right\}\,,
\end{equation}
\begin{equation}
\phi_\ell\partial_m {\cal D}_\gamma(\bar\phi)|_{\ell\ne m} = \gamma_\ell\left[\theta_{\gamma_m,2}{\cal D}_{\gamma_{\ell^+m^-}}(\bar\phi) - \theta_{\gamma_Q,2} {\cal D}_{\gamma_{\ell^+Q^-}}(\bar\phi)\right]\,,
\end{equation}
\vskip -0.2cm
\begin{align}
\partial_\ell^2{\cal D}_\gamma(\bar\phi) & = s^{-2}(|\gamma|-1)(|\gamma|-2)\cdot \left\{\theta_{\gamma_\ell,3}{\cal D}_{\gamma_{\ell^{--}}}(\bar\phi) \right.\nonumber\\[1.0ex]
& \left. - 2\theta_{\gamma_\ell,2}\theta_{\gamma_Q,2}{\cal D}_{\gamma_{\ell^-Q^-}}(\bar\phi) + \theta_{\gamma_Q,3}{\cal D}_{\gamma_{Q^{--}}}(\bar\phi)\right\}\,,
\end{align}
\vskip -0.2cm
\begin{align}
\phi_\ell\partial_\ell^2{\cal D}_\gamma(\phi) & = s^{-1}(|\gamma|-1)\cdot \left\{(\gamma_\ell-2)\theta_{\gamma_\ell,3}{\cal D}_{\gamma_{\ell^-}}(\bar\phi)\right.\nonumber\\[1.0ex]
& \hskip -0.5cm \left.-2(\gamma_\ell-1)\theta_{\gamma_\ell,2}\theta_{\gamma_Q,2}{\cal D}_{\gamma_{Q^-}}(\bar\phi) +\gamma_\ell\theta_{\gamma_Q,3}{\cal D}_{\gamma_{\ell^+Q^{--}}}(\bar\phi)\right\}\,,
\end{align}
\vskip -0.2cm
\begin{align}
\phi_m\partial_\ell^2{\cal D}_\gamma(\bar\phi)|_{\ell\ne m} & = s^{-1}\gamma_m(|\gamma|-1)\cdot \left\{\theta_{\gamma_\ell,3}{\cal D}_{\gamma_{m^+\ell^{--}}}(\bar\phi)\right.\nonumber\\[1.0ex]
& \hskip -1.38cm \left.-2\theta_{\gamma_\ell,2}\theta_{\gamma_Q,2}{\cal D}_{\gamma_{m^+\ell^-Q^-}}(\bar\phi) +\theta_{\gamma_Q,3}{\cal D}_{\gamma_{m^+Q^{--}}}(\bar\phi)\right\}\,,
\end{align}
\vskip -0.2cm
\begin{align}
\phi_\ell^2\partial_\ell^2{\cal D}_\gamma(\bar\phi) & = (\gamma_\ell-2)(\gamma_\ell-1)\theta_{\gamma_\ell,3}{\cal D}_{\gamma}(\bar\phi) - 2(\gamma_\ell-1)\gamma_\ell\theta_{\gamma_\ell,2}\theta_{\gamma_Q,2}{\cal D}_{\gamma_{\ell^+Q^-}}(\bar\phi) \nonumber\\[1.0ex]
& \hskip -0.4cm +\gamma_\ell(\gamma_\ell+1)\theta_{\gamma_Q,3}{\cal D}_{\gamma_{\ell^{++}Q^{--}}}(\bar\phi)\,,
\end{align}
\vskip -0.2cm
\begin{align}
\phi_m^2\partial_\ell^2{\cal D}_\gamma(\bar\phi)|_{\ell\ne m} & = \gamma_m(\gamma_m+1)\left\{\theta_{\gamma_\ell,3}{\cal D}_{\gamma_{m^{++}\ell^{--}}}(\bar\phi)\right. \nonumber\\[1.0ex]
& \hskip -1.26cm \left. - 2\theta_{\gamma_\ell,2}\theta_{\gamma_Q,2}{\cal D}_{\gamma_{m^{++}\ell^-Q^-}}(\bar\phi)
+\theta_{\gamma_Q,3}{\cal D}_{\gamma_{m^{++}Q^{--}}}(\bar\phi)\right\}\,,
\end{align}
\vskip -0.2cm
\begin{align}
\phi_\ell\phi_m\partial_\ell^2{\cal D}_\gamma(\bar\phi)|_{\ell\ne m} & =
\gamma_m\left\{(\gamma_\ell-2)\theta_{\gamma_\ell,3}{\cal D}_{\gamma_{m^{+}\ell^{-}}}(\bar\phi)\right. \nonumber\\[1.0ex]
& \hskip -1.76cm \left. -
2(\gamma_\ell-1)\theta_{\gamma_\ell,2}\theta_{\gamma_Q,2}{\cal D}_{\gamma_{m^{+}Q^-}}(\bar\phi)
+\gamma_\ell\theta_{\gamma_Q,3}{\cal D}_{\gamma_{m^{+}\ell^+Q^{--}}}(\bar\phi)\right\}\,,
\end{align}
\vskip -0.2cm
\begin{align}
\partial_\ell\partial_m{\cal D}_\gamma(\bar\phi)|_{\ell\ne m} & = s^{-2}(|\gamma|-1)(|\gamma|-2)\nonumber\\[1.0ex]
& \hskip -1.15cm \cdot
\left[\theta_{\gamma_\ell,2}\theta_{\gamma_m,2}{\cal D}_{\gamma_{(\ell^-)(m^-)}}(\bar\phi)
-\theta_{\gamma_\ell,2}\theta_{\gamma_Q,2}{\cal D}_{\gamma_{(\ell^-)(Q^-)}}(\bar\phi)\right.\nonumber\\[1.0ex]
& \left. \hskip -1.11cm -\theta_{\gamma_m,2}\theta_{\gamma_Q,2}
{\cal D}_{\gamma_{(m^-)(Q^-)}}(\bar\phi)
+ \theta_{\gamma_Q,3}{\cal D}_{\gamma_{Q^{--}}}(\bar\phi)\right]\,,
\end{align}
\vskip -0.2cm
\begin{align}
\phi_\ell\partial_\ell\partial_m{\cal D}_\gamma(\bar\phi)|_{\ell\ne m} & = s^{-1}(|\gamma|-1)\nonumber\\[1.0ex]
& \hskip -1.55cm \cdot \left\{(\gamma_\ell-1)
\left[\theta_{\gamma_\ell,2}\theta_{\gamma_m,2}{\cal D}_{\gamma_{m^-}}(\bar\phi)-
\theta_{\gamma_\ell,2}\theta_{\gamma_Q,2}{\cal D}_{\gamma_{Q^-}}(\bar\phi)\right]\right. \nonumber\\[1.0ex]
&
\hskip -1.55cm
\left.-\gamma_\ell\left[\theta_{\gamma_m,2}\theta_{\gamma_Q,2}{\cal D}_{\gamma_{\ell^+m^-Q^-}}(\bar\phi)
-
\theta_{\gamma_Q,3}{\cal D}_{\gamma_{\ell^+Q^{--}}}(\bar\phi)\right]\right\}\,,
\end{align}
\vskip -0.2cm
\begin{align}
\phi_\ell\phi_m\partial_\ell\partial_m{\cal D}_\gamma(\bar\phi)|_{\ell\ne m} & =\nonumber\\[1.0ex]
& \hskip -2.0cm
(\gamma_\ell-1)(\gamma_m-1)\theta_{\gamma_\ell,2}\theta_{\gamma_m,2}{\cal D}_\gamma(\bar\phi)
-
\gamma_m(\gamma_\ell-1)\theta_{\gamma_\ell,2}\theta_{\gamma_Q,2}{\cal D}_{\gamma_{m^+Q^-}}(\bar\phi)\nonumber\\[1.0ex]
& \hskip -2.0cm -
\gamma_\ell(\gamma_m-1)\theta_{\gamma_m,2}\theta_{\gamma_Q,2}{\cal D}_{\gamma_{\ell^+Q^-}}(\bar\phi)
+
\gamma_\ell\gamma_m\theta_{\gamma_Q,3}{\cal D}_{\gamma_{\ell^+m^+Q^{--}}}(\bar\phi)\,,
\label{eq:lastalg}
\end{align}
\vskip 0.2cm
\hrule
\vskip 0.3cm
\noindent where
\begin{equation}
\theta_{a,b} = \left\{\begin{array}{ll}1 & \quad \text{if}\quad a \ge b\,,\\[2.0ex]
0 & \quad \text{otherwise}\,.\end{array}\right.
\end{equation}
Now, projecting -- by way of example -- eq.~(\ref{eq:delDgamma}) onto $V_\alpha$ yields
\begin{equation}
\langle V_\alpha, \partial_\ell {\cal D}_\gamma\rangle_\kappa =
s^{-1}(|\gamma|-1)[\theta_{\gamma_\ell,2}\,\chi_{\alpha\gamma_{\ell^-}}
- \theta_{\gamma_Q,2}\,\chi_{\alpha\gamma_{Q^-}}]\,.
\end{equation}
Analogously it be can done for all
eqs.~(\ref{eq:firstalg})--({\ref{eq:lastalg}); we see indeed that
projecting the whole function ${\cal L}_{\scriptscriptstyle{\rm FP}}\cdot{\cal P}_n$ onto $V_\alpha$ is
just a matter of tedious yet simple algebra. We conclude that
$\psi_{\alpha\gamma}$ can be expanded as a self--contained sum of
contributions, each being proportional to some matrix element
of~$\chi$. However, we observe that
\begin{equation}
\left[\oplus_1^{m_1}\ldots\oplus_{Q-1}^{m_{Q-1}}\right]\cdot
\left[\ominus_1^{\ell_1}\ldots\ominus_{Q-1}^{\ell_{Q-1}}\right]\cdot\gamma\,\in\Omega_{n+\sum_km_k-\sum_k\ell_k}\quad
\text{ for } \gamma\in\Omega_n\, \text{ and }\ \{\ell_k<\gamma_k\}\,.
\end{equation}
If $\max_k\deg\{A_k\} = m_A$, then the action of the drift term on
${\cal P}_n$ mixes Dirichlet distributions with index arrays in the \emph{bucket spaces}
$\Omega_{n-1},\Omega_n,\ldots,\Omega_{n+m_A-1}$. Likewise, if
$\max_{i,k}\deg\{B_{ik}\}=m_B$, then the action of the diffusion term
on ${\cal P}_n$ mixes Dirichlet distributions with index arrays in the \emph{bucket spaces}
$\Omega_{n-2},\Omega_{n-1},\ldots,\Omega_{n+m_B-2}$. Accordingly,
in order to compute the matrix $\psi$, we need to compute
$\chi_{\alpha\gamma}$ for $|\alpha|\le n$ and for $\gamma\in\Omega_k$ for some
$k\in\{n-2,\ldots,n_\text{max}\}$, with $n_\text{max} = \max\{n+m_A-1,n+m_B-2\}$.
\vskip 0.2cm
\noindent $v$) In order to work out
eqs.~(\ref{eq:bcorth})--(\ref{eq:bcskew}), we first insert
eq.~(\ref{eq:polexp}) into $J_k$ and extract the coefficient
multiplying each $c_\gamma$, namely
\begin{equation}
J_k(\bar\phi) =
\sum_{\gamma\in\Omega_n}c_\gamma\left\{A_k(\bar\phi){\cal D}_\gamma(\bar\phi)
-
\frac{1}{2}\sum_{i=1}^{Q-1}\partial_i[B_{ik}(\bar\phi){\cal D}_\gamma(\bar\phi)]\right\}
\equiv
\sum_{\gamma\in\Omega_n}\Upsilon_{k\gamma}(\bar\phi)c_\gamma\,.
\end{equation}
We need to compute each matrix coefficient
$\Upsilon_{k\gamma}(\bar\phi)$ just for two sets of boundary points, namely
$\Upsilon_{k\gamma}(\bar\phi_{k,m})$ (in order to impose the boundary
conditions on ${\mathfrak H}_k$) and $\Upsilon_{k\gamma}(\bar\phi_{Q,m})$ (in
order to impose the boundary conditions on ${\mathfrak H}_Q$). The reader should
notice that ${\cal D}_\gamma(\bar\phi_{km})=0$ unless $\gamma_k=1$ as well
as ${\cal D}_\gamma(\bar\phi_{Qm})=0$ unless $\gamma_Q=1$. Since
$\Upsilon_{k\gamma}$ depends on both ${\cal D}_\gamma$ and
$\{\partial_i{\cal D}_\gamma\}$, this means that
$\Upsilon_{k\gamma}(\bar\phi_{km})=0$ unless $\gamma_k=1,2$ and
equally $\Upsilon_{k\gamma}(\bar\phi_{Qm})=0$ unless
$\gamma_Q=1,2$. Therefore, we conclude that the only unknowns taking
part in the boundary equations are those $c_\gamma$ which have at
least one component $\gamma_k=1,2$ with $k=1,\ldots,Q$.
\vskip 0.6cm
\begin{myremark}
By now, it should be sufficiently clear what the pros and cons of
projecting ${\cal P}$ onto a set of Dirichlet distributions are. We find
it worthwhile summarizing them:
\begin{itemize}
\item{the Dirichlet distributions $\{{\cal D}_\gamma\}_{\gamma\in\Omega_n}$
are not orthogonal polynomials with respect to the scalar product
$\langle\cdot,\cdot\rangle_\kappa$, yet they form a basis of $\bar
{\cal H}_n$;}
\item{while the zeros of the orthogonal polynomials $\{V_\alpha\}$ and
$\{U_\alpha\}$ are non--trivial algebraic varieties, the Dirichlet
distributions $\{{\cal D}_\gamma\}_{\gamma\in\Omega_n}$ are
non--negative on $\bar T_Q(s)$. This means that the positiveness of
${\cal P}_n$ relies entirely on the signs of the expansion coefficients
$\{c_\gamma\}$. If $c_\gamma\ge 0$ $\ \forall\gamma\in\Omega_n$,
then ${\cal P}_n$ can be statistically interpreted as a distributional
mixture;}
\item{if for too small values of $n$ the RG approximation gives
${\cal P}_n(\bar\phi)<0$ for $\bar\phi$ in some positive--measure subset
of $\bar T_Q(s)$, it is anyway possible to obtain a decent
(non--quasi--best) approximation of ${\cal P}$ by changing the sign of
some coefficient $c_\gamma$ and by subsequently renormalizing the
whole vector $c$;}
\item{the differentiation rules of ${\cal D}_\gamma$ generate
self--contained algebraic expressions involving Dirichlet
distributions with different indices. Although $\psi$ is a dense
matrix, it can be easily computed. Notice, however, that not only
$|\Omega_n|$ inflates almost exponentially with $Q$, but also the
CPU time needed to compute $\chi_{\alpha\gamma}$ for a given pair
$(\alpha,\gamma)$ blows up, since eq.~(\ref{eq:scalprod}) contains a
non--factorizable multiple sum $\sum_{\beta\le\alpha} =
\sum_{\beta_1=0}^{\alpha_1}\ldots
\sum_{\beta_{Q-1}=0}^{\alpha_{Q-1}}$\,;}
\item{the Dirichlet distribution ${\cal D}_\gamma$ vanishes on ${\mathfrak H}_k$
unless $\gamma_k=1$. It is therefore very simple to keep track of
which terms are responsible for the behaviour of ${\cal P}_n$ on
$\partial \bar T_Q(s)$. Such a task would be a nightmare with any
other polynomial basis. \qed}
\end{itemize}
\vskip 0.2cm
\end{myremark}
\section{Example~1: binary voter model with zealots}
The binary voter model, introduced in \cite{Clifford1973,Holley1975},
can be considered as an archetype of agent--based models for opinion
dynamics. Owing popularity to its exact solvability on a lattice in
any dimension, the model has been studied in a number of variants. We
refer the reader to~\cite{FortCastLor} for a comprehensive review of
the relevant literature. The microscopic dynamics of the model is
simply defined. Agents carry a binary variable $v\in\{+1,-1\}$ and are
selected at random for transitions. When an agent is selected, she
flips her variable to that of a neighbour agent, also chosen at
random. In a certain time the system collapses to a consensus state
(all agents eventually share the same opinion), unless a stabilization
mechanism is turned on. One possibility is to perturb the system by
introducing {\it zealots} among the agents, {\it i.e.\ } special individuals
who never change their opinion. Zealots in the context of the binary
voter model have been originally proposed in \cite{Mobilia2003}. If
competing zealots with opposite opinions are present, consensus states
are prevented as discussed in \cite{Mobilia2007}. As far as we are
concerned here, the binary voter model with zealots is of interest
because
\begin{itemize}
\item{it is a one--dimensional model, {\it i.e.\ } $Q=2$;}\\[-2.0ex]
\item{the FPE of the model can be solved exactly.}
\end{itemize}
Both these features make it a simple case study to test the RG
method. Let $N$ denote the total number of agents, $Z_\pm$ the number
of zealots with $v=\pm 1$ and $N_\pm$ the number of dynamic agents
with $v=\pm 1$. Along with \cite{Mobilia2007}, we define $\phi=N_+/N$,
$z_\pm = Z_\pm/N$ and $s=1-z_+-z_-$. Accordingly, it must be
$0\le\phi\le s$, {\it i.e.\ } $\bar T_Q(s)$ is just an interval in this
case. The FPE of the model reads
\begin{equation}
0 = -{\partial_\phi}[A(\phi)P(\phi)] +
\frac{1}{2}{\partial^2_\phi}[B(\phi)P(\phi)] =
\partial_\phi\left\{-A(\phi)P(\phi) +
\frac{1}{2}{\partial_\phi}[B(\phi)P(\phi)]\right\} = \partial_\phi
J(\phi)\,,
\label{eq:bvmfpe}
\end{equation}
with $\partial_\phi=\partial/\partial\phi$. The drift and diffusion
coefficients are given by
\begin{align}
A(\phi) & = [z_+s - \phi(1-s)]\,,\\[1.0ex]
B(\phi) & = N^{-1}[(\phi+z_+)(s-\phi) + \phi(s+z_--\phi)]\,.
\end{align}
If one introduces the auxiliary variables $\delta=z_+-z_-$,
$r=\sqrt{\delta^2+4s}$ and $u_\pm = s/2-\delta/4\pm r/4$, then the
exact solution of the FPE \cite{Mobilia2007} reads
\begin{equation}
{\cal P}(\phi) =
W\cdot[(\phi-u_+)(\phi-u_-)]^{(Z_++Z_--2)/2}\left[1+\frac{r}{2\phi-s-\frac{r-\delta}{2}}\right]^{(\delta/2r)(2N-Z_+-Z-)}\,,
\label{vmsol}
\end{equation}
with $W$ being a normalization constant such that
$\int_0^s\text{d}\phi\,{\cal P}(\phi)=1$. When $z_\pm=z$, the solution collapses
to
\begin{equation}
{\cal P}_\text{sym}(\phi) = W\cdot[zs+2\phi(s-\phi)]^{Nz-1}\,.
\end{equation}
Notice that ${\cal P}_\text{sym}$ is a polynomial with
$\deg\{{\cal P}_\text{sym}\}=2Nz-2$, while ${\cal P}$ is a rational function for
$z_+\ne z_-$. It should be observed that the condition $\hat
n(\bar\phi)\cdot J(\bar\phi)=0$ is meaningless for $Q=2$ since $\hat
n$ is not defined at all. Indeed, $Q=2$ is a degenerate case:
$J(\phi)=\text{const.}$ is a first integral of eq.~(\ref{eq:bvmfpe})
and eq.~(\ref{eq:bc}) is simply fulfilled provided we choose the
constant to be zero. Now, a Dirichlet distribution with $Q=2$ is
actually a beta distribution
\begin{equation}
{\cal D}_{(\gamma_1,\gamma_2)}(\phi) =
\frac{\Gamma(\gamma_1+\gamma_2)}{\Gamma(\gamma_1)\Gamma(\gamma_2)}s^{1-\gamma_1-\gamma_2}\phi^{\gamma_1-1}(s-\phi)^{\gamma_2-1}
= B_{\gamma_1\gamma_2}(\phi)\,,
\end{equation}
and the \emph{bucket space} amounts in this case to
\begin{equation}
\Omega_n =
\biggl\{(n+1,1),\,(n,2),\ldots,(2,n),\,(1,n+1)\biggr\}\,,
\qquad |\Omega_n| = {n+1\choose n} = n+1\,.
\end{equation}
\begin{figure}[t!]
\centering
\includegraphics[width=0.95\textwidth]{bvm_16_32_k1=2_k2=2.pdf}
\vskip 0.9cm
\includegraphics[width=0.95\textwidth]{bvm_16_16_k1=2_k2g2.pdf}
\vskip 0.9cm
\includegraphics[width=0.95\textwidth]{bvm_asym_16_16_32_k1=2_k2=2.pdf}
\vskip -0.0cm
\caption{ \footnotesize RG approximations of the probability density
of the binary voter model for various parameter sets.}
\label{fig:bvmone}
\end{figure}
\begin{figure}[t!]
\centering
\hskip -0.5cm
\includegraphics[width=0.95\textwidth]{bvm_spct_16_32_k1=2_k2=2.pdf}
\vskip 0.1cm
\caption{ \footnotesize Dirichlet spectra of the RG approximations.}
\label{fig:bvmspectra}
\end{figure}
\vskip -0.4cm
By using the differentiation formulae reported in sect.~3 and some
scratch paper, we can easily work out the matrix coefficients
$\psi_{\alpha\gamma} = \langle V_\alpha, -A{\cal D}_\gamma +
2^{-1}\partial_\phi[B{\cal D}_\gamma]\rangle_\kappa$, namely
\begin{align}
\langle V_\alpha, -A{\cal D}_\gamma\rangle_\kappa & =
-z_+s\,\chi_{\alpha(\gamma_1,\gamma_2)}
+s(1-s)\frac{\gamma_1}{\gamma_1+\gamma_2}\,\chi_{\alpha(\gamma_1+1,\gamma_2)}\,, \end{align}
and
\begin{align}
\langle V_\alpha,\frac{1}{2}\partial_\phi[B{\cal D}_\gamma]\rangle_\kappa & = \frac{1}{N}\biggl\{
-2s\frac{\gamma_1}{\gamma_1+\gamma_2}\,\chi_{\alpha(\gamma_1+1,\gamma_2)}
+\frac{1}{2}(2s-z_++z_-)\,\chi_{\alpha(\gamma_1,\gamma_2)}\nonumber\\[3.0ex]
& \hskip -1.6cm
-\frac{s}{\gamma_1+\gamma_2}\left[\gamma_1(\gamma_1-1)\theta_{\gamma_1,2}\,\chi_{\alpha(\gamma_1,\gamma_2)}
-\gamma_1(\gamma_1+1)\theta_{\gamma_2,2}\,\chi_{\alpha(\gamma_1+1,\gamma_2-1)}\right]\nonumber
\end{align}
\begin{align}
& \hskip 1.0cm
+\frac{1}{2}(2s-z_++z_-)\left[(\gamma_1-1)\theta_{\gamma_1,2}\,\chi_{\alpha(\gamma_1,\gamma_2)}
-
\gamma_1\theta_{\gamma_2,2}\,\chi_{\alpha(\gamma_1+1,\gamma_2-1)}\right]\nonumber\\[3.0ex]
& \hskip 1.0cm
+\frac{z_+}{2}(\gamma_1+\gamma_2-1)\left[\theta_{\gamma_1,2}\chi_{\alpha(\gamma_1-1,\gamma_2)}
-\theta_{\gamma_2,2}\chi_{\alpha(\gamma_1,\gamma_2-1)}\right]\biggr\}\,,
\end{align}
with $\gamma\in\Omega_n$ and $\alpha=0,1,\ldots,n$. Having coded
$\psi$, we checked numerically that $\nu_0=1$ independently of
$n$. The RG problem is therefore well posed: it is sufficient to
impose the normalization condition eq.~(\ref{eq:polnorm}) to guarantee
that $\Psi$ has maximal rank.
Numerical results are illustrated in Fig.~\ref{fig:bvmone} for both
symmetric and asymmetric cases: the two plots on top show the exact
solution ${\cal P}_\text{sym}$ and its RG approximations for $N=1000$,
$Z_\pm=$ $Z\in\{16,32\}$, $\kappa=(2,2)$ and a bunch of values of
$n$; the central plots show results for the same physical setup, yet
with asymmetric choices of the weight index array $\kappa$; finally,
those at the bottom show the exact solution ${\cal P}$ and its RG
approximations for $N=1000$, $Z_+=4$, $Z_-\in\{16,32\}$ and
$\kappa=(2,2)$. In Fig.~\ref{fig:bvmspectra}, we report plots of the
Dirichlet spectra obtained in the symmetric case with
$\kappa=(2,2)$. A few comments are in order:
\begin{itemize}
\item{positiveness is violated at small values of $n$ for essentially
all physical setups, but soon recovered at larger $n$;}
\item{distributional convergence is reached at $n=2Nz-2$ in the
symmetric case. It could not be otherwise: in order to exactly
represent a polynomial $P$ with $\deg\{P\}=n$ by another polynomial
$Q$, it must be $\deg\{Q\}\ge n$;}
\item{convergence deteriorates in the symmetric case for $\kappa_1\ne
\kappa_2$. Broadly speaking, the measure weight ${\cal D}_\kappa$
overlaps with both $V_\alpha$ and ${\cal L}_{\scriptscriptstyle{\rm FP}}\cdot {\cal D}_\gamma$ in
$\psi_{\alpha\gamma}$. If the mass of ${\cal P}$ concentrates in a given
subset of $\bar T_Q(s)$, it is recommendable not to use a weight
function whose mass concentrates elsewhere. Although this suggestion
is only useful once ${\cal P}$ is known, symmetries should be taken into
account in order to properly choose $\kappa$;}
\item{the RG method works well also in the asymmetric case, where
${\cal P}$ is a rational function. Here, exact convergence is expected to
be reached only asymptotically;}
\item{the Dirichlet spectra look rather localized. Instead of
ordering $c_{(\gamma_1,\gamma_2)}$ according to the RLO, in
Fig.~\ref{fig:bvmspectra} we plot data against the relative 1--norm
distance of $(\gamma_1,\gamma_2)$ from}
\begin{equation}
(\bar\gamma_1,\bar\gamma_2)=
\underset{(\gamma_1,\gamma_2)\in\Omega_n}{\operatorname{argmax}}
\left\{c_{(\gamma_1,\gamma_2)}\right\} =
\left(\frac{n}{2}+1,\frac{n}{2}+1\right)\,,\qquad n\text{
even}\,.
\end{equation}
\item[]{The outcome is evidently an exponential decrease for
$n=2Nz-2$, with an increasingly marked bending at larger values of
$n$. An exponential behaviour is not surprising in consideration
that ${\cal P}_\text{sym}$ is essentially a centered Gaussian
distribution;}
\item{the bending at $n>2Nz-2$ is clearly due to the
non--orthogonality of $\{{\cal D}_\gamma\}$.}
\end{itemize}
\section{Example~2: multi--state voter model with zealots}
As a second case study for the RG method, we examine the multi--state
voter model with zealots, a generalization of the binary version
considered so far, where both dynamic agents and zealots carry an
opinion $v\in\{1,\ldots,Q\}$. The ordering dynamics of the model with
no zealots has been discussed in~\cite{Starnini}, while a variant with
committed agents on a weighted network has been more recently studied
in~\cite{Chen}. Here, we are interested in a simple formulation of the
model with $N$ agents on the complete graph, for which we expect the
mean field description to work well. Let $N_k$ and $Z_k$
denote respectively the number of dynamic agents and zealots with
$v=k$. We define $\phi_k = N_k/N$, $z_k=Z_k/N$ and
$s=1-\sum_{k=1}^{Q-1}z_k$. The FPE reads
\begin{equation}
0 = -\sum_{\ell=1}^{Q-1}\partial_\ell [A_\ell(\phi)P(\phi)] +
\frac{1}{2}\sum_{\ell,m=1}^{Q-1}\partial_\ell\partial_m[B_{\ell
m}(\phi)P(\phi)]
\end{equation}
with drift and diffusion coefficients given by
\begin{align}
A_\ell(\bar\phi) & = z_\ell s - (1-s)\phi_\ell\,,\\[2.0ex]
B_{\ell m}(\bar\phi) & = \frac{\delta_{\ell
m}}{N}[(\phi_\ell+z_\ell)(s-\phi_\ell) + \phi_\ell
(1-z_\ell-\phi_\ell)] - \frac{1-\delta_{\ell m}}{N}[2\phi_\ell\phi_m
+ z_\ell\phi_m + z_m\phi_\ell]\,.
\end{align}
To the best of our knowledge, no analytic solution of the FPE is
known in the literature. Therefore, the results of the RG method can
be only compared to numerical simulations. Similar to the previous
section, the derivation of $\psi_{\alpha\gamma}$ requires a modest
algebraic effort. We have indeed
\begin{align}
& \langle
V_\alpha,-\sum_{\ell=1}^{Q-1}\partial_\ell[A_\ell{\cal D}_\gamma]\rangle_\kappa
= (Q-1)(1-s)\langle V_\alpha,{\cal D}_\gamma\rangle_\kappa +
(1-s)\sum_{\ell=1}^{Q-1}\langle
V_\alpha,\phi_\ell\partial_\ell{\cal D}_\gamma\rangle_\kappa
\nonumber\\[1.0ex]
& \hskip 1.0cm -s\sum_{\ell=1}^{Q-1}z_\ell\langle
V_\alpha,\partial_\ell {\cal D}_\gamma\rangle_\kappa\,,
\end{align}
\begin{align}
& \langle
V_\alpha,\frac{1}{2}\sum_{m,\ell=1}^{Q-1}\partial_\ell\partial_m[B_{\ell
m}{\cal D}_\gamma]\rangle_\kappa = -\frac{Q(Q-1)}{N}\langle
V_\alpha,{\cal D}_\gamma\rangle_\kappa -
\frac{2Q}{N}\sum_{\ell=1}^{Q-1}\langle
V_\alpha,\phi_\ell\partial_\ell {\cal D}_\gamma\rangle_\kappa
\nonumber\\[1.0ex]
& \hskip 1.0cm +\frac{1}{N}\sum_{\ell=1}^{Q-1}(1+s-Qz_\ell)\langle
V_\alpha,\partial_\ell{\cal D}_\gamma \rangle_\kappa -
\frac{1}{N}\sum_{\ell=1}^{Q-1}\langle
V_\alpha,\phi_\ell^2\partial_\ell^2{\cal D}_\gamma \rangle_\kappa
\nonumber
\end{align}
\begin{align}
& \hskip 1.0cm + \frac{1}{2N}\sum_{\ell=1}^{Q-1}(1+s-2z_\ell)\langle
V_\alpha,\phi_\ell\partial_\ell^2{\cal D}_\gamma \rangle_\kappa +
\frac{s}{2N}\sum_{\ell=1}^{Q-1}z_\ell\langle V_\alpha,
\partial_\ell^2{\cal D}_\gamma \rangle_\kappa\nonumber\\[1.0ex]
& \hskip 1.0cm -\frac{1}{2N}\sum_{\ell\ne m}^{1\ldots Q}\bigl[2\langle
V_\alpha,\phi_\ell\phi_m\partial_\ell\partial_m{\cal D}_\gamma
\rangle_\kappa + z_\ell\langle
V_\alpha,\phi_m\partial_\ell\partial_m{\cal D}_\gamma\rangle_\kappa +
z_m\langle
V_\alpha,\phi_\ell\partial_\ell\partial_m{\cal D}_\gamma\rangle_\kappa
\bigr]\,,
\end{align}
and we simply need to express the various scalar products in terms of
the matrix elements of $\chi$ via
eqs.~(\ref{eq:firstalg})--(\ref{eq:lastalg}). To give a feeling of the
goodness of the approximation, in Fig.~\ref{fig:msvm} we qualitatively
compare the histogram of the probability density obtained
from Monte Carlo simulations of the model (top left) and the RG
approximation (top right) for a physical setup with $Q=3$, $N=1000$,
$Z_1=Z_2=Z_3=4$ and RG parameters $n=12$, ${n}_{\scriptscriptstyle{\rm bc}}=20$ and
$\kappa=(2,2,2)$.
It is interesting to examine how much ${\cal P}_n$ depends upon the number
${n}_{\scriptscriptstyle{\rm bc}}$ of boundary conditions. 2--norm distances can be easily
evaluated once the RG coefficients are known. If ${\cal P}_n^{(1)} =
\sum_{\gamma\in\Omega_n}c^{(1)}_\gamma{\cal D}_\gamma$ and ${\cal P}_n^{(2)} =
\sum_{\gamma\in\Omega_n}c^{(2)}_\gamma{\cal D}_\gamma$, then it can be
shown that
\begin{equation}
||{\cal P}_n^{(1)} - {\cal P}_n^{(2)}||^2_2 = s^{1-Q}
\sum_{\gamma,\eta\in\Omega_n}\left[c^{(1)}_\gamma-c^{(2)}_\gamma\right]\left[c^{(1)}_\eta-c^{(2)}_\eta\right]
\frac{\Gamma(|\gamma|)\Gamma(|\eta|)}{\Gamma(|\gamma|+|\eta|-Q)}\prod_{k=1}^{Q}
\frac{\Gamma(\gamma_k+\eta_k-1)}{\Gamma(\gamma_k)\Gamma(\eta_k)}\,.
\end{equation}
\begin{figure}[t!]
\centering
\begin{minipage}[t]{0.45\textwidth}
\includegraphics[width=0.95\textwidth]{hist.pdf}
\end{minipage}
\begin{minipage}[t]{0.45\textwidth}
\includegraphics[width=0.95\textwidth]{ortho.pdf}
\end{minipage}
\vskip 0.5cm
\begin{minipage}[t]{0.45\textwidth}
\hskip 0.5cm
\includegraphics[width=0.95\textwidth]{dist.pdf}
\end{minipage}
\begin{minipage}[t]{0.45\textwidth}
{
\small
\hskip 2.0cm
\begin{tabular}{r|c}
\\[-6.4cm]
\hline\hline\\[-2.0ex]
$n$ & $\log_{10}\text{cond}\{\trans{\Psi}\cdot\Psi\}$ \\[0.2ex]
\hline\\[-2.3ex]
6 & 10.67(1) \\
8 & 14.85(1) \\
10 & 19.17(1) \\
12 & 23.69(1) \\
14 & 28.24(1) \\
16 & 32.73(1) \\
18 & 37.16(1) \\[0.6ex]
\hline\hline
\end{tabular}
}
\end{minipage}
\caption{ \footnotesize (top left) Probability density of the
multi--state voter model from Monte Carlo simulations (axes have
been rescaled so as to host a 30--bins histogram with unitary bin
size). (top right) RG probability density with parameters $n=12$,
${n}_{\scriptscriptstyle{\rm bc}}=20$ and
$\kappa=(2,2,2)$. (bottom left) Sensitivity of ${\cal P}_n$ to the
number ${n}_{\scriptscriptstyle{\rm bc}}$ of boundary conditions, see
eq.~(\ref{eq:Xn}). (bottom right) condition number of
$\trans{\Psi}\cdot\Psi$. All plots refer to a physical setup with
$Q=3$, $N=1000$, $Z_1=Z_2=Z_3=4$.}
\label{fig:msvm}
\end{figure}
\noindent In order to assess the sensitivity of ${\cal P}_n$ to ${n}_{\scriptscriptstyle{\rm bc}}$, we
could look at $||{\cal P}_n|_{{n}_{\scriptscriptstyle{\rm bc}}+2}-{\cal P}_n|_{{n}_{\scriptscriptstyle{\rm bc}}}||_2^2$ as a function of
${n}_{\scriptscriptstyle{\rm bc}}$. Unfortunately, this quantity depends strongly on the polynomial
degree $n$, hence it becomes difficult to compare distances
corresponding to different values of $n$. A smoother behaviour is
displayed by the distance ratio
\begin{equation}
X_n({n}_{\scriptscriptstyle{\rm bc}}) =
\sqrt{\frac{||{\cal P}_n|_{{n}_{\scriptscriptstyle{\rm bc}}+2}-{\cal P}_n|_{{n}_{\scriptscriptstyle{\rm bc}}}||_2^2}{||{\cal P}_n|_{{n}_{\scriptscriptstyle{\rm bc}}}-{\cal P}_n|_{{n}_{\scriptscriptstyle{\rm bc}}-2}||_2^2}}\qquad
\left(\text{notice that} \lim_{{n}_{\scriptscriptstyle{\rm bc}}\to\infty}X_n({n}_{\scriptscriptstyle{\rm bc}}) = 1 \right)\,,
\label{eq:Xn}
\end{equation}
which we plot against ${n}_{\scriptscriptstyle{\rm bc}}$ in Fig.~5 (bottom left), once more for
$Q=3$, $N=1000$, $Z_1=Z_2=Z_3=4$ and $\kappa=(2,2,2)$. For $n>12$,
limits of our computer implementation emerge: $X_n({n}_{\scriptscriptstyle{\rm bc}})$ becomes
numerically unstable due to large cancellations occurring when
subtracting the coefficients $c_\gamma$, hence we give up reporting
it. Anyway, the plot shows that solutions of higher degree are more
sensitive to the number of boundary conditions. This looks natural if
one considers that the larger $n$ the more ${\cal P}_n$ fluctuates on the
boundary hypersurfaces: in order to {\it gentle} the orthogonal
probability flux crossing the boundary, this must be forced to
vanish at more and more boundary points. By construction
$X_n$ carries no information about the overall scale of the 2--norm
distances. This turns out to be very small for all $n$ and
${n}_{\scriptscriptstyle{\rm bc}}\gtrsim 10$ (this estimate is likely to increase for larger values of $Q$).
It is likewise interesting to look at the condition number of
$\trans{\Psi}\cdot\Psi$. As the table in Fig.~5 (bottom right)
shows, this blows up exponentially as $n$
increases, while it is rather insensitive to ${n}_{\scriptscriptstyle{\rm bc}}$ (the uncertainty reported
in the table measures the variation range for $20\le{n}_{\scriptscriptstyle{\rm bc}}\le
200$). The exponential enhancement with $n$ requires a robust
algorithm in order to perform the matrix inversion, as already
observed in sect.~3.
Finally, the above discussion concerns only the analytic properties of
${\cal P}_n$. In order to make a quantitative comparison between ${\cal P}_n$
and the empirical probability density obtained from the Monte Carlo
(MC) simulation of the multi--state voter model, we can look at the respective
distributional moments. Those of ${\cal P}_n$ can be easily worked--out and
exactly expressed as functions of the coefficients $\{c_\gamma\}$. In
particular, the first two moments are given by
\begin{align}
& \mathds{E}\left[\phi_k\, |\, {\cal P}_n\right] =
s\sum_{\gamma\in\Omega_n}c_\gamma\frac{\gamma_k}{|\gamma|}\,,\\[1.0ex]
& \mathds{E}\left[\phi_k^2\, |\, {\cal P}_n\right] =
s^2\sum_{\gamma\in\Omega_n}c_\gamma\frac{\gamma_k(\gamma_k+1)}{|\gamma|(|\gamma|+1)}\,,
\qquad \mathds{E}\left[\phi_j\phi_k\, |\, {\cal P}_n\right]_{j\ne k} =
s^2\sum_{\gamma\in\Omega_n}c_\gamma\frac{\gamma_j\gamma_k}{|\gamma|(|\gamma|+1)}\,.
\end{align}
Numerical estimates look rather stable against changes of $n$ and
${n}_{\scriptscriptstyle{\rm bc}}$. For physical parameters as above all RG approximations give
$\mathds{E}[\phi_k]_\text{RG} = 0.32933\ldots = s/3$. However, the first
moment is not indicative, as it just results from the symmetry of the
setup. In order to make a real comparison, we have to look at the second
moments. Our best estimates from RG approximations are
$\mathds{E}[\phi_k^2]_\text{RG} = 0.1253\ldots$ to be compared to
$\mathds{E}[\phi_k^2]_\text{MC} = 0.124(1)$ and for $j\ne k$, $\mathds{E}[\phi_j\phi_k]_\text{RG} =
0.1000\ldots$ to be compared to
$\mathds{E}[\phi_j\phi_k]_\text{MC}=0.1004(5)$. As can be seen, results are in
very good agreement.
\section{Symmetry considerations}
The symmetry group of the simplex is the symmetric group. Since the
implementation of the RG method becomes numerically demanding at
large $Q$, it is worthwhile discussing if and how permutational
symmetries can help reduce the computational work load.
\subsection{Permutational symmetry of the coefficients $\chi_{\alpha\gamma}$}
We first observe that the choice of the weight index array $\kappa$ is
totally arbitrary, yet different values of it correspond to different
orthogonal bases. A convenient option is the isotropic one, namely
\begin{equation}
{\kappa_{\scriptscriptstyle\text{iso}}} = (\underbrace{\hat \kappa,\ldots ,\hat \kappa}_{Q-1\text{ times}},\bar\kappa)\,,
\end{equation}
which for any $Q$ depends only on two integer values $\hat\kappa$ and
$\bar\kappa$. If we denote by $S_{Q-1}$ the set of permutations of
$\{1,\ldots,Q-1\}>$ and for
$\sigma\in S_{Q-1}$ we define $\sigma\cdot\bar\phi \equiv
(\phi_{\sigma(1)},\ldots,\phi_{\sigma(Q-1)})$, $\sigma\cdot\alpha =
(\alpha_{\sigma(1)},\ldots,\alpha_{\sigma(Q-1)})$ for $\alpha\in
\mathds{N}_0^{Q-1}$ and $\sigma\cdot\gamma
= (\gamma_{\sigma(1)},\ldots,\ldots,\gamma_{\sigma(Q-1)},\gamma_Q)$
for $\gamma\in\mathds{N}^Q$, then we immediately see that
${\cal D}_{\kappa_{\scriptscriptstyle\text{iso}}}(\sigma\cdot\bar\phi) = {\cal D}_{\kappa_{\scriptscriptstyle\text{iso}}}(\bar\phi)$. For any other
choice of the index array, it holds
\begin{align}
{\cal D}_\gamma(\sigma\cdot\bar\phi) & = \frac{\Gamma(|\gamma|)}{\prod_{m=1}^Q\Gamma(\gamma_m)}s^{1-|\gamma|}\left[\prod_{m=1}^{Q-1}\phi_{\sigma(m)}^{\slashed{\gamma}_m}\right](s-|\bar\phi|)^{\slashed{\gamma_Q}} \nonumber\\[2.0ex]
& = \frac{\Gamma(|\gamma|)}{\prod_{m=1}^Q\Gamma(\gamma_m)}s^{1-|\gamma|}\left[\prod_{m=1}^{Q-1}\phi_m^{\slashed{\gamma}_{\sigma^{-1}(m)}}\right](s-|\bar\phi|)^{\slashed{\gamma}_Q} = {\cal D}_{\sigma^{-1}\cdot\gamma}(\bar\phi)\,,
\label{eq:dircov}
\end{align}
with $\sigma^{-1}$ denoting the inverse permutation of
$\sigma$. Remarkably, a property analogous to eq.~(\ref{eq:dircov}) is
also fulfilled by the orthogonal polynomials provided $\kappa={\kappa_{\scriptscriptstyle\text{iso}}}$, namely
\begin{myprop}
If $\kappa= {\kappa_{\scriptscriptstyle\text{iso}}}$, then $V_\alpha(\sigma\cdot\bar\phi) = V_{\sigma^{-1}\cdot\alpha}(\bar\phi)$, $U_\alpha(\sigma\cdot\bar\phi) = U_{\sigma^{-1}\cdot\alpha}(\bar\phi)$ and $f_{\sigma\cdot\alpha} = f_\alpha$ for any $\alpha\in\mathds{N}_0^{Q-1}$, $\sigma\in S_{Q-1}$ and $\bar\phi\in \bar T_Q(s)$.
\end{myprop}
\noindent\emph{Proof}. With regard to $V_\alpha$, we first notice by
direct inspection that
\begin{equation}
v_{\alpha(\sigma\cdot\beta)}(\kappa) = v_{(\sigma^{-1}\cdot{\alpha})\beta}(\sigma^{-1}\cdot\kappa)\,.
\label{eq:vsymm}
\end{equation}
Therefore, we have
\begin{align}
V_\alpha\left(\sigma\cdot \bar\phi\right) & =
\sum_{\beta\le\alpha}v_{\alpha\beta}(\kappa)(\sigma\cdot \bar\phi)^\beta = \sum_{\beta\le\alpha}v_{\alpha\beta}(\kappa)\bar\phi^{\sigma^{-1}\cdot\beta} = \sum_{\sigma\cdot\beta\le\alpha}v_{\alpha(\sigma\cdot\beta)}(\kappa)\bar\phi^\beta \nonumber\\[2.0ex]
& = \sum_{\sigma\cdot \beta\le\alpha}v_{\sigma^{-1}\cdot\alpha\beta}(\sigma^{-1}\cdot\kappa)\bar\phi^\beta\,.
\end{align}
Moreover,
\begin{align}
\sum_{\sigma\cdot\beta\le\alpha} =
\sum_{\beta_{\sigma(1)}=0}^{\alpha_1}\ldots
\sum_{\beta_{\sigma(d)}=0}^{\alpha_d} =
\sum_{\beta_1=0}^{\alpha_{\sigma^{-1}(1)}}\ldots
\sum_{\beta_d=0}^{\alpha_{\sigma^{-1}(d)}} = \sum_{\beta\le \sigma^{-1}\cdot \alpha}
\end{align}
whence we conclude
\begin{equation}
V_\alpha\left(\sigma\cdot\bar\phi\right) = \sum_{\beta\le\sigma^{-1}\cdot \alpha}v_{(\sigma^{-1}\cdot\alpha)\beta}(\sigma\cdot\kappa)\bar\phi^\beta = V_{\sigma^{-1}\cdot\alpha}(\bar\phi)\quad\text{if}\quad \sigma\cdot\kappa = \kappa\,.
\end{equation}
Analogously, we have
\begin{align}
U_\alpha(\sigma\cdot\bar\phi) & = \frac{1}{{\cal D}_\kappa(\sigma\cdot\bar\phi)}\frac{\partial^{|\alpha|}}{\partial\phi_{\sigma(1)}^{\alpha_1}\ldots\partial\phi_{\sigma(d)}^{\alpha_d}}\left\{\prod_{m=1}^{Q-1}\phi_{\sigma(m)}^{\alpha_m+\kappa_m-1}(1-|\bar\phi|)^{|\alpha|+\kappa_{Q}-1}\right\} \nonumber\\[2.0ex]
& = \frac{1}{{\cal D}_{\sigma^{-1}\cdot\kappa}(\bar\phi)}\partial^{|\alpha|}_{\sigma^{-1}\cdot\alpha}\left\{\prod_{m=1}^{Q-1}\phi_m^{\alpha_{\sigma^{-1}(m)}+\kappa_{\sigma^{-1}(m)}-1}(1-|\bar\phi|)^{|\alpha|+\kappa_{Q}-1}\right\}\nonumber\\[2.0ex] & = U_{\sigma^{-1}\cdot \alpha}(\bar\phi)\quad\text{if}\quad\sigma^{-1}\cdot\kappa = \kappa\,.
\end{align}
It is trivially clear that $f_{\sigma\cdot\alpha}=f_\alpha$ if $\kappa
= {\kappa_{\scriptscriptstyle\text{iso}}}$.\qed
\begin{figure}[t!]
\small
\vskip 0.3cm
\centering
\colorbox{myblue}
{
\parbox{0.4\textwidth}
{
\vbox{
\texttt{for}\ k \texttt{from } $0$ \texttt{ to } $n$ \texttt{ do}\\[0.5ex]
\hbox{\texttt{\ \ for } $\alpha$ \texttt{\,in }${\mathfrak P}_{k,Q-1}$\texttt{ do } }\\[0.5ex]
\hbox{\texttt{\ \ \ \ for } $\gamma$ \texttt{\,in }$\Omega_n$\texttt{ do }} \\[0.5ex]
\hbox{\texttt{\ \ \ \ \ \ compute }
$\chi_{\alpha\gamma} = \langle V_\alpha,{\cal D}_\gamma\rangle_{{\kappa_{\scriptscriptstyle\text{iso}}}}$} \\[0.5ex]
\hbox{\texttt{\ \ \ \ \ \ for } $\beta$ \texttt{\,in } $\Pi(\alpha)/\{\alpha\}$\texttt{ do }} \\[0.5ex]
\hbox{\texttt{\ \ \ \ \ \ \ \ find } $\sigma\in S_{Q-1}: \sigma\cdot\beta = \alpha$} \\[0.5ex]
\hbox{\texttt{\ \ \ \ \ \ \ \ assign } $\chi_{\beta(\sigma^{-1}\cdot\gamma)} \leftarrow \chi_{\alpha\gamma}$} \\[0.5ex]
\hbox{ \texttt{\ \ \ \ \ \ end do } }\\[0.5ex]
\hbox{ \texttt{\ \ \ \ end do } }\\[0.5ex]
\hbox{ \texttt{\ \ end do } }\\[0.5ex]
\hbox{ \texttt{end do } }\\[-3.0ex]
}
}
}
\vskip 0.2cm
\caption{A convenient recipe to compute $\chi$ when $\kappa={\kappa_{\scriptscriptstyle\text{iso}}}$.}
\label{fig:chialg}
\end{figure}
\noindent From Prop.~3 it follows
\begin{align}
\chi_{\alpha\gamma} & = \int_{\bar T_Q(s)}\text{d}\bar\phi\, V_\alpha(\bar\phi){\cal D}_\gamma(\bar\phi){\cal D}_{\kappa_{\scriptscriptstyle\text{iso}}}(\bar\phi) \!\!\!\! \stackrel {\phantom{\int_{a_a}} \bar\phi \to \sigma\cdot\bar\phi'\ \ }{=} \!\!\! \int_{\bar T_Q(s)}\text{d}\bar\phi'\, V_\alpha(\sigma\cdot \bar\phi'){\cal D}_\gamma(\sigma\cdot\bar\phi'){\cal D}_{\kappa_{\scriptscriptstyle\text{iso}}}(\sigma\cdot\bar\phi')\nonumber\\[2.0ex]
& = \ \int_{\bar T_Q(s)}\text{d}\bar\phi'\, V_{\sigma^{-1}\cdot\alpha}(\bar\phi') {\cal D}_{\sigma^{-1}\cdot \gamma}(\bar\phi'){\cal D}_{\kappa_{\scriptscriptstyle\text{iso}}}(\bar\phi') = \chi_{(\sigma^{-1}\cdot\alpha)( \sigma^{-1}\cdot\gamma)}\,\qquad \text{if} \ \kappa = {\kappa_{\scriptscriptstyle\text{iso}}}\,.
\end{align}
Accordingly, many matrix elements of $\chi$ are exactly the same,
which explains the little--square structure of Fig.~\ref{fig:chi}. In
order to establish a convenient way of computing $\chi$, we introduce
the partition set
\begin{equation}
{\mathfrak P}_{n,Q-1} = \{\alpha\in \mathds{N}_0^{Q-1}: |\alpha|=n \ \text{ and }
\ \alpha_1\ge\alpha_2\ge\ldots\ge\alpha_{Q-1}\}\,,
\label{eq:partone}
\end{equation}
and for $\alpha\in \mathds{N}_0^{Q-1}$ the permutation set
\begin{equation}
\Pi(\alpha) = \left\{\eta\in\mathds{N}_0^{Q-1}: \ \ \eta=\sigma\cdot\alpha
\text{ \ \ for some \ }\sigma\in S_{Q-1}\right\}\,.
\label{eq:permutone}
\end{equation}
Obviously, if $\alpha$ has $m_1$ components equal to $a_1$, \ldots,
$m_r$ components equal to $a_r$, such that
\begin{equation}
m_1 + \ldots + m_r = Q-1\,,\qquad a_1m_1+\ldots+a_rm_r = |\alpha|\,,
\end{equation}
then it holds
\begin{equation}
|\Pi(\alpha)| = \frac{(Q-1)!}{m_1!\ldots m_r!} = {Q-1\choose m_1,\ldots,m_r}\,.
\end{equation}
Partitions and permutations allow to decompose the index space of the
orthogonal polynomials as a union of disjoint sets, namely
\begin{equation}
\{\alpha\in\mathds{N}_0^{Q-1}: |\alpha|\le n\} = \bigcup_{k=0}^n
\bigcup_{\alpha\in{\mathfrak P}_{k,Q-1}}\Pi(\alpha)\,.
\label{eq:domdecomp}
\end{equation}
In Fig.~\ref{fig:chialg} we provide a recipe to compute $\chi$, which
is based on the above set decomposition and works correctly since
$\Omega_n$ is permutationally closed. Moreover, it is well known since
Euler's age \cite{knuth4a} that $p_{n,Q-1} \equiv
|\mathfrak{P}_{n,Q-1}|$ can be obtained from the generating function
\begin{equation}
f(x) = \prod_{n=1}^{Q-1}\frac{1}{1-x^k} = \sum_{n=0}^\infty p_{n,Q-1}x^k\,.
\end{equation}
If we define the truncated Taylor expansion
\begin{equation}
F_n(x) = \sum_{k=0}^np_{k,Q-1}x^k\,,
\end{equation}
then $q_{n,Q-1} = F_n(1)$ represents the total number of matrix rows
$\alpha$ for which $\chi_{\alpha\gamma}$ really needs to be
computed. In Table~\ref{tab:qkd}, we report $q_{n,Q-1}$ for the first
few values of $n$ and $Q$.
\begin{table}[!t]
\small
\centering
\begin{tabular}{c|cccccccccccccc}
\hline\hline\\[-2.0ex]
$n\backslash Q$ & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 & 11 & 12 & 13
& 14 & 15 \\[0.2ex]
\hline\\[-2.3ex]
0 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 \\
1 & 2 & 2 & 2 & 2 & 2 & 2 & 2 & 2 & 2 & 2 & 2 & 2 & 2 & 2 \\
2 & 4 & 4 & 4 & 4 & 4 & 4 & 4 & 4 & 4 & 4 & 4 & 4 & 4 & 4 \\
3 & 6 & 7 & 7 & 7 & 7 & 7 & 7 & 7 & 7 & 7 & 7 & 7 & 7 & 7 \\
4 & 9 & 11 & 12 & 12 & 12 & 12 & 12 & 12 & 12 & 12 & 12 & 12 & 12 & 12 \\
5 & 12 & 16 & 18 & 19 & 19 & 19 & 19 & 19 & 19 & 19 & 19 & 19 & 19 & 19 \\
6 & 16 & 23 & 27 & 29 & 30 & 30 & 30 & 30 & 30 & 30 & 30 & 30 & 30 & 30 \\
7 & 20 & 31 & 38 & 42 & 44 & 45 & 45 & 45 & 45 & 45 & 45 & 45 & 45 & 45 \\
8 & 25 & 41 & 53 & 60 & 64 & 66 & 67 & 67 & 67 & 67 & 67 & 67 & 67 & 67 \\
9 & 30 & 53 & 71 & 83 & 90 & 94 & 96 & 97 & 97 & 97 & 97 & 97 & 97 & 97 \\
10 & 36 & 67 & 94 & 113 & 125 & 132 & 136 & 138 & 139 & 139 & 139 & 139 & 139 & 139 \\
11 & 42 & 83 & 121 & 150 & 169 & 181 & 188 & 192 & 194 & 195 & 195 & 195 & 195 & 195 \\
12 & 49 & 102 & 155 & 197 & 227 & 246 & 258 & 265 & 269 & 271 & 272 & 272 & 272 & 272 \\
13 & 56 & 123 & 194 & 254 & 298 & 328 & 347 & 359 & 366 & 370 & 372 & 373 & 373 & 373 \\
14 & 64 & 147 & 241 & 324 & 388 & 433 & 463 & 482 & 494 & 501 & 505 & 507 & 508 & 508 \\
15 & 72 & 174 & 295 & 408 & 498 & 564 & 609 & 639 & 658 & 670 & 677 & 681 & 683 & 684 \\[2.0ex]
\hline\hline
\end{tabular}
\caption{First coefficients $q_{n,Q-1}$.}
\label{tab:qkd}
\end{table}
\subsection{Symmetric solutions of the Fokker--Planck equation}
If the FPE is symmetric under a subset of index permutations of
$\bar\phi$, then its solution is expected to have the same symmetry
(we assume here that no spontaneous symmetry breaking occurs). A
permutationally symmetric FPE describes a system where no
physical state is \emph{a priori} favoured with
respect to the others, a frequent case in phenomenological applications.
For simplicity's sake, we assume that the FPE is maximally
symmetric, {\it i.e.\ } it is symmetric under
$\bar\phi\to\sigma\cdot\bar\phi$, $\forall \sigma\in S_{Q-1}$. As an
example, the reader could consider a multi--state voter model with
zealots, where $Z_k=Z$ for $k=1,\ldots,Q$ and the network topology
preserves the symmetry (for a non--trivial instance, see ref.~\cite{palombi}).
Imposing that ${\cal P}_n$ is permutationally invariant yields
\begin{align}
{\cal P}_n(\sigma\cdot\bar\phi) & =\sum_{\gamma\in\Omega_n}c_\gamma{\cal D}_\gamma(\sigma\cdot\bar\phi) =\sum_{\gamma\in\Omega_n}c_\gamma{\cal D}_{\sigma^{-1}\cdot\gamma}(\bar\phi)\nonumber\\[1.0ex] & =\sum_{\gamma\in\Omega_n}c_{\sigma\cdot\gamma}{\cal D}_{\gamma}(\bar\phi)
=\sum_{\gamma\in\Omega_n}c_{\gamma}{\cal D}_{\gamma}(\bar\phi) = {\cal P}_n(\bar\phi)\,,
\end{align}
and since $\{{\cal D}_\gamma\}$ is a polynomial basis, we infer
$c_{\sigma\cdot\gamma}=c_\gamma$,
$\forall \gamma\in\Omega_n$ and $\forall \sigma\in S_{Q-1}$. Of
course, this result can be fruitfully used as a check of the numerical
implementation of the RG method. Nevertheless, in this section we
would like to discuss whether we can include such information even in
the theoretical construction of the weak solution. In analogy with
eqs.~(\ref{eq:partone})--(\ref{eq:permutone}), we introduce the
partition set
\begin{equation}
{\mathfrak L}_n = \{\gamma\in \Omega_n: \ \gamma_1\ge\gamma_2\ge\ldots\ge\gamma_{Q-1}\}\,,
\label{eq:parttwo}
\end{equation}
and for $\gamma\in\Omega_n$ the permutation set
\begin{equation}
\Lambda(\gamma) = \left\{\eta\in\Omega_n: \ \ \eta=\sigma\cdot\gamma
\text{ \ \ for some \ }\sigma\in S_{Q-1}\right\}\,.
\label{eq:permuttwo}
\end{equation}
If ${\cal P}_n$ is permutationally symmetric, then it can be written as
\begin{equation}
{\cal P}_n(\bar\phi) =
\sum_{\gamma\in{\mathfrak L}_n}c_\gamma\sum_{\eta\in\Lambda(\gamma)}{\cal D}_\eta(\bar\phi)
= \sum_{\gamma\in{\mathfrak L}_n}\hat c_\gamma
\frac{1}{|\Lambda(\gamma)|}\sum_{\eta\in\Lambda(\gamma)}{\cal D}_\eta(\bar\phi)
= \sum_{\gamma\in{\mathfrak L}_n}\hat c_\gamma \,{\cal S}\cdot{\cal D}_\gamma(\bar\phi)\,,
\label{eq:Psym}
\end{equation}
where we have introduced the rescaled coefficient $\hat c_\gamma =
|\Lambda(\gamma)|c_\gamma$ and
the symmetrized Dirichlet distribution ${\cal S}\cdot{\cal D}_\gamma =
|\Lambda(\gamma)|^{-1}\sum_{\eta\in\Lambda(\gamma)}{\cal D}_\eta$. Eq.~(\ref{eq:Psym})
tells us that a symmetric RG solution can be represented as a linear
combination of symmetrized Dirichlet distributions (which are
symmetric!). Can we reformulate the whole RG method so as to only make
use of symmetrized Dirichlet distributions? The answer is affirmative, yet the
reader should not undervalue technicalities.
\noindent\emph{i}) As a preliminary observation, we argue that a
symmetrized Dirichlet distribution faithfully decomposes into a basis
of symmetrized orthogonal polynomials. To this aim, we first need to
examine the permutational properties of the coefficients
$\{d_{\gamma\beta}\}$ connecting the Dirichlet basis $\{{\cal D}_\gamma\}$
to the Appel basis $\{U_\beta\}$, see eq.~(\ref{eq:Dirproj}). Under
the isotropic assumption, a permutation of the components of
$\bar\phi$ results in
\begin{align}
{\cal D}_\gamma(\sigma\cdot\bar\phi) &
=\sum_{|\beta|\le|\slashed{\gamma}|}d_{\gamma\beta}\,U_\beta(\sigma\cdot\bar\phi)
=\sum_{|\beta|\le|\slashed{\gamma}|}d_{\gamma\beta}\,U_{\sigma^{-1}\cdot\beta}(\bar\phi)
=\sum_{|\beta|\le|\slashed{\gamma}|}d_{\gamma(\sigma\cdot\beta)}\,U_{\beta}(\bar\phi)\,.
\end{align}
However,
\begin{equation}
{\cal D}_\gamma(\sigma\cdot\bar\phi) =
{\cal D}_{\sigma^{-1}\cdot\gamma}(\bar\phi) \ =
\sum_{|\beta|\le|\slashed{\gamma}|}d_{(\sigma^{-1}\cdot\gamma)\beta}\,U_\beta(\bar\phi)\,,
\end{equation}
whence we infer
\begin{equation}
\sum_{|\beta|\le|\slashed{\gamma}|}d_{(\sigma^{-1}\cdot\gamma)\beta}\,U_\beta(\bar\phi)
\ =\sum_{|\beta|\le|\slashed{\gamma}|}d_{\gamma(\sigma\cdot\beta)}\,U_{\beta}(\bar\phi)\,.
\end{equation}
Since $\{U_\beta\}$ is a polynomial basis, we conclude that
$d_{\gamma(\sigma\cdot\beta)} =
d_{(\sigma^{-1}\cdot\gamma)\beta}$ (of course, this could have
been equivalently obtained via the identity $\chi_{\alpha\gamma} =
\chi_{(\sigma^{-1}\cdot\alpha)(\sigma^{-1}\cdot\gamma)}$, discussed in
sect.~6.1). Then, we apply the symmetrization
operator ${\cal S}$ to both sides of eq.~(\ref{eq:Dirproj}), namely
\begin{align}
{\cal S}\cdot{\cal D}_\gamma(\bar\phi) & =
\frac{1}{|\Lambda(\gamma)|}\,\sum_{\eta\in\Lambda(\gamma)}\sum_{|\beta|\le
|\slashed{\gamma}|} d_{\eta\beta}
U_\beta(\bar\phi)\nonumber\\[2.0ex]
& = \sum_{|\beta|\le
|\slashed{\gamma}|}\left[\frac{1}{|\Lambda(\gamma)|}\,\sum_{\eta\in\Lambda(\gamma)}d_{\eta\beta}\right]
U_\beta(\bar\phi) = \sum_{|\beta|\le |\slashed{\gamma}|}
e_{\gamma\beta}U_\beta(\bar\phi)\,,
\end{align}
with $e_{\gamma\beta}\equiv
|\Lambda(\gamma)|^{-1}\,\sum_{\eta\in\Lambda(\gamma)}d_{\eta\beta} =
{\cal S}\cdot d_{\gamma\beta}$. We can easily work out the permutational
properties of the coefficients $\{e_{\gamma\beta}\}$. We have indeed
\begin{align}
e_{\gamma(\sigma\cdot\beta)}
=\frac{1}{|\Lambda(\gamma)|}\sum_{\eta\in\Lambda(\gamma)}d_{\eta(\sigma\cdot\beta)}
=
\frac{1}{|\Lambda(\gamma)|}\sum_{\eta\in\Lambda(\gamma)}d_{(\sigma^{-1}\cdot\eta)\beta}
= e_{\gamma\beta}\,,
\end{align}
as $\Lambda(\gamma) = \Lambda(\sigma\cdot\gamma)$ for any $\sigma\in S_{Q-1}$. For the same reason, it holds
\begin{equation}
e_{(\sigma\cdot\gamma)\beta}
=\frac{1}{|\Lambda(\sigma\cdot\gamma)|}\sum_{\eta\in\Lambda(\sigma\cdot\gamma)}d_{\eta\beta}
=
\frac{1}{|\Lambda(\gamma)|}\sum_{\eta\in\Lambda(\gamma)}d_{\eta\beta}
= e_{\gamma\beta}\,.
\end{equation}
Therefore, we conclude that $e_{\gamma(\sigma\cdot\beta)} =
e_{(\sigma\cdot\gamma)\beta} = e_{\gamma\beta}$. By decomposing
the set $\{|\beta|\le|\slashed{\gamma}|\}$ according to
eq.~(\ref{eq:domdecomp}), we finally obtain
\begin{align}
{\cal S}\cdot{\cal D}_\gamma(\bar\phi) & =
\sum_{k=0}^{|\slashed{\gamma}|}\sum_{\beta\in{\mathfrak P}_{k,Q-1}}\sum_{\alpha\in\Pi(\beta)}e_{\gamma\alpha}\,U_\alpha(\bar\phi)
=
\sum_{k=0}^{|\slashed{\gamma}|}\sum_{\beta\in{\mathfrak P}_{k,Q-1}}e_{\gamma\beta}\sum_{\alpha\in\Pi(\beta)}\,U_\alpha(\bar\phi)
\nonumber\\[2.0ex]
& = \sum_{k=0}^{|\slashed{\gamma}|}\sum_{\beta\in{\mathfrak P}_{k,Q-1}}\hat e_{\gamma\beta}\,{\cal S}\cdot U_\beta(\bar\phi)\,,
\label{eq:symdecomp}
\end{align}
with $\hat e_{\gamma\beta} \equiv
|\Pi(\beta)|e_{\gamma\beta}$. Eq.~(\ref{eq:symdecomp}) tells us that
since ${\cal S}\cdot{\cal D}_\gamma$ is a symmetric function, it
decomposes faithfully into a set of symmetric polynomials $\{{\cal S}\cdot
U_\beta\}$. In addition, we notice that if
$\alpha,\beta\in\mathds{N}_0^{Q-1}$, then either
$\Pi(\alpha)\cap\Pi(\beta)=\emptyset$ or $\Pi(\alpha)=\Pi(\beta)$.
Therefore, it makes sense to define
\begin{equation}
\delta_{\Pi(\beta)\Pi(\gamma)} = \left\{\begin{array}{ll} 1 & \text{
if }\ \Pi(\beta) = \Pi(\gamma)\,,\\[1.0ex] 0 & \text{ if
}\ \Pi(\beta)\cap\Pi(\gamma) = \emptyset\,.\end{array}\right.
\end{equation}
Finally, we observe that
\begin{align}
\langle{\cal S}\cdot V_\alpha,{\cal S}\cdot U_\beta\rangle_\kappa &
=\frac{1}{|\Pi(\alpha)}\frac{1}{|\Pi(\beta)|}\sum_{\epsilon\in\Pi(\alpha)}\sum_{\eta\in\Pi(\beta)}
\langle V_\epsilon,U_\eta\rangle_\kappa \nonumber\\[2.0ex]
& =
\frac{1}{|\Pi(\alpha)}\frac{1}{|\Pi(\beta)|}\sum_{\epsilon\in\Pi(\alpha)}\sum_{\eta\in\Pi(\beta)}
f_\epsilon \delta_{\epsilon\eta} =
\frac{f_\alpha}{|\Pi(\alpha)|}\delta_{\Pi(\alpha)\Pi(\beta)}\,,
\end{align}
since $f_\alpha$ is permutationally invariant under the isotropic
assumption. Now, projecting ${\cal S}\cdot{\cal D}_\gamma$ onto ${\cal S}\cdot
V_\alpha$ with $|\alpha|\le|\slashed{\gamma}|$ yields
\begin{align}
\langle {\cal S}\cdot V_\alpha,{\cal S}\cdot{\cal D}_\gamma\rangle_\kappa &
=\sum_{k=0}^{|\slashed{\gamma}|}\sum_{\beta\in{\mathfrak P}_{k,Q-1}}\hat
e_{\gamma\beta}\,\langle {\cal S} V_\alpha,{\cal S} U_\beta\rangle_\kappa
\nonumber\\[1.0ex]
& = \sum_{k=0}^{|\slashed{\gamma}|}\sum_{\beta\in{\mathfrak P}_{k,Q-1}}
\frac{\hat
e_{\gamma\beta}f_\alpha}{|\Pi(\alpha)|}\delta_{\Pi(\alpha)\Pi(\beta)}
= e_{\gamma\alpha}f_\alpha\,,
\end{align}
while the above scalar product vanishes for
$|\alpha|>|\slashed{\gamma}|$. We conclude
\begin{equation}
\langle {\cal S}\cdot V_\alpha,{\cal S}\cdot{\cal D}_\gamma\rangle_\kappa =
\left\{\begin{array}{ll} \chi_{\alpha\gamma}\cdot\overleftarrow{{\cal S}} & \text{ if
}\ |\alpha|\le|\slashed{\gamma}|\,,\\[2.0ex] 0 & \text{
otherwise\,,}\end{array}\right.
\label{eq:symscalprod}
\end{equation}
where $\overleftarrow{{\cal S}}$ symmetrizes from the right over the index array $\gamma$.
\noindent \emph{ii}) We have already observed
that the action of ${\cal L}_{\scriptscriptstyle{\rm FP}}$ mixes Dirichlet distributions with index
arrays in different \emph{bucket spaces}, corresponding to different
polynomial degrees. Since the symmetrization operator ${\cal S}$ averages over
permutations, we need to clarify how it behaves with
respect to a shift of degree. This is needed in order to generate a
dictionary of reference formulae analogous to
eqs.~(\ref{eq:firstalg})--(\ref{eq:lastalg}). From a theoretical point
of view, the problem originates from the fact that the index
raising/lowering operators do not commute with permutations. Indeed,
they fulfill the relations
\begin{align}
\label{eq:opluscommut}
\oplus_\ell\cdot \sigma & = \sigma\cdot \oplus_{\sigma(\ell)}\\[1.0ex]
\ominus_\ell\cdot \sigma & = \sigma\cdot
\ominus_{\sigma(\ell)}\,,\qquad \ell=1,\ldots,Q-1\,,\qquad
\sigma\in S_{Q-1}\,.
\label{eq:ominuscommut}
\end{align}
As an example, we notice that the action of $\phi_k^m\partial_k^n$ on
${\cal S}\cdot{\cal D}_\gamma$ breaks the permutational symmetry of the symmetrized
Dirichlet distribution and shifts its degree for $m\ne n$. If we then sum
over $k$, the symmetry is recovered, but in general the result cannot be
written anymore as a permutational average over
$\Lambda(\gamma)$. With the same spirit by which we wrote down
eqs.~(\ref{eq:firstalg})--(\ref{eq:lastalg}), we consider some specific
cases. Two very simple ones are
\begin{align}
& \sum_{\ell=1}^{Q-1}\phi_\ell\,{\cal S}\cdot{\cal D}_\gamma(\bar\phi) =
\frac{1}{|\Lambda(\gamma)|}\sum_{\eta\in\Lambda(\gamma)}|\bar\phi|\,{\cal D}_\eta(\bar\phi)
=
\frac{1}{|\Lambda(\gamma)|}\sum_{\eta\in\Lambda(\gamma)}(s+|\bar\phi|-s)\,{\cal D}_\eta(\bar\phi)
\nonumber\\[1.0ex]
& \hskip 0.2cm = s\,{\cal S}\cdot{\cal D}_\gamma(\bar\phi) - s\frac{\gamma_Q}{|\gamma|}\,{\cal S}\cdot{\cal D}_{\gamma_{Q^+}}(\bar\phi)\,,
\label{eq:firstanalytic}
\end{align}
and
\begin{align}
& \sum_{\ell=1}^{Q-1}\phi_\ell
\partial_\ell\,{\cal S}\cdot{\cal D}_\gamma(\bar\phi) =
\frac{1}{|\Lambda(\gamma)|}\sum_{\eta\in\Lambda(\gamma)}\sum_{\ell=1}^{Q-1}\left[\theta_{\eta_\ell,2}(\eta_\ell-1){\cal D}_\eta(\bar\phi)-\theta_{\eta_Q,2}\eta_\ell{\cal D}_{\eta_{\ell^+Q^-}}(\bar\phi)\right]\nonumber\\[1.0ex]
& \hskip 0.2cm =
\left[\sum_{\ell=1}^{Q-1}\theta_{\gamma_\ell,2}(\gamma_\ell-1)+\theta_{\gamma_Q,2}(\gamma_Q-1)\right]\,{\cal S}\cdot{\cal D}_\gamma(\bar\phi)
-\theta_{\gamma_Q,2}(|\gamma|-1)\,{\cal S}\cdot{\cal D}_{\gamma_{Q^-}}(\bar\phi)\,.
\label{eq:secondanalytic}
\end{align}
As can be seen, in both cases the result is still a permutational
average over
$\Lambda(\gamma)$. However, let us consider the action of
$\sum_{k=1}^{Q-1}\partial_k$ on ${\cal S}\cdot{\cal D}_\gamma$. From
eq.~(\ref{eq:delDgamma}), we have
\begin{align}
& \sum_{\ell=1}^{Q-1}\partial_\ell\,{\cal S}\cdot{\cal D}_\gamma(\bar\phi) =
\frac{s^{-1}(|\gamma|-1)}{|\Lambda(\gamma)|}\sum_{\eta\in\Lambda(\gamma)}
\sum_{\ell=1}^{Q-1}[\theta_{\eta_\ell,2}{\cal D}_{\eta_{\ell^-}}(\bar\phi)
- \theta_{\eta_Q,2}{\cal D}_{\eta_{Q^-}}(\bar\phi)]\nonumber\\[2.0ex]
& \hskip 0.2cm =
s^{-1}(|\gamma|-1)\left\{\left[\frac{1}{|\Lambda(\gamma)|}\sum_{\eta\in\Lambda(\gamma)}\sum_{\ell=1}^{Q-1}\theta_{\eta_\ell,2}{\cal D}_{\eta_{\ell^-}}(\bar\phi)\right]
- \theta_{\gamma_Q,2}(Q-1)\,{\cal S}\cdot{\cal D}_{\gamma_{Q^-}}(\bar\phi)\right\}\,.
\end{align}
In order to check that the sum in square brackets is
permutationally invariant, it is sufficient to make use of
eq.~(\ref{eq:ominuscommut}) and observe that
$\theta_{\eta_\ell,2}{\cal D}_{\eta_{\ell^{-}}}(\sigma\cdot\bar\phi) =
\theta_{\gamma_{\sigma(\ell)},2}{\cal D}_{\sigma\cdot\gamma_{\sigma(\ell)^{-}}}(\sigma\cdot\bar\phi)
=
\theta_{\gamma_{\sigma(\ell)},2}{\cal D}_{\gamma_{\sigma(\ell)^-}}(\bar\phi)$,
which upon summing over $\ell$ becomes manifestly invariant. Since
$|\slashed{\eta}_{\ell^-}|=n-1$ for $\eta\in\Lambda(\gamma)$, we can
write
\begin{equation}
\frac{1}{|\Lambda(\gamma)|}\sum_{\eta\in\Lambda(\gamma)}\sum_{\ell=1}^{Q-1}\theta_{\eta_\ell,2}{\cal D}_{\eta_{\ell^-}}(\bar\phi)
= \sum_{\eta\in{\cal L}_{n-1}}\hat g_\eta\,{\cal S}\cdot{\cal D}_\eta(\bar\phi)\,,
\label{eq:deelsymexp}
\end{equation}
since we have already shown in eq.~(\ref{eq:Psym}) --- by an argument that
could be here repeated --- that a polynomial, which we know to be
symmetric, can be expanded as a linear combination of symmetrized
Dirichlet distributions (in this case the index arrays live on
${\cal L}_{n-1}$ due to the degree shift produced by the
differentiation). Determining the coefficients $\{\hat g_\eta\}$
analytically is non--trival and beyond the aims of this
paper. Nevertheless, by projecting eq.~(\ref{eq:deelsymexp}) onto the
symmetrized orthogonal polynomials and by making use of
eq.~(\ref{eq:symscalprod}), we obtain
\begin{equation}
\sum_{\eta\in{\cal L}_{n-1}} (e_{\eta\alpha}f_\alpha)\hat g_\eta =
\frac{1}{|\Lambda(\gamma)|}\sum_{\eta\in\Lambda(\gamma)}\sum_{\ell=1}^{Q-1}\theta_{\eta_\ell,2}\langle
{\cal S}\cdot V_\alpha,{\cal D}_{\eta_{\ell^-}}(\bar\phi)\rangle_\kappa\,,
\label{eq:firstnonanal}
\end{equation}
which can be numerically inverted.
Along the same line, all the symmetric counterparts of
eqs.~(\ref{eq:firstalg})--(\ref{eq:lastalg}) can be worked
out. Analytic expressions such as
eqs.~(\ref{eq:firstanalytic})--(\ref{eq:secondanalytic})
effectively help save CPU time, while formulae requiring numerical
inversions such as eq.~(\ref{eq:firstnonanal}) are of no benefit. Such
cases require more sophisticated analyses, which we do not
attempt here.
\section{Conclusions}
In this paper we have explored the possibility of representing
the solution of the Fokker--Planck equation for many--variable
steady--state birth--death systems as a linear combination of
Dirichlet distributions. This idea was first suggested in
\cite{palombi}, where a variant of the multi--state
voter model with zealots over a community--based network
\cite{Girvan} was studied. We have shown here that \emph{quasi}--optimal
coefficients for such a linear expansion can be generally obtained
from a variant of the Ritz--Galerkin method for partial differential
equations. As a test, we have applied the Dirichlet expansion
successfully to the binary and multi--state voter models with zealots
on a complete graph. Although Ritz--Galerkin techniques are widely
employed in engineering applications, no adaptation to systems defined
on the simplex has been ever considered in the literature, to the best
of our knowledge.
We expect the domain of applicability of the method to go beyond that
of voter models and to extend to a \emph{positive--measure subset} of
statistical physics. Applications could include variants of SIS model
for epidemic spreading, naming games (a variant with committed agents
has been recently studied in \cite{Xie2}) and other complex
systems, only subject to the conditions that \emph{i}) a steady--state
distribution with positive variances exists and \emph{ii}) the drift
and diffusion coefficients of the Fokker--Planck equation are
polynomials. With regard to condition \emph{i}), our proposal could be
generalized by considering an expansion in Dirichlet distributions
with linear coefficients depending on time, so as to allow for a
treatment of the time--dependent Fokker--Planck equation. This would
permit to describe the system while it relaxes to
equilibrium. Nevertheless, systems with consensus--like exit states,
which have
attracted much attention in recent years, are anyway ruled out as
finite--degree polynomials can never approximate a Dirac delta
distribution. Concerning condition~\emph{ii}), cases where the drift
and diffusion coefficients are non--polynomial analytic functions
could be maybe faced by expanding them in power series to some finite
degree $m$ and by subsequently applying the Ritz-Galerkin method; this
would generate a sequence $({\cal P}_{n,m})$ of solutions, whose
convergence for $n,m\to\infty$ should be studied.
We conclude by recalling that in this work we have provided no
theoretical arguments to show that the coerciveness condition of the
Lax--Milgram theorem is generally fulfilled by a Fokker--Planck
operator with polynomial diffusion matrix on the simplex. Without a
general proof, the applicability of the method has to be checked on a
case--by--case basis.
|
\section{Introduction}
\section{Introduction}
It is highly desirable to electronically manipulate the photonic spectrum of a multi-level emitter such as an atom or quantum dot (QD). While it is well-known that the spectrum is influenced by the photon emitter's electromagnetic environment (e.g., via the Purcell effect \cite{Purcell}), engineering the environment to obtain desirable characteristics often results in a fixed structure that is not actively tunable. Surface plasmon polaritons (SPPs) on graphene \cite{Wunsch}-\cite{Han2008} are highly tunable, and offer a promising way to achieve electronic control over an emitter's spectrum through interactions with graphene SPPs. Recent scattering-type scanning near field optical microscopy (SNOM) imaging experiments \cite{Chen} have demonstrated in real space the excitation of graphene SPPs on finite graphene structures, and in \cite{Yan} excitation and damping of SPPs on graphene structures was investigated experimentally for graphene on several substrates. Additionally, graphene quantum plasmonics has been considered in \cite{Koppens} where vacuum Rabi splitting was shown, and in \cite{Man} where active control over a quantum state via biasing was demonstrated.
When placed in the vicinity of a multi-level emitter, graphene, along with the vacuum density of electromagnetic field modes, forms the photonic reservoir with which the emitter interacts. The spectral and statistical properties of such a system is strongly dependent on the reservoir mode density via the local density of states (LDOS) \cite{Carmichael1}. In \cite{Vlack}, the reservoir of electromagnetic modes is altered by the presence of a metal nano-particle, and the resonance fluorescence was examined in the vicinity of the nano-particle plasmon resonance. Several disadvantages of this system is that the LDOS is not tunable, and placing a photon emitter at the desired spatial position is challenging. From a practical viewpoint, one desires a spatial position that is translationally invariant, e.g., near a surface, with an LDOS that can be tuned
in a controllable manner.
In this work we use translationally invariant graphene, which is electronically-tunable, to alter the reservoir for a two-level artificial atom (hereafter referred to as a QD) in a controllable manner. Furthermore, it is known that a graphene support structure consisting of a dielectric layer can play a role in shaping the LDOS \cite{Hanson}, and so we also consider the effect of a substrate on the resonance fluorescence.
Since the plasmon response of graphene exhibits a strong dependance on bias in the low-mid THz (meV) frequency regime, we model pump fields and QD excitons at these frequencies \cite{Zibik, Hash}. After obtaining the LDOS properties of the medium, we derive and solve a quantum master equation to demonstrate control over the Mollow triplet \cite{Mollow, Loudon} of a QD by a nearby graphene sheet.
The complex reservoir including the graphene constitutes a lossy inhomogeneous environment for the QD, and here we use a rigorous
photon Green function theory applicable for arbitrary lossy media \cite{Welsch}.
The Mollow triplet is caused by coherent Rabi oscillations and quantum fluctuations and is of fundamental importance.
In addition to exploring how the Mollow triplet changes with a tunable graphene layer, we also show that thermal bath effects are important for low THz frequencies.
\begin{figure}[t]
\centering
\includegraphics[width=0.35\textwidth]{1_a.eps}
\caption{Schematic of a driven QD above supported graphene.} \label{fig:1}
\end{figure}
\section{quantum master equation and Green function theory}
Figure \ref{fig:1} shows the geometry under consideration, consisting of a QD which is a distance $d$ from an infinite graphene sheet. The polarization of the pump field is aligned with the dipole moment, perpendicular to the graphene surface. The Hamiltonian of the coupled system is the sum of QD, pump, reservoir (graphene+vacuum), and their interaction,
\begin{align}
H_{S}& =\hslash\omega_{x}\sigma^{+}\sigma^{-},
H_{\text{pump}} \!=\!\frac{\hslash\Omega}{2}\left(\sigma^{+}e^{-i\omega_{\text{L}}t}
+\sigma^{-}e^{+i\omega_{\text{L}}t}\right),\nonumber \\
\quad H_{R}& =
\hslash\int d\mathbf{r}\intop_{0}^{\infty}\omega\mathbf{f}^{\dagger}\left(\mathbf{r},\omega\right)
\mathbf{f}\left(\mathbf{r},\omega\right)d\omega, \nonumber \\
\quad H_{I}& =-\left[\sigma^{+}\intop_{0}^{\infty}\mathbf{d}\cdot\mathbf{E}\left(\mathbf{r}_{d},
\omega\right)d\omega+\mathrm{H.c.}\right],
\end{align}
where
$\omega_{\text{L}}$ is the THz laser frequency of the pump field, $\sigma^{+}/\sigma_{-}$ are the Pauli operators of the QD exciton, $\Omega=\braket{\mathbf{E}_{\text{pump}}(\mathbf{r}_d)}\cdot \mathbf{d}/\hslash$ is the effective Rabi frequency of the pump source ($\mathbf{d}$ and ${\bf r}_d$ are the QD dipole moment and position),
$\mathbf{f}^{\dagger}/\mathbf{f}$ are bosonic field operators, $\omega_{x}$ is the exciton resonance, and $\mathbf{E}\left(\mathbf{r}_{d},\omega\right)$ is the total electric field operator at the QD position \cite{Welsch},
\begin{equation}
\mathbf{E}\left( \mathbf{r},\omega\right) \! =\!i\sqrt
{\frac{\hslash}{\pi\varepsilon_{0}}}\int\!\sqrt{\operatorname{Im}\!\left(
\varepsilon\left( \mathbf{r}^{\prime},\omega\right)
\right) }\mathbf{G}\left( \mathbf{r},\mathbf{r}^{\prime},\omega
\right) \cdot\mathbf{f}\left( \mathbf{r}^{\prime}
,\omega\right) d\mathbf{r}^{\prime},\label{EF}
\end{equation}
where the permittivity ($\varepsilon$) and Green function (${\bf G}$) describe the photonic environment (reservoir) of the graphene and dielectric background. The Green tensor in the quantum field operator is the classical Green function (propagator) that provides the electromagnetic response at $\mathbf{r}$ due to an excitation at $\mathbf{r}^{\prime}$. All material parameters may be complex-valued.
This Hamiltonian is used to form a quantum master equation as described in Ref.~\cite{Vlack}. However, since we are interested in THz operation we do not make the usual zero-temperature bath approximation. Using the traces
$\mathrm{Tr}_{\text{R}}\![{\mathbf{f}}(\mathbf{r},\omega)
{\mathbf{f}}^\dagger (\mathbf{r}^{\prime},\omega^{\prime})
\rho_{R}] =[\overline{n}(\omega)+1]
\delta ( \mathbf{r}-\mathbf{r}^{\prime})
\delta ( \omega-\omega ^{\prime}),
$
and
$
\mathrm{Tr}_{\text{R}}\left[{\mathbf{f}}^\dag \left( \mathbf{r
,\omega \right) {\mathbf{f}}\left( \mathbf{r}^{\prime },\omega
^{\prime }\right) \rho _{R}\right]=\overline{n}(\omega)
\delta( \mathbf{r}-\mathbf{r}^{\prime})
\delta(\omega -\omega^{\prime})
$,
where the average thermal photon number is $\overline{n}=\left( e^{\hslash \omega /k_{B}T}-1\right) ^{-1}$ and $\rho _{R}=\rho _{R}\left( 0\right) $ is the density operator of the
reservoir which is assumed to initially be in thermal equilibrium,
we obtain the master equation for the time-evolution of the density operator
($\rho = \rho(t)$),
\begin{align}
&\frac{d}{dt}\rho =-\frac{i}{\hslash }\left[ {H}_{\text{S}},\rho
\right] -\int_{0}^{t}d\tau \Big\{ J_{\rm ph}^{\overline{n}+1}\left( \tau
\right)[ {\sigma }_{+}{\sigma }_{-}\left( -\tau \right) \rho
\nonumber \\
& - {\sigma }_{-}(-\tau) \rho {\sigma}_{+}]+\text{H.C.} \Big\}
+L_\text{pure}\,-
\nonumber
\\
& \!\int_{0}^{t}\!d\tau \left\{ J_{\rm ph}^{\overline{n}}\left( -\tau \right)\left [ {\sigma
_{-}{\sigma }_{+}\left( -\tau \right) \rho
- {\sigma }_{+}\left( -\tau \right)
\rho {\sigma }_{-}\right ] +\text{H.c.}\right\} ,
\end{align}
\noindent where $L_{\text{pure}}$ is a pure dephasing term defined in \cite{Vlack}, $\sigma _{\pm }\left( -\tau \right) =e^{-iH_{\text{S}}\tau /\hslash }\widehat
\sigma }_{\pm }e^{iH_{\text{S}}\tau /\hslash }$,
$
\widetilde{J}_{\rm ph}^{\overline{n}}\left( \tau \right) =\int_{0}^{\infty
}d\omega J_{\rm ph}\left( \omega \right) \overline{n}\left( \omega \right)
e^{-i\left( \omega -\omega _{\rm L}\right) \tau },
$
and the photon reservoir function is related to the Green function through
\begin{equation}
J_\text{ph}\left( \omega \right) =\frac{\mathbf{d}\cdot \text{Im}\left( \mathbf{G
\left( \mathbf{r},\mathbf{r},\omega \right) \right) \cdot \mathbf{d}}{\pi
\hslash \varepsilon _{0}}, \label{LDOS}
\end{equation}
which gives a a measure of the QD-environment coupling. Importantly, although at room temperature the average number of phonons at visible
frequencies is negligible, in the THz range $\overline{n}=O(1)$, and so thermal photon effects are required in general.
We assume laterally-infinite graphene modeled as an infinitesimally-thin, local, two-sided surface characterized by a surface conductivity $\sigma$. The Green functions for a graphene sheet at the interface between two dielectrics are given in \cite{Han2008}, and in \cite{Hanson} for graphene on a finite-thickness dielectric support. Considering the graphene sheet in the plane $y=0$, with material described by $\varepsilon _{1}$ for $y>0$ and $\varepsilon _{2}$ for $y<0$, the Green tensor for points in region $n$ is
\begin{equation}
\underline{\mathbf{G}}\left( \mathbf{r}, \mathbf{r'} \right)=\left(\underline{\mathbf{I}} \, k_{n}^{2}+\mathbf{\nabla\nabla\cdot}\right)
\left\{ \underline{\mathbf{g}}
^{\text{p}}\left( \mathbf{r,r}^{\prime}\right) +\underline{\mathbf{g}}^{\text{s}}\left( \mathbf{r,r}^{\prime}\right) \right\},\label{hpt1}
\end{equation}
where $k_{n}=\omega\sqrt{\mu_{0}\varepsilon_{n}}$ is the wavenumber.
The principle (p) and scattered (s) Green's function components are
\begin{align}
&\underline{\mathbf{g}}^{\text{p}}\left( \mathbf{r,r}^{\prime }\right)
\underline{\mathbf{I}}\,\frac{e^{\,ik_{1}R}}{4\pi R},
\ \ \underline{\mathbf{g}}^{\text{s}}\left( \mathbf{r,r}^{\prime }\right) =
\widehat{\mathbf{y}}\widehat{\mathbf{y}}~g_{n}^{\text{s}}\left( \mathbf{r,r
^{\prime }\right) + \nonumber
\\
&
\ \ \left( \widehat{\mathbf{y}}\widehat{\mathbf{x}}\frac
\partial }{\partial x}
+\widehat{\mathbf{y}}\widehat{\mathbf{z}}\frac
\partial }{\partial z}\right) g_{c}^{\text{s}}\left( \mathbf{r,r}^{\prime
}\right) +\left( \widehat{\mathbf{x}}\widehat{\mathbf{x}}~\mathbf{+~
\widehat{\mathbf{z}}\widehat{\mathbf{z}}\right) g_{t}^{\text{s}}\left(
\mathbf{r,r}^{\prime }\right) ,
\end{align
where $\underline{\mathbf{I}}$ is the unit dyadic, $k_{\rho}$ is a radial
wavenumber, $p_{n}^{2}=k_{\rho}^{2}-k_{n}^{2}$, $r=\sqrt{\left( x-x^{\prime}\right)
^{2}+\left( z-z^{\prime}\right)^{2}}$, and $R=\left\vert \mathbf{r-r
^{\prime}\right\vert =\sqrt{\left( y-y^{\prime}\right) ^{2}+r^{2}}$. The Sommerfeld integrals are
\begin{equation}
g_{\beta}^{\text{s}}\left( \mathbf{r,r}^{\prime}\right) =\frac{1}{2\pi
\int_{-\infty}^{\infty}C_{\beta}\frac{H_{0}^{\left( 1\right) }\left(
k_{\rho} r\right) e^{-p\left( y+y^{\prime}\right) }}{4p}k_{\rho
}dk_{\rho},
\end{equation}
where $\beta=t,n,c$ depends on the graphene and dielectric support layers . For a pump polarized perpendicular to the graphene surface we only need $G_{zz}$ and $\beta=n$, with
\begin{equation}
C_{n}=\frac{\left( \frac{\varepsilon _{2}}{\varepsilon _{1}
p_{1}-p_{2}\right) i\omega \varepsilon _{1}-\sigma p_{1}p_{2}}{\left( \frac
\varepsilon _{2}}{\varepsilon _{1}}p_{1}+p_{2}\right) i\omega \varepsilon
_{1}-\sigma p_{1}p_{2}} . \label{C}
\end{equation
\noindent For more complex geometries, such as graphene on a multi-layered dielectric, only the coefficient $C_{n}$ changes.
The wave parameter $p_{n}=\sqrt{k_{\rho}^{2}-k_{n}^{2}}$, leads to
branch points at $k_{\rho}=\pm k_{n}$, and thus the $k_{\rho}$-plane is a
four-sheeted Riemann surface. The standard hyperbolic branch cuts
\cite{Ishimaru} that separate the proper sheet (where $\operatorname{Re
\left( p_{n}\right) >0$, such that the radiation\ condition as $\left\vert
y\right\vert \rightarrow\infty$ is satisfied) and the improper sheet
are the same as in the absence of surface conductivity $\sigma$. The zeros of
the denominators of $C_{\beta}$ lead to pole singularities in the spectral plane associated with surface plasmon polaritons (SPPs).
Using complex-plane analysis, the scattered Green's function can be written as discrete pole (SPP) contributions plus a branch cut integral over the continuum of radiation modes. For $\varepsilon _{1}=\varepsilon _{2}=\varepsilon$, setting the denominator of (\ref{C}) to zero leads to the (TM) SPP wavenumber
$
k_{\rho }=k\sqrt{1-\left( \frac{2}{\sigma \eta }\right) ^{2}}, \label{tmsw}
$
where $\eta =\sqrt{\mu _{0}/\varepsilon }$. In this case, the vertical wavenumber parameter in the Sommerfeld integrals
becomes $p=\sqrt{k_{\rho }^{2}-k^{2}}=i2\omega \varepsilon /\sigma $ and if $\sigma $ is real-valued then $\mathrm{Re}\left( p\right) >0$ is violated and
the TM mode is on the improper Riemann sheet. Assuming complex-valued
conductivity $\sigma =\sigma ^{\prime }+i\sigma ^{\prime \prime }$,
$
p=\frac{i2\omega \varepsilon }{\sigma }=\frac{2\omega \varepsilon }
\left\vert \sigma \right\vert ^{2}}\left( \sigma ^{\prime \prime }+i\sigma
^{\prime }\right) ,
$
and therefore if $\sigma ^{\prime \prime }>0$ (as shown below, when the intraband conductivity
dominates) the mode is a surface wave on the proper sheet, whereas if
\sigma ^{\prime \prime }<0$ (interband conductivity dominates) the mode is
on the improper sheet, assuming an $\text{exp}(-i \omega t)$ reference \cite{MZPRL2007,Han2008}. Assuming the dipole moment is perpendicular to the graphene surface, only TM SPPs can be excited.
The graphene surface conductivity is \cite{GSC2007}
\begin{align}
\sigma \left( \omega \right) =& \frac{ie^{2}k_{B}T}{\pi \hslash ^{2}\left(
\omega +i\gamma \right) }\left( \frac{\mu _{c}}{k_{B}T}+2\ln \left( e^{
\frac{\mu _{c}}{k_{B}T}}+1\right) \right) \nonumber \\
&\!\! + \frac{ie^{2}\left( \omega +i\Gamma \right) }{\pi \hslash ^{2}
\int_{0}^{\infty }\frac{f_{d}\left( -\varepsilon \right) -f_{d}\left(
\varepsilon \right) }{\left( \omega +i\Gamma \right) ^{2}-4\left(
\varepsilon /\hslash \right) ^{2}}d\varepsilon , \label{gen}
\end{align}
where $\mu_{c}$ is chemical potential, $\gamma$ and $\Gamma$ are phenomenological intraband and interband scattering rates, respectively ($\tau=1/\gamma$ is the scattering time), $e$ is the charge of an electron, and $f_{d}\left( \varepsilon\right) =\left( e^{\left(\varepsilon-\mu_{c}\right) /k_{B}T}+1\right) ^{-1}$ is the Fermi-Dirac distribution. The first and second terms in the conductivity are due to intraband and interband contributions, respectively. For $k_{B}T\ll\left\vert \mu_{c}\right\vert,\hslash\omega$ \cite{GSC2007sum
\begin{equation}
\sigma \left( \omega \right) = \frac{ie^{2} \left\vert \mu_{c}\right\vert}{\pi \hslash\left(
\omega +i\gamma \right) } + \frac{ie^{2}
}{4\pi\hslash}\ln\left( \frac{2\left\vert \mu_{c}\right\vert -\left(
\omega+i\Gamma\right) \hslash}{2\left\vert \mu_{c}\right\vert +\left(
\omega+i\Gamma\right) \hslash}\right) . \label{T0}
\end{equation}
\noindent In the following we use (\ref{gen}) for $T=300$ K and (\ref{T0}) for $T=0$ K calculations. We consider a local (momentum independent) conductivity since the main effect considered here is the nontrivial DOS provided by the graphene plasmon energy dispersion.
The Drude form of the conductivity has been verified in the far-infrared \cite{Daw}-\cite{Kim}, and in the near infrared and visible the interband behavior has been verified in \cite{Li}. In the absence of scattering and bias the high-frequency optical conductivity is $\sigma=\sigma_\textrm{min}=e^2/4\hslash$, which has been verified in optical experiments \cite{Nair}.
Absorption is associated with both scattering and interband transitions. Since realistic values of $\Gamma$ will have a negligible effect on the results we will ignore interband scattering. For $\hslash \omega < 2\left\vert \mu_{c}\right\vert $ interband absorption is blocked, otherwise interband absorption will often dominate Re($\sigma$).
For the intraband term the value of $\gamma$ generally depends on temperature via phonon interactions, the method of growth/fabrication (e.g., epitaxial, chemical vapor deposition, exfoliation), the presence of impurities, and the presence of a substrate. Measured values of scattering times at room temperature ranged from a few fs \cite{Daw}-\cite{Choi} to several tens of fs \cite{Daw}-\cite{Tan} to several hundred fs ($\sim$ 0.35 ps \cite{Lee}-\cite{Kim}), and at low temperature scattering times on the order of a few ps have been measured (1.1 ps \cite{Li} and 5 ps \cite{Tan}). Short scattering times are usually associated with impurities and defects since the room-temperature electron-phonon scattering time is estimated to be a few ps \cite{Berger}. In the following we assume $\tau=5$ ps for $T=0$ and $\tau=0.35$ ps for room temperature results. We assume lossless non-dispersive dielectrics to focus on graphene's electrodynamic response rather than on the substrate response.
\section{Purcell factors}
The partial LDOS projected normal to the graphene surface, $\rho _{\text{LDOS}}=\left( 6/\pi \omega \right) \mathrm{Im}\left(
G_{zz}\left( \mathbf{r},\mathbf{r},\omega \right) \right) $, normalized
by the free-space value $\rho _{\text{LDOS}}^{0}=\omega ^{2}/\left( \pi
^{2}c^{3}\right) $ gives the Purcell factor \cite{Purcell} (i.e., the enhanced spontaneous emission factor of a single photon emitter
\begin{equation}
\mathrm{PF}=\frac{\rho _{\text{LDOS}}}{\rho _{\text{LDOS}}^{0}}=\frac{6\pi }{k_{0}^{3}
\mathrm{Im}\left( G_{zz}\left( \mathbf{r},\mathbf{r},\omega \right)
\right) .
\end{equation}
In the following we consider both suspended graphene, where vacuum exists on either side of the graphene sheet, and supported graphene on a dielectric layer. Figure \ref{fig:2}(a) shows the tunability of the Purcell factor \cite{H} at THz frequencies for a single suspended graphene layer over a range of chemical potentials at a distance of 10 nm from the graphene surface for $\mu_{c}/k_B T\gg 1$ and $\omega / \gamma \gg 1$ (at room temperature the results of Fig.~\ref{fig:2} will hold with minor quantitative changes). It is clear that in the low THz regime the LDOS and Purcell factor can be tuned considerably by an external bias. Figure \ref{fig:2}(b) shows the Purcell factor for supported graphene on a $d_s=10$ nm thin substrate having relative permittivity $\varepsilon_r=4$ (the approximate permittivity of ${\rm Si0}_{2}$). The presence of the substrate perturbs the SPP of the suspended graphene sheet, and clearly red-shifts the Purcell factor maximums (larger values of $\varepsilon_r$ would further red-shift the Purcell factor). As discussed in \cite{Hanson}, the presence of a substrate tends to confine the SPP mode, leading to higher attenuation as energy concentrates at the lossy graphene surface.
Figure \ref{fig:2}(c) shows the Purcell factor for two layers of graphene separated by a 10 nm, $\varepsilon_r=4$ substrate. This case closely resembles the result of Fig.~\ref{fig:2}(b), although the bottom graphene layer leads to a parallel-plate like waveguide structure \cite{Hanson_PP}, which tends to further concentrate energy in the dielectric, narrowing and shifting the PF peaks. Figure \ref{fig:2}(d) shows the Purcell factor as a function of position and frequency for supported graphene on a 10 nm, $\varepsilon_r=$4 substrate. Clearly, to be able to tune the QD resonance florescence the QD needs to be located sufficiently close to the graphene surface to strongly couple to the graphene SPP, due to the strong confinement of the SPP mode. However, one of the advantages of using graphene sheets is that this coupling is translationally invariant in the $x$ and $z$ directions.
\begin{figure}[t]
\centering
\subfloat[]{\includegraphics[width=0.23\textwidth]{2_a}}~
\subfloat[]{\includegraphics[width=0.23\textwidth]{2_b}}\vspace{0.02cm}
\subfloat[]{\includegraphics[width=0.23\textwidth]{2_c}}~
\subfloat[]{\includegraphics[width=0.25\textwidth]{2_d}}
\caption{ (a) The Purcell factor for suspended graphene in vacuum, (b) for graphene on a $d_s=10$ nm, $\varepsilon=4$ substrate, (c) for two layers of graphene separated by a 10 nm, $\varepsilon=4$ substrate, and (d) as a function of position ($y_0$ is the dot position) and frequency for supported graphene on a 10 nm, $\varepsilon=$4 substrate. }
\label{fig:2}
\end{figure}
\begin{figure}[b]
\centering
\subfloat[]{\includegraphics[width=0.23\textwidth]{3_a}}~
\subfloat[]{\includegraphics[width=0.23\textwidth]{3_b}}\vspace{0.1cm}
\subfloat[]{\includegraphics[width=0.25\textwidth]{3_c}}~
\subfloat[]{\includegraphics[width=0.25\textwidth]{3_d}}\vspace{0.02cm}
\caption{ Purcell factor for suspended graphene for three different chemical potentials at $T=0$ K (a) and $T=300$ K (b). (c)-(d) Real and imaginary part of graphene conductivity at $\mu_c=80$ meV normalized by $\sigma_{\textrm{min}}=e^2/4 \hslash =6.085\times 10^{-5}$ S. }
\label{fig:3}
\end{figure}
\begin{figure}[b]
\centering
\subfloat[]{\includegraphics[width=0.158\textwidth]{4_a}}~
\subfloat[]{\includegraphics[width=0.158\textwidth]{4_b}}~
\subfloat[]{\includegraphics[width=0.16\textwidth]{4_c}}
\caption{ Incoherent spectrum of the QD 10 nm above graphene, pumped at 25 THz with (a) $\mu_{c}=60\,$meV, (b) $\mu_{c}=80\,$meV, (c) $\mu_{c}=100\,$meV.}
\label{fig:4}
\end{figure}
Figure \ref{fig:3} shows the Purcell factor for suspended graphene for three different values of chemical potential at $T=0$ (a) and $T=300$ K (b). It can be seen that at the higher temperature the peaks are broadened due to higher absorption compared to the $T=0$ result. Figure \ref{fig:3}(c)-(d) show the conductivity for the $\mu_c=$80 meV case. As frequency increases from a low value, the Drude (intraband) term drops off at $\gamma$ and the interband contribution becomes more important, with a sharp transition in the real-part when $ \alpha=\hslash \omega / 2 \mu_c=$1. The imaginary part undergoes a cusp-discontinuity (for $T=0$ K) at $\alpha=$1. When Im$(\sigma)<0$, as occurs in the vicinity of the cusp, the TM SPP can not propagate (generally, when this occurs a TE SPP can propagate, although for a vertical dipole excitation the TE SPP will not be excited), and this is associated with the drop-off of the Purcell factor. The peak in the Purcell factor corresponds approximately to the frequency where $ \textrm{Im}(\sigma)=\textrm{Re}(\sigma)$.
\section{Spectrum of a driven Quantum Dot}
The incoherent spectrum is defined as
\begin{align}
S_{0}\left( \omega \right) = \lim_{t\rightarrow 0}\mathrm{Re}&\int_{0}^{\infty
}d\tau \left\langle \sigma _{+}\left( t+\tau \right) \sigma
_{-}\left( t\right) \right\rangle \nonumber \\
&- \left\langle \sigma _{+}\left( t\right)
\right\rangle \left\langle \sigma _{-}\left( t\right) \right\rangle e^{i\left( \omega _{L}-\omega \right) \tau }d\tau, \label{IS}
\end{align}
\noindent where the second term subtracts the coherent scattering from the pump field. The incoherent spectrum of the QD is shown in Figs.~\ref{fig:4}(a-c) for a Rabi frequency of 10 meV at 25 THz, assuming $T=0$ (see the Purcell factor of Fig.~\ref{fig:3}(a)); in all subsequent results the dipole moment is taken to be $30$ Debye. Note that the pump field will naturally be efficiently increased by the coupling to the SPP. It is evident that by changing the chemical potential of graphene the weights of the sidebands can be substantially changed. That is, the dominant peak of the incoherent spectrum can be shifted, for example, by varying the bias voltage on graphene. As Fig. 3 shows, by varying the bias we can shift the peak of the LDOS; when the LDOS peak aligns with the peak of one of the Mollow triplets, the corresponding triplet is enhanced. Commensurately, a small value of the LDOS at the position of a triplet peak suppresses that peak (due to closure of the plasmon decay channel). This is one of the key results of the paper: the quantum coupling between QDs and graphene can be profoundly influenced by simply changing the bias field.
Note that in all three cases in Fig.~\ref{fig:4} the exciton-laser detuning is zero ($\omega_{x}=\omega_{L}$) and it is solely the change in the LDOS with bias that is responsible for the significant spectrum tuning. As seen in Fig.~\ref{fig:3}(b), at T=300 K the peaks are broadened but are not significantly shifted compared to the $T=0$ case. Therefore, at room temperature the Mollow triplet can also be controlled. However, in this case since the peaks overlap more than for $T=0$ it is advantageous to choose the Rabi frequency (which controls the separation of the triplet's peaks) and chemical potential values (which control the Purcell factor peaks) to further separate the peaks to achieve similar control over the triplet as in the low temperature case.
The far-field detectable spectrum at position ${\bf r}_D$ is defined as \cite{Vlack}
$S_{p}\left( \mathbf{r}_D,\omega \right) = \frac{2} {\varepsilon_0} \vert \mathbf{d} \cdot \underline{\mathbf{G}}\left( \mathbf{r}_D, \mathbf{r}_d, \omega \right) \vert ^2 S_0 \left( \omega \right)$.
The factor that multiplies $S_0$ has some features in the vicinity of $\textbf{r}_D=\lambda_\textrm{SPP}$, but is otherwise dominated by the homogeneous-space part of the Green function and is fairly featureless, and so the Mollow triplet of the detectable spectrum will resemble $S_0$.
\begin{figure}[tbh]
\centering
\subfloat[]{\includegraphics[width=0.23\textwidth]{5_a}}
\subfloat[]{\includegraphics[width=0.23\textwidth,height=0.18\textwidth]{5_b}}\vspace{0.1cm}
\subfloat[]{\includegraphics[width=0.26\textwidth]{5_c}}
\subfloat[]{\includegraphics[width=0.26\textwidth]{5_d}}
\caption{ (a) Purcell factor for suspended graphene at $x=10\,$nm and $\mu_{c}=10\,$meV at $T=0\,$K and $T=300\,$K, (b) The incoherent spectrum of the QD above graphene for a Rabi frequency of 10 meV and pump resonance at $4$ THz, with and without including $\overline{n}$, (c)-(d) show the corresponding conductivity. }
\label{fig:5}
\end{figure}
In Fig.~\ref{fig:3} the considered values of $\mu_c$ lead to peaks of the Purcell factor in the range 20-30 THz. For these bias values and frequencies interband absorption is blocked, and the only damping of the conductivity is due to scattering. At several tens of THz, but for room temperature, $\mu_c /k_B T\gg 1$ is at least weakly satisfied, and the Purcell factor results do not change qualitatively, although enhanced absorption broadens the curves (Fig.~\ref{fig:3}(b)). Furthermore, at several tens of THz, even at room temperature the average photon number $\overline{n}$ is negligible, and so the zero-temperature bath approximation holds. However, for small-enough bias the Purcell factor peaks below a few THz, in which case both $\mu_c /k_B T\gg 1$ and $\omega / \gamma \gg 1$ are violated at room temperature. At frequencies of a few THz and for small bias, temperature plays an important role in both the Purcell factor via the graphene conductivity (interband transitions will not be not blocked, leading to enhanced absorption, and the location of PF peak blueshifts due to the second inequality being violated) and in the incoherent spectrum via the effect of the average photon number being non-negligible. To examine this effect, Figure \ref{fig:5}(a) shows the Purcell factor at $x=10\,$nm and $\mu_{c}=10\,$ meV at $T=0\,$K and $T=300\,$K. Figure \ref{fig:5}(b) shows the effect on the resonance fluorescence spectrum. Also shown is the effect of including $\overline{n}$ in the $T=300\,$K calculation, where it can be seen that the inclusion of the room temperature thermal bath is important at these low THz frequencies. Figure \ref{fig:4}(c)-(d) show the normalized conductivity; compared to the 300 K result in Fig.~\ref{fig:3}(c)-(d) for $\mu_c=80$ meV, here the intraband contribution is still important when the interband term becomes active, significantly perturbing the LDOS from the $T=0$ K case (the Drude fall-off is set by $\tau$ and the onset of interband absorption is set by $\mu_c$, and so these two effects can be independently controlled, although $\mu_c$ also governs the amplitude of the intraband contribution). Since we keep the same pump frequency for $T=0$ K and $T=300$ K, in the latter case the LDOS is relatively flat at the pump frequency.
\section{Conclusions}
In conclusion,
we have shown that the
Purcell effect and the
Mollow triplet of a two-level emitter can be tuned by varying the chemical potential of a nearby graphene layer.
We have modeled this effect using an exact Green function theory for
the LDOS and exploited a quantum master equation to model the quantum dynamics.
We have also demonstrated the important influence of temperature.
This novel QD-graphene system
allows considerable spectral control at the quantum level via altering an easily-assessable external parameters of the system.
\section*{Acknowledgements}
This work was supported by the Natural Sciences and
Engineering Research Council of Canada. We thank Ronchun Ge for useful discussions.
|
\section{Introduction}
Mutual information is fundamentally important for measuring statistical
dependence between variables \cite{Linfoot1957,Joe1989,Brillinger2007,Reshef2011},
and for quantifying information transfer by engineered or naturally
occurring communication systems \cite{Shannon1948,Brennan2012}.
Statistical analysis using mutual information has been particularly
influential in neuroscience \cite{BialekSpikes1999}, and is becoming
so in systems biology for studying the biomolecular signaling networks
used by cells to detect, process and act upon the chemical signals
they receive \cite{Cheong2011,Bowsher2012,Bowsher2013}.
It can, however, be challenging to estimate mutual information reliably
with available sample sizes \cite{Panzeri2012}, and difficult to
derive mutual information and capacity exactly using mechanistic
models of the `channels' via which signals are conveyed. Furthermore, connections between mutual information and better
known statistical procedures such as regression, and their associated dependence measures, are still poorly understood. In order to address these challenges, the relationship between mutual information and the error incurred by estimation (or `prediction') using the conditional mean is now receiving attention. The focus has been on minimum mean square estimation error or, more generally, its average across the elements of the vector being estimated \cite{Prelov2004,Guo2005,Guo2008}. Instead, we focus on
connections between mutual information and regression-based dependence measures, $\nu^{-1}$, that utilise the determinant of the second-moment matrix of the conditional mean prediction error. We examine convergence
properties as $\nu\rightarrow0$, and establish sharp lower bounds on
mutual information of the form $\mathrm{log}(\nu^{-1/2})$. The bounds are tighter than
lower bounds based on the Pearson correlation and ones derived using average mean square-error rate distortion arguments.
The \emph{mutual information} between 2 random vectors $X$ and $Z$, written
$I(X;Z),$ is the Kullback-Leibler divergence between their joint
distribution and the product of their marginal distributions \cite{CoverThomasBook2006}. Mutual information thus measures the divergence between the joint distribution of $(X,Z)$ and the distribution in which $X$ and $Z$ are independent but have the same marginals. $I(X;Z)$ has desirable properties
as a measure of statistical dependence: it satisfies (after monotonic
transformation) all 7 of R\'{e}nyi's postulates \cite{Renyi1959} for
a dependence measure between 2 random variables, and underlies the
recently introduced maximal information coefficient \cite{Reshef2011}
for detecting pairwise associations in large data sets. Importantly, mutual information captures nonlinear dependence and dependence arising from moments higher than the conditional mean.
A decision-theoretic interpretation is indicative of the broad applicability of mutual information as a summary measure of statistical dependence. It can be shown \cite{Bernardo1979} that $I(X;Z)$ is equal to the increase in expected utility from reporting the posterior distribution of either of the two random vectors based on observation of the other, compared to reporting its marginal distribution---for example, reporting $p(Z|X=x)$ instead of $p(Z)$. This holds when the utility
function is a smooth, proper, local score function---as appropriate
for scientific reporting of distributions as `pure inferences' \cite{BernardoSmithBayesianTheoryBook2000}---because the logarithmic score function is the only score function having all these properties.
In information theory, the supremum of $I(X;Z)$ over the set of allowed
input distributions $\mathcal{F}$, termed the \emph{information capacity}, equates to the maximal
errorless rate of information transmission over a noisy
channel when the channel is used for long times \cite{Shannon1948}.
The above discussion makes clear that, from a rich variety of perspectives, mutual information is fundamentally important for measuring statistical
dependence and for quantifying information
transfer by signaling and communication mechanisms.
Our general setting may be depicted as
\begin{eqnarray}
X \rightarrow Y \rightarrow Z,
\label{picture}
\end{eqnarray}
with $(X,Y,Z)$ a real-valued random vector. Here $Z$ is conditionally independent of $X$ given $Y$. The conditional distribution of $Z$ given $X$ is determined by some physical mechanism, whose `internal' variables are denoted by $Y$ in Eq. \ref{picture}. Such a mechanism is often termed a channel in information theory, although we do not restrict attention to signaling and communication channels here. Often we have in mind situations where $X$ causes $Z$ but not \emph{vice versa}, and the conditional distribution of $Z$ given $X$ does not depend on the experimental `regime' giving rise to the distribution of $X$ \cite{Dawid2010}. There is always an asymmetry between $X$ and $Z$ in the general setting we consider. We term $X$ the \emph{input} or \emph{treatment} because its marginal distribution can in principle be any distribution (although we may wish to restrict attention to particular classes thereof). In contrast, the \emph{output} or \emph{response} $Z$ is the realisation of the mechanism given the input $X$. In general, not all marginal distributions for $Z$ can be obtained for a given mechanism by appropriate choice of the marginal of $X$. When analysing the probabilistic properties of physical models of mechanisms, the distribution (or set of distributions) for the input $X$ is given, but the marginal distribution of $Z$ is often unknown. In experimental settings, the input distribution is taken not to affect the conditional distribution of $Z$ given $X$, or else can sometimes be directly manipulated.
Examples of our general setting include experimental design with $X$ as the treatment and $Z$ the response of interest; and signaling or communication channels with $X$ as the input signal and $Z$ its noisy representation. An example of a scientific area of application is the current effort to understand the biomolecular signaling mechanisms
used by living cells to relay the chemical signals
they receive from their environment \cite{Cheong2011}. Here, the interest is both in understanding why some biomolecular mechanisms perform better than others, and in measuring experimentally in the laboratory the mutual information between $X$ and $Z$ or the information capacity. Broadly speaking, the first involves deriving dependence measures between $X$ and $Z$ for different stochastic mechanisms (given a particular input distribution). The second might involve, for example, nonparametric estimation of the mutual information between the concentration of a chemical treatment applied to the cells and the level of an intracellular, biochemical output.
We note, however, that the formal statements of our results do not require any particular interpretation of $X$ and $Z$. Rather, the general setting just described motivates the results and places them in context. A sequential reading of Equations 2 to 12 inclusive provides a convenient preview of our theoretical results establishing lower bounds on mutual information and information capacity.
\section{Setup and Notation}
For random vectors $X$ and $Z$,
we define
$\nu(Z|X)=\mathbb{\mathrm{\det}\left(E\mathit{\left\{ \mathbb{V}\left[Z|X\right]\right\} }\right)}\mathrm{/\det}\left(\mathbb{V}\left[\mathit{Z}\right]\right)$, where $\mathbb{V}$ denotes a covariance matrix.
In general, $\nu(Z|X)$ is not equal
to $\nu(X|Z)$. For 2 scalar random variables, $\nu(Z|X)$ is equal
to the minimum mean square estimation error or minimum MSEE for estimation
of $Z$ using $X$, normalised by the variance of $Z$ (since $\mathbb{E}\left[Z|X\right]$
is the optimal estimator). We denote the optimal estimation
or `prediction' error by $e(Z|X)=Z-\mathbb{E}\left[Z|X\right]$. In general, $\nu(Z|X)$ is the ratio of the determinant of the second-moment matrix of the
error $e(Z|X)$ and the determinant of the variance matrix of $Z$, that is
\begin{eqnarray}
\nu(Z|X)=\frac{ \mathbb{\mathrm{\det}\left(E\mathit{\left[e(Z|X)e(Z|X)^{\mathrm{T}}\right]}\right)} } {\mathrm{\det}\left(\mathbb{V}\left[\mathit{Z}\right]\right)}.
\end{eqnarray}
We will show that $\nu(Z|X)^{-1}$
provides a generalised measure of `signal-to-noise', applicable to
non-Gaussian settings, that relies on first and second conditional
moments of $Z$ given $X$ (via the law of total variance) rather than on all features of the joint distribution. We make few assumptions about the conditional density describing the mechanism or channel $f(z|x)$, except that the conditional mean $m(x)=\mathbb{E}\left[Z|X=x\right]$ is an invertible, continuously differentiable function of $x$.
A central result of the paper (see Theorem \ref{thm:ZgXlwrbnd} and Corollary \ref{rdarg}) is then that
\begin{equation}
I(X;Z)\geq\mathrm{\log}\left\{\nu(Z|X)^{-1/2}\right\}
\geq \frac{-d_Z}{2}\mathrm{\log}\left\{ \frac{d_{Z}^{-1}\mathrm{tr}(\mathbb{E}\{\mathbb{V}[Z|X]\})}{\mathrm{[\det}(\mathbb{V}[Z])]^{d_{Z}^{-1}}}\right\} ,
\label{main}
\end{equation}
where all terms are evaluated under the joint density for $(X,Z)$ implied by the channel $f(z|x)$ and a Gaussian density for the transformed input, $m(X)$. Here $d_Z$ is the dimension of the vector $Z$. The second term in Eq. \ref{main} is our lower bound utilising the determinant of the second-moment matrix of the prediction error of the conditional mean $\mathbb{E}[Z|X]$, while the third term instead utilises the average mean square error of that conditional mean. We discuss the relation of the third term to rate distortion arguments later in the paper. Notice that characterising the first and second conditional moments of the mechanism, $\mathbb{E}[Z|X]$ and $\mathbb{V}[Z|X]$, is enough (via the law of total variance) to evaluate the lower bound $\mathrm{\log}\left\{\nu(Z|X)^{-1/2}\right\}$ for a given Gaussian distribution of $m(X)$. Maximising the bound over such distributions then also yields a useful bound on the information capacity.
As a first step in analysing connections between mutual information and our regression-based measures,
we explore the relationship between the convergence to zero of $\nu(Z|X)$
or $\nu(X|Z)$, and the convergence of mutual information. For simplicity,
we analyse the bivariate case where the variable $X$ has finite support, for
example a finite collection of treatment concentrations in a cell
signaling experiment. We write $I$ for mutual information,
$H$ for discrete entropy and $h$ for differential entropy.
\section{Convergence properties}
\begin{theorem}
\label{thm:ZgXasymp}Let $(X_{n},Z_{n})$ be a sequence of pairs of real-valued
random variables, with the support of $X_{n}$ given by a finite set $\mathcal{X}_{n}$
($|\mathcal{X}_{n}|\geq2$ and bounded above by a constant $\forall n$).
Write $m_{n}(X_{n})$ for the function $\mathbb{E}[\breve{Z}_{n}|X_{n}]$,
where $\breve{Z}_{n}\triangleq Z_{n}\mathbb{V}[Z_{n}]^{-\frac{1}{2}}.$
Denote its support by $\mathcal{M}_{n} = \{m_{n}(x);x\in\mathcal{X}_{n}\}\subset\mathbb{R}$.
Let $\epsilon_{n}^{\ast}$ be 1/2 of the minimum distance between
any two points in $\mathcal{M}_{n}$ and $\epsilon^{\ast}\triangleq\inf\{\epsilon_{n}^{\ast}\}.$
Suppose that:
(1) $ \nu(Z_{n}|X_{n})\rightarrow0$ as $n\rightarrow\infty$; (2) the functions $m_{n}$ are one-to-one mappings from $\mathcal{X}_{n}$
to the real line; and (3) $\epsilon^{\ast}>0$, so that any pair of
points in a support $\mathcal{M}_{n}$ are separated by at least $2\epsilon^{\ast}$.
Then as $n\rightarrow\infty$ , $H(X_{n}|Z_{n})\rightarrow0$ and
therefore $H(X_{n})-I(X_{n};Z_{n})\rightarrow0$.
\end{theorem}
In Theorem \ref{thm:ZgXasymp}, the response variable $Z$ is real-valued
and can, for example, be either a continuous or discrete random variable.
Theorem \ref{thm:ZgXasymp} establishes the convergence of the mutual
information $I(X;Z)$ to the entropy of $X$ as $\nu(Z|X)\rightarrow0$, under the condition that the conditional mean $\mathbb{E}[Z|X]$ is
an invertible function of $X$. By definition, $I(X;Z)$ cannot be greater than the entropy of $X$.
The intuition for the result in Theorem \ref{thm:ZgXasymp} is that the convergence of $\nu(Z|X)$ to zero enables construction
of a point estimator of $X$ (utilising the conditional expectation
$\mathbb{E}[Z|X]$) whose performance becomes `perfect' in this limit. The condition requiring invertibility of the conditional mean function would be needed even in the case where $Z$ is a deterministic function of $X$, otherwise $I(X;Z) \leq H(Z) < H(X)$.
We have rescaled the output $Z_{n}$
so that $\mathbb{V}[\breve{Z}_{n}]$ is constant at $1$ for all $n$.
In particular, Theorem \ref{thm:ZgXasymp} does \textit{not} require
the minimum mean square estimation error, $\mathbb{E}\{(\mathit{Z-\mathbb{E}[Z|X]})^{2}\},$
to converge to zero as $n\rightarrow\infty$. A physical example where the minimum MSEE diverges but Theorem \ref{thm:ZgXasymp} applies is given by a molecular signaling system, with $Z$ the number of output molecules, which is operating in the macroscopic (or large system size) limit where the dynamics of chemical concentrations are deterministic, conditional on the input $X$.
\begin{proof}
We have that $\nu(Z_{n}|X_{n})=\nu(\breve{Z}_{n}|X_{n})\rightarrow0$.
Since $\nu(\breve{Z}_{n}|X_{n})=\mathbb{E}\{\mathbb{V}[\breve{Z}_{n}|X_{n}]\}$
$=\mathbb{E}\{(\mathit{\breve{Z}_{n}-\mathbb{E}[\breve{Z}_{n}|X_{n}]})^{2}\}$,
it follows that $\breve{Z}_{n}-\mathbb{E}[\breve{Z}_{n}|X_{n}]$ converges
to zero in mean square (in $L^{2}$) and therefore $\breve{Z}_{n}-\mathbb{E}[\breve{Z}_{n}|X_{n}]\rightarrow^{pr}0$
(where $\rightarrow^{pr}$ denotes convergence in probability). Since
$H(X_{n}|\breve{Z}_{n})=H(X_{n}|Z_{n})$ $=H(X_{n})-I(X_{n},Z_{n}),$
we must establish that $H(X_{n}|\breve{Z}_{n})\rightarrow0$ as $n\rightarrow\infty$.
Consider estimating $X_{n}$ based on observation of $\breve{Z}_{n}$
using the estimator, $\widehat{X}_{n}(\breve{Z}_{n})$, which is equal
to the a point $x\in\mathcal{X}_{n}$ that minimises the distance
$|\breve{Z}_{n}-m_{n}(x)|$. Condition (3) applies not to $\mathbb{E}[Z_{n}|X_{n}=x]$
but to $\mathbb{E}[\breve{Z}_{n}|X_{n}=x].$ Let $z\in\mathbb{R},m_{n}\mathcal{\in M}_{n}$
and notice that if $|\mathit{z-}m_{n}|<\epsilon^{*},$ then $\mathit{|z-}m_{n}|<\epsilon_{n}^{*}<\mathit{|z-}m_{n}'|$,
that is $m_{n}$ is closer to $z$ than is any other point $m_{n}'$
in $\mathcal{M}_{n}$. Therefore, if $|\mathbb{\mathit{\breve{Z}_{n}-}E}[\breve{Z}_{n}|X_{n}]|<\epsilon^{*},$
then $\mathbb{E}[\breve{Z}_{n}|X_{n}]$ is located at the unique point
$m_{n}\mathcal{\in M}_{n}$ that minimises $|\breve{Z}_{n}-m_{n}|$
and, using condition (2), $\widehat{X}_{n}(\breve{Z}_{n})=X_{n}$,
that is, the estimator recovers $X_{n}$ without error$.$ The probability
of estimation error, $p_{\mathrm{error}},$ satisfies
\[
\mathbb{\mathit{p\mathrm{_{error}}=}P\left\{ \mathit{\hat{X}_{n}\left(\breve{Z}_{n}\right)\neq X_{n}}\right\} }\leq\mathbb{P}\left\{ |\mathbb{\mathit{\breve{Z}_{n}-}E}[\breve{Z}_{n}|X_{n}]|\geq\epsilon^{*}\right\} ,
\]
and therefore $p_{\mathrm{error}}\rightarrow0$ as $n\rightarrow\infty.$
Fano's Inequality \cite{CoverThomasBook2006} then implies that
$H(X_{n}|\breve{Z}_{n})\rightarrow0$ as required.
\end{proof}
We have thus shown in the context of Theorem \ref{thm:ZgXasymp} that the limit $\nu(Z|X)\rightarrow0$
characterises a regime of large signal-to-noise for the input
$X$, without the need to impose further conditions on the joint distribution of $X$ and $Z$. In \cite{Prelov2004}, the authors consider the opposite regime of low
signal-to-noise for non-linear channels with additive Gaussian noise.
They obtain an asymptotic expansion of the mutual information, $I(X;Z)$, whose leading term
is a decreasing function of a variable which, in their setting, is
equal to $\nu(Z|X)$. In Theorem \ref{thm:XgZasymp} below, we consider
the case where the conditioning is on the response $Z$ instead of on the input $X$ (see Section \ref{depreg} for further discussion of regression on the response variable).
\begin{theorem}
\label{thm:XgZasymp}Let $(X_{n},Z_{n})$ be a sequence of pairs of real-valued
random variables, with the support of $X_{n}$ given by a finite set $\mathcal{X}_{n}$
($|\mathcal{X}_{n}|\geq2$ and bounded above by a constant $\forall n$).
Define $\breve{X}_{n}\triangleq X_{n}\mathbb{V}[X_{n}]^{-\frac{1}{2}}$,
with support \textup{$\mathcal{\breve{X}}_{n}$, }and let $\epsilon_{n}^{\ast}$
be 1/2 of the minimum distance between any two points in $\breve{\mathcal{X}}_{n}$.
Suppose that $\epsilon^{\ast}\triangleq\inf\{\epsilon_{n}^{\ast}\}>0.$
If $\nu(X_{n}|Z_{n})\rightarrow0$ as $n\rightarrow\infty$, then
$H(X_{n}|Z_{n})\rightarrow0$ and $H(X_{n})-I(X_{n};Z_{n})\rightarrow0$. \end{theorem}
\begin{proof}
Given in the Appendix, using an argument similar to the proof of Theorem \ref{thm:ZgXasymp}.
\end{proof}
\noindent Again, the mutual
information $I(X;Z)$ converges to the entropy of $X$, this time as $\nu(X|Z)\rightarrow0$.
(We note in passing that, if both $X$ and $Z$ have finite support, a corollary
of Theorem \ref{thm:XgZasymp} is that: either $H(X_{n})-H(Z_{n})\rightarrow0,$
or $\nu (X_{n}|Z_{n})$ and $\nu (Z_{n}|X_{n})$
do not simultaneously converge to zero).
Theorems \ref{thm:ZgXasymp} and \ref{thm:XgZasymp} establish
connections between regression-based measures of dependence $\nu ^{-1}$ and mutual information,
without imposing strong assumptions about the joint distribution of
$(X,Z).$ We conjecture that similar theorems will hold for random variables
$X$ and $Z$ having a joint density with respect to Lebesgue measure. These theorems indicate the possibility, explored below, of lower bounding the mutual information using $\nu ^{-1}$.
A consequence of our Theorem \ref{thm:ZgXlwrbnd}
is that, for the general multivariate, absolutely continuous case,
$I(X;Z)\rightarrow\infty$ when $\nu(Z|X)\rightarrow 0$
(and $\mathbb{E}[Z|X=x]$
is an invertible mapping), and $\nu$ is evaluated under the appropriate
marginal distribution for the input $X$.
\section{Lower bounds on mutual information and capacity}
Our aim in this and the subsequent section is to establish lower bounds on mutual information, $I(X;Z)$,
for an absolutely continuous random vector $(X,Z)$ with finite dimension
$d\geq2$. These bounds will hold under certain marginal distributions for the input $X$, and also provide lower bounds on the information capacity. We first show that the following Lemma holds when the marginal
distribution of $X$ is normal. The Lemma relates the mutual information of $X$ and $Z$ to their variances and covariance.
\begin{lem}
\label{lem:lbndCorr}Let $(X,Z)$ be a random vector in $\mathbb{R^{\mathrm{\mathit{d}}}}$,
$d\geq2$, with joint density $f(x,z)$ with respect to Lebesgue
measure. Suppose the density of $X$, $f(x),$ is multivariate normal and that
$(X,Z)$ has finite variance matrix under $f$. Then
\begin{eqnarray}
I_{f}(X;Z) &\geq& \mathrm{\mathrm{\log}\left\{ \frac{\mathrm{\det}(\Sigma_{\mathit{XX}})}{det(\Sigma_{\mathit{XX}}-\Sigma_{\mathit{XZ}}\Sigma\mathit{_{ZZ}^{-\mathrm{1}}}\Sigma_{\mathit{ZX}})}\right\}_{\mathit{f}} ^{1/2}} \label{eq:lbndCorr} \\
&=&\mathrm{\log}\left\{ \frac{\mathrm{\det}(\Sigma_{\mathit{ZZ}})}{\mathrm{det}(\Sigma_{\mathit{ZZ}}-\Sigma_{\mathit{ZX}}\Sigma\mathit{_{XX}^{-\mathrm{1}}}\Sigma_{\mathit{XZ}})}\right\}_{\mathit{f}} ^{1/2} \notag
\end{eqnarray}
where, for example, $\Sigma_{ZX}=\mathrm{Cov}(X,Z)=\mathbb{E}\{(X-\mathbb{E}[X])(Z-\mathbb{E}[Z])^{\mathrm{T}}\},$
and subscript $\mathit{f}$ indicates that the mutual information and the covariance matrices are those under
the joint density $f(x,z).$\end{lem}
\begin{proof}
Given in the Appendix.
\end{proof}
When $d=2$ (and, again, $f(x)$ is normal), Lemma \ref{lem:lbndCorr}
simplifies to give the lower bound \[I(X;Z)\geq\mathrm{log\{[1-Corr^{2}(\mathit{X,Z})]^{-1/2}\}},\]
where $\mathrm{Corr}$ denotes the Pearson correlation. This inequality shows how to relate perhaps the best known measure of association to mutual information. An outline
proof is given for the $d=2$ case in \cite{Mitra2001}.
The multivariate,
$d>2$, case has not to our knowledge appeared previously, although
it is a reasonably straightforward generalisation. We provide a complete
proof for $d\geq2$ in the Appendix.
\begin{rem}
The lower bounds in Eq. \ref{eq:lbndCorr} are the mutual information
of a multivariate normal density, $g$, with identical covariance
matrix to that of $f,$ the true joint density of $X$ and $Z$. We want to obtain a lower bound for $I_{\mathit{f}}(X;Z)$
in terms of $\nu_{f}(Z|X).$ It might appear that one way to do so would
be to adopt a similar strategy, attempting to show that $I_{\mathit{f}}(X;Z)$
is bounded below by the mutual information of a multivariate normal,
$g$, having identical $\nu(Z|X)$ to $f$ (and with the marginal
density $g(x)=f(x)$). However, this strategy fails. Although $I_{g}(X;Z)$
still depends only on $\nu(Z|X),$ the equality $\mathbb{E_{\mathit{f}}}\{\mathrm{log}[g(X,Z)/f(X)g(Z)]\}=\mathbb{E_{\mathit{g}}}\{\mathrm{log}[g(X,Z)/f(X)g(Z)]\}$
no longer holds in general (see Eq. \ref{eq:careMS-1} and subsequent argument in the Appendix).
\end{rem}
Our general strategy to obtain lower bounds on mutual information and capacity is as follows. First, we specify a channel with suitable `pseudo-output' and Gaussian `pseudo-input'. These are transformed versions of $Z$ and $X$ respectively (sometimes we transform $X$ alone). The transformations are given by the conditional mean functions: for example, the pseudo-input may be $m(X) = \mathbb{E}[Z|X]$, as in Theorem \ref{thm:ZgXlwrbnd} below. Second, we apply Lemma \ref{lem:lbndCorr} with the pseudo-input and pseudo-output in place of $X$ and $Z$ there. We then
make use of the following relationship that holds for any random vector
$(U,W)$:
\begin{eqnarray}
\mathrm{Cov}(U,\mathbb{E}[U|W]) & = & \mathbb{E}\left\{\mathbb{E}\left[(U-\mathbb{E}[U])(\mathbb{E}[U|W]-\mathbb{E}[U])^{\mathrm{T}}|W\right]\right\} \label{eq:CovVE} \\
& = & \mathbb{V}\{\mathbb{E}[U|W]\}. \notag
\end{eqnarray}
We now use this strategy to obtain a lower bound for the mutual information, $I_{\mathit{f}}(X;Z)$,
in terms of $\nu_{f}(Z|X).$
\begin{theorem}
\label{thm:ZgXlwrbnd}Let $(X,Z)$ be a random vector in $\mathbb{R^{\mathrm{\mathit{d}}}},d\geq2$.
Consider the conditional density (with respsect to Lebesgue measure), $f(z|x)$,
and suppose that $m(x)=\mathbb{E}[Z|X=x]$ is a one-to-one, continuously
differentiable mapping (whose domain is an open set and a support
of $X$). Let $m(X)$ be normally distributed and denote by $f(x)$ the implied density of $X$. Then
\begin{equation}
I_{f}(X;Z)\geq\mathrm{\log}\left\{\nu_{f}(Z|X)^{-1/2}\right\},\label{eq:ZgXlwrbound}
\end{equation}
where we assume moments are finite such that $\nu_{f}(Z|X)$ is well defined
and that $\mathbb{E}\mathit{\left\{ \mathbb{V}\left[Z|X\right]\right\} }$
is non-singular under the joint density, $f$. This lower bound is sharp since $I_{f}(X;Z)=\log\left\{\nu_{f}(Z|X)^{-1/2}\right\}$
when $f(x,z)$ is a multivariate normal density. Furthermore, the information capacity $C_{Z|X}$ satisfies
\begin{equation}
C_{Z|X}\geq\mathrm{\log}\left\{\nu_{f}(Z|X)^{-1/2}\right\}, \label{capacitybnd}
\end{equation}
provided $f(x)$ is the density of an allowed input distribution, that is a distribution in $\mathcal{F}$.
\end{theorem}
\begin{proof}
Consider the mechanism $M\rightarrow X\rightarrow Z$, where $X=m^{-1}(M)$
and $X$ is then applied to the channel $f(Z|X).$ Here $M=m(X)$ is the transformed or pseduo-input. Notice that $\mathbb{E}[Z|M]=\mathbb{E}[Z|X]=M$ because $\sigma(X) = \sigma(M)$,
and therefore $\mathrm{Cov}(M,Z)=\mathrm{Cov}(Z,\mathbb{E}[Z|M])=\mathbb{V}\{\mathbb{E}[Z|M]\}=\mathbb{V}[M]$,
by Eq. \ref{eq:CovVE} (and using twice that $\mathbb{E}[Z|M]=M$). Under the conditions of Theorem
\ref{thm:ZgXlwrbnd}, $M$ has a Gaussian distribution, with the distribution
of $X$ that implied by the mapping $X=m^{-1}(M).$ Given the properties
of the one-to-one mapping $m(x)$, we also have that $I_{f}(M;Z)=I_{f}(X;Z)$ (see, for example, \cite{Kraskov2004}).
Applying Lemma \ref{lem:lbndCorr} to the Gaussian, pseudo-input $M$ and the output $Z$ yields
\begin{eqnarray*}
I_{f}(X;Z)\geq\frac{1}{2}\mathrm{\log}\left\{ \frac{\mathrm{\det}(\mathbb{V}[Z])}{\mathrm{det}(\mathbb{V}[Z]-\mathrm{Cov}(Z,M)\mathbb{V}[M]^{-1}\mathrm{Cov}(Z,M)^{\mathrm{T}})}\right\}_{\mathit{f}} \\
=\frac{1}{2}\mathrm{\log}\left\{ \frac{\mathrm{\det}(\mathbb{V}[Z])}{\mathrm{det}(\mathbb{V}[Z]-\mathbb{V}\{\mathbb{E}[Z|M]\})}\right\}_{\mathit{f}} ,
\end{eqnarray*}
where the second line again uses $\mathrm{Cov}(Z,M)=\mathbb{V}[M]=\mathbb{V}\{\mathbb{E}[Z|M]\}$. The result then follows directly because $\mathbb{V}[Z]=\mathbb{V}\{\mathbb{E}[Z|M]\}+\mathbb{E}\{\mathbb{V}[Z|M]\},$
and because we have that $\mathbb{V}[Z|M]=\mathbb{V}[Z|X]$ since $\sigma(M)=\sigma(X).$
Eq. \ref{capacitybnd} follows from Eq. \ref{eq:ZgXlwrbound} and the definition of capacity as the supremum of mutual information over the collection of allowed input distributions, $\mathcal{F}$.
\end{proof}
Notice that characterising the first and second conditional moments of the mechanism, $\mathbb{E}[Z|X]$ and $\mathbb{V}[Z|X]$, is enough (using the law of total variance) to evaluate the lower bound $\mathrm{\log}\left\{\nu(Z|X)^{-1/2}\right\}$ for a given Gaussian distribution of $m(X)$. Maximising the bound over such distributions then yields the largest lower bound on the information capacity. The approach is applicable when
experimental data have been generated under some other
input distribution, provided the first and second conditional moments are carefully estimated.
We now discuss the relationship of our lower bounds in Eqs. \ref{eq:ZgXlwrbound} and \ref{capacitybnd} to average mean-square error, rate distortion-type arguments with a Gaussian source.
\begin{cor}
Let $d_{Z}$ be the dimension of $Z$ and suppose that the conditions of Theorem \ref{thm:ZgXlwrbnd} apply. It follows from Eq. \ref{eq:ZgXlwrbound},
scaling by the reciprocal of $d_{Z}$, that
\begin{eqnarray}
d_{Z}^{-1}I_{f}(X;Z)\geq\frac{1}{2}d_{Z}^{-1}\mathrm{\log}\{\nu_{f}(Z|X)^{-1/2}\} & = & \frac{1}{2}\mathrm{\log}\left\{ \frac{\mathrm{[\det}(\mathbb{V}[Z])]^{d_{Z}^{-1}}}{[\mathrm{det}(\mathbb{E}\{\mathbb{V}[Z|X]\})]^{d_{Z}^{-1}}}\right\}_{\mathit{f}} \nonumber \\
\geq & & \frac{1}{2}\mathrm{\log}\left\{ \frac{\mathrm{[\det}(\mathbb{V}[Z])]^{d_{Z}^{-1}}}{d_{Z}^{-1}\mathrm{tr}(\mathbb{E}\{\mathbb{V}[Z|X]\})}\right\}_{\mathit{f}} ,\label{eq:rdtype}
\end{eqnarray}
since $\mathbb{E}_{f}\{\mathbb{V}[Z|X]\}$ is positive definite, where $\mathit{f}$ indicates evaluation under the joint density $f(x,z)$. We have used that $\mathrm{0<[\det}(M)]^{1/m}\leq m^{-1}\mathrm{tr}(M)$
for a positive definite, $m\times m$ matrix, $M$ \cite{CoverThomasBook2006}.
\label{rdarg}
\end{cor}
Notice that $d_{Z}^{-1}\mathrm{tr}(\mathbb{E}\{\mathbb{V}[Z|X]\})$
is the average minimum MSEE given $X$, the average being across the scalar components of the vector
$Z$.
Eq. \ref{eq:rdtype} establishes that our lower bound, $\mathrm{\log}\{\nu_{f}(Z|X)^{-1/2}\}$,
is tighter than the lower bound based on the average minimum MSEE, $\frac{1}{2}\mathrm{\log\{\mathrm{\det}(\mathbb{V}[Z])}^{d_{Z}^{-1}}/$ $d_{Z}^{-1}\mathrm{tr}(\mathbb{E}\{\mathbb{V}[Z|X]\})\}_{\mathit{f}}$. The two lower bounds clearly coincide for the bivariate case, $d=2$. The bound based on the average minimum MSEE might, at first sight, appear to have the form that would be obtained by a rate distortion-type argument \cite[Section 4.5.2]{BergerBook} with $Z$ as the Gaussian `source'. However, this is not the case because the bound would in general be evaluated under a joint density for $(X,Z)$ different than $\mathit{f}$. (However, see Lemma \ref{thm:XgZlwrbnd} and the subsequent discussion for the case of regression on the response variable). Notice also that in our setting of Eq. \ref{picture}, we may not be able to adjust the input distribution in order to obtain a Gaussian marginal for $Z$.
Instead, one might treat $M=\mathbb{E}[Z|X]$ as the Gaussian source in a rate distortion-type argument. Consider the case $d=2$. One obtains the result $I_{f}(X;Z)\geq\frac{1}{2}\mathrm{log}(\mathbb{V}\{\mathbb{E}[Z|X]\}/\mathbb{E}\{\mathbb{V}[Z|X]\})_{f}$,
where the numerator is the variance of the source and the denominator
is the expected square-error distortion between the source and its
estimate, here $Z$. The right-hand side of this inequality is strictly
less than our lower bound, $\mathrm{\log}\{\nu_{f}(Z|X)^{-1/2}\}=\frac{1}{2}\mathrm{log}(\mathbb{V}[Z]/\mathbb{E}\{\mathbb{V}[Z|X]\})_{f},$
since $\mathbb{V}[Z]-\mathbb{V}\{\mathbb{E}[Z|X]\}=\mathbb{E}\{\mathbb{V}[Z|X]\}>0$.
An analogous argument applies to the case $d>2$, since $\mathrm{\det}(\mathbb{V}[Z]) > \mathrm{\det}($ $\mathbb{V}\{\mathbb{E}[Z|X]\})$. (For the case where $X$ itself is Gaussian, see Lemma \ref{thm:XgZlwrbnd}).
We conclude that
our lower bounds in Eqs. \ref{eq:ZgXlwrbound} and \ref{capacitybnd} are tighter than lower bounds derived using average mean-square error rate distortion-type arguments with a Gaussian source.
\section{\label{depreg} Regression on the response variable}
We have so far considered lower bounds on mutual information that rely on the error in estimation of the response variable $Z$, using the conditional mean of $Z$ given $X$. In this section we instead consider lower bounds that utilise the regression of $X$ on the response variable, $Z$.
Using the novel proof strategy adopted for Theorem \ref{thm:ZgXlwrbnd}, we can show that the lower bound $\mathrm{\log}\{\nu_{f}(\tilde{X}|Z)^{-1/2}\}$ improves upon the one based on the Pearson correlation, $\mathrm{log\{[1-Corr_{\mathit{f}}^{2}(\mathit{\tilde{X},Z})]^{-1/2}\}}$, where $\tilde{X}$ is the result of transforming the input to have a Gaussian marginal distribution. We can thus always (weakly) improve upon the lower bound for the bivariate case given by Lemma \ref{lem:lbndCorr}. Intuitively, the improvement arises because $\nu_{f}(\tilde{X}|Z)^{-1}$ captures dependence from non-linearity in the conditional mean, $\mathbb{E}[\tilde{X}|Z]$, whereas the Pearson correlation does not.
\begin{theorem}
\label{thm:CfMS}Let $(X,Z)$ be a random vector in $\mathbb{R^{\mathrm{\mathit{2}}}}$
with joint density $f(x,z)$ with respect to Lebesgue measure. Suppose
there exists a one-to-one mapping $s:X\rightarrow\tilde{X}$ (whose
domain is an open set and a support of $X$) such that $\tilde{X}$
has a Gaussian density. We assume that the mapping $s$ is continuously
differentiable with derivative that is everywhere non-zero, and that
$\nu_{f}(\tilde{X}|Z)^{-1/2}$ exists. Then\textup{
\begin{equation}
I_{f}(X;Z)\geq\mathrm{\log}\{\nu_{f}(\tilde{X}|Z)^{-1/2}\}\geq\mathrm{log}\{[1-\mathrm{Corr_{\mathit{f}}^{2}}(\tilde{X},Z)]^{-1/2}\},\label{eq:CfMS}
\end{equation}
}where the third term is the lower bound on $I_{f}(\tilde{X};Z)=I_{f}(X;Z)$
given by Lemma \ref{lem:lbndCorr} in the case $d=2$. We assume
$\mathrm{Corr}_{\mathit{f}}(\mathit{\tilde{X}},Z)$ is well defined and less than $1$. Subscript $f$ indicates evaluation under the joint distribution implied by $f(x,z)$.
\end{theorem}
As we show in the proof below, the lower
bound $\mathrm{\log}\{\nu_{f}(\tilde{X}|Z)^{-1/2}\}$ can be understood
as the result of first transforming $Z$ to the `pseudo-output' $t(Z)=\mathbb{E}[\tilde{X}|Z],$
and then basing the bound on $\mathrm{Corr_{\mathit{f}}^{2}}(\tilde{X},t(Z))$,
using Lemma \ref{lem:lbndCorr}. We then establish
that using another (measurable) transformation,
$t(Z)$, cannot yield a greater bound (given some choice of the Gaussian
variable $\tilde{X}$). This includes, in particular, the lower bound
based on the squared correlation of $\tilde{X}$ and $Z$ itself.
Notice that we cannot construct a lower bound using the maximal correlation
of $(X,Z)$ \cite{Gebelein1941,Renyi1959}, because the implied transformation
of $X$ need not result in the Gaussian distribution needed to apply
Lemma \ref{lem:lbndCorr}.
\begin{proof}
Given the properties of the one-to-one mapping $s(x)$, we have that
$I_{f}(X;Z)=I_{f}(\tilde{X};Z)$.
Consider the mechanism $\tilde{X}\rightarrow X \rightarrow Z\rightarrow\mathbb{E}[\tilde{X}|Z]$,
in which we first transform $\tilde{X}$ to $X$ and then apply $X$
to the `channel' $f(Z|X).$ The pseudo-output here is $\mathbb{E}[\tilde{X}|Z]$. By the data processing inequality, $I_{f}(\tilde{X};Z)\geq I_{f}(\tilde{X};\mathbb{E}[\tilde{X}|Z]).$
Applying Lemma \ref{lem:lbndCorr} to the Gaussian pseudo-input $\tilde{X}$ and the pseudo-output $\mathbb{E}[\tilde{X}|Z]$ yields
\[
I_{f}(\tilde{X};\mathbb{E}[\tilde{X}|Z])\geq\frac{1}{2}\mathrm{\mathrm{\log}\{[1-\mathrm{Corr^{2}}(\tilde{X},\mathbb{E}[\tilde{X}|Z])]^{-1/2}}\}.
\]
By Eq. \ref{eq:CovVE}, $\mathrm{Cov}(\tilde{X},\mathbb{E}[\tilde{X}|Z])=\mathbb{V}\{\mathbb{E}[\tilde{X}|Z]\}.$
Hence $\mathrm{Corr^{2}}(\tilde{X},\mathbb{E}[\tilde{X}|Z])=\mathbb{V}\{\mathbb{E}[\tilde{X}|Z]\}$ $/\mathbb{V}[\tilde{X}]$,
and
\[
I_{f}(\tilde{X};\mathbb{E}[\tilde{X}|Z])\geq\frac{1}{2}\mathrm{\mathrm{\log}\left\{ \mathbb{E\mathit{\left\{ \mathbb{V}\left[\tilde{X}|Z\right]\right\} }}/\mathbb{V}\left\{\mathit{\tilde{X}}\right\}\right\} }=\mathrm{\log}\{\nu_{f}(\tilde{X}|Z)^{-1/2}\},
\]
since $\mathbb{V}[\tilde{X}]=\mathbb{V}\{\mathbb{E}[\tilde{X}|Z]\}+\mathbb{E}\{\mathbb{V}[\tilde{X}|Z]\}$
by the law of total variance. This establishes the first inequality
in Eq. \ref{eq:CfMS} and, importantly, does so in a way that enables
us to establish the second. It is a direct consequence of \cite{Renyi1959}
that $\mathbb{V}\{\mathbb{E}[\tilde{X}|Z]\}/\mathbb{V}[\tilde{X}]$
is equal to $\mathrm{sup_{\mathit{t}}\mathrm{Corr}^{2}(\mathit{t\left(Z\right),\tilde{X}})}$,
where the supremum is over all Borel measurable functions $t$ such
that $\mathrm{Corr}(\mathit{t\left(Z\right),\tilde{X}})$ is well
defined. It follows that
\begin{equation}
1-\nu_{f}(\tilde{X}|Z)=\mathbb{V}\{\mathbb{E}[\tilde{X}|Z]\}/\mathbb{V}[\tilde{X}]\geq\mathrm{Corr}^{2}(\mathit{t(Z),\tilde{X}})\geq\mathrm{Corr}^{2}(\mathit{Z,\tilde{X}}), \label{eq:biv}
\end{equation}
for all $t(\cdot)$, which implies the second inequality in Eq. \ref{eq:CfMS}.
\end{proof}
An analogue of Theorem \ref{thm:ZgXlwrbnd} when the conditioning is on the response variable $Z$ is given by the following Lemma. The proof
is straightforward. A related lower bound is given
without proof in the frequency domain by \cite{Rieke1999}, the bound being on the mutual information
rate in continuous-time.
\begin{lem}
\label{thm:XgZlwrbnd}Let $(X,Z)$ be a random vector in $\mathbb{R^{\mathrm{\mathit{d}}}}$,
$d\geq2,$ with joint density with respect to Lebesgue measure, $f(x,z)$.
Suppose there exists a one-to-one mapping $s:X\rightarrow\tilde{X}$
(whose domain is an open set and a support of $X$) such that $\tilde{X}$
has a Gaussian density. We assume that the mapping $s$ is continuously
differentiable with a Jacobian that is everywhere non-zero, and that
$\nu_{f}(\tilde{X}|Z)^{-1/2}$ exists. Then, scaling by the reciprocal
of the dimension of $X$,
\begin{equation}
d_{X}^{-1}I_{f}(X;Z)\geq\frac{1}{2}d_{X}^{-1}\mathrm{\log}\{\nu_{f}(\tilde{X}|Z)^{-1/2}\}\geq\frac{1}{2}\mathrm{\log}\left\{ \frac{\mathrm{[\det}(\mathbb{V}[\tilde{X}])]^{d_{X}^{-1}}}{d_{X}^{-1}\mathrm{tr}(\mathbb{E}\{\mathbb{V}[\tilde{X}|Z]\})}\right\}_{\mathit{f}} ,\label{eq:XgZlwrbound}
\end{equation}
and
\begin{equation}
C_{Z|X}\geq I_{f}(X;Z) \geq \mathrm{\log}\left\{\nu_{f}(\tilde{X}|Z)^{-1/2}\right\}, \label{capacitybnd2}
\end{equation}
provided $f(x)$ is the density of an allowed input distribution, that is a distribution in $\mathcal{F}$.
\end{lem}
\begin{proof}
A concise proof of the first inequality in Eq. \ref{eq:XgZlwrbound} uses that $h(\tilde{X}|Z=z)\leq\frac{1}{2}\mathrm{log}[(2\pi e)^{d_{X}}\mathrm{det}(\mathbb{V}[\tilde{X}|Z=z])]$,
which follows from the maximum entropy property of the multivariate
Gaussian distribution for a given covariance matrix. Then $h(\tilde{X}|Z)\leq\frac{1}{2}\mathrm{log}[(2\pi e)^{d_{X}}\mathrm{det}(\mathbb{E}\{\mathbb{V}[\tilde{X}|Z=z])\}]$,
where we have applied Jensen's
inequality, using the concavity of the function $\mathrm{log}\{\mathrm{det}(\Sigma)\}$
for symmetric, non-negative definite matrices, $\Sigma$ \cite{Cover1988}.
The result follows since $I_{f}(X;Z)=I_{f}(\tilde{X};Z)=h(\tilde{X})-h(\tilde{X}|Z)$
and the entropy of the Gaussian $\tilde{X}$ is given by $h(\tilde{X})=\frac{1}{2}\mathrm{log}[(2\pi e)^{d_{X}}\mathrm{det}(\mathbb{V}[\tilde{X}])]$.
The second inequality in Eq. \ref{eq:XgZlwrbound} follows because we
take the covariance matrix $E\{\mathbb{V}[\tilde{X}|Z]\}$ to be positive
definite and $\mathrm{0<[\det}(M)]^{1/m}\leq m^{-1}\mathrm{tr}(M)$
for a positive definite, $m\times m$ matrix, $M$.
\end{proof}
\noindent The existence of the mapping $s(x)$ in Lemma \ref{thm:XgZlwrbnd} is
not unduly restrictive. For example, when $d=2$ and the input $X$ has a strictly
increasing distribution function taking values in $(0,1)$, then $F_{X}(X)$
is uniformly distributed and can be invertibly transformed to a Gaussian
random variable. Similar comments apply when $d>2$, using the multivariate transformation
of \cite{Rosenblatt1952} to independent uniform r.v.'s on $(0,1)$.
Rate distortion-type arguments using average mean-square error distortion and $\tilde{X}$ as the Gaussian source cannot establish Eq. \ref{eq:XgZlwrbound} for $d_X > 1$. Such arguments \cite[Section 4.5.2]{BergerBook} show only that $\frac{1}{2}\mathrm{\log\{\mathrm{\det}(\mathbb{V}[\tilde{X}])}^{d_{X}^{-1}}/d_{X}^{-1}$ $\mathrm{tr}(\mathbb{E}\{\mathbb{V}[\tilde{X}|Z]\})\}_{f}$
is a lower bound for $d_{X}^{-1}I_{f}(X;Z)$ under the conditions of
Lemma \ref{thm:XgZlwrbnd}. As we have shown, our lower bound in Eq. \ref{eq:XgZlwrbound}, $\mathrm{\log}\{\nu_{f}(\tilde{X}|Z)^{-1/2}\}$, is tighter than this one derived using average mean-square error rate distortion arguments with Gaussian source.
\section{Applications}
The lower bounds on mutual information and capacity derived in previous sections will prove useful in at least two types of application: analysing the dependence between input and response vectors using empirical data; and analysing the information capacity of signaling and communication mechanisms for which physical models are available. An illustration of the first type of application using simulated data and further discussion are given immediately below. An existing example of the second type is given in \cite{Mitra2001} which examines the information capacity of optical fiber communication by employing a lower bound based on the Pearson correlation. Indeed, physical models of a communication mechanism can often be solved for their moments when distributional results are not feasible. For example, models of biomolecular signaling mechanisms are stochastic kinetic models of biochemical reaction networks \cite{BowsherAOS} that can be solved approximately using system-size expansions of the master equation \cite{GrimaEMREJCPpaper2010}. Such expansions can be used to provide fast, computational evaluation of $\mathbb{E}[Z|X]$ and $\mathbb{V}[Z|X]$ for many (rate) parameter vectors describing the network \cite{iNApaper}. Theorem \ref{thm:ZgXlwrbnd}, Eq. \ref{capacitybnd} can then be used to approximate the capacity of the signaling mechanism and explore its parameter sensitivity.
\subsection{Lower bound estimation\label{MCarlo}}
Estimation of mutual information rapidly becomes problematic as the dimension, $d$, of $(X,Z)$ grows. A distinct advantage of the lower bounds in Eqs. \ref{eq:ZgXlwrbound} and \ref{eq:XgZlwrbound} based on $\mathrm{log}(\nu^{-1/2})$
is that they are amenable to inference by using nonparametric
regression to estimate the relevant conditional mean. Nonparametric regression and covariance matrix estimation methods for higher dimensions \cite{Friedman1991,Bickel2008} should break down more slowly than mutual information estimation methods as $d$ grows. This is valuable for applications, including
those in systems biology where multiple inputs and outputs
often need to be considered. Estimation of our lower bounds should therefore prove useful for analysing the dependence between input and response in higher dimensions.
The following simulation study demonstrates that use of the lower bounds can substantially improve inference about mutual information when the sample size becomes limited for a given value of $d$. Here we use $d=2$. Inferential procedures for the $d>2$ setting lie beyond the scope of the present paper and will be explored in future work. The $k$-nearest neighbour point estimator \cite{Kraskov2004}, $\hat{I}_{knn}$, is widely regarded as the leading method for estimation of mutual information using continuously distributed data. We employ a \emph{composite estimator} defined as the maximum of $\hat{I}_{knn}$ and the lower limit of our bootstrap confidence interval for $\mathrm{log}(\nu^{-1/2})$.
This composite
estimator makes use of our lower bounds to correct erroneous point estimates. We find that the lower bounds are able to provide substantial improvements to the downward bias and root mean square error
(rmse) we report for the nearest-neighbour estimator.
We assume that we are given data $\left\{ (X_{i},Z_{i});i=1,...,N\right\} $
for independent and identically distributed units and that the distribution of the input $X$ is known (see the discussion following Eq. \ref{picture}). We obtain confidence
intervals, for example, for the lower bound in Eq. \ref{eq:CfMS}
based on $\nu(\tilde{X}|Z)$ as follows. (1)
Obtain fitted values, $\hat{X}_{i}$, for the transformed, Gaussian
input $\widetilde{X}$ by nonparametric estimation of $\mathbb{E}[\tilde{X}|Z]$
using a smoothing spline; (2) Obtain the estimate $1-\hat{\nu}_{{\scriptscriptstyle \tilde{X}|Z}}$
as the ratio of the sample variance of $\hat{X}_{i}$ to the known variance of $\tilde{X}_{i}$ (see Eq. \ref{eq:biv}); (3)
Obtain bias-corrected, accelerated ($\mathrm{BC_{a}}$) bootstrap
confidence intervals \cite{Efron1994} using the estimator $\mathrm{\log}(\hat{\nu}_{{\scriptscriptstyle \tilde{X}|Z}}^{-1/2})$.
Details of the proposed procedure are given in the Appendix.
\begin{figure}[t!]
\begin{centering}
\includegraphics[width=1\textwidth]{figure1AOS.pdf}
\par\end{centering}
\caption{ \textbf{Bootstrap confidence intervals of the lower bounds $\mathrm{log}(\nu^{-1/2})$
can substantially improve inference about mutual information through use of a composite estimator. } {{Simulation
results are shown for the bivariate normal (a--c), and for
a mixture model incorporating the same linear regression
model but with a mixture of normals distribution for $X$ (d--f). $N$ is the number of i.i.d. observations.
a) and d): for $N=25$, the $\mathrm{BC_{a}}$, nominally
$90\%$ confidence intervals for our lower bounds in Eqs. \ref{eq:ZgXlwrbound}
and \ref{eq:CfMS} resp., together with the $k$-nearest
neighbour estimates, $\hat{I}_{knn}$, with $k=3$.
(Analogous confidence intervals for the lower bounds based on $\mathrm{Corr}(\tilde{{X}},Z)^{2}$
are shown in d) for the mixture model when $I(X;Z)>4$ bits).
Parameter vectors for each model were sampled independently from their
parameter spaces. A single data set is generated for each
parameter vector in a) and d). Coverage probability
(grey crosses, b) and c)) gives frequency with
which the $\mathrm{BC_{a}}$ interval covers the true $I(X;Z)$; Exceedance
probability (grey crosses, e) and f)) gives frequency
with which $I(X;Z)$ exceeds the lower limit of the $\mathrm{BC_{a}}$
interval (nominally $>0.9$). Root mean square errors (rmse) are plotted
for $\hat{I}_{knn}$ (filled circles), and the composite
estimators (see text) based on $\nu_{{\scriptscriptstyle {Z|X}}}^{-1}$
or $\nu_{{\scriptscriptstyle \tilde{X}|Z}}^{-1}$ (black crosses)
and $\mathrm{Corr}(\tilde{X},Z)^{2}$ (diamonds). Results based
on $500$ Monte Carlo replications.}} }
\label{fig1} \vspace{-0.25in}
\end{figure}
Figures 1 and 2 present simulation results for a range of true
values of the mutual information and for two types of data generation
mechanism: a bivariate normal distribution and a mixture model. In
both, $Z=\alpha+\beta X+\varepsilon$, with $\varepsilon$ normally
distributed conditional on $X$ (with constant variance independent
of $X$). In the first, $X$ has a marginal normal distribution (under the data generating density, $f$),
hence $(X,Z)$ has a bivariate normal distribution, $X=\tilde{X}$, and the bounds $\mathrm{log}(\nu^{-1/2})$ in Eqs. \ref{eq:ZgXlwrbound} and \ref{eq:CfMS}
hold with equality. In the second, $X$
is specified to be an equally-weighted mixture of 2 normals, and we obtain the pseudo-input $\tilde{X}$
by first transforming to uniformity using the probability integral transform
and then transforming to normality. We adopt the second specification because $\mathbb{E}[\tilde{X}|Z]$
becomes non-linear (and sigmoidal), but the true value of $I(X;Z)$ is
still known with precision through the use of a Monte Carlo average for $h(Z)$ (see Appendix). Details of the parameterisations of the models used are also given in the Appendix.
\begin{figure}[tbh]
\includegraphics[width=1\textwidth]{figure2AOS.pdf}
\caption{\textbf{The lower bounds $\mathrm{log}(\nu^{-1/2})$
can substantially reduce bias through use of a composite estimator}. The lower bounds based on the Peason correlation do not reduce estimation bias for the mixture model with non-linear conditional mean. Biases are plotted for the $k$-nearest neighbour estimates, $\hat{I}_{knn}$,
with $k=3$ (filled circles), and for the composite estimators (see
text) based on $\nu_{{\scriptscriptstyle Z|X}}^{-1}$ or $\nu_{{\scriptscriptstyle \tilde{X}|Z}}^{-1}$
(bivariate normal and mixture models respectively; crosses) and $\mathrm{Corr}(X,Z)^{2}$
(diamonds). Parameter vectors for each model were sampled independently
from their parameter spaces. Results are shown for the
bivariate normal (a--c), and for the mixture model (d--f). $N$ is the number of \emph{i.i.d.} observations in each
data set. All results based on $500$ Monte Carlo replications. }
\end{figure}
Panels a) and d) of Fig. 1 show our $\mathrm{BC_{a}}$ confidence
intervals for the lower bounds based on Eq. \ref{eq:ZgXlwrbound}
and \ref{eq:CfMS} respectively, together with the point estimates
$\hat{I}_{knn}$, for independently generated data sets corresponding
to different true values of $I(X;Z)$ and for sample size $N=25$. In \cite{Kraskov2004},
the authors recommend in practice to use values of $k$ between 2 and 4. We therefore
calculate $\hat{I}_{knn}$ using $k=3$ nearest neighbours ($\hat{I}_{knn}=I^{(2)}(X,Z;k=3)$ in \cite{Kraskov2004}). The poor
performance of $\hat{I}_{knn}$ with this sample size is evident for
both models, particularly for mutual information in excess of 3 bits,
where substantial, growing bias and rmse are evident (see Fig. 2 for plots of bias).
Higher values of $k$ result in worse bias and rmse of $\hat{I}_{knn}$ (not shown). The remaining panels of Fig. 1 depict,
for sample sizes $N=25$ and $N=50$, various properties under repeated
sampling: the frequency with which the $\mathrm{BC_{a}}$ interval
for the lower bound covers {[}b) and c){]} and has a lower limit exceeded
by {[}e) and f){]} the true mutual information; the rmse of $\hat{I}_{knn}$;
and the rmse of our composite estimator, given by the maximum of $\hat{I}_{knn}$
and the lower limit of the $\mathrm{BC_{a}}$ interval. For both sample sizes, the nonparametric confidence
intervals perform well under repeated sampling and provide substantial
reductions in bias and rmse when comparing $\hat{I}_{knn}$ to the composite
estimator (see also Fig. 2). Finally, in the mixture model where $\mathbb{E}[\tilde{X}|Z]$
is non-linear in $Z$, the lower bound based on $\mathrm{Corr}(\tilde{X},Z)^{2}$
performs considerably worse than that based on ${\nu}_{{\scriptscriptstyle \tilde{X}|Z}}$, as shown in panels d) to f) of Figs. 1 and 2. The corresponding $\mathrm{BC_{a}}$ intervals lie well below those based on ${\nu}_{{\scriptscriptstyle \tilde{X}|Z}}$ and have lower limits below $\hat{I}_{knn}$ in all cases shown in panel d). The associated composite estimator consequently fails to reduce either the bias or the rmse of estimation.
\section{Appendix}
\subsection{Additional proofs}
\begin{proof}
(Lemma \ref{lem:lbndCorr}) Let $g(x,z)$ be the multivariate Gaussian
density with the same unconditional first and second moments as $f(x,z),$
and with marginal Gaussian density $g(x)=f(x).$ Thus, $\mathrm{\mathbb{V}_{\mathit{g}}}[\left(X,Z\right)]=\mathrm{\mathbb{V}_{\mathit{f}}}[\left(X,Z\right)].$
We use subscripts to identify the relevant joint density throughout.
Notice that
\[
I_{\mathit{f}}(X;Z)=\mathbb{E_{\mathit{f}}}\left\{ \mathrm{log}\frac{g(X,Z)}{f(X)g(Z)}\right\} -\mathbb{E_{\mathit{f}}}\left\{ \mathrm{log}\frac{g(X,Z)f(Z)}{f(X,Z)g(Z)}\right\} \geq\mathbb{E_{\mathit{f}}}\left\{ \mathrm{log}\frac{g(X,Z)}{f(X)g(Z)}\right\} ,
\]
where the second expectation of the equality is seen to be non-positive
by applying Jensen's inequality and then integrating first with respect
to $x$. Furthermore,
\begin{equation}
\mathbb{E_{\mathit{f}}}\left\{ \mathrm{log}\frac{g(X,Z)}{f(X)g(Z)}\right\} =\mathbb{E_{\mathit{g}}}\left\{ \mathrm{log}\frac{g(X,Z)}{f(X)g(Z)}\right\} =I_{\mathit{g}}(X;Z),\label{eq:careMS-1}
\end{equation}
because $\mathbb{E}_{f}[\mathrm{log\{}g(\cdot)\}]=\mathbb{E}_{g}[\mathrm{log}\{g(\cdot)\}].$
For example, $\mathbb{E}_{f}[\mathrm{log}\{g(X,Z)\}]=\mathbb{E}_{g}[\mathrm{log}\{g(X,Z)\}]$
because
\begin{eqnarray*}
\mathbb{E}_{f}\left\{ \left(\begin{array}{c}
X-E[X]\\
Z-E[Z]
\end{array}\right)^{\mathrm{T}}\mathrm{\mathbb{V}_{\mathit{g}}}[\left(X,Z\right)]^{-1}\left(\begin{array}{c}
X-E[X]\\
Z-E[Z]
\end{array}\right)\right\} & =\\
\mathrm{tr}\left\{ \mathrm{\mathbb{V}_{\mathit{g}}}[\left(X,Z\right)]^{-1}\mathbb{E}_{f}\left[\left(\begin{array}{c}
X-E[X]\\
Z-E[Z]
\end{array}\right)\left(\begin{array}{c}
X-E[X]\\
Z-E[Z]
\end{array}\right)^{\mathrm{T}}\right]\right\} & =d,
\end{eqnarray*}
since $\mathrm{\mathbb{V}_{\mathit{g}}}[\left(X,Z\right)]=\mathrm{\mathbb{V}_{\mathit{f}}}[\left(X,Z\right)].$
Evaluating $I_{\mathit{g}}(X;Z)$=$h_{g}(X)+h_{g}(Y)-h_{g}(X,Y)$
is straightforward since the marginal and joint densities under $g$
are all Gaussian. We find
\[
I_{\mathit{g}}(X;Z)=\frac{1}{2}\mathrm{log\left\{ \frac{det(\mathbb{V_{\mathit{g}}}\left[X\right])det(\mathbb{V_{\mathit{g}}}\left[\mathit{Z}\right])}{det(\mathbb{V_{\mathit{g}}}\left[(X,Z)\right])}\right\} }=\frac{1}{2}\mathrm{log\left\{ \frac{det(\mathbb{V_{\mathit{f}}}\left[X\right])det(\mathbb{V_{\mathit{f}}}\left[\mathit{Z}\right])}{det(\mathbb{V_{\mathit{f}}}\left[(X,Z)\right])}\right\} },
\]
since $g$ and $f$ have identical second moments by construction.
The stated results are then obtained by partitioning of the matrix
$\mathbb{V}_{f}\left[(X,Z)\right].$
\end{proof}
\begin{proof}
(Theorem \ref{thm:XgZasymp}). We have that $\nu(X_{n}|Z_{n})=\nu(\breve{X}_{n}|Z_{n})\rightarrow0$.
Since $\nu(\breve{X}_{n}|Z_{n})=\mathbb{E}\{\mathbb{V}[\breve{X}_{n}|Z_{n}]\}$
$=\mathbb{E}\{(\mathit{\breve{X}_{n}-\mathbb{E}[\breve{X}_{n}|Z_{n}]})^{2}\}$,
it follows that $\breve{X}_{n}-\mathbb{E}[\breve{X}_{n}|Z_{n}]$ converges
to zero in mean square (in $L^{2}$) and therefore $\breve{X}_{n}-\mathbb{E}[\breve{X}_{n}|Z_{n}]\rightarrow^{pr}0$.
Consider estimating $\breve{X}_{n}$ based on observation of $Z_{n}$
as follows: the estimator $\hat{X}_{n}(Z_{n})$ is equal to a point
in the support of $\breve{X}_{n}$ which minimises the Euclidean distance
from $\mathbb{E}[\breve{X}_{n}|Z_{n}]$. Let $x\in\mathbb{R},\breve{x}_{n}\mathcal{\in\breve{X}}_{n}$
and notice that if $|\breve{x}_{n}-x|<\epsilon^{*},$ then $\mathit{|}\breve{x}_{n}-x|<\epsilon_{n}^{*}<\mathit{|}\breve{x}_{n}'-x|$,
that is $x$ is closer to $\breve{x}_{n}$ than to any other point
$\breve{x}_{n}'$ in $\mathcal{\breve{X}}_{n}$. Therefore, if $|\mathit{\breve{X}_{n}-}E[\breve{X}_{n}|Z_{n}]|<\epsilon^{*},$
$\mathbb{E}[\breve{X}_{n}|Z_{n}]$ is closer to $\breve{X}_{n}$ than
to any other point in the support, the estimator $\hat{X}_{n}(Z_{n})$
is uniquely defined, and that estimator recovers $\breve{X}_{n}$
without error $\left(\hat{X}_{n}(Z_{n})=\breve{X}_{n}\right).$ Thus,
the probability of estimation error, $p_{\mathrm{error}},$ satisfies
\[
p\mathrm{_{error}}=P\left\{ \mathit{\hat{X}_{n}\left(Z_{n}\right)\neq\breve{X}_{n}}\right\} \leq\mathbb{P}\left\{ |\mathit{\breve{X}_{n}-}E[\breve{X}_{n}|Z_{n}]|\geq\epsilon^{*}\right\} .
\]
Since $\breve{X}_{n}-\mathbb{E}[\breve{X}_{n}|Z_{n}]\rightarrow^{pr}0$,
$p_{\mathrm{error}}$ must therefore tend to zero as $n\rightarrow\infty.$
Fano's Inequality gives
\[
H(p\mathrm{_{error}})+p_{\mathrm{error}}\mathrm{log}|\mathcal{X}_{\mathit{n}}|\geq H(\breve{X}_{n}|Z_{n})=H(X_{n}|Z_{n}),
\]
since $|\mathcal{\breve{X}}_{\mathit{n}}|=|\mathcal{X}_{\mathit{n}}|$
and the rescaling does not change the conditional entropy. Therefore
$H(X_{n}|Z_{n})\rightarrow0$ as $n\rightarrow\infty.$
\end{proof}
\subsection*{Models, parametrisations and algorithms used in the simulation study of Section \ref{MCarlo}}
Figures 1 and 2 present simulation results for two data generation
mechanisms. In both, $Z=\beta X+\varepsilon$ with $\varepsilon\sim N(0,\sigma_{\varepsilon}^{2})$
and $\varepsilon$ independent of $X.$ The two models, together with
the schemes used to generate parameter vectors for the results shown
in Figures 1 and 2, are as follows:
\begin{enumerate}
\item \textit{Bivariate normal }\emph{model}: $X\sim N(0,\sigma_{{\scriptscriptstyle{X}}}^{2}).$
Model parameters were sampled as follows: i) $\beta$ uniformly distributed on $(1,10)$; ii) $\sigma_{\varepsilon}^{2}=10^{\theta_{1}}$
with $\theta_{1}$ uniformly distributed on $(-2,2)$; and iii) $\sigma_{{\scriptscriptstyle{X}}}^{2}=10^{\theta_{2}}$
with $\theta_{2}$ uniformly distributed on $(-2,2)$.
\item \textit{Mixture model}: $X$ is an equally-weighted mixture of 2 normal
distributions, that is $f_{{\scriptscriptstyle{X}}}(x)=\frac{1}{2}N(\mu_{1},\sigma_{1}^{2})+\frac{1}{2}N(\mu_{2},\sigma_{2}^{2})$. Model parameters were sampled as
follows: i) $\beta=10^{\theta_{1}}$, with $\theta_{1}$ uniformly
distributed on $(-1,1)$; ii) $\sigma_{\varepsilon}^{2}=10^{\theta_{2}}$
with $\theta_{2}$ uniformly distributed on $(-2.5,2.5)$. We set
$\mu_{1}=-\mu_{2}=5$ and $\sigma_{1}^{2}=\sigma_{2}^{2}=25/4.$
\end{enumerate}
The mixture model allows precise evaluation of $I(X;Z)=h(Z)-h(Z|X)$
via Monte Carlo sampling. We have $h(Z|X)=\frac{1}{2}\mathrm{log(2\pi e\sigma_{\varepsilon}^{2})}$.
Note that the marginal density $f(Z)$ is also an equally-weighted
mixture of 2 normals which we can express in closed form. Hence, we
can also estimate $h(Z)=-\mathbb{E}[\mathrm{log}f(Z)]$ as the Monte
Carlo average of $\mathrm{log}f(Z_{m})$ where $Z_{m}$ ($m=1,...,M$)
is a draw from the mixture model. For our numerical calculations we
set $M=10^{5},$ and monitored convergence of the Monte Carlo average.
Computations were implemented in R (version 2.12.2). The non-parametric
estimation of $\mathbb{E}[\tilde{X}|Z]$ was performed using the `smooth.spline'
function (an implementation of smoothing splines \cite{Green1994}) with
the number of knots set to $10$; the smoothing parameter was
chosen using cross-validation on the original dataset; all other parameters
were set to their default values. $90\%$ $\mathrm{BC_{a}}$ confidence
intervals were calculated from $B=2000$ bootstrap replications (using
the `boot' package). For the $k$-nearest neighbour estimation of mutual information \cite{Kraskov2004}
we used the authors' `MIxnyn' function within their MILCA suite
(available at \url{http://www.klab.caltech.edu/~kraskov/MILCA/}).
\bibliographystyle{nar}
|
\section{Introduction}
\label{}
Neutrinos with energies below few tens of MeV are expected to undergo a large enhancement to the cross-section for their elastic scattering off heavy nuclei \cite{freedman,drukier}. During this neutral-current interaction, all nucleons are expected to contribute coherently to the scattering process\footnote{The contribution from protons is markedly reduced by the numerical value of the weak mixing angle, resulting in a CENNS cross-section effectively proportional to the square of the number of neutrons in the target nucleus \cite{drukier}.}. Coherent elastic neutrino-nucleus scattering (CENNS) is expected to be the dominant mechanism in neutrino transport within supernovae and neutron stars \cite{wilson}, but it remains undetected forty years after its first description. This somewhat puzzling absence of experimental evidence for the largest low-energy neutrino cross-section is due to the modest energy of the nuclear recoils induced, and to the limited intensity of available neutrino sources in this energy range. The coherent nuclear mechanism behind this mode of neutrino interaction is assumed to be taking place also during the spin-independent scattering of favored dark matter candidates (WIMPs), adding to the interest of its experimental verification. Unfulfilled proposals for the detection of CENNS abound \cite{drukier,ieee,steve,ppc,red}.\\
The development of smallish coherent neutrino detectors may eventually enable technological applications \cite{leocastle}. An example is the non-intrusive monitoring of nuclear reactors using compact neutrino detectors, improving on existing methods \cite{bernstein1}. A number of tests for new physics should also be possible with a CENNS-sensitive neutrino detector. For instance, a neutral-current detector responds almost identically to all known neutrino types \cite{all}: observation of neutrino oscillations in such a device
would provide unequivocal direct evidence for sterile neutrino(s) \cite{drukier}. In addition to this, the differential cross section for this process is strongly dependent on a finite value of the neutrino magnetic moment \cite{dodd}. Recent studies have described the sensitivity of CENNS to non-standard neutrino interactions with quarks \cite{new}, to the effective neutrino charge radius \cite{bernabeu}, and to neutron density distributions in the nucleus \cite{patton}. A precise measurement of the CENNS cross section
would also provide a sensitive appraisal of the weak nuclear charge \cite{larry}.\\
In a seminal paper \cite{drukier}, Drukier and Stodolsky examined the prospects for CENNS detection from a variety of low-energy neutrino sources (solar, terrestrial, supernova, reactor, and spallation source). The opportunity provided by the high-luminosity SNS spallation source at Oak Ridge National Laboratory has been treated in \cite{frankyuri,kate}. While the main use for the SNS is as an intense neutron source, its protons-on-target (POT) produce pions, which decay at rest generating a monochromatic flash of 30 MeV prompt $\nu_{\mu}$. This is followed by delayed $\nu_{e}$ and $\bar{\nu}_{\mu}$ emissions with a broad energy (Michel spectrum, few tens of MeV), over the $\sim$2.2 $\mu$s timescale characteristic of $\mu$ decay. For present running conditions, the neutrino flux at 20 m from the SNS source is 1.7$\cdot10^{7}$ cm$^{-2}$ s$^{-1}$, per flavor. An attractive feature of this intense neutrino source is its time structure, consisting of 60 Hz of sub-$\mu$s pulses. This provides a $\sim6\cdot10^{-4}$ background rejection factor, when looking for CENNS signals in a sensitive detector, using only the $10~\mu$s window following POT. As will be discussed in Sec.\ 5, unfortunately most POT-associated backgrounds are expected to reach a CENNS detector with a time profile similar to that from the sought neutrino signals. It is therefore of the utmost importance to understand and abate these beam-associated backgrounds. Steady-state backgrounds unrelated to the SNS beam can be characterized during time periods preceding the POT signal. In the absence of beam-related backgrounds, the residual obtained by subtracting pre-POT from post-POT energy spectra is expected to display a characteristic CENNS-induced excess (Sec.\ 5).
\section{Advantages of CsI[Na] as a CENNS detector}
\label{}
The inorganic CsI[Na] scintillator presents several important advantages for CENNS detection at a spallation source. These can be listed:
\begin{itemize}
\item The high mass of both recoiling species, cesium and iodine, provides a very large coherent enhancement to the CENNS cross-section. This results into a substantial signal event rate (tens of events per kg per year, 20 m from the SNS source), as long as a sufficiently-low detection threshold can be achieved (few keV of nuclear recoil energy, keVnr, Fig.\ 1).
\item Both recoiling species are essentially indistinguishable due to their very similar mass (Fig.\ 2), greatly simplifying the understanding of the response of the detector. This is of particular importance in the presence of competing backgrounds, like those described in Sec.\ 5.
\item The light yield from low-energy nuclear recoils in CsI[Na] is sufficiently large to expect a realistic $\sim$7 keVnr threshold when employing conventional bialkali photomultipliers (PMTs) in combination with this scintillator (Sec.\ 3). In the near future, this threshold can be further reduced down to $\sim$4.5 keVnr with the adoption of ultra-bialkali (UBA) photocathodes \cite{UBA}, once this relatively new technology has been extended to large surface area PMTs. A 7 keVnr threshold is demonstrated here, under background conditions similar to those expected at the SNS (Sec.\ 5).
\item It may be possible to perform statistical discrimination (as opposed to event-by-event discrimination) between low-energy nuclear and electron recoils for $\gtrsim$1,000 accumulated events. This is based on a difference of $\sim$60 ns between the fast scintillation decay time noticed for these two families of events (Sec.\ 3). Similar features have been observed and exploited in NaI[Tl] scintillators dedicated to WIMP detection \cite{nai60}.
\item CsI[Na] crystals are naturally low in internal radioactivity. When grown from screened salts confirmed to contain low levels of U, Th, $^{40}$K, $^{87}$Rb, and $^{134,137}$Cs \cite{radiopurekims}, as is done here, they provide a sufficient CENNS signal-to-background ratio. A steady-state background subtraction made possible by the pulsed nature of the CENNS signal at the SNS further improves this situation (Sec.\ 5).
\item A convenient low-energy signal at 57.6 keV arises from the inelastic scattering of fast neutrons, $^{127}$I(n,n'$\gamma$). This signal can be monitored to provide an {\it in situ} measurement of neutron backgrounds competing with the CENNS signal (Sec.\ 5).
\item CsI[Na] lacks the excessive long-lived afterglow (phosphorescence) that is characteristic of CsI[Tl] and other scintillators (\cite{afterglow1}, Fig.\ 3). This is a crucial feature in an application involving the detection of faint scintillation signals, in particular for a detector operated in a location lacking significant overburden. In such a site, the frequent large energy depositions from cosmic-ray interactions ($\sim$1 / cm$^{2}$ min) can lead to an intolerable continuum of delayed afterglow emissions in other scintillators, resulting in an increase of the effective threshold.
\item Several other practical advantages exist: CsI[Na] exhibits a high light yield ($\sim$45 photons/keVee at 662 keVee \cite{lyield}). With a light emission peaking at 420 nm, it features the best match to the response curve of bialkali photocathodes of any scintillator. This results in almost twice the photoelectron yield of CsI[Tl], for similar energy depositions. CsI[Na] exhibits a short scintillation decay time ($\tau_{f}\sim0.6
~\mu$s, $\tau_{f}\sim1~\mu$s for CsI[Tl]), suitable for the timing characteristics of the neutrino emission at a spallation source. It is a rugged detector material, and exhibits a negligible light yield variation in the temperature range 15 C-30 C. It is also relatively inexpensive ($\sim$ \$1,000/kg), permitting an eventual detector mass large enough (few hundreds of kg) to attempt the interesting CENNS physics studies mentioned in Sec.\ 1.
\end{itemize}
In this paper we describe the development of a 14 kg CsI[Na] detector in a dedicated radiation shield, intended for installation at the SNS during 2014 within the framework of the CSI:SNS collaboration\footnote{Recently renamed as COHERENT collaboration.} \cite{white}. This device is expected to be sufficient for a first measurement of CENNS at the SNS, while providing operational experience towards a larger-mass detector.
\section{Characterization of the response of CsI[Na] to low-energy nuclear recoils}
\label{}
Nuclear recoils are not as efficient in producing scintillation or ionization as electron recoils of the same energy. This is true of any detector medium. The ratio between the scintillation yields for these two recoiling species is referred to as ``quenching factor", and is typically of order few percent at low energy. Therefore, a crucial first step before a CENNS measurement is a careful characterization of this quenching factor, in the relevant few keVnr to few tens of keVnr range. Its present measurement for CsI[Na] uses a dedicated setup described in detail elsewhere \cite{myq2}, employed in that publication for the characterization of NaI[Tl]. The reader is referred to \cite{myq2} for a full description of the method, which features several improvements with respect to previously-employed techniques. \\
Briefly stated, a small 14.5 cm$^{3}$ CsI[Na] crystal was exposed to the monochromatic 2.2 MeV emissions from a DD neutron generator, capturing the scattered neutrons with a second scintillator capable of neutron/gamma discrimination. The scattering angle is varied by changing the position of this second detector on a goniometric table, scanning in the process a range of nuclear recoil energies (8 keVnr-30 keVnr) in the crystal under test. Separately, the response to electron recoils was explored using Compton scattering from a collimated beam of 356 keV $^{133}$Ba gammas, captured by a germanium detector following their interaction with CsI[Na]. Again, the Compton scattering angle was scanned, leading to a measurement of the response to electron recoils in the range 3 keVee-50 keVee (``ee" makes reference to ``electron equivalent", i.e, ionization energy). This small CsI[Na] detector was obtained from the same source \protect\cite{proteus} as the 2 kg and 14 kg crystals discussed below, using the same sodium dopant concentration. Scintillation was measured using a high-quantum efficiency UBA PMT \cite{myq2}. \\
Fig.\ 4 shows example distributions of light yield for several choices of neutron scattering angle (i.e., nuclear recoil energy). Cs and I recoils are indistinguishable due to their very similar nuclear mass. This figure is equivalent to Fig.\ 8 in \cite{myq2}. The light yield obtained for electron recoils via Compton scattering is also displayed in the inset, similar to Fig.\ 4 in \cite{myq2}. Following \cite{myq2}, the ratio between these yields is employed to extract the quenching factor in the energy region of interest for a CENNS exploration at the SNS (Fig.\ 5). \\
Similar to the situation described in \cite{myq2,myq3} for NaI[Tl], the low-energy quenching factor found here is considerably smaller than in previous attempts by others \cite{qfkims}. A discussion of systematic effects able to artificially increase low-energy quenching factor measurements is provided in \cite{myq2,myq4}. Several measures designed to avoid these systematics were in place during the present measurement \cite{myq2}. While the adoption of our smaller quenching factor is very conservative vis-a-vis a calculation of CENNS rates at the SNS, a separate measurement using a dedicated monochromatic pulsed neutron beam line at TUNL's tandem accelerator is in progress: preliminary results using NaI[Tl] \cite{private} reproduce the trend towards a monotonically decreasing quenching factor with decreasing energy, found here and in \cite{myq2,myq3}. A confirmation of the same behavior for CsI[Na] and CsI[Tl] would have an important impact on the claimed sensitivity of a recent dark matter search using cesium iodide \cite{newkims}. \\
Fig.\ 6 displays the scintillation decay time observed during present calibrations, for both electron recoils (ER) and nuclear recoils (NR). The evolution with energy evident in the figure has been observed before for CsI[Tl] \cite{qfkims} and NaI[Tl] \cite{nai60,myq2,othernai}. Small differences in decay time have been exploited for the statistical discrimination between NR's and ER's in dark matter searches, once a sufficiently large number of events is gathered \cite{nai60}. In view of Fig.\ 6, it may be possible to apply a similar statistical discrimination approach to eventually demonstrate that POT-coincident events at the SNS do indeed originate dominantly in nuclear recoils. In order to further explore this possibility, two subsets of the NR and ER available data were selected, each containing 1,000 events, under the criterion of having distributions as close as possible in number of photoelectrons (PE) registered per event (Fig.\ 7, inset). Waveforms for these 1,000 events were then added together, paying attention to assigning a zero-time coordinate to the position of the first photoelectron in the waveform\footnote{Also removing from the analysis the artificial initial spike in PMT current generated when performing this action.} (Fig.\ 7). The average low-energy NR and ER waveforms thus obtained are then fitted allowing for a fast and slow scintillation decay component \cite{decaytime} plus a (negligible) constant term. Keeping in mind that 10 PE corresponds here to $\sim$0.55 keVee and to $\sim$12 keVnr, these fits confirm the evolution with energy observed in Fig.\ 6. This indicates that it may indeed be possible to distinguish a group of $\gtrsim$1,000 NR events from its ER equivalent, through a measurement of the fast decay time constant of their ensemble. The planned approach to data-acquisition at the SNS (Sec.\ 5) would allow to test this interesting possibility, by comparing the cumulative of NR-dominated events in the few $\mu$s following POT, to those in the few $\mu$s prior to POT, the latter being representative of ER-dominated steady-state backgrounds not associated to pulsed SNS emissions. However, this test would require a substantial further improvement to the signal-to-background ratio presently achieved (Sec.\ 5).
\section{Detector and shielding design}
\label{}
In order to assess the steady-state environmental and internal backgrounds affecting a deployment at the SNS, we developed a 2 kg prototype of the 14 kg CsI[Na] CENNS detector. The complete shielding intended for SNS installation was populated with this prototype, and operated in a 6 m.w.e. laboratory at the University of Chicago, an overburden similar to that available at the SNS basement location discussed below. The two detectors are identical in all respects, except for their length. The expectation is that the larger device will exhibit a slightly reduced low-energy background per unit detector mass, due to its enhanced peak-to-Compton ratio.\\
A $\sim$9.9 PE/keVee yield is measured for few hundred keVee gamma interactions in the 2 kg prototype\footnote{This light yield is expected to increase by 30\% at 10 keVee, due to the intrinsic non-linearity characteristic of CsI[Na] \cite{nonlinearity}. We observe an increase of precisely this magnitude in our low-energy Compton scattering calibrations (Fig.\ 4), after accounting for the difference in quantum efficiency between UBA (33\%) and standard bialkali PMTs (22\%) for CsI[Na] scintillation wavelengths.}. Adopting the 22\% quantum efficiency expected from the standard bialkali photocathode employed and the CsI[Na] spectral emission, this corresponds to the nominal $\sim$45 photon/keVee light yield from CsI[Na] \protect\cite{lyield}, indicating an optimal light collection efficiency. Based on existing information \cite{uniformity} and manufacturer data, only a small decrease in overall light collection efficiency is expected from the larger crystal, by limiting its length to 33 cm and using a large active area (11.5 cm diameter) PMT for the readout. This efficiency will be characterized along the long axis of the final crystal using a collimated gamma-ray scan. \\
Figs.\ 8 and 9 show different aspects of detector and shielding. Samples from the crystal boule were screened against internal radioactive contaminants using low-background germanium gamma spectroscopy at SNOLAB, and ICP-MS. The concentration of $^{238}$U, $^{235}$U and $^{232}$Th was below the $\sim$1 ppb sensitivity of the measurements, however showing a presence\footnote{It is worth mentioning that the presence of a small $^{40}$K contamination provides a convenient built-in low-energy calibration reference at 3.2 keVee. This is the binding energy of K-shell electrons in the argon daughter following $^{40}$K electron capture, detectable whenever a 1,461 keV gamma de-excitation escapes the crystal.} of 17$\pm$16 mBq/kg of $^{40}$K. The material contained detectable levels of radioactive Cs and Rb isotopes (26$\pm$2 mBq/kg of $^{134}$Cs, 28$\pm$3 mBq/kg of $^{137}$Cs, and 72 ppb of natural Rb -i.e., 20 ppb $^{87}$Rb-). These are known to be responsible for most of the internal low-energy backgrounds in this scintillator \cite{radiopurekims}. These concentrations were determined to be sufficiently low for the present application (Sec.\ 5). Crystals are wrapped in PFTE expanded-membrane reflector. Low-background Suprasil synthetic silica windows are employed. Epoxies and optical couplants were screened against excessive internal radioactivity. The crystals are encapsulated in electroformed OFHC copper cans custom-grown at the Pacific Northwest National Laboratory (PNNL). A low-background Electron Tubes bialkali 9390UFL PMT, responsible for just 0.5$\pm$0.3 Bq of gamma emissions above 0.1 MeV, is used. A custom-designed OFHC copper structure holds crystal and PMT in place, also providing support for two inches of ancient lead, placed between detector and high-activity electronics in its voltage divider (Fig.\ 8). This ultra-low background (ULB) lead, also present in a 1" innermost layer surrounding the detector, was measured to contain $<$0.02 Bq/kg of $^{210}$Pb using radiochemical extraction and high-sensitivity alpha spectroscopy at PNNL. This inner lead layer results in a negligible low-energy background from $^{210}$Pb bremsstrahlung \cite{vojtila}. The gamma shielding is completed with contemporary lead ($\sim$100 Bq $^{210}$Pb/kg), resulting in a minimum of 18 cm of lead surrounding the CsI[Na] crystals in any direction.\\
Shielding against thermal neutrons is present in the form of borated silicone (sides, top) or cadmium sheet (bottom) layers immediately outside of the lead castle. Surrounding these is a set of seven plastic scintillator muon veto panels (sides and top), two inches thick. Each panel is read out by four summed-output 3/4" PMTs, inserted into the panel so as to obtain a compact assembly devoid of protruding light guides. Each set of four PMTs is gain-matched, and their common high-voltage bias is adjusted to obtain an optimal separation between muon interactions and lower-energy signals from environmental gammas, by avoiding veto discriminator triggers on the second. The efficiency of this active shield against events generated by cosmic rays interactions in detector or shielding is measured at $>$99.6\% by monitoring the spectral reduction above 10 MeV, an energy regime dominated by such events. The triggering rate from a Lecroy 620BL 8-channel discriminator used to provide a common logic output from the ensemble of veto panels is just $\sim$300 Hz for this veto efficiency. This rate is in good agreement with expectations based on the muon flux at sea level, and the surface area of this veto. This modest triggering rate is expected to generate a negligible incidence of false veto coincidences during operation at the SNS. \\
Fig.\ 9 shows this assembly inserted into a steel pipe, which supports additional external polyethylene neutron moderator. This pipe is intended for an eventual installation in a shallow (few meter deep) underground well outside of the SNS building. Prompt fast neutrons with energies of up to several hundred MeV are able to escape the massive shielding monolith surrounding the SNS spallation target, with a time-profile of arrival to the detector that can approximate that for the neutrino emissions. This time profile for different beam-associated backgrounds is illustrated in Fig.\ 10. Therefore, highly-penetrating hard neutrons able to induce nuclear recoils constitute the main background of concern in a search for CENNS at a spallation source\footnote{Nevertheless, their recoil energy spectrum and time profile should allow for their discrimination at the expense of some neutrino signal loss: the fastest neutrons induce large recoil energies at early times (Fig.\ 10), whereas the most energetic recoils from CENNS are due to $\nu_{e}$ and $\bar{\nu}_{\mu}$ from $\mu$ decay, which can arrive at later times. Preliminary data obtained at the SNS \cite{private2} indicate only a marginal neutron flux $\sim 2~\mu$s following a POT spill, leading to a strong reduction in the backgrounds affecting CENNS from delayed $\nu_{e}$ and $\bar{\nu}_{\mu}$.} \cite{steve,white}. A shallow well is expected to provide sufficient shielding in the direct line-of-sight to the source to remove this concern. An effort to characterize neutron and gamma backgrounds and their time profile at different locations within the SNS facility is underway within the COHERENT collaboration \cite{white}. Preliminary results using arrays of large liquid scintillator detectors with neutron/gamma discrimination capability indicate that an existing basement corridor, 20-25 m away from the SNS target, may be an ideal emplacement for CENNS detectors. This location provides $\sim$17 meters of concrete and gravel shielding in addition to the target monolith, resulting in a further extinction of the fast neutron flux by some five orders of magnitude \cite{private2}. This may be sufficient to entirely remove their associated backgrounds. Background measurements and simulation efforts continue at the time of this writing.
\section{Background measurements, data acquisition strategy, and prospects for CENNS detection.}
\label{}
Fig.\ 11 displays the steady-state backgrounds measured with the shielded 2 kg crystal under a 6 m.w.e. overburden. No neutron moderator was present during these measurements. This spectrum of energy was obtained using a preamplifier and shaping amplifier, with signals registered by a multi-channel analyzer. This approach is limited: while it is useful to help identify remaining contaminations and to assess the veto efficiency, it can underestimate very small energy depositions due to their limited triggering efficiency, or overestimate them by a lack of discrimination against PMT afterpulses, phosphorescence, etc. \\
In order to properly assess the steady-state background rate down to single photoelectron (SPE) emissions, we mimicked the 60 Hz POT trigger that will be available at the SNS through the use of an electronic pulser. This logic signal provides an external trigger to a National Instruments PCI 5153 digitizer, set to sample the PMT output at 0.5 Gs/s, sufficient to reconstruct SPEs and the integrated current from small few-photoelectron signals. A second available digitizer channel monitors the ``OR" output from the 8-channel discriminator merging the muon veto panel signals, allowing for the rejection of events associated to cosmic ray interactions in detector or shielding. The acquired waveforms are 60 $\mu$s long, of which 50 $\mu$s are pre-trigger information. Steady-state backgrounds not-related to the SNS beam can be measured, and eventually subtracted, by inspecting the PMT current present in the 10 $\mu$s immediately preceding the POT trigger\footnote{A concern must be considered, which is that the pre-POT spectrum might contain energy depositions on the average slightly larger than for the post-POT equivalent, due to the monotonically decreasing phosphorescent yield from earlier energy depositions in the crystal. However, we observe statistically-indistinguishable low-energy ($<$20 PE) background spectra from the 10 $\mu$s pre- and post-POT trigger intervals during these tests. This absence of a measurable difference over a 10 $\mu$s time shift is to be expected for few-PE signals, in view to the comparatively very long time constants involved in delayed phosphorescent de-excitations able to generate such faint signals (Fig.\ 3).}. The chosen waveform length allows to impose the same data cuts to pre-POT and post-POT signals. A first data cut is based on the presence of an excessive number of SPEs in the 40 $\mu$s preceding the 10 $\mu$s region being inspected. We find that imposing this condition allows for an efficient rejection of low-energy events associated to phosphorescence from a previous large energy deposition, while introducing only a modest dead time\footnote{The SPE rate observed from this PMT when matched to the 2 kg prototype detector is 6,000 Hz, most of it associated to phosphorescence, i.e., not uniformly distributed in time. The irreducible few-PE background rate in Fig.\ 13 is obtained by imposing the condition of zero PE in the 40 $\mu$s pre-trigger trace, which results in a 75\% live-time. Relaxing this condition to accept one PE in the pre-trigger trace increases this background rate by a factor of two.}. \\
Surviving events involving less than 20 PE in the 10 $\mu$s post-trigger window can be further analyzed by examining the pulse shape of their cumulative (i.e., time-integrated) PMT current distribution. A vast majority of these events display cumulative currents characteristic of a random grouping of their constituent photoelectrons, dissimilar to a radiation-induced scintillation pulse like those shown in Fig.\ 7. This is evidenced by their broad, featureless distribution of cumulative current rise-times, shown in Fig.\ 12. Additional data cuts can be imposed by exploiting this observation. A Monte Carlo simulation of the expected cumulative current from neutrino-induced CENNS events is used to define and optimize these cuts. The simulation employs as input the characteristic fast and slow scintillation components studied in Fig.\ 7. It also includes the measured dispersion in SPE current, as well as the post-trigger time-of-arrival of neutrino signals from the different neutrino flavors, prompt and delayed. A basic set of pulse-shape cuts can be imposed resulting in 46\% CENNS signal acceptance and 96\% rejection of backgrounds from randomly-grouped PEs. The resulting irreducible low-energy spectrum and accepted CENNS signal rate is shown in Fig.\ 13. We expect a further improved signal-to-background ratio from a more sophisticated treatment of this pulse-shape rejection (e.g., using neural networks \cite{neural} or wavelet analysis \cite{wavelets}).\\
A labVIEW-based DAQ software is optimized to write waveform bundles to file with 100\% throughput at the 60 Hz SNS trigger rate. The software inspects incoming waveforms in real time, not saving to file the $\sim$2/3 of events containing blank (zero PE) traces. The same program periodically compresses raw binary files, discarding uncompressed originals, resulting in an expected manageable 40 Gb/day data stream at the SNS. \\
The steady-state background achieved during these tests is sufficiently small to expect a statistically-significant evidence for a CENNS-induced excess in the residual between the energy spectra from 10 $\mu$s post-POT and 10 $\mu$s pre-POT time periods, following a reasonable (few year) exposure (Fig.\ 13, inset). A comparison between the presently measured concentration of $^{87}$Rb and $^{134,137}$Cs, and simulations of their contribution \cite{radiopurekims}, indicates that just a $\sim$5\% of the low-energy steady-state background measured in the 2 kg prototype originates from these internal radioisotopes\footnote{Accounting for the duty-cycle imposed by mimicking the SNS trigger and for the measured photoelectron yield per keV, the low-energy background shown in Fig.\ 13 is equivalent to $\sim$600 counts / keV kg day.}. A more exhaustive control of these, as is done in \cite{radiopurekims}, would therefore not lead to a significant enhancement of the CENNS signal-to-background ratio in a future larger CsI[Na] detector array. On the other hand, the fast (0.1-10 MeV) neutron flux in the shallow underground laboratory where these tests were performed has been measured at $\sim7\cdot10^{-4}$ n/cm$^{2}$s, using a combination of Bonner spectrometers. The exact energy distribution of these neutrons is unknown. Signals from their elastic scattering are expected to accumulate at few PE, due to the limited maximum recoil energy imparted to a heavy nucleus like Cs or I, and the few-percent quenching factor involved. A rough simulation of their contribution to the spectrum in Fig.\ 13 indicates that a significant further reduction in low-energy background may be achieved through the addition of neutron moderator against these environmental neutrons.\\
Fig.\ 14 shows the CENNS signal together with two beam-coincident neutron-induced signals able to produce competing nuclear recoils in the CsI[Na] crystal. The first is the high-energy neutron component escaping the SNS monolith already mentioned in the previous section, calculated here using a detailed GEANT simulation of neutron production at the SNS target, and its transport to the basement location described above. This simulation requires parallelization, event biasing, and up to $\sim$25,000 CPU-hours in a computer cluster, due to the long distance ($\sim 24$ m) of moderating materials traversed\footnote{Along the direct line of sight from the SNS target to the basement location considered, these are 5.2 m of monolith steel, 3.5 m of high-density concrete, 2 m of regular concrete, 13.2 m of compacted road base stone, and 70 cm of regular concrete \cite{private2}.}. This background calculation will be better informed by the ongoing measurements of neutron flux at the SNS, but it can already be observed to have negligible impact on a CENNS measurement in this basement location, or its equivalent. \\
The second competing signal in Fig.\ 14 arises from the charged-current reaction $^{208}$Pb($\nu_{e},e^{-}$)$^{208}$Bi in the lead shielding around the detector. This reaction is of great interest for supernova neutrino detection in this same energy range \cite{nspectra,halo}, and specifically for the ongoing HALO experiment \cite{halo2}. Its value has not been experimentally assessed yet. The excited $^{208}$Bi daughter is expected to rapidly decay generating up to three neutrons, detectable by the CsI[Na] crystal. The planned CENNS measurement provides an excellent opportunity to parasitically measure this separate neutrino cross-section. Neutrino-induced neutron spallation is expected to play a dominant role in heavy-element nucleosynthesis in supernovae: this additional measurement will contribute experimental information to an area that relies heavily on theoretical models \cite{snspa}.\\
The signal from $^{208}$Pb($\nu_{e},e^{-}$)$^{208}$Bi shown in Fig.\ 14 is calculated as follows\footnote{The cross-section for inelastic neutral-current scattering of $\nu_{x}$ off $^{208}$Pb is one order of magnitude lower than for this charged-current, and is neglected here. Accounting for contributions from all neutrino flavors, the neutral current reaction rate in Pb is expected to add to the charged current background studied here by $\sim$15\% \cite{katepr}.}. Neutron production rate in the 2,300 kg of lead shielding surrounding the detector is derived from Table II in \cite{nmultipl}, adjusting for the $\sim1.1\cdot10^{7}$ cm$^{-2}$ s$^{-1}$ $\nu_{e}$ flux expected 25 m away from the SNS target, at the farthest presently considered basement location. This input includes the expected multiplicity in neutron generation\footnote{Table II in \cite{nmultipl} contains two typos: it should read 16350 instead of 1630 (``Pb+1n" line, column ``10 m"), and 2400 instead of 1140 (``Pb+2n" line, column ``20 m") \cite{lazauskas}. These corrections have been taken into account.}. Following \cite{nspectra}, a simple evaporation energy spectrum terminated at 5 MeV is adopted for the initial neutron energies\footnote{An improved assessment of this emission spectrum would be desirable, in view of the factor 2-3 difference between neutrino energies at the SNS and those considered in \cite{nspectra}. However, based on the modest dependence on neutrino energy indicated in \cite{nspectra}, the estimates of Fig.\ 14 are not expected to change by much.}. These neutrons are then transported through a detailed MCNP-Polimi \cite{polimi} geometry of the detector assembly, resulting in 3.6\% of them producing nuclear recoils in the CsI[Na] crystal. For instances of multiple scattering in the crystal, recoil energies are added\footnote{This slightly overestimates the contribution from this process above detector threshold, since the quenching factor (Fig.\ 5) is expected to be monotonically decreasing with decreasing recoil energy.}. It should be kept in mind that an uncertainty by a factor $\sim$3 exists in the calculated cross-section for this charged-current process, specifically for the decay-at-rest $\nu_{e}$ energies expected at the SNS, depending on the nuclear model employed. Even if the $^{208}$Pb($\nu_{e},e^{-}$)$^{208}$Bi cross-section adopted here is one of the largest \cite{nmultipl,xsections,indian}, the CENNS signal is observed to dominate at energies near the detector threshold (Fig.\ 14). \\
While the spectral shape of the beam-coincident excess should be sufficient to disentangle $^{208}$Pb($\nu_{e},e^{-}$)$^{208}$Bi and CENNS contributions, there are additional strategies available to show the presence of two separate components. For instance, the innermost 1" of ULB lead shielding can be trivially replaced by low-background polyethylene neutron moderator during a second running period\footnote{This action is expected to bring up the observed steady-state low-energy backgrounds by just a few percent, due to the predictable increase in $^{210}$Pb bremsstrahlung \cite{vojtila} and reduction in external gamma shielding \cite{leadthickness}.}. The resulting considerable softening in energy of neutrons from the charged-current reaction in lead, leading to lower energy recoils, can be noticed in Fig.\ 14. The observation of the dissimilar signal rate changes predicted for the $\sim$7-25 keVnr and for the $\gtrsim$25 keVnr energy regions following this inner shield substitution (Fig.\ 14) can be used to confirm the presence of two separate neutrino-induced signals, each dominating in one of these energy ranges. In addition to this, a convenient fast neutron monitor internal to a CsI[Na] detector is provided by the inelastic scattering signal\footnote{This characteristic signal is observed to arise not only from muon-induced neutrons (Fig.\ 11), but also with a very high rate of incidence during DD neutron generator irradiations of CsI[Na] and NaI[Tl] \cite{myq2} (elastic and inelastic scattering cross-sections for 1-10 MeV neutrons in $^{127}$I differ only by a factor of $\sim$2).} at 57.6 keV (Fig.\ 11). Its rate in coincidence with the SNS beam can be used to provide an independent measure of the neutron flux from $^{208}$Pb($\nu_{e},e^{-}$)$^{208}$Bi reaching the CsI[Na] crystal. Another possibility is to measure this $^{208}$Pb($\nu_{e},e^{-}$)$^{208}$Bi channel independently and upfront, by placing a one-liter Bicron-501A liquid scintillator cell within the lead shield in lieu of the CsI(Na) crystal. Exploiting the neutron/gamma discrimination characteristic of this scintillator, it should be possible to identify approximately 40 events per month at the SNS basement from this neutrino-induced reaction, removing the uncertainties presently associated to this background.
\section{Conclusions}
\label{}
The COHERENT collaboration \cite{white} uses a multi-target approach to CENNS detection at the SNS, in this way maximizing the sensitivity of new investigations in neutrino physics, reachable through this unexploited channel. The collaboration presently contemplates the use of p-type point contact (PPC) germanium detector arrays \cite{ppc}, a two-phase liquid xenon detector \cite{red}, and CsI[Na] detectors like those described here. In addition to these, a recently decommissioned multi-ton array comprising 1,000 NaI[Tl] detectors previously used for national security applications, has recently become available at PNNL. An interesting use for these crystals, beyond the mentioned CENNS-related studies, would be an attempt to observe the pattern of $\nu_{\mu}$ oscillations into sterile neutrinos within the detector array itself, a possibility reachable for $\Delta m^{2}\gtrsim 7$ eV$^{2}$ \cite{osc}. The characteristics of these NaI[Tl] detectors (background, light yield, quenching factor) are being presently assessed by the collaboration.\\
To summarize, we have demonstrated the potential of CsI[Na] as a suitable target for CENNS studies at a spallation source. Necessary conditions of detector threshold, target mass, and background level, have been shown to be in place for a first measurement of this intriguing mode of neutrino interaction, following a short (few years) exposure with a 14 kg CsI[Na] crystal. It should be possible to extract a first measurement of the $^{208}$Pb($\nu_{e},e^{-}$)$^{208}$Bi cross-section, of interest for future supernova detection, from the same effort. The present feasibility study relies strongly on the availability of a detector location at the SNS with sufficient neutron moderating material in the direct line-of-sight to the target. The COHERENT collaboration is presently performing background studies in a number of detector emplacements potentially meeting this requirement.
\section{Acknowledgements}
\label{}
We are indebted to our colleagues from the COHERENT collaboration for many useful exchanges, in particular to P.S. Barbeau, Yu.\ Efremenko and K. Scholberg, and to I. Lawson for facilitating many sample measurements at SNOLAB. N.E.F. was supported by the NNSA Stewardship Science Graduate Fellowship program under grant number DE-FC52-08NA28752. T.W.H. was supported by the Intelligence Community Postdoctoral Research Fellowship Program. Additional support was received from the Kavli Institute for Cosmological Physics at the University of Chicago through grant NSF PHY-1125897, and an endowment from the Kavli Foundation and its founder Fred Kavli. J.I.C. acknowledges the hospitality of the International Institute of Physics at the Universidade Federal do Rio Grande do Norte during the completion of this manuscript. This work was completed in part with resources provided by the University of Chicago Research Computing Center.
\bibliographystyle{elsarticle-num}
|
\section{INTRODUCTION}
\subsection{Asymptotic Giant Branch Stars and M32}
The observational properties of stars near
the main sequence turn-off (MSTO) have a natural primacy in studies
of the star-forming history (SFH) of galaxies. However,
with the exception of stars that formed within the past few hundred Myr,
crowding and the intrinsic faintness of the MSTO restrict its use
to Local Group galaxies. Core helium-burning stars provide an alternate means of
probing the stellar content of more distant systems (e.g. McQuinn et al. 2011),
although these stars can only be resolved in relatively low
surface brightness areas of galaxies in the nearest groups.
Stars evolving near the tip of the Asymptotic Giant Branch (AGB) are among the
most luminous objects in a galaxy, and so offer another means of
investigating stellar content. The stars on the
upper few magnitudes of the AGB can have a wide
range of ages, extending back $10+$ Gyr for solar
metallicities, and so have the potential to provide constraints on the
SFH over a large fraction of the age of the Universe. The red colors of the most
luminous AGB stars make them well-suited for study with adaptive
optics (AO)-equipped ground-based telescopes, since the
corrections delivered by AO systems improve towards longer wavelengths.
Because they are very bright and have colors that are much
redder than the majority of other luminous stars in a galaxy,
AGB stars will be resolved in the near and mid-infrared with future space-based
facilites -- such as the JWST -- in systems that are well outside of the Local Group.
Still, given their highly evolved nature, models of AGB stars
can be uncertain (e.g. discussion by Marigo et al. 2008), and there has been mixed
success matching observations with model predictions (e.g. Zibetti et al. 2013;
Melbourne et al. 2012; Girardi et al. 2010). This has motivated recent efforts to
characterize the behaviour of AGB stars in dwarf galaxies that span a broader range of
metallicities than is sampled by the Magellanic Clouds and the Milky-Way, which
have heretofore been the key calibrators of AGB properties (e.g. Rosenfield
et al. 2014; Melbourne et al. 2010).
The Local Group compact elliptical (cE) galaxy
M32 is a fundamental laboratory for probing AGB evolution
and investigating the role that these stars play
in stellar content studies. M32 is close enough to allow
the MSTO stars that probe the early phases of its evolution to be resolved,
so that its SFH can be probed in detail (e.g. Monachesi et al. 2012). At the same time,
M32 is also far enough away and has a high enough stellar density that integrated
spectra that are not subject to stochastic sampling effects can be recorded
in its high surface brightness central regions. That a rich population of
AGB stars in M32 can be resolved over a wide range of radii (e.g.
Davidge \& Jensen 2007) also means that the problems associated with
low number statistics that plague studies of AGB stars in more diffuse and/or
less massive systems are avoided.
The utility of M32 for testing techniques to probe stellar content has
been amply demonstrated in the past. M32 has been the target of numerous
spectroscopic studies, and these have revealed signatures of a substantial --
possibly even dominant -- population with an age of a few Gyr.
Absorption lines originating from the Balmer series of Hydrogen
-- most noteably H$\beta$ (e.g. O'Connell 1980; Bica et al. 1990; Rose
1994; Schiavon et al. 2004) -- and the first-overtone CO bands (Davidge 1990)
have depths in the integrated spectrum of M32 that are indicative of a
large intermediate age population. This has been confirmed by the subsequent
study of resolved AGB (e.g. Davidge 2000; Davidge \& Jensen 2007) and MSTO (Monachesi
et al. 2012) stars. Spectroscopic studies also reveal a near solar
luminosity-weighted metallicity. This is consistent with the metallicities of
individual AGB stars in the central regions of M32 (Davidge et al. 2010) and the
mean photometric properties of red giant branch (RGB) stars at large
radii (Grillmair et al. 1996; Monachesi et al. 2012).
\subsection{The Evolution of M32}
The origins of M32 are enigmatic.
While the structural characterics of M32 are reminiscent of a
classical elliptical galaxy (e.g. Kormendy 1985), its near-solar
metallicity is consistent with it being the remnant of what was once a much
larger galaxy. In addition to having a relatively high mean
metallicity for its mass, the location of M32 on the Fe $vs.$ Mg$_2$ plane
is indicative of a [Mg/Fe] that is lower than in most early-type galaxies (Worthey
et al. 1992). This hints that SNe I played a significant role in
the chemical enrichment of M32. This is not expected for
spheroidal systems that experienced rapid chemical enrichment, but is associated
with disks that are stable over time scales of at least a few hundred Myr.
M32 and M31 are in close physical proximity at present, with a projected
separation of only a few kpc (e.g. Evans et al. 2000; Sarajedini et al.
2012). This raises the prospect of past interactions between them, and
Bekki et al. (2001) suggest that M32 is the remnant of
a disk-dominated galaxy that was threshed by tidal interactions with M31.
If the progenitor of M32 was a disk galaxy then this could explain the chemical
mixture of M32. The tidal threshing model could also explain
other properties of M32, such as the lack of a globular cluster system, its low gas
content (Sage et al. 1998; Grcevich \& Putman 2009), and the uniform mixing of RR
Lyraes with other stellar types (Sarajedini et al. 2012).
There may have been more recent encounters between M31 and M32. Block et al.
(2006) argue that a satellite of M31 -- presumably M32 -- passed through the inner
regions of its disk a few hundred Myr ago. Such an encounter is consistent with
the complex dynamical state of the ISM near the center of M31 (e.g. Melchior \& Combes
2011). Smith et al. (2012) find that the dust properties of the central few kpc of
M31 may also have been affected by an interaction, and note that the
dust mass within the central 3 kpc is comparable to that expected in a dwarf galaxy.
The concentration of bright AGB stars along one arm of the M31 major axis
(Davidge et al. 2012; Davidge 2012; Dalcanton et al. 2012) may also have formed during
this event. A basic prediction of the Block et al. (2006) model
is that the interaction with M32 triggered disturbances in the ISM
of M31 that are propogating outwards, spurring star-formation as they go. The presence
of such a propogating density wave is not consistent with the long-lived nature
of the star-forming rings (Davidge et al. 2012; Dalcanton et al. 2012), and
may have difficulties explaining related shell-like structures in the outermost
regions of the disk (Fritz et al. 2012).
Monachesi et al. (2012) use deep images of the outer regions of M32
to resolve stars near the MSTO and examine the SFH. They conclude that
M32 is not coeval. There is a dominant component that is older than 5 Gyr, and
a significant population with an age 2 -- 5 Gyr. The MSTO of the oldest stars
is not detected, and so firm constraints can not be placed
on the mass fraction that formed during very early epochs. A population of stars that
is younger than 2 Gyr and accounts for a few percent of the total stellar mass
was also detected. While MSTO stars are primary probes of stellar age, blue
stragglers, and contamination from clusters of young stars in the outer disk of M31 add
uncertainties to age estimates obtained from the MSTO, and may account for
at least some fraction of this relatively young component.
\subsection{The Present Study}
Obtaining agreement between the SFHs deduced from AGB and
MSTO stars in nearby galaxies is an important step if
AGB stars are to be considered as probes of more distant systems.
There are challenges inherent to detecting the most evolved
AGB stars in the visible, even though this is a wavelength region
that has been explored in detail and is where space-based
facilities typically deliver their best angular resolution. The most
evolved RGB and AGB stars with moderate and higher metallicities have cool photospheric
temperatures that are conducive to the formation of
deep absorption bands of TiO and ZrO at visible and red wavelengths,
and blanketing from these species forces the RGB and
AGB sequences in the color-magnitude diagrams (CMDs) at visual wavelengths to slump
over (e.g. see discussion in Ortolani et al. 1990). This complicates efforts to measure
the luminosities of the brightest AGB stars, while also making AGB-tip stars more
susceptible to blending with sources that have lower intrinsic luminosities. The
detection of highly evolved AGB stars at visible wavelengths may also be affected by
obscuration from circumstellar dust.
Many of the challenges that are inherent to the detection of bright AGB stars
at visible wavelengths are eased when observing in the near- and mid-infrared
(NIR and MIR). The effects of line blanketing and circumstellar
obscuration are significantly reduced at these wavelengths.
In addition, cool AGB stars are significantly brighter
in the infrared than the bulk of the underlying population that contributes most of the
stellar mass. When observed in the NIR and MIR, the coolest AGB stars are thus
less susceptible to blending and can be used to probe regions in galaxies where MSTO
stars can not be detected. In the particular case of M32, AGB stars have been resolved
in $H$ and $K$ by AO-equipped ground-based telescopes to within a few arcsec
of the galaxy center (e.g. Davidge et al. 2000, Davidge \& Jensen 2007;
Davidge et al. 2010). This is a part of the galaxy where there is no hope of resolving
stars near the MSTO in the foreseeable future.
The high emissivity associated with ground-based facilities is a formidable
obstacle for studying all but the brightest sources in the MIR. While deep MIR
observations are the exclusive purview of space-based facilities, these
same facilities have poorer angular resolution than ground-based facilities due
to their modest apertures. The superior angular resolution offered by
ground-based facilities provides a niche that can be exploited, at least at the
shortest MIR wavelengths.
In the present study, data obtained with the
Canada-France-Hawaii Telescope (CFHT), Gemini North (GN), and the SPITZER satellite
are used to investigate the NIR and MIR photometric properties of (1) the most luminous
stars in M32 and (2) the isophotal properties of the integrated light.
The data sample wavelengths where a significant fraction of the light
from the coolest stars is emitted. The sample of resolved stars is comprised mainly
of objects evolving on the upper AGB, as crowding in the MIR
data effectively prevents the detection of RGB stars (Section 4.2).
The GN observations are of particular interest for isophotal
studies as they explore the central regions of the galaxy
between 3 and $5\mu$m with sub-arcsec angular resolution, which is not possible
with the current generation of space-based infrared facilities.
Wavelength-dependent differences in isophotal characteristics such as eccentricities
and fiducial radii could point to age gradients and/or distinct structural components.
A distance of $770 \pm 40$ kpc for M32, measured from
the brightness of the RGB-tip by Evans et al. (2000) and
corresponding to a distance modulus $\mu_0 = 24.43 \pm 0.11$,
is adopted here. Our conclusions would not change
if the distance modulus were varied by $\sim \pm 0.1$ magnitude. A foreground reddening
of A$_V = 0.17$, taken from the Schlafly \& Finkbeiner (2011) re-calibration of the
Schlegel, Finkbeiner, \& Davis (1998) maps, is also used throughout this paper.
The absence of warm and cool interstellar dust (Gordon et al.
2006) and the colors of RR Lyraes (Sarajedini et al. 2012)
suggests that there is no internal reddening. Finally, the `average' reddening
law listed in Table 1 of Indebetouw et al. (2005) is adopted for MIR wavelengths.
Details of the observations and their reduction
are described in Section 2, while the procedures used to construct, calibrate, and
characterize the photometric measurements are discussed in Section 3. The CMDs obtained
from these data are presented in Section 4, and the NIR and MIR
luminosity functions (LFs) are compared with models in Section 5.
The isophotal properties of M32 in the NIR and MIR
are examined in Section 6. The paper closes in Section 7
with a brief discussion and summary of the results.
\section{OBSERVATIONS \& DATA REDUCTION}
\subsection{SPITZER IRAC Data}
Post Basic Calibrated Dataset (PBCD) mosaics of IRAC (Fazio
et al. 2004) images in [3.6], [4.5], [5.8], and [8.0] from
program ID 3400 (PI Rich) were downloaded from the NASA/IPAC Infrared Science
archive\footnote[1]{http://irsa.ipac.caltech.edu/Missions/spitzer.html}.
The [5.8] mosaic of M32 and the surroundings is shown in Figure 1.
The data were recorded with on-sky offsets to cover a field
that is larger than that sampled by a single exposure, and the total
exposure time varies across the mosaiced field.
Only the area in the mosaics that has the largest effective
exposure time, and hence the deepest and most uniform
photometric characteristics, was considered for this study. The extracted M32 field
covers a $5.0 \times 5.4$ arcmin$^2$ area that is
centered near $\alpha$ = 00:42:56, $\delta$ = 40:45:48 (E2000).
The images in all four IRAC channels were used to obtain the
brightnesses of individual bright stars in and around M32, thereby
allowing MIR CMDs and LFs to be constructed. In addition, the
[3.6] and [4.5] images were used to examine the isophotal properties of M32.
The [5.8] and [8.0] images were not considered for this latter analysis as
the isophotal properties of M32 could not be traced
to radii $> 1$ arcmin in those images.
The geometry of the IRAC detector arrays allows parallel [3.6] and [5.8]
observations to be recorded of a field that samples the outer disk
of M31 directly to the east of M32. This field also overlaps with moderately
deep archival WIRCam data (Section 2.2). A sub-section of this field that covers
a $5.1 \times 5.2$ arcmin$^2$ area and is centered near $\alpha$ = 00:42:56,
$\delta$ = +40:45:48.3 (E2000) was used to assess contamination from
the M31 disk in the M32 IRAC data. The location of the M31 outer disk field
is indicated in Figure 1.
Parallel observations that sample a field to the west
of M32 were also recorded in [4.5] and [8.0]. This parallel field samples
an area that is closer to the main body of M31 than M32. The stellar
content is less homogeneous in this field because of increased contamination
from the main disk of M31, raising the concern that it
may not be representative of the M31 disk near M32. Therefore, this
parallel field was not considered when assessing disk contamination.
\subsection{CFHT WIRCam Data}
$J$ and $K$ photometry of stars in M32 and the outer
disk of M31 were obtained from archival CFHT WIRCam (Puget et al. 2004) images.
The detector in WIRCam is a mosiac of four $2048 \times 2048$ HgCdTe arrays
arranged in a $2 \times 2$ format with 45 arcsec gaps between arrays.
Each WIRCam pointing covers a $21 \times 21$ arcmin$^2$ field
with 0.3 arcsec pixel$^{-1}$ sampling.
The images used here are of pointings M31-401 and
M31-405 from programs 2009BC29 (PI: Sick) and 2009BH52 (PI: Tully).
Twenty-five 20 sec $J$ exposures and twenty-six 20 sec $Ks$ exposures
that had been processed with the I'IWI pipeline were downloaded from
the CFHT archive at the Canadian Astronomical Data Centre
\footnote[2]{http://www1.cadc-ccda.hia-iha.nrc-cnrc.gc.ca/en/cfht/}. These
were aligned to correct for on-sky offsets and the
median of the pixel intensity distribution, which serves as a proxy for the
mean sky level, was subtracted from each frame. The results were then stacked by
finding the median intensity at each pixel location, trimmed
to the area of common angular coverage, and transformed to the
orientation of the IRAC images using the GEOMAP and GEOTRANS tasks in IRAF.
Stars in the final images have 0.7 arcsec FWHM.
\subsection{Gemini NIRI Data}
Ground-based observations in the MIR are hampered by high levels
of emissivity, which mushrooms in size at wavelengths longward of $2.5\mu$m to produce
noise of monumental proportions. The high background levels hinder efforts to
study all but comparatively bright objects in the MIR with ground-based telescopes.
This severe handicap notwithstanding, the large apertures of ground-based facilities
results in angular resolutions that surpass those delivered by
space-based telescopes, providing a niche for ground-based MIR observing.
Ground-based observations are at present the only means of obtaining
sub-arcsec angular resolution images in the MIR.
The central regions of M32 were observed with NIRI (Hodapp et al. 2003)
through $Ks$, $L'$, and $M'$ filters as part of queue program GN-2006B-Q-64
(PI: Davidge). The NIRI detector is a $1024 \times 1024$ InSb array.
The f/32 camera was used for these observations, and each exposure thus samples a $22.5
\times 22.5$ arcsec$^2$ field with 0.022 arcsec pixel$^{-1}$ sampling.
\subsubsection{$Ks$}
The $Ks$ images were corrected for wavefront distortions using the GN
facility AO system ALTAIR (Herriot et al. 2000). ALTAIR was used in natural
guide star (NGS) mode, with the semi-stellar nucleus of M32 serving as the NGS.
Twenty images, each consisting of three coadded 5 sec exposures,
were recorded. A dither pattern that samples the corners of a $3 \times
3$ arcsec square was used to facilitate the identification and suppression of
cosmic rays and cosmetic defects. Five exposures were recorded at
each dither position. The telescope was offset to one of four background fields
that are located 90 arcsec to the North, South, East, and West of the galaxy
nucleus after each dither position was observed. These background fields monitor
the sky level, and allow a calibration frame to be constructed that can be used
to remove thermal emission from warm sources along the optical path (see below).
The reduction of the $Ks$ images followed standard procedures for NIR
imaging, and utilized three calibration frames. The first of these
monitors the dark current, and this was constructed by median-combining
frames taken of the unilluminated array that have the same exposure times as
the science data. The resulting dark frame was subtracted from each image.
The second calibration frame monitors flat-field variations, and
was constructed from a series of exposures that involve the Gemini
facility calibration unit (GCAL). A series of images with the GCAL light
source turned on were recorded, and these were followed
by another sequence in which the GCAL light source was turned off. The first set of
images contain signatures of the flat-field pattern (multiplicative in nature)
combined with thermal emission (additive in nature), whereas the second set contains
only signatures from thermal emission. Taking the difference between the two sets of
exposures thus removes the thermal emission, leaving behind the flat-field
information. The difference between the two sets of images was normalized to
unity and the dark-subtracted images were divided by the resulting flat-field frame.
A third calibration frame was constructed to remove thermal
signatures that are produced by warm objects along the optical path, such
as dust on exposed optical surfaces. Such objects glow in the infrared and are out
of focus, producing diffuse signatures that usually become evident only after
flat-fielding. Flat-fielded images of the background fields were median-combined
after the mean sky level was subtracted from each image to remove variations
in the sky brightness to construct a thermal contamination calibration frame.
The result was subtracted from the flat-fielded M32 images.
The images that were corrected for thermal emission were aligned
and combined by taking the median intensity at each
pixel location. The result was trimmed to the area common to all exposures. Stars in
the final $Ks$ image have FWHM = 0.09 arcsec.
\subsubsection{$L'$ and $M'$}
ALTAIR could not be used for the $L'$ and $M'$ observations
as it has transmissive optical components that introduce background
signal at wavelengths longward of $2.5\mu$m that saturates the NIRI
detector. Even without ALTAIR, the high ambient background in $L'$ and $M'$
necessitates the use of short exposures to prevent saturation.
A single observation at each dither position in $L'$ consisted of
30 co-added 0.6 sec exposures, while a single observation in $M'$ consisted of
nine exposures, each of which was made up of 34 co-added 0.5 sec exposures.
The $L'$ and $M'$ observations were recorded with a 5 point dither pattern that
samples the four corners and the middle of a $3 \times 3$ arcsec square.
The dither pattern was repeated seven times in each filter, and the
total on-source exposure times are 630 sec in $L'$ and 5355 sec in $M'$.
One of four background fields located along the cardinal axes was observed following
each complete observation at a dither position.
The construction of flat-field frames in $L'$ and $M'$ is
problematic given the high levels of thermal emission in the MIR.
In any event, thermal noise dominates in these filters, and so the
primary processing step is the removal of thermal emission signatures. The mean
intensity and noise pattern in the background field observations varies from
exposure-to-exposure, with the largest variations in $M'$.
To account for this variability, a calibration image was constructed
for each M32 exposure by linearly interpolating between successive background field
observations. The interpolated images were then subtracted from the M32 exposures.
The resulting background-subtracted images have substantially lower residual
noise than those constructed using either the preceeding or subsequent background
observation. Even so, residual thermal signal at the few percent level remains,
indicating that the variations in the thermal background
vary in a manner that is more complicated than the linear trend assumed here.
The background-subtracted $L'$ and $M'$ frames were aligned and the median
intensity at each pixel was found after a residual mean sky level, which was
measured near the edges of the field, was subtracted from each image.
The $M'$ images were aligned using the nucleus of M32, which is the only source
detected in individual frames. The relatively modest signal from the
center of M32 in $M'$ introduces $\leq 0.1$ arcsec (i.e. a few pixels)
uncertainties in the alignment of the individual $M'$ exposures.
The final processing step was to trim the stacked $L'$ and $M'$ images to the
area that is common to all pointings. Individual stars are
seen in the final $L'$ image, and these have FWHM = 0.18 arcsec.
As for the $M'$ images, these were recorded over two nights during only fair
seeing conditions, and individual stars are not detected.
The angular resolution of these data, measured from observations of standard
stars that were observed throughout the M32 observing sequence, is 0.40 arcsec FWHM.
\section{PHOTOMETRIC MEASUREMENTS}
\subsection{SPITZER Data}
The IRAC Instrument Handbook \footnote[3]
{http://irsa.ipac.caltech.edu/data/SPITZER/docs/irac/iracinstrumenthandbook/}
advises against the use of point spread function (PSF)-fitting to extract photometry
from IRAC images, with non-uniformities in pixel sensitivity and the
undersampling of the PSF cited as concerns. Still, numerous studies of the stellar
contents of nearby galaxies have used PSF-fitting to extract photometry from
IRAC images (e.g. Meixner et al. 2006, McQuinn et al. 2007; Mould et al. 2008), as it
is an accepted means of performing photometry in crowded
environments. To assess the best means of measuring stellar
brightnesses in the current study, two sets of photometric measurements
were made from the [3.6] and [5.8] images: (1) aperture photometry with brightnesses
obtained from the PHOT routine in the DAOPHOT (Stetson 1987) package, and (2)
PSF-fitting photometry using ALLSTAR (Stetson \& Harris 1988).
Sources within 90 arcsec of the center of M32 were not
photometered, as surface brightness fluctuations in this high-density region
can masquerade as individual stars. Both sets of photometry were calibrated using
flux zeropoints calculated by Reach et al. (2005).
Crowding is an obvious concern when conducting aperture photometry in
external galaxies. To mitigate against crowding, the aperture measurements were made
with a comparatively narrow 3 pixel (1.8 arcsec) radius. Final
brightnesses were then obtained by applying an aperture correction that was
measured from stars in uncrowded parts of the images. An additional concern
for crowded extragalactic environments is that non-uniformities in the integrated
surface brightness can occur over angular scales that are comparable to the area
where the local sky level is measured, and these introduce errors into
both aperture and PSF-fitting measurements. To correct for localized
sky structure, a running boxcar median filter with dimensions that are sufficient to
suppress stars was applied to the images, and the resulting smoothed image
was subtracted from the data.
PSFs were constructed using tasks in DAOPHOT (Stetson 1987).
A single PSF was constructed for each filter, and faint objects close to
PSF stars were subtracted out using progressively improved versions of the PSF.
Once a final PSF was obtained then the final photometric measurements were
made with the PSF-fitting routine in ALLSTAR (Stetson \& Harris 1988).
The PHOT and ALLSTAR routines estimate the uncertainty in photometric
measurements using the properties of the source and the surrounding background.
The photometric measurements of objects that are in crowded environments and/or
that are non-stellar in appearance have higher uncertainties than those in
uncrowded environments that have a star-like light profile.
To remove sources with problematic photometry, objects that
depart from the general trend on the error $vs.$ magnitude relation for each filter
were removed from the photometric catalogue. This parallels the procedure
applied by Davidge (2010) to PSF-fitting photometry, where objects were rejected
using the uncertainty computed by ALLSTAR. Davidge (2013) found that the properties
of the final sample culled using photometric error are similar to those that
result when other object characteristics, such as image sharpness, are
used to identify sources with problematic photometry.
The $([5.8], [3.6] - [5.8])$ CMDs obtained from aperture and PSF-fitting
measurements for sources in M32 and the M31 disk field are compared in Figure 2.
The standard deviation about the mean [3.6]--[5.8]
for objects with [5.8] between 15.75 and 16.25, $\sigma_{AGB}$,
is shown for each CMD. This brightness interval samples AGB stars in a
heavily populated part of each CMD, and is well above the faint limit. An
iterative $2.5\sigma$ rejection filter was applied to suppress outliers when
computing $\sigma_{AGB}$.
The CMDs constructed with PSF-fitting sample a larger number of
objects and go significantly fainter than those obtained from
aperture measurements. While the differences between $\sigma_{AGB}$
computed from the aperture and PSF-fitting photometric measurements are modest,
the AGB sequence when [5.8] $< 15$ in the PSF-fitting CMD of M32
is better defined than in the aperture photometry CMD. Given (1) the
differences in the number of objects between the two sets of CMDs, and (2) the
tightness of the bright end of the M32 CMD obtained with PSF-fitting,
then PSF-fitting is adopted as the prefered photometry technique.
\subsection{WIRCam and Gemini NIRI Data}
The PSF-fitting program ALLSTAR was used to obtain photometric measurements
from the WIRCam and NIRI images. As with the IRAC observations, the PSFs were
constructed from bright, isolated stars, and an iterative scheme was applied to suppress
faint neighboring stars. No attempt was made to photometer objects within 90
arcsec of the galaxy center in the WIRCam images, as crowding confounds efforts to
resolve individual stars there unless the data have an angular resolution
$\lesssim 0.1$ arcsec FWHM (e.g. Davidge et al. 2000).
The WIRCam photometry was calibrated using zeropoints obtained from
standard star observations that were recorded during the same semester, and
this information is available on the CFHT web site \footnote[4]
{http://www.cfht.hawaii.edu/Instruments/Imaging/WIRCam/WIRCamStandardstars.html}.
The NIRI photometry was calibrated using zeropoints obtained from standard star
observations that were recorded as part of the baseline calibration set for
this program.
The IRAC observations have angular resolutions that are significantly
poorer than the WIRCam data. The WIRCam data can then be used to assess in
a preliminary way the effect of crowding on the MIR data. Emphasis is placed on the
[3.6] data for these experiments, as (1) this passband is closest in wavelength
to the WIRCam observations, and (2) at longer wavelengths crowding is expected to
become less of an issue because of the increased contrast between the brightest AGB
stars and the vast majority of other stars in M32.
Stars in the IRAC [3.6] observations have a FWHM of 1.8 arcsec, and
so the $J$ and $Ks$ WIRCam images were convolved with a Gaussian that degrades
the angular resolution to match that of the [3.6] observations.
The $(K, J-K)$ CMDs of objects between 1.5 and 2.5 arcmin distance from the
center of M32 obtained from the unsmoothed and smoothed WIRCam images are compared in
Figure 3. While there are obvious differences near the faint end, there is reasonable
agreement at the bright end, most noteably near the AGB-tip.
The contrast between AGB stars and the vast body of
fainter stars increases towards longer wavelengths, and so
the method described above is a conservative approach for assessing
the impact of crowding. It is thus significant that the comparisons in
Figure 3 suggest that the angular resolution of the [3.6] data should not skew
the properties of the brightest stars in the outer regions of M32. Additional evidence
to support this claim is presented in Section 4.2.
\section{RESULTS: CMDs}
\subsection{Near-Infrared CMDs}
The $(K, J-K)$ CMDs of the M31 outer disk and
of three annuli centered on M32 that sample equal areas on the sky
are shown in Figure 4. The inner and outer boundaries of the region where
resolved stars in M32 are investigated are indicated in Figure 1.
The dominant plume of objects in each CMD is populated by
AGB stars, and the $J-K$ color and peak brightness of the AGB sequence are
consistent with those measured by Davidge \& Jensen (2007).
It should be recalled that the WIRCam CMDs sample a much larger portion of M32 than
those presented by Davidge \& Jensen (2007). The peak $K$ brightness in the Inner
annulus is $K \sim 15.5$, which is similar to the peak brightness near
the center of the galaxy (e.g. Davidge et al. 2000). The absence of a
radial trend in the peak $K$ brightness is one indicator that the
photometry of the brightest stars is not affected by crowding.
The M31 CMD extends to fainter magnitudes than the M32 CMDs, owing
to the lower stellar density in that field.
A significant fraction of the brightest AGB stars near the center of M32 are
long period variables (LPVs) with light curve amplitudes in $K$ that may exceed two
magnitudes (Davidge \& Rigaut 2004). Such variability complicates efforts to determine
ages from CMDs using the brightest stars. To mitigate the impact of such variability, in
the remainder of this section emphasis is placed on the photometric properties of stars
that are midway down the AGB on the CMD, where the affects of
LPV-like variations should average out. The SFH of M32 is investigated
in Section 5 using model LFs that account for LPV behaviour.
Basic insights into the stellar contents of M32 and the outer disk of M31 can
be gleaned by comparing the mean colors of the AGB sequences. To this end, mean
$J-K$ colors were computed for objects with $K$ between 16.75 and 17.25. This
magnitude interval samples an area of the CMD that is well above the faint limit and is
richly populated in both M32 and M31. A $2.5\sigma$
iterative rejection filter was applied to suppress outliers.
The mean $J-K$ colors in the CMDs of the Inner, Middle, and Outer annuli
of M32 are $1.255 \pm 0.010$, $1.277 \pm 0.011$, and $1.292 \pm 0.015$,
where the uncertainties are the standard error of the mean.
Modelling of long-slit spectra suggests that the
stellar content of M32 changes with radius (Worthey 2004; Rose et al. 2005; Coelho et
al. 2009). Still, the systematic absorption line gradients that are typical of
large pressure-supported galaxies (e.g. Davidge 1992; Davies et al.
1993; Sanchez-Blazquez et al. 2006) are not seen
(Davidge 1991; Davidge et al. 2008, but see also Hardy et al. 1994).
Some of the features that do vary with radius in M32 show behaviour that is
contrary to what might be expected when compared with classical elliptical galaxies
(e.g. Davidge, de Robertis \& Yee 1990).
The progressive increase in $J-K$ towards larger radius in M32
is contrary to what is expected in the absence of large-scale spectroscopic
gradients. Rather than sampling an instrinsic radial change in the properties of M32,
it is likely that the radial increase in mean $J-K$ is due to contamination from
stars in the M31 disk. The mean $J-K$ color in the M31 disk field is $1.343 \pm 0.010$,
and the fractional contamination from M31 disk stars in the M32 CMDs increases towards
larger radii. Indeed, $K-$band surface brightness measurements of M32 from the 2MASS
Large Galaxy Atlas (Jarrett et al. 2003) indicate that the number of stars in the
Outer annulus that belong to M32 is 75\% lower than in the Inner annulus. However, the
Outer annulus contains 40\% fewer stars than in the Inner annulus, indicating that
the fractional contribution from M31 disk stars grows with increasing radius.
Adding an AGB component with a color like that in the M31 disk field to a
population with a mean color like that in the Inner annulus will produce a
radial trend in the mean M32 $J-K$ colors in the same sense as that found here.
The mean $J-K$ color of AGB stars in M32 is bluer on average than in the
M31 disk, suggesting differences in mean metallicity and/or mean age.
Quantitative insights into the stellar contents of M32 and the M31
outer disk can be obtained by making comparisons with isochrones, and
the CMDs of the M31 outer disk and the M32 Middle annulus are compared
with model sequences from Marigo et al. (2008) in Figure 5. These
sequences, and all other models used in this paper, were
downloaded from the Padova Observatory website \footnote[5]
{http://stev.oapd.inaf.it/cgi-bin/cmd}, and are based on the most recent versions
of the models constructed by this group that include the thermally-pulsing (TP)
phase of AGB evolution. This phase of evolution is critical for understanding the
MIR properties of the most evolved AGB stars.
The models shown in Figure 5 have Z = 0.016 and
ages 1 Gyr, 3 Gyr, and 10 Gyr. Based on previous estimates of the metallicity of
M32, it is anticipated that the number of C stars will be modest, and
the majority of stars will be oxygen-rich M giants. Thus, the
60\% AlOx $+$ 40\% silicate composition model discussed by Groenewegen (2006)
is assumed for circumstellar dust. While the chemical composition of the circumstellar
disk is of only secondary importance for the NIR photometric properties of AGB stars,
this is not the case in the MIR (Section 4.2).
The 3 Gyr isochrone tracks the ridgeline of the M32 sequence in the right hand
panel of Figure 5, while the 1 and 10 Gyr isochones more-or-less follow the
blue and red envelopes of the AGB sequence. In contrast, the majority of bright stars
in the M31 outer disk have isochrone-based ages $\geq 3$ Gyr when compared with the
solar metallicity sequences. That the AGB stars in the M31 field tend to have colors
that are suggestive of a sizeable old population is consistent with the SFH measured in
other fields in the outer disk of M31 (e.g. Davidge et al. 2012; Bernard et al. 2012),
including close to M32 (Monachesi et al. 2012).
A fiducial Galactic RSG sequence, constructed from the entries in Table 5
of Koornneef (1983), is included in the left hand panel of Figure 5.
The objects that populate the plume with $J-K \sim 0.6$ and
$K < 16$ in the M31 outer disk CMD have colors that are consistent
with them being G and K supergiants, indicating that the outer disk
field contains moderately young stars. Relatively young stars are expected
in this part of M31 as (1) there are areas of recent star formation
in nearby fields (van den Bergh 1964; Hodge 1979), and (2) members of other
groupings may drift into this field due to the motions that
these objects attain as they interact with large interstellar
clouds. Significant mixing of stars in the low-density outer regions of disks
can occur over time scales of only a few tens of Myr (e.g. Davidge et al. 2011).
\subsection{Mid-Infrared CMDs}
The $([5.8],[3.6]-[5.8])$ CMDs of objects in M32 and the M31 outer disk
are shown in Figure 6, while the $([8.0], [4.5]-[8.0])$ CMDs of M32 are
shown in Figure 7. The majority of stars in these CMDs are evolving on the AGB,
and there are few -- if any -- RGB stars. The Marigo et al. (2008)
models predict that the red giant branch (RGB)-tip in solar metallicity
systems with ages $> 3$ Gyr has an absolute magnitude $\geq -7$ in each IRAC filter,
which corresponds to apparent magnitudes $\geq 17.5$ in M32.
Given that the number density of RGB stars is $\sim 4 - 5 \times$
higher than that of AGB stars, then it is perhaps not unexpected that
the faint limit of the $([5.8],[3.6]-[5.8])$ CMDs coincides with the expected
location of the RGB-tip, as the resulting increase in crowding near the RGB-tip
will impose a limiting magnitude if individual RGB stars can not be resolved.
The vertical sequence with [3.6] -- [5.8] $\sim 0$ in the left
hand panel of Figure 6 contains a mix of foreground main sequence stars, evolved
stars in M31 and M32 that do not have circumstellar envelopes, and extragalactic
sources. This plume is less well populated in the M32 $([5.8], [3.6]-[5.8])$ CMDs
due to the smaller sky coverage of each annulus. The number of
extragalactic sources in the CMDs can be estimated from
the source counts given by Fazio et al. (2004). Assuming a
uniform distribution on the sky then there should be $\sim 74$ extragalactic
objects in the M31 outer disk field with [5.8] between 15.5
and 16.5, and 56 objects towards M32. For comparison, there are 359 objects detected
in this magnitude range in the M31 outer disk field, and 432
towards M32. Background galaxies thus account for 10 -- 20\% of the objects in
the $([5.8],[3.6]-[5.8])$ CMDs. These are upper limits to extragalactic
contamination as ALLSTAR rejects sources that are obviously non-stellar.
The M31 outer disk field contains objects with [5.8] $= 13.5 - 14.0$ that
span a broader range of $[3.6]-[5.8]$ colors than those in the M32 CMDs.
There is evidence that many of the brightest objects in the M32 Middle and Outer
annulus CMDs likely belong to the M31 disk. In particular, outside of the region within
90 arcsec of the center of M32 where individual stars are most susceptible to
blending, objects with $[5.8] \geq 14$ appear to have a more-or-less uniform
distribution, as expected if they belong to the outer disk of M31.
The $([5.8],[3.6]-[5.8])$ and $([8.0],[4.5]-[8.0]$) CMDs of the Middle annulus
of M32 are compared with Z=0.016 isochrones from Marigo et al. (2008) in Figure 8.
While the 1 and 3 Gyr isochrones match the ridgeline of the M32 $([5.8],[3.6]-[5.8])$
CMD, the agreement with the brightest objects is poor.
The isochrones on the $([8.0], [4.5]-[8.0])$ CMD also fall short of the
brightest stars in the Middle annulus of M32, although
the ridgeline of $[4.5]-[8.0]$ colors is consistent with ages $\sim 3$ Gyr.
The inability of the models to match the observations near the bright end may be
because some of the brightest AGB stars are LPVs. Comparisons with the
$([3.6]-[4.5],[4.5]-[5.8])$ two color diagram (TCD) -- discussed below -- indicates that
the predicted spectral-energy distribution (SEDs) disagree with the observations
by 0.1 -- 0.2 magnitudes.
The isochrones also indicate that old, metal-rich AGB stars
should be detected with these data. The 10 Gyr isochrone more-or-less skirts
the red envelope of objects in M32 when $[5.8] < 17$, with LPVs possibly accounting for
the modest number of objects with [3.6]--[5.8] colors that are redder than the 10 Gyr
sequence. A similar situation holds for the $([8.0], [4.5]-[8.0])$ CMD. There is thus
photometric evidence in the MIR for an old metal-rich population, and in Section 6
it is demonstrated that these objects contribute significantly to the number counts.
The $([5.8], [3.6]-[5.8])$ CMD of the M31 outer disk field is compared
with isochrones in the left hand panel of Figure 8. The
main locus of the $([5.8], [3.6]-[5.8])$ CMD of the M31
outer disk when $[5.8] < 16$ is consistent with a mean age that is
similar to M32. The isochrones suggest that the
objects with [3.6]--[5.8] $> 1$ near [5.8] $\sim 14$
may be highly evolved AGB stars with ages between 3 and 10 Gyr. Still, some
of these may be LPVs, in which case their location on the CMD may not reflect
their true age and metallicity. The $([5.8], [3.6]-[5.8])$ CMD of the M31 outer disk
field is consistent with stars being present that formed over a broad range of ages.
TCDs are a means of comparing model SEDs with those that are observed.
The $([3.6]-[4.5],[4.5]-[5.8])$ TCDs of sources in M32 with [4.5] $< 17$ are
compared with post-RGB models from Marigo et al. (2008) in Figure 9. There is a
red plume in each radial interval that has a dispersion of a few tenths of
a magnitude. Even though the stars on this plume are highly evolved and
many expected to be LPVs, the comparatively modest scatter is likely a consequence of
the simultaneous nature of the IRAC observations, coupled with the
smaller amplitude of LPV variations towards longer wavelengths. Still, the red
plumes predicted by the models systematically fall 0.1 -- 0.2 magnitudes below
the observations.
The properties of isochrones on the upper reaches of MIR CMDs depends on
the composition of the circumstellar dust, as this material can be a significant
source of emission at these wavelengths. The comparisons in Figures 8 and 9 assume
that circumstellar dust around M stars has a 60\% ALOx $+ 40$\% silicate mix.
Do the ages obtained above change if different dust compositions are adopted? The
role that dust composition plays on MIR isochrones is investigated
in Figure 10, where the $([5.8],[3.6]-[5.8])$ and $([8.0], [4.5]-[8.0])$
CMDs of the M32 Middle annulus are compared with Z=0.016 isochrones from Marigo
et al. (2008) that assume 100\% silicate (solid lines) and
100\% ALOx (dashed lines) compositions.
At a fixed age, the comparisons in Figure 10 indicate that model-to-model
differences in peak brightness amount to only a few tenths of a magnitude.
The mean [3.6]--[5.8] color of the 3 Gyr isochrone is also not
sensitive to the chemical properties of the dust, although this
is not the case for the 10 Gyr [3.6]--[5.8] colors.
Dust composition has a greater impact on isochrones in the $([8.0],
[4.5]-[8.0])$ CMD, due to the greater contribution from thermal emission near [8.0]
and the differences in absorption efficiencies that grow with
wavelength (e.g. Figure 1 of Groenewegen 2006). Indeed, the 3
and 10 Gyr isochrones that assume a 100\% ALOx composition overlap on
the $([8.0], [4.5]-[8.0])$ CMD. Neither 100\% ALOx sequence gives a satisfactory match
to the M32 CMD, with both falling redward of the M32 ridgeline. This suggests
that some silicates must be present to reproduce the [4.5]--[8.0] colors of evolved AGB
stars in M32.
\subsection{The $([3.6], K-[3.6])$ CMD}
The NIRI $K'$ and $L'$ images are of special interest as they have an
angular resolution that allows the brightest stars to be resolved to
within a few arcsec of the galaxy nucleus. The $L'$ magnitudes can
be transformed into the IRAC [3.6] photometric system, allowing the behaviour of
objects on the $([3.6], K-[3.6])$ CMD to be examined over a wider range
of radii than is possible with the IRAC data alone. The
$([3.6],K-[3.6])$ CMDs of stars in M32 and the M31 outer disk are shown in Figure 11.
The M32 Center CMD is constructed exclusively from the NIRI observations. Objects
within 2 arcsec of the nucleus of M32 have been excluded, as the
photometry in that region is compromised by crowding and/or surface brightness
fluctuations. The CMDs in the other panels use $K$ magnitudes from
the WIRCam observations and [3.6] magnitudes from the IRAC images.
The peak [3.6] magnitudes in the M32 Center, Inner, and Middle annuli agree to
within $\pm 0.2$ magnitude. Thus, the AGB peak brightness does not change
greatly with radius, and there is not a central concentration of luminous AGB stars.
With the exception of the outer annulus CMD, the peak [3.6] brightness of objects
in the M31 outer disk field is $\sim 0.5$ magnitude brighter than in the M32 CMDs.
If the brightest stars in the M31 outer disk were uniformly distributed then
similar objects might be expected in the inner and middle annuli, although
a population of bright field stars would not be expected in the M32 Center field
due to the modest area covered on the sky when NIRI is used in f/32 mode.
\section{RESULTS: LFs AND THE SFH DURING INTERMEDIATE EPOCHS}
The LFs of stars in the inner, middle, and outer annuli of M32 are compared
in Figure 12. Contamination from foreground stars, background galaxies, and stars in
the disk of M31 is a concern, and so the M31 outer disk LFs were
subtracted from the M32 LFs after adjusting for differences
in sky coverage. A potential source of uncertainty in this
correction is that the stellar content of the M31 disk field
may differ from that of the disk near M32 (e.g. Figure 11).
The entries in Figure 12 are specific frequencies that
give the number of objects per 0.2 magnitude interval in a system with M$_K = -16$.
The scaling to M$_K = 16$ was done using integrated annular brightnesses computed
from the M32 surface brightness measurements in the
2MASS Large Galaxy Atlas (Jarrett et al. 2003). It can be seen from
Figure 12 that the specific frequency in $K$ does not vary significantly
from annulus to annulus, in agreement with the results found by Davidge \&
Jensen (2007). However, there is a tendency for the specific frequencies
of stars in the [5.8] LF of the outer annulus to fall below those
in the inner and middle annuli, although the differences for individual bins are not
statistically significant. It should be recalled that the $K$ and [5.8] passbands
sample different sources of emission, with the [5.8] LFs being more sensitive to
AGB stars that have circumstellar envelopes than the $K$ LFs.
The interpretation of the LF of AGB stars is complicated by the rapid pace
of evolution on the upper AGB, so that small number statistics may make
signatures associated with changes in the SFR difficult to detect.
Photometric variablity will also blur the LFs, smoothing features that
might otherwise provide insight into the SFH. Finally, models of AGB
evolution are prone to uncertainties in the stellar physics (e.g. Marigo et al. 2008).
The expectation of uncertainties in the models notwithstanding, it is encouraging that
there is consistency between the isochrones and observations on the
CMDs and TCDs (Section 4).
Rather than attempt to deduce an independent SFH from the LFs, the present
analysis is restricted to comparisons with models that follow pre-defined
SFHs. The simplest models are those that assume a simple stellar population (SSP),
and in the top panel of Figure 13 the mean $K$ LFs of AGB stars in the inner and middle
annuli are compared with 2 and 10 Gyr SSP sequences constructed from the Marigo et al.
(2008) models. The models assume Z=0.016, a 60\% silicate $+$ 40\% AlOx
circumstellar dust mix, and that none of the stars are photometric variables.
The models in this panel have been scaled to match
the number counts in the two faintest bins.
The 10 Gyr model covers only the faintest part of the plotted LF.
In contrast, the 2 Gyr model in the top panel of Figure 13 provides a
reasonable match over many magnitudes to the $K-$band LF. Still,
the 2 Gyr model predicts a peak $K-$band magnitude that is $\sim 1$ magnitude
fainter than observed. Comparisons involving the [5.8] LF yield
similar results.
The majority of bright AGB stars in M32 are LPVs, with light
curve amplitudes that approach $\pm 1$ magnitude in $K$ (Davidge \& Rigaut 2004), and
the light distribution in the models changes significantly if LPV-like photometric
variability is added to the models. This is demonstrated in the middle panel
of Figure 13, which shows the 2 Gyr model $K$ LF convolved with the $\Delta K$
distribution that Davidge \& Rigaut (2004) obtained from
the light curves of Galactic bulge LPVs tabulated by Glass et al. (1995).
After convolution with this amplitude kernel, the 2 Gyr model $K$ LF matches the
peak of the observed LF as well as the approximate shape of the LF near the bright end.
Photometric variability can also be folded into models that
cover the IRAC passbands. The amplitude of LPVs near $5\mu$m is $\sim 0.8
\times$ that in $K$ (e.g. Table 4 of Le Bertre 1993), and so
a modified $K$ kernel, that was compressed along the magnitude axis by a
factor $1/0.8 = 1.25 \times$, is applied to the [5.8] models.
The 2 Gyr model constructed in this manner is shown
in the bottom panel of Figure 13, and there is reasonable agreement with the [5.8] LF.
While the comparisons with the 2 Gyr SSP model in Figure 13 are consistent
with a large intermediate age population, in reality M32
is a composite stellar system containing stars that span a range of ages
and metallicities. Monachesi et al. (2012) have investigated the SFH of M32
using stars near the MSTO, and model LFs that are based on the
SFH shown in their Figure 13a are compared with the observations in the middle
and lower panels of Figure 13. The models follow a Chabrier (2001) mass function,
and were convolved with the $K$ and [5.8] amplitude kernels described above.
Tacit assumptions are that: (1) all AGB stars regardless of mass have the same
variability characteristics, and (2) all of the AGB stars in M32 are LPVs.
The models further assume that the brightest AGB stars have a solar
metallicity, which is not unreasonable given the metallicity distribution
function and age-metallicity relation found by Monachesi et al. (2012).
The composite model matches the $K$ number counts when $K > 15.8$, and
is a better match to the observations in this magnitude range than the 2 Gyr SSP LF.
The improved agreement with the observations is because old AGB stars in the
composite model bolster the number counts at the faint end, making the composite
model steeper than the SSP model. The agreement between
the composite model and observations degrades when $K < 15.7$, although the
differences are at less than the $2\sigma$ level.
The agreement with the [5.8] LF is similarly good when
[5.8] $> 14.4$, with the overall shape of the observed LF reproduced
at these magnitudes. The agreement is also better than
was achieved with the 2 Gyr SSP model. However, as with the $K$ LF, the
composite model does not match the bright end of the LF as well as the 2 Gyr SSP model.
Despite the promising agreement between the composite model and the
observations, the model falls short near the bright ends of the $K$ and [5.8] LFs.
There are uncertainties in the structure models of
the most evolved AGB stars, and the cumulative impact
of any uncertainties are expected to be greatest near the AGB-tip.
Uncertainties in the LPV amplitude distribution function may also
contribute to difficulties matching the numbers of very bright objects.
A single amplitude kernel that is based on one population of objects (LPVs in
the Galactic bulge) was adopted for this study. The agreement between the
models and the bright end of the LFs in Figure 13 would improve if the amplitude
distribution function were to be extended to larger amplitudes and/or was skewed
to contain a larger fraction of objects at the high amplitude end.
The differences between the observations and the
composite models might also indicate a mix of ages that differs from that predicted by
the Monachesi et al. (2012) SFH. A larger fractional contribution from moderately
young stars should boost the number of stars near the observed AGB-tip. To investigate
this possibility, a model LF that follows the Monachesi et al.
(2012) SFH from the present day to 7 Gyr, but that has no stars older than 7 Gyr,
was constructed and the result is shown in Figure 13.
The suppression of star formation more than 7 Gyr
in the past has a significant impact on the agreement with the observations.
The differences are more pronounced in [5.8] than in $K$ as the AGB sequences of old
populations extend to brighter intrinsic magnitudes in the MIR.
The comparison between the [5.8] LF and the truncated SFH model thus indicates
that M32 must contain a substantial metal-rich population that formed
during relatively early epochs. This demonstrates that MIR observations provide a
direct means of probing the SFH of a system over a wide range of epochs.
How sensitive is the [5.8] LF to changes in the
size of star-forming events during intermediate epochs?
The [5.8] LF of M32 is compared with models that contain an old population
and an intermediate age component in Figure 14.
Two flavors of intermediate age component are investigated: a discreet
burst, which is represented by a 2 Gyr SSP, and
a star-forming event with a constant SFR that has a 1 Gyr duration, and
occured between 2 and 3 Gyr in the past. A solar metallicity is assumed.
Models in which the intermediate age component contributes
three different stellar mass fractions are shown in Figure 14.
It can be seen that (1) the composite models provide a
better fit to the LFs than SSPs, and (2) models that involve an extended
period of star-forming activity during intermediate epochs give a better match to the
LF in the [5.8] interval between 14.5 and 16.0. In fact, the slope of the LF in this
brightness interval, which corresponds to M$_{[5.8]}$ between --8.5 and
--10.0, is sensitive to the age mix of the system. Models that
assume a dominant intermediate age component yield a better match to the bright
end of the [5.8] LF, although this is where the evolutionary models are
least reliable and there is susceptibility to uncertainties in the nature of
the amplitude distributions of LPVs (see above).
To summarize this section, model LFs that assume that the
brightest AGB stars in M32 are LPVs better match the
LFs in $K$ and [5.8] than those that assume no variability. Models that contain an
intermediate age component -- including those
based on the SFH constructed by Monachesi et al. (2012) -- match the shape
of the LFs in M32 in $K$ and [5.8] over a range of magnitudes.
It is also shown that the presence of an old population
is required to explain the overall shape of the [5.8] LF of M32. These results highlight
the potential utility of MIR observations of AGB stars to probe the stellar contents
of moderately distant galaxies.
\section{RESULTS: SURFACE BRIGHTNESS PROFILES AND STAR COUNTS}
\subsection{The Main Body of M32}
The light profile of a galaxy forms part of its fossil record
and can be used to gain insights into its evolution. In this section, the structural
characteristics of M32 are examined over a range of wavelengths to search for
possible radial differences in stellar content that might
provide clues into the past history of the galaxy.
To this end, the structure of M32 in $J, K, [3.6]$, and [4.5] was investigated using
the isophote-fitting task {\it ellipse} (Jedrzejewski 1987), as implemented in STSDAS.
The analysis was performed on star-subtracted images. In addition to
removing resolved stars that belong to M32, star subtraction also removes bright
stars that belong to M31, but does not account for unresolved light from
M31. To account for this, residual background sky levels were measured near the
edges of the observed field and the result was subtracted prior to the
isophotal analysis. This procedure assumes that the light from the
M31 disk is uniformly distributed across the field, with the areas where the
background light levels are measured being representative of this part of M31.
The light profiles are compared in the top panel of Figure 15.
A cluster of bad pixels causes a gap in the radial coverage
of the WIRCam observations, and the affected interval
is indicated. The surface brightness profile in the central 150 arcsec of M32
can be fit with an R$^{1/4}$ law (e.g. Choi et al. 2002). The effective radius, R$_e$,
and surface brightness, $\mu_e$, obtained by fitting an R$^{1/4}$ law to the WIRCam and
SPITZER images are listed in Table 1. The R$_e$ measurements made by Kent (1987) and
Choi et al. (2002) at visible/red wavelengths are also listed for comparison.
The errors in R$_e$ are $1\sigma$ random uncertainties that reflect
the scatter about the fitted relation. Systematic effects, which are
not included in the error estimates, have the potential to introduce uncertainties
that are much larger than the random errors. For example,
Choi et al. (2002) and Kent (1987) measure R$_e$
at similar wavelengths, but their results differ by 3 arcsec. This
suggests that a more realistic total uncertainty in R$_e$ may be
$1 - 2$ arcsec, which is $2 - 3 \times$ larger than the random uncertainties.
Still, it is encouraging that the R$_e$ values measured in $K$, [3.6], and [4.5] agree
within their random uncertainties, while the [3.6] and [4.5] R$_e$ measurements agree
with the R$_e$ measured in the $R$ filter by Kent (1987) at the $2\sigma$ level.
Given the uncertainties in R$_e$, the entries in Table 1 suggest that R$_e$ is constant
with wavelength, at least to within a few arcsec. The spatial distribution of the
components that dominate the light at visible wavelengths, which includes a substantial
contribution from main sequence stars, is thus not greatly different from the
spatial distribution of the red stars that contribute significantly to the
unresolved NIR and MIR light.
Additional insights into the evolution of M32 can be gleaned from isophote
shapes. The ellipticities measured from the WIRCam and Spitzer images are compared
with ellipticity measurements from Table III of Kent (1987) in
the middle panel of Figure 15. The error bars show $1\sigma$ uncertainties that
reflect the scatter about the fitted ellipses. These are
internal errors, and the point-to-point jitter in the eccentricity measurements in a
given filter suggests that at large radii the actual uncertainties may be a few
times larger than those computed by {\it ellipse}.
There are no systematic differences between the NIR, MIR and
R-band ellipticities at radii $< 16$ arcsec. However, between
16 and 45 arcsec the R-band ellipticities consistently fall
below those measured at longer wavelengths, suggesting that the NIR and MIR isophotes
at these radii are flatter than at visible wavelengths. The uncertainties in the
ellipticities computed by {\it ellipse} would have to be in error by almost
an order of magnitude to overlap with the $R$ eccentricity
measurements. In fact, the ellipticities measured
in $J$ and $K$ from the WIRCam data agree with those obtained by Peletier
(1993) in the same filters, while the ellipticities measured by Kent (1987) agree
with those measured by Choi et al. (2002). Hence, the differences found here
are consistent with other independent measurements.
The ellipticity measurements suggest that at least some of the
unresolved NIR and MIR light originates from objects that have a flatter distribution
on the sky than the objects that dominate the light at shorter wavelengths.
Such a distribution is suggestive of a disk, and evidence of a disk in the
outermost regions of M32 has been found by Choi et al. (2002) and Graham (2002).
The coefficient of the fourth order cosine term in the fourier
expansion of the isophotes is a sensitive indicator of isophote
shape (e.g. Carter 1978). If this coefficient $< 0$ then the isophote
is `boxy, whereas if it is $> 0$ then the isophote is
`disky'. The fourth order cosine coefficients -- B4 --
measured from the NIR and MIR data are plotted in the lower panel of Figure 15. The
B4 values typically have only modest departures from zero, and to suppress
random errors the points that are plotted in the lower panel of Figure 15
are the averages of coefficients from three adjacent isophotes.
The entries in the lower panel of Figure 15 indicate that when
$r < 16$ arcsec the B4 coefficient tends to be predominantly
negative, indicating a boxy morphology. However, at larger radii there is a
higher incidence of positive B4 values. The mean B4 values at $r < 16$ arcsec
and at $r > 48$ arcsec are compared in Table 2, and it is evident that B4 changes with
radius in $J$, [3.6], and [4.5]. In fact, (1) the mean B4 entries in these filters
in a given radial interval all agree, and (2) at $r > 48$
arcsec the B4 coefficient in these filters exceeds zero at the $2\sigma$
or greater level. Still, the B4 measurements made from the
$Ks$ images differ from those in the other three filters. Those measurements
notwithstanding, the B4 values from the other three filters provide
tantalizing evidence that the structural properties of M32
when $r < 16$ arcsec is different from that at $r > 48$ arcsec.
\subsection{The Central Regions of M32}
The center of M32 samples the deepest part of the gravitational potential
well, and so may harbor signatures of past events that shaped its evolution.
The angular resolution of the NIRI $K', L'$ and $M'$ observations
allow the light distribution near the center of M32
to be investigated on spatial scales $\leq 1$ parsec.
The central regions of M32 have been examined previously at sub-arcsec
resolutions, although not at wavelengths $> 2.5\mu$m.
Lauer et al. (1998) and Corbin et al. (2001) find no evidence of radial
color variations in the visible and NIR, while Peletier (1993) and Davidge (2000)
find that the narrow-band CO index in M32 may strengthen with decreasing radius in the
central 3 arcsec.
The $K'$ and $L'$ NIRI observations were smoothed to the
angular resolution of the $M'$ data, and the resulting light, color,
and ellipticity profiles are shown in Figure 16.
The central structural properties of M32 change substantially over
small angular scales, and subtle variations in the character of the outer
wings of the PSF may significantly affect colors in these regions. Therefore,
no effort is made to examine the isophotal properties of the central 0.2 arcsec of M32.
The light profiles in the top panel of Figure 16 have similar slopes,
and the $K'-L'$ and $K'-M'$ colors are more-or-less constant. The
ellipticities measured from the $K'$ and $L'$ images are similar out to 1.3
arcsec. However, between 0.5 and 0.7 arcsec the $M'$ ellipticities fall below those in
$K'$ and $L'$. While the centroiding of the individual $M'$ images was subject to
$\pm 0.1$ arcsec jitter (Section 2), this can not explain the dispersion in
ellipticities at these radii. Uncertainties due to centroiding errors were assessed by
conducting simulations involving a mock $M'$ image that was constructed using
the isohophotal analysis of the actual $M'$ image. As the final $M'$ science
image was the result of median-combining 35 individual images,
this mock image was replicated 35 times, with each replicant
randomly offset by $\pm 0.1$ arcsec in radius.
The 35 images with simulated image jitter were then median-combined, and
$ellipse$ was run on the result. A number of such realizations were
run, and a comparison of the ellipticities measured from all of these
indicates that a $\pm 0.1$ arcsec centroiding error introduces
$\pm 0.03$ error in the ellipticities at a radius of 0.2
arcsec, and $\pm 0.02$ error at 0.4 arcsec.
Uncertainties in image centroiding thus can not explain the difference between the
$M'$ and shorter wavelength ellipticity measurements near 0.5 arcsec
radius. It would be of interest to obtain additional images near $5\mu$m to
investigate further the MIR ellipticities between 0.5 and 0.7 arcsec.
The isohotal properties of the $K'$ and $L'$ images,
with the $K'$ image smoothed to the angular resolution of the $L'$ image,
are compared in Figure 17. The angular resolution of the data used to construct
Figure 17 is thus 0.2 arcsec, as opposed to the 0.4 arcsec resolution of the data
used for Figure 16. The $K'$ and $L'$ surface brightness profiles have
similar slopes, and this is reflected in the flat $K'-L'$ profile, shown in
the middle panel. The $K'$ and $L'$ ellipticity measurements are also
similar at radii $> 0.2$ arcsec. The tendency for ellipticity to decrease
at radii $< 0.6$ arcsec is likely due -- at least in part -- to seeing effects
(e.g. Erwin \& Sparke 2003). The comparisons in Figures 16 and 17 indicate
that the `stubborn normalcy' in stellar content near the center of M32
noted by Lauer et al. (1998) is evident at angular resolutions 0.2 -- 0.4 arcsec
in the NIR and MIR.
\section{DISCUSSION \& SUMMARY}
The stellar content of the Local Group cE galaxy M32 has been investigated
using ground and space-based images that span the $1 - 8\mu$m wavelength
region. A goal of this study is to use M32 as a foil to
consider the utility of MIR observations for investigating the stellar contents of
more distant galaxies. M32 is a favorable target for checking models and developing
techniques that might be applied to trace the evolution of more distant galaxies
as it is relatively nearby, allowing stars near
the MSTO to be resolved, and has been extensively studied.
It is demonstrated that accounting for the LPV nature of bright
AGB stars in M32 is an essential element of successfully
modelling the NIR and MIR photometric properties of these objects. In the current study
this has been done by convolving model LFs with an amplitude distribution
that is based on Galactic LPVs. While variability smooths out
features in the LF, it does not completely obliterate information that can be
used to probe the SFH. Indeed, the overall shape of the smoothed [5.8] LF can be
used to investigate the contributions made by stars with intermediate and old ages.
Comparisons between the NIR and MIR LFs and models constructed from the Bressan
et al. (2012) sequences indicate that M32 is not a SSP, and must contain an
old, metal-rich stellar component. Model $K$ and [5.8] LFs that
assume the SFH found by Monachesi et al. (2012) match the observations in both filters.
The potential limitations imposed by stellar variability notwithstanding, the results
in this paper indicate that observations of more distant systems in the MIR, such as
those that will be conducted with facilities such as the JWST, will prove useful for
constraining the fraction of old and intermediate age stars in galaxies where other
moderately bright age indicators, such as core helium burning stars, can not be resolved.
It would be of interest to investigate the
NIR and MIR LFs of Local Group galaxies that contain AGB stars
with ages and/or metallicities that differ from those in M32.
Models predict that the rate of mass loss on the
upper AGB, and hence the size of circumstellar envelopes, depends on
progenitor mass and metallicity (e.g. Bowen \& Willson 1991), although there
is still not extensive observational evidence to support a metallicity-dependence
(e.g. Lagadec 2010). The models predict that more massive, metal-rich
AGB stars will have thicker, dustier circumstellar envelopes than
less massive, more metal-poor objects, and the former are expected to be brighter
in the MIR than the latter. Hence, a different fraction of the MIR
light might be expected to originate from circumstellar sources
in galaxies with mean metallicities that differ from that of M32.
We close the paper by briefly considering the origins of M32, as the
analysis of the NIR and MIR data provides information that is relevant to this issue.
The NIR and MIR images suggest that M32 may be structurally complex.
The isophotal analysis discussed in Section 6 reveals a flattened
structure in the unresolved NIR and MIR light that can be
traced to within 16 arcsec of the galaxy center.
The coefficient of the fourth order cosine expansion of the MIR isophotes
is also indicative of a disky morphology at $r > 48$ arcsec.
These results are interesting given that Graham (2002) and Choi et al. (2002)
find evidence for a diffuse disk-like distribution in the outer regions of M32.
If a disk is present then clues to its origins will be
found in its stellar content. For example, a
diffuse remnant of a once-larger progenitor disk might be expected to show
a SFH that was truncated at the time of the events that disrupted
the disk, and there might be evidence for elevated galaxy-wide
star-forming activity. On the other hand, a diffuse disk that
formed during cosmologically recent epochs from the accretion
of high angular momentum gas that was able to cool and form
stars might show a star-forming history that is distinct from the main body
of the galaxy, and that continued to relatively recent epochs.
The changes in structural characteristics found here are not
accompanied by an obvious radial change in stellar content. There
are no large-scale gradients in absorption line
strengths in M32, although fortuitous gradients in age and metallicity that conspire
to negate absorption line gradients can not be
ruled out (Davidge 1991; Worthey 2004; Rose et al. 2005).
While in Section 4 it is shown that there is a radial gradient
in the mean $J-K$ color of the M32 AGB sequence, this trend
is consistent with progressively larger fractional contamination from stars in
the outer disk of M31 towards larger distances from
the center of M32, rather than a gradient in -- say --
mean metallicity within M32. More rigorous constraints
on stellar content variations are placed by the specific frequency of bright
AGB stars, which does not vary with radius out to very large radii when
normalized according to either visible or NIR surface brightnesses (Davidge \&
Jensen 2007). The peak $3.5\mu$m brightness of AGB stars in M32 also does not change
over offsets from 2 arcsec to 176 arcsec from the galaxy center (Section 4),
underscoring previous studies at shorter wavelengths that have found that the
brightest AGB stars are distributed throughout the galaxy. If there is a diffuse
disk around M32 then it has a stellar content that is not greatly different
from that of the underlying body of the galaxy.
Could M32 be a fossil remnant of a major merger?
Hammer et al. (2010) investigate the possibility of a major merger between M31
and another galaxy a few Gyr in the past. In addition to
producing multiple tidal features around M31, such an event can
also explain episodes of elevated star formation that are
evident in the age distribution of M31 globular clusters (e.g. Puzia et al. 2005),
and the diverse nature of the M31 globular cluster system (e.g. Beasley et al. 2005).
It has been demonstrated in this paper that the NIR and MIR photometric
properties of AGB stars in M32 are consistent with significant extended periods of
star-forming activity in M32 during intermediate epochs, and this
is in qualitative agreement with the interaction modelled by Hammer et al. (2010).
It would be of interest to determine if there are complex debris trails in
the vicinity of M31 of the type predicted by the Hammer et al. (2010) simulations
that are populated by AGB stars with NIR and MIR properties like those studied here.
\acknowledgements{Thanks are extended to the anonymous referee for a timely and
comprehensive report that greatly improved the presentation of the results.}
\parindent=0.0cm
\clearpage
\begin{table*}
\begin{center}
\begin{tabular}{lccl}
\tableline\tableline
Filter & R$_e^1$ & $\mu_e^2$ & Reference \\
& (arcsec) & (mag arcsec$^{-2}$) & \\
\tableline
B$+$I & 29 & 19.43 (B), 17.53 (I) & Choi et al. (2002) \\
R & 32 & 18.79 & Kent (1987) \\
J & $36.6 \pm 0.6$ & $16.92 \pm 0.01$ & This paper \\
Ks & $34.1 \pm 0.6$ & $15.86 \pm 0.01$ & This paper \\
$[3.6]$ & $33.6 \pm 0.7$ & $20.34 \pm 0.02$ & This paper \\
$[4.5]$ & $33.1 \pm 0.8$ & $20.41 \pm 0.02$ & This paper \\
\tableline
\end{tabular}
\tablenotetext{1}{Effective radius.}
\tablenotetext{2}{Effective surface brightness.}
\caption{Light Profile Measurements}
\end{center}
\end{table*}
\clearpage
\begin{table*}
\begin{center}
\begin{tabular}{lrr}
\tableline\tableline
Filter & $<B4>$ & $<B4>$ \\
& (r $< 16$ arcsec) & (r $> 48$ arcsec) \\
$J$ & --0.0013 & 0.0024 \\
& $\pm 0.0024$ & $\pm 0.0011$ \\
& & \\
$Ks$ & --0.0045 & -0.0010 \\
& $\pm 0.0026$ & $\pm 0.0013$ \\
& & \\
$[3.6]$ & --0.0007 & 0.0024 \\
& $\pm 0.0015$ & $\pm 0.0006$ \\
& & \\
$[4.5]$ & --0.0006 & 0.0033 \\
& $\pm 0.0022$ & $\pm 0.0006$ \\
& & \\
\tableline
\tableline
\end{tabular}
\caption{Mean B4 Coefficients}
\end{center}
\end{table*}
\clearpage
|
\section{Introduction}\label{intro}
\subsection{Background and Definitions}
\noindent Understanding which primes divide the Fermat numbers (i.e. numbers of the form $2^{2^n} + 1$) has been a question of interest for nearly four centuries. Progress toward answering related questions has been made with the tools of arithmetic dynamics, a field whose objective is to study the behavior of \emph{dynamical sequences} $\{a_n\}$ that are defined by $a_n = \phi(a_{n-1})$ for some choice of a zeroth term $a_0 \in \BZ$ and a polynomial $\phi \in \BZ[x]$. In the case of the Fermat numbers, the relevance of arithmetic dynamics is evident, because the sequence $\{a_n\}$ of Fermat numbers can be expressed in recursive form as $a_0 = 3$ and $a_n = \phi(a_{n-1})$, where $\phi \in \BZ[x]$ is given by $\phi(x) = (x-1)^2 + 1 = x^2 - 2x + 2$. Although it is difficult to give explicit descriptions of the primes that divide the terms of dynamical sequences, we may still be able to obtain qualitative results regarding the distribution of these prime factors in the set of all primes. In particular, given a choice of $a_0 \in \BZ$ and $\phi \in \BZ[x]$, we may attempt to determine the \emph{natural density} $\on{nd}(\phi, a_0)$ of the set $$P_\phi(a_0) = \{\text{prime } p : p \mid a_n \text{ for at least one } n \geq 0\}$$ as a subset of the set of all primes, where $\on{nd}(\phi, a_0)$ is defined by the limit
$$\on{nd}(\phi,a_0) = \limsup_{x \rightarrow \infty} \frac{|\{p \in P_\phi(a_0) : p \leq x\}|}{|\{\text{prime } p \leq x \}|}.$$
In~\cite{odo1} and~\cite{odo2}, Odoni showed that $\on{nd}(\phi, a_0) = 0$ when the Galois groups $G_n(\phi)$ of iterates $\phi^n$ of $\phi$ satisfy the equality
$$\lim_{n \rightarrow \infty} \frac{|\{g \in G_n(\phi) : g \text{ fixes at least one root of } \phi^n\}|}{|G_n(\phi)|} = 0.$$
In the case of \emph{Sylvester's sequence}, which is the sequence $\{a_n\}$ obtained by taking $a_0 = 2$ and $\phi(x) = x^2 - x + 1$, Odoni proved that $\on{nd}(\phi,2) = 0$ by showing that for each $n$ we have $G_n(\phi) \simeq \Aut(T_n)$, where $T_n$ is the complete, rooted binary tree of height $n$. To obtain this result for Sylvester's sequence, Odoni expressed $T_n$ as the tree $T_n(\phi,0)$ of iterated preimages of $0$ under $\phi$ (our notation for a preimage tree of height $n$ is $T_n($``function'',``root''$)$). More precisely, taking $\phi^0(x) = x$ and $\phi^{-n}(x) = \phi^{-1}(\phi^{-(n-1)}(x))$ for each $n \geq 1$, we can construct a binary tree $T(\phi,0)$ of infinite height rooted at $0$ such that the nodes on level $i$ are the elements of $\phi^{-i}(0)$ and such that for each $y \in \phi^{-i}(0)$, the parent of $y$ in level $i-1$ is $\phi(y) \in \phi^{-(i-1)}(0)$. We may then take $T_n(\phi, 0)$ to be the rooted binary subtree of $T(\phi, 0)$ obtained by deleting all levels of $T(\phi, 0)$ beyond level $n$. (Notice that $T(\phi,0)$ is a complete tree if and only if the discriminant of each iterate of $\phi$ does not vanish; when $\phi(x) = x^2 - x + 1$, the tree $T(\phi,0)$ is in fact complete.) As a consequence of Odoni's success with Sylvester's sequence, it is natural to study maps of the form $G_n(\phi) \to \Aut(T_n(\phi,0))$, which are called \emph{arboreal Galois representations}, for various choices of $\phi \in \BZ[x]$.
In~\cite{jone} and~\cite{JM}, Jones and Manes consider a more general situation. Let $K$ be a global field (i.e., an algebraic number field or function field over a finite field), let $\phi \in K(x)$ be a rational function of degree $d \geq 2$ (the case of $d = 1$ is well-studied), and let $\alpha \in \BP^1(K)$ (the projective line of the field $K$). Furthermore, consider only those pairs $(\phi, \alpha)$ such that the equation $\phi^n(x) = \alpha$ has $d^n$ distinct solutions for each $n$, where $\phi^n$ denotes the $n^\mathrm{th}$ iterate of $\phi$ (this condition is easily verified by showing that the orbit of each critical point of $\phi$ avoids $\alpha$). Then the tree $T(\phi, \alpha)$ of iterated preimages of $\alpha$ under $\phi$ is complete and $d$-ary, and its nodes are elements of the separable closure $K^{\ksep}$ of the field $K$. We observe that the elements of the absolute Galois group $\Gal(K^{\ksep}/K)$ induce tree automorphisms of $T(\phi, \alpha)$, so we obtain a homomorphism
$$\rho : \Gal(K^{\ksep}/K) \to \Aut(T(\phi, \alpha)),$$
which is again an arboreal Galois representation. Define $G(\phi, \alpha) \defeq \im \rho$. Notice that $G_n(\phi, \alpha) = \Gal(K(\phi^{-n}(\alpha))/K)$ is the quotient of $G(\phi, \alpha)$ by the equivalence relation $a \sim b$ if $a$ and $b$ act in the same way on $T_n(\phi, \alpha)$. Thus, we may view $G(\phi, \alpha)$ as the limit of the inverse system $\{G_n(\phi, \alpha)\}$. It suffices to take $\alpha = 0$: if $g(x) = x + \alpha$, then we have the equality of Galois groups
$$G_n(\phi, \alpha) = G_n(g^{-1} \circ \phi \circ g, 0),$$
so we may replace $\phi$ with $g^{-1} \circ \phi \circ g$ and $\alpha$ with $0$.
We then write $T(\phi, 0)$ as $T(\phi)$ and $G(\phi, 0)$ as $G(\phi)$ for short. In light of Odoni's result for Sylvester's sequence, it is natural to try to characterize the structure of $G(\phi)$ as a subgroup of $\Aut(T(\phi))$.
\subsection{Main Results}\label{mainresults}
\noindent In this paper, we extend work of Jones and Manes (see~\cite{jone} and~\cite{JM}) on describing the group $G(\phi)$ for various choices of rational functions $\phi \in K(x)$. Particularly, we consider the following two cases for which the structure of $G(\phi)$ is not fully understood.
\medskip
\noindent {\bf Problem (1):} The sequence $\{a_n\}$ defined by $a_0 = 0$ and $a_n = \phi(a_{n-1})$ is periodic. (Actually, we want to study this problem for arbitrary $a_0 \in K$. Nonetheless, if $a_0 \neq 0$, we may conjugate $\phi$ by $g(x) = x + a_0$ to set $a_0 = 0$.)\\
\noindent {\bf Problem (2):} The function $\phi$ commutes with a nontrivial M\"{o}bius transformation that fixes the root $\alpha = 0$ of the tree $T(\phi)$.
\medskip
As exemplified by Odoni's work, understanding the structure of $G(\phi)$ may be useful for obtaining zero-density results for $P_\phi(a_0)$. In~\cite{jone}, Jones describes a stochastic method that can in some cases translate results about the size of $G(\phi)$ into zero-density results for $P_\phi(a_0)$. One key step in Jones' stochastic method is to find ``natural'' subgroups $N(\phi) \subset \Aut(T(\phi))$ that satisfy $G(\phi) \subset N(\phi)$ and $[N(\phi) : G(\phi)] < \infty$. By ``natural,'' we mean subgroups whose action on $T(\phi)$ can be stated explicitly in the language of tree automorphisms, without any reference to elements of $G(\phi)$. In the case of Sylvester's sequence, the desired subgroup $N(\phi)$ happens to be $\Aut(T(\phi))$ itself. But it is easy to show that in the two problems enumerated above, $[\Aut(T(\phi)) : G(\phi)] = \infty$. Consequently, one of our objectives in describing $G(\phi)$ is to find and test candidates for the group $N(\phi)$ when $\phi$ belongs to either of the two problems enumerated above.
The rest of the paper is organized as follows. Section~\ref{period} discusses our results regarding Problem (1) above. The primary results proved in Section~\ref{period} are as follows. We first show that $G(\phi)$ is a small subgroup of the stabilizer $\Stab(\phi) \subset \Aut(T(\phi))$ of the branch whose nodes are precisely the terms of the sequence $\{a_n\}$, thereby resolving a question of Jones (Question 3.4 on p. 18 of~\cite{jone}).
\begin{proposition*}[Proposition~\ref{prop3}]
For every $\phi \in K(x)$ such that $\phi^m(0) = 0$ for some positive integer $m$, $[\Stab(\phi) : G(\phi)]$ is necessarily infinite.
\end{proposition*}
Let $S(\phi) \subset \Stab(\phi)$ be the group of automorphisms of $T(\phi)$ that act in the same way on the subtrees rooted at $a_r$ and $a_s$ whenever $r \equiv s$ modulo the period of the sequence. For the family $\phi \in \BZ[x]$ defined by $\phi(x) = x^2 + kx$, we obtain the following characterization of $G(\phi)$.
\begin{theorem*}[Theorems~\ref{thm8} and~\ref{thm9}]
If $k \in \BZ \setminus \{-2, 0, 2, 4\}$ and $\phi(x) = x^2 + kx$, then $[S(\phi) : G(\phi)] < \infty$. On the other hand, if $k \in \{-2,0,2,4\}$, then $[S(\phi) : G(\phi)] = \infty$, but for each $k$ we can compute a valid \mbox{choice for $N(\phi)$}.
\end{theorem*}
We obtain a similar but more restricted result in the case when $\phi(x) = x^2 - (k+1)x + k$.
Sections~\ref{trees} and~\ref{squares} treat our results regarding Problem (2) above. In Section~\ref{trees}, we first prove a general lemma about centralizers of subgroups of $\Aut(T)$ for infinite, rooted $d$-ary trees $T$ (see Lemma~\ref{cor1} for the precise statement). Using this lemma, we resolve Conjecture 3.5 of~\cite{jone}, which is proven as Theorem~\ref{thm5} in the present paper. Define $C(\phi) \subset \Aut(T(\phi))$ to be the centralizer of the group $A(\phi)$ of M\"{o}bius transformations that commute with $\phi$ and fix $0$ (the group $A(\phi)$ embeds into $\Aut(T(\phi))$). Then following theorem tells us that $C(\phi)$ is in general a small subgroup of $\Aut(T(\phi))$.
\begin{theorem*}[Theorem~\ref{thm5}]
When $A(\phi)$ is nontrivial we have $[\Aut(T(\phi)) : C(\phi)] = \infty$.
\end{theorem*}
In Section~\ref{squares}, we present progress made toward a proof of Conjecture 1.3 of~\cite{JM}, a restricted version of which states that $[C(\phi) : G(\phi)] < \infty$ when $K = \BQ$ and $\phi$ is a quadratic rational function of the form $\phi(x) = \frac{k_0(x^2 + 1)}{x}$ where $k_0 \in \BZ \setminus \{0\}$. In~\cite{JM} it is shown that $[C(\phi) : G(\phi)]$ is finite when a certain sequence $\{\delta_n(k_0) : n = 2, 3, \ldots\}$ contains no squares. Our first result extends a lower bound from~\cite{JM} on $k_0$-values for which $[C(\phi) : G(\phi)] < \infty$.
\begin{theorem*}[Theorem~\ref{thm6}]
When $1 \leq |k_0| \leq 10^6$, the sequence $\{\delta_n(k_0)\}$ contains no squares. Thus, when $\phi(x) = \frac{k_0(x^2 + 1)}{x}$ and $1 \leq |k_0| \leq 10^6$, we have $[C(\phi) : G(\phi)] < \infty$.
\end{theorem*}
The lower bound of $10^6$ on $k_0$-values was proven by reducing $\delta_n(k_0)$ modulo various primes. Our final results in Section~\ref{squares} explore an algebraic method of showing that $\delta_n(k_0)$ is not a square. In particular, we prove the following theorem, which provides significant evidence toward the conjecture that $\delta_n(k_0)$ is not a square for all integers $n \geq 2$ and $k_0 \neq 0$.
\begin{theorem*}[Theorem~\ref{thm4}]
Fix $n \geq 2$. Then $\delta_n(k_0)$ is a square for at most finitely many nonzero integers $k_0$.
\end{theorem*}
\section{When the Sequence $\{a_n\}$ is Periodic}\label{period}
In this section, we discuss our results regarding Problem (1) of Section~\ref{mainresults}.
\subsection{The Size of the Stabilizer of $\{a_n\}$}
\noindent Suppose that $\phi^m(0) = 0$ for some positive integer $m$. Consider the sequence $\{a_n\}$ defined by $a_0 = 0 \text{ and } a_n = \phi(a_{n-1})$. Clearly $\{a_n\}$ is periodic, since $a_m = 0 = a_0$, and its elements belong to $K$. Also, notice that $a_i \in \phi^{i-m}(0)$ for each $i \in \{0, \dots, m-1\}$. Thus, in the tree $T(\phi)$, the sequence $\{a_n\}$ appears as an infinite branch, beginning with the root $a_0$ and proceeding $a_{m-1}, a_{m-2}, \ldots, a_1$ at levels $1, 2, \ldots, m,$ and then repeating.
We now define $\Stab(\phi)$ to be the stabilizer of the action of $\Aut(T(\phi))$ on this branch. It is clear that $G(\phi) \subset \Stab(\phi)$ because the elements of $G(\phi)$ must fix elements of $K$. According to~\cite{jone}, it is hoped that $G(\phi)$ is often a large subgroup of $\Stab(\phi)$; i.e., it is hoped that the index $[\Stab(\phi) : G(\phi)]$ is finite for many possibilities of $\phi$, so that we could take $N(\phi) = \Stab(\phi)$ for these functions $\phi$. One way of checking this is to study the \emph{Hausdorff dimension} of $G(\phi)$ in $\Stab(\phi)$, which is defined as follows.
\begin{defn}[Equation (1) of~\cite{hd}]\label{def1}
Let $T$ be a complete rooted $d$-ary tree of infinite height, and for each $n \in \BN$ let $T_n$ denote the subtree of height $n$ whose root is that of $T$. The \emph{Hausdorff dimension} $\on{hd}(H,G)$ for a subgroup $H$ in a group $G \subset \Aut(T)$ is given by
$$\on{hd}(H, G) = \lim_{n \rightarrow \infty} \frac{\log |H_n|}{\log |G_n|},$$
where for each $n$ we denote by $H_n$ and $G_n$ the quotients of $H$ and $G$, respectively, by the equivalence relation $a \sim b$ if \mbox{$a$ and $b$ act in the same way on $T_n$.}
\end{defn}
To check if a subgroup $H$ of a group $G \subset \Aut(T)$ has infinite index in $G$ (i.e. $[G : H] = \infty$), it suffices to check that $\on{hd}(H, G) < 1$. We now show that, contrary to what might be expected, $\on{hd}(G(\phi),\Stab(\phi)) < 1$, so in fact $[\Stab(\phi) : G(\phi)]$ is necessarily infinite.
\begin{proposition}\label{prop3}
For every $\phi \in K(x)$ such that $\phi^m(0) = 0$ for some positive integer $m$, we have that $\on{hd}(G(\phi),\Stab(\phi)) = 1 - d^{-m}$, and thus $[\Stab(\phi) : G(\phi)]$ is infinite.
\end{proposition}
\begin{proof}
Suppose $n \geq m$. For each $n$, we write $G_n(\phi)$ for the Galois group of the $n^\mathrm{th}$ iterate of $\phi$ and $\Stab_n(\phi)$ for the quotient of $\Stab(\phi)$ by the equivalence relation $a \sim b$ if $a$ and $b$ act in the same way on $T_n(\phi)$. We first compute $|\Stab_n(\phi)|$. Notice that the elements of $\Stab_n(\phi)$ can be enumerated as follows: $\Stab_n(\phi)$ is the product over all $i \in \{1, \dots, n\}$ of the groups $A_i$ of all automorphisms of the forest of $d-1$ subtrees of $T_n(\phi)$ rooted at the children of $a_{i-1}$ other than $a_i$. It is easy to see that $$|A_i| = (d-1)! \cdot \left[(d!)^{\frac{d^{n-i} - 1}{d-1}}\right]^{d-1} = \frac{(d!)^{d^{n-i}}}{d}.$$
We then have that
$$|\Stab_n(\phi)| = \prod_{i = 1}^n |A_i| = \frac{(d!)^{\sum_{i = 1}^n d^{n-i}}}{d^n} = \frac{(d!)^{\frac{d^n - 1}{d-1}}}{d^n}.$$
We next compute an upper bound on $|G_n(\phi)|$. For each $n$, define $S_n(\phi) \subset \Stab_n(\phi)$ to be the group of automorphisms of $T_n(\phi)$ that act in the same way on the subtrees rooted at $a_r$ and $a_s$ whenever $r \equiv s \pmod m$. It is clear that $G_n(\phi) \subset S_n(\phi)$. To show that $\on{hd}(G(\phi),\Stab(\phi)) < 1$, it suffices to show that $\on{hd}(S(\phi),\Stab(\phi)) < 1$, where $S(\phi)$ is the inverse limit of the groups $S_n(\phi)$. We now compute $|S_n(\phi)|$ for each $n$. Noticing that $S_n(\phi)$ is isomorphic to the product $\prod_{i = 1}^m A_i$, we have that
$$|S_n(\phi)| = \prod_{i = 1}^m |A_i| = \frac{(d!)^{\sum_{i = 1}^m d^{n-i}}}{d^m} = \frac{(d!)^{\frac{d^n - d^{n-m}}{d-1}}}{d^m}.$$
We then compute the Hausdorff dimension $\on{hd}(S(\phi),\Stab(\phi))$ in $\Stab(\phi)$ as
\begin{eqnarray*}
\on{hd}(S(\phi),\Stab(\phi)) & = & \lim_{n \rightarrow \infty} \frac{\log |S_n(\phi)|}{\log |\Stab_n(\phi)|} \\
& = & \lim_{n \rightarrow \infty} \frac{(\log d!) \cdot \frac{d^n - d^{n-m}}{d-1} - m}{(\log d!) \cdot \frac{d^n - 1}{d-1} - n} \\
& = & 1 - d^{-m}.
\end{eqnarray*}
Since $m \geq 1, d \geq 2$, we have that $\on{hd}(S(\phi),\Stab(\phi)) = 1 - d^{-m} < 1$, so we have that $\on{hd}(G(\phi),\Stab(\phi)) < 1$. It follows that $[\Stab(\phi) : G(\phi)]$ is necessarily infinite.
\end{proof}
\begin{remark}
It follows from Proposition~\ref{prop3} that $[\Aut(T(\phi)) : G(\phi)] = \infty$ for $\phi$ satisfying the conditions of Problem (1), as stated in Section~\ref{intro}. Also, from the proof of Proposition~\ref{prop3}, one can easily determine that
$$[\Stab_n(\phi) : S_n(\phi)] = \frac{(d!)^{\frac{d^{n-m}-1}{d-1}}}{d^{n-m}},$$
which is trivially bounded below by
$$(d!)^{\frac{d^{n-m}-1}{d-1} - (n-m)}.$$
\end{remark}
\subsection{The Size of $G(\phi)$ as a Subgroup of $S(\phi)$}
\noindent One question that arises from the proof of Proposition~\ref{prop3} is whether $[S(\phi) : G(\phi)]$ is finite, a situation that would allow us to take $N(\phi) = S(\phi)$. Restricting to the case of $K = \BQ$, we provide an example of when $[S(\phi) : G(\phi)]$ is in fact equal to $1$. Notice that \emph{very} few examples exist where one can precisely determine $G(\phi)$; one can refer to Section 2.1 of~\cite{jone} for more discussion on this point.
\begin{example}\label{ex2}
Take $K = \BQ$ and $\phi(x) = x^2 + x \in \BQ(x)$. Then, $m = 1$ and we have $G_n(\phi) \hookrightarrow S_n(\phi) \simeq \Aut(T_n'(\phi)),$ where $T_n'(\phi)$ denotes the subtree of $T_n(\phi)$ rooted at $-1$, the nonzero child of the root of $T_n(\phi)$. Observe that for each $n \geq 1$ we have $G_n(\phi) = \Gal(\BQ(\phi^{-n}(0))/\BQ) = \Gal(\BQ(\phi^{-(n-1)}(-1))/\BQ)$. Now let $\mu$ be the M\"{o}bius transformation defined by $\mu(x) = x - 1$. Then taking $\psi = \mu^{-1} \circ \phi \circ \mu$, we have an equality of Galois groups
\begin{eqnarray*}
\Gal(\BQ(\phi^{-(n-1)}(-1))/\BQ) & = & \Gal(\BQ(\psi^{-(n-1)}(\mu^{-1}(-1)))/\BQ) \\
& = & \Gal(\BQ(\psi^{-(n-1)}(0))/\BQ).
\end{eqnarray*}
But notice that $\psi$ is precisely the polynomial corresponding to Sylvester's sequence, which was studied in~\cite{odo2}. Indeed, we have
$$\psi(x) = [(x-1)^2 + (x-1)] + 1 = x^2 - x + 1.$$
From~\cite{odo2}, we know that for each $n$ we have an isomorphism
$$\Gal(\BQ(\psi^{-(n-1)}(0))/\BQ) \simeq \Aut(T_{n-1}(\psi)) = \Aut(T_n'(\phi)).$$
It follows that for each $n \geq 1$ we have the \emph{equality}, not just an injection,
$$G_n(\phi) = S_n(\phi).$$
Because $G(\phi)$ is the inverse limit of $\{G_n(\phi)\}$ and because $S(\phi)$ is the inverse limit of $\{S_n(\phi)\}$, we have that $G(\phi) = S(\phi)$. Thus, $[S(\phi) : G(\phi)] = 1$.
\end{example}
In the case of $K = \BQ$, the method presented in Example~\ref{ex2} can be used to analyze any map of the form $\phi(x) = x^2+ kx \in \BZ[x]$. We find that the following result, which can be deduced from p.~7 of~\cite{jone} and from Theorem 1.2 of~\cite{jone2}, is useful for studying rational functions of the form $\phi(x) = x^2 + kx$.
\begin{theorem}[\cite{jone2,jone}]\label{thm7}
Let $K = \BQ$, and let $\phi(x) \in \BZ[x]$ be
$$\phi(x) = x^2 - kx + k.$$
Then if $k \in \BZ\setminus \{-2,0,2,4\}$, we have that $[\Aut(T(\phi)) : G(\phi)] < \infty$. Moreover, for all $k, a_0 \in \BZ$, we have that $\on{nd}(\phi, a_0) = 0$.
\end{theorem}
Using the first part of Theorem~\ref{thm7}, we obtain the following result.
\begin{theorem}\label{thm8}
Let $K = \BQ$, and let $\phi(x) \in \BZ[x]$ \mbox{be of the form}
$$\phi(x) = x^2 + kx.$$
Then if $k \in \BZ\setminus \{-2,0,2,4\}$, we have that $[S(\phi) : G(\phi)] < \infty$.
\end{theorem}
\begin{proof}
We have $G_n(\phi) \hookrightarrow S_n(\phi) \simeq \Aut(T_n'(\phi)),$ where $T_n'(\phi)$ denotes the subtree of $T_n(\phi)$ rooted at $-k$, the nonzero child of the root of $T_n(\phi)$. For each $n \geq 1$, we have $G_n(\phi) = \Gal(\BQ(\phi^{-n}(0))/\BQ) = \Gal(\BQ(\phi^{-(n-1)}(-k))/\BQ)$. Let $\mu$ be the M\"{o}bius transformation defined by $\mu(x) = x - k$. Then taking $\psi = \mu^{-1} \circ \phi \circ \mu$ as in Example~\ref{ex2}, we have that $\Gal(\BQ(\phi^{-(n-1)}(-k))/\BQ) = \Gal(\BQ(\psi^{-(n-1)}(\mu^{-1}(-k)))/\BQ) =\Gal(\BQ(\psi^{-(n-1)}(0))/\BQ)$, from which we deduce that $$G_n(\phi) = G_{n-1}(\psi) \hookrightarrow \Aut(T_{n-1}(\psi)) \simeq \Aut(T_n'(\phi)) \simeq S_n(\phi).$$
Consider the inverse systems $\{G_n(\phi)\} \to G(\phi)$, $\{G_{n-1}(\psi)\}$ $\to G(\psi)$, $\{\Aut(T_{n-1}(\psi))\}$ $\to \Aut(T(\psi))$, and $\{S_n(\phi)\} \to S(\phi)$, and observe that we have the equalities $[G(\psi) : G(\phi)] = [S(\phi) : \Aut(T(\psi))] = 1$. Now suppose $k \notin \{-2,0,2,4\}$. Notice that $\psi(x) = x^2 - kx + k$, so since $k \notin \{-2,0,2,4\}$, Theorem~\ref{thm7} implies that $[\Aut(T(\psi)) : G(\psi)] < \infty$. It follows from the multiplicativity of the group index that $[S(\phi) : G(\phi)] < \infty$.
\end{proof}
It is natural to consider what can be deduced in the case that $k \in \{-2,0,2,4\}$. If $k = 0$, then $G(\phi)$ is trivial, so necessarily we have $[S(\phi) : G(\phi)] = \infty$. The question of whether $[S(\phi) : G(\phi)]$ is finite is more difficult to answer when $k \in \{-2,2,4\}$. The following result handles each of these three cases individually.
\begin{theorem}\label{thm9}
Let $K = \BQ$, and let $\phi(x) \in \BZ[x]$ \mbox{be of the form}
$$\phi(x) = x^2 + kx.$$
Then if $k \in \{-2,2,4\}$, we have that $[S(\phi) : G(\phi)] = \infty$. Specifically:
\begin{enumerate}
\item When $k = -2$, $G(\phi)$ has index $2$ in a subgroup of $\Aut(T(\phi))$ that is isomorphic to $\underset{\leftarrow n}\lim \, (\BZ/(3 \cdot 2^n)\BZ)^*$.
\item When $k = 2$, $G(\phi) \simeq \BZ/2\BZ \times \BZ_2$, where $\BZ_2$ is the ring of $2$-adic integers.
\item When $k = 4$, $G(\phi)$ has index $2$ in a subgroup of $\Aut(T(\phi))$ that is isomorphic to $\BZ/2\BZ \times \BZ_2$.
\end{enumerate}
\end{theorem}
\begin{proof}
Much of the argument used to prove Theorem~\ref{thm8} still applies: our method still consists of reducing the question to computing the Galois action on a subtree of $T(\phi)$, and then employing a direct analysis. Indeed, regardless of our choice of $k$, we have $G_n(\phi) \hookrightarrow S_n(\phi) \simeq \Aut(T_n'(\phi)),$ where $T_n'(\phi)$ denotes the subtree of $T_n(\phi)$ rooted at the ``$-k$'' child of the root of $T_n(\phi)$. For each $n \geq 1$, we have $G_n(\phi) = \Gal(\BQ(\phi^{-n}(0))/\BQ) = \Gal(\BQ(\phi^{-(n-1)}(-k))/\BQ)$.
\begin{enumerate}
\item Suppose $k = -2$. Then $\phi(x) = x^2 - 2x$, and conjugating $\phi$ by the M\"{o}bius transformation $\mu(x) = x +1$, we obtain $\psi(x) = (\mu^{-1}\circ\phi\circ\mu)(x) = x^2 - 2$. We then have the equality of Galois groups $G_n(\phi) = \Gal(\BQ(\phi^{-(n-1)}(2))/\BQ) = \Gal(\BQ(\psi^{-(n-1)}(1))/\BQ)$. Notice that the roots of the equation $\psi^n(x) = 1$ are always real and that $\zeta_{3 \cdot 2^{n+1}} + \zeta_{3 \cdot 2^{n+1}}^{-1}$ is a root of $\psi^n(x) = 1$ for each $n$, where $\zeta_{m}$ is a primitive $m^{\mathrm{th}}$ root of unity. Further observe that we have the inclusions
$$\BQ(\zeta_{3 \cdot 2^n} + \zeta_{3 \cdot 2^n}^{-1}) \hookrightarrow \BQ(\psi^{-(n-1)}(1))/\BQ \hookrightarrow \BQ(\zeta_{3 \cdot 2^n})$$
and that the degree of the composite extension is $[\BQ(\zeta_{3 \cdot 2^n}) : \BQ(\zeta_{3 \cdot 2^n} + \zeta_{3 \cdot 2^n}^{-1})] = 2$. Since $\BQ(\zeta_{3 \cdot 2^n})$ is not a real extension of $\BQ$, we have that $\BQ(\zeta_{3 \cdot 2^n} + \zeta_{3 \cdot 2^n}^{-1}) = \BQ(\psi^{-(n-1)}(1))/\BQ$. Then $[\BQ(\zeta_{3 \cdot 2^n}) : \BQ(\psi^{-(n-1)}(1))/\BQ] = 2$, so we find that
$$|G_n(\phi)| = |\Gal(\BQ(\psi^{-(n-1)}(1))/\BQ)| = \frac{|\Gal(\BQ(\zeta_{3 \cdot 2^n}))|}{2} = \frac{2^n}{2} = 2^{n-1}.$$
Thus, we can compute the Hausdorff dimension of $G(\phi)$ in $S(\phi)$ to be
$$\on{hd}(G(\phi), S(\phi)) = \lim_{n \rightarrow \infty} \frac{\log 2^{n-1}}{\log 2^{2^{n-1}-1}} = \lim_{n \rightarrow \infty} \frac{n-1}{2^{n-1} - 1} = 0 < 1,$$
so indeed $[S(\phi) : G(\phi)] = \infty$. We also have that $G(\phi)$ embeds with index $2$ into the inverse limit of the inverse system $\{\Gal(\BQ(\zeta_{3 \cdot 2^n}))\} \simeq \{(\BZ/(3 \cdot 2^n)\BZ)^*\}$, since for each $n$ we have
$$\Gal(\BQ(\zeta_{3 \cdot 2^n})) \simeq (\BZ/(3 \cdot 2^n)\BZ)^*.$$
\item Suppose $k = 2$. Then $\phi(x) = x^2 + 2x$, and conjugating $\phi$ by the M\"{o}bius transformation $\mu(x) = x - 1$, we obtain $\psi(x) = (\mu^{-1} \circ \phi \circ \mu)(x) = x^2$. We then have the equality of Galois groups $G_n(\phi) = \Gal(\BQ(\phi^{-(n-1)}(-2))/\BQ) = \Gal(\BQ(\psi^{-(n-1)}(-1))/\BQ)$. Thus, to determine $G_n(\phi)$, it suffices to determine $\Gal(\BQ(\psi^{-(n-1)}(-1))/\BQ)$, which is the Galois group of the polynomial $x^{2^{n-1}} + 1$. Notice that the Galois group of $x^{2^n} - 1$ is $\BZ/2\BZ \times \BZ/2^{n-2}\BZ$ when $n \geq 2$, and a splitting field for both polynomials $x^{2^{n-1}} + 1$ and $x^{2^n} - 1$ is $\BQ(\zeta_{2^n})$. It follows that we have the isomorphism $$G_n(\phi) \simeq \BZ/2\BZ \times \BZ/2^{n-2}\BZ$$
for each $n$. Thus $|G_n(\phi)| = 2^{n-1}$, and from the proof of Proposition~\ref{prop3} we have that $|S_n(\phi)| = 2^{2^{n-1} -1}$. Thus, we compute the Hausdorff dimension of $G(\phi)$ in $S(\phi)$ to be
$$\on{hd}(G(\phi), S(\phi)) = \lim_{n \rightarrow \infty} \frac{\log 2^{n-1}}{\log 2^{2^{n-1}-1}} = \lim_{n \rightarrow \infty} \frac{n-1}{2^{n-1} - 1} = 0 < 1,$$
so indeed $[S(\phi) : G(\phi)] = \infty$. We also have that $G(\phi)$ is the inverse limit of the system $\{G_n(\phi)\} \simeq \{\BZ/2\BZ \times \BZ/2^{n-2}\BZ\}$, which is $\BZ/2\BZ \times \BZ_2$, where $\BZ_2$ is the ring of $2$-adic integers.
\item Suppose $k = 4$. Then $\phi(x) = x^2 + 4x$, and the nonperiodic branch of $T(\phi)$ is rooted at $-4$, the nonzero child of the root of $T(\phi)$. The only child of $-4$ is $-2$, so we have that $G_n(\phi) = \Gal(\BQ(\phi^{-(n-2)}(-2))/\BQ)$. Conjugating $\phi$ by the M\"{o}bius transformation $\nu(x) = x +2$, we obtain $\theta(x) = (\nu^{-1}\phi\nu)(x) = x^2 - 2$. We then have the equality of Galois groups $G_n(\phi) = \Gal(\BQ(\theta^{-(n-2)}(0))/\BQ)$. By an argument analogous to the one used in the proof of the first case above, we find that $|G_n(\phi)| = 2^{n-2}$. Thus, we compute the Hausdorff dimension of $G(\phi)$ in $S(\phi)$ to be
$$\on{hd}(G(\phi), S(\phi)) = \lim_{n \rightarrow \infty} \frac{\log 2^{n-2}}{\log 2^{2^{n-1}-1}} = \lim_{n \rightarrow \infty} \frac{n-2}{2^{n-1} - 1} = 0 < 1,$$
so indeed $[S(\phi) : G(\phi)] = \infty$. We also have that $G(\phi)$ embeds with index $2$ into the inverse limit of the system $\{\Gal(\BQ(\zeta_{2^n}))\} \simeq \{\BZ/2\BZ \times \BZ/2^{n-2}\BZ\}$, which is $\BZ/2\BZ \times \BZ_2$.
\end{enumerate}
This concludes the proof of Theorem~\ref{thm9}.
\end{proof}
\begin{remark}
The third case in Theorem~\ref{thm9} serves as a companion to results presented in~\cite{reu}, where the Galois groups of iterates of polynomials $\phi(x) = (x + t)^2 - t - 2$ are studied for $t \in \BZ$ and $|t| > 2$. Indeed, notice that $\phi(x) = (x + t)^2 - t - 2 = x^2 + 4x$ when $t = 2$.
\end{remark}
It remains to be shown whether the stochastic methods discussed in~\cite{jone} can be extended to obtain zero-density results for $P_\phi(a_0)$ when the forward orbit of $0$ under $\phi$ is periodic. If such a generalization is possible, then Theorems~\ref{thm8} and~\ref{thm9} can be utilized to obtain zero-density results for $P_\phi(a_0)$ in the case of $\phi(x) = x^2 + kx$. Nevertheless, the fact that $x^2 + kx$ is conjugate to $x^2 - kx + k$ taken together with the fact that we already have zero-density results for $P_\phi(a_0)$ when $\phi(x) = x^2 - kx + k$ (see the second part of Theorem~\ref{thm7}) suggests that it may be easy to deduce zero-density results for the family of functions $\phi(x) = x^2 + kx$, where $k \in \BZ$. Indeed, we now state and prove that $\on{nd}(\phi, a_0) = 0$ for this family.
\begin{theorem}\label{thm13}
Let $K = \BQ$, and let $\phi(x) \in \BZ[x]$ be of the form
$$\phi(x) = x^2 + kx.$$
Then for all $a_0 \in \BZ$, we have that $\on{nd}(\phi, a_0) = 0$.
\end{theorem}
\begin{proof}
Observe that $\phi^n(x)$ factorizes over the integers as
$$\phi^n(x) = x(x+k)\prod_{i = 1}^{n-1} (\phi^i(x) + k).$$
Fix $a_0 \in \BZ$. Let $P'_\phi(a_0)$ be the set of all $p \in P_\phi(a_0)$ such that $p \mid \phi^i(a_0) + k$ for some $i \in \BN$, and fix $p \in P'_\phi(a_0)$. Also, let $\mu$ be the M\"{o}bius transformation defined by $\mu(x) = x - k$, and define $\psi \in \BZ[x]$ by $\psi = \mu^{-1} \circ \phi \circ \mu$. Notice that $\psi^i(x) = \phi^i(x-k) + k$. Thus, we have that $p \mid \psi^i(a_0 + k)$, which implies that $p \in P_\psi(a_0 + k)$, and so $P'_\phi(a_0) \subset P_\psi(a_0 + k)$. But, as observed in the proof of Theorem~\ref{thm8}, $\psi(x) = x^2 - kx + k$, so by the second part of Theorem~\ref{thm7}, $P_\psi(a_0 + k)$ has density $0$ in the set of all primes. It follows that $P'_\phi(a_0)$ has density $0$ in the set of all primes.
Now observe that if $p \mid \phi^n(a_0)$ for some $n \in \BN$ but $p \not\in P'_\phi(a_0)$, then $p \mid a_0^2 + ka_0$. Therefore, $P_\phi(a_0) - P'_\phi(a_0)$ is a finite set, so because $P'_\phi(a_0)$ has density $0$ in the set of all primes, we have that $\on{nd}(\phi, a_0) = 0$.
\end{proof}
Theorems~\ref{thm8} and~\ref{thm9} handle the family of functions $\phi(x) = x^2 + kx$ for $k \in \BZ$. A similar process may be used to analyze the family of functions $\phi(x) = x^2 - (k + 1)x + k \in \BZ[x]$, for which $d = 2$ and $m = 2$ (recall with the family $\phi(x) = x^2 + kx$, we had $d = 2$ and $m = 1$). We find that the following result, which can be deduced from p.~7 of~\cite{jone} and from Theorem 1.2 of~\cite{jone2}, is useful for studying rational functions of the form $\phi(x) = x^2 - (k+1)x + k$.
\begin{theorem}[\cite{jone2, jone}]\label{thm10}
Let $K = \BQ$, and let $\phi(x) \in \BZ[x]$ be
$$\phi(x) = x^2 + kx - 1.$$
Then if $k \in \BZ\setminus \{-1,0,2\}$, we have that $[\Aut(T(\phi)) : G(\phi)] < \infty$.
\end{theorem}
Using Theorem~\ref{thm10}, we obtain the following result, which notably bears more restrictions than its counterpart Theorem~\ref{thm8}.
\begin{theorem}\label{thm11}
Let $K = \BQ$, and let $\phi(x) \in \BZ[x]$ be of the form
$$\phi(x) = x^2 - (k+1)x + k.$$
Observing that $G(\phi) \subset G(\phi, 1) \times G(\phi, k+1)$, suppose $[G(\phi, 1) \times G(\phi, k+1) : G(\phi)] < \infty$. Then if $k \in \BZ\setminus \{-1,1,2\}$, we have that $[S(\phi) : G(\phi)] < \infty$.
\end{theorem}
\begin{proof}
In the proof of Proposition~\ref{prop3}, we found that $|S_n(\phi)| = 2^{2^n - 2^{n-2}-2}$. It is in fact difficult to compute $|G_n(\phi)|$ for each $n$. Observe that the vertices of level $1$ are $k$ and $1$ and that the children of $k$ (which live on level $2$) are $0$ and $k + 1$. Taking $n > 2$, we observe that automorphisms of $T_n(\phi)$ that belong to $G_n(\phi)$ are entirely determined by their actions on the complete subtrees rooted at the ``$1$'' vertex of level $1$ and the ``$k+1$'' vertex of level $2$. But what makes $|G_n(\phi)|$ difficult to compute is that perhaps for some $\phi$ we might have $G_n(\phi) \not\simeq G_{n-1}(\phi, 1) \times G_{n-2}(\phi, k+1)$, even though we always have $G_n(\phi) \subset G_{n-1}(\phi, 1) \times G_{n-2}(\phi, k+1)$. In the case where $[G(\phi, 1) \times G(\phi, k+1) : G(\phi)] < \infty$, we may resolve this problem by simply replacing $G_n(\phi)$ with $G_{n-1}(\phi, 1) \times G_{n-2}(\phi, k+1)$.
Let $G''(\phi) = G(\phi, 1) \times G(\phi, k+1)$, which is the inverse limit of the system $\{G_{n-1}(\phi, 1) \times G_{n-2}(\phi, k+1)\}$. Let $\mu$ be the M\"{o}bius transformation defined by $\mu(x) = x + 1$. Then taking $\psi = \mu^{-1} \circ \phi \circ \mu$, we have an equality of Galois groups $\Gal(\BQ(\phi^{-(n-1)}(1))/\BQ) = \Gal(\BQ(\psi^{-(n-1)}(0))/\BQ)$, from which we deduce that $G_{n-1}(\phi,1) = G_{n-1}(\psi) \hookrightarrow \Aut(T_{n-1}(\psi))$. Denote by $T_n'(\phi)$ the subtree of $T_n(\phi)$ rooted at the ``$1$'' child of the root $0$. Similarly, let $\nu$ be the M\"{o}bius transformation defined by $\nu(x) = x + (k+1)$. Then taking $\pi = \nu^{-1} \circ \phi \circ \nu$, we have an equality of Galois groups
$\Gal(\BQ(\phi^{-(n-2)}(k+1))/\BQ) = \Gal(\BQ(\pi^{-(n-2)}(0))/\BQ)$,
from which we deduce that $G_{n-2}(\phi,k+1) = G_{n-2}(\pi) \hookrightarrow \Aut(T_{n-2}(\pi))$. Denote by $T_n''(\phi)$ the subtree of $T_n(\phi)$ rooted at the ``$k+1$'' child of the level $1$ vertex $k$. Then $\Aut(T_{n-2}(\pi)) \simeq \Aut(T_n''(\phi))$. Now we have that $\Aut(T_n'(\phi)) \times \Aut(T_n''(\phi)) \simeq S_n(\phi)$, so we obtain the embeddings
$$G_{n-1}(\phi,1) \times G_{n-2}(\phi, k_1) = G_{n-1}(\psi) \times G_{n-2}(\pi) \hookrightarrow$$ $$\Aut(T_n'(\phi))\times \Aut(T_n''(\phi)) \simeq S_n(\phi).$$
Let the inverse limit of the system $\{\Aut(T_n'(\phi)) \times \Aut(T_n''(\phi))\}$ be $A'(\phi)$. Then consider the inverse systems
$$\{G_{n-1}(\phi,1) \times G_{n-2}(\phi, k_1)\} \to G''(\phi), \{G_{n-1}(\psi) \times G_{n-2}(\pi)\} \to G(\psi) \times G(\pi),$$ $$\{\Aut(T_n'(\phi)) \times \Aut(T_n''(\phi))\} \to A'(\phi), \{S_n(\phi)\} \to S(\phi),$$
and observe that we have the equality $[G(\psi) \times G(\phi): G''(\phi)] = [S(\phi) : A'(\phi)] = 1$. Now suppose $k \notin \{-1,1,2\}$. Notice that $\psi(x) = x^2 + (1-k)x -1$, so since $k \notin \{-1,1,2\}$, Theorem~\ref{thm10} implies that $[A'(\phi) : G(\psi) \times G(\pi)] < \infty$. It follows by \mbox{multiplicativity of the group index that $[S(\phi) : G(\phi)] < \infty$.}
\end{proof}
In light of Theorem~\ref{thm11}, it is again natural to wonder what can be deduced in the case that $k \in \{-1,1,2\}$. The question of whether $[S(\phi) : G(\phi)]$ is finite is more difficult to answer when $k \in \{-1,1,2\}$, because we cannot simply rely on Theorem~\ref{thm10}. Although it is not easy to obtain an analogue of Theorem~\ref{thm9}, we can at least show that the cases of $k = \pm 1$ are equivalent. Indeed, if we denote by $\phi_-$ and $\phi_+$ the polynomials $\phi_-(x) = x^2 - 1$ and $\phi_+(x) = x^2 - 2x + 1$, then one can check that $G(\phi_+) = G(\phi_-) = G(\phi_-, 1)$.
\section{The Size of $C(\phi)$ as a Subgroup of $\Aut(T(\phi))$}\label{trees}
\noindent We now consider the case where $\phi$ commutes with a nontrivial M\"{o}bius transformation that fixes $0$. Let $A(\phi)$ be the group of all such M\"{o}bius transformations (together with the identity). Observe that $A(\phi) \hookrightarrow \Aut(T(\phi))$, and let $C(\phi)$ be the centralizer of $A(\phi)$ in $\Aut(T(\phi))$. Further observe that $G(\phi) \subset C(\phi)$. Since $C(\phi)$ is the obvious candidate for $N(\phi)$ in the case when $A(\phi)$ is nontrivial, we want to better understand $C(\phi)$ as a subgroup of $\Aut(T(\phi))$. In this regard, it is conjectured in~\cite{jone} that the index $[\Aut(T(\phi)) : C(\phi)]$ is infinite.
\begin{conjecture}[Conjecture 3.5 of~\cite{jone}]\label{conj2}
When $A(\phi)$ is nontrivial, $[\Aut(T(\phi)) : C(\phi)] = \infty$.
\end{conjecture}
On p. 19 of~\cite{JM}, Conjecture~\ref{conj2} is proved to be true for quadratic rational functions that have the form
$$\phi(x) = \frac{k(x^2 + 1)}{x} \text{ for } k \in K\setminus\{-1/2,0,1/2\}.$$
The method used in~\cite{JM} is to show that the Hausdorff dimension (recall Definition~\ref{def1}) of $C(\phi)$ in $\Aut(T(\phi))$ is less than $1$. We now modify and extend this method to prove Conjecture~\ref{conj2} in the general case.
Let $T_n$ be a complete rooted $d$-ary tree of height $n$, and for each $i \in \{0, \dots, n\}$ let $T_i$ denote the subtree of height $i$ whose root is that of $T_n$. We extend Proposition 4.1 of~\cite{JM}, which relates the centralizers of some subgroups of $\Aut(T_n)$ to automorphism groups of smaller trees in the case where $d = 2$, as follows.
\begin{proposition}\label{prop1}
Fix $n \in \BN$ and a subgroup $H \subset \Aut(T_n)$, and for each $i \leq n$, let $C_i \subset \Aut(T_i)$ be the centralizer of $H|_{T_i} = \{g|_{T_i} : g \in H\} \subset \Aut(T_i)$. Consider the natural map $C_n \to C_i$ defined by restriction. Then there is an injective homomorphism $$h : \ker(C_n \to C_i) \rightarrow \Aut(T_{n-i})^{m(i)},$$ where $m(i)$ is the number of orbits in the action of $H$ on the vertices of level $i$. Moreover, the injective map $h$ is an isomorphism if the action of $H$ on the vertices of level $i$ is free.\footnote{Recall that the action of a group $G$ on a set $S$ is free if $gx = x \Rightarrow g = 1$ for all $g \in G, x \in X$.}
\end{proposition}
\begin{proof}
Let the vertices of level $i$ be denoted $x_1, \dots, x_{d^i}$. Suppose the action of $H$ on $\{x_1, \dots, x_{d^i}\}$ has orbits $O_1, \dots, O_{m(i)}$, and pick a representative $y_k \in O_k$ for each $k \in \{1, \dots, m(i)\}$. Define the map
$$h : \ker(C_n \to C_i) \rightarrow \Aut(T_{n-i})^{m(i)} \text{ by } \tau \mapsto \left(\tau|_{T_{y_1}}, \dots, \tau|_{T_{y_{m(i)}}}\right),$$
where for each $k \in \{1, \dots, {m(i)}\}$ we denote by $T_{y_k}$ the complete $d$-ary subtree of $T_n$ rooted at $y_k$ (each $T_{y_k}$ has height $n-i$, so each $\tau|_{T_{y_k}}$ can be viewed as an element of $\Aut(T_{n-i})$).
It is clear that $h$ is a homomorphism. We now show that $h$ is injective. Suppose $\tau \in \ker(h)$; i.e., $\tau \in \ker(C_n \to C_i)$ acts by the identity on each of the trees $T_{y_k}$. By definition, $\tau(g(x)) = g(\tau(x))$ for any $g \in H$ and $x \in T_n$, since $\tau \in C_n$. If $x \in T_{y_k}$ for any $k$, then we also have that $\tau(g(x)) = g(x)$. Now suppose $z \in T_{x_j}$ for some $j \in \{1, \dots, d^i\}$, and suppose $x_j \in O_k$. Since the action of $H$ on $O_k$ is transitive, there exists $g \in H$ such that $g(x_j) = y_k$. Observe that $g(z) \in T_{y_k}$, so $\tau(z) = \tau(g^{-1}(g(z))) = g^{-1}(g(z)) = z$. Since $\tau \in \ker(C_n \to C_i)$, we have that $\tau$ also fixes $T_i$. It follows that $\tau = \id$, so $h$ is injective.
Suppose the action of $H$ on the vertices of level $i$ is free. We then show that $h$ is surjective. Take a list of automorphisms $$(\tau_1, \dots, \tau_{m(i)}) \in \Aut(T_{y_1}) \times \dots \times \Aut(T_{y_{m(i)}}) \simeq \Aut(T_{n-i})^{m(i)}.$$ Define $\wt{\tau} \in \Aut(T_n)$ as follows. Because the action of $H$ on the vertices of level $i$ is free, for each $j \in \{1, \dots, d^i\}$ there exists a unique $g_j \in H$ such that $g_j x_j \in \{y_1, \dots, y_{m(i)}\}$. Then let $\wt{\tau}|_{T_{i-1}} = \id$, and if $x_j \in O_\ell$ let $\wt{\tau}|_{T_{x_j}} = g_j^{-1} \tau_\ell g_j$ for $j \in \{1, \dots, d^i\}$. It is clear that $\wt{\tau}$ is a well-defined tree automorphism, so indeed $\wt{\tau} \in \Aut(T_n)$. We now check that $\wt{\tau} \in C_n$; i.e., we check that $g^{-1} \wt{\tau} g = \wt{\tau}$ for each $g \in H$. Fix $g \in H$. Then if $v \in T_{i-1}$, we have $\wt{\tau}(v) = v$ and $\wt{\tau}(g(v)) = g(v)$ since $g(v) \in T_{i-1}$, so clearly $(g^{-1}\wt{\tau}g)(v) = \wt{\tau}(v)$. If $v \in T_{x_j}$ for some $j \in \{1, \dots, d^i\}$, then since $g$ is a tree automorphism, we have that $g(v) \in T_{g(x_j)}$. Let $g(x_j) = x_k$. Then, $g_k = g_j g^{-1}$, and so, $(g^{-1}\wt{\tau}g)(v) = (g^{-1}(gg_j^{-1}\tau_\ell g_j g^{-1})g)(v) = (g_j^{-1}\tau_\ell g_j)(v) =\wt{\tau}(v)$, as desired. We finally notice that $\wt{\tau} \in \ker(C_n \to C_i)$ because $\wt{\tau}|_{T_i}$ acts by the identity.
\end{proof}
\begin{remark}
Proposition 4.1 of~\cite{JM} handles the case where $d = 2$, $i = 1$, and $H$ is generated by an involution that swaps the vertices of level $1$.
The isomorphism of Proposition~\ref{prop1} may seem easy to deal with in the case that $H$ acts simply transitively on the vertices of level $i$. In this regard, it is natural to want to describe the subgroups of $\Aut(T_n)$ that act simply transitively on the vertices of level $i$ for some $i$, but there are some obstacles to doing so. Indeed, suppose $H$ acts simply transitively on the vertices of level $i$. Then it is \emph{not necessarily} true that the image of $H$ under the restriction map from $\Aut(T_i)$ to $\Aut(T_{i-1})$ acts simply transitively on the vertices of level $i-1$. As an example, consider the case of $d = 2$ and $i = 3$, and let $x_1, x_2, x_3, x_4$ be the vertices of level $2$. Notice that $\Aut(T_2)$ is a nonabelian group of order $8$ that contains a Klein 4-subgroup. It follows from the classification of groups of order $8$ that $\Aut(T_2)$ is isomorphic to the dihedral group of order $8$. This isomorphism is given explicitly by $\Aut(T_2) \simeq \langle a,x : a^4 = x^2 = 1, xax = a^3\rangle$, where $a$ is the automorphism that sends $(x_1, x_2, x_3, x_4) \mapsto (x_3, x_4, x_2, x_1)$ and $x$ is unique nontrivial element of $\Aut(T_1) \subset \Aut(T_2)$. Now, let $H \subset \Aut(T_3)$ be the subgroup defined by $H \simeq \langle \wt{a},x \rangle$, where $\wt{a}$ is the automorphism that acts by $a$ on $T_2$ and that also swaps the children of each of the nodes $x_1, x_2, x_3, x_4$. Notice that we have the relations $\wt{a}^4 = x^2 = 1, x\wt{a}x = \wt{a}^3$, so $H$ is isomorphic to the dihedral group of order $8$. It is now easy to check that $H \subset \Aut(T_3)$ is a subgroup that acts simply transitively on the vertices of level $3$ but whose restriction to $\Aut(T_2)$ does not act freely on the vertices of level $2$, because the restriction of $H$ to $\Aut(T_2)$ is $\Aut(T_2)$ itself.
\end{remark}
Now let $T$ be a complete rooted $d$-ary tree of infinite height, and for each $n \in \BN$ let $T_n$ denote the subtree of height $n$ whose root is that of $T$. Corollary 4.2 of~\cite{JM} uses the result of Proposition 4.1 of~\cite{JM} to describe the size of the centralizer of a subgroup of $\Aut(T)$, in the case where $d = 2$, $i = 1$, and $H$ is generated by an involution that swaps the vertices of level $1$. The proof of the following lemma does not use the method presented in Corollary 4.2 of~\cite{JM} but still gives an analogous result.
\begin{lemma}\label{cor1}
Let $H \subset \Aut(T)$ be a subgroup, and let $C$ denote the centralizer of $H$ in $\Aut(T)$. Assume the notation of Proposition~\ref{prop1}. Then the Hausdorff dimension of $C$ in $\Aut(T)$ is at most ${m(i)}/d^i$ for each $i \in \BN$, with equality when the \mbox{action of $H$ on the vertices of level $i$ is free.}
\end{lemma}
\begin{proof}
The Hausdorff dimension $\on{hd}(C, \Aut(T))$ of $C$ in $\Aut(T)$ is given by the limit
$$\on{hd}(C,\Aut(T)) = \lim_{n \rightarrow \infty} \frac{\log |C_n|}{\log |\Aut(T_n)|},$$ because $C_n$ is the restriction of $C$ to $\Aut(T_n)$ for each $n$.
Notice that
$$\log |\Aut(T_n)| = \log\left((d!)^{\frac{d^{n}-1}{d-1}}\right) = \frac{d^{n}-1}{d-1} \cdot \log d!.$$
Next, from Proposition~\ref{prop1}, we deduce that $|\ker(C_n \to C_i)| \leq |\Aut(T_{n-i})^{m(i)}|$. We then compute the order of the product of the automorphism groups $\Aut(T_{n-i})^{m(i)}$ to be
$$|\Aut(T_{n-i})^{m(i)}| = (d!)^{{m(i)} \cdot \frac{d^{n-i} - 1}{d-1}}.$$
Thus, we obtain an upper bound on $|C_n|$ since the restriction map $C_n \to C_i$ is a surjection, namely
\begin{eqnarray*}
\log |C_n| & = & \log \big[|\im(C_n \to C_i)| \cdot |\ker(C_n \to C_i)|\big] \\
& \leq & \log |C_i| + \log |\Aut(T_{n-i})^{m(i)}| \\
& = & \log |C_i| + \left({m(i)} \cdot \frac{d^{n-i} - 1}{d-1}\right) \cdot \log d!.
\end{eqnarray*}
Combining our results, we compute an upper bound on the Hausdorff dimension of $C$, namely
\begin{eqnarray*}
\on{hd}(C,\Aut(T)) & = & \lim_{n \rightarrow \infty} \frac{\log|C_n|}{\log |\Aut(T_n)|} \\
& \leq & \lim_{n \rightarrow \infty} \frac{\log |C_i| + \left({m(i)} \cdot \frac{d^{n-i} - 1}{d-1}\right)\cdot \log d!}{\frac{d^n-1}{d-1}\cdot \log d!}\\
& = & \frac{{m(i)}}{d^i},
\end{eqnarray*}
which is the desired result.
\end{proof}
\begin{remark}
Lemma~\ref{cor1} is surprising, because the bound on $\on{hd}(C,\Aut(T))$ holds \emph{for each} $i \in \BN$. As a consequence, we have that $m(i+1) = d \cdot m(i)$ when the action of $H$ on the vertices of levels $i$ and $i+1$ is free. This consequence may be readily verified in a given example.
\end{remark}
We now use Lemma~\ref{cor1} to prove Conjecture~\ref{conj2}.
\begin{theorem}\label{thm5}
When $A(\phi)$ is nontrivial we have $[\Aut(T(\phi)) : C(\phi)] = \infty$.
\end{theorem}
\begin{proof}
It follows immediately from Lemma 12 that $\on{hd}(C(\phi), \Aut(T(\phi)) < 1$ if the action of any $\mu \in A(\phi)$ on the tree $T(\phi)$ is nontrivial. But since $\mu$ is a M\"{o}bius transformation, it can have at most two fixed points, so our assumption that $\phi^n$ has $d^n$ distinct roots already implies $\mu$ acts nontrivially on the roots of $\phi^2$.
\end{proof}
The following is an example that illustrates the proof of Theorem~\ref{thm5}.
\begin{example}\label{ex1}
Let $K = \BQ(\zeta_n)$, where $a \in \BQ_{>0}$ and $\zeta_n$ is a primitive $n^\mathrm{th}$ root of $\pm 1$ (for $n > 1$). Consider rational functions $\phi_n(x) \in K(x)$ defined as follows:
$$\phi_n(x) = \pm \frac{k(x^n \mp a)}{x^{n-1}},$$
where $k \in K$. Notice that $\phi_n$ commutes with the M\"{o}bius transformation $m_n$ defined by $m_n(x) = \zeta_nx$, and also observe that $m_n(0) = 0$ for all $n$. Consider the nontrivial subgroup $A'(\phi_n) \subset A(\phi_n)$ generated by $m_n$, and notice that $A'(\phi_n)$ acts simply transitively on level $1$ of $T$. We have that the centralizer $C'(\phi_n)$ of $A'(\phi_n)$ in $\Aut(T(\phi_n))$ contains $C(\phi_n)$. By Lemma~\ref{cor1}, the Hausdorff dimension of $C'(\phi_n)$ is $1/n$, so $[\Aut(T(\phi_n)) : C'(\phi_n)] = \infty$, implying that $[\Aut(T(\phi_n)) : C(\phi_n)] = \infty$.
\end{example}
\section{The Size of $G(\phi)$ as a Subgroup of $C(\phi)$}\label{squares}
\noindent We retain the setting of Section~\ref{trees}; i.e., $\phi$ commutes with a nontrivial M\"{o}bius transformation that fixes $0$. Theorem~\ref{thm5} provides evidence that $C(\phi)$ is a small subgroup of $\Aut(T(\phi))$. But to check that $C(\phi)$ is a good candidate for $N(\phi)$, we want to show that $[C(\phi) : G(\phi)] < \infty$. To further simplify the situation, we restrict to the case of $d = 2$. As shown in Section 2 of~\cite{JM}, it then suffices to consider functions $\phi$ that have the form
$$\phi(x) = \frac{k_0(x^2+1)}{x}.$$
We further restrict to the case of $K = \BQ$ and $k_0 \in \BZ \setminus \{0\}$. In this case, Conjecture 3.9 of~\cite{jone} (alternatively, Conjecture 1.3 of~\cite{JM}) takes the following form.
\begin{conjecture}\label{conj1}
Let $\phi(x) \in \BQ(x)$ be a rational function of the form
$$\phi(x) = \frac{k_0(x^2+1)}{x},$$ where $k_0 \in \BZ \setminus\{0\}$. Then $[C(\phi) : G(\phi)] < \infty$.
\end{conjecture}
The question of resolving Conjecture~\ref{conj1} reduces to a seemingly unrelated question regarding squares in sequences. Let the pair of polynomials $(\delta_n, \varepsilon_n)_{n = 1, 2, \ldots}$ be given by the recursion
\begin{eqnarray*}
(\delta_1(k), \varepsilon_1(k)) & = & (2k^2,k) \text{ and }\\ (\delta_n(k), \varepsilon_n(k)) & = & \left(\delta_{n-1}(k)^2 + \varepsilon_{n-1}(k)^2, \frac{\delta_{n-1}(k)\varepsilon_{n-1}(k)}{k}\right) \text{ for } n \geq 2.
\end{eqnarray*}
(Notice that $\delta_n$ and $\varepsilon_n$ are indeed polynomials for all $n$ because of the values assigned to $\delta_1, \varepsilon_1$.) Jones and Manes then obtain the following result.
\begin{theorem}[Theorem 5.3 of~\cite{JM}]\label{thm1}
Let $\phi(x) \in \BQ(x)$ be a rational function of the form
$$\phi(x) = \frac{k_0(x^2+1)}{x},$$
where $k_0 \in \BZ \setminus\{0\}$. If $\delta_n(k_0)$ is not a square in $\BZ$ for all $n \geq 2$ , then $[C(\phi) : G(\phi)] < \infty$.
\end{theorem}
We therefore make it our objective to prove that $\delta_n(k_0)$ is not a square for all integers $n \geq 2$ and $k_0 \neq 0$. In what follows, we discuss two methods of achieving this objective, namely a congruence method and an algebraic method.
\subsection{Congruence Method}
\noindent In~\cite{JM}, congruence methods are used to make progress toward this objective. Indeed, Jones and Manes obtain the following result by reducing $\delta_n(k_0)$ modulo the primes $2$, $3$, $5$, and $7$.
\begin{theorem}[Theorem 5.8 of~\cite{JM}]\label{thm2}
If one of the following congruences hold, then $\delta_n(k_0)$ is not a square for all $n \geq 2$:
\begin{itemize}
\item $k_0 \equiv 1 \pmod 2$,
\item $k_0 \equiv 1,2 \pmod 3$,
\item $k_0 \equiv 2, 3 \pmod 5$, or
\item $k_0 \equiv 1, 2, 5, 6 \pmod 7$.
\end{itemize}
\end{theorem}
\begin{comment}
\begin{proof}
Notice that $\delta_1(k) \equiv 2 \pmod{k-1}$ and $\varepsilon_1(k) \equiv 1 \pmod{k-1}$. Thus, $\delta_1(k) \equiv 2 \pmod 3$ and $\varepsilon_1(k) \equiv 1 \pmod 3$. Let $n$ be odd, and suppose $\delta_n(k) \equiv 2 \pmod 3$ and that $\varepsilon_n(k) \equiv 1 \pmod 3$. Then,
\begin{eqnarray*}
\delta_{n+1}(k) & = & \delta_n(k)^2 + \varepsilon_n(k)^2 \equiv 1 + 1 \equiv 2 \pmod 3, \\
\varepsilon_{n+1}(k) & = & \delta_n(k)\varepsilon_n(k) \equiv 2 \cdot 1 \equiv 2 \pmod 3,\\
\delta_{n+2}(k) & = & \delta_{n+1}(k)^2 + \varepsilon_{n+1}(k)^2 \equiv 1 + 1 \equiv 2 \pmod 3, \\
\varepsilon_{n+2}(k) & = & \delta_{n+1}(k)\varepsilon_{n+1}(k) \equiv 2 \cdot 2 \equiv 1 \pmod 3.
\end{eqnarray*}
It follows by induction that $\delta_n(k) \equiv 2 \pmod 3$ for all $n$, so $\delta_n(k)$ cannot be a square since $2$ is not a quadratic residue modulo 3.
\end{proof}
\end{comment}
In~\cite{JM} it is shown using such congruence methods that if $\delta_n(k_0)$ is a square in $\BZ$ for some $n,k_0$, then $|k_0| > 10,000$. The method of reducing modulo various primes also results in the following useful theorem.
\begin{theorem}[Theorem 5.8 of~\cite{JM}]\label{thm3}
For all odd $n$ we have that $\delta_n(k_0)$ is not a square for all $k_0 \neq 0$.
\end{theorem}
It is possible to obtain a result analogous to Theorem~\ref{thm2} for every prime, and using an exhaustive search, we can find the congruence classes of $k_0$ modulo a given prime for which $\delta_n(k_0)$ is not a square for all $n \geq 2$. We extend Theorem~\ref{thm2} as follows.
\begin{theorem}\label{thm6}
If one of the following congruences hold, then $\delta_n(k_0)$ is not a square for all $n$. It follows that $\delta_n(k_0)$ is not a square for all $1 \leq |k_0| \leq 10^6$.
\begin{itemize}
\item $k_0 \equiv 1 \pmod 2$,
\item $k_0 \equiv 1,2 \pmod 3$,
\item $k_0 \equiv 2, 3 \pmod 5$,
\item $k_0 \equiv 1, 2, 5, 6 \pmod 7$,
\item $k_0 \equiv 1,2,5,6,9,10 \pmod{11}$,
\item $k_0 \equiv 3,6,7,10 \pmod{13}$,
\item $k_0 \equiv 1,3,14,16 \pmod{17}$,
\item $k_0 \equiv 6,7,9,10,12,13 \pmod{19}$,
\item $k_0 \equiv 1,6,8,9,14,15,17,22\pmod{23}$,
\item $k_0 \equiv 2,11,12,14,15, 17,18,27\pmod{29}$,
\item $k_0 \equiv 6,10,11,14,17,20,21,25\pmod{31}$,
\item $k_0 \equiv 6,8,10,11,14,17,18,19,20,23,26,27,29,31\pmod{37}$,
\item $k_0 \equiv 2,11,13,14,15,26,27,28,30,39\pmod{41}$,
\item $k_0 \equiv 4,5,12,14,17,21,22,26,29,31,38,39\pmod{43}$,
\item $k_0 \equiv 4,7,11,12,15,19,21,26,28,32,35,36,40,43\pmod{47}$,
\item $k_0 \equiv 2,19,22,25,26,27,28,31,34,51\pmod{53}$,
\item $k_0 \equiv 1,3,7,8,51,25,29,30,34,52,56,58\pmod{59}$,
\item $k_0 \equiv 1,2,3,12,17,24,29,30,31,32,37,44,49,58,59,60\pmod{61}$,
\item $k_0 \equiv 5,6,9,15,18,22,27,30,32,33,34,35,37,40,45,49,$\\$52,58,61,62\pmod{67}$,
\item $k_0 \equiv 6,7,11,12,16,20,27,28,30,33,38,41,43,44,51,55,$\\$59,60,64,65\pmod{71}$,
\item $k_0 \equiv 6,7,12,13,14,20,24,29,32,33,40,41,44,49,53,59,$\\$60,61,66,67\pmod{73}$,
\item $k_0 \equiv 20,68\pmod{79}$,
\item $k_0 \equiv 12,16,21,62,71\pmod{83}$,
\item $k_0 \equiv 9,50\pmod{103}$,
\item $k_0 \equiv 106\pmod{107}$,
\item $k_0 \equiv 92 \pmod{109}$,
\item $k_0 \equiv 26,105\pmod{131}$,
\item $k_0 \equiv 89\pmod{149}$,
\item $k_0 \equiv 24\pmod{157}$,
\item $k_0 \equiv 19\pmod{173}$, or
\item $k_0 \equiv 52,145 \pmod{197}$.
\end{itemize}
\end{theorem}
\begin{proof}
For each one of the above pairs of congruence class $k_0$ and prime $p$, one can check that a full period of the (periodic) sequence $(\delta_n(k_0), \varepsilon_n(k_0))$ modulo $p$ is such that $\delta_n(k_0)$ is never a quadratic residue when $n$ is even (by Theorem~\ref{thm3}, we may assume that $n$ is even), implying that $\delta_n(k_0)$ is never a square. In the case that the sequence $(\delta_n(k_0), \varepsilon_n(k_0))$ modulo $p$ is preperiodic (i.e., the sequence behaves nonperiodically before it becomes periodic), one can check that each term of the ``preperiod'' is such that $\delta_n(k_0)$ (when not reduced modulo $p$) is never a square when $n$ is even. It follows from the methods described in~\cite{JM} that $\delta_n(k_0)$ is not a square for all $|k_0| \leq 10^6$ (in fact, these congruences work up to $1,056,575$).
\end{proof}
Theorem~\ref{thm6} provides good evidence that Conjecture~\ref{conj1} is true. Indeed, we would expect that pairs $(n,k_0)$ for which $\delta_n(k_0)$ is a square are such that $n$ and $k_0$ are both small (e.g. recall that the largest square in the Fibonacci sequence is $144$), and Theorem~\ref{thm6} indicates that $\delta_n(k_0)$ is not a square even for small $n$ and $k_0$. By repeating the method described in the proof of Theorem~\ref{thm6}, one can presumably raise the lower bound on $k_0$-values even higher than $10^6$.
\subsection{Algebraic Method}
\noindent We now present a more algebraic study of the polynomials $\delta_n, \varepsilon_n$ (i.e., one that does not involve reducing modulo various primes). To begin with, we state and prove some algebraic properties of the polynomials $\delta_n, \varepsilon_n$ as follows.
\begin{lemma}\label{lem1}
Let $\deg(f)$ denote the degree of a polynomial $f$, and let $\low(f)$ denote the degree of the lowest-degree (nonzero) term of $f$. Then we have the following results:
$$\deg(\delta_n) = 2^n,\quad \low(\delta_n) = \frac{2^n}{3} - \frac{(-1)^n}{3} + 1,$$ $$\deg(\varepsilon_n) = 2^n - n, \quad \text{ and } \quad \low(\varepsilon_n) = \frac{2^n}{3} + \frac{(-1)^n}{6} + \frac{1}{2}.$$
\end{lemma}
\begin{proof}
We first make the following three claims: (1) $\deg(\delta_n) > \deg(\varepsilon_n)$ for all $n$; (2) $\low(\delta_n) = \low(\varepsilon_n)$ for all even $n$; and (3) $\low(\delta_n) = \low(\varepsilon_n) + 1$ for all odd $n$. Claim (1) clearly holds for $n = 1$. Now suppose Claim (1) holds for some $n \geq 1$. From the equation $\delta_{n+1}(k) = \delta_n(k)^2 + \varepsilon_n(k)^2$, we find that $\deg(\delta_{n+1}) = 2\max(\deg(\delta_n), \deg(\varepsilon_n)) = 2\deg(\delta_n)$ . Also, from the equation $\varepsilon_{n+1}(k) = \frac{\delta_n(k)\varepsilon_n(k)}{k}$, we find that $\deg(\varepsilon_{n+1}) = \deg(\delta_n) + \deg(\varepsilon_n) - 1 < 2\deg(\delta_n) = \deg(\delta_{n+1})$, from which Claim (1) follows by induction.
We prove Claims (2) and (3) simultaneously. These two claims clearly hold for $n = 1$ and $n = 2$. Then suppose the claim holds for $n$ where $n$ is even, and let $a = \low(\delta_n) = \low(\varepsilon_n)$. Notice that $\low(\delta_{n+1}) = 2 \min(\low(\delta_{n}), \low(\varepsilon_{n})) = 2a$
and that $\low(\varepsilon_{n+1}) = \low(\delta_n) + \low(\varepsilon_n) -1 = 2a-1$. Also notice that $\low(\delta_{n+2}) = 2 \min(\low(\delta_{n+1}), \low(\varepsilon_{n+1})) = 4a-2$
and that $\low(\varepsilon_{n+2}) = \low(\delta_{n+1}) + \low(\varepsilon_{n+1}) -1 = 4a - 2$. Claims (2) and (3) then follow by induction.
It follows from Claims (1), (2), and (3) that we have these recursions:
\begin{align*}
\deg(\delta_n) & = 2\deg(\delta_{n-1}), \deg(\varepsilon_n) = \deg(\delta_{n-1}) + \deg(\varepsilon_{n-1}) - 1,\\
\low(\delta_n) & = 2\low(\varepsilon_{n-1}), \low(\varepsilon_n) = \low(\delta_{n-1}) + \low(\varepsilon_{n-1}) - 1.
\end{align*}
Solving these recursions then yields the lemma.
\end{proof}
\begin{lemma}\label{lem2}
The leading coefficient of $\delta_n$ is a square for all $n \geq 2$. Moreover, the degree of each term in $\delta_n$ is even for all $n$.
\end{lemma}
\begin{proof}
The first statement clearly holds for $n = 2$. From the proof of Lemma~\ref{lem1}, we have that $\deg(\delta_n) > \deg(\varepsilon_n)$ for each $n$, so the leading term of $\delta_{n+1}$ is the square of the leading term of $\delta_n$.
The second statement clearly holds for $n = 1$. From the recursion
\begin{eqnarray*}
(\delta_1(k), \varepsilon_1(k)) & = & (2k^2,k) \text{ and }\\ (\delta_n(k), \varepsilon_n(k)) & = & \left(\delta_{n-1}(k)^2 + \varepsilon_{n-1}(k)^2, \frac{\delta_{n-1}(k)\varepsilon_{n-1}(k)}{k}\right) \text{ for } n \geq 2,
\end{eqnarray*}
we find that $\varepsilon_{n-1}(k)^2 = \delta_{n-1}(k) - \delta_{n-2}(k)^2$,
and solving the recursion in terms of $\delta$'s alone yields that
$$\delta_n(k) = \delta_{n-1}(k)^2 + \frac{\delta_{n-1}(k)\delta_{n-2}(k)^2 - \delta_{n-2}(k)^4}{k^2}.$$
Suppose the second statement holds for all $n \leq m$ where $m \geq 1$, and set $n = m+1$ in the above equation. Then, the right-hand-side of the above equation is easily seen to be such that each term has even degree, so each term of $\delta_{m+1}$ has even degree. Thus, the second statement holds by induction.
\end{proof}
\begin{remark}
It follows from the second statement in Lemma~\ref{lem2} that we need only consider positive $k$ in our search for squares in the sequence $\{\delta_n(k) : n = 1, 2, \ldots\}$.
\end{remark}
\begin{lemma}\label{newlem}
For all even $n \geq 2$, the polynomial $\delta_n$ is not the square of any polynomial (in $\BC[k]$).
\end{lemma}
\begin{proof}
Fix even $n \geq 2$. We claim that $\frac{\low(\delta_n)}{2}$ is odd. Indeed, since we took $n$ to be even, we can write $n = 2m$ for some $m \geq 1$, and so by Lemma~\ref{lem1}, we have that
\begin{eqnarray*}
\frac{\low(\delta_n)}{2} & = & \frac{1}{2}\left(\frac{2^n}{3} - \frac{(-1)^n}{3} + 1\right) \\
& = & \frac{2^{2m-1}}{3} - \frac{1}{6} + \frac{1}{2} \\
& = & \frac{2^{2m-1} + 1}{3} \\
& \equiv & 1 \pmod 2.
\end{eqnarray*}
Now, let $g$ be defined by
$$g(k) = \delta_n(k) \cdot k^{-\low(\delta_n)}.$$
Observe that $g \in \BZ[k]$, and since $\low(\delta_n)$ is even, we have that $\delta_n$ is the square of a polynomial if and only if $g$ is the square of a polynomial. For the sake of contradiction, suppose $g$ is the square of a polynomial $f$. Notice that $\deg f = \frac{\deg g}{2}$ is odd because $\frac{\low(\delta_n)}{2}$ is odd and $\deg \delta_n \equiv 0 \pmod 4$. Let $m$ be the largest nonnegative even integer smaller than $\deg f$ such that the coefficient of $k^m$ in the polynomial $f$ is nonzero (such an $m$ exists because the constant term of $g$ is nonzero, so the constant term of $f$ must also be nonzero), and take this coefficient to be $b$. If $b \neq 0$, then the polynomial $g$ has a nonzero term of odd degree. Since $\low(\delta_n)$ is even, it follows that $\delta_n$ has a nonzero term of odd degree, which contradicts the result of Lemma~\ref{lem2}. It follows that $\delta_n$ is not the square of any polynomial.
\end{proof}
We have now provided a partial characterization of the polynomials $\delta_n$ and $\varepsilon_n$. The following result is a more general statement about polynomials and is not specific to the case of $\delta_n$.
\begin{proposition}\label{newprop}
Let $f \in \BZ[k]$ be a polynomial of positive even degree whose leading coefficient is a square. If $f(k_0)$ is a square for infinitely many integers $k_0$, then $f$ is the square of a polynomial.\footnote{The proof of this proposition was obtained by generalizing an argument presented in~\cite{noam}.}
\end{proposition}
\begin{proof}
Write $f(k) = a^2k^{2n} + \sum_{i = 0}^{2n-1} a_ik^i$, where $a \neq 0$. If we take $g(k) = \sqrt{f(k)}$, then the Taylor expansion of $g(k)$ about the point $k = \infty$ can be expressed as $g(k) = ak^n + \sum_{i = 0}^{n-1} b_ix^i + h(k)$, where the coefficients $b_0, \dots, b_{n-1}$ are rational numbers and where $h(k) = O(1/k)$. Notice that $\lim_{k \to \pm\infty} h(k) = 0$, and suppose for the sake of contradiction that $h$ is not identically $0$. Then for every $\varepsilon > 0$, there exists $M$ such that $0 < |h(k)| < \varepsilon$ whenever $|k| > M$. By taking $\varepsilon$ sufficiently small, we can ensure that $[(b_0 - \varepsilon, b_0) \cup (b_0,b_0 + \varepsilon)] \cap \BZ = \varnothing$, which implies that $b_0 + h(k) \not\in \BZ$ whenever $|k| > M$. Thus, by assumption, there exists an integer $k_0$ so large in absolute value that the following conditions hold: (1) $g(k_0) - b_0 - h(k_0) \in \BZ$, (2) $b_0 + h(k_0) \not\in \BZ$, and (3) $g(k_0) \in \BZ$. These three conditions are clearly contradictory, so we must have that $h$ is identically $0$. It follows that $g$ is a polynomial and that $f(k) = g(k)^2$, so $f$ is indeed the square of a polynomial.
\end{proof}
Combining Proposition~\ref{newprop} with the description of the polynomials $\delta_n$ provided in Lemmas~\ref{lem1},~\ref{lem2}, and~\ref{newlem} yields the following result, which provides evidence toward the conjecture that $\delta_n(k_0)$ is not a square for all integers $n \geq 2$ and $k_0 \neq 0$.
\begin{theorem}\label{thm4}
Fix $n \geq 2$. Then $\delta_n(k_0)$ is a square for at most finitely many nonzero integers $k_0$.
\end{theorem}
\begin{proof}
By Theorem~\ref{thm3}, we may restrict to the case of $n$ even. In this case, Lemma~\ref{newlem} tells us that $\delta_n$ is not the square of a polynomial. By Lemma~\ref{lem1}, $\deg \delta_n$ is positive and even, and by Lemma~\ref{lem2}, the leading coefficient of $\delta_n$ is a square. The theorem now follows immediately from Proposition~\ref{newprop}.
\end{proof}
\begin{remark}
The result of Theorem~\ref{thm4} forms a companion to a result presented on p.~25 of~\cite{JM}, where it is shown that for fixed $k_0$, $\delta_n(k_0)$ is not a square for all but finitely many $n$.
\end{remark}
\begin{comment}
Another way to show that a number $m \in \BN$ is not a square is to find $m' \in \BN$ such that $m'^2 < m < (m' + 1)^2$. This idea motivates the following analysis, in which we attempt to find polynomials $q \in \BZ[k]$ such that for a fixed integer $n \geq 2$, we have $q(k)^2 < p(k) < \big(q(k) + 1\big)^2$. The next result is a more general statement about polynomials and is not specific to the case of $\delta_n$.
\begin{proposition}\label{prop2}
Let $p \in \BZ[k]$ be a polynomial of degree $2^n$ whose coefficients are nonnegative and whose leading coefficient is a square. Further suppose that the degree of each term in $p$ is even. We write $p(k)$ as
$$p(k) = a_{2^n} k^{2^n} + a_{2^n - 2} k^{2^n - 2} + \dots + a_2k^2 + a_0.$$
Then there exists a unique polynomial $q \in \BQ[k]$ that satisfies the following properties:
\begin{enumerate}
\item The degree of $q$ is $2^{n-1}$, and the degree of each term in $q(k)$ is even.
\item The leading coefficient of $q$ is positive.
\item If $ck^m$ is a term of $q(k)^2$ such that $2^{n-1} < m \leq 2^n$, then $c = a_m$.\footnote{Alternatively, we may require that $p(k) - q(k)^2$ is of degree at most $2^{n-1}$.}
\item The constant term of $q$ is an integer.
\item For sufficiently large $k_0$ we have $q(k_0)^2 \leq p(k_0) < \big(q(k_0) + 1\big)^2$.
\end{enumerate}
\end{proposition}
\begin{proof}
We will first compute a polynomial $\wt{q}(k)$, from which we will obtain $q(k)$. Suppose that $\wt{q}(k)$ satisfies conditions (1) and (2) above \emph{and} that $\wt{q}(k)$ satisfies the following modified version of condition (3): If $ck^m$ is a term of $\wt{q}(k)^2$ such that $2^{n-1} \leq m \leq 2^n$, then $c = a_m$. Then, we can express $\wt{q}(k)$ as $$\wt{q}(k) = b_{2^{n-1}} k^{2^{n-1}} + b_{2^{n-1} - 2} k^{2^{n-1} - 2} + \dots + b_2k^2 + b_0.$$
By squaring this expression of $\wt{q}(k)$, we find that $b_{2^{n-1}} = \sqrt{a_{2^n}} \in \BZ_{>0}$. Condition (3) above implies that $$a_{2^n - 2} = 2b_{2^{n - 1}}b_{2^{n-1} - 2},$$
and since $b_{2^{n-1}} > 0$, we obtain a unique rational solution for $b_{2^{n-1} - 2}$. Similarly, condition (3) implies that we have the following equality for each $m \geq 1$:
$$a_{2^n - 2m} = 2b_{2^{n-1}}b_{2^{n-1}-2m} + [\text{polyn. with int. coeffs. in } b_{2^{n-1}}, \dots, b_{2^{n-1} - 2m + 2}],$$
which is a linear equation in $b_{2^{n-1} - 2m}$. By inducting on $m$, we find that $b_{2^{n-1} - 2m}$ has a unique rational solution for each $m$. We have thus uniquely determined the coefficients $b_0, \dots, b_{2^{n-1}}$ of $\wt{q}(k)$, and so we have uniquely determined $\wt{q}(k)$.
Suppose $r(k) = p(k) - \wt{q}(k)^2$ is such that its leading coefficient is nonnegative. Then take $q(k) = \wt{q}(k) - (b_0 - \lfloor b_0\rfloor)$. Notice that the constant term of $q(k)$ is an integer, so condition (4) is satisfied. Clearly we have the inequality $q(k_0)^2 \leq p(k_0)$ for sufficiently large $k_0$. Moreover, $\big(q(k) + 1\big)^2 - p(k)$ is a polynomial of degree $2^{n-1}$ with positive leading coefficient, so for sufficiently large $k_0$, we have the other inequality $p(k_0) < \big(q(k_0) + 1\big)^2$. On the other hand, suppose $r(k)$ is such that its leading coefficient is negative. Then take $q(k) = \wt{q}(k) - (b_0 - \lceil b_0 - 1\rceil)$. Clearly we have the inequality $\big(q(k_0) + 1\big)^2 \geq \wt{q}(k_0)^2 > p(k_0)$ for sufficiently large $k_0$. Moreover, $q(k)^2 - p(k)$ is a polynomial of degree $2^{n-1}$ with negative leading coefficient, so for sufficiently large $k_0$, we have the other inequality $q(k_0)^2 < p(k_0)$.
We now establish uniqueness of $q(k)$. Condition (3) above uniquely determines all but the constant term of $q(k)$. And if $q(k)$ had a different constant term (which condition (4) stipulates must be an integer), condition (5) would fail to hold. Thus, $q(k)$ is uniquely determined by the conditions in the proposition.
\end{proof}
By Lemmas~\ref{lem1} and~\ref{lem2}, we have that $\delta_n$ satisfies the conditions imposed upon the polynomial $p$ in the statement of Proposition~\ref{prop2} when $n \geq 2$. Fixing $n \geq 2$ and taking $p = \delta_n$, we obtain a unique polynomial $q$ that satisfies the properties given in Proposition~\ref{prop2}. In this case, the coefficients of $q$ may not be integers, although computing $q$ for $n = 2, \dots, 10$ suggests that $q$ may indeed have integer coefficients for all $n$. The method of ``sandwiching'' $\delta_n(k)$ between consecutive squares can be used to obtain the following result.
\begin{theorem}\label{thm12}
Fix $n \geq 2$. Let $S$ denote the set of nonzero integers such that $q(S) \subset \BZ$ (observe that $S$ is an infinite set with positive density in $\BZ$). Then there exist at most finitely many $k \in S$ such that $\delta_n(k)$ is a square.
\end{theorem}
\begin{proof}
By Theorem~\ref{thm3}, we may restrict to the case of $n$ even. We claim that $\frac{\low(\delta_n)}{2}$ is odd. Indeed, since we took $n$ to be even, we can write $n = 2m$ for some $m \geq 1$, and so by Lemma~\ref{lem1}, we have that
\begin{eqnarray*}
\frac{\low(\delta_n)}{2} & = & \frac{1}{2}\left(\frac{2^n}{3} - \frac{(-1)^n}{3} + 1\right) \\
& = & \frac{2^{2m-1}}{3} - \frac{1}{6} + \frac{1}{2} \\
& = & \frac{2^{2m-1} + 1}{3} \\
& \equiv & 1 \pmod 2.
\end{eqnarray*}
Let $k_0$ be so large so that conditions listed in the statement of Proposition~\ref{prop2} hold, and notice that $q(k_0) \in \BZ$, and assume temporarily that $\delta_n(k)$ is a square for more than $2^{n-1}$ values of $k_0$. By interpolating, it follows from condition (5) that $q(k)^2 = \delta_n(k)$, where this equality is an equality of \emph{polynomials}. So, $\frac{\low(\delta_n)}{2} = \low(q)$, and by condition (1), we have that $\low(q)$ is even. We now have contradictory results regarding the parity of $\frac{\low(\delta_n)}{2}$, so our temporary assumption is false and $\delta_n(k_0)$ is not a square for all but finitely many $k_0$.
\end{proof}
In Theorem~\ref{thm12}, the set $S$ contains numbers that are divisible by some large power of 2. Thus, Theorem~\ref{thm12} provides evidence that $\delta_n(k_0)$ is not a square even when $k_0$ is even (which Theorem~\ref{thm6} was not able to do).
\end{comment}
\section*{Open Problems}
\noindent In the case where the sequence $\{a_n\}$ defined by $a_0 = 0$ and $a_n = \phi(a_{n-1})$ is periodic and $d > 2$, the Galois theoretic properties of iterates of $\phi$ have not yet been studied. Moreover, it remains to be studied for a wider selection of rational functions $\phi$ whether the index $[S(\phi) : G(\phi)]$ is finite. In the case where $\phi$ commutes with a nontrivial M\"{o}bius transformation that fixes $0$, Conjecture~\ref{conj1} remains open; i.e. it is yet to be determined whether $[C(\phi) : G(\phi)]$ is finite for all rational functions $\phi$ of the form $\phi(x) = \frac{k_0(x^2 + 1)}{x}$, and the situation where $d > 2$ has not yet been studied. It would also be interesting to investigate whether one can use standard conjectures, such as the $abc$ conjecture or conjectures on bounds for heights of rational points on curves, to prove that $\delta_n(k_0)$ is not a square. Finally, a number of additional questions and conjectures relating to arboreal Galois representations are posed in~\cite{jone}.
\section*{Acknowledgments}
\noindent This research was supervised by Joe Gallian at the University of Minnesota Duluth REU and is
supported by the National Science Foundation (grant number DMS-1062709) and the
National Security Agency (grant number H98230-11-1-0224). I would like to thank Joe
Gallian for his advising and support. I would also like to thank Rafe Jones for introducing me to the field of arithmetic dynamics, explaining his work, and providing numerous helpful comments on my research. I would like to thank Noam Elkies for many fruitful discussions on polynomials and squarefree sequences, and for directing me to the source~\cite{noam}. I would finally like to thank Adam Hesterberg, Noah Arbesfeld, Simon Rubinstein-Salzedo, and Daniel Kane for helpful discussions and
valuable suggestions on the paper.
|
\section{\label{sec:level1}Introduction\protect\\}
Tremendous efforts have been devoted to the production of ultracold samples of heteronuclear molecules over recent years.
Promising applications range from quantum computation \cite{DeMilleQC} to ultracold chemistry \cite{JunYeUltracoldChemistry}, quantum dipolar physics \cite{Krems2009,JunYe2013}, and tests of fundamental physics \cite{Krems2009,DeMilleReview2009}.
However, access to dense samples of ground state ultracold polar molecules has so far been limited to a few outstanding experiments \cite{JunYeSTIRAP,InnsbruckRbCs2014}, which utilized magneto-association followed by transfer to the ground state by stimulated Raman adiabatic passage \cite{RevModPhys.70.1003}.
A less experimentally complex path toward a sample of ultracold molecules is short-range photoassociation (PA) \cite{Weidemuller2008}.
Recent discoveries of short-range PA transitions in RbCs \cite{Gabbanini2011}, NaCs \cite{PhysRevA.84.061401}, KRb \cite{PhysRevA.86.053428} and Rb$_2$ \cite{PhysRevA.87.053404,Bellos2011} proved the general applicability of the process to alkali dimers.
Short-range PA is a convenient means to produce ultracold molecules due to its simplicity and possible continuous operation. Although PA leads to an unavoidable distribution of molecules over many vibrational, rotational, and hyperfine levels, it has been argued that simple measures may allow removal of excited states following PA \cite{ERHudson2008,Bruzewicz14}.
However, to date short-range PA has predominantly produced molecules in rotationally excited states: $\sim 2$\% of molecules were observed in $J=0$ for LiCs \cite{Weidemuller2008} and no $J=0$ molecules were observed for NaCs \cite{PhysRevA.84.061401}.
In this paper, we perform high-resolution depletion spectroscopy \cite{WangPRA2007,Weidemuller2008} to measure the distribution of rotational levels in RbCs molecules produced via short-range PA to the $2^3\Pi_0$ electronic state.
We confirm that a large fraction of these $X(v=0)$ molecules, up to 33 \%, are in their rovibronic ground states, i.e., $X$($v=0$, $J=0$).
We also show that the formation pathway to $X(v=0)$ is a two-photon cascade decay as shown in Fig. \ref{fig:formationdetectionpathway}(a),
as opposed to a direct one-photon decay as was previously supposed \cite{Bruzewicz14}.
\begin{figure*}[htp]
\includegraphics[scale=0.32]{formationdetectionpathways.eps}
\caption{\label{fig:formationdetectionpathway} (Color online) (a) Short-range photoassociation followed by two-photon cascade decay to rovibronic ground state molecules.
(b) Detection of vibronic ground state molecules through two-color REMPI. Also shown is the depletion transition used for rotational spectroscopy. Here, a narrow-linewidth laser depopulates given rotational levels from $X(v=0)$ and thus modulates the REMPI signal. Potential energy curves are from Ref. \cite{Allouche2002}.
}
\end{figure*}
\section{\label{sec:level1}Experimental setup\protect\\}
Most of our experimental setup has been previously described \cite{Bruzewicz14}.
$^{85}$Rb and Cs atoms are laser-cooled and trapped in a dual-species forced dark spontaneous force optical trap (dark SPOT) \cite{KetterleDarkSPOT} loaded by alkali dispensers.
The overlap of the two atom clouds is optimized using absorption imaging from two orthogonal directions.
The typical density $n$ and total atom number $N$ for Rb and Cs are $n_{{\rm Rb}}\sim 6 \times 10^{10}$ cm$^{-3}$, $N_{{\rm Rb}} \sim 5 \times 10^6$, $n_{{\rm Cs}} \sim 8 \times 10^{10}$ cm$^{-3}$, and $N_{{\rm Cs}} \sim 1 \times 10^7$.
The translational temperature of the atoms is measured by time-of-flight imaging to be $T \sim 100\ \mu$K.
For the PA transition, we use up to 300 mW of light from a Ti:sapphire laser, focused onto the atom clouds with a beam waist ($1/e^2$ power radius) of 100 $\mu$m.
We use four distinct PA transitions in this work, namely to the first two rotational levels of the $2^3\Pi_{0^+}(v=10)$ state ($J^P=0^+$ and $1^-$ near 11817.1 cm$^{-1}$) \cite{Bruzewicz14} and of the $2^3\Pi_{0^-}(v=11)$ state ($J^P=0^-$ and $1^+$ near 11803.9 cm$^{-1}$) \cite{Gabbanini2013}. Here $J$ corresponds to the total angular momentum excluding nuclear spin, and $P$ to the parity.
We detect molecules using two-color resonance-enhanced multi-photon ionization (1+1 REMPI) as shown in Fig. \ref{fig:formationdetectionpathway}(b).
The first photon from a nanosecond pulse dye laser resonantly excites the vibronic ground state to a given excited state ($X(v=0)$ $\rightarrow $ 2$^1\Pi_1 (v=12)$) at 15342.0 cm$^{-1}$ \cite{Gustavsson1988},
where an intense ($\sim 2$ mJ/pulse) 532 nm pulse ionizes the molecules 10 ns later.
The energy of the resonant pulse is kept below 0.1 mJ to minimize power broadening and off-resonance excitation.
The molecular ions are accelerated by the electric field inside the vacuum chamber ($\sim 100$ V/cm) toward a Channeltron detector, where they are converted into electrical signals. The RbCs$^+$ signals are separated from other atomic and molecular signals based on time-of-flight measurements.
The experiment is repeated at 100 Hz, the repetition rate of the REMPI lasers.
Since the linewidth of the dye laser ($\sim 6$ GHz) is larger than the spacing of rotational levels, it is impossible to address molecules in a single rotational level.
To circumvent this difficulty, we use the technique of high-resolution depletion spectroscopy \cite{WangPRA2007, Weidemuller2008}.
An additional CW diode laser is used to deplete the population of any given rotational level in the $X(v=0)$ manifold of states.
The $2 ^3\Pi_{0^+}(v=8)$ level is chosen as the upper state of this depletion transition,
as all molecular constants for the upper and lower states of this transition are accurately known \cite{JTKim2008,Fellows1999}, making rotational line assignments unambiguous.
Following the excitation, molecules decay predominantly into other vibrational states before the REMPI detection pulse arrives. The molecular ion signal is monitored while the frequency of the depletion laser is scanned; the population in a particular rotational level appears as a decrease in the molecular ion signal as shown in Fig. \ref{fig:depletionspectra}.
To reduce the effect of drifts in the ion signal, we cycle the depletion laser on and off for consecutive detection pulses. We separately average signals with and without the depletion laser after 400 cycles
and take a ratio of these two averages to obtain a normalized depletion signal (Fig. \ref{fig:depletionspectra}).
\begin{figure
\includegraphics[scale=0.4]{depletionspectra.eps
\caption{\label{fig:depletionspectra} (Color online) Depletion spectra with the PA laser tuned to (a) the $|\Omega_{{\rm PA}}=0^+,\ J_{{\rm PA}}^P=0^+ \rangle$ line and (b) the $|\Omega_{{\rm PA}}=0^-,\ J_{{\rm PA}}^P=0^- \rangle$ line.
The normalized depletion signal is taken at 2 MHz frequency steps with an average over 400 pulses and shown after smoothing \cite{Fig2comment}.
Labels indicate the rotational quantum numbers $J_X \rightarrow J'$ of the depletion transition. The sum of depletion depths is over 90\%,
which indicates that the ion signal originating from off-resonant excitation or from excited vibrational levels is small in our REMPI spectroscopy. Shown here are several close-lying $J \rightarrow J' = J+1$ lines; we also observed all associated $J \rightarrow J' = J-1$ lines, at their predicted frequencies.}
\end{figure}
\section{\label{sec:level1}Results\protect\\}
Initially, we tuned the PA laser to the $2 ^3\Pi_{0^+}(v=10)$ state, which we found in \cite{Bruzewicz14} to be one of the most efficient states for producing $X(v=0)$ state molecules.
For PA to the $J^P_{{\rm PA}}=0^+$ level, one-photon decay would lead to the production of exclusively $J^P_X=1^-$ molecules due to the $\Delta J \leq 1,\ J=0 \not\leftrightarrow 0$, and $P_iP_f=-1$ selection rules [where $P_i$ $(P_f)$ is the parity of the initial (final) state of the transition].
However, in a scan of the depletion laser, we observed two clear peaks, which we identified as originating from $J^P_X=0^+$ and $J^P_X=2^+$ molecules in the ground state (Fig. \ref{fig:depletionspectra}(a)).
The absolute positions of, and rotational splittings between, all observed lines agree (within our experimental uncertainties of $\sim$300 MHz and $\sim$30 MHz respectively) with the values predicted from the molecular constants of Ref. \cite{JTKim2008,Fellows1999}.
We obtained a similar result with the PA laser tuned to the $J^P_{{\rm PA}}=1^-$ line;
here we observed two peaks originating from $J^P_X=1^-$ and $J^P_X=3^-$.
These observations suggest that two-photon decays are solely responsible for the population of the $X(v=0)$ state.
Then, we tuned the PA laser to the $2 ^3\Pi_{0^-}$ state observed by Fioretti and Gabbanini \cite{Gabbanini2013}.
As is discussed below, our previous assignment \cite{Bruzewicz14} of this PA series to an $\Omega=0^-$ potential is consistent with our results presented here.
For the $J^P_{{\rm PA}}=0^-$ line, we mainly observed $J^P_X = 1^-$ molecules,
which would be the only state populated if the parity changes in each step of a two-photon cascade.
However, we also observed small peaks corresponding to $J^P_X=0^+$ and $J^P_X=2^+$ as shown in Fig. \ref{fig:depletionspectra}(b).
Similarly, for the $J^P_{{\rm PA}} =1^+$ line, we observed population in all of the rotational levels possible by a two-photon decay but without obeying the usual parity selection rule: $J^P_X=0^+,1^-,2^+$, and $3^-$.
\begin{figure*}[htp]
\includegraphics[scale=0.32]{rotationaldistribution.eps}
\caption{\label{fig:Rotationaldistribution} (Color online) Decay pathways and ground state rotational distribution starting from four distinct PA states.
Molecular states are labeled by their $\Omega$ and $J^P$ quantum numbers.
Black squares are the measured population distribution and colored histograms show calculated distributions
based on H${\rm \ddot{o}}$nl-London factors.
Shaded bars are obtained assuming parity selection rules are followed throughout the decay, i.e. no Stark mixing in intermediate levels, and match the results for the $\Omega=0^+$ PA state.
For the $\Omega_{{\rm PA}} =0^-$ PA state, solid bars are calculated assuming complete Stark mixing of $\Omega$-doublets in the intermediate state.
The measured values fall between the two predictions for the extreme assumptions; hence Stark mixing of the intermediate $\Omega$-doublet states is substantial, but not complete.
}
\end{figure*}
The qualitative difference in the decay of the $\Omega =0^+$ and $\Omega =0^-$ PA states arises from the different intermediate states of the cascade.
Due to the $\Omega=0^+ \not\leftrightarrow \Omega=0^-$ selection rule,
$\Omega=0^-$ states cannot decay directly to $\Omega = 0^+$ states, including the ground $X^1\Sigma^+$ state.
Thus, the only possible decay path for $\Omega = 0^-$ molecules to the $X^1\Sigma^+$ state is a two-step decay via intermediate states with $\Omega =1$.
These intermediate $\Omega = 1$ states have $\Omega$-doublet substructure, such that for each $J$ there is a pair of nearly-degenerate levels of opposite parity. In our experiment, these levels are likely to be mixed due to the electric field used to extract ions for detection. Hence, for decays via such $\Omega =1$ intermediate states, the usual parity selection rules are no longer valid under our experimental conditions.
By contrast, the $\Omega = 0^+$ excited state can decay to the $X$ state via a cascade with intermediate states that have either $\Omega = 0^+$ or $\Omega = 1$. Our observations indicate that the overwhelmingly dominant pathway is through an $\Omega = 0^+$ intermediate state.
We summarize the possible decay paths in Fig. \ref{fig:Rotationaldistribution},
along with the distribution of the rotational levels obtained by analyzing depletion spectra.
We renormalize the sum of the experimental population fractions to 1 to simplify the comparison with theoretical predictions.
The observed rotational distributions are consistent with predictions from H$\ddot{{\rm o}}$nl-London factors \cite{HonlLondon1974} for decays of the $\Omega = 0^+$ PA state, assuming that each molecule decays only through an $\Omega = 0^+$ intermediate state. However, for $\Omega=0^-$ PA states, the unknown degree of parity mixing in the intermediate $\Omega=1$ states due to the Stark effect makes a quantitative prediction of rotational distributions difficult.
Instead, we calculate rotational distributions for two limiting cases: (1) assuming that the parity selection rule is exactly satisfied throughout the decay (the case of no Stark mixing) and (2) assuming that the two intermediate states are fully mixed, resulting in maximal population of otherwise parity-forbidden levels. The experimental values for the rotational distributions fall between these two limiting cases as shown in Fig. \ref{fig:Rotationaldistribution} (c) and (d).
Similar two-photon cascade decays have been observed in Cs$_2$ \cite{PhysRevA.79.021402,PhysRevA.85.030502}, where some of the one-photon direct decays are strictly forbidden by the ungerade-gerade selection rule applicable to homonuclear diatomic molecules.
In RbCs, the intercombination transitions from the $2 ^3\Pi_{0^+}$ state (a triplet state) to the $X$ state (a singlet state) should be allowed due to singlet-triplet mixing by the spin-orbit interaction.
However, we did not observe any one-photon decay to $J^P_X=1^-$ molecules for the $|\Omega_{{\rm PA}}=0^+,
J^P_{{\rm PA}}=0^+ \rangle$ line (Fig. \ref{fig:Rotationaldistribution}(a)), which indicates that the one-photon decay path is much less probable than the two-photon cascade.
We suspect that the two-photon cascade for $\Omega_{{\rm PA}} = 0^+$ occurs via states in the $A^1\Sigma^+[0+] / b^3\Pi[0+]$ complex,
where singlet-triplet mixing could be very efficient due to the avoided crossing between these two potential curves \cite{PhysRevA.81.042511}.
However, we do not have a direct way to confirm this hypothesis.
To independently verify that the one-photon spontaneous emission probability is small, we determine the transition dipole matrix element $\langle \mu \rangle$ between the $X$ and $2 ^3 \Pi_{0^+}$ states by measuring the saturation behavior of the depletion transition as shown in Fig. \ref{fig:SaturationBehavior}.
This behavior can be understood by modeling the depletion transition as an open two-level system.
Since the decay rate from the upper state is much higher than the excitation rate by the depletion laser, we can ignore coherent effects such as Rabi oscillations. We also assume a single average interaction time $\tau \sim 10$ ms, consistent with the transit time through the depletion laser beam determined by thermal velocities or by free fall in gravity.
In this model, the depletion lineshape $L$ is given by
\begin{equation}
L(\Delta, \Omega_R ) =A \left\{ 1- \exp \left( -\frac{(\frac{\gamma}{2})^2}{\Delta^2 + (\frac{\gamma}{2})^2} \frac{\Omega_{R}^2}{\gamma} \tau \right) \right\},
\end{equation}
where $\Delta$ is the detuning of the depletion laser from resonance, $\Omega_{R} = E \cdot \langle \mu \rangle \chi / \hbar$ is the Rabi frequency for a transition dipole moment $\langle \mu \rangle$ driven by the electric field amplitude $E$ of the depletion laser, $\chi \approx 0.12$ is the factor arising from the nuclear wavefunctions (both the Franck-Condon factor and the angular factor),
and $\gamma$ is the natural linewidth of the transition.
With this model, we analyzed the power dependence of the depletion depth and the linewidth of the transition as shown in Fig. \ref{fig:SaturationBehavior}. By fitting the data with Eq. (1), we obtain $\langle \mu \rangle \approx 2 \times 10^{-2} \ e a_0$.
We estimate this value to have an overall uncertainty of a factor of $\sim 2$, dominated by the uncertainty in our estimate of the interaction time.
In Ref. \cite{Bruzewicz14}, a crude argument led to an estimated value of this matrix element of $0.3 \ ea_0$.
Our result suggests that the one-photon direct decay is less efficient than we previously expected in \cite{Bruzewicz14}.
This is consistent with our non-observation of any single-photon decay.
Finally, in \cite{Bruzewicz14}, we estimated the production rate of $X(v=0)$ molecules to be $\sim 6\times 10^3$ molecules/s starting from atoms in a dark SPOT. By considering the branching ratio (1/3) for the newly identified pathway, we conclude that the production rate of $X(v=0,J=0)$ molecules via the $|\Omega=0^+, J^P_{{\rm PA}} = 0 \rangle$ PA line is $\sim 2 \times 10^3$ molecules/s. This is the highest production rate of molecules in the rovibronic ground state via PA to date.
\begin{figure
\includegraphics[scale=0.4]{SaturationBehavior.eps
\caption{\label{fig:SaturationBehavior} (Color online) Power dependence of the depletion signal depth and linewidth of the depletion feature for the $J_X=2 \rightarrow J' = 3$ transition.
Data points and error bars are extracted from curve fits of corresponding depletion scans.
Lines are fits to Eqn.(1) with $\langle \mu \rangle$ as a free parameter.
The best fit values are $ \langle \mu \rangle = 1.7 \times 10^{-2} \ e a_0 \ ( 1.5 \times 10^{-2} \ e a_0)$ for the depletion depth (depletion linewidth) data. The deviation of the data points of the depletion depth from the fit curve at high depletion laser intensities likely arises from the imperfect assumption of a single interaction time.}
\end{figure}
\section{\label{sec:level1}Conclusion\protect\\}
We performed depletion spectroscopy to resolve the rotational level distribution of RbCs molecules produced in the $X(v=0)$ state following short-range PA to the $2 ^3 \Pi_{0^+}$ and $2 ^3 \Pi_{0^-}$ states.
In contrast to our expectations (as discussed in Ref. \cite{Bruzewicz14}), we found that the $X$ state population arises from a two-photon cascade decay in both cases, rather than a direct one-photon decay from the $2^3\Pi_{0^+}$ state.
We also confirmed that PA to the $2 ^3 \Pi_{0^+}, J^P_{{\rm PA}}=0$ lines produce a substantial fraction ($\approx$33\%) of $X(v=0)$ state molecules in the rotational ground state at a rate of $\sim 2 \times 10^3$ molecules/s.
This represents the strongest demonstrated pathway to the production of rovibronic ground state molecules via PA to date.
This demonstration is a starting point for accumulating ultracold molecules in the rovibronic ground state by PA in an optical trap.
There, the higher densities and lower temperatures achievable should allow substantially higher production rates; moreover, collisions with co-trapped atoms should remove metastable excited states. Hence our results point to the possibility of a simple procedure for producing a pure trapped sample of rovibronic ground state molecules.
\begin{acknowledgments}
This work was supported by DOE, AFSOR MURI, and ARO MURI.
\end{acknowledgments}
|
\section{Introduction}
Cosmological simulations have greatly improved our understanding of the physics of galaxy formation and are widely used to guide the interpretation of observations and the design of new observational campaigns and instruments. Simulations enable astronomers to ``turn the knobs'' much as experimental physicists are able to in the laboratory. While such numerical experiments can be valuable even if the simulations fail to reproduce observations, in general our confidence in the conclusions drawn from simulations, and the number of applications they can be used for, increases with the level of agreement between the best-fit model and the observations.
For many years the overall agreement between hydrodynamical simulations and observations of galaxies was poor. Most simulations produced galaxy mass functions with the wrong shape and normalisation, the galaxies were too massive and too compact, and the stars formed too early. Star formation in high-mass galaxies was not quenched and the models could not simultaneously reproduce the stellar masses and the thermodynamic properties of the gas in groups and clusters (e.g.\ \citealt{Scannapieco2012Aquila} and references therein).
Driven in part by the failure of hydrodynamical simulations to reproduce key observations, semi-analytic and halo-based models have become the tools of choice for detailed comparisons between galaxy surveys and theory (see \citealt{Baugh2006SAMReview} and \citealt{Cooray2002HaloModelReview} for reviews). Thanks to their flexibility and relatively modest computational expense, these approaches have proven valuable for many purposes. Examples include the interpretation of observations of galaxies within the context of the cold dark matter framework, relating galaxy populations at different redshifts, the creation of mock galaxy catalogues to investigate selection effects or to translate measurements of galaxy clustering into information concerning the occupation of dark matter haloes by galaxies.
However, hydrodynamical simulations have a number of important advantages over these other approaches. The risk that a poor or invalid approximation may lead to over-confidence in an extrapolation, interpretation or application of the model is potentially smaller, because they do not need to make as many simplifying assumptions. Although the subgrid models employed by current hydrodynamical simulations often resemble the ingredients of semi-analytic models, there are important parts of the problem for which subgrid models are no longer required. Since hydrodynamical simulations evolve the dark matter and baryonic components self-consistently, they automatically include the back-reaction of the baryons on the collisionless matter, both inside and outside of haloes. The higher resolution description of the baryonic component provided by hydrodynamical simulations also enables one to ask more detailed questions and to compare with many more observables. Cosmological hydrodynamical simulations can be used to model galaxies and the intergalactic medium (IGM) simultaneously, including the interface between the two, which may well be critical to understanding the fuelling and feedback cycles of galaxies.
The agreement between hydrodynamical simulations of galaxy formation and observations has improved significantly in recent years. Simulations of the diffuse IGM already broadly reproduced quasar absorption line observations of the \lya\ forest two decades ago \citep[e.g.][]{Cen1994LyaForest,Zhang1995LyaForest,Hernquist1996LyaForest,Theuns1998LyaForest,Dave1999LyaForest}. The agreement is sufficiently good that comparisons between theory and observation can be used to measure cosmological and physical parameters \citep[e.g.][]{Croft1998LyaPowerSpectrum,Schaye2000IGMTemp,Viel2004LyaPowerSpectrum,McDonald2005LyaPowerSpectrum}. More recently, simulations that have been re-processed using radiative transfer of ionizing radiation have succeeded in matching key properties of the high-column density \HI\ absorbers \citep[e.g.][]{Pontzen2008DLAs,Altay2011HICDDF,McQuinn2011LLS,Rahmati2013HICDDF}.
Reproducing observations of galaxies and the gas in clusters of galaxies has proven to be more difficult than matching observations of the low-density IGM, but several groups have now independently succeeded in producing disc galaxies with more realistic sizes and masses
(e.g.\ \citealt{Governato2004Disk,Governato2010Disk,Okamoto2005Disks,Agertz2011Disk,Guedes2011Eris,McCarthy2012RotSize,Brook2012MagicDisks,Stinson2013MAGICC,Munshi2013StellarHaloMass,Aumer2013Disks,Hopkins2013FIRE,Vogelsberger2013IllustrisModel,Vogelsberger2014Illustris,Marinacci2014Disks}).
For the thermodynamic properties of groups and clusters of galaxies the progress has also been rapid \citep[e.g.][]{Puchwein2008AGN,McCarthy2010AGN,Fabjan2010AGN,LeBrun2014CosmoOWLS}. The improvement in the realism of the simulated galaxies has been accompanied by better agreement between simulations and observations of the metals in circumgalactic and intergalactic gas \citep[e.g.][]{Stinson2012CGM,Oppenheimer2012IGMMetals}, which suggests that a more appropriate description of galactic winds may have been responsible for much of the progress.
Indeed, the key to the increase in the realism of the simulated galaxies has been the use of subgrid models for feedback from star formation that are more effective in generating galactic winds and, at the high-mass end, the inclusion of subgrid models for feedback from active galactic nuclei (AGN). The improvement in the resolution afforded by increases in computing power and code efficiency has also been important, but perhaps mostly because higher resolution has helped to make the implemented feedback more efficient by reducing spurious, numerical radiative losses. Improvements in the numerical techniques to solve the hydrodynamics have also been made \citep[e.g.][]{Price:2008kx,Springel2010Arepo,Read:2010uq,Saitoh:2013uq,Hopkins:2013lr} and may even be critical for particular applications \citep[e.g.][]{Agertz:2007fk,Bauer2012SubsonicTurbulence}, but overall their effect appears to be small compared to reasonable variations in subgrid models for feedback processes \citep[][]{Scannapieco2012Aquila}.
Here we present the EAGLE\ project\footnote{EAGLE\ is a project of the Virgo consortium for cosmological supercomputer simulations.}, which stands for Evolution and Assembly of GaLaxies and their Environments. EAGLE\ consists of a suite of cosmological, hydrodynamical simulations of a standard $\Lambda$CDM universe.
The main models were run in volumes of 25 to 100 comoving Mpc (cMpc) on a side and employ a resolution that is sufficient to marginally resolve the Jeans scales in the warm ($T\sim 10^4\,{\rm K}$) interstellar medium (ISM). The simulations use state-of-the-art numerical techniques and subgrid models for radiative cooling, star formation, stellar mass loss and metal enrichment, energy feedback from star formation, gas accretion onto, and mergers of, supermassive black holes (BHs), and AGN feedback. The efficiency of the stellar feedback and the BH accretion were calibrated to broadly match the observed $z\sim 0$ galaxy stellar mass function (GSMF) subject to the constraint that the galaxy sizes must also be reasonable, while the efficiency of the AGN feedback was calibrated to the observed relation between stellar mass and BH mass. The goal was to reproduce these observables using, in our opinion, simpler and more natural prescriptions for feedback than used in previous work with similar objectives.
By ``simpler'' and ``more natural'', which are obviously subjective terms, we mean the following. Apart from stellar mass loss, we employ only one type of stellar feedback, which captures the collective effects of processes such as stellar winds, radiation pressure on dust grains, and supernovae. These and other feedback mechanisms are often implemented individually, but we believe they cannot be properly distinguished at the resolution of $10^2$--$10^3\,{\rm pc}$ that is currently typical for simulations that sample a representative volume of the universe. Similarly, we employ only one type of AGN feedback (as opposed to e.g.\ both a ``radio'' and ``quasar'' mode). Contrary to most previous work, stellar (and AGN) feedback is injected in thermal form without turning off radiative cooling and without turning off hydrodynamical forces. Hence, galactic winds are generated without specifying a wind direction, velocity, mass loading factor, or metal mass loading factor. We also do not need to boost the BH Bondi-Hoyle accretion rates by an ad-hoc factor. Finally, the amount of feedback energy (and momentum) that is injected per unit stellar mass depends on local gas properties rather than on non-local or non-baryonic properties such as the dark matter velocity dispersion or halo mass.
The EAGLE\ suite includes many simulations that will be presented elsewhere. It includes higher-resolution simulations that zoom into individual galaxies or galaxy groups \citep[e.g.][]{Sawala2014EagleZooms}. It also includes variations in the numerical techniques \citep{Schaller2014EagleSPH} and in the subgrid models \citep{Crain2014EagleModels} that can be used to test the robustness of the predictions and to isolate the effects of individual processes.
This paper is organised as follows. We begin in \S\ref{sec:general} with a discussion of the use and pitfalls of cosmological hydrodynamical simulations in light of the critical role played by subgrid processes. We focus in particular on the implications for the interpretation and the predictive power of the simulations, and the role of numerical convergence. In \S\ref{sec:simulations} we describe the simulations and our definition of a galaxy. This section also briefly discusses the numerical techniques and subgrid physics. The subgrid models are discussed in depth in \S\ref{sec:subgrid}; readers not interested in the details may wish to skip this section. In \S\ref{sec:cal_obs} we show the results for observables that were considered in the calibration of the subgrid models, namely the $z\sim 0$ GSMF, the related relation between stellar mass and halo mass, galaxy sizes, and the relations between BH mass and stellar mass. We also consider the importance of the choice of aperture used to measure stellar masses and investigate both weak and strong convergence (terms that are defined in \S\ref{sec:general}). In \S\ref{sec:otherobs} we present a diverse and representative set of predictions that were not used for the calibration, including specific star formation rates and passive fractions, the Tully-Fisher relation, the mass-metallicity relations, various properties of the intracluster medium, and the column density distributions of intergalactic metals. All results presented here are for $z\sim 0$. We defer an investigation of the evolution to \citet{Furlong2014EagleEvolution} and other future papers. We summarize and discuss our conclusions in \S\ref{sec:summary}. Finally, our implementation of the hydrodynamics and our method for generating the initial conditions are summarized in Appendices~\ref{app:hydro} and \ref{app:ics}, respectively.
\section{Implications of the critical role of subgrid models for feedback}
\label{sec:general}
In this section we discuss what, in our view, the consequences of our reliance on subgrid models for feedback are for the predictive power of the simulations (\S\ref{sec:need}) and for the role of numerical convergence (\S\ref{sec:conv_discussion}).
\subsection{The need for calibration}
\label{sec:need}
Because the recent improvement in the match between simulated and observed galaxies can, for the most part, be attributed to the implementation of more effective subgrid models for feedback, the success of the hydrodynamical simulations is subject to two important caveats that are more commonly associated with semi-analytic models.
First, while it is clear that effective feedback is required, the simulations can only provide limited insight into the nature and source of the feedback processes. For example, suppose that the implemented subgrid model for supernovae is too inefficient because, for numerical reasons, too much of the energy is radiated away, too much of the momentum cancels out, or the energy/momentum are coupled to the gas at the wrong scale. If we were unaware of such numerical problems, then we might erroneously conclude that additional feedback processes such as radiation pressure are required. The converse is, of course, also possible: the implemented feedback can also be too efficient, for example because the subgrid model underestimates the actual radiative losses. The risk of misinterpretation is real, because it can be shown that many simulations underestimate the effectiveness of feedback due to excessive radiative losses \citep[e.g.][]{DallaVecchia2012Winds}, which themselves are caused by a lack of resolution and insufficiently realistic modelling of the ISM.
Second, the \emph{ab initio} predictive power of the simulations is currently limited when it comes to the properties of galaxies. If the efficiency of the feedback processes depends on subgrid prescriptions that may not be good approximations to the outcome of unresolved processes, or if the outcome depends on resolution, then the true efficiencies cannot be predicted from first principles. Note that the use of subgrid models does not in itself remove predictive power. If the physical processes that operate below the resolution limit and their connection with the physical conditions on larger scales are fully understood and can be modelled or observed, then it may be possible to create a subgrid model that is sufficiently realistic to retain full predictive power. However, this is currently not the case for feedback from star formation and AGN. As we shall explain below, this implies that simulations that appeal to a subgrid prescription for the generation of outflows are unable to predict the stellar masses of galaxies. Similarly, for galaxies whose evolution is controlled by AGN feedback, such simulations cannot predict the masses of their central BHs.
To illustrate this, it is helpful to consider a simple model. Let us assume that galaxy evolution is self-regulated, in the sense that galaxies tend to evolve towards a quasi-equilibrium state in which the gas outflow rate balances the difference between the gas inflow rate and the rate at which gas is locked up in stars and BHs. The mean rate of inflow (e.g.\ in the form of cold streams) evolves with redshift and tracks the accretion rate of dark matter onto haloes, which is determined by the cosmological initial conditions. For simplicity, let us further assume that the outflow rate is large compared to the rate at which the gas is locked up. Although our conclusions do not depend on the validity of this last assumption, it simplifies the arguments because it implies that the outflow rate balances the inflow rate, when averaged
over appropriate length and time scales. Note that the observed low
efficiency of galaxy formation (see Fig.~\ref{fig:eta} in \S\ref{sec:eta}) suggests that this may actually be a reasonable approximation, particularly for low-mass galaxies.
This toy model is obviously incorrect in detail. For example, it ignores the re-accretion of matter ejected by winds, the recycling of stellar mass loss, and the interaction of outflows and inflows. However, recent numerical experiments and analytic models provide some support for the general idea \citep[e.g.][]{Finlator2008MZ,Schaye2010OWLS,Booth2010DMHaloesBHs,Dave2012EquilModel,Haas2013OwlsI,Haas2013OwlsII,Feldmann2013SFLaw,Dekel2013ToyModel,Altay2013DLAs,Lilly2013EquilModel,Sanchez2014Review}. This idea in itself is certainly not new and follows from the existence of a feedback loop \citep[e.g.][]{White1991GF}, as can be seen as follows. If the inflow rate exceeds the outflow rate, then the gas fraction will increase and this will in turn increase the star formation rate (and/or, on a smaller scale, the BH accretion rate) and hence also the outflow rate. If, on the other hand, the outflow rate exceeds the inflow rate, then the gas fraction will decrease and this will in turn decrease the star formation rate (and/or the BH accretion rate) and hence also the outflow rate.
In this self-regulated picture of galaxy evolution the outflow rate is determined by the inflow rate. Hence, the outflow rate is \emph{not} determined by the efficiency of the implemented feedback. Therefore, if the outflow is driven by feedback from star formation, then the star formation rate will adjust until the outflow rate balances the inflow rate, irrespective of the (nonzero) feedback efficiency. However, the star formation rate for which this balance is achieved, and hence also ultimately the stellar mass, do depend on the efficiency of the implemented feedback. If the true feedback efficiency cannot be predicted, then neither can the stellar mass. Similarly, if the outflow rate is driven by AGN feedback, then the BH accretion rate will adjust until the outflow rate balances the inflow rate (again averaged over appropriate length and time scales). The BH accretion rate, and hence the BH mass, for which this balance is achieved depend on the efficiency of the implemented feedback, which has to be assumed. According to this toy model, which appears to be a reasonable description of the evolution of simulated galaxies, the stellar and BH masses are thus determined by the efficiencies of the (subgrid) implementations for stellar and AGN feedback, respectively.
The simulations therefore need to be calibrated to produce the correct stellar and BH masses. Moreover, if the true efficiency varies systematically with the physical conditions on a scale resolved by the simulations, then the implemented subgrid efficiency would also have to be a function of the local physical conditions in order to produce the correct mass functions of galaxies and BHs.
A similar story applies to the gas fractions of galaxies or, more precisely, for the amount of gas above the assumed star formation threshold, even if the simulations have been calibrated to produce the correct GSMF. We can see this as follows. If the outflow rate is determined by the inflow rate, then it is \emph{not} determined by the assumed subgrid star formation law. Hence, if we modify the star formation law,\footnote{The argument breaks down if the gas consumption time scale becomes longer than the Hubble time.} then the mean outflow rate should remain unchanged. And if the outflow rate remains unchanged, then so must the star formation rate because for a fixed feedback efficiency the star formation rate will adjust to the rate required for outflows to balance inflows. If the star formation rate is independent of the star formation law, then the galaxies must adjust the amount of star-forming gas that they contain when the star formation law is changed.
Hence, to predict the correct amount of star-forming gas, we need to calibrate the subgrid model for star formation to the observed star formation law.
Fortunately, the star formation law is relatively well characterised observationally on the $\sim 10^2 - 10^3\,{\rm pc}$ scales resolved by large-volume simulations, although there are important unanswered questions, e.g.\ regarding the dependence on metallicity. Ultimately the star formation law must be predicted by simulations and will probably depend on the true efficiency of feedback processes within the ISM, but resolving such processes is not yet possible in simulations of cosmological volumes.
It is not obvious how the efficiency of feedback from star formation should be calibrated. We could choose to calibrate to observations of outflow rates relative to star formation rates. However, those outflow rates are highly uncertain and may be affected by AGN feedback. It is also unclear on what scale the outflow rate should be calibrated. In addition, the outflow velocity and the wind mass loading may be individually important. Moreover, unless the interaction of the wind with the circumgalactic medium is modelled correctly and resolved, then obtaining a correct outflow rate on the scale used for the calibration does not necessarily imply that it is also correct for the other scales that matter.
We choose to calibrate the feedback efficiency using the observed present-day GSMF, as is also common practice for semi-analytic models. We do this mostly because it is relatively well constrained observationally and because obtaining the correct stellar mass - halo mass relation, and hence the correct GSMF if the cosmological initial conditions are known, is a pre-condition for many applications of cosmological simulations. For example, the physical properties of the circumgalactic medium (CGM) are likely sensitive to the halo mass, but because halo mass is difficult to measure, observations and simulations of the CGM are typically compared for galaxies of the same stellar mass.
One may wonder what the point of hydrodynamical simulations (or, indeed, semi-analytic models) is if they cannot predict stellar masses or BH masses. This is a valid question for which there are several answers. One is that the simulations can still make predictions for observables that were not used for the calibration, and we will present such predictions in \S\ref{sec:otherobs} and in subsequent papers. However, which observables are unrelated is not always unambiguous. One way to proceed, and an excellent way to learn about the physics of galaxy formation, is to run multiple simulations with varying subgrid models. It is particularly useful to have multiple prescriptions calibrated to the same observables. EAGLE\ comprises many variations, including several that reproduce the $z\sim 0$ GSMF through different means \citep{Crain2014EagleModels}.
A second answer is that making good use of simulations of galaxy formation does not necessarily mean making quantitative predictions for observables of the galaxy population. We can use the simulations to gain insight into physical processes, to explore possible scenarios, and to make qualitative predictions. How does gas get into galaxies? What factors control the size of galaxies? What is the origin of scatter in galaxy scaling relations? What is the potential effect of outflows on cosmology using weak gravitational lensing or the \lya\ forest? The list of interesting questions is nearly endless.
A third answer is that cosmological, hydrodynamical simulations can make robust, quantitative predictions for more diffuse components, such as the low-density IGM and perhaps the outer parts of clusters of galaxies.
A fourth answer is that calibrated simulations can be useful to guide the interpretation and planning of observations, as the use of semi-analytic and halo models has clearly demonstrated. In this respect hydrodynamical simulations can provide more detailed information on both the galaxies and their gaseous environments.
\subsection{Numerical convergence}
\label{sec:conv_discussion}
The need to calibrate the efficiency of the feedback and the associated limits on the predictive power of the simulations call the role of numerical convergence into question. The conventional point of view is that subgrid models should be designed to yield numerically converged predictions. Convergence is clearly a necessary condition for predictive power. However, we have just concluded that current simulations cannot, in any case, make \emph{ab initio} predictions for some of the most fundamental observables of the galaxy population.
While it is obvious that we should demand convergence for predictions that are relatively robust to the choice of subgrid model, e.g.\ the statistics of the \lya\ forest, it is less obvious that the same is required for observables that depend strongly and directly on the efficiency of the subgrid feedback. One could argue that, instead, we only need convergence after recalibration of the subgrid model. We will call this ``weak convergence'', as opposed to the ``strong convergence'' that is obtained if the results do not change with resolution when the model is held fixed.
If only weak convergence is required, then the demands placed on the subgrid model are much reduced, which has two advantages:
First, we can take better advantage of increases in resolution. The subgrid scale can now move along with the resolution limit, so we can potentially model the physics more faithfully if we adopt higher resolution.
A second advantage of demanding only weak convergence is that we do not have to make the sacrifices that are required to improve the strong convergence and that might have undesirable consequences. We will provide three examples of compromises that are commonly made.
Simulations that sample a representative volume currently lack the resolution and the physics to predict the radiative losses to which outflows are subject within the ISM. Strong convergence can nevertheless be achieved if these losses are somehow removed altogether, for example, by temporarily turning off radiative cooling and calibrating the criterion for switching it back on \citep[e.g.][]{Gerritsen1997PhD,Stinson2006Winds}. However, it is then unclear for which gas the cooling should be switched off. Only the gas elements into which the subgrid feedback was directly injected? Or also the surrounding gas that is subsequently shock-heated?
Other ways to circumvent radiative losses in the ISM are to generate the outflow outside the galaxy or to turn off the hydrodynamic interaction between the wind and the ISM \citep[e.g.][]{Springel2003Multiphase,Oppenheimer2006Wpot,Oppenheimer2010GSMF,Puchwein2013GSMF,Vogelsberger2013IllustrisModel,Vogelsberger2014Illustris}. This is a valid choice, but one that eliminates the possibility of capturing any aspect of the feedback other than mass loss, such as puffing up of discs, blowing holes, driving turbulence, collimating outflows, ejecting gas clouds, generating small-scale galactic fountains, etc. Furthermore, it necessarily introduces new parameters that control where the outflow is generated and when the hydrodynamics is turned back on. These parameters may directly affect results of interest, including the state of gas around galaxies, and may also re-introduce resolution effects. A potential solution to this problem is to never re-couple and hence to evaluate all wind interactions using a subgrid model, even outside the galaxies, as is done in semi-analytic models.
However, bypassing radiative losses in the ISM is not by itself sufficient to achieve strong convergence. In addition, the feedback must not depend on physical conditions in the ISM since those are unlikely to be converged. Instead, one can make the feedback depend on properties defined by the dark matter, such as its local velocity dispersion or halo mass \citep[e.g.][]{Oppenheimer2006Wpot,Okamoto2010Sats,Oppenheimer2010GSMF,Puchwein2013GSMF,Vogelsberger2013IllustrisModel,Vogelsberger2014Illustris}, which are generally better converged than the properties of the gas. As was the case for turning off cooling or hydrodynamic forces, this choice makes the simulations less ``hydrodynamical'', moving them in the direction of more phenomenological approaches, and it also introduces new problems. How do we treat satellite galaxies given that their subhalo mass and dark matter velocity dispersion are affected by the host halo? Or worse, what about star clusters or tidal dwarf galaxies that are not hosted by dark matter haloes?
In practice, however, the distinction between weak and strong convergence is often unclear. One may surmise that keeping the physical model fixed is equivalent to keeping the code and subgrid parameters fixed (apart from the numerical parameters controlling the resolution), but this is not necessarily the case because of the reliance on subgrid prescriptions and the inability to resolve the first generations of stars and BHs. For typical subgrid prescriptions, the energy, the mass, and the momentum involved in individual feedback events, and the number or intermittency of feedback events do not all remain fixed when the resolution is changed. Any such changes could affect the efficiency of the feedback. Consider, for example, a star-forming region and assume that feedback energy from young stars is distributed locally at every time step. If the resolution is increased, then the time step and the particle mass will become smaller. If the total star formation rate remains the same, then the feedback energy that is injected per time step will be smaller because of the decrease in the time step. If the gas mass also remains the same, then the temperature increase per time step will be smaller. A lower post-feedback temperature often leads to larger thermal losses. If, instead, the subgrid model specifies the temperature jump (or wind velocity), then the post-feedback temperature will remain the same when the resolution is increased, but the number of heating events will increase because the same amount of feedback energy has to be distributed over lower-mass particles. There is no guarantee that more frequent, lower-energy events drive the same outflows as less frequent, higher-energy events.
Moreover, for cosmological initial conditions, higher resolution implies resolving smaller haloes, and hence tracing the progenitors of present-day galaxies to higher redshifts. If these progenitors drive winds, then this may impact the subsequent evolution.
In \S\ref{sec:gsmf} we investigate both the weak and strong convergence of our simulations, focusing on the GSMF.
We test the weak convergence for a wide variety of predictions in sections~\ref{sec:cal_obs} and \ref{sec:otherobs}.
\section{Simulations}
\label{sec:simulations}
\begin{table}
\caption{The cosmological parameters used for the EAGLE\ simulations: $\Omega_{\rm m}$,
$\Omega_\Lambda$, and $\Omega_{\rm b}$ are the average densities
of matter, dark energy and baryonic matter
in units of the critical density at redshift zero; $H_0$ is the Hubble
parameter, $\sigma_8$ is the square root of the linear variance of the
matter distribution when smoothed with a top-hat filter of radius
$8~h^{-1}\,{\rm cMpc}$, $n_{\rm s}$ is the scalar power-law index of the power
spectrum of primordial adiabatic perturbations, and $Y$ is the primordial abundance of helium.}
\begin{center}
\begin{tabular}{|l|r|}
\hline
Cosmological parameter & Value \\
\hline
$\Omega_{\rm m}$ & 0.307\phantom{00} \\
$\Omega_\Lambda$ & 0.693\phantom{00} \\
$\Omega_{\rm b}$ & 0.04825\phantom{} \\
$h \equiv H_0$/(100 km\,s$^{-1}$\,Mpc$^{-1})$ & 0.6777\phantom{0} \\
$\sigma_8$ & 0.8288\phantom{0} \\
$ n_{\rm s}$ & 0.9611\phantom{0} \\
$Y$ & 0.248\phantom{00}\\
\hline
\end{tabular}
\end{center}
\label{tbl:cosmo_params}
\end{table}
EAGLE\ was run using a modified version of the $N$-Body Tree-PM smoothed particle hydrodynamics (SPH) code \textsc{gadget}~3, which was last described in \citet{Springel2005Gadget2}. The main modifications are the formulation of SPH, the time stepping and, most importantly, the subgrid physics.
The subgrid physics used in EAGLE\ is based on that developed for OWLS\ \citep{Schaye2010OWLS}, and used also in GIMIC\ \citep{Crain2009GIMIC} and cosmo-OWLS\ \citep{LeBrun2014CosmoOWLS}. We include element-by-element radiative cooling for 11 elements, star formation, stellar mass loss, energy feedback from star formation, gas accretion onto and mergers of supermassive black holes (BHs), and AGN feedback. As we will detail in \S\ref{sec:subgrid}, we made a number of changes with respect to OWLS. The most important changes concern the implementations of energy feedback from star formation (which is now thermal rather than kinetic), the accretion of gas onto BHs (which now accounts for angular momentum), and the star formation law (which now depends on metallicity).
In the simulations presented here the amount of feedback energy that is injected per unit stellar mass decreases with the metallicity and increases with the gas density. It is bounded between one third and three times the energy provided by supernovae and, on average, it is about equal to that amount. The metallicity dependence is motivated by the fact that we expect greater (unresolved) thermal losses when the metallicity exceeds $\sim 10^{-1}\,{\rm Z}_\odot$, the value for which metal-line cooling becomes important. The density dependence compensates for spurious, numerical radiative losses which, as expected, are still present at our resolution even though they are greatly reduced by the use of the stochastic prescription of \citet{DallaVecchia2012Winds}. The simulations were calibrated against observational data by running a series of high-resolution 12.5~cMpc and intermediate resolution 25~cMpc test runs with somewhat different dependencies on metallicity and particularly density. From the models that predicted reasonable physical sizes for disc galaxies, we selected the one that best fit the $z\sim 0$ GSMF. For more details on the subgrid model for energy feedback from star formation we refer the reader to \S\ref{sec:snii}.
As described in more detail in Appendix~\ref{app:hydro}, we make use of the conservative pressure-entropy formulation of SPH derived by \citet{Hopkins:2013lr}, the artificial viscosity switch from \citet{Cullen:2010qy}, an artificial conduction switch similar to that of \citet{Price:2008kx}, the $C^2$ \citet{Wendland:1995} kernel and the time step limiters of \citet{Durier:2012fj}. We will refer to these numerical methods collectively as ``Anarchy''. Anarchy will be described in more detail by Dalla Vecchia (in preparation), who also demonstrates its good performance on standard hydrodynamical tests (see \citealt{Hu2014SPHGal} for tests of a similar set of methods). In \citet{Schaller2014EagleSPH} we will show the relevance of the new hydrodynamical techniques and time stepping scheme for the results of the EAGLE\ simulations. Although the Anarchy implementation yields dramatic improvements in the performance on some standard hydrodynamical tests as compared to the original implementation of the hydrodynamics in \textsc{gadget}~3, we generally find that the impact on the results of the cosmological simulations is small compared to those resulting from reasonable variations in the subgrid physics (see also \citealt{Scannapieco2012Aquila}).
\begin{table*}
\begin{center}
\caption{Box sizes and resolutions of the main EAGLE\ simulations. From
left-to-right the columns show: simulation name suffix; comoving box size;
number of dark matter particles (there is initially an equal number of baryonic particles); initial baryonic particle mass; dark matter
particle mass; comoving, Plummer-equivalent gravitational
softening length; maximum proper softening length.}
\label{tbl:sims}
\begin{tabular}{lrrrrrrr}
\hline
Name & $L$ & $N$ & $m_{\rm g}$ & $m_{\rm dm}$ &
$\epsilon_{\rm com}$ & $\epsilon_{\rm prop}$ \\
& (comoving Mpc) & & (${{\rm M}_\odot}$) & (${{\rm M}_\odot}$) & (comoving kpc) & (proper kpc)\\
\hline
L025N0376 & 25 & $376^3$ & $1.81\times 10^6$ & $9.70\times 10^6$ & 2.66 &0.70\\
L025N0752 & 25 & $752^3$ & $2.26\times 10^5$ & $1.21\times 10^6$ & 1.33 & 0.35\\
L050N0752 & 50 & $752^3$ & $1.81\times 10^6$ & $9.70\times 10^6$ & 2.66 &0.70\\
L100N1504 & 100 & $1504^3$ & $1.81\times 10^6$ & $9.70\times 10^6$ & 2.66 &0.70\\
\hline
\end{tabular}
\end{center}
\end{table*}
The values of the cosmological parameters used for the EAGLE\ simulations are taken from the most recent Planck results \citep[][Table 9]{PlanckI} and are listed in Table~\ref{tbl:cosmo_params}. A transfer function with these
parameters was generated using CAMB \citep[version Jan\_12]{CAMB}.
The linear matter power spectrum was generated by multiplying a
power-law primordial power spectrum with an index of $n_{\rm s} = 0.9611$ by the
square of the dark matter transfer function evaluated at redshift
zero\footnote{The CAMB input parameter file and the linear power spectrum are available at \url{http://eagle.strw.leidenuniv.nl/}.}. Particles arranged in a glass-like initial configuration were displaced according to 2nd-order Lagrangian perturbation theory using the method of \citet{Jenkins20102lpt} and the public Gaussian white noise field \emph{Panphasia} \citep{Jenkins2013ICs,Jenkins2013Panphasia}. The methods used to generate the initial conditions are described in detail in Appendix~\ref{app:ics}.
\begin{figure*}
\resizebox{\textwidth}{!}{\includegraphics{highresblowoutofgalaxies_reduced.png}}
\caption{A $100\times 100\times 20$ cMpc slice through the Ref-L100N1504 simulation at $z=0$. The intensity shows the gas density while the colour encodes the gas temperature using different colour channels for gas with $T<10^{4.5}\,{\rm K}$ (blue), $10^{4.5}\,{\rm K}<T<10^{5.5}\,{\rm K}$ (green), and $T>10^{5.5}\,{\rm K}$ (red). The insets show regions of 10 cMpc and 60 ckpc on a side and zoom into an individual galaxy with a stellar mass of $3\times 10^{10}\,{{\rm M}_\odot}$. The 60 ckpc image shows the stellar light based on monochromatic u, g and r band SDSS filter means and accounting for dust extinction. It was created using the radiative transfer code {\sc skirt}\ \citep{Baes2011SKIRT}.}
\label{fig:zoom}
\end{figure*}
\begin{figure*}
\resizebox{\textwidth}{!}{\includegraphics{HS_flip.png}}
\caption{Examples of galaxies taken from simulation Ref-L100N1504 illustrating the $z=0$ Hubble sequence of galaxy morphologies. The images were created with the radiative transfer code {\sc skirt}\ \citep{Baes2011SKIRT}. They show the stellar light based on monochromatic u, g and r band SDSS filter means and accounting for dust extinction. Each image is 60 ckpc on a side. For disc galaxies both face-on and edge-on projections are shown. Except for the 3rd elliptical from the left, which has a stellar mass of $1\times 10^{11}\,{{\rm M}_\odot}$, and the merger in the bottom-left, which has a total stellar mass of $8\times 10^{10}\,{{\rm M}_\odot}$, all galaxies shown have stellar masses of 5--$6\times 10^{10}\,{{\rm M}_\odot}$.}
\label{fig:morphology}
\end{figure*}
Table~\ref{tbl:sims} lists box sizes and resolutions of the main EAGLE\ simulations. All simulations were run to redshift $z=0$. Note that contrary to convention, box sizes, particles masses and gravitational softening lengths are \emph{not} quoted in units of $h^{-1}$. The gravitational softening was kept fixed in comoving units down to $z=2.8$ and in proper units thereafter. We will refer to simulations with the same mass and spatial resolution as L100N1504 as intermediate resolution runs and to simulations with the same resolution as L025N0752 as high-resolution runs.
Particle properties were recorded for 29 snapshots between redshifts 20 and 0. In addition, we saved a reduced set of particle properties (``snipshots'') at 400 redshifts between 20 and 0. The largest simulation, L100N1504, took about 4.5~M CPU hours to reach $z=0$ on a machine with 32~TB of memory, with the EAGLE\ subgrid physics typically taking less than 25 per cent of the CPU time.
The resolution of EAGLE\ suffices to marginally resolve the Jeans scales in the warm ISM. The Jeans mass and length for a cloud with gas fraction, $f_{\rm g}$, are, respectively, $M_{\rm J} \approx 1 \times 10^7\,{{\rm M}_\odot}\, f_{\rm g}^{3/2} (n_{\rm H} /10^{-1}\,{\rm cm}^{-3})^{-1/2} (T / 10^4\,{\rm K})^{3/2}$ and $L_{\rm J} \approx 2~{\rm kpc} ~f_{\rm g}^{1/2} (n_{\rm H} / 10^{-1}\,{\rm cm}^{-3})^{-1/2} (T / 10^4\,{\rm K})^{1/2}$, where $n_{\rm H}$ and $T$ are the total hydrogen number density and the temperature, respectively. These Jeans scales can be compared to the gas particle masses and maximum proper gravitational softening lengths listed in columns 4 and 7 of Table~\ref{tbl:sims}.
Simulations with the same subgrid physics and numerical techniques as used for L100N1504 were carried out for all box sizes (12.5 -- 100 cMpc) and particles numbers ($188^3$ -- $1504^3$). We will refer to this physical model as the reference model and will indicate the corresponding simulations with the prefix ``Ref-'' (e.g.\ Ref-L100N1504). As detailed in \S\ref{sec:subgrid}, we re-ran the high-resolution simulations with recalibrated parameter values for the subgrid stellar and AGN feedback to improve the match to the observed $z\sim 0$ GSMF. We will use the prefix ``Recal-'' when referring to the simulations with this alternative set of subgrid parameters (e.g.\ Recal-L025N0752). Note that in terms of weak convergence, Ref-L100N1504 is more similar to model Recal-L025N0752 than to model Ref-L025N0752 (see \S\ref{sec:conv_discussion} for a discussion of weak and strong convergence). In addition, we repeated the L050N0752 run with adjusted AGN parameters in order to further improve the agreement with observations for high-mass galaxies. We will refer to this model with the prefix ``AGNdT9''. Table~\ref{tbl:subgridpars} summarizes the values of the four subgrid parameters that vary between the models presented here. \citet{Crain2014EagleModels} and \citet{Schaller2014EagleSPH} will present the remaining EAGLE\ simulations, which concern variations in the subgrid physics and the numerical techniques, respectively. Finally, \citet{Sawala2014EagleZooms} present very high-resolution zoomed simulations of Local Group like systems run with the EAGLE\ code and a physical model that is nearly identical to the one used for the Ref-L100N1504 model described here.
\begin{table}
\begin{center}
\caption{Values of the subgrid parameters that vary between the models presented here. The parameters $n_{\rm H,0}$ and $n_n$ control, respectively, the characteristic density and the power-law slope of the density dependence of the energy feedback from star formation (see equation \ref{eq:f(Z,n)} in \S\ref{sec:calibration}). The parameter $C_{\rm visc}$ controls the sensitivity of the BH accretion rate to the angular momentum of the gas (see equation \ref{eq:mdotaccr} in \S\ref{sec:bh_accretion}) and $\Delta T_{\rm AGN}$ is the temperature increase of the gas during AGN feedback (see \S\ref{sec:AGNfeedback}). }
\label{tbl:subgridpars}
\begin{tabular}{lrrrl}
\hline
Prefix & $n_{\rm H,0}$ & $n_n$ & $C_{\rm visc}$ & $\Delta T_{\rm AGN}$ \\
& (${\rm cm}^{-3}$) & & & (K) \\
\hline
Ref & 0.67 & $2/\ln10$ & $2\pi$ & $10^{8.5}$ \\
Recal & 0.25 & $1/\ln10$ & $2\pi \times 10^3$ & $10^9$ \\
AGNdT9 & 0.67 & $2/\ln10$ & $2\pi \times 10^2$ & $10^9$ \\
\hline
\end{tabular}
\end{center}
\end{table}
Figure~\ref{fig:zoom} illustrates the large dynamic range of EAGLE. It shows the large-scale gas distribution in a thick slice through the $z=0$ output of the Ref-L100N1504 run, colour-coded by the gas temperature. The insets zoom in on an individual galaxy. The first zoom shows the gas, but the last zoom shows the stellar light after accounting for dust extinction. This image was created using three monochromatic radiative transfer simulations with the code {\sc skirt}\ \citep{Baes2011SKIRT} at the effective wavelengths of the Sloan Digital Sky Survey (SDSS) u, g \& r filters. Dust extinction is implemented using the metal distribution predicted by the simulations and assuming that 30 per cent of the metal mass is locked up in dust grains. Only material within a spherical aperture with a radius of 30 pkpc is included in the radiative transfer calculation. More examples of {\sc skirt}\ images of galaxies are shown in Figure~\ref{fig:morphology}, in the form of a Hubble sequence. This figure illustrates the wide range of morphologies present in EAGLE. Note that \citet{Vogelsberger2014IllustrisNature} showed a similar figure for their Illustris simulation. In future work we will investigate how morphology correlates with other galaxy properties. More images, as well as videos, can be found on the EAGLE\ web sites at Leiden, \url{http://eagle.strw.leidenuniv.nl/}, and Durham, \url{http://icc.dur.ac.uk/Eagle/}.
We define galaxies as gravitationally bound subhaloes identified by the \textsc{subfind} algorithm \citep{Springel2001Subfind,Dolag2009Substructure}. The procedure consists of three main steps. First we find haloes by running the Friends-of-Friends (FoF; \citealt{Davis1985FoF}) algorithm on the dark matter particles with linking length 0.2 times the mean interparticle separation. Gas and star particles are assigned to the same, if any, FoF halo as their nearest dark matter particles. Second, \textsc{subfind} defines substructure candidates by identifying overdense regions within the FoF halo that are bounded by saddle points in the density distribution. Note that whereas FoF considers only dark matter particles, \textsc{subfind} uses all particle types within the FoF halo. Third, particles that are not gravitationally bound to the substructure are removed and the resulting substructures are referred to as subhaloes. Finally, we merged subhaloes separated by less than the minimum of 3~pkpc and the stellar half-mass radius. This last step removes a very small number of very low-mass subhaloes whose mass is dominated by a single particle such as a supermassive BH.
For each FoF halo we define the subhalo that contains the particle with the lowest value of the gravitational potential to be the central galaxy while any remaining subhaloes are classified as satellite galaxies. The position of each galaxy is defined to be the location of the particle belonging to the subhalo for which the gravitational potential is minimum.
The stellar mass of a galaxy is defined to be the sum of the masses of all star particles that belong to the corresponding subhalo and that are within a 3-D aperture with radius 30~pkpc. Unless stated otherwise, other galaxy properties, such as the star formation rate, metallicity, and half-mass radius, are also computed using only particles within the 3-D aperture.
In \S\ref{sec:aperture} we show that this aperture gives a nearly identical GSMF as the 2-D Petrosian apertures that are frequently used in observational studies.
We find the effect of the aperture to be negligible for $M_\ast < 10^{11}\,{{\rm M}_\odot}$ for all galaxy properties that we consider. However, for more massive galaxies the aperture reduces the stellar masses somewhat by cutting out intracluster light. For example, at a stellar mass $M_\ast = 10^{11}\,{{\rm M}_\odot}$ as measured using a 30 pkpc aperture, the median subhalo stellar mass is 0.1~dex higher (see \S\ref{sec:aperture} for the effect on the GSMF). Without the aperture, metallicities are slightly lower and half-mass radii are slightly larger for $M_\ast > 10^{11}\,{{\rm M}_\odot}$, but the effect on the star formation rate is negligible.
\section{Subgrid physics}
\label{sec:subgrid}
In this section we provide a thorough description and motivation for the subgrid physics implemented in EAGLE: radiative cooling (\S\ref{sec:cooling}), reionisation (\S\ref{sec:reionisation}), star formation (\S\ref{sec:sf}), stellar mass loss and metal enrichment (\S\ref{sec:chemo}), energy feedback from star formation (\S\ref{sec:snii}), and supermassive black holes and AGN feedback (\S\ref{sec:BHs}). These subsections can be read separately. Readers who are mainly interested in the results may skip this section.
\subsection{Radiative cooling}
\label{sec:cooling}
Radiative cooling and photoheating are implemented element-by-element following \citet{Wiersma2009Cooling}, including all 11 elements that they found to be important: H, He, C, N, O, Ne, Mg, Si, S, Ca, and Fe. \citet{Wiersma2009Cooling} used \textsc{cloudy} version\footnote{Note that OWLS\ used tables based on version 05.07.} 07.02 \citep{Ferland1998Cloudy} to tabulate the rates as a function of density, temperature, and redshift assuming the gas to be in ionisation equilibrium and exposed to the cosmic microwave background (CMB) and the \citet{Haardt2001UVB} model for the evolving UV/X-ray background from galaxies and quasars. By computing the rates element-by-element, we account not only for variations in the metallicity, but also for variations in the relative abundances of the elements.
We caution that our assumption of ionisation equilibrium and the neglect of local sources of ionizing radiation may cause us to overestimate the cooling rate in certain situations, e.g.\ in gas that is cooling rapidly \cite[e.g.][]{Oppenheimer2013Nonequil} or that has recently been exposed to radiation from a local AGN \citep{Oppenheimer2013AGNFossils}.
We have also chosen to ignore self-shielding, which may cause us to underestimate the cooling rates in dense gas. While we could have accounted for this effect, e.g.\ using the fitting formula of \citet{Rahmati2013HICDDF}, we opted against doing so because there are other complicating factors. Self-shielding is only expected to play a role for $n_{\rm H} > 10^{-2}\,{\rm cm}^{-3}$ and $T\la 10^4\,{\rm K}$ \citep[e.g.][]{Rahmati2013HICDDF}, but at such high densities the radiation from local stellar sources, which we neglect here, is expected to be at least as important as the background radiation \citep[e.g.][]{Schaye2001MaxHI,Rahmati2013LocalSources}.
\subsection{Reionization}
\label{sec:reionisation}
Hydrogen reionization is implemented by turning on the time-dependent, spatially-uniform ionizing background from \citet{Haardt2001UVB}. This is done at redshift $z=11.5$, consistent with the optical depth measurements from \citet{PlanckI}. At higher redshifts we use net cooling rates for gas exposed to the CMB and the photo-dissociating background obtained by cutting the $z = 9$ \citet{Haardt2001UVB} spectrum above 1~Ryd.
To account for the boost in the photoheating rates during reionization relative to the optically thin rates assumed here, we inject 2~eV per proton mass. This ensures that the photoionised gas is quickly heated to $\sim 10^4\,{\rm K}$. For H this is done instantaneously, but for \HeII\ the extra heat is distributed in redshift with a Gaussian centred on $z=3.5$ of width $\sigma(z)=0.5$. \citet{Wiersma2009Chemo} showed that this choice results in broad agreement with the thermal history of the intergalactic gas as measured by \citet{Schaye2000IGMTemp}.
\subsection{Star formation}
\label{sec:sf}
Star formation is implemented following \citet{Schaye2008SF}, but with the metallicity-dependent density threshold of \citet{Schaye2004SF} and a different temperature threshold, as detailed below. Contrary to standard practice, we take the star formation rate to depend on pressure rather than density. As demonstrated by \citet{Schaye2008SF}, this has two important advantages. First, under the assumption that the gas is self-gravitating, we can rewrite the observed Kennicutt-Schmidt star formation law \citep{Kennicutt1998Law}, $\dot{\Sigma}_\ast = A (\Sigma_{\rm g}/1~{{\rm M}_\odot}\,{\rm pc}^{-2})^n$, as a pressure law:
\begin{equation}
\dot{m}_\ast = m_{\rm g} A \left (1~{{\rm M}_\odot}\,{\rm pc}^{-2}\right
)^{-n} \left ({\gamma \over G} f_{\rm g} P\right )^{(n-1)/2},
\label{eq:sflaw}
\end{equation}
where $m_{\rm g}$ is the gas particle mass, $\gamma=5/3$ is the
ratio of specific heats, $G$ is the gravitational constant, $f_{\rm g}$ is the mass fraction in gas (assumed to be unity), and $P$ is the total pressure. Hence, the free parameters $A$ and $n$ are determined by observations of the gas and star formation rate surface densities of galaxies and no tuning is necessary. Second, if we impose an equation of state, $P=P_{\rm eos}(\rho)$, then the observed Kennicutt-Schmidt star formation law will still be reproduced without having to change the star formation parameters. In contrast, if star formation is implemented using a volume density rather than a pressure law, then the predicted Kennicutt-Schmidt law will depend on the thickness of the disc and thus on the equation of state of the star forming gas. Hence, in that case the star formation law not only has to be calibrated, it has to be recalibrated if the imposed equation of state is changed. In practice, this is rarely done.
Equation (\ref{eq:sflaw}) is implemented stochastically. The probability that a gas particle is converted into a collisionless star particle during a time step $\Delta t$ is $\min(\dot{m}_\ast \Delta t/m_{\rm g},1)$.
We use $A=1.515\times10^{-4}~{\rm M}_\odot\,{\rm yr}^{-1}\,{\rm kpc}^{-2}$ and $n=1.4$, where we have decreased the amplitude by a factor 1.65 relative to the value used by \citet{Kennicutt1998Law} because we use a Chabrier rather than a Salpeter stellar initial mass function (IMF). We increase $n$ to 2 for $n_{\rm H} > 10^3\,{\rm cm}^{-3}$, because there is some evidence for a steepening at high densities \citep[e.g.][]{Liu2011KSLaw,Genzel2010SFLaw}, but this does not have a significant effect on the results since only $\sim 1$\% of the stars form at such high densities in our simulations.
Star formation is observed to occur in cold ($T\ll 10^4\,{\rm K}$), molecular gas. Because simulations of large cosmological volumes, such as ours, lack the resolution and the physics to model the cold, interstellar gas phase, it is appropriate to impose a star formation threshold at the density above which a cold phase is expected to form. In OWLS\ we used a constant threshold of $n_{\rm H}^* = 10^{-1}\,{\rm cm}^{-3}$, which was motivated by theoretical considerations and yields a critical gas surface density $\sim 10~{{\rm M}_\odot}\,{\rm pc}^{-2}$ \citep{Schaye2004SF,Schaye2008SF}. The critical volume density, $n_{\rm H} = 0.1~{\rm cm}^{-3}$, is also similar to the value used in other work of comparable resolution \citep[e.g.][]{Springel2003Multiphase,Vogelsberger2013IllustrisModel}. Here we instead use the metallicity-dependent density threshold of \citet{Schaye2004SF} as implemented in OWLS\ model ``SFTHRESZ'' (eq.\ 4 of \citealt{Schaye2010OWLS}; equations 19 and 24 of
\citealt{Schaye2004SF}),
\begin{equation}
\label{eq:sfthresz}
n_{\rm H}^*(Z)=10^{-1}\,{\rm cm}^{-3} \left ({Z \over 0.002}\right )^{-0.64},
\end{equation}
where $Z$ is the gas metallicity (i.e.\ the fraction of the gas mass in elements heavier than helium). In the code the threshold is evaluated as a mass density rather than a total hydrogen number density. To prevent an additional dependence on the hydrogen mass fraction (beyond that implied by equation \ref{eq:sfthresz}), we convert $n_{\rm H}$ into a mass density assuming the initial hydrogen mass fraction, $X=0.752$. Because the \citet{Schaye2004SF} relation diverges at low metallicities, we impose an upper limit of $n_{\rm H}^*=10~{\rm cm}^{-3}$. To prevent star formation in low overdensity gas at very high redshift, we also require the gas density to exceed 57.7 times the cosmic mean, but the results are insensitive to this value.
The metallicity dependence accounts for the fact that the transition from a warm, neutral to a cold, molecular phase occurs at lower densities and pressures if the metallicity, and hence also the dust-to-gas ratio, is higher. The phase transition shifts to lower pressures if the metallicity is increased due to the higher formation rate of molecular hydrogen, the increased cooling due to metals and the increased shielding by dust \citep[e.g.][]{Schaye2001MaxHI,Schaye2004SF,Pelupessy2006H2,Krumholz2008HIH2,Gnedin2009H2form,Richings2014Shielding}. Our metallicity-dependent density threshold causes the critical gas surface density below which the Kennicutt-Schmidt law steepens to decrease with increasing metallicity.
Because our simulations do not model the cold gas phase, we impose a temperature floor, $T_{\rm eos}(\rho_{\rm g})$, corresponding to the equation of state $P_{\rm eos}\propto \rho_{\rm g}^{4/3}$, normalised to\footnote{For the purpose of imposing temperature floors, $T_{\rm eos}(\rho_{\rm g})$ is converted into an entropy assuming a fixed mean molecular weight of 1.2285, which corresponds to an atomic, primordial gas. Other conversions in the code use the actual mean molecular weight and hydrogen abundance, but we keep them fixed here to prevent particles with different abundances from following different effective equations of state.} $T_{\rm eos} = 8\times 10^3\,{\rm K}$ at $n_{\rm H} = 10^{-1}\,{\rm cm}^{-3}$, a temperature that is typical for the warm ISM \citep[e.g.][]{Richings2014Shielding}. The slope of $4/3$ guarantees that the Jeans mass, and the ratio of the Jeans length to the SPH kernel, are independent of the density, which prevents spurious fragmentation due to the finite resolution \citep{Schaye2008SF,Robertson2008SFlaw}. Following \citet{DallaVecchia2012Winds}, gas is eligible to form stars if $\log_{10} T < \log_{10}T_{\rm eos} + 0.5$ and $n_{\rm H} > n_{\rm H}^*$, where $n_{\rm H}^*$ depends on metallicity as specified above.
Because of the existence of a temperature floor, the temperature of star forming (i.e.\ interstellar) gas in the simulation merely reflects the effective pressure imposed on the unresolved, multiphase ISM, which may in reality be dominated by turbulent rather than thermal pressure. If the temperature of this gas needs to be specified, e.g.\ when computing neutral hydrogen fractions in post-processing, then one should assume a value based on physical considerations rather than use the formal simulation temperatures at face value.
In addition to the minimum pressure corresponding to the equation of state with slope $4/3$, we impose a temperature floor of 8000~K for densities $n_{\rm H}>10^{-5}\,{\rm cm}^{-3}$ in order to prevent very metal-rich particles from cooling to temperatures characteristic of cold, interstellar gas. This constant temperature floor was not used in OWLS\ and is unimportant for our results. We impose it because we do not wish to include a cold interstellar phase since we do not model all the physical processes that are needed to describe it. We only impose this limit for densities $n_{\rm H}>10^{-5}\,{\rm cm}^{-3}$, because we should not prevent the existence of cold, adiabatically cooled, intergalactic gas, which our algorithms can model accurately.
\subsection{Stellar mass loss and type Ia supernovae}
\label{sec:chemo}
Star particles are treated as simple stellar populations (SSPs) with a \citet{Chabrier2003IMF} IMF in the range $0.1-100~{{\rm M}_\odot}$. The implementation of stellar mass loss is based on \citet{Wiersma2009Chemo}. At each time step\footnote{To reduce the computational cost associated with neighbour finding for stars, we implement the enrichment every 10 gravitational time steps for star particles older than 0.1 Gyr; for the high-resolution run, Recal-L025N0752, this is further reduced to once every 100 time steps for star particles older than 1 Gyr. We have verified that our results are unaffected by this reduction in the sampling of stellar mass loss from older SSPs.} and for each stellar particle, we compute which stellar masses reach the end of the main sequence phase using the metallicity-dependent lifetimes of \citet{Portinari1998Chemo}. The fraction of the initial particle mass reaching this evolutionary stage is used, together with the initial elemental abundances, to compute the mass of each element that is lost through winds from AGB stars, winds from massive stars, and core collapse supernovae using the nucleosynthetic yields from
\citet{Marigo2001AGByields} and \citet{Portinari1998Chemo}. The elements H, He, C, N, O, Ne, Mg, Si, and Fe are tracked individually, while for Ca and S we assume fixed mass ratios relative to Si of 0.094 and 0.605, respectively \citep{Wiersma2009Chemo}. In addition, we compute the mass and energy lost through supernovae of type Ia.
The mass lost by star particles is distributed among the neighbouring SPH particles using the SPH kernel, but setting the mass of the gas particles equal to the constant initial value, $m_{\rm g}$. Each SPH neighbour $k$ that is separated by a distance $r_k$ from a star particle with smoothing length $h$ then receives a fraction $\frac{m_{\rm g}}{\rho_k} W(r_k,h)/\Sigma_i \frac{m_{\rm g}}{\rho_i}W(r_i,h)$ of the mass lost during the time step, where $W$ is the SPH kernel and the sum is over all SPH neighbours. To speed up the calculation, we use only 48 neighbours for stellar mass loss rather than the 58 neighbours used for the SPH.
In \citet{Wiersma2009Chemo} and OWLS\ we used the current gas particle masses rather than the constant, initial gas particle mass when computing the weights. The problem with that approach is that gas particles that are more massive than their neighbours, due to having received more mass lost by stars, carry more weight and therefore become even more massive relative to their neighbours. We found that this runaway process can cause a very small fraction of particles to end up with masses that far exceed the initial particle mass. The fraction of very massive particles is always small, because massive particles are typically also metal rich and relatively quickly converted into star particles. Nevertheless, it is still undesirable to preferentially direct the lost mass to relatively massive gas particles. We therefore removed this bias by using the fixed initial particle mass rather than the current particle mass, effectively taking the dependence on gas particle mass out of the equation for the distribution of stellar mass loss.
We also account for the transfer of momentum and energy associated with the transfer of mass from star to gas particles. We refer here to the momentum and energy related to the difference in velocity between the star particle and the receiving gas particles, in addition to that associated with the mass loss process itself (e.g.\ winds or supernovae). We assume that winds from AGB stars have a velocity of $10~{\rm km}\,{\rm s}^{-1}$ \citep{Bergeat2005AGBMassLoss}. After adjusting the velocities of the receiving gas particles to conserve momentum, energy conservation is achieved by adjusting their entropies. Momentum and energy transfer may, for example, play a role if the differential velocity between the stellar and gas components is similar to or greater than the sound speed of the gas, although we should keep in mind that the change in the mass of a gas particle during a cooling time is typically small.
As in \citet{Wiersma2009Chemo}, the abundances used to evaluate the radiative cooling rates are computed as the ratio of the mass density of an element to the total gas density, where both are calculated using the SPH formalism. Star particles inherit their parent gas particles' kernel-smoothed abundances\footnote{Note that this implies that metal mass is only approximately conserved. However, \citet{Wiersma2009Chemo} demonstrated that the error in the total metal mass is negligible even for simulations that are much smaller than EAGLE.} and we use those to compute their lifetimes and yields. The use of SPH-smoothed abundances, rather than the mass fractions of the elements stored in each particle, is consistent with the SPH formalism. It helps to alleviate the symptoms of the lack of metal mixing that occurs when metals are fixed to particles. However, as discussed in \citet{Wiersma2009Chemo}, it does not solve the problem that SPH may underestimate metal mixing. The implementation of diffusion can be used to increase the mixing \citep[e.g.][]{Greif2009Mixing,Shen2010IGMEnrichment}, but we have opted not to do this because the effective diffusion coefficients that are appropriate for the ISM and IGM remain unknown.
The rate of supernovae of type Ia (SNIa) per unit initial stellar mass is given by,
\begin{equation}
\dot{N}_{\rm SNIa} = \nu \frac{e^{-t/\tau}}{\tau},
\label{eq:snia}
\end{equation}
where $\nu$ is the total number of SNIa per unit initial stellar mass and $\exp(-t/\tau)/\tau$ is a normalised, empirical delay time distribution function.
We set $\tau = 2$~Gyr and $\nu = 2\times 10^{-3}\,{{\rm M}_\odot}^{-1}$. Figure~\ref{fig:snia} shows that these choices yield broad agreement with the observed evolution of the SNIa rate density for the intermediate resolution simulations, although the AGNdT9-L050N0752 may overestimate the rate by $\sim 30$ per cent for lookback times of 4--7 Gyr. The high-resolution model, Recal-L025N0752, is consistent with the observations at all times.
\begin{figure}
\resizebox{\colwidth}{!}{\includegraphics{snia.pdf}}
\caption{The evolution of the supernova Ia rate density. Data points show observations from SDSS Stripe 82 \citep{Dilday2010SNIa}, SDSS-DR7 \citep{Graur2013SNIa}, SNLS \citep{Perrett2012SNIa}, GOODS \citep{Dahlen2008SNIa}, SDF \citep{Graur2011SNIa}, and CLASH \citep{Graur2014SNIa}, as compiled by \citet{Graur2014SNIa}. Only data classified by \citealt{Graur2014SNIa} as the ``most accurate and precise measurements'' are shown. The $1\sigma$ error bars account for both statistical and systematic uncertainties. The simulations assume that the rate is a convolution of the star formation rate density with an exponential delay time distribution (eq.~\ref{eq:snia}) with e-folding time $\tau = 2~$Gyr, normalised to yield $\nu = 2\times 10^{-3}\,{{\rm M}_\odot}^{-1}$ supernovae Ia per unit stellar mass when integrated over all time.}
\label{fig:snia}
\end{figure}
At each time step for which the mass loss is evaluated, star particles transfer the mass and energy associated with SNIa ejecta to their neighbours. We use the SNIa yields of the W7 model of \citet{Thielemann2003SNIayields}. Energy feedback from SNIa is implemented identically as for prompt stellar feedback using the stochastic thermal feedback model of \citet{DallaVecchia2012Winds} summarized in \S\ref{sec:snii}, using $\Delta T = 10^{7.5}\,{\rm K}$ and $10^{51}\,{\rm erg}$ per SNIa.
\subsection{Energy feedback from star formation}
\label{sec:snii}
Stars can inject energy and momentum into the ISM through stellar winds, radiation, and supernovae. These processes are particularly important for massive and hence short-lived stars. If star formation is sufficiently vigorous, the associated feedback can drive large-scale galactic outflows \citep[e.g.][]{Veilleux2005WindsReview}.
Cosmological, hydrodynamical simulations have traditionally struggled to make stellar feedback as efficient as is required to match observed galaxy masses, sizes, outflow rates and other data. If the energy is injected thermally, it tends to be quickly radiated away rather than to drive a wind \citep[e.g.][]{Katz1996TreeSPH}. This ``overcooling'' problem is typically attributed to a lack of numerical resolution. If the simulation does not contain dense, cold clouds, then the star formation is not sufficiently clumpy and the feedback energy is distributed too smoothly. Moreover, since in reality cold clouds contain a large fraction of the mass of the ISM, in simulations without a cold interstellar phase the density of the warm, diffuse phase, and hence its cooling rate, is overestimated.
While these factors may well contribute to the problem, \citet[][see also \citealt{DallaVecchia2008Winds}, \citealt{Creasey2011Overcooling} and \citealt{Keller2014Winds}]{DallaVecchia2012Winds} argued that the fact that the energy is distributed over too much mass may be a more fundamental issue.
For a standard IMF there is $\sim 1$ supernova per 100~${{\rm M}_\odot}$ of SSP mass and, in reality, all the associated mechanical energy is initially deposited in a few solar masses of ejecta, leading to very high initial temperatures (e.g.\ $\sim 2\times 10^8\,{\rm K}$ if $10^{51}\,{\rm erg}$ is deposited in $10~{{\rm M}_\odot}$ of gas). In contrast, in SPH simulations that distribute the energy produced by a star particle over its SPH neighbours, the ratio of the heated mass to the mass of the SSP will be much greater than unity. The mismatch in the mass ratio implies that the maximum temperature of the directly heated gas is far lower than in reality, and hence that its radiative cooling time is much too short. Because the mass ratio of SPH to star particles is independent of resolution, to first order this problem is independent of resolution. At second order, higher resolution does help, because the thermal feedback can be effective in generating an outflow if the cooling time is large compared with the sound crossing time across a resolution element, and the latter decreases with increasing resolution (but only as $m_{\rm g}^{1/3}$).
Thus, subgrid models are needed to generate galactic winds in large-volume cosmological simulations. Three types of prescriptions are widely used: injecting energy in kinetic form \citep[e.g.][]{Navarro1993KineticFeedback,Springel2003Multiphase,DallaVecchia2008Winds,Dubois2008Winds} often in combination with temporarily disabling hydrodynamical forces acting on wind particles \citep[e.g.][]{Springel2003Multiphase,Okamoto2005Disks,Oppenheimer2006Wpot}, temporarily turning off radiative cooling \citep[e.g.][]{Gerritsen1997PhD,Stinson2006Winds}, and explicitly decoupling different thermal phases (also within single particles) \citep[e.g.][]{Marri2003Multiphase,Scannapieco2006Multiphase,Murante2010Multiphase,Keller2014Winds}. Here we follow \citet[][see also \citealt{Kay2003XrayGroups}]{DallaVecchia2012Winds} and opt for a different type of solution: stochastic thermal feedback. By making the feedback stochastic, we can control the amount of energy per feedback event even if we fix the mean energy injected per unit mass of stars formed. We specify the temperature jump of gas particles receiving feedback energy, $\Delta T$, and use the fraction of the total amount of energy from core collapse supernovae per unit stellar mass that is injected on average, $f_{\rm th}$, to set the probability that an SPH neighbour of a young star particle is heated. We perform this operation only once, when the stellar particle has reached the age $3\times 10^7\,{\rm yr}$, which corresponds to the maximum lifetime of stars that explode as core collapse supernovae.
The value $f_{\rm th}=1$ corresponds to an expectation value for the injected energy of $8.73\times 10^{15}\,{\rm erg}\,{\rm g}^{-1}$ of stellar mass formed, which corresponds to the energy available from core collapse supernovae for a Chabrier IMF if we assume $10^{51}\,{\rm erg}$ per supernova and that stars with mass $6-100~{{\rm M}_\odot}$ explode ($6-8~{{\rm M}_\odot}$ stars explode as electron capture supernovae in models with convective overshoot; e.g.\ \citealt{Chiosi1992Overshoot}).
If $\Delta T$ is sufficiently high, then the initial (spurious, numerical) thermal losses will be small and we can control the overall efficiency of the feedback using $f_{\rm th}$. This freedom is justified, because there will be \emph{physical} radiative losses in reality that we cannot predict accurately for the ISM. Moreover, because the true radiative losses likely depend on the physical conditions, we may choose to vary $f_{\rm th}$ with the relevant, local properties of the gas.
By considering the ratio of the cooling time to the sound crossing time across a resolution element, \citet{DallaVecchia2012Winds} derive the maximum density for which the thermal feedback can be efficient (their equation 18),
\begin{equation}
n_{{\rm H},t_{\rm c}} \sim 10~{\rm cm}^{-3} \left (\frac{T}{10^{7.5}\,{\rm K}}\right )^{3/2} \left (\frac{m_{\rm g}}{10^6\,{{\rm M}_\odot}}\right )^{-1/2},
\label{eq:nhtc}
\end{equation}
where $T> \Delta T$ is the temperature after the energy injection and we use $\Delta T = 10^{7.5}\,{\rm K}$.
This expression assumes that the radiative cooling rate is dominated by free-free emission and will thus significantly overestimate the
value of $n_{{\rm H},t_{\rm c}}$ when line cooling dominates, i.e.\ for $T \ll 10^7\,{\rm K}$. In our simulations some stars do, in fact, form in gas that far exceeds the critical value $n_{{\rm H},t_{\rm c}}$, particularly in massive galaxies. Although the density of the gas in which the stars inject their energy will generally be lower than that of the gas from which the star particle formed, since the star particles move relative to the gas during the $3\times 10^7\,{\rm yr}$ delay between star formation and feedback, this does mean that for stars forming at high gas densities
the radiative losses may well exceed those that would occur in a simulation that has the resolution and the physics required to resolve the small-scale structure of the ISM. As we calibrate the total amount of energy that is injected per unit stellar mass to achieve a good match to the observed GSMF, this implies that we may overestimate the required amount of feedback energy. At the high-mass end AGN feedback controls the efficiency of galaxy formation in our simulations. If the radiative losses from stellar feedback are overestimated, then this could potentially cause us to overestimate the required efficiency of AGN feedback.
The critical density, $n_{{\rm H},t_{\rm c}}$, increases with the numerical resolution, but also with the temperature jump, $\Delta T$. We could therefore reduce the initial thermal losses by increasing $\Delta T$. However, for a fixed amount of energy per unit stellar mass, i.e.\ for a fixed value of $f_{\rm th}$, the probability that a particular star particle generates feedback is inversely proportional to $\Delta T$. \citet{DallaVecchia2012Winds} show that, for the case of equal mass particles, the expectation value for the number of heated gas particles per star particle is (their equation 8)
\begin{equation}
\left < N_{\rm heat}\right > \approx 1.3 f_{\rm th} \left (\frac{\Delta T}{10^{7.5}\,{\rm K}}\right )^{-1}
\end{equation}
for our Chabrier IMF and only accounting for supernova energy (assuming that supernovae associated with stars in the range 6-100 ${{\rm M}_\odot}$ each yield $10^{51}\,{\rm erg}$). Hence, using $\Delta T \gg 10^{7.5}\,{\rm K}$ or $f_{\rm th} \ll 1$ would imply that most star particles do not inject any energy from core collapse supernovae into their surroundings, which may lead to poor sampling of the feedback cycle. We therefore keep the temperature jump set to $\Delta T = 10^{7.5}\,{\rm K}$. Although the stochastic implementation enables efficient thermal feedback without the need to turn off cooling, the thermal losses are unlikely to be converged with numerical resolution for simulations such as EAGLE. Hence, recalibration of $f_{\rm th}$ may be necessary when the resolution is changed.
\subsubsection{Dependence on local gas properties}
\label{sec:calibration}
We expect the true thermal losses in the ISM to increase when the metallicity becomes sufficiently high for metal-line cooling to become important. For temperatures of $10^5\,{\rm K} < T < 10^7{\rm K}$ this happens when $Z \ga 10^{-1}\,{\rm Z}_\odot$ \citep[e.g.][]{Wiersma2009Cooling}. Although the exact dependence on metallicity cannot be predicted without full knowledge of the physical conditions in the ISM, we can capture the expected, qualitative transition from cooling losses dominated by H and He to losses dominated by metals by making $f_{\rm th}$ a function of metallicity,
\begin{equation}
f_{\rm th} = f_{\rm th,min} + \frac{f_{\rm th,max} - f_{\rm th,min}}
{1 + \left (\frac{Z}{0.1{\rm Z}_\odot}\right )^{n_Z}},
\label{eq:f(Z)}
\end{equation}
where ${\rm Z}_\odot = 0.0127$ is the solar metallicity and $n_Z>0$. Note that $f_{\rm th}$ asymptotes to $f_{\rm th,max}$ and $f_{\rm th,min}$ for $Z \ll 0.1{\rm Z}_\odot$ and $Z\gg 0.1{\rm Z}_\odot$, respectively.
Since metallicity decreases with redshift at fixed stellar mass, this physically motivated metallicity dependence tends to make feedback relatively more efficient at high redshift.
As we show in \citet{Crain2014EagleModels}, this leads to good agreement with the observed, present-day GSMF. In fact, \citet{Crain2014EagleModels} show that using a constant $f_{\rm th} =1$ appears to yield even better agreement with the low-redshift mass function, but we keep the metallicity dependence because it is physically motivated: we do expect larger radiative losses for $Z \gg 0.1Z_\odot$ than for $Z \ll 0.1Z_\odot$. If we were only interested in the GSMF, then equation (\ref{eq:f(Z)}) (or $f_{\rm th} =1$) would suffice. However, we find that pure metallicity dependence results in galaxies that are too compact, which indicates that the feedback is too inefficient at high gas densities. As discussed above, this is not unexpected given the resolution of our simulations. Indeed, we found that increasing the resolution reduces the problem.
We therefore found it desirable to compensate for the excessive initial, thermal losses at high densities by adding a density dependence to $f_{\rm th}$:
\begin{equation}
f_{\rm th} = f_{\rm th,min} + \frac{f_{\rm th,max} - f_{\rm th,min}}
{1 + \left (\frac{Z}{0.1{\rm Z}_\odot}\right )^{n_Z} \left (\frac{n_{\rm H,birth}}{n_{{\rm H},0}}\right )^{-n_n}},
\label{eq:f(Z,n)}
\end{equation}
where $n_{\rm H,birth}$ is the density inherited by the star particle, i.e.\ the density of its parent gas particle at the time it was converted into a stellar particle. Hence, $f_{\rm th}$ increases with density at fixed metallicity, while still respecting the original asymptotic values. We use $n_Z = n_n = 2/\ln10$. The seemingly unnatural value $2/\ln10 \approx 0.87$ of the exponent is a leftover from an equivalent, but more complicated expression that was originally used in the code. Using the round number 1 instead of 0.87 would have worked equally well. We use $n_{{\rm H},0} = 0.67~{\rm cm}^{-3}$, a value that was chosen after comparing a few test simulations to the observed present-day GSMF and galaxy sizes. The higher resolution simulation Recal-L025N0752 instead uses $n_{{\rm H},0} = 0.25~{\rm cm}^{-3}$ and a power-law exponent for the density term of $-1/\ln10$ rather than $-2/\ln10$ (see Table~\ref{tbl:subgridpars}), which we found gives better agreement with the GSMF. Note that a density dependence of $f_{\rm th}$ may also have a physical interpretation. For example, higher mean densities on $10^2-10^3\,{\rm pc}$ scales may result in more clustered star formation, which may reduce thermal losses. However, we stress that our primary motivation was to counteract the excessive thermal losses in the high-density ISM that can be attributed to our limited resolution.
We use the asymptotic values $f_{\rm th,max}=3$ and $f_{\rm th,min}=0.3$, where the high asymptote $f_{\rm th,max}$ is reached at low metallicity and high density, and vice versa for the low asymptote. As discussed in \citet{Crain2014EagleModels}, where we present variations on the reference model, the choice of the high asymptote is the more important one. Using a value of $f_{\rm th,max}$ greater than unity enables us to reproduce the GSMF down to lower masses.
Values of $f_{\rm th}$ greater than unity can be motivated on physical grounds by appealing to other sources of energy than supernovae, e.g.\ stellar winds, radiation pressure, or cosmic rays, or if supernovae yield more energy per unit mass than assumed here (e.g.\ in case of a top-heavy IMF). However, we believe that a more appropriate motivation is again the need to compensate for the finite numerical resolution. Galaxies containing few star particles tend to have too high stellar fractions \citep[e.g.][]{Haas2013OwlsI}, which can be understood as follows. The first generations of stars can only form once the halo is resolved with a sufficient number of particles to sample the high-density gas that is eligible to form stars. We do not have sufficient resolution to resolve the smallest galaxies that are expected to form in the real Universe. Hence, the progenitors of the galaxies in the simulations started forming stars, and hence driving winds, too late. As a consequence, our galaxies start with too high gas fractions and initially form stars too efficiently. As the galaxies grow substantially larger than our resolution limit, this initial error becomes progressively less important. Using a higher value of $f_{\rm th,max}$ counteracts this sampling effect as it makes the feedback from the first generations of stars that form more efficient.
The mean and median values of $f_{\rm th}$ that were used for the feedback from the stars present at $z=0.1$ in Ref-L100N1504 are 1.06 and 0.70, respectively. For Recal-L025N0752 these values are 1.07 and 0.93. Hence, averaged over the entire simulation, the total amount of energy is similar to that expected from supernovae alone. A more detailed discussion of the effects of changing the functional form of $f_{\rm th}$ is presented in \citet{Crain2014EagleModels}. In that work we also present models in which $f_{\rm th}$ is constant or depends on halo mass or dark matter velocity dispersion.
\subsection{Black holes and feedback from AGN}
\label{sec:BHs}
In our simulations feedback from accreting, supermassive black holes (BHs) quenches star formation in massive galaxies, shapes the gas profiles in the inner parts of their host haloes, and regulates the growth of the BHs.
Models often make a distinction between ``quasar-'' and ``radio-mode'' BH feedback \citep[e.g.][]{Croton2006SA,Bower2006SA,Sijacki2007AGN}, where the former occurs when the BH is accreting efficiently and comes in the form of a hot, nuclear wind, while the radio mode operates when the accretion rate is low compared to the Eddington rate and the energy is injected in the form of relativistic jets. Because cosmological simulations lack the resolution to properly distinguish these two feedback modes and because we want to limit the number of feedback channels to the minimum required to match the observations of interest, we choose to implement only a single mode of AGN feedback with a fixed efficiency. The energy is injected thermally at the location of the BH at a rate that is proportional to the gas accretion rate. Our implementation may therefore be closest to the process referred to as quasar-mode feedback. For OWLS\ we found that this method led to excellent agreement with both optical and detailed X-ray observations of groups and clusters \citep{McCarthy2010AGN,McCarthy2011AGN,LeBrun2014CosmoOWLS}.
Our implementation consists of two parts: i) prescriptions for seeding low-mass galaxies with central BHs and for their growth via gas accretion and merging (we neglect any growth by accretion of stars and dark matter); ii) a prescription for the injection of feedback energy. Our method for the growth of BHs is based on the one introduced by \citet{Springel2005AGN} and modified by \citet{Booth2009AGN} and \citet{Rosas2013BHs}, while our method for AGN feedback is close to the one described in \citet{Booth2009AGN}. Below we summarize the main ingredients and discuss the changes to the methods that we made for EAGLE.
\subsubsection{BH seeds}
The BHs ending up in galactic centres may have originated from the direct collapse of (the inner parts of) metal-free dwarf galaxies, from the remnants of very massive, metal-free stars, or from runaway collisions of stars and/or stellar mass BHs (see e.g.\ \citealt{Kocsis2013BHReview} for a recent review). As none of these processes can be resolved in our simulations, we follow \citet{Springel2005AGN} and place BH seeds at the centre of every halo with total mass greater than $10^{10}\,{{\rm M}_\odot}/h$ that does not already contain a BH. For this purpose, we regularly run the friends-of-friends (FoF) finder with linking length 0.2 on the dark matter distribution. This is done at times spaced logarithmically in the expansion factor $a$ such that $\Delta a = 0.005a$.
The gas particle with the highest density is converted into a collisionless BH particle with subgrid BH mass $m_{\rm BH} = 10^5\,{{\rm M}_\odot}/h$. The use of a subgrid BH mass is necessary because the seed BH mass is small compared with the particle mass, at least for our default resolution. Calculations of BH properties such as its accretion rate are functions of $m_{\rm BH}$, whereas gravitational interactions are computed using the BH particle mass. When the subgrid BH mass exceeds the particle mass, it is allowed to stochastically accrete neighbouring SPH particles such that BH particle and subgrid masses grow in step.
Since the simulations cannot model the dynamical friction acting on BHs with masses $\la m_{\rm g}$, we force BHs with mass $< 100 m_{\rm g}$
to migrate towards the position of the minimum of the gravitational potential in the halo. At each time step the BH is moved to the location of the particle that has the lowest gravitational potential of all the neighbouring particles whose velocity relative to the BH is smaller than $0.25 c_{\rm s}$, where $c_{\rm s}$ is the speed of sound, and whose distance is smaller than three gravitational softening lengths. These two conditions prevent BHs in gas poor haloes from jumping to nearby satellites.
\subsubsection{Gas accretion}
\label{sec:bh_accretion}
The rate at which BHs accrete gas depends on the mass of the BH, the local density and temperature, the velocity of the BH relative to the ambient gas, and the angular momentum of the gas with respect to the BH. Specifically, the gas accretion rate, $\dot{m}_{\rm accr}$, is given by the minimum of the Eddington rate,
\begin{equation}
\dot{m}_{\rm Edd} = \frac{4\pi G m_{\rm BH} m_{\rm p}}{\epsilon_{\rm r} \sigma_{\rm T} c},
\end{equation}
and
\begin{equation}
\dot{m}_{\rm accr} = \dot{m}_{\rm Bondi} \times \min\left (C_{\rm visc}^{-1}(c_{\rm s}/V_\phi)^3,1 \right ),
\label{eq:mdotaccr}
\end{equation}
where $\dot{m}_{\rm Bondi}$ is the Bondi-Hoyle (\citeyear{Bondi1944Accr}) rate for spherically symmetric accretion,
\begin{equation}
\dot{m}_{\rm Bondi} = \frac{4\pi G^2 m_{\rm BH}^2 \rho}{(c_{\rm s}^2 + v^2)^{3/2}}.
\end{equation}
Here $m_{\rm p}$ is the proton mass, $\sigma_{\rm T}$ the Thomson cross section, $c$ the speed of light, $\epsilon_{\rm r}=0.1$ the radiative efficiency of the accretion disc, and $v$ the relative velocity of the BH and the gas. Finally, $V_\phi$ is the rotation speed of the gas around the BH computed using equation (16) of \citet{Rosas2013BHs} and $C_{\rm visc}$ is a free parameter related to the viscosity of the (subgrid) accretion disc.
The mass growth rate of the BH is given by \begin{equation} \dot{m}_{\rm BH} = (1-\epsilon_{\rm r}) \dot{m}_{\rm accr}. \end{equation}
The factor $(c_{\rm s}/V_\phi)^3/C_{\rm visc}$ by which the Bondi rate is multiplied in equation~(\ref{eq:mdotaccr}) is equivalent to the ratio of the Bondi and the viscous time scales (see \citealt[][]{Rosas2013BHs}). We set $C_{\rm visc} = 2\pi$ for Ref-L100N1504, but increase the value of $C_{\rm visc}$ by a factor $10^3$ for the recalibrated high-resolution model, Recal-L025N0752, and by a factor $10^2$ for AGNdT9-L050N0752 (see Table~\ref{tbl:subgridpars}). Since the critical ratio of $V_\phi/c_{\rm s}$ above which angular momentum is assumed to reduce the accretion rate scales with $C_{\rm visc}^{-1/3}$, angular momentum is relatively more important in the recalibrated simulations, delaying the onset of quenching by AGN to larger BH masses.
As demonstrated by \citet{Rosas2013BHs}, the results are only weakly dependent on $C_{\rm visc}$ because the ratio of $V_\phi/c_{\rm s}$ above which the accretion rate is suppressed, which scales as $C_{\rm visc}^{-1/3}$, is more important than the actual suppression factor, which scales as $C_{\rm visc}$.
Our prescription for gas accretion differs from previous work in two respects. First, the Bondi rate is not multiplied by a large, ad-hoc factor, $\alpha$. \citet{Springel2005AGN} used $\alpha=100$ while OWLS\ and \citealt{Rosas2013BHs} used a density dependent factor that asymptoted to unity below the star formation threshold. Although the use of $\alpha$ can be justified if the simulations underestimate the gas density or overestimate the temperature near the Bondi radius, the correct value cannot be predicted by the simulations. We found that at the resolution of EAGLE, we do not need to boost the Bondi-Hoyle rate for the BH growth to become self-regulated. Hence, we were able to reduce the number of free parameters by eliminating $\alpha$.
Second, we use the heuristic correction of \citet{Rosas2013BHs} to account for the fact that the accretion rate will be lower for gas with more angular momentum (because the accretion is generally not spherically symmetric as assumed in the Bondi model, but proceeds through an accretion disc).
\subsubsection{BH mergers}
BHs are merged if they are separated by a distance that is smaller than both the smoothing kernel of the BH, $h_{\rm BH}$, and three gravitational softening lengths, and if their relative velocity is smaller than the circular velocity at the distance $h_{\rm BH}$, $v_{\rm rel} < \sqrt{G m_{\rm BH}/h_{\rm BH}}$, where $h_{\rm BH}$ and $m_{\rm BH}$ are, respectively, the smoothing length and subgrid mass of the most massive BH in the pair. The limit on the allowed relative velocity prevents BHs from merging during the initial stages of galaxy mergers.
\subsubsection{AGN feedback}
\label{sec:AGNfeedback}
AGN feedback is implemented thermally and stochastically, in a manner analogous to energy feedback from star formation. The energy injection rate is $\epsilon_{\rm f} \epsilon_{\rm r} \dot{m}_{\rm accr} c^2$, where $\epsilon_{\rm f} = 0.15$ is the fraction of the radiated energy that is coupled to the ISM. As was the case for the stellar feedback efficiency, $f_{\rm th}$, the value of $\epsilon_{\rm f}$ must be chosen by calibrating to observations, in this case the normalisation of the relation between BH mass and stellar mass. As demonstrated and explained by \citet[][see also \citealt{Booth2009AGN}]{Booth2010DMHaloesBHs}, the value of $\epsilon_{\rm f}$ \emph{only} affects the BH masses, which are inversely proportional to $\epsilon_{\rm f}$. In particular, the outflow rate generated by the AGN and hence also the factor by which the star formation is reduced, are highly insensitive to $\epsilon_{\rm f}$ provided it is nonzero. This can be explained by self-regulation: the BH accretion rate adjusts until the rate at which energy is injected is sufficient for outflows to balance inflows.
We use the same value for the AGN efficiency as in OWLS, $\epsilon_{\rm f}=0.15$ and $\epsilon_{\rm r}=0.1$, which implies that a fraction $\epsilon_{\rm f}\epsilon_{\rm r}=0.015$ of the accreted rest mass energy is returned to the local ISM. As was the case for stellar feedback, the required value will depend on the radiative losses in the ISM, which may depend on the resolution and the precise manner in which the energy is injected. We do not implement a dependence on metallicity, because metals are not expected to dominate the radiative losses at the high temperatures associated with AGN feedback. As shown in Figure~\ref{fig:bh}, a constant value of $\epsilon_{\rm f}=0.15$ yields broad agreement with observations of the relation between BH mass and stellar mass.
Each BH carries a ``reservoir'' of feedback energy, $E_{\rm BH}$. After each time step $\Delta t$, we add $\epsilon_{\rm f} \epsilon_{\rm r} \dot{m}_{\rm accr} c^2 \Delta t$ to this reservoir. If the BH has stored sufficient energy to heat at least $n_{\rm heat}$ particles of mass $m_{\rm g}$, then the BH is allowed to stochastically heat each of its SPH neighbours by increasing their temperature by $\Delta T_{\rm AGN}$. For each neighbour the heating probability is
\begin{equation}
P = \frac{E_{\rm BH} }{ \Delta\epsilon_{\rm AGN} N_{\rm ngb} \left <m_{\rm g}\right > },
\end{equation}
where $\Delta\epsilon_{\rm AGN}$ is the change in internal energy per unit mass corresponding to the temperature increase, $\Delta T_{\rm AGN}$ (we convert the parameter $\Delta T_{\rm AGN}$ into $\Delta\epsilon_{\rm AGN}$ assuming a fully ionised gas with primordial composition), $N_{\rm ngb}$ is the number of gas neighbours of the BH and $\left <m_{\rm g}\right >$ is their mean mass. We then reduce $E_{\rm BH}$ by the expectation value for the injected energy. We use $n_{\rm heat}=1$ and limit the time step of the BHs such that we expect\footnote{Because the expected probability is based on the accretion rate in the previous time step, limiting the BH time step does not guarantee that $P<0.3$. If the probability exceeds 0.3, then we limit it to 0.3 and store the unused energy in $E_{\rm BH}$.} $P<0.3$ (see \S\ref{sec:injection}).
The most important parameter for the AGN feedback is the temperature increase $\Delta T_{\rm AGN}$. Larger values will make individual feedback events more energetic, generally resulting in smaller radiative losses in the ISM. However, larger values will also make the feedback more intermittent. We set $\Delta T_{\rm AGN} = 10^{8.5}\,{\rm K}$ in the L100N1504 reference model, but use $10^9\,{\rm K}$ for our recalibrated high-resolution model Recal-L025N0752 and model AGNdT9-L050N0752 (see Table~\ref{tbl:subgridpars}). These temperatures exceed the value of $10^8\,{\rm K}$ used in OWLS\ and the $\Delta T = 10^{7.5}\,{\rm K}$ that we use for stellar feedback. As can be seen from equation~(\ref{eq:nhtc}), the critical density above which the feedback energy is expected to be radiated away increases with the value of $\Delta T$. Because the density of the ambient gas around the BH tends to increase with resolution, we found that we need to increase $\Delta T$ when increasing the resolution. Similarly, because the gas density around the BH often reaches values that are much higher than is typical for star-forming gas, we require higher temperature jumps for AGN feedback than for stellar feedback.
\section{Comparison with observables considered during the calibration of the feedback}
\label{sec:cal_obs}
In this section we will compare the main EAGLE\ simulations to $z\sim 0$ observations of the GSMF, the related stellar mass - halo mass relation, galaxy sizes, and the relation between BH mass and stellar mass. Since these observables were considered during the calibration of the subgrid models for feedback, we cannot consider the EAGLE\ results reported in this section to be ``predictions''. However, note that we had no control over the slope of the $M_{\rm BH}-M_\ast$ relation and that galaxy sizes were only used to rule out strongly discrepant models (i.e.\ models without a density dependence of the energy feedback from star formation).
\subsection{The galaxy stellar mass function}
\label{sec:gsmf}
\begin{figure}
\resizebox{\colwidth}{!}{\includegraphics{mf.pdf}}
\caption{The galaxy stellar mass function at $z=0.1$ for the EAGLE\ simulations Ref-L100N1504 (blue), AGNdT9-L050N0752 (red), and Recal-L025N0752 (green-blue). The curves switch from solid to dashed at the high-mass end when there are fewer than 10 objects per (0.2~dex) stellar mass bin. At the low-mass end the curves become dotted when the stellar mass falls below that corresponding to 100 baryonic particles. Data points show measurements with $1\sigma$ error bars from the GAMA survey (open circles; $z<0.06$; \citealt{Baldry2012GSMF}) and from SDSS (filled circles; $z\sim 0.07$; \citealt{Li2009GSMF}). The high-resolution model Recal-L025N0752 is noisier because of its small box size. The intermediate-resolution models slightly underestimate the galaxy number density at the knee of the mass function and slightly overestimate the abundance at $M_\ast \sim 10^{8.5}\,{{\rm M}_\odot}$. The galaxy number density agrees with the data to $\la 0.2$ dex.}
\label{fig:gsmf}
\end{figure}
Figure~\ref{fig:gsmf} shows the $z=0.1$ galaxy stellar mass function (GSMF) from EAGLE. The dark blue curve shows Ref-L100N1504, the green curve shows the high-resolution simulation Recal-L025N0752, and the red curve shows AGNdT9-L050N0752. Recall that AGNdT9-L050N0752 employs a higher heating temperature for AGN feedback than the reference model, which makes the feedback more efficient. While this is unimportant for the GSMF, we will see in \S\ref{sec:groups} that it offers a significant improvement for the intracluster medium. At the high-mass end the curves switch from a solid to a dashed line style where there are fewer than 10 objects per (0.2~dex) stellar mass bin.
At the low-mass end the curves become dotted when the stellar mass falls below that corresponding to 100 baryonic particles, where sampling effects associated with the limited resolution become important, as can be seen by comparing the intermediate- and high-resolution simulations.
The GSMF of the high-resolution simulation Recal-L025N0752 is noisier because the box size is too small to provide a representative sample. Note that the main problem is not Poisson noise due to the small number of objects per bin, but the small number of large-scale modes that modulate the local number density of galaxies of various masses.
Indeed, Fig.~\ref{fig:gsmf_conv} shows that the GSMF of Recal-L025N0752 has the same wiggles as that of Ref-L025N0376, which uses the same box size and, apart from the change in resolution, the same initial conditions. The wiggles that are present for Ref-L025N0376 are absent for model Ref-L100N1504, even though these two simulations use identical resolutions and (subgrid) parameter values. This confirms that the wiggles in the GSMF of Recal-L025N0752 are caused by the small size of its simulation volume. We will therefore focus on the larger volume simulations when comparing the simulated and observed GSMFs.
The simulation results are compared with observations from the Galaxy And Mass Assembly (GAMA) survey (\citealt{Baldry2012GSMF}; open circles) and from SDSS (\citealt{Li2009GSMF}; filled circles).
For the intermediate-resolution simulations the galaxy number densities agree with the observations to $\la 0.2$~dex over the full mass range for which the resolution and box size are adequate, i.e.\ from $2\times 10^8\,{{\rm M}_\odot}$ to over $10^{11}\,{{\rm M}_\odot}$ (slightly below $10^{11}\,{{\rm M}_\odot}$ for Recal-L025N0752). The observed shape of the GSMF is thus reproduced well.
At fixed number density, the differences in stellar mass between the simulations and observations are smaller than 0.3~dex for Ref-L100N1504 and AGNdT9-L050N0752. Given that even for a fixed IMF, uncertainties in the stellar evolution models used to infer stellar masses are $\sim 0.3$~dex \citep[e.g.][]{Conroy2009SPSSUncertainty,Behroozi2010Uncertainties,Pforr2012SPSSUncertainty,Mitchell2013MassUncertainty}, there is perhaps little point in trying to improve the agreement between the models and the data further.
The subgrid models for energy feedback from star formation and for BH accretion have been calibrated to make the simulated GSMF fit the observed one, so the excellent agreement with the data cannot be considered a successful prediction. However, success was by no means guaranteed given that the computational expense of hydrodynamical simulations severely limits the number of test runs that can be performed and, more importantly, because the freedom built into the model is rather limited. For example, while the mass scale above which AGN feedback becomes dominant is sensitive to the parameter $C_{\rm visc}$ of the subgrid model for BH accretion (see equation \ref{eq:mdotaccr} in \S\ref{sec:bh_accretion}), the efficiency of the AGN feedback was calibrated to the observed relation between BH mass and stellar mass and does not affect the shape of the GSMF \citep{Booth2009AGN,Booth2010DMHaloesBHs}.
\begin{figure*}
\resizebox{\colwidth}{!}{\includegraphics{mf_sam_comp.pdf}}
\resizebox{\colwidth}{!}{\includegraphics{mf_hydro_comp.pdf}}
\caption{Comparisons of the GSMF from EAGLE's Ref-L100N1504 with the semi-analytic models of \citet{Gonzalez-Perez2014Galform}, \citet{Henriques2013SAM}, and \citet{Porter2014SAM} (left panel) and with the large hydrodynamical simulations of \citet{Oppenheimer2010GSMF}, \citet{Puchwein2013GSMF}, the Illustris simulation \citep[][data taken from \citealt{Genel2014IllustrisEvolution}]{Vogelsberger2014Illustris}, and the MassiveBlack-II simulation \citep{Khandai2014MassiveBlack} (right panel). All models are for a Chabrier IMF (\citealt{Gonzalez-Perez2014Galform} and \citealt{Khandai2014MassiveBlack} have been converted from Kennicutt and Salpeter IMFs, respectively). The EAGLE\ curve is dotted when galaxies contain fewer than 100 stellar particles and dashed when there are fewer than 10 galaxies per stellar mass bin.
Except for \citet{Oppenheimer2010GSMF}, all simulations include AGN feedback. Apart from MassiveBlack-II, all models were calibrated to the data (the Galform semi-analytic model of \citealt{Gonzalez-Perez2014Galform} was calibrated to fit the K-band galaxy luminosity function). The agreement with the data is relatively good for both EAGLE\ and the semi-analytic models, but EAGLE\ fits the data substantially better than the other hydrodynamical simulations do.}
\label{fig:gsmf_other}
\end{figure*}
Figure~\ref{fig:gsmf_other} shows that the level of correspondence between the data and EAGLE\ is close to that attained for semi-analytic models (left panel) and is unprecedented for large, hydrodynamical simulations (right panel). As can be seen from the right panel, even though \citet{Oppenheimer2010GSMF}, \citet{Puchwein2013GSMF}, and Illustris \citep{Vogelsberger2014IllustrisNature,Genel2014IllustrisEvolution} all adjusted their subgrid feedback models to try to match the data, the fits to the data are substantially less good than for EAGLE. In particular, their models all produce mass functions that are too steep below the ``knee'' of the Schechter function and too shallow for larger masses. It is worth noting that each of these three groups implemented the feedback from star formation kinetically, scaled the wind velocity with the velocity dispersion of the dark matter, determined the dependence of the wind mass loading on the dark matter velocity dispersion by assuming a constant wind energy, and temporarily turned off the hydrodynamical forces on wind particles to allow them to escape the galaxies. This contrasts with EAGLE, where the feedback was implemented thermally rather than kinetically, the feedback energy varied with local gas properties, and the hydrodynamical forces were never turned off.
Hence, contrary to the other models shown, EAGLE's subgrid model does not impose any particular wind velocity or mass loading or any dependence on dark matter or halo properties. The injected energy does depend on the local metallicity and gas density, but the relation between the outflow properties and the energy injected at the star formation site is an outcome of the simulation. \citet{Crain2014EagleModels} will show that while varying the feedback energy with local gas properties is necessary to obtain reasonable galaxy sizes, the $z\sim 0$ GSMF is actually also reproduced by the EAGLE\ model that injects a constant energy per unit stellar mass (equal to the energy from supernovae) without any calibration.
While the excellent fit to the low-$z$ GSMF is encouraging, the success of the model can only be judged by comparing to a wide range of observables and redshifts, particularly those that were not considered during the calibration. We will consider a diverse selection of observables in \S\ref{sec:otherobs} and will investigate their evolution in \citet{Furlong2014EagleEvolution} and other future papers.
\subsubsection{Effect of the choice of aperture}
\label{sec:aperture}
\begin{figure}
\resizebox{\colwidth}{!}{\includegraphics{mf_ref100_only_apertures.pdf}}
\caption{The effect of the choice of aperture on the GSMF. Curves show the $z=0.1$ GSMF from Ref-L100N1504 for different 3-D apertures: radii of 30, 50, and 100 proper kiloparsec, a 2-D Petrosian aperture, and no aperture at all. In all cases only stellar mass bound to a subhalo is considered. The simulation curves are dotted where galaxies contain fewer than 100 stellar particles and dashed where there are fewer than 10 galaxies per stellar mass bin. Data points indicate observations. The \citet{Li2009GSMF} and \citet{Bernardi2013GSMF} data points are both for SDSS, but use Petrosian magnitudes and integrals of S\'ersic plus exponential fits, respectively. The \citet{Baldry2012GSMF} data points are for the GAMA survey and use integrals of single S\'ersic fits. The choice of aperture is important for $M_\ast > 10^{11}\,{{\rm M}_\odot}$, both for the simulation and the observations.}
\label{fig:gsmf_aperture}
\end{figure}
For the simulations we chose to define a galaxy's stellar mass as the sum of the mass of the stars that are part of a gravitationally bound subhalo and that are contained within a 3-D aperture of radius 30 proper kiloparsec (see \S\ref{sec:simulations}). Figure~\ref{fig:gsmf_aperture} shows the effect of the choice of aperture for Ref-L100N1504. For $M_\ast < 10^{11}\,{{\rm M}_\odot}$ the results are insensitive to the aperture, provided it is $\ga 30$ pkpc. However, for $M_\ast > 10^{11}\,{{\rm M}_\odot}$ the aperture does become important, with larger apertures giving larger masses.
The same is true for the observations, as can be seen by comparing the data from \citet{Li2009GSMF} with the re-analysis of SDSS data
by \citet{Bernardi2013GSMF} (open triangles in Fig.~\ref{fig:gsmf_aperture}). \citet{Baldry2012GSMF} and \citet{Li2009GSMF} are in good agreement, but \citet{Bernardi2013GSMF} find a much shallower bright-end slope than previous analyses. For $M_\ast > 10^{11}\,{{\rm M}_\odot}$ \citet{Bernardi2013GSMF} attribute substantially more mass to galaxies than \citet{Li2009GSMF} and \citet{Baldry2012GSMF}.
Part of the difference is due to the assumed mass-to-light ratios (even though all studies assume a Chabrier IMF) and the way in which the background is subtracted (see e.g.\ \citealt{Bernardi2013GSMF} and \citealt{Kravtsov2014GalaxyHalo} for discussion). Most of the difference between \citet{Li2009GSMF} and \citet{Bernardi2013GSMF}
can probably be attributed to the way in which a galaxy's light is measured. \citet{Li2009GSMF} integrate the light within a 2-D aperture of size twice the Petrosian radius, defined to be the radius at which the mean local surface brightness is 0.2 times the mean internal surface brightness. \citet{Bernardi2013GSMF} on the other hand, estimate the total amount of light by integrating S\'ersic plus exponential profile fits. Hence, the \citet{Bernardi2013GSMF} mass function potentially includes intracluster light and the discrepancy between different authors is related to the fact that it is unclear where cD galaxies end. \citet{Baldry2012GSMF} integrate single S\'ersic fits to the light profiles, which we would expect includes less intracluster light than the S\'ersic plus exponential fits of \citet{Bernardi2013GSMF} but more than the Petrosian apertures of \citet{Li2009GSMF}. However, \citet{Bernardi2013GSMF} find that the high-mass end of the \citet{Baldry2012GSMF} mass function is affected by their redshift cut ($z<0.06$).
We believe the \citet{Baldry2012GSMF} and \citet{Li2009GSMF} data to be the most suitable for comparison to our results, since our definition of a galaxy excludes intracluster light. For \citet{Li2009GSMF} this is confirmed by our finding that a 3-D aperture of 30 pkpc gives nearly identical results to a 2-D Petrosian cut, as can be seen from Figure~\ref{fig:gsmf_aperture}.
Thus, for masses $> 10^{11}\,{{\rm M}_\odot}$ comparisons of the GSMF with observations would benefit from mimicking the particular way in which the mass is estimated for real data. This would, however, have to be done separately for each survey. For our present purposes this is unnecessary, also because our simulation volume is in any case too small to study the GSMF at masses $\gg 10^{11}\,{{\rm M}_\odot}$.
\subsubsection{Numerical convergence}
\label{sec:convergence}
\begin{figure*}
\resizebox{\colwidth}{!}{\includegraphics{mf_strong_convergence.pdf}}
\resizebox{\colwidth}{!}{\includegraphics{mf_weak_convergence.pdf}}
\caption{Strong (left panel) and weak (right panel) tests of the convergence of the GSMF with numerical resolution. Models L025N0752 have a better mass and spatial resolution than L025N0376 by factors of 8 and 2, respectively. The strong convergence test compares models with identical subgrid parameter values, while the weak convergence test compares the original, intermediate-resolution model Ref-L025N0376 with a high-resolution model Recal-L025N0752 for which the parameters of the subgrid models for feedback from star formation and for gas accretion onto BHs were recalibrated in order to reproduce the observed GSMF. For comparison, the thin curves in the left panel show the strong convergence test for the galaxy formation model used for the Illustris simulation as reported by \citet{Vogelsberger2013IllustrisModel}. The EAGLE\ curves are dotted where galaxies contain fewer than 100 stellar particles and dashed where there are fewer than 10 galaxies per stellar mass bin.}
\label{fig:gsmf_conv}
\end{figure*}
The left panel of Figure~\ref{fig:gsmf_conv} compares the GSMFs for model Ref-L025N0376, which has the same resolution as the largest EAGLE\ volume Ref-L100N1504, and the higher-resolution model Ref-L025N0752. The two Ref-L025 simulations use identical subgrid parameters, but the mass and spatial resolution differ by factors of 8 and 2, respectively. In \S\ref{sec:conv_discussion} we termed a comparison between models with identical parameters a ``strong convergence test''. Below $10^9\,{{\rm M}_\odot}$ the mass function is substantially flatter in the high-resolution model. However, at $M_\ast \sim 10^9\,{{\rm M}_\odot}$ its GSMF is up to 0.4~dex higher than for the fiducial resolution, leading to disagreement with the data. The largest discrepancy is the stellar mass corresponding to a number density of $\sim 2\times 10^{-2}\,{\rm cMpc}^{-3}$, which is about an order of magnitude higher than observed.
The thin curves in Figure~\ref{fig:gsmf_conv} show the strong convergence test of \citet{Vogelsberger2013IllustrisModel} using the galaxy formation model that was also used for Illustris. Clearly, the strong convergence is similarly poor. This is somewhat surprising, since Illustris uses a subgrid model for feedback from star formation that was designed to give good strong convergence. In particular, the parameters of the subgrid wind model vary with the velocity dispersion of the dark matter rather than with the properties of the gas and hydrodynamical interactions between the wind and the ISM are not modelled.
That the strong convergence is not particularly good for EAGLE\ is unsurprising for the reasons discussed in \S\ref{sec:conv_discussion} and \S\ref{sec:snii}. For $M_\ast < 2\times 10^8\,{{\rm M}_\odot}$ galaxies in Ref-L025N0376 contain fewer than 100 star particles, which is insufficient to properly sample the feedback from star formation in the context of EAGLE's subgrid model. Because the feedback can be modelled down to lower masses in Ref-L025N0752, galaxies with $M_\ast \sim 10^9\,{{\rm M}_\odot}$ have had systematically different histories than galaxies of a similar mass in Ref-L025N0376. In addition, higher resolution enables the gas density distribution to be populated by particles up to higher densities, where our fiducial implementation of thermal feedback becomes inefficient (equation~\ref{eq:nhtc} in \S\ref{sec:snii}).
In \S\ref{sec:conv_discussion} we argued that hydrodynamical simulations such as EAGLE\ should recalibrate the efficiency of the subgrid feedback when the resolution is changed substantially. In general, keeping the subgrid parameters fixed does not imply that the physical model remains unchanged, since the energy, mass or intermittency associated with the feedback events changes. Moreover, the efficiency of the feedback cannot, in any case, be predicted from first principles, even if the convergence were perfect.
Recal-L025N0752 is our recalibrated high-resolution simulation. As detailed in \S\ref{sec:calibration} and Table~\ref{tbl:subgridpars}, the dependence of the feedback energy per unit stellar mass on the gas density is somewhat different between the different resolutions. However, the mean values of $f_{\rm th}$, which is equal to the expectation value of the amount of injected energy in units of the energy available from core collapse supernovae, are nearly identical: 1.06 at intermediate resolution (for stars formed at $z>0.1$ in Ref-L100N1504) and 1.07 at high resolution (for stars formed at $z>0.1$ in Recal-L025N0752). The asymptotic maximum of $f_{\rm th}$, reached at low metallicity and low gas density, is 3 in both cases. As detailed in \S\ref{sec:bh_accretion} and Table~\ref{tbl:subgridpars}, Recal-L025N0752 also uses a different value for the parameter that controls the importance of angular momentum in suppressing accretion onto BHs, making the accretion rate more sensitive to the angular momentum of the accreting gas. Without this change, AGN feedback would become important at too low masses. Finally, the high-resolution model uses a higher AGN feedback temperature, $\Delta T_{\rm AGN} = 10^9\,{\rm K}$ rather than $10^{8.5}\,{\rm K}$, which helps to suppress the increase in the cooling losses that would otherwise occur due to the higher gas densities that are resolved in the higher resolution model. Without this change the AGN feedback would be insufficiently effective.
The right panel of Figure~\ref{fig:gsmf_conv} shows a ``weak convergence test'', i.e.\ a comparison of the GSMFs of the calibrated intermediate resolution model Ref-L025N0376 and the recalibrated high-resolution model Recal-L025N0752. The two curves show some of the same bumps and wiggles, because the initial conditions used for the two simulations share the same large-scale modes. In the mass range for which galaxies in the intermediate-resolution model are resolved with more than 100 star particles ($M_\ast > 2\times 10^8\,{{\rm M}_\odot}$) the difference in the galaxy number density is smaller than 0.2~dex. We conclude that the weak convergence is good.
\subsection{The relation between stellar mass and halo mass}
\label{sec:eta}
\begin{figure*}
\resizebox{\colwidth}{!}{\includegraphics{eta_m200_2scat.pdf}}
\resizebox{\colwidth}{!}{\includegraphics{eta_mstar_2scat.pdf}}
\caption{The ratio of the stellar to halo mass, relative to the universal baryon fraction, as a function of halo mass (left panel) and stellar mass (right panel) for central galaxies. The simulation curves are dotted where there are fewer than 100 stellar particles per galaxy. Filled circles show individual objects where there are fewer than 10 objects per bin. The shaded regions show the $1\sigma$ scatter in the simulations. For clarity we only show the scatter in Recal-L025N0752 for $M_\ast < 10^{10}\,{{\rm M}_\odot}$ and in Ref-L100N1504 for $M_\ast > 10^{10}\,{{\rm M}_\odot}$. The EAGLE\ results agree with results inferred from observations through the technique of abundance matching (grey, solid curves; \citealt{Moster2013AbundMatching,Behroozi2013AbundMatching}). The small difference between EAGLE\ and the abundance matching in the location and height of the peak are consistent with EAGLE's small underestimate of the GSMF around the knee (see Fig.~\ref{fig:gsmf}).}
\label{fig:eta}
\end{figure*}
The GSMF can be thought of as a convolution between the mass function of dark matter haloes and a function describing the galaxy content of the haloes as a function of their mass.
The halo mass function can be predicted accurately when the cosmology is known, but the galaxy content of haloes is very sensitive to the baryonic processes involved in the formation of galaxies. As modelling galaxy formation is EAGLE's primary goal, it is of interest to compare the relation between stellar mass and halo mass in the simulations to the relation inferred from observations. Because the subgrid model for feedback was calibrated to fit the $z\sim 0$ GSMF, the relation between stellar and halo mass can hardly be considered a prediction. We therefore discuss this relation in this section, even though we did not calibrate the simulations to fit the relation inferred from observations.
Figure~\ref{fig:eta} shows the ``galaxy formation efficiency'', $(M_\ast / M_{200}) / (\Omega_{\rm b}/\Omega_{\rm m})$, for central galaxies as a function of either the mass of their host halo (left panel) or their stellar mass (right panel). Here the halo mass, $M_{200}$, is defined as the total mass contained within the virial radius $R_{200}$, defined to be the radius within which the mean internal density is 200 times the critical density, $3H^2/8\pi G$, centred on the dark matter particle of the corresponding FoF halo with the minimum gravitational potential (see \S\ref{sec:simulations}). If the baryon fraction in the halo were equal to the cosmic average of $\Omega_{\rm b}/\Omega_{\rm m} \approx 0.16$, then an efficiency of unity would indicate that the stellar mass accounts for all the halo's share of baryons.
We focus on central galaxies because the strong tidal stripping to which satellite haloes are subject obscures the underlying relation between galaxy formation efficiency and halo mass.
The simulation clearly shows that galaxy formation is most efficient in haloes with mass $\sim 10^{12}\,{{\rm M}_\odot}$, as has been found by many others. In fact, it would be more appropriate to say that this is the mass where galaxy formation is ``least inefficient'' as the efficiency is only $\sim 10$\% at the peak. The efficiency is sharply peaked at a stellar mass of $\sim 10^{10.4}\,{{\rm M}_\odot}$, which corresponds to the onset of the knee in the GSMF (Fig.~{\ref{fig:gsmf}). As is the case for most models of galaxy formation, in EAGLE\ the sharp reduction at lower masses is mostly due to stellar feedback, while the drop off at higher masses can in part be attributed to inefficient cooling, but is mostly caused by AGN feedback.
Although halo masses can be measured observationally, e.g.\ from gravitational lensing or satellite kinematics, the errors are still relatively large and it is difficult to disentangle central and satellite galaxies. In Figure~\ref{fig:eta} we therefore compare with results obtained through the abundance matching technique. In its most basic form abundance matching relates central galaxies to haloes by matching the observed GSMF to the halo mass function predicted from a collisionless simulation, assuming that the stellar masses of galaxies increase monotonically with the masses of their host haloes \citep[e.g.][]{Vale2004AbundanceMatching}. Modern versions allow for scatter and evolution, and assume that the masses of satellite galaxies are set at the last time they were centrals.
Figure~\ref{fig:eta} compares EAGLE\ to the abundance matching results of \citet{Behroozi2013AbundMatching} and \citet{Moster2013AbundMatching}. Note that the abundance matching studies assumed the WMAP7 cosmology, whereas we assume the Planck cosmology. For EAGLE\ we use the total mass of the halo in the hydrodynamical simulation, whereas abundance matching studies use collisionless simulations. Because feedback processes reduce halo masses, we expect $M_{\rm 200}$ to be biased high by $\sim 10$\% for the abundance matching results \citep[e.g.][]{Sawala2013HaloMass,Cui2014HaloMass,Velliscig2014HaloMass,Cusworth2014ClusterMass,Martizzi2014ClusterMass,Sawala2014EagleZooms,Vogelsberger2014Illustris}, but this effect is small compared to the dynamic range shown\footnote{For $M_{200} \ll 10^{10}\,{{\rm M}_\odot}$ the systematic errors in the abundance matching results are likely to be much greater because only a small fraction of such low-mass haloes may host galaxies \citep{Sawala2013HaloMass,Sawala2014EagleZooms}.}. Beyond the peak the results become increasingly sensitive to the aperture used to measure the galaxy's light. For example, \citet{Kravtsov2014GalaxyHalo} show that using the \citet{Bernardi2013GSMF} GSMF as input increases the efficiency by $\sim 0.5$ dex at $M_{\rm 200} = 10^{14}\,{{\rm M}_\odot}$ relative to the values of \citet{Behroozi2013AbundMatching} and \citet{Moster2013AbundMatching}. However, as discussed in \S\ref{sec:aperture}, our use of a fixed 30~pkpc aperture means that comparison to \citet{Bernardi2013GSMF} is inappropriate at the high-mass end. In \S\ref{sec:groups} we will show that a more robust comparison with observations of the total stellar content of massive galaxies reveals good agreement with EAGLE.
The convergence with resolution is good and the galaxy formation efficiency in EAGLE\ is very close to that inferred from abundance matching. This was of course to be expected, given the good convergence and the good agreement with the observations for the GSMF.
The peak efficiency is 0.1--0.2 dex lower in EAGLE\ and is reached at a slightly ($\sim 0.2$~dex) higher stellar mass, which is consistent with the fact that EAGLE\ slightly undershoots the observed GSMF at the knee (see Fig.~\ref{fig:gsmf}).
\subsection{Galaxy sizes}
\label{sec:sizes}
\begin{figure}
\resizebox{\colwidth}{!}{\includegraphics{size50_ns_LT_2p5_2scat.pdf}}
\caption{Galaxy size as a function of stellar mass for galaxies at $z=0.1$ in proper kiloparsec. The coloured curves show the median, projected half-mass radii for the simulations and the shaded regions show the $1\sigma$ scatter. For clarity we only show the scatter in Recal-L025N0752 for $M_\ast < 10^{10}\,{{\rm M}_\odot}$ and in Ref-L100N1504 for $M_\ast > 10^{10}\,{{\rm M}_\odot}$. The simulation curves are dotted below the resolution limit of 600 stellar particles. Where there are fewer than 10 galaxies per bin, individual objects are shown as filled circles. The models are compared with S\'ersic half-light radii from SDSS (\citealt{Shen2003Sizes}; the grey, solid line shows the median and the grey dotted lines indicate 1$\sigma$ scatter) and GAMA (\citealt{Baldry2012GSMF}; data points with error bars indicate the $1\sigma$ scatter, shown separately for blue and red galaxies). The simulations and \citet{Shen2003Sizes} only include late-type galaxies, i.e.\ a S\'ersic index $n_{\rm s} < 2.5$.}
\label{fig:sizes}
\end{figure}
The parameters of the subgrid model for feedback from star formation and AGN were calibrated to observations of the $z\sim 0$ GSMF. The parameter that controls the importance of the angular momentum of the gas in suppressing BH accretion was set to a value for which AGN feedback causes the GSMF to turn over at a mass similar to what is observed.
As will be shown in \citet{Crain2014EagleModels}, we found that for EAGLE, calibration of the stellar feedback is actually unnecessary to reproduce the GSMF. Fixing the amount of energy injected per unit stellar mass to that available in the form of core collapse supernovae, i.e.\ $f_{\rm th}=1$, works well, as does the physically motivated dependence on the gas metallicity that we use (eq.~\ref{eq:f(Z)}). However, such models produce galaxies that are far too compact because of excessive radiative losses at high gas densities, and we can show analytically that these spurious cooling losses are caused by our limited numerical resolution (see \S\ref{sec:snii}).
We consider it reassuring that the breakdown of the subgrid model for feedback from star formation at high density is understood and leads to a clear conflict with observations. On the other hand, the fact that such an unrealistic model has no trouble matching the observed GSMF emphasizes the importance of comparing to a wide range of observables.
To counteract the numerical radiative losses occurring at high gas densities, we introduced a dependence of the feedback energy from star formation on the gas density, while keeping both the maximum and mean amounts of energy reasonable (see \S\ref{sec:calibration}).
Although we could not afford the computational expense of calibrating the models to fit both the $z\sim 0$ GSMF and the size distribution in detail, we did reject models that produced galaxies that were far too small. As a consequence of this strategy, the $z\sim 0$ galaxy sizes cannot be regarded as true predictions.
Figure~\ref{fig:sizes} plots the median value of the half-mass radius, $R_{50}$, i.e.\ the radius that encloses 50 per cent of the stellar mass in projection, as a function of galaxy stellar mass.
The half-mass radii were determined by fitting S\'ersic laws to the projected, azimuthally averaged surface density profiles, as in \citet{McCarthy2012RotSize}. Following \citet{Shen2003Sizes}, only galaxies with S\'ersic index $n_{\rm s}<2.5$ are included. For Ref-L1001504, 94\% of the galaxies with more than 600 star particles have $n_{\rm s}<2.5$.
The high-resolution Recal-L025N0752 agrees very well with the intermediate-resolution models for $M_\ast > 10^{9}\,{{\rm M}_\odot}$, which corresponds to about 600 star particles for the intermediate-resolution runs. For this mass the median $R_{50}$ is about three and a half times the maximum gravitational softening length (see Table~\ref{tbl:sims}). Hence, we take the stellar mass $600 m_{\rm g}$ as the minimum value for which we can measure half-mass radii. We thus require six times more stellar particles to measure sizes than we need to measure mass.
The simulations are compared to data from SDSS \citep{Shen2003Sizes} and GAMA \citep{Baldry2012GSMF}. Note that the observations fit surface brightness profiles and provide half-light radii rather than half-mass radii, so the comparison with the models is only fair if the stellar mass-to-light ratio does not vary strongly with radius. As mentioned above, \citet{Shen2003Sizes} select galaxies with $n_{\rm s}<2.5$, as we have done here. \citet{Baldry2012GSMF} on the other hand present results separately for red and blue galaxies, finding that the latter are $\sim 0.2$ dex more extended at fixed stellar mass. \citet{Shen2003Sizes} use Petrosian apertures, which we expect to yield results similar to the 3-D apertures of 30 pkpc that we use for the simulations (see \S\ref{sec:aperture}).
For $M_\ast \gg 10^8\,{{\rm M}_\odot}$ \citet{Shen2003Sizes} agree better with the \citet{Baldry2012GSMF} results for red galaxies, even though $n_{\rm s} < 2.5$ should pick out more disky and hence bluer galaxies. The differences between the two data sets are indicative of the level of correspondence between independent measurements of observed galaxy sizes.
For $10^9 < M_\ast/{{\rm M}_\odot} < 10^{10}$ the simulation results fall in between those of \citet{Baldry2012GSMF} for red and blue galaxies. For $M_\ast< 10^9\,{{\rm M}_\odot}$ and $M_\ast> 10^{10}\,{{\rm M}_\odot}$ the simulations agree very well with the sizes of blue and red galaxies, respectively. At $10^{11}\,{{\rm M}_\odot}$ the red sample of \citet{Baldry2012GSMF} gives sizes that are about 0.1--0.2 dex larger than found for both the simulations and the data from \citet{Shen2003Sizes}. This difference may be due to the fact that \citet{Shen2003Sizes} use Petrosian sizes, whereas \citet{Baldry2012GSMF} do not. Indeed, if we do not impose any 3-D aperture, then the simulation curve follows the results of the red sample nearly exactly for $M_\ast \ga 10^{11}\,{{\rm M}_\odot}$, while the sizes of lower-mass galaxies remain unchanged (not shown). The agreement with \citet{Shen2003Sizes} is excellent: the difference with the simulations is $\le 0.1$ dex for all models and for the full range of stellar mass.
For $M_\ast > 10^{10}\,{{\rm M}_\odot}$ the scatter in the sizes of the simulated galaxies is similar to the observed dispersion, but at lower masses it appears to be smaller. This could be due to a lack of resolution or some other deficiency in the simulations or halo finder, but it could also be due to observational errors or to the fact that we have ignored variations in the stellar mass-to-light ratio and dust extinction.
\subsection{The relation between BH mass and stellar mass}
\label{sec:magorrian}
Figure~\ref{fig:bh} shows the mass of the central supermassive BH as a function of the galaxy's stellar mass. The simulation results are compared with the compilation of observations from \citet{McConnell2013BH}. The observed stellar mass was obtained by extrapolating a fit to the mass profile of the bulge inferred from kinematic data. Because the observed galaxies were selected to be early-type, the bulge likely dominates the stellar mass, at least for the massive systems.
The three EAGLE\ simulations give nearly identical results, indicating good convergence. For $M_\ast \ll 10^{10}\,{{\rm M}_\odot}$ the BH mass asymptotes to $10^5\,{{\rm M}_\odot}/h$, which is the mass of the seed BHs that are inserted into FoF haloes with mass $>10^{10}\,{{\rm M}_\odot}/h$ that do not already contain BHs. As can be seen from Fig.~\ref{fig:eta}, a halo mass of $10^{10}\,{{\rm M}_\odot}$ corresponds to $M_\ast\sim 10^8\,{{\rm M}_\odot}$. Above $M_\ast \sim 10^{10}\,{{\rm M}_\odot}$ the relation between BH mass and stellar mass steepens, but it quickly flattens off to a relation that agrees very well with the observations for $M_\ast \ga 10^{11}\,{{\rm M}_\odot}$. The rapid growth of the BHs between $M_\ast = 10^{10}$ and $10^{11}\,{{\rm M}_\odot}$ coincides with the steepening of the GSMF (compare Fig.~\ref{fig:gsmf}) and the sharp increase in the fraction of galaxies that are passive (right panel of Fig.~\ref{fig:ssfr}). This is understandable, as the AGN feedback associated with the rapid BH growth quenches star formation.
The agreement with the observations is good, although the observed scatter is larger. In terms of the normalisation of the $M_{\rm BH}$-$M_\ast$ relation the good agreement is perhaps not a surprise. The normalisation is determined by the assumed efficiency of the AGN feedback, $\epsilon_{\rm f}\epsilon_{\rm r}$, i.e.\ the amount of energy that is injected per unit of accreted mass \citep[e.g.][]{Booth2009AGN,Booth2010DMHaloesBHs}}. We used the same value ($\epsilon_{\rm f}\epsilon_{\rm r} = 0.015$) as was used for OWLS\ and cosmo-OWLS, which \citet{Booth2009AGN} and \citet{LeBrun2014CosmoOWLS} found to give agreement with the observed $M_{\rm BH}$-$M_\ast$ relation. Fig.~\ref{fig:bh} shows that this efficiency also works for EAGLE, even though the mass resolution of EAGLE\ is nearly two orders of magnitude better than for OWLS\ and about 3 orders of magnitude better than for cosmo-OWLS. Note, however, that we used higher AGN heating temperatures than the $\Delta T_{\rm AGN} = 10^8\,{\rm K}$ that was used in OWLS\ (see Table~\ref{tbl:subgridpars}).
It would clearly be desirable to extend the comparison to observations to lower masses, but in this regime a more careful analysis is required. This is because of the importance of systematic and selection effects for the observations \citep[e.g.][]{Lauer2007BHBias,Schulze2011BHBias} and because a bulge-to-disc decomposition would be necessary for the simulations since most low-mass galaxies are disky. The same issues likely also affect the comparison of the scatter.
\begin{figure}
\resizebox{\colwidth}{!}{\includegraphics{bh_2scat_Stars_030kpc_BH_MostMassive_MERGED.pdf}}
\caption{The relation between the mass of the central supermassive black hole and the stellar mass of galaxies. The coloured curves show the median relations for the simulations and the shaded regions show the $1\sigma$ scatter. For clarity we only show the scatter in Recal-L025N0752 for $M_\ast < 10^{10}\,{{\rm M}_\odot}$ and in Ref-L100N1504 for $M_\ast > 10^{10}\,{{\rm M}_\odot}$. Where there are fewer than 10 objects per bin, individual objects are shown as filled circles.
Data points with $1\sigma$ error bars show the compilation of observations from \citet{McConnell2013BH}. The simulations show the total stellar mass (within a 3-D aperture of 30~pkpc), while observations show bulge masses. However, the observed galaxies were selected to be early-type. The simulations agree with the observations, although the observed scatter is larger.}
\label{fig:bh}
\end{figure}
\section{Comparison with other observations}
\label{sec:otherobs}
In this section we will compare the results of EAGLE\ to a diverse set of low-$z$ observations of galaxies, galaxy clusters, and the IGM. The results reported in this section were not used to calibrate the subgrid models for feedback and can therefore be considered predictions that can be used as independent consistency checks. During the testing phase, we did look at earlier, more basic versions of some of the plots shown here, so most of the predictions cannot be considered blind. However, we have not adjusted any model parameters to improve the results shown in this section.
There are two exceptions to the above statements. First, we plotted the metal column density distributions (\S\ref{sec:cddf}) for the first time after the simulations had finished, so this was a truly blind prediction. Second, the discrepancy between the gas fraction in clusters predicted by Ref-L100N1504 and inferred from X-ray observations that will be discussed in \S\ref{sec:groups} was the motivation for running model AGNdT9-L050N0752. This model represents an educated guess in terms of the modifications to the subgrid AGN feedback, because we could only afford to calibrate models using volumes of 25 cMpc on a side, which are too small to contain clusters of galaxies.
The observables presented in this section were not selected because the models reproduce them accurately. They were selected because they give a broad overview of the $z\sim 0$ EAGLE\ universe, because we had the tools to compute them, and because we are currently not preparing separate papers on them. Future papers will present more observables as well as results for higher redshifts.
\subsection{Specific star formation rates and passive fractions}
\label{sec:ssfr}
\begin{figure*}
\resizebox{\colwidth}{!}{\includegraphics{ssfr_2scat.pdf}}
\resizebox{\colwidth}{!}{\includegraphics{passive.pdf}}
\caption{\emph{Left panel:} Specific star formation rate, $\dot{M}_\ast/M_\ast$, for actively star-forming galaxies as a function of stellar mass at $z=0.1$. Galaxies are classified as star-forming if their SSFR $> 10^{-2}\,{\rm Gyr}^{-1}$, indicated by the horizontal, dashed line. The coloured curves show simulation medians and the shaded regions show the $1\sigma$ scatter. For clarity we only show the scatter in Recal-L025N0752 for $M_\ast < 10^{10}\,{{\rm M}_\odot}$ and in Ref-L100N1504 for $M_\ast > 10^{10}\,{{\rm M}_\odot}$, all at $z=0.1$. The higher and lower diagonal lines correspond to 10 star-forming gas particles (assuming $n_{\rm H} = 0.1~{\rm cm}^{-3}$) at intermediate and high resolution, respectively. To the left of these lines the curves are dotted to indicate that the results are unreliable due to sampling effects. In particular, the sharp upturns at the lowest masses trace lines of fixed numbers of star-forming gas particles.
The data points show observations from GAMA ($0.05 < z < 0.32$; \citealt{Bauer2013ssfr}) with the error bars indicating the $1\sigma$ scatter.
\emph{Right panel:} Fraction of passive galaxies, i.e.\ galaxies with SSFR $< 10^{-2}\,{\rm Gyr}^{-1}$, as a function of stellar mass at $z=0.1$. In both panels the simulation curves are dotted where they are unreliable due to poor resolution ($< 10$ star-forming gas particles) and dashed were there are $<10$ objects per bin. Data points show observations from
\citet{Bauer2013ssfr} and \citet{Moustakas2013GSMF}.}
\label{fig:ssfr}
\end{figure*}
The left panel of Figure~\ref{fig:ssfr} shows the specific star formation rate (SSFR), $\dot{M}_\ast/M_\ast$, of actively star-forming galaxies as a function of stellar mass. Here, galaxies are classified to be star-forming if the ${\rm SSFR} > 0.01~{\rm Gyr}^{-1}$, which is indicated by the horizontal, dashed line in the left panel.
The higher and lower diagonal lines in the left panel indicate the SSFR corresponding to 10 star-forming gas particles (assuming a gas density of $n_{\rm H}=10^{-1}\,{\rm cm}^{-3}$, the star formation threshold that we impose at the metallicity $Z=0.002$) at intermediate and high resolution, respectively. To the left of these curves resolution effects become important, which we indicate by using dotted lines. In particular, the increase in the SSFR at low stellar mass that is clearly visible for the intermediate resolution simulations is a numerical effect: the curves trace lines of constant numbers of star-forming particles. Compared with the intermediate-resolution models, the high-resolution simulation Recal-L025N0752 predicts slightly higher SSFRs. The difference is 0.2 dex at $M_\ast = 10^9\,{{\rm M}_\odot}$ and less than 0.1 dex above $10^{10}\,{{\rm M}_\odot}$.
The models are compared with observations from \citet{Bauer2013ssfr}, who measured the SSFRs of $\sim 73,000$ galaxies from the GAMA survey using spectroscopic H$\alpha$ measurements and dust corrections based on Balmer decrements. The intermediate-resolution simulations agree with the data at the high-mass end, but underpredict the SSFR at low masses, reaching a maximum discrepancy of 0.3 -- 0.4 dex at $10^9\,{{\rm M}_\odot}$. The high-resolution model also underpredicts the SSFR, but the discrepancy is less than 0.2 dex. These differences are comparable to the systematic uncertainty in the data. For example, even for a fixed IMF the systematic uncertainty in the stellar mass, which shifts the data parallel to the diagonal lines, is $\sim 0.3$~dex \citep{Conroy2009SPSSUncertainty,Behroozi2010Uncertainties,Pforr2012SPSSUncertainty,Mitchell2013MassUncertainty} and the systematic error in the star formation rate, which shifts the data vertically, is likely to be at least as large \citep[e.g.][]{Moustakas2006SfrIndicators}. The scatter in the simulations is $\sim 50$\% smaller than observed, but the observed scatter includes measurement and systematic uncertainties.
The right panel of Figure~\ref{fig:ssfr} shows the fraction of galaxies that are passive as a function of stellar mass. For the simulations we classify galaxies as passive if they have ${\rm SSFR} < 0.01~{\rm Gyr}^{-1}$, but the observational papers use somewhat different and varying criteria. We leave a more precise comparison for future work, e.g.\ using colours and accounting for dust extinction for the simulated galaxies. At low stellar masses the curves become dashed where there are, on average, fewer than 10 star-forming gas particles in a galaxy with ${\rm SSFR} = 0.01~{\rm Gyr}^{-1}$. These parts of the curves are unreliable and the upturn of the passive fraction at low mass is thus due to the limited resolution of the simulations. This interpretation is confirmed by the fact that the upturn shifts to eight times lower masses if the particle mass is decreased by a factor of eight, switching from the intermediate resolution Ref-L100N1504 to the high-resolution Recal-L025N0752.
For $M_\ast \gg 10^9\,{{\rm M}_\odot}$, where the simulations are close to converged, both the simulations and the observations show a strong increase of the passive fraction with mass, from $\sim 10$ per cent at $10^9\,{{\rm M}_\odot}$ to $\sim 90$ per cent at $10^{11.5}\,{{\rm M}_\odot}$. Relative to the data, the simulation curves are shifted towards higher stellar masses by about 0.3 dex. This difference is similar to the systematic uncertainty in the observed stellar masses. We also find shifts of similar magnitudes if we vary the critical SSFR below which simulated galaxies are classified as passive by a factor of two.
We conclude that in the regime where the simulations can be trusted, the predicted SSFRs and passive fractions are slightly lower than the observations but agree with them to within the expected (systematic) errors.
\subsection{Tully-Fisher relation}
\label{sec:tf}
\begin{figure}
\resizebox{\colwidth}{!}{\includegraphics{tf_vmax_2scat.pdf}}
\caption{The relation between the maximum of the rotation curve and stellar mass, i.e.\ an analogue of the Tully-Fisher relation, for late-type galaxies at $z=0.1$. The coloured curves show the medians for the simulations. The curves are dotted below the resolution limit of 100 stellar particles. Where there are fewer than 10 galaxies per bin, individual objects are shown as filled circles. The shaded regions show the $1\sigma$ scatter in the simulations. For clarity we only show the scatter in Recal-L025N0752 for $M_\ast < 10^{10}\,{{\rm M}_\odot}$ and in Ref-L100N1504 for $M_\ast > 10^{10}\,{{\rm M}_\odot}$. The simulation results only include galaxies with S\'ersic index $n_{\rm s} < 2.5$ and are based on maximum circular velocities. The data points with $1\sigma$ error bars correspond to the set of homogenised observations of disc galaxies compiled by \citet{Avila-Reese2008TF} and the grey line indicates the median. The model predictions are in remarkable agreement with the data.}
\label{fig:tf}
\end{figure}
Figure~\ref{fig:tf} shows the relation between the maximum of the rotation curve and stellar mass for disc galaxies, i.e.\ a close relative of the Tully-Fisher relation \citep{Tully1977TF}.
For the simulations we classify galaxies with S\'ersic index $n_{\rm s}<2.5$ as late-type, as we did when considering galaxy sizes (\S\ref{sec:sizes}). We use circular velocities ($v_{\rm c}=\sqrt{GM(<r)/r}$) rather than trying to estimate rotation velocities, since the latter become noisy for galaxies that are not resolved with many particles.
The data points with $1\sigma$ error bars correspond to the set of homogenised observations of disc galaxies compiled by \citet{Avila-Reese2008TF} and the grey line indicates the median.
The stellar masses have been reduced by 0.15 dex, which is necessary to convert to a Chabrier IMF (Avila-Reese, private communication). In addition, following \citet{McCarthy2012RotSize} and \citet{Dutton2011TF}, we applied a small correction to the stellar masses using the expression given in the appendix of \citet{Li2009GSMF} to improve the consistency with those derived from more accurate five-band SDSS data.
All simulations track each other very closely, implying excellent numerical convergence. The simulations are in excellent agreement with the data. Over the mass range $10^9 \la M_\ast/{{\rm M}_\odot} <10^{11}$ the difference in velocity between the models and the data compiled by \citet{Avila-Reese2008TF} is less than 0.03~dex, which is smaller than the 0.1~dex $1\sigma$ error on the fit to the observations. At higher masses, which are only probed by Ref-L100N1504, the difference with the observations increases, reaching 0.12 dex at $M_\ast = 10^{11.3}\,{{\rm M}_\odot}$. However, most of these very massive galaxies do not look disky and would probably not be selected by \citet{Avila-Reese2008TF}.
Note that we have not attempted to analyse the simulations and the data in the same manner, because this would go beyond the scope of the current study. As mentioned above, we use maximum circular velocities, whereas the observations are based on maximum gas rotation velocities, which may show more scatter if the orbits are not all circular. In addition, the observations probe only the inner parts of the halo, whereas we consider the entire halo. \citet{McCarthy2012RotSize} found that for the GIMIC\ simulations the maximum circular velocities are nearly always reached within two effective radii for $M_\ast \ga 10^{9.5}\,{{\rm M}_\odot}$, and should therefore be easily accessible to the observations, but it is possible that for smaller masses the observations underestimate the maximum rotation velocity.
\subsection{Mass-metallicity relations}
\label{sec:Zm}
The left panel of Figure~\ref{fig:mz} shows the metallicity of the ISM, which we take to be star-forming gas for the simulations, as a function of stellar mass. For both the intermediate- and the high-resolution models the gas metallicity increases with stellar mass and flattens off for $M_\ast > 10^{10}\,{{\rm M}_\odot}$. However, the high-resolution simulation, Recal-L025N0752, predicts systematically lower metallicities. For $M_\ast \ga 10^{10}\,{{\rm M}_\odot}$ the difference is less than 0.15 dex, but it increases with decreasing mass, reaching a maximum of 0.4 dex at $M_\ast \sim 10^{8.5}\,{{\rm M}_\odot}$. Because there is no clear mass below which the two resolutions diverge, it is unclear where to put the resolution limit and we therefore have not dotted any part of the curves.
Interestingly, model Ref-L025N0752 (not shown) yields a mass-metallicity relation that agrees better with Ref-L100N1504 than the prediction of Recal-L025N0752 does, particularly for $M_\ast < 10^9\,{{\rm M}_\odot}$. The high-resolution run again predicts lower metallicities than the intermediate-resolution version, but the maximum difference is smaller than 0.2~dex. For $M_\ast < 10^{7.5}\,{{\rm M}_\odot}$ the metallicity is actually lower at intermediate resolution than at high resolution. Hence, for the mass-metallicity relation the strong convergence is considerably better than one might infer from the comparison of Ref-L025N0752 and Recal-L025N0752. Recall that the latter was recalibrated to fit the GSMF, which meant the efficiency of feedback had to be increased relative to the reference model, particularly at $M_\ast \sim 10^9\,{{\rm M}_\odot}$ (see Fig.~\ref{fig:gsmf_conv}). Apparently, the stronger outflows in Recal-L025N0752 reduce the metallicity of the ISM.
Thus, the ``strong convergence'' is better than the ``weak convergence''. This is possible because in this case the weak convergence test compares simulations that were each calibrated to fit the GSMF, not the mass-metallicity relation.
The two sets of observations that are shown in the left panel of Figure~\ref{fig:mz} are both derived from SDSS data.
\citet{Tremonti2004Zgas} estimated the metallicity statistically based
on theoretical model fits to various strong emission lines, while \citet{Zahid2014Zgas} derived metallicities using the R23 strong line method as calibrated by \citet{Kobulnicky2004Zgas}. The two studies do not agree with each other. In particular, while \citet{Tremonti2004Zgas} and \citet{Zahid2014Zgas} agree at $M_\ast \sim 10^{11}\,{{\rm M}_\odot}$, the former find a steeper relation than the latter, resulting in metallicities that are about 0.2 dex lower for $10^9 - 10^{10}\,{{\rm M}_\odot}$. The difference is due to the uncertain calibration of the emission-line diagnostics. In fact, as shown by \citet{Kewley2008ZCalibration}, the systematic uncertainty is even larger than suggested by this plot. For example, the empirical calibration of \citet{Pilyugin2005ZCalibration} yields a metallicity that is 0.75~dex lower than that of \citet{Tremonti2004Zgas} at $10^{11}\,{{\rm M}_\odot}$ and an almost flat relation with stellar mass, dropping by only 0.2 dex when the stellar mass decreases to $10^9\,{{\rm M}_\odot}$. Besides the calibration issues, the gas phase abundance likely underestimates the total metallicity of the ISM because a non-negligible fraction of the metals may condense onto dust-grains \citep[e.g.][]{Dwek1998Dust,Mattsson2012Dust}.
Finally, the systematic uncertainty in the stellar mass, for a fixed IMF, is about 0.3~dex \citep[e.g.][]{Conroy2009SPSSUncertainty}.
The metallicities predicted by the simulations are also subject to significant systematic uncertainties unrelated to the galaxy formation physics. Even for a fixed IMF, the nucleosynthetic yields are uncertain at the factor of two level \cite[e.g.][]{Wiersma2009Chemo}. However, we choose not to simply re-scale the simulation metallicities within this uncertainty because that would make them inconsistent with the radiative cooling rates used during the simulation.
\begin{figure*}
\resizebox{\colwidth}{!}{\includegraphics{mzr_gas_2scat.pdf}}
\resizebox{\colwidth}{!}{\includegraphics{mzr_star_2scat.pdf}}
\caption{The metallicity of the ISM (left panel) and of stars (right panel) as a function of stellar mass. The conversion from the absolute oxygen abundances shown along the left $y$-axis in the left panel to the metallicities relative to solar shown along the right $y$-axis assumes $12 + \log_{10}({\rm O}/{\rm H})_\odot = 8.69$ \citep{AllendePrieto2001Solar}. Note that the two panels show the same range in metallicity. Curves show the median relations for the simulations at $z=0.1$, where we take ISM to be all star-forming gas, and the shaded regions show the $1\sigma$ scatter. For clarity we only show the scatter in Recal-L025N0752 for $M_\ast < 10^{10}\,{{\rm M}_\odot}$ and in Ref-L100N1504 for $M_\ast > 10^{10}\,{{\rm M}_\odot}$. Where there are fewer than 10 galaxies per bin, individual objects are shown as filled circles. The high-mass galaxies with very low gas metallicities correspond to objects that are nearly devoid of gas, leading to sampling problems in the simulations. The data points show observations reported by \citet{Zahid2014Zgas} and \citet{Tremonti2004Zgas} for gas, and by \citet{Gallazzi2005Zstars} and \citet{Kirby2013Zstars} for stars (converted to solar abundances assuming ${\rm Z}_\odot = 0.0127$ and $12 + \log_{10}({\rm Fe}/{\rm H})_\odot = 7.52$, respectively). The dashed line in the right panel shows the best-fit relation given by \citet{Kirby2013Zstars}, which includes also lower-mass galaxies than shown here.}
\label{fig:mz}
\end{figure*}
Given the large systematic uncertainties in both the normalisation and the shape of the observed mass-metallicity relation, and the systematic uncertainties in the yields adopted in the simulations, care needs to be taken when comparing the models and the data. We will nevertheless proceed to make such a comparison.
The median mass-metallicity relations predicted by the intermediate-resolution simulations agree with \citet{Zahid2014Zgas} to better than 0.2 dex at all masses and to better than 0.1 dex for $M_\ast > 10^{9.5}\,{{\rm M}_\odot}$, but the observed relation is steeper at lower masses. The predicted scatter is larger than observed by \citet{Tremonti2004Zgas}, particularly for the highest masses. The scatter in the gas metallicity of these massive objects is large in the simulations because they typically contain very few star-forming gas particles. This causes strong sampling effects and large variations in time following AGN outbursts.
The median metallicity predicted by the high-resolution model Recal-L025N0752 matches \citet{Tremonti2004Zgas} to better than 0.2 dex over the full mass range covered by both the simulation and the observations ($10^{8.5} < M_\ast/{{\rm M}_\odot} < 10^{11}$) and to better than 0.1 dex for $M_\ast > 10^{9.2}\,{{\rm M}_\odot}$. Apparently, the increase in the efficiency of energy feedback from star formation that is required to make the GSMF fit the observations (and which was implemented by changing the density dependence of the efficiency, see \S\ref{sec:calibration}), simultaneously decreases the metallicity of the ISM of low-mass galaxies to the values observed by \citet{Tremonti2004Zgas}.
The predicted relations between stellar metallicity and mass are shown in the right panel of Figure~\ref{fig:mz} and compared with observations from SDSS from \citet{Gallazzi2005Zstars} and for dwarf galaxies from \citet{Kirby2013Zstars}. The trends and differences largely parallel those seen for the gas-phase abundances in the left panel. For $M_\ast \ga 10^9\,{{\rm M}_\odot}$ simulation Recal-L025N0752 is relatively close to the data, but at lower masses all models predict higher metallicities than observed by \citet{Kirby2013Zstars}. As was the case for the gas metallicity, the (strong) convergence is actually much better than suggested by this figure. For $M_\ast > 10^{7.5}\,{{\rm M}_\odot}$ simulation Ref-L025N0752 (not shown) predicts a stellar metallicity that is lower, but within 0.1 dex of the metallicity predicted by Ref-L100N1504. Model AGNdT9-L050N0752 predicts slightly higher metallicities than Ref-L100N1504 for $M\gg 10^{10}\,{{\rm M}_\odot}$, which agrees better with the data.
The main difference between the conclusions that can be drawn from the gas and stellar metallicities concerns the scatter. While the scatter in the gas phase abundances was overestimated in the simulations, the scatter in the stellar abundances appears to be strongly underestimated. However, it would be surprising for the scatter in the observed stellar metallicity to be so much larger than the observed scatter in the gas phase metallicity, which suggests that the scatter in the observed stellar metallicities may be dominated by errors. Indeed, while the mean relation from the CALIFA integral field survey is close to that of \citet{Gallazzi2005Zstars}, the scatter is about a factor of two smaller \citep{Gonzalez2014CalifaZM}.
\subsection{X-ray observations of the intracluster medium}
\label{sec:groups}
In this section we will consider some parameters that are commonly measured from X-ray observations of the intragroup and intracluster gas. The comparison to observations is more like-for-like than in previous sections, because all simulation results are derived by applying observational analysis techniques to virtual X-ray observations of the simulations. Simulation Recal-L025N0752 is not considered here because the simulation box is too small to produce clusters of galaxies.
The methods used to generate the plots are identical to those employed for cosmo-OWLS\ in \citet{LeBrun2014CosmoOWLS} and we refer the reader to \S2.2 of that paper for details. Briefly, gas density, temperature and metallicity profiles are determined by fitting single temperature, single metallicity ``Astrophysical Plasma Emission Code'' (APEC) \citep{Smith2001APEC} models to synthetic \emph{Chandra} X-ray spectra in three-dimensional radial bins centred on the minimum of the gravitational potential in the halo. Mass profiles are obtained by fitting the functions proposed by \citet{Vikhlinin2006} to the density and temperature profiles and assuming hydrostatic equilibrium. We then determine the radius within which the mean internal density equals 500 times the critical density, $R_{500, {\rm hse}}$ , and the corresponding spherical overdensity mass,$M_{500, {\rm hse}}$. We will use the subscript ``hse'' to indicate that the quantity has been inferred from virtual observations under the assumption of hydrostatic equilibrium (which holds only approximately, see \citealt{LeBrun2014CosmoOWLS} and references therein). Mean X-ray temperatures and elemental abundances within $R_{500, {\rm hse}}$ are determined by fitting APEC models to a single radial bin. We include all $z=0$ haloes with FoF mass $ > 10^{12.5}\,{{\rm M}_\odot}$ but plot only results for haloes with $M_{500, {\rm hse}} > 10^{13}\,{{\rm M}_\odot}$ for which the correspondence between $M_{\rm 500}$ and $M_{500, {\rm hse}}$ is good for most objects, except that $M_{500, {\rm hse}}$ is systematically biased low by $\sim 20$ per cent \citep[see Fig.~B1 of][]{LeBrun2014CosmoOWLS}.
\begin{figure}
\resizebox{\colwidth}{!}{\includegraphics{Lopt_M500.eps}}
\caption{\textit{I}-Band luminosity within $R_{500,{\rm hse}}$ as a function of $M_{500,{\rm hse}}$ at $z=0$. Black data points represent observations of \citet{Sanderson2013}, \citet{Gonzalez2013}, and \citet{Kravtsov2014GalaxyHalo}, and the dashed black line represents the SDSS image stacking results of \citet{Budzynski2014}. Where necessary, observations were converted to the $I$-band following \citet{LeBrun2014CosmoOWLS}. The observational studies and the simulations both include contributions from satellites and diffuse intracluster light (ICL). The simulations agree well with the data.}
\label{fig:Lopt_groups}
\end{figure}
Figure~\ref{fig:Lopt_groups} shows the (Cousins) $I$-band luminosity within $R_{500,{\rm hse}}$ as a function of $M_{500,{\rm hse}}$. Each point corresponds to a single simulated or observed object.
The predicted luminosity-mass relation matches the observations very well. As the $I$-band luminosity is a proxy for stellar mass and the simulations were calibrated to the observed GSMF, this may at first sight not be surprising. However, the high-mass tail of the GSMF was not calibrated to any observations, because the test simulations were too small to contain such rare objects. Moreover, here we plot the total luminosity within $R_{500}$, a radius that exceeds the aperture used for the GSMF by more than an order of magnitude. Hence, the results shown here include contributions from satellites and the intracluster light, both for the observations and simulations.
\begin{figure}
\resizebox{\colwidth}{!}{\includegraphics{fgas_M500.eps}}
\caption{The $z=0$ gas mass fraction within $R_{500,{\rm hse}}$ as a function of $M_{500,{\rm hse}}$. All quantities are inferred from (virtual) X-ray observations. Black data points represent observations of \citet{Vikhlinin2006}, \citet{Maughan2008}, \citet{Allen2008}, \citet{Pratt2009}, \citet{Sun2009}, and \citet{Lin2012}. The reference model overpredicts the gas fractions. Model AGNdT9-L050N0752, which employs a higher heating temperature for AGN feedback, performs well for groups of galaxies, but may also overpredict the gas fraction in higher mass ($\ga 10^{14}\,{{\rm M}_\odot}$) clusters.}
\label{fig:fgas_groups}
\end{figure}
Figure~\ref{fig:fgas_groups} shows the gas mass fraction, $M_{\rm gas,500,hse}/M_{\rm 500,hse}$ as a function of mass $M_{\rm 500,hse}$. Because the gas mass is derived from the (virtual) X-ray data, it only correctly accounts for gas that has a temperature similar to that of the gas that dominates the X-ray emission. For the reference model the gas mass inferred from X-ray observations, under the assumption of hydrostatic equilibrium, is about 0.2~dex higher than observed, except perhaps for the two most massive objects.
\citet{LeBrun2014CosmoOWLS} have shown that the gas fraction is particularly sensitive to the temperature to which the AGN heat the surrounding gas in our subgrid prescription for AGN feedback. In particular, higher heating temperatures, which correspond to more energetic but less frequent bursts, eject the gas more effectively, yielding lower gas fractions. This was the motivation for running model AGNdT9-L050N0752, which uses a heating temperature $\Delta T_{\rm AGN}$ of $10^9\,{\rm K}$, compared with $10^{8.5}\,{\rm K}$ for the reference model. Before running this model, we used a 25~cMpc version to (approximately) recalibrate the BH accretion model so as to maintain the good match with the GSMF, in particular the location of the knee. We could, however, not afford to run multiple 50~cMpc models and could therefore not calibrate to observations of groups of galaxies.
As can be seen from Figure~\ref{fig:fgas_groups}, contrary to model Ref-L100N1504, model AGNdT9-L050N0752 does appear to reproduce the observations of group gas fractions. That is, for $M_{500,{\rm hse}} < 10^{13.5}\,{{\rm M}_\odot}$ the simulation points agree with an extrapolation of the observations for high-mass systems.
There is a strong hint that the gas fraction may again become too high for more massive clusters, although with only 1 object with $M_{500,{\rm hse}}> 10^{13.5}\,{{\rm M}_\odot}$ it is hard to judge the significance of this deviation.
\citet{LeBrun2014CosmoOWLS} found that the cosmo-OWLS\ simulations, which use $2\times 1024^3$ particles in $400~h^{-1}\,{\rm cMpc}$ volumes, reproduce these and many other observations of groups and clusters over the full mass range of $10^{13}-10^{15}\,{{\rm M}_\odot}$ for $\Delta T_{\rm AGN} = 10^8\,{\rm K}$. This may seem surprising given that EAGLE\ requires higher values of $\Delta T_{\rm AGN}$. Note, however, that because the particle mass in cosmo-OWLS\ is more than 3 orders of magnitudes larger than for EAGLE, the energy in individual AGN feedback events in cosmo-OWLS\ is still much larger than that in AGNdT9-L050N0752.
Figure~\ref{fig:Lx_groups} shows the X-ray luminosity in the 0.5--2.0 keV band as a function of the temperature measured from the (virtual) X-ray data. For the reference model the agreement with the observations is reasonably good at low temperatures (the lack of simulated points with $L \ll 10^{42}\,{\rm erg}\,{\rm s}^{-1}$ is due to the fact that we only selected systems with $M_{500,{\rm hse}}> 10^{13}\,{{\rm M}_\odot}$), but the predicted luminosity is about a factor of three too high above 1 keV. Model AGNdT9-L050N0752 appears to match the data well, but more objects with $k_{\rm B}T> 1~{\rm keV}$ are needed to better assess the degree of correspondence.
\begin{figure}
\resizebox{\colwidth}{!}{\includegraphics{Lx_Tx.eps}}
\caption{The soft (0.5--2.0 keV) X-ray luminosity as a function of the X-ray temperature at $z=0$. Only points for which $M_{500,{\rm hse}}> 10^{13}\,{{\rm M}_\odot}$ are shown. Black data points represent observations of \citet{Horner2001}, \citet{Osmond2004}, \citet{Pratt2009}, and \citet{Mehrtens2012}. The reference model predicts too high X-ray luminosities for clusters above 1 keV, but simulation AGNdT9-L050N0752 is consistent with the data.}
\label{fig:Lx_groups}
\end{figure}
\subsection{Column density distributions of intergalactic metals}
\label{sec:cddf}
The galactic outflows that we invoke to reproduce observations of galaxies also disperse heavy elements into the IGM. Furthermore, the winds shock-heat the gas, which may, in turn, change its ionisation balance. Hence, it is interesting to compare the predicted distribution of intergalactic metal ions to the observations. This is a strong test for the model, since the subgrid feedback was only calibrated to match the stellar properties of galaxies.
\begin{figure*}
\resizebox{\colwidth}{!}{\includegraphics{column_EAGLEsims_CIV_lowz.eps}}
\resizebox{\colwidth}{!}{\includegraphics{column_EAGLEsims_OVI_lowz.eps}}
\caption{The column density distribution functions of \ion{C}{iv}\ (left panel) and \ion{O}{vi}\ (right panel). The coloured curves show the simulation predictions and the data points with $1\sigma$ error bars indicate observations taken with STIS/FUSE \citep{Danforth2008LowzAbs}, COS \citep{Danforth2014LowzAbs}, STIS/FUSE/GHRS \citep{Cooksey2010CIV}, and STIS \citep{Thom2008OVI, Tripp2008OVI}. The redshift ranges of the observations vary and are indicated in the legend. For clarity we only show the simulation results for $z=0.27$. The predictions are consistent with the data.}
\label{fig:cddf}
\end{figure*}
Figure~\ref{fig:cddf} compares the predicted column density distribution functions (CDDFs) of \ion{C}{iv}\ (left panel) and \ion{O}{vi}\ (right panel) with measurements derived from quasar absorption line observations, mainly from the Hubble Space Telescope (HST). Note that this prediction was completely blind.
The CDDF is conventionally defined as the number of absorbers per unit column density, $N$, and per unit absorption distance, $dX$. The number of absorbers per unit absorption distance is obtained from the quantity that is actually observed, the number of absorbers per unit redshift, via $dX = dz (H_0/H(z))(1+z)^2$. The redshift ranges of the observations vary and are indicated in the legend. All observations are for $z<1$ and most for much lower redshift. For clarity we only show the simulation results for our $z=0.27$ snapshots. However, limiting the comparison to $z=0.27$ does not affect our conclusions because the evolution is weak.
For the simulations we
compute ion fractions for each gas particle using \textsc{cloudy} photoionisation models, assuming the gas is in ionisation equilibrium and exposed to the \citet{Haardt2001UVB} model for the UV/X-ray background from galaxies and quasars. We then obtain the CDDF by projecting the simulation cube onto a 2-D grid and applying SPH interpolation to compute the ion column density in each cell. We use a grid cell size of 10~ckpc, which is sufficiently small to obtain convergence, and have verified that projection effects are negligible by comparing results obtained from simulations using different box sizes.
Observationally, the CDDF is obtained by decomposing the identified absorption features into Voigt profiles and grouping those into systems using criteria that differ between observers and that are not always well defined. We intend to mimic the observational procedures more closely in future work. From Figure~\ref{fig:cddf} it is clear that the differences between different sets of observations exceed the reported statistical uncertainties, suggesting the presence of significant systematic errors. Particularly for \ion{O}{vi}, the analysis of COS spectra by \citet{Danforth2014LowzAbs} yields systematically more absorbers than the earlier analyses of STIS/FUSE/GHRS data by \citet{Danforth2008LowzAbs}, \citet{Thom2008OVI}, and \citet{Tripp2008OVI}.
As discussed in \S\ref{sec:Zm}, even for a fixed IMF the nucleosynthetic yields are uncertain at the factor of two level \cite[e.g.][]{Wiersma2009Chemo}. This suggests that we are free to rescale the metal column densities, i.e.\ to shift the curves in Figure~\ref{fig:cddf} horizontally by up to 0.3~dex. However, doing so would break the self-consistency of the simulations as the metal abundances determine the cooling rates.
The simulation predictions generally agree well with the data, falling in between the different sets of observations, both for \ion{C}{iv}\ and \ion{O}{vi}. The simulations appear to produce too few ultra-strong absorbers, i.e.\ systems with column densities $\sim 10^{15}\,{\rm cm}^{-2}$. However, the frequency of these extremely rare systems is particularly sensitive to systematics and hence requires a more careful comparison.
For \ion{O}{vi}\ the difference between Ref-L100N1504 and Recal-L025N0752 is substantial for $N_{\ionsubscript{O}{VI}} \sim 10^{14}\,{\rm cm}^{-2}$ with the high-resolution model yielding up to a factor of 3 more absorbers. However, this does not lead to any disagreement with the data as all simulations fall in between the different sets of observations. Recall that in low-mass galaxies feedback from star formation is more effective in the recalibrated, high-resolution model Recal-L025N0752 than in the reference model. It is interesting that while this boost in the feedback efficiency decreases the metallicity of the ISM (Fig.~\ref{fig:mz}), it boosts the abundance of metal ions in the IGM. It is tempting to conclude that the more effective feedback transports more metals from galaxies into the IGM. However, whether this is the case is not clear from the results presented here due to the importance of ionisation corrections.
In future work we will compare with high-redshift data and with absorption line observations of the gas around galaxies of known mass. For now, we are encouraged by the fact that a model that was calibrated to the GSMF and galaxy sizes, also yields good agreement with observations of intergalactic metals.
\section{Summary and discussion}
\label{sec:summary}
We have introduced the EAGLE\ project, where EAGLE\ stands for ``Evolution and Assembly of GaLaxies and their Environments''. EAGLE\ consists of a suite of large, hydrodynamical cosmological simulations. In this introductory paper we have focused on a set of simulations for which the subgrid parameters for feedback were calibrated to match the observed $z\sim 0$ galaxy stellar mass function (GSMF), subject to the constraint that the galaxy sizes must also be reasonable. \citet{Crain2014EagleModels} will present models in which the subgrid physics is varied.
The largest EAGLE\ simulation, Ref-L100N1504, uses nearly 7 billion ($2\times 1504^3$) particles in a 100 cMpc box. This corresponds to an initial baryonic particle mass of $1.8\times 10^6\,{{\rm M}_\odot}$ and a force resolution of 0.7~proper kpc (smaller at high redshift), which we refer to as ``intermediate resolution''. The resolution was chosen to marginally resolve the Jeans scales in the warm ($T\sim 10^4\,{\rm K}$) ISM. The high-resolution model, Recal-L025N0752, has eight times better mass resolution and two times better spatial resolution, thus resolving a galaxy like the Milky Way with $\sim 10^6$ particles.
The simulations were run with the code \textsc{gadget}~3, but with a modified implementation of SPH, the time stepping, and the subgrid models. The simulations include subgrid prescriptions for (element-by-element) radiative cooling, star formation, stellar evolution and mass loss, energy feedback from star formation, the growth of supermassive BHs, and AGN feedback. The prescription for star formation accounts for the observation that stars form from molecular clouds and that the \HI-H$_2$ transition depends on metallicity. The subgrid model for accretion onto BHs accounts for the fact that angular momentum suppresses the accretion rate.
The most critical parts of the model are the implementations of energy feedback from star formation and AGN. We argued that present-day simulations of representative volumes cannot predict the efficiency of the feedback processes from first principles because of their reliance on subgrid models, because of spurious radiative losses due to the limited resolution, and because they lack the resolution and do not include all the physics necessary to model the structure of the ISM.
We discussed some of the implications of the inability to predict the efficiency of the feedback from first principles. We argued that current cosmological simulations can predict neither BH nor stellar masses, which implies that the subgrid models for feedback need to be calibrated to observations. Another consequence is that it is difficult to distinguish different physical feedback mechanisms that operate nearly simultaneously, such as winds driven by supernovae and radiation pressure. Furthermore, unless one can demonstrate that the model does not suffer from overcooling due to limited numerical resolution, one cannot conclude that there is a need for a new, physical feedback process just because the implemented feedback is insufficiently effective.
Because the spurious radiative losses depend on the resolution, one may have to recalibrate when the resolution is changed. We termed this ``weak convergence'' as opposed to the ``strong convergence'' that corresponds to the same physical model giving consistent results at different resolutions. However, we argued that most subgrid models for feedback effectively change with resolution even if the subgrid parameters are kept constant.
The quest for strong convergence of simulations that lack the resolution to model the ISM has led to significant sacrifices, which generally involve disabling aspects of the hydrodynamics during feedback. Examples include temporarily turning off radiative cooling, temporarily turning off hydrodynamical forces, and making the feedback efficiency dependent on dark matter velocity dispersion rather than on local properties of the gas. However, until the cooling losses can be predicted, even fully converged simulations will be unable to predict stellar and BH masses from first principles. We therefore prefer to minimize the sacrifices and to opt for weak convergence. Nevertheless, we demonstrated that the strong convergence of our model is reasonably good (Fig.~\ref{fig:gsmf_conv}).
Motivated by the above considerations, we chose to keep our subgrid models for feedback as simple as possible. We employ only one type of stellar feedback and hence we do not distinguish between stellar winds, radiation pressure, and core collapse supernovae. Similarly, we include only one type of AGN feedback and therefore do not implement separate ``quasar'' and ``radio modes''. We find that a more complex approach is not required to match observational data.
We implement both feedback from star formation and AGN thermally using the stochastic prescription of \citet{DallaVecchia2012Winds}. By injecting the energy stochastically rather than at every time step, we can specify both the temperature jump of the heated gas and the expectation value for the amount of energy that is injected. This enables us to better mimic the physical conditions associated with observed feedback processes, in particular the high heating temperatures that suppress the initial radiative losses, than would otherwise be possible given the limited resolution of the simulations. The velocities and mass loading factors of galactic winds are thus not imposed, but are an outcome of the simulation.
The temperature jump associated with feedback events is chosen to balance the need to minimize both the initial, radiative losses (which are largely numerical) and the time between feedback events (to allow for self-regulation). The probability of heating events then needs to be calibrated by comparing the simulation results for some observable to real data.
The subgrid efficiency of the AGN feedback, i.e.\ the expectation value for the amount of energy that is injected into the ISM per unit of accreted gas mass, is constant and was chosen to match the normalisation of the observed relation between the masses of galaxies and their central supermassive BHs. This parameter is, however, unimportant for observables other than the masses of BHs.
The subgrid efficiency of the feedback from star formation, $f_{\rm th}$, i.e.\ the expectation value for the amount of energy that is injected into the ISM in units of the energy available from core collapse supernovae, was chosen to reproduce the observed GSMF for $M_\ast < 10^{10.5}\,{{\rm M}_\odot}$, i.e.\ below the knee of the Schechter function. Finally, the value of the parameter that controls the sensitivity of the BH accretion rate to the angular momentum of the surrounding gas was adjusted to make the mass function turn over at the onset of the exponential drop of the observed GSMF.
We made $f_{\rm th}$ a function of both metallicity and density.
We use a physically motivated metallicity dependence with $f_{\rm th}$ dropping when the metallicity is increased from values $\ll 0.1Z_\odot$ to $\gg 0.1 Z_\odot$. This reduction in the efficiency is meant to capture the increase in radiative losses that is expected when metal-line cooling becomes important, which happens for $Z > 0.1 Z_\odot$ at the temperatures relevant for gas shock-heated in galactic winds \citep[e.g.][]{Wiersma2009Cooling}.
While a constant value of $f_{\rm th}=1$, or a pure metallicity dependence, each give an excellent fit to the GSMF, they result in galaxies that are far too compact \citep{Crain2014EagleModels}. This happens because, at the resolution of EAGLE, the stochastic implementation for stellar feedback is still subject to numerical radiative losses at high gas densities, as we demonstrated analytically. To compensate for these spurious losses, we increase $f_{\rm th}$ at high gas densities. However, $f_{\rm th}$ never exceeds 3 and the mean value is smaller than 1.1.
We compared EAGLE\ to a diverse set of observations of the low-redshift Universe, carefully distinguishing between observations that were considered during the calibration (the GSMF and thus also the directly related $M_\ast - M_{200}$ relation, galaxy sizes, and the $M_{\rm BH} - M_\ast$ relation) and those that were not. We came to the following conclusions:
\begin{itemize}
\item The observed GSMF is reproduced over the range $10^8 < M_\ast/{{\rm M}_\odot} \la 10^{11}$. At fixed mass, the difference in number density relative to the data is $\la 0.2$ dex. At fixed number density, the difference in mass is smaller than 0.3 dex (Fig.~\ref{fig:gsmf}). Even for a fixed IMF, this discrepancy is comparable to the systematic uncertainty in the observed masses due to stellar evolution alone. This level of agreement with the data is close to that obtained by semi-analytic models and is unprecedented for hydrodynamical simulations (Fig.~\ref{fig:gsmf_other}).
\item Three-dimensional apertures of 30 proper kpc, which we used throughout the paper, give results close to the Petrosian masses that are often used for observations, e.g.\ by SDSS. For $M_\ast > 10^{11}\,{{\rm M}_\odot}$ larger apertures yield higher masses (Fig.~\ref{fig:gsmf_aperture}).
\item The stellar mass - halo mass relation for central galaxies is close to that inferred from abundance matching. The efficiency of galaxy formation, $M_\ast/M_{200}$, peaks at the halo mass $M_{200} \sim 10^{12}\,{{\rm M}_\odot}$ and at the stellar mass $M_\ast \sim 10^{10.4}\,{{\rm M}_\odot}$ (Fig.~\ref{fig:eta}).
\item Disc galaxy sizes are well matched to the observations. Over the full range of stellar mass, $10^8 < M_\ast/{{\rm M}_\odot} < 10^{11.5}$, the median stellar half-mass radii of late-type galaxies agree with the observed half-light radii to within 0.1 dex (Fig.~\ref{fig:sizes}).
\item The median relation between BH mass and stellar mass agrees with the observations, but the scatter in the model is smaller than observed. The simulations predict that galaxies with total stellar masses of $10^9-10^{10}\,{{\rm M}_\odot}$ typically host BHs with masses that fall below the extrapolation of the high-mass power-law relation (Fig.~\ref{fig:bh}).
\item The predicted relation between the median specific star formation rate ($\dot{M}_\ast/M_\ast$; SSFR) and stellar mass for star-forming galaxies, i.e.\ the ``main sequence of star formation'', agrees with the observations to within 0.2 dex over the observed range of $10^9 < M_\ast/{{\rm M}_\odot} <10^{11}$ at high-resolution and to within 0.35 dex at intermediate resolution (Fig.~\ref{fig:ssfr}, left panel).
\item The predicted fraction of galaxies that are passive, which we define as SSFR $< 10^{-2}\,{\rm Gyr}^{-1}$ for the simulations, increases sharply with stellar mass between $10^{10}$ and $10^{11.5}\,{{\rm M}_\odot}$, in agreement with the observations (Fig.~\ref{fig:ssfr}, right panel).
\item The predicted median relation between the maximum of the rotation curve and stellar mass of late-type galaxies, i.e.\ a close analogue of the Tully-Fisher relation, agrees with the observations to better than 0.03 dex over the observed mass range of $10^9 \la M_\ast/{{\rm M}_\odot} <10^{11}$ (Fig.~\ref{fig:tf}).
\item The relations between ISM metallicity and stellar mass and between stellar metallicity and stellar mass are predicted to flatten with stellar mass. For the gas the predicted median metallicities agree with the observed values to within 0.1 dex for $M_\ast > 10^{9.5}\,{{\rm M}_\odot}$ at intermediate resolution and down to the lowest observed mass, $M_\ast \sim 10^{8.5}\,{{\rm M}_\odot}$, at high resolution. At lower masses the predicted relations are less steep than extrapolations of the observed trends. For the stellar metallicities the discrepancies are larger. For $M_\ast > 10^{10}\,{{\rm M}_\odot}$ all simulations agree with the data to better than 0.2~dex, but the difference increases with decreasing mass. At $M_\ast \sim 10^8\,{{\rm M}_\odot}$ the stellar metallicities in the intermediate- and high-resolution simulations are higher than observed by about 0.7 and 0.3~dex, respectively.
\item For the mass-metallicity relations the strong convergence is significantly better than the weak convergence, i.e.\ simulations that keep the subgrid parameters fixed converge better with numerical resolution than simulations for which the feedback is (re)calibrated to the $z\sim 0$ GSMF at each resolution. Hence, the increase in the efficiency of the feedback from star formation that was applied at high resolution in order to match the observed GSMF, simultaneously steepens the $Z(M_\ast)$ relations, improving the agreement with the data.
\item A comparison to observations of groups and clusters of galaxies with $M_{500,{\rm hse}} > 10^{13}\,{{\rm M}_\odot}$, where the subscript ``hse'' indicates that the quantity was estimated from virtual observations under the assumption of hydrostatic equilibrium, revealed that:
\begin{itemize}
\item The predicted relation between the total $I$-band light within $R_{500,{\rm hse}}$ and $M_{500,{\rm hse}}$ agrees with the data. Note that this includes contributions from satellites and intracluster light (Fig.~\ref{fig:Lopt_groups}).
\item The gas mass fractions, $M_{\rm gas,500,hse}/M_{\rm 500,hse}$, are overestimated by about 0.2~dex in the reference model. For $M_{\rm 500,hse} < 10^{13.5}\,{{\rm M}_\odot}$ this can be remedied by increasing the subgrid AGN heating temperature, as implemented in model AGNdT9-L050N0752. At higher masses this change may be insufficient, although larger simulation volumes are needed to confirm this (Fig.~\ref{fig:fgas_groups}).
\item The reference model predicts soft X-ray luminosities that are about 0.5~dex higher than observed for clusters with spectroscopic temperatures $\sim 1$~keV. However, model AGNdT9-L050N0752 is consistent with the observations (Fig.~\ref{fig:Lx_groups}).
\end{itemize}
\item The column density distributions of intergalactic \ion{C}{iv}\ and \ion{O}{vi}\ are in good agreement with the data, falling in between the results obtained by different surveys (Fig.~\ref{fig:cddf}).
\end{itemize}
Hence, in the resolved mass range, which spans $10^9 \la M_\ast/{{\rm M}_\odot} \la 10^{11}$ for some observables and $10^8 \la M_\ast/{{\rm M}_\odot} \la 10^{11}$ for others, EAGLE\ agrees with a diverse set of low-redshift observations of galaxies. At the same time, EAGLE\ reproduces some key observations of intergalactic metals. The only discrepancies found in this work that substantially exceed observational uncertainties concern the gas and stellar metallicities of dwarf galaxies, which are too high, and
the predictions of the reference model for X-ray observations of the intracluster medium. The metallicity problem is only substantial at intermediate resolution, so it is possible that it can be resolved simply by increasing the resolution further. We already demonstrated that the problem with groups of galaxies can be remedied by increasing the heating temperature used in the subgrid model for AGN feedback, as implemented in model AGNdT9-L050N0752, without compromising the successes of the reference model. However, larger volumes are needed to judge whether the increase in the heating temperature that was implemented in this model suffices to obtain agreement with the data for massive ($M_{500} \ga 10^{14}\,{{\rm M}_\odot}$) clusters of galaxies.
In future papers we will test many more predictions of EAGLE. Although we will undoubtedly uncover problems, so far we have no reason to believe that the results shown here are unrepresentative. We will show that the success of EAGLE\ extends to other areas that have in the past proven to be challenging for hydrodynamical simulations, such as the bimodal distribution of galaxies in colour-magnitude diagrams. We will also demonstrate that the relatively good agreement with the data is not limited to low redshift. In addition to further exploring the models that have been presented here, we plan to use the larger suite of physical models presented in \citet{Crain2014EagleModels} to gain insight into the physical processes underlying the observed phenomena. Finally, we have already begun to carry out higher-resolution resimulations of individual structures \citep[e.g.][]{Sawala2014EagleZooms,Sawala2014ChosenFew} with the code used for EAGLE.
Although the relatively good agreement between EAGLE\ and the observations, as well as that between other recent, hydrodynamical simulations of representative volumes and the data \citep[e.g.][]{Vogelsberger2014IllustrisNature}, is encouraging, we should keep in mind that we have not attempted to model many of the physical processes that may be important for the formation and evolution of galaxies. For example, EAGLE\ does not include a cold interstellar gas phase, radiation transport, magnetohydrodynamics, cosmic rays, conduction, or non-equilibrium chemistry, and EAGLE\ does not distinguish between different forms of energy feedback from star formation and between different forms of AGN feedback. We argued that at present there are good reasons for such omissions, but many of those arguments would no longer apply if the numerical resolution were increased by several orders of magnitude. While it will take some time for simulations of representative volumes to attain the resolution that is required to model the cold ISM, simulations of individual objects can already do much better. Ultimately, simulations should be able to predict the efficiencies of the most important feedback processes and hence to predict, rather than calibrate to, the GSMF.
We hope that EAGLE\ will prove to be a useful resource for the community.\footnote{We intend to make the simulation output public in due course, starting with galaxy properties, which we will make available using the SQL interface that was also used for the Millennium simulation \citep{Lemson2006Database,Springel2005Millennium}. Details will be provided on the EAGLE\ web site, see \url{http://eagle.strw.leidenuniv.nl/}.}
The agreement with observations is sufficiently good for the simulations to be used in ways that have so far been reserved for semi-analytic models of galaxy formation. At the same time, because hydrodynamical simulations provide much more detailed 3-D information, make fewer simplifying assumptions, and simultaneously model the galaxies and the IGM, EAGLE\ enables one to ask many questions that are difficult to investigate with semi-analytic models.
\section*{Acknowledgments}
We would like to thank Volker Springel for sharing his codes and for his advice and discussions. We gratefully acknowledge discussions with Jarle Brinchmann, Shy Genel, Justin Read, Debora Sijacki. We are also thankful to Martin Bourne and Laura Sales for their contributions to the initial phase of the project, Amandine Le Brun for her help with the X-ray plotting routines, Peter Draper and Lydia Heck for their help with the computing resources in Durham, and to Wojciech Hellwing for help with computing in Poland. We are also grateful to all the people working on the analysis of the EAGLE\ simulations and would like to thank the anonymous referee for a constructive report.
This work used the DiRAC Data Centric system at Durham University, operated by the Institute for Computational Cosmology on behalf of the STFC DiRAC HPC Facility (www.dirac.ac.uk). This equipment was funded by BIS National E-infrastructure capital grant ST/K00042X/1, STFC capital grant ST/H008519/1, and STFC DiRAC Operations grant ST/K003267/1 and Durham University. DiRAC is part of the National E-Infrastructure. We also gratefully acknowledge PRACE for awarding us access to the resource Curie based in France at Tr\`es Grand Centre de Calcul. This work was sponsored by the Dutch National Computing Facilities
Foundation (NCF) for the use of supercomputer facilities, with financial support from the Netherlands Organization for Scientific
Research (NWO) and by the HPC Infrastructure for Grand Challenges of Science and Engineering Project, co-financed by the European Regional Development Fund under the Innovative Economy Operational Programme and conducted at the Institute for Mathematical and Computational Modelling at University of Warsaw. The research was supported in part by the European Research Council under the European Union's Seventh Framework
Programme (FP7/2007-2013) / ERC Grant agreements 278594-GasAroundGalaxies, GA 267291 Cosmiway, and 321334 dustygal, the Interuniversity Attraction Poles Programme initiated by the Belgian Science Policy Office ([AP P7/08 CHARM]), the National Science Foundation under Grant No.\ NSF PHY11-25915, the UK Science and Technology Facilities
Council (grant numbers ST/F001166/1 and ST/I000976/1), Rolling and Consolodating Grants to the ICC, Marie Curie Reintegration Grant PERG06-GA-2009-256573.
\bibliographystyle{mn2e}
|
\section{Results: Liquid-drop impact dynamics}
In our experiments, we release a stationary water drop of diameter, $D$, ranging from 1.4 mm to 4.6 mm from a height $h$. $D$ is chosen to represent the size range of natural raindrops \cite{5,27}. The drop falls vertically in air onto a granular bed comprised of $d_{sand} = 90\pm15$ $\mu$m glass beads with volume fraction $\phi = 0.60$. To adjust the range of impact energy, $E$, we vary $h$ from 1.8 mm up to 12 m, allowing a 4.6 mm drop to reach $98\%$ of its terminal velocity (see Materials and Methods).
The dynamics of liquid-drop impact on a granular surface are captured using high-speed photography as illustrated in Fig.~\ref{Figure1} for the strike of a water drop at three different $E$ (Supplementary Movie S1, S2 and S3). Upon impact, the drop penetrates into the top layer of the granular bed (Fig.~\ref{Figure1}A, F, K). After the initial impact, the drop can be treated as an incompressible fluid. The downward motion of the top part of the drop causes drop deformation and spreading.
At low $E$, the spreading liquid lamella moves horizontally along the surface of a shallow crater (Fig.~\ref{Figure1}B). The lamella retracts after reaching the maximum spreading diameter and entrains a layer of granular particles on its surface. Since the lamella's surface-to-volume ratio reduces as it recedes, particles at the interface are gradually pushed into the liquid bulk, resulting in a ``liquid marble'' armored with a thick layer of granular particles (Fig.~\ref{Figure1}C, D)\cite{24}. Above $E = 1.9\times10^{-6}$ J, the marble can even bounce off the granular bed (Fig.~\ref{Figure1}D). The jumping height of the marble is non-monotonic with increasing $E$. As the spreading diameter increases, the lamella traps more particles, which increases the weight of the marble and reduces the jumping height.
Increasing $E$ further, the rim of the spreading lamella develops a fingering instability (Fig.~\ref{Figure1}G). After reaching the maximum spreading diameter, the fingers start to retract and gradually push particles at interface into the bulk (Fig.~\ref{Figure1}H, I). The process continues until the concentration of particles within the retracting lamella becomes so high that the receding motion is completely arrested due to the jamming of particles. The jamming transition occurs before the fingers can fully retract back to a sphere, which leads to an asymmetric liquid marble with finger protrusions on its surface (Fig.~\ref{Figure1}I, J). The length of these protrusions increases with $E$. At this intermediate $E$, the marble stops bouncing off the surface.
At even larger $E$, a water crown is formed along the wall of a deep crater (Fig.~\ref{Figure1}L, M). The crown detaches from the granular surface at the edge of the crater. Above $E = 9.7\times10^{-5}$ J, the rim of the crown becomes unstable and disintegrates into secondary droplets (Fig.~\ref{Figure1}N). This violent splashing process effectively mixes granular particles with the liquid. Finally, the crown, fully loaded with particles, retracts and falls flat on the surface (Fig.~\ref{Figure1}O).
The dynamics of liquid-drop impacts illustrated by high-speed photography provide essential information for understanding the morphology of liquid-drop impact craters. Based on the dynamics, we will develop a simple theoretical understanding of various features of liquid-drop impact craters in the Theory section. Before that, we shall first show our experimental results on the morphology of liquid-drop impact craters.
\section{Results: Morphology of impact craters}
After impact, water gradually drains into the granular bed and various fascinating crater topologies are observed at the end (Fig.~\ref{Figure2}A-F). To fully characterize the morphology of liquid-drop imprints, we need to consider three main features of impact craters, i.e., the diameter of impact craters, the depth of impact craters and the granular residues left in the center of impact craters.
\begin{figure}[ht]
\centerline{\includegraphics[width=0.95\linewidth]{Fig2.eps}}
\caption{Morphology of liquid-drop impact craters. Impact craters from the strike of a 3.1 mm water drop with $E = 9.7\times10^{-7}$ J (A), $6.5\times10^{-6}$ J (B), $3.2\times10^{-5}$ J (C), $6.0\times10^{-5}$ J (D), $8.2\times10^{-5}$ J (E) and $3.0\times10^{-4}$ J (F). Scale bars: 3.0 mm. (A) and (B) Ring-shaped granular residues at low $E$. (C) Solid pellet-shaped granular residue at intermediate $E$. (D) and (E) Asymmetric granular residues. (D) marks the transition between the low and high $E$ regime in Fig.~\ref{Figure5}. (F) Splash pattern at high $E$. $D_c$ and $D_g$ are defined in (C).}
\label{Figure2}
\end{figure}
\subsection{Diameter of impact craters}
We characterize the size of an impact crater by measuring its diameter, $D_c$ (Fig.~\ref{Figure2}C). Plotting $D_c$ versus $E$ reveals a power-law scaling with an exponent of $0.17\pm0.01$ (Fig.~\ref{Figure3}A), consistent with Nefzaoui and Skurtys's result \cite{20}. This scaling is visibly different from the 1/4 power-law scaling associated with the impact craters created by low-speed solid spheres. The 1/4 scaling of solid-sphere impacts arises when $E$ lifts granular particles of volume $\sim D_c^3$ to a height of $\sim D_c$ against the gravity \cite{10,11,23}.
Surprisingly, the 0.17 scaling is quantitatively similar to the Schmidt-Holsapple (S-H) scaling from hypervelocity impact cratering associated with asteroid strikes \cite{23}:
\begin{eqnarray}
\label{equ1} D_c & \sim & g^{-0.17}\cdot D^{0.83}\cdot U^{0.34} \nonumber \\
& \sim & (\rho g)^{-0.17}\cdot D^{0.32}\cdot E^{0.17},
\end{eqnarray}
where $U$ is the impact velocity of the projectile, $\rho$ is the density of projectile and $g$ is the gravitational acceleration. $\rho$ emerges in Eq.~\ref{equ1} when we convert $U$ into the impact energy $E$. Eq.~\ref{equ1} inspired us to apply the full S-H scaling to our data. Remarkably, we find that the variation of $D_c$ with different $D$ collapses to a constant, $C=D_c/\left((\rho g)^{-0.17}D^{0.32}E^{0.17}\right)=1.74\pm 0.15$ (Fig.~\ref{Figure3}B). Moreover, we tested the scaling using nine different liquids and seven different granular particles at two different ambient pressures. The results all conform to Eq.~\ref{equ1} (Supporting Information (SI) Fig. S1). Particularly, the $D_c$ scaling is independent of or only weakly depends on liquid properties such as density, viscosity or surface tension.
\subsection{Depth of impact craters}
The quantitative similarity between liquid-drop impact cratering and asteroid impact cratering also extends to the aspect ratio of their impact craters, $\alpha \equiv d_c/D_c$. Here, $d_c$ is the depth of crater, defined as the vertical distance between the rim and the bottom floor of the crater.
Previous studies reported the depth of crater in the presence of granular residues \cite{17,21}, which, however, does not reflect the true bottom of a crater underneath the granular residues. Here, to detect $d_c$ without the optical obstruction of granular residues in the center of crater, we focus on the range of $E$ where the liquid marble bounces off the surface (Fig.~\ref{Figure1}D). The landing of the marble does not trigger further granular avalanche and, therefore, does not modify the crater depth at later times (Movie S1). Even though $E$ with jumping marbles does not cover the full dynamic range of our experiments, it still extends for almost two decades, allowing us to measure $d_c$ in a sufficient range comparable to other impact cratering experiments \cite{11,28}.
Within this $E$ range, $d_c$ increases linearly with $D_c$, which leads to a constant crater aspect ratio $\alpha=0.20\pm 0.01$ (Fig.~\ref{Figure4}). As a comparison, simple craters from the Moon, Mars and Mercury also show an aspect ratio $\alpha=0.20 \pm 0.03$ \cite{29}. Even though there is a seven-order-of-magnitude difference in lengths, liquid-drop impact craters and planetary craters show the same aspect ratio within experimental errors (Fig.~\ref{Figure4}). The angle of repose of granular materials, $\theta_r$, sets an upper limit for $\alpha$. For $\theta_r = 26^\circ$---the angle measured from our experiments---we have $\alpha \lesssim \tan\theta_r/2 = 0.24$. However, the geometrical factor alone is not sufficient to explain the similarities and differences between impact cratering processes. The aspect ratio of impact craters from low-speed solid-sphere impacts is 0.12, substantially smaller than that of liquid-drop impact craters (Fig.~\ref{Figure4} inset). With strong scattering around 0.16, the aspect ratio of impact craters from hypervelocity solid-sphere impact experiments partially overlaps with that of liquid-drop impact craters (Fig.~\ref{Figure4}).
A theoretical understanding of the scaling of the diameter and the depth of impact craters and a discussion on the similarity between liquid-drop impact cratering and asteroid impact cratering will be presented below in the Theory section.
\subsection{Granular residues}
Finally, we also measure the size of granular residues, $D_g$, in the center of impact craters (Fig.~\ref{Figure2}C). $D_g$ as a function of $E$ exhibits two different regimes (Fig.~\ref{Figure5}). At low $E$, $D_g$ slowly increases. Above certain threshold impact energy, $E^*$, it starts to enlarge strongly and merges into a master curve.
\begin{figure}[ht]
\centerline{\includegraphics[width=0.9\linewidth]{Fig3.eps}}
\caption{Scaling of liquid-drop impact craters. (A) $D_c$ versus $E$ for different drop sizes. Solid lines indicate the 0.17 scaling. The dashed line indicates the 1/4 scaling. (B) Scaled $D_c$ following the S-H scaling rule (Eq.~\ref{equ1}): $D_c/((\rho g)^{-0.17}D^{0.32}E^{0.17})$. The dashed line indicates 1.74.}
\label{Figure3}
\end{figure}
\begin{figure}[ht]
\centerline{\includegraphics[width=1.0\linewidth]{Fig4.eps}}
\caption{Aspect ratio of liquid-drop impact craters. $d_c$ versus $D_c$ for four different impact cratering processes. Group (1) is from astronomical observations of asteroid impact craters on different planetary bodies \cite{29}. Group (2) is from hypervelocity solid-sphere impact experiments \cite{28}. Group (3) is from low-speed solid-sphere impact experiments \cite{11}. Group (4) is from liquid-drop impacts with $D$ = 3.9 mm (blue diamonds), 3.1 mm (green triangles), 2.6 mm (red disks), 1.4 mm (dark yellow squares). Circles are for $D$ = 3.1 mm impacting at one tenth of the atmospheric pressure. Insets show $d_c/D_c$ of liquid-drop impact craters. The upper and lower line indicate the aspect ratio of planetary impact craters (0.20) and low-speed solid-sphere impact craters (0.12) respectively.}
\label{Figure4}
\end{figure}
The trend of $D_g$ can be qualitatively understood based on the impact dynamics. As shown in Fig.~\ref{Figure1}, a liquid marble coated with a layer of granular particles is formed at low $E$ during impacts. The thickness of the granular layer depends on the number of entrained particles. The liquid phase of the marble eventually drains into the granular bed and particles are left as a granular residue. With a small $E$, particles cover only the surface of the marble. Hence, when the liquid drains into the bed, a ring of particles is left (Fig.~\ref{Figure2}A). Since the maximal spreading diameter of the impinging drop increases with $E$, at larger $E$, an increasing number of particles are entrained at the lamella interface and pushed into the bulk of the liquid marble, which leads to a liquid marble with a thicker layer of granular particles. As a result, the hole at the center of the ring-shaped residues gradually fills up (Fig.~\ref{Figure2}B). At $E$ close to the transition impact energy $E^*$, particles completely saturate the marble, which leaves a solid pellet-shaped residue in the crater (Fig.~\ref{Figure2}C). Increasing $E$ above $E^*$, the receding lamella cannot fully restore back to a spherical shape due to the jamming of particles inside the marble (Fig.~\ref{Figure1}F-J), which gives rise to a flat asymmetric granular residue quickly enlarging the measurement of $D_g$ (Fig.~\ref{Figure2}D-F). A model based on the above picture will be constructed in the next section to quantitatively describe the size of granular residues.
\section{Theory and discussion}
\subsection{Understanding the S-H scaling}
Theoretical understanding of the S-H scaling in asteroid impacts is solely based on similarity and dimensional analyses independent of detailed dynamics of impact cratering processes \cite{23,30,31}, which can thus be equally applied for liquid-drop impact cratering. However, dimensional analysis alone cannot reveal physical mechanisms associated with the liquid-drop impact process. Hence, it is more useful to look into the physical picture derived from the studies of asteroid impact cratering, which may help to explain the origin of the S-H scaling in liquid-drop impact cratering. During asteroid impacts, a large fraction of $E$ (over $97\%$ for high-velocity impacts) dissipates into heat rather than transferring to the kinetic energy of ejecta \cite{30}. The conversion efficiency of $E$ into the kinetic energy is determined by the impact pressure \cite{30,31,32}. Such a large energy partitioning is believed to give rise to the S-H scaling, which expresses a mixture of energy and momentum scaling with the power exponent between 1/4 and 1/7 \cite{30,31,32}.
Large energy partitioning also occurs during liquid-drop impacts. Only a small fraction of $E$ converts into the kinetic energy of particles, while the rest turns into the surface energy and viscous dissipation of spreading lamella. Since both the surface energy and the viscous dissipation increase with the maximal contact surface between the lamella and the granular bed ($\sim \pi D_c^2$), a larger $\pi D_c^2$ leads to a lower energy conversion. Thus, we propose a simple formula for the coefficient of energy conversion: $f = (\pi D^2/\pi D_c^2)$, where $\pi D^2$ provides the only relevant area for normalization. The fraction of energy for ejecting particles is then $E_{eject} \equiv f\cdot E$ with $f < 1$ automatically satisfied by construction. $E_{eject}$ is consistent with recent experiments that estimate the momentum of ejected particles \cite{21}. Finally, an energy scaling argument similar to that used for solid-sphere impacts can be applied: instead of $E$, $E_{eject}$ lifts granular particles in a crater of volume $V_c$ to a height determined by $d_c$, i.e., $E_{eject} \approx \phi \rho_{sand}V_cgd_c$, where $\phi = 0.60$ is the volume fraction of the bed and $\rho_{sand}$ is the particle density. If we approximate the crater as a paraboloid and replace $d_c=\alpha D_c$, then $V_c = \pi \alpha D_c^3/8$. Taken together, we successfully recover the S-H scaling:
\begin{eqnarray}
D_c \approx \left(\frac{\pi}{8} \alpha^2\phi \frac{\rho_{sand}}{\rho}\right)^{-1/6} [(\rho g)^{-1/6} D^{1/3} E^{1/6}].
\label{equ2}
\end{eqnarray}
Moreover, with $\alpha = 0.20 \pm 0.01$ for liquid-drop impact craters (Fig.~\ref{Figure4}), we have the dimensionless prefactor $C \equiv \left(\frac{\pi}{8} \alpha^2\phi \frac{\rho_{sand}}{\rho}\right)^{-1/6} = 1.86 \pm 0.04$, quantitatively matching our measurement $C = 1.74 \pm 0.15$ (Fig.~\ref{Figure3}B). Hence, the scaling analysis provides a quantitative description for both the diameter and the depth of the liquid-drop impact craters.
\begin{figure}[h]
\centerline{\includegraphics[width=0.9\linewidth]{Fig5.eps}}
\caption{Morphology of granular residues, $D_g$ versus $E$. Crater morphologies shown in Fig.~\ref{Figure2}A-F are indicated. Stars mark the transition impact energy $E^*$ between the low and high $E$ regime for each drop size. Solid lines are from the liquid marble model (Eq.~\ref{equ3}). The dashed line is $D_g(E^*)$ calculated by combining the liquid marble model (Eq.~\ref{equ3}) with the jamming criterion (Eq.~\ref{equ5}).}
\label{Figure5}
\end{figure}
\subsection{Discussion on the analogy between liquid-drop impact cratering and asteroid impact cratering}
It should be clear from the above derivation that the energy partitioning of liquid-drop impact cratering and asteroid impact cratering shares a quantitative similarity. The forms of energy dissipation in the two processes are obviously different. For asteroid impacts, the impact energy is primarily dissipated by shock-wave heating of the asteroid and the target during the initial stage of the impact event \cite{32}. For liquid-drop impacts, it dissipates mainly through the deformation and viscous dissipation of the liquid drops. However, the ratio of the energy dissipation over the total impact energy seems to follow the same quantitative trend in the two processes. Hence, it would be interesting to check if the energy conversion coefficient of asteroid impact cratering is also inversely proportional to the surface area of impact craters, i.e., $f\sim 1/D_c^2$. Without shock-wave heating or projectile deformation, a large energy partitioning does not occur in low-speed solid-sphere impact cratering. Most of the impact energy is thus directly converted into the kinetic energy of granular particles for creating impact craters, which leads to the 1/4 power as dictated by the energy scaling \cite{10,11}.
Finally, it is also interesting to compare liquid-drop impact cratering and asteroid impact cratering more generally in terms of hydrodynamic similarity and the states of matter. Firstly, it is known that the important dimensionless number governing asteroid impact cratering is the inverse Froude number, Fr$^{-1}=gD/2U^2$ \cite{23}. For typical asteroid impacts, $10^{-6}<$ Fr$^{-1}$ $<10^{-2}$, which overlaps well with our liquid-drop impact experiments $2\times10^{-4}<$ Fr$^{-1}$ $<0.1$. Secondly, in studying asteroid impact cratering, the impacted surface is frequently modeled as a Bingham fluid \cite{32}. On the other hand, granular materials typically display Bingham fluid behavior \cite{33}. More importantly, during asteroid strikes, the impact pressure can rise as high as $10^3$ GPa and the temperature may increase above 2000 $^\circ$C. Under such extreme conditions, asteroids of normal composition have already been liquefied if not vaporized \cite{32}. Hence, liquid drops provide a better model than solid spheres for high-energy asteroids. This important analogy has been overlooked in many previous attempts in search of the link between asteroid impact cratering and low-speed solid-spheres impact cratering \cite{11,12,15,16,34,35}.
\subsection{Model for granular residues}
The model for granular residues can be divided into two parts: (1) Based on the liquid marble model, we will show a quantitative understanding of the size of granular residues at low $E$. (2) Employing the concept of the jamming transition, we will calculate the transition energy $E^*$ between the low and high energy regimes (Fig.~\ref{Figure5}).
(1) As shown previously, the slow increase of $D_g$ at low $E$ is due to the formation of liquid marbles (Fig.~\ref{Figure2}A-C). $D_g$ in this regime is equal to the diameter of liquid marbles. A simple model can thus be constructed based on the liquid marble model proposed by Aussillous and Qu{\'e}r{\'e} \cite{24}. Firstly, the number of entrained particles at the lamella-granular bed interface, $N$, is proportional to the maximal contact area between the lamella and the bed. Therefore, $N\approx (\pi D_c^2/\pi d_{sand}^2)$. The volume of the liquid marble is simply the sum of the volume of the drop and the volume of entrained particles: $V_{m}=V_{drop}+V_{sand}=\pi D^3/6+N\pi d_{sand}^3/6=\pi D^3/6+\pi d_{sand}D_c^2/6$. If we assume the marble is spherical, then the effective diameter of the liquid marble is $D_m = (6V_m/\pi)^{1/3}$. For $D_m \ll \kappa^{-1}$, the liquid marble maintains a spherical shape, where $\kappa^{-1} = (\gamma/\rho_m g)^{1/2}$ is the capillary length, $\gamma$ is the surface tension of the liquid and $\rho_m$ is the density of the liquid-granular mixture. The diameter of the liquid marble and, therefore, the diameter of the granular residue is simply $D_g = D_m = (6V_m/\pi)^{1/3}$. However, for $D_m \gg \kappa^{-1}$, the marble deforms into a puddle under the force of gravity. The thickness of the puddle is given by $2\kappa^{-1}$. If we approximate the shape of the puddle as an oblate ellipsoid, then the diameter of the marble is given by $D_g = (3V_m/\pi\kappa^{-1})^{1/2}$. In summary, we have
\begin{equation}
D_g = \left\{
\begin{array}{rl}
C_1\cdot(6V_m/\pi)^{1/3} & \text{if } D_m \ll \kappa^{-1},\\
C_2\cdot(3V_m/\pi\kappa^{-1})^{1/2} & \text{if } D_m \gg \kappa^{-1}.
\end{array} \right.
\label{equ3}
\end{equation}
where we add two proportionality constants $C_1$ and $C_2$ to account for the fact that $D_m$ is close to $\kappa^{-1}$ between the two limiting cases and the approximation taken for the shape of the puddle. Replacing $D_c$ in $V_m$ using the S-H scaling (Eq.~\ref{equ2}), we finally reach $D_g(E)$. The results quantitatively agrees with our measurements (solid lines in Fig.~\ref{Figure5}) with the fitting parameters $C_1 = 1.1$ and $C_2 = 1.55 \pm 0.15$ on the order of one.
(2) Increasing $E$ further, at the transition impact energy $E^*$, the retraction of lamella is arrested before it can fully restore back to a sphere, which leads to asymmetric granular residues with quickly enlarging $D_g$ and results in a crossover from the low-energy ``liquid marble'' regime to the high-energy regime (Fig.~\ref{Figure5}). As discussed previously, the resistance against the capillary retraction comes from the jamming of entrained particles. Thus, we can identify $E^*$ as the ``jamming energy'' of liquid-drop impact process. Note that the particles entrained at the liquid interface are gradually pushed into the interior of the receding liquid lamella due to the strong capillary retraction. Hence, the jamming occurs in the bulk of liquid marble rather than only at its interface \cite{36}.
\begin{figure}[h]
\centerline{\includegraphics[width=0.7\linewidth]{Fig6.eps}}
\caption{Transition energy $E^*$ versus drop size $D$. Solid line is based on the jamming criterion (Eq.~\ref{equ5}).}
\label{Figure6}
\end{figure}
A simple analysis based on the liquid marble model can show that the number of entrained particles at the lamella-granular interface is not sufficient to jam the liquid marble at $E^*$. To reach the jamming transition, the effect of liquid imbibition during the impact needs to be considered \cite{19}. We estimate the volume of imbibed liquid into the bed, $V_{imb}$, based on the well-established Washburn-Lucas equation \cite{37}, which leads to the following equation specifically for liquid-drop impact cratering (SI):
\begin{eqnarray}
V_{imb}=0.058A\cdot (\eta^2\gamma)^{-1/4}\cdot \rho^{1/4}d_{sand}\cdot D^{5/4}\cdot E^{1/2},
\label{equ4}
\end{eqnarray}
where $\eta$ is the liquid viscosity and $A$ is a proportionality constant of order one. The jamming transition at $E^*$ can then be expressed as:
\begin{eqnarray}
\frac{V_{sand}}{V_m-V_{imb}} = \frac{\pi d_{sand}D_c^2/6}{\pi D^3/6 +\pi d_{sand}D_c^2/6 - V_{imb}} = \phi_c
\label{equ5}
\end{eqnarray}
with the jamming volume fraction $\phi_c \approx 0.55$ \cite{25,26}. Using the S-H scaling for $D_c$, Eq.~\ref{equ5} quantitatively agrees with our measurement on $E^*$ for different $D$ with the fitting constant $A = 1.19 \pm 0.22$ (Fig.~\ref{Figure6}). Finally, in combination with the liquid marble model, we also reach $D_g(E^*)$ (dashed line in Fig.~\ref{Figure5}) (see SI for additional comments).
\section{Conclusions}
When a liquid drop impacts on a granular surface, the impact energy is converted into the surface energy of the deformed drop, the internal energy of liquid and particles, and the kinetic energy of the spreading lamella and ejected particles. The process is notoriously complicated, involving high Reynolds hydrodynamics, shock compression in the impinging drop, fast granular flows and capillary interactions between fluid and granular particles. Given the complexity, it is surprising that the simple model presented here can quantitatively captures the morphology of liquid-drop impact craters over a large range of impact energy. Such a model will be considerable useful for predicting the outcome of raindrop impacts on granular media---a ubiquitous process occurring in numerous natural, agricultural and industrial circumstances.
Moreover, our study reveals a quantitative similarity between raindrop impact cratering and asteroid impact cratering in terms of both the energy scaling and the aspect ratio of their impact craters. Comparing with extensively-studied low-speed solid-sphere impact cratering, liquid-drop impact cratering provides a better analogy to high-energy asteroid impact cratering. Apparently, one should be very cautious when drawing a close link between the two processes. $E$ of an asteroid is on the order of $10^{15}$ J \cite{32}, while the maximal $E$ of liquid drops is $10^{-3}$ J. The 18-order-of-magnitude energy difference undoubtedly activates different physical processes. Nevertheless, the remarkable similarity between the two processes indicates that they may share common mechanisms that are worth further investigation.
\begin{materials}
For all the experiments and data presented in the main text, we used deionized water as our liquid drops and 90 $\pm$ 15 $\mu$m soda-lime glass beads ($\rho = $2.52 g/cm$^3$, MoSci) as our granular particles. In Supporting Information (SI), we also tested glass particles of different sizes (45.5 $\pm$ 7.5 $\mu$m, 215 $\pm$ 35 $\mu$m, 427 $\pm$ 73 $\mu$m, 725 $\pm$ 125 $\mu$m) and wetting properties to verify the S-H scaling. Moreover, in SI, we used several different liquids including methanol, ethylene glycol, mineral oil, water-glycerin mixtures and sodium dodecyl sulfate solutions as liquid drops to probe the effect of liquid viscosity and surface tension on the impact cratering process. A brief discussion of liquid-drop impact cratering on wet granular bed is also presented in SI.
A Photron SA-X2 camera was used for high-speed imaging of drop impact dynamics. The morphology of impact craters were measured using a high-precision laser profilometer (Kenyence LJ-V7060) with the resolution in the x-y plane at 20 $\mu$m and the resolution in depth at 0.4 $\mu$m. The camera and the profilometer was further combined to monitor the depth of crater during impacts. Experiments in the terminal velocity regime were conducted in an indoor laboratory with a high-height experimental platform. To prevent perturbation from air flows that lead to uncontrollable impact positions, we set up a PVC tube of 11.5 m in length and 20 cm in diameter. A free falling drop travels inside the tube before it impacts on a granular bed underneath the bottom opening of the tube. The release heights in previous investigations are all below 3 m \cite{17,18,19,20,21}, which seriously limits the dynamic range of impact energy and thus the accuracy of the scaling relationship. Finally, we also performed one set of experiments at one tenth of the atmospheric pressure to test possible effects of ambient air on the dynamics of liquid-drop impact cratering. The ambient air has been shown to play a significant role in liquid-drop impacts on solid surfaces \cite{8,9}.
\end{materials}
\begin{acknowledgments}
We thank J. Hong and W. Suszynski for the help with experiments and J. Melosh for the comments on asteroid impact cratering. We also thank F. Bates, K. Dorfman, L. Francis, S. Kumar, L. Xu and L.-N. Zou for suggestions on the paper. R.Z and Q.Z. acknowledge support from UMN UROP program.
\end{acknowledgments}
|
\section{Introduction} \label{sec-ku}
Multivariate phase-type distributions have been a topic of research interest for quite some time. The first constructive proposal of such a class (henceforth denoted by MPH) can be found in \cite{AL84}. This class was later extended to MPH$^*$ in \cite{Ku89}, while the latest (and perhaps final) proposal of a definition (denoted by MVPH) is given in \cite{BN10}. All of the proposals carry their distinctive problems, be it that they seem too limited (as MPH) or that even elementary descriptions like distribution functions are not given explicitly (for MPH$^*$).
The purpose of the present paper is the derivation of semi-explicit expressions for the density function of bivariate MPH$^*$ - distributed random variables. More exactly, for $(Z_1,Z_2) \in \text{MPH}^*$ we shall derive a simple expression for $\mathbb{E} (e^{-sZ_2} 1_{ \{ Z_1 \in dy \} })$, i.e.\ a density function for $Z_1$ joint with a Laplace transform for $Z_2$. As a univariate Laplace transform, this can be readily inverted to yield the bivariate density function $\prob( Z_1 \in dy, Z_2 \in dx)$ for $x,y > 0$, see e.g.\ \cite{AW95}.
For ease of reference, we shall use the remainder of this introduction to restate the pertinent results in Kulkarni's construction of the class MPH$^*$. The main result along with some remarks are then presented in section 2.
Let ${\cal J} = ( J_t: t \geq 0)$ denote a Markov process on a finite state space $E' := \{ 1, \ldots, m+1 \}$ with $m \in \mathbb{N}$, having generator matrix
\[ \begin{pmatrix} Q & - Q {\bf 1} \\ {\bf 0} & 0
\end{pmatrix}
\]
where $Q$ is invertible, i.e.\ the states $i \in E := \{ 1, \ldots, m \}$ are transient. The initial distribution of ${\cal J}$ is denoted by $(\alpha, \alpha_{m+1})$ with $\alpha = (\alpha_1, \ldots, \alpha_m)$ and $\alpha_i := \prob (J_0 = i)$ for $i \in E'$. We assume of course $\alpha_{m+1} < 1$. Let $R = (r_{ij})_{i \leq k, j \leq m}$ denote a reward matrix of dimension $k \times m$, with $r_{ij} \geq 0$ for all $i,j$. Write also $r_i(j) := r_{ij}$ whenever it is more convenient. Define the time of absorption of ${\cal J}$ by
\begin{equation} \label{def-tau}
\tau := \min \{ t \geq 0: J_t = m+1 \}
\end{equation}
and further the random variables
\begin{equation} \label{def-Z}
Z_i := \int_0^\tau r_i(J_t) dt
\end{equation}
for $i \in \{ 1, \ldots, k \}$. Then we say that $(Z_1, \ldots, Z_k) \in MPH^*$. The distribution of $(Z_1, \ldots, Z_k)$ shall be denoted by $MPH^*( \alpha, Q, R)$. To avoid trivial singularities later on, we assume that $\sum_{j=1}^m r_{ij} >0$, i.e.\ $\prob(Z_i >0) >0$ for all $i \in \{ 1, \ldots, k \}$.
\section{The bivariate phase-type distribution}
From now on we set $k=2$, i.e.\ we consider bivariate distributions in MPH$^*$ only. The plan is the following: By theorem 1 in \cite{Ku89} the marginal distribution of $Z_1$ is phase-type. As indicated in remark 1 of \cite{Br12a}, it can be characterised in terms of the first passage times for a fluid flow. To be a bit more precise, let
\begin{equation} \label{def-fpt}
\tau(y) := \inf \{ t \geq 0: Y_t > y \}
\end{equation}
denote the first passage times for a suitable fluid flow model $({\cal J}, {\cal Y})$. Then
\[ \prob( Z_1 > y ) = \prob_\alpha \left( \tau(y) < \tau | Y_0 = 0 \right)
\]
where $\tau$ is the same as in (\ref{def-tau}) and $\prob_\alpha$ denotes the conditional probability given that $\prob( J_0 = i ) = \alpha_i$ for $i \in E$. Now we attach a phase-dependent time devaluation along the path of ${\cal Y}$ up to $\tau(y)$ to obtain an expression for
\[ \mathbb{E} \left( e^{- s \int_0^{\tau(y)} r_2(J_s) ds} \right)
\]
which is the Laplace transform of $Z_2$ (with argument $s$) on the set of paths that satisfy $\tau(y) < \tau$, i.e.\ $Z_1 > y$. From here it is only a small step to obtain an expression for $\mathbb{E} ( e^{-s Z_2} 1_{ \{ Z_1 \in dy \} })$.
Two-dimensional fluid flow models have been analysed in detail in \cite{BR13}. We shall make use of some of the results therein, adapted to the question investigated here. In order to do so, we need to introduce some more notation. First we define the fluid flow models $({\cal J}, {\cal Y})$ and $({\cal J}, {\cal X})$ by
\begin{equation} \label{def-XY}
Y_t := \int_0^t r_1(J_s) \; ds \qquad \text{and} \qquad X_t := \int_0^t r_2(J_s) \; ds
\end{equation}
for all $t \geq 0$, where the phase process ${\cal J}$ is the same as in section \ref{sec-ku} and
\[ r_1(m+1) := r_2(m+1) := 0
\]
Partition the set $E$ of transient states into $E=E_0 \cup E_+$, where
\[ E_0 := \{ i \in E: r_{1i} =0 \} \qquad \text{and} \qquad E_+ := \{ i \in E: r_{1i} > 0 \}
\]
According to this partition, write $Q$ and $\alpha$ in block form, i.e.\
\[ Q = \begin{pmatrix} Q_{00} & Q_{0+} \\ Q_{+0} & Q_{++} \end{pmatrix} \qquad \text{and} \qquad \alpha = (\alpha_0, \alpha_+)
\]
Further write $(\eta_0, \eta_+)' := \eta := -Q {\bf 1}$. Finally, define the diagonal matrices
\[ R_+ := diag( r_{1i}: i \in E_+),
\]
\[ D_+ := diag( r_{2i}: i \in E_+) \qquad \text{and} \qquad D_0 := diag( r_{2i}: i \in E_0).
\]
Now we can state the main result:
\begin{Thm}
Let $(Z_1, Z_2) \sim MPH^*( \alpha, Q, R)$. Then
\[ \mathbb{E} ( e^{-s Z_2} 1_{ \{ Z_1 =0 \} }) = \alpha_0 (s D_0 - Q_{00})^{-1} \eta_{0}
\]
for $s \geq 0$ and
\[ \mathbb{E} ( e^{-s Z_2} 1_{ \{ Z_1 \in dy \} }) = \alpha(s) e^{W(s) y} \eta(s) \; dy
\]
for $y > 0$ and $s \geq 0$, where
\begin{align*}
\alpha(s) &:= \alpha_0 (s D_0 - Q_{00})^{-1} Q_{0+} + \alpha_+ \\
W(s) &:= R_+^{-1} \left( (Q_{++} - s D_+) - Q_{+0} (Q_{00} - s D_0)^{-1} Q_{0+} \right) \\
\eta(s) &:= R_+^{-1} \left( Q_{+0} (s D_0 - Q_{00})^{-1} \eta_0 + \eta_+ \right)
\end{align*}
\end{Thm}
\begin{Prf}
Due to the construction in (\ref{def-Z}) and (\ref{def-XY}), the representations $Z_1 = Y_\tau$ and $Z_2 = X_\tau$ hold, where $\tau$ is defined in (\ref{def-tau}).
This means that on the set $\{ Z_1 =0 \}$, the phase process ${\cal J}$ lives only on $E_0$ before it gets absorbed. Define $\sigma := \min \{ t \geq 0: J_t \notin E_0 \}$. Clearly, $\sigma \leq \tau < \infty$ and $\{ \sigma = \tau \} = \{ Z_1 = 0 \}$. Thus
\[ \mathbb{E} ( e^{-s Z_2} 1_{ \{ Z_1 =0 \} }) = \mathbb{E} ( e^{-s \int_0^\tau r_2(J_s) \; ds} 1_{ \{ \sigma = \tau \} })
\]
Theorem 1 in \cite{BR13} states that
\[ \mathbb{E} ( e^{-s \int_0^t r_2(J_s) \; ds} 1_{ \{ t < \sigma < \tau \} }) = \alpha_0 e^{(Q_{00} - s D_0) t} {\bf 1}
\]
for $s \geq 0$. Hence,
\[ \mathbb{E} ( e^{-s \int_0^t r_2(J_s) \; ds} 1_{ \{ \sigma = \tau \in dt \} }) = \alpha_0 e^{(Q_{00} - s D_0) t} \eta_0
\]
for all $t > 0$. Now integrating over $t \in ]0, \infty[$ yields the first statement. For the second statement, theorem 2 in \cite{BR13} states that
\[ \mathbb{E} \left( e^{-s \int_0^{\tau(y)} r_2(J_s) \; ds} 1_{ \{ \tau(y) < \tau \} } \right) = e^{W(s) y}
\]
for $s \geq 0$, where $\tau(y)$ is defined in (\ref{def-fpt}). Given our construction of ${\cal Y}$ and $Z_1$, this is equivalent to
\[ \mathbb{E} \left( e^{-s \int_0^{\tau(y)} r_2(J_s) \; ds} 1_{ \{ Z_1 > y \} } \right) = e^{W(s) y}
\]
From here we obtain for small $h > 0$
\begin{multline*}
\mathbb{E} \left( e^{-s \int_0^{\tau(Z_1)} r_2(J_s) \; ds} 1_{ \{ y < Z_1 < y+h \} } \right) \\
= e^{W(s) y} \left( h R_+^{-1} \eta_+ + h R_+^{-1} Q_{+0} (s D_0 - Q_{00})^{-1} \eta_0 + o(h) \right)
\end{multline*}
and hence
\begin{align*}
\mathbb{E} ( \left. e^{-s Z_2} 1_{ \{ Z_1 \in dy \} } \right| J_0 = i) &= e_i' e^{W(s) y} \eta(s) \; dy
\end{align*}
for $y > 0$ and ascending phases $i \in E_+$. Considering all possible initial phases, we obtain by the same reasoning as for the first statement
\[ \mathbb{E} ( e^{-s Z_2} 1_{ \{ Z_1 \in dy \} }) = \left( \alpha_0 (s D_0 - Q_{00})^{-1} Q_{0+} + \alpha_+ \right) e^{W(s) y} \eta(s) \; dy
\]
for $y > 0$, which is the second statement.
\end{Prf}
\begin{Rem}
Let $Z=(Z_1, \ldots, Z_k) \in \text{MPH}^*$ with $k \geq 3$. According to theorem 6 in \cite{Ku89}, every pair $(Z_i, Z_j)$ with $i \neq j$ has a bivariate MPH$^*$ distribution. Thus we can use theorem 1 to determine the two-dimensional marginal distributions of a $k$-variate MPH$^*$ distribution.
\end{Rem}
\begin{Rem}
For $s=0$ we obtain the marginal distribution of $Z_1$, which is given as follows. Let $k := | E_0 |$ and $n := | E_+ |$, where $|M|$ denotes the cardinality of a set $M$. $Z_1$ has a PH($\beta, T$) distribution of order $n$ with
\[ \beta_i = \alpha_{k+i} + \alpha_0 (- Q_{00}^{-1}) Q_{0+} e_i
\]
for $i \in \{ 1, \ldots, n \}$ and
\[ \beta_{n+1} = \alpha_{m+1} + \alpha_0 (- Q_{00}^{-1}) \eta_0
\]
The rate matrix $T$ is given by $T = W(0) = R_+^{-1} \left( Q_{++} - Q_{+0} Q_{00}^{-1} Q_{0+} \right)$ such that
\begin{align*}
- T {\bf 1} &= - R_+^{-1} \left( Q_{++} {\bf 1} - Q_{+0} Q_{00}^{-1} Q_{0+} {\bf 1} \right) \\
&= R_+^{-1} \left( \eta_+ + Q_{+0} {\bf 1} + Q_{+0} Q_{00}^{-1} (\eta_0 - Q_{00} {\bf 1}) \right) \\
&= R_+^{-1} \left( \eta_+ + Q_{+0} Q_{00}^{-1} \eta_0 \right) \\
&= \eta(0)
\end{align*}
as to be expected.
\end{Rem}
\begin{Rem}
Theorem 4 in \cite{Ku89} states that the joint Laplace transform of $(Z_1, Z_2)$ is given by
\begin{equation} \label{eq-ku}
\mathbb{E} ( e^{- s_1 Z_1} e^{- s_2 Z_2} ) = - \alpha ( \Delta - Q)^{-1} Q {\bf 1}
\end{equation}
where $\Delta := diag( s_1 r_1(j) + s_2 r_2(j): j \in E)$. A relatively arduous way to arrive at this result is
\begin{align*}
\mathbb{E} ( e^{- s_1 Z_1} e^{- s_2 Z_2} ) &= \int_0^\infty e^{- s_1 y} \mathbb{E} ( e^{-s_2 Z_2} 1_{ \{ Z_1 \in dy \} }) \; dy + \mathbb{E} ( e^{-s_2 Z_2} 1_{ \{ Z_1 =0 \} }) \\
&= \left( \alpha_0 (Q_{00} - s_2 D_0)^{-1} Q_{0+} + \alpha_+ \right) \int_0^\infty e^{- s_1 y} e^{ W(s_2) y} \eta(s_2) \; dy \\
& \qquad - \alpha_0 (Q_{00} - s_2 D_0)^{-1} \eta_{0} \\
&= - \left( \alpha_0 (Q_{00} - s_2 D_0)^{-1} Q_{0+} + \alpha_+ \right) (W(s_2) - s_1 I)^{-1} \eta(s_2) \\
& \qquad - \alpha_0 (Q_{00} - s_2 D_0)^{-1} \eta_{0}
\end{align*}
First we observe that
\begin{align*}
W(s_2) - s_1 I &= R_+^{-1} \left( (Q_{++} - s_2 D_+) - Q_{+0} (Q_{00} - s_2 D_0)^{-1} Q_{0+} - s_1 R_+ \right) \\
&= R_+^{-1} \left( (Q_{++} - s_1 R_+ -s_2 D_+) - Q_{+0} (Q_{00} - s_1 R_0 - s_2 D_0)^{-1} Q_{0+} \right)
\end{align*}
since $R_0 = {\bf 0}$ by definition. To shorten notations, we write $W:= W(s_2) - s_1 I $. Further, we write
\[ R_+^{-1} \left( Q_{+0} (sD_0 - Q_{00})^{-1} \eta_0 + \eta_+ \right) = R_+^{-1} \left( - Q_{+0} (Q_{00} - s D_0)^{-1}, I \right) \begin{pmatrix} \eta_0 \\ \eta_+ \end{pmatrix}
\]
To arrive at (\ref{eq-ku}), we need to show that
\begin{align*}
( \Delta - Q)^{-1} &= \begin{pmatrix} (Q_{00} - s_2 D_0)^{-1} Q_{0+} \\ I \end{pmatrix} (-W)^{-1} R_+^{-1} \begin{pmatrix} - Q_{+0} (Q_{00} - s_2 D_0)^{-1} & I \end{pmatrix} \\
& \qquad + \begin{pmatrix} -(Q_{00} - s_2 D_0)^{-1} & {\bf 0} \\ {\bf 0} & {\bf 0} \end{pmatrix}
\end{align*}
In block form we can write
\[ ( \Delta - Q) = \begin{pmatrix} s_2 D_0 - Q_{00} & Q_{0+} \\ Q_{+0} & s_1 R_+ + s_2 D_+ - Q_{++} \end{pmatrix}
\]
since $R_0 = {\bf 0}$. Thus
\begin{align*}
( \Delta - Q) & \begin{pmatrix} (Q_{00} - s_2 D_0)^{-1} Q_{0+} \\ I \end{pmatrix} (-W)^{-1} \begin{pmatrix} - Q_{+0} (Q_{00} - s_2 D_0)^{-1} & I \end{pmatrix} \\
&= \begin{pmatrix} {\bf 0} \\ - R_+ W \end{pmatrix} (-W)^{-1} R_+^{-1} \begin{pmatrix} - Q_{+0} (Q_{00} - s_2 D_0)^{-1} & I \end{pmatrix} \\
&= \begin{pmatrix} {\bf 0} & {\bf 0} \\ - Q_{+0} (Q_{00} - s_2 D_0)^{-1} & I \end{pmatrix}
\end{align*}
and further
\[ ( \Delta - Q) \begin{pmatrix} -(Q_{00} - s_2 D_0)^{-1} & {\bf 0} \\ {\bf 0} & {\bf 0} \end{pmatrix} = \begin{pmatrix} I & {\bf 0} \\ Q_{+0} (Q_{00} - s_2 D_0)^{-1} & {\bf 0} \end{pmatrix}
\]
Together this yields the desired result.
\end{Rem}
\begin{Rem} The most important ingredient to compute the covariance is $\mathbb{E}( Z_1 Z_2)$. Corollary 1 in \cite{Ku89} provides an iteration scheme to compute joint moments. An explicit formula is obtained via
\[ \mathbb{E}( Z_1 Z_2) = \left. \frac{d}{ds} \mathbb{E}( e^{-s Z_2} Z_1) \right|_{s=0}
\]
To this aim,
\begin{align*}
\mathbb{E}( e^{-s Z_2} Z_1) &= \int_0^\infty y \mathbb{E} ( e^{-s Z_2} 1_{ \{ Z_1 \in dy \} }) = \alpha(s) \int_0^\infty y e^{W(s) y} \; dy \; \eta(s)
\end{align*}
and
\begin{align*}
\int_0^\infty y e^{W(s) y} \; dy &= W(s)^{-1} \left[ y e^{W(s) y} \right]_{y=0}^\infty - W(s)^{-1} \int_0^\infty e^{W(s) y} \; dy = W(s)^{-2}
\end{align*}
yield
\begin{align*}
\mathbb{E}( Z_1 Z_2) &= \left. - \frac{d}{ds} \alpha(s) W(s)^{-2} \eta(s) \right|_{s=0}
\end{align*}
This can be readily evaluated using the differentiation rule
\[ \frac{d}{ds} (M(s)^{-1}) = M(s)^{-1} \frac{d}{ds} M(s) M(s)^{-1}
\]
for matrix-valued functions $M(s)$, see sections I.1.3-4 in \cite{Bo04}.
\end{Rem}
\begin{Rem}
The special case MPH as described in \cite{AL84} is obtained as follows. Using the decomposition of state space $E$ and generator matrix $A$ as on p.692 therein, we can translate $E_+ = \Gamma_2^c$, $E_0 = \Gamma_2$, and
\[ Q_{++} = \begin{pmatrix} A^{(1,2)} & B^{(1)} \\ {\bf 0} & A^{(1)} \end{pmatrix}, \qquad Q_{+0} = \begin{pmatrix} B^{(2)} \\ {\bf 0} \end{pmatrix}, \qquad Q_{0+} = {\bf 0}, \qquad Q_{00} = A^{(2)}
\]
The construction in \cite{AL84} further specifies $R_+ = I$, $D_0 = I$, and
\[ D_+ = \begin{pmatrix} I^{1,2} & {\bf 0} \\ {\bf 0} & {\bf 0} \end{pmatrix}
\]
where $I^{1,2}$ denotes the identity matrix on $ \Gamma_1^c \cap \Gamma_2^c$. This yields $\alpha(s) = \alpha_+$,
\[ W(s) = Q_{++} - s D_+ = \begin{pmatrix} A^{(1,2)} - s I & B^{(1)} \\ {\bf 0} & A^{(1)} \end{pmatrix}
\]
and
\[ \eta(s) = \begin{pmatrix} B^{(2)} (sI - A^{(2)})^{-1} (- A^{(2)} {\bf 1}) \\ {\bf 0} \end{pmatrix} + \begin{pmatrix} -( A^{(1,2)} {\bf 1} + B^{(1)} {\bf 1} + B^{(2)} {\bf 1}) \\ - A^{(1)} {\bf 1} \end{pmatrix} .
\]
\end{Rem}
\begin{Rem} If $r_{1j} > 0$ for all $j \in E$, then $E=E_+$, hence $\mathbb{E} ( e^{-s Z_2} 1_{ \{ Z_1 =0 \} }) =0$, and
\[ \mathbb{E} ( e^{-s Z_2} 1_{ \{ Z_1 \in dy \} }) = \alpha_+ e^{W(s) y} \eta_+
\]
where
\[ W(s) = R_+^{-1} \left( Q_{++} - s D_+ \right)
\]
for all $s \geq 0$. If further $r_{1i}=r_{2i}$ for all $i \in E$ with $r_{2i} > 0$ and $q_{ij} = 0$ for $r_{2i}=0$ and $r_{2j} > 0$, then we obtain the special case of the class MPH where $Z_1 \geq Z_2$ almost surely. This specifies to $\Gamma_2 = {\bf 0}$, $\Gamma_1 = \{ i \in E: r_{2i} = 0 \}$, as well as
\[ A^{(1,2)} = \left( \frac{q_{ij}}{r_{1i}} \right)_{i,j \in \Gamma_1^c}, \quad B^{(1)} = \left( \frac{q_{ij}}{r_{1i}} \right)_{i \in \Gamma_1^c, j \in \Gamma_1} \quad \text{and} \quad A^{(1)} = \left( \frac{q_{ij}}{r_{1i}} \right)_{i,j \in \Gamma_1}.
\]
\end{Rem}
\begin{Rem}
With no additional effort, the current framework can be extended to allow $r_{2i} < 0$ for some $i \in E$. One needs to take care of the range of $s$ for the Laplace transform $\mathbb{E} ( e^{-s Z_2} 1_{ \{ Z_1 \in dy \} })$ to converge (but there is such one, see lemma 2 in \cite{BR13}) or consider Fourier transforms. Then $Z_2$ has a so-called bilateral phase-type distribution, i.e.\ it is the mixture of two random variables $Z_2^+$ and $Z_2^-$ where $Z_2^+$ and $- Z_2^-$ have phase-type distributions. In particular, $Z_2$ may also assume negative values now. Theorem 2.3.2 in \cite{AR05} states that bilateral phase-type distributions are (weakly) dense in the class of all distributions on $\mathbb{R}$. For the marginal distribution of $Z_2$ see \cite{As04}, for more on bilateral phase-type distributions see \cite{AR05}.
\end{Rem}
|
\section{QED Effective Action on the Worldline}
Worldline numerics is built on the worldline formalism which was
initially invented by Feynman~\cite{PhysRev.80.440, PhysRev.84.108}.
Much of the recent interest in this formalism is based on the work of
Bern and Kosower, who derived it from the infinite string-tension
limit of string theory and demonstrated that it provided an efficient
means for computing amplitudes in QCD~\cite{PhysRevLett.66.1669}.
For this reason, the worldline formalism is often referred to as
`string inspired'. However, the formalism can also be obtained
straight-forwardly from first-quantized field
theory~\cite{1992NuPhB.385..145S}, which is the approach we will adopt
here. In this formalism the degrees of freedom of the field are
represented in terms of one-dimensional path integrals over an
ensemble of closed trajectories.
\begin{widetext}
We begin with the QED effective action expressed in the proper-time
formalism \cite{Schwinger:1951},
\begin{equation}
\label{eqn:trln}
\mathrm{Tr~ln}\left[\frac{\slashed{p}+e\slashed{A}_\mu^0
-m}{\slashed{p}-m}\right] = -\frac{1}{2}\int d^4x \int_0^\infty
\frac{dT}{T}e^{-iTm^2}
\times \mathrm{tr}\biggr( \bra{x}e^{iT(\slashed{p}
+e\slashed{A}^0_\mu)^2}\ket{x}
- \bra{x}e^{iTp^2}\ket{x}\biggr).
\end{equation}
To evaluate $\bra{x}e^{iT(\slashed{p}_\mu +
e\slashed{A}_\mu)^2}\ket{x}$, we recognize that it is simply the
propagation amplitude $\braket{x,T}{x,0}$ from ordinary quantum
mechanics with $(\slashed{p}_\mu + e\slashed{A}_\mu)^2$ playing the
role of the Hamiltonian. We therefore express this factor in its path
integral form:
\begin{equation}
\bra{x}e^{iT(\slashed{p}_\mu + e\slashed{A}_mu)^2}\ket{x} = \mathcal{N}
\int \mathcal{D}x_\rho(\tau) e^{-\int_0^T d\tau \left[\frac{\dot{x}^2(\tau)}{4}
+ i A_\rho x^\rho(\tau)\right]}
\times \frac{1}{4} {\rm tr}e^{\frac{1}{2}\int_0^T d\tau \sigma_{\mu \nu}F^{\mu \nu}(x_{\rm CM}+x(\tau))}.
\end{equation}
$\mathcal{N}$ is a normalization constant that we can fix by using
our renormalization condition that the fermion determinant should
vanish at zero external field:
\begin{equation}
\bra{x}e^{iTp^2}\ket{x} = \mathcal{N}\int \mathcal{D}
x_p(\tau)e^{-\int_0^T d\tau\frac{\dot{x}^2(\tau)}{4}}
= \int \frac{d^4p}{(2\pi T)^4}\bra{x}e^{iTp^2}\ket{p}\braket{p}{x}
= \frac{1}{(4\pi T)^2},
\end{equation}
We may now write
\begin{equation}
\mathcal{N}\int \mathcal{D}x_\rho(\tau) e^{-\int_0^T d\tau[\frac{\dot{x}^2(\tau)}{4}+iA_\rho x^\rho(\tau)]}
\frac{1}{4} {\rm tr}e^{\frac{1}{2}\int_0^T d\tau \sigma_{\mu \nu}F^{\mu \nu}(x_{\rm CM}+x(\tau))}
= \frac{\left\langle e^{-i\int_0^T d\tau A_\rho x^\rho(\tau)}\frac{1}{4} {\rm tr}e^{\frac{1}{2}\int_0^T d\tau \sigma_{\mu \nu}F^{\mu \nu}(x_{\rm CM}+x(\tau))}\right\rangle_x}{(4\pi T)^2} ,
\end{equation}
where
\begin{equation}
\label{eqn:meandef}
\mean{\hat{\mathcal{O}}}_x = \frac{\int \mathcal{D}x_\rho(\tau) \hat{\mathcal{O}}
e^{-\int_0^T d\tau\frac{\dot{x}^2(\tau)}{4}}}{\int \mathcal{D}x_\rho(\tau)
e^{-\int_0^T d\tau\frac{\dot{x}^2(\tau)}{4}}}
\end{equation}
is the weighted average of the operator $\hat{\mathcal{O}}$ over an ensemble of closed particle
loops with a Gaussian velocity distribution.
Finally, combining all of the equations from this section results in the
renormalized one-loop effective action for \ac{QED} on the worldline:
\begin{equation}
\label{eqn:QEDWL}
\Gamma^{(1)}[A_\mu] = \frac{2}{(4\pi)^2}\int_0^\infty
\frac{dT}{T^3}e^{-m^2T}\int d^4x_\mathrm{CM} \times
\left[\left\langle e^{i\int_0^Td\tau A_\rho(x_\mathrm{CM}+x(\tau))\dot{x}^\rho(\tau)}
\frac{1}{4} {\rm tr}e^{\frac{1}{2}\int_0^T d\tau \sigma_{\mu \nu}F^{\mu \nu}(x_{\rm CM}+x(\tau))}\right\rangle _x -1\right].
\end{equation}
\end{widetext}
\section{Worldline Numerics}
The averages, $\mean{\hat{\mathcal{O}}}$, defined by equation (\ref{eqn:meandef})
involve
functional integration over every possible closed path through spacetime
which has a Gaussian velocity distribution.
The prescription of the worldline numerics technique is to compute
these averages approximately using a finite set of $N_l$ representative loops
on a computer. The worldline average is then approximated as the mean of
an operator evaluated along each of the worldlines in the ensemble:
\begin{equation}
\mean{\hat{\mathcal{O}}[x(\tau)]} \approx
\frac{1}{N_l} \sum_{i=1}^{N_l} \hat{\mathcal{O}}[x_i(\tau)].
\end{equation}
\subsection{Loop Generation}
\label{sec:loopgen}
The velocity distribution for the loops depends on
the proper time, $T$. However, generating a separate ensemble of loops for
each value of $T$ would be very computationally expensive. This problem is alleviated by generating
a single ensemble of loops, $\vec{y}(\tau)$, representing unit proper time,
and scaling those loops accordingly for different values of $T$:
\begin{equation}
\vec{x}(\tau) = \sqrt{T}\vec{y}(\tau/T) ,
\end{equation}
\begin{equation} \int_0^T d\tau \vec{\dot{x}}^2(\tau) \rightarrow \int_0^1 dt
\vec{\dot{y}}^2(t).
\end{equation}
There is no way to treat the integrals as continuous as we generate our loop
ensembles. Instead, we treat the integrals as sums over discrete points
along the proper-time interval $[0,T]$. This is fundamentally different
from space-time discretization, however. Any point on the worldline loop
may exist at any position in space, and $T$ may take on any value. It is
important to note this distinction because the worldline method retains
Lorentz invariance while lattice techniques, in general, do not.
The challenge of loop cloud generation is in generating a discrete set
of points on a unit loop which obeys the prescribed velocity distribution.
There are a number of different algorithms for achieving this goal that have
been discussed in the literature. Four possible algorithms are compared
and contrasted in \cite{Gies:2003cv}. In this work, we choose a more
recently developed algorithm, dubbed ``d-loops", which was first described
in \cite{Gies:2005sb}. To generate a ``d-loop", the number of points is iteratively
doubled, placing the new points in a Gaussian distribution between the existing neighbour points.
We quote the algorithm directly:
\begin{quote}
\begin{itemize}
\item[(1)] Begin with one arbitrary point
$N_0=1$, $y_{N}$.
\item[(2)] Create an $N_1=2$ loop, i.e., add a point $y_{N/2}$ that is
distributed in the heat bath of $y_N$ with
\begin{equation}
e^{-\frac{N_1}{4} 2 (y_{N/2} -y_{N})^2}. \label{yn2}
\end{equation}
\item[(3)] Iterate this procedure, creating an $N_k=2^k$ points
per line (ppl) loop by adding $2^{k-1}$ points
$y_{{qN}/{N_k}}$, $q=1,3,\dots, N_k-1$ with distribution
\begin{equation}
e^{-\frac{N_k}{4} 2 [y_{qN/N_k} -\frac{1}{2}(y_{(q+1)N/N_k}+
y_{(q-1)N/N_k})]^2}. \label{ynk}
\end{equation}
\item[(4)] Terminate the procedure if $N_k$ has reached $N_k=N$ for
unit loops with $N$ ppl.
\item[(5)] For an ensemble with common center of mass, shift each
whole loop accordingly.
\end{itemize}
\end{quote}
The above d-loop algorithm was selected since it is simple and about
10\% faster than previous algorithms, according to its developers,
because it requires fewer algebraic operations. The generation of the
loops is largely independent from the main program. Because of this,
it was simpler to generate the loops using a Matlab script.
This function was used to produce text files containing the worldline data for
ensembles of loops. Then, these text files were read into memory at the
launch of each calculation. The results of this generation routine can
be seen in figure \ref{fig:worldlineplot}.
When the \ac{CUDA} kernel is called\footnote{Please see appendix
\ref{ch:cudafication} for an overview of \ac{CUDA}.}, every thread
in every block executes the kernel function with its own unique
identifier. Therefore, it is best to generate worldlines in integer
multiples of the number of threads per block. The Tesla C1060 device
allows up to 512 threads per block.
\begin{figure}
\centering
\includegraphics[width=\linewidth,clip,trim=2cm 1cm 2cm 1cm]{images/worldlineplot}
\caption[Discretization of the worldline loop]
{A single discrete worldline loop shown at several levels
of discretization. The loops form fractal patterns and have a strong
parallel with Brownian motion. The colour
represents the phase of a particle travelling along the loop, and
begins at dark blue, progresses in a random walk through yellow,
and ends at dark
red. The total flux through this particular worldline at $T=1$ and
$B=B_k$ is about $0.08 \pi/e$.}
\label{fig:worldlineplot}
\end{figure}
\section{Cylindrical Worldline Numerics}
We now consider cylindrically symmetric external magnetic fields.
In this case, we may simplify (\ref{eqn:QEDWL}),
\begin{eqnarray}
\label{eqn:cylEA}
\frac{\Gamma^{(1)}_\mathrm{ ferm}}{T
L_z} &=& \frac{1}{4\pi} \int_0^\infty \rho_\mathrm{ cm}
\biggr[ \int_0^\infty \frac{dT}{T^3}e^{-m^2T} \times \\
& & ~~~~~~ \left\{\langle
W\rangle_{\vec{r}_\mathrm{ cm}} - 1 -\frac{1}{3}(eB_\mathrm{
cm}T)^2\right\}\biggr]d\rho. \nonumber
\end{eqnarray}
\subsection{Cylindrical Magnetic Fields}
We have $\vec{B} =
B(\rho)$\bhattext{z} with
\begin{equation} \label{eqn:BWLN} B(\rho) = \frac{A_\phi(\rho)}{\rho} +
\frac{dA_\phi(\rho)}{d\rho} \end{equation} if we make the gauge choice that $A_0 =
A_\rho = A_z = 0$.
We begin by considering $A_\phi(\rho)$ in the form
\begin{equation} A_\phi(\rho) = \frac{F}{2\pi \rho}f_\lambda(\rho) \end{equation} so that \begin{equation}
B_z(\rho)=\frac{F}{2\pi\rho}\frac{df_\lambda(\rho)}{d\rho} \end{equation} and the total
flux is \begin{equation} \Phi=F(f_\lambda(L_\rho)-f_\lambda(0)). \end{equation} It is convenient
to express the flux in units of $\frac{2 \pi}{e}$ and define a dimensionless
quantity
\begin{equation}
\mathcal{F}=\frac{e}{2 \pi} F.
\end{equation}
\subsection{Wilson Loop}
The quantity inside the angled brackets in equation (\ref{eqn:QEDWL}) is a
gauge invariant observable called a Wilson loop. We note that the proper time
integral provides a natural path ordering for this operator.
The Wilson loop expectation value is
\begin{equation}
\label{eqn:wilsonloop}
\langle W\rangle_{\vec{r}_\mathrm{cm}} \!\! = \biggl \langle e^{ie\int_0^T d\tau
\vec{A}(\vec{r}_{\mathrm{ cm}} + \vec{r}(\tau)) \cdot \dot{\vec{r}}}
\frac{1}{4} \mathrm{ tr} e^{\frac{e}{2}\int_0^T d\tau
\sigma_{\mu \nu}F_{\mu \nu}(\vec{r}_{\mathrm{ cm}} + \vec{r}(\tau))}\biggr
\rangle_{\vec{r}_\mathrm{ cm}} \!\!\!\!\!\!\!\!,
\end{equation}
which we look at as a product between a scalar part ($e^{ie\int_0^T d\tau
\vec{A}(\vec{r}_{\mathrm{ cm}} + \vec{r}(\tau)) \cdot \dot{\vec{r}}}$)
and a fermionic part ($\frac{1}{4} \mathrm{ tr} e^{\frac{e}{2}\int_0^T d\tau
\sigma_{\mu \nu}F_{\mu \nu}(\vec{r}_{\mathrm{ cm}} + \vec{r}(\tau))}$).
\subsubsection{Scalar Part}
In a magnetic field, the scalar part is related to the flux through
the loop, $\Phi_B$, by Stokes theorem:
\begin{eqnarray}
e^{ie\int_0^T d\tau
\vec{A} \cdot \dot{\vec{r}}} &=&
e^{ie\oint \vec{A}\cdot d\vec{r}} = e^{ie\int_{\vec{\Sigma}} \vec{\nabla}\times\vec{A} \cdot d\vec{\Sigma}}\\
& = & e^{ie\int_{\vec{\Sigma}} \vec{B} d\vec{\Sigma}} = e^{ie\Phi_B}.
\end{eqnarray}
Consequently, this factor accounts for the Aharonov-Bohm phase acquired by
particles in the loop.
The loop discretization results in the following approximation of the
scalar integral:
\begin{equation}
\oint \vec{A}(\vec{r})\cdot d\vec{r} = \sum_{i=1}^{N_\mathrm{ppl}}
\int_{\vec{r}^i}^{\vec{r}^{i+1}}\vec{A}(\vec{r})\cdot d\vec{r}.
\end{equation}
Using a linear parameterization of the positions, the line integrals are
\begin{equation}
\int_{\vec{r}^i}^{\vec{r}^{i+1}}\vec{A}(\vec{r})\cdot d\vec{r} =
\int_0^1dt \vec{A}(\vec{r}(t))\cdot(\vec{r}^{i+1} - \vec{r}^i).
\end{equation}
Using the same gauge choice outlined above ($\vec{A}=A_\phi \hat{\phi}$),
we may write
\begin{equation}
\vec{A}(\vec{r}(t)) = \frac{\mathcal{F}}{e\rho^2}
f_\lambda(\rho^2)(-y,x,0),
\end{equation}
where we have chosen $f_\lambda(\rho^2)$ to depend on $\rho^2$ instead
of $\rho$ to simplify some expressions and to
avoid taking many costly square roots in the worldline numerics.
We then have
\begin{equation} \int_{\vec{r}^i}^{\vec{r}^{i+1}}\vec{A}(\vec{r})\cdot
d\vec{r} = \mathcal{F} (x^iy^{i+1}-y^i x^{i+1})\int_0^1
dt\frac{f_\lambda(\rho_i^2(t))}{\rho_i^2(t)}. \end{equation} The linear interpolation
in Cartesian coordinates gives
\begin{equation}
\label{eqn:rhoi} \rho_i^2(t) = A_i + 2B_it + C_i t^2,
\end{equation}
where
\begin{eqnarray}
A_i &=& (x^i)^2 + (y^i)^2 \\
B_i&=&x^i(x^{i+1} - x^i) + y^i(y^{i+1}-y^i)\\
C_i &= &(x^{i+1}-x^i)^2 + (y^{i+1}-y^i)^2.
\label{eqn:Ci}
\end{eqnarray}
In performing the integrals along the straight lines connecting
each discretized loop point, we are in danger of violating gauge invariance.
If these integrals can be performed analytically, than gauge invariance
is preserved exactly. However, in general, we wish to compute these integrals
numerically. In this case, gauge invariance is no longer guaranteed, but
can be preserved to any numerical precision that's desired.
\subsubsection{Fermion Part}
For fermions, the Wilson loop is modified by a factor,
\begin{eqnarray}
W^\mathrm{ ferm.} &=& \frac{1}{4}\mathrm{ tr}\left(e^{\frac{1}{2}
e\int_0^T d\tau \sigma_{\mu \nu}F^{\mu \nu}}\right)\\
&=& \frac{1}{4}\mathrm{ tr}\left(e^{\sigma_{x y}
e\int_0^T d\tau B\left(x(\tau)\right)}\right) \\
&=& \cosh{\left(e \int_0^T d\tau B\left(x(\tau)\right)\right)}\\
&=& \cosh{\left(2\mathcal{F}\int_0^T
d\tau f'_\lambda(\rho^2(\tau))\right)},
\label{eqn:Wfermfpl}
\end{eqnarray}
where we have used the
relation
\begin{equation}
eB = 2\mathcal{F}\frac{d f_\lambda(\rho^2)}{d \rho^2} =
2\mathcal{F}f'_\lambda(\rho^2).
\end{equation}
This factor represents an additional contribution to the
action because of the spin interaction with the magnetic field.
Classically, for a particle with a magnetic moment $\vec{\mu}$
travelling through a magnetic field in a time $T$, the
action is modified by a term given by
\begin{equation}
\Gamma^0_\mathrm{ spin} = \int_0^T \vec{\mu} \cdot \vec{B}(\vec{x}(\tau)) d\tau.
\end{equation}
The magnetic moment is related to the electron spin
$\vec{\mu} = g\left(\frac{e}{2m}\right)\vec{\sigma}$,
so we see that the integral in the above quantum fermion factor is
very closely related to the classical action
associated with transporting a magnetic moment through a magnetic field:
\begin{equation}
\Gamma^0_\mathrm{ spin} = g\left(\frac{e}{2m}\right) \sigma_{x y} \int_0^T B_z(x(\tau))d\tau.
\end{equation}
Qualitatively, we could write
\begin{equation}
W^\mathrm{ ferm} \sim \cosh{\left(\Gamma^0_\mathrm{ spin}\right)}.
\end{equation}
As a possibly useful aside, we may want to express
the integral in terms of $f_\lambda(\rho^2)$ instead of its derivative.
We can do this by integrating by parts:
\begin{eqnarray}
\int_0^T d\tau f'_\lambda(\rho^2(\tau))&=&
\frac{T}{N_\mathrm{ ppl}}\sum_{i=1}^{N_\mathrm{ ppl}}\int_0^1
dt f'_\lambda(\rho^2_i(\tau)) \\
& = &
\frac{T}{N_\mathrm{ ppl}}\sum_{i=1}^{N_\mathrm{ ppl}}\biggr[
\frac{f_\lambda(\rho^2_i(t))}{2(B_i+C_it)}\biggr|^{t=1}_{t=0} \nonumber \\
& & +\frac{C_i}{2} \!\!\int_0^1\!\!\!\!
\frac{f_\lambda(\rho^2_i(t))}{(B_i+C_i t)^2} dt\biggr] \\
& = & \frac{T}{N_\mathrm{ ppl}}\sum_{i=1}^{N_\mathrm{
ppl}}\frac{C_i}{2}\!\!\int_0^1\!\!\!\! \frac{f_\lambda(\rho^2_i(t))}{(B_i+C_i t)^2} dt,
\end{eqnarray}
with $\rho_i^2(t)$ given by equations (\ref{eqn:rhoi}) to (\ref{eqn:Ci}).
The second equality is obtained from integration-by-parts. In the third
equality, we use the loop sum to cancel the boundary terms in pairs:
\begin{equation}
\label{eqn:Wfermfl}
W^{\mathrm{ ferm.}} = \cosh{\left(\frac{\mathcal{F}T}{N_\mathrm{
ppl}}\sum_{i=1}^{N_{\mathrm{ ppl}}}
C_i \int_0^1 dt \frac{f_\lambda(\rho^2_i(t))}{(B_i+C_i t)^2}\right)}.
\end{equation}
In most cases, one would use equation (\ref{eqn:Wfermfpl}) to compute the fermion factor
of the Wilson loop. However,
equation (\ref{eqn:Wfermfl}) may be useful in cases where $f'_\lambda(\rho^2(\tau))$
is not known or is difficult to compute.
\subsection{Renormalization}
The field strength renormalization counter-terms result from the small $T$
behaviour of the worldline integrand. In the limit where $T$ is very small,
the worldline loops are very localized around their center of mass. So,
we may approximate their contribution as being that of a constant field
with value $\vec{A}(\vec{r}_{\mathrm{ cm}})$. Specifically, we require that the field
change slowly on the length scale defined by $\sqrt{T}$. This condition on
$T$ can be written
\begin{equation}
T \ll \left|\frac{m^2}{e B'(\rho^2)}\right|
= \left| \frac{m^2}{2\mathcal{F}f''_\lambda(\rho^2_\mathrm{ cm})}\right|.
\end{equation}
When this limit is satisfied, we may use the exact expressions for the
constant field Wilson loops to determine the small $T$ behaviour of the
integrands and the corresponding counter terms.
The Wilson loop averages for constant magnetic fields in scalar and
fermionic \ac{QED} are
\begin{equation}
\mean{W}_\mathrm{ ferm} = eBT\coth{(eBT)}
\end{equation}
and
\begin{equation}
\mean{W}_\mathrm{ scal} = \frac{eBT}{\sinh{(eBT)}}.
\end{equation}
\begin{widetext}
Therefore, the integrand for fermionic \ac{QED} in the limit of small $T$ is
\begin{eqnarray} \label{eqn:fermI} I_\mathrm{ ferm}(T) &=&
\frac{e^{-m^2T}}{T^3}\left[eB(\vec{r}_{cm})T\coth{(eB(\vec{r}_{cm})T)}
- 1 -\frac{e^2}{3} B^2(\vec{r}_{cm})T^2\right] \nonumber \\ &\approx&-\frac{(eB)^4
T}{45}+\frac{1}{45} (eB)^4 m^2 T^2+\left( \frac{2 (eB)^6}{945}-\frac{(eB)^4
m^4}{90}\right)T^3
+\frac{(7 (eB)^4 m^6-4 (eB)^6 m^2)T^4 }{1890}+O(T^5).
\end{eqnarray}
For scalar QED we have
\begin{eqnarray}
\label{eqn:scalI} I_\mathrm{ scal}(T) &=&
\frac{e^{-m^2T}}{T^3}\left[\frac{eB(\vec{r}_{cm})T}{\sinh{(eB(\vec{r}_{cm})T)}}
- 1 +\frac{1}{6} (eB)^2(\vec{r}_{cm})T^2\right] \nonumber \\ &\approx&\frac{7
(eB)^4 T}{360}-\frac{7 (eB)^4 m^2 T^2}{360}+\frac{(147 (eB)^4 m^4-31
(eB)^6)T^3 }{15120} +\frac{ (31 (eB)^6 m^2-49 (eB)^4
m^6)T^4}{15120}+O(T^5).
\end{eqnarray}
\end{widetext}
Beyond providing the renormalization conditions, these expansions can
be used in the small $T$ regime to avoid a problem with the Wilson
loop uncertainties in this region. Consider the uncertainty in the
integrand arising from the uncertainty in the Wilson loop:
\begin{equation}
\delta I(T) = \frac{\partial I}{\partial W} \delta W = \frac{e^{-m^2 T}}{T^3} \delta W.
\end{equation}
In this case, even though we can compute the Wilson loops for small $T$
precisely, even a small uncertainty is magnified by a divergent factor when
computing the integrand for small values of $T$. So, in order to perform
the integral, we must replace the small $T$ behaviour of the integrand with
the above expansions (\ref{eqn:fermI}) and (\ref{eqn:scalI}). Our worldline
integral then proceeds by analytically computing the integral for the small
$T$ expansion up to some small value, $a$, and adding this to the remaining
part of the integral~\cite{MoyaertsLaurent:2004}:
\begin{equation}
\int_0^{\infty} I(T) dT = \underbrace{\int_0^a I(T) dT}_\mathrm{ small ~T}
+ \underbrace{\int_a^\infty I(T)dT}_\mathrm{ worldline ~numerics}.
\end{equation}
Because this normalization procedure uses the constant field expressions for small values of
$T$, this scheme introduces a small systematic uncertainty. To improve on the
method outlined here, the derivatives of the background field can
be accounted for by using the analytic forms of the heat kernel expansion to perform the
renormalization~\cite{Gies:2001tj}.
\section{Uncertainty analysis in worldline numerics}
\label{ch:WLError}
So far in the worldline numerics literature, the discussions of uncertainty
analysis have been unfortunately brief. It has been suggested that the
standard deviation of the worldlines provides a good measure of the
statistical error in the worldline method~\cite{Gies:2001zp,
Gies:2001tj}. However, the distributions produced by the worldline
ensemble are highly non-Gaussian (see figure \ref{fig:hists}),
and therefore the standard error in
the mean is not a good measure of the uncertainties
involved. Furthermore, the use of the same worldline ensemble to
compute the Wilson loop multiple times in an integral results in
strongly correlated uncertainties. Thus, propagating uncertainties
through integrals can be computationally expensive due to the
complexity of computing correlation coefficients.
The error bars on worldline calculations impact the conclusions that
can be drawn from calculations, and also have important implications
for the fermion problem, which limits the domain of applicability of
the technique (see section \ref{sec:fermionproblem}). It is therefore
important that the error analysis is done thoughtfully and
transparently. The purpose of this chapter is to contribute a more
thorough discussion of uncertainty analysis in the worldline numerics technique
to the literature in hopes of avoiding any confusion associated with
the above-mentioned subtleties.
There are two sources of uncertainty in the worldline technique: the
discretization error in treating each continuous worldline as a set of
discrete points, and the statistical error of sampling a finite number
of possible worldlines from a distribution. In this section, we
discuss each of these sources of uncertainty.
\subsection{Estimating the Discretization Uncertainties}
\label{sec:discunc}
The discretization error arising from the integral over $\tau$ in the
exponent of each Wilson loop (see equation (\ref{eqn:wilsonloop})) is
difficult to estimate since any number of loops could be represented
by each choice of discrete points. The general strategy is to make
this estimation by computing the Wilson loop using several different
numbers of points per worldline and observing the convergence
properties.
The specific procedure adopted for this work involves dividing each discrete worldline into several
worldlines with varying levels of discretization.
Since we are using the d-loop
method for generating the worldlines (section \ref{sec:loopgen}),
a $\frac{N_\mathrm{ppl}}{2}$ sub-loop consisting of every other
point will be guaranteed to contain the prescribed distribution of velocities.
To look at the convergence for the loop discretization,
each worldline is divided into three groups. One group of $\frac{N_\mathrm{ppl}}{2}$ points, and two groups of
$\frac{N_\mathrm{ppl}}{4}$.
This permits us to compute the average holonomy factors at three levels of
discretization:
\begin{equation}
\mean{W}_{N_\mathrm{ppl}/4} = \mean{e^{\frac{i}{2}\triangle} e^{\frac{i}{2}\Box}},
\end{equation}
\begin{equation}
\mean{W}_{N_\mathrm{ppl}/2} = \mean{e^{i\circ}},
\end{equation}
and
\begin{equation}
\mean{W}_{N_\mathrm{ppl}} = \mean{e^{\frac{i}{2}\circ} e^{\frac{i}{4}\Box} e^{\frac{i}{4}\triangle}},
\end{equation}
where the symbols $\circ$, $\Box$, and $\triangle$ denote the worldline integral,
$\int_0^T d\tau A(x_{CM}+x(\tau))\cdot \dot x$, computed using the sub-worldlines
depicted in figure \ref{fig:Division}.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{images/DiscretizationDivision}
\caption[Illustration of convergence testing scheme]
{Diagram illustrating the division of a worldline into three smaller interleaved worldlines}
\label{fig:Division}
\end{figure}
We may put these factors into the equation of a parabola to extrapolate the result to an infinite
number of points per line (see figure \ref{fig:DiscErr}):
\begin{equation}
\mean{W}_{\infty} \approx \frac{8}{3}\mean{W}_{N_\mathrm{ppl}} - 2 \mean{W}_{N_\mathrm{ppl}/2} +\frac{1}{3}\mean{W}_{N_\mathrm{ppl}/4}.
\end{equation}
So, we estimate the discretization uncertainty to be
\begin{equation}
\delta \mean{W}_{\infty} \approx |\mean{W}_{N_\mathrm{ppl}} - \mean{W}_{\infty}|.
\end{equation}
Generally, the statistical uncertainties are the limitation in the precision of the
worldline numerics technique. Therefore, $N_\mathrm{ppl}$ should be chosen to be
large enough that the discretization uncertainties are small relative to the
statistical uncertainties.
\begin{figure}
\centering
\includegraphics[width=\linewidth,clip,trim=1.8cm 0.3cm 1.8cm 1cm]{images/DiscErrPlot}
\caption[Illustration of extrapolation to infinite points per loop]
{This plot illustrates the method used to extrapolate the Wilson loop to
infinite points per loop and the uncertainty estimate in the approximation.}
\label{fig:DiscErr}
\end{figure}
\subsection{Estimating the Statistical Uncertainties}
We can gain a great deal of insight into the nature of the statistical uncertainties
by examining the specific case of the uniform magnetic field since we know the
exact solution in this case.
\subsubsection{The Worldline Ensemble Distribution is not Normal.}
A reasonable first instinct for estimating the error bars is to use the standard
error in the mean of the collection of individual worldlines:
\begin{equation}
\mathrm{ SEM}( W ) = \sqrt{\sum_{i=1}^{N_l}\frac{(W_i - \mean{W})^2}{N_l(N_l-1)}}.
\end{equation}
This approach has been promoted in early papers on worldline numerics~\cite{Gies:2001zp, Gies:2001tj}.
In figure \ref{fig:resids}, we have plotted the residuals and the corresponding
error bars for several values of the proper time parameter, $T$. From this plot,
it appears that the error bars are quite large in the sense that we appear to
produce residuals which are considerably smaller than would be implied by the
sizes of the error bars. This suggests that we have overestimated the size of
the uncertainty.
\begin{figure}
\centering
\includegraphics[width=\linewidth,clip,trim=0.8cm 0.3cm 1.8cm 1cm]{images/resids}
\caption[Standard errors in the mean for worldline numerics]
{Residuals of worldline calculations and the corresponding standard
errors in the mean. For reasons discussed in this section, these error bars
overestimate the uncertainties involved.}
\label{fig:resids}
\end{figure}
We can see why this is the case by looking more closely at the
distributions produced by the worldline technique. An exact
expression for these distributions can be derived in the case of the
constant magnetic field~\cite{MoyaertsLaurent:2004}:
\begin{eqnarray}
\label{eqn:exactdist}
w(y)&=&\frac{W_\mathrm{ exact}}{\sqrt{1-y^2}}\sum_{n=-\infty}^{\infty}\biggl[f(\arccos(y)+2n\pi)+\nonumber \\
& & ~~~~~~~~~~~~ f(-\arccos(y)+2n\pi)\biggr]
\end{eqnarray}
with
\begin{equation}
f(\phi)=\frac{\pi}{4BT\cosh^2(\frac{\pi \phi}{2BT})}.
\end{equation}
Figure \ref{fig:hists} shows histograms of the worldline results along with the
expected distributions. These distributions highlight a significant hurdle in
assigning error bars to the results of worldline numerics.
\begin{figure}
\centering
\includegraphics[width=\linewidth,clip,trim=0.8cm 0.3cm 1.8cm 1cm]{images/hist}
\caption[Histograms showing the worldline distributions]
{Histograms showing the worldline distributions of the residuals
for three values of $T$ in the constant magnetic field case.
Here, we are neglecting the fermion factor. The dark
line represents the exact distribution computed using equation
\ref{eqn:exactdist}. The worldline means are indicated with an arrow,
while the exact mean in each case is 0. There are 5120 worldlines in
each histogram. The vertical axes are normalized to a total area of unity.}
\label{fig:hists}
\end{figure}
Due to their highly non-Gaussian nature, the standard error in the mean is not a good
characterization of the distributions that are produced. We should not interpret each individual
worldline as a measurement of the mean value of these distributions; for large values of $BT$,
almost all of our worldlines will produce answers which are far away from the mean of the
distribution. This means that the variance of the distribution will be very large, even
though our ability to determine the mean of the distribution is relatively precise
because of the increasing symmetry about the mean as $T$ becomes large.
\subsubsection{Correlations between Wilson Loops}
Typically, numerical integration is performed by replacing the integral with a sum over a finite set
of points from the integrand. We will begin the present discussion by considering the uncertainty
in adding together two points (labelled $i$ and $j$) in our integral over $T$. Two terms of the
sum representing the numerical integral will involve a function of $T$ times the two
Wilson loop factors,
\begin{equation}
I = g(T_i)\mean{W(T_i)} + g(T_j)\mean{W(T_j)}
\end{equation}
with an uncertainty given by
\begin{eqnarray}
\delta I &=& \left | \pderiv{I}{\mean{W(T_i)}} \right|^2 (\delta \mean{W(T_i)})^2 \\ \nonumber
& &
~~ + \left | \pderiv{I}{\mean{W(T_j)}} \right|^2 (\delta
\mean{W(T_j)})^2 \\ \nonumber
& & ~~ + 2 \left | \pderiv{I}{\mean{W(T_i)}}
\pderiv{I}{\mean{W(T_j)}} \right | \times \\ \nonumber
& & ~~\rho_{ij}
(\delta \mean{W(T_i)}) (\delta \mean{W(T_j)}) \\
&=& g(T_i)^2 (\delta \mean{W(T_i)})^2 + g(T_j)^2 (\delta \mean{W(T_j)})^2+ \nonumber \\
& & ~ 2 \left | g(T_i)g(T_j) \right | \rho_{ij} (\delta \mean{W(T_i)}) (\delta \mean{W(T_j)})
\end{eqnarray}
and the correlation coefficient $\rho_{ij}$ given by
\begin{equation}
\label{eqn:corrcoef}
\rho_{ij} = \frac{\mean{(W(T_i) - \mean{W(T_i)})(W(T_j)-\mean{W(T_j)})}}{\sqrt{(W(T_i)
-\mean{W(T_i)})^2}\sqrt{(W(T_j)-\mean{W(T_j)})^2}}.
\end{equation}
The final term in the error propagation equation takes into account correlations between the
random variables $W(T_i)$ and $W(T_j)$. Often in a Monte Carlo computation, one can
treat each evaluation of the integrand as independent, and neglect the uncertainty
term involving the correlation coefficient. However, in worldline numerics,
the evaluations are related because the same worldline ensemble is reused
for each evaluation of the integrand.
The correlations are significant (see figure \ref{fig:corr}), and this term
can't be neglected. Computing each correlation coefficient takes
a time proportional to the square of the number of worldlines. Therefore, it may
be computationally expensive to formally propagate uncertainties through an
integral.
\begin{figure}
\centering
\includegraphics[width=\linewidth,clip,trim=0.8cm 0.3cm 1.8cm 1cm]{images/corrrand}
\caption[Correlation coefficients between different evaluations of the integrand]
{Correlation coefficients, equation (\ref{eqn:corrcoef}), between $\mean{W(T)}$
and $\mean{W(T=3)}$ computed using individual worldlines, groups of worldlines, and
shuffled groups of worldlines.}
\label{fig:corr}
\end{figure}
The point-to-point correlations were originally pointed out by Gies
and Langfeld who addressed the problem by updating (but not replacing
or regenerating) the loop ensemble in between each evaluation of the
Wilson loop average~\cite{Gies:2001zp}. This may be a good way of
addressing the problem. However, in the following section, we promote a
method which can bypass the difficulties presented by the correlations
by treating the worldlines as a collection of worldline groups.
\subsubsection{Grouping Worldlines}
Both of the problems explained in the previous two subsections can be overcome
by creating groups of worldline loops within the ensemble. Each group of worldlines
then makes a statistically independent measurement of the Wilson loop average
for that group. The statistics between the groups of measurements are normally
distributed, and so the uncertainty is the standard error in the mean of the
ensemble of groups (in contrast to the ensemble of worldlines).
For example, if we divide the $N_l$ worldlines into $N_G$ groups of $N_l/N_G$
worldlines each, we can compute a mean for each group:
\begin{equation}
\mean{W}_{G_j} = \frac{N_G}{N_l}\sum_{i=1}^{N_l/N_G}W_i.
\end{equation}
Provided each group contains the same number of worldlines,
the average of the Wilson loop is unaffected by this grouping:
\begin{eqnarray}
\mean{W} & = & \frac{1}{N_G} \sum_{j=1}^{N_G} \mean{W}_{G_j} \\
& = & \frac{1}{N_l} \sum_{i=1}^{N_l} W_i.
\end{eqnarray}
However, the uncertainty is the standard error in the mean of
the groups,
\begin{equation}
\delta \mean{W} = \sqrt{\sum_{i=1}^{N_G}
\frac{(\mean{W}_{G_i} - \mean{W})^2}{N_G(N_G-1)}}.
\end{equation}
Because the worldlines are unrelated to one another, the choice of how to
group them to compute a particular Wilson loop average is arbitrary. For example,
the simplest choice is to group the loops by the order they were generated, so that
a particular group number, $i$, contains worldlines $iN_l/N_G$ through $(i+1)N_l/N_G -1$.
Of course, if the same worldline groupings are used to compute different Wilson
loop averages, they will still be correlated. We will discuss this problem in a moment.
The basic claim of the worldline technique is that the mean of the
worldline distribution approximates the holonomy factor. However, from
the distributions in figure \ref{fig:hists}, we can see that the
individual worldlines themselves do not approximate the holonomy
factor. So, we should not think of an individual worldline as an
estimator of the mean of the distribution. Thus, a resampling
technique is required to determine the precision of our statistics. We
can think of each group of worldlines as making an independent
measurement of the mean of a distribution. As expected, the groups of
worldlines produce a more Gaussian-like distribution (see figure
\ref{fig:uncinmean}), and so the standard error of the groups is a
sensible measure of the uncertainty in the Wilson loop value.
\begin{figure}
\centering
\includegraphics[width=\linewidth,clip,trim=0.8cm 0.3cm 1.8cm 1cm]{images/uncinmean}
\caption[Histogram for reproducing measurements with groups of worldlines]
{The histogram demonstrating the precision with which we can reproduce
measurements of the mean using different groups of 100 worldlines at $BT=6.0$.
In this case, the distribution is Gaussian-like and meaningful error bars can
be placed on our measurement of the mean.}
\label{fig:uncinmean}
\end{figure}
We find that the error bars are about one-third as large as those
determined from the standard error in the mean of the individual
worldlines, and the smaller error bars better characterize the size of
the residuals in the constant field case (see figure
\ref{fig:resids2}). The strategy of using subsets of the available
data to determine error bars is called jackknifing. Several previous
papers on worldline numerics have mentioned using jackknife analysis to
determine the uncertainties, but without an explanation of the
motivations or the procedure employed \cite{2005PhRvD..72f5001G,
PhysRevLett.96.220401, Dunne:2009zz, PhysRevD.84.065035}.
\begin{figure}
\centering
\includegraphics[width=\linewidth,clip,trim=0.8cm 0.3cm 1.8cm 1cm]{images/residsboth}
\caption[Comparison of error bars between standard error in the mean and
jackknife analysis]
{The residuals of the Wilson loops for a constant magnetic field showing
the standard error in the mean (thin error bars) and the uncertainty in determining
the mean (thick blue error bars). The standard error in the mean overestimates the
uncertainty by more than a factor of 3 at each value of $T$.}
\label{fig:resids2}
\end{figure}
The grouping of worldlines alone does not address the problem of
correlations between different evaluations of the integrands. Figure
\ref{fig:corr} shows that the uncertainties for groups of worldlines
are also correlated between different points of the integrand.
However, the worldline grouping does provide a tool for bypassing the
problem. One possible strategy is to randomize how worldlines are
assigned to groups between each evaluation of the integrand. This
produces a considerable reduction in the correlations, as is shown in
figure \ref{fig:corr}. Then, errors can be propagated through the
integrals by neglecting the correlation terms. Another strategy is to
separately compute the integrals for each group of worldlines, and
then consider the statistics of the final product to determine the
error bars. This second strategy is the one adopted for the work
presented in this paper. Grouping in this way reduces the amount of
data which must be propagated through the integrals by a factor of the
group size compared to a delete-1 jackknife scheme, for example. In
general, the error bars quoted in the remainder of this paper are obtained by
computing the standard error in the mean of groups of worldlines.
\subsection{Uncertainties and the Fermion Problem}
\label{sec:fermionproblem}
The fermion problem of worldline numerics is a name given to an enhancement of
the uncertainties at large $T$~\cite{Gies:2001zp,
MoyaertsLaurent:2004}. It should not be confused with the
fermion-doubling problem associated with lattice methods. In a
constant magnetic field, the scalar portion of the calculation
produces a factor of $\frac{BT}{\sinh{(BT)}}$, while the fermion
portion of the calculation produces an additional factor
$\cosh{(BT)}$. Physically, this contribution arises as a result of the
energy required to transport the electron's magnetic moment around the
worldline loop. At large values of $T$, we require subtle
cancellation between huge values produced by the fermion portion with
tiny values produced by the scalar portion. However, for large $T$,
the scalar portion acquires large relative uncertainties which make
the computation of large $T$ contributions to the integral very
imprecise.
This can be easily understood by examining the worldline distributions shown in figure
\ref{fig:hists}. Recall that the scalar Wilson loop average for these histograms is given
by the flux in the loop, $\Phi_B$:
\begin{equation}
\mean{W} = \left<\exp{\left(ie\int_0^Td\tau \vec{A}(\vec{x}_\mathrm{ cm}
+ \vec{x}(\tau))\cdot d\vec{x}(\tau)\right)}\right> = \left< e^{ie\Phi_B}\right>.
\end{equation}
For constant fields, the flux through the worldline loops obeys the distribution function
\cite{MoyaertsLaurent:2004}
\begin{equation}
f(\Phi_B) = \frac{\pi}{4BT\cosh^2\left(\frac{\pi \Phi_B}{2BT}\right)}.
\end{equation}
For small values of $T$, the worldline loops are small and the
amount of flux through the loop is correspondingly small. Therefore, the
flux for small loops is narrowly distributed about $\Phi_B = 0$. Since zero
maximizes the Wilson loop ($e^{i0}=1$),
this explains the enhancement to the right of the distribution for small values of $T$.
As $T$ is increased, the flux through any given worldline becomes very large and the
distribution of the flux becomes very broad.
For very large $T$, the width of the distribution is many
factors of $2\pi/e$. Then, the phase ($e \Phi_B\mod{2\pi}$) is nearly
uniformly distributed, and the
Wilson loop distribution reproduces the Chebyshev distribution ({\em i.e.}\
the distribution obtained from projecting uniformly distributed points on
the unit circle onto the horizontal axis),
\begin{equation}
\lim_{T\to\infty}w(y) = \frac{1}{\pi\sqrt{1-y^2}}.
\end{equation}
The mean of the Chebyshev distribution is zero due to its symmetry.
However, this symmetry is not
realized precisely unless we use a very large number of loops. Since the
width of the distribution is already 100$\times$ the value of the mean at
$T=6$, any numerical asymmetries in the distribution result in very large
relative uncertainties of the scalar portion. Because of these uncertainties,
the large contribution from the fermion factor are not cancelled precisely.
This problem makes it very difficult to compute
the fermionic effective action unless the fields are well localized
\cite{MoyaertsLaurent:2004}. For example, the fermionic factor for
non-homogeneous magnetic fields oriented along the z-direction is
\begin{equation}
\cosh{\left(e\int_0^T d\tau B(x(\tau))\right)}.
\end{equation}
For a homogeneous field, this function grows exponentially with $T$ and
is cancelled by the exponentially vanishing scalar Wilson loop.
For a localized field,
the worldline loops are very large for large values of $T$, and they primarily
explore regions far from the field. Thus, the fermionic factor grows more slowly
in localized fields, and is more easily cancelled by the rapidly vanishing scalar part.
In this section, we have identified two important considerations in
determining the uncertainties associated with worldline numerics computations.
Firstly, the computed points within the integrals over proper time,
$T$, or center of mass, $\vec{x}_\mathrm{ cm}$, are highly correlated
because one typically reuses the same ensemble of worldlines to
compute each point. Secondly, the statistics of the worldlines are
not normally distributed and each individual worldline in the ensemble
may produce a result which is very far from the mean value. So, in
determining the uncertainties in the worldline numerics technique, one should
consider how precisely the mean of the ensemble can be measured from
the ensemble and this is not necessarily given by the standard error
in the mean.
These issues can be addressed simultaneously using a scheme where the
worldlines from the ensemble are placed into groups, and the effective
action or the effective action density is evaluated separately for
each group. The uncertainties can then be determined by the statistics
of the groups. This scheme is less computationally intensive than a
delete-1 jackknife approach because less data (by a factor of the
group size) needs to be propagated through the integrals. It is also
less computationally intensive than propagating the uncertainties
through the numerical integrals because it avoids the computation of
numerous correlation coefficients.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=\linewidth,clip,trim=0.8cm 0.3cm 1.8cm 1cm]{images/Integrand}
\caption[small $T$ behaviour of worldline numerics]
{The small $T$
behaviour of worldline numerics. The data points represent the numerical
results, where the error bars are determined from the jackknife analysis described
in chapter \ref{ch:WLError}.
The solid line represents the exact solution while the dotted line represents
the small $T$ expansion of the exact solution. Note the amplification of
the uncertainties.}
\end{center}
\end{figure}
\section{Computing an Effective Action}
The ensemble average in the effective action is simply the sum over the
contributions from each worldline loop, divided by the number of loops in
the ensemble. Since the computation of each loop is independent of the
other loops, the ensemble average may be straightforwardly parallelized by
generating separate processes to compute the contribution from each loop.
For this parallelization, four Nvidia Tesla C1060 \acp{GPU} were used through
the \ac{CUDA} parallel programming framework. Because \acp{GPU} can spawn
thousands of parallel processing threads\footnote{Each Tesla C1060 device has 30 multiprocessors
which can each support 1,024 active threads. So, the number of threads available at a time
is 30,720 on each of the Tesla devices. Billions of threads may be scheduled on each
device~\cite{cudazone}.}
with much less computational overhead
than an \ac{MPI} cluster, they excel at handling a very large number of parallel
threads, although the clock speed is slower and fewer memory resources are typically available.
In contrast, parallel computing on a cluster using \acs{CPU}
tends to have a much higher speed per thread, but there are typically fewer
threads available. The worldline technique is exceedingly parallelizable,
and it is a straightforward matter to divide the task into thousands or tens of thousands of
parallel threads. In this case, one should expect excellent performance
from a \ac{GPU} over a parallel \ac{CPU} implementation, unless thousands
of \acp{CPU} are available for the program. The \ac{GPU} architecture has
recently been used by another group for computing Casimir forces using
worldline numerics~\cite{2011arXiv1110.5936A}.
Figure \ref{fig:coprocessing}
illustrates the co-processing and parallelization scheme used here for the
worldline numerics.
\begin{figure}
\centering
\includegraphics[width=\linewidth,clip,trim=-0.5cm 0 0 0]{images/CoprocessingDiagram}
\caption[Heterogeneous processing scheme for worldline numerics]
{The \ac{CPU} manages the loops which compute the
integrals over center of mass and proper time.
For each proper time integral, we require the results
from a large number of individual worldlines. The \ac{GPU}
is used to compute the integral
over each worldline in parallel, and the
results are returned to the \ac{CPU} for use in the effective action
calculation.}.
\label{fig:coprocessing}
\end{figure}
In an
informal test, a Wilson loop average was computed using an ensemble of 5000
loops in 4.7553 seconds using a serial implementation
while the \ac{GPU} performed the same calculation
in 0.001315 seconds using a \ac{CUDA} implementation.
A parallel \ac{CPU} code would require a cluster with thousands of
cores to achieve a similar speed, even if we assume linear (ideal)
speed-up. So, for worldline numerics computations,
a relatively inexpensive \ac{GPU} can be expected to outperform a small
or mid-sized cluster. This increase in computation speed has enabled the
detailed parameter searches discussed in this dissertation.
Of course, there are also trade-offs from using the
\ac{GPU} architecture with the worldline numerics technique. The most significant
of these is the limited availability of memory on the device. A \ac{GPU} device
provides only a few GB of global memory (4GB on the Tesla C1060). This limit
forces compromises between the number of points-per-loop and the number of
loops to keep the total size of the loop cloud data small. The limited availability of memory
resources also limits the number of threads which can be executed concurrently on the device.
Because of the overhead associated with transferring data to and from the device,
the advantages of a \ac{GPU} over a \ac{CPU} cluster are most pronounced
on problems which can be divided into several hundred or thousands of processes.
Therefore, the \ac{GPU} may not offer great performance advantages
for a small number of loops. Finally, there is some additional complexity involved
in programming for the \ac{GPU} in terms of learning specialized libraries and
memory management on the device. This means that the code may take longer to develop
and there may be a learning curve for researchers. However, this problem
is not much more pronounced with \ac{GPU} programming than with other
parallelization strategies.
Once the ensemble average of the Wilson loop has been computed,
computing the effective action is a straightforward matter of
performing numerical integrals. The effective action density is
computed by performing the integration over proper time, $T$. Then,
the effective action is computed by performing a spacetime integral
over the loop ensemble center of mass. In all cases where a numerical
integral was performed, Simpson's method was
used~\cite{burden2001numerical}. Integrals from 0 to $\infty$ were
mapped to the interval $[0,1]$ using substitutions of the form $x =
\frac{1}{1+T/T_\mathrm{ max}}$, where $T_\mathrm{ max}$ sets the scale
for the peak of the integrand. In the constant field case, for the
integral over proper time, we expect $T_\mathrm{ max} \sim 3/(eB)$ for
large fields and $T_\mathrm{ max} \sim 1$ for fields of a few times
critical or smaller. In section~\ref{ch:WLError}, we presented a
detailed discussion of how the statistical and discretization
uncertainties can be computed in this technique.
In this implementation, the numerical integrals are done using a
serial CPU computation. This serial portion of the algorithm tends to
limit the speedup that can be achieved with the large number of
parallel threads on the \ac{GPU} device\footnote{By Amdahl's law, for
a program with a ratio, $P$, of parallel to serial computations on
$N$ processors, the speedup is given by $S <
\frac{1}{(1-P)+\frac{P}{N}}$. This law predicts rapidly diminishing
returns from increasing the number of processors when $P>0$ for a
fixed problem size.}. However, an important benefit of the large
number of threads available on the \ac{GPU} is that the number of
worldlines in the ensemble can be increased without limit, as long as
more threads are available, without significantly increasing the
computation time. If perfect occupancy could be achieved on a Tesla
C1060 device, an ensemble of up to 30,720 worldlines could be computed
concurrently. Thus, the \ac{GPU} provides an excellent architecture
for improving the statistical uncertainties. Additionally, there is
room for further optimization of the algorithm by parallelizing the
serial portions of the algorithm to achieve a greater speedup.
More details about implementing this algorithm on the \ac{CUDA}
architecture can be found in appendix~\ref{ch:cudafication}. A listing
of the \ac{CUDA} worldline numerics code appears in appendix of
\cite{2012PhDT........21M}.
\section{Verification and Validation}
The worldline numerics software can be validated and verified by making sure that it
produces the correct results where the derivative expansion is a good approximation,
and that the results are consistent with previous numerical calculations of flux tube
effective actions. For this reason, the validation was done primarily with flux tubes with
a profile defined by the function
\begin{equation}
f_\lambda(\rho^2) = \frac{\rho^2}{(\lambda^2 + \rho^2)}.
\end{equation}
For large values of $\lambda$, this function varies slowly on the Compton wavelength
scale, and so the derivative expansion is a good approximation. Also, flux tubes
with this profile were studied previously using worldline numerics
\cite{Moyaerts:2003ts, MoyaertsLaurent:2004}.
Among the results presented in~\cite{MoyaertsLaurent:2004} is a comparison of
the derivative expansion and worldline numerics for this magnetic field
configuration. The result is that the next-to-leading-order term
in the derivative expansion is only a small correction to the the
leading-order term for $\lambda \gg \lambda_e$, where the derivative
expansion is a good approximation. The derivative expansion breaks
down before it reaches its formal validity limits
at $\lambda \sim \lambda_e$. For this reason,
we will simply focus on the leading order derivative expansion (which we call
the \ac{LCF} approximation).
The effective action of \ac{QED} in the \ac{LCF} approximation is given
in cylindrical symmetry by
\begin{eqnarray}
\label{eqn:LCFferm}
\Gamma^{(1)}_\mathrm{ ferm} &=& \frac{1}{4\pi}\int_0^\infty dT
\int_0^\infty \rho_\mathrm{ cm} d\rho_\mathrm{ cm}\frac{e^{-m^2T}}{T^3} \times \nonumber \\
& &\biggl\{eB(\rho_\mathrm{ cm})T\coth{(eB(\rho_\mathrm{ cm})T)} \\
& &
- 1 -\frac{1}{3}(eB(\rho_\mathrm{ cm})T)^2\biggr\}. \nonumber
\end{eqnarray}
\begin{figure}
\centering
\includegraphics[width=\linewidth,clip,trim=0.8cm 0.3cm 1.2cm 1cm]{images/igrandvsT4_F10l21}
\caption[Comparison with derivative expansion for $T$ integrand]
{The integrand of the proper time, $T$, integral for three different values
of the radial coordinate, $\rho$ for a $\lambda = 1$ flux tube. The solid lines
represent the zeroth-order derivative expansion, which, as expected, is a good approximation
until $\rho$ becomes too small.}
\label{fig:igrandvsT}
\end{figure}
Figure \ref{fig:igrandvsT} shows a comparison between the proper time integrand,
\begin{equation}
\frac{e^{-m^2T}}{T^3}\left[\langle W\rangle_{\vec{r}_\mathrm{ cm}}
- 1 -\frac{1}{3}(eB_\mathrm{ cm}T)^2\right],
\end{equation}
and the \ac{LCF} approximation result for a flux tube with $\lambda = \lambda_e$ and
$\mathcal{F} = 10$. In this case, the \ac{LCF} approximation is only appropriate far from the
center of the flux tube, where the field is not changing very rapidly. In the figure, we can
begin to see the deviation from this approximation, which gets more pronounced closer to the
center of the flux tube (smaller values of $\rho$).
\begin{figure}
\centering
\includegraphics[width=\linewidth,clip,trim=0.2cm 0.3cm 1.8cm 1cm]{images/vsxcm_ferm_smooth}
\caption[Comparison with \acs{LCF} approximation for action density]
{The fermion term of the effective action density as a function
of the radial coordinate for a flux tube with width $\lambda = 10 \lambda_e$. }
\label{fig:vsrhocm}
\end{figure}
The effective action density for a slowly varying flux tube is plotted in figure
\ref{fig:vsrhocm} along with the \ac{LCF} approximation. In this case,
the \ac{LCF} approximation agrees within the statistics of the worldline numerics.
\section{Conclusions}
In this paper, we have reviewed the worldline numerics numerical technique with
a focus on computing the effective action of \ac{QED} in
non-homogeneous, cylindrically symmetric magnetic fields. The method
uses a Monte Carlo generated ensemble of worldline loops to
approximate a path integral in the worldline formalism. These
worldline loops are generated using a simple algorithm and encode the
information about the magnetic field by computing the flux through the
loop and the action acquired from transporting a magnetic moment
around the loop. This technique preserves Lorentz symmetry exactly
and can preserve gauge symmetry up to any required precision.
We have discussed implementing this technique on \ac{GPU} architecture using
\ac{CUDA}. The main advantage of this architecture is that it allows for a
very large number of concurrent threads which can be utilized with
very little overhead. In practice, this means that a large ensemble
of worldlines can be computed concurrently, thus allowing for a considerable
speedup over serial implementations, and allowing for the precision of the
numerics to increase according to the number of threads available.
This work was supported by the Natural Sciences and Engineering Research
Council of Canada, the Canadian Foundation for Innovation, the British
Columbia Knowledge Development Fund. It has made used of the NASA ADS
and arXiv.org.
\bibliographystyle{prsty}
|
\section*{Introduction}
Automatic text classification is an important text mining task, due to the huge number of text documents that we have to manage daily. Text classification has a wide variety of applications such as Web document and email classification. Indeed, most of the Web news services daily provide a large number of articles making them impossible to be organized manually~\cite{lang_ICML_1995}. Automatic subject classification~\cite{cohen_AAAI_1996} and SPAM filtering~\cite{shami_etal_AAAI_1998} are two additional examples of the interest of automatic text classification.
Automatic text classification can be defined as below. Given a set of documents such that each document is labeled with a class value, learn a model that assigns a document with unknown class to one or more particular classes. This can also be done by assigning a probability value to each class or by ranking the classes.
A wide variety of classical machine learning techniques have been used for text classification. Indeed, texts may be represented by word frequencies vectors, and thus most of the quantitative data methods can be used directly on the notorious “bag-of-words” model (cf.~\cite{sebastiani_ACM-CS_2002,aggarwal_zhai_MTD_2012}).
Choosing a classifier is a multicriteria problem. In particular one has often to make a trade-off between accuracy and comprehensibility. In this paper, we are interested in both criteria with a deeper interest in comprehensibility. We are thus interested in rule-based approaches and especially in class association rules algorithms. Several studies have already successfully considered association rule-based approaches in text mining (e.g., \cite{ahonen_etal_TR_1997}, \cite{zaiane_antonie_ADC_2002}, \cite{cherfi_etal_PMAR_2009}, \cite{roche_etal_IIPWM_2004}). This framework is suitable for considering some statistical characteristics (e.g., high-dimensionality, sparsity\ldots) of the bag-of-words model where a document is represented as a set of words with their associated frequency in the document.
However a text is more than a set of words and their frequencies. Enhancing the bag-of-words approach with linguistic features has also attracted several works (e.g., \cite{jaillet_etal_IDA_2006,do_Master_2012,Kovacs:2008da,OrdonezSalinas:2010db}, \cite{Pado:2007bu,Lowe:2001wx,Curran:2002vm}, \cite{nivre_TR_2005,FerreriCancho:2004wd}).
We here propose a class association rules based approach enriched by linguistic knowledge. The paper is organized as follows: after introducing the techniques we are going to use (class association rules \S\,\ref{car-sec}, dependencies \S\,\ref{dar}, hyperonymization \S\,\ref{hc}) we describe our main algorithms (for training \S\,\ref{train-sec}, classifying \S\,\ref{classify-sec} and evaluating \S\,\ref{evaluate-sec}); follows the experimental section, where we give results obtained by tfidf pruning \S\,\ref{tfidf}, dependency-based pruning \S\,\ref{dep-sec} and hyperonymization \S\,\ref{hyper-sec}, and, finally, we end up by a conclusion and perspectives for future work \S\,\ref{con}.
\section{Proposed model for text classification}
Let a \emph{corpus} be a set $\mathbf{C}=\{D_1,\ldots,D_n\}$ of documents. Let $\mC$ be a set of classes. An \emph{annotated corpus} is a pair $(\mathbf{C},\mathrm{class})$ where $\mathrm{class}:\mathbf{C}\to\mC$ is a function that maps each document $D_i$ to a (predefined) class of $\mC$.
A document $D\in\mathbf{C}$ is a set of sentences $S$. The corpus $\mathbf{C}$ can be considered as a set of sentences $\mathbf{S}=\{S_1,\ldots,S_m\}$ if we go through the forgetful functor (which forgets the document to which the sentence belongs). Repeated sentences in the same document, or identical sentences in different documents are considered as distinct, i.e., there is a function $\iota\colon\mathbf{S}\to\mathbf{C}$ which restores the forgotten information. We extend the $\mathrm{class}$ function to $\mathbf{S}$ by $\mathrm{class}(S):=\mathrm{class}(\iota(S))$.
A sentence $S$ is a sequence of words $w$ (sometimes we will consider $S$ simply as a set, without changing the notation). Let $\mW=\bigcup_{S\in\mathbf{S}} \bigcup_{w\in S}\{w\}$ be the set of all words of~$\mathbf{C}$.
\medskip
\subsection{Class association rules and text classification}\label{car-sec}
Let $\mI$ be a set of objects called \emph{items} and $\mC$ a set of classes. A \emph{transaction} $T$ is a pair $(\{i_1,\ldots,i_n\},c)$, where $\{i_1,\ldots,i_n\}\subseteq\mI$ and $c\in\mC$. We denote by $\mT$ the set of transactions, by $\mathrm{items}(T)$ the set of items (or “itemset”) of $T$ and by $\mathrm{class}(T)$ the class of $T$.
Let $I$ be an itemset. The \emph{support} of $I$ is defined by
$$\textstyle\mathrm{supp}(I):=\frac{\#\{T\in\mT\mid I\subseteq \mathrm{items}(T)\}}{\#\mT}.$$
Let $\sigma\in[0,1]$ be a value called \emph{minimum support}. An itemset $I$ is called \emph{frequent} if its \emph{support} exceeds $\sigma$.
The \emph{confidence} of a transaction $t$ is defined as
$$\textstyle\mathrm{conf}(t):=\frac{\#\{T\in\mT\mid \mathrm{items}(t)\subseteq\mathrm{items}(T)\wedge\mathrm{class}(t)=\mathrm{class}(T)\}}{\#\{T\in\mT\mid \mathrm{items}(t)\subseteq\mathrm{items}(T)\}}.$$
Let $\kappa\in[0,1]$ be a value called \emph{minimum confidence}. A \emph{class association rule} (or “CAR”) $r=(\{i_1,\ldots,i_n\},c)$ \cite{CAR} is a transaction with frequent itemset and a \emph{confidence} exceeding $\kappa$.
To classify text with CARs, we consider words as being items, documents as being itemsets and pairs of documents and classes as being transactions. The advantage of this technique is that CARs can be easily understood and hence potentially improved by the user, especially if the classifier is tuned so that it produces humanly reasonable number of rules. Once the classifier is trained, to classify a new sentence we first find all CARs whose items are contained in the sentence, and then use an aggregation technique to choose a predominant class among those of the CARs we found.
An important issue of CARs is that the complexity is exponential with respect to the itemset size, and hence we need to keep it bounded in specific ranges, independently of the size of documents to classify. Using entire documents as transactions is computationally out of reach, therefore pruning techniques play an important rôle. Our approach consists in (a) restricting CARs to the sentence level, (b) prune sentences by using morphosyntactic information (cf. \S\,\ref{dar}) and modifying itemsets using semantic information (cf. \S\,\ref{hc}).
\subsection{Itemset pruning using dependencies}\label{dar}
One can prune sentences either by using word frequencies (cf. \S\,\ref{tfidf}) or by using information obtained by morphosyntactic parsing (cf. \S\,\ref{dep-sec}). In this paper we introduce the latter approach, in the frame of dependency grammar.
\emph{Dependency grammar} \cite{tesniere1959,melcuk} is a syntactic theory, alternative to \emph{phrase-structure analysis} \cite{chomsky1957} which is traditionally taught in primary and secondary education. In phrase-structure syntax, trees are built by grouping words into “phrases” (with the use of intermediate nodes NP, VP, etc.), so that the root of the tree represents the entire sentence and its leaves are the actual words. In dependency grammar, trees are built using solely words as nodes (without introducing any additional “abstract” nodes). A single word in every sentence becomes the \emph{root} (or \emph{head}) of the tree. An oriented edge between two words is a \emph{dependency} and is tagged by a representation of some (syntactic, morphological, semantic, prosodic, etc.) relation between the words. For example in the sentence “John gives Mary an apple,” the word “gives” is the head of the sentence and we have the following four dependencies:
\begin{center}\begin{dependency}
\begin{deptext}[row sep=-5pt,column sep=1cm]
John \& gives \& Mary \& an \& apple.\\
\end{deptext}
\deproot{2}{head}
\depedge{5}{2}{dobj}
\depedge{1}{2}{nsubj}
\depedge{3}{2}{iobj}
\depedge{4}{5}{det}
\end{dependency}
\end{center}
\noindent where tags nsubj, dobj, iobj, det denote “noun subject,” “direct object,” “indirect object” and “determinant.”
Let $S$ be a sentence and $\mD$ be the set of dependency tags: \{nsubj, ccomp, prep, dobj, \ldots\} A \emph{dependency} is a triple $(w_1,w_2,d)$ where $w_1,w_2\in S$ and $d\in\mD$. Let $\mathrm{Dep}(S)$ denote the set of dependencies of $S$ and $\mathrm{root}(S)$ the head of~$S$. Pruning will consist in defining a \emph{morphosyntactic constraint} $\phi$ i.e. a condition on dependencies (and POS tags) of words, the fulfillment of which is necessary for the word to be included in the itemset.
But before describing pruning algorithms and strategies, let us first present a second technique used for optimizing itemsets. This time we use semantic information. We propose to replace words by their hyperonyms, expecting that the frequencies of the latter in the itemsets will be higher than those of the former, and hence will improve the classification process.
\subsection{Hyperonymization}\label{hc}
The WordNet lexical database \cite{miller1995} contains sets of words sharing a common meaning, called \emph{synsets}, as well as semantic relations between synsets, which we will use to fulfill our goal. More specifically, we will use the relations of \emph{hyperonymy} and of \emph{hyperonymic instance}. The graph having synsets as nodes, and hyperonymic relations as edges, is connected and rooted: starting with an arbitrary synset, one can iterate these two relations until attaining a sink. {Note that} in the case of nouns it will invariably be the synset \petitid{00001740} \{entity\} while for verbs there are approx.\ 550 different verb sinks.
Let $\mathbb W$ be the WordNet lexical database, $s\in \mathbb W$ a synset and $h:\mathbb W\to 2^{\mathbb W}$ the hyperonymic or hyperonymic instance relation. We define an \emph{hyperonymic chain} $\mathrm{CH}(s)$ as a sequence $(s_i)_{i\geq0}$ where $s_0=s$ and $s_i\in h(s_{i-1})$, for all $i\geq 1$. Hyperonymic chains are not unique since a given synset can have many hyperonyms. To replace a word by the most pertinent hyperonym, we have to identify the most significant hyperonymic chains of it.
The \emph{wn-similarity} project \cite{pedersen} has released synset frequency calculations based on various corpora. Let $\mathrm{lf}(s)$ denote the logarithmic frequency of synset~$s$ in the BNC English language corpus~\cite{bnc} and let us arbitrarily add infinitesimally small values to the frequencies so that they become unique ($s\ne s'\Rightarrow\mathrm{lf}(s)\ne\mathrm{lf}(s')$). We use frequency as the criterion for selecting a single hyperonymic chain to represent a given synset, and hence define the \emph{most significant hyperonymic chain} $\mathrm{MSCH}(s)$ as the hyperonymic chain $(s_i)_{i\geq0}$ of $s$ such that $s_i=\argmax_{s\in h(s_{i-1})}\mathrm{lf}(s)\text{, for all $i\geq1$}.$ The chain $\mathrm{MSCH}(s)$ is unique thanks to the uniqueness of synset frequencies.
Our CARs are based on words, not synsets. Hence we need to extend MSCHs to words. Let $w$ be a lemmatized word. We denote by $\mathrm{Synsets}(w)\subset\mathbb W$ the set of synsets containing $w$. If the cardinal $\#(\mathrm{Synsets}(w))>1$ then we apply a standard disambiguation algorithm to find the most appropriate synset $s_w$ for $w$ in the given context. Then we take $(s_i)_i=\mathrm{MSCH}(s_w)$ and for each synset $s_i$ in this chain we define $h_i(w)=\mathrm{proj}_1(s_i)$ $(i>0)$, that is the projection of $s_i$ to its first element, which by WordNet convention is the most frequent word in the synset. The function vector $h_*:\mW\to\mW$ (with $h_0\equiv\mathrm{Id}$) is called \emph{hyperonymization}, and $h_i(w)$ is the \emph{$i$-th order hyperonym of $w$}.
\section{Operational implementations for document classification}
Our text classifier operates by first training the classifier on sentences and then classifying the documents by aggregating sentence classification. These two procedures are described in Sections~\ref{train-sec} and~\ref{classify-sec} respectively. Specific evaluation procedure is presented in Section~\ref{evaluate-sec}.
\subsection{Training}\label{train-sec}
\begin{algorithm}[tb]
\SetKwFunction{train}{Train}
\SetKwFunction{prune}{Prune}
\SetKwFunction{hyper}{Hyperonymize}
\SetKwFunction{lemma}{Lemmatize}
\SetKwFunction{apriori}{Apriori}
\SetKw{KwTo}{:=}
\SetKwProg{Fn}{}{:}{end}
\KwData{An annotated corpus $\mathbf{C}$, values of minimum support $\sigma$ and minimum confidence $\kappa$}
\KwResult{A set of CARs $\mR=(\{R_1,\ldots,R_N\},\mathrm{conf})$ where $\mathrm{items}(R_i)\subset \mW$, $\mathrm{class}(R_i)\in\mC$, and $\mathrm{conf}(R_i)$ is the confidence of rule $R_i$}
\Fn(){\train{$\mathbf{C}$, $\sigma$, $\kappa$}}{
$\mathbf{S}$ \KwTo $\mathrm{forgetful}(\mathbf{C})$;
$\mathbf{S}'$ \KwTo $\emptyset$\;
\For{$S\in\mathbf{S}$}
{
$S'$ \KwTo \hyper(\prune(\lemma($S$)))\;
$\mathrm{class}(S')$ \KwTo $\mathrm{class}(\iota(S))$\;
$\mathbf{S}'$ \KwTo $\mathbf{S}'\cup \{S'\}$\;
}
$\mR$ \KwTo \apriori($\mathbf{S}',\sigma,\kappa$)\;
}
\caption{Training}\label{train}
\end{algorithm}
The \texttt{Train} algorithm (cf. Alg.~\ref{train}) takes as input an annotated corpus $\mathbf{C}$ and values of minimum support $\sigma$ and minimum confidence $\kappa$. It returns a set of CARs together with their confidence values.
The first part of the algorithm consists in processing the corpus, to obtain efficient and reasonably sized transactions. Three functions are applied to every sentence:
\begin{enumerate}
\item \texttt{Lemmatize} is standard lemmatization: let $\mP$ be the set of POS tags of the \emph{TreeTagger} system \cite{treetagger} (for example, NP stands for “proper noun, singular”, VVD stands for “verb, past tense”, etc.), and let $\mW'$ be the set of lemmatized forms of $\mW$ (for example, “say” is the lemmatized form of “said”); then we define
$\lambda:\mW\to(\mW\cup\mW')\times\mP$,
which sends a word $w$ to the pair $(w',p)$ where $w'$ is the lemmatized form of $w$ (or $w$ itself, if the word is unknown to \emph{TreeTagger}) and $p$ is its POS tag.
\item \texttt{Prune} is a function which prunes the lemmatized sentence so that only a small number of (lemmatized) words (and POS tags) remains. Several sentence pruning strategies will be proposed and compared (cf. \S\,\ref{tfidf} and \ref{dep-sec}).
\item \texttt{Hyperonymize} is a function which takes the words in the pruned itemset and replaces them by the members of their most significant hyperonymic chains. Several strategies will also be proposed and compared (cf. \S\,\ref{hyper-sec}).
\end{enumerate}
The second part of Alg.~\ref{train} uses the \emph{apriori} algorithm \cite{apriori} with the given values of minimum support and minimum confidence and output restrictions so as to generate only rules with item $c\in\mC$ in the consequent. It returns a set $\mR$ of CARs and their confidence.
It should be noted that this algorithm operates on individual sentences, hereby ignoring the document level.
\subsection{Classification}\label{classify-sec}
\begin{algorithm}[tb]
\SetKwFunction{classify}{Classify}
\SetKwFunction{prune}{Prune}
\SetKwFunction{hyper}{Hyperonymize}
\SetKwFunction{lemma}{Lemmatize}
\SetKwFunction{apriori}{Apriori}
\SetKwFunction{sort}{Sort}
\SetKw{KwTo}{:=}
\SetKwProg{Fn}{}{:}{end}
\KwData{A set of CARs $\mR$, a document $D_0$}
\KwResult{The predicted class $\mathrm{predclass}(D_0)$, variety $\beta$, dispersion $\Delta$}
\Fn(){\classify{$\mR$, $D_0$}}{
\For{$S\in D_0$}
{
\If{$\exists r\in \mR$ such that $\mathrm{items}(r)\subset S$}{
$R_{S}$ \KwTo $\displaystyle\argmax_{r\in\mR\wedge \mathrm{items}(r)\subset S}\mathrm{conf}(r)$\;}
}
$\mathrm{predclass}(D_0)$ \KwTo $\displaystyle\argmax_{c\in\mC}\sum_{\substack{S\in D_0\\\mathrm{class}(R_{S})=c}}\mathrm{conf}(R_{S})$\;
$\beta$ \KwTo $\#\{c\in\mC\mid (\mathrm{class}(R_{S})=c)\wedge (\mathrm{conf}(R_{S})>0)\}$\;
$\Delta$ \KwTo $\displaystyle\max_{c\in\mC}\sum_{\substack{S_i\in D_0\\\mathrm{class}(R_{S_i})=c}}\mathrm{conf}(R_{S_i})-\min_{c\in\mC}\sum_{\substack{S_i\in D_0\\\mathrm{class}(R_{S_i})=c}}\mathrm{conf}(R_{S_i})$\;
}
\caption{Classification}\label{classify}
\end{algorithm}
The \texttt{Classify} algorithm (cf. Alg.~\ref{classify}) uses the set of CARs produced by \texttt{Train} to predict the class of a new document $D_0$ and furthermore provides two values measuring the quality of this prediction: variety $\beta$ and dispersion $\Delta$.
The first part of the algorithm takes each sentence $S$ of the document $D_0$ and finds the most confident CAR that can be applied to it (i.e., such that the itemset of the rule is entirely contained in the itemset of the sentence). At this stage we have, for every sentence: a rule, its predicted class and its confidence.
Our basic unit of text in \texttt{Train} is sentence, therefore CARs generated by Alg.~\ref{train} produce a class for each sentence of $D_0$. An aggregation procedure is thus needed in order to classify the document.
This is done by taking class by class the sum of confidence of rules and selecting the class with the highest sum.
Although this simple class-weighted sum decision strategy is reasonable, it is not perfect and may lead to wrong classification. This strategy will be \textit{optimally} sure and robust if (a) the number of classes is minimal, and (b) the values when summing up confidence of rules are sufficiently spread apart. The degree of fulfillment of these two conditions is given by the parameters \emph{variety}~$\beta$ (the number of classes for which we have rules), and \emph{dispersion} $\Delta$ (the gap between the most confident class and least confident one). These parameters will contribute to comparison among the different approaches we will investigate.
\subsection{Evaluation}\label{evaluate-sec}
\begin{algorithm}[tb]
\SetKwFunction{traincomparable}{TrainComparable}
\SetKwFunction{classify}{Classify}
\SetKwFunction{evaluate}{Evaluate}
\SetKwFunction{singleevaluate}{SingleEvaluate}
\SetKwFunction{prune}{Prune}
\SetKwFunction{findoptimal}{FindOptimal}
\SetKwFunction{hyper}{Hyperonymize}
\SetKwFunction{lemma}{Lemmatize}
\SetKwFunction{apriori}{Apriori}
\SetKwFunction{partition}{Partition}
\SetKwFunction{shuffle}{Shuffle}
\SetKwFunction{sort}{Sort}
\SetKw{KwTo}{:=}
\SetKwProg{Fn}{}{:}{end}
\KwData{An annotated corpus $\mathbf{C}$, initial values of minimal support $\sigma_0$ and confidence $\kappa_0$, standard number of rules $\rho_0$}
\KwResult{Values of average precision $\overline P$, recall $\overline R$, F-measure $\overline F$. Values of average number of rules $\rho$, variety ${\beta}$ and dispersion ${\Delta}$}
\Fn(){\singleevaluate{$\mathbf{C}$, $\sigma$, $\kappa$}}{
$(\mathbf{C}_1,\ldots,\mathbf{C}_{10})$ \KwTo \partition(\shuffle($\mathbf{C}$),10)\;
\tcc{tenfold cross validation}
\For{$I\in \{1,\ldots,10\}$}
{
($\mR_I$, $\beta_I$, $\Delta_I$) \KwTo \train($\mathbf{C}\setminus\mathbf{C}_1$, $\sigma$, $\kappa$)\;
\For{$D\in\mathbf{C}_I$}{
$\mathrm{predclass}(D)$ \KwTo \classify($\mR_I$, $D$)\;
}
\For{$c\in\mC$}{
$R_I(c)$ \KwTo $\frac{\#\{d\in\mathbf{C}_I\mid (\mathrm{predclass}(d)=c)\wedge(\mathrm{class}(d)=c)\}}{\#\{d\in\mathbf{C}_I\mid \mathrm{class}(d)=c\}}$\;
$P_I(c)$ \KwTo $\frac{\#\{d\in\mathbf{C}_I\mid (\mathrm{predclass}(d)=c)\wedge(\mathrm{class}(d)=c)\}}{\#\{d\in\mathbf{C}_I\mid \mathrm{predclass}(d)=c\}}$;
$F_I(c)$ \KwTo $\frac{2R_I(c)P_I(c)}{R_I(c)+P_I(c)}$\;
}
}
\For{$c\in\mC$}{
$(R(c),P(c),F(c))$ \KwTo $\frac{1}{10}\sum_{I=1}^{10}(R_I(c),P_I(c),F_I(c))$\;
}
$(\rho,\beta,\Delta)$ \KwTo $\frac{1}{10}\sum_{I=1}^{10}(\#\mR_I,\beta_I,\Delta_I)$\;
$(\overline R,\overline P,\overline F)$ \KwTo $\frac{1}{\#\mC}\sum_{c\in\mC}(R(c),P(c),F(c))$\;
}
\Fn(){\evaluate{$\mathbf{C}$, $\sigma_0$, $\kappa_0$, $\rho_0$}}{
$(\sigma,\kappa)$ \KwTo $\findoptimal(\mathbf{C},\sigma_0,\kappa_0,\rho_0)$\;
$(\overline R,\overline P,\overline F,\rho,\beta,\Delta)$ \KwTo $\singleevaluate(\mathbf{C},\sigma,\kappa)$\;
}
\caption{Evaluation}\label{evaluate}
\end{algorithm}
We evaluate the classifier (Alg.~\ref{evaluate}), by using 10-fold cross validation to obtain average values of recall, precision, F-measure, variety and dispersion. This is done by algorithm \texttt{SingleEvaluate}, once we specify values of minimal support and minimal confidence.
Comparing rule-based classification methods is problematic because one can always increase F-measure performance by increasing the number of rules, which results in overfitting them. To avoid this phenomenon and compare methods in a fair way, we fix a number of rules $\rho_0$ (we have chosen $\rho_0=1{,}000$ in order to produce a humanly reasonably readable set of rules) and find values of minimal support and confidence so that \mbox{F-measure} is maximal under this constraint.
Function \texttt{FindOptimal} will launch \texttt{SingleEvaluate} as many times as necessary on a dynamic grid of values $(\sigma,\kappa)$ (starting with initial values $(\sigma_0,\kappa_0)$), so that, at the end, the number of rules produced by \texttt{Train} is as close as possible to $\rho_0$ (we have used $\#\mR\in[\rho_0-2,\rho_0+2]$) and $\overline F$ is maximal.
\section{Experimental results on Reuters corpus}\label{corpus}
In this section, we investigate three methods: (a) pruning through a purely frequentist method, based on tfidf measure (\S\,\ref{tfidf}); (b) pruning using dependencies (\S\,\ref{dep-sec}); (c) pruning using dependencies followed by hyperonymic extension~(\S\,\ref{hyper-sec}).
\subsection{Preliminaries}
In the Reuters \cite{reuters} corpus we have chosen the 7 most popular topics (GSPO~= sports, E12~= monetary/economic, GPOL~= domestic politics, GVIO~= war, civil war, GDIP~= international relations, GCRIM~= crime, law enforcement, GJOB~= labor issues) and extracted the 1,000 longest texts of each.
The experimental document set is thus a corpus of 7,000 texts of length between 120 and 3,961 words (mean $398.84$, standard variation $169.05$). The texts have been analyzed with the Stanford Dependency Parser \cite{sdp} in collapsed mode with propagation of conjunct dependencies.
\subsection{Tfidf-based corpus pruning}\label{tfidf}
Tfidf-based corpus pruning consists in using a \textit{classical} \texttt{Prune} function as defined in Alg.~\ref{tfidf-alg}. It will be our baseline for measuring performance of dependency- and hyperonymy-based methods.
\begin{algorithm}[tb]
\SetKwFunction{classify}{Classify}
\SetKwFunction{evaluate}{Evaluate}
\SetKwFunction{prune}{Prune}
\SetKwFunction{hyper}{Hyperonymize}
\SetKwFunction{lemma}{Lemmatize}
\SetKwFunction{apriori}{Apriori}
\SetKwFunction{partition}{Partition}
\SetKwFunction{shuffle}{Shuffle}
\SetKwFunction{tfidf}{Tfidf}
\SetKwFunction{sort}{Sort}
\SetKw{KwTo}{:=}
\SetKwProg{Fn}{}{:}{end}
\KwData{An annotated corpus (considered as a set of sentences) $\mathbf{S}$}
\KwResult{The pruned corpus $\mathbf{S}'$}
\Fn(){\prune{$\mathbf{S}$, $N$}}{
$\mathbf{S}'$ \KwTo $\emptyset$\;
\For{$S\in \mathbf{S}$}{
\For{$w\in S$}{
$\tfidf_S(w)$ \KwTo $\freq_S(w)\cdot\log\left(\frac{\#\{S\in\mathbf{C}\}}{\#\{S\in\mathbf{C}\mid w\in S\}}\right)$
}
$S'$ \KwTo $\emptyset$; $S_0$ \KwTo $S$\;
\For{$i\in\{1,\ldots,N\}$}{
$w'$ \KwTo $\displaystyle\argmax_{S_0}\tfidf_S(w)$\;
$S_0$ \KwTo $S_0\setminus \{w'\}$; $S'$ \KwTo $S'\cup\{w'\}$\;
}
$\mathbf{S}'$ \KwTo $\mathbf{S}'\cup\{S'\}$\;
}
}
\caption{Tfidf-based corpus pruning}\label{tfidf-alg}
\end{algorithm}
Note that this definition of the tfidf measure diverges from the legacy one by the fact that we consider not documents but sentences as basic text units. This is because we compare tfidf-generated CARs to those using syntactic information, and syntax is limited to the sentence level. Therefore, in order to obtain a fair comparison, we have limited term frequency to the sentence level and our “document frequency” is in fact a sentence frequency.
Having calculated $\tfidf_S(w)$ for every $w\in S\in\mathbf{S}$, we take $N$ words from each sentence with the highest tfidf values, and use them as transaction items. The performance of this method depends on the value of $N$. On Fig.~\ref{tfidf-fig} the reader can see the values of three quantities as functions of $N$:
\begin{enumerate}
\item F-measure: we see that F-measure increases steadily and reaches a maximum value of 83.99 for $N=10$. Building transactions of more than 10 words (in decreasing tfidf order) deteriorates performance, in terms of F-measure;
\item variety: the number of predicted classes for sentences of the same document progressively increases but globally remains relatively low, around 3.1, except for $N=12$ and $N=13$ where it reaches $4.17$;
\item dispersion: it increases steadily, with again a small outlier for $N=12$, probably due to the higher variety obtained for that value of $N$.
\end{enumerate}
\begin{figure}[ht]
\centering
\resizebox{\textwidth}{!}{\includegraphics{tfidf1.pdf}~\includegraphics{tfidf3.pdf}~\includegraphics{tfidf4.pdf}}
\caption{F-measure, variety and dispersion of tfidf-based pruning methods as a function of the number of words kept in the corpus\label{tfidf-fig}}
\end{figure}
Furthermore, each investigated method will generate transactions of various sizes. It is fair to compare them with tfidf-based methods with similar transactions sizes. Therefore we will use the results displayed in Fig.~\ref{tfidf-fig} to compare the performance of subsequent methods with the one of the tfidf-based method of similar transaction size. Table~\ref{tab:results_tdfif_based_pruning_N1} presents the results obtained by applying the tdfif-based pruning method, with a single word per transaction\enlargethispage{.5\baselineskip} ($N=1$).
\begin{table}[h!]
\caption{Tfidf-based pruning, keeping a single word per transaction}
\noindent\begin{tabularx}{\textwidth}{l|CCCCCCC|C}
&E12&GCRIM&GDIP&GJOB&GPOL&GSPO&GVIO&AVG\\\hline
Recall&69.30&44.48&55.44&45.75&52.54&82.90&67.98&59.77\\
Precision&70.09&77.81&71.25&79.76&71.62&80.78&73.35&74.95\\
F-measure&69.69&56.60&62.36&58.15&60.61&81.83&70.56&65.69\\\hline
\multicolumn{9}{l}{\footnotesize MinSupp=0.006, MinConf=67.6, Var.=1.36, Disp.=21.53, AvgTransSize=1.00}
\end{tabularx}
\label{tab:results_tdfif_based_pruning_N1}
\end{table}
\subsection{Methods based on dependencies}\label{dep-sec}
\begin{algorithm}[tb]
\SetKwFunction{classify}{Classify}
\SetKwFunction{evaluate}{Evaluate}
\SetKwFunction{prune}{Prune}
\SetKwFunction{hyper}{Hyperonymize}
\SetKwFunction{lemma}{Lemmatize}
\SetKwFunction{apriori}{Apriori}
\SetKwFunction{partition}{Partition}
\SetKwFunction{shuffle}{Shuffle}
\SetKwFunction{tfidf}{Tfidf}
\SetKwFunction{sort}{Sort}
\SetKw{KwTo}{:=}
\SetKwProg{Fn}{}{:}{end}
\KwData{An annotated corpus $\mathbf{S}$ and a morphosyntactic contraint $\phi\colon S\to\{\mathrm{true},\mathrm{false}\}$}
\KwResult{The pruned corpus $\mathbf{S}'$}
\Fn(){\prune{$\mathbf{S}$, $\phi$}}{
$\mathbf{S}'$ \KwTo $\emptyset$\;
\For{$S\in \mathbf{S}$}{
$S'$ \KwTo $\emptyset$\;
\For{$w\in S$}{
\If{$\phi(w)=\mathrm{true}$}{$S'$ \KwTo $S'\cup\{w\}$}
}
$\mathbf{S}'$ \KwTo $\mathbf{S}'\cup\{S'\}$\;
}
}
\caption{Dependency-based corpus pruning}\label{dep-algo}
\end{algorithm}
\begin{subtables}
\begin{table}[ht]
\caption{Strategy I$_0$: Pruning by keeping only heads of sentences}
\kern-5pt
\noindent\begin{tabularx}{\textwidth}{l|CCCCCCC|C}
&E12&GCRIM&GDIP&GJOB&GPOL&GSPO&GVIO&AVG\\\hline
Recall&38.54&57.46&26.88&17.43&31.06&88.49&51.90&44.54\\
Precision&49.13&66.18&49.61&65.73&39.07&42.18&60.31&53.17\\
F-measure&43.19&61.51&34.87&27.55&34.61&57.13&55.79&44.95\\\hline
\multicolumn{9}{l}{\footnotesize MinSupp=0.004, MinConf=36.6, Var.=2.14, Disp.=47.84, AvgTransSize=1.00}
\end{tabularx}
\label{tab:dependency_head_sentence}
\medskip
\caption{Strategy I$_1$: Pruning by keeping only nsubj $\to$ head dependencies}
\kern-5pt
\noindent\begin{tabularx}{\textwidth}{l|CCCCCCC|C}
&E12&GCRIM&GDIP&GJOB&GPOL&GSPO&GVIO&AVG\\\hline
Recall&70.55&60.44&66.97&58.27&63.22&78.76&69.92&66.88\\
Precision&73.05&80.84&72.87&86.98&71.71&85.29&76.29&78.15\\
F-measure&71.78&69.17&69.80&69.79&67.19&81.89&72.97&71.80\\\hline
\multicolumn{9}{l}{\footnotesize MinSupp=0.007, MinConf=60.4, Var.=1.43, Disp.=41.71, AvgTransSize=1.04}
\end{tabularx}
\label{tab:dependency_nominal_subject}
\end{table}
\begin{table}[t]
\caption{Strategy I$'_1$: Pruning by keeping only ccomp $\to$ head dependencies}
\kern-5pt
\noindent\begin{tabularx}{\textwidth}{l|CCCCCCC|C}
&E12&GCRIM&GDIP&GJOB&GPOL&GSPO&GVIO&AVG\\\hline
Recall&57.33&33.82&25.31&16.96&21.74&47.62&59.60&37.48\\
Precision&37.83&47.55&38.50&42.06&34.62&57.05&54.17&44.54\\
F-measure&45.59&39.53&30.54&24.17&26.71&51.91&56.75&39.31\\\hline
\multicolumn{9}{l}{\footnotesize MinSupp=0.008, MinConf=34.4, Var.=1.97, Disp.=19.59, AvgTransSize=1.15}
\end{tabularx}
\label{tab:dependency_clausal_complement}
\medskip
\caption{Strategy I$_2$: Pruning by keeping only nouns at distance~1 from head}
\noindent\begin{tabularx}{\textwidth}{l|CCCCCCC|C}
&E12&GCRIM&GDIP&GJOB&GPOL&GSPO&GVIO&AVG\\\hline
Recall&80.75&75.92&73.24&68.59&70.59&95.55&77.96&77.51\\
Precision&73.21&83.51&75.35&89.86&73.67&80.52&77.36&79.07\\
F-measure&76.80&79.53&74.28&77.80&72.09&87.39&77.66&77.94\\\hline
\multicolumn{9}{l}{\footnotesize MinSupp=0.016, MinConf=51.6, Var.=2.43, Disp.=244.82, AvgTransSize=2.70}
\end{tabularx}
\label{tab:dependency_pos_tags}
\medskip
\caption{Strategy I$'_2$: Pruning by keeping only verbs at distance~1 from head}
\kern-5pt
\noindent\begin{tabularx}{\textwidth}{l|CCCCCCC|C}
&E12&GCRIM&GDIP&GJOB&GPOL&GSPO&GVIO&AVG\\\hline
Recall&54.32&62.50&44.37&20.41&27.58&91.39&67.68&52.61\\
Precision&49.58&65.98&48.44&78.39&46.69&43.57&63.84&56.64\\
F-measure&51.84&64.19&46.32&32.39&34.68&59.01&65.70&50.59\\\hline
\multicolumn{9}{l}{\footnotesize MinSupp=0.019, MinConf=30, Var.=4.00, Disp.=175.41, AvgTransSize=2.01}
\end{tabularx}
\label{tab:dependency_verbs_distance1_head}
\end{table}
\end{subtables}
In this section we investigate several strategies using the dependency structure of sentences. Our general approach (cf. Alg.~\ref{dep-algo}) keeps only words of $S$ that fulfill a given morphosyntactic constraint $\phi$. The following strategies correspond to various definitions of $\phi$.
\subsubsection{Strategy I$_0$}
Our first strategy will be to keep only the head of each sentence (which, incidentally, is a verb in 85.37\% of sentences of our corpus).
This corresponds to the constraint $\phi(w)\equiv(w=\mathrm{root}(S))$. Results are given on Table~\ref{tab:dependency_head_sentence}.
Although the recall of GSPO is quite high (a possible interpretation could be that sports use very specific verbs), F-measure is quite low when we compare it to the one of the tfidf-based method of the same average itemset length, namely~65.69\%.
\subsubsection{Strategy I$_1$} The second strategy consists in keeping words connected to the head by a (single) dependency of type nsubj (=~nominal subject). This
occurs in 79.84\% of sentences of our corpus. The constraint is then $\phi(w)\equiv(\exists(w,\mathrm{root}(S),$ $\mathrm{nsubj})\in \mathrm{Dep}(S))$. Results are given on Table~\ref{tab:dependency_nominal_subject}.
Note that the slightly higher than~1 transaction size is probably due to the rare cases where there are more than one nsubj dependencies pointing to the head. The scores rise dramatically when compared to those of the strategy based only on the head of the sentence. The average F-measure (71.80\%) is significantly higher than the tfidf-based performance for the same average transaction size (65.69\%). \emph{This shows that using a dependency property to select a word is a better choice than the one provided by the frequentist tfidf-based method.} Note that the case of nsubj is unique: if we take ccomp (=~clausal complement) instead of nsubj, the performance falls even below the level of strategy I$_0$ (Table~\ref{tab:dependency_clausal_complement}).
\subsubsection{Strategy I$_2$}
The third strategy considers all nouns (POS tags starting with \texttt{N}) at distance~1 from the head in the dependency graph. Such dependencies occur in 59.24\% of the sentences of our corpus. This corresponds to $\phi(x)\equiv((\exists(x,\mathrm{root}(S),d)\in \mathrm{Dep}(S))\wedge(\mathrm{POS}(x)=\mathtt{N*}))$. Results are given on Table~\ref{tab:dependency_pos_tags}.
The result seems better than the one of strategy I$_1$ (Table~\ref{tab:dependency_nominal_subject}). However, if we take transaction size into account, it is in fact merely equivalent to---and hence not better than, as it was the case for I$_1$---the tfidf-based method with the same transaction size. Once again we see a very high recall rate for the sports category.
One could be tempted to check the performance of taking verbs (instead of nouns) at distance~1 from the head. Indeed, verbs at that position are more frequent than nouns: they occur in 62.94\% of the sentences of our corpus. Nevertheless, the results are not as good (Table~\ref{tab:dependency_verbs_distance1_head}). This shows that despite their high frequency, verbs contain less pertinent information than nouns at the same distance from the head.
\subsection{Methods based on dependencies and hyperonyms}\label{hyper-sec}
\begin{algorithm}[tb]
\SetKwFunction{classify}{Classify}
\SetKwFunction{evaluate}{Evaluate}
\SetKwFunction{prune}{Prune}
\SetKwFunction{hyper}{Hyperonymize}
\SetKwFunction{lemma}{Lemmatize}
\SetKwFunction{apriori}{Apriori}
\SetKwFunction{partition}{Partition}
\SetKwFunction{shuffle}{Shuffle}
\SetKwFunction{tfidf}{Tfidf}
\SetKwFunction{sort}{Sort}
\SetKw{KwTo}{:=}
\SetKwProg{Fn}{}{:}{end}
\KwData{A dependency-pruned corpus $\mathbf{S}'$, an hyperonymic function $\mathrm{MSCH}\colon \mW\to \mW^{\mathbb{N}}$, the hyperonymic order $N$}
\KwResult{The hyperonymically extended corpus $\mathbf{S}''$}
\Fn(){\hyper{$\mathbf{S}'$, $\mathrm{MSCH}$, $N$}}{
$\mathbf{S}''$ \KwTo $\emptyset$\;
\For{$S'\in \mathbf{S}'$}{
$S''$ \KwTo $\emptyset$\;
\For{$w\in S'$}{
\eIf{$\mathrm{proj}_N(\mathrm{MSCH}(w))\ne\emptyset$}{$S''$ \KwTo $S''\cup\{\mathrm{proj}_N(\mathrm{MSCH}(w))\}$}{$S''$ \KwTo $S''\cup\{w\}$}
}
$\mathbf{S}''$ \KwTo $\mathbf{S}''\cup\{S''\}$
}
}
\caption{Corpus hyperonymization}\label{hyper-algo}
\end{algorithm}
In this section we add semantic information by the means of hyperonyms, using the hyperonymization function $h$ (\S\,\ref{hc}).
The preprocessing is done by Alg.~\ref{hyper-algo}: $h_i(w)$ is an $N$-th order hyperonym of $w$, if it exists in WordNet. In case there is no $N$-th order hyperonym, the word remains unchanged. We call $N$ the \emph{hyperonymic factor} of our itemset transformation.
\subsubsection{Strategy II$_1$} This strategy considers hyperonymic factor $N=1$. We thus first apply strategy I$_1$ and then hyperonymization $h_1$. Results are presented on Table~\ref{tab:hyperonymic_factor_N1}.
\begin{subtables}
\begin{table}[t]
\caption{Strategy II$_1$: I$_1$ followed by first-order hyperonymization}
\kern-5pt
\noindent\begin{tabularx}{\textwidth}{l|CCCCCCC|C}
&E12&GCRIM&GDIP&GJOB&GPOL&GSPO&GVIO&AVG\\\hline
Recall&72.39&56.04&71.32&48.96&59.02&82.20&70.42&65.76\\
Precision&66.21&75.42&64.33&82.21&67.71&73.20&70.73&71.40\\
F-measure&69.16&64.30&67.64&61.37&63.07&77.44&70.57&67.65\\\hline
\multicolumn{9}{l}{\footnotesize MinSupp=0.010, MinConf=44.8, Var.=1.92, Disp.=78.47, AvgTransSize=1.04}
\end{tabularx}
\label{tab:hyperonymic_factor_N1}
\medskip
\caption{Strategy II$_2$: I$_1$ followed by second-order hyperonymization}
\kern-5pt
\noindent\begin{tabularx}{\textwidth}{l|CCCCCCC|C}
&E12&GCRIM&GDIP&GJOB&GPOL&GSPO&GVIO&AVG\\\hline
Recall&69.02&52.78&71.97&47.77&54.24&80.80&65.67&63.18\\
Precision&64.94&74.57&61.23&80.64&65.81&69.96&72.33&69.93\\
F-measure&66.92&61.81&66.16&60.00&59.47&74.99&68.84&65.46\\\hline
\multicolumn{9}{l}{\footnotesize MinSupp=0.008, MinConf=44.8, Var.=1.85, Disp.=70.51, AvgTransSize=1.04}
\end{tabularx}
\label{tab:hyperonymic_factor_N2}
\end{table}
\end{subtables}
The performance is globally inferior to the one of Strategy I$_1$ (in which, F-measure attained 71.80\%). It is interesting to note that the recall of class GJOB has decreased significantly (48.96\% vs. 58.27\%): in other words, using hyperonyms when dealing with labor issues results into failure to recognize 9.31\% of the documents as belonging to the domain; one could say that terms used in GJOB lose their “labor specificity” already at first-order hyperonymization. On the other hand, the (already high in I$_1$) recall of GSPO has increased even more, compared to I$_1$ (from 78.76\% to 82.20\%): it seems that sports terminology remains in the domain even after hyperonymization, and replacing specific terms by more general ones has increased their frequency as items, and hence improved recall. We have the same phenomenon with the recall of GDIP (which increased from 66.97\% to 71.32\%), and also slightly with the recalls of E12 and GVIO.
\subsubsection{Strategy II$_2$} This strategy is similar to strategy II$_1$ but uses hyperonymic factor $N=2$. Results are presented on Table~\ref{tab:hyperonymic_factor_N2}.
The performance is globally inferior to the one of II$_1$ (where we used first-order hyperonyms), with two minor exceptions: the recall of GDIP that increased by 0.65\% and the precision of GVIO that increased by 1.6\%. What is noteworthy however, is the fact that the recalls of GDIP and GSPO are still higher than the ones of strategy I$_1$ (no hyperonyms).
To better understand the behavior of the system when climbing the hyperonymic chain by replacing words by hyperonyms of increasingly higher order (and returning to the original word when there are no hyperonyms left) we calculated the performance for $N$-th order hyperonyms for $1\leq N\leq12$. Note that when $N>12$ the amount of remaining hyperonyms is negligible and the strategy is similar to strategy I$_1$ (no hyperonyms). On Fig.~\ref{hypers-fig}, the reader can see the evolution of recall (black), precision (red) and F-measure (blue) for the average of all class, and then specifically for GSPO and for GDIP. Dashed lines represent the recall, precision and F-measure of strategy I$_1$.
\begin{figure}[ht]
\centering
\resizebox{\textwidth}{!}{\includegraphics{hypers1.pdf}~\includegraphics{hypers2.pdf}~\includegraphics{hypers3.pdf}}
\caption{F-1 measure for hyperonymization of orders $1\leq N\leq12$: the average case, class GSPO, class GDIP\label{hypers-fig}}
\end{figure}
In the average case, the effect of hyperonymization of orders 1--4 is to decrease performance. After $N=5$, the global number of hyperonyms available in WordNet rapidly decreases so that the situation gradually returns to the one of I$_1$ (no hyperonyms) and we see curves asymptotically converging to I$_1$ lines from underneath.
Not so for GSPO, the GSPO recall curve of which is above the I$_1$ value for most $N$ ($N=1$, 2, 6--8 and 10--12).
The phenomenon is even better illustrated in the case of GDIP: as the reader can see on the figure, the \emph{complete GDIP recall curve is located above the I$_1$ one}.
It seems that in these two cases (GDIP and, to a lesser extent, GSPO), hyperonyms of all orders have a positive impact on the classifier. Unfortunately this impact only concerns recall and is compensated by bad precision, so that F-measure is still inferior to the I$_1$ case.
\section{Conclusion and future work}\label{con}
In this paper we have investigated the use of association rules for text classification by applying two new techniques: (a) we reduce the number of word features through the use of morphosyntactic criteria in the framework of dependency syntax; for that we keep words dependent from the head by specific dependencies and/or having specific POS tags (b) we replace words by their hyperonyms of different orders, which we have calculated out of WordNet using frequencies and, in some cases, disambiguation. We have obtained positive results for case (a), in particular when we compare dependency-based single-item rules with tfidf-based ones. In case (b) the results we share in this paper are less efficient but still interesting, especially we found classes for which hyperonymization significantly improves recall.
This work opens several perspectives, among which:
--- examine why these particular classes are favorable to hyperonymization, whether this is related to the structure of WordNet or to linguistic properties of the domain;
--- explore \emph{partial hyperonymization} i.e., is it possible to hyperonymize only specific items according to the needs of the classifier?\footnote{Indeed, by hyperonymizing all words one wins on one side and loses on the other: for example “dalmatian” and “poodle” will both be replaced by “dog”, but “dog” occurrences will be replaced by “canid”. It would be more preferable to keep the word “dog” in the second case, so that we have a real increase in frequency.} How do we choose, on the word level, if we should rather keep the original word (to increase precision) or switch to some hyperonym (to increase recall)?
--- we have used only recall and precision as quality measures of our rules, and our evaluation is strongly dependent on these measures since the selection of the 1,000 rules we keep is entirely based upon them. There are other quality measures available, how do they apply and how can they be compared and combined? How robust are the results?
--- and finally: how can we optimize the distinctive feature of association rules, namely the fact of being intelligible by the user? How can the user's experience (and linguistic knowledge) be incorporated in the enhancement of rules to obtain the best possible result from his/her point of view?
\bibliographystyle{splncs03}
|
\section*{Introduction}
Let $M^n$ be a complex analytic manifold and $\Lambda_{\tau}$ a family of
subvarieties in $M^n$ depending analytically on the parameter $\tau$. For
almost all values of $\tau$ the corresponding pairs $(M^n,\Lambda_\tau)$ are
homeomorphic to one another, and the fundamental group of the set of such
generic $\tau$ acts on different homology groups related to such pairs. This
action is responsible for the ramification and qualitative analytic features of
all known functions given by integral representations (such as Fourier and
Radon transforms, fundamental solutions of hyperbolic and parabolic equations,
Newton--Coulomb potentials, hypergeometric functions, Feynman integrals etc.).
Explicit formulae expressing this action are called (generalized)
Picard--Lefschetz formulae, see e.g. \cite{AVG}, \cite{AVGL}, \cite{Ph1,Ph2},
\cite{giventh}, \cite{V1}.
The classical Picard--Lefschetz formula describes the case when $\Lambda_\tau$
is the family of irreducible hypersurfaces in $M^n$, non-singular for generic
$\tau$ and having a Morse singularity for $\tau$ running over a hypersurface in
the parameter space, the investigated loop is a small circle embracing this
hypersurface, and the homology group in question is $H_{n-1}(\Lambda_\tau)$.
Similar formulae for many other possible degenerations of the variety
$\Lambda_\tau$ and other homology groups were studied in \cite{Ph1} (see also
\cite{giventh}) and in \cite{V2}. In the present paper we consider the most
general situation: as in \cite{V1}, \cite{V2} we investigate the {\it
stratified Morse singularities} of $\Lambda_\tau$ (which include all
degenerations of \cite{Ph1} as special cases), and as in \cite{Ph1}, \cite{Ph2}
we consider the homology groups of $M^n \setminus \Lambda_\tau$ (with twisted
coefficients, and with both closed and compact supports), to which the study of
almost all other homology groups related with the pair $(M^n, \Lambda_\tau)$
can be reduced (except may be for the intersection homology of $\Lambda_\tau$
studied in the last part of \cite{V2}).
\medskip
The Arnold's ``complexification'' functor (see e.g. \cite{A}), establishing an
informal analogy between objects of ``real'' and ``complex'' worlds, maps the
Morse theory into the Picard--Lefschetz theory; certainly, the analogue of the
stratified Morse theory of \cite{GM} should be a stratified Picard--Lefschetz
theory. On the other hand, recently it became clear that instead of the
homology groups of (sub)varieties, usually considered in all these theories, in
the ``complex'' situation it is natural to deal with the homology, generally
with coefficients in non-constant local systems, of their complements, cf.
\cite{Ph2}, \cite{gvz}, \cite{giventh}. Thus the matter of our paper seems to
be the most adequate ``complexification'' of that of \cite{GM}.
\medskip
{\bf Agreements.} In what follows we assume that $M^n = \C^n$ (which is not
restricting because all our considerations are local), and families
$\Lambda_\tau$ (which appear later as $A \cup X_t$) depend on one parameter. We
consider only homology groups reduced modulo a point in the case of absolute
homology and modulo the fundamental cycle in the case of relative homology of a
complex analytic variety. We often use a short notation of type $H_*(X,Y)$
instead of a more rigorous $H_*(X, Y \cap X)$ or $H_*(X \cup Y, Y)$. The sign
$\Box$ denotes the end or absence of a proof.
\section{Main characters and stating the problem}
Let $A$ be a complex analytic subvariety in $\C^n$ with a fixed analytic
Whitney stratification (see e.g. \cite{GM}, \cite{V1}), let $\sigma \subset A$
be a stratum of dimension $k$, $a$ a point of $\sigma$, and $B \equiv
B_\varepsilon$ a small closed disc in $\C^n$ centred at $a$, so that $B$ has
nonempty intersections only with the strata adjoining $\sigma$ (or $\sigma$
itself) and $\partial B$ is transversal to the stratification. In particular
(see \cite{Che}, \cite{GM}) the induced partition of the pair $(B, \partial B)$
is again a Whitney stratification of $B$.
Let $f: (B,a) \to (\C,0)$ be a holomorphic function such that $df \ne 0$ in
$B$.
\medskip
{\bf Definition 1} (cf. \cite{GM}). The function $f$ has a {\it Morse
singularity} on $A$ at the point $a$ if its restriction on the manifold
$\sigma$ has a Morse singularity at $a$, and for any stratum $\tau \ne \sigma$
and any sequence of points $b_i \in \tau$ converging to $a$ and such that the
planes tangent to $\tau$ at $b_i$ (considered as the points in the associated
Grassmann bundle $G^{\dim \tau}(T_* \C^n)$) converge to a plane in $T_a \C^n$,
this limit plane does not lie in the hyperplane $\{df|_a = 0\} \subset
T_a\C^n$. If this condition is satisfied, we say that the one-parametric family
of varieties $A \cup f^{-1}(t),$ $t \in \C,$ has a {\it stratified Morse
singularity} at $a$.
\medskip
{\bf Notation.} For any $t \in \C$ denote by $X_t$ the set $f^{-1}(t)$ and by
$AX_t$ the set $ A\cup X_t$.
\medskip
\begin{figure}
\unitlength 1.00mm \linethickness{0.4pt}
\begin{center}
\begin{picture}(126.00,74.00)
\bezier{180}(40.00,41.00)(40.00,21.00)(56.00,9.00)
\bezier{144}(40.00,41.00)(40.00,26.00)(25.00,12.00)
\bezier{180}(110.00,41.00)(110.00,21.00)(126.00,9.00)
\bezier{180}(56.00,9.00)(71.00,24.00)(91.00,24.00)
\bezier{180}(126.00,9.00)(111.00,24.00)(91.00,24.00)
\bezier{60}(95.00,12.00)(101.67,17.33)(104.33,22.67)
\bezier{20}(105.67,25.00)(107.00,27.33)(107.67,29.67)
\bezier{20}(108.67,31.67)(109.33,33.67)(109.67,36.33)
\bezier{80}(95.00,12.00)(86.33,20.67)(79.00,23.00)
\bezier{36}(76.00,24.33)(71.67,26.33)(68.00,26.67)
\bezier{20}(65.33,26.80)(61.00,27.20)(57.00,27.00)
\bezier{28}(54.33,26.67)(50.33,26.33)(47.33,25.33)
\bezier{92}(44.00,24.33)(34.33,21.33)(25.00,12.00)
\put(75.00,56.00){\circle*{1.33}} \put(76.00,59.00){\makebox(0,0)[cc]{$a$}}
\put(91.00,56.00){\makebox(0,0)[cc]{$\sigma$}}
\put(119.00,20.00){\makebox(0,0)[cc]{$A$}}
\put(35.00,42.00){\line(1,0){79.00}}
\multiput(114.00,42.00)(0.12,0.48){42}{\line(0,1){0.48}}
\put(119.00,62.00){\line(-1,0){79.00}}
\multiput(40.00,62.00)(-0.12,-0.48){42}{\line(0,-1){0.48}}
\bezier{36}(86.67,48.67)(82.67,46.00)(78.00,46.00)
\bezier{40}(86.67,48.67)(91.00,50.33)(96.00,51.00)
\bezier{48}(78.00,46.00)(73.00,46.00)(67.00,48.67)
\bezier{52}(67.00,48.67)(61.67,50.67)(54.00,51.00)
\bezier{88}(96.00,51.00)(86.00,56.00)(75.00,56.00)
\bezier{88}(75.00,56.00)(64.00,56.00)(54.00,51.00)
\bezier{16}(51.25,50.00)(48.33,48.67)(46.67,47.33)
\bezier{12}(45.00,46.00)(43.50,45.00)(41.33,43.00)
\bezier{16}(108.67,43.00)(106.50,45.00)(105.00,46.00)
\bezier{20}(103.33,47.33)(101.67,48.67)(98.75,50.00)
\multiput(40.00,41.00)(0.11,0.17){6}{\line(0,1){0.17}}
\multiput(110.00,41.00)(-0.11,0.17){6}{\line(0,1){0.17}}
\bezier{24}(63.80,52.00)(61.80,51.50)(57.80,51.25)
\bezier{20}(73.00,54.00)(70.50,53.80)(67.00,53.00)
\bezier{20}(76.00,53.90)(78.67,53.70)(80.67,53.10)
\bezier{24}(83.33,52.30)(85.33,51.70)(88.50,51.43)
\bezier{12}(90.67,51.25)(91.80,51.15)(96.00,51.05)
\bezier{116}(12.00,56.00)(11.00,39.33)(0.00,34.00)
\bezier{116}(12.00,56.00)(13.00,39.33)(24.00,34.00)
\put(1.00,45.00){\line(1,0){21.00}}
\put(12.00,1.00){\makebox(0,0)[cc]{{\large a}}}
\put(75.00,1.00){\makebox(0,0)[cc]{{\large b}}}
\put(31.00,54.00){\vector(0,1){20.00}}
\put(28.00,69.00){\makebox(0,0)[cc]{$f$}}
\put(120.00,52.00){\makebox(0,0)[cc]{$X$}}
\end{picture}
\caption{A Morse function on a stratified variety}
\end{center}
\end{figure}
For example, in Fig.~1b (the complexification of) the linear function defining
the plane $X$ has a stratified Morse singularity on (the complexification of)
$A$.
\medskip
Let $L_{\al}$ be the linear (i.e. with fibre $\C^1$) local system on $\C^n
\setminus AX_t$ with the set $\al = (\al_1, \ldots, \al_\nu)$ of ramification
indices: the cardinality $\nu$ of this set is equal to the number of
irreducible $(n-1)$-dimensional components of $AX_t$, and any small loop going
around a smooth piece of the $i$-th component in the positive direction (with
respect to the natural complex structure in the normal bundle) acts on the
fibre as multiplication by $\al_i$. We assume that $X_t$ is the first component
of $AX_t,$ so that $\al_1$ is responsible for the rotations around $X_t$.
The dual local system will be denoted by $L_{\al^*},$ $\al^* = (\al_1^{-1},
\ldots, \al_\nu^{-1})$. For all $t \in \C \setminus 0$ sufficiently close to
$0$ (say, satisfying the condition $|t| \le \delta$ with sufficiently small
$\delta$) all corresponding sets $\C^n \setminus AX_t$ are homeomorphic to one
another and form a locally trivial bundle over the set of such $t$.
Denote by $C$ the loop $\{\delta e^{i\tau} \}, \ \tau \in [0,2\pi],$ generating
the group $\pi_1(\C^1 \setminus 0)$; the inverse loop $\{\delta e^{-i\tau}\}$
is denoted by $C^*$, and the closed disc bounded by any of these loops by
$D_{\delta}$.
The monodromy action of the loop $C$ (and $C^*$ as well) in two groups
$$ H_*(\C^n \setminus AX_t, \La) \quad and \quad
H^{lf}_*(\C^n \setminus AX_t, \La)$$ is well defined, where the letters $lf$ in
the last expression denote the homology groups of locally finite chains. The
study of this action is our main aim.
\medskip
These monodromy operators can be localized in the standard way, see e.g.
\cite{Ph2}, \cite{V1}. We suppose that $a$ is the unique point of
non-transversality of the manifold $X_0$ and the stratified variety $A$, and
the number $\delta$ participating in the definition of the loop $C$ is so small
that for any $t \in D_{\delta} \setminus 0$ the manifold $X_t$ is smooth and
transversal to the naturally stratified variety $A \cup \partial B $.
We choose $\{\delta\}$ as the distinguished point in $D_{\delta}\setminus 0$
and redenote $X_{\delta}$ and $AX_{\delta}$ simply by $X$ and $AX$. For any
non-negative integer $i $ we consider four groups
\begin{equation}
\label{groups}
\begin{array}{rclrcl}
\bar {\mathcal H}_{i,\al} & \equiv & H_i^{lf}(B \setminus AX,\partial B;L_\al),
\quad & {\mathcal H}_{i,\al} & \equiv & H_i^{lf}(B \setminus AX,L_\al), \\
\bar \chi_{i,\al} & \equiv & H_i(B \setminus AX,\partial B;L_\al), \quad &
\chi_{i,\al} & \equiv & H_i(B \setminus AX,L_\al).
\end{array}
\end{equation}
There are obvious homomorphisms
\begin{equation}
\label{triple}
\begin{array}{rclrcl}
\tilde i_{\al}: {\mathcal H}_{*,\al} & \to & H^{lf}_*(\C^n \setminus AX, \La),
& \quad \tilde j_{\al}: H^{lf}_*(\C^n \setminus AX, \La) & \to & \bar {\mathcal H}_{*,\al}, \\
i_{\al}: \chi_{*,\al} & \to & H_*(\C^n \setminus AX, \La), & \quad j_{\al}:
H_*(\C^n \setminus AX, \La) & \to & \bar \chi_{*,\al}.
\end{array}
\end{equation}
Also, any loop $\lambda \in \pi_1(D_\delta \setminus 0)$ defines in the
standard way (see \cite{AVGL}, \cite{V1}) the {\em local variation operators}
\begin{equation}
\label{varrs} \widetilde{Var}_{(\lambda)}: \bar {\mathcal H}_{*,\al} \to
{\mathcal H}_{*,\al} \, , \quad Var_{(\lambda)}: \bar \chi_{*,\al} \to
\chi_{*,\al} \ ,
\end{equation}
so that the monodromy action of the loop $\lambda$ in the group $
H^{lf}_*(\C^n \setminus AX, \La) $ (respectively, $ H_*(\C^n \setminus AX,
\La)$) is equal to $\hbox{Id} + \tilde i_{\al}\circ \widetilde{Var}_{(\lambda)}
\circ \tilde j_{\al}$ (respectively, $\hbox{Id} + i_{\al}\circ Var_{(\lambda)}
\circ j_{\al}). $
\medskip
The composition operators $\tilde J_\alpha \equiv \tilde j_\alpha \circ \tilde
i_\alpha : {\mathcal H}_{*,\al} \to \bar {\mathcal H}_{*,\al}$ and $J_\alpha
\equiv j_\alpha \circ i_\alpha : \chi_{*,\al} \to \bar \chi_{*,\al}$ allow us
to express the {\it local monodromy} action of the loop $\lambda$ on four
groups $\bar {\mathcal H}_{*,\alpha}, \bar \chi_{*,\alpha}, {\mathcal
H}_{*,\alpha}, \chi_{*,\alpha}$ as
\begin{equation}
\label{mons}
\begin{array}{rr}
Id+ \tilde J_\alpha \circ \widetilde{Var}_{(\lambda)}, \quad &
Id+ J_\alpha \circ Var_{(\lambda)}, \\
Id+ \widetilde{Var}_{(\lambda)} \circ \tilde J_\alpha, \quad & Id+
Var_{(\lambda)} \circ J_\alpha
\end{array}
\end{equation}
respectively.
The groups (\ref{groups}) are related by non-degenerate Poincar\'e--Lefschetz
pairings
\begin{equation}
\label{poinc}
\begin{array}{rcl}
\bar {\mathcal H}_{i,\al} \otimes \chi_{2n-i,\al^*} & \to & \C , \\
{\mathcal H}_{i,\al} \otimes \bar \chi_{2n-i,\al^*} & \to & \C.
\end{array}
\end{equation}
{\bf Proposition 1.} {\it For any $i$ and $\alpha$, the operators
\begin{equation}
\label{vars}
\begin{array}{rccl}
\widetilde{Var}_{(C)}: & \bar {\mathcal H}_{i,\al} & \to & {\mathcal H}_{i,\al} \ , \\
Var_{(C^*)}: & \bar \chi_{2n-i,\al^*} & \to & \chi_{2n-i,\al^*}
\end{array}
\end{equation}
are conjugate with respect to the pairings (\ref{poinc}), i.e., for any
elements $ x \in \bar {\mathcal H}_{i,\al} $ and $ y \in \bar \chi_{2n-i,\al^*}
$ we have}
$$ \langle x, Var_{(C^*)}y \rangle =
\langle \widetilde{Var}_{(C)} x, y \rangle.$$
Proof (cf. \cite{sabir}). Let $x$ and $y$ be some elements of the groups $ \bar
{\mathcal H}_{i,\al} $ and $ \bar \chi_{2n-i,\al^*} $ respectively, and ${\bf
x}$ and ${\bf y}$ some relative cycles representing them and intersecting
generically in $B$ (in particular having no common points in $\partial B$).
Then the intersection number $\langle{\bf x},{\bf y}\rangle$ is well defined
(although it is not an invariant of the homology classes $x$ and $y$). Let
$h_C{\bf x}$ and $h_{C^*}{\bf y}$ be two relative cycles in $B \setminus AX$
obtained from ${\bf x}$ and ${\bf y}$ by the monodromy over the loops $C$ and
$C^*$ as in the construction of variation operators (i.e., fixed close to the
boundary of $B$). Then $\langle\widetilde{Var}_{(C)}(x),y\rangle - \langle x,
Var_{(C^*)}(y)\rangle = \langle h_C {\bf x} - {\bf x},{\bf y}\rangle - \langle
{\bf x},h_{C^*} {\bf y} - {\bf y} \rangle = \langle h_C{\bf x}, {\bf y} \rangle
- \langle {\bf x}, h_{C^*} {\bf y} \rangle = \langle h_C{\bf x}, {\bf y}
\rangle - \langle h_{C^*}h_C{\bf x}, h_{C^*} {\bf y} \rangle =0$ because these
intersection pairings are invariant under the monodromy action. \quad $\Box$
\medskip
Set $k=\dim \sigma$, $m=n-k,$ and let $T$ be the $m$-dimensional complex plane
through $a$ transversal to $\sigma$. Then, replacing the sets $B, A$ and $X_t$
by their intersections with $T,$ we get all structures as above (homology
groups, local variation and monodromy action, intersection indices and maps
$\tilde J_\al,$ $J_\al$) placed in the space of reduced dimension $m$. In the
next section we show how, knowing all these objects for this reduced space, we
can reconstruct them for the initial objects in entire space $\C^n.$ This
reconstruction will de done by induction over a flag of planes connecting $T$
and $\C^n.$
\medskip
\section{Main results}
Consider the flag of complex planes (or, more generally, smooth complex
submanifolds)
\begin{equation} \label{flag}
T \equiv T^m \subset T^{m+1} \subset \cdots \subset T^n \equiv \C^n,
\end{equation}
of dimensions $m, m+1, \ldots, n$ respectively, all of which intersect $\sigma$
transversally at the point $a$. Then the intersections with the strata of $A$
define a Whitney stratification on any of varieties $A \cap B \cap T^r$. If the
flag (\ref{flag}) is generic, then for any $r = m, \ldots, n$ the restriction
of $f$ on $A \cap T^r$ is a Morse function in $B$; we shall suppose that this
condition is satisfied for all $r$. Any loop $\lambda$ in $D_\delta \setminus
0$, considered as family of planes $X_t \cap T^r$, $t \in \lambda$, defines
then the variation operators
\begin{equation}\label{varr}
\widetilde{\hbox{Var}}_{(\lambda),r}: H_i^{lf}(B \cap T^r \setminus AX,\partial
B;L_\al) \to H_i^{lf}(B \cap T^r \setminus AX,L_\al),
\end{equation}
\begin{equation}\label{varrr}
\hbox{Var}_{(\lambda),r}: H_i(B \cap T^r \setminus AX,\partial B;L_\al) \to
H_i(B \cap T^r \setminus AX,L_\al),
\end{equation}
in particular $\widetilde{\hbox{Var}}_{(\lambda),n} =
\widetilde{\hbox{Var}}_{(\lambda)}$, $\hbox{Var}_{(\lambda),n} =
\hbox{Var}_{(\lambda)}$.
Everywhere below we consider only operators (\ref{varr}), (\ref{varrr}) defined
by the loop $\lambda = C$ and denote these operators simply by $\hbox{Var}_r$.
Denote the groups participating in these operators as follows:
\begin{equation}
\label{grflag}
\begin{array}{ll}
\bar {\mathcal H}_{i,\al}(r) \equiv H_i^{lf}(B \cap T^r \setminus AX,\partial
B;L_\al), &
{\mathcal H}_{i,\al}(r) \equiv H_i^{lf}(B \cap T^r \setminus AX,L_\al), \\
\bar \chi_{i,\al}(r) \equiv H_i(B \cap T^r \setminus AX,\partial B;L_\al), &
\chi_{i,\al}(r) \equiv H_i(B \cap T^r \setminus AX,L_\al),
\end{array}
\end{equation}
and local monodromy operators defined by the loop $C$ on four groups $\bar
{\mathcal H}_{*,\alpha}(r),$ ${\mathcal H}_{*,\alpha}(r), $ $\bar
\chi_{*,\alpha}(r), $ $\chi_{*,\alpha}(r)$ by $\bar M_r, M_r, \bar \mu_r,
\mu_r$ respectively.
Of course, the analogues of the Poincar\'e duality (\ref{poinc}) and
Proposition 1 are valid for any $r$.
\bigskip
{\bf Theorem 1.} {\it For any $l=0, \ldots, n-m,$ and any $\al,$ there are
almost canonical isomorphisms}
\begin{equation}
\label{stab}
\begin{array}{rccl}
\Sigma^l : &{\mathcal H}_{i,\al}(m) & \to & {\mathcal H}_{i+l,\al}(m+l), \\
\dar^l : & \bar \chi_{i,\al}(m) & \to & \bar \chi_{i+l,\al}(m+l).
\end{array}
\end{equation}
The construction of these maps will be described in the next section. For the
explanation of the word ``almost'' see Remark in \S \ 3.3.
\medskip
The stabilization of groups $\bar {\mathcal H}_{i+l,\al}(m+l)$ and
$\chi_{i+l,\al}(m+l)$ depends very much on the fact, is the coefficient $\al_1$
(responsible for the ramification of $L_{\al}$ close to the plane $X_t$) equal
to 1 or not. To unify the corresponding formulae, set $\Phi_{i,\al}\equiv
\psi_{i,\al} \equiv 0$ if $\al_1 \ne 1;$ if $\al_1=1,$ we consider the local
system $L_{\hat \al}$ on $B \setminus A,$ isomorphic to $L_{\al}$ in a small
closed disc $ b \subset B \setminus X_t$ with center $a$ (i.e. having the same
monodromy coefficients $\al_2, \ldots, \al_\nu$ related to all components of
$A$) and set $\Phi_{i,\al}\equiv H_i^{lf}(B \cap T^m\setminus A, \partial B;
L_{\hat \al}) \simeq H_{i+2l}^{lf}(B \cap T^{m+l}\setminus A, \partial B;
L_{\hat \al}),$ $\psi_{i,\al}\equiv H_i(B \cap T^m\setminus A, L_{\hat \al})
\simeq H_i(B \cap T^{m+l}\setminus A, L_{\hat \al}).$
\bigskip
{\bf Theorem 1$'$.} {\it For any $l=1, \ldots, n-m,$ and any $\al,$ there are
almost canonical isomorphisms}
\begin{equation}
\label{stab2}
\begin{array}{rcl}
\bar {\mathcal H}_{i+l,\al}(m+l) & \simeq & {\mathcal H}_{i,\al}(m)
\oplus \Phi_{i-l,\al} \oplus \Phi_{i-l+1,\al} , \\
\chi_{i+l,\al}(m+l) & \simeq & \bar \chi_{i,\al}(m)
\oplus \psi_{i+l,\al} \oplus \psi_{i+l-1,\al}. \\
\end{array}
\end{equation}
For any $l=1, \ldots, n-m,$ we denote by $\rlh_l$ (respectively, $\infty_l$)
the injection ${\mathcal H}_{i,\al}(m) \to \bar {\mathcal H}_{i+l,\al}(m+l)$
(respectively, $\bar \chi_{i,\al}(m) \to \chi_{i+l,\al}(m+l)$) defined by the
first summands in the right-hand parts of (\ref{stab2}).
\bigskip
{\bf Theorem 2.} {\it For any $l = 1, \ldots, n-m,$
a) the obvious homomorphism $ \tilde J_{\al,m+l} \equiv \tilde j_{\al}\circ
\tilde i_{\al}: {\mathcal H}_{i+l,\al}(m+l) \to \bar {\mathcal
H}_{i+l,\al}(m+l)$ (i.e., the reduction mod $\partial B$) maps any element
$\Sigma^l(x),$ $x \in {\mathcal H}_{i,\al}(m),$ to
\smallskip
$\rlh_{l} \circ \widetilde{Var}_m \circ \tilde J_{\al,m}(x) \equiv \, \rlh_{l}
\circ (M_m - Id)(x)$ \quad if $l$ is even \quad and to
\smallskip
$- \rlh_{l} (2x + \widetilde{Var}_m \circ \tilde J_{\al,m}(x)) \equiv -
\rlh_{l} \circ (M_m + Id)(x)$ \quad if $l$ is odd.
\smallskip
b) the similar homomorphism $ J_{\al,m+l} \equiv j_{\al}\circ i_{\al}:
\chi_{i+l,\al}(m+l) \to \bar \chi_{i+l,\al}(m+l)$ is equal to zero on the last
two summands in (\ref{stab2}) and maps any element $\infty_l(y)$, $y \in \bar
\chi_{i,\al}(m),$ to
\smallskip
$\dar^l \circ J_{\al,m} \circ Var_m(y) \equiv \, \dar^l \circ (\bar
\mu_m-Id)(y) $ \quad if $l$ is even \quad and to
\smallskip
$-\dar^l (2y + J_{\al,m}\circ Var_m(y)) \equiv - \dar^l \circ (\bar \mu_m +
Id)(y) $ \quad if $l$ is odd.}
\bigskip
{\bf Theorem 3.} {\it For any $\al$ and any $l = 1, \ldots, n-m,$
a) the operator $\widetilde{Var}_{m+l}: \bar {\mathcal H}_{i+l,\al}(m+l) \to
{\mathcal H}_{i+l,\al}(m+l)$ is equal to zero on two last summands in the first
row of (\ref{stab2}) and maps the element $\rlh_l(x)$, $x \in {\mathcal
H}_{i,\al}(m),$ of the first summand to $\Sigma^l(x)$;
b) the operator $Var_{m+l}: \bar \chi_{i+l,\al}(m+l) \to \chi_{i+l,\al}(m+l)$
maps any element $\dar^l(y),$ $y \in \bar \chi_{i,\al}(m),$ to $\infty_l(y)$.}
\bigskip
{\bf Corollary.} {\it The local monodromy operators $M_{m+l}, \bar M_{m+l},
\bar \mu_{m+l}$ and $\mu_{m+l}$ respectively map the elements $\Sigma^l(x),
\rlh_l(x), \dar^l(y)$ and $\infty_l(y)$ (where $x \in {\mathcal
H}_{*,\alpha}(m),$ $ y \in \bar \chi_{*,\alpha}(m)$) into the elements $(-1)^l
\Sigma^l(M_m(x))$, $(-1)^l \rlh_l(M_m(x))$, $(-1)^l \dar^l(\bar \mu_m(y))$ and
$(-1)^l \infty_l(\bar \mu_m(y))$ respectively. Operators $\bar M_{m+l}$ and
$\mu_{m+l}$ act trivially (i.e. as the identity operators) on the last two
summands in both formulae (\ref{stab2}).}
\medskip
{\bf Theorem 4.} {\it For any $l=1, \ldots, n-m,$
\noindent A) for any $x \in {\mathcal H}_{i,\al}(m)$ and $y^* \in \bar
\chi_{i,\al^*}(m),$
$$\langle \Sigma^l(x), \dar^l(y^*) \rangle =
(-1)^{l(i+1+(l-1)/2)}\langle x,y^* \rangle ; $$
\noindent B) for any $\xi \in \bar {\mathcal H}_{i+l,\al}(m+l) $ and $ \zeta^*
\in \chi_{2m-i+l,\al^*}(m+l)$,
\begin{enumerate}
\item
$\langle \xi,\zeta^* \rangle =0$ if $\xi$ and $\zeta^*$ belong to summands in
the right-hand part of (\ref{stab2}) placed not one over the other;
\item If $\xi = \rlh_l(x),$
$x \in {\mathcal H}_{i,\al}(m)$, and $\zeta^* = \infty_l(y^*)$, $ y^* \in \bar
\chi_{i,\al^*}(m)$, then $\langle \xi,\zeta^* \rangle= (-1)^{il+1+l(l-1)/2}
\langle x, \bar \mu_m(y^*) \rangle .$
\end{enumerate}
}
\medskip
{\bf Corollary} (periodicity theorem). {\it Suppose that the groups
$\Phi_{*,\al},$ $\psi_{*,\al}$ are trivial (e.g. $\al_1 \ne 1$). Then all maps
$\rlh_l,$ $\infty_l$ are isomorphisms, and the entire structure depending on
$A,$ $\al$ and $r$ and consisting of four graded groups $\bar {\mathcal
H}_{*,\al}(r),$ ${\mathcal H}_{*,\al}(r),$ $\bar \chi_{*,\al}(r),$
$\chi_{*,\al}(r),$ similar four groups with $\al$ replaced by $\al^*,$ all
possible intersection pairings between them, and the operators $\tilde J_\al,$
$J_\al,$ $\widetilde{Var}_r,$ $Var_r$, $M_{r}, \bar M_{r}, \mu_{r}$ and $\bar
\mu_{r}$ is periodic in $r$ with period 2.}
\bigskip
The ideas of proofs of some of these results are essentially contained in
\cite{V2}, however the {\it answers} are sometimes different and should have
been written explicitly: this it the main purpose of the present article.
\medskip
We shall often use the following well-known fact (see e.g. \cite{GM},
\cite{M}).
\medskip
{\bf Proposition 2.} {\it For any point $a$ of an analytic subset $A \subset
\C^n$ there exists a small disc $B$ centred at $a$ such that the pair $(A \cap
B, A \cap \partial B)$ is homeomorphic to the cone over $A \cap \partial B$,
and this homeomorphism is identical on $A \cap \partial B$ and maps $a$ into
the vertex of the cone. Moreover, the same is true for any smaller concentric
disc.} $\quad \Box$
\section{Realization of formulae (\protect\ref{stab}), (\protect\ref{stab2})}
The stabilization of groups (\ref{grflag}) will be constructed by induction
over the flag of planes $T^{m+l}$. Namely, for any $r = m, \ldots ,n-1$ we
construct the maps
\begin{equation}
\label{stab3}
\begin{array}{rccl}
\Sigma : & {\mathcal H}_{j,\al}(r) & \to & {\mathcal H}_{j+1,\al}(r+1), \\
\dar : & \bar \chi_{j,\al}(r) & \to & \bar \chi_{j+1,\al}(r+1), \\
\rlh : & {\mathcal H}_{j,\al}(r) & \to & \bar {\mathcal H}_{j+1,\al}(r+1), \\
\infty : & \bar \chi_{j,\al}(r) & \to & \chi_{j+1,\al}(r+1).
\end{array}
\end{equation}
Two first of them (and also two last if $\Phi_{*,\alpha} = \psi_{*,\alpha}=0$)
are isomorphisms.
\medskip
Here are some preliminary constructions and reductions, cf. \cite{V2}.
\subsection{Adapted coordinates and polydisk $B'$}
\noindent Let $A, a, \sigma, B, f, D_\delta$ and $X_t$ be the same as in the
previous sections. \smallskip
{\bf Definition.} A local analytic coordinate system $\{z_1, \ldots, z_n\}$ in
$\C^n$ with origin at $a$ is called {\it adapted} if $z_n \equiv f$, the
tangent space to $\sigma$ at $a$ is spanned by the vectors $\partial/\partial
z_1, \ldots, \partial/\partial z_k$ (so that the restrictions of the functions
$z_1, \ldots, z_k$ constitute a local coordinate system on $\sigma$), and in
restriction to $\sigma$
$$z_n \equiv z_1^2 + \cdots + z_k^2 . $$
In Fig. 1b a real version of this situation is shown, where $n=3$, the plane $X
\equiv X_\delta$ is given by the equation $z_3 \equiv \delta$, and $z_1$ is the
coordinate along the stratum $\sigma$. The transversal slice of this picture by
the plane $\{z_1 = 0\}$ is shown in Fig. 1a. \smallskip
By the Morse lemma, adapted coordinates always exist; let us fix such a
coordinate system. Without loss of generality we can define the flag
(\ref{flag}) by the conditions $T^r = \{z \mid \nolinebreak z_1 = \cdots =
z_{n-r} =0\}$.
We can assume that $B \equiv B_\varepsilon$ is a closed disc of radius
$\varepsilon$ with respect to the standard Hermitian metric defined by these
coordinates $z_1, \ldots, z_n$. Moreover, in our considerations we can replace
it by a closed polydisk defined by these coordinates. Namely, let $B' \subset
B$ be the polydisk $\{z \mid |z_i| \le \varepsilon/n \hbox{ for all } i\}$ and
suppose that the number $\delta$ (participating in the definition of the disc
$D \equiv D_\delta \subset \C^1$) is sufficiently small with respect to
$\varepsilon$ and $\varepsilon^2$.
\medskip
{\bf Lemma 1.} {\it For every surface $X_\lambda \subset B$ defined by the
equation $z_n \equiv \lambda, \ \lambda \in D$, the pair $(B' \setminus
AX_\lambda, \partial B' \setminus AX_\lambda)$ is homeomorphic to the pair $(B
\setminus AX_\lambda, \partial B \setminus AX_\lambda).$ These homeomorphisms
depend continuously on the parameter $\lambda$, and the induced isomorphisms of
all groups (\ref{groups}) onto similar groups in whose definition $B$ is
replaced by $B'$ coincide with the morphisms induced by the identical embedding
$B' \setminus AX_\lambda \to B \setminus AX_\lambda$ in the case of groups
${\mathcal H}_{i,\al}, \chi_{i,\al}$ and to the morphisms induced by the
obvious map $(B \setminus AX_\lambda)/(\partial B \setminus AX_\lambda) \to (B'
\setminus AX_\lambda)/(\partial B' \setminus AX_\lambda)$ (reduction mod $B
\setminus B'$) in the case of groups $\bar {\mathcal H}_{i,\al}, \bar
\chi_{i,\al}.$
Moreover, for every $r = m, \ldots, n$ and every $\lambda \in D$, all these
statements remain valid if we replace both $B$ and $B'$ by $B \cap T^r$ and $B'
\cap T^r$ respectively.}
\bigskip
The proof repeats the proof of Proposition 5 in \cite{V2}.
\medskip
Thus, everywhere in the proof of Theorems 1--4 we can replace the disc $B$ by
the polydisk $B'$.
\subsection{Fibre bundle \protect$z_{n-r}$}
\noindent For any $r = m, \ldots, n$ denote the $2r$-dimensional polydisk $B'
\cap T^r$ by $B^{(r)}$ and the characteristic radius $\varepsilon /n$ of all
such polydisks by $\epsilon$. Remember the notation $X \equiv X_\delta$ and $AX
\equiv A \cup X_\delta$. Denote the $\epsilon$-disc $\{w \in \C^1 \mid |w| \le
\epsilon \}$ by $\Omega$. For arbitrary $r < n$ consider the projection
\begin{equation} \label{przet}
z_{n-r}: (B^{(r+1)} \setminus AX) \to \Omega.
\end{equation}
\begin{figure}
\unitlength=1.00mm \special{em:linewidth 0.4pt} \linethickness{0.4pt}
\begin{picture}(113.00,53.00)
\put(39.00,29.00){\rule{2.00\unitlength}{2.00\unitlength}}
\put(30.00,30.00){\circle*{2.00}} \put(50.00,30.00){\circle*{2.00}}
\put(51.00,30.00){\vector(1,0){12.00}} \put(29.00,30.00){\vector(-1,0){12.00}}
\put(31.00,30.00){\vector(1,0){18.00}} \put(40.00,53.00){\vector(0,-1){45.00}}
\put(67.00,30.00){\makebox(0,0)[cc]{$\epsilon$}}
\put(13.00,30.00){\makebox(0,0)[cc]{$-\epsilon$}}
\put(30.00,33.00){\makebox(0,0)[cc]{$-\sqrt{\delta}$}}
\put(50.00,26.00){\makebox(0,0)[cc]{$\sqrt{\delta}$}}
\put(24.00,30.00){\line(4,-5){8.00}} \put(32.00,20.00){\line(4,5){16.00}}
\put(48.00,40.00){\line(4,-5){8.00}} \put(95.00,9.00){\makebox(0,0)[cc]{{\large
$b$}}} \put(21.00,9.00){\makebox(0,0)[cc]{{\large $a$}}}
\put(99.00,29.00){\rule{2.00\unitlength}{2.00\unitlength}}
\put(110.00,30.00){\circle*{2.00}} \put(90.00,30.00){\circle*{2.00}}
\put(98.00,31.00){\line(-1,0){6.00}} \put(92.00,29.00){\line(1,0){5.00}}
\put(97.00,29.00){\line(3,1){6.00}} \put(103.00,31.00){\line(1,0){5.00}}
\put(108.00,29.00){\vector(-1,0){6.00}} \put(110.50,31.00){\oval(5.00,2.00)[t]}
\put(110.50,29.00){\oval(5.00,2.00)[b]} \put(89.50,29.00){\oval(5.00,2.00)[b]}
\put(89.50,31.00){\oval(5.00,2.00)[t]} \put(87.00,31.00){\line(0,-1){2.00}}
\put(113.00,29.00){\line(0,1){2.00}}
\end{picture}
\caption{Base \protect$\Omega$ of the fibre bundle
\protect$z_{n-r}$}\label{omega}
\end{figure}
For any $t \in \Omega$ denote by ${\mathcal F}_t$ the fibre $(B^{(r+1)}
\setminus AX ) \cap \{z|z_{n-r}=t\}$ of this projection over the point $t$, and
by $\partial {\mathcal F}_t$ the boundary ${\mathcal F}_t \cap
\{z|\max(|z_{n-r+1}|, \ldots, |z_n|)=\delta\} \subset \partial B^{(r+1)}$ of
this fibre. In particular, ${\mathcal F}_0 \equiv B^{(r)}\setminus AX$,
$H_*^{lf}({\mathcal F}_0,
\partial {\mathcal F}_0; L_\alpha) \simeq \tilde {\mathcal H}_{*, \alpha}(r),$
$H_*^{lf}({\mathcal F}_0, L_\alpha) \simeq {\mathcal H}_{*, \alpha}(r),$
$H_*({\mathcal F}_0, \partial {\mathcal F}_0; L_\alpha) \simeq \tilde \chi_{*,
\alpha}(r),$ and $H_*({\mathcal F}_0, L_\alpha) \simeq \chi_{*, \alpha}(r).$
Any of two planes $\{z|z_{n-r}= \pm \sqrt{\delta}\}$ contains a distinguished
point $\Delta^{\pm}$, the critical point of the restriction of $z_{n-r}$ to the
manifold $\sigma \cap X \cap T^{r+1}$.
\smallskip
{\bf Lemma 2.} {\it If the radius $\epsilon$ of the polydisk $B'$ is
sufficiently small and $\delta$ is sufficiently small with respect to
$\epsilon^2$, then
a) the projection (\ref{przet}) defines a locally trivial fibre bundle over the
disc $\Omega$ with two points $\pm \sqrt{\delta}$ removed, the standard fibre
$({\mathcal F}_t, \partial {\mathcal F}_t)$ of which is homeomorphic to
$({\mathcal F}_0,
\partial {\mathcal F}_0)$;
b) the fibres over the exceptional points $\sqrt{\delta}$ and $-\sqrt{\delta}$
are homeomorphic to the direct product $\partial {\mathcal F}_0 \times (0,1]$
(or, more transparently, to the disc $B^{(r)}$ from which a cone over $\partial
B^{(r)} \setminus \partial {\mathcal F}_0$ is removed);
c) the restriction of this projection on $\partial B^{(r+1)}$ defines a
trivializable bundle with typical fibre $\partial B^{(r)} \setminus AX$ over
the interior part of $\Omega$.}
\smallskip
This lemma follows directly from the construction, from Thom's isotopy theorem
and from Proposition 2, cf. \cite{V2}. $\quad \Box$
\smallskip
Consider two variation operators $\tilde V_{+,-}: H^{lf}_*({\mathcal F}_0,
\partial {\mathcal F}_0;L_\al) \to H^{lf}_*({\mathcal F}_0,L_\al)$ and two similar
operators $V_{+,-}: H_*({\mathcal F}_0, \partial {\mathcal F}_0;L_\al) \to
H_*({\mathcal F}_0,L_\al)$ defined by the simple loops in $\Omega$
corresponding to the segments $[0, \sqrt{\delta}]$ and $[0, -\sqrt{\delta}]$.
(For instance the $\infty$-shaped loop in Fig. \ref{omega}b is a composition of
the second of these loops and the loop inverse to the first of them.)
\smallskip
{\bf Lemma 3.} {\it The operators $\tilde V_+$, $\tilde V_-$ are equal to one
another and to the operator $\widetilde{Var}_r: \bar {\mathcal H}_*(r) \to
{\mathcal H}_*(r).$ The operators $V_+$, $V_-$ are equal to one another and to
the operator} $Var_r: \bar \chi_*(r) \to \chi_*(r).$
\medskip
Indeed, any of these operators is defined by a loop in the space of pairs of
complex hyperplanes in $T^{r+1}$ in general position with $A$: for $V_+$ and
$V_-$ the first plane of this pair is fixed and coincides with $X_\delta \cap
T^{r+1}$, while the second is distinguished by the condition $z_{n-r}=t$ where
$t$ runs over the corresponding simple loop in $\Omega$; for $Var_r$ the {\it
second} plane is fixed and coincides with the one distinguished by $z_{n-r}=0$,
and the first moves and coincides with the planes $X_\lambda \cap T^{r+1}$
where $\lambda$ runs over the circle $C \equiv \partial D_\delta$.
It is easy to see that all these three loops are homotopic to one another in
the space of pairs of hyperplanes generic with respect to $A$, and Lemma 3 is
proved. $\quad \Box$
\subsection{Construction of maps (\protect\ref{stab3}).}
Let $x$ be any element of the group ${\mathcal H}_{j,\al}(r)$. The class
$\Sigma(x) \in {\mathcal H}_{j+1,\al}(r+1)$ realizing the first isomorphism in
(\ref{stab3}) is obtained from $x$ by a sort of the suspension operation.
Namely, using the fibre bundle structure (\ref{przet}), we transport the
realizing $x$ cycle ${\bf x} \subset {\mathcal F}_0$ over the segment
$[-\sqrt{\delta}, \sqrt{\delta}]$. This one-parametric family of cycles sweeps
out a $(j+1)$-dimensional chain in $B^{(r+1)} \setminus AX$ oriented by the
pair of orientations, the first of which is the orientation of the base segment
chosen as in Fig. \ref{omega}a, and the second is induced by the original
orientation of ${\bf x}$. The boundary of this chain lies in the marginal
fibres over the endpoints $-\sqrt{\delta}$ and $\sqrt{\delta}$. By the
statement b) of Lemma 2 the homology groups with closed supports of these
fibres are trivial, thus we can contract these boundaries inside these fibres
and get a cycle; $\Sigma(x)$ is defined as its homology class.
To construct the cycle $\rlh(x)$ we first transport the cycle ${\bf x}$ over
the S-shaped path in Fig. \ref{omega}a and get similar cycles in the fibres
over some two interior points of the segments $[\sqrt{\delta}, \epsilon]$ and
$[-\epsilon, -\sqrt{\delta}]$, then sweep out some $(j+1)$-dimensional chains
over these segments oriented as is shown in the same Fig. \ref{omega}a, and
again span the boundaries of the obtained chains inside the fibres over the
points $\pm \sqrt{\delta}$. The class $\rightleftharpoons(x) \in \bar {\mathcal
H}_{j+1,\al}(r+1)$ corresponding to $x$ is defined by half the difference of
these two cycles.
Given a homology class $y \in \bar \chi_{j,\al}(r)$, the corresponding cycle
$\dar(y) \in \bar \chi_{j+1,\al}(r+1)$ is swept out by the similar
one-parametric family of cycles obtained from a realizing $y$ compact cycle
${\bf y}$ transported over the axis $\{Re\ z_{n-r} =0\}$ oriented downwards.
To obtain the class $\infty(y) \in \chi_{j+1, \al}(r+1)$ we transport the same
cycle ${\bf y}$ along the $\infty$-shaped path in Fig. \ref{omega}b. The
boundary of the $(j+1)$-dimensional chain swept out by it belongs to $\partial
B'$ and is homeomorphic there (via the trivializing homeomorphism from
statement c) of Lemma 2) to the direct product of $\partial {\bf y}$ and this
path. In particular, it is the boundary of the $(j+1)$-dimensional chain in
$\partial B'$, homeomorphic to the direct product of $\partial {\bf y}$ and the
2-chain in $\C^1$ bounded by this path. Thus the difference of these two
$(j+1)$-dimensional chains is an absolute compact cycle in $B^{(r+1)} \setminus
AX$; $\infty(y)$ is defined as its homology class.
\medskip
Operations $\Sigma^l$ and $\dar^l$ participating in Theorem 1 are just the
$l$-fold iterations of these homomorphisms $\Sigma$ and $\dar$. The
homomorphisms $\rlh_l$ and $\infty_l$ defining the first summands in
(\ref{stab2}) are defined as $\rlh \circ \Sigma^{l-1}$ and $\infty \circ
\dar^{l-1}$ respectively.
\medskip
{\bf Remark.} If we replace the coordinate $z_{n-r}$ by $-z_{n-r}$, all
homomorphisms $\Sigma,$ $\dar,$ $\rlh$ and $\infty$ will be multiplied by $-1$.
The operations $\Sigma^l,$ $\dar^l,$ $\rlh_l$ and $\infty_l$ are thus defined
only up to a (common) sign switching with any involution $z_q \to z_{-q},$
$q=k, k-1, \ldots, k-l+1$.
\medskip
{\bf Conjecture.} {\it There is a continuous involution of the pair
$(B^{(r+1)}, AX)$ commuting with the involution $z_{n-r} \to -z_{n-r}$ of
$\Omega$ and identical on the fibre $\{z|z_{n-r}=0\}$ over the fixed point of
the latter involution.}
\medskip
Although we do not prove this conjecture, in the next section we use its
``homological shadow'', i.e. the involution in the homology groups, existing
(as we shall see) independently on this conjecture. \medskip
Finally, we realize two last summands in both equations (\ref{stab2}). Let
$\Delta \in B^{(m+l)}$ be any intersection point of the plane $X_{\delta}$ and
the stratum $\sigma$ of $A$. Let $\beta \subset B^{(m+l)}$ be a very small
closed disc centred at $\Delta$. Then by the K\"unneth formula for arbitrary
$\al$
\begin{equation}
\label{kunn} H_*(\beta \setminus AX, L_{\al}) \simeq \psi_{*,\al} \otimes
H_*(\C^1 \setminus 0, L_{\al_1}),
\end{equation}
\begin{equation}
\label{kunn1} H^{lf}_*(\beta \setminus AX, \partial \beta ; L_{\al}) \simeq
\Phi_{*-2l,\al} \otimes H^{lf}_*(\C^1 \setminus 0, L_{\al_1})
\end{equation}
(where the groups $H_*(\C^1 \setminus 0, L_{\al_1})$ and $H^{lf}_*(\C^1
\setminus 0, L_{\al_1})$ are trivial if $\al_1 \ne 1$ and are isomorphic to
$H_*(S^1)$ and $H_{*-1}(S^1)$ respectively if $\al_1 = 1$, and $\Phi_{*-2l}$
is the graded group obtained from $\Phi_*$ by the shift of grading). The sum of
two last summands in the second row of (\ref{stab2}) is isomorphic to the
$(i+l)$-dimensional component of the group (\ref{kunn}) and is realized as
follows. \medskip
{\bf Lemma 4.} {\it The homomorphism
\begin{equation}
\label{inj} H_*(\beta \setminus AX, L_{\al}) \to \chi_{*,\al}(m+l)
\end{equation}
induced by the identical embedding is injective.}
\medskip
If $l>1$ (so that the set $B^{(m+l)} \cap X_\delta \cap \sigma$ of possible
points $\Delta$ is path-connected) then the sum of two last summands in the
second row of (\ref{stab2}) coincides with the image of this injection. If
$l=1$ and this set consists of two points $\Delta^+$ and $\Delta^-$ (see \S \
3.2), then the corresponding groups $H_*(\beta^+ \setminus AX, L_{\al})$,
$H_*(\beta^- \setminus AX, L_{\al})$ are naturally identified to one another.
(This identification is transparent on the dual cohomological level, because
both cohomology groups dual to two factors in (\ref{kunn}) are induced from
groups $H^*(B^{(m+l)} \setminus A)$ and $H^*(B^{(m+l)} \setminus X)$
respectively.)
Consider the similar homomorphism $H_*(\beta^+ \setminus AX, L_{\al}) \oplus
H_*(\beta^- \setminus AX, L_{\al}) \to \chi_{*,\al}(m+l)$ defined by the
embedding $(\beta^+ \cup \beta^-) \to B^{(m+l)}$ and take the subgroup of its
image invariant under this identification; this subgroup is again isomorphic to
either of $H_*(\beta^\pm \setminus AX, L_{\al})$ and realizes the last
summands in the second row of (\ref{stab2}). Two last summands in the first row
are Poincar\'e dual to them, let us describe them explicitly.
The pair $(B^{(m+l)}, A)$ is homeomorphic to the product $(B^{(m)}, B^{(m)}
\cap A) \times \C^l,$ therefore for any cycle $w \in \Phi_{i-l}$ the cycle $w
\times \C^l$ defines an element of the group $H^{lf}_{i+l}(B^{(m+l)} \setminus
A, \partial B; L_{\hat \alpha})$ and, since $\alpha_1 =1,$ also of the group
$H^{lf}_{i+l}(B^{(m+l)} \setminus AX, \partial B; L_{\alpha}) \equiv \bar
{\mathcal H}_{i+l,\al}(m+l) $. We choose this cycle in general position with a
cycle generating the group $H^{lf}_{2(m+l)-1}(B^{(m+l)} \setminus X, \partial
B)$ (say, with the cycle given by the condition $z_n < \delta$), then the
intersection of these two cycles gives us also an element of the group $\bar
{\mathcal H}_{i+l-1,\al}(m+l) $; this element corresponds to the element $z$ of
the third summand of the decomposition (\ref{stab2}) of this group.
\section{Proof of main theorems}
\begin{figure}
\unitlength=1.00mm \special{em:linewidth 0.4pt} \linethickness{0.4pt}
\begin{picture}(137.00,116.00)
\put(15.00,105.00){\circle*{2.00}} \put(25.00,105.00){\circle*{2.00}}
\put(20.00,105.00){\oval(30.00,20.00)[]} \put(39.00,105.00){\vector(1,0){7.00}}
\put(65.00,105.00){\oval(30.00,20.00)[]} \put(84.00,105.00){\vector(1,0){7.00}}
\put(110.00,105.00){\oval(30.00,20.00)[]} \put(115.00,45.00){\circle*{2.00}}
\put(105.00,45.00){\circle*{2.00}} \put(114.83,45.67){\oval(4.33,4.00)[t]}
\put(112.67,45.67){\line(-1,0){6.67}} \put(104.67,44.33){\oval(4.67,4.00)[b]}
\put(107.00,44.33){\line(1,0){7.00}} \put(20.00,75.00){\oval(30.00,20.00)[]}
\put(39.00,75.00){\vector(1,0){7.00}} \put(65.00,75.00){\oval(30.00,20.00)[]}
\put(84.00,75.00){\vector(1,0){7.00}} \put(110.00,75.00){\oval(30.00,20.00)[]}
\put(15.00,75.00){\circle*{2.00}} \put(25.00,75.00){\circle*{2.00}}
\put(20.00,75.00){\vector(-1,1){5.00}} \put(15.00,80.00){\vector(-1,-1){5.00}}
\put(10.00,75.00){\vector(1,-1){5.00}} \put(15.00,70.00){\vector(1,1){10.00}}
\put(25.00,80.00){\vector(1,-1){5.00}} \put(30.00,75.00){\vector(-1,-1){5.00}}
\put(25.00,70.00){\vector(-1,1){5.00}} \put(60.00,75.00){\circle*{2.00}}
\put(70.00,75.00){\circle*{2.00}} \put(65.00,75.00){\vector(-1,2){5.67}}
\put(59.33,86.00){\vector(-1,0){10.33}} \put(49.00,86.00){\vector(0,-1){22.00}}
\put(49.00,64.00){\vector(1,0){10.00}} \put(59.00,64.00){\vector(1,2){11.00}}
\put(70.00,86.00){\vector(1,0){11.00}} \put(70.00,64.00){\vector(-1,2){5.67}}
\put(81.00,86.00){\vector(0,-1){22.00}} \put(81.00,64.00){\vector(-1,0){11.00}}
\put(105.00,75.00){\circle*{2.00}} \put(115.00,75.00){\circle*{2.00}}
\put(109.00,75.00){\vector(0,1){11.00}}
\put(109.00,86.00){\vector(-1,0){15.00}}
\put(94.00,86.00){\vector(0,-1){22.00}} \put(94.00,64.00){\vector(1,0){17.00}}
\put(111.00,64.00){\vector(0,1){22.00}} \put(111.00,86.00){\vector(1,0){15.00}}
\put(126.00,86.00){\vector(0,-1){23.00}}
\put(126.00,63.00){\vector(-1,0){17.00}}
\put(109.00,63.00){\vector(0,1){12.00}} \put(15.00,45.00){\circle*{2.00}}
\put(25.00,45.00){\circle*{2.00}} \put(20.00,45.00){\oval(30.00,20.00)[]}
\put(39.00,45.00){\vector(1,0){7.00}} \put(65.00,45.00){\oval(30.00,20.00)[]}
\put(65.00,50.00){\circle*{2.00}} \put(65.00,40.00){\circle*{2.00}}
\put(65.00,40.00){\line(-1,1){5.00}} \put(65.00,50.00){\line(1,-1){5.00}}
\put(84.00,45.00){\vector(1,0){7.00}} \put(110.00,45.00){\oval(30.00,20.00)[]}
\put(15.00,16.00){\circle*{2.00}} \put(25.00,16.00){\circle*{2.00}}
\put(20.00,16.00){\oval(30.00,20.00)[]} \put(39.00,16.00){\vector(1,0){7.00}}
\put(65.00,16.00){\oval(30.00,20.00)[]} \put(65.00,11.00){\circle*{2.00}}
\put(84.00,16.00){\vector(1,0){7.00}} \put(110.00,16.00){\oval(30.00,20.00)[]}
\put(20.00,27.00){\vector(0,-1){22.00}} \put(65.00,21.00){\circle*{2.00}}
\put(65.00,27.00){\line(-1,-2){3.00}} \put(62.00,21.00){\line(3,-5){6.00}}
\put(68.00,11.00){\vector(-1,-2){3.00}} \put(105.00,16.00){\circle*{2.00}}
\put(115.00,16.00){\circle*{2.00}} \put(110.00,27.00){\line(0,-1){10.00}}
\put(110.00,17.00){\line(-5,2){5.00}} \put(105.00,13.00){\line(5,3){10.00}}
\put(115.00,13.00){\line(-5,2){5.00}} \put(110.00,15.00){\vector(0,-1){10.00}}
\put(137.00,104.00){\makebox(0,0)[cc]{{\large a)}}}
\put(137.00,74.00){\makebox(0,0)[cc]{{\large b)}}}
\put(137.00,44.00){\makebox(0,0)[cc]{{\large c)}}}
\put(137.00,15.00){\makebox(0,0)[cc]{{\large d)}}}
\put(105.00,16.00){\oval(6.00,6.00)[l]} \put(115.00,16.00){\oval(6.00,6.00)[r]}
\put(115.00,105.00){\circle*{2.00}} \put(105.00,105.00){\circle*{2.00}}
\put(70.00,105.00){\circle*{2.00}} \put(60.00,105.00){\circle*{2.00}}
\put(15.00,106.00){\vector(1,0){10.00}} \put(15.00,104.00){\vector(1,0){10.00}}
\put(60.00,106.00){\line(1,2){5.00}} \put(65.00,116.00){\vector(1,-2){5.00}}
\put(60.00,104.00){\line(1,-2){5.00}} \put(65.00,94.00){\vector(1,2){5.00}}
\put(105.00,104.00){\vector(-1,0){11.00}}
\put(94.00,104.00){\vector(0,-1){10.00}} \put(94.00,94.00){\vector(1,0){32.00}}
\put(126.00,94.00){\vector(0,1){10.00}}
\put(126.00,104.00){\vector(-1,0){11.00}}
\put(105.00,106.00){\vector(-1,0){11.00}}
\put(94.00,106.00){\vector(0,1){10.00}} \put(94.00,116.00){\vector(1,0){32.00}}
\put(126.00,116.00){\vector(0,-1){10.00}}
\put(126.00,106.00){\vector(-1,0){11.00}}
\put(14.00,45.00){\vector(-1,0){9.00}} \put(25.00,45.00){\vector(1,0){10.00}}
\put(60.00,45.00){\vector(-1,0){10.00}} \put(70.00,45.00){\vector(1,0){10.00}}
\put(117.00,45.33){\vector(1,0){8.00}} \put(102.00,44.67){\vector(-1,0){7.00}}
\end{picture}
\caption{Proofs of Lemmas 5 and 6} \label{defs}
\end{figure}
{\bf Lemma 5.} {\it For any $x \in {\mathcal H}_{i,\al}(r)$, the homomorphism $
\tilde J_{\al,r+1} \equiv \tilde j_{\al}\circ \tilde i_{\al}: {\mathcal
H}_{j+1,\al}(r+1) \to \bar {\mathcal H}_{j+1,\al}(r+1)$ (the reduction mod
$\partial B^{(r+1)}$) maps any element $\Sigma(x),$ $x \in {\mathcal
H}_{j,\al}(r),$ to $- \rlh (2x + \widetilde{Var}_{r} \circ \tilde J_{\al,r}(x))
\equiv -\rlh \circ (M_r+Id)(x)$.
The similar homomorphism $ J_{\al,r+1} \equiv j_{\al}\circ i_{\al}:
\chi_{j+1,\al}(r+1) \to \bar \chi_{j+1,\al}(r+1)$ maps any element $\infty
(y)$, $y \in \bar \chi_{j,\al}(r),$ to $- \dar (2y + J_{\al,r} \circ
Var_{r}(y)) \equiv - \dar \circ (\bar \mu_r + Id)(y)$. }
\medskip
{\it Proof.} See Figs. \ref{defs}a), b).
\medskip
{\bf Lemma 6.} {\it For any $r = m, \ldots, n-1,$ and any elements} $x \in
{\mathcal H}_{j,\al}(r)$ and $y \in \bar \chi_{j,\al}(r),$
$\widetilde{Var}_{r+1}(\rlh(x)) = \Sigma(x)$, $Var_{r+1} (\dar(y)) =
\infty(y)$.
\medskip
{\it Proof.} When $\xi$ moves along the circle $\delta \cdot e^{i\tau}, \tau
\in [0,2\pi],$ the ramification points $\pm \sqrt{\xi} \subset \Omega$ of the
fibre bundle $z_{n-r}: (B^{(r+1)},AX_\xi) \to \Omega$ move as is shown in Fig.
\ref{defs} c). We move the segments $[-\epsilon, -\sqrt{\xi}]$, $[\sqrt{\xi},
\epsilon]$ connecting the ramification points with ``infinity'' in such a way
that they coincide with the parts of the real axis close to the boundary of
$\Omega$. Let us also move the cycle realizing $\rightleftharpoons(\alpha)$ in
such a way that at any instant of this deformation it forms a fibre bundle with
standard fibre $\alpha/2$ over the union of two corresponding segments (except
for their endpoints). At the final instant of the monodromy each of these two
parts of $\rightleftharpoons(\alpha)$ will increase by the cycle
$\Sigma(\alpha)/2$ swept by the formal half of the initial cycle $\alpha$ in
transport over the added part of the resulting segment, so that $
\widetilde{Var}_{r+1}(\rightleftharpoons(\alpha)) = \Sigma(\alpha)$. For the
similar proof of the formula for the variation of $\dar(y)$, see Fig.
\ref{defs}d).
\medskip
This lemma proves statement b) of Theorem 3 and assertion of statement a)
concerning the variation of cycles $\rlh_l(x)$.
\medskip
Statement a) of Theorem 2 and assertion of statement b) concerning the action
on elements $\infty_l(y)$ follow from Lemmas 5 and 6 by induction over $r$.
\medskip
{\bf Corollary.} {\it The local monodromy operators $M_{r+1}, \bar M_{r+1},
\bar \mu_{r+1}$ and $\mu_{r+1}$ respectively map the elements $\Sigma(x),
\rlh(x), \dar(y)$ and $\infty(y)$ (where $x \in {\mathcal H}_{*,\alpha}(r), y
\in \bar \chi_{*,\alpha}(r)$) into $- \Sigma(M_r(x))$, $- \rlh(M_r(x))$, $-
\dar(\bar \mu_r(y))$ and $- \infty(\bar \mu_r(y))$ respectively.}
\medskip
{\bf Lemma 7.} {\it For any $x \in {\mathcal H}_{j,\al}(r)$ and} $y^* \in \bar
\chi_{2r-j,\al^*}(r)$,
$$\langle\Sigma(x), \dar(y^*)\rangle = (-1)^{1+j}\langle x, y^*\rangle ,$$
$$\langle\rlh(x), \infty(y^*)\rangle = (-1)^{1+j}\langle x, \bar
\mu_r(y^*)\rangle.$$
This lemma follows immediately from constructions. Statements A) and B2) of
Theorem 4 follow from it by induction over $l$.
\medskip
{\bf Proposition 3.} {\it The homomorphisms $\Sigma : {\mathcal H}_{j,\al}(r)
\to {\mathcal H}_{j+1,\al}(r+1)$ and $\dar : \bar \chi_{j,\al}(r) \to \bar
\chi_{j+1,\al}(r+1)$ are isomorphisms for any $\al$ and $r=m, \ldots, n-1$.}
\medskip
{\it Proof.} Consider any smooth deformation contracting the disc $\Omega$ onto
the segment $[-\sqrt{\delta}, \sqrt{\delta}]$. Using the fibre bundle
(\ref{przet}), we can lift this deformation to the space $B^{(r+1)}\setminus
AX$. This lifted field allows us to realize any element of the group ${\mathcal
H}_{j+1,\al}(r+1)$ by a locally finite chain lying in this pre-image, and hence
also by one of the form $\Sigma(x),$ $x \in {\mathcal H}_{j,\al}(r)$. Thus the
map $\Sigma$ is epimorphic.
Our proposition follows now from the first equation of Lemma 7 and from the
fact that both Poincar\'e pairings ${\mathcal H}_{j,\al}(r) \otimes \bar
\chi_{2r-j,\al^*}(r) \to \C$ and ${\mathcal H}_{j+1,\al}(r+1) \otimes \bar
\chi_{2r-j+1,\al^*}(r+1) \to \C$ are non-degenerate.
\medskip
Theorem 1 is a direct corollary of this proposition.
\medskip
Denote by $\partial^- B^{(r+1)}$ the union of the usual boundary $\partial
B^{(r+1)}$ and the ``left half'' of $B^{(r+1)}$ consisting of points $z$ with
$z_{n-r} \le 0$.
\medskip
{\bf Proposition 4.} {\it For any $r=m, \ldots, n-1$ there is short exact
sequence}
\begin{equation}
\label{exact} 0 \to {\mathcal H}_{j}(r) \to \bar {\mathcal H}_{j+1}(r+1)
\stackrel{\rho}{\to} H^{lf}_{j+1}(B^{(r+1)}\setminus AX,
\partial^- B^{(r+1)}; L_\al) \to 0,
\end{equation}
{\it where the injection ${\mathcal H}_{j}(r) \to \bar {\mathcal H}_{j+1}(r+1)$
is the map $\rlh,$ and the epimorphism $\rho$ is induced by the reduction
modulo $\{z|z_{n-r} \le 0 \}$.}
\medskip
{\it Proof} (cf. \S \ 4.2 in \cite{V2}). We filter the disc $\Omega$ by the
sets $\{ \Psi_0 \subset \Psi_1 \subset \Psi_2 \}$ where $\Psi_0$ consists of
two points $\pm \sqrt{\delta}$, $\Psi_1$ consists of two segments
$[\sqrt{\delta}, \epsilon]$ and $[-\epsilon, -\sqrt{\delta}]$ (so that the set
$\Omega \setminus \Psi_1$ is a 2-cell), and $\Psi_2 \equiv \Omega$. Using the
projection (\ref{przet}) we lift this filtration onto the set
$B^{(r+1)}\setminus AX$; let $E_{p,q}^r$ be the spectral sequence generated by
this filtration and calculating the group $\bar {\mathcal H}_{*,\al}(r+1)$.
\medskip
{\bf Proposition 5.} {\it For any $q$, the elements $E_{p,q}^1$ of this
spectral sequence are as shown in the second row of Table 1.}
\medskip
\begin{table}
\caption{The groups $E^1_{p,q}$ for the main, anti-invariant and invariant
spectral sequences}
\begin{center}
\begin{tabular}{|c|cccc|}
\hline
$p$ & 0 & 1 & 2 & $\ge 3$ \\
\hline $E^1_{p,q}$ & $(H^{lf}_{q-1}(\partial {\mathcal F}_0),L_\al)^2$ & $(\bar
{\mathcal H}_{q,\al}(r))^2$ &
$\bar {\mathcal H}_{q,\al}(r)$ & 0 \\
\hline $E_{p,q}^{1,-}$ & $H^{lf}_{q-1}(\partial {\mathcal F}_0,L_\al)$ &
$\bar {\mathcal H}_{q,\al}(r)$ & 0 & 0 \\
\hline $E_{p,q}^{1,+}$ & $H^{lf}_{q-1}(\partial {\mathcal F}_0,L_\al)$ & $\bar
{\mathcal H}_{q,\al}(r)$ &
$\bar {\mathcal H}_{q,\al}(r)$ & 0 \\
\hline \end{tabular} \end{center} \end{table}
Indeed, the assertion concerning the group $E_{0,q}^1$ follows from the
statement b) of Lemma 2. For any element $z \in \bar {\mathcal H}_*(r)$, two
corresponding elements in two summands $\bar {\mathcal H}_*(r)$ of $E_{1,*}^1$
are the homology classes of two cycles (mod $z_{n-r}^{-1}(\Psi_0)$) swept out
by the copies of the cycle realizing $z$ transported over two segments as in
the definition of the operation $\rlh$, see \S \ 3.3, and the corresponding
element of the group $E_{2,*}^1$ is swept out by the two-parametric family of
cycles obtained from the cycle realizing $z$ by the transportation to all
points of the 2-cell $\Omega \setminus \Psi_1$.
\medskip
Consider an involution acting on the term $E^1$ of our spectral sequence: it
permutes the elements of terms $E_{0,q}$, $E_{1,q}$, corresponding to the same
elements of groups $\tilde H^{lf}_{q-1}(\partial {\mathcal F}_0,L_\al)$, $\bar
{\mathcal H}_{q,\al}(r)$, and does not touch the terms $E_{2,q}$. Denote by
$E_{p,q}^{+,r}$ and $E_{p,q}^{-,r}$ the invariant and anti-invariant parts of
this involution, respectively.
\medskip
{\bf Lemma 8.} {\it The splitting of the term $E^1$ into the invariant and
anti-invariant parts is compatible with all further differentials and thus
defines the splitting of the entire spectral sequence. The groups $E^{+,1}$ and
$E^{-,1}$ of the invariant (respectively, anti-invariant) subsequence are as
shown in the fourth (respectively, third) row of the Table 1. The unique
nontrivial differential $\partial_1: \bar {\mathcal H}_{q,\al}(r) \to
H^{lf}_{q-1}(\partial {\mathcal F}_0,L_\al)$ of the anti-invariant subsequence
coincides with the boundary operator in ${\mathcal F}_0$, so that this
subsequence is nothing but the exact sequence of the pair $({\mathcal F}_0,
\partial {\mathcal F}_0)$ calculating the group $H^{lf}_*({\mathcal F}_0, L_\al) \equiv
{\mathcal H}_{q,\al}(r)$. The invariant subsequence is isomorphic to the
spectral sequence generated by the same filtration on the quotient space
$(B^{(r+1)}\setminus AX)/ (\partial^- B^{(r+1)}\setminus AX)$ and calculating
its homology group $H^{lf}_{i+1}(B^{(r+1)}\setminus AX,
\partial^- B^{(r+1)}\setminus AX; L_\al) $; this isomorphism is defined
by the reduction modulo $\{z|z_{n-r} \le 0\}$.}
\medskip
All this follows immediately from the construction and proves Proposition 4.
\medskip
By the construction, the induced involution of the limit homology group $\bar
{\mathcal H}_{i+1}(r+1)$ acts as multiplication by $-1$ on the subgroup
$\rlh({\mathcal H}_{i}(r))$ and acts trivially on the corresponding quotient
group, in particular the invariant subspace of this involution is canonically
isomorphic to $H^{lf}_{i+1}(B^{(r+1)}\setminus AX,
\partial^- B^{(r+1)}\setminus AX; L_\al) $.
\medskip
{\bf Lemma 9.} {\it Let $\beta$ be a sufficiently small closed disc in
$B^{(r+1)}$ centred at the distinguished point $\Delta^+$ (or $\Delta^-$) of
the fibre $\{z|z_{n-r} = \sqrt{\delta}\}$ (respectively, $-\sqrt{\delta}$).
Then the group $H^{lf}_{*}(B^{(r+1)}\setminus AX,
\partial^- B^{(r+1)}\setminus AX; L_\al) $
is isomorphic to $H_*^{lf}(\beta \setminus AX, \partial \beta; L_\al)$; this
isomorphism is induced by the reduction modulo the closure of the complement of
$\beta$.}
\medskip
This lemma follows from Lemma 2. Together with the K\"unneth decomposition
(\ref{kunn}) it proves the statement of Theorem 1$'$ concerning the group $\bar
{\mathcal H}_*$, and also the fact that the last summands in (\ref{stab2}) are
actually generated by the cycles described between Lemmas 4 and 5. Three
remaining assertions of Theorems 2--4 (triviality of the action of
homomorphisms $J_{\al, (m+l)}$ (in Theorem 2) and $\widetilde{Var}_{(m+l)}$ (in
Theorem 3) on the last summands in (\ref{stab2}) and statement B1 in Theorem 4)
follow almost immediately from the construction of all these cycles. Lemma 4 is
just the fact Poincar\'e dual to the surjectivity of the map $\rho$ in
(\ref{exact}).
|
\section{Introduction}
Since the discovery of conduction at the interface between the two band-insulating perovskite oxides SrTiO$_3$ and LaAlO$_3$\cite{Ohtomo2004} a plethora of new effects have been found, ranging from superconductivity\cite{Reyren2007} to magnetism\cite{Brinkman2007,BenShalom2009, Ariando2011, Dikin2011, Li2011, Bert2011, Lee2013, Huijben2006} and tunable switching of high mobility interface conductivity,\cite{Thiel2006, Caviglia2008, Guduru2013} depending on the ground state the sample reaches. The nature of the ground state present in the systems depends closely on the growth parameters of the LaAlO$_3$-layer,\cite{Brinkman2007} the LaAlO$_3$ layer thickness\cite{Kozuka2010, Wong2010, Bell2009} and configuration of the heterostructures with different capping layers on top of LaAlO$_3$.\cite{Huijben2006, McCollam2014, Huijben2009, Huijben2013}\\
A number of possible mechanisms are proposed to be responsible for the conduction at the interface,\cite{Nakagawa2006, Yoshimatsu2008, Sing2009, Siemons2007, Herranz2007, Willmott2007} which can lead to multiple charge carrier conduction if several are active. Indeed multiple carrier conduction has been found in SrTiO$_3$/LaAlO$_3$-interfaces over a wide range of growth condictions.\cite{McCollam2014, Guduru2013a,Popovic2008a, Lerer2011,BenShalom2009, Rakhmilevitch2013, Seo2009}\\
Until present, most investigations on the electronic properties of LaAlO$_3$/SrTiO$_3$ interfaces have been done using transport experiments,\cite{Hwang2012, Mannhart2010, Ohtomo2004, Reyren2007, Brinkman2007, BenShalom2009, Ariando2011, Dikin2011, Li2011, Bert2011, Lee2013, Huijben2006, Thiel2006, Caviglia2008,Kozuka2010, Wong2010, Huijben2013, McCollam2014, Huijben2009, Guduru2013, Guduru2013a, BenShalom2009, Rakhmilevitch2013, Seo2009, Bell2009} while measurements of thermoelectric power are still sparse in SrTiO3/LaAlO3-interfaces.\cite{Pallecchi2010, Lerer2011, Filippetti2012} Whereas transport experiments are generally dominated by the charge carrier mobility, contributions of lower mobility can be accessed in thermoelectric power measurements. Additionally, the thermoelectric power is known to be sensitive to magnetic scattering, thus an ideal tool to investigate multiple charge carrier contributions in samples with magnetic signatures.\\
In the present work we report on our investigation of the interface electronic structure of one specific type of LAO/STO heterostructure, with a 10\,nm (26 unit cells) LAO film, which is known to exhibit magnetic signatures.\cite{Brinkman2007} Magnetotransport and thermoelectric-power measurements have been performed in a large temperature and magnetic field range. We apply a two carrier model to the magnetotransport data and find two different charge carriers with different densities and mobilities. By combining transport and thermopower data, we extend this model to the thermoelectric tensor $\epsilon$ at zero field and develop a preliminary description for its behaviour in magnetic field.\\
The paper is organized as follows: in section II we describe the details of sample growth and the experimental setup. In section III we show our transport results, which we describe in terms of a classical two-carrier model. In section IV we present our results of the thermopower-measurements and extend our model to thermopower. In section V we discuss our results and draw conclusions in section VI.
\section{Experimental details}
The sample is grown by pulsed laser deposition and has a 10\,nm thick (26 unit cells) LAO film on a TiO$_2$-terminated single crystal STO [001] substrate (treatment described in Ref. \onlinecite{Koster1998}). The LAO film was deposited at a substrate temperature of 850$\,^{\circ}\mathrm{C}$ and an oxygen pressure of $2\cdot10^{-3}$\,mbar, in order to minimize oxygen vacancies, using a single-crystal LaAlO$_3$ target. The growth of the LAO film was monitored using in-situ reflection high-energy electron diffraction. After the growth, the samples were cooled to room temperature at the deposition pressure.\\
Electrical contacts to the sample were made using an ultrasonic wire-bonder to punch through the LAO-layers and attaching manganine-wires to the holes with silver-paint.\\
The thermoelectric power was measured in a home-build apparatus in a standard one heater, two thermometer-geometry similar to the one used in Ref. \onlinecite{Fletcher1994}. The resistivity was measured on the same sample and in the same apparatus using a conventional low-noise lock-in technique. Both quantities were measured in positive and negative field directions and the data shown is symmetrized (antisymmetrized) to obtain the diagonal (off-diagonal) components of the resistivity- and thermoelectric power tensors. We use the historical sign convention for the Nernst-Ettingshausen effect throughout the paper (positive Nernst signal along $y$-direction, when the field is in $z$-direction and gradient in $x$-direction).
\section{Magnetotransport}
\label{sec:magnetotransport}
The temperature dependence of the sheet resistance $R_S$, the magnetoresistance $R_{xx}$ and the Hall-data $R_{xy}$ are shown in figure \ref{fig:figure1}a, c and d, respectively. The analysis of the transport data follows closely our previous work \cite{Guduru2013a}. From the transport data (figure \ref{fig:figure1}), we can distinguish three different regions (from low to high temperature): region I (up to $\sim$\,8\,K) with logarithmically decreasing sheet resistance and linear Hall effect, region II ($\sim$\,8\,K to $\sim$\,50\,K) with strongly decreasing sheet resistance and strongly non-linear Hall-resistance and region III (above $\sim$\,50\,K) showing an increase in the sheet resistance and a linear Hall-effect with respect to the magnetic field.
\begin{figure}[htb]
\centering
\includegraphics[width=\linewidth]{Figure1.pdf}%
\caption{(color online) a) Temperature dependence of sheet resistance (line). Sheet resistance from two-charge carrier model (dots).\\
b) Sheet resistance (black line) and Hall-resistance (dashed red line) at 4\,K showing the negative magnetoresistance and the linear Hall-effect.\\
c) Measured sheet resistance $R_{xx}$ for temperature between 4\,K and 100\,K.\\
d) Measured Hall resistivity for temperature between 4\,K and 100\,K (empty dots). Fits with the two-charge-carrier model (lines).}%
\label{fig:figure1}
\end{figure}
The transport data can be analyzed using a two-charge-carrier model for two independent, electron-like channels. For T\;$>$\;11\,K the magnetic field dependent Hall-resistance $R_{xy}$ was fitted with a restriction to the zero-field resistance $R_{S0}$ using the equations\cite{AshcroftMermin1976}:
\begin{equation}
\label{eq:2chargecarrier}
\begin{aligned}
R_{xy} &=
\frac{B}{e}\frac{(n_1\mu_1^2+n_2\mu_2^2)+(\mu_1\mu_2B)^2(n_1+n_2)}{
(n_1\mu_1+n_2\mu_2)^2+(\mu_1\mu_2B)^2(n_1+n_2)^2}\\
R_{S0}&=\frac{1}{e(n_1\mu_1+n_2\mu_2)}
\end{aligned}
\end{equation}
with the $e$ the electron charge, $B$ the magnetic field and the independent fitting parameters $n_{1,2}$ and $\mu_{1,2}$ as charge carrier density and mobility for the two channels, respectively. For the lowest and highest
temperature, where the Hall-resistance is completely linear, the density for one type of charge carrier was obtained from a linear fit to $R_{xy}=B/(en)$ and the mobility from $R_{S0}=1/(en\mu)$.
\begin{figure}[htb]
\centering
\includegraphics[width=0.8\linewidth]{Figure2.pdf}%
\caption{(color online) a) Charge carrier densities of the two electron-like bands $n_1$, $n_2$ from the 2-charge-carrier model (inset: Arrhenuis-plot for the n$_2$).\\
b) Mobilities of the two electron-like bands $\mu_1$, $\mu_2$ from the
2-charge-carrier model.
}%
\label{fig:figure2}
\end{figure}
The obtained fitting parameters are shown in figure \ref{fig:figure2}.\footnote{The data-point at 7.5\,K violates the Mott-Ioffe-Regel criterion ($k_fl<1$). We still included it in the figure for completeness.} The first type of charge carriers ($n_1$, $\mu_1$) has a relatively high and temperature independent charge carrier density of about $n_1 \sim 10^{14}$\,cm$^{-2}$. The mobility of these charge carriers is low with $\mu_1 \sim 40$\,cm$^{2}/$Vs and increases at higher temperatures to $\mu_1 \sim 1$\,cm$^{2}/$Vs at 4\,K (therefore referred to as low mobility charge carriers). This decrease in mobility to lower temperatures was attributed to magnetic, Kondo-like scattering due to the negative magneto-resistance at low temperatures and is described in more detail in our previous work\cite{Brinkman2007, Guduru2013a}. The slight decrease in mobility at higher temperatures is probably due to electron-phonon scattering.\\
The second type of charge carriers has a much lower charge carrier density which is increasing with a thermally activated behaviour $n_2 \propto e^{-\Delta_2/k_BT}$ from $n_2 \sim 2\cdot10^{9}$\,cm$^{-2}$ at 11\,K to $n_2 \sim 4\cdot10^{11}$\,cm$^{-2}$ at 65\,K, i.e. we deal with charge carrier, that have to be treated as a non-degenerate electron gas. To derive the energy gap between the non-degenerate, high mobility band at higher energy and the low lying, low mobility band, we can use an Arrhenius-plot (inset of figure \ref{fig:figure2}a)), which gives a gap of $\Delta_2 = (6.4\pm0.4)$\,meV, comparable to the results on other samples\cite{Guduru2013a, Huijben2006, Huijben2013}. The mobility is high at low temperatures ($\mu_2 \sim 2000$\,cm$^{2}$/Vs) and decreases at higher temperatures by one order of magnitude (high mobility charge carriers).
\section{Thermoelectric Power}
The temperature dependence of the zero-field thermoelectric power $-S_{xx}$ is shown in figure \ref{fig:figure3}a. First, the thermoelectric power is negative, confirming that our carriers are electrons. Second, we point out the absence of a clear phonon-drag peak.\cite{Fletcher1994} Only a faint deviation from the observed dependence is visible around 15\,K (marked by an asterisk *). Thus we assume that we are in the regime of diffusion-dominated thermopower. The three regions identified in the transport measurements are indicated by the dashed lines. Region I cannot be clearly identified in thermopower due to the lack of data-points. In region II, the thermopower is increasing monotonously, proportional to approximately between $T^{0.3}$ and $T^{0.4}$. We note that the temperature dependence is also in agreement with thermoelectric power by variable range hopping which gives a $T^{1/3}$ dependence for a 2D electron gas.\cite{Burns1985}\\
If multiple charge carriers are present, the individual contributions of the charge carrier have to be weighted by their conductivities and we can write for $S_{xx}$ at zero magnetic field\footnote{for the thermoelectric tensor we can write as explained below: $\epsilon_{xx} = \epsilon_{xx}^{(1)} + \epsilon_{xx}^{(2)}$. With $\epsilon_{xx}^{(i)} = \sigma_{xx}^{(i)}S_{xx}^{(i)}$ we get immediately $S_{xx}(\sigma_{xx}^{(1)} + \sigma_{xx}^{(2)}) = \sigma_{xx}^{(1)}S_{xx}^{(1)} + \sigma_{xx}^{(2)}S_{xx}^{(2)}$ and equation (\ref{eq:2bandTEP}) follows.}:
\begin{equation}
\label{eq:2bandTEP}
S_{xx} = \frac{\sigma_{xx}^{(1)} S_{xx}^{(1)} + \sigma_{xx}^{(2)}S_{xx}^{(2)}}{\sigma_{xx}^{(1)}+\sigma_{xx}^{(2)}}.
\end{equation}
The combination of both contributions explains the observed temperature dependence and will be described later in this paper using the thermoelectric tensor (see figure \ref{fig:figure4}).\\
In region III, the thermopower is constant up to about 120\,K and increases monotonously at higher temperatures, again proportional to approximately $T^{0.4}$.
\begin{figure}[htb]
\centering
\includegraphics[width=0.8\linewidth]{Figure3.pdf}%
\caption{(color online) a) Zero-field thermoelectric power $S_{xx}$ in a semi-logarithmic plot for clarity. A small deviation from the observerd $T^{0.4}$ dependence around 15\,K is marked by an asterisk *.\\
b) Magnetothermopower \textit{MS} (see text) for selected temperatures.\\
c) Nernst effect $S_{xy}$ for selected temperatures.
}%
\label{fig:figure3}
\end{figure}\\
When a magnetic field is applied, the magnitude of the thermoelectric power decreases. We can define the magneto-thermopower MS as
\begin{equation}
MS(B)=\frac{S_{xx}(B)-S_{xx}(B=0)}{S_{xx}(B=0)},
\label{magneto-TEP}
\end{equation}
shown in figure \ref{fig:figure3}b. In region I (4\,K), the magneto-thermopower is weak and linear with a decrease by about 11\% at 16\,T. In region II, the MS becomes stronger and non-linear with a maximal suppression of 25\% at 15\,K. Above 15\,K the magneto-thermopower becomes weaker again but remains non-linear. In region III, the magneto-thermopower is reduced even more and becomes linear again.\\
The Nernst effect $S_{xy}$ (figure \ref{fig:figure3}c) shows similar behaviour. In region I it is nearly linear and small, with $\nu = S_{xy}/B \approx -12$\,nV/(TK). We note that this is still enhanced compared to the classical Fermi-liquid picture\cite{Behnia2009}, which is approximately:
\begin{equation}
\nu=-\frac{\pi^2}{3}\frac{k_B}{e}\frac{k_BT}{\epsilon_F}\mu \approx -[0.1...1] \frac{\mbox{nV}}{\mbox{TK}}
\end{equation}
depending on the actual Fermi-energy.
At higher temperature (region II), $S_{xy}$ becomes non-linear and increases by an order of magnitude at 40\,K and 16\,T in comparison to the value at 4\,K. When entering region III, the Nernst signal becomes linear again and decreases towards higher temperatures.\\
The overall behaviour of the thermoelectric signals in magnetic field resembles the transport signals, i.e. linear and small in regions I and III and non-linear and large in region II. It is therefore beneficial to derive an appropriate two-charge-carrier model for the thermopower, which can be done following the work of S. Cao \textit{et al.}\cite{Cao1996a}\\
The thermopower tensor is defined by $S=E/(\nabla T)$ under the condition of $J_q =0$, with $E$ the electric field, $J_q$ the charge current density and $\nabla T$ the temperature gradient. We can use an extended Ohm's law
\begin{equation}
J_q = \sigma E - \epsilon \nabla T
\end{equation}
with $\sigma$ the conductivity tensor, $\epsilon$ the thermoelectric tensor and $\rho=\sigma^{-1}$ the resistivity, to find an expression for the thermopower tensor:
\begin{equation}
\label{eq:thermopower}
S = E/(\nabla T)= \sigma^{-1}\epsilon = \rho\epsilon
\end{equation}
By rewriting equation (\ref{eq:thermopower}), we get $\epsilon = \sigma S$ for the thermoelectric tensor. Using $\sigma$ and $S$ from our measured data we obtain $\epsilon_{xx}$ and $\epsilon_{xy}$, shown in figure \ref{fig:figure4}.\\
\begin{figure}[ht]
\centering
\includegraphics[width=0.94\linewidth]{Figure4.pdf}%
\caption{(color online) a) Temperature dependence of $\epsilon_{xx}$ (black dots). For illustration typical curves for degenerate and non-degenerate electrons are shown (see text).\\
b) Magnetic field dependence of $\epsilon_{xx}$ for selected temperatures between 4\,K and 100\,K. \\
c) Magnetic field dependence of $\epsilon_{xy}$ for selected temperatures between 4\,K and 100\,K.\\
}%
\label{fig:figure4}
\end{figure}
\noindent The temperature dependence of the thermoelectric tensor $\epsilon_{xx}$ at zero magnetic field (figure \ref{fig:figure4}a) is characterized by a negative quantitity (typical for electrons), which magnitude increases linearly with temperature for T\;$<$\;25\,K. At 25\,K, the thermoelectric tensor starts to increase rapidly up to a plateau, which extends between 50\,K and 90\,K. This increase originates from two simultaneous effects: the increase in mobility of the low mobility (degenerate) charge carriers and the thermally activated population of the high mobility (non-degenerate) band. Above 90\,K, the thermoelectric tensor starts to decrease, probably due to an increase in phonon scattering.\\
To describe the zero-field data, the contributions of the degenerate and the non-degenerate band have to be added up: $\epsilon = \epsilon_d + \epsilon_{nd}$. The degenerate, low mobility band can be described by the the well-known Mott-formula:\cite{Mott1971}
\begin{equation}
\epsilon_d=-\sigma_d\frac{\pi^2k_B}{3e}\frac{k_BT}{\epsilon_F}(1+p)
\label{eq:degenerate}
\end{equation}
with the Fermi-energy $\epsilon_F$ and the scatter parameter $p=(\partial ln\,\tau/\partial ln\,\varepsilon)|_{\varepsilon_F}$. The temperature dependence of the thermopower of a non-degenerate twodimensional electron-gas (as are the thermally excited high mobility charge carriers) is given by:\cite{Ziman1960}
\begin{equation}
\epsilon_{nd} = \sigma_{nd}\frac{k}{e}\left[(p + 2) + \frac{\epsilon_e}{kT} \right]
\label{eq:non-degenerate}
\end{equation}
with $\epsilon_e$ the Fermi-energy measured from the lower edge of the conduction band and $p=(\partial ln\,\tau/\partial ln\,\varepsilon)$. Thus, the thermopower of a non-degenerate electron gas is proportional to $1/T$.\\
To illustrate the behaviour of the temperature dependence of these bands, we plotted typical curves for the thermoelectric tensor of degenerate and non-degenerate electron systems $\epsilon_d$ and $\epsilon_{nd}$ in figure \ref{fig:figure4}a using the conductivities of the two bands as determined by the transport measurements.\\
As parameters for the curves we used $p=-0.5$, which is the theoretical value for hard sphere scattering,\cite{Gallagher1992} $\epsilon_e = 6.4$\,meV for the non-degenerate gas, as extracted from the thermal activation analysis, and $\epsilon_F=20$\,meV for the degenerate electron gas, in order to archieve the right magnitude compared to the data measured. We note that $\epsilon_F=20$\,meV is in the right order of magnitude for a 2D electron gas with the density measured. Fitting the temperature dependence is not possible because the scatter parameters and energies in equations (\ref{eq:degenerate}) and (\ref{eq:non-degenerate}), respectively, are not independent from each other. The actual values of the scatter parameter $p^{(i)}$ and the energies $\epsilon_e$ and $\epsilon_F$ can be determined by magnetic field dependent measurements, using an appropriate model for the magnetic field dependence of the two bands.\\
The magnetic field dependence of the diagonal ($\epsilon_{xx}/\epsilon_{xx}(B=0)$) and off-diagonal ($\epsilon_{xy}$) components of the thermoelectric tensor are shown in figure \ref{fig:figure4}b and c, respectively. The magnitude of the diagonal component of the thermoelectric tensor $\epsilon_{xx}/\epsilon_{xx}(B=0)$ decreases with field at all measured temperatures. However the details of their magnetic field dependence change drastically. At low temperatures, $\epsilon_{xx}/\epsilon_{xx}(B=0)$ decreases steeply at low fields, turning into saturation towards higher fields. The total decrease is thereby the largest at 20\,K. Starting from 30\,K, $\epsilon_{xx}/\epsilon_{xx}(B=0)$ is flat at low fields, getting steaper towards high fields and saturate again at highest fields. This saturation vanishes at 100\,K within the measured field range.\\
The off-diagonal component $\epsilon_{xy}$ increases linearly at low fields, changing with a kink to a lower, still linear slope at higher fields. The field where the slope changes, increases linearly with temperature from about 2\,T at 4\,K to 4.2\,T at 20\,K. Due to the sharpeness of the kink and his $B$-$T$-dependence, we attribute this kink to be a remnant of magnetic scattering. At 30\,K and 40\,K, a similar change in the slope exists, but at higher fields and smeared out over several Teslas. Therefore, we attribute this behaviour to the existence of two different types of charge carriers. At 65\,K and 100\,K, the transition either vanishes or is so much broadened and shifted to higher fields, that it is not visible anymore in our measurement.\\
To model the field dependence of the thermoelectric tensor of the first type charge carriers, we can use an expression for the diffusion thermoelectric tensor in the classical, degenerate limit given by:\cite{Fletcher1994}
\begin{equation}
\label{eq:thermotensor}
\begin{aligned}
\epsilon_{xx} &=
-\sigma_{xx}\frac{L_0eT}{\varepsilon_F}\left[1+p\frac{1-\mu^2B^2}{1+\mu^2B^2}
\right]\\
\epsilon_{xy} &=
-\sigma_{xy}\frac{L_0eT}{\varepsilon_F}\left[1+\frac{2p}{1+\mu^2B^2}\right]
\end{aligned}
\end{equation}
with $L_0=\pi^2k_B^2 /3e^2$ the Lorenz-number, $k_B$ the Boltzmann-constant, $\varepsilon_F$ the Fermi energy, $p=(\partial ln\,\tau/\partial ln\,\varepsilon)|_{\varepsilon_F}$ and $\tau$ the transport lifetime. The conductivities $\sigma_{xx}$ and $\sigma_{xy}$ are calculated with
\begin{equation}
\label{eq:sigma}
\begin{aligned}
\sigma_{xx} &= \frac{ne\mu}{1+\mu^2B^2}\\
\sigma_{xy} &= \frac{ne\mu^2B}{1+\mu^2B^2}
\end{aligned}
\end{equation}
using the charge-carrier densities $n$ and the mobilties $\mu$ obtained from the transport data (figure \ref{fig:figure2}).\\
The second type of charge carriers ($n_1$, $\mu_1$) is non-degenerate. Therefore it cannot be modeled with the same equations. To the best of our knowledge, a theoretical model for the magnetic field dependent thermopower of a non-degenerate electron gas is still missing. A development of such a model is beyond the scope of this work and remains as a future challenge.
\section{Discussion}
The main part of our data, i.e. for temperatures above 30\,K, seems to be well described by a two-charge carrier model, also for thermoelectric power. However, below 30\,K some peculiar features are observed: first, the magnitude of the Nernst-signal is enhanced compared to the Fermi-liquid picture. Second, the kink observed in the off-diagonal component of the thermoelectric tensor $\epsilon_{xy}$, which has a linear $B$-$T$-dependence.\\
A similar anomaly is observed in transport measurements at the same temperature, namely the observation of a negative magneto-resistance shown in figure \ref{fig:figure1}b and described elsewhere.\cite{Guduru2013a, Brinkman2007} There, the negative magneto-resistance is attributed to magnetic, Kondo-like scattering. Indeed, the thermoelectric power in Kondo-lattices shows similar behaviour as in the LaAlO$_3$/SrTiO$_3$ interface.\\
In three-dimensional Kondo-lattices with one type of charge carrier, the Seebeck coefficient $S_{xx}$ is decreasing strongly with magnetic field and saturating at high fields, and the Nernst coefficient $\nu=S_{xy}/B$ is large at small fields and decreasing to higher fields.\cite{Bel2004, Falkowski2011} In other words, the Nernst effect is increasing strongly at low magnet fields and saturates at high fields. In our measurements, we observe the same behaviour for low temperatures; a strong decrease in the Seebeck effect with saturation to high fields and a strong increase in Nernst at low field saturating at high fields (see figure \ref{fig:figure3}). Since we observe the same signatures as in Kondo-lattice materials, we suggest that a similar mechanism could play a role in our sample and we attribute the strong magnetic field dependence of the thermoelectric tensor to an additional magnetic scattering acting on the low mobility charge carriers ($n_1, \mu_1$) at low temperature.\\
A possible route to magnetic scattering can be explained by the polar-catastrophe scenario,\cite{Nakagawa2006} where charge is transfered to the interface due to the polarity of the LaAlO$_3$-layers. This additional charge can change the electric state of the non-magnetic Ti$^{4+}$-ions in SrTiO$_3$ to magnetic Ti$^{3+}$-ions.\cite{Pentcheva2007, Lee2013} These magnetic Ti$^{3+}$-ions then can act as scattering-partners for the electrons at the interface and the low mobility charge carrier are located close to the interface at the Ti$^{3+}$-ions. Over the location of the high mobility charge carriers we cannot give any conclusion.\\
We lack an appropriate model to determine the density of magnetic moments at the interface from our data. However, we can place an upper limit by assuming that they are created by the electrons arising from the polar-catastrophe. This would give an upper limit of $1/2$ a moment per unit cell or $3.4 \times 10^{-14}$\,cm$^{-2}$.
\section{Conclusion}
We have measured a complete set of transport and thermoelectric power data in a temperature range from 4\,K to 100\,K in fields up to 18\,T. We find two different electron-like charge carrier with different densities and mobilities: a degenerate band, with a low mobility and a high carrier density and a non-degenerate band with a higher mobility, which vanishes at low temperatures. The temperature dependence of the thermoelectric tensor can be described by this two-band picture, but for the magnetic field dependence an appropriate model for the non-degenerate band is still missing. We identify anomalies in the thermopower data, which cannot be readily explained by the two-band picture. We attribute them to an additional strongly magnetic field dependent scattering mechanism of the low mobility charge carriers located close to the Ti$^{3+}$-atoms at the interface.
\section{Acknowledgments}
This work has been performed at the HFML-RU/FOM member of the European Magnetic Field Laboratory (EMFL) and is part of the InterPhase research program of the Foundation for Fundamental Research on Matter (FOM, financially supported by the Netherlands Organization for Scientific Research (NWO)).
|
\section{#1}\setcounter{equation}{0}}
\def\vskip0cm\noindent{\vskip0cm\noindent}
\newcommand{\textcolor{red}}{\textcolor{red}}
\newcommand{\cvd}{$\hfill \sqcap \hskip-6.5pt \sqcup$}
\newcommand{\Arrowvert}{\Arrowvert}
\newcommand{\delta_{\lambda}}{\delta_{\lambda}}
\newcommand{\tilde{u}_{\lambda}}{\tilde{u}_{\lambda}}
\newcommand{\tilde{v}_{\lambda}}{\tilde{v}_{\lambda}}
\def{\rm I\mskip -3.5mu R}{{\rm I\mskip -3.5mu R}}
\def{\rm I\mskip -3.5mu N}{{\rm I\mskip -3.5mu N}}
\def{\rm I\mskip -3.5mu Z}{{\rm I\mskip -3.5mu Z}}
\defi^{*}{i^{*}}
\def\varepsilon{\varepsilon}
\def\mathcal{D}^{1,2}(\R^n){\mathcal{D}^{1,2}({\rm I\mskip -3.5mu R}^n)}
\def\rightarrow{{\rightarrow}}
\def\nabla{\nabla}
\def\delta{\delta}
\def\gamma{\gamma}
\def\Delta{\Delta}
\def{\lambda}{{\lambda}}
\def{\Lambda}{{\Lambda}}
\def{\alpha}{{\alpha}}
\def{\beta}{{\beta}}
\def{\sigma}{{\sigma}}
\def{\tau}{{\tau}}
\def\omega{\omega}
\def_{i,\l}{_{i,{\lambda}}}
\defp_{\e}{p_{\varepsilon}}
\def_{k,\l}{_{k,{\lambda}}}
\def_{0,\l}{_{0,{\lambda}}}
\def_{1,\l}{_{1,{\lambda}}}
\def_{2,\l}{_{2,{\lambda}}}
\def_{3,\l}{_{3,{\lambda}}}
\def_{4,\l}{_{4,{\lambda}}}
\defu_\l{u_{\lambda}}
\def\Phi_{\e,\l}{\Phi_{\varepsilon,{\lambda}}}
\defr_\l{r_{\lambda}}
\def\tilde y{\tilde y}
\def\tilde a{\tilde a}
\def\tilde r{\tilde r}
\def\gamma_\l{\gamma_{\lambda}}
\defx_\l{x_{\lambda}}
\defK(x){K(x)}
\defI_\l{I_{\lambda}}
\defu^\e{u^\varepsilon}
\defu^\e_{\d,y}{u^\varepsilon_{\delta,y}}
\def\phi ^\e_{\d,y}{\phi ^\varepsilon_{\delta,y}}
\defT ^\e_{\d,y}{T ^\varepsilon_{\delta,y}}
\defw_\l{w_{\lambda}}
\defz_\l{z_{\lambda}}
\def\xi_{\l}{\xi_{{\lambda}}}
\def\partial{\partial}
\def\frac {2n}{n+2}{\frac {2n}{n+2}}
\def\frac {2n}{n-2}{\frac {2n}{n-2}}
\def\frac {ns}{n+2s}{\frac {ns}{n+2s}}
\deff_{\e}{f_{\varepsilon}}
\deff_0{f_0}
\def\Omega{\Omega}
\def\longrightarrow{\longrightarrow}
\def\rightarrow{\rightarrow}
\def\mathcal{E}{\mathcal{E}}
\title[On a generalized Toda System]
{On a general $SU(3)$ Toda System}
\thanks{The first two authors are supported by PRIN-2009-WRJ3W7 grant and
the third author is supported by NSERC of Canada}
\author[Gladiali]{Francesca Gladiali}
\address{Francesca Gladiali, Dipartimento Polcoming, Universit\`a di Sassari - Via Piandanna 4, 07100 Sassari - Italy.}
\email{<EMAIL>}
\author[Grossi]{Massimo Grossi}
\address{Massimo Grossi, Dipartimento di Matematica, Universit\`a di Roma
La Sapienza, P.le A. Moro 2 - 00185 Roma- Italy.}
\email{<EMAIL>}
\author[Wei]{Juncheng Wei}
\address{Juncheng Wei, Department of Mathematics, University of British Columbia, Vancouver, BC V6T 1Z2, Canada, and Department of Mathematics, Chinese University of Hong Kong, Shatin, NT, Hong Kong}
\email{<EMAIL>}
\begin{abstract}
We study the following generalized $SU(3)$ Toda System
$$
\left\{\begin{array}{ll}
-\Delta u=2e^u+\mu e^v & \hbox{ in }{\rm I\mskip -3.5mu R}^2\\
-\Delta v=2e^v+\mu e^u & \hbox{ in }{\rm I\mskip -3.5mu R}^2\\
\int_{{\rm I\mskip -3.5mu R}^2}e^u<+\infty,\ \int_{{\rm I\mskip -3.5mu R}^2}e^v<+\infty
\end{array}\right.
$$
where $\mu>-2$. We prove the existence of radial solutions bifurcating from the radial solution $(\log \frac{64}{(2+\mu) (8+|x|^2)^2}, \log \frac{64}{ (2+\mu) (8+|x|^2)^2})$ at the values $\mu=\mu_n=2\frac{2-n-n^2}{2+n+n^2},\ n\in{\rm I\mskip -3.5mu N} $.
\end{abstract}
\maketitle
\sezione{Introduction and statement of main results}
Let us consider the following generalized Toda system,
\begin{equation}\label{1}
\left\{\begin{array}{ll}
-\Delta u=2e^u+\mu e^v & \hbox{ in }{\rm I\mskip -3.5mu R}^2\\
-\Delta v=2e^v+\mu e^u & \hbox{ in }{\rm I\mskip -3.5mu R}^2\\
\int_{{\rm I\mskip -3.5mu R}^2}e^u<+\infty,\ \int_{{\rm I\mskip -3.5mu R}^2}e^v<+\infty
\end{array}\right.
\end{equation}
with $\mu>-2$. This system is the limiting equation for the so-called Gudnason model in non-abelian Chern-Simons theory
\begin{equation}\label{1.2n}
\left\{\begin{array}{l} \Delta U= \alpha^2 (e^{U+V}+ e^{U-V})(e^{U+V}+ e^{U-V} -2) +\alpha \beta (e^{U+V}- e^{U-V})^2 + 4\pi \sum_{j=1}^n \delta_{p_j} (x),\\
\Delta V= \alpha \beta (e^{U+V}-e^{U-V})(e^{U+V} + e^{U-V} -2) +\beta^2 (e^{2U+2V}- e^{2U-2V}) +4\pi \sum_{j=1}^n \delta_{p_j} (x)
\end{array}
\right.
\end{equation}
where $x \in {\rm I\mskip -3.5mu R}^2, \alpha, \beta>0$ and $n$ is an nonnegative integer. In fact, as in \cite{ALW} and \cite{CI}, we rescale (\ref{1.2n}) as $ U(x)= \tilde{U} (\epsilon x) +2\log \epsilon, V(x)= \tilde{V} (\epsilon x)$. Letting $\epsilon \to 0$ we obtain the following system
\begin{equation}
\label{1.2m}
\left\{\begin{array}{l}
-\Delta \tilde{U}=2 \alpha^2 (e^{\tilde{U}+\tilde{V}}+ e^{\tilde{U}-\tilde{V}})+ 4n\pi \delta_0 \\
-\Delta \tilde{V}=2 \alpha \beta (e^{\tilde{U}+\tilde{V}}-e^{\tilde{U}-\tilde{V}})+ 4n \pi \delta_0
\end{array}
\right.
\end{equation}
If $n=0$, then by suitable linear transformation (\ref{1.2m}) is equivalent to (\ref{1}) with $\mu= 2 \frac{\alpha^2-\alpha\beta}{ \alpha^2+\alpha \beta}$. Clearly it holds that $\mu>-2$.
We refer to \cite{HLTY} for the background on Gudnason model and the references therein. \\
This problem is the natural extension, in the case of systems, of the known Liouville equation
\begin{equation}\label{i1}
\left\{\begin{array}{ll}
-\Delta u=e^u & \hbox{ in }{\rm I\mskip -3.5mu R}^2\\
\int_{{\rm I\mskip -3.5mu R}^2}e^u<+\infty.
\end{array}\right.
\end{equation}
All solutions of \eqref{i1} have been completely classified in \cite{CL1}.\\
Contrary to the case of \eqref{i1}, it is not available a classification result for solutions of the problem \eqref{1} and the structure of the space of the solutions is far from being fully understood. \\
{ Note that if $u=v$ then system \eqref{1} reduces to the equation
\begin{equation}\nonumbe
-\Delta u=(2+\mu)e^u \quad\hbox{ in }{\rm I\mskip -3.5mu R}^2
\end{equation}
which admits, according to \cite{CL1}, the $3$-parameter family of solutions
\begin{equation}\label{3}
U_{\mu,\delta,y}(x)=\log\frac{64\delta}{(2+\mu)\left(8\delta+|x-y|^2\right)^2},
\end{equation}
with $\delta>0$ and $y\in{\rm I\mskip -3.5mu R}^2$ and so \eqref{1} always admits the solution $(u,v)=(U_{\mu,\delta,y}, U_{\mu,\delta,y})$ for any $\mu>-2$, $\delta>0$ and $y\in{\rm I\mskip -3.5mu R}^2$.}\\
A first general result to \eqref{1} is given in \cite{CK} where the authors consider the problem
\begin{equation}\label{i2}
\left\{\begin{array}{ll}
-\Delta u=\gamma_{11}e^u+\gamma_{12}e^v & \hbox{ in }{\rm I\mskip -3.5mu R}^2\\
-\Delta v=\gamma_{21}e^u+\gamma_{22}e^v & \hbox{ in }{\rm I\mskip -3.5mu R}^2\\
\int_{{\rm I\mskip -3.5mu R}^2}e^u<+\infty,\ \int_{{\rm I\mskip -3.5mu R}^2}e^v<+\infty.
\end{array}\right.
\end{equation}
Under some assumption on the coefficients $\gamma_{ij}$ including $\gamma_{ij}\ge0$, $$\gamma_{11}+\gamma_{12}=1$$ and $$\gamma_{21}+\gamma_{22}=1$$ the authors prove that all solutions to \eqref{i2} are radially symmetric and of the type as in \eqref{3}.\\
Always in the case where $\gamma_{ij}\ge0$, in \cite{CSW} some milder assumption on $\gamma_{ij}$ are given in order to obtain the radial symmetry of the solutions. A complete classification of radial solutions to (\ref{1}) in the case of $\mu>0$ can be found in \cite{LZ1, LZ2}.\\
On the other hand, if some of $\gamma_{ij}$ are negative, we will see that the system can have {both radial and }nonradial solutions.\\
In this setting, the most studied case is that of the Cartan matrix, where $\gamma_{ij}$, $i,j=1,2$ are given by
\begin{equation}\label{i3}
\begin{pmatrix}
2&-1\\
-1&2
\end{pmatrix}.
\end{equation}
In this case Jost and Wang (see \cite{JW2}) provided the following beautiful classification result,\\
{\bf Theorem} (see \cite{JW2}).\\
{\em All solutions of the problem
\begin{equation}\label{i4}
\left\{\begin{array}{ll}
-\Delta u=2e^u-e^v & \hbox{ in }{\rm I\mskip -3.5mu R}^2\\
-\Delta v=2e^v-e^u & \hbox{ in }{\rm I\mskip -3.5mu R}^2\\
\int_{{\rm I\mskip -3.5mu R}^2}e^u<+\infty,\ \int_{{\rm I\mskip -3.5mu R}^2}e^v<+\infty
\end{array}\right.
\end{equation}
are given by (in complex notation)
\begin{equation}\label{i5}
\left\{\begin{array}{ll}
u(z)=\log\frac{4\left(a_1^2a_2^2+a_1^2|2z+c|^2+a_2^2|z^2+2bz+bc-d|^2\right)}
{\left(a_1^2+a_2^2|z+b|^2+|z^2+cz+d|^2\right)^2}\\
v(z)=\log\frac{16a_1^2a_2^2\left(a_1^2+a_2^2|z+b|^2+|z^2+cz+d|^2\right)}
{\left(a_1^2a_2^2+a_1^2|2z+c|^2+a_2^2|z^2+2bz+bc-d|^2\right)^2}
\end{array}\right.
\end{equation}
for some real constants $a_1,a_2>0$ and $b,c,d\in\mathbb{C}$. (See also \cite{WZZ}.) }\\[.4cm]
\noindent This shows that when some of $\gamma_{ij}'s$ are negative system \eqref{i2} can have {both radial and} nonradial solutions.\\
This result was extended when in \eqref{i4} appear singular sources (see\cite{LWY}).\\
The purpose of this paper is to investigate more general cases than those of the Cartan matrix and to understand for which matrices $\gamma_{ij}$ can be expected results of existence of solutions. So we will consider system \eqref{1} with $\mu<0$.\\
Note that variational methods seem difficult to apply because of the lack of compactness of the domain. Even the choice of a suitable functional space is quite unclear. For this reason we use the {\em bifurcation theory}.\\
{Let us denote by
\begin{equation}\label{4}
U_\mu(x)=U_{\mu,1,O}(x)=\log\frac{64}{(2+\mu)\left(8+|x|^2\right)^2}.
\end{equation}}
A first step in our proof is the {\em linearization }of the system \eqref{1} at the trivial solution $u=v
U_\mu$. We have the following
\begin{lemma}\label{6}
Let us suppose that $(w_1,w_2)\ne(0,0)$ is a solution to
\begin{equation}\label{5}
\left\{\begin{array}{ll}
-\Delta w_1=2e^{U_\mu} w_1+\mu e^{U_\mu} w_2 & \hbox{ in }{\rm I\mskip -3.5mu R}^2\\
-\Delta w_2=2e^{U_\mu} w_2+\mu e^{U_\mu} w_1 & \hbox{ in }{\rm I\mskip -3.5mu R}^2
\end{array}\right.
\end{equation}
Set
\begin{equation}\label{a1}
\mu_n=2\frac{2-n-n^2}{2+n+n^2}\quad n\in{\rm I\mskip -3.5mu N} \cup\{0\},\end{equation}
We have the following alternative,\\
$i)$ If $\mu>-2$ but $\mu\ne\mu_n$ then all the bounded solutions to \eqref{5} are given, in radial coordinates, by
\begin{equation}\label{a2}
(w_1,w_2)=\left(a_1\frac{rcos\theta}{8+r^2}+a_2\frac{r\sin\theta}{8+r^2}+a_3\frac{8-r^2}{8+r^2},a_1\frac{rcos\theta}{8+r^2}+a_2\frac{r\sin\theta}{8+r^2}+a_3\frac{8-r^2}{8+r^2}\right)
\end{equation}
for some real constants $a_1,a_2,a_3$.\\
$ii)$ if $\mu=\mu_n$ then all the bounded solutions to \eqref{5} are given, in radial coordinates, by
\begin{eqnarray}\label{a3}
&&(w_1,w_2)=\left(
\sum\limits_{m=0}^n \left[A_m\cos(m\theta)+B_m\sin(m\theta)\right]P^m_n\left(\frac{8-r^2}{8+r^2}\right)+a_1\frac{rcos\theta}{8+r^2}\right.\nonumber\\
&&\,\,\,\,\,\,\,+a_2\frac{r\sin\theta}{8+r^2}+a_3\frac{8-r^2}{8+r^2} \,\,,\,\
-\sum\limits_{m=0}^n \left[A_m\cos(m\theta)+B_m\sin(m\theta)\right]P^m_n\left(\frac{8-r^2}{8+r^2}\right)\nonumber\\
&&\left.\,\,\,\,\,\,\,+a_1\frac{rcos\theta}{8+r^2}+a_2\frac{r\sin\theta}{8+r^2}+a_3\frac{8-r^2}{8+r^2}\right)
\end{eqnarray}
for some real constants $a_1,a_2,a_3$, $A_m$, $m=0,\dots,n$ and $B_m$, $m=1,\dots,n$, where
\begin{equation}\label{a4}
P_n^m(r)=(-1)^m(1-r^2)^\frac m2 \frac {d^m}{dr^m}P_n(r)
\end{equation}
and $P_n(r)$ are the classical Legendre polynomials defined as
\begin{equation}\label{a5}
P_n(r)=\frac1{2^nn!} \frac {d^n}{dr^n}(r^2-1)^n.
\end{equation}
\end{lemma}
\noindent From the previous lemma we see that the linearized operator \eqref{5} has two types of degeneracies. The first one, which holds for every $\mu$, is due to the invariance of the problem \eqref{1} for dilations and translations, while the second one appears only at the values $\mu_n$ defined in \eqref{a1}.\\
In Section \ref{s3} we will show that this second type of degeneracy is due to the existence of other solutions of \eqref{1} which bifurcate from the solution $(U_{\mu},U_{\mu})$ for $\mu=\mu_n$ and $n\in {\rm I\mskip -3.5mu N}$.\\
This is our main result,
\begin{theorem}\label{i6}
Let $n\in{\rm I\mskip -3.5mu N}$ and $\mu_n$ be as defined in \eqref{a1}. The points $(\mu_n,U_{\mu_{n}},U_{\mu_{n}})$ are radial bifurcation points for the curve of solutions $(\mu,U_{\mu},U_{\mu})$, i.e. for $\varepsilon>0$ small enough there exist radial solutions $(u_\varepsilon,v_\varepsilon)$ to \eqref{1} which satisfy
\begin{equation}\label{a6}
\left\{\begin{array}{ll}
u_\varepsilon(|x|)=U_{\mu_\varepsilon}+\varepsilon P_n\left(\frac{8-|x|^2}{8+|x|^2}\right)+\varepsilon(z_{1,\varepsilon}+z_{2,\varepsilon})\\
v_\varepsilon(|x|)=U_{\mu\varepsilon}+\varepsilon P_n\left(\frac{8-|x|^2}{8+|x|^2}\right)+\varepsilon(z_{1,\varepsilon}-z_{2,\varepsilon})
\end{array}\right.
\end{equation}
with $\mu_\varepsilon$ such that $\mu_0=\mu_n$ and some functions $(z_{1,\varepsilon},z_{2,\varepsilon})$ uniformly bounded in $L^{\infty}({\rm I\mskip -3.5mu R}^2)$, with $z_{i,0}=0$ for $i=1,2$.
\end{theorem}
\begin{remark}
When $n\neq 1$ since $P_n(t)\neq t$ then $u_\varepsilon(|x|)\neq U_{\mu,\delta,O}$ and this provides a second radial solution to the problem.
\end{remark}
\begin{remark}
If $n=1$ in Theorem \ref{i6} then $\mu_1=0$ and our system decouples in two Liouville equations. On the other hand, in this case, $P_1(t)=t$ and the solutions in \eqref{a6} coincide with $U_{\mu,\delta,O}$. If $n=2$ then $\mu_2=-1$ and we fall in case of the Cartan matrix. Of course in these cases we get much weaker information of the complete classification results of \cite{CL1} and \cite{JW2}. In fact our results are significative for $n>2$.
\end{remark}
\noindent From Lemma \ref{6} we conjecture that our system admits $2n+4$ families of solutions bifurcating from $\mu=\mu_n$ as in the case $n=2$. This conjecture is supported by the following perturbation result from $\mu_2=-1$:
\begin{theorem}
\label{thm2}
Let $ (u_{a_1, a_2}, v_{a_1, a_2})$ be the radial solution of the Toda system (\ref{i4}) given by (\ref{i5}) with $b=0, c=0, d=0$. Then there exists $\epsilon >0$ such that for any $ \mu \in (-1-\epsilon, -1+\epsilon)$ there exists a radial solution $ (u_{\epsilon}, v_{\epsilon})$ to \eqref{1} which satisfies
\begin{equation}
u_\epsilon= u_{a_1, a_2} + \epsilon z_1, \ v_\epsilon= v_{a_1, a_2} +\epsilon z_2
\end{equation}
with some radial smooth and logrithmatically growing functions $ z_1$ and $z_2$.
\end{theorem}
\noindent
\begin{remark}
Theorem \ref{thm2} provides two families of radial solutions to \eqref{5} such that $ u \not = v$. Even when $a_1=a_2$ the two types of solutions constructed in Theorems \ref{i6} and \ref{thm2} are different. An open question is whether or not non-radial solutions can be perturbed.
\end{remark}
\noindent The proof of Theorem \ref{i6} uses the bifurcation result of \cite{CRJFA} (see Section \ref{s3} for the exact statement). One of the main hypothesis of this result is that the kernel of the linearized operator has to be one-dimensional. From Lemma \ref{6} we know that the linearized operator has instead a very rich kernel. Then, to overcame this problem, we consider only the case of radial solutions. In this way the kernel is reduced to be two-dimensional at the values $\mu=\mu_n$ defined in \eqref{a1}. \\
As said before, the degeneracy responsable of the bifurcation is that generated by the radial eigenfunctions $P_n\left(\frac{8-|x|^2}{8+|x|^2}\right)$ and by the dilation invariance of the system. We solve this second problem projecting on a suitable subspace which excludes the presence of the function $\frac{8-|x|^2}{8+|x|^2}$ (the eigenfunction given by the scaling invariance of the problem).\\
There is another reason why we consider only the radial case: in fact the presence of the radial solution $ P_n\left(\frac{8-|x|^2}{8+|x|^2}\right)$ in the kernel of the linearized operator, for $\mu=\mu_n$, does not allow us to say that the solutions that we could find using other topological methods (in a nonradial setting) are nonradial.
We want to emphasize that this is a bifurcation for an operator which is strongly degenerate. The case of the bifurcation in a degenerate setting (radial degeneracy) has been considered in \cite{GGN}. In that paper the authors rule out the degeneracy of the corresponding operator considering a nondegenerate approximating problem in expanding balls and showing that the bifurcating branches of solutions converge to branches of bifurcating solutions of the original problem. Here, however, we prefer to use the projection.\\
Of course when we apply the bifurcation theory to find a solution, a Lagrange multiplier appears in the system (see Theorem \ref{3.10}).\\
The final step of the proof is to show that the Lagrange multiplier is $zero$. A crucial point is to prove the following {\em mass quantizaton} of the solutions $u_\varepsilon$, $v_\varepsilon$ (given in \eqref{a6}), i.e. that it holds
$$\int_{{\rm I\mskip -3.5mu R}^2}e^{u_\varepsilon}=\int_{{\rm I\mskip -3.5mu R}^2}e^{v_\varepsilon}=\frac{8\pi}{2+\mu_\varepsilon}.$$
The choice of the functional spaces plays a very important role here; we will consider some suitable weighted subspaces of $W^{2,2}\left({\rm I\mskip -3.5mu R}^2\right)$ already used in \cite{CI}. Moreover, to overcame some problems given by the presence of the Lagrange multiplier, we look for solutions which belong to $L^{\infty}({\rm I\mskip -3.5mu R}^2)$ also. (This choice is suggested by the form of the eigenfunctions given in Lemma \ref{6}). The interested reader can find all the definitions of the spaces and their main properties in Section \ref{s3}. \\
The paper is organized as follows: In Section \ref{s2} we prove Lemma \ref{6}. In Section \ref{s3} we solve system \eqref{1} up to a Lagrange multiplier and in Section \ref{s4} we prove Theorem \ref{i6} showing that the Lagrange multiplier is $zero$. Finally in Section \ref{s5} we use a perturbation approach to prove Theorem \ref{thm2}.
\sezione{The linearized operator}\label{s2}
\noindent In this Section we consider the linearized system \eqref{5} and we prove Lemma \ref{6}.
\begin{proof}[Proof of Lemma \ref{6}]
Let $w_1,w_2$ be solutions to \eqref{5} and set $z=w_1-w_2$. We have the following alternatives:\\
$i) z\equiv0$. In this case \eqref{5} reduces to the equation
\begin{equation}\label{7}
-\Delta w_1=(2+\mu)e^{U_\mu}w_1
\end{equation}
which admits, in radial coordinates, the solutions
\begin{equation}\label{8}
w_1=w_2=\left.\sum\limits_{i=1}^2a_i\frac{\partial U_{\mu,\delta,y}}{\partial y_i}+a_3\frac{\partial U_{\mu,\delta,y}}{\partial\delta}\right|_{\delta=1,y=O}
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!=a_1\frac{rcos\theta}{8+r^2}+a_2\frac{r\sin\theta}{8+r^2}+a_3\frac{8-r^2}{8+r^2}
\end{equation}
for some real constants $a_1,a_2,a_3$.\\
$ii) z\not\equiv0$. Normalizing $z$ with its $||.||_\infty$ we get the $z$ is a solution to
\begin{equation}\label{9}
\left\{\begin{array}{ll}
-\Delta z=(2-\mu)e^{U_\mu}z\hbox{ in }{\rm I\mskip -3.5mu R}^2 \\
||z||_\infty=1
\end{array}\right.
\end{equation}
which is equivalent to
\begin{equation}\label{10}
-\Delta z=\frac{2-\mu}{2+\mu}\frac{64}{\left(8+|x|^2\right)^2} z\hbox{ in }{\rm I\mskip -3.5mu R}^2
\end{equation}
The eigenvalues of \eqref{10} (and the relative $L^{\infty}$-eigenfunctions) were computed in \cite{GG}, Theorem 11.1. So we know that \eqref{10} has a nontrivial bounded solution if and only if
\begin{equation}\label{11}
\frac{2-\mu}{2+\mu}=\frac{n(n+1)}2\quad n\in{\rm I\mskip -3.5mu N}\cup\{0\},
\end{equation}
and
\begin{equation}\label{12}
z(r,\theta)=\sum\limits_{m=0}^n \left[A_m\cos(m\theta)+B_m\sin(m\theta)\right]P^m_n\left(\frac{8-r^2}{8+r^2}\right),
\end{equation}
where $P^m_n$ are the associated Legendre polynomials as defined in \eqref{a4}-\eqref{a5}. Finally each eigenvalue $c_n=\frac{n(n+1)}2$ has multiplicity $2n + 1$. From \eqref{11} we deduce that
\begin{equation}\label{13}
\mu=\mu_n=2\frac{2-n-n^2}{2+n+n^2}\quad n\in{\rm I\mskip -3.5mu N}\cup\{0\}.
\end{equation}
Let us compute $w_1$ and $w_2$. Summing the equations in \eqref{5} we get that
\begin{equation}\label{14}
-\Delta(w_1+w_2)=(2+\mu)e^{U_\mu}(w_1+w_2)
\end{equation}
which implies
\begin{equation}\label{15}
w_1+w_2=\left.\sum\limits_{i=1}^2a_i\frac{\partial U_{\mu,\delta,y}}{\partial y_i} +a_3\frac{\partial U_{\mu,\delta,y}}{\partial\delta}\right|_{\delta=1,y=O}
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!=a_1\frac{rcos\theta}{8+r^2}+a_2\frac{r\sin\theta}{8+r^2}+a_3\frac{8-r^2}{8+r^2}.
\end{equation}
Since by \eqref{12} we derive that
\begin{equation}\label{16}
w_1-w_2= \sum\limits_{m=0}^n \left[A_m\cos(m\theta)+B_m\sin(m\theta)\right]P^m_n\left(\frac{8-r^2}{8+r^2}\right),
\end{equation}
by \eqref{15} and \eqref{16} we get \eqref{a3}.
\end{proof}
\begin{remark}
{As said in the proof of Lemma \ref{6} the linearized operator has, for every $\mu\neq \mu_n$, a $3-$dimensional kernel. For $\mu=\mu_n$ we have to add the $2n+1$ solutions of equation \eqref{10}
whose one radial and $2n$ nonradial. This implies that the Morse index of the solution $(U_\mu,U_\mu)$, in a suitable space, changes in $\mu=\mu_n$ by $2n+1$. The fact that the change of the Morse index in $\mu=\mu_n$ is odd is a good indication to obtain the bifurcation.}\\
\end{remark}
\sezione{An existence result}\label{s3}
\noindent From here on, all the functions that we consider will be radial, i.e. such that $w(x)=w(|x|)$ for any $x\in {\rm I\mskip -3.5mu R}^2$.\\
Applying Lemma \ref{6} we have that \eqref{5} has, for every $\mu>-2$ and $\mu\ne\mu_n$, the solution (up to a constant)
\begin{equation}\label{n1}
(w_1,w_2)=\left(\frac{8-|x|^2}{8+|x|^2},\frac{8-|x|^2}{8+|x|^2}\right),
\end{equation}
and for $\mu=\mu_n$ and $n\in {\rm I\mskip -3.5mu N}$,
\begin{equation}\label{n2}
(w_1,w_2)=\left(A P_n\left(\frac{8-|x|^2}{8+|x|^2}\right)+a\frac{8-|x|^2}{8+
|x|^2},-A P_n\left(\frac{8-|x|^2}{8+|x|^2}\right) +a\frac{8-|x|^2}{8+|x|^2},\right)
\end{equation}
for some real constants $a$ and $A$, where $P_n=P_n^0$ is the $n-$th Legendre Polynomial, as defined in \eqref{a5}.\\
To obtain the bifurcation result we will use the transversality condition of Crandall and Rabinowitz (see Theorem 1.7 in \cite{CRJFA}). Although this result is very well known, we report it for completeness. \\
Before stating it, let us introduce some notations. Let $G=G(t,x):{\rm I\mskip -3.5mu R}\times X\rightarrow Y$ be an operator from the Banach spaces $X$ and $Y$. \\
We will denote by $G'_x$ the Fr\'echet derivative of $G$ with respect to $x$ and by $G''_{t,x}$ the Fr\'echet derivative of $G'_x$ with respect to $t$.
\begin{theorem}[\cite{CRJFA}, Theorem 1.7]\label{CR}
Let $X,Y$ be Banach spaces, $V$ a neighborhood of $0$
in $X$ and
$F:(-2,2)\times V\to Y$
have the properties:
\begin{itemize}
\item[a)]$F(t,0)=0$ for $|t|<2$,
\item[b)]The partial derivatives $F'_t$, $F'_x$ and $F''_{t,x}$ exist and are continuous,
\item[c)]$Ker(F'_x(0,0))$ and $Y\setminus R(F'_x(0,0))$ are one-dimensional,
\item[d)]$F''_{t,x}(0,0)w_0\notin R(F'_x(0,0))$, where $Ker(F'_x(0,0))=span\{w_0\}$.
\end{itemize}
If $Z$ is any complement of $Ker(F'_x(0,0))$ in $X$, then there is a neighborhood $U$
of $(0,0)$ in ${\rm I\mskip -3.5mu R}\times X$, an interval $(-\varepsilon_0, \varepsilon_0)$, and continuous functions
$\eta:(-\varepsilon_0, \varepsilon_0)\to {\rm I\mskip -3.5mu R}$, $z:(-\varepsilon_0, \varepsilon_0)\to Z$ such that $\eta(0)=0$, $z(0)= 0$ and
$$F^{-1}(0)\cap U=\{(\eta(\varepsilon),\varepsilon w_0+\varepsilon z(\varepsilon))\, :\,|\varepsilon|<\varepsilon_0\}\cup \{(t,0)\,:\,(t,0)\in U\}.$$
\end{theorem}
\noindent To apply the bifurcation result we have to {\em rule out} the degeneracy due to the invariance for dilations of problem \eqref{1}. To do this we ask that
the functions are orthogonal, in some sense, to the eigenfunction $(\frac{8-|x|^2}{8+|x|^2},\frac{8-|x|^2}{8+|x|^2})$. \\[.1cm]
As we did in the study of the linearized operator in Section 2, we consider the sum and difference of the solutions $u$ and $v$ of \eqref{1}. Let
\begin{equation}\label{n3}
\left\{\begin{array}{ll}
\tilde \phi=u+v\\
\psi=u-v.
\end{array}\right.
\end{equation}
Then problem \eqref{1} is equivalent to
\begin{equation}\label{3.1}
\left\{\begin{array}{ll}
-\Delta \tilde\phi=(2+\mu) \left(e^{\frac{\tilde \phi+\psi}2}+ e^{\frac{\tilde \phi-\psi}2}\right) & \hbox{ in }{\rm I\mskip -3.5mu R}^2\\
-\Delta \psi=(2-\mu) \left(e^{\frac{\tilde \phi+\psi}2}- e^{\frac{\tilde \phi-\psi}2}\right) & \hbox{ in }{\rm I\mskip -3.5mu R}^2\\
\int_{{\rm I\mskip -3.5mu R}^2}e^{\frac{\tilde \phi+\psi}2}<+\infty,\ \int_{{\rm I\mskip -3.5mu R}^2}e^{\frac{\tilde \phi-\psi}2}<+\infty
\end{array}\right.
\end{equation}
and $(\mu,2U_{\mu},0)$ is a curve of radial solutions of \eqref{3.1}.\\
First we shift these solutions in the origin, i.e. we let
\begin{equation}\label{n4}
\tilde\phi=2U_{\mu}+\phi.
\end{equation}
Then, finding solutions for \eqref{3.1} is equivalent to find solutions for the problem
\begin{equation}\label{3.2}
\left\{\begin{array}{ll}
-\Delta\phi=(2+\mu)e^{U_{\mu}} \left(e^{\frac{ \phi+\psi}2}+ e^{\frac{ \phi-\psi}2}-2\right) & \hbox{ in }{\rm I\mskip -3.5mu R}^2\\
-\Delta \psi=(2-\mu)e^{U_{\mu}} \left(e^{\frac{ \phi+\psi}2}- e^{\frac{ \phi-\psi}2}\right) & \hbox{ in }{\rm I\mskip -3.5mu R}^2\\
\int_{{\rm I\mskip -3.5mu R}^2}e^{U_{\mu}}e^{\frac{ \phi+\psi}2}<+\infty,\ \int_{{\rm I\mskip -3.5mu R}^2}e^{U_{\mu}}e^{\frac{ \phi-\psi}2}<+\infty
\end{array}\right.
\end{equation}
and $(\mu,0,0)$ is a solution of \eqref{3.2} for every value of $\mu$. Moreover $u=U_\mu+\frac 12(\phi+\psi)$ and $v=U_\mu+\frac 12(\phi-\psi)$ are solutions of \eqref{1}.\\
Let us introduce the spaces in which we apply Theorem \ref{CR}. Let ${\alpha}\in(0,1)$ be fixed, and
\begin{equation}\label{3.3}
X_{{\alpha}}=\left\{u\in L^{2}_{loc}({\rm I\mskip -3.5mu R}^2)\,|\, \int_{{\rm I\mskip -3.5mu R}^2}(1+|x|^{2+{\alpha}})u^2\, dx<+\infty\right\},
\end{equation}
equipped with the inner product $(u,v)_{X_{\alpha}}=\int_{{\rm I\mskip -3.5mu R}^2}(1+|x|^{2+{\alpha}})uv\, dx$,
\begin{equation}\label{3.4}
Y_{{\alpha}}=\left\{u\in W^{2,2}_{loc}({\rm I\mskip -3.5mu R}^2)\,|\, \Arrowvert \Delta u\Arrowvert^2_{X_{\alpha}}+\Arrowvert\frac u{(1+|x|^{1+\frac\a2})}\Arrowvert^2_{L^2({\rm I\mskip -3.5mu R}^2)}<+\infty\right\},
\end{equation}
equipped with the inner product
$$(u,v)_{Y_{\alpha}}=(\Delta u,\Delta v)_{X_{\alpha}}+\int_{{\rm I\mskip -3.5mu R}^2}\frac {uv}{(1+|x|^{2+{\alpha}})}\, dx.$$
The spaces $X_{\alpha}$ and $Y_{\alpha}$ were introduced by Chae and Imanuvilov in \cite{CI} to find Non-Topological Multivortex Solutions for a nonlinear problem with exponential nonlinearity in ${\rm I\mskip -3.5mu R}^2$.\\
We recall some properties of the spaces $X_{\alpha}$ and $Y_{\alpha}$ that we need in the sequel.
\begin{proposition}[\cite{CI}]\label{CIa-properties}
Let ${\alpha}\in(0,1)$, then
\begin{itemize}
\item[i)] $Y_{{\alpha}}\subset C^0_{loc}({\rm I\mskip -3.5mu R}^2)$ and $X_{\alpha}\subset L^1({\rm I\mskip -3.5mu R}^2)$.
\item[ii)] There exists $C_1>0$ such that for any $g\in Y_{\alpha}$
\begin{equation}\label{3.6}
|g(x)|\leq C_1 \Arrowvert g\Arrowvert_{Y_{\alpha}}\left( \log^+|x|+1\right)\hbox{ for any }x\in {\rm I\mskip -3.5mu R}^2
\end{equation}
where $log^+|x|=\max\{0,\log|x|\}$.
\item[iii)] If $g\in Y_{{\alpha}}$ then $\Delta g\in X_{{\alpha}}$.
\item[iv)] If $\Delta g=0$ we get that $g=c$ for some $c\in{\rm I\mskip -3.5mu R}$.
\end{itemize}
\end{proposition}
Let us introduce the function
\begin{equation}\label{3.5}
M(f)(x):=-\frac 1{2\pi}\int_{{\rm I\mskip -3.5mu R}^2}\log|x-y|f(y)\,dy.
\end{equation}
Let us recall some properties of $M$,
\begin{proposition}[\cite{CI}]\label{CIb-properties}
We have that,
\begin{itemize}
\item[i)] $M:X_{{\alpha}}\rightarrow Y_{{\alpha}}$.
\item[ii)] $-\Delta M(f)=f$ in ${\rm I\mskip -3.5mu R}^2$.
\item[iii)] $M\in L(X_{{\alpha}},Y_{{\alpha}})$ and
\begin{equation}\label{3.5-bis}
|M(g)(x)|\leq C_2 \Arrowvert g\Arrowvert_{X_{\alpha}}\left( \log^+|x|+1\right)\hbox{ for any }x\in {\rm I\mskip -3.5mu R}^2,
\end{equation}
\end{itemize}
where $C_2$ depends only on ${\alpha}$.
\end{proposition}
\noindent
Since all eigenfunctions in \eqref{n2} are bounded in the $L^{\infty}$-norm, it seems appropriate to use as a starting space the subset of functions of $Y_{\alpha}$
which are bounded in the $L^{\infty}$-norm. (Recall that functions in $Y_{\alpha}$ may have logaritmic growth at infinity.)
Then we define
\begin{equation}\label{nuovo-spazio-Y}
\widetilde Y_{\alpha}:=Y_{\alpha}\cap L^{\infty}({\rm I\mskip -3.5mu R}^2).
\end{equation}
$\widetilde Y_{\alpha}$ is a Banach space with the norm
$$\Arrowvert g\Arrowvert_{\widetilde Y_{\alpha}}=\max \{ \Arrowvert g\Arrowvert_{Y_{\alpha}}\, ,\,\Arrowvert g\Arrowvert_{\infty}\}$$
where $\Arrowvert g\Arrowvert_{\infty}$ denotes the usual norm in $L^{\infty}({\rm I\mskip -3.5mu R}^2)$.
\noindent Using the operator $M$ defined in \eqref{3.5},
we look for solutions of \eqref{3.2} as zeros of the operator $\widetilde T:(-2,2)\times \widetilde Y_{\alpha}\times\widetilde Y_{\alpha}$
\begin{equation}\label{3.7}
\widetilde T(\mu,\phi,\psi)=\left(
\begin{array}{l}
\phi-M\left((2+\mu)e^{U_{\mu}} \left(e^{\frac{ \phi+\psi}2}+ e^{\frac{ \phi-\psi}2}-2\right)\right)\\
\psi-M\left((2-\mu)e^{U_{\mu}} \left(e^{\frac{ \phi+\psi}2}- e^{\frac{ \phi-\psi}2}\right)\right)
\end{array}
\right)
\end{equation}
Since $\phi$ and $\psi$ are bounded and $U_{\mu}\in X_{\alpha}$, by
Propositions \ref{CIa-properties} and \ref{CIb-properties}, the operator $\widetilde T(\mu,\phi,\psi)$ maps $(-2,2)\times \widetilde Y_{\alpha} \times \widetilde Y_{\alpha} \mapsto Y_{\alpha}\times Y_{\alpha}$.
Hence the zeros of $\widetilde T(\mu,\phi,\psi)$
are bounded solutions of \eqref{3.2}.\\
Observe that $(\mu,0,0)$ is a curve of solutions of \eqref{3.2} in this space, i.e. $\widetilde T(\mu,0,0)=0$ for $|\mu|<2$.\\
Let us consider the linerized operator at $(\mu_n,0,0)$, i.e.
\begin{equation}\label{3.8}
\widetilde T'_{(\phi,\psi)}(\mu_n,0,0)\left(\begin{array}{l}
w_1\\
w_2
\end{array}\right)=\left(
\begin{array}{l}
w_1-M\left( (2+\mu_n)e^{U_{\mu_n}}w_1\right)\\
w_2-M\left((2-\mu_n)e^{U_{\mu_n}}w_2\right).
\end{array}
\right)
\end{equation}
As shown in the proof of Proposition \ref{6}, the solutions $w_1,w_2\in \widetilde Y_{{\alpha}}$ of
$$\widetilde T'_{(\phi,\psi)}(\mu_n,0,0)\left(\begin{array}{l}
w_1\\
w_2
\end{array}\right)=\left(\begin{array}{l} 0\\0\end{array}\right)$$
are given by
$$\left(\begin{array}{l}
w_1\\
w_2
\end{array}\right)=\left(\begin{array}{c}
a\frac{8-r^2}{8+r^2} \\
A P_n\left(a \frac{8-r^2}{8+r^2}\right)\end{array}\right)$$
for $a,A\in {\rm I\mskip -3.5mu R}$.\\
Here we want to find a new radial solution of \eqref{3.2} which is due to the presence of the second component $w_2$ in the linearized equation (which appears only at the values $\mu=\mu_n$).
We are going to verify the assumption of the bifurcation result of Crandall and Rabinowitz, (Theorem \ref{CR}). To this end we define
\begin{equation}\label{K}
K:=\{g\in Y_{\alpha}\, :\, \int_{{\rm I\mskip -3.5mu R}^2}(\Delta g) \frac{8-|x|^2}{8+|x|^2}\, dx=0\}.
\end{equation}
From the definition of $Y_{\alpha}$ we get that $K$ is a linear closed subspace of $Y_{{\alpha}}$. We let $P_{K}$ be the projection of $Y_{{\alpha}}$ on $K$.
Finally we consider the operator $T(\mu,\phi,\psi)$ defined by
\begin{equation}\label{3.9}
T(\mu,\phi,\psi)=\left(
\begin{array}{l}
\phi -P_{K}\left(M\left((2+\mu)e^{U_{\mu}} \left(e^{\frac{ \phi+\psi}2}+ e^{\frac{ \phi-\psi}2}-2\right)\right)\right)\\
\psi-M\left((2-\mu)e^{U_{\mu}} \left(e^{\frac{ \phi+\psi}2}- e^{\frac{ \phi-\psi}2}\right)\right)
\end{array}
\right)
\end{equation}
Letting $
\widetilde K:=\left\{g\in \widetilde Y_{\alpha}\, :\, \int_{{\rm I\mskip -3.5mu R}^2}(\Delta g) \frac{8-|x|^2}{8+|x|^2}\, dx=0\right\}\subset \widetilde Y_{\alpha}\cap K$,
the operator $T(\mu,\phi,\psi)$ maps $(-2,2)\times \widetilde K \times \widetilde Y_{\alpha} $ into $K\times Y_{{\alpha}}$. \\
The zeros of the operator $T(\mu,\phi,\psi)$ then satisfy
\begin{equation}\label{3.10}
\left\{\begin{array}{ll}
-\Delta\phi=(2+\mu)e^{U_{\mu}} \left(e^{\frac{ \phi+\psi}2}+ e^{\frac{ \phi-\psi}2}-2\right)+64 L\frac{8-|x|^2}{(8+|x|^2)^3} & \hbox{ in }{\rm I\mskip -3.5mu R}^2\\
-\Delta \psi=(2-\mu)e^{U_{\mu}} \left(e^{\frac{ \phi+\psi}2}- e^{\frac{ \phi-\psi}2}\right) & \hbox{ in }{\rm I\mskip -3.5mu R}^2\\
\int_{{\rm I\mskip -3.5mu R}^2}e^{U_{\mu}}e^{\frac{ \phi+\psi}2}<+\infty,\ \int_{{\rm I\mskip -3.5mu R}^2}e^{U_{\mu}}e^{\frac{ \phi-\psi}2}<+\infty
\end{array}\right.
\end{equation}
where $L=L(\phi,\psi)\in {\rm I\mskip -3.5mu R}$ and $\phi,\psi\in L^{\infty}$. Once we prove the existence of $\phi$ and $\psi$ verifying \eqref{3.10},
the final step will be to show that $L=0$ so that $\phi$ and $\psi$ are indeed solutions of \eqref{3.2}. This will be done in Section \ref{s4}.\\[.3cm]
To apply Theorem \ref{CR} to the operator $T(\mu,\phi,\psi):(-2,2)\times \widetilde K \times \widetilde Y_{\alpha} \to K\times Y_{{\alpha}}$ at the point $(\mu_n,0,0)$, we have to check hypotheses $a),\dots,d)$.
This will be done in a series of technical lemmas.\\
Even if the statement and the proofs of these lemmas is quite long we prefer to include them in this section since this is the most delicate part of the proof.\\
We start proving the following
\begin{lemma}\label{lem3.1} We have that
$ker\left(T'_{(\phi,\psi)}(\mu_n,0,0)\right)=span\left\{\left( \begin{array}{c} 0\\P_n(\frac{8-|x|^2}{8+|x|^2})\end{array}\right)\right\}$ and then it is one-dimensional.
\end{lemma}
\begin{proof}
Let us consider the linearized operator of $T$ in $(\mu_n,0,0)$. We have that
\begin{equation}\nonumber
T'_{(\phi,\psi)}(\mu_n,0,0)\left(\begin{array}{l}
w_1\\
w_2
\end{array}\right)=\left(
\begin{array}{l}
w_1-P_{K}\left(M\left( (2+\mu_n)e^{U_{\mu_n}}w_1\right)\right)\\
w_2-M\left((2-\mu_n)e^{U_{\mu_n}}w_2\right)
\end{array}\right)
=\left(\begin{array}{l}0\\0\end{array}\right)
\end{equation}
if and only if $(w_1,w_2)\in \widetilde K\times \widetilde Y_{{\alpha}} $ satisfies
\begin{equation}\label{3.12-b}
\left\{\begin{array}{ll}
-\Delta w_1-\frac{64}{(8+|x|^2)^2}w_1-64L\frac{8-|x|^2}{(8+|x|^2)^3}=0 & \hbox{ in }{\rm I\mskip -3.5mu R}^2\\
-\Delta w_2-\frac{n(n+1)}2\frac{64}{(8+|x|^2)^2}w_2=0 & \hbox{ in }{\rm I\mskip -3.5mu R}^2
\end{array}\right.
\end{equation}
for some $L=L(w_1)\in {\rm I\mskip -3.5mu R}$. Observe that, in the second equation we used the definition of $\mu_n$.\\
The first equation in \eqref{3.12-b} can be explicitly solved and gives the solutions
\begin{eqnarray}\label{3.12-c}
&&w_1(|x|)=C_1\frac{8-|x|^2}{8+|x|^2}+C_2\frac{|x|^2\log |x|-8\log |x|-16}{8+|x|^2}+\nonumber\\
&&\frac23L\frac{2(|x|^2-8)\log |x|-(|x|^2-8)\log(8+|x|^2)-16}{8+|x|^2},\nonumber
\end{eqnarray}
which are bounded if and only if $C_2=L=0$. Hence $w_1(|x|)=C_1\frac{8-|x|^2}{8+|x|^2}$ and again $C_1=0$ because
$w_1\in \widetilde K$.\\
The second equation, as said in Lemma \ref{6}, has the nontrivial radial bounded solution $P_n\left(\frac{8-|x|^2}{8+|x|^2}\right)$, so that
\begin{equation}\label{ker}
Ker\left(T'_{(\phi,\psi)}(\mu_n,0,0)\right)=span <\left(\begin{array}{c}
0\\
P_n\left(\frac{8-|x|^2}{8+|x|^2}\right)
\end{array}\right)>
\end{equation}
is one dimensional.
\end{proof}
\begin{lemma}\label{lem3.2}
Let $g\in X_{\alpha}$ be a radial function. Then the ordinary differential equation
\begin{equation}\label{ordinary}
-w''-\frac 1r w'-\frac{n(n+1)}2 \frac{64}{(8+r^2)^2}w=g(r) \quad \hbox{ in }(0,+\infty)
\end{equation}
has the solution $w$ given by
\begin{equation}\label{sol}
w(r)=-P_n\left(\frac{8-r^2}{8+r^2}\right)\left\{\int_0^r\frac 1{s\left(P_n\left(\frac{8-s^2}{8+s^2}\right)\right)^2} \int_0^s tP_n\left(\frac{8-t^2}{8+t^2}\right)g(t)\, dtds +C\right\}.
\end{equation}
\end{lemma}
\begin{proof}
As explained in the proof of Lemma \ref{6} the homogeneous equation $-w''-\frac 1r w'-\frac{n(n+1)}2 \frac{64}{(8+r^2)^2}w=0$ has only the bounded solution $P_n\left(\frac{8-r^2}{8+r^2}\right)$. Then \eqref{sol} follows by a standard use of the variation of constants method. Let us show that $w$ is well defined at a zero of the Legendre function $P_n$.
Denote by $z_n\in (-1,1)$ a point where $P_n(z_n)=0$. Since the zeros of the Legendre function are {\em simple} we get $P_n(z)=\gamma_n(z-z_n)+O\left(|z-z_n|^2\right)$ in a neighborhood of $z_n$ and then it is not difficult to show that $\lim_{z\to z_n}w(z)<+\infty$. This ends the proof.
\end{proof}
\begin{lemma}\label{n5}
If $g\in X_{\alpha}$ then the function $w$ given in \eqref{sol} for $n=1$ satisfies the following asymptotic estimate
\begin{equation}\label{n6}
w(r)=-\left(\int_0^{+\infty}\frac{8-t^2}{8+t^2}tg(t)dt\right)\log r+O(1)\quad\hbox{as }r\rightarrow+\infty
\end{equation}
\end{lemma}
\begin{proof}
Note that since $g\in X_{\alpha}$ then the integral in \eqref{n6} is well defined.
Since $n=1$ the solution $w$ is given by
$$w(r)=-\frac{8-r^2}{8+r^2}\left(I(r)+C\right)$$
with
\begin{equation}\label{I(r)}
I(r)=\int_0^r\frac {(8+s^2)^2}{s(8-s^2)^2} \int_0^s t\frac{8-t^2}{8+t^2}g(t)\, dtds.
\end{equation}
Because $ \frac{8-r^2}{8+r^2}$ is bounded, we just consider the term $I(r)$ for $r$ large. Using that $ \frac {(8+s^2)^2}{s(8-s^2)^2}=\frac 1s+O\left(\frac 1{(1+s)^3}\right)$ for $s$ large enough we have that
\begin{eqnarray}\label{n7}
&I(r)=& -\int_0^r \frac 1s\int_0^s t\frac{8-t^2}{8+t^2}g(t)\, dtds+\int_0^r O\left(\frac 1{(1+s)^3}\right) \int_0^s t\frac{8-t^2}{8+t^2}g(t)\, dtds\nonumber\\
&=&\hbox{(integrating by parts)}\nonumber\\
&=&-\log r\int_0^r t\frac{8-t^2}{8+t^2}g(t)\, dt+ \int_0^r t\log t \frac{8-t^2}{8+t^2}g(t)\, dt +\int_0^r O\left(\frac 1{(1+s)^3}\right) \int_0^s t\frac{8-t^2}{8+t^2}g(t)\, dtds\nonumber\\
&=&-\log r\int_0^r t\frac{8-t^2}{8+t^2}g(t)\, dt+I_1(r)+I_2(r)
\end{eqnarray}
Since $g\in X_{\alpha}$,
\begin{equation}\label{n8}
\int_0^r t\frac{8-t^2}{8+t^2}g(t)\, dt=\int_0^{+\infty} t\frac{8-t^2}{8+t^2}g(t)\, dt+O\left(\frac1{(1+r)^\frac\a2}\right)
\end{equation}
and
\begin{equation}\label{n9}
|I_1(r)|,|I_2(r)|=O(1).
\end{equation}
Finally by \eqref{n7}, \eqref{n8} and \eqref{n9} we have the claim.
\end{proof}
\begin{lemma}\label{n10}
If $g\in X_{\alpha}$ then the function $w$ given in \eqref{sol} for $n>1$ satisfies the following asymptotic estimate
\begin{equation}\label{n11}
w(r)=-\left(\int_0^{+\infty}P_n\left(\frac{8-t^2}{8+t^2}\right)tg(t)dt\right)\log r+O(1)\quad\hbox{as }r\rightarrow+\infty
\end{equation}
\end{lemma}
\begin{proof}
We can repeat step by step the proof of lemma \ref{n5} once we have proved the following estimate,
\begin{equation}\label{comp-asint}
\frac {1}{s\left(P_n\left(\frac{8-s^2}{8+s^2}\right)\right)^2}=\frac 1s+O\left(\frac 1{(1+s)^3}\right)\quad \hbox{ for }s\hbox{ large enough.}
\end{equation}
First observe that $\lim_{s\to +\infty}\left(P_n\left(\frac{8-s^2}{8+s^2}\right)\right)^2=\left(P_n(-1)\right)^2=1$ from the definition of the Legendre Polynomials. Then we have that
$$\frac {1}{s\left(P_n\left(\frac{8-s^2}{8+s^2}\right)\right)^2}-\frac 1s=\frac {1-\left(P_n\left(\frac{8-s^2}{8+s^2}\right)\right)^2}{s\left(P_n\left(\frac{8-s^2}{8+s^2}\right)\right)^2}$$
and, using that $P_n$ and $P'_n$ are both bounded in $(-1,1)$,
$$\frac {1-\left(P_n\left(\frac{8-s^2}{8+s^2}\right)\right)^2}{s}=O\left(P_n\left(\frac{8-s^2}{8+s^2}\right)P'_n\left(\frac{8-s^2}{8+s^2}\right)\left(\frac{8-s^2}{8+s^2}\right)'\right)=O\left(\frac 1{(1+s)^3}\right)$$
for $s$ large enough.
This proves \eqref{comp-asint} and completes the proof of the lemma.
\end{proof}
\begin{corollary}\label{lem3.3}
If $g\in X_{\alpha}$ is a radial function and $w$ is the solution to \eqref{ordinary} given in \eqref{sol}, we have that $w\in L^{\infty}_{loc}([0,+\infty))$ and
\begin{itemize}
\item[i)] if $ \int_{{\rm I\mskip -3.5mu R}^2}P_n\left(\frac{8-|x|^2}{8+|x|^2}\right) g(|x|)\, dx=0$ then $w\in L^\infty([0,+\infty))$,
\item[ii)] if $ \int_{{\rm I\mskip -3.5mu R}^2}P_n\left(\frac{8-|x|^2}{8+|x|^2}\right) g(|x|)\, dx\ne0$ then $w$ has a logarithmic growth at infinity.
\end{itemize}
\end{corollary}
\begin{proof}
Since $g\in X_{\alpha}$ then $\int_0^st P_n\left(\frac{8-t^2}{8+t^2}\right)g(t)\, dt=O(s)$ for $s$ small enough. Then $\frac 1{s\left(P_n\left(\frac{8-s^2}{8+s^2}\right)\right)^2} \int_0^s tP_n\left(\frac{8-t^2}{8+t^2}\right)g(t)\, dt$ is uniformly bounded in a neighborhood of $0$ and so $w(r)$ is regular at the origin. Moreover, it is regular at a zero of the Legendre function $P_n$ and so it belongs to $L^{\infty}_{loc}([0,+\infty))$.
Finally $i)$ and $ii)$ follow by \eqref{n6} and \eqref{n11}.
\end{proof}
\begin{lemma}\label{N1}
We have that for any $g\in X_{\alpha}$ then $w\in Y_{\alpha}$ for any $n\ge1$.
\end{lemma}
\begin{proof}
First we observe that, by assumption $g(|x|)\in L^{2}_{loc}({\rm I\mskip -3.5mu R}^2)$ so that $w(|x|)\in W^{2,2}_{loc}({\rm I\mskip -3.5mu R}^2)$.
Next, by Corollary \ref{lem3.3} we have that there exists $C$ such that $|w(r)|\leq C(1+\log(1+r))$.
Then
$$\int_{{\rm I\mskip -3.5mu R}^2}\frac{w^2}{\left( 1+|x|^{1+\frac {{\alpha}}2}\right)^2}\, dx<+\infty$$
and
$$\int_{{\rm I\mskip -3.5mu R}^2}\left(\Delta w\right)^2\left(1+|x|^{2+{\alpha}}\right)\le C\left( \int_{{\rm I\mskip -3.5mu R}^2}\left( g(|x|)\right)^2\left(1+|x|^{2+{\alpha}}\right)+
\int_{{\rm I\mskip -3.5mu R}^2}\frac{w^2}{\left(8+r^2\right)^4}\right)<+\infty$$
since $g\in X_{\alpha}$. This ends the proof.\\
\end{proof}
\begin{lemma}\label{lem3.6}
Set $w\in \widetilde K$. Then we have that
\begin{equation}\label{q-fin}
\int_{{\rm I\mskip -3.5mu R}^2} \frac {8-|x|^2}{(8+|x|^2)^3}w \, dx=0.
\end{equation}
\end{lemma}
\begin{proof}
Let us denote by $B_R$ the ball centered at the origin and radius $R$. Integrating by parts twice we get
\begin{equation}\label{***}
\int_{B_R}\Delta w \frac {8-|x|^2}{8+|x|^2}\, dx=2\pi\left( Rw'(R)+Rw(R)\frac{16R}{(8+R^2)^2}\right)-\int_{B_R}\frac {64(8-|x|^2)}{(8+|x|^2)^3}w \, dx.
\end{equation}
We claim that there exists $R_k\to +\infty$ as $k\to +\infty$ such that
\begin{equation}\label{****}
R_kw'(R_k)\to 0.
\end{equation}
By contradiction let us assume that there exists ${\alpha}>0$ such that $|w'(R)|\geq \frac{{\alpha}}R$ for any $R\geq R_0$. This implies that $w$ is not bounded and so \eqref{****} holds.\\
Putting $R=R_k$ in \eqref{***} we obtain (since $w\in \widetilde K$)
\begin{eqnarray}
&&0=\int_{{\rm I\mskip -3.5mu R}^2} \Delta w \frac {8-|x|^2}{8+|x|^2}\, dx=\lim_{k\to +\infty}\int_{B_{R_k}}\Delta w \frac {8-|x|^2}{8+|x|^2}\, dx\nonumber\\
&&=\lim_{k\to +\infty}2\pi\left( R_kw'(R_k)+w(R_k)\frac{16R^2_k}{(8+R_k^2)^2}\right)-\int_{B_{R_k}}\frac {64(8-|x|^2)}{(8+|x|^2)^3}w \, dx\nonumber\\
&&=-\int_{{\rm I\mskip -3.5mu R}^2} \frac {64(8-|x|^2)}{(8+|x|^2)^3}w \, dx\nonumber
\end{eqnarray}
and the proof is complete.
\end{proof}
\begin{lemma}\label{lem3.5}
We have that the range $R\left(T'_{(\phi,\psi)}(\mu_n,0,0)\right)\subset K\times Y_{{\alpha}}$ has codimension one and it is given by the functions $(f_1,f_2)\in K\times Y_{{\alpha}}$ that satisfies
\begin{equation}\label{condition}
\int_{{\rm I\mskip -3.5mu R}^2}P_n\left(\frac{8-|x|^2}{8+|x|^2}\right)\Delta f_2\, dx=0.
\end{equation}
\end{lemma}
\begin{proof}
Let $(f_1,f_2)\in K\times Y_{{\alpha}}$. Our claim is equivalent to prove that
\begin{equation}\label{range}
T'_{(\phi,\psi)}(\mu_n,0,0)\left(\begin{array}{l}
w_1\\
w_2
\end{array}\right)=\left(\begin{array}{l}
f_1\\
f_2
\end{array}\right)
\end{equation}
has a solution $(w_1,w_2)\in \widetilde K\times \widetilde Y_{{\alpha}}$ for any $f_1\in K$ and $f_2$ satisfying \eqref{condition}.
On the other hand \eqref{range} is equivalent to find $(w_1,w_2)\in \widetilde K\times \widetilde Y_{{\alpha}}$ that solves
\begin{eqnarray}
&-\Delta w_1-(2+\mu_n)e^{U_{\mu_n}}w_1=-\Delta f_1 +L\frac {8-|x|^2}{(8+|x|^2)^3}& \hbox{ in }{\rm I\mskip -3.5mu R}^2\label{eq-1}\\
&-\Delta w_2-(2-\mu_n)e^{U_{\mu_n}}w_2=-\Delta f_2 & \hbox{ in }{\rm I\mskip -3.5mu R}^2\label{eq-2}
\end{eqnarray}
for any $(f_1,f_2)\in K\times Y_{{\alpha}}$
and for some $L=L(w_1)\in {\rm I\mskip -3.5mu R}$. From the definition of the space $K$ and of the operator $M$ we have that
$$L(w_1)=-\frac{\int_{{\rm I\mskip -3.5mu R}^2}\frac {8-|x|^2}{(8+|x|^2)^3}w_1\, dx}{\int_{{\rm I\mskip -3.5mu R}^2}\frac {(8-|x|^2)^2}{(8+|x|^2)^4}\, dx}.$$
Let us consider the function $\widetilde w_1(r)$ which solves \eqref{ordinary} with $n=1$ and $g(r)=-\Delta f_1$. Since $f_1\in K$ we get from Corollary \ref{lem3.3} and Lemma \ref{N1} that $\widetilde w_1\in \widetilde Y_{\alpha}$.\\
Finally, setting $w_1=\widetilde w_1+A\frac {8-|x|^2}{8+|x|^2}$ with $A=-\frac{\int_{{\rm I\mskip -3.5mu R}^2}\Delta \widetilde w_1 \frac {8-|x|^2}{8+|x|^2}\, dx}{\int_{{\rm I\mskip -3.5mu R}^2}\frac {64(8-|x|^2)^2}{(8+|x|^2)^4}\, dx}$ we get that $w_1\in \widetilde K$ and satisfies $-\Delta w_1-(2+\mu_n)e^{U_{\mu_n}}w_1=-\Delta f_1$ in ${\rm I\mskip -3.5mu R}^2$. On the other hand, by Lemma \ref{lem3.6}, since $w_1\in \widetilde K$ we have that $\int_{{\rm I\mskip -3.5mu R}^2} \frac {8-|x|^2}{(8+|x|^2)^3}w_1
\, dx=0$ and then $L(w_1)=0$. So $w_1$ satisfies \eqref{eq-1}.\\
For what concerns equation \eqref{eq-2}, using again Corollary \ref{lem3.3}, we have that the solution $w_2$ belongs to $\widetilde Y_{\alpha}$ if and only if $f_2$ satisfies \eqref{condition}.
This concludes the proof.
\end{proof}
\begin{lemma}\label{lem3.7}
We have that
\begin{equation}\label{T''}
T''_{\mu,(\phi,\psi)}(\mu_n,0,0)\left(\begin{array}{c} 0\\
P_n\left(\frac{8-r^2}{8+r^2}\right)\end{array}\right)\notin R\left(T'_{(\phi,\psi)}(\mu_n,0,0)\right).
\end{equation}
\end{lemma}
\begin{proof}
We have that $T''_{\mu,(\phi,\psi)}(\tilde \mu, \tilde\phi,\tilde \psi)$ is given by
\begin{eqnarray}
&\left(\begin{array}{cc}
0 & 0\\
M\left(-\frac{128}{(2+\tilde \mu)^2(8+|x|^2)^2}\left(e^{\frac{ \tilde\phi+\tilde\psi}2}- e^{\frac{ \tilde\phi-\tilde\psi}2}\right)I\right) & M\left(-\frac{128}{(2+\tilde \mu)^2(8+|x|^2)^2}\left(e^{\frac{ \tilde\phi+\tilde\psi}2}+ e^{\frac{ \tilde\phi-\tilde\psi}2}\right)I\right)
\end{array}\right) &\nonumber
\end{eqnarray}
so that
$$T''_{\mu,(\phi,\psi)}(\mu_n,0,0)\left(\begin{array}{c} 0\\
P_n\left(\frac{8-|x|^2}{8+|x|^2}\right)\end{array}\right)=\left(\begin{array}{c} 0\\
M\left(-\frac{256}{(2+\tilde \mu)^2(8+|x|^2)^2}P_n\left(\frac{8-|x|^2}{8+|x|^2}\right)\right)\end{array}\right).$$
This proves \eqref{T''} since
\begin{eqnarray}
&&\int_{{\rm I\mskip -3.5mu R}^2}\left(-\Delta \left(M\left(-\frac{256}{(2+\tilde \mu)^2(8+|x|^2)^2}P_n\left(\frac{8-|x|^2}{8+|x|^2}\right)\right)\right)\right)P_n\left( \frac{8-|x|^2}{8+|x|^2}\right)\, dx \nonumber\\
&&=\frac{-256}{(2+\mu_n)^2}\int_{{\rm I\mskip -3.5mu R}^2}\frac{1}{(8+|x|^2)^2}\left(P_n\left(\frac{8-|x|^2}{8+|x|^2}\right)\right)^2\, dx\neq 0\nonumber
\end{eqnarray}
and the Range of $T'_{(\phi,\psi)}(\mu_n,0,0)$ is characterized by \eqref{condition}.
\end{proof}
Now we have all the ingredients to apply Theorem \ref{CR} and we have the following result:
\begin{theorem}\label{t3.2}
Let us fix $n \in {\rm I\mskip -3.5mu N}$ and let $\mu_n$ be as defined in \eqref{a1}. Then the
points $(\mu_n,0,0)$ are radial bifurcation points for the curve of solutions $(\mu,0,0)$, i.e. for $\varepsilon>0$ small enough there exist $u_\varepsilon$, $v_\varepsilon$
satisfying
\begin{equation}\label{c1}
\left\{\begin{array}{ll}
-\Delta u_\varepsilon=2e^{u_\varepsilon}+\mu_\varepsilon e^{v_\varepsilon}+64L_\varepsilon\frac{8-|x|^2}{\left(8+|x|^2\right)^3} & \hbox{ in }{\rm I\mskip -3.5mu R}^2\\
-\Delta v_\varepsilon=2e^{v_\varepsilon}+\mu_\varepsilon e^{u_\varepsilon}+64L_\varepsilon\frac{8-|x|^2}{\left(8+|x|^2\right)^3} & \hbox{ in }{\rm I\mskip -3.5mu R}^2\\
\int_{{\rm I\mskip -3.5mu R}^2}e^{u_\varepsilon}\le C,\ \int_{{\rm I\mskip -3.5mu R}^2}e^{v_\varepsilon}\le C
\end{array}\right.
\end{equation}
for some Lagrange multiplier $L_\varepsilon$. Moreover we have that, for $\varepsilon$ small enough,
\begin{equation}\label{fin}
\left\{\begin{array}{ll}
u_\varepsilon=U_{\mu_{\varepsilon,n}}+\varepsilon P_n\left(\frac{8-|x|^2}{8+|x|^2}\right)+\varepsilon(z_{1,\varepsilon}+z_{2,\varepsilon})\\
v_\varepsilon=U_{\mu_{\varepsilon,n}}-\varepsilon P_n\left(\frac{8-|x|^2}{8+|x|^2}\right)+\varepsilon(z_{1,\varepsilon}-z_{2,\varepsilon})
\end{array}\right.
\end{equation}
with $\mu_\varepsilon$ such that $\mu_0=\mu_n$ and some functions $z_{1,\varepsilon},z_{2,\varepsilon}$ such that $z_{1,0}=z_{2,0}=0$ and $z_{1,\varepsilon},z_{2,\varepsilon}$ are uniformly bounded in $L^\infty({\rm I\mskip -3.5mu R}^2)$.
\end{theorem}
\begin{proof}
We apply Theorem \ref{CR} with $X=\widetilde K\times \widetilde Y_{{\alpha}}$, $Y=K\times Y_{{\alpha}}$ and $F=T(\mu,\phi,\psi)$ as defined in \eqref{3.9}.
We have that $a)$ clearly holds.\\
Concerning $b)$ let us consider a sequence $(\phi_k,\psi_k)\rightarrow(\phi,\psi)$ in $\widetilde K\times \widetilde Y_{{\alpha}}$. Then we have that $\phi_k\rightarrow\phi$ and $\psi_k\rightarrow\psi$ in $Y_{\alpha}$ and {\em uniformly} in ${\rm I\mskip -3.5mu R}^2$. This is enough to conclude that $T'_\mu$, $T'_{(\phi,\psi)}$ and $T''_{\mu,(\phi,\psi)}$ are continuous. We skip the tedious details.\\
Assumption $c)$ follows by Lemma \ref{lem3.1} and Lemma \ref{lem3.5}.\\
Finally $d)$ follows by Lemma \ref{lem3.7}.\\
From Theorem \ref{CR} we have that for every $n\in {\rm I\mskip -3.5mu N}$ there is a branch of nontrivial radial solutions $(\phi_\varepsilon,\psi_\varepsilon)$ of $T(\mu,\phi,\psi)=0$ bifurcating from $(\mu_n,0,0)$. Moreover, we know that, for $\varepsilon$ small enough,
$(\phi_\varepsilon,\psi_\varepsilon)=\left(\varepsilon z_{1,\varepsilon},\varepsilon P_n\left(\frac{8-|x|^2}{8+|x|^2}\right)+\varepsilon z_{2,\varepsilon}\right)$ with $(z_{1,\varepsilon},z_{2,\varepsilon})$ belonging to a neighborhood of the origin in $\widetilde K \times \widetilde Y_{\alpha} $ such that $z_{i,0}=0$ for $i=1,2$.
Then $(\phi_\varepsilon,\psi_\varepsilon)$ satisfy, for $\varepsilon$ small enough,
\begin{equation}
\left\{\begin{array}{ll}
\phi_\varepsilon=\varepsilon z_{1,\varepsilon} \nonumber\\
\psi_\varepsilon=\varepsilon P_n\left(\frac{8-|x|^2}{8+|x|^2}\right)+\varepsilon z_{2,\varepsilon}
\end{array}\right.
\end{equation}
with $z_{1,\varepsilon},z_{2,\varepsilon}$ uniformly bounded in $L^\infty({\rm I\mskip -3.5mu R}^2)$. Recalling the definition of $\phi$ and $\psi$ in \eqref{n3}, \eqref{n4} we get the claim.
\end{proof}
\sezione{The Lagrange multiplier is zero}\label{s4}
\noindent In this section we will show that the Lagrange multiplier $L_\varepsilon$ appearing in \eqref{c1} is zero.
In the sequel we denote by $C$ a generic constant (independent of $n$) which can change from line to line.\\
Let us start this section proving a bound on $L_\varepsilon$.
\begin{lemma}\label{c2}
We have that
\begin{equation}\label{d2}
|L_\varepsilon|\le C.
\end{equation}
\end{lemma}
\begin{proof}
Let us multiply the first equation of \eqref{c1} by $\frac{8-|x|^2}{8+|x|^2}$ and integrate over ${\rm I\mskip -3.5mu R}^2$. We get
$$64 L_\varepsilon\int_{{\rm I\mskip -3.5mu R}^2}\frac{\left(8-|x|^2\right)^2}{\left(8+|x|^2\right)^4}\, dx=\int_{{\rm I\mskip -3.5mu R}^2}\left(-\Delta u_\varepsilon\right)\frac{8-|x|^2}{8+|x|^2}\, dx-\int_{{\rm I\mskip -3.5mu R}^2}\left(2e^{u_\varepsilon}+\mu_{\varepsilon,n} e^{v_\varepsilon}\right)\frac{8-|x|^2}{8+|x|^2}\, dx.$$
We know that $||u_\varepsilon||_{Y_{\alpha}}\leq ||U_{\mu_{\varepsilon,n}} ||_{Y_{\alpha}}+C\leq C$
since $\mu_{\varepsilon,n}>-2+\delta$ for $\varepsilon$ small enough. Then we have that
\begin{equation}
\int_{{\rm I\mskip -3.5mu R}^2}\left(-\Delta u_\varepsilon\right)\frac{8-|x|^2}{8+|x|^2}\leq||u_\varepsilon||_{Y_{\alpha}}\left(\int_{{\rm I\mskip -3.5mu R}^2}\frac 1{1+|x|^{2+{\alpha}}}\left(\frac{8-|x|^2}{8+|x|^2}\right)^2\right)^\frac12\leq C.
\end{equation}
Then the uniform bound on $\int_{{\rm I\mskip -3.5mu R}^2}e^{u_\varepsilon}\,,\int_{{\rm I\mskip -3.5mu R}^2}e^{v_\varepsilon}$ in \eqref{c1} gives that
\begin{equation}
|L_\varepsilon|\le C
\end{equation}
which gives the claim.
\end{proof}
\noindent Now we prove some useful estimates.
\begin{proposition}\label{c2-bis}
If $u_\varepsilon$ and $v_\varepsilon$ satisfy \eqref{c1} we have that
\begin{equation}\label{c3}
\left\{\begin{array}{ll}
u_\varepsilon(x)=-\frac1{2\pi}\left(2\int_{\mathbb{R}^2}e^{u_\varepsilon}+\mu_{\varepsilon,n}\int_{\mathbb{R}^2}e^{v_\varepsilon}\right)\log|x|+O(1)
\\ &\hbox{as }|x|\rightarrow+\infty\\
v_\varepsilon(x)=-\frac1{2\pi}\left(2\int_{\mathbb{R}^2}e^{v_\varepsilon}+\mu_{\varepsilon,n}\int_{\mathbb{R}^2}e^{u_\varepsilon}\right)\log|x|+O(1)
\end{array}\right.
\end{equation}
\end{proposition}
\begin{proof}
Recalling that $u_\varepsilon$ and $v_\varepsilon$ are radial, we write the ODE corresponding to \eqref{c1},
\begin{equation}\label{n13}
\left\{\begin{array}{ll}
- u_\varepsilon''-\frac1r u_\varepsilon'=2e^{u_\varepsilon}+\mu_{\varepsilon,n} e^{v_\varepsilon}+64L_\varepsilon
\frac{8-r^2}{\left(8+r^2\right)^3} & \hbox{ in }(0,+\infty)\\
- v_\varepsilon''-\frac1r v_\varepsilon'=2e^{v_\varepsilon}+\mu_{\varepsilon,n} e^{v_\varepsilon}+64L_\varepsilon
\frac{8-r^2}{\left(8+r^2\right)^3} & \hbox{ in }(0,+\infty)
\end{array}\right.
\end{equation}
Let us prove our estimates for $u_\varepsilon$ (the case of $v_\varepsilon$ is the same). Integrating \eqref{n13} twice we get
\begin{eqnarray}\label{n14}
&&-u_\varepsilon(r)+u_\varepsilon(0)=\int_0^r\left(2e^{u_\varepsilon}+\mu_{\varepsilon,n} e^{v_\varepsilon}\right)sds\log r-
\int_0^r\left(2e^{u_\varepsilon}+\mu_{\varepsilon,n} e^{v_\varepsilon}\right)s\log sds\nonumber\\
&&+2L_\varepsilon\frac {r^2}{8+r^2}
\end{eqnarray}
Then, from \eqref{fin} we deduce that
\begin{equation}\label{n20}
e^{u_\varepsilon},e^{v_\varepsilon}\le \frac C{1+|x|^4}
\end{equation}
So from \eqref{n14} we derive
\begin{equation}\label{n21}
-u_\varepsilon(r)+u_\varepsilon(0)=\left(\int_0^{+\infty}\left(2e^{u_\varepsilon}+\mu_{\varepsilon,n} e^{v_\varepsilon}\right)sds\right)\log r+O(1),
\end{equation}
which gives the claim.
\end{proof}
Next proposition states the important {\em mass quantizaton} of the solutions $u_\varepsilon$ and $v_\varepsilon$,
\begin{proposition}\label{c4}
We have that
\begin{equation}\label{c5}
\int_{\mathbb{R}^2}e^{u_\varepsilon}=\int_{\mathbb{R}^2}e^{v_\varepsilon}=\frac{8\pi}{2+\mu_{\varepsilon,n}}
\end{equation}
\end{proposition}
\begin{proof}
Using \eqref{fin} and the $L^\infty$-boundedness of $z_1$ and $z_2$ we have that $\left|u_\varepsilon(x)+4\log x\right|\le C $ and $\left|v_\varepsilon(x)+4\log x\right|\le C \hbox{ if }|x|\ge R$ for some $R\in {\rm I\mskip -3.5mu R}$ sufficiently large.
By Proposition \ref{c2-bis} we deduce that
\begin{equation}
\left\{\begin{array}{ll}
-\frac1{2\pi}\left(2\int_{\mathbb{R}^2}e^{u_\varepsilon}+\mu_{\varepsilon,n}\int_{\mathbb{R}^2}e^{v_\varepsilon}\right)+4=0
\\
-\frac1{2\pi}\left(2\int_{\mathbb{R}^2}e^{v_\varepsilon}+\mu_{\varepsilon,n}\int_{\mathbb{R}^2}e^{u_\varepsilon}\right)+4=0
\end{array}\right.
\end{equation}
which gives the claim.
\end{proof}
\begin{proposition}\label{c6}
Let us suppose that $u$ and $v$ are smooth solutions of
\begin{equation}\label{c7}
\left\{\begin{array}{ll}
-\Delta u=2e^u+\mu e^v+H(x) & \hbox{ in }{\rm I\mskip -3.5mu R}^2\\
-\Delta v=2e^v+\mu e^u+H(x) & \hbox{ in }{\rm I\mskip -3.5mu R}^2\\
\end{array}\right.
\end{equation}
with $H\in C^0({\rm I\mskip -3.5mu R}^2)$.
Then we have the following Pohozaev identity,
\begin{eqnarray}\label{c8}
&&R\mu\int_{\partial B_R}\nabla u\cdot\nabla v-2R\mu\int_{\partial B_R}\frac{\partial u}{\partial\nu}\frac{\partial v}{\partial\nu}-R\int_{\partial B_R}\left(|\nabla u|^2+|\nabla v|^2 \right)+\nonumber\\
&&2R\int_{\partial B_R}\left(\left(\frac{\partial u}{\partial\nu} \right)^2+\left(\frac{\partial v}{\partial\nu} \right)^2\right)+R(4-\mu^2)\int_{\partial B_R}\left(e^u+e^v\right)=
\nonumber\\
&&
2(4-\mu^2)\int_{B_R}\left(e^u+e^v\right)+(\mu-2)\int_{B_R}H(x)(x\cdot\nabla u+x\cdot\nabla v)
\end{eqnarray}
\end{proposition}
\begin{proof}
It is the same as in \cite{CK}(see also \cite{JW1})
\end{proof}
\begin{corollary}\label{cor9}
Assume that $H(x)$ is a radial function and that that $u=u(r)$ and $v=v(r)$ are radial solutions of \eqref{c7}.
Let us suppose there exists $\tilde R$ such that $\int_0^{\tilde R} H(r)rdr=0$. Then we have that
\begin{eqnarray}\label{c10}
&&(4-\mu^2)\left[\left(\int_0^{\tilde R}e^urdr\right)^2+\left(\int_0^{\tilde R} e^vrdr\right)^2+\mu\left(\int_0^{\tilde R} e^urdr\right)\left(\int_0^{\tilde R} e^vrdr\right)
\right]+\nonumber\\
&&(4-\mu^2)\tilde{R}^2\left(e^{u(\tilde R)}+e^{v(\tilde R)}\right)=
2(4-\mu^2)\left(\int_0^ {\tilde R}e^urdr+\int_0^ {\tilde R}e^vrdr\right)+\nonumber\\
&&(\mu-2)\int_0^ {\tilde R}H(r)(u'+v')r^2dr.
\end{eqnarray}
\end{corollary}
\begin{proof}
Using that $u$ and $v$ are radial solutions of \eqref{c7}, then \eqref{c8} becomes
\begin{eqnarray}\label{c7-bis}
&&-\mu R^2 u'(R)v'(R)+R^2\left(u'(R)^2+v'(R)^2\right)+(4-\mu^2)R^2\left(e^{u(R)}+e^{v(R)}\right)=\nonumber\\
&&2(4-\mu^2)\int_0^R\left(e^u+e^v\right)rdr+(\mu-2)\int_0^RH(r)(u'+v')r^2dr
\end{eqnarray}
for any $R$.
Integrating \eqref{c7} and recalling that $\int_0^{\tilde R} H(r)rdr=0$ we get that
\begin{equation}\label{c10-bis}
\left\{\begin{array}{ll}
- \tilde Ru'(\tilde R)=2\int_0^{\tilde R}e^urdr+\mu\int_0^{\tilde R}e^vrdr\\
- \tilde Rv'(\tilde R)=2\int_0^{\tilde R} e^vrdr+\mu\int_0^{\tilde R}e^urdr\\
\end{array}\right.
\end{equation}
Setting $\alpha=\int_0^{\tilde R} e^urdr$, $\beta=\int_0^{\tilde R} e^vrdr$ and recalling \eqref{c5} we have that \eqref{c7-bis} becomes
\begin{eqnarray}\label{c9}
&&-\mu(2{\alpha}+\mu{\beta})(2{\beta}+\mu{\alpha})+(2{\alpha}+\mu{\beta})^2+(2{\beta}+\mu{\alpha})^2+(4-\mu^2)\tilde R^2\left(e^{u(\tilde R)}+e^{v(\tilde R)}\right)
=\nonumber\\
&& 2(4-\mu^2)({\alpha}+{\beta})+(\mu-2)\int_0^{\tilde R}H(r)(u'+v')r^2dr,
\end{eqnarray}
which gives the claim.
\end{proof}
\begin{proposition}\label{c11} Let $u_\varepsilon$ and $v_\varepsilon$ be the solutions of \eqref{c1}.
We have that $L_\varepsilon=0$ in \eqref{c1} for $\varepsilon$ small enough.
\end{proposition}
\begin{proof}
From \eqref{fin}
we deduce that
\begin{equation}\label{c12}
R^2\left(e^{u_\varepsilon(R)}+e^{v_\varepsilon(R)}\right)\rightarrow0\hbox{ as }R\rightarrow+\infty.
\end{equation}
Moreover, let us observe that integrating \eqref{c1} between $0$ and $r$ we have, using \eqref{c5},\\
\begin{equation}\label{c15}
|u_\varepsilon'(r)r|\le 4+CL_\varepsilon
\end{equation}
and since by Lemma \ref{c2} $L_\varepsilon$ is uniformly bounded, we derive that $u_\varepsilon'(r)$ is bounded for $r$ large enough.\\
Then, we apply the Pohozaev identity \eqref{c8} with $H(x)=64L_\varepsilon\frac{8-|x|^2}{\left(8+|x|^2\right)^3}$ as in \eqref{c1} and $\mu=\mu_{\varepsilon,n}$. Using that
$\int_0^{+\infty} \frac{8-r^2}{\left(8+r^2\right)^3}r dr=0$ and that $e^{u_\varepsilon}$, $e^{v_\varepsilon}$ and $\frac{8-r^2}{\left(8+r^2\right)^3}(u_\varepsilon'+v_\varepsilon')r^2$ belong to $L^1({\rm I\mskip -3.5mu R}^2)$,
we can apply \eqref{c7-bis} with $\tilde R=+\infty$ getting that
\begin{eqnarray}\label{c13}
&&(4-\mu_{\varepsilon,n}^2)\left[\left(\int_0^\infty e^{u_\varepsilon}rdr\right)^2+\left(\int_0^\infty e^{v_\varepsilon}rdr\right)^2+\mu_{\varepsilon,n}^2\left(\int_0^\infty e^{u_\varepsilon}rdr\right)\left(\int_0^\infty e^{v_\varepsilon}rdr\right)
\right]=\nonumber\\
&&2(4-\mu_{\varepsilon,n}^2)\int_0^\infty \left(e^{u_\varepsilon}+e^{v_\varepsilon}\right) r dr + 64(\mu_{\varepsilon,n}^2-2)L_\varepsilon\int_0^\infty\frac{8-r^2}{\left(8+r^2\right)^3}(u_\varepsilon'+v_\varepsilon')r^2dr.
\end{eqnarray}
Then using the mass quantization property of Proposition \ref{c4} we have that $\int_0^\infty e^{u_\varepsilon}rdr=\int_0^\infty e^{v_\varepsilon}rdr=\frac{4}{2+\mu_{\varepsilon,n}}$. So \eqref{c13} becomes
\begin{equation}\label{c14}
0=L_\varepsilon\int_0^\infty\frac{8-r^2}{\left(8+r^2\right)^3}(u_\varepsilon'+v'_\varepsilon)r^2dr.
\end{equation}
In the previous section we showed that $u_\varepsilon,v_\varepsilon\rightarrow U_{\mu_n}$ as $\varepsilon\rightarrow0$. This implies that $u_\varepsilon',v_\varepsilon'\rightarrow-\frac{4r}{8+r^2}$ pointwise in ${\rm I\mskip -3.5mu R}$.\\
Passing to the limit as $\varepsilon\rightarrow0$ in \eqref{c14} we get that $L_\varepsilon$=0 for $\varepsilon$ small.
This ends the proof.
\end{proof}
{\bf Proof of Theorem \ref{i6}}\\
It follows by Theorem \ref{t3.2} and Proposition \ref{c11}.
\sezione{A perturbation result: proof of Theorem \ref{thm2}}\label{s5}
The proof of Theorem \ref{thm2} follows simply from a perturbation argument. To this end, we fix $a_1>0, a_2>0$ and $\mu= -1 + \delta$. Denote $ (U_0, V_0)= (u_{a_1, a_2},v_{a_1, a_2})$. We write $ (u, v)= ( U_0, V_0) + (\phi, \psi)$ and substitute into the system (\ref{1}) to obtain
\begin{equation}
\label{nm1}
\left\{\begin{array}{l}
\Delta \phi + 2e^{U_0} \phi - e^{V_0} \psi = -\mu (e^{V_0+\psi} -e^{V_0} -e^{V_0} \psi ) -(\mu+1) e^{V_0} -2 (e^{U_0+\phi}-e^{U_0}- e^{U_0}\phi) \\
\Delta \psi + 2e^{V_0} \psi - e^{U_0} \phi = -\mu (e^{U_0+\phi} -e^{U_0} -e^{U_0} \phi ) -(\mu+1) e^{U_0} -2 (e^{V_0+\phi}-e^{V_0}- e^{V_0}\phi) \\
\phi= \phi (r),\psi= \psi (r)
\end{array}
\right.
\end{equation}
Here we work in the radial class. As in Section \ref{s2} we consider the invertibility of the linearized operator in some suitable Sobolev spaces. To this end, recall $X_\alpha$ and $ Y_\alpha$ introduced in Section \ref{s2}. We denote $ X_{\alpha, r}$ and $Y_{\alpha, r}$ as the radial functions in $X_\alpha$ and $Y_\alpha$ respectively.
On $X_{\alpha, r}$ and $Y_{\alpha,r}$, we equip with two norms respectively:
\begin{equation}
\label{norm12}
\| f \|_{**}= \sup_{y \in {\rm I\mskip -3.5mu R}^2} (1+|y|)^{2+\alpha}|f(y)|, \ \ \| h \|_{*}=\max (\| h\|_{Y_\alpha}, \sup_{y \in {\rm I\mskip -3.5mu R}^2} (\log (2+|y|))^{-1} |h(y)|).
\end{equation}
The proof of Theorem \ref{thm2} follows from the following invertibility of the linearized operator (\ref{nm1}) in the radial class and contraction mapping theorem.
\begin{lemma}
\label{lemma2}
Assume that $h=(h_1, h_2) \in X_{\alpha, r}$. Then one can find a unique solution $ (\phi, \psi)=T^{-1}(h) \in Y_{\alpha, r}$ such that
\begin{equation}\label{e305}
\left\{\begin{array}{c}
\Delta \phi+ 2 e^{U_0}\phi- e^{V_0}\psi=h_1\\
\Delta \psi+ 2 e^{V_0} \psi - e^{U_0} \phi=h_2\\
\phi= \phi (r), \ \ \psi =\psi (r)
\end{array},
\right. \,
\| \phi \|_{*} +\| \psi\|_{*} \leq C ( \| h_1 \|_{**} + \| h_2 \|_{**}) .
\end{equation}
Moreover, the map $ (h_1, h_2) \stackrel{T} {\longrightarrow} (\phi, \psi)$ can be made continuous and smooth.
\end{lemma}
\begin{proof} This follows directly from Lemma 2.3 of \cite{ALW} by restricting to the radial class. We omit the details.
\end{proof}
|
\section{\label{sec:level1}Introduction \lowercase{} }
The reaction diffusion system studied in this paper is described using an established analytical approach,\cite{hagberg1997dynamics, barkley1991model, mikhailov1990foundations, PhysRevLett.60.1880, PhysRevE.51.1899, hagberg1994complex}. The nonuniform motion of the solution front is also investigated. The solution of front is given by a relationship between the normal velocity of the front vector and its curvature $k$ consisting of an integrodifferential equation for the front curve.\cite{hagberg1997dynamics} The model is a simple, two-component model that gives rise to a fast propagator (activator) and recovery variable known as an inhibitor.\cite{winfree2001geometry, murray2002mathematical, field1985oscillations} This is due to presence of nonequilibrium Ising-Bloch bifurcation. \newline
Adding the kinetic coefficients requires detailed reaction schemes of incoming waves from the energetic material into the lens. The basic mechanisms are: phase change to solid intermediates and the formation of gaseous products. Because the material decomposes below its melting point, this results in a complex pathway involving NO, nitroso, and Non-Violent Resistance (NVR) cycles. One pathway involves elimination of HONO and HON, leading to the formation of HCN (OST; C$_3$H 3N$_3$O) under high confinement conditions. The other cycle includes the formation of H$_2$O, NO and NO$_2$. The next reaction leads to NO cycles, NO$_2$ with nitroso cycles and the formation of minor products at higher pressures,\cite{behrens1991thermali} Interfacing the material with nitroguanidine leads to diffusion of decomposition products and the creation of spiral waves in the chemical. Figure \ref{arbode} demonstrates the behaviour of nitroguanidine on a model developed for this purpose based on a nonlinear system.
\begin{figure}[ht!]
\centering
\includegraphics[width=90mm]{figureone.eps}
\caption{Diffusion in nitroguanidine for averaged $N_2$ and C$-$N molecules. At $t=0.01$ to $t=0.08$, there is an increase temperature due to incoming waves; At $t=0.08$ to $t=0.5$, there is a phase change from solid to vapor, where at $t=0.5$ outburst occurs ($t=0.5$ is defined outside of the Dirichlet boundaries.)}
\label{arbode}
\end{figure}
The stars show the diffusion of $N_2$ molecules into nitroguanidine, which causes a temperature increase creating an unstable condition within the system. This leads to an increase in kinetic energy and molar volume, and ultimately to the explosion of the chemical. This work will lead to important new research in the fields of guided stress/shock waves.
\section{\label{sec:level1}The Model \lowercase{} }
The model developed can be applied to any diffusive products in organic reactant. In this paper, the diffusive products are dealt with as external control parameters which affect the wave behaviour in the lens. Specifically, these parameters are the crystal temperature and partial pressures of the decomposition products. Hence, an understanding of the decomposition is essential. In studying the decomposition of such a system, it is always assumed that it enters the vapor phase before dissociating. In a series of experiments, It was found to release about 500 kJ.mol$^{-1}$ of energy, independent of the heating rate and mass. The majority of the decomposition products form via the scission of $N_2$, N-NO$_2$ and C$-$N bonds. The decomposition is largely independent of the heating rate, which is due to the total heat release (due to the exothermic decomposition reactions). Finding bifurcation requires knowledge of the partial pressures of the reactants and products, as well as their kinetic rates.\cite{park1993kinetics}. According to several study the kinetic behaviour of molecular dissociation for $N_2$ bond dissociation is 200 kJ/mol and the energy of decomposition of the triple bond-fission reaction products of CH$_2$NNO$_2$ are 155, 234 and 297 kJ.mol$^{-1}$.
\newline
For the lens, the reaction-diffusion system is introduced through the bond-fission products of $N_2$ and C$-$N at Dirichlet boundaries. Because this is a bistable system, a front-like structure connects the two fronts in a homogeneous steady state. A small perturbation known as an Ising--Bloch bifurcation\cite{yadav2004interaction} exists, where the fronts exchange stability with a pair of counter-propagating Bloch fronts. This front motion is broken by imposing a fixed chemical concentration at the boundary of the reactor. Here, the nitroguanidine molecule is considered as a stationary spot to bifurcate the incoming fronts. The spiral waves and breaking spots in an excitable\cite{meron1992pattern} and bistable medium are examples of oscillatory behaviour; the oscillations often result from the underlying oscillating dynamics of local chemical kinetics.\cite{haim1996breathing}\newline
\begin{figure*}
\centering
\begin{minipage}[ht!]{.5\textwidth}\centering
\includegraphics[scale=0.50]{figure2a.eps}
\end{minipage}\hfill
\begin{minipage}[ht!]{.5\textwidth}\centering
\includegraphics[scale=.90]{figure2b.eps}
\end{minipage}\hfill
\caption{a) Solution to Eqs. 1 and 2 for different values of the diffusion parameter $D$. The data show the increase in temperature in the lattice. b) Compression front at the nitroguanidine interface boundary.}
\label{difphase1}
\end{figure*}
The model presented in this paper was used to study the wave propagation under three dynamic boundary conditions. This propagation can influence the state of reactants by displacing them. This is due to the increase in specific volume, which generates compression waves and sets the reactants and flame front into motion following precursor shock. Thus, the three important, overriding factors are the wave speed, specific volume ratio and boundary conditions. The boundary condition in this model is therefore set at a critical speed of $(S/\alpha)$ at which the initial wave speed is lower than that of the shockwave-induced compression ($M_s$), such that an acceleration mechanism can be introduced through an external field. Here, $S$ is the burn velocity, and $\alpha$ is sound speed in front of the shock. \newline
Under this condition, the chemical reaction is affected by the compression shock wave and shock diffusion replaces diffusion-controlled flame propagation.\cite{spitzer2010energetic,krischer1992oscillatory,hagberg1994pattern}
Having established the following parameters: crystal pressure (T$_c$), partial pressures for $N_2$ and C$-$N resulting from bond dissociation as P$_{N_2}$ and P$_{CN}$, kinetic rates ($k$), surface area of the reaction diffusion system (S$_a$) and the external field ($h$), the reaction diffusion system is given by:
\begin{equation}
\dot{u}=\epsilon^{-1}(-k_d u- a_0 u^3-k_r u)+ \delta_t \rho_{xx}
\label{bif1}
\end{equation}
and
\begin{equation}
\dot{v}=u- a_1( k_r v - \delta v_{xx} ) + h
\label{bif2}
\end{equation}
where $u$ and $v$ are the two variables for the propagator and inhibitor.\cite{bar2009spiral,bode1994measurement,haas1995observation} The parameter $a_0$ represents the kinetic expression for $k_u. S_a. P_{N_2}$ and $a_1$ represents $k_u. S_{v}. P_{CN}$, where $k_u$ is given by the rate of molecules hitting the surface for P$_{N_2}$ and P$_{CN}$ (note that $u$ is given in units of saturation coverage), $P_{N_2}$ and $P_{CN}$ are the partial pressures induced by bond dissociation and $S_a$ and $S_{v}$ are the lattice cross-sections. $k_d$ and $k_r$ represent the desorption and reaction rates. The reaction rate and other parameters were taken from mechanistic studies based on a comparison between experimental data and theoretical modeling using the Arrhenius equation.\cite{huang1987detection, maharrey2005thermal,melius1987j}
In Eq. 2, $h$ represents the induced compression wave. The system has two solutions for the up and down states. The localized functions $a_0$ and $a_1$ are defined through the decomposition properties of the material. However, by arbitrary definition, if $a_0=0$, the stationary front solution loses its stability to Pitchfork bifurcation for $\epsilon > \epsilon_c( \delta)$.\cite{hagberg1996controlling} \newline
Arbitrary analysis of Eq. \ref{bif1} and \ref{bif2} reveals a bistability.
The parameter $\delta$ affects the relative spatial extent of the front.
$D$ in Eq. 3 is composed of the diffusion parameter of diffusive products in the nitroguanidine crystal and is given in a nonlinear form through a Boltzmann--Gibbs distribution:
\begin{equation}
\delta_t \rho_{xx} =D\bigtriangledown^2 v_{xx}
\label{bif3}
\end{equation}
The anomalous diffusion equation is dependent on how it is applied. Here, the diffusion model was developed through a nonlinear system.\cite{plastino1995non} In this case because of the introduction of external forces to the reaction-diffusion system, the same approach has to be applied to the diffusion parameter. The nonlinear system defined is based on a Tsallis formalism or a standard thermostatic Boltzmann--Gibbs equation\cite{tsallis1996anomalous, huang1987statistical} for diffusion of the decomposition products in lenses. Therefore, $v_{xx}$ is defined by\cite{lenzi2005nonlinear}:
\begin{equation}
v_{xx}(r)=r^{-2} exp_q[-\L(r)/z]
\label{bifdif1}
\end{equation}
The term $exp_q$ exists for two conditions of $q=1$ and $q>1$ \cite{plastino1995non}. If right-hand-side divergence of the equation $v_{xx}(r)$ is assumed, the function at $q=1$ converges to the ordinary differential equation exponential function.\cite{naudts2010q}
The function $\L(r)$ is given in ref. \cite{assis2006nonlinear}. Assuming external forces are present, the function becomes:
\begin{equation}
v_{xx}(r)=r^{-2} exp((\frac{1}{D}) \int_0^r d\hat{r}-kr)
\label{bifdif2}
\end{equation}
With the exertion of external forces, the equation above needs to satisfy the developed boundary condition for the explicit time-dependent equation given by:
\begin{equation}
v_{xx}(r,t)= (\frac{1}{\phi(t)})(\frac{r}{\phi(t)})^{\frac{D}{2} -1} exp[(-\frac{1}{0.65D^2})(\frac{r}{\phi(t)})^{1.3D}]
\label{bifdif3}
\end{equation}
The solution to $\phi(t)$ is given in ref. \cite{assis2006nonlinear}. The diffusion parameter $D$ affects the total system behaviour, where $1<D<2$, $2<D<3$ or $D>3$, and leads to subdiffusive behaviour at the boundary and superdiffusive behaviour far from the edges. A simplified phase diagram for the relative diffusion parameter is given by figure \ref{difphase1}.\newline
The effect of external field can be defined for three regions: $(i)$ at the front corresponding to the down state invading the up state, $(ii)$ at the front corresponding to the up state invading the down state and $(iii)$ where both fronts exist. For a quantitative study of the front dynamics and relative pattern characteristics for the inner region, the front is defined at $x \rightarrow r= x-x_f(t)$, stretching the spatial coordinates according to $z=r/ \sqrt{2}$ and expanding $u$ and $v$ according to $\sqrt{\epsilon}$ steps. Here, $\epsilon$ corresponds to the ratio across the fronts. The initial value of $v$ was measured at the material interface with the nitroguanidine lens and $u$ was measured using $-tanh(z \sqrt{2})$. With a step size of $\sqrt{\epsilon}$ we have:
\begin{equation}
\delta^2_z u_1+u-3u_0^2 u_1=v_1-x_fu_0z
\label{bif4}
\end{equation}
The solution to Eq. \ref{bif4} yields the nonequilibrium Ising-Bloch bifurcation (NIB) velocity, which is calculated for the energetic diffusive molecules using:
\begin{equation}
C_0=\frac{-3}{\eta \sqrt{2}}v_f
\label{bif5}
\end{equation}
The total normal front velocity can be found using:
\begin{equation}
C_n=C_0-k
\label{bif6}
\end{equation}
where $k$ is the curvature induced by compression waves during outburst. By calculating the front velocity ($C_n$), the effect of an external field at the front residue can also be analysed. As previously mentioned, the compression wave is treated as an external field which accelerates the flame velocity and diffusion in the lens and forms a Pitchfork bifurcation. The compression wave front in a heterogeneous solid is more complex than in a homogeneous solid because of the presence of voids, grains and flaws.\cite{tarver1993energy} A good approximation of the compression wave shape function in a homogeneous system can be given using a similar front to those developed for diffusion systems. Figure \ref{difphase1}b demonstrates the homogeneous compression fronts introduced to the system.
The compression waves interact with the lower and upper boundaries at $h \approx 0$, which is the termination point. A dynamic equation for $v_f$ requires the branches to be scaled spatially. By assuming un-stretched coordinates:
\begin{equation}
v_t-C_nv_y=u \pm (v) - a_1( k_r (v + v_{yy})- \beta(y+ \frac{x_f \sqrt{\epsilon}}{\eta} )+h
\label{bif7}
\end{equation}
Here, $\beta$ represents the compression wave front. By substituting Eq. \ref{bif6} into Eq. \ref{bif7}:
\begin{equation}
v_t=u \pm (v) - a_1( k_r (v + v_{yy})- \beta(y+ \frac{x_f \sqrt{\epsilon}}{\eta} )+h +C_n
\label{bif8}
\end{equation}
Initially $C_n$ is approximated at $v(0,t)$.
\begin{figure}
\centering
\begin{minipage}[ht]{.5\textwidth}\centering
\includegraphics[trim= 5mm 0 0 0 , scale=0.50]{figure3a.eps}
\end{minipage}\hfill
\begin{minipage}[ht]{.5\textwidth}\centering
\includegraphics[trim= 5mm 0 0 0 , scale=0.48]{figure3b.eps}
\end{minipage}\hfill
\caption{a) Illustration of the front solution for Ising--Bloch (diamonds) and NIB bifurcation. The compression wave drives the system into instability through a metastable phase. b) Front solution as a function of time for an oscillatory unstable front.}
\label{difphase}
\end{figure}
The front solution to Eq. \ref{bif8}a is given for different diffusivity phases. Regions $(A)$ and $(D)$ show opposing behaviour of decreasing and increasing kinetics of the Bloch fronts through a metastable region $(B)$. Introducing the compression waves leads the system to instability (region $(D)$) causing faster energy release. \newline
Finally, oscillation in the model is produced so that the validity of the system for an unstable nonlinear condition can be approximated using Fig. \ref{difphase}b.
\section{\label{sec:level1}Discussion \lowercase{} }
The production of wave fronts in the nitroguanidine lens shows inhibitor production giving rise to two additional fronts at the chemical fixed point (nitroguanidine). This provides a predictive method for the future design of such a system where guided wave propagation in the medium is necessary. It was also shown that the front-like structure connects the two fronts in a homogeneous steady state along the fixed chemical point such that bifurcation occurs in the medium. However, the kinetics of phase transition in nitroguanidine are too complex to be described by a steady state form of equation because turbulent waves are formed upon the diffusion of decomposition at microseconds of diffusion, and fronts lose their stability to lattice deformation. However, an unstable solution could be obtained by introducing a sticking variable to Eq. \ref{bif2} to provide phase transformation instability to the solution, as demonstrated in Fig. \ref{difphase}a. Phase transformation therefore changes the faceting by influencing the adsorption rate, lattice heat increase, and extended induced partial bond dissociation. This also causes further external periodic perturbations in a metastable condition. \newline
Introducing the nonlinear form of diffusion increased the efficiency of such an approach if compared with the standard model.
\section{\label{sec:level1}Conclusion \lowercase{} }
The diffusion of the decomposition products in a nitroguanidine lens was studied in a chaotic system. The diffusion of the composites in a wave shaper was described by a two-variable model with inhibition and production in the reaction diffusion system. The calculation of wave propagation showed a good predictive behaviour for microseconds after the disruption in nitroguanidine.
\nocite{*}
\section{\label{sec:level1}References \lowercase{} }
\providecommand{\noopsort}[1]{}\providecommand{\singleletter}[1]{#1}%
|
\section{Introduction}\label{sec1}
An $m\times n$ row-Latin rectangle is an $m\times n$ array in which each of the numbers $1,\ldots,n$ appear exactly once in each row. A \emph{partial transversal of size} $k$ in an $m\times n$ row-Latin rectangle $R$ is a set of $k$ entries of $R$ such that no two of them are in the same row or in the same column. If all the elements in a partial transversal are distinct we call it a \emph{partial rainbow transversal}.
Drisko \cite{Drisko98} proved the following elegant result:
\begin{thm}\label{thm1:1}
Any $(2n-1)\times n$ row-Latin rectangle has a partial rainbow transversal of size $n$.
\end{thm}
In the same paper Drisko also gave an example of a $(2n-2)\times n$ row-Latin rectangle in which there is no partial rainbow transversal of size $n$.
Given sets $F_1, \ldots ,F_m$, a {\em partial rainbow function} is a partial choice function of the sets $F_i$.
A {\em partial rainbow set} is the range of a partial rainbow function. If the sets $F_1, \ldots ,F_m$ are matchings in a given bipartite graph, then a {\em partial rainbow matching} is a partial rainbow set which is also a matching. Drisko's result asserts that any $2n-1$ matchings of size $n$ in a bipartite graph with $2n$ vertices have a partial rainbow matching of size $n$.
In this paper we generalize Drisko's theorem to matroids.
We provide some definitions and notation concerning matroids. For a set $A$ and an element $x$ we use the notation $A+x$ for $A\cup\set{x}$ and $A-x$ for $A\setminus\set{x}$. A collection $\M$ of subsets of a \emph{ground set} $\s$ is a \emph{matroid} if it is hereditary and it satisfies an augmentation property: If $A, B\in\M$ and $|B|>|A|$, then there exists $x\in B\setminus A$ such that $A+x\in \M$. Sets in $\M$ are called {\em independent} and subsets of $\s$ not belonging to $\M$ are called {\em dependent}.
An element $x\in \s$ is \emph{spanned} by $A$ if either $x\in A$ or $I+x\not\in \M$ for some independent set $I\subseteq A$.
The set of elements that are spanned in $\M$ by $A$ is denoted by $\spn_\M(A)$.
A \emph{circuit} is a minimal dependent set. For more background on matroid theory the reader is referred to in the books of Oxley~\cite{Oxley11} and Welsh~\cite{Welsh76}.
Chappell \cite{Chappell99} proved the following generalization of Drisko's theorem:
\begin{thm}\label{thm1:2}
Any $(2n-1)\times n$ array whose entries are taken from the ground set of a matroid $\M$, and whose rows are independent sets in $\M$, has a partial transversal of size $n$ which is an independent set of $\M$.
\end{thm}
In this paper we prove the following:
\begin{thm}\label{thm1:3}
Let $\M$ and $\N$ be two matroids on the same ground set $\s$. Any $2n-1$ sets of size $n$, each in $\M\cap\N$, have a partial rainbow set of size $n$ in $\M\cap\N$.
\end{thm}
Note that Theorem~\ref{thm1:1} is obtained from Theorem~\ref{thm1:3} by taking both $\M$ and $\N$ to be partition matroids, and Theorem~\ref{thm1:2} follows in the case that one of the matroids is a partition matroid.
\section{Proof of Theorem~\ref{thm1:3}}
We shall use the following basic facts on matroids:
\begin{fact}\label{fact1}
If $I\in\M$ and $I+x\not\in\M$, there exists a unique minimal subset of $I$, which we denote by $C_\M(I,x)$, that spans $x$, and for each $a\in C_\M(I,x)$, we have $I+x-a\in\M$ and $\spn_\M(I+x-a)=\spn_\M(I)$.
\end{fact}
Fact~\ref{fact2} is an immediate consequence of the augmentation property:
\begin{fact}\label{fact2}
Let $I$ and $J$ be independent sets in $\M$. If $|I|<|J|$, then there exists $J_1\subseteq J\setminus I$ such that $|J_1|\ge |J|-|I|$ and $I\cup J_1\in\M$.
\end{fact}
Fact~\ref{fact3} is also known as the \emph{circuit elimination axiom}:
\begin{fact}\label{fact3}
If $C_1$ and $C_2$ are circuits with $e\in C_1\cap C_2$ and $f\in C_1\setminus C_2$ then there exists a circuit $C_3$ such that $f\in C_3\subseteq (C_1\cup C_2)-e$.
\end{fact}
We shall also need the following lemma from \cite{AKZ14}. The proof is repeated here in order to make the discussion self-contained.
\begin{lem}\label{lem3}
Let $\M$ be a matroid. Let $I\in\M$ and $X=\set{x_1,\ldots,x_k}\subseteq I$ and $Y=\set{y_1,\ldots,y_k}\subseteq \spn_\M(I)\setminus I$ be such that $\spn_\M\left(\left(I\setminus X\right) \cup Y\right)=\spn_\M(I)$.
Suppose $y_{k+1}\in \spn_\M(I)\setminus I$ and $x_{k+1}$ are such that $x_{k+1}\in C_\M(I,y_{k+1})\setminus X$ and $x_{k+1}\not\in C_\M(I,y_i)$ for all $i=1,\ldots,k$. Then $x_{k+1}\in C_\M(\left(I\setminus X\right) \cup Y,y_{k+1})$.
\end{lem}
\begin{proof}
Suppose, for contradiction, that $x_{k+1}\not\in C_\M(\left(I\setminus X\right) \cup Y,y_{k+1})$.
Let $C_1= C_\M(I,y_{k+1})+y_{k+1}$ and $C_2= C_\M(\left(I\setminus X\right) \cup Y,y_{k+1})+y_{k+1}$. Then, by Fact~\ref{fact3}, there exits a circuit $C^1\subseteq C_1\cup C_2$, such that $x_{k+1}\in C^1$ and $y_{k+1}\not\in C^1$.
Since $I$ is independent, $C^1$ must contain at least one element $y_j\in Y$. Let $C_3=C_\M(I,y_j)+y_j$. Since $x_{k+1}\not\in C_\M(I,y_j)$, there exists a circuit $C^2\subseteq C^1\cup C_3$ such that $x_{k+1}\in C^2$ and $y_{j}\not\in C^2$, by Fact\ref{fact3}. Since $C^2\cap Y\subset C^1\cap Y$ we must have $|C^2\cap Y|<|C^1\cap Y|$. We proceed this way until we obtain a circuit whose intersection with $Y$ is empty. This will contradict the independence of $I$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm1:3}]
Let $A_1,\ldots,A_{2n-1}\in \M\cap \N$, each of size $n$. Let $R\in \M\cap\N$ be a partial rainbow set of maximal size. Assume, for contradiction, that $|R|<n$. Without loss of generality we may assume that $R\cap A_i=\emptyset$ for $i=1,\dots,n$. We define,
\begin{defn}
A \emph{colorful alternating trail} (CAT) of length $k$ ($1\le k\le n-1$) relative to $R$ consists of
a set $\{a_1,\ldots, a_k\}$, where distinct $a_i$'s belong to distinct $A_j$'s ($j\in \{1,\ldots,n-1\}$) and a set $\{r_1,\ldots, r_k\}\subset R$, such that
\begin{enumerate}
\item [(P$_\M$)]
$R+a_1-r_1+a_2-r_2+\cdots -r_{i-1}+a_i\in \M$ for all $i=1,\ldots,k$.
\item [(P$_\N$)]
$R+a_1-r_1+a_2-r_2+\cdots +a_i-r_i\in \N$ and $\spn_\N(R+a_1-r_1+a_2-r_2+\cdots+a_i-r_i)=\spn_\N(R)$ for all $i=1,\ldots,k$.
\end{enumerate}
If, in addition, $R+a_1-r_1+a_2-r_2+\cdots +a_{k-1}-r_{k-1}+a_k\in \N$ then the CAT is called \emph{augmenting} (in this case the condition (P$_\N$) for $i=k$ is redundant).
\end{defn}
Note that since $R$ is of maximal size, no augmenting CAT relative to $R$ exists. The theorem will be proved by showing that the assumption $|R|<n$ yields a partial rainbow set of size $|R|+1$ in $\M\cap\N$.
\begin{claim}\label{claim:1}
For each $k=1,\ldots, n-1$, the CATs involving only elements of $A_1,\ldots,A_k$ contain at least $k$ distinct elements of $R$.
\end{claim}
\begin{proof}[Proof of Claim~\ref{claim:1}]
\renewcommand{\qedsymbol}{}
We prove Claim~\ref{claim:1} by induction on $k$. Since $|R|<n$ and $|A_1|=n$ there exists an element $a_1\in A_1$ such that $R+a_1\in \M$. By the maximality property of $R$ we must have $R+a_1\not\in \N$. By Fact~\ref{fact1}, there exists an element $r_1\in R$ such that $R+a_1-r_1\in \N$ and $\spn_\N(R+a_1-r_1)=\spn_\N(R)$. Thus, $\{a_1\}$ and $\{r_1\}$ form a CAT of length 1 and Claim~\ref{claim:1} holds for $k=1$.
Now suppose Claim~\ref{claim:1} holds for $k-1$ for some $k\ge 2$. Without loss of generality we may assume that the CATs involving only elements of $A_1,\ldots,A_{k-1}$ contain the elements $r_1,\ldots,r_{k-1}\in R$. Let $R_{k-1}=\{r_1,\ldots,r_{k-1}\}$.
Since $|R|<n$, Fact~\ref{fact2} implies that the set $A_k$ contains at least $k$ elements that are not in $\spn_\M(R\setminus R_{k-1})$. Since $|R_{k-1}|=k-1$, at least one of these elements is not in $\spn_\N(R_{k-1})$. Let $a$ be such an element. That is, $a\in A_k$ and
\begin{equation}\label{eq1:1}
a\not\in\spn_\M(R\setminus R_{k-1})
\end{equation}
and
\begin{equation}\label{eq1:2}
a\not\in\spn_\N(R_{k-1}).
\end{equation}
First assume $a\not\in\spn_\M(R)$. Then $R+a\in \M$. Since $R$ is of maximal size it follows that $R+a\not\in \N$. By (\ref{eq1:2}), $C_\N(R,a)\not\subseteq R_{k-1}$ and thus, there exists an element $r\in R\setminus R_{k-1}$ such that $R+a-r\in \N$ and $\spn_\N(R+a-r)=\spn_\N(R)$. The CAT consisting of $\{a\}$ and $\{r\}$ contains the extra element $r\not\in R_{k-1}$, and thus Claim~\ref{claim:1} holds for $k$.
Now assume that $a\in\spn_\M(R)$. By (\ref{eq1:1}), there exists $r'\in C_\M(R,a)\cap R_{k-1}$. By the definition of $R_{k-1}$, there exists a CAT containing $r'$, say
\begin{equation}\label{eq1:3}
\{a_{i_1},\ldots, a_{i_l}\} \Text{ and } \{ r_{i_1},\ldots,r_{i_l}=r'\}.
\end{equation}
We may assume that none of $r_{i_1},\ldots,r_{i_{l-1}}$ is in $C_\M(R,a)$ (otherwise we take $r'$ to be the first of $r_{i_1},\ldots,r_{i_{l-1}}$ that belongs to $C_\M(R,a)$). Let $R'=R+a_{i_1}-r_{i_1}+\cdots-r_{i_{l-1}}+a_{i_l}$. Since $R'\in\M$ and no element of $C_\M(R,a)$ was discarded along the trail from $R$ to $R'$, we have $C_\M(R',a)=C_\M(R,a)$. Hence, $r'\in C_\M(R',a)$. By Fact~\ref{fact1} we have
\begin{equation}\label{eq1:4}
R'-r'+a\in \M.
\end{equation}
If $a\not\in\spn_\N(R)$ then $R+a\in \N$. By Property (P$_\N$) we have $\spn_\N(R'-r')=\spn_\N(R)$. Thus, $R'-r'+a\in \N$, which, together with (\ref{eq1:4}), yields an augmenting CAT, contrary to the maximality of $|R|$.
Hence, we may assume that $a\in\spn_\N(R)$. By (\ref{eq1:2}), there exists an element $r''\in C_\N(R,a)\setminus R_{k-1}$.
If none of the $a_{i_j}$ ($j=1,\ldots,l$) in (\ref{eq1:3}) satisfies $r''\in C_\N(R,a_{i_j})$, then, by Lemma~\ref{lem3}, $r''\in C_\N(R'-r',a)$, and thus, the sets $\{a_{i_1}, \ldots, a_{i_l},a\}$ and $\{r_{i_1},\ldots,r',r''\}$ make a CAT which contains the extra element $r''\not\in R_{k-1}$, proving Claim~\ref{claim:1} for $k$.
Otherwise, let $j$ be minimal such that $r''\in C_\N(R,a_{i_j})$ and let $R''=R+a_{i_1}-r_{i_1}+\cdots+a_{i_{j-1}}-r_{i_{j-1}}$. Then, by Lemma~\ref{lem3}, it follows that $r''\in C_\N(R'',a_{i_j})$. Hence, $R''+a_{i_j}-r''\in\N$ and $\spn_\N(R''+a_{i_j}-r'')=\spn_\N(R'')=\spn_\N(R)$. We obtain a CAT consisting of $\{a_{i_1}, \ldots, a_{i_{j-1}}, a_{i_j}\}$ and $\{r_{i_1},\ldots,r_{i_{j-1}},r''\}$. This CAT involves the extra element $r''\not\in R_{k-1}$. Thus, the CATs involving only elements of $A_1,\ldots,A_{k-1},A_k$ contain at least $k$ elements of $R$, proving Claim~\ref{claim:1} for $k$. This completes the proof of Claim~\ref{claim:1}.
\end{proof}
To conclude the proof of Theorem~\ref{thm1:3}, note that by Claim~\ref{claim:1}, the CATs involving only elements of $A_1,\ldots, A_{n-1}$ contain all the elements of $R$. Since $|R|<n$ there exists an element $a_n\in A_n$ such that $R+a_n\in \N$. By the maximality property of $R$, we must have that $a_n\in \spn_\M(R)$. Let $r\in R$ satisfy $R+a_n-r\in \M$.
Applying Claim~\ref{claim:1} to the case $k=n-1$, it follows that $r$ belongs to a CAT consisting of some sets $\{a_{i_1}, \ldots, a_{i_l}\}$ and $\{r_{i_1},\ldots,r_{i_{l-1}},r_{i_l}=r\}$, where each of $a_{i_1}, \ldots, a_{i_l}$ belongs to a different set among $A_1,\ldots,A_{n-1}$.
Let $R'=R+a_{i_1}-r_{i_1}+\cdots-r_{i_{l-1}}+a_{i_l}$. We may assume that $r$ is the first element in this trail satisfying $r\in C_\M(R,a_n)$ (otherwise we take $r$ to be the first element in the trail belonging to $C_\M(R,a_n)$) and thus, $C_\M(R',a_n)=C_\M(R,a_n)$.
Hence, $R'-r+a_n\in \M$. Since $\spn_\N(R'-r)=\spn_\N(R)$, by Property ($P_\N$), we also have $R'-r+a_n\in \N$. Thus, $R'-r+a_n$ is a rainbow matching of size $|R|+1$.
\end{proof}
\providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}
\providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR }
\providecommand{\MRhref}[2]
\href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2}
}
\providecommand{\href}[2]{#2}
|
\section{Introduction}
In their seminal 1986 paper~\cite{KPZ86}, Kardar, Parisi and Zhang (KPZ) proposed the stochastic evolution equation for a height function $\mathcal{H}(T,X)\in \R$ ($T\in \Rplus$ is time and $X\in\R$ is space)
\begin{equation*}
\partial_T \mathcal{H}(T,X)= \tfrac12 \partial_X^2 \mathcal{H}(T,X)+\tfrac12\left(\partial_X \mathcal{H}(T,X)\right)^2+\xi(T,X).
\end{equation*}
The randomness $\xi$ models the deposition mechanism and it is taken to be space-time Gaussian white noise, so that formally $\EE[\xi(T,X)\xi(S,Y)] = \delta(T-S)\delta(X-Y)$.
The Laplacian reflects the smoothing mechanism and the non-linearity reflects the slope-dependent growth velocity of the interface. Using earlier physical work of Forster, Nelson and Stephen~\cite{FNS77}, KPZ predicted that for large time $T$, the height function $\mathcal H(T,X)$ has fluctuations around its mean of order $T^{1/3}$ with spatial correlation length of order $T^{2/3}$. Since then, the exact nature of these fluctuations has been a subject of extensive study. For additional background, see the reviews~\cite{Cor11,Q14,SS10c}.
For general initial data, it is expected that the solutions to the KPZ equation are locally Brownian in space~\cite{QR12, CH13,Hai11}. Therefore, making direct sense of the non-linearity in the equation is a challenge~\cite{BC95,Hai11}. The physically relevant notion~\cite{BG97,ACQ10,Cor11,Q14,Hai11,Dem13,SS10b} of a solution to the KPZ equation is therefore defined indirectly via the well-posed stochastic heat equation (SHE) with multiplicative noise,
\begin{equation*}
\partial_T \mathcal{Z}(T,X)=\tfrac12\partial_X^2 \mathcal{Z}(T,X)+\mathcal{Z}(T,X)\xi(T,X)
\end{equation*}
with initial condition $\mathcal{Z}(0,X)=\mathcal{Z}_0(X)=e^{\mathcal{H}(0,X)}$. The Cole--Hopf solution of the KPZ equation is then defined as $\mathcal{H}(T,X)=\ln(\mathcal{Z}(T,X))$. On account of this definition, we will talk about the SHE and KPZ equation interchangeably, stating most of our main results (with the exception of those in this first section) in terms of the SHE.
By a version of the Feynman--Kac formula, the solution of the SHE can be formally written as
\begin{equation*}
\mathcal{Z}(T,X) =\EE_{T,X}\left[\mathcal{Z}_0(b(0))\, : \exp:\left\{-\int_0^T \xi(b(S),S) \d S\right\}\right]
\end{equation*}
where the expectation $\EE_{T,X}$ is over a Brownian motion $b(\cdot)$ going backwards in time from $b(T)=X$, and where $:\exp:$ is the Wick ordered exponential~\cite[Section~4.2]{Cor11}. This provides an interpretation for $\mathcal{Z}(T,X)$ as the partition function of the continuum directed random polymer (CDRP)~\cite{ACQ10, AKQ12b}.
Formally, the spatial derivative $\mathcal{U}(T,X)=\partial_X \mathcal{H}(T,X)$ of the KPZ equation satisfies the stochastic Burgers equation
\begin{equation*}
\partial_T \mathcal{U}(T,X)= \tfrac12 \partial_X^2 \mathcal{U}(T,X)+\tfrac12 \partial_X \big(\mathcal{U}(T,X)\big)^2+\partial_X\xi(T,X),
\end{equation*}
which can be thought of as a continuum version of an interacting particle system~\cite{BG97,BQS11}.
Let $B(X)$ be a two-sided Brownian motion with $B(0)=0$ and zero drift. Stationary (zero drift) initial data $\mathcal{H}(0,X)=B(X)$ for the KPZ equation corresponds with SHE initial data $\mathcal{Z}(0,X) = e^{B(X)}$ and stochastic Burgers equation initial data $\mathcal{U}(0,X)=\partial_X B(X)$. This is called stationary, because for any later time $T$, $\mathcal{U}(T,\cdot)$ is marginally distributed as another spatial Gaussian white-noise. In terms of the KPZ equation, for fixed $T>0$, $\mathcal{H}(T,\cdot)$ is marginally distributed as $\tilde{B}(\cdot)+ \mathcal{H}(T,0)$ where $\tilde{B}(\cdot)$ is a two-sided Brownian motion (though not independent of $B$ or $\mathcal H(T,0)$).
The first rigorous confirmation of the $T^{1/3}$ fluctuation scale prediction for the KPZ equation was provided by~\cite{BQS11}, showing that there exist constants $c_0>0$ and $0<c_1<c_2<\infty$ such that for all $T>c_0$,
\begin{equation*}
c_1 T^{2/3} \leq \var\big(\mathcal{H}(T,0)\big) \leq c_2 T^{2/3}.
\end{equation*}
A similar fluctuation scale result was demonstrated recently in~\cite{CH13} (and applies equally well for a broad class of KPZ initial data) based on the KPZ line ensemble construction.
The present work provides an exact formula for the one-point probability distribution of the stationary solution to the KPZ equation, and a limit theorem for $\mathcal{H}(T,X)$ after proper centering and scaling by $T^{1/3}$. The following theorem and corollary are special cases (drift $b=0$ and position $X=0$) of Theorem~\ref{ThmFormulaStationary}, Proposition~\ref{PropInverseMelling} and Theorem~\ref{CorUniversality}.
\begin{theorem}\label{ThmFormulaStationaryIntroVersion}
Let $\mathcal{H}(T,X)$ be the stationary (zero drift) solution to the KPZ equation and let $\BesselK_0$ denote the modified Bessel function~\cite{AS84}. Then, for $T>0$, $\sigma=(2/T)^{1/3}$ and $S>0$,
\begin{equation*}
\EE\left[2\sigma \BesselK_0\left(2\sqrt{S \,\exp\big\{\tfrac{T}{24} +\mathcal{H}(T,0)\big\}}\right)\right]=\Xi\left(S,0,\sigma\right),
\end{equation*}
where the function $\Xi$ is given in Definition~\ref{longdef}.
Equivalently, for any $r\in\R$, we have
\begin{equation*}
\PP\left(\frac{\mathcal{H}(T,0)+\frac{T}{24}}{(T/2)^{1/3}}\le r\right)
=\frac{1}{\sigma^2}\frac{1}{2\pi\I}\int_{-\delta+\I\R} \frac{\d \xi}{\Gamma(-\xi)\Gamma(-\xi+1)} \int_{\R} \d x\, e^{x\xi/\sigma} \Xi\left(e^{-\frac{x+r}\sigma},0,\sigma\right)
\end{equation*}
for any $\delta>0$.
\end{theorem}
\begin{theorem}\label{CorUniversalityIntroVersion}
For any $r\in\R$,
\begin{equation*}
\lim_{T\to\infty}\PP\left(\frac{\mathcal{H}(T,0)+\frac{T}{24}}{(T/2)^{1/3}}\le r\right)=F_0(r),
\end{equation*}
where $F_0$ is given in Definition~\ref{anotherlongone} with $\tau=0$.
\end{theorem}
Inherent in the work of KPZ was the premise that a larger class of growth processes than just their eponymous equation should display the same $T^{1/3}$ and $T^{2/3}$ scaling exponents. The class of such models is referred to as the KPZ universality class. Generally speaking, the universality belief is that a growth model will belong to the KPZ class if it has the same physical properties as the KPZ equation, namely local growth dynamics, a smoothing mechanism and irreversibility arising from the condition that the speed of growth as a function of the slope has non-zero second derivative.
It took a quarter of a century to prove that the KPZ equation was in the KPZ universality class itself (via demonstrating the $1/3$ and $2/3$ exponents)~\cite{BQS11,ACQ10,CQ10,BC11,BCF12,CH13,SS10b}. Before this, starting with the 1999 work of~\cite{BDJ99,Jo00b}, a few growth models in the KPZ universality class were rigorously analyzed. These models were the polynuclear growth model (PNG), totally asymmetric simple exclusion process (TASEP) and last passage percolation (LPP) with special exponential, geometric or Bernoulli weights. Beyond the $T^{1/3}$ and $T^{2/3}$ scaling, the limit distributions and spatial processes for these models were determined. These statistical properties agreed between the models, but depended non-trivially on the type of initial data or geometry for the growth models, such as curved~\cite{BDJ99,Jo00b,BR99,PS02,Jo03,BF08,BF07}, flat~\cite{BR99,Sas05,
BFPS06,BF07,Fer04,BFP06,FSW13,BFS07b} or stationary~\cite{BR00,SI04,PS02b,FS05a,BFP09,FSW14,BFP12}. All these results strongly used the underlying determinantal structure that these models all enjoy (see the reviews~\cite{Fer07,Fer10b,Cor11,BG12,Fer13} for further references and details).
The KPZ equation does not seem to have a full-blown determinantal structure (as opposed to PNG, TASEP and LPP). However, in the last few years a number of new exactly solvability methods have been developed which have led to explicit formulas for the one-point marginal distribution of the solution to the KPZ equation with specific types of initial data and verified the $1/3$ exponent for general initial data (also the $2/3$ exponent has been verified for specific initial data). With the exception of the non-rigorous replica method (method 2, below), the other (rigorous) methods have all proceeded via analysis of exactly solvable discretizations or regularizations of the KPZ equation such as the (partially) asymmetric simple exclusion process (ASEP), the $q$-deformed totally asymmetric simple exclusion process ($q$-TASEP), or the O'Connell-Yor semi-discrete directed random polymer (see the review~\cite{Cor14} and references therein). These stochastic processes converge to the KPZ equation under special {\it weakly asymmetric} or {\it weak noise} scaling. It should be emphasized that the developed methods are presently only adapted to study certain types of initial data (except in the case of method 5, the KPZ line ensemble). As we summarize them below (for a partial list of references to subsequent developments and extensions, see~\cite{Cor14}), we will first focus on narrow wedge initial data for the KPZ equation, which means starting the SHE with $\mathcal{Z}(0,X)=\delta_{X=0}$.
\begin{enumerate}
\item~\cite{TW08,TW08b,TW08c} used Bethe ansatz to compute transition probabilities for the $N$-particle ASEP, extracted a one-point marginal distribution formula suitable for the $N$ to infinity limit corresponding with step initial data, and manipulated the resulting formula into a Fredholm determinant formula amenable to asymptotic analysis. This served as the starting point for the rigorous derivation in~\cite{ACQ10} of the one-point distribution for the KPZ equation with narrow-wedge initial data (see also~\cite{SS10b} for a parallel and independent, though non-rigorous, derivation of this).
\item~\cite{Dot10} and~\cite{CDR10} computed exact formulas for moments of the SHE with $\mathcal{Z}(0,X)=\delta_{X=0}$ using the connection with the delta Bose gas and the Bethe ansatz. From these moments they derived a formula for the Laplace transform of $\mathcal{Z}(T,X)$ and hence, by inverting the transform, the distribution of $\mathcal{H}(T,X)$. This physics replica method derivation suffers from being quite non-rigorous since the moments, in fact, grow too fast to determine the Laplace transform and distribution.
\item~\cite{BC11} introduced Macdonald processes and connected them to certain 2d growth processes (and 1d marginals like $q$-TASEP) as well as provided exact Fredholm determinant formulas for one-point distributions amenable to asymptotic analysis. A limit transition connects these processes to the Whittaker processes which, in~\cite{OCon09}, had been introduced and related to the O'Connell--Yor semi-discrete directed random polymer via a geometric lifting of the RSK correspondence. This method was used in~\cite{BCF12} to rederive the narrow wedge KPZ one-point distribution formula.
\item~\cite{BCS12} used Markov dualities of ASEP and $q$-TASEP, as well as the Bethe ansatz to compute explicit formulas for expectations of a large class of observables of these models, when started from step initial data. From these expectations, they derived a formula for a $q$-deformed Laplace transform of the one-point distribution. This provides an alternative to the methods of~\cite{TW08,TW08b,TW08c} as well as a rigorous regularization of the replica method used in~\cite{Dot10} and~\cite{CDR10}.
\item~\cite{CH13} constructed a line ensemble extension to the fixed time $T$ solution to the narrow wedge initial data KPZ equation which enjoys a distributional invariance called the $H$-Brownian Gibbs property as well as certain uniform regularity under $T^{1/3}, T^{2/3}$ scaling as $T$ goes to infinity. From this they proved the validity of the $2/3$ spatial exponent for the narrow wedge initial data KPZ equation and proved the $1/3$ fluctuation exponent for a wide class of KPZ initial data.
\end{enumerate}
A brief review and comparison of methods 1, 2 and 4 can be found in~\cite{ICtwowaystosolveASEP}, whereas some aspects of method 3 are reviewed in~\cite{BG12,BP13}. See also the review~\cite{Cor14}.
Besides the narrow-wedge initial data, there are a few other types of initial data for which these methods have proved successful in computing exact formulas for KPZ equation one-point distributions.
\begin{enumerate}
\item Half Brownian KPZ initial data corresponds with $\mathcal{Z}(X,0)=e^{B(X)}\Id_{X\geq 0}$, $B(\cdot)$ being a one-sided Brownian motion. It was rigorously analyzed via method 1 in~\cite{CQ10} and method 3 in~\cite{BCF12}, as well as non-rigorously analyzed via method 2 in~\cite{SI11}. A family generalizing half Brownian initial data was further rigorously analyzed via method 3 in~\cite{BCF12}.
\item Flat and half-flat KPZ initial data corresponds with $\mathcal{Z}(X,0) = 1$. It was non-rigorously analyzed via method 2 in~\cite{CLD11,CLD12,LD14}. No rigorous confirmation of these results have appeared yet.
\item Stationary KPZ initial data, the subject of this paper, corresponds with \mbox{$\mathcal{Z}(X,0)= e^{B(X)}$}, $B(\cdot)$ being a two-sided Brownian motion fixed at $B(0)=0$. It was non-rigorously analyzed via method 2 in~\cite{IS12}. In Remark~\ref{comparerem} we address the question of comparing the formula derived therein to that proved in Theorem~\ref{ThmFormulaStationaryIntroVersion}.
\end{enumerate}
Using these exact one-point formulas, it has further been confirmed in all of the above cases of initial data that the large $T$ one-point distribution converges to the same distribution as observed in the determinantal models of PNG, TASEP and LPP. Presently it is only for determinantal models that multi-point distributions and limit processes have been computed (see, however, nonrigorous work of~\cite{SP11,Dot13}). Besides the specific types of initial data discussed above, using method 5,~\cite{CH13} proved that up to certain rather weak hypothesis on initial data, the KPZ equation always has order $T^{1/3}$ fluctuations as $T$ goes to infinity.
\subsection{Outline}
In this paper we build on method 3, Macdonald processes, in order to prove Theorem~\ref{ThmFormulaStationaryIntroVersion}. It is not clear presently how to arrive at this result via the other rigorous methods (1, 4, or 5). Let us outline the main steps to prove Theorem~\ref{ThmFormulaStationaryIntroVersion} as well as make note of some of the other results of interest which we attain herein:
\medskip
\noindent {\bf Section~\ref{sectmodel}:} We introduce the O'Connell--Yor semi-discrete directed random polymer with log-gamma boundary sources and the associated multi-path extensions to its partition functions. Theorem~\ref{ThmFormulaSemiDiscrete} provides a Fredholm determinant formula for the Laplace transform of the partition function of the polymer model. Theorem~\ref{ThmFormulaContinuous} gives the analogue of Theorem~\ref{ThmFormulaSemiDiscrete}, but for the SHE/KPZ equation; Theorem~\ref{ThmFormulaStationary} gives a corresponding formula for the stationary version of the model; and Theorem~\ref{CorUniversality} demonstrates the KPZ universality ($T^{1/3}$ scaling and limiting one-point probability distribution) of the stationary model.
Theorem~\ref{ThmFormulaStationaryIntroVersion} is, in fact, a special case of Theorem~\ref{ThmFormulaStationary}. As such, the rest of the paper divides naturally into two parts. The first part, comprised of Sections~\ref{SectMacdonald},~\ref{SectConvWhitt}, and~\ref{SectSemiDirected}, provides a proof of Theorem~\ref{ThmFormulaSemiDiscrete}. The second part, comprised of Sections~\ref{SectCDRP},~\ref{SectStatSHE}, and~\ref{SectUniversality}, provides asymptotic analysis of our semi-discrete directed random polymer results to prove the SHE/KPZ equation results of Theorems~\ref{ThmFormulaContinuous} and~\ref{ThmFormulaStationary}, as well as Theorem~\ref{CorUniversality}.
\smallskip
\noindent {\bf Section~\ref{SectMacdonald}:} We introduce the $q$-Whittaker processes (equivalently, $t=0$ Macdonald processes) with $q\in(0,1)$. These are measures on interlacing partitions or Gelfand--Tsetlin patterns $\big\{\lambda^{(k)}_j\}_{1\leq j\leq k\leq N}$. For a certain class of $q$-Whittaker {\it nonnegative specializations} of the processes (indexed by $\tilde \alpha,\tilde \beta$ and $\tilde \gamma$ parameters) we prove Theorem~\ref{qFredDetThm}, a Fredholm determinant formula for the $e_q$-deformed Laplace transform of the random variable $q^{-\lambda^{(N)}_1}$. This is done following the general approach introduced in~\cite{BC11} and used there to prove a similar type of formula for $q^{\lambda^{(N)}_N}$. Unlike for $q^{\lambda^{(N)}_N}$, studied in~\cite{BC11}, $q^{-\lambda^{(N)}_1}$ is an unbounded random variable which only has finitely many moments. This would appear to be a major impediment in implementing the approach of~\cite{BC11} since it relies upon taking a generating
function of explicit formulas for moments $\EE\Big[\big(q^{-\lambda^{(N)}_1}\big)^k\Big]$ in order to recover the distribution. This issue of moment divergence does not arise for the so-called pure $\tilde \beta$ specializations, and so in that case we can follow the approach of~\cite{BC11} to prove this special case of Theorem~\ref{qFredDetThm} (this is recorded as Proposition~\ref{qFredDetThmbeta}). It is the $\tilde \alpha,\tilde \gamma$ specialization (for which moments diverge) which, however, we are really after due to its relationship with the semi-discrete directed random polymer with log-gamma boundary sources. In order to extend Theorem~\ref{qFredDetThm} to those specializations as well, we observe that the equality of the $\tilde \beta$ specialization $e_{q}$-Laplace transform with the corresponding Fredholm determinant actually implies a formal series (in Newton power sum symmetric polynomials) identity. The $\tilde \alpha,\tilde \gamma$ specialization of this identity yields convergent series on
both sides and
hence proves the equivalence of the $\tilde \alpha,\tilde \beta,\tilde \gamma$ specialized $e_{q}$-Laplace transform with the claimed Fredholm determinant in Theorem~\ref{qFredDetThm}. In this way, we see the power of relating our observable of interest, $q^{-\lambda^{(N)}_1}$, to the larger structure of $q$-Whittaker processes and symmetric polynomials.
This rigorous $e_q$-deformed Laplace transform derivation should be contrasted to the non-rigorous derivations (in method 2) of the Laplace transform of $\mathcal{Z}(T,0)$ from the moments $\EE\Big[\big(\mathcal{Z}(T,0)\big)^k\Big]$ which grow too quickly to uniquely identity the distribution. Under the various limit transitions which relate $q^{-\lambda_1^{(N)}}$ to $\mathcal{Z}(T,0)$ (i.e.\ Theorems~\ref{ThmConnectWhitpoly} and~\ref{ThmDiscreteToContinuous}) we lose the tools of symmetric functions which saved us. In particular, it is not clear how the $\tilde\beta$ specialization behaves under these limit transitions, and the notion of formal series identities seems to be lost.
\smallskip
\noindent {\bf Section~\ref{SectConvWhitt}:} We introduce Whittaker processes, measures on $\big\{T^{(k)}_j\}_{1\leq j\leq k\leq N}$. Theorem~\ref{ThmConnectWhitpoly} uses results of~\cite{OCon09,COSZ11} to relate these processes to the O'Connell--Yor semi-discrete directed random polymer with log-gamma boundary sources from Section~\ref{sectmodel}. In particular, this implies that the random variable $e^{T^{(N)}_1}$ (for a suitable Whittaker process specialization) and the polymer partition function $\mathbf{Z}^{N,M}(\tau)$ have the same distribution. Theorem~\ref{thmweakconv} shows how the $\tilde \alpha,\tilde \gamma$ specialized $q$-Whittaker processes converges as $q\to 1$ to the Whittaker processes (under special scaling), and thus (up to scaling) how $q^{-\lambda^{(N)}_1}$ converges to $e^{T^{(N)}_1}$. The pure $\tilde \gamma$ specialization version of this convergence result was proved as~\cite[Theorem~4.1.21]{BC11}, and the pure $\tilde \alpha$ specialization version was proved (modulo a decay
estimate which was not checked) as~\cite[Theorem~4.2.4]{BC11}. By combining these specializations, it becomes unnecessary to check the omitted decay estimate from~\cite[Theorem~4.2.4]{BC11}. So as not to be too obtuse, we provide the steps in this proof, even though they closely mimic those from~\cite{BC11}.
\smallskip
\noindent {\bf Section~\ref{SectSemiDirected}:}
We prove Theorem~\ref{ThmFormulaSemiDiscrete} by combining Theorem~\ref{ThmConnectWhitpoly} with Theorem~\ref{ThmFormulaWhit}. Theorem~\ref{ThmFormulaWhit} provides a Fredholm determinant formula for the Laplace transform under Whittaker processes of $e^{T^{(N)}_1}$. It is proved in this section by asymptotic analysis of the corresponding Fredholm determinant formula for the $e_q$-Laplace transform under $q$-Whittaker processes of $q^{-\lambda^{(N)}_1}$ (given as Proposition~\ref{propstep2}) along with the process convergence result of Theorem~\ref{thmweakconv}.
\smallskip
\noindent {\bf Section~\ref{SectCDRP}:} We turn here to studying the asymptotic behavior of the semi-discrete directed random polymer with boundary sources, as relates to the SHE/KPZ equation. Theorem~\ref{ThmDiscreteToContinuous} records a result of~\cite{QMR12} showing how the semi-discrete model converges to the SHE/KPZ equation. Theorem~\ref{ThmKbbeta} then provides the corresponding asymptotic analysis of the Fredholm determinant for the semi-discrete model with log-gamma boundary sources coming from Theorem~\ref{ThmFormulaSemiDiscrete}. These considerations prove Theorem~\ref{ThmFormulaContinuous} which gives the Laplace transform of $\mathcal{Z}_{b,\beta}(T,X)$, SHE/KPZ equation solution with initial data $\mathcal{Z}_0(X)=\exp(B(x))$ where $B(X)$ is a two-sided Brownian motion with drift $\beta$ on the left of $0$ and drift $b$ on the right of $0$ for $\beta>b$.
\smallskip
\noindent {\bf Section~\ref{SectStatSHE}:} We now take the limit as $\beta\searrow b$ in order to recover $\mathcal{Z}_{b}(T,X)$, the solution to the SHE/KPZ equation with stationary initial data $\mathcal{Z}_0(X)=\exp(B(x))$ where $B(X)$ is a two-sided Brownian motion with drift $b$ (on both sides). Taking the corresponding limit of Theorem~\ref{ThmFormulaContinuous} requires some care (in particular an analytic continuation argument similar to that used previously in the rigorous analysis of stationary TASEP in~\cite{FS05a} and in the non-rigorous replica analysis of the KPZ equation in~\cite{IS12}) and is given as Theorem~\ref{ThmFormulaStationary}.
\smallskip
\noindent {\bf Section~\ref{SectUniversality}:}
We prove Theorem~\ref{CorUniversality}, which demonstrates a universality result, namely that in the large time limit we recover the $T^{1/3}$ fluctuation scaling and the one-point probability distribution function previously obtained previously for stationary PNG and TASEP~\cite{BR00,SI04,PS02b,FS05a}.
\subsubsection*{Acknowledgments:}
The authors extend thanks to N. O'Connell, J. Quastel and T. Sasamoto for helpful discussions on this subject. A. Borodin was partially supported by the NSF grant DMS-1056390. I. Corwin was partially supported by the NSF grant DMS-1208998 as well as by Microsoft Research and MIT through the Schramm Memorial Fellowship, by the Clay Mathematics Institute through the Clay Research Fellowship and through the Institute Henri Poincare through the Poincare Chair. P.L. Ferrari was supported by the German Research Foundation via the SFB 1060--B04 project.
B.\ Vet\H o is grateful for the support of the Humboldt Research Fellowship for Postdoctoral Researchers during his stay at the University of Bonn and for the Postdoctoral Fellowship of the Hungarian Academy of Sciences.
His work was partially supported by OTKA (Hungarian National Research Fund) grant K100473.
\section{Models and main results}\label{sectmodel}
\subsection{Semi-discrete directed random polymer with boundary sources}
To obtain our main result, the one-point probability distribution functions for the stationary KPZ equation (Theorem~\ref{ThmFormulaStationaryIntroVersion} and more generally, Theorem~\ref{ThmFormulaStationary}), we start by studying a semi-discrete directed random polymer model with log-gamma boundary sources. This is a mixture of models introduced by O'Connell and Yor~\cite{OCY01} and Sepp\"{a}l\"{a}inen~\cite{Sep09}. Indeed, taking $M=0$ and $\tau>0$ recovers the semi-discrete directed random polymer of~\cite{OCon09} while taking $M>0$ and $\tau=0$ recovers the log-gamma discrete directed random polymer of~\cite{Sep09}.
\begin{figure}
\begin{center}
\psfrag{w11}[cc]{$\omega_{-1,1}$}
\psfrag{w21}[cc]{$\omega_{-2,1}$}
\psfrag{wM1}[cc]{$\omega_{-M,1}$}
\psfrag{w1N}[bc]{$\,\omega_{-1,N}$}
\psfrag{w2N}[bc]{$\!\omega_{-2,N}$}
\psfrag{wMN}[bc]{$\omega_{-M,N}$}
\psfrag{tn}[bc]{$(\tau,N)$}
\psfrag{B1}[lc]{$B_1$}
\psfrag{B2}[lc]{$B_2$}
\psfrag{B3}[lc]{$B_3$}
\psfrag{BN}[lc]{$B_N$}
\psfrag{s2}[cc]{$s_2$}
\psfrag{s3}[cc]{$s_3$}
\psfrag{sNm1}[cc]{$s_{N-1}$}
\psfrag{t}[cc]{$\tau$}
\psfrag{phi}[bc]{$\phi$}
\includegraphics[height=5cm]{FigSemiDiscreteDP}
\caption{Illustration of the semi-discrete directed random polymer with log-gamma boundary sources. The thick solid line is a possible directed random polymer path $\phi$ from $(-M,1)$ to $(\tau,N)$. Its energy is given by \eqref{eqEnergy}. The random variables $\omega_{-k,n}$ are distributed as $-\ln\Gamma(\alpha_k-a_n)$, while the Brownian motions $B_1,\ldots,B_N$ have drifts $a_1,\ldots,a_N$ respectively.}
\label{FigSemiDiscreteDP}
\end{center}
\end{figure}
For $\theta>0$, a random variable $X$ is distributed as $\Gamma(\theta)$ (written $X\sim \Gamma(\theta)$) if it has density with respect to Lebesgue measure given by
$$
\frac{\d}{\d x}\PP(X\le x) = \Id_{\{x>0\}} \frac{1}{\Gamma(\theta)}\, x^{-\theta-1}e^{-x}
$$
and a random variable $W$ is distributed as $-\ln\Gamma(\theta)$ (written $W\sim -\ln\Gamma(\theta)$) if $W = -\ln X$ for $X\sim \Gamma(\theta)$.
Fix $N\geq 1$ and $M\geq 0$. Let $a=(a_1,\ldots,a_N)\in \R^N$ and $\alpha = (\alpha_1,\ldots,\alpha_M)\in \big(\R_{>0}\big)^M$ be such that $\alpha_m-a_n>0$ for all $1\leq n\leq N$ and $1\leq m\leq M$.
Consider the setting as in Figure~\ref{FigSemiDiscreteDP}, where the horizontal axis is discrete on the left of $0$ and continuous on the right of $0$, while the vertical axis is discrete. In this semi-discrete setting we introduce randomness in the following way. For all $1\leq m\leq M$ and $1\leq n\leq N$ let $\omega_{-m,n}\sim -\ln\Gamma(\alpha_m-a_n)$ be independent log-Gamma random variables specified by the parameters $a,\alpha$; and for all $1\leq n\leq N$ let $B_n$ be independent Brownian motions with drift $a_n$. The $\omega_{-m,n}$ can be thought of as sitting at the lattice points $(-m,n)$ while the $B_n$ can be thought of as sitting along the horizontal rays from $(0,n)$. We denote by $\PP$ and $\EE$ the probability measure and expectation with respect to these random variables.
A discrete up-right path $\phi^d$ from $(i_1,j_1)$ to $(i_{\ell},j_{\ell})$ (written as $\phi^d:(i_1,j_1)\nearrow(i_{\ell},j_{\ell})$) is an ordered set of points $\big((i_1,j_1),(i_2,j_2),\ldots, (i_{\ell},j_{\ell})\big)$ with each $(i_k,j_k)\in \Z^2$ and each increment $(i_k,j_k) -(i_{k-1},j_{k-1})$ either $(1,0)$ or $(0,1)$. A semi-discrete up-right path $\phi^{sd}$ from $(0,n)$ to $(\tau,N)$ (written as $\phi^{sd}:(0,n)\nearrow (\tau,N)$) is a union of horizontal line segments $\big((0,n)\to (s_n,n)\big) \cup \big((s_n,n+1)\to (s_{n+1},n+1)\big)\cup \cdots \big((s_{N-1},N)\to (\tau,N)\big)$ where \mbox{$0\leq s_n<s_{n+1}<\cdots <s_{N-1}\leq \tau$}. It is convenient to think of $\phi^{sd}$ as a surjective non-decreasing function from $[0,\tau]$ onto $\{n,\ldots, N\}$.
As we are working with a mixture of a discrete and semi-discrete lattice, our up-right paths $\phi$ will be composed of discrete portions $\phi^d$ adjoined to a semi-discrete portions $\phi^{sd}$ in such a way that for some $1\leq n\leq N$, $\phi^d:(-M,1)\nearrow (-1,n)$ and $\phi^{sd}:(0,n)\nearrow (\tau,N)$. To such a path we associate an energy
\begin{equation}\label{eqEnergy}\begin{aligned}
E(\phi)&=\sum_{(i,j)\in \phi^d} \omega_{i,j} + \int_{0}^{\tau} dB_{\phi^{sd}(s)}(s) \\
&= \sum_{(i,j)\in \phi^d} \omega_{i,j} + B_n(s_n)+\big(B_{n+1}(s_{n+1})-B_{n+1}(s_n)\big)+\ldots+\big(B_N(\tau)-B_N(s_{N-1})\big).
\end{aligned}\end{equation}
This energy is random, as it is a function of the $\omega_{i,j}$ and $B_k$ random variables. We associate a Boltzmann weight $e^{E(\phi)}$ to each path $\phi$. The polymer measure on $\phi$ is proportional to this weight. The normalizing constant, or polymer partition function, is written as $\mathbf{Z}^{N,M}(\tau)$ and is equal to the integral of the Boltzmann weight over the background measure on the path space $\phi$. Explicitly it can be written as
$$
\mathbf{Z}^{N,M}(\tau) =\mathbf{Z}_1^{N,M}(\tau) = \sum_{n=1}^{N} \sum_{\phi^d:(-M,1)\nearrow (-1,n)} \, \int_{\phi^{sd}:(0,n)\nearrow(\tau,N)} e^{E(\phi)} \d\phi^{sd}
$$
where $\d\phi^{sd}$ represents the Lebesgue measure on the simplex $0\leq s_n<s_{n+1}<\cdots <s_{N-1}\leq \tau$ with which $\phi^{sd}$ is identified. Though we do not pursue it, let us note that for $M$ fixed, as a function of $\tau$ and $N$, $\mathbf{Z}^{N,M}(\tau)$ satisfies a semi-discrete SHE (for more on this, see~\cite[Section~5.2]{BC11} or~\cite[Section~6]{BCS12}).
In the spirit of the geometric lifting of the Robinson--Schensted--Knuth correspondence considered in~\cite{OCon09,COSZ11} (and for later use in the statement of Theorem \ref{ThmConnectWhitpoly}) we define a multi-path extension of this polymer and its partition function. For $M\geq 0$ fixed and $1\leq j\leq k\leq N$ define
$$
\mathbf{Z}_j^{k,M}(\tau) = \sum_{1\leq n_1<\cdots<n_j\leq k} \sum_{\substack{\phi_1^d,\ldots,\phi_j^d \\ \phi_a^d\cap \phi_b^d=\emptyset\textrm{ for }a\neq b \\ \phi_a^d:(-M,a)\nearrow (0,n_a)}} \, \int_{(\phi_1^{sd},\ldots \phi_j^{sd})\in D_j^{k,\tau}(n_1,\ldots, n_j)} e^{E(\phi_1)+\cdots+E(\phi_j)} \d\phi_1^{sd}\cdots \d\phi_j^{sd}
$$
where $D_j^{k,\tau}(n_1,\ldots, n_j)$ is the set of $(\phi_1^{sd},\ldots \phi_j^{sd})$ with $\phi_a^{sd}:(0,n_a)\nearrow (\tau,k-j+a)$ such that for all $a\neq b$ and $s\in [0,\tau]$, $\phi_a^{sd}(s)\neq \phi_b^{sd}(s)$ (i.e.\ the paths are non-intersecting). Each $\phi_a^{sd}$ can be identified via the jumping times $0\leq s^{(a)}_{n_a}<\cdots< s^{(a)}_{k-j+a}\leq \tau$, and $\d\phi_1^{sd}\cdots \d\phi_j^{sd}$ is the Lebesgue measure on the Euclidean set $\big(s^{(1)}_{n_1},\ldots, s^{(1)}_{k-j+1}, s^{(2)}_{n_2},\ldots, s^{(2)}_{k-j+2},\ldots, s^{(j)}_{n_j},\ldots, s^{(j)}_{N}\big)$. Note that $\mathbf{Z}^{N,M}(\tau) = \mathbf{Z}_1^{N,M}(\tau)$.
Finally, for $M\geq 0$ fixed and $1\leq j\leq k\leq N$ define
\begin{equation}\label{eqfreeenergy}
\mathbf{F}_j^{k,M}(\tau) = \ln\left(\frac{\mathbf{Z}_j^{k,M}(\tau)}{\mathbf{Z}_{j-1}^{k,M}(\tau)}\right)
\end{equation}
with the convention that $\mathbf{Z}_0^{k,M}(\tau)\equiv 1$.
The following Fredholm determinant formula for the Laplace transform of $\mathbf{Z}^{N,M}(\tau)$, proven in Section~\ref{SectSemiDirected}, is based on the developments of Sections~\ref{SectMacdonald} and~\ref{SectConvWhitt}. The restriction that $N\geq 9$ is likely purely technical and arises in the proof of Proposition~\ref{qFredDetThmbeta} as helpful in establishing certain convergence bounds. Since all of our asymptotics based off of this theorem involve sending $N$ to infinity, this restriction becomes inconsequential.
\begin{theorem}\label{ThmFormulaSemiDiscrete}
Fix $N\geq 9$, $M\geq 0$ and $\tau> 0$. Let $a=(a_1,\ldots,a_N)\in \R^N$ and \mbox{$\alpha = (\alpha_1,\ldots,\alpha_M)\in \big(\R_{>0}\big)^M$} be such that $\alpha_m-a_n>0$ for all $1\leq n\leq N$ and $1\leq m\leq M$.
For $1\leq m\leq M$ and $1\leq n\leq N$ let $\omega_{-m,n}\sim -\ln\Gamma(\alpha_m-a_n)$ be independent log-Gamma random variables and for all $1\leq n\leq N$ let $B_n$ be independent Brownian motions with drift $a_n$. Then for all $u\in \C$ with positive real part
\begin{equation*}
\EE\left[ e^{-u \mathbf{Z}^{N,M}(\tau)} \right] = \det(\Id+ K_{u})_{L^2(\Cv{a;\alpha;\varphi})}
\end{equation*}
where the operator $K_u$ is defined in terms of its integral kernel
\begin{equation*}
K_{u}(v,v') = \frac{1}{2\pi \I}\int_{\Cs{v}}\d s\, \Gamma(-s)\Gamma(1+s) \prod_{n=1}^{N}\frac{\Gamma(v-a_n)}{\Gamma(s+v-a_n)} \prod_{m=1}^M\frac{\Gamma(\alpha_m-v-s)}{\Gamma(\alpha_m-v)}
\frac{ u^s e^{v\tau s+\tau s^2/2}}{v+s-v'}.
\end{equation*}
The contour $\Cv{a;\alpha;\varphi}$ is given in Definition~\ref{DefCaCsdefBis} with any $\varphi\in(0,\pi/4)$, as is the contour $\Cs{v}$.
\end{theorem}
\begin{remark}
Let us make clear our usage of the notion of a Fredholm determinant. Fix a Hilbert space $L^2(X,\mu)$ where $X$ is a measure space and $\mu$ is a measure on $X$. When $X=\Gamma$, a simple (anticlockwise oriented) smooth contour in $\C$, we write $L^2(\Gamma)$ where for $z\in \Gamma$, $\d\mu(z)$ is understood to be $\tfrac{\d z}{2\pi \I}$. When $X$ is the product of a discrete set $D$ and a contour $\Gamma$, $\d\mu$ is understood to be the product of the counting measure on $D$ and $\frac{\d z}{2\pi \I}$ on $\Gamma$. Let $K$ be an {\it integral operator} acting on $f(\cdot)\in L^2(X)$ by $Kf(x) = \int_{X} K(x,y)f(y)\,\d\mu(y)$. $K(x,y)$ is called the {\it kernel} of $K$ and we will assume throughout that $K(x,y)$ is continuous in both $x$ and $y$. Assuming its convergence, the {\it Fredholm determinant expansion} of $\Id+K$ is defined as
\begin{equation*}
\det(\Id+K)_{L^2(X)} = 1+\sum_{n=1}^{\infty} \frac{1}{n!} \int_{X} \cdots \int_{X} \det\left[K(x_i,x_j)\right]_{i,j=1}^{n} \prod_{i=1}^{n} \d\mu(x_i).
\end{equation*}
Note that we do not require $K$ to be trace-class, and only use the notation $\det(\Id+K)_{L^2(X)}$ as a shorthand for the right-hand side of the above equation.
\end{remark}
\begin{remark}
The condition that $\tau>0$ is important to ensure that the integral defining the kernel $K_{u}$ is finite (cf.\ the estimates in Section \ref{prop3sec}). It seems that as long as $M\geq N$, it is possible to take the limit $\tau\to 0$. By continuity of the function $\mathbf{Z}^{N,M}(\tau)$ in $\tau$, this provides a Fredholm determinant formula for $\mathbf{Z}^{N,M}(0)$, or in other words, the log-gamma polymer partition function. A similar formula to this appeared in \cite{BCR13}, though involving a small (finite) contour in place of $\Cv{a;\alpha;\varphi}$ (see also \cite{LeDoussalThierry}). That formula was used therein for asymptotics of the free energy, though only for a certain range of parameters. The large (infinite) contour formula we arrive at here may be useful in removing that parameter range restriction in a parallel manner as \cite{BCF12} used such contours to remove similar restrictions present in \cite{BC11}.
\end{remark}
The contours in Theorem~\ref{ThmFormulaSemiDiscrete} are defined as follows.
\begin{definition}\label{DefCaCsdefBis}
Let $a=(a_1,\ldots,a_N)\in \R^N$ and \mbox{$\alpha = (\alpha_1,\ldots,\alpha_M)\in \big(\R_{>0}\big)^M$} be such that $\alpha_m-a_n>0$ for all $1\leq n\leq N$ and $1\leq m\leq M$. Set $\mu=\tfrac{1}{2}\max(a)+\tfrac{1}{2}\min(\alpha)$ and $\eta = \tfrac{1}{4}\max(a)+\tfrac{3}{4}\min(\alpha)$. Then, for all $\varphi\in (0,\pi/4)$, we define the contour \mbox{$\Cv{a;\alpha;\varphi}=\{\mu+e^{\I (\pi+\varphi)}y\}_{y\in \Rplus}\cup \{\mu+e^{\I(\pi-\varphi)}y\}_{y\in \Rplus}$}. The contours are oriented so as to have increasing imaginary part. For every \mbox{$v\in \Cv{a;\alpha;\varphi}$} we choose $R=-\Re(v)+\eta$, $d>0$, and define a contour $\Cs{v}$ as follows: $\Cs{v}$ goes by straight lines from $R-\I \infty$, to $R-\I d$, to $1/2-\I d$, to $1/2+\I d$, to $R+\I d$, to $R+\I\infty$. The parameter $d$ is taken small enough so that $v+\Cs{v}$ does not intersect $\Cv{a;\alpha;\varphi}$. See Figure~\ref{FigContoursSemiDiscrete} for an illustration.
\end{definition}
\begin{figure}
\begin{center}
\psfrag{Cv}[lb]{$\Cv{a;\alpha;\varphi}$}
\psfrag{v+Cs}[lb]{$v+\Cs{v}$}
\psfrag{Cs}[lb]{$\Cs{v}$}
\psfrag{mu}[cb]{$\mu$}
\psfrag{eta}[cb]{$\eta$}
\psfrag{v}[cb]{$v$}
\psfrag{R}[cb]{$R$}
\psfrag{2d}[lb]{$2d$}
\psfrag{alpha}[cb]{$\alpha$'s}
\psfrag{a}[cb]{$a$'s}
\psfrag{0}[cb]{$0$}
\psfrag{phi}[lb]{$\varphi$}
\includegraphics[height=5cm]{ContoursStat}
\end{center}
\caption{(Left) The contour $\Cv{\eta;\varphi}$ (dashed) where the
black dots symbolize the set of singularities of $K_u(v,v')$ in $v$ at $\cup_{1\leq n\leq N}\{a_n,a_n-1,\dots\}$ coming from the factors $\Gamma(v-a_n)$.
The contour $v+\Cs{v}$ is the solid line.
(Right) The contour $\Cs{v}$ where the light gray dots are the singularities at $\{1,2,\dots\}$ and the dark gray dots are those at $\cup_{1\leq m \leq M}\{\alpha_m-v,\alpha_m+1-v,\dots\}$ coming from \mbox{$\Gamma(\alpha_m-v-s)$}.}
\label{FigContoursSemiDiscrete}
\end{figure}
To eventually access the stationary KPZ equation, we need to choose our $a$ and $\alpha$ parameters appropriately.
\begin{definition}\label{defnearstat}
For what follows, we set $M=1$, $a_1=a$, $a_n\equiv 0$ for $n>1$, $\alpha_1=\alpha>a$ and define $\Zsd(\tau,N)$ as the semi-discrete directed random polymer partition function in which the weight $\omega_{-1,1}$ is replaced by zero.
\end{definition}
\begin{corollary}\label{CorollaryForStationarity}
For $\alpha>a$,
\begin{equation}\label{Besseltransform}
\EE\left[2\big(u\,\Zsd(\tau,N)\big)^{\frac{\alpha-a}2}\BesselK_{-(\alpha-a)}\Big(2\sqrt{u\,\Zsd(\tau,N)}\Big)\right]=\Gamma(\alpha-a)\,\EE\left[e^{-u \mathbf{Z}^{N,1}(\tau)} \right]
\end{equation}
where $\BesselK_\nu$ is the modified Bessel function of order $\nu$, cf.~\cite{AS84}.
\end{corollary}
\begin{proof}
Since all polymer paths $\phi$ must go through the point $(-1,1)$, it follows that
\begin{equation}\label{relZsd}
\mathbf{Z}^{N,1}(\tau)=e^{\omega_{-1,1}}\Zsd(\tau,N).
\end{equation}
By the definition of the log-gamma distribution,
$e^{-\omega_{-1,1}}$ has gamma distribution with parameter $\alpha-a$ and density $x^{\alpha-a-1}e^{-x}/\Gamma(\alpha-a)$ on $\R_+$. Using this along with the independence of $\omega_{-1,1}$ and $\Zsd(\tau,N)$, we may rewrite the Laplace transform of $\mathbf{Z}^{N,1}(\tau)$ using \eqref{relZsd} as
\begin{equation*}
\begin{aligned}
\EE\left[e^{-u\mathbf{Z}^{N,1}(\tau)}\right]&=\EE\left[e^{-ue^{\omega_{-1,1}}\Zsd(\tau,N)}\right]\\
&=\EE\left[\int_0^\infty\frac{e^{-u\Zsd(\tau,N)x^{-1}}x^{\alpha-a-1}e^{-x}}{\Gamma(\alpha-a)}\d x\right]\\
&=\frac1{\Gamma(\alpha-a)}\EE\left[2\big(u\Zsd(\tau,N)\big)^{\frac{\alpha-a}2}\BesselK_{-(\alpha-a)}\left(2\sqrt{u\Zsd(\tau,N)}\right)\right].
\end{aligned}
\end{equation*}
The last equation follows from the identity
$$\int_0^\infty e^{-x-cx^{-1}}x^{-\nu-1}\d x=2c^{d/2}\BesselK_{\nu}\big(2\sqrt c\big),\qquad c>0,$$
which can be derived from the integral representation 9.6.24 of the modified Bessel function in~\cite{AS84}.
\end{proof}
\begin{remark}\label{remstat}
If we further specialize $\Zsd(\tau,N)$ so that $\alpha=a$, we arrive at a model which is stationary. This fact can be gathered from the results of~\cite{SV10} and is explicitly explained in Appendix~\ref{AppStationary}.
\end{remark}
At this point we have a choice to make. We seek to study the stationary SHE/KPZ equation. One way to access that is through a suitable scaling limit of the stationary semi-discrete directed random polymer with log-gamma boundary sources, described in Remark~\ref{remstat}. Alternatively, we could take a suitable limit of the semi-discrete directed random polymer with log-gamma boundary sources with $\alpha>a$. This leads the SHE/KPZ equation with nearly stationary initial data (in fact, two sided Brownian initial data with drifts $\beta>b$). Subsequently, we can take $\beta \to b$ to recover the stationary SHE/KPZ equation. We opt for taking the second route. In either case, there is a technical challenge which we must overcome. Let us presently illustrate this issue for taking the limit $\alpha\to a$, even though it is the other route which we actually pursue. The expectation in the right-hand side of \eqref{Besseltransform} is given by a Fredholm determinant in Theorem~\ref{ThmFormulaSemiDiscrete}. In the
limit $\alpha\to a$, this Fredholm determinant goes to zero linearly in $\alpha-a$, compensating the divergence of $\Gamma(\alpha-a)$ so as to have a non-trivial limit. To take this limit, however, it is necessary to analytically continue our formulas in the quantity $\alpha-a$ (initially in $\R_{>0}$) and use uniqueness of analytic continuations to justify the extension to $\alpha-a=0$.
\begin{remark}
In principle, Theorem~\ref{ThmFormulaSemiDiscrete} could be utilized for a variety of other asymptotics which we do not pursue here. For instance, it should be possible to access some of the one-point probability distribution functions which were previously studied in the case of last passage percolation with boundary conditions in~\cite{BP07}. It should also be possible to take limits to study analogous situations for the SHE/KPZ equation which would involve two-sided version of the initial data considered in~\cite{BCF12} with the inclusion of extra log-gamma weights (see also,~\cite{IS12} for a special case of such initial data).
\end{remark}
\subsection{SHE/KPZ equation with two-sided Brownian initial data}
Theorem~\ref{ThmDiscreteToContinuous} (a result quoted from~\cite{QMR12}) describes the special scaling under which the ($M=1$, $a_1=a$, $a_n\equiv 0$ for $n>1$, and $\alpha_1=\alpha>a$) semi-discrete directed random polymer with log-gamma boundary sources converges to the SHE/KPZ equation with two-sided Brownian motion initial data. The following analogue of Corollary~\ref{CorollaryForStationarity} is proven in Section~\ref{SectCDRP}.
\begin{theorem}\label{ThmFormulaContinuous}
Let us denote by $\mathcal Z_{b,\beta}(T,X)$ the solution to the SHE/KPZ equation with initial data $\mathcal{Z}_0(X)= \exp(B(X))$, where $B(X)$ is a two-sided Brownian motion with drift $\beta$ to the left of $0$ and drift $b$ to the right of $0$, with $\beta>b$, that is, \mbox{$B(X)=\mathbf{1}_{X\leq 0} \big(B^l(X)+\beta X\big) + \mathbf{1}_{X>0} \big(B^r(X)+b X\big)$} where $B^l:(-\infty,0]\to \R$ is a Brownian motion without drift pinned at $B^l(0)=0$, and $B^r:[0,\infty)\to \R$ is an independent Brownian motion pinned at $B^r(0)=0$.
Then, for $S>0$,
\begin{multline}\label{Bessel=Fredholm}
\EE\left[2\left(Se^{\frac{X^2}{2T}+\frac T{24}}\mathcal Z_{b,\beta}(T,X)\right)^{\frac{\beta-b}2}\BesselK_{-(\beta-b)}\left(2\sqrt{Se^{\frac{X^2}{2T}+\frac T{24}}\mathcal Z_{b,\beta}(T,X)}\right)\right]\\
=\Gamma(\beta-b)\det(\Id-K_{b+X/T,\beta+X/T})_{L^2(\R_+)}
\end{multline}
where $\BesselK_\nu(z)$ is the modified Bessel function of order $\nu$ and the kernel on the right-hand side is given by
\begin{equation}\label{defKbbeta}
K_{b,\beta}(x,y)=\frac1{(2\pi\I)^2}\int\d w\int\d z \frac{\sigma\pi S^{\sigma(z-w)}}{\sin(\sigma\pi(z-w))} \frac{e^{z^3/3-zy}}{e^{w^3/3-wx}}
\frac{\Gamma(\beta-\sigma z)}{\Gamma(\sigma z-b)} \frac{\Gamma(\sigma w-b)}{\Gamma(\beta-\sigma w)}
\end{equation}
where
\begin{equation}\label{defsigma}
\sigma=(2/T)^{1/3}.
\end{equation}
The integration contour for $w$ is from $-\frac1{4\sigma}-\I\infty$ to $-\frac1{4\sigma}+\I\infty$ and crosses the real axis between $b$ and $\beta$.
The other contour for $z$ goes from $\frac1{4\sigma}-\I\infty$ to $\frac1{4\sigma}+\I\infty$, it also crosses the real axis between $b$ and $\beta$ and it does not intersect the contour for $w$.
\end{theorem}
\begin{remark}\label{comparerem}
Let us compare the above result to that derived (non-rigorously via the replica method) in~\cite[Proposition 1]{IS12}. The initial data considered therein is two-sided Brownian, plus a log-gamma distributed (independent) height shift. We may use Theorem~\ref{ThmFormulaContinuous} and reverse the proof of Corollary~\ref{CorollaryForStationarity} so as to prove a one-point formula for this initial data. Inspection reveals that the resulting formula matches that of~\cite{IS12}. As we soon explain, in order to go from this formula to the stationary initial data formula requires work and the final formula shown in~\cite{IS12} is not as readily compared to the final formula proved herein as Theorem~\ref{ThmFormulaStationary}.
\end{remark}
\subsection{SHE/KPZ equation with stationary initial data}
The stationary initial conditions for the KPZ equation are the two-sided Brownian motions with a fixed drift. For the SHE this means to let $\mathcal{Z}_0(X)=\exp(B(X))$ with $B$ a two-sided Brownian motion with drift $b\in \R$. Call the resulting solution to the SHE $\mathcal{Z}_b(T,X)$. We can get the result by carefully taking the $\beta\to b$ limit in Theorem~\ref{ThmFormulaContinuous}. This limit is accomplished by analytically continuing the expressions on both sides of \eqref{Bessel=Fredholm}. In order to be able to state our main result of the paper we need a few notations.
\begin{definition}\label{longdef}
For $b\in\left(-\frac14,\frac14\right)$, define on $\R_+$ the function
\begin{equation}\label{defqss}\begin{aligned}
q_{b}(x)&=\frac1{2\pi\I}\int_{-\frac1{4\sigma}+\I\R}\d w \frac{\sigma\pi S^{b-\sigma w}}{\sin(\pi(b-\sigma w))} e^{-w^3/3+wx}\frac{\Gamma(\sigma w-b)}{\Gamma(b-\sigma w)}
\end{aligned}\end{equation}
and for $b\in\R$, let
\begin{equation}\label{defrs}
r_b(x)=e^{b^3/(3\sigma^3)-bx/\sigma}.
\end{equation}
Further, for $b\in\left(-\frac14,\frac14\right)$, define the kernel
\begin{equation}\label{defbarKbb}
\bar K_{b}(x,y)=\frac1{(2\pi\I)^2}\int_{-\frac1{4\sigma}+\I\R}\d w\int_{\frac1{4\sigma}+\I\R}\d z \frac{\sigma\pi S^{\sigma(z-w)}}{\sin(\sigma\pi(z-w))} \frac{e^{z^3/3-zy}}{e^{w^3/3-wx}}
\frac{\Gamma(b-\sigma z)}{\Gamma(\sigma z-b)} \frac{\Gamma(\sigma w-b)}{\Gamma(b-\sigma w)}.
\end{equation}
Finally, letting $\gamma_{\rm E}=0.577\ldots$ represent the Euler--Mascheroni constant, define
\begin{equation}\label{defz}
\begin{aligned}
\Xi(S,b,\sigma)=&-\det(\Id-\bar K_{b})\Big[b^2/\sigma^2+\sigma(2\gamma_{\rm E}+\ln S)\\
&+\big\langle(\Id-\bar K_{b})^{-1}(\bar K_{b}r_{-b}+q_{b}),r_b\big\rangle +\big\langle(\Id-\bar K_{b})^{-1}(r_{-b}+q_{b}),q_{-b}\big\rangle\Big].
\end{aligned}
\end{equation}
where the determinants and scalar products are all meant in $L^2(\R_+)$.
\end{definition}
\begin{remark}
By using the general identity
$$\det(\Id-K)\big\langle(\Id-K)^{-1}f,g\big\rangle=\det(\Id-K)-\det(\Id-K-f\otimes g),$$
it is also possible to write $\Xi$ as a linear combination of Fredholm determinants:
\begin{multline*}
\Xi(S,b,\sigma)=\det\big(\Id-\bar K_{b}-(\bar K_{b}r_{-b}+q_{b})\otimes r_b\big)+\det\big(\Id-\bar K_{b}-(r_{-b}+q_{b})\otimes q_{-b}\big)\\
-\det(\Id-\bar K_{b})\big[2+b^2/\sigma^2+\sigma(2\gamma_{\rm E}+\ln S)\big].
\end{multline*}
Note that $\Xi$ also depends on $S$ implicitly through $\bar K_{b}$ and $q_{\pm b}$.
The right-hand side of \eqref{defz} is well-defined for any admissible choice of the parameters, see Remark~\ref{rem:barKproduct}.
\end{remark}
The following result (which implies Theorem~\ref{ThmFormulaStationaryIntroVersion} when $b=X=0$) is proven in Section~\ref{SectStatSHE}.
\begin{theorem}\label{ThmFormulaStationary}
Let $\mathcal{Z}_b(T,X)$ be the solution to the SHE with initial data $\mathcal{Z}_0(X)=e^{B(X)}$ with $B$ a two-sided Brownian motion with $B(0)=0$ and drift $b\in \R$.
Let $\BesselK_0$ denote the modified Bessel function and consider $X\in\R$, $T>0$ and $b\in \R$ such that $b+\frac XT\in\left(-\frac14,\frac14\right)$. For $S>0$,
\begin{equation}\label{K0transform}
\EE\left[2\sigma \BesselK_0\left(2\sqrt{Se^{\frac{X^2}{2T}+\frac T{24}}\mathcal Z_b(T,X)}\right)\right]=\Xi\left(S,b+\frac XT,\sigma\right)
\end{equation}
where the function $\Xi$ is defined in \eqref{defz}.
\end{theorem}
We remark that the condition $b+\frac XT\in\left(-\frac14,\frac14\right)$ could be weakened to $b+\frac XT\in(-1,1)$ in a slightly more technical way, but the formulation \eqref{defz} is not convergent outside the latter regime. See Remark~\ref{rem:bcondition} for more details.
The integral transform that appears on the left-hand side of \eqref{K0transform} is the Mellin transform~\cite{NY92} of the stationary (drift $b$) KPZ equation solution
\begin{equation}\label{defF}
\mathcal{H}_b(T,X)=\ln\mathcal{Z}_b(T,X).
\end{equation}
It is possible to recover the distribution function from \eqref{K0transform} using a double inverse Mellin transform (proven in Appendix~\ref{AppMellin}).
\begin{proposition}\label{PropInverseMelling}
Consider $T>0$, $X\in \R$, and $b\in \R$ such that $b+\frac XT\in\left(-\frac14,\frac14\right)$. For any $r\in\R$,
\begin{multline*}
\PP\left(\frac{\mathcal H_b(T,X)+\frac T{24}+\frac{X^2}{2T}}{(T/2)^{1/3}}\le r\right)\\
=\frac{1}{\sigma^2}\frac{1}{2\pi\I}\int_{-\delta+\I\R} \frac{\d \xi}{\Gamma(-\xi)\Gamma(-\xi+1)} \int_{\R} \d x\, e^{x\xi/\sigma} \Xi\left(e^{-\frac{x+r}\sigma},b+\frac XT,\sigma\right)
\end{multline*}
for any $\delta>0$.
\end{proposition}
\begin{proof}
This follows from applying Proposition~\ref{prop:mellin} with $R=\sigma(\mathcal H_b(T,X)+T/24+X^2/(2T))$ and $x=-\sigma\ln S$.
The finite negative exponential moment $\EE(\exp(-\delta R/\sigma))$ is ensured by Lemma~\ref{lem:Ftailbound} for any $\delta>0$.
\end{proof}
This formula should be compared to~\cite[Theorem~2]{IS12} in which the non-rigorous replica method was utilized to study the stationary KPZ equation (see Remark \ref{comparerem}).
\begin{remark}\label{rem:shiftarg}
Comparing \eqref{K0transform} for different values of $b$ and $X$ shows
\begin{equation}\label{shiftarg}
\mathcal Z_{b-X/T}(T,X)=e^{-\frac{X^2}{2T}}\mathcal Z_b(T,0)
\end{equation}
in distribution. This rotational invariance property can be explained directly from the definition of the SHE, as in~\cite[Section~3.2]{BCF12}.
\end{remark}
In the large $T$ limit one expects, by the universality belief, that limiting one-point probability distribution functions for the KPZ equation should converge to those previously determined in the context of TASEP or in the polynuclear growth model for analogous types of initial data~\cite{BR00,SI04,PS02b,FS05a}. Here we use the same notations as in~\cite[Theorem~1.2]{BFP09} specialized to the one-point setting.
\begin{definition}\label{anotherlongone}
Recall the Airy function $\Ai$, cf.~\cite{AS84}. For $\tau,s\in\R$, define
\begin{align*}
\mathcal R&=s+e^{-\frac23\tau^3}\int_s^\infty\d x\int_0^\infty\d y \Ai(x+y+\tau^2)\,e^{-\tau(x+y)},\\
\Psi(y)&=e^{\frac23\tau^3+\tau y}-\int_0^\infty\d x \Ai(x+y+\tau^2)\,e^{-\tau x},\\
\Phi(x)&=e^{-\frac23\tau^3}\int_0^\infty\d\lambda\int_s^\infty dy \Ai(x+\tau^2+\lambda)\Ai(y+\tau^2+\lambda)\,e^{-\tau y}
-\int_0^\infty\d y\Ai(y+x+\tau^2)\,e^{\tau y}.
\end{align*}
Let $P_s$ be the projection operator $P_s(x)=\Id_{\{x>s\}}$, the Airy kernel with shifted entries by
\begin{equation}\label{defKAi}
\wh
K_{\Ai}(x,y)=\int_0^\infty\d\lambda\Ai(x+\lambda+\tau^2)\Ai(y+\lambda+\tau^2),
\end{equation}
and define the function
\begin{equation}\label{defg}
g(\tau,s)=\mathcal R-\big\langle(\Id-P_s\wh K_{\Ai}P_s)^{-1}P_s\Phi,P_s\Psi\big\rangle.
\end{equation}
Finally, let
\begin{equation}\label{defBFP}
F_\tau(r)=\frac{\partial}{\partial r}\left(g(\tau,r)\det\left(\Id-P_r\wh K_{\Ai}P_r\right)_{L^2(\R)}\right).
\end{equation}
\end{definition}
In the large $T$ limit, the fluctuations of $\mathcal H_b(T,X)$ are governed by $F_\tau$, as shown in the following result (proven in Section~\ref{SectUniversality}).
\begin{theorem}\label{CorUniversality}
Let $b\in(-\frac14,\frac14)$ be fixed and consider any $\tau\in\R$. Define $\sigma=(2/T)^{1/3}$ and consider the scaling
\begin{equation}\label{XTscaling}
X=-bT+\frac{2\tau}{\sigma^2}.
\end{equation}
Then, for any $r\in\R$,
\begin{equation*}
\lim_{T\to\infty}\PP\left(\frac{\mathcal H_b(T,X)+\frac T{24}(1+12b^2)-2^{1/3}b\tau T^{2/3}}{(T/2)^{1/3}}\le r\right)=F_\tau(r).
\end{equation*}
\end{theorem}
\section{Ascending $q$-Whittaker processes}\label{SectMacdonald}
\subsection{Defining the processes}
The ascending $q$-Whittaker processes $\M_{\tilde a;\rho}$ are special cases of the ascending Macdonald processes~\cite{BC11} in which the Macdonald parameters $t=0$ and $q\in (0,1)$. The $q$-Whittaker measures $\MM_{\tilde a;\rho}$ are marginals of the ascending $q$-Whittaker processes. We provide a brief account of these objects as well as the $q$-Whittaker (or Macdonald $t=0$) symmetric functions used to define them. For a more involved discussion and background, see~\cite[Sections~2.2 and 3.1]{BC11}.
Fix $N\geq 1$. The {\it $q$-Whittaker process} $\M_{\tilde a;\rho}$ is a probability measure on sequences of interlacing partitions
\begin{equation*}
\varnothing \prec \lambda^{(1)}\prec \lambda^{(2)}\prec\cdots \prec\lambda^{(N)}
\end{equation*}
(equivalently Gelfand--Tsetlin patterns, or column-strict Young tableaux) parameterized by positive reals\footnote{The reason we use tildes for the parameters of this measure is because they will eventually be expressed in terms of parameters without tildes, when we perform a $q\to 1$ scaling limit to their Whittaker counterparts.} $\tilde a=\{\tilde a_1,\ldots,\tilde a_N\}$, a single $q$-Whittaker nonnegative specialization $\rho$ of the algebra of symmetric function, and the Macdonald parameter $q\in (0,1)$. The probability measure is given by
\begin{equation*}
\M_{\tilde a;\rho}\big(\la^{(1)},\dots,\la^{(N)}\big)= \frac{P_{\la^{(1)}}(\tilde a_1)P_{\la^{(2)}/\la^{(1)}}(\tilde a_2)\cdots P_{\la^{(N)}/\la^{(N-1)}}(\tilde a_N) Q_{\la^{(N)}}(\rho)}{\Pi(\tilde a_1,\ldots, \tilde a_N;\rho)}\,.
\end{equation*}
We write $\EE_{\M_{\tilde a,\rho}}$ for the expectation with respect to this measure (though sometimes may drop the $\M_{\tilde a,\rho}$ subscript when it is clear).
Some explanation of notation is in order. A partition $\lambda$ is an integer sequence \mbox{$\lambda=(\lambda_1\geq \lambda_2\geq \ldots \geq 0)$} with finitely many nonzero entries, and we say that $\mu \prec \lambda$ if the two partitions interlace: $\mu_i\leq \lambda_i \leq \mu_{i-1}$ for all meaningful $i$'s. In Young diagram terminology, $\mu\prec\lambda$ is equivalent to saying that the skew partition $\lambda/\mu$ is a horizontal strip.
The functions $P_{\bullet}$ and $Q_{\bullet}$ are $q$-Whittaker symmetric functions (i.e.\ Macdonald symmetric functions with parameter $t=0$) which are indexed by (skew) partitions and implicitly depend on the Macdonald parameter $q\in (0,1)$. The remarkable properties of Macdonald symmetric functions are developed in~\cite[Section~VI]{Mac79}, and all of the relevant facts to which we appeal are also reviewed in~\cite[Section~2.1]{BC11}. The evaluation of a $q$-Whittaker symmetric function on a positive variable $\tilde a$ (as in $P_{\lambda/\mu}(\tilde a)$) means to restrict the function to a single nonzero variable and then substitute the value $\tilde a$ in for that variable. This is a special case of a $q$-Whittaker nonnegative specialization $\rho$ which is an algebra homomorphism of the algebra of symmetric functions $\Sym\to \C$ that takes skew $q$-Whittaker symmetric functions to nonnegative real numbers (notation: $P_{\lambda/\mu}(\rho)\geq 0$, $Q_{\lambda/\mu}(\rho)\geq 0$ for any partitions $\
\lambda$ and $\mu$). Restricting the $q$-Whittaker symmetric
functions to a finite number of nonzero variables (i.e.\ considering $q$-Whittaker polynomials) and then substituting nonnegative numbers for these variables constitutes such a specialization.
We will work with a more general class of specializations which can be thought of as unions and limits of such finite length specializations as well as {\it dual} specializations. Let $\tilde \alpha=\{\tilde \alpha_i\}_{i\ge 1}$, $\tilde \beta=\{\tilde \beta_i\}_{i\ge 1}$, and $\tilde \gamma$ be nonnegative reals such that $\sum_{i=1}^\infty(\tilde \alpha_i+\tilde \beta_i)<\infty$. Let $\rho=\rho(\tilde \alpha;\tilde \beta;\tilde \gamma)$ be a specialization of $\Sym$ defined by
\begin{equation}\label{specdefeqn}
\sum_{n\ge 0} g_n(\rho) u^n= e^{\tilde \gamma u} \prod_{i\ge 1} \frac{1+\tilde \beta_i u}{(\tilde \alpha_i u;q)_\infty}=: \Pi\big(u;\rho(\tilde \alpha;\tilde \beta;\tilde \gamma)\big).
\end{equation}
Here $u$ is a formal variable, $g_n=Q_{(n)}$ is the $q$-analog of the complete homogeneous symmetric function $h_n$, and $(\tilde a;q)_n=\prod_{i=0}^{n-1} (1-q^i \tilde a)$ is the $q$-Pochhammer symbol (with obvious extension when $n=\infty$). Since $\{g_n\}_{n\geq 0}$ form an algebraic basis of $\Sym$, this uniquely defines the specializations $\rho$. Such $\rho$ are $q$-Whittaker nonnegative (see~\cite[Section~2.2.1]{BC11} for more details). Alternatively, one can specify the above specializations $\rho(\tilde \alpha;\tilde \beta;\tilde \gamma)$ in terms of the values they take on the Newton power sum symmetric functions $p_k = \sum_i (x_i)^k$ via
\begin{equation*}
\begin{aligned}
p_1\big(\rho(\tilde \alpha;\tilde \beta;\tilde \gamma)\big) &\mapsto (1-q)\tilde \gamma + \sum_i \left(\tilde \alpha_i+(1-q)\tilde \beta_i\right),\\
p_k\big(\rho(\tilde \alpha;\tilde \beta;\tilde \gamma)\big) &\mapsto \sum_i \left((\tilde \alpha_i)^k+(-1)^{k-1}(1-q^k)(\tilde \beta_i)^k\right), \qquad k\geq 2.
\end{aligned}
\end{equation*}
We can also express $\Pi(u;\rho)$ in terms of these Newton power sum symmetric functions as
$$
\Pi(u;\rho) = \exp\left(\sum_{k=1}^{\infty} \frac{u^k\, p_k(\rho)}{(1-q^k)k}\right).
$$
When it is clear which specialization we are discussing, we will just write $\rho$ rather than $\rho(\tilde \alpha;\tilde \beta;\tilde \gamma)$.
The normalization for the ascending $q$-Whittaker process is given by
\begin{equation*}
\sum_{\la^{(N)}}P_{\la^{(N)}}(\tilde a) Q_{\la^{(N)}}(\rho) = \Pi(\tilde a;\rho) = \prod_{n=1}^{N} \Pi(\tilde a_n;\rho),
\end{equation*}
as follows from a generalization of Cauchy's identity for Schur functions (corresponding to the case $q=0$). It is not hard to see that for $\rho=\rho(\tilde \alpha;\tilde\beta;\tilde\gamma)$ the condition of the partition function $\Pi(\tilde a;\rho)$ to be finite is equivalent to $\tilde a_n\tilde \alpha_m<1$ for all $n,m$, and hence we will always assume that this holds.
The projection of $\M_{a;\rho}$ to a single partition $\lambda^{(k)}$, $k\in\{1,\dots,N\}$, is the {\it $q$-Whittaker measure} given by
\begin{equation*}
\MM_{\tilde a;\rho}\big(\lambda^{(k)}\big)= \frac{P_{\lambda^{(k)}}(\tilde a_1,\ldots,\tilde a_k) Q_{\lambda^{(k)}}(\rho)} {\Pi(\tilde a_1,\ldots,\tilde a_k;\rho)}\,.
\end{equation*}
In what follows we will be concerned primarily with the marginal distribution of $\lambda^{(N)}_1$.
\subsection{Fredholm determinant formula}
In order to state the main theorem of the section, we must specify parameters for the $q$-Whittaker measure as well as various contours which participate.
\begin{definition}\label{defParams}
For $N\geq 1$ consider non-negative reals $\tilde a=\{\tilde a_1,\ldots, \tilde a_N\}$. We will work with $q$-Whittaker non-negative specializations $\rho = \rho(\tilde \alpha;\tilde \beta;\tilde \gamma)$ as in \eqref{specdefeqn} where $\tilde \alpha=\{\tilde \alpha_1,\ldots,\tilde \alpha_{M_{\alpha}}\}$, $\tilde \beta=\{\tilde \beta_1,\ldots,\tilde \beta_{M_{\beta}}\}$ and $\tilde \gamma$ satisfy that for all $i$, $\tilde \alpha_i,\tilde \beta_i,\tilde\gamma \geq 0$ and $\max(\tilde \alpha),\max(\tilde \beta) <\min(\tilde a^{-1})$, where $\tilde a^{-1} = \{\tilde a_1^{-1},\ldots, \tilde a_N^{-1}\}$.
\end{definition}
\begin{definition}\label{CwPredef}
For $\tilde a$, $\tilde \alpha$ and $\tilde \beta$ as in Definition~\ref{defParams} and an angle $\varphi\in (0,\pi/2]$ define $\CwPre{\tilde a; \tilde \alpha,\tilde \beta; \varphi}=\{\tilde \mu+e^{-\I \varphi\sign(y)}y,y\in \R\}$ (oriented so as to have decreasing imaginary part) where $\tilde \mu = \tfrac{1}{2}\max(\tilde \alpha\cup \tilde \beta) +\tfrac{1}{2}\min(\tilde a^{-1})$.
For $w\in \CwPre{\tilde a; \tilde \alpha,\tilde \beta;\varphi}$, we choose $R,d>0$ and the contour $\CsPre{w}$ as follows:
$\CsPre{w}$ goes by straight lines from $R-\I \infty$, to $R-\I d$, to $1/2-\I d$, to $1/2+\I d$, to $R+\I d$, to $R+\I\infty$. We choose $R$ and $d$ such that the following holds: For all $s\in\CsPre{w}$, $q^s w$ lies to the left of $\CwPre{\tilde a;\tilde \alpha,\tilde \beta; \varphi}$ and encloses all $\tilde \alpha$ and $\tilde\beta$; and for $|w|$ large, $R\approx \ln|w|$ and $d\approx |w|^{-1}$ (here $\approx$ means up to a positive constant bounded from zero and infinity). See Figure~\ref{FigqWhitContours} for an illustration of these contours.
\end{definition}
\begin{figure}
\begin{center}
\psfrag{w}[cc]{$w$}
\psfrag{qsw}[lc]{$q^s w$}
\psfrag{beta}[cc]{$-\tilde\beta$}
\psfrag{alpha}[lc]{$\tilde\alpha$}
\psfrag{ainv}[cc]{$\tilde a^{-1}$}
\psfrag{phi}[lc]{$\varphi$}
\psfrag{Dw}[cc]{}
\psfrag{C}[lc]{$\CwPre{\tilde a;\tilde \alpha,\tilde \beta; \varphi}$}
\includegraphics[height=7cm]{FigContoursQlevel.eps}
\end{center}
\caption{The contour $\CwPre{\tilde a;\tilde \alpha,\tilde \beta; \varphi}$ (from Definition~\ref{CwPredef}) is depicted along with the contour corresponding to $q^s w$ for $w\in \CwPre{\tilde a;\tilde \alpha,\tilde \beta; \varphi}$ and $s\in \CsPre{w}$.}
\label{FigqWhitContours}
\end{figure}
We are prepared to state the central result of this section. The (most likely technical) condition that $N\geq 9$ (which comes from some convergence estimates used in the proof of Proposition~\ref{qFredDetThmbeta}) is not much of a limitation since we will ultimately be concerned in studying the large $N$ limit of this formula.
\begin{theorem}\label{qFredDetThm}
Fix $N\geq 9$ and $\tilde a,\tilde\alpha,\tilde\beta,\tilde\gamma$ as in Definition~\ref{defParams}. Then for all $\zeta\in \C\setminus \Rplus$
\begin{equation}\label{thmlaplaceeqn}
\EE_{\M_{\tilde a,\rho(\tilde \alpha;\tilde \beta;\tilde \gamma)}}\Bigg[ \frac{1}{\big(\zeta q^{-\lambda^{(N)}_1};q\big)_{\infty}}\Bigg] = \det(\Id+\tilde K_{\zeta})_{L^2(\CwPre{\tilde a; \tilde \alpha,\tilde\beta;\varphi})}
\end{equation}
where $\CwPre{\tilde a; \tilde \alpha,\tilde\beta;\varphi}$ as in Definition~\ref{CwPredef} with any $\varphi\in (0,\pi/2]$. The operator $\tilde K_{\zeta}$ is defined in terms of its integral kernel
\begin{equation}\label{eqnkzetakernel}
\tilde K_{\zeta}(w,w') = \frac{1}{2\pi \I}\int_{\CsPre{w}} \Gamma(-s)\Gamma(1+s)(-\zeta)^s g_{w,w'}(q^s)\,\d s
\end{equation}
where
\begin{equation}\label{gwwprimeeqn}\begin{aligned}
g_{w,w'}(q^s) &=\frac{1}{q^s w - w'}\, \frac{\Pi(w;a)}{\Pi(q^s w;a)}\, \frac{\Pi\big((q^s w)^{-1};\rho(\tilde\alpha;\tilde\beta;\tilde\gamma)\big)}{\Pi\big((w)^{-1};\rho(\tilde\alpha;\tilde\beta;\tilde\gamma)\big)}\\
&=\frac{\exp\big(\gamma w^{-1}(q^{-s}-1)\big)}{q^s w - w'} \,\prod_{i=1}^{N} \frac{(q^{s}w a_i;q)_{\infty}}{(w a_i;q)_{\infty}}\, \prod_{i=1}^{M_{\alpha}} \frac{\big((w)^{-1}\alpha_i;q\big)_{\infty}}{\big((q^{s}w)^{-1}\alpha_i ;q\big)_{\infty}}\,\prod_{i=1}^{M_{\beta}} \frac{1+ (q^s w)^{-1} \beta_i}{1+ (w)^{-1} \beta_i},
\end{aligned}\end{equation}
the contour $\CsPre{w}$ is as in Definition~\ref{CwPredef} and the function $\Pi$ is defined as in \eqref{specdefeqn}.
\end{theorem}
\begin{remark}
This formula bares many similarities to that for the $e_q$-Laplace transform for $q^{\lambda^{N}_N}$ in \cite[Theorem 3.2.11]{BC11} and \cite[Theorem 4.13]{BCF12}. The $\lambda^{N}_N$ are closely related to the particle system $q$-TASEP \cite[Section 3.3.2]{BC11}, and hence these formulas served as the starting point for large time asymptotic analysis of $q$-TASEP~\cite{FV13,Bar14}. It was shown in \cite{BP13b} that $\lambda^{N}_1$ relates to the particle system $q$-PushTASEP. The above theorem may (in a similar manner as in \cite{FV13,Bar14}) be of use in asymptotic analysis of this system.
\end{remark}
The remainder of this section is devoted to the proof of this theorem. The starting point for this proof is the approach described in~\cite[Section~3.2]{BC11} to compute the $e_{q}$-Laplace transform of $q^{\lambda^{(N)}_N}$. Rather quickly, though, we encounter challenges not previously present requiring new ideas. The approach from~\cite[Section~3.2]{BC11} for the random variable $q^{\lambda^{(N)}_{N}}$ starts by utilizing Macdonald difference operators to compute nested contour integral formulas for the moments $\EE\Big[q^{k \lambda^{(N)}_{N}}\Big]$ (the subscript $\M_{\tilde a,\rho(\tilde \alpha;\tilde \beta;\tilde \gamma)}$ is suppressed here). Since the random variable $q^{\lambda^{(N)}_{N}}\in (0,1]$, its moments determine its distribution, and its $e_q$-Laplace transform can be computed via a suitable moment generating series. Plugging the explicit formulas for these moments into the $e_q$-Laplace transform moment generating series results (after further manipulations) in a Fredholm determinant.
For $q^{-\lambda^{(N)}_{1}}$, this approach (of~\cite[Section~3.2]{BC11}) runs into a major issue in the first step. The random variable $\lambda^{(N)}_1$ is part of a partition and hence a non-negative integer. Since $q\in (0,1)$, it follows that $q^{-\lambda^{(N)}_{1}}\geq 1$. Moreover, if any of the specialization parameters $\tilde\alpha$ are non-zero (i.e.\ $\tilde\alpha_i>0$ for some $i$) then $q^{-\lambda^{(N)}_{1}}$ will only have a finite number of finite moments (Lemma~\ref{lemMomentsexist}). Therefore, recovering the distribution or $e_{q}$-Laplace transform from these finitely many moments is impossible. But we need the case where some of these $\tilde{\alpha}$ parameters are strictly positive as it relates after various limit transition to the stationary KPZ equation. Therefore, we must overcome this apparent obstacle.
In this case (where some $\alpha_i>0$), for $k$ small enough $\EE\Big[q^{-k\lambda^{(N)}_{1}}\Big]<\infty$ and there still exist nested contour integral moment formulas (coming from Macdonald difference operators). These formulas involve $k$ in a straightforward manner and one can try to extend the formula for $k$ to be an arbitrary integer. For those $k$ for which the moments are infinite, there fail to exist suitable choices of contours upon which to integrate. If one neglects this (important) impediment, it is possible to mimic the approach of~\cite[Section~3.2]{BC11}. The outcome of this formal calculation is the statement of Theorem~\ref{qFredDetThm} (with some additional guess to work out which contours to use in the final answer). Of course, this is not a mathematically justified calculation since it involves summing infinitely many terms which are themselves infinite and ill-defined. The outcome, however, is an equality between two finite quantities.
The challenge is to turn this into a meaningful rigorous result, and hence prove Theorem~\ref{qFredDetThm}. This is done in two steps:
\smallskip
\noindent {\bf Step 1:} Apply the approach of~\cite[Section~3.2]{BC11} to prove a Fredholm determinant formula for the $e_q$-Laplace transform of $q^{-\lambda^{(N)}_1}$ in the special case where all $\tilde \alpha_i\equiv 0$ (we also take $\tilde\gamma=0$ for this step). By studying the $q$-Whittaker measure under the pure $\tilde\beta$ specialization, we can prove a priori that $\lambda^{(N)}_1 \leq M_{\beta}$ (recall, $M_{\beta}$ is the number of non-zero entries in $\tilde\beta$). This bound shows that $\EE\Big[q^{-k\lambda^{(N)}_{1}}\Big]\leq q^{-kM_{\beta}}$, and hence we may adapt the approach from~\cite[Section~3.2]{BC11} to prove the pure $\tilde\beta$ specialization Fredholm determinant.
\smallskip
\noindent {\bf Step 2:} Interpret the pure $\tilde\beta$ specialization Fredholm determinant formula for the $e_q$-Laplace transform as a formal series identity in the Newton power-sum symmetric functions. Then, apply the $\rho(\tilde\alpha;\tilde\beta;\tilde\gamma)$ specialization to this formal series identity and observe that both sides of the identity form convergent series, hence proving the desired numerical identity which is Theorem~\ref{qFredDetThm}.
\smallskip
The key fact which lets us succeed here is that we are working with symmetric functions. This situation should be compared to the non-rigorous physics replica method. There, the problem is to compute the distribution (via the Laplace transform) of the solution to the stochastic heat equation $\mathcal{Z}(T,X)$. It is possible to compute similar sorts of moment formulas as those above for $\EE\big[\mathcal{Z}(T,X)^k\big]$. These moments remain finite for all $k$, but grow like $e^{c k^3}$ for some constant $c>0$. This means that the moment problem is ill-posed and these moments do not determine the distribution of $\mathcal{Z}(T,X)$. However, in the replica method calculations (e.g.~\cite{Dot10,CDR10}) one can still try to compute $\EE\big[e^{\zeta \mathcal{Z}(T,X)}\big]$ through these moments via expanding the exponential and interchanging the expectation and infinite summation. Though the Laplace transform is necessarily finite, the associated moment generating series (coming from the mathematically
unjustifiable interchange of expectation and the summation in the Taylor expansion of the exponential) is divergent. After some additional manipulations, this divergent generating function is `summed' to a finite expression -- a Fredholm determinant. At least in the case of $\mathcal{Z}_0(X)=\delta_{X=0}$, the resulting formulas agree with those proved in~\cite{ACQ10}.
In light of these similarities, one might hope to find a way to implement a variant of the rigorous justification we provide in the study of $q^{-\lambda^{(N)}_{1}}$ into the setting of the SHE. However, this seems unlikely. The justification we provide draws heavily upon the relationship between our observable $q^{-\lambda^{(N)}_{1}}$ and the $q$-Whittaker processes / symmetric functions. Such structures do not clearly survive the limit transitions which eventually relate to the SHE (see however~\cite{CH13,OCW11} for some trace of these structures). Furthermore, the pure $\tilde\beta$ specialization which is used to justify the formal identity we prove, does not (as of yet) have any analog (or limit) in the SHE (or even semi-discrete directed polymer) setting.
\subsection{Step 1: Pure $\tilde \beta$ Fredholm determinant formula}
For this step, we will focus on the $q$-Whittaker proceses with specializations $\rho=\rho(0;\tilde\beta;0)$ and $\tilde\beta= (\tilde\beta_1,\ldots\tilde\beta_{M})$ with $M\geq 1$ arbitrary (note that in Step 1b and 1c we return to considering general specializations to provide moment formulas). For these specialization, the $q$-Whittaker function $Q_{\lambda}(\rho)=0$ unless $\lambda_1\leq M$ (see~\cite[Section~2.2.1]{BC11}). This provides the a priori bound that under the $q$-Whittaker process, $q^{-\lambda^{(N)}_1} \leq q^{-M}$. Due to this bound, we will assume in Steps 1a--1e that
\begin{equation}\label{eqnzetabdd}
|\zeta|< (1-q) q^M,
\end{equation}
though in Steps 1d--1e we will impose an additional condition on $|\zeta|$.
\subsubsection{Step 1a: Relating $e_q$-Laplace transform to moments generating series}
Observe that for $\zeta$ satisfying \eqref{eqnzetabdd}, the function
$$
\zeta \mapsto \frac{1}{\big(\zeta q^{-\lambda^{(N)}_1};q)_{\infty}}
$$
(which can be rewritten as $e_q\big(\zeta (1-q)^{-1} q^{-\la^{(N)}_1}\big)$, cf.\ Appendix~\ref{qSec})
is always finite and analytic in $\zeta$. Using the $q$-Binomial theorem (cf.\ Appendix~\ref{qSec}) we may expand
\begin{equation}\label{eqexpandmgs}
\frac{1}{\big(\zeta q^{-\lambda^{(N)}_1};q)_{\infty}} = \sum_{k=0}^{\infty} \frac{\big(\zeta/(1-q)\big)^k}{k_q!} q^{-k\lambda^{(N)}_1}
\end{equation}
where the $q$-deformed factorial is defined as
$$
k_q! = \frac{(q;q)_k}{(1-q)^k}.
$$
Due to \eqref{eqnzetabdd}, it follows that each summand on the right-hand side of \eqref{eqexpandmgs} can be bounded deterministically by a corresponding summand of a convergent geometric series. This justifies the interchange of expectation and summation necessary to establish the equality
\begin{equation}\label{eqninterchanged}
\EE_{\M_{\tilde a,\rho(0;\tilde \beta;0)}}\Bigg[\frac{1}{\big(\zeta q^{-\lambda^{(N)}_1};q)_{\infty}}\Bigg] = \sum_{k=0}^{\infty} \frac{\big(\zeta/(1-q)\big)^k}{k_q!} \EE_{\M_{\tilde a,\rho(0;\tilde \beta;0)}}\Big[q^{-k\lambda^{(N)}_1}\Big]
\end{equation}
for those $\zeta$ satisfying \eqref{eqnzetabdd}.
\subsubsection{Step 1b: Nested contour integral formulas for moments}
In Step 1b--1c, we temporarily return to the general $\rho(\tilde\alpha;\tilde\beta;\tilde\gamma)$ specialization and prove nested contour integral formulas for moments (when they exist).
Let us first describe conditions on $\tilde a,\tilde\alpha,\tilde\beta$ under which moments of $q^{-\lambda^{(N)}_1}$ are finite, or infinite.
\begin{lemma}\label{lemMomentsexist}
Fix $N\geq 1$ and $\tilde a,\tilde\alpha,\tilde\beta,\tilde\gamma$ as in Definition~\ref{defParams}. For $k\geq 1$, if $\max_{i,j} \tilde a_i \tilde \alpha_j <q^{k}$, then
$$
\EE_{\M_{\tilde a,\rho(\tilde\alpha;\tilde \beta;\tilde\gamma)}}\Big[q^{-k\lambda^{(N)}_1}\Big] <+\infty.
$$
On the other hand, if $\max_{i,j} \tilde a_i \tilde \alpha_j >q^{k}$, then
$$
\EE_{\M_{\tilde a,\rho(\tilde\alpha;\tilde \beta;\tilde\gamma)}}\Big[q^{-k\lambda^{(N)}_1}\Big] = +\infty.
$$
\end{lemma}
\begin{proof}
We first address the case that $\max_{i,j} \tilde a_i \tilde \alpha_j <q^{k}$. We can bound
\begin{align*}
\EE_{\M_{\tilde a,\rho(\tilde\alpha;\tilde \beta;\tilde\gamma)}}\big[q^{-k\lambda^{(N)}_1}\big] &= \sum_{\lambda^{(N)}} q^{-k \lambda^{(N)}_{1}} \M_{\tilde a,\rho(\tilde\alpha;\tilde \beta;\tilde\gamma)}(\lambda^{(N)})\\
&\leq \sum_{\lambda^{(N)}} q^{-k |\lambda^{(N)}|} \M_{\tilde a,\rho(\tilde\alpha;\tilde \beta;\tilde\gamma)}(\lambda^{(N)})\\
&= \sum_{\lambda^{(N)}} \M_{q^{-k}\tilde a,\rho(\tilde\alpha;\tilde \beta;\tilde\gamma)}(\lambda^{(N)})<\infty.
\end{align*}
The equality on the first line is by definition; the inequality on the second line is from the fact that $\lambda^{(N)}_1 \leq |\lambda^{(N)}|$, where $|\lambda| = \sum \lambda_i$; the equality on the third line comes from the fact that $c P_{\lambda}(\rho)= P_{\lambda}(c\rho)$; the final inequality on the third line comes from the fact that the $q$-Whittaker process $\M_{q^{-k}\tilde a,\rho(\tilde\alpha;\tilde \beta;\tilde\gamma)}$ is well-defined as long as $q^{-k} \tilde a_i \tilde \alpha_j< 1$.
Turning now to the case that $\max_{i,j} \tilde a_i \tilde \alpha_j >q^{k}$, assume (without loss of generality) that $\tilde a_1 = \max (\tilde a)$ and $\tilde\alpha_1 = \max(\alpha)$. By the interlacing inequalities, $\lambda^{(1)}_1 \leq \lambda^{(N)}_1$. This means that if we show $\EE_{\M_{\tilde a,\rho(\tilde\alpha;\tilde \beta;\tilde\gamma)}}\Big[q^{-k\lambda^{(1)}_1}\Big]=+\infty$, then so too must $\EE_{\M_{\tilde a,\rho(\tilde\alpha;\tilde \beta;\tilde\gamma)}}\Big[q^{-k\lambda^{(N)}_1}\Big]=+\infty$. To further simplify considerations, observe that the $q$-Whittaker process $\M_{\tilde a_1,\rho(\tilde\alpha_1;0;0)}$ is stochastically dominated by the $q$-Whittaker process $\M_{\tilde a_1,\rho(\tilde\alpha;\tilde \beta;\tilde\gamma)}$. This can be seen, for instance, in light of dynamics~\cite[Section~2.3]{BC11} which maps the first process to the second process by only increasing coordinates. Owing to this stochastic domination, it suffices to prove that $\EE_{\M_{\tilde a_1,\rho(\tilde\alpha_1;0;
0)
}}\Big[q^{-k\lambda^{(1)}_1}\Big]=+\infty$. This expectation, however, is computable quite explicitly since $P_{\lambda^{(1)}_1}(\tilde a_1) = (\tilde a_1)^{\lambda^{(1)}_1}$, $Q_{\lambda^{(1)}_1}(\tilde \alpha_1) = (\tilde\alpha_1)^{\lambda^{(1)}_1} (q;q)_{\lambda^{(1)}_1}^{-1}$ and $\Pi(\tilde a_1;\tilde \alpha_1) = (\tilde a_1\tilde\alpha_1;q)_{\infty}^{-1}$. Therefore, we find that
\begin{align*}
\EE_{\M_{\tilde a_1,\rho(\tilde\alpha_1;0;0)}}\Big[q^{-k\lambda^{(1)}_1}\Big] &= \sum_{\lambda^{(1)}_1\geq 0} q^{-k \lambda^{(1)}_{1}} \M_{\tilde a_1,\rho(\tilde\alpha_1;0;0)}(\lambda^{(1)}_{1})\\
&= \sum_{\lambda^{(1)}_1\geq 0} q^{-k \lambda^{(1)}_{1}} (\tilde a_1\tilde\alpha_1)^{\lambda^{(1)}_1} \frac{(\tilde a_1\tilde\alpha_1;q)_{\infty}}{(q;q)_{\lambda^{(1)}_1}},
\end{align*}
which is seen to diverge to $+\infty$ under the condition that $q^{-k} \tilde a_1\tilde\alpha_1 >1$.
\end{proof}
The following proposition provides explicit formulas for those moments which are necessarily finite due to Lemma~\ref{lemMomentsexist}. It should be observed that suitable contours in the statement of the proposition exist under the same conditions as the finiteness of moments.
\begin{proposition}\label{momentprop}
Fix $N\geq 1$ and $\tilde a,\tilde\alpha,\tilde\beta,\tilde\gamma$ as in Definition~\ref{defParams}. For $k\geq 1$, if \mbox{$\max_{i,j} \tilde a_i \tilde \alpha_j <q^{\eta k}$} for some $\eta>1$, then
\begin{equation}\label{eqmoments}
\EE_{\M_{\tilde a,\rho(\tilde \alpha;\tilde \beta;\tilde \gamma)}}\Big[ q^{-k \lambda^{(N)}_1}\Big] = \frac{(-1)^k q^{\frac{k(k-1)}{2}}}{(2\pi \I)^k} \int_{C_1} \cdots \int_{C_k} \prod_{1\leq A<B\leq k}\frac{z_A-z_B}{z_A-q z_B} \prod_{i=1}^{k} \frac{g(z_i)}{g(qz_i)} \frac{\d z_i}{z_i}
\end{equation}
where
\begin{equation}\label{eqg}
g(z) =\prod_{i=1}^{N} \frac{1}{(\tilde a_i z;q)_{\infty}} \, \frac{1}{\Pi\big((z)^{-1};\rho(\tilde \alpha;\tilde \beta;\tilde \gamma)\big)} = \prod_{i=1}^{N} \frac{1}{(\tilde a_i z;q)_{\infty}} \,
e^{\tilde \gamma z^{-1}} \prod_{i=1}^{M_{\alpha}} (\tilde \alpha_i z^{-1};q)_{\infty}\, \prod_{i=1}^{M_{\beta}} \frac{1}{1+\tilde \beta_i z^{-1}}.
\end{equation}
The contours $C_1,\ldots,C_k$ are defined by $C_i= q^{\eta(k-i)} C_k$ where $C_k=\{ c+e^{- \I \varphi \sgn(y)}, y\in \R\}$ (oriented so as to have decreasing imaginary part) with any $\varphi\in (0,\pi/2]$ and with any $c\in \R$ satisfying $q^{-\eta k} \max (\alpha) < c<\min (a^{-1})$.
\end{proposition}
\begin{proof}
We provide a brief proof, as this result has essentially appeared before in~\cite[Remark~2.2.11]{BC11} and~\cite[Theorem~4.6]{BCGS13} (in the more general Macdonald processes language).
The proof is based on a simple observation. Assume we have a linear operator $\mathcal{D}$ on the space of functions in $N$ variables whose restriction to the space of symmetric polynomials acts diagonally in the basis of $q$-Whittaker polynomials: $\mathcal{D} P_\lambda=d_\lambda P_\lambda$ for any partition $\lambda$ with length $\ell(\lambda)\le N$. Then we can apply $\mathcal{D}$ to both sides of the identity (acting in the $\tilde a$ variables)
\begin{equation*}
\sum_{\lambda} P_{\lambda}(\tilde a) Q_{\lambda}(\rho)=\Pi(\tilde a;\rho).
\end{equation*}
Dividing the result by $\Pi(\tilde a;\rho)$, we obtain
\begin{equation*}
\EE_{\M_{\tilde a,\rho}}\left[ d_\lambda \right]=\frac{\mathcal{D}\Pi(\tilde a;\rho)} {\Pi(\tilde a;\rho)}\,.
\end{equation*}
This equality is valid so long as the expectation on the left-hand side is finite. If we apply $\mathcal{D}$ several times, we obtain
\begin{equation*}
\EE_{\M_{\tilde a,\rho}}\left[ (d_\lambda)^k \right]=\frac{\mathcal{D}^k\Pi(\tilde a;\rho)} {\Pi(\tilde a;\rho)}\,.
\end{equation*}
If we have several possibilities for $\mathcal{D}$ we can obtain formulas for averages of the observables equal to products of powers of the corresponding eigenvalues. One of the remarkable features of Macdonald polynomials is that there exists a large family of such operators. These are the Macdonald difference operators.
We will consider a slight variant of the $t=0$ $(N-1)$-st difference operator. For any $u\in \R$ and $1\leq i\leq N$ define the shift operator $T_{u,x_i}$ via its action
\begin{equation*}
\left(T_{u,x_i}F\right)(x_1,\ldots, x_N) = F(x_1,\ldots, u x_i,\ldots, x_{N}).
\end{equation*}
The operator $\tilde{D}$ which we utilize is given by
\begin{equation*}
\tilde{D} = \sum_{j=1}^{N} \prod_{\substack{i=1\\i\neq j}}^{N} \frac{x_j}{x_j-x_i} T_{q^{-1},x_j}.
\end{equation*}
The $q$-Whittaker polynomials diagonalize this operator~\cite[Remark~2.2.11]{BC11}, so that for all $\lambda$ of length $\ell(\lambda)\leq N$,
\begin{equation*}
\tilde{D} P_{\lambda}(x_1,\ldots, x_N) = q^{-\lambda_1} P_{\lambda}(x_1,\ldots, x_N).
\end{equation*}
Thus, using the procedure described above, we find that
\begin{equation}\label{eqmomentsabove}
\EE_{\M_{\tilde a,\rho}}\Big[q^{-k\lambda^{(N)}_1} \Big]=\frac{\tilde{D}^k\Pi(\tilde a;\rho)} {\Pi(\tilde a;\rho)}\,.
\end{equation}
This equality is true assuming that the left-hand side is finite. Lemma~\ref{lemMomentsexist} provides the conditions for $\rho=\rho(\tilde\alpha;\tilde\beta;\tilde\gamma)$ such that this expectation is finite.
To complete the proof we must identify the right-hand side of \eqref{eqmoments} with the right-hand side of \eqref{eqmomentsabove}. This identification follows from residue calculus. The contour $C_k$ (along which $z_k$ is integrated) can be deformed to cross the set $\tilde a^{-1}$. This deformation crosses poles and the integral is thus expanded as a sum of residues at these poles and a remaining integral over a new contour which lies to the right of the $\tilde a^{-1}$. The remaining integral evaluates to zero, as can be seen by using Cauchy's theorem and the at least quadratic decay of the integrand (as $z_k$ goes to infinity in the right half of $\C$). Each of the residue terms involves $k-1$ integrals, and this procedure can be repeated for the $z_{k-1}$ through $z_1$ integrals. The resulting residue expansion of the integrals match exactly the formula on the right-hand side of~\eqref{eqmomentsabove}. See~\cite[Section~2.2.3]{BC11} for more details of this type of residue bookkeeping.
\end{proof}
\subsubsection{Step 1c: Unnesting the integrals}
Proposition~\ref{momentprop} provides a nested contour integral formula for the moments of $q^{-\lambda^{(N)}_1}$ under the general $\rho(\tilde\alpha;\tilde\beta;\tilde\gamma)$ specialization. Here we record the effect of deforming all of the contours to lie upon the same fixed contour.
\begin{proposition}\label{propunnested}
Fix $N\geq 1$ and $\tilde a,\tilde\alpha,\tilde\beta,\tilde\gamma$ as in Definition~\ref{defParams}. For $k\geq 1$, if \mbox{$\max_{i,j} \tilde a_i \tilde \alpha_j <q^{\eta k}$} for some $\eta>1$, then
\begin{equation*}
\EE_{\M_{\tilde a,\rho(\tilde \alpha;\tilde \beta;\tilde \gamma)}}\big[ q^{-k \lambda^{(N)}_1}\big] = \sum_{\mu\vdash k} \frac{1}{m_1! m_2! \cdots} \frac{(1-q)^k}{(2\pi \I)^{\ell(\mu)}} \int_{C}\cdots \int_{C} \det\left[\frac{1}{w_i q^{\mu_i} - w_j}\right]_{i,j=1}^{\ell(\mu)} \prod_{j=1}^{\ell(\mu)} \frac{g(w_j)}{g(q^{\mu_j}w_j)} \d w_j
\end{equation*}
where $\mu$ is a partition of $k$ (hence the notation $\mu\vdash k$) of length $\ell(\mu)$ and multiplicities \mbox{$m_i = |\{j:\mu_j=i\}|$}, the function $g$ is given by \eqref{eqg}, and the contour \mbox{$C=\{ c+e^{- \I \varphi \sgn(y)}, y\in \R\}$} (oriented so as to have decreasing imaginary part) with any $\varphi\in (0,\pi/2]$ and with any $c\in \R$ satisfying $\max (\alpha) < c<\min (a^{-1})$.
\end{proposition}
\begin{proof}
This is essentially proved as~\cite[Proposition~3.2.1]{BC11}, or~\cite[Proposition~7.4]{BCPS13}. The contours $C_{k-1}$ through $C_{1}$ (on the right-hand side of \eqref{eqmomentsabove}) are sequentially deformed to lie along $C_k$. This deformation involves crossing certain strings of poles (recorded by the partition $\mu$). The resulting formula comes from bookkeeping these residues. Note that once all integration contours coincide with $C_k$, these contours can be simultaneously deformed (without crossing any poles or changing the value of the integrals) to any choice of contour $C$ as specified in the statement of the proposition.
\end{proof}
\subsubsection{Step 1d: Summing the moment generating series}
We return now to studying the case of the pure $\tilde\beta$ specialization. Equation \eqref{eqninterchanged} of Step 1a shows that the $e_q$-Laplace transform of $q^{-\lambda^{(N)}_1}$ is equal to a generating series of the moments $\EE\Big[q^{-k\lambda^{(N)}_1}\Big]$ provided $|\zeta|< (1-q) q^M$. Proposition~\ref{propunnested} of Step 1c (in the pure $\tilde\beta$ specialization) provides explicit formulas for these moments. The following proposition shows how plugging these formulas into the moment generating series results in a Fredholm determinant formula. In order for this sum to converge, we must impose some further restrictions on $|\zeta|$. We also assume $N\geq 2$ here as it is helpful for the technical aspects of the argument of the proof to proceed.
\begin{proposition}\label{propFirstFred}
Fix $N\geq 2$, $\tilde a, \tilde\beta$ as in Definition~\ref{defParams}, $\varphi\in (0,\pi/2]$, and a contour $\CwPre{\tilde a;0,\tilde\beta;\varphi}$ as in Definition~\ref{CwPredef}. Specialize $g$ from \eqref{eqg} to the pure $\tilde\beta$ specialization as
\begin{equation*}
g(w) =\prod_{i=1}^{N} \frac{1}{(\tilde a_i w;q)_{\infty}} \, \frac{1}{\Pi\big((w)^{-1};\rho(0;\tilde \beta;0)\big)} = \prod_{i=1}^{N} \frac{1}{(\tilde a_i w;q)_{\infty}} \,
\prod_{i=1}^{M} \frac{1}{1+\tilde \beta_i w^{-1}}.
\end{equation*}
Define $f(w) = \frac{g(w)}{g(qw)}$ and the constant
\begin{equation}\label{eqncconst}
C_1= \sup_{\ell\geq 1, w\in \CwPre{\tilde a; 0,\tilde\beta;\varphi}} |f(q^\ell w)|.
\end{equation}
Then for all $\zeta\in \C\setminus \Rplus$, such that $|\zeta| < \min\big\{ (1-q)q^M, C_1^{-1}\big\}$,
\begin{equation*}
\EE_{\M_{\tilde a,\rho(0;\tilde \beta;0)}}\Bigg[\frac{1}{\big(\zeta q^{-\lambda^{(N)}_1};q\big)_{\infty}}\Bigg] = \det(\Id+K_{\zeta})_{L^2(\Z_{>0}\times \CwPre{\tilde a; 0,\tilde\beta;\varphi})}.
\end{equation*}
The operator $K_{\zeta}$ is defined in terms of its kernel
\begin{equation*}
K_{\zeta}(n_1,w_1;n_2,w_2) = \frac{\zeta^{n_1}}{q^{n_1} w_1-w_2}\, \frac{g(w_1)}{g(q^{n_1}w_1)}.
\end{equation*}
\end{proposition}
\begin{proof}
In light of \eqref{eqninterchanged} and the bound $|\zeta|<(1-q) q^M$, it suffices to prove that
\begin{equation*}
\sum_{k=0}^{\infty} \frac{\big(\zeta/(1-q)\big)^k}{k_q!} \EE_{\M_{\tilde a,\rho(0;\tilde \beta;0)}}\Big[q^{-k\lambda^{(N)}_1}\Big] = \det(\Id+K_{\zeta})_{L^2(\Z_{>0}\times \CwPre{\tilde a; 0,\tilde\beta;\varphi})}.
\end{equation*}
Using Proposition~\ref{propunnested} for the pure $\tilde\beta$ specialization, we can rewrite (see~\cite[Proposition~3.2.8]{BC11} or~\cite[Proposition~4.12]{BCF12}) the $k$-th term in this moment generating series as
\begin{multline*}
\frac{\big(\zeta/(1-q)\big)^k}{k_q!} \EE_{\M_{\tilde a,\rho(0;\tilde \beta;0)}}\Big[q^{-k\lambda^{(N)}_1}\Big] \\
=\sum_{L\geq 0} \frac{1}{L!} \int_{\CwPre{\tilde a; 0,\tilde\beta;\varphi}}\cdots \int_{\CwPre{\tilde a; 0,\tilde\beta;\varphi}} \sum_{\stackrel{n_1,\ldots,n_L\geq 1}{\sum n_i=k}}\det\left[\frac{1}{w_i q^{n_i}-w_j}\right]_{i,j=1}^{L} \prod_{j=1}^{L} \zeta^{n_j} \frac{g(w_j)}{g(q^{n_j}w_j)} \frac{\d w_j}{2\pi \I}.
\end{multline*}
Summing over $k$ yields
\begin{multline}\label{eqfredexpC}
\sum_{k=0}^{\infty} \frac{\big(\zeta/(1-q)\big)^k}{k_q!} \EE_{\M_{\tilde a,\rho(0;\tilde \beta;0)}}\Big[q^{-k\lambda^{(N)}_1}\Big] \\
= \sum_{L\geq 0} \frac{1}{L!} \int_{\CwPre{\tilde a; 0,\tilde\beta;\varphi}}\cdots \int_{\CwPre{\tilde a; 0,\tilde\beta;\varphi}} \sum_{n_1,\ldots, n_L\geq 1} \det\left[\frac{1}{w_i q^{n_i}-w_j}\right]_{i,j=1}^{L} \prod_{j=1}^{L} \zeta^{n_j} \frac{g(w_j)}{g(q^{n_j}w_j)} \frac{\d w_j}{2\pi \I}
\end{multline}
which is the Fredholm determinant expansion of $\det(\Id+K_{\zeta})_{L^2(\Z_{>0}\times \CwPre{\tilde a; 0,\tilde\beta;\varphi})}$.
The convergence of this Fredholm determinant expansion, as well as the manipulations used in reaching it require some justifications. (After all, the manipulations involved rearranging an infinite summation.) Note that by assumptions on the contour $\CwPre{\tilde a; 0,\tilde\beta;\varphi}$, the function $q^{n_i} w_i/w_j-1$ remains bounded from 0 uniformly as $w_i,w_j\in \CwPre{\tilde a; 0,\tilde\beta;\varphi}$ and $n_i\geq 1$ vary. It follows then from Hadamard's inequality that there exists a constant $B_1>0$ such that for all $w_i,w_j\in \CwPre{\tilde a; 0,\tilde\beta;\varphi}$ and $L,n_1,\ldots, n_L\geq 1$
$$
\left| \det\left[\frac{1}{w_i q^{n_i}-w_j}\right]_{i,j=1}^{L} \right| \leq B_1^L L^{L/2}.
$$
We may also show that for all $w_j\in \CwPre{\tilde a; 0,\tilde\beta;\varphi}$ and $n_j\geq 1$,
$$
\left|\frac{g(w_j)}{g(q^{n_j}w_j)}\right|\leq C_1^{n_j} C_2 w_j^{-N}
$$
where $C_1$ is defined in \eqref{eqncconst} and
$$
C_2 = C_1^{-1} \sup_{w\in \CwPre{\tilde a; 0,\tilde\beta;\varphi}} f(w) w^N.
$$
This is shown by writing (recall $f(w)=\tfrac{g(w)}{g(qw)}$)
$$
\frac{g(w_j)}{g(q^{n_j}w_j)} = f(w_j) f(qw_j) \cdots f(q^{n_j-1} w_j)
$$
and using the definition of $C_1$ and $C_2$. The finiteness of these constants is easily verified.
Combining these observations, we may bound in absolute value the $L$-th summand in \eqref{eqfredexpC} by
$$
\frac{1}{L!} B_1^L L^{L/2} \left(\sum_{n\geq 1} C_2 (C_1\zeta)^n \int_{\CwPre{\tilde a; 0,\tilde\beta;\varphi}} \frac{|\d w|}{2\pi} |w|^{-N}\right)^L.
$$
As we have assumed $N\geq 2$, the integral in $|\d w|$ is bounded by a constant $B_2>0$. Since, by hypothesis, $|\zeta|<C_1^{-1}$ we can also bound the summation in $n$ by another constant $B_3>0$. Therefore, the above expression is bounded by
$$
\frac{1}{L!} (C_2 B_1 B_2 B_3)^L L^{L/2}.
$$
Since the summation of this over $L\geq 0$ is finite, the Fredholm determinant expansion \eqref{eqfredexpC} is absolutely convergent. Using similar bounds as described above, we can also justify all of the interchanging of summands necessary to complete the proof of the proposition.
\end{proof}
\subsubsection{Step 1e: Rewriting as a Fredholm determinant of desired type}
We will now prove Theorem~\ref{qFredDetThm} in the case of pure $\tilde\beta$ specialization subject to the condition that $\zeta\in \C\setminus \Rplus$ satisfies $|\zeta| < \min\big\{ (1-q)q^M, C_1^{-1}\big\}$, with $C_1$ from \eqref{eqncconst}.
\begin{proposition}\label{qFredDetThmbetarestricted}
Fix $N\geq 2$ and $\tilde a,\tilde\beta$ as in Definition~\ref{defParams}. Then for all $\zeta\in \C\setminus \Rplus$ satisfying $|\zeta| < \min\big\{ (1-q)q^M, C_1^{-1}\big\}$, with $C_1$ from \eqref{eqncconst}, \begin{equation*}
\EE_{\M_{\tilde a,\rho(0;\tilde \beta;0)}}\Bigg[ \frac{1}{\big(\zeta q^{-\lambda^{(N)}_1};q\big)_{\infty}}\Bigg] = \det(\Id+\tilde K_{\zeta})_{L^2(\CwPre{\tilde a;0,\tilde\beta;\varphi})}
\end{equation*}
with $\CwPre{\tilde a; 0,\tilde\beta;\varphi}$ as in Definition~\ref{CwPredef} with any $\varphi\in (0,\pi/2]$. The operator $\tilde K_{\zeta}$ is defined in terms of its integral kernel given in \eqref{eqnkzetakernel} with $g_{w,w'}(q^{s})$ from \eqref{gwwprimeeqn} explicitly given in the pure $\tilde\beta$ specialization by
\begin{equation}\label{gwwprimeprime}
g_{w,w'}(q^s) = \frac{1}{q^s w-w'} \prod_{i=1}^{N} \frac{ (\tilde a_i q^s w;q)_{\infty}}{(\tilde a_i w;q)_{\infty}} \prod_{i=1}^{M} \frac{ 1+ \tilde\beta_i (q^s w)^{-1}}{ 1+ \tilde\beta_i (w)^{-1}}.
\end{equation}
\end{proposition}
\begin{proof}
The convergent Fredholm expansion of $\det(\Id+K_{\zeta})_{L^2(\Z_{>0}\times \CwPre{\tilde a; 0,\tilde\beta;\varphi})}$ can be written as
$$
\det(\Id+K_{\zeta})_{L^2(\Z_{>0}\times \CwPre{\tilde a; 0,\tilde\beta;\varphi})} = \sum_{L\geq 0} \frac{1}{L!} \int_{\CwPre{\tilde a; 0,\tilde\beta;\varphi}} \frac{\d w_1}{2\pi \I}\cdots \int_{\CwPre{\tilde a; 0,\tilde\beta;\varphi}}\frac{\d w_L}{2\pi \I} \det\left[ \sum_{n=1}^{\infty} \zeta^n g_{w_i,w_j}(q^n)\right]_{i,j=1}^{L}
$$
where $g_{w,w'}(q^s)$ is from the statement of the proposition.
We will have proved the proposition if we can show that
\begin{equation}\label{eqsummingmellin}
\sum_{n=1}^{\infty} \zeta^n g_{w,w'}(q^n) = \frac{1}{2\pi \I}\int_{\CsPre{w}} \Gamma(-s)\Gamma(1+s)(-\zeta)^s g_{w,w'}(q^s)\,\d s.
\end{equation}
To show this, we will apply the following {\it Mellin--Barnes} representation.
\begin{lemma}\label{gammasumlemma}
For all functions $g$, all negatively oriented (with respect to the points $1,2,\ldots$) contours $C_{1,2,\ldots}$ and all $\zeta\in \C\setminus \Rplus$ which satisfy the conditions below, we have the identity
\begin{equation*}
\sum_{n=1}^{\infty} g(q^n) \zeta^n = \frac{1}{2\pi \I} \int_{C_{1,2,\ldots}} \Gamma(-s)\Gamma(1+s)(-\zeta)^s g(q^s)\,\d s
\end{equation*}
where the function $\zeta\mapsto (-\zeta)^s$ on the right-hand side is defined with respect to a branch cut along $\zeta \in \R_{+}$. The conditions which must be satisfied are as follows: for $k$ large, there must exist positively oriented contours $C_k$ which enclose the points $1,2,\ldots, k$, which do not enclose any singularities of $g(q^s)$, and which are such that the integral above taken along the symmetric difference of $C_{1,2,\ldots}$ and $C_k$ goes to zero as $k$ goes to infinity.
\end{lemma}
\begin{proof}
The identity follows from $\Res{s=k}\,\Gamma(-s)\Gamma(1+s) = (-1)^{k+1}$.
\end{proof}
\begin{figure}
\begin{center}
\psfrag{0}[cb]{$0$}
\psfrag{k1}[rt]{$k$}
\psfrag{k2}[lt]{$k+1$}
\psfrag{Cs}[lb]{$C_{k}$}
\psfrag{Carc}[lb]{$C^{arc}_k$}
\psfrag{Cseg}[lb]{$C^{seg}_k$}
\psfrag{R}[cb]{$R$}
\psfrag{2d}[lb]{$2d$}
\psfrag{12}[cb]{$\frac12$}
\includegraphics[height=6cm]{SymDiffContours}
\caption{Left: The contour $C_{k}$ composed of the union of two parts -- the first part is the portion of the contour $\CsPre{w}$ which lies within the ball of radius $k+1/2$ centered at the origin; the second part is the arc of that ball which causes the union to be a closed contour which encloses $\{1,2,\ldots, k\}$ and no other integers. Right: The symmetric difference between $C_k$ and $\CsPre{w}$ is given by two parts: a semi-circle arc which we call $C^{arc}_k$ and a portion of $R+\I \R$ with magnitude exceeding $k+1/2$ which we call $C^{seg}_k$.}\label{SymDiffContours}
\end{center}
\end{figure}
We apply Lemma~\ref{gammasumlemma} to prove \eqref{eqsummingmellin}. Let $C_{1,2,\ldots} = \CsPre{w}$ and let $C_{k}$ be composed of the union of two parts -- the first part is the portion of the contour $\CsPre{w}$ which lies within the ball of radius $k+1/2$ centered at the origin; the second part is the arc of the boundary of that ball which causes the union to be a closed contour which encloses $\{1,2,\ldots, k\}$ and no other integers. The contours $C_{k}$ are oriented positively and illustrated in the left-hand side of Figure~\ref{SymDiffContours}. We may assume $k$ is large enough so that the circle of radius $k+1/2$ intersects $\CsPre{w}$ on its vertical component. By the definition of the contours $\CwPre{\tilde a; 0,\tilde\beta;\varphi}$ and $\CsPre{w}$ we are assured that the contours $C_k$ do not contain any poles of $(-\zeta)^s g_{w,w'}(q^s)$.
This is due to the fact that the contours have been chosen such that as $s$ varies, $q^sw$ stays entirely to the left of $\CwPre{\tilde a; 0,\tilde\beta;\varphi}$ and hence does not touch $w'$.
In order to apply the above lemma we must estimate the integral along the symmetric difference of $C_{1,2,\ldots}$ and $C_k$. Identify the part of the symmetric difference given by the circular arc as $C^{arc}_k$ and the part given by the portion of $R+\I \R$ with magnitude exceeding $k+1/2$ as $C^{seg}_k$ (see the right-hand side of Figure~\ref{SymDiffContours}). We will estimate the integrand on each of these contours.
Concerning the term $(-\zeta)^s$, we may write $-\zeta = r e^{i\sigma}$ with $\sigma\in (-\pi,\pi)$ and $r>0$. Then we have $(-\zeta)^s =r^s e^{\I s\sigma}$. Writing $s=x+\I y$ we have $|(-\zeta)^s| = r^{x}e^{-y\sigma}$. Note that our assumptions on $\zeta$ imply $r<\min\big\{ (1-q)q^M, C_1^{-1}\big\}$, with $C_1$ from \eqref{eqncconst}, and $\sigma\in(-\pi,\pi)$.
Concerning the product of Gamma functions, one readily confirms that there exists $c>0$ such that for all $s$ with $\dist(s,\Z)\geq 1/2$
\begin{equation*}
\Big| \Gamma(-s)\Gamma(1+s) \Big| \leq \frac{c}{e^{\pi |\Im(s)|}}.
\end{equation*}
Focusing on $s\in C^{seg}_k$, the above bounds imply
$$
\Big|\Gamma(-s)\Gamma(1+s)(-\zeta)^s\Big| \leq c r^R e^{-y\sigma - \pi |y|}
$$
as $\dist(s,\Z)\geq 1/2$ and $x=R$ along $C^{seq}_k$ (recall $s=x+\I y$). As $s$ varies along $C^{seq}_k$, $q^s w$ cycles around a circle of fixed radius containing the origin and hence we may bound $|g_{w,w'}(q^s)|<C$ for some constant $C>0$ independent of $s\in C^{seq}_k$ and $k$. That implies
$$
\left|\frac{1}{2\pi \I} \int_{C^{seq}_{k}} \Gamma(-s)\Gamma(1+s)(-\zeta)^s g_{w,w'}(q^s)\,\d s \right| \leq \int_{C^{seq}_{k}} \frac{|\d s|}{2\pi} C c r^R e^{-y\sigma - \pi |y|}.
$$
Since $\sigma\in (-\pi,\pi)$, the integrand decays exponentially as $|y|$ increases (recall $s=x+\I y$). This means that as $k$ goes to infinity, the above integral converges to zero.
Focusing on $s\in C^{arc}_k$, the earlier bounds imply
$$
\left|\Gamma(-s)\Gamma(1+s)(-\zeta)^s\right| \leq c r^x e^{-y\sigma - \pi |y|}
$$
as $\dist(s,\Z)\geq 1/2$ along $C^{seq}_k$. By inspection, we may bound $|g_{w,w'}(q^s)|<C' q^{-x M}$ for some constant $C'>0$ independent of $s\in C^{arc}_k$ and $k$. To see this, observe that as \mbox{$s\in C^{arc}_k$} and $k$ varies, $(q^s w-w')^{-1}$ stays uniformly bounded, and since $|a_iq^s w|$ stays uniformly bounded strictly below 1, each term $(a_i q^s w;q)_{\infty}$ remains uniformly bounded by a constant (the denominator $(a_i w;q)_{\infty}$ remains constant as $s$ and $k$ vary). The term $1+ \frac{\beta_i}{q^s w}$ is bounded by a constant times $q^{-x}$ where $s= x+\I y$ (and the corresponding term in the denominator $1+\frac{\beta_i}{w}$ remains constant as $s$ and $k$ vary). Combining these considerations leads to the claimed bound.
This bound implies
$$
\left|\frac{1}{2\pi \I} \int_{C^{arc}_{k}} \Gamma(-s)\Gamma(1+s)(-\zeta)^s g_{w,w'}(q^s)\,\d s \right| \leq \int_{C^{arc}_{k}} \frac{|\d s|}{2\pi} C' c r^x q^{-x M} e^{-y\sigma - \pi |y|}.
$$
By assumption, $rq^{-M}< e^{-\nu}$ for some $\nu>0$, implying that $r^x q^{-xM}\leq e^{-\nu x}$. Plugging this in, and using the fact that $\sigma\in (-\pi,\pi)$, we see that the integrand decays exponentially as both $|y|$ and $x$ increase. This means that as $k$ goes to infinity, the above integral converges to zero. This completes the verification necessary to apply Lemma~\ref{gammasumlemma} and hence \eqref{eqsummingmellin} follows and the proposition is proved.
\end{proof}
\subsubsection{Step 1f: Analytic continuation}
We show that for $\tilde a, \tilde\beta$ fixed, we may use analytic continuation to extend the domain of applicability of Proposition~\ref{qFredDetThmbetarestricted} to hold for all $\zeta\in \C\setminus \Rplus$. This proves Theorem~\ref{qFredDetThm} in the case of a pure $\tilde\beta$ specialization.
\begin{proposition}\label{qFredDetThmbeta}
Fix $N\geq 9$ and $\tilde a,\tilde\beta$ as in Definition~\ref{defParams}. Then for all $\zeta\in \C\setminus \Rplus$
\begin{equation}\label{eqnmmthd}
\EE_{\M_{\tilde a,\rho(0;\tilde \beta;0)}}\Bigg[ \frac{1}{\big(\zeta q^{-\lambda^{(N)}_1};q\big)_{\infty}}\Bigg] = \det(\Id+\tilde K_{\zeta})_{L^2(\CwPre{\tilde a;0,\tilde\beta;\varphi})}
\end{equation}
with $\CwPre{\tilde a; 0,\tilde\beta;\varphi}$ as in Definition~\ref{CwPredef} with any $\varphi\in (0,\pi/2]$. The operator $\tilde K_{\zeta}$ is defined in terms of its integral kernel given in \eqref{eqnkzetakernel} with $g_{w,w'}(q^{s})$ from \eqref{gwwprimeeqn} explicitly given in the pure $\tilde\beta$ specialization by \eqref{gwwprimeprime}.
\end{proposition}
\begin{proof}
In order to prove this result, we demonstrate that both sides of \eqref{eqnmmthd} are analytic in $\zeta$ as it varies within $\C\setminus\Rplus$. The identity for $|\zeta|$ small enough follows from Proposition~\ref{qFredDetThmbetarestricted} and the general $\zeta$ result then follows from uniqueness of analytic continuations. Throughout, let $\tilde a,\tilde\beta$ be fixed and let $D\subset \C\setminus\Rplus$ be any compact domain bounded away from $\R_{+}$. Also, let $\varphi\in (0,\pi/2]$ be fixed as well as the contours $\CwPre{\tilde a;0,\tilde\beta;\varphi}$ and $\CsPre{w}$ for each $w\in \CwPre{\tilde a;0,\tilde\beta;\varphi}$.
To establish the analyticity of the right-hand side of \eqref{eqnmmthd} observe that
$$
\EE_{\M_{\tilde a,\rho(0;\tilde \beta;0)}}\Bigg[ \frac{1}{\big(\zeta q^{-\lambda^{(N)}_1};q\big)_{\infty}}\Bigg] = \sum_{n=0}^{\infty}
\frac{\M_{\tilde a,\rho(0;\tilde \beta;0)}\big(\lambda^{(N)}_1 = n\big)}{(\zeta q^{-n};q)_{\infty}}
$$
is analytic over $\zeta\in\C\setminus\{q^{\ell}\}_{\ell\in\Z}$. This follows from the fact that for any region of $\C$ bounded away from $\{q^{\ell}\}_{\ell\in \Z}$, the function $\zeta\to (\zeta;q)_{\infty}$ is uniformly bounded from zero and analytic. This means that the above series is uniformly convergent in any such region. Since each term is analytic in $\zeta$, the series is as well.
To establish the analyticity of the left-hand side of \eqref{eqnmmthd}, we show that \mbox{$\zeta\mapsto \det(\Id+\tilde K_{\zeta})_{L^2(\CwPre{\tilde a; 0,\tilde\beta;\varphi})}$} is an analytic function of $\zeta$ in any domain $D$ bounded away from $\R_{+}$. Consider the Fredholm determinant expansion
\begin{equation}\label{eqnfredexpansionseries}
\det(\Id+\tilde K_{\zeta})_{L^2(\CwPre{\tilde a;0,\tilde\beta;\varphi})} = \sum_{L\geq 0} \frac{1}{L!} \int_{\CwPre{\tilde a;0,\tilde\beta;\varphi}}\frac{\d w_1}{2\pi \I} \cdots \int_{\CwPre{\tilde \alpha,\varphi}} \frac{\d w_L}{2\pi \I} \det\big[\tilde K_{\zeta}(w_i,w_j)\big]_{i,j=1}^L.
\end{equation}
We will show that for each $L\geq 0$, the corresponding summand is an analytic function of $\zeta\in D$ and that uniformly over $\zeta\in D$, the above series in $L$ is absolutely convergent. From this it will follow that the series itself is also analytic. Let us write $F_{L}(\zeta)$ for the $L^{th}$ summand in \eqref{eqnfredexpansionseries}:
\begin{multline*}
F_L(\zeta) = \frac{1}{L!} \int_{\CwPre{\tilde a;0,\tilde\beta;\varphi}}\frac{\d w_1}{2\pi \I}\cdots \int_{\CwPre{\tilde a;0,\tilde\beta;\varphi}}\frac{\d w_L}{2\pi \I} \int_{\CsPre{w_1}} \frac{\d s_1}{2\pi \I} \cdots \int_{\CsPre{w_n}} \frac{\d s_L}{2\pi \I} \det\left[\frac{1}{q^{s_i}w_i -w_j}\right]_{i,j=1}^{L}\\
\times \prod_{j=1}^{L} \left(\Gamma(-s_j)\Gamma(1+s_j) (-\zeta)^{s_j} \prod_{i=1}^{N}\frac{(q^{s_j}w_j \tilde a_i;q)_{\infty}}{(w_j \tilde a_i;q)_{\infty}}\prod_{i=1}^{M}\frac{1+\tilde\beta_i (q^{s_j}w_j)^{-1}}{1+\tilde\beta_i (w_j)^{-1}} \right).
\end{multline*}
We utilize the following readily proved estimates for the integrand in \eqref{eqnfredexpansionseries}. There exists $C_0>0$ such that for all $w\in \CwPre{\tilde a;0,\tilde\beta;\varphi}$ and all $s\in \CsPre{w}$
$$
\left|\prod_{i=1}^{N}\frac{(q^{s_j}w_j \tilde a_i;q)_{\infty}}{(w_j \tilde a_i;q)_{\infty}}\prod_{i=1}^{M}\frac{1+\tilde\beta_i (q^{s_j}w_j)^{-1}}{1+\tilde\beta_i (w_j)^{-1}}\right| \leq C_0\big(|w| q\big)^{-N x/2} q^{-M x}
$$
where we recall the notation $s=x+\I y$ and $M=M_{\beta}$ is the number of non-zero elements of $\tilde\beta$. Note that the constant $C_0$ depends on $\tilde a,\tilde\beta$ and the exact choice of contours. We may bound $\Gamma(-s)\Gamma(1+s) (-\zeta)^{s}$ as in Step 1e. For $s$ on the vertical portion of $\CsPre{w}$ (with real part $R$), there exists a constant $C_1>0$ such that
$$
\Big\vert \Gamma(-s)\Gamma(1+s) (-\zeta)^{s}\Big\vert \leq C_1 y^{-1} r^R e^{-y\sigma -\pi |y|}
$$
where we write $\zeta = r e^{i \sigma}$ with $\sigma\in (-\pi,\pi)$.
For $s$ on the rest of $\CsPre{w}$, there exists a constant $C_2>0$ such that
$$
\Big\vert \Gamma(-s)\Gamma(1+s) (-\zeta)^{s}\Big\vert \leq C_2 d^{-1} r^x,
$$
where $d$ comes from Definition~\ref{CwPredef}.
Finally, from Hadamard's inequality and the conditions we have imposed on $\CsPre{w_j}$ there exists a constant $C_3>0$ such that
\begin{equation*}
\left| \det\left[\frac{1}{q^{s_i}w_i -w_j}\right]_{i,j=1}^{L} \right| \leq C_3^L L^{L/2}.
\end{equation*}
Let us see how these bounds imply the analyticity of the fixed $L$ summand in \eqref{eqnfredexpansionseries} as well as the uniform absolute convergence of the series. The integrand in \eqref{eqnfredexpansionseries} is clearly analytic in $\zeta\in \C\setminus\Rplus$. Likewise holds true for the integral in \eqref{eqnfredexpansionseries} when $w_1,\ldots, w_L$ and $s_1,\ldots, s_L$ are restricted to compact portions of their respective contours. To show that the entire integral in \eqref{eqnfredexpansionseries} is analytic over $\zeta\in D$ (for some compact domain $D$ bounded away from $\R_{+}$) it suffices to show uniform integrability of the integrand as $\zeta\in D$ varies.
See also Lemma~\ref{lem:series_anal}.
Writing $\zeta = r e^{\I \sigma}$, let $r^{*}$ represent the maximal $r$ over $\zeta\in D$ and $\sigma^{*}$ represent the $\sigma$ which is closest to $\pm \pi$ over $\zeta \in D$. Then (with possibly modified values for $C_1,C_2,C_3$ to account for the approximations that for $|w|$ large, $R \approx \ln|w|$ and $d\approx |w|^{-1}$) we find that for $s$ on the vertical portion of $\CsPre{w}$ (with real part $R$) there is a constant $C_4>0$ such that
\begin{multline*}
\left|\Gamma(-s)\Gamma(1+s) (-\zeta)^{s} \prod_{i=1}^{N}\frac{(q^{s}w \tilde a_i;q)_{\infty}}{(w \tilde a_i;q)_{\infty}}\prod_{i=1}^{M}\frac{1+\tilde\beta_i (q^{s}w)^{-1}}{1+\tilde\beta_i (w)^{-1}} \right| \\
\leq C_0 C_1 y^{-1} e^{-|y|(\pi -\sigma^{*})} \big( r^{*} |w|^{-N/2} q^{-N/2 -M}\big)^{C_4 \ln |w|},
\end{multline*}
whereas for $s$ on the rest of $\CsPre{w}$
\begin{equation*}
\left|\Gamma(-s)\Gamma(1+s) (-\zeta)^{s} \prod_{i=1}^{N}\frac{(q^{s}w \tilde a_i;q)_{\infty}}{(w \tilde a_i;q)_{\infty}}\prod_{i=1}^{M}\frac{1+\tilde\beta_i (q^{s}w)^{-1}}{1+\tilde\beta_i (w)^{-1}} \right| \leq C_0 C_2 |w| \big(r^{*} |w|^{-N/2} q^{-N/2 -M}\big)^{x}.
\end{equation*}
Performing the integrals over $s_1,\ldots, s_L$ we find that some constants $C_5,C_6>0$,
\begin{equation*}\begin{aligned}
&\int_{\CsPre{w_1}} \frac{\d s_1}{2\pi \I} \cdots \int_{\CsPre{w_n}} \frac{\d s_L}{2\pi \I} \Bigg|\det\left[\frac{1}{q^{s_i}w_i -w_j}\right]_{i,j=1}^{L}\\
&\times \prod_{j=1}^{L} \left(\Gamma(-s_j)\Gamma(1+s_j) (-\zeta)^{s_j} \prod_{i=1}^{N}\frac{(q^{s_j}w_j \tilde a_i;q)_{\infty}}{(w_j \tilde a_i;q)_{\infty}}\prod_{i=1}^{M}\frac{1+\tilde\beta_i (q^{s_j}w_j)^{-1}}{1+\tilde\beta_i (w_j)^{-1}} \right)\Bigg|\\
&\leq C_3^L L^{L/2} \prod_{i=1}^L \left( C_5 \ln|w| \big(r^{*} |w_i|^{-N/2} q^{-N/2 -M}\big)^{C_4 \ln |w|} + C_6 |w| \frac{\big(r^{*} |w_i|^{-N/2} q^{-N/2 -M}\big)^{1/2}}{\ln(r^{*} |w_i|^{-N/2} q^{-N/2 -M}\big)}\right).
\end{aligned}\end{equation*}
The right-hand side above decays in large $|w|$ like $|w|^{1-N/4}$ (up to logarithmic corrections). Thus, given that $N\geq 9$, we find that for some constant $C_7>0$,
\begin{multline*}
\frac{1}{L!} \int_{\CwPre{\tilde a;0,\tilde\beta;\varphi}}\frac{\d w_1}{2\pi \I}\cdots \int_{\CwPre{\tilde a;0,\tilde\beta;\varphi}}\frac{\d w_L}{2\pi \I} \int_{\CsPre{w_1}} \frac{\d s_1}{2\pi \I} \cdots \int_{\CsPre{w_n}} \frac{\d s_L}{2\pi \I} \,\Bigg|\det\left[\frac{1}{q^{s_i}w_i -w_j}\right]_{i,j=1}^{L}\\
\times \prod_{j=1}^{L} \left(\Gamma(-s_j)\Gamma(1+s_j) (-\zeta)^{s_j} \prod_{i=1}^{N}\frac{(q^{s_j}w_j \tilde a_i;q)_{\infty}}{(w_j \tilde a_i;q)_{\infty}}\prod_{i=1}^{M}\frac{1+\tilde\beta_i (q^{s_j}w_j)^{-1}}{1+\tilde\beta_i (w_j)^{-1}} \right)\Bigg| \leq \frac{C_3^L L^{L/2}}{L!} C_7.
\end{multline*}
This implies $F_L(\zeta)$ is analytic, and it also shows that $|F_{L}(\zeta)| \leq \frac{C_3^L L^{L/2}}{L!} C_7$ uniformly over $\zeta \in D$. Hence follows the uniform absolute convergence and analyticity of the full series $\det(\Id+\tilde K_{\zeta})_{L^2(\CwPre{\tilde a;0,\tilde\beta;\varphi})}$ as well by Lemma~\ref{lem:series_anal}.
\end{proof}
\subsection{Step 2: Formal series identity}
In this step, we will complete the proof of Theorem~\ref{qFredDetThm} for general $\tilde\alpha,\tilde\beta,\tilde\gamma$ specializations. Recall that the algebra of symmetric functions $\Sym$ in formal variables $X=(x_1,x_2,\ldots)$ is algebraically generated by Newton power sums (for more background, see~\cite[Chapter~1]{Mac79})
$$
p_k(X) = \sum_i (x_i)^{k}, \qquad k=1,2,\ldots.
$$
For any partition $\lambda$, set $p_{\lambda}(X) = \prod_i p_{\lambda_i}(X)$. These form a linear basis for $\Sym$.
In order to prove the theorem, we must show that the identity in \eqref{thmlaplaceeqn} holds for general specializations satisfying the conditions of Definition~\ref{defParams}. We do this in three lemmas. In Lemma~\ref{lemstep2a} we establish that the left-hand side of \eqref{thmlaplaceeqn} can be expanded into a series in the $\rho(\tilde \alpha;\tilde \beta;\tilde \gamma)$ specialization of $p_{\lambda}(X)$ functions with coefficients $\ell_{\lambda}(\zeta,q,\tilde{a})$ which are independent of said specialization. In Lemma~\ref{lemstep2b} we do the same for the right-hand side of \eqref{thmlaplaceeqn} with coefficients $r_{\lambda}(\zeta,q,\tilde{a})$. In Lemma~\ref{lemstep2c} we observe that since Proposition~\ref{qFredDetThmbeta} amounts to the identity \eqref{thmlaplaceeqn} for all pure $\tilde\beta$ specializations, this implies the equality of all coefficients $\ell_{\lambda}(\zeta,q,\tilde{a})=r_{\lambda}(\zeta,q,\tilde{a})$ in the $p_k$ expansions. This along with the two previous lemmas, however,
implies that \eqref{thmlaplaceeqn} holds for all specializations for which the $p_k$ expansions are absolutely convergent -- in particular for the general $\rho(\tilde \alpha;\tilde \beta;\tilde \gamma)$ specialization satisfying Definition~\ref{defParams}. This completes the proof of Theorem~\ref{qFredDetThm}.
What remains, therefore, is to state and prove the three lemmas.
\begin{lemma}\label{lemstep2a}
There exist coefficients $\ell_{\lambda}=\ell_{\lambda}(\zeta,q,\tilde{a})$, depending on $\zeta\in \C\setminus\Rplus$, $q\in(0,1)$ and $\tilde{a}$ (but not $\rho(\tilde \alpha;\tilde \beta;\tilde \gamma)$) such that, for all specializations $\rho(\tilde \alpha;\tilde \beta;\tilde \gamma)$ satisfying Definition~\ref{defParams},
$$
\EE_{\M_{\tilde a,\rho(\tilde \alpha;\tilde \beta;\tilde \gamma)}}\Bigg[ \frac{1}{\big(\zeta q^{-\lambda^{(N)}_1};q\big)_{\infty}}\Bigg] = \sum_{\lambda} \ell_{\lambda}\, p_{\lambda}\big(\rho(\tilde \alpha;\tilde \beta;\tilde \gamma)\big).
$$
Moreover, the right-hand side is an absolutely convergent series for all such $\rho(\tilde \alpha;\tilde \beta;\tilde \gamma)$.
\end{lemma}
\begin{proof}
Let us first work in terms of the formal variables $X$ and the algebra $\Sym$. Since the $p_{\lambda}$ form a linear basis of $\Sym$, there are coefficients $c_{\lambda,\mu}$ with $|\lambda|=|\mu|$ such that
$$
Q_{\lambda}(X) = \sum_{\mu:|\mu|=|\lambda|} c_{\lambda,\mu} p_{\lambda}(X).
$$
Similarly, we can express
$
\Pi(\tilde a;X) = \sum_{\lambda} P_{\lambda}(\tilde a) Q_{\lambda}(X)
$
via the $p_{\lambda}(X)$ basis as
\begin{equation}\label{pieq}
\Pi(\tilde a;X) = \exp\left(\sum_{k=1}^{\infty} \frac{p_{k}(\tilde{a}) p_{k}(X)}{(1-q^k)k}\right).
\end{equation}
Using these expansions in $p_{\lambda}(X)$ functions we can write
\begin{equation}\label{eqnretg}
\sum_{\lambda^{(N)}} \frac{1}{\big(\zeta q^{-\lambda^{(N)}_1};q\big)_{\infty}} \frac{P_{\lambda}(\tilde a) Q_{\lambda}(X)}{\Pi(\tilde a;X)} = \sum_{\lambda} \ell_{\lambda}(\zeta,q,\tilde{a})\, p_{\lambda}(X).
\end{equation}
To establish the equality, we have used the above expansions of $Q$ and $\Pi$ into the $p_{\lambda}$ functions. It is easy to see that for each $p_{\lambda}(X)$ there are only a finite number of terms from these expansions which combine to form the coefficient $\ell_{\lambda}(\zeta,q,\tilde{a})$.
This determines the value of the coefficients $\ell_{\lambda}(\zeta,q,\tilde{a})$. It is, moreover, evident that the expansions are absolutely convergent for specializations $\rho(\tilde \alpha;\tilde \beta;\tilde \gamma)$ satisfying Definition~\ref{defParams}. The specialization of the right-hand side of \eqref{eqnretg} is identified with
$$\EE_{\M_{\tilde a,\rho(\tilde \alpha;\tilde \beta;\tilde \gamma)}}\Bigg[ \frac{1}{\big(\zeta q^{-\lambda^{(N)}_1};q\big)_{\infty}}\Bigg],$$ hence completing the proof of the lemma.
\end{proof}
\begin{lemma}\label{lemstep2b}
There exist coefficients $r_{\lambda}=r_{\lambda}(\zeta,q,\tilde{a})$, depending on $\zeta\in \C\setminus \Rplus$, $q\in (0,1)$ and $\tilde{a}$ (but not $\rho(\tilde \alpha;\tilde \beta;\tilde \gamma)$) such that, for all specializations $\rho(\tilde \alpha;\tilde \beta;\tilde \gamma)$ satisfying Definition~\ref{defParams},
$$
\det(\Id+\tilde K_{\zeta})_{L^2(\CwPre{\tilde a; \tilde \alpha,\tilde\beta;\varphi})} = \sum_{\lambda} r_{\lambda} \,p_{\lambda}\big(\rho(\tilde \alpha;\tilde \beta;\tilde \gamma)\big).
$$
Moreover, the right-hand side is an absolutely convergent series for all such $\rho(\tilde \alpha;\tilde \beta;\tilde \gamma)$.
\end{lemma}
\begin{proof}
Recall that the Fredholm determinant means the expansion
$$
\det(\Id+\tilde K_{\zeta})_{L^2(\CwPre{\tilde a; \tilde \alpha,\tilde\beta;\varphi})} = \sum_{L\geq 0}\frac{1}{L!} \int_{\CwPre{\tilde a; \tilde \alpha,\tilde\beta;\varphi}}\frac{\d w_1}{2\pi \I}\cdots \int_{\CwPre{\tilde a; \tilde \alpha,\tilde\beta;\varphi}}\frac{\d w_L}{2\pi \I} \det\left[ \tilde K_{\zeta}(w_i,w_j)\right]_{i,j=1}^L.
$$
The kernel is defined in \eqref{eqnkzetakernel} as
$$
\tilde K_{\zeta}(w,w') = \frac{1}{2\pi \I}\int_{\CsPre{w}} \Gamma(-s)\Gamma(1+s)(-\zeta)^s g_{w,w'}(q^s)\,\d s
$$
where, as in \eqref{gwwprimeeqn},
$$
g_{w,w'}(q^s) = \frac{1}{q^s w - w'}\, \frac{\Pi(w;a)}{\Pi(q^s w;a)}\, \frac{\Pi\big((q^s w)^{-1};\rho(\tilde\alpha;\tilde\beta;\tilde\gamma)\big)}{\Pi\big((w)^{-1};\rho(\tilde\alpha;\tilde\beta;\tilde\gamma)\big)}.
$$
For specializations $\rho(\tilde \alpha;\tilde \beta;\tilde \gamma)$ satisfying Definition~\ref{defParams}, we can choose the contours $\CwPre{\tilde a; \tilde \alpha,\tilde\beta;\varphi}$ and $\CsPre{w}$ as in Definition~\ref{CwPredef} in such as way that
$$\bigg|\frac{\tilde\alpha}{q^s w}\bigg|, \bigg|\frac{\tilde\alpha}{w}\bigg|, \bigg|\frac{\tilde\beta}{q^s w}\bigg|, \bigg|\frac{\tilde\beta}{w}\bigg|<1
$$ for all $w\in \CwPre{\tilde a; \tilde \alpha,\tilde\beta;\varphi}$ and $s\in \CsPre{w}$.
These conditions imply that the $\rho(\tilde \alpha;\tilde \beta;\tilde \gamma)$ specialization of the identities in \eqref{pieq} remain valid (with convergent right-hand sides). In particular, we find that the term in the formula for $g_{w,w'}(q^s)$ is
$$
\frac{\Pi\big((q^s w)^{-1};\rho(\tilde\alpha;\tilde\beta;\tilde\gamma)\big)}{\Pi\big((w)^{-1};\rho(\tilde\alpha;\tilde\beta;\tilde\gamma)\big)}
= \exp\left(\sum_{k=1}^{\infty} \frac{p_{k}\big((q^s w)^{-1}\big) p_{k}\big(\rho(\tilde\alpha;\tilde\beta;\tilde\gamma)\big)}{(1-q^k)k} - \frac{p_{k}\big((w)^{-1}\big) p_{k}\big(\rho(\tilde\alpha;\tilde\beta;\tilde\gamma)\big)}{(1-q^k)k} \right).
$$
We can substitute this convergent expansion into $g_{w,w'}(q^s)$ and thus express the Fredholm determinant expansion in terms of a convergent series in the $p_{\lambda}\big(\rho(\tilde\alpha;\tilde\beta;\tilde\gamma)\big)$:
$$
\det(\Id+\tilde K_{\zeta})_{L^2(\CwPre{\tilde a; \tilde \alpha,\tilde\beta;\varphi})} = \sum_{\lambda} r_{\lambda}(\zeta,q,\tilde{a})\, p_{\lambda}\big(\rho(\tilde\alpha;\tilde\beta;\tilde\gamma)\big).
$$
It is easily checked that the $L^{th}$ term in the Fredholm determinant expansion contributes to $p_{\lambda}$'s with $|\lambda|\geq L$, and hence each coefficient is well-defined and finite. The convergence of this sum follows from the convergence of the expansion into the $p_{k}\big(\rho(\tilde\alpha;\tilde\beta;\tilde\gamma)\big)$ as well as the convergence of the Fredholm determinant expansion.
\end{proof}
\begin{lemma}\label{lemstep2c}
For any $\zeta\in \C\setminus\Rplus$, $q\in(0,1)$ and $\tilde{a}$, we have
$$
\ell_{\lambda}(\zeta,q,\tilde{a})=r_{\lambda}(\zeta,q,\tilde{a})
$$
for all partitions $\lambda$.
\end{lemma}
\begin{proof}
Proposition~\ref{qFredDetThmbeta} implies that the left-hand sides of the identities in Lemmas~\ref{lemstep2a} and~\ref{lemstep2b} are equal for all pure $\tilde\beta$ specializations. The right-hand sides of the identities in Lemmas~\ref{lemstep2a} and~\ref{lemstep2b} are therefore also equal under these specializations:
$$
\sum_{\lambda} \ell_{\lambda}(\zeta,q,\tilde a)\, p_{\lambda}\big(\rho(0;\tilde \beta;0)\big)=\sum_{\lambda} r_{\lambda}(\zeta,q,\tilde a) \, p_{\lambda}\big(\rho(0;\tilde \beta;0)\big).
$$
View both sides as power series in individual $\tilde{\beta}_j$'s. The equality implies equality of parts of fixed degree. Moreover, $p_{\lambda}\big(\rho(0;\tilde \beta;0)\big)$ for fixed $|\lambda|$ are linearly independent for sufficiently many nonzero $\tilde{\beta}_j$'s. Therefore, since this equality holds for general $\tilde\beta$ specializations, it implies equality of the expansion coefficients.
\end{proof}
\section{Whittaker processes}\label{SectConvWhitt}
\subsection{Defining the processes}
Let us introduce some notation. Write $T$ for the triangular array $\big(T^{(k)}_j\big)_{1\leq j\leq k\leq N}$ with entries in $\R$. Alternatively, write $T=\big(T^{(1)},\ldots, T^{(N)}\big)$ with \mbox{$T^{(k)} = \big(T^{(k)}_1,\ldots, T^{(k)}_k\big)$}. Also, write $\nu= (\nu_1,\ldots,\nu_N)\in \R^N$.
\begin{definition}\label{defWhitfunc}
As shown by Givental~\cite{Giv97}, the class-one $\mathfrak{gl}_{N}$-Whittaker functions admit the following integral representation:
$$
\psi_{\nu}(T^{(N)}) = \int_{\R^{\frac{N(N-1)}{2}}} e^{\mathcal{F}_{\nu}(T)} \prod_{1\leq j\leq k\leq N-1} \d T^{(k)}_j
$$
where the integral is over all $T$ with fixed $T^{(N)}$, and where
$$
\mathcal{F}_{\nu}(T)=\I\sum_{n=1}^{N} \nu_n\left(\sum_{i=1}^n T^{(n)}_i-\sum_{i=1}^{n-1} T^{(n-1)}_i\right)-\sum_{n=1}^{N-1}\sum_{i=1}^n \left(e^{T^{(n)}_i-T^{(n+1)}_i}+e^{T^{(n+1)}_{i+1}-T^{(n)}_i}\right).
$$
\end{definition}
We now define a class of Whittaker processes which are composites of those which arose in~\cite{OCon09,COSZ11}.
\begin{definition}
Fix integers $N\geq 1$, $M\geq 0$, vectors $a=(a_1,\ldots, a_N)$, $\alpha=(\alpha_1,\ldots, \alpha_M)$, and $\tau\geq 0$, such that $\alpha_m>0$ for all $1\leq m\leq M$ and $\alpha_m+a_n>0$ for all $1\leq n\leq N$ and $1\leq m\leq M$. The {\it Whittaker process} corresponding to these parameters is a probability measure on $\R^{\frac{N(N+1)}{2}}$ with density function (with respect to Lebesgue measure) given by
$$
\W{a;\alpha,\tau}(T) = e^{-\tau \sum_{n=1}^{N} \frac{a_n^2}{2}} \prod_{n=1}^{N}\prod_{m=1}^{M} \frac{1}{\Gamma(\alpha_m+a_n)} e^{\mathcal{F}_{\I a}(T)} \theta_{\alpha,\tau}(T^{(N)})
$$
with
$$
\theta_{\alpha,\tau}(T^{(N)}) = \int_{\R^N}\psi_{\nu}(T^{(N)}) e^{-\tau \sum_{n=1}^{N} \frac{\nu_n^2}{2}} \prod_{n=1}^{N}\prod_{m=1}^{M} \Gamma(\alpha_m - \I \nu_n) m_N(\nu)\,\d\nu_1\ldots\d\nu_N,
$$
and the Skylanin measure $m_N$ defined as
$$
m_N(\nu) = \frac{1}{(2\pi)^N N!} \prod_{1\leq j\neq k\leq N} \frac{1}{\Gamma(\I \nu_k - \I \nu_j)}.
$$
The Whittaker measure $\WM{a;\alpha,\tau}(T^{(N)})$ is the marginal of the Whittaker process $\W{a;\alpha,\tau}(T)$ on $T^{(N)}$ as defined in~\cite[Definition 4.1.16]{BC11}.
\end{definition}
\subsection{Whittaker processes and the semi-discrete directed random polymer}
The following result connects the developments of Sections~\ref{SectMacdonald} and~\ref{SectConvWhitt} with the study of the partition function for the semi-discrete directed random polymer with log-gamma boundary sources.
\begin{theorem}\label{ThmConnectWhitpoly}
Fix integers $N\geq 1$, $M\geq 0$ and $\tau\geq 0$. Let $a=(a_1,\ldots,a_N)\in \R^N$ and \mbox{$\alpha = (\alpha_1,\ldots,\alpha_M)\in \big(\R_{>0}\big)^M$} be such that $\alpha_m-a_n>0$ for all $1\leq n\leq N$ and $1\leq m\leq M$. Recall $\mathbf{F}_j^{k,M}(\tau)$ defined in \eqref{eqfreeenergy}. Then $\big\{\mathbf{F}_j^{k,M}(\tau)\big\}_{1\leq j\leq k\leq N}$ is distributed according to the Whittaker process $\W{-a;\alpha,\tau}$, where $-a = (-a_1,\ldots, -a_N)$.
\end{theorem}
\begin{proof}
This result relies on a combination of the work of~\cite{OCon09} on the semi-discrete directed random polymer and of~\cite{COSZ11} on the log-gamma discrete directed polymer. Those papers use geometric liftings of the Robinson--Schensted--Knuth correspondence to relate the polymer partition functions to pure $\tilde\gamma$ and pure $\tilde\alpha$ specialized Whittaker processes. The present result follows by combining~\cite[Theorems~3.7 and 3.9]{COSZ11} with~\cite[Theorem~3.1]{OCon09}. See also~\cite[Section~5.2.1]{BC11} for the $M=0$ case.
\end{proof}
\begin{remark}\label{remint}
It follows from Theorem~\ref{ThmConnectWhitpoly} that the Whittaker process is positive and integrates to 1. It should also be possible to show this directly in the manner of~\cite[Proposition~4.1.18]{BC11}.
\end{remark}
\subsection{Convergence of $q$-Whittaker processes to Whittaker processes}
We start by recording how $q$-Whittaker polynomials limit to Whittaker functions. Note that in the scalings which we describe below, it is understood that when it is necessary to work with integers, we take the integer part of $\e$ dependent expressions.
\begin{proposition}[Theorem~4.1.7 of~\cite{BC11}]\label{qwhitconvTHM}
For $N\geq 1$, consider the scalings
$$q=e^{-\e},\qquad \mathcal{A}(\e) = -\e^{-1}\frac{\pi^2}{6} - \frac{1}{2}\ln\frac{\e}{2\pi}$$
and for $1\leq n\leq N$
$$z_n =e^{\I \e \nu_n},\qquad \lambda^{(N)}_n = (N-2n)\e^{-1}\log\e^{-1} + \e^{-1} T^{(N)}_n.$$
Define rescaled (and index, variable flipped) $q$-Whittaker functions by
$$
\psi^{\e}_{\nu}(T^{(N)}) = \e^{\frac{N(N-1)}{2}} e^{\frac{N(N+2)}{2} \mathcal{A}(\e)} P_{\lambda^{(N)}}(z).
$$
Then, for all $\nu\in \R^N$, we have the following:
\begin{enumerate}
\item For each $\sigma\subset \{1,\ldots,N-1\}$, there exists a polynomial $R_{N,\sigma}$ of $N$ variables (chosen independently of $\nu_1,\ldots,\nu_N$ and $\e$) such that for all $T^{(N)}\in \R^N$ with
$$\sigma = \sigma(T^{(N)}) := \big\{n\in \{1,\ldots,N-1\}: T^{(N)}_n-T^{(N)}_{n+1}\leq 0\big\},$$
we have the following estimate: for some $c^*>0$
\begin{equation*}
\big|\psi^{\e}_{\nu}(T^{(N)})\big| \leq R_{N,\sigma(T^{(N)})}(T^{(N)}) \prod_{n\in \sigma(T^{(N)})} \exp\{-c^* e^{-(T^{(N)}_n-T^{(N)}_{n+1})/2}\}.
\end{equation*}
\item For $(T^{(N)})$ varying in a compact domain of $\R^{N}$, $\psi^{\e}_{\nu}(T^{(N)})$ converges (as $\e$ goes to zero) uniformly to $\psi_{\nu}(T^{(N)})$.
\end{enumerate}
\end{proposition}
\begin{theorem}\label{thmweakconv}
Fix integers $N\geq 1$, $M\geq 0$, vectors $a=(a_1,\ldots, a_N)$, $\alpha=(\alpha_1,\ldots, \alpha_M)$ and $\tau> 0$ such that $\alpha_m>0$ for all $1\leq m\leq M$ and $\alpha_m+a_n>0$ for all $1\leq n\leq N$ and $1\leq m\leq M$. Introduce the following $\e>0$ dependent scalings:
\begin{equation}\label{pgfourfour}\begin{gathered}
q= e^{-\e},\qquad \tilde \gamma= \tau \e^{-2},\qquad \tilde{a}_n= e^{-a_n\e},\, 1\leq n\leq N,\qquad \tilde{\alpha}_m= e^{-\alpha_m\e},\, 1\leq m\leq M,\\
\lambda^{(k)}_{j} = \tau \e^{-2} + M \e^{-1}\ln\e^{-1} + (k+1-2j)\e^{-1}\ln\e^{-1} + T^{(k)}_{j} \e^{-1},\, 1\leq j\leq k\leq N.
\end{gathered}\end{equation}
The $q$-Whittaker process $\M_{\tilde a,\rho(\tilde \alpha;0;\tilde \gamma)}\big(\la^{(1)},\ldots,\la^{(N)}\big)$ induces an $\e$-indexed measure on $T$ which converges weakly, as $\e\to 0$, to the Whittaker process
$\W{a;\alpha,\tau}(T)$.
\end{theorem}
\begin{remark}
The above theorem only deals with convergence of the $\tilde\alpha,\tilde\gamma$ specialized \mbox{$q$-Whittaker} process. It is presently unclear whether the $\tilde\beta$ specialized process admits a non-trivial limit as $q\to 1$.
\end{remark}
\begin{proof}
This proof is quite similar to that of~\cite[Theorems~4.1.12 and 4.2.4]{BC11} which work with (respectively) the pure $\tilde \gamma$ and pure $\tilde \alpha$ cases. It should be noted, however, that the pure $\tilde \alpha$ case~\cite[Theorem~4.2.4]{BC11} was stated modulo a decay estimate which was not checked. By combining the $\tilde \gamma$ with the $\tilde \alpha$ specialization, the necessary decay is easily shown to hold. On account of the similarities to those theorems, we only include the steps of the proof and refer to the proofs from~\cite{BC11} for the justification of the estimates.
The $q$-Whittaker process which we seek to asymptotically analyze is given as
$$
\M_{\tilde a,\rho(\tilde \alpha;0;\tilde \gamma)}\big(\la^{(1)},\ldots,\la^{(N)}\big) = \frac{P_{\la^{(1)}}(\tilde a_1)P_{\la^{(2)}/\la^{(1)}}(\tilde a_2)\cdots P_{\la^{(N)}/\la^{(N-1)}}(\tilde a_N) Q_{\la^{(N)}}\big(\rho(\tilde \alpha;0;\tilde \gamma)\big)}{\Pi\big(\tilde a;\rho(\tilde \alpha;0;\tilde \gamma)\big)}.
$$
Through the association of the $\la^{(k)}_{j}$ with the $T^{(k)}_j$ given in \eqref{pgfourfour}, this measure is pushed forward to one on $T$. It suffices to show that for any compact set $D\in \R^{\frac{N(N+1)}{2}}$, the convergence (as $\e$ goes to zero) is uniform as $T$ varies in $D$. This is due to the positivity of the measure and our independent knowledge (see Remark~\ref{remint}) that the limiting density integrates to 1. In order to estimate the behavior of the $q$-Whittaker process, we split it up into three lemmas, the combination of which proves the theorem.
\begin{lemma}
Fix any compact subset $D\in \R^{\frac{N(N+1)}{2}}$. Then
$$
P_{\la^{(1)}}(\tilde a_1)P_{\la^{(2)}/\la^{(1)}}(\tilde a_2)\cdots P_{\la^{(N)}/\la^{(N-1)}}(\tilde a_N) = \Big(e^{-\frac{(N-1)(N-2)}{2} \mathcal{A}(\e)} e^{-\e^{-1} \tau \sum_{n=1}^{N} a_n} \e^{M\sum_{n=1}^{N} a_n}\Big) \mathcal{F}_{\I a}(T) e^{o(1)}
$$
where the $o(1)$ error goes uniformly (with respect to $T\in D$) to zero as $\e$ goes to zero.
\end{lemma}
\begin{proof}
This is proved by combining the computations of~\cite[Lemmas~4.1.23 and 4.2.5]{BC11}.
\end{proof}
\begin{lemma}
We have
$$
\Pi\big(\tilde a;\rho(\tilde \alpha;0;\tilde \gamma)\big) = \Big(e^{\tau N \e^{-2}} e^{-\e^{-1}\tau\sum_{n=1}^{N} a_n} \prod_{n=1}^{N}\prod_{m=1}^{M}\frac{1}{e^{\mathcal{A}(\e)}\e^{1-\alpha_m-a_n}}\Big) e^{\tau \sum_{n=1}^{N}a_n^2/2} \prod_{n=1}^{N}\prod_{m=1}^{M} \Gamma(\alpha_m+a_n) e^{o(1)}
$$
where the $o(1)$ error goes to zero as $\e$ goes to zero.
\end{lemma}
\begin{proof}
This is proved by combining the computations of~\cite[Lemmas~4.1.24 and 4.2.6]{BC11}.
\end{proof}
\begin{lemma}
Fix any compact subset $D\in \R^{\frac{N(N+1)}{2}}$. Then
$$
Q_{\la^{(N)}}\big(\rho(\tilde \alpha;0;\tilde \gamma)\big) = \Big(e^{\frac{(N-1)(N-2)}{2}\mathcal{A}(\e)} e^{\tau N \e^{-2}} \e^{\frac{N(N+1)}{2}} \prod_{m=1}^{M} \e^{N(\alpha_m-1)}\Big) \theta_{\alpha,\tau}(T^{(N)}) e^{o(1)}
$$
where the $o(1)$ error goes uniformly (with respect to $T\in D$) to zero as $\e$ goes to zero.
\end{lemma}
\begin{proof}
This is proved by combining the computations of~\cite[Lemmas~4.1.25 and 4.2.7]{BC11}. However, since the result of~\cite[Lemma~4.2.7]{BC11} was stated modulo a decay estimate, we will provide the steps to prove the above result. That decay estimate is readily confirmed in the present case because of the presence of the $\tilde \gamma$ specialization, which provides ample decay. In~\cite{BC11}, the proof of these analogous lemmas split into four steps. It is only in the fourth step where a bit more justification is needed, which we give.
We employ the torus scalar product~\cite[Section~2.1.5]{BC11} with respect to which the Macdonald polynomials are orthogonal (we keep $t=0$ and use the notation $\T^N$ to represent the torus $\{z:|z_1|,\ldots, |z_N|=1\}$):
\begin{equation*}
\langle f,g\rangle'_N = \int_{\T^N} f(z) \overline{g(z)} m_N^{q}(z) \prod_{n=1}^{N} \frac{\d z_n}{z_n}, \qquad m_N^q(z) = \frac{1}{(2\pi \I)^{N} N!}\prod_{1\leq j\neq k\leq N} (z_jz_k^{-1};q)_{\infty}.
\end{equation*}
Note that taking $t=0$~\cite[equation~(2.8)]{BC11} yields
\begin{equation*}
\langle P_{\lambda^{(N)}},P_{\lambda^{(N)}}\rangle'_N = \prod_{n=1}^{N-1} (q^{\lambda^{(N)}_n-\lambda^{(N)}_{n+1}+1};q)_{\infty}^{-1}.
\end{equation*}
Recalling the definition of $\Pi$, we may write
\begin{equation*}
Q_{\lambda^{(N)}}(\rho) = \frac{1}{\langle P_{\lambda^{(N)}},P_{\lambda^{(N)}}\rangle'_N} \big\langle \Pi(z_1,\ldots, z_N;\rho),P_{\lambda^{(N)}}(z_{1},\ldots, z_{N})\big\rangle'_N.
\end{equation*}
Therefore, in order to study the asymptotic behavior of $Q_{\la^{(N)}}\big(\rho(\tilde \alpha;0;\tilde \gamma)\big)$, we will study the torus scalar product above. Let us introduce one additional scaling to those in \eqref{pgfourfour} that for $1\leq n\leq N$, $z_n = e^{\e \I\nu_n}$.
In Step 1 we show that $\langle P_\lambda,P_\lambda\rangle'_N = e^{o(1)}$ where the $o(1)$ error goes uniformly (with respect to $T\in D$) to zero as $\e$ goes to zero. The proof from~\cite[Lemma~4.1.25]{BC11} applies just as well here.
In Step 2 we find that for any compact subset $V\subset \R^{N}$,
\begin{align*}
\Pi\big(z_1,\ldots, z_N;\rho(\tilde \alpha;0;\tilde \gamma)\big) &= E_{\Pi} e^{-\tau \sum_{n=1}^{N} \nu_n^2/2} \prod_{n=1}^{N}\prod_{m=1}^{M}\Gamma(\alpha_m-\I\nu_n) e^{o(1)},\\
E_{\Pi}&= e^{\tau N \e^{-2}} e^{\tau \e^{-1} \I \sum_{n=1}^{N} \nu_n} \prod_{n=1}^{N}\prod_{m=1}^{M}\frac{1}{e^{\mathcal{A}(\e)}\e^{1-\alpha_m+\I \nu_n}},
\end{align*}
and, using Proposition~\ref{qwhitconvTHM}, we find
\begin{align*}
P_{\lambda^{(N)}}(z_1,\ldots, z_N) &= E_{P}\,\psi_{\nu}(T^{(N)}) e^{o(1)},\\
E_{P}&=\e^{-\frac{N(N-1)}{2}} e^{-\frac{(N-1)(N+2)}{2}\mathcal{A}(\e)}e^{\tau \e^{-1}\I\sum_{n=1}^{N} \nu_n} \e^{-M\sum_{n=1}^{N} \I \nu_n}
\end{align*}
where the $o(1)$ error goes uniformly (with respect to $T\in D$ and $\nu\in V$) to zero as $\e$ goes to zero. The proof from~\cite[Lemma~4.1.25]{BC11} applies just as well here.
In Step 3 we find that for any compact subset $V\subset \R^{N}$,
\begin{equation*}
m_N^{q}(z)\prod_{n=1}^{N} \frac{\d z_n}{z_n} = E_m m_N(\nu) \prod_{n=1}^{N} \d\nu_n e^{o(1)}, \qquad E_{m}=\e^{N^2}e^{N(N-1)\mathcal{A}(\e)}
\end{equation*}
where the $o(1)$ error goes uniformly (with respect to $\nu\in V$) to zero as $\e$ goes to zero. The proof from~\cite[Lemma~4.1.25]{BC11} applies just as well here.
In Step 4 we find that
\begin{multline*}
\big\langle \Pi(z_1,\ldots, z_N;\rho),P_{\lambda^{(N)}}(z_{1},\ldots, z_{N})\big\rangle'_N \\
= \left(e^{\frac{(N-1)(N-2)}{2}\mathcal{A}(\e)} e^{\tau N \e^{-2}} \e^{\frac{N(N+1)}{2}}\prod_{m=1}^{M} \frac{1}{\e^{N(1-\alpha_m)}} \right)\theta_{\tau}(T^{(N)}) e^{o(1)}
\end{multline*}
where the $o(1)$ error goes uniformly (with respect to $T\in D$) to zero as $\e$ goes to zero.
The proof from~\cite[Lemma~4.1.25]{BC11} applies just as well here, though we need to check that the following inequality still holds: for all $\nu_n\in [-\e^{-1}\pi,\e^{-1}\pi]$, $1\leq n\leq N$,
\begin{equation}\label{decaynueqn}
\left|\frac{\Pi(z_1,\ldots, z_N;\rho)}{E_{\Pi}}\right| \leq e^{-\tau \sum_{n=1}^{N} \nu_n^2/6}.
\end{equation}
This was checked in Step 4 of the proof of~\cite[Lemma~4.1.25]{BC11} for $\rho =\rho(0;0;\tilde \gamma)$. It is, however, easily confirmed that including the $\tilde\alpha$ specialization as well as the $\tilde \gamma$ one does not increase the left-hand side of \eqref{decaynueqn}. In particular, this amounts to showing that for all $\nu\in [-\e^{-1}\pi,\e^{-1}\pi]$ (and $\alpha>0$ fixed),
$$
\frac{e^{\mathcal{A}(\e)} \e^{1-\alpha+\I \nu}}{(e^{-\e \alpha}e^{\e \I \nu};e^{-\e})_{\infty}} = \Gamma_q(\alpha- \I \nu)
$$
is bounded by a constant (cf.\ Appendix~\ref{qSec} for more on the $q$-Gamma function). This is easily checked, hence Step 4 proceeds and the lemma is proved as in~\cite[Lemma~4.1.25]{BC11}.
\end{proof}
As in the proof of~\cite[Theorems~4.1.12 and 4.2.4]{BC11}, the above three lemmas (along with the Jacobian factor of $\e^{\frac{N(N+1)}{2}}$ coming from the rescaling of the $q$-Whittaker process) implies Theorem~\ref{thmweakconv}.
\end{proof}
\section{Semi-discrete polymer with boundary sources -- Proof of Theorem~\ref{ThmFormulaSemiDiscrete}}\label{SectSemiDirected}
\begin{theorem}\label{ThmFormulaWhit}
Fix $N\geq 9$, $M\geq 0$, $\tau>0$ and vectors $a=(a_1,\ldots, a_N)\in \R^N$ and \mbox{$\alpha=(\alpha_1,\dots,\alpha_M)\in\R^M$} so that $\alpha_m-a_n>0$ for all $1\leq m\leq M$, $1\leq n\leq N$. Then for all $u\in \C$ with positive real part
\begin{equation*}
\EE_{\W{-a;\alpha,\tau}}\Big[e^{-u e^{T^{(N)}_{1}}} \Big] = \det(\Id+ K_{u})_{L^2(\Cv{a;\alpha;\varphi})}\, ,
\end{equation*}
where the operator $K_u$ is defined as in Theorem~\ref{ThmFormulaSemiDiscrete} and the contour $\Cv{a;\alpha;\varphi}$ is given in Definition~\ref{DefCaCsdefBis} with any $\varphi\in(0,\pi/4)$.
\end{theorem}
Before proving this theorem, let us see how, combined with the earlier result of Theorem~\ref{ThmConnectWhitpoly}, the above theorem yields Theorem~\ref{ThmFormulaSemiDiscrete}.
\begin{proof}[Proof of Theorem~\ref{ThmFormulaSemiDiscrete}]
Theorem~\ref{ThmConnectWhitpoly} implies that $\mathbf{Z}_1^{N,M}(\tau)$ is equal in distribution to $e^{T^{(N)}_1}$ where $T$ is distributed according to the Whittaker process $\W{-a;\alpha,\tau}$. Theorem~\ref{ThmFormulaWhit} provides a Fredholm determinant formula for the Laplace transform of $e^{T^{(N)}_1}$ which implies Theorem~\ref{ThmFormulaSemiDiscrete}.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{ThmFormulaWhit}]
The proof of Theorem~\ref{ThmFormulaWhit} follows a similar line as that of~\cite[Theorem~4.5]{BCF12}. We proceed in two steps. In the first step we prove:
\begin{lemma}\label{lemstep1}
Under the scalings from \eqref{pgfourfour} and with $\zeta = -\e^{M+N}e^{-\e^{-1} \tau} u$, for all
$u\in \C$ with positive real part
\begin{equation}\label{thmlaplaceeqnlemma}
\lim_{\e\to 0}\EE_{\M_{\tilde a^{-1},\rho(\tilde \alpha;0;\tilde \gamma)}}\Bigg[ \frac{1}{\big(\zeta q^{-\lambda^{(N)}_1};q\big)_{\infty}}\Bigg] = \EE_{\W{-a;\alpha,\tau}}\Big[e^{-u e^{T^{(N)}_{1}}} \Big]
\end{equation}
\end{lemma}
In the second step we prove:
\begin{proposition}\label{propstep2}
Under the scalings from \eqref{pgfourfour} and with $\zeta = -\e^{M+N}e^{-\e^{-1} \tau} u$, for all
$u\in \C$ with positive real part
$$\lim_{\e\to 0} \det(\Id+\tilde K_{\zeta})_{L^2(\CwPre{\tilde a^{-1}; \tilde \alpha,0;\varphi})} = \det(\Id+ K_{u})_{L^2(\Cv{a;\alpha;\varphi})}.$$
\end{proposition}
Combining these two results along with Theorem~\ref{qFredDetThm} (which shows the equivalence of the left-hand sides of these two results) immediately yields Theorem~\ref{ThmFormulaWhit}.
\end{proof}
\subsection{Step 1: Proof of Lemma~\ref{lemstep1}}
Rewrite the left-hand side of equation \eqref{thmlaplaceeqnlemma} as
\begin{equation*}
\lim_{\e\to 0}\EE_{\M_{\tilde a^{-1},\rho(\tilde \alpha;0;\tilde \gamma)}}\Bigg[ \frac{1}{\big(\zeta q^{-\lambda^{(N)}_1};q\big)_{\infty}}\Bigg] = \lim_{\e\to 0}\EE_{\M_{\tilde a^{-1},\rho(\tilde \alpha;0;\tilde \gamma)}}\big[e_q(x_q)\big]
\end{equation*}
where
\begin{equation*}
x_q=(1-q)^{-1}\zeta q^{\lambda^{(N)}_1} =-ue^{-T^{(N)}_1} \e/(1-q)
\end{equation*}
and
$$
e_q(x) = \frac{1}{\big((1-q)x;q\big)_{\infty}}
$$
is a $q$-exponential (cf.\ Appendix~\ref{qSec}). Combine this with the fact that $e_q(x)\to e^{x}$ uniformly on $x\in~(-\infty,0)$ to show that, considered as a function of $T^{(N)}_1$, $e_q(x_q)\to e^{-u e^{-T^{(N)}_1}}$ uniformly over $T^{(N)}_1\in \R$. By Theorem~\ref{thmweakconv}, the measure on $T$ induced from the $q$-Whittaker process on $\la^{(1)},\ldots,\la^{(N)}$ converges weakly in distribution to the Whittaker process $\W{-a;\alpha,\tau}$. Combining this weak convergence with the uniform convergence of $e_q(x_q)$ and Lemma~\ref{problemma2} completes the proof of Lemma~\ref{lemstep1}.
\subsection{Step 2: Proof of Proposition~\ref{propstep2}}
Employing the change of variables $w=q^v$ and $w'=q^{v'}$, the kernel in the left-hand side of Proposition~\ref{propstep2} can be rewritten as
\begin{equation*}
\det(\Id+\tilde K_{\zeta})_{L^2(\CwPre{\tilde a^{-1}; \tilde \alpha,0;\varphi})} = \det(\Id+K_u^{\e})_{L^2(\CvEps{a;\alpha;\varphi})}.
\end{equation*}
Here, the kernel $K_u^{\e}$ is given by
\begin{equation}\label{445}
K_u^{\e}(v,v') = \frac{1}{2\pi \I}\int_{\CsPre{q^v}}h^q(s)\,\d s
\end{equation}
where (cf.\ Appendix~\ref{qSec} for the definition of $\Gamma_{q}$)
\begin{multline}\label{446}
h^q(s)=\Gamma(-s)\Gamma(1+s)\left(\frac{-\zeta}{(1-q)^{M+N}}\right)^s \frac{q^v \ln q}{q^{s+v} - q^{v'}} e^{\tilde \gamma q^{-v}(q^{-s}-1)}\\
\times \prod_{n=1}^{N} \frac{\Gamma_q(v- a_n)}{\Gamma_q(s+v- a_n)} \prod_{m=1}^M \frac{\Gamma_q(\alpha_m-s-v)}{\Gamma_q(\alpha_m-v)}.
\end{multline}
The contour on which this kernel $K_u^{\e}$ acts is the image of the contour $\CwPre{\tilde a^{-1}; \tilde \alpha;\varphi}$ under the map $x\mapsto \ln_q x$ and the contour $\CsPre{q^v}$ is as in Definition~\ref{CwPredef}.
There was some freedom in specifying the contour $\CwPre{\tilde a^{-1}; \tilde \alpha;\varphi}$. It will be convenient for us to fix a particular contour in performing asymptotics. Let $\mu = \tfrac{1}{2}\max(a)+\tfrac{1}{2}\min(\alpha)$.
Then we define the contour $\CvEps{a;\alpha;\varphi}$ as the image of \mbox{$q^\mu + e^{\pm \varphi\I}\Rplus $} under the map $x\mapsto \ln_{q} x$.
This contour is illustrated in Figure~\ref{whitasymcontours}.
Note that as $\e\to 0$ this contour converges locally uniformly to $\Cv{a;\alpha;\varphi}$ from Definition~\ref{DefCaCsdefBis}, as can readily be seen by Taylor expanding the map $x\mapsto \ln_{q} x$.
It follows from the above observation that the contour on which the kernel $K_u^{\e}$ is defined converges as $\e\to 0$ to the contour $\Cv{a;\alpha;\varphi}$
on which the kernel in Theorem~\ref{ThmFormulaWhit} is defined.
Let us now likewise demonstrate the pointwise convergence of the integrand in the integral \eqref{445} defining kernel $K_u^{\e}$ to that of the kernel $K_u$.
Consider the behavior of each term as $q\to 1$ (or equivalently as $\e\to 0$ as $q=e^{-\e}$):
\begin{align}
e^{\tau s \e^{-1}}\left(\frac{-\zeta}{(1-q)^{N+M}}\right)^s & \to u^s\,,\label{pwlimits1}\\
\frac{q^v \ln q}{q^{s+v} - q^{v'}} &\to \frac{1}{v+s-v'}\,, \label{pwlimits2}\\
\frac{\Gamma_q(v-a_m)}{\Gamma_q(v+s-a_m)} & \to \frac{\Gamma(v-a_m)}{\Gamma(s+v-a_m)}\,,\label{pwlimits3}\\
\frac{\Gamma_q(\alpha_m-s-v)}{\Gamma_q(\alpha_m-v)} & \to \frac{\Gamma(\alpha_m-s-v)}{\Gamma(\alpha_m-v)}\,,\label{pwlimits3.5}\\
e^{-\tau s \e^{-1}} \exp\left(\tilde \gamma q^{-v}(q^{-s}-1)\right) & \to e^{v \tau s + \tau s^2/2}\,.\label{pwlimits4}
\end{align}
Combining these pointwise limits together gives the integrand of the kernel $K_u$ given in Theorem~\ref{ThmFormulaSemiDiscrete}.
In order to prove convergence of the Fredholm determinant, one needs more than just pointwise convergence.
There are four things we must do to complete \emph{Step 2} and prove convergence of the determinants.
In proving convergence of Fredholm determinants it is convenient to have the contour on which the operators act be fixed as $\e$ varies.
\smallskip
In \emph{Step 2a} we deform $\CvEps{a;\alpha;\varphi}$ to a contour $\CvEpsR{a;\alpha;\varphi;r}$ with a portion $\CvRLeq{a;\alpha;\varphi;<r}$ (of distance $<r$ to the origin)
which coincides with the limiting contour $\Cv{a;\alpha;\varphi}$.
\smallskip
Then in \emph{Step 2b} we show that for any fixed $\kappa>0$, by choosing $\e_0$ small enough and $r_0$ large enough,
for all $\e<\e_0$ and $r>r_0$ the determinant restricted to $L^2(\CvRLeq{a;\alpha;\varphi;<r})$ differs from the entire determinant on $L^2(\CvEpsR{a;\alpha;\varphi;r})$ by less than $\kappa$.
Thus, at an arbitrarily small cost of $\kappa$, we can restrict to a sufficiently large radius on which the contour is independent of $\e$.
\smallskip
In \emph{Step 2c} we show that for any $\kappa>0$, for $\e$ small, the Fredholm determinant of $K_u^{\e}$ restricted to $L^2(\CvRLeq{a;\alpha;\varphi;<r})$
differs by at most $\kappa$ from the Fredholm determinant of $K_u$ restricted to the same space.
\smallskip
Finally, \emph{Step 2d} shows that for $r_0$ large enough, for all $r>r_0$ the Fredholm determinant of $K_u$ restricted to $L^2(\CvRLeq{a;\alpha;\varphi;<r})$
differs from the Fredholm determinant of $K_u$ on $L^2(\Cv{a;\alpha;\varphi})$ by at most $\kappa$.
Summing up the steps, we deform the contour, cut the contour to be finite, take the $\e\to 0$ limit, and then repair the contour to its final form
-- all with error at most $3\kappa$ for $\kappa$ arbitrarily small.
\subsubsection*{Step 2a:} We must define the contour to which we want to deform $\CvEps{a;\alpha;\varphi}$, and then justify that this deformation does not change the value of the Fredholm determinant.
\begin{definition}\label{cpctcontdef}
Fix $\varphi\in (0,\pi/4)$, $r>0$, and real numbers $a=\{a_1,\dots,a_N\}$ and \mbox{$\alpha=\{\alpha_1,\dots,\alpha_M\}$} such that $\alpha_m-a_n>0$ for all $1\leq n\leq N$ and $1\leq m\leq M$.
Define the finite contour $\CvRLeq{a;\alpha;\varphi;<r}$ to be $\{\mu+te^{(\pi+ \varphi)\I}:0\leq t\leq r\}\cup \{\mu+te^{(\pi- \varphi)\I}:0\leq t\leq r\}$ where we have set $\mu = \tfrac{1}{2}\max(a)+\tfrac{1}{2}\min(\alpha)$.
The maximal imaginary part along $\CvRLeq{a;\alpha;\varphi;<r}$ is $r\sin(\varphi)$.
Define the infinite contour $\CvEpsR{a;\alpha;\varphi;r}$ (oriented with increasing imaginary part) to be the union of $\CvRLeq{a;\alpha;\varphi;<r}$ with $\CvEpsRGeq{a;\alpha;\varphi;>r}$ and $\CvEpsReq{a;\alpha;\varphi;=r}$.
Here, the contour $\CvEpsRGeq{a;\alpha;\varphi;>r}$ is the portion of the contour $\CvEps{a;\alpha;\varphi}$ which has imaginary part exceeding $r\sin(\varphi)$ in absolute value;
and the contour $\CvEpsReq{a;\alpha;\varphi;=r}$ is composed of the two horizontal line segments which join $\CvRLeq{a;\alpha;\varphi;<r}$ with $\CvEpsRGeq{a;\alpha;\varphi;>r}$.
These contours are illustrated in Figure~\ref{whitasymcontours}.
\end{definition}
\begin{figure}
\begin{center}
\psfrag{alpha}[cb]{$\mu$}
\psfrag{C}[lb]{$\Cv{a;\alpha;\varphi}$}
\psfrag{Ce}[lb]{$\!\!\CvEps{a;\alpha;\varphi}$}
\psfrag{C1}[lb]{$\CvEpsR{a;\alpha;\varphi;>r}$}
\psfrag{C2}[lb]{$\CvEpsR{a;\alpha;\varphi;=r}$}
\psfrag{C3}[lb]{$\Cv{a;\alpha;\varphi;<r}$}
\includegraphics[height=5cm]{PathsWithAsymptCont}
\caption{Left: The infinite contour $\CvEps{a;\alpha;\varphi}$ and the limiting contour $\Cv{a;\alpha;\varphi}$.
Right: The infinite contour $\CvEpsR{a;\alpha;\varphi;r}$ (which we deform from $\CvEps{a;\alpha;\varphi}$).}
\label{whitasymcontours}
\end{center}
\end{figure}
Now we justify replacing the contour $\CvEps{a;\alpha;\varphi}$ by $\CvEpsR{a;\alpha;\varphi;r}$.
\begin{lemma}
For any $r>0$ there exists $\e_0>0$ such that for all $\e<\e_0$,
\begin{equation*}
\det(\Id+K_u^{\e})_{L^2(\CvEps{a;\alpha;\varphi})} = \det(\Id+K_u^{\e})_{L^2(\CvEpsR{a;\alpha;\varphi;r})}.
\end{equation*}
\end{lemma}
\begin{proof}
The two contours differ only by a finite length modification. We can continuously deform between the two contours.
We will employ Lemma~\ref{TWprop1} which says that as long as the kernel is analytic in a neighborhood of the contour as we continuously deform
then the Fredholm determinant remains unchanged throughout the deformation.
The only things which could threaten the analyticity of the kernel are the poles coming from the left-hand side terms of \eqref{pwlimits2}, \eqref{pwlimits3} and \eqref{pwlimits3.5}.
On account of the condition satisfied by the contour $\CsPre{q^v}$ (see Definition~\ref{CwPredef}), it follows that these poles are avoided.
By choosing $\e$ small enough, the two contours we are deforming between can be made as close as desired.
Taking them close enough ensures it is possible then to deform between them while avoiding poles of the kernel in $v$ or $v'$ -- hence proving the lemma.
\end{proof}
\subsubsection*{Step 2b:}
We must now show that we can, with small error, restrict our Fredholm determinant to acting on the finite, fixed contour $\CvRLeq{a;\alpha;\varphi;<r}$.
This requires us choosing $r>r_0$ for $r_0$ large enough, and also choosing $\e<\e_0$ for $\e_0$ small enough.
\begin{proposition}\label{compactifyprop}
Fix $\varphi\in (0,\pi/4)$. For any $\kappa>0$ there exist $r_0>0$ and $\e_0>0$ such that for all $r>r_0$ and $\e<\e_0$
\begin{equation*}
\left|\det(\Id+ K_u^{\e})_{L^2(\CvEpsR{a;\alpha;\varphi;r})} - \det(\Id+ K_u^{\e})_{L^2(\CvRLeq{a;\alpha;\varphi;<r})}\right| \leq \kappa.
\end{equation*}
\end{proposition}
The proof of this proposition is fairly technical and is given in Section~\ref{prop2sec}.
\subsubsection*{Step 2c:}
Having restricted our attention to the finite contour $\CvRLeq{a;\alpha;\varphi;<r}$ which does not change with $\e$, we may now take the limit of Fredholm determinants on the restricted $L^2$ space as $\e\to 0$.
\begin{proposition}\label{finiteSprop}
Fix $\varphi\in(0,\pi/4)$. For any $\kappa>0$ and any $r>0$ there exists $\e_0>0$ such that for all $\e<\e_0$
\begin{equation*}
\left|\det(\Id+ K_u^{\e})_{L^2(\CvRLeq{a;\alpha;\varphi;<r})} - \det(\Id+ K_u)_{L^2(\CvRLeq{a;\alpha;\varphi;<r})}\right|\leq \kappa
\end{equation*}
where $K_u(v,v')$ is given in Theorem~\ref{ThmFormulaSemiDiscrete}.
\end{proposition}
The proof of this proposition is also fairly technical and is given in Section~\ref{prop1sec}.
\subsubsection*{Step 2d:} Finally, we show that post-asymptotics we can return to the simple infinite contour $\Cv{a;\alpha;\varphi}$.
\begin{proposition}\label{postasymlemma}
Fix $\varphi\in(0,\pi/4)$. For any $\kappa>0$ there exists $r_0>0$ such that for all $r>r_0$
\begin{equation*}
\left| \det(\Id+ K_u)_{L^2(\CvRLeq{a;\alpha;\varphi;<r})} - \det(\Id+ K_u)_{L^2(\Cv{a;\alpha;\varphi})}\right|\leq \kappa.
\end{equation*}
\end{proposition}
The proof of this proposition is given in Section~\ref{prop3sec}. It is a fair amount more straightforward than the previous two proofs and hence is given first.
Having completed the four substeps, we may combine Propositions~\ref{compactifyprop},~\ref{finiteSprop} and~\ref{postasymlemma} to show that for any $\kappa>0$,
there exists $\e_0>0$ such that for all $\e<\e_0$,
\begin{equation*}
\left|\det(\Id+\tilde K_{\zeta})_{L^2(\CwPre{\tilde a^{-1};\tilde\alpha,0;\varphi})} - \det(\Id+ K_u)_{L^2(\Cv{a;\alpha;\varphi})} \right| \leq 3\kappa
\end{equation*}
where $\det(\Id+\tilde K_{\zeta})$ is as in the right-hand side of Proposition~\ref{propstep2}.
Since $\kappa$ is arbitrary this shows that
\begin{equation*}
\lim_{\e \to 0} \det(\Id+\tilde K_{\zeta})_{L^2(\CwPre{\tilde a^{-1};\tilde\alpha,0;\varphi})} = \det(\Id+ K_u)_{L^2(\Cv{a;\alpha;\varphi})}.
\end{equation*}
The above result completes the proof of Proposition~\ref{propstep2} modulo proving Propositions~\ref{compactifyprop},~\ref{finiteSprop} and~\ref{postasymlemma}.
\subsection{Proof of Proposition~\ref{postasymlemma}}\label{prop3sec}
By virtue of Lemma~\ref{exponentialdecaycutoff}, it suffices to show that for some $c,C>0$,
\begin{equation}\label{kuineq1}
\big|K_u(v,v')\big|\leq Ce^{-c|v|}
\end{equation}
as $v,v'$ varies along $\Cv{a;\alpha;\varphi}$.
Before proving this let us recall the contours with which we are dealing.
The variable $v$ lies on $\Cv{a;\alpha;\varphi}$ and hence can be written as $v=\mu -\kappa \cos(\varphi) \pm \I \kappa \sin(\varphi)$, for $\kappa\in \R_+$
where the $\pm$ represents the two rays of the contour.
The $s$ variables lie on $\Cs{v}$ which depends on $v$ and has two parts:
The portion (which we denote by $\Cs{v;\sqsubset}$) with real part bounded between $1/2$ and $R$ and imaginary part between $\pm d$ for $d$ sufficiently small,
and the vertical portion (which we denote by $\Cs{v;\vert}$) with real part $R$.
Recall that $R=-\Re(v)+\eta$ where $\eta =\tfrac{1}{4}\max(a)+\tfrac{3}{4}\min(\alpha)$.
Let us denote by $h(s)$ the integrand through which $K_u(v,v')$ is defined.
We split the proof into two steps. \emph{Step 1:} We show that the integral of $h(s)$ over $s\in \Cs{v;\sqsubset}$ is bounded by an expression with exponential decay in $|v|$, uniformly over $v'$.
\emph{Step 2:} We show the integral of $h(s)$ over $s\in \Cs{v;\vert}$ is bounded by an expression with exponential decay in $|v|$, uniformly over $v'$. The combination of these two steps imply the inequality \eqref{kuineq1} and hence complete the proof.
\subsubsection*{Step 1:}
We deal with the various terms in $h(s)$ separately and develop bounds for each.
Write $s=x+\I y$ and note that along the contour $\Cs{v;\sqsubset}$, $y\in [-d,d]$ for $d$ small, and $x\in [1/2,R]$.
Let us start with $e^{v\tau s + \tau s^2/2}$.
The norm of this is bounded by the exponential of the real part of the exponent. For $s$ along $\Cs{v;\sqsubset}$
\begin{equation*}
\Re(v s + s^2/2) = x\Re(v)+\frac{x^2}{2}-y\Im(v)-\frac{y^2}{2} .
\end{equation*}
Given our choice of $R=-\Re(v)+\eta$, by taking $d$ sufficiently small and using the bound $\Re(v)\leq \tilde c'-c' |v|$ for some constants $c',\tilde c'$ (depending on $\varphi$), we may deduce that
\begin{equation*}
\Re(v s + s^2/2) \leq \tilde c -c |v| x
\end{equation*}
for some constants $c,\tilde c>0$. From this it follows that
\begin{equation*}
|e^{v\tau s + \tau s^2/2}| \leq C e^{-c\tau |v|x}.
\end{equation*}
Turning to the other terms in $h(s)$, we have that
\begin{equation*}
|u^s|\leq e^{x\ln |u| - y\arg(u)}
\end{equation*}
and we may also bound
\begin{equation}\label{threebounds}
\left|\frac{\Gamma(v-a_m)}{\Gamma(s+v-a_m)}\right|, \quad \left|\frac{\Gamma(\alpha_m-v-s)}{\Gamma(\alpha_m-v)}\right|,
\quad \left|\frac{1}{v+s-v'}\right|,\quad |\Gamma(-s)\Gamma(1+s)|\quad \leq \textrm{const}
\end{equation}
for some constant $\textrm{const}>0$.
The first two bounds come from the functional equation for the Gamma function, and the last from the fact that $s$ is bounded away from $\Z$. Let us explain in some further detail the first bound (the second follows in a similar manner). Just for this argument, call $\tilde{v} = v+a_m$. It follows that $\tilde{v} = \tilde{\mu} - \kappa\cos(\varphi) \pm \I \kappa \sin(\varphi)$ with $\tilde\mu$ real and strictly positive. We can write $s = t + r$ where $t\in \Z_{\geq 0}$ and $r$ has real part bounded in $[0,1)$ and imaginary part bounded between $\pm d$. The functional equation for the Gamma function then implies that
$$
\frac{\Gamma(\tilde{v})}{\Gamma(s+\tilde{v})} = \frac{1}{s-1+\tilde v}\frac{1}{s-2+\tilde v}\cdots\frac{1}{r+\tilde v} \frac{\Gamma(\tilde v)}{\Gamma(r+\tilde v)}.
$$
As $\tilde v$ varies along its contour, all of the factors $\frac{1}{s-j+\tilde{v}}$, $j\in \Z_{\geq 1}$, are bounded in norm by a constant (uniform as $\tilde{v}$ and $s$ vary along their contours), and, in fact, all but two of those factors are bounded in norm by 1. This implies that the product of these factors is bounded by a constant (uniform as $\tilde{v}$ and $s$ vary along their contours). As for the remaining factor $\frac{\Gamma(\tilde v)}{\Gamma(r+\tilde v)}$, as $r$ varies with real part in $[0,1)$ and imaginary part in $[-d,d]$, and as $\tilde v$ varies along its contour, this ratio remains uniformly bounded by a constant. This implies the first bound in \eqref{threebounds}. The second follows in a similar manner.
Combining the bounds from \eqref{threebounds} together shows that for $|v|$ large, the portion of the integral of $h(s)$ for $s$ in $\Cs{v;\sqsubset}$ is bounded by (recall $s=x+\I y$)
\begin{equation*}
\int_{\Cs{v;\sqsubset}}|\d s| \textrm{const}\cdot e^{-c\tau |v|x + x\ln |u| -\arg(u)y} \leq C e^{-c|v|}
\end{equation*}
for some constants $c,C>0$.
\subsubsection*{Step 2:}
As above, we consider the various terms in $h(s)$ separately and develop bounds for each.
Let us write $s=R+\I y$ and note that $s\in \Cs{v;\vert}$ corresponds to $y$ varying over all $|y|\geq d$.
As in \emph{Step 1}, the most important bound will be that of $e^{v\tau s + \tau s^2/2}$.
Observe that
\begin{equation*}
\Re(v s + s^2/2) = \Re(v)R - \Im(v) y + \frac{R^2}{2} -\frac{y^2}{2} = -\frac{(y+\Im(v))^2}{2} + \frac{\Im(v)^2}{2} +\frac{R^2}{2} +\Re(v)R.
\end{equation*}
Observe that because $\varphi\in (0,\pi/4)$ and $R=-\Re(v)+\eta$,
\begin{equation*}
\frac{\Im(v)^2}{2} +\frac{R^2}{2} +\Re(v)R \leq \tilde c -c |v|^2
\end{equation*}
for some constants $c,\tilde c>0$. Thus
\begin{equation}\label{step2bdd}
\Re(v s + s^2/2) \leq -\frac{(y+\Im(v))^2}{2} +\tilde c-c |v|^2.
\end{equation}
Let us now turn to the other terms in $h(s)$. We bound
\begin{equation*}
|u^s|\leq e^{R\ln |u| - y\arg(u)}.
\end{equation*}
By standard bounds for the large imaginary part behavior, we can show
\begin{equation*}
\left|\frac{\Gamma(v-a_m)}{\Gamma(s+v-a_m)}\right| \leq C e^{\frac{\pi}{2} |y|}, \qquad \left|\frac{\Gamma(\alpha_m-v-s)}{\Gamma(\alpha_m-v)}\right| \leq C e^{-(\frac{\pi}{2}-\e)|y|}\le C
\end{equation*}
for some constant $C>0$ sufficiently large and $\e>0$ small enough. Also, $|1/(v+s-v')|\leq C$ for a fixed constant. Finally, the term
\begin{equation*}
|\Gamma(-s)\Gamma(1+s)|\leq Ce^{-\pi |y|}
\end{equation*}
for some constant $C>0$.
Combining these together shows that the integral of $h(s)$ over $s$ in $\Cs{v;\vert}$ is bounded by a constant times
\begin{equation}\label{eqnest}
\int_{\R} \exp\left(-\tau \frac{(y+\Im(v))^2}{2} -\tau c |v|^2 + R\ln|u| -y\arg(u) -\pi |y| + N\frac{\pi}{2}|y|\right)\d y.
\end{equation}
We can factor out the terms above which do not depend on $y$, giving
\begin{equation*}
\exp\left(-\tau c |v|^2 + R\ln|u| \right) \int_{\R} \exp\left(-\tau \frac{(y+\Im(v))^2}{2} -y \arg(u) + N\frac{\pi}{2}|y|\right)\d y.
\end{equation*}
Notice that the prefactors on $y$ and $|y|$ in the integrand's exponential are fixed constants.
We can therefore use the following bound that for $a$ fixed and $b\in \R$, there exists a constant $C$ such that
\begin{equation}\label{usethefollowingbound}
\int_{\R} e^{-\nu(y+b)^2 + a|y|}\d y \leq C e^{|ab|},\quad \nu>0.
\end{equation}
For $a<0$ this inequality is obvious, so let us assume $a>0$ and consider which $y$ maximizes the exponential in the integrand on the left-hand side of the inequality. Without loss of generality, we may take $b>0$ as well. It is clear that the maximizing $y$ will be negative, so we are looking for the maximum over $y<0$ of $-\nu(y+b)^2 - ay$. This is achieved when $y+b = -\frac{a}{2\nu}$ which means that the maximal argument of the exponential is $\frac{a^2}{4\nu} + ab$. It is easy to see that there is rapid decay away from this maximal value and hence the integral is bounded by a constant time $e^{\frac{a^2}{4\nu}+ab}$. Since $a$ is fixed, this is itself like a constant time $e^{ab}$. The argument for $b<0$ likewise produces a bound by a constant times $e^{-ab}$, hence inequality \eqref{usethefollowingbound} follows.
Using inequality \eqref{usethefollowingbound}, we find that we can upper-bound \eqref{eqnest} by
\begin{equation*}
\exp\left(-\tau c |v|^2 + R\ln|u|+c'|v|\right).
\end{equation*}
For $|v|$ large enough, the Gaussian decay in the above bound dominates, and hence integral of $h(s)$ over $s$ in $\Cs{v;\vert}$ is bounded by
\begin{equation*}
C e^{-c|v|}
\end{equation*}
for some constants $c,C>0$.
\subsection{Proof of Proposition~\ref{finiteSprop}}\label{prop1sec}
Fix $\kappa,r>0$. We are presently considering the Fredholm determinant of the kernels $K_u^{\e}$ and $K_u$ restricted to the fixed finite contour $\CvRLeq{a;\alpha;\varphi;<r}$.
By Lemma~\ref{uniformptconvergence}, we only need to show convergence of the kernel $K_u^{\e}(v,v')$ to $K_u(v,v')$ as $\e\to0$, uniformly over $v,v'\in \CvRLeq{a;\alpha;\varphi;<r}$.
This is achieved via showing that for all $\kappa'>0$ there exists $\e_0>0$ such that for all $\e<\e_0$ and for all $v,v'\in \CvRLeq{a;\alpha;\varphi;<r}$,
\begin{equation}\label{etapeqn}
\left|K_u^{\e}(v,v')-K_u(v,v')\right|\leq \kappa'.
\end{equation}
The kernels $K_u^{\e}$ and $K_u$ are both defined via integrals over $s$.
The contour on which $s$ is integrated can be fixed for ($\e<\e_0$) to equal $\Cs{v}$, which is the $s$ contour used to define $K_u$. The fact that the $s$ contours are the same for $K_u^{\e}$ and $K_u$ is convenient. The proof of \eqref{etapeqn} will follow from three lemmas. The first deals with the uniformity of convergence of the integrand defining $K_u^{\e}$ to the integrand defining $K_u$ for $s$ restricted to any fixed compact set.
Before stating this lemma, let us define some notation.
\begin{definition}\label{CsMdef}
Let $\Cs{v;>L}= \{s\in \Cs{v}: |s|\geq L\}$ be the portion of $\Cs{v}$ of magnitude greater than $L$ and similarly let $\Cs{v;<L}= \{s\in \Cs{v}: |s|<L\}$.
Let us assume $L$ is large enough so that $\Cs{v;>L}$ is the union of two vertical rays with fixed real part \mbox{$R=-\Re(v)+\eta$} (recall $\eta = \tfrac{1}{4}\max(a)+\tfrac{3}{4}\min(\alpha)$).
Assuming this, we will write $s=R+\I y$. Then for $y_L=(L^2- R^2)^{1/2}$, the contour \mbox{$\Cs{v;>L}=\{R+\I y: |y|\geq y_L\}$}.
\end{definition}
\begin{lemma}\label{etappclaim}
For all $\kappa''>0$ and $L>0$ there exists $\e_0>0$ such that for all $\e<\e_0$, for all $v,v'\in \CvRLeq{a;\alpha;\varphi;<r}$, and for all $s\in \Cs{v;<L}$,
\begin{equation}\label{uniflemmaeqn}
\left| h^q(s) - \Gamma(-s)\Gamma(1+s) \prod_{n=1}^{N}\frac{\Gamma(v-a_n)}{\Gamma(s+v-a_n)} \prod_{m=1}^M\frac{\Gamma(\alpha_m-v-s)}{\Gamma(\alpha_m-v)}
\frac{ u^s e^{v\tau s+\tau s^2/2}}{v+s-v'}\right|\leq \kappa''
\end{equation}
where $h^q$ is given in \eqref{446}.
\end{lemma}
\begin{proof}
This is a strengthened version of the pointwise convergence in \eqref{pwlimits1} through \eqref{pwlimits4}.
It follows from the uniform convergence of the $\Gamma_q$ function to the $\Gamma$ function on compact regions away from the poles (cf.\ Appendix~\ref{qSec}, as well as standard Taylor series estimates.
The choice of contours is such that the pole arising from $1/(v+s-v')$ is uniformly avoided in the limiting procedure as well.
\end{proof}
It remains to show that for $L$ large enough, the integrals defining $K_u^{\e}(v,v')$ and $K_u(v,v')$ restricted to $s$ in $\Cs{v;>L}$,
have negligible contribution to the kernel, uniformly over $v,v'$ and $\e$. This must be done separately for each of the kernels and hence requires two lemmas.
\begin{lemma}
For all $\kappa'>0$ there exist $L_0>0$ and $\e_0>0$ such that for all $\e<\e_0$, for all $v,v'\in \CvRLeq{a;\alpha;\varphi;<r}$, and for all $L>L_0$,
\begin{equation*}
\bigg|\int_{\Cs{v;>L}}\d s h^q(s)\bigg|\leq \kappa'.
\end{equation*}
\end{lemma}
\begin{proof}
We will use the notation introduced in Definition~\ref{CsMdef} and assume $L_0$ is large enough so that $\Cs{v;>L}$ is only comprised of two vertical rays.
Let us first consider the behavior of the left-hand side of \eqref{pwlimits4}. The magnitude of this term is bounded by the exponential of
\begin{equation*}
\Re(-\tau \e^{-1} s + \e^{-2} \tau q^{-v} (q^{-s}-1)).
\end{equation*}
This quantity is periodic in $y$ (recall $s=R+\I y$) with a fundamental domain $y\in[-\pi\e^{-1},\pi \e^{-1}]$.
For $\e^{-1}\pi>|y|>y_0$ for some $y_0$ which can be chosen uniformly in $v$ and $\e$, the following inequality holds
\begin{equation*}
\Re(-\tau \e^{-1} s + \e^{-2} \tau q^{-v} (q^{-s}-1)) \leq -\tau y^2/6.
\end{equation*}
This can is proved by careful Taylor series estimation and the inequality that for $x\in [-\pi,\pi]$, $\cos(x)-1\leq -x^2/6$.
This provides Gaussian decay in the fundamental domain of $y$.
Turning to the ratio of $q$-Gamma functions in \eqref{pwlimits3}, observe that away from its poles, the denominator
\begin{equation}\label{festatement}
\left|\frac{1}{\Gamma_q(s+v-a_m)}\right|\leq c e^{c' {\rm dist}(\Im(s),2\pi \e^{-1} \Z)}
\end{equation}
where $c,c'$ are positive constants independent of $\e$ and $v$ (as it varies in its compact contour).
This establishes a periodic bound on this denominator, which grows at most exponentially in the fundamental domain.
The numerator $\Gamma_q(v-a_m)$ in \eqref{pwlimits3} is bounded uniformly by a constant.
This is because the $v$ contour was chosen to avoid the poles of the Gamma function, and the convergence of the $q$-Gamma function to the Gamma function is uniform on compact sets away from those poles.
Similarly,
\[|\Gamma_q(\alpha_m-s-v)|\le c e^{-c''{\rm dist}(\Im(s),2\pi \e^{-1} \Z)}\le c\]
where $c, c''$ are positive constants.
This is from the uniform convergence of the $q$-Gamma function to the Gamma function which implies that $\Gamma_q(\alpha_m-v)$ remains uniformly bounded from below as $v\in\CvRLeq{a;\alpha;\varphi;<r}$ varies.
Finally, the magnitude of \eqref{pwlimits1} corresponds to $|u^s|$ and behaves like $e^{-R\ln{|u|} + y\arg(u)}$.
Thus, we have established the following inequality which is uniform in $v,v'$ and $\e$ as $y$ varies:
\begin{multline}\label{perbdd}
\left|\left(\frac{-\zeta}{(1-q)^N}\right)^s \frac{q^v \ln q}{q^{s+v} - q^{v'}} e^{\tilde \gamma q^v(q^{s}-1)} \prod_{i=1}^{N} \frac{\Gamma_q(v+\ln_q(\tilde a_i^{-1}))}{\Gamma_q(s+v+\ln_q(\tilde a_i^{-1}))}
\prod_{j=1}^M\frac{\Gamma_q(\ln_q(\tilde\alpha_j)-s-v)}{\Gamma_q(\ln_q(\tilde\alpha_j)-v)}\right|\\
\leq \tilde c\, e^{-\big({\rm dist}(\Im(s),2\pi \e^{-1} \Z)\big)^2/6+c'N\big|{\rm dist}(\Im(s),2\pi \e^{-1} \Z)\big|}
\end{multline}
for some constant $\tilde c>0$. Notice that this inequality is periodic with respect to the fundamental domain for \mbox{$y\in[-\pi\e^{-1},\pi \e^{-1}]$}.
The last term to consider is $\Gamma(-s)\Gamma(1+s)=\frac{-\pi}{\sin(\pi s)}$ which is not periodic in $y$ and decays like $e^{-\pi |y|}$ for $y\in \R$.
Since $\Cs{v;>L}$ is only comprised of two vertical rays, we must control the integral of $h^q(s)$ for $s=R+\I y$ and $|y|>y_L$.
By making sure $L$ is large enough, we can use the periodic bound \eqref{perbdd} to show that the integral over $y_L<|y|<\e^{-1} \pi$ is less than $\kappa$ (with the desired uniformity in $v,v'$ and $\e$).
For the integral over $|y|>\e^{-1}\pi$, we can use the above exponential decay of $\Gamma(-s)\Gamma(1+s)$.
On shifts by $2\pi \e^{-1}\Z$ of the fundamental domain, the exponential decay of $\Gamma(-s)\Gamma(1+s)$ can be compared to the boundedness of the other terms
(which is certainly true considering the bounds we established above). The integral of each shift can be bounded by a term in a convergent geometric series.
Taking $\e_0$ small then implies that the sum can be bounded by $\kappa'$ as well.
\end{proof}
\begin{lemma}
For all $\kappa'>0$ there exists $L_0>0$ such that for all $v,v'\in \CvRLeq{a;\alpha;\varphi;<r}$, and for all $L>L_0$,
\begin{equation*}
\left|\int_{\Cs{v;>L}} \d s \Gamma(-s)\Gamma(1+s) \prod_{n=1}^{N}\frac{\Gamma(v-a_n)}{\Gamma(s+v-a_n)} \prod_{m=1}^M\frac{\Gamma(\alpha_m-v-s)}{\Gamma(\alpha_m-v)}
\frac{u^s e^{v\tau s+\tau s^2/2}}{v+s-v'}\right|\leq\kappa'.
\end{equation*}
\end{lemma}
\begin{proof}
The desired decay here comes easily from the behavior of $vs+s^2/2$ as $s$ varies along $\Cs{v;>L}$.
As before, assume that $L_0$ is large enough so that this contour is only comprised of two vertical rays and set $s=R+ \I y$ for $y\in \R$ for $|y|>y_L$.
As in the proof of Proposition~\ref{postasymlemma} given in Section~\ref{prop3sec}, one shows that
\begin{equation*}
|e^{v\tau s+\tau s^2/2}|\leq C e^{-cy^2}
\end{equation*}
uniformly over $v,v'\in \CvRLeq{a;\alpha;\varphi;<R}$, and for all $L>L_0$.
This behavior should be compared to that of the other terms: $|\Gamma(-s)\Gamma(1+s)|\approx e^{-\pi |y|}$; $|u^s|= e^{-R\ln{|u|} + y\arg(u)}$;
$\left|\frac{\Gamma(v-a_n)}{\Gamma(s+v-a_n)}\right|\leq C e^{|y| \pi/2}$; $\left|\frac{\Gamma(\alpha_m-v-s)}{\Gamma(\alpha_m-v)}\right|\leq C e^{|y|(\pi/2-\e)}$; and $|1/(v+s+v')|\leq C$ as well.
Combining these observations we see that the integral decays in $|y|$ at worst like $C e^{-cy^2+c' |y|}$.
Thus, by choosing $L$ large enough so that $y_L\gg 1$, we can be assured that the integral over $|y|>y_L$ is as small as desired, proving the lemma.
\end{proof}
Let us now combine the above three lemmas to finish the proof of the Proposition~\ref{finiteSprop}.
Choose $\kappa'=\kappa/3$ and fix $L_0$ and $\e_0'$ as specified by the second and third of the above lemmas.
Fix some $L>L_0$ and let $\ell$ equal the length of the finite contour $\Cs{v;<L}$. Set $\kappa''=\frac{\kappa'}{3\ell}$ and apply Lemma~\ref{etappclaim}.
This yields an $\e_0$ (which we can assume is less than $\e_0'$) so that \eqref{uniflemmaeqn} holds. This implies that for $\e<\e_0$, and for all $v,v'\in \CvRLeq{\alpha,\varphi;<r}$,
\begin{multline*}
\bigg| \int_{\Cs{v;<L}} h^q(s)\,\d s\\
-\int_{\Cs{v;<L}} \Gamma(-s)\Gamma(1+s) \prod_{n=1}^{N}\frac{\Gamma(v-a_n)}{\Gamma(s+v-a_n)} \prod_{m=1}^M\frac{\Gamma(\alpha_m-v-s)}{\Gamma(\alpha_j-v)}\frac{ u^s e^{v\tau s+\tau s^2/2}}{v+s-v'}\d s \bigg|\leq \kappa'/3.
\end{multline*}
From the triangle inequality and the three factors of $\kappa'/3$ we arrive at the claimed result of \eqref{etapeqn} and thus complete the proof of Proposition~\ref{finiteSprop}.
\subsection{Proof of Proposition~\ref{compactifyprop}}\label{prop2sec}
The proof of this proposition is essentially a finite $\e$ (recall $q=e^{-\e}$) perturbation of the proof of Proposition~\ref{postasymlemma} given in Section~\ref{prop3sec}.
The estimates presently are a little more involved since the functions involved are $q$-deformations of classic functions.
However, by careful Taylor approximation with remainder estimates, all estimates can be carefully shown. By virtue of Lemma~\ref{exponentialdecaycutoff}, it suffices to show that for some $c,C>0$,
\begin{equation}\label{kueineq}
|K_u^{\e}(v,v')|\leq C e^{-c|v|}.
\end{equation}
Before proving this, let us recall from Definition~\ref{cpctcontdef} the contours with which we are dealing.
The variable $v$ lies on $\CvEpsR{a;\alpha;\varphi;r}$ for $\varphi\in (0,\pi/4)$. The $s$ variables lies on $\Cs{v}$ from Definition~\ref{DefCaCsdefBis} which depends on $v$ and can be divided into two parts:
The portion (which we denote by $\Cs{v,\sqsubset}$) with real part bounded between $1/2$ and $R$ and imaginary part between $-d$ and $d$ for $d$ sufficiently small;
and the vertical portion (which we denote by $\Cs{v,\vert}$) with real part $R$ where $R=-\Re(v)+\eta$ and $\eta=\tfrac{1}{4}\max(a)+\tfrac{3}{4}\min(\alpha)$.
Let us recall that the integrand in \eqref{445}, through which $K_u^{\e}(v,v')$ is defined, is denoted by $h^q(s)$. We split the proof into two steps.
\emph{Step 1:}~We show that the integral of $h^q(s)$ over $s\in \Cs{v,\sqsubset}$ is bounded for all $\e<\e_0$ by an expression with exponential decay in $|v|$, uniformly over $v'$.
\emph{Step 2:}~We show that the integral of $h^q(s)$ over $s\in \Cs{v,\vert}$ is bounded for all $\e<\e_0$ by an expression with exponential decay in $|v|$, uniformly over $v'$. The combination of these two steps implies the inequality \eqref{kueineq} and hence completes the proof.
\subsubsection*{Step 1:}
We consider the various terms in $h^q(s)$ separately (in particular we consider the left-hand sides of \eqref{pwlimits1} through \eqref{pwlimits4})
and develop bounds for each which are valid uniformly for $\e<\e_0$ and $\e_0$ small enough.
Let us write $s=x+\I y$ and note that along the contour $\Cs{v,\sqsubset}$, $y\in [-d,d]$ for $d$ small, and $x\in [1/2,R]$.
Let us start with the left-hand side of \eqref{pwlimits4} which can be rewritten as
\begin{equation*}
\exp\left(\tau (-\e^{-1}s + \e^{-2} q^{-v}(q^{-s}-1))\right).
\end{equation*}
The norm of the above expression is bounded by the exponential of the real part of the exponent.
For $\varphi\in (0,\pi/4)$, one shows (as a perturbation of the analogous estimate in \emph{Step 1} of the Proof of Proposition~\ref{postasymlemma}) via Taylor expansion with remainder estimates that
\begin{equation*}
\tau \Re(-\e^{-1}s + \e^{-2} q^{-v}(q^{-s}-1))\leq \tilde c- \tau c|v| x
\end{equation*}
for some constants $c,\tilde c$.
The above bound implies
\begin{equation*}
\left|\exp\left(\tau(-\e^{-1}s + \e^{-2} q^{-v}(q^{-s}-1))\right)\right|\leq C e^{-\tau c|v|x}.
\end{equation*}
Let us now turn to the other terms in $h^q(s)$. We bound the left-hand side of \eqref{pwlimits1} as
\begin{equation*}
\left|e^{\tau s \e^{-1}}\left(\frac{-\zeta}{(1-q)^{M+N}}\right)^s\right| \leq C |u^s| \leq C e^{x\ln|u| - y\arg(u)}.
\end{equation*}
We may also bound the left-hand sides of \eqref{pwlimits2}, \eqref{pwlimits3} and \eqref{pwlimits3.5}, as well as the remaining product of Gamma functions by constants:
\begin{equation*}
\left|\frac{\Gamma_q(v+\ln_q(\tilde a_m^{-1}))}{\Gamma_q(s+v+\ln_q(\tilde a_m^{-1}))} \right|, \quad
\left|\frac{\Gamma_q(\ln_q(\tilde\alpha_m)-s-v)}{\Gamma_q(\ln_q(\tilde\alpha_m)-v)} \right|, \quad
\left|\frac{q^v \ln q}{q^{s+v} - q^{v'}}\right|, \quad
|\Gamma(-s)\Gamma(1+s)|\quad\leq \textrm{const}
\end{equation*}
for some constant $\textrm{const}>0$ (which may be different in each case).
The first two bounds come from the functional equation for the $q$-Gamma function (cf.\ Appendix~\ref{qSec}), and the last from the fact that $s$ is bounded away from $\Z$.
Combining these together shows that for $|v|$ large,
\begin{equation*}
\left|\int_{\Cs{v,\sqsubset}}h^q(s)\,\d s\right| \leq \int_{\Cs{v,\sqsubset}} C e^{-\tau c|v| \Re(s) + x\ln|u|-y \arg(u)} |\d s| \leq C' e^{-c'|v|}
\end{equation*}
for some constants $c',C'>0$, while for bounded $|v|$ the integral is just bounded as well.
\subsubsection*{Step 2:}
As above, we consider the various terms in $h^q(s)$ separately and develop bounds for each.
Let us write $s=R+\I y$ and note that $s\in \Cs{v,\vert}$ corresponds to $y$ varying over all $|y|\geq d$.
Four of the terms we consider (corresponding to the left-hand sides of \eqref{pwlimits2}, \eqref{pwlimits3}, \eqref{pwlimits3.5} and \eqref{pwlimits4})
are periodic functions in $y$ with fundamental domain $y\in [-\pi\e^{-1},\pi\e^{-1}]$.
We will first develop bounds on these four terms in this fundamental domain, and then turn to the non-periodic terms.
We start by controlling the behavior of the left-hand side of \eqref{pwlimits4} as $y$ varies in its fundamental domain.
For each $\varphi<\pi/4$ there exists a sufficiently small (yet positive) constant $c'$ such that as $y$ varies in its fundamental domain
\begin{equation*}
\tau \Re(-\e^{-1}s + \e^{-2} q^{-v}(q^{-s}-1)) \leq c' \tau \Re(vs+s^2/2).
\end{equation*}
On account of this, we can use the bound \eqref{step2bdd} from the proof of Proposition~\ref{postasymlemma}. This implies that
\begin{equation*}
\tau \Re(-\e^{-1}s + \e^{-2} q^{-v}(q^{-s}-1)) \leq c' \tau \left(-\frac{(y+\Im(v))^2}{2} -c |v|^2\right).
\end{equation*}
Let us now turn to the other $y$-periodic terms in $h^q(s)$. By bounds for the large imaginary part behavior of the $q$-Gamma function, we can show
\begin{equation*}
\left|\frac{\Gamma_q(v+\ln_q(\tilde a_m^{-1}))}{\Gamma_q(s+v+\ln_q(\tilde a_m^{-1}))} \right| \leq C e^{c\cdot{\rm dist}(\Im(s+v),2\pi \e^{-1} \Z)}
\end{equation*}
for some constants $c,C>0$.
Note that as opposed to \eqref{festatement} when $|v|$ was bounded, in the above inequality, we write ${\rm dist}(\Im(s+v),2\pi \e^{-1} \Z)$ in the exponential on the right-hand side.
This is because we are presently considering unbounded ranges for $v$.
One has similarly the bound
\begin{equation*}
\left|\frac{\Gamma_q(\ln_q(\tilde\alpha_m)-v-s)}{\Gamma_q(\ln_q(\tilde\alpha_m)-v)} \right| \leq C e^{-c' {\rm dist}(\Im(s+v),2\pi \e^{-1} \Z)}
\end{equation*}
for other positive constants $C$ and $c'$.
Also, we can bound
\begin{equation*}
\left|\frac{q^v \ln q}{q^{s+v} - q^{v'}}\right|\leq C
\end{equation*}
for some constant $C>0$.
The parts of $h^q(s)$ which are not periodic in $y$ can easily be bounded. We bound the left-hand side of \eqref{pwlimits1} as in \emph{Step 1} by
\begin{equation*}
\left|e^{-\tau s \e^{-1}}\left(\frac{-\zeta}{(1-q)^N}\right)^s\right| \leq C |u^s| \leq C e^{x\ln|u| - y\arg(u)}.
\end{equation*}
Finally, the term
\begin{equation*}
|\Gamma(-s)\Gamma(1+s)|\leq Ce^{-\pi |y|}
\end{equation*}
for some constant $C>0$.
We may now combine the estimates above.
The idea is to first prove that the integral on the fundamental domain $y\in [-\pi\e^{-1},\pi\e^{-1}]$ is exponentially small in $|v|$.
Then, by using the decay of the two non-periodic terms above, we can get a similar bound for the integral as $y$ varies over all of $\R$.
For $j\in \Z$, define the $j$ shifted fundamental domain as $D_j=j\e^{-1}2\pi + [-\e^{-1}\pi,\e^{-1}\pi]$. Let
\begin{equation*}
I_j:= \int_{D_j} h^q(R+\I y)\,\d y
\end{equation*}
and observe that combining all of the bounds developed above, we have that
\begin{equation*}
|I_j|\leq C \int_{-\e^{-1}\pi}^{\e^{-1}\pi} F_1(y) F_2(y)\,\d y
\end{equation*}
where
\begin{equation*}
\begin{aligned}
F_1(y) &= \exp\left(c' \tau \left(-\frac{(y+\Im(v))^2}{2} -c |v|^2\right) +c'' {\rm dist}(\Im(s+v),2\pi \e^{-1} \Z) +x\ln|u|\right),\\
F_2(y) &= \exp\left(- (y+j\e^{-1}2\pi)\arg(u) -\pi |y+j\e^{-1}2\pi| \right).
\end{aligned}
\end{equation*}
The term $F_1(y)$ is from the periodic bounds while $F_2(y)$ from the non-periodic terms (hence explaining the $j\e^{-1}2\pi$ shift in $y$).
By assumption on $u$, we have $-\arg(u)-\pi=\delta\leq c$ for some $\delta$. Therefore
$
F_2(y) \leq C e^{-c\e^{-1} |j|}
$
for some constants $c,C>0$. Thus
\begin{equation*}
|I_j|\leq C e^{-c\e^{-1} |j|} \int_{-\e^{-1}\pi}^{\e^{-1}\pi} F_1(y)\,\d y.
\end{equation*}
Just as in the end of \emph{Step 2} in the proof of Proposition~\ref{postasymlemma}, we can estimate the integral
\begin{equation*}
\int_{-\e^{-1}\pi}^{\e^{-1}\pi} F_1(y)\,\d y \leq \hat C e^{-\hat c|v|}
\end{equation*}
for some constants $\hat C,\hat c>0$. This implies
$|I_j|\leq \hat C C e^{-c\e^{-1} |j|} e^{-\hat c|v|}$.
Finally, observe that
\begin{equation*}
\left|\int_{\CsPre{v,\vert}} h^{q}(s)\,\d s\right| \leq \sum_{j\in \Z} |I_j| \leq \hat C C e^{-\hat c|v|} \sum_{j\in \Z}e^{-c\e^{-1} |j|} \leq C' e^{-\hat c|v|}
\end{equation*}
where $C'$ is independent of $\e$ as long as $\e<\e_0$ for some fixed $\e_0$. This is the bound desired to complete this step.
\section{SHE/KPZ equation with two-sided Brownian initial data -- Proof of Theorem~\ref{ThmFormulaContinuous}}\label{SectCDRP}
\subsection{Convergence of the Laplace transforms}
Recall the special semi-discrete directed random polymer considered in Definition~\ref{defnearstat} in which $M=1$ and $a_1=a$, $a_n\equiv 0$ for $n>1$, and $\alpha_1=\alpha>0$. We denoted by $\Zsd(\tau,N)$ the semi-discrete directed random polymer partition function in which the weight $\omega_{-1,1}$ is replaced by zero. We will now observe how, by scaling $\tau, N, a,\alpha$ accordingly, it is possible to show convergence of this partition function to the solution to the SHE with initial data related to the scalings of $a,\alpha$. Towards this end, let $T>0$ and $X\in \R$ represent the limiting time and space variables for the SHE and define the following $N$ dependent scalings
\begin{align}
\kappa&=\sqrt{\frac TN}+\frac XN,\label{defkappa}\\
\tau&=\kappa N=\sqrt{TN}+X\label{deftau}.
\end{align}
\begin{definition}\label{digammadef}
Let $\Psi(z) = \frac{d}{dz}\ln \Gamma(z)$ be the digamma function. For a given
$\theta\in \R_+$, define
\begin{equation*}
\kappa(\theta):=\Psi'(\theta),\quad f(\theta):=\theta\Psi'(\theta)-\Psi(\theta),\quad c(\theta):=(-\Psi''(\theta)/2)^{1/3}.
\end{equation*}
We may alternatively parameterize $\theta\in \R_+$ in terms of $\kappa \in\R_+$ as
\begin{equation*}
\theta_\kappa:=(\Psi')^{-1}(\kappa)\in\R_+,\quad f_\kappa:=\inf_{t>0} (\kappa t - \Psi(t))\equiv f(\theta_\kappa),\quad c_\kappa:=c(\theta_\kappa).
\end{equation*}
As given at the beginning of Section 6 in~\cite{BCF12}, the large $\theta$ asymptotics of $\kappa$ and $f$ are
\begin{align}
\kappa(\theta)&=\frac{1}{\theta}+\frac{1}{2 \theta^2}+\frac{1}{6\theta^3}+\Or(\theta^{-5}),\label{kappaseries}\\
f(\theta)&=1-\ln(\theta)+\frac{1}{\theta}+\frac{1}{4\theta^2}+\Or(\theta^{-4}).\label{fseries}
\end{align}
\end{definition}
\begin{theorem}[\cite{QMR12}]
\label{ThmDiscreteToContinuous}
Fix $T>0$, $X\in\R$ and real numbers $b<\beta$.
With Definition~\ref{digammadef}, let \mbox{$\vartheta=\theta_{\sqrt{T/N}}\simeq\sqrt{N/T}+\frac12$}.
Consider the semi-discrete directed random polymer in Definition~\ref{defnearstat} with partition function $\Zsd(\tau,N)$. Let the $a$ and $\alpha$ parameters of the polymer be defined as
\begin{equation}
a=\vartheta+b,\quad \alpha=\vartheta+\beta.\label{awt}
\end{equation}
Define the scaling factor
\begin{equation*}
C(N,T,X)=\exp\left(N+\frac12(N-1)\ln(T/N)+\frac12\left(\sqrt{TN}+X\right)+X\sqrt{N/T}\right).
\end{equation*}
Then, as $N$ goes to infinity,
\[\frac{\Zsd(\sqrt{TN}+X,N)}{C(N,T,X)}\Rightarrow\mathcal Z_{b,\beta}(T,X).\]
The convergence is in distribution and $\mathcal Z_{b,\beta}(T,X)$ is the solution to the SHE given in the statement of Theorem~\ref{ThmFormulaContinuous}.
\end{theorem}
Instead of $\vartheta$ in Theorem~\ref{ThmDiscreteToContinuous}, we choose our scaling parameter for the analysis and for \eqref{awt} to be
\begin{equation}\label{deftheta}
\theta=\theta_\kappa\simeq \sqrt{\frac NT}-\frac XT+\frac12
\end{equation}
which is the $\theta_\kappa$ given in Definition~\ref{digammadef} that corresponds to $\kappa$ given in \eqref{defkappa}.
We rewrite \eqref{awt} as
\begin{equation*}
a=\theta+b+X/T,\quad \alpha=\theta+\beta+X/T.
\end{equation*}
The scaling factor that appears in Theorem~\ref{ThmKbbeta} below is
\begin{equation}\label{defu}
u=Se^{-N-\frac12(N-1)\ln\frac TN-\frac12\sqrt{TN}-X\sqrt{\frac NT}+\frac T{24}-\frac X2+\frac{X^2}{2T}}.
\end{equation}
By comparing the exponents of $C(N,T,X)$ and $u$ and by Theorem~\ref{ThmDiscreteToContinuous},
\begin{equation}\label{uZconvdistr}
u\Zsd(\tau,N)\to Se^{\frac{X^2}{2T}+\frac T{24}}\mathcal Z_{b,\beta}(T,X)
\end{equation}
in distribution as $N\to\infty$ where $\mathcal Z_{b,\beta}(T,X)$ is the partition function of the continuous directed random polymer (CDRP) with boundary drift $b$ and $\beta$.
The convergence of the Fredholm determinant is the following.
\begin{theorem}\label{ThmKbbeta}
Fix $S$ with positive real part, $T>0$, $b<\beta$ real numbers and assume that $X=0$.
Set $\kappa$ and $\tau$ as in \eqref{defkappa} and \eqref{deftau}, $\sigma$ as in \eqref{defsigma} and $\theta$ as in \eqref{deftheta}.
Define the parameters
\begin{equation}
a=\theta+b,\quad \alpha=\theta+\beta,\label{a2nd}
\end{equation}
and use $u$ given in \eqref{defu}. Then
\begin{equation}\label{eq6.15}
\lim_{N\to\infty} \det(\Id+ K_{u})_{L^2(\Cv{a_+;\alpha;\pi/4})}=\det(\Id-K_{b,\beta})_{L^2(\R_+)}
\end{equation}
where $a_+=\max\{a,0\}$ and $K_{b,\beta}$ is defined in \eqref{defKbbeta}.
\end{theorem}
\begin{remark}
The Fredholm determinant in the left-hand side of \eqref{eq6.15}
is a special case of the one in Theorem~\ref{ThmFormulaSemiDiscrete} where we specialized to $a_1=a$, $a_2=a_3=\ldots=0$, $\alpha_1=\alpha$, $M=1$.
The condition $\varphi\in (0,\pi/4)$ in Theorem~\ref{ThmFormulaSemiDiscrete} is to ensure that the Fredholm series converges.
$\varphi=\pi/4$ is the borderline and depending on where the line crosses the real axis, the series might converge or not.
In Theorem~\ref{ThmKbbeta}, we use $\varphi=\pi/4$ and the crossing at the axis is chosen to be the critical point.
For this case, as one can see from the estimates in the proof (see e.g.\ Lemma~\ref{lemma:wtg}), the Fredholm series converges.
\end{remark}
In order to keep the notation simpler, we prove the theorem above for $X=0$; in the $X\neq0$ case, one can simply substitute $b$ by $b+\frac XT$.
The condition on the parameter $S$ comes from its appearance in the argument of the logarithm and as base of powers with complex exponent.
In order to avoid the different branches, we restrict it to the halfplane with positive real part.
In order to prove Theorem~\ref{ThmFormulaContinuous} and~\ref{ThmFormulaStationary}, we need some bounds on the modified Bessel function which are contained in the following lemma.
\begin{lemma}\label{LemmaModifiedBessel}
For $\nu>0$ and $x\in\R_+$, it holds:
\begin{enumerate}
\item[(a)] $x\mapsto x^\nu \BesselK_{-\nu}(x)$ is positive, continuous and decreasing in $x\in\R_+$,
\item[(b)] $0\leq x^\nu \BesselK_{-\nu}(x)\leq C(\nu)$ with $C(\nu)=2^\nu \Gamma(\nu)\sim \nu^{-1}$ as $\nu\to 0$,
\item[(c)] $0\leq -\frac{\d}{\d x}x^\nu \BesselK_{-\nu}(x) = x^\nu \BesselK_{1-\nu}(x) \leq C|1-\nu| x^\beta$ with $\beta=\max\{1,2\nu-1\}$,
\item[(d)] $\BesselK_{-\nu}(x)\simeq C e^{-x} x^{-1/2}$ as $x\to\infty$ where $C$ is independent of $\nu$.
\end{enumerate}
\end{lemma}
\begin{proof}
By the integral representation 9.6.24 of~\cite{AS84},
\begin{equation*}
\BesselK_{\nu}(x)=\int_{\R_+}\d t \cosh(\nu t) e^{-x \cosh(t)},
\end{equation*}
properties of (a) are trivially verified. To get (b), we bound the $\cosh$ by simple exponential and obtain
\begin{equation*}
x^\nu \BesselK_{-\nu}(x)\leq \int_{\R_+} \d t x^\nu e^{\nu t} e^{-x e^{t}/2}=2^\nu \Gamma(\nu,x/2)
\end{equation*}
where $\Gamma(a,z)$ is the incomplete Gamma function (where the last equality is obtained by the change of variable $\tau=xe^t/2$).
In particular, $\Gamma(\nu,0)=\Gamma(\nu)$. As $x^\nu \BesselK_{-\nu}(x)$ is monotone, (b) is shown.
The bound in (c) is obtained using (b), subdividing the cases $\nu\in (0,1]$ and $\nu>1$ taking into account that $\BesselK_\nu(x)=\BesselK_{-\nu}(x)$.
Finally, the bound (d) is formula 9.7.2 of~\cite{AS84}.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{ThmFormulaContinuous}]
We start with \eqref{Besseltransform}. By Lemma~\ref{LemmaModifiedBessel}, $x^\nu \BesselK_{-\nu}(x)$ is a continuous and bounded function. Then, the convergence in distribution \eqref{uZconvdistr} implies
that the left-hand side of \eqref{Besseltransform} converges to that of \eqref{Bessel=Fredholm}.
The convergence of the right-hand side of \eqref{Besseltransform} to that of \eqref{Bessel=Fredholm} is exactly Theorem~\ref{ThmKbbeta} which is proved below.
\end{proof}
The rest of this section is devoted to the proof of Theorem~\ref{ThmKbbeta}.
\subsection{Formal critical point asymptotics}
By using \eqref{deftheta} and comparing \eqref{defu} with \eqref{fseries}, we have
\[u=\frac S\theta e^{-Nf_\kappa+\Or(N^{-1/2})},\]
that is, we can work with
\begin{equation}\label{usimple}
u=\frac S\theta e^{-Nf_\kappa}
\end{equation}
instead to get the same limit.
To rewrite the kernel $K_u$, first we apply the identity
$\Gamma(-s)\Gamma(1-s)=-\pi/\sin(\pi s)$. Then, we do a change of variable
$\tilde z=s+v$, to get
\begin{equation}\label{kvvztilde}
K_u(v,v')
=\frac{1}{2\pi \I}\int_{v+\Cs{v}}\!\! \d\tilde z \frac\pi{\sin(\pi(\tilde z-v))} \frac{\Gamma(v)^{N-1}}{\Gamma(\tilde z)^{N-1}}
\frac{\exp\left(-\frac12\tau v^2-v\ln u\right)}{\exp\left(-\frac12\tau \tilde z^2-\tilde z\ln u\right)}
\frac1{\tilde z-v'} \frac{\Gamma(v-a)}{\Gamma(\tilde z-a)} \frac{\Gamma(\alpha-\tilde z)}{\Gamma(\alpha-v)}.
\end{equation}
Let
\begin{equation}\label{defG}
G(z)=\ln\Gamma(z)-\kappa\frac{z^2}2+f_\kappa z.
\end{equation}
We are looking for the critical point of $G$, hence we are to solve the equation
\[G'(z)=\Psi(z)-\kappa z+f_\kappa=0.\]
It follows from Definition~\ref{digammadef} that $\theta_\kappa$ is a double critical point, i.e.\ $G'(\theta_\kappa)=G''(\theta_\kappa)=0$ and the Taylor series is
\[G(z)=G(\theta_\kappa)-\frac{(c_\kappa)^3}3\left(z-\theta_\kappa\right)+\Or\left(\left(z-\theta_\kappa\right)^4\right).\]
With the present choice of $\kappa$, we have $c_\kappa=\sigma^{-1}N^{-1/3}$, hence
\begin{equation}\label{Gtaylor}
NG(\theta+\sigma w)=NG(\theta)-\frac13 w^3+\Or\left(\frac{w^4}\theta\right).
\end{equation}
We can rewrite the kernel in \eqref{kvvztilde} using \eqref{usimple} and \eqref{defG} as
\[K_u(v,v') = -\frac{1}{2\pi \I}\int_{v+\Cs{v}}\d\tilde z \frac{\pi S^{\tilde z-v}}{\sin(\pi(\tilde z-v))} \frac{e^{NG(v)-NG(\tilde z)}}{\tilde z-v'}
\frac{\Gamma(v-a)}{\Gamma(\tilde z-a)} \frac{\Gamma(\alpha-\tilde z)}{\Gamma(\alpha-v)} \frac{\Gamma(\tilde z)}{\Gamma(v)}\frac{\theta^v}{\theta^{\tilde z}}.\]
We do the change of variables $v=\theta+\sigma w$, $v'=\theta+\sigma w'$ and $\tilde z=\theta+\sigma z$ and substitute \eqref{a2nd} to get
\begin{multline*}
K_\theta(w,w') = -\frac{1}{2\pi \I}\int_{\Cv{z}}\d z \frac{\sigma\pi S^{\sigma(z-w)}}{\sin(\sigma\pi(z-w))} \frac{e^{NG(\theta+\sigma w)-NG(\theta+\sigma z)}}{z-w'}\\
\times\frac{\Gamma(\sigma w-b)}{\Gamma(\sigma z-b)} \frac{\Gamma(\beta-\sigma z)}{\Gamma(\beta-\sigma w)}
\frac{\Gamma(\theta+\sigma z)}{\Gamma(\theta+\sigma w)} \frac{\theta^{\sigma w}}{\theta^{\sigma z}}.
\end{multline*}
As $\theta$ goes to infinity with $N$, for the last two factors,
\[\frac{\Gamma(\theta+\sigma z)}{\Gamma(\theta+\sigma w)} \frac{\theta^{\sigma w}}{\theta^{\sigma z}}\to1.\]
Along with the Taylor expansion in \eqref{Gtaylor}, we get that
\[K_\theta(w,w')\to-\wt K_{b,\beta}(w,w')\]
where
\begin{equation}\label{defwtKbbeta}
\wt K_{b,\beta}(w,w')=\frac1{2\pi \I}\int_{\Cv{z}}\d z \frac{\sigma\pi S^{\sigma(z-w)}}{\sin(\sigma\pi(z-w))} \frac{e^{z^3/3-w^3/3}}{z-w'}
\frac{\Gamma(\sigma w-b)}{\Gamma(\sigma z-b)}\frac{\Gamma(\beta-\sigma z)}{\Gamma(\beta-\sigma w)}.
\end{equation}
The Fredholm determinant of this kernel is rewritten in terms of a Fredholm determinant on $L^2(\R_+)$ in Lemma~\ref{lem:PFLemKPZreformuation}.
\subsection{Rigorous asymptotic analysis}
Let us first assume that
\begin{equation}\label{bbetaassumption}
b<-\frac12\quad\mbox{and}\quad\beta>\frac12.
\end{equation}
We will relax this assumption at the end of Section~\ref{ss:conv_bounds}.
We follow the lines of the proof of Theorem~3.3 in~\cite{BCF12}.
We have to determine \mbox{$\lim_{N\to\infty} \det(\Id+K_u)_{L^2(\mathcal{C}_v)}$} where the contour $\mathcal C_v$ is defined below and it is different from the contour given in Theorem~\ref{ThmFormulaContinuous}.
This change of notation only applies for this section, so it will not cause difficulties.
The contour $\mathcal{C}_v$ is chosen as
\[\mathcal{C}_v=\{\theta-1/4+\I r, |r|\leq r^*\}\cup\{\theta e^{\I t}, t^*\leq |t| \leq \pi/2\}\cup\{\theta-|y|+\I y,|y|\geq\theta\}\]
where
\begin{equation}\label{defrt*}
r^*=\sqrt{\frac\theta2-\frac1{16}},\qquad t^*=\arcsin\left(\sqrt{\frac1{2\theta}-\frac1{16\theta^2}}\right).
\end{equation}
The contour $\mathcal{C}_{\tilde z}$ is set as
\begin{equation}\label{eq5.11}
\mathcal{C}_{\tilde z}=\{\theta+p/4+\I \tilde y, \tilde y\in\R\}\cup\bigcup_{k=1}^{\ell} B_{v+k}
\end{equation}
where $B_{z}$ is a small circle around $z$ clockwise oriented and $p\in\{1,2\}$ depending on the value of $v$, see Figure~\ref{PFFigPathsKPZ}.
More precisely, for given $v$, we consider the sequence of points $S=\{\Re(v)+1,\Re(v)+2,\ldots\}$ and we choose $p=p(v)$ and $\ell=\ell(v)$ as follows:
\begin{itemize}
\item[(1)] If the sequence $S$ does not contain points in $[\theta,\theta+1/2]$, then let $\ell\in \N_0$ be such that $\Re(v)+\ell \in [\theta-1,\theta]$ and we set $p=1$.
\item[(2)] If the sequence $S$ contains a point in $[\theta,\theta+3/8]$, then let $\ell\in \N$ such that \mbox{$\Re(v)+\ell \in [\theta,\theta+3/8]$} and set $p=2$.
\item[(3)] If the sequence $S$ contains a point in $[\theta+3/8,\theta+1/2]$, then let $\ell\in \N$ such that $\Re(v)+\ell \in [\theta-5/8,\theta-1/2]$ and set $p=1$.
\end{itemize}
With this choice, the singularity of the sine along the line $\theta+p/4+\I\R$ is not present, since the poles are at a distance at least $1/8$ from it.
Also, the leading contribution of the kernel will come from situation (1) with $\ell=0$ and $p=1$.
This choice of the contours is identical to the one made in the unperturbed case in~\cite{BCF12}.
If condition \eqref{bbetaassumption} holds, these contours can be used, since the extra singularities coming from
$\Gamma(v-a)$ are on the left-hand side of $\Cv{v}$ and the poles coming from $\Gamma(\alpha-\tilde z)$ are on the right-hand side of $\Cv{\tilde z}$.
Otherwise the contours have to be locally modified. This is made precise later.
\begin{figure}
\begin{center}
\psfrag{Phi1}[cb]{$\Phi^{-1}$}
\psfrag{0}[cb]{$0$}
\psfrag{theta}[cb]{$\theta$}
\psfrag{-thetat}[ct]{$-\frac{\theta}{\sigma}$}
\psfrag{+p4}[lt]{$\frac{p}{4\sigma}$}
\psfrag{-14}[ct]{$\frac{1}{4\sigma}$}
\psfrag{theta+p4}[lt]{$\theta+\frac p4$}
\psfrag{theta-14}[rt]{$\theta-\frac14$}
\psfrag{p}[lb]{$p$}
\psfrag{w}[cb]{$w$}
\psfrag{Cw}[lb]{$\mathcal{C}_w$}
\psfrag{v}[cb]{$v$}
\psfrag{Cv}[lb]{$\mathcal{C}_v$}
\psfrag{Cztilde}[lb]{$\mathcal{C}_{\tilde z}$}
\psfrag{Cz}[lb]{$\mathcal{C}_z$}
\psfrag{a}[cb]{$a$}
\psfrag{b}[cb]{$\frac b\sigma$}
\psfrag{alpha}[cb]{$\alpha$}
\psfrag{beta}[cb]{$\frac\beta\sigma$}
\includegraphics[height=7cm]{PathsKPZStat}
\caption{Left: Integration contours $\mathcal{C}_v$ (dashed) and $\mathcal{C}_{\tilde z}$ (the solid line plus circles at $v+1,\ldots,v+\ell$)
where the small black dots are poles either of the sine or of the gamma functions.
Right: Integration contours after the change of variables $\mathcal{C}_w$ (dashed) and $\mathcal{C}_z$ (the solid line plus circles at $w+1,\ldots,w+\ell$), with $p=p(w)\in\{1,2\}$.}
\label{PFFigPathsKPZ}
\end{center}
\end{figure}
With $\sigma$ defined in \eqref{defsigma}, we do the change of variables
\begin{equation*}
\{v,v',\tilde z\}=\{\Phi(w),\Phi(w'),\Phi(z)\}\quad\textrm{with}\quad\Phi(z):=\theta+z\sigma
\end{equation*}
and
\begin{multline}\label{defKtheta}
K_\theta(w,w'):=\sigma K_u(\Phi(w),\Phi(w'))=\\
-\frac{1}{2\pi \I}\int_{\Cv{z}}\d z \frac{\sigma\pi S^{\sigma(z-w)}}{\sin(\sigma\pi(z-w))}
\frac{e^{NG(\theta+\sigma w)-NG(\theta+\sigma z)}}{z-w'}
\times\frac{\Gamma(\sigma w-b)}{\Gamma(\sigma z-b)}\frac{\Gamma(\beta-\sigma z)}{\Gamma(\beta-\sigma w)}
\frac{\Gamma(\theta+\sigma z)}{\Gamma(\theta+\sigma w)} \frac{\theta^{\sigma w}}{\theta^{\sigma z}}.
\end{multline}
After this change of variables, the contours $\mathcal{C}_w=\Phi^{-1}(\mathcal{C}_v)$ and $\mathcal{C}_z=\Phi^{-1}(\mathcal{C}_{\tilde z})$ are given by
\begin{equation}\label{PFeqCwKPZ}
\mathcal{C}_w=\{-1/(4\sigma)+\I r/\sigma, |r|\leq r^*\}\cup\{(e^{\I t}-1)\theta/\sigma, t^*\leq |t|\leq \pi/2\}\cup\{-|y|+\I y,|y|\geq \theta/\sigma\}
\end{equation}
and
\begin{equation*}
\mathcal{C}_z=\{p/(4\sigma)+\I y, y\in\R\}\cup \bigcup_{k=1}^{\ell} B_{w+k/\sigma}
\end{equation*}
with $r^*$ and $t^*$ defined in \eqref{defrt*}, and $B_{z}$ is a small circle around $z$ clockwise oriented. Then we have
\begin{equation*}
\det(\Id+K_u)_{L^2(\mathcal{C}_v)} = \det(\Id+K_\theta)_{L^2(\mathcal{C}_w)}.
\end{equation*}
Thus, we need to prove that
\begin{equation*}
\lim_{N\to\infty}\det(\Id+K_\theta)_{L^2(\mathcal{C}_w)}=\det(\Id-K_{b,\beta})_{L^2(\R_+)}
\end{equation*}
with $K_{b,\beta}$ given in \eqref{defKbbeta}. The convergence of the kernel follows by Proposition~\ref{PFPropPtConvKPZ} and the exponential bound by Proposition~\ref{PFPropBoundKPZ}.
We then obtain
\begin{equation}\label{KthetatoK}
\lim_{N\to\infty} \det(\Id+K_\theta)_{L^2(\mathcal{C}_w)} = \det(\Id-\wt K_{b,\beta})_{L^2(\mathcal{C}_w)}
\end{equation}
with $\wt K_{b,\beta}$ given in \eqref{defwtKbbeta}.
Note that by definition \eqref{PFeqCwKPZ}, the contour $\Cv{w}$ itself depends on $\theta$.
With a slight abuse of notation, we will denote by $\Cv{w}$ also the contour on the right-hand side of \eqref{KthetatoK} that appears in the $\theta\to\infty$ limit,
which is $-\frac14+\I\R$ with a possible local perturbation close to $0$ that will be given later.
Lemma~\ref{lem:PFLemKPZreformuation} shows that the limiting Fredholm determinant is equal to $\det(\Id-K_{b,\beta})_{L^2(\R_+)}$
and thus completes the proof of Theorem~\ref{ThmKbbeta} for the case of \eqref{bbetaassumption}.
\subsection{Pointwise convergence and bounds}\label{ss:conv_bounds}
Proposition~\ref{PFPropPtConvKPZ} and~\ref{PFPropBoundKPZ} in this section are the analogues of Propositions~6.1 and~6.2 in~\cite{BCF12}.
For the sake of completeness, we give the proof of them putting emphasis on the new factors that appear in the kernel.
These are the following gamma ratios in the definition \eqref{defKtheta} of $K_\theta$:
\[\frac{\Gamma(\sigma w-b)}{\Gamma(\sigma z-b)}
\frac{\Gamma(\beta-\sigma z)}{\Gamma(\beta-\sigma w)}
\frac{\Gamma(\theta+\sigma z)}{\Gamma(\theta+\sigma w)} \frac{\theta^{\sigma w}}{\theta^{\sigma z}}.\]
As in~\cite{BCF12}, the scale of the steep descent analysis is $N\theta=\Or(N^{3/2})$.
The main contribution of the Fredholm determinant $\det(\Id+K_\theta)_{L^2(\Cv{w})}$ comes from the regime
when the variables $v,v'$ and $\tilde z$ are in the neighbourhood of $\theta$, i.e.\ $w,w'$ and $z$ are in the neighbourhood of $0$.
The function that gives the leading contribution to the integral in the steep descent analysis is
\[\wt G(z)=\frac{G(\theta+\theta z)}\theta.\]
It has a double critical point at $0$, and for further derivatives, it holds
\begin{equation}\label{wtG'}
\begin{aligned}
\wt G^{(3)}(0)&=-1+\Or(\theta^{-1}),\\
\wt G^{(n)}(0)&=\Or(1),\qquad n\ge4.
\end{aligned}\end{equation}
We will denote the real part of $\wt G$ by
\begin{equation}\label{defwtg}
\wt g(x,y):=\Re(\wt G(x+\I y)).
\end{equation}
The statements of the following lemma are completely taken from Lemma~6.3, 6.4 and~6.5 of~\cite{BCF12}.
\begin{lemma}\label{lemma:wtg}\mbox{}
\begin{enumerate}
\item For any fixed $X\ge0$, the function $Y\mapsto\wt g(X,Y)$ is strictly increasing for $Y>0$ with $\partial_Y\wt g(X,Y)\ge\partial_Y\wt g(0,Y)$.
\item For $X\ge0$,
\[\wt g(X,Y)\ge\wt g(X,0)+Y^4/12+\Or(Y^4/\theta,Y^6).\]
\item The function $t\mapsto\wt g\big(\cos (t)-1,\sin (t)\big)$ is strictly decreasing for $t\in(0,\pi/2]$. For \mbox{$t\in[0,\pi/2]$} and $\theta$ large enough,
\[\wt g\big(\cos(t)-1,\sin(t)\big)- \wt g(0,0)\le - \sin(t)^4/16.\]
\item The function $Y\mapsto\wt g(-Y,Y)$ is strictly decreasing for $Y>0$. For $Y\to\infty$, we have
\[\partial_Y\wt g(-Y,Y)\simeq-\ln Y.\]
\end{enumerate}
\end{lemma}
Notational remark: $\Or(Y^4/\theta,Y^6)$ in Lemma~\ref{lemma:wtg} is the error term coming from Taylor expansion around $Y=0$.
We will also use the following properties of the gamma function.
\begin{lemma}\label{gammalemma}$ $
\begin{enumerate}
\item For any $u,v\in\R$,
\begin{equation}\label{gammalemma1}
\left|\frac{\Gamma(u+\I v)}{\Gamma(u)}\right|^2=\frac{\Gamma(u+\I v)\Gamma(u-\I v)}{\Gamma(u)^2} =\prod_{n=0}^\infty\left(1+\frac{v^2}{(u+n)^2}\right)^{-1}.
\end{equation}
\item For any $u,v,w\in\R$,
\begin{equation}\label{gammalemma2}
\left|\frac{\Gamma(u+\I w)}{\Gamma(v\pm\I w)}\right|\simeq|w|^{u-v}\quad\mbox{as}\quad |w|\to\infty
\end{equation}
where $\simeq$ means that the ratio of the two sides converges to $1$.
\end{enumerate}
\end{lemma}
\begin{proof}
Part (1) is Formula~6.1.25 in~\cite{AS84}.
To get (2), we use Formula~6.1.45 in~\cite{AS84}
\[\lim_{|y|\to\infty}(2\pi)^{-1/2}|\Gamma(x+\I y)|e^{|y|\pi/2}|y|^{1/2-x}=1\]
for $x$ and $y$ real. \eqref{gammalemma2} is a straightforward consequence.
\end{proof}
\begin{proposition}\label{PFPropPtConvKPZ}
Uniformly for $w,w'$ in a bounded set of $\mathcal{C}_w$,
\begin{equation*}
\lim_{N\to\infty} K_\theta(w,w') = - \wt K_{b,\beta}(w,w')
\end{equation*}
where $\wt K_{b,\beta}$ is given by \eqref{defwtKbbeta}.
\end{proposition}
\begin{proof}
The dependence on $N$ of the kernel $K_\theta$ in \eqref{defKtheta} appears in the factors
\begin{equation}\label{expNG}\begin{aligned}
e^{NG(\theta+\sigma w)-NG(\theta+\sigma z)}\frac{\Gamma(\theta+\sigma z)}{\Gamma(\theta+\sigma w)}\frac{\theta^{\sigma w}}{\theta^{\sigma z}}&=\frac{e^{(N-1)G(\theta+\sigma w)-\frac\kappa2(\theta+\sigma w)^2+f_\kappa(\theta+\sigma w)+\sigma w\ln\theta}}{e^{(N-1)G(\theta+\sigma z)
-\frac\kappa2(\theta+\sigma z)^2+f_\kappa(\theta+\sigma z)+\sigma z\ln\theta}}\\
&=\frac{e^{(N-1)\theta \wt G(\frac{w\sigma}\theta)-\frac{\kappa\theta^2}2\left(1+\frac{w\sigma}\theta\right)^2+f_\kappa\theta\left(1+\frac{w\sigma}\theta\right)
+\theta \frac{w\sigma}\theta\ln\theta}}{e^{(N-1)\theta \wt G(\frac{z\sigma}\theta)-\frac{\kappa\theta^2}2\left(1+\frac{z\sigma}\theta\right)^2+f_\kappa\theta\left(1+\frac{z\sigma}\theta\right)
+\theta \frac{z\sigma}\theta\ln\theta}}.
\end{aligned}\end{equation}
One can already see that the scale of the steep descent analysis is $N^{3/2}$.
By \eqref{kappaseries} and \eqref{fseries}, we have $\kappa\theta^2/2=\Or(\theta)$ and $f_\kappa\theta=\Or(\theta)$,
which shows that we have to investigate the real part of $\wt G$ along the contour $\Cv{z}$.
For $N$ large enough and for $w$ in a fixed bounded subset of $\Cv{w}$, $\Re(w\sigma+1)>1/2$ and $\Re((z-w)\sigma)\in (0,1)$ so that we have $\ell=0$ and $p=1$,
i.e.\ in this case $\Cv{z}=\{\frac{1}{4\sigma}+\I y,y\in\R\}$. Taylor expansion around $w=0$ give us
\begin{equation}\label{NGTaylor}
\begin{aligned}
N G(\theta+\sigma w)=N\theta\wt G\left(\frac{w\sigma}\theta\right)&=N\theta\wt G(0)+\frac{N\theta}{6}\wt G^{(3)}(0)\left(\frac{\sigma w}\theta\right)^3+\Or(N\theta w^4/\theta^4)\\
&=N\theta\wt G(0)-\frac{N\theta \sigma^3}{2\theta^3}\frac{w^3}3+\Or(w^4/\theta,N\theta w^3/\theta^4)\\
&=N\theta\wt G(0)-\frac{w^3}{3}+\Or(w^4/\theta)
\end{aligned}
\end{equation}
where we used \eqref{wtG'}.
We divide the integral over $z$ into two parts: (a) $|\Im(z)|>\theta^{1/3}$ and (b) $|\Im(z)|\leq \theta^{1/3}$.
\smallskip
(a) \emph{Contribution of the integration over $|\Im(z)|>\theta^{1/3}$.}
We will show that the integral can be bounded as
\begin{equation}\label{largezintegral}
\int_{|z|>\theta^{1/3}}\d ze^{N\theta\left(\wt g(0,0)-\wt g\left(\frac1{4\sigma\theta},\frac{\Im(z)}\theta\right)\right)+\Or(1/\theta)}
=\Or(\theta)\int_{\theta^{-2/3}}^\infty\d y e^{N\theta\left(\wt g(0,0)-\wt g\left(\frac1{4\sigma\theta},\frac y\theta\right)\right)+\Or(1/\theta)}.
\end{equation}
This can be seen as follows. We have to work with the $z$-dependent part of the left-hand side of \eqref{expNG}. Therefore,
\[e^{NG(\theta)-NG(\theta+\sigma z)}=e^{N\theta\left(\wt G(0)-\wt G\left(\frac{z\sigma}\theta\right)\right)}
=e^{N\theta\left(\wt g(0,0)-\wt g\left(\frac1{4\sigma\theta},\frac{\Im(z)}\theta\right)\right)}\]
by the definition \eqref{defwtg}. Then
\[\frac{\Gamma(\theta+\sigma z)}{\Gamma(\theta+\sigma w)}\frac{\theta^{\sigma w}}{\theta^{\sigma z}}
=\frac{\Gamma(\theta+\sigma z)}{\Gamma(\theta+1/4)}\frac{\theta^{1/4}}{\theta^{\sigma z}}
\frac{\Gamma(\theta+1/4)}{\Gamma(\theta+\sigma w)}\frac{\theta^{\sigma w}}{\theta^{1/4}}.\]
By \eqref{gammalemma1},
\[\left|\frac{\Gamma(\theta+\sigma z)}{\Gamma(\theta+p/4)}\frac{\theta^{p/4}}{\theta^{\sigma z}}\right|\le1,\]
and from \eqref{gammalemma2},
\[\frac{\Gamma(\theta+1/4)}{\Gamma(\theta+\sigma w)}\frac{\theta^{\sigma w}}{\theta^{1/4}}=1+\Or\left(\frac{(1/4-w)^2}\theta\right)\]
as $\theta\to\infty$ which can be controlled by $\Or(1/\theta)$ in the exponent in \eqref{largezintegral}.
The remaining $z$-dependent factor $\Gamma(\beta-\sigma z)/\Gamma(\sigma z-b)$ is only polynomial in $\Im(z)$ along $\Cv{z}$ by \eqref{gammalemma2}.
Hence using the first part of Lemma~\ref{lemma:wtg} about the decay of $y\mapsto\wt g(1/(4\sigma\theta),y)$,
we see that the integral in \eqref{largezintegral} can be bounded by
\begin{equation}\label{largezbound}
e^{N\theta\left(\wt g(0,0)-\wt g\left(\frac1{4\sigma\theta},\theta^{-2/3}\right)\right)+\Or(\theta^{-1})}\leq e^{N\theta\left(\wt g(0,0)-\wt g\left(\frac1{4\sigma\theta},0\right)-\frac{\theta^{-8/3}}{12}+\Or(\theta^{-11/3})\right)+\Or(\theta^{-1})}.
\end{equation}
Exploiting the relation $\wt g(1/(4\sigma\theta),0)=\wt g(0,0)+\Or(\theta^{-3})$, we get the exponential decay
\[\eqref{largezbound}\le\Or(1)\exp(-cN^{1/6})\]
with some $c>0$.
\smallskip
(b) \emph{Contribution of the integration over $|\Im(z)|\leq \theta^{1/3}$.}
As in~\cite{BCF12}, one can see that in the expansion
\[-N\theta\wt G\left(\frac{z\sigma}\theta\right)=-N\theta\wt G(0)+\frac{z^3}3+\Or(z^4/\theta)\]
for $z=1/(4\sigma)+\I y$, the real part
\[\Re\left(\frac{z^3}{3}\right)=-\frac{y^2}{4\sigma^2}+\frac{1}{192\sigma^3}\]
dominates the error term $\Or(z^4/\theta)$ for large $\theta$.
\smallskip
(b.1) $\theta^{1/6}\leq |\Im(z)|\leq \theta^{1/3}$.
From the previous observation,
\begin{multline*}
\left|\frac{1}{2\pi \I}\int_{\theta^{1/6}\le|\Im(z)|\le\theta^{1/3}}\d z
\frac{\sigma\pi S^{\sigma(z-w)}}{\sin(\sigma\pi(z-w))} \frac{e^{z^3/3-w^3/3+\Or(w^4/\theta,z^4/\theta,1/\theta)}}{z-w'}
\frac{\Gamma(\sigma w-b)}{\Gamma(\sigma z-b)}\frac{\Gamma(\beta-\sigma z)}{\Gamma(\beta-\sigma w)}\right|\\
\le\Or\left(e^{-c\theta^{1/3}}\right)=\Or\left(e^{-c'N^{1/6}}\right)
\end{multline*}
for some $c,c'>0$.
\smallskip
(b.2) $|\Im(z)|\le\theta^{1/6}$.
This contribution of the integral is
\begin{equation}\label{verysmallzintegral}
-\frac{1}{2\pi \I}\int_{\frac1{4\sigma}+\I y,|y|\le\theta^{1/6}}\d z \frac{\sigma\pi S^{\sigma(z-w)}}{\sin(\sigma\pi(z-w))}
\frac{e^{z^3/3-w^3/3+\Or(w^4/\theta,z^4/\theta)}}{z-w'}\frac{\Gamma(\sigma w-b)}{\Gamma(\sigma z-b)}\frac{\Gamma(\beta-\sigma z)}{\Gamma(\beta-\sigma w)}.
\end{equation}
For $|y|\leq \theta^{1/6}$, $\Or(z^4/\theta)=\Or(\theta^{-1/3})$.
Using $|e^x-1|\leq |x| e^{|x|}$ for $x=\Or(z^4/\theta)$ and then for $x=\Or(w^4/\theta)$,
we can delete the error term by making an error of order $\Or(\theta^{-1/3})=\Or(N^{-1/6})$. Thus,
\[\eqref{verysmallzintegral}=\Or(N^{-1/6})-\frac{1}{2\pi \I}\int_{\frac{1}{4\sigma}+\I y,|y|\leq\theta^{1/6}}\d z \frac{\sigma \pi S^{\sigma(z-w)}}{\sin(\sigma\pi(z-w))}
\frac{e^{z^3/3-w^3/3}}{z-w'}\frac{\Gamma(\sigma w-b)}{\Gamma(\sigma z-b)}\frac{\Gamma(\beta-\sigma z)}{\Gamma(\beta-\sigma w)}.\]
Finally, extending the last integral to $\frac{1}{4\sigma}+\I\R$, we make an error of order $\Or(e^{-c\theta^{1/3}})$ for some constant $c>0$.
Putting all the above estimates together we obtain that, for $w,w'\in \mathcal{C}_w$ in a bounded set around $0$,
\[K_\theta(w,w')=\Or(N^{-1/6})-\frac{1}{2\pi \I}\int_{\frac{1}{4\sigma}+\I \R}\d z \frac{\sigma \pi S^{\sigma(z-w)}}{\sin(\sigma\pi(z-w)} \frac{e^{z^3/3-w^3/3}}{z-w'}
\frac{\Gamma(\sigma w-b)}{\Gamma(\sigma z-b)}\frac{\Gamma(\beta-\sigma z)}{\Gamma(\beta-\sigma w)}\]
which completes the proof.
\end{proof}
\begin{proposition}\label{PFPropBoundKPZ}
For any $w,w'$ in $\mathcal{C}_w$, uniformly for all $N$ large enough,
\begin{equation*}
|K_\theta(w,w')|\leq C e^{-|\Im(w)|}
\end{equation*}
for some positive constant $C$.
\end{proposition}
\begin{proof}
We follow the lines of the proof of Proposition~6.2 in~\cite{BCF12}. First, we
can rewrite the kernel as
\begin{multline}\label{Ktheta_separated}
K_\theta(w,w')=S^{-\sigma w}e^{NG(\theta+\sigma w)-NG(\theta)}\frac{\Gamma(\sigma w-b)}{\Gamma(\beta-\sigma w)}
\frac{\Gamma(\theta+p/4)}{\Gamma(\theta+\sigma w)} \frac{\theta^{\sigma w}}{\theta^{p/4}}\\
\times\frac{-1}{2\pi \I}\int_{\Cs{v}}\d z \frac{\sigma\pi S^{\sigma z}}{\sin(\sigma\pi(z-w))} \frac{e^{NG(\theta)-NG(\theta+\sigma z)}}{z-w'}
\frac{\Gamma(\sigma z-b)}{\Gamma(\beta-\sigma z)} \frac{\Gamma(\theta+\sigma z)}{\Gamma(\theta+p/4)}\frac{\theta^{p/4}}{\theta^{\sigma z}}
\end{multline}
where we have separated the dependence on $w$ and $z$.
The dependence on $w'$ is marginal because we can choose the integration variable $z$ such that $|z-w'|\geq 1/(4\sigma)$ and
because we will get the bound through evaluating the absolute value of the integrand of \eqref{Ktheta_separated}.
\smallskip
\emph{Case 1: $w\in \{-1/(4\sigma)+\I y, |y|\leq r^*/\sigma\}$ with $r^*$ given in \eqref{defrt*}.}
The integration contour for $z$ is $1/(4\sigma)+\I\R$ ($p=1$) and no extra contributions from poles of the sine are present.
The factor $1/\sin(\sigma\pi(z-w))$ is uniformly bounded from above. By taking $z=\frac1{4\sigma}+\I\frac{Y\theta}\sigma$, we get
\begin{equation}\label{smallwestimate}\begin{aligned}
|K_\theta(w,w')|&\le\Or(1)\left|e^{N\theta\left(\wt G\left(\frac{w\sigma}\theta\right)-\wt G(0)\right)}
\frac{\Gamma(\theta+1/4)}{\Gamma(\theta+\sigma w)}\frac{\theta^{\sigma w}}{\theta^{1/4}}\right|
\int_{\R}\d Y \frac{e^{N\theta \left(\widetilde g(0,0)-\widetilde g(\tilde \e,Y)\right)} \theta}{(1+|\Im(w)|)^{(b+\beta)+\frac12}}\\
&\le\Or(1)\left|e^{N\theta\left(\wt G\left(\frac{w\sigma}\theta\right)-\wt G(0)\right)}
\frac{\Gamma(\theta+1/4)}{\Gamma(\theta+\sigma w)}\frac{\theta^{\sigma w}}{\theta^{1/4}}\right|(1+|\Im(w)|)^{-(b+\beta)-\frac12}
\end{aligned}\end{equation}
where $\tilde\e=1/(4\sigma\theta)$. The integral over $Y$ is finite by Proposition~\ref{PFPropPtConvKPZ}.
The last factor above $(1+|\Im(w)|)^{-(b+\beta)-1/2}$ is due to
\[\left|\frac{\Gamma(\sigma w-b)}{\Gamma(\beta-\sigma w)}\right|\simeq(\sigma\Im(w))^{-(b+\beta)-\frac12}\]
as $|\Im(w)|\to\infty$ which follows from \eqref{gammalemma2}.
In order to avoid the possible divergence of this bound around $\Im(w)=0$, we use $(1+|\Im(w)|)$ instead of $|\Im(w)|$ in \eqref{smallwestimate}.
The factor $(1+|\Im(w)|)^{-(b+\beta)-1/2}$ will be negligible since we prove exponential decay in $|\Im(w)|$.
We rewrite the estimate \eqref{smallwestimate} as in \eqref{expNG}:
\begin{equation}\label{exp(N-1)G}
\left|e^{N\theta\left(\wt G\left(\frac{w\sigma}\theta\right)-\wt G(0)\right)}
\frac{\Gamma(\theta+1/4)}{\Gamma(\theta+\sigma w)}\frac{\theta^{\sigma w}}{\theta^{1/4}}\right|=\frac{e^{(N-1)\theta\Re \wt G(\frac{w\sigma}\theta)-\frac{\kappa\theta^2}2\left(1+\frac{w\sigma}\theta\right)^2+f_\kappa\theta\left(1+\frac{w\sigma}\theta\right)}}
{e^{(N-1)\theta \wt G(0) -\frac{\kappa\theta^2}2+f_\kappa\theta}}
\left|\frac{\Gamma(\theta+1/4)}{\Gamma(\theta)}\theta^{-1/4}\right|.
\end{equation}
Since $|w\sigma/\theta|=\Or(\theta^{-1/2})$, we use the Taylor expansion of \eqref{NGTaylor} to get that
\[(N-1)\theta\wt G\left(\frac{w\sigma}\theta\right)=(N-1)\theta\wt G(0)-\frac{w^3}3+\Or(w^4/\theta).\]
Substituting $w=-1/(4\sigma)+\I y$ and taking real part, we get
\[(N-1)\theta\Re\left(\wt G\left(\frac{w\sigma}\theta\right)-\wt G(0)\right)=-\frac1{4\sigma} y^2+\Or(1)+\Or(y^4/\theta).\]
For $|y|\le r^*/\sigma=\Or(\theta^{1/2})$, the error term $\Or(y^4/\theta)$ is dominated by the $y^2$ term for $\theta$ large enough. Hence we can write
\[(N-1)\theta\Re\left(\wt G\left(\frac{w\sigma}\theta\right)-\wt G(0)\right)\le-\frac1{8\sigma}y^2+\Or(1).\]
For the rest of the terms in the exponent of \eqref{exp(N-1)G} after substituting $w=-1/(4\sigma)+\I y$, we find that
\begin{multline*}
\Re\left(-\frac{\kappa\theta^2}2\left(1+\frac{w\sigma}\theta\right)^2+\frac{\kappa\theta^2}2+f_\kappa\theta\left(1+\frac{w\sigma}\theta\right)-f_\kappa\theta\right)\\
=\frac{\kappa\theta^2}2\left(-\left(1-\frac1{4\theta}\right)^2+1+\frac{y^2}{\theta^2}\right)-\frac{f_\kappa}4=\frac\kappa2y^2+\Or(1)=\Or(y^2/\theta,1).
\end{multline*}
Putting these bounds together yields
\[|K_\theta(w,w')|\le\Or(1) e^{-\frac{1}{8\sigma} |\Im(w)|^2}\leq C e^{-|\Im(w)|}.\]
\smallskip
\emph{Case 2: $w\in\{(e^{\I t}-1)\theta/\sigma, t^*\leq |t|\leq \pi/2\}\cup\{-|y|+\I y,|y|\geq \theta/\sigma\}$}.
We divide the estimation of the bound by separating into the contributions from
(a) integration over $\frac{p}{4\sigma}+\I\R$ with $p\in\{1,2\}$ depending on $w$ (see the definitions after \eqref{eq5.11}) and
(b) integration over the circles $B_{w+k/\sigma}$, $k=1,\ldots,\ell$.
\smallskip
\emph{Case 2(a).} First notice that the estimate \eqref{smallwestimate} of \emph{Case 1} still holds with the minor difference that
$\tilde\e=p/(4\theta)$ where $p\in\{1,2\}$ depending on the value of $w$, so that we only need to estimate the exponent.
For $w\in\{(e^{\I t}-1)\theta/\sigma, t^*\leq |t|\leq \pi/2\}$, the third part of Lemma~\ref{lemma:wtg} shows that $\widetilde g(\cos(t)-1,\sin(t))-\widetilde g(0,0)\leq -\sin(t)^4/16$.
Replacing $\Im(w)=\sin(t) \theta/\sigma$ and using $|\Im(w)|\geq \sqrt{\theta/2-1/16}$ we obtain
\[(N-1)\theta\Re\left(\wt G\left(\frac{w\sigma}\theta\right)-\wt G(0)\right)\le-c_1|\Im(w)|^4/\theta\le-c_2|\Im(w)|^2\]
if $\theta$ is large enough and for $c_1,c_2>0$.
Then we take $w=(e^{\I t}-1)\theta/\sigma$ in \eqref{exp(N-1)G} for the other terms of the exponent to get
\begin{multline*}
\Re\left(\frac\kappa2(-(\theta+\sigma w)^2+\theta^2)+f_\kappa(\theta+\sigma w-\theta)\right)
=\Re\left(\frac\kappa2(-\theta^2e^{2\I t}+\theta^2)+f_\kappa(\theta e^{\I t}-\theta)\right)\\
=\frac{\kappa\theta^2}2(1-\cos(2t))+f_\kappa\theta(\cos t-1)\le\kappa\theta^2\sin^2 t.
\end{multline*}
Since $\kappa\simeq1/\theta$, this term becomes small compared to $|\Im(w)|^2=\theta^2\sin^2 t/\sigma^2$ as $\theta$ gets large.
It follows from Lemma~\ref{lemma:perturbedgamma} below that the ratio $\Gamma(\sigma w-b)/\Gamma(\beta-\sigma w)$ decays along the semicircle
$\{(e^{\I t}-1)\theta/\sigma, t^*\leq |t|\leq \pi/2\}$.
For $w\in\{-|y|+\I y,|y|\geq \theta/\sigma\}$, it follows from the last statement of Lemma~\ref{lemma:wtg} that
$\partial_Y\wt g(-Y,Y)\sim-\ln Y$ meaning that $\wt g(-Y,Y)\simeq-Y\ln Y$ for $Y$ large.
What we show is that the rest in the exponent is of smaller order.
For $w=-y+\I y$, we have
\begin{multline*}
\Re\left(\frac\kappa2(-(\theta+\sigma w)^2+\theta^2)+f_\kappa(\theta+\sigma w-\theta)\right)\\
=\Re\left(\frac\kappa2(-(\theta-\sigma y+\I\sigma y)^2+\theta^2)+f_\kappa(\theta-\sigma y+\I\sigma y-\theta)\right)
=\kappa\theta\sigma y-f_\kappa\sigma y
\end{multline*}
which simplifies in the leading order by \eqref{kappaseries}--\eqref{fseries}, but it is certainly controlled by the decay $-Y\ln Y$ in the exponent.
The factor $\Gamma(\sigma w-b)/\Gamma(\beta-\sigma w)$ decreases as $|\Im(w)|$ increases along $\{-|y|+\I y,|y|\geq \theta/\sigma\}$ by Lemma~\ref{lemma:perturbedgamma}.
This shows that for $\theta$ large enough, we have the bound
$$|K_\theta(w,w')|\le Ce^{-|\Im(w)|}.$$
\smallskip
\emph{Case 2(b).} The contribution of the integration over $B_{w+k/\sigma}$ is
(up to a $\pm$ sign depending on $k$) given by
\[\frac{S^k e^{N G(\Phi(w))-N G(\Phi(w+k/\sigma))}}{w+k/\sigma-w'}
\frac{\Gamma(\beta-\sigma w-k)}{\Gamma(\sigma w-b+k)}
\frac{\Gamma(\sigma w-b)}{\Gamma(\beta-\sigma w)} \frac{\Gamma(\theta+\sigma w+k)}{\Gamma(\theta+\sigma w)}
\frac{\theta^{\sigma w}}{\theta^{\sigma w+k}}.\]
It is shown in the last part of the proof of Proposition~6.2 in~\cite{BCF12} that the first ratio can be bounded by
\begin{equation}\label{firstresidueratio}
\left|\frac{S^k e^{N G(\Phi(w))-N G(\Phi(w+k/\sigma))}}{w+k/\sigma-w'}\right|\le e^{-c|\Im(w)|}
\end{equation}
for an arbitrary $c>0$ if $N$ is large enough uniformly in $k$. For the rest of the factors, we have
\begin{equation}\label{uparrow1}
\frac{\Gamma(\beta-\sigma w-k)}{\Gamma(\sigma w-b+k)}\frac{\Gamma(\sigma w-b)}{\Gamma(\beta-\sigma w)}
=\frac1{(\sigma w-b)_{k\uparrow}(\beta-\sigma w-k)_{k\uparrow}}
\end{equation}
and
\begin{equation}\label{uparrow2}
\frac{\Gamma(\theta+\sigma w+k)}{\Gamma(\theta+\sigma w)} \frac{\theta^{\sigma w}}{\theta^{\sigma w+k}}
=\frac{(\theta+\sigma w)_{k\uparrow}}{\theta^k}
\end{equation}
where $(x)_{k\uparrow}=x(x+1)\dots(x+k-1)$ is the rising factorial.
For a fixed $w$, \eqref{uparrow2} goes to $1$ as $\theta\to\infty$, but the error is not uniform in $\theta$.
Hence we regard \eqref{uparrow2} as a degree $k$ polynomial in $w$.
Since along the contour \mbox{$\{(e^{\I t}-1)\theta/\sigma, t^*\leq |t|\leq \pi/2\}\cup\{-|y|+\I y,|y|\geq \theta/\sigma\}$}, $|\Re(w)|\le|\Im(w)|$,
the absolute value of \eqref{uparrow2} is also at most a degree $k$ polynomial in $|\Im(w)|$.
The leading coefficient is uniformly small for large $\theta$.
On the other hand, the denominator of \eqref{uparrow1} is independent of $\theta$, and the imaginary part of each of the factors of the products is $\sigma\Im(w)$, hence
\[\left|\frac1{(\sigma w-b)_{k\uparrow}(\beta-\sigma w-k)_{k\uparrow}}\right|\le\frac1{\sigma^{2k}|\Im(w)|^{2k}}.\]
This cancels the polynomial coming from \eqref{uparrow2}, and since $k\le\sigma|\Im(w)|$, by choosing $c$ in \eqref{firstresidueratio} large enough,
the product is still exponentially small. The sum of the $k$ residues is also bounded by $e^{-|\Im(w)|}$ as required.
\end{proof}
\begin{lemma}\label{lemma:perturbedgamma}
For the function
\[w\mapsto\left|\frac{\Gamma(\sigma w-b)}{\Gamma(\beta-\sigma w)}\right|,\]
the following holds:
\begin{enumerate}
\item Along the semicircle $w(t)=(e^{\I t}-1)\theta/\sigma$, it decreases for $t^*\le t\le\pi/2$
and increases for $-\pi/2\le t\le-t^*$ if $\theta$ is large enough.
\item Along the halflines $w(y)=-y\pm\I y$, it decreases for $y\ge\theta/\sigma$ if $\theta$ is large enough.
\end{enumerate}
\end{lemma}
\begin{proof}
Let us call
\[f(x,y):=\Re(\ln\Gamma(x+\I y))=\sum_{n=0}^\infty \left(\frac{x}{n+1}-\frac12 \ln\left((x+n)^2+y^2\right)+\ln(n)\Id_{n\geq 1}\right)-\gamma_{\rm E} x\]
where the second equation appears at the beginning of Section 5.2 in~\cite{BCF12}. It follows that
\begin{align*}
\frac{\partial f(x,y)}{\partial x}&=\sum_{n=0}^\infty \left(\frac1{n+1}-\frac{x+n}{(x+n)^2+y^2}\right)-\gamma_{\rm E}\\
\frac{\partial f(x,y)}{\partial y}&=\sum_{n=0}^\infty -\frac{y}{(x+n)^2+y^2}.
\end{align*}
\begin{enumerate}
\item Let $w(t)=(e^{\I t}-1)\theta/\sigma$. It is elementary to see that
\begin{equation}\label{gammaratio'1}\begin{aligned}
&\frac\partial{\partial t}\Re\left(\ln\frac{\Gamma(\sigma w(t)-b)}{\Gamma(\beta-\sigma w(t))}\right)\\
&\qquad=\theta\sin t\bigg(\sum_{n=0}^\infty\bigg(\frac{-\theta-b+n}{(\theta(\cos t-1)-b+n)^2+\theta^2\sin^2 t}\\
&\qquad\qquad+\frac{\theta+\beta+n}{(\theta(1-\cos t)+\beta+n)^2+\theta^2\sin^2 t}-\frac2{n+1}\bigg)+2\gamma_{\rm E}\bigg).
\end{aligned}\end{equation}
If we consider the above sum for $n\ge\theta$, then it is not hard to show by dominated convergence that
\begin{multline*}
\sum_{n=\lfloor\theta\rfloor}^\infty\left(\frac{-\theta-b+n}{(\theta(\cos t-1)-b+n)^2+\theta^2\sin^2 t}
+\frac{\theta+\beta+n}{(\theta(1-\cos t)+\beta+n)^2+\theta^2\sin^2 t}-\frac2{n+1}\right)\\
\to\int_1^\infty\d x\left(\frac{x-1}{(\cos t-1+x)^2+\sin^2 t}+\frac{x+1}{(1-\cos t+x)^2+\sin^2 t}-\frac2{x+1}\right)
\end{multline*}
as $\theta\to\infty$ for a fixed $t\in(0,\pi/2]$. The integrand on the right-hand side is $\Or(x^{-2})$ as $x\to\infty$, hence the integral is finite.
On the other hand, the sum
\[\sum_{n=0}^{\lfloor\theta\rfloor}\left(\frac{-\theta-b+n}{(\theta(\cos t-1)-b+n)^2+\theta^2\sin^2 t}
+\frac{\theta+\beta+n}{(\theta(1-\cos t)+\beta+n)^2+\theta^2\sin^2 t}\right)\]
remains bounded as $\theta\to\infty$, since it converge to the corresponding integral on $[0,1]$.
But $\sum_{n=0}^{\lfloor\theta\rfloor}\frac2{n+1}\simeq2\ln\theta$ which goes to infinity.
This shows that the derivative in \eqref{gammaratio'1} is negative for $\theta$ large enough if $t\in(0,\pi/2]$.
For negative $t$, the argument is identical. The factor $\sin t$ on the right-hand side of \eqref{gammaratio'1} makes the derivative positive.
This is sufficient for the first assertion of the lemma.
\item We set $w(y)=-y+\I y$. A straightforward computation yields
\begin{equation}\label{gammaratio'2}\begin{aligned}
&\frac1\sigma\frac\partial{\partial y}\Re\left(\ln\frac{\Gamma(\sigma w(y)-b)}{\Gamma(\beta-\sigma w(y))}\right)\\
&\qquad=\sum_{n=0}^\infty\left(\frac{-2\sigma y-b+n}{(-\sigma y-b+n)^2+(\sigma y)^2}
+\frac{2\sigma y+\beta+n}{(\sigma y+\beta+n)^2+(\sigma y)^2}-\frac2{n+1}\right)+2\gamma_{\rm E}.
\end{aligned}\end{equation}
As in the previous part of the proof, we have
\begin{multline*}
\sum_{n=\lfloor y\rfloor}^\infty\left(\frac{-2\sigma y-b+n}{(-\sigma y-b+n)^2+(\sigma y)^2}
+\frac{2\sigma y+\beta+n}{(\sigma y+\beta+n)^2+(\sigma y)^2}-\frac2{n+1}\right)\\
\to\int_1^\infty\d x\left(\frac{-2\sigma+x}{(-\sigma+x)^2+\sigma^2}+\frac{2\sigma+x}{(\sigma+x)^2+\sigma^2}-\frac2{x+1}\right)
\end{multline*}
as $y\to\infty$ and the integral is finite, because the integrand is $\Or(x^{-2})$.
Similarly to the first part of this proof, the first two summands on the right-hand side of \eqref{gammaratio'2} summed over $n\in[0,\lfloor y\rfloor]$ are finite,
because the corresponding integral on $[0,1]$ is finite.
Hence the sum $\sum_{n=0}^{\lfloor y\rfloor}\frac2{n+1}$ makes the derivative negative for $y$ large enough.
The statement for the other branch of the contour can be proved in the same way.
\end{enumerate}
\end{proof}
Proposition~\ref{PFPropPtConvKPZ} and \ref{PFPropBoundKPZ} together imply the convergence of the Fredholm determinants
\begin{equation}\label{restrictedKuconv}
\lim_{N\to\infty}\det(\Id+K_u)_{L^2(\Cv{(2a+\alpha)/3,\varphi})}=\det(\Id-\wt K_{b,\beta})_{L^2(\Cv{w})}
\end{equation}
if assumption \eqref{bbetaassumption} holds.
If \eqref{bbetaassumption} does not hold,
then we modify locally the contours $\Cv{v}$ and $\Cv{\tilde z}$ around the critical point $\theta$ such that they cross the real axis strictly between the poles $a$ and $\alpha$.
The contours $\Cv{w}$ and $\Cv{z}$ are similarly modified.
From now on, we focus on the $\theta\to\infty$ limit of these contours, that is,
we explain how to modify the contour $\Cv{w}$ starting from $-\frac1{4\sigma}+\I\R$ and $\Cv{z}$ starting from $\frac 1{4\sigma}+\I\R$ (since $p=1$).
In order to keep the factor $1/\sin(\sigma\pi(z-w))$ bounded in the limiting kernel $\wt K_{b,\beta}$,
the contour $\Cv{z}$ has to be confined between $\e+\Cv{w}$ and $1/\sigma-\e+\Cv{w}$ for an arbitrarily small but fixed $\e>0$ which might depend on $b$ and $\beta$.
If $\beta-b\ge1$, then the distance of the poles at $\frac b\sigma$ and $\frac\beta\sigma$ is enough to let the two contours run parallelly between them.
In this case, for $b>-\frac12$, replace the $|\Im(w)|\le\frac{2b+1}{2\sigma}$ and $|\Im(z)|\le\frac{2b+1}{2\sigma}$ part of $\Cv{w}$ and $\Cv{z}$ by the parallel semicircles
$\{\frac{2b+1}{2\sigma}e^{\I t}-\frac1{4\sigma},-\frac\pi2\le t\le\frac\pi2\}$ and $\{\frac{2b+1}{2\sigma}e^{\I t}+\frac1{4\sigma},-\frac\pi2\le t\le\frac\pi2\}$ respectively.
The case $\beta<\frac12$ is handled symmetrically.
If $\beta-b<1$, then the contours will come closer together between the poles, we chose them such that they intersect the real axis at $(2b+\beta)/(3\sigma)$ and at $(b+2\beta)/(3\sigma)$.
This choice of the modified contours is shown on Figure~\ref{PFFigPathsKPZproof}.
\begin{figure}
\begin{center}
\psfrag{-thetat}[cb]{$-\theta c_\theta N^{1/3}$}
\psfrag{w}[cb]{$w$}
\psfrag{Cw}[lb]{$\mathcal{C}_w$}
\psfrag{Cz}[lb]{$\mathcal{C}_z$}
\psfrag{bbar}[cb]{$\bar b$}
\psfrag{O1}[cb]{$\Or(1)$}
\psfrag{0}[cb]{$0$}
\psfrag{b}[cb]{$\frac b\sigma$}
\psfrag{beta}[cb]{$\frac\beta\sigma$}
\psfrag{-thetat}[ct]{$-\frac{\theta}{\sigma}$}
\psfrag{+p4}[lt]{$\frac{p}{4\sigma}$}
\psfrag{-14}[ct]{$-\frac{1}{4\sigma}$}
\psfrag{c1}[ct]{$\tfrac{2b+\beta}{3\sigma}$}
\psfrag{c2}[ct]{$\tfrac{b+2\beta}{3\sigma}$}
\includegraphics[height=7cm]{PathsKPZproofStat}
\caption{A possible perturbation of the integration contours, compare with Figure~\ref{PFFigPathsKPZ} (right).
The dots are the singularities of $\Gamma(\sigma w-b)$ at $b/\sigma,(b-1)/\sigma,\dots$ and those of $\Gamma(\beta-\sigma z)$ at $\beta/\sigma,(\beta+1)/\sigma,\dots$.}\label{PFFigPathsKPZproof}
\end{center}
\end{figure}
The local modification of the contours has no influence on the bounds for large $z$ and/or for large $w$.
This is because $N G(\theta+\sigma b)-N G(\theta) =\Or(1)$ and the contour for $z$ is the same away from a distance $\Or(1)$ from the origin.
This shows that \eqref{restrictedKuconv} remains valid for any $b<\beta$.
\subsection{Reformulation of the kernel}
The following lemma about the reformulation along with its proof is the analogue of Lemma~8.6 in~\cite{BCF12}.
\begin{lemma}\label{lem:PFLemKPZreformuation}
For the kernels $\wt K_{b,\beta}$ defined in \eqref{defwtKbbeta} and for $K_{b,\beta}$ given in \eqref{defKbbeta}, it holds
\begin{equation}\label{reformulation}
\det(\Id-\wt K_{b,\beta})_{L^2(\Cv{w})}=\det(\Id-K_{b,\beta})_{L^2(\R_+)}.
\end{equation}
\end{lemma}
\begin{proof}
Assume first that \eqref{bbetaassumption} holds.
For this choice of $b$ and $\beta$, if $w'\in\Cv{w}$ and $z\in\Cv{z}$, then $\Re(z-w')>0$ and one can write
\[\frac1{z-w'}=\int_{\R_+}\d\lambda e^{-\lambda(z-w')}.\]
Using this equation, we have
\[\wt K_{b,\beta}(w,w')=\int_{\R_+}\d\lambda A(w,\lambda)B(\lambda,w')\]
where $A:L^2(\Cv{w})\to L^2(\R_+)$ with
\[A(w,\lambda)=\frac1{2\pi \I}\int_{\Cv{z}}\d z \frac{\sigma\pi S^{\sigma(z-w)}}{\sin(\sigma\pi(z-w))} e^{z^3/3-w^3/3-\lambda z}
\frac{\Gamma(\sigma w-b)}{\Gamma(\sigma z-b)}\frac{\Gamma(\beta-\sigma z)}{\Gamma(\beta-\sigma w)}\]
and $B:L^2(\R_+)\to
L^2(\Cv{w})$ with $B(\lambda,w')=e^{\lambda w'}$.
One checks easily that
\[BA(x,y)=\frac1{2\pi\I}\int_{\Cv{w}}\d w B(x,w)A(w,y)=K_{b,\beta}(x,y),\]
and \eqref{reformulation} follows since $\det(\Id-AB)_{L^2(\Cv{w})}=\det(\Id-BA)_{L^2(\R_+)}$.
It remains to relax condition \eqref{bbetaassumption}.
By Lemma~\ref{lem:Fredholm_anal}, both sides of \eqref{reformulation} are analytic functions of the parameters $b$ and $\beta$ for $b<\beta$.
We have proved above that the two analytic functions coincide if \eqref{bbetaassumption} holds,
therefore \eqref{reformulation} follows by analytic continuation for any $b<\beta$.
\end{proof}
\section{SHE/KPZ equation with stationary initial data -- Proof of Theorem~\ref{ThmFormulaStationary}}\label{SectStatSHE}
To prove Theorem~\ref{ThmFormulaStationary} using the formula of Theorem~\ref{ThmFormulaContinuous}, we need the following lower tail estimate of the solution to the SHE proven in~\cite{CH13}.
\begin{lemma}\label{lem:Ftailbound}
Fix $T>0$ and $X\in \R$ and consider $\mathcal Z_{b,\beta}(T,X)$. For any $b\in \R$ and $\delta>0$, there exist constants $c_1,c_2,c_3>0$ such that for all $\beta\in (b-\delta,b+\delta)$, and all $s\geq 1$
$$\PP\left(\mathcal{Z}_{b,\beta}(T,X)< e^{-c_3 s}\right)\le c_1 e^{-c_2 s^{3/2}}.$$
\end{lemma}
\begin{proof}
This follows the lower tail bound of~\cite[Corollary~1.13]{CH13}.
That result, however, is stated in such a way that the constants $c_1,c_2,c_3$ depend on $\beta$.
The lemma we are proving asks for constants which are independent as $\beta$ varies in $(b-\delta,b+\delta)$.
However, the desired uniformity follows via a simple coupling argument.
The stochastic heat equation is attractive, it means that if we couple our initial data
\mbox{$\mathcal{Z}_{b,\beta}(0,X)=\mathbf{1}_{X\leq 0} \big(B^l(X)+\beta X\big) + \mathbf{1}_{X>0} \big(B^r(X)+b X\big)$} to the same Brownian motions $B^l$ and $B^r$
(here $B^l:(-\infty,0]\to \R$ is a Brownian motion without drift pinned at $B^l(0)=0$, and $B^r:[0,\infty)\to \R$ is an independent Brownian motion pinned at $B^r(0)=0$),
then for $\beta>\beta'$ since $\mathcal{Z}_{b,\beta'}(0,X)\geq \mathcal{Z}_{b,\beta}(0,X)$ for all $X\in R$,
it follows that $\mathcal{Z}_{b,\beta'}(T,X)\geq \mathcal{Z}_{b,\beta}(T,X)$ for all $X\in R$ and $T>0$.
This immediately implies that to bound the lower tail of $\mathcal{Z}_{b,\beta}(T,X)$ as $\beta$ varies in $(b-\delta,b+\delta)$,
it suffices to choose $c_1,c_2,c_3>0$ corresponding to $\beta = b+\delta$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{ThmFormulaStationary}]
We need to show the convergence of both sides of \eqref{Bessel=Fredholm} to the $\beta=b$ expression. The convergence of the right-hand side of \eqref{Bessel=Fredholm} to that of \eqref{K0transform} up to a factor of $\sigma$ follows from Theorem~\ref{thm:limK0delta} since $\lim_{\beta\to b}\Gamma(\beta-b)(\beta-b)=1$. Now consider the left-hand side of~\eqref{Bessel=Fredholm}. Denote by $c_0=2\sqrt{Se^{\frac{X^2}{2T}+\frac T{24}}}$, $x=c_0\sqrt{\mathcal Z_{b,\beta}(T,X)}$, and set $\nu=\beta-b$. Then
\begin{equation}\label{eq7.1}
\textrm{l.h.s.~of}~\eqref{Bessel=Fredholm} = \EE(x^\nu \BesselK_{-\nu}(x)) = -\int_{\R_+}\d\xi \xi^\nu \BesselK_{1-\nu}(\xi)\PP(x\leq \xi)
\end{equation}
where we used integration by parts. Let $c_1$ be the constant in Lemma~\ref{lem:Ftailbound} and denote by $\xi_0=c_0 e^{-c_1/2}$. Decomposing the integral into $[0,\xi_0)$ and $[\xi_0,\infty)$ and making the change of variable $\xi(s)=c_0 e^{-c_1 s/2}$, we obtain
\begin{equation*}
\begin{aligned}
\eqref{eq7.1}=&-\int_{\xi_0}^\infty \d\xi \xi^\nu \BesselK_{1-\nu}(\xi)\PP(x\leq \xi) \\
&-\frac{c_0 c_1}{2}\int_{1}^\infty \d s e^{-c_1 s/2}\xi(s)^\nu \BesselK_{1-\nu}(\xi(s))\PP(\mathcal Z_{b,\beta}(T,X)\leq e^{-c_1 s}).
\end{aligned}
\end{equation*}
Using Lemma~\ref{LemmaModifiedBessel}~(c) and (d), the first integral is bounded uniformly in $\nu$ and we can therefore take $\nu\to 0$.
The same lemma implies also that $e^{-c_1 s/2}\xi(s)^\nu \BesselK_{1-\nu}(\xi(s))$ is bounded by $e^{c s}$ for some constant $c$ independent of $\nu$ and
by the bound on the tail of the probability of Lemma~\ref{lem:Ftailbound}, we also have that the second integrand is uniformly bounded by an integrable function.
Thus by dominated convergence, we can take the $\nu\to 0$ limit inside and we obtain
\begin{equation*}
\lim_{\beta\to b}\textrm{l.h.s.~of}~\eqref{Bessel=Fredholm} = -\int_{\R_+}d\xi \BesselK_{1}(\xi)\PP(c_0 \sqrt{\mathcal Z_{b}(T,X)} \leq \xi) = \EE(\BesselK_0(c_0 \sqrt{\mathcal Z_{b}(T,X)}))
\end{equation*}
where in the last step we integrated by parts.
\end{proof}
Later in this section, we will work in $L^2(\R_+)$, so the functions are defined on $\R_+$ and the scalar product of two functions is meant as
\[\langle f,g\rangle=\int_{\R_+}\d xf(x)g(x).\]
To extend the definition \eqref{defqss} for
$u,v\in\left(-\frac14,\frac14\right)$, let us define on $\R_+$ the functions
\begin{equation}\label{defquv}\begin{aligned}
q_{u,v}(x)&=\frac1{2\pi\I}\int_{-\frac1{4\sigma}+\I\R}\d w \frac{\sigma\pi S^{v-\sigma w}}{\sin(\pi(v-\sigma w))} e^{-w^3/3+wx} \frac{\Gamma(\sigma w-u)}{\Gamma(v-\sigma w)}\\
&=\frac1{2\pi\I}\int_{\frac1{4\sigma}+\I\R}\d z \frac{\sigma\pi S^{\sigma z+v}}{\sin(\pi(\sigma z+v))} e^{z^3/3-zx} \frac{\Gamma(-u-\sigma z)}{\Gamma(\sigma z+v)},
\end{aligned}\end{equation}
and recall \eqref{defrs}. Note that $r_s\not\in L^2(\R_+)$ if
$s\le0$.
Further, to extend \eqref{defbarKbb} for
$b,\beta\in\left(-\frac14,\frac14\right)$, we introduce the kernel
\begin{equation}\label{defbarK}
\bar K_{b,\beta}(x,y)=\frac1{(2\pi\I)^2}\int_{-\frac1{4\sigma}+\I\R}\d w\int_{\frac1{4\sigma}+\I\R}\d z \frac{\sigma\pi S^{\sigma(z-w)}}{\sin(\sigma\pi(z-w))}
\frac{e^{z^3/3-zy}}{e^{w^3/3-wx}} \frac{\Gamma(\beta-\sigma z)}{\Gamma(\sigma z-b)} \frac{\Gamma(\sigma w-b)}{\Gamma(\beta-\sigma w)}.
\end{equation}
Note that the only difference between $K_{b,\beta}$ and $\bar K_{b,\beta}$ is the two integration contours.
The one for $K_{b,\beta}$ is shown on the left-hand side of Figure~\ref{fig:KbarKcontours}, for $\bar K_{b,\beta}$, they are vertical lines.
Note also that $\bar K_{b,b}=\bar K_b$ of \eqref{defbarKbb} and $q_{b,b}=q_b$ of \eqref{defqss}.
Finally recall the function $\Xi$ defined in \eqref{defz}.
\begin{figure}
\psfrag{0}[c]{$\frac b\sigma$}
\psfrag{d}[c]{$\frac\beta\sigma$}
\psfrag{Cw}[cc]{$\Cv{w}$}
\psfrag{Cz}[cc]{$\Cv{z}$}
\psfrag{Cw1}[cc]{$\Cv{w}'$}
\psfrag{Cz1}[cc]{$\Cv{z}'$}
\psfrag{Cw2}[cc]{$\Cv{w}''$}
\psfrag{Cz2}[cc]{$\Cv{z}''$}
\includegraphics[height=4cm]{KbarKcontours}
\caption{The integration contours $\Cv{w}$ and $\Cv{z}$ for $K_{b,\beta}$ are on the left.
The other contours are: $\Cv{w}'=-\frac{1}{4\sigma}+\I\R$, $\Cv{z}'=\frac{1}{4\sigma}+\I\R$, $\Cv{w}''$ a small circle around $b/\sigma$ and $\Cv{z}''$ a small circle around $\beta/\sigma$.
By modifying the contours as shown here and applying the residue theorem, one gets \eqref{applyresidue}.
The dots show the poles of the integrands validating the manipulations of the contours.}\label{fig:KbarKcontours}
\end{figure}
\begin{remark}\label{rem:barKproduct}
We prove in Lemma~\ref{LemmaInvertibilityOfK00} that
\begin{equation}\label{detneq0}
\det(\Id-\bar K_{b,b})_{L^2(\R_+)}\neq0.
\end{equation}
Together with Lemma~\ref{lem:barKproperties} below, it shows that the right-hand side of \eqref{defz} is finite as follows.
In the first scalar product, if $b\le0$, then $r_b\not\in L^2(\R_+)$ (if $b\ge0$, then $r_{-b}\not\in L^2(\R_+)$), but by Lemma~\ref{lem:barKproperties}, $\bar K_{b,b}r_{-b}$ decays
exponentially with a faster rate than $r_b$ might blow up.
The second one is obviously finite. For the third one, we can write
\begin{equation}\label{barKproductrewrite}
\big\langle(\Id-\bar K_{b,b})^{-1}q_{b,b},r_b\big\rangle=\langle
q_{b,b},r_b\rangle+\big\langle\bar K_{b,b}(\Id-\bar
K_{b,b})^{-1}q_{b,b},r_b\big\rangle.
\end{equation}
Using Lemma~\ref{lem:barKproperties} again, $q_{b,b}$ decays exponentially with a faster rate than $r_b$ might blow up, hence $\langle q_{b,b},r_b\rangle$ is finite.
On the other hand,
\[\big\langle\bar K_{b,b}(\Id-\bar K_{b,b})^{-1}q_{b,b},r_b\big\rangle=\big\langle(\Id-\bar K_{b,b})^{-1}q_{b,b},\bar K_{-b,-b}r_b\big\rangle,\]
since the adjoint of $\bar K_{b,b}$ in the real $L^2(\R_+)$ is $\bar K_{-b,-b}$ which can also be seen from the representation \eqref{eqA9}.
The function $\bar K_{-b,-b}r_b$ is already in $L^2(\R_+)$, so the second term on the right-hand side of \eqref{barKproductrewrite} is also well-defined.
A similar argument works for the last scalar product in \eqref{defz}.
\end{remark}
Fix $b\in\left(-\frac14,\frac14\right)$. The kernel $K_{b,\beta}$ is defined for all $\beta\in\left(b,\frac14\right)$ by \eqref{defKbbeta}.
The following theorem describes the behaviour of the corresponding Fredholm determinant in the decreasing $\beta\to b$ limit which is that it goes to $0$ linearly in $\beta-b$.
\begin{theorem}\label{thm:limK0delta}
Let $b\in\left(-\frac14,\frac14\right)$ be fixed. For the kernel $K_{b,\beta}$, we have
\begin{equation}\label{limK0delta}
\lim_{\beta\to b}\frac1{\beta-b}\det(\Id-K_{b,\beta})=\frac1\sigma\Xi(S,b,\sigma)
\end{equation}
with $\Xi$ defined in \eqref{defz}. Recall that the notation $q_{b}$ and $\bar{K}_{b}$ from Definition~\ref{longdef} are related to that above via $q_b = q_{b,b}$ and $\bar{K}_b = \bar{K}_{b,b}$.
\end{theorem}
\begin{remark}\label{rem:bcondition}
Theorem~\ref{thm:limK0delta} is proved with the condition $b\in\left(-\frac14,\frac14\right)$ for technical convenience. One could likely extend the proof with minor modifications up to the range $b\in (-1,1)$. Beyond this range the integration contours which appear implicitly on the right-hand side of \eqref{limK0delta} in the definitions of $q_{b}$, $q_{-b}$ and $\bar K_{b}$ have to depend on $b$ so that
the $w$ contours cross the real axis on the right of the singularities of $\Gamma(\sigma w-b)$ at $(b-1)/\sigma,(b-2)/\sigma,\dots$ whereas the $z$ contours cross on the left of the poles $(b+1)/\sigma,(b+2)/\sigma,\dots$ coming from $\Gamma(b-\sigma z)$.
On the other hand, if $b$ is not in $(-1,1)$, then the kernel function $\bar K_{b}(x,y)$ does not decay in $x$ or $y$, hence the right-hand side of \eqref{limK0delta} is not well-defined via \eqref{defz}.
\end{remark}
Before proving Theorem~\ref{thm:limK0delta}, we give the following decay estimates.
\begin{lemma}\label{lem:barKproperties}
For each $\varepsilon>0$ fixed, there is a $C$ which only depends on $\varepsilon$ such that for all $b,\beta\in\left(-\frac14+\varepsilon,\frac14-\varepsilon\right)$, we have the following bounds:
\begin{align}
\left|\bar K_{b,\beta}(x,y)\right|&\le Ce^{-\frac1{4\sigma}(x+y)},\label{K00decay}\\
\left|q_{b,\beta}(x)\right|&\le Ce^{-\frac1{4\sigma}x}.\label{q00decay}
\end{align}
\end{lemma}
\begin{proof}
We can estimate by taking the absolute value of the integrand in
\eqref{defbarK} as follows
\begin{multline*}
\left|\bar K_{b,\beta}(x,y)\right|\\
\le\frac1{(2\pi)^2}\int_{-\frac1{4\sigma}+\I\R}|\d w|\int_{\frac1{4\sigma}+\I\R}|\d z| \left|\frac{\sigma\pi S^{\sigma(z-w)}}{\sin(\sigma\pi(z-w))} \frac{e^{z^3/3}}{e^{w^3/3}}
\frac{\Gamma(\beta-\sigma z)}{\Gamma(\sigma z-b)} \frac{\Gamma(\sigma w-b)}{\Gamma(\beta-\sigma w)}\right|\cdot\left|e^{-zy+wx}\right|.
\end{multline*}
The factor $e^{z^3/3}/e^{w^3/3}$ has Gaussian decay in $|\Im(w)|$ and in $|\Im(z)|$, the other factors are slower, in particular, see \eqref{gammalemma2} for the gamma ratio.
For any value of $w\in-\frac14+\I\R$ and $z\in\frac14+\I\R$,
\[\left|e^{-zy+wx}\right|=e^{-\frac1{4\sigma}(x+y)}.\]
Since the integration paths pass at least $\varepsilon$ far from the singularities of the integrand for any given $\varepsilon$, one can choose a uniform constant $C$ so that \eqref{K00decay} holds.
\eqref{q00decay} can be proved similarly.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:limK0delta}]
The first step is to rewrite the kernel $K_{b,\beta}$, since in the original form given in \eqref{defKbbeta},
the two integration contours intersect the real axis between the pole at $b/\sigma$ and the pole at $\beta/\sigma$, so the two contours would
collide in the $\beta\to b$ limit, see also Figure~\ref{fig:KbarKcontours}.
Hence by using the residue theorem, we cross the pole at $b/\sigma$ with the $w$ integration contour and cross the pole at $\beta/\sigma$
with the $z$ integration contour, both manipulations resulting in a residue term.
If we assume that
\[-\frac14<b<\beta<\frac14,\]
then the new integration contours can be chosen to be $-\frac1{4\sigma}+\I\R$ for $w$ and $\frac1{4\sigma}+\I\R$ for $z$ as shown on Figure~\ref{fig:KbarKcontours}.
That is, with the notation \eqref{defbarK}, we can write
\begin{equation}\label{applyresidue}\begin{aligned}
K_{b,\beta}(x,y)=\bar K_{b,\beta}(x,y)
&+\frac1{2\pi\I}\int_{-\frac1{4\sigma}+\I\R}\d w \frac{\sigma\pi S^{\beta-\sigma w}}{\sin(\pi(\beta-\sigma w))}\frac{e^{\beta^3/(3\sigma^3)-\beta y/\sigma}}{e^{w^3/3-wx}}
\frac{\Gamma(\sigma w-b)}{\Gamma(\beta-\sigma w)} \frac1{\sigma\Gamma(\beta-b)}\\
&+\frac1{2\pi\I}\int_{\frac1{4\sigma}+\I\R}\d z \frac{\sigma\pi S^{\sigma z-b}}{\sin(\pi(\sigma z-b))}\frac{e^{z^3/3-zy}}{e^{b^3/(3\sigma^3)-bx/\sigma}}
\frac{\Gamma(\beta-\sigma z)}{\Gamma(\sigma z-b)} \frac1{\sigma\Gamma(\beta-b)}\\
&+\frac{\sigma\pi S^{\beta-b}}{\sin(\pi(\beta-b))} \frac{e^{\beta^3/(3\sigma^3)-\beta y/\sigma}}{e^{b^3/(3\sigma^3)-bx/\sigma}} \frac1{\sigma^2\Gamma(\beta-b)^2}
\end{aligned}\end{equation}
by the residue theorem.
Using the functions defined in \eqref{defquv} and \eqref{defrs}, we have
\begin{multline}\label{Kresidues}
K_{b,\beta}(x,y)=\bar K_{b,\beta}(x,y)
+q_{b,\beta}(x)r_\beta(y)\frac1{\sigma\Gamma(\beta-b)}+r_{-b}(x)q_{-\beta,-b}(y)\frac1{\sigma\Gamma(\beta-b)}\\
+\frac{\sigma\pi S^{\beta-b}}{\sin(\pi(\beta-b))} r_{-b}(x)r_\beta(y)\frac1{\sigma^2\Gamma(\beta-b)^2}.
\end{multline}
The last equation shows that $K_{b,\beta}$ is a finite rank perturbation of
$\bar K_{b,\beta}$, i.e.\ we could write
\[K_{b,\beta}(x,y)=\bar K_{b,\beta}(x,y)+\sum_{i=1}^3 f_i(x)g_i(y)\]
with appropriate $f_i$ and $g_i$. In this case, for the Fredholm determinants, the following holds
\begin{equation}\label{rank3pert}
\det\bigg(\Id-\bar K_{b,\beta}-\sum_{i=1}^3 f_i\otimes g_i\bigg)_{L^2(\R_+)}
=\det\left(\Id-\bar K_{b,\beta}\right)_{L^2(\R_+)}
\det\left[\delta_{ij}-\left\langle\left(\Id-\bar K_{b,\beta}\right)^{-1}f_i,g_j\right\rangle\right]_{i,j=1}^{3}
\end{equation}
where $\delta_{ij}$ is the Kronecker's delta provided that $\det(\Id-\bar K_{b,\beta})_{L^2(\R_+)}\neq0$.
For $\beta$ close enough to $b$, this follows by continuity from \eqref{detneq0}.
By \eqref{Kresidues}, we define
\begin{equation}\label{deffigi}\begin{aligned}
f_1(x)&=\frac{q_{b,\beta}(x)}{\sigma\Gamma(\beta-b)},\qquad & g_1(y)&=r_\beta(y),\\
f_2(x)&=\frac{r_{-b}(x)}{\sigma\Gamma(\beta-b)},\qquad & g_2(y)&=q_{-\beta,-b}(y),\\
f_3(x)&=\frac{r_{-b}(x)}{\sigma\Gamma(\beta-b)},\qquad & g_3(y)&=\frac{\pi S^{\beta-b}}{\sin(\pi(\beta-b))}\frac{r_\beta(y)}{\Gamma(\beta-b)}.
\end{aligned}\end{equation}
With this choice of $f_i$ and $g_i$, the Fredholm determinant of $K_{b,\beta}$ is equal to \eqref{rank3pert}.
Since we are to take the limit of $(\beta-b)^{-1}\det(\Id-K_{b,\beta})_{L^2(\R_+)}$ as $\beta\to b$,
it is enough to consider the Taylor series up to first order in the second determinant on the right-hand side of \eqref{rank3pert}.
With the choice \eqref{deffigi}, this second determinant is equal to
\begin{equation}\label{detseries}
\det\left[\begin{matrix}
1-\frac{\beta-b}\sigma\langle Rq_{b,\beta},r_\beta\rangle
& -\frac{\beta-b}\sigma\langle Rq_{b,\beta},q_{-\beta,-b}\rangle
& -\frac{\beta-b}\sigma\langle Rq_{b,\beta},r_\beta\rangle\\
-1+\Or(\beta-b)
& 1-\frac{\beta-b}\sigma\langle Rr_{-b},q_{-\beta,-b}\rangle
& -1-\frac{\beta-b}\sigma(\xi+\langle R\bar K_{b,\beta}r_{-b},r_\beta\rangle\\
-1+\Or(\beta-b)
& -\frac{\beta-b}\sigma\langle Rr_{-b},q_{-\beta,-b}\rangle
& -\frac{\beta-b}\sigma(\xi+\langle R\bar K_{b,\beta}r_{-b},r_\beta\rangle)
\end{matrix}\right]
\end{equation}
where we neglect all the $\Or((\beta-b)^2)$ and higher order terms and we write $R=(\Id-\bar K_{b,\beta})^{-1}$ for simplicity in the above formula.
The value
\[\xi=b^2/\sigma^2+\sigma(2\gamma_{\rm E}+\ln S).\]
To obtain the first column of \eqref{detseries}, we use that
\begin{align*}
\big\langle(\Id-\bar K_{b,\beta})^{-1}f_2,g_1\big\rangle
&=\langle f_2,g_1\rangle+\big\langle K_{b,\beta}(\Id-\bar K_{b,\beta})^{-1}f_2,g_1\big\rangle\\
&=\frac{e^{\beta^3/(3\sigma^3)-b^3/(3\sigma^3)}}{(\beta-b)\Gamma(\beta-b)}
+\frac1{\sigma\Gamma(\beta-b)}\big\langle K_{b,\beta}(\Id-\bar K_{b,\beta})^{-1}r_{-b},r_\beta\big\rangle\\
&=1+\Or(\beta-b).
\end{align*}
The scalar product $\langle(\Id-\bar K_{b,\beta})^{-1}f_3,g_1\rangle$ is the same.
To get the last two entries in the third column of \eqref{detseries}, we do the separation
\[\big\langle(\Id-\bar K_{b,\beta})^{-1}f_2,g_3\big\rangle=\langle f_2,g_3\rangle+\big\langle\bar K_{b,\beta}(\Id-\bar K_{b,\beta})^{-1}f_2,g_3\big\rangle\]
where the first scalar product on right-hand side is of order $1$ whereas the rest is $\Or(\beta-b)$:
\begin{align*}
\langle f_2,g_3\rangle&=\frac{\sigma\pi S^{\beta-b}}{\sin(\pi(\beta-b))}\frac{e^{\beta^3/(3\sigma^3)-b^3/(3\sigma^3)}}{\Gamma(\beta-b)^2}\int_0^\infty\d x e^{-(\beta-b)x/\sigma}\\
&=1+(b^2/\sigma^2+2\sigma\gamma_{\rm E}+\sigma\ln S)\frac{\beta-b}\sigma+\Or((\beta-b)^2).
\end{align*}
This argument works again for $f_3$ instead of $f_2$.
The other terms in the determinant \eqref{detseries} are computed easily by \eqref{rank3pert} and \eqref{deffigi}.
Furthermore, all of these terms are finite which can be seen by using the idea of Remark~\ref{rem:barKproduct}.
By expanding the determinant in \eqref{detseries} and considering the terms up to first order, one can see that
\begin{equation}\label{detK0delta}\begin{aligned}
&\frac1{\beta-b}\det(\Id-K_{b,\beta})_{L^2(\R_+)}\\
&\quad=-\frac1\sigma\det(\Id-\bar K_{b,\beta})_{L^2(\R_+)}\big[b^2/\sigma^2+\sigma(2\gamma_{\rm E}+\ln S)
+\big\langle\bar K_{b,\beta}(\Id-\bar K_{b,\beta})^{-1}r_{-b},r_\beta\big\rangle\\
&\qquad+\big\langle(\Id-\bar K_{b,\beta})^{-1}q_{b,\beta},q_{-\beta,-b}\big\rangle
+\big\langle(\Id-\bar K_{b,\beta})^{-1}q_{b,\beta},r_\beta\big\rangle
+\big\langle(\Id-\bar K_{b,\beta})^{-1}r_{-b},q_{-\beta,-b}\big\rangle\big]\\
&\qquad+\Or(\beta-b).
\end{aligned}\end{equation}
What remains to show is that the right-hand side of
\eqref{detK0delta} converges to $\Xi(S,b,\sigma)/\sigma$. To see
that
\begin{equation}\label{convFred}
\det(\Id-\bar K_{b,\beta})_{L^2(\R_+)}\to\det(\Id-\bar K_{b,b})_{L^2(\R_+)},
\end{equation}
we use the Fredholm series expansion
\begin{equation}\label{KbarFredholm}
\det(\Id-\bar K_{b,\beta})_{L^2(\R_+)}=\sum_{n=0}^\infty\frac{(-1)^n}{n!}\int_{\R_+}\d x_1\dots\int_{\R_+}\d x_n\det\left[\bar K_{b,\beta}(x_i,x_j)\right]_{i,j=1}^{n}.
\end{equation}
The $n\times n$ determinant on the right-hand side of \eqref{KbarFredholm} is bounded by
\begin{equation}\label{boundntimesn}
\left|\det\left[\bar K_{b,\beta}(x_i,x_j)\right]_{i,j=1}^{n}\right|\le C^n e^{-\frac1{2\sigma}(x_1+\dots+x_k)} n^{n/2}
\end{equation}
using Lemma~\ref{lem:barKproperties} and the Hadamard bound on determinants with bounded entries.
Hence the integrand in the $n$th term on the right-hand side of \eqref{KbarFredholm} can be dominated uniformly as $\beta$ varies in $\left(-\frac14+\varepsilon,\frac14-\varepsilon\right)$.
In particular, by dominated convergence, as $\beta\to b$, the $n$th term of the expansion in \eqref{KbarFredholm} converges to the corresponding term of the expansion of $\det(\Id-\bar K_{b,b})$.
Further, by integrating the bound \eqref{boundntimesn}, the absolute value of the $n$th term of the series \eqref{KbarFredholm} is at most $(2\sigma C)^n n^{n/2}/n!$.
Since it is summable, a repeated application of the dominated convergence yields the convergence of Fredholm determinants \eqref{convFred}.
The last step is to show the convergence of the scalar products in \eqref{detK0delta}.
For this end, we first show that the resolvents converge in operator norm in $L^2(\R_+)$.
\begin{lemma}\label{lemma:resolventconv}
If $\lim_{\beta\to b}\|\bar K_{b,\beta}-\bar K_{b,b}\|=0$ and $\|(\Id-\bar K_{b,b})^{-1}\|<\infty$, then
\[\lim_{\beta\to b}\left\|(\Id-\bar K_{b,\beta})^{-1}-(\Id-\bar K_{b,b})^{-1}\right\|=0.\]
\end{lemma}
\begin{proof}
\begin{align*}
\left\|(\Id-\bar K_{b,\beta})^{-1}-(\Id-\bar K_{b,b})^{-1}\right\|
&=\left\|\left[\left(\Id-(\Id-\bar K_{b,b})^{-1}(\bar K_{b,\beta}-\bar K_{b,b})\right)^{-1}-1\right](\Id-\bar K_{b,b})^{-1}\right\|\\
&\le\left\|(\Id-\bar K_{b,b})^{-1}\right\|\sum_{n\ge1}\left\|(\Id-\bar K_{b,b})^{-1}(\bar K_{b,\beta}-\bar K_{b,b})\right\|^n
\end{align*}
which goes to $0$ as $\beta\to b$.
\end{proof}
To finish the proof of Theorem~\ref{thm:limK0delta}, we first check the conditions of Lemma~\ref{lemma:resolventconv}
and then we show that the right-hand side of \eqref{detK0delta} goes to $\Xi(S,b,\sigma)/\sigma$.
To verify the convergence condition of Lemma~\ref{lemma:resolventconv}, one can write $\bar K_{b,\beta}-\bar K_{b,b}$ as a common double integral with a difference of gamma ratios.
This difference goes to $0$ pointwise as $\beta\to b$, hence the Hilbert--Schmidt norm of $\bar K_{b,\beta}-\bar K_{b,b}$ goes to $0$ by dominated convergence as $\beta\to b$.
The finite norm condition for the resolvent is a direct consequence of \eqref{detneq0}.
Since
\[q_{b,\beta}\to q_{b,b}\qquad\mbox{and}\qquad q_{-\beta,-b}\to q_{-b,-b}\]
as $\beta\to b$ in $L^2(\R_+)$ by dominated convergence, we have
\begin{equation}\label{easiestscalarprod}
\big\langle(\Id-\bar K_{b,\beta})^{-1}q_{b,\beta},q_{-\beta,-b}\big\rangle\to\big\langle(\Id-\bar K_{b,b})^{-1}q_{b,b},q_{-b,-b}\big\rangle.
\end{equation}
In the other scalar products, the functions $r_b$ and $r_{-b}$ appear in the limit which may not be in $L^2(\R_+)$, therefore, as in Remark~\ref{rem:barKproduct}, we use again the identity
\[(\Id-\bar K_{b,b})^{-1}=\Id+(\Id-\bar K_{b,b})^{-1}\bar K_{b,b}\]
and note that $\bar K_{b,b}r_b$ and $\bar K_{b,b}r_{-b}$ are already in $L^2(\R_+)$, so the rest of the argument is the same as for \eqref{easiestscalarprod} and the use of dominated convergence.
This completes the proof of Theorem~\ref{thm:limK0delta}.
\end{proof}
\section{SHE/KPZ equation universality -- Proof of Theorem~\ref{CorUniversality}}\label{SectUniversality}
In this section, we prove Theorem~\ref{CorUniversality}. Let
\begin{equation}\label{deff}
f(t)=\BesselK_0(2e^t)
\end{equation}
where $\BesselK_0$ is the modified Bessel function and set
\begin{equation}\label{Sscaling}
S=e^{-\frac{\tau^2+r}\sigma}.
\end{equation}
Then using the scaling \eqref{XTscaling} and the notation \eqref{defF}, we have
\begin{equation}\label{freeenergytransform}\begin{aligned}
&\frac{\partial}{\partial r}\EE\left[2\sigma \BesselK_0\left(2\sqrt{Se^{\frac{X^2}{2T}+\frac T{24}}\mathcal Z_b(T,X)}\right)\right]\\
&\qquad=\frac{\partial}{\partial r}\EE\left[2\sigma f\left(\frac1{2\sigma}\left(\frac{\mathcal H_b(T,X)+\frac T{24}(1+12b^2)-2^{1/3}b\tau T^{2/3}}{(T/2)^{1/3}}-r\right)\right)\right]\\
&\qquad=\EE\left[2\sigma\frac{\partial}{\partial r}f\left(\frac1{2\sigma}\left(\frac{\mathcal H_b(T,X)+\frac T{24}(1+12b^2)-2^{1/3}b\tau T^{2/3}}{(T/2)^{1/3}}-r\right)\right)\right].
\end{aligned}\end{equation}
In the last step, we used the following property of the function $f$ given in \eqref{deff} along with Lemma~\ref{lem:Ftailbound}.
\begin{lemma}
If $H$ is a random variable with exponential negative tail, that is, if $\EE[\exp(-uH)]$ is finite for some $u>0$, then
\begin{equation}\label{derivativelemma}
\frac\partial{\partial r}\EE f(-r+H)=\EE\left[\frac\partial{\partial r}f(-r+H)\right]
\end{equation}
for the function $f$ given in \eqref{deff}.
\end{lemma}
\begin{proof}
The left-hand side of \eqref{derivativelemma} is written as
\[\lim_{h\to0}\EE\left[\frac1h(f(-(r+h)+H)-f(-r+H))\right].\]
Since $f'\in[0,1]$ and $f'(t)$ has a double exponential decay for large $t$ and it is bounded,
the dominated convergence applies and yields the equality with the right-hand side of \eqref{derivativelemma}.
\end{proof}
To conclude the proof of Theorem~\ref{CorUniversality}, we use the rotational invariance formula \eqref{shiftarg} with $b\to b+X/T$ and with the parameter setting \eqref{XTscaling}:
\begin{equation*}
Se^{\frac{X^2}{2T}+\frac T{24}}\mathcal Z_b(T,X)\stackrel{\d}=Se^{\frac T{24}}\mathcal Z_{\tau\sigma}(T,0).
\end{equation*}
Substituting this on the left-hand side of \eqref{freeenergytransform}, we see that
\[\eqref{freeenergytransform}=\frac{\partial}{\partial r}\EE\left[2\sigma \BesselK_0\left(2\sqrt{Se^{\frac T{24}}\mathcal Z_{\tau\sigma}(T,0)}\right)\right]
=\frac{\partial}{\partial r}\Xi\left(S=e^{-\frac{\tau^2+r}\sigma},\tau\sigma,\sigma\right)\]
where we applied Theorem~\ref{ThmFormulaStationary} in the last step with $S=e^{-(\tau^2+r)/\sigma}$ and $b=\tau\sigma$ (which is in $\left(-\frac14,\frac14\right)$ for $T$ large enough).
Then using Lemma~\ref{lem:TinflimitK} below and the definition \eqref{defBFP}, we see that the right-hand side of \eqref{freeenergytransform} converges to $F_\tau(r)$ for each $r\in\R$.
The functions $\{f_T\}_{T>0}$ with $f_T(x)=-f'(x/(2\sigma))$ are strictly decreasing in $x$ with a limit of $1$ at $x=-\infty$ and $0$ at $x=\infty$,
and for each $\delta>0$, on $\R\setminus[-\delta,\delta]$, $f_T$ converges uniformly to $\Id(x\le0)$.
Hence by rewriting the right-hand side of \eqref{freeenergytransform} with $f_T$ which converges to $F_\tau(r)$ and by using a continuous version of Lemma~\ref{problemma1},
we conclude the proof of Theorem~\ref{CorUniversality}.
\begin{lemma}\label{lem:TinflimitK}
We have
\begin{equation}\label{TinflimitK}
\lim_{T\to\infty}\Xi\left(e^{-\frac{\tau^2+r}\sigma},\tau\sigma,\sigma\right)=g(\tau,r)\det(\Id-P_r\wh K_{\Ai}P_r)_{L^2(\R)}
\end{equation}
where $g$ is defined by \eqref{defg}, the shifted Airy kernel $\wh K_{\Ai}$ is given by \eqref{defKAi}, and $P_s(x)=\Id_{\{x>s\}}$.
\end{lemma}
\begin{proof}
Let us introduce the following notation for this proof. Consider the operator
\[B_r(x,y)=\Ai(x+y+r)\]
acting on $L^2(\R_+)$ as an integral operator and the functions
\[e_\alpha(x)=e^{\alpha x}\]
for $\alpha\in\R$. Note that $e_\alpha\not\in L^2(\R_+)$ if $\alpha\le0$, but since it will always appear in this proof together with $B_r$,
the fast decay of the Airy function makes all the integrals convergent.
We take the $T\to\infty$ limit on the left-hand side of \eqref{TinflimitK} by using \eqref{defz} with $S=e^{-(\tau^2+r)/\sigma}$ and $b=\tau\sigma$.
If we take the $T\to\infty$ limit of $\bar K_{\tau\sigma,\tau\sigma}$, we observe that
\[\frac{\sigma\pi}{\sin(\sigma\pi(z-w))}\to\frac1{z-w}\]
since $\sigma\to0$. On the other hand, because $\Gamma(z)\simeq z^{-1}$ as $z\to0$, we also have
\[\frac{\Gamma(\tau\sigma-\sigma z)}{\Gamma(\sigma z-\tau\sigma)}\frac{\Gamma(\sigma w-\tau\sigma)}{\Gamma(\tau\sigma-\sigma w)}\to1.\]
Substituting \eqref{Sscaling}, one obtains
\begin{align*}
\bar K_{\tau\sigma,\tau\sigma}(x,y)&\to\frac1{(2\pi\I)^2}\int_{-\frac1{4\sigma}+\I\R}\d w\int_{\frac1{4\sigma}+\I\R}\d z \frac{e^{z^3/3-z(y+r+\tau^2)}}{e^{w^3/w-w(x+r+\tau^2)}}\frac1{z-w}\\
&=K_{\Ai}(x+r+\tau^2,y+r+\tau^2)=B_r^2(x+\tau^2,y+\tau^2)
\end{align*}
pointwise.
By using the same argument as in the proof of \eqref{convFred}, we see that we also have convergence in trace norm and hence the convergence of Fredholm determinants.
Moreover, by applying the analogue of Lemma~\ref{lemma:resolventconv}, we obtain the convergence of the resolvents.
Similarly, we have
\begin{align*}
q_{\tau\sigma,\tau\sigma}(x)&\to-e^{-\tau^3-\tau r}\frac1{2\pi\I}\int_{-\frac1{4\sigma}+\I\R}\d w\,\frac{e^{-w^3/3+w(x+r+\tau^2)}}{\tau-w}\\
&=-e^{-\tau^3-\tau r}\int_0^\infty\d\lambda\Ai(x+r+\tau^2+\lambda)\,e^{-\lambda \tau}
=-e^{-\tau^3-\tau r}(B_r e_{-\tau})(x+\tau^2)
\end{align*}
and
\[q_{-\tau\sigma,-\tau\sigma}(x)\to-e^{\tau^3+\tau r}(B_r e_\tau)(x+\tau^2).\]
The convergence holds also in $L^2(\R_+)$ by the dominated convergence theorem.
By writing $r_{\tau\sigma}(x)=e^{\frac43\tau^3}e_{-\tau}(x+\tau^2)$ and $r_{-\tau\sigma}(x)=e^{-\frac43\tau^3}e_\tau(x+\tau^2)$, we get
\begin{multline}\label{scalarprodcalc}
\big\langle(\Id-\bar K_{\tau\sigma,\tau\sigma})^{-1} q_{\tau\sigma,\tau\sigma},q_{-\tau\sigma,-\tau\sigma}\big\rangle\\
\to\int_0^\infty\d x\int_0^\infty\d y(\Id-B_r^2)^{-1}(x+\tau^2,y+\tau^2)(B_re_{-\tau})(y+\tau^2)(B_re_\tau)(x+\tau^2),
\end{multline}
\begin{multline}
\big\langle(\Id-\bar K_{\tau\sigma,\tau\sigma})^{-1} q_{\tau\sigma,\tau\sigma},r_{\tau\sigma}\big\rangle\\
\to -e^{\tau^3/3-\tau r}\int_0^\infty\d x\int_0^\infty\d y(\Id-B_r^2)(x+\tau^2,y+\tau^2)(B_re_{-\tau})(y+\tau^2)e_{-\tau}(x+\tau^2),
\end{multline}
\begin{multline}
\big\langle(\Id-\bar K_{\tau\sigma,\tau\sigma})^{-1} r_{-\tau\sigma},q_{-\tau\sigma,-\tau\sigma}\big\rangle\\
\to -e^{-\tau^3/3+\tau r}\int_0^\infty\d x\int_0^\infty\d y(\Id-B_r^2)(x+\tau^2,y+\tau^2)e_\tau(y+\tau^2)(B_re_\tau)(x+\tau^2)
\end{multline}
and
\begin{multline}
\big\langle\bar K_{\tau\sigma,\tau\sigma}(\Id-\bar K_{\tau\sigma,\tau\sigma})^{-1} r_{-\tau\sigma},r_{\tau\sigma}\big\rangle\\
\to\int_0^\infty\d x\int_0^\infty\d y\int_0^\infty\d z B_r^2(x+\tau^2,y+\tau^2)(\Id-B_r^2)^{-1}(y+\tau^2,z+\tau^2)e_\tau(z+\tau^2)e_{-\tau}(x+\tau^2).
\end{multline}
Finally observe that by \eqref{Sscaling}
\begin{equation}\label{firsttermslimit}
(\tau\sigma)^2/\sigma^2+\sigma(2\gamma_{\rm E}+\ln S)\to-r.
\end{equation}
To get a similar formulation on the right-hand side of \eqref{TinflimitK}, we substitute $s=r$ to the ingredients defining the function $g$:
\begin{equation}\label{BFPcalc}\begin{aligned}
\wh K_{\Ai}(x+r,y+r)&=B_r^2(x+\tau^2,y+\tau^2),\\
\mathcal R&=r+e^{\tau^3/3-\tau r}\int_0^\infty\d x(B_re_{-\tau})(x+\tau^2)e_{-\tau}(x+\tau^2),\\
\Phi(x+r)&=e^{\tau^3/3-\tau r}\int_0^\infty\d y B_r^2(x+\tau^2,y+\tau^2)e_{-\tau}(y+\tau^2)-(B_re_\tau)(x+\tau^2),\\
\Psi(y+r)&=e^{-\tau^3/3+\tau r}e_\tau(y+\tau^2)-(B_re_{-\tau})(y+\tau^2).
\end{aligned}\end{equation}
To obtain \eqref{TinflimitK} one needs to substitute the limits \eqref{scalarprodcalc}--\eqref{firsttermslimit} to \eqref{defz},
comparing the result with the right-hand side of \eqref{TinflimitK} using the expressions of \eqref{BFPcalc},
and rewriting the scalar product in the definition of the function $g$ in \eqref{defg} as an integral.
\end{proof}
|
\section{Introduction}
Soon after discovery of general relativity (GR) first solutions
corresponding to spherical symmetric black holes were found
\cite{Schwarzschild_16,Reissner_16,Nordstrom_18}, however, initially
people were rather sceptical about possible astronomical
applications of the solutions corresponding to black holes
\cite{Einstein_39} (see, for instance, also one of the first
textbooks on GR \cite{Bergmann_42}). Even after an introduction of
the black hole concept by Wheeler \cite{Wheeler_68} (he used the
term in his public lecture in 1967 \cite{Frolov_11}) we don't know
not too many examples where we really need GR models with strong
gravitational fields which arise near black hole horizons to explain
observational data. The cases where we need strong field
approximation are very important since they give an opportunity to
check GR predictions in a strong field limit, therefore, one could
significantly constrain of alternative theories of gravity.
One of the most important options to test a gravity in the strong
field approximation is analysis of relativistic line shape as it was
shown in \cite{Fabian_89} with assumptions that a line emission is
originated at a circular ring area of a flat accretion disk. Later
on, such signatures of the Fe $K\alpha$-line have been found in the
active galaxy MCG-6-30-15 \cite{Tanaka_95}. Analyzing the spectral
line shape the authors concluded the emission region is so close to
the black hole horizon that one has to use Kerr metric approximation
\cite{Kerr_63} to fit observational data \cite{Tanaka_95}. Results
of our simulations of iron $K\alpha$ line formation are given in
\cite{Zakharov_99} (where we used our approach \cite{Zakharov_94}),
see also \cite{Fabian_10} for a more recent review of the subject.
Now there are two basic observational techniques to investigate a
gravitational potential at the Galactic Center, namely, a)
monitoring the orbits of bright stars near the Galactic Center to
reconstruct a gravitational potential \cite{Ghez_00} (see also a
discussion about an opportunity to evaluate black hole dark matter
parameters in \cite{Zakharov_07} and an opportunity to constrain
some class of an alternative theory of gravity \cite{Dusko_PRD_12});
b) in mm-band with VLBI-technique measuring a size and a shape of
shadows around black hole giving an alternative possibility to
evaluate black hole parameters. The formation of retro-lensing
images (also known as mirage, shadows or "faces" in the literature)
due to the strong gravitational field effects nearby black holes has
been investigated by several authors
\citep{Holz02,Geralico03,ZNDI05,ZNDI05b}.
Theories with extra dimensions admit astrophysical objects
(supermassive black holes in particular) which are rather different
from standard ones. Tests have been proposed when it would be
possible to discover signatures of extra dimensions in supermassive
black holes since the gravitational field may be different from the
standard one in the GR approach. So, gravitational lensing features
are different for alternative gravity theories with extra dimensions
and general relativity.
Recently, Bin-Nun \cite{Bin_Nun10,Bin_Nun10a,Bin_Nun10b} discussed
an opportunity that the black hole at the Galactic Center is
described by the tidal Reissner-- Nordstr\"om metric which may be
admitted by the Randall--Sundrum II braneworld scenario
\cite{Dadhich_00}. Bin-Nun suggested an opportunity of evaluating
the black hole metric analyzing (retro-)lensing of bright stars
around the black hole in the Galactic Center. Doeleman et al.
evaluated a size of the smallest spot for the black hole at the
Galactic Center
with VLBI technique in mm-band \cite{Doeleman_08} (see, constraints done from previous
observations \cite{Shen_05}). Theoretical
studies showed that the size of the smallest spot near a black hole
practically coincides with shadow size because the spot is
the envelope of the shadow \cite{Falcke00,ZNDI05,ZNDI05b}.
As it was shown \cite{ZNDI05,ZNDI05b}, measurements of the shadow
size around the black hole may help to evaluate parameters of black
hole metric \footnote{One of the first calculations of shapes of
orbits visible by a distant observer have been done in
\cite{Cunningham_73}. An apparent shape of Kerr black hole was
discussed in \cite{Bardeen_73} (see also a very similar picture in
monograph \cite{chandra}). Later on, the apparent shapes of black
holes are called shadows \cite{Falcke00}.}. Sizes and shapes of
shadows are calculated for different types of black holes and
gravitational lensing in strong gravitational field has been
analyzed in a number of papers \cite{Shadows}.
We derive an analytic expression for the
black hole shadow size as a function of the tidal charge for the
Reissner-- Nordstr\"om metric. We conclude that observational data
concerning shadow size measurements are not consistent with
significant negative charges, in particular, the significant tidal
charge $q=\mathcal{Q}^{2}/M^2=-6.4$ \footnote{A dimensional tidal
charge $Q=q*M^2=\mathcal{Q}^2$, where $Q$ is a tidal charge,
$\mathcal{Q}$ is the the standard charge in the Reissner--
Nordstr\"om metric. One can consider also negative tidal charges, it
means that a class of metrics with tidal charges is an extension of
the class of the standard Reissner-- Nordstr\"om metrics.},
discussed in
\cite{Bin_Nun10,Bin_Nun10a,Bin_Nun10b}, where the author used
slightly different notations, namely $q'=q/4$, is practically ruled
out with a very high probability (the tidal charge is roughly
speaking is far beyond $9\sigma$ confidence level). We also show a
smaller shadow sizes in respect to estimates obtained with the
Schwarzschild black hole model can be explained with the Reissner --
Nordstr\"om metric with a significant charge. It was found a
critical $q$ value for shadow existence, namely for $q \leq 9/8$,
Reissner -- Nordstr\"om black holes have shadows while for $q
> 9/8$ the shadows do not exist. Interestingly, the same critical value is responsible for
a qualitative different behavior of quasinormal modes for the
scattering \cite{Chirenti_12} and for existence of circular orbits
of neutral test particles \cite{Pugliese_11}.
As J. A. Wheeler coined "Black holes have no hair": it means that a
black hole is characterized by only three parameters, its mass $M$,
angular momentum $J$ and charge $\mathcal{Q}$ (see, e.g.
\cite{MTW,Wald84}, or \cite{Hertog_06} for a more recent review).
Therefore, in principle, charged black holes can be formed, although
astrophysical conditions that lead to their formation may look
rather problematic. Nevertheless, one could not claim that their
existence is forbidden by theoretical or observational arguments.
Moreover, we will show below that observations give a hint about an
existence of a significant charge, but its origin is not clear at
the moment.
Charged black holes are also object of intensive studies in quantum
gravity, since a static, spherically
symmetrical solution of Yang-Mills-Einstein equations with fairly natural
requirements on asymptotic behavior of the solutions gives a Reissner-Nordstr\"{o}m
metric
\cite{Gal1}. Thus, the metric describes a spherically symmetric
black hole with a color charge (and (or) a magnetic monopole).
Later on, color charges have been found for rotating black holes as
well
\cite{Kleihaus_97}.
\section{Basic Equations}
The expression for the Reissner - Nordstr\"{o}m metric in natural
units ($G=c=1$) has the form
\begin {equation}
ds^{2}=-\left(1-\frac{2M}{r}+\frac{\mathcal{Q}^{2}}{r^{2}}\right)dt^{2}+\left(1-\frac{2M}{r}+\frac{\mathcal{Q}^{2}}{r^{2}}\right)^{-1}dr^{2}+
r^{2}(d{\theta}^{2}+{\sin}^{2}\theta d{\phi}^{2}).
\label{RN_Lecce_0
\end {equation}
Applying the Hamilton-Jacobi method to the problem of
geodesics in the Reissner - Nordstr\"{o}m metric,
the motion of a test particle in the $r$-coordinate can be described
by following equation (see, for example, \cite{MTW})
\begin {eqnarray}
r^{4}(dr/d\lambda)^{2}=R(r),\label{RN_D 1}
\end {eqnarray}
where $\lambda$ is the affine parameter \cite{MTW} and
\begin {eqnarray}
&& R(r)=P^{2}(r)-\Delta({\mu}^{2}r^{2}+L^{2}), \nonumber\\
&& P(r)=Er^{2}-e\mathcal{Q}r, \label{RN_D_2}\\
&& \Delta=r^{2}-2Mr+\mathcal{Q}^{2}. \nonumber
\end {eqnarray}
Here, the constants $\mu, E, L$ and $e$ are associated with the
particle, i.e. $\mu$ is its mass, $E$ is energy at infinity, $L$ is
its angular momentum at infinity and $e$ is the particle's charge.
We shall consider the motion of uncharged particles $(e=0)$ below.
In this case, the expression for the polynomial $R(r)$ takes the
form
\begin {eqnarray}
R(r)=(E^{2}-{\mu}^{2})r^{4}+
2M{\mu}r^{3}-(\mathcal{Q}^{2}{\mu}^{2}+L^{2})r^{2}
+2ML^{2}r-\mathcal{Q}^{2}L^{2}. \label{RN_D_3}
\end {eqnarray}
Depending on the multiplicities of the roots of the
polynomial $R(r)$, we can have three types of motion in the $r$ -
coordinate \citep{Zakharov_86}. In particular, by defining the outer
event horizon as usual $r_{+}=M+\sqrt{M^2-\mathcal{Q}^{2}}$
\cite{MTW}, we have:
(1) if the polynomial $R(r)$
has no roots for $r\geq r_{+}$, a test particle is captured by the
black hole;
(2) if $R(r)$ has roots and $(\partial{R}/\partial{r})(r_{max})\neq
0$ with $r_{max} > r_+$ ($r_{max}$ is the maximal root), a particle
is scattered after approaching the black hole;
(3) if $R(r)$ has a root and $R(r_{max}) = (\partial{R}/
\partial{r})(r_{max})=0$,
the particle now takes an infinite proper time to approach the
surface $r = const$.
If we are considering a photon ($\mu = 0$), its motion in the $r$-coordinate depends
on the root multiplicity of the polynomial $\hat{R}(\hat{r})$
\begin {eqnarray}
\hat{R}(\hat{r})={R(r)}/({M^{4}E^{2}})={\hat{r}}^{4}-\xi^{2}{\hat{r}}^{2}+2{\xi}^{2}\hat{r}
-{\hat{\mathcal{Q}}}^{2}{\xi}^{2}.\label{RN_D_4}
\end {eqnarray}
where $\hat {r}=r/M, \xi=L/(ME)$ and $\hat
{\mathcal{Q}}=\mathcal{Q}/M.$ Below we do not write the hat symbol
for these quantities.
One could see from Eq. (\ref{RN_D_4}) and Eqs. (\ref{RN_D_2}) as
well that the black hole charge may influence substantially the
photon motion at small radii ($r \approx 1$), while the charge
effect is almost negligible at large radial coordinates of photon
trajectories ($r
>> 1$). In the last case we should keep in mind that the charge may cause
only small corrections on photon motion.
\section{Derivation of shadow size as a function of charge}
Let us consider the problem of the capture cross section of
a photon by a charged black hole. It is clear that the critical value of the
impact parameter for a photon to be captured by a Reissner - Nordstr\"{o}m black hole
depends on the multiplicity root condition of the polynomial $R(r)$.
This requirement is equivalent to the vanishing discriminant
condition \cite{Kostrikin_82}. To find the critical value of impact
parameter for Schwarzschild and RN metrics the condition has been
used for corresponding cubic and quartic equations
\cite{Zakharov_88,Zakharov_91,Zakharov_94b}. In particular, it was
shown that for these cases the vanishing discriminant condition
approach is more powerful in comparison with the procedure excluding
$r_{max}$ from the following system
\begin {eqnarray}
R(r_{max}) = 0, \quad \\
\label{RN_D_4a}
\dfrac{\partial{R}}{\partial{r}}(r_{max})=0,
\label{RN_D_4b}
\end {eqnarray}
as it was done, for example, by Chandrasekhar \cite{chandra} (and
earlier by Darwin \cite{Darwin_58}) to solve similar problems,
because $r_{max}$ is automatically excluded in the condition for
vanishing discriminant.
Introducing the notation $\xi^{2}=l, \mathcal{Q}^{2}=q$, we obtain
\begin {eqnarray}
R(r)=r^{4}-lr^{2}+2lr-ql.\label{RN_D_5}
\end {eqnarray}
We remind basic algebraic definitions and relations. If we consider
an arbitrary polynomial $f(X)$ with degree $n$
\begin {eqnarray}
f(X) = X^n+a_1X^{n-1}+...+a_{n-1}X+a_n,
\end {eqnarray}
the elementary symmetric polynomials $s_{k}$ have the following
form, where $X_1,...X_n$ are roots of the polynomial $f(X)$
\cite{Kostrikin_82}
\begin {eqnarray}
s_{k} (X_1,...X_n) = \sum_{1 \leqslant i_1 < i_2 <...<i_k
\leqslant n} X_{i_{1}}X_{i_{2}}...X_{i_k},
\end {eqnarray}
where $k=1,2,..., n$. The symmetrical $k$-power sum polynomial $p_k$
have the following expression
\begin {eqnarray}
p_{k} (X_1,...X_n) = X_{1}^k+X_{2}^k+...+X_{n}^k, \quad {\rm for}
\quad k \geq 0.
\end {eqnarray}
To express $p_k$ through $s_k$ one can use Newton's equations
\begin {eqnarray}
p_{k}-p_{k-1}s_1+p_{k-2}s_2+...+(-1)^{k-1}p_1s_{k-1}+(-1)^k ks_k=0,
\quad {\rm for} \quad 1 \leqslant k \leqslant n;
\end {eqnarray}
\begin {eqnarray}
p_{k}-p_{k-1}s_1+p_{k-2}s_2+...+(-1)^{n-1}p_{k-n+1}s_{n-1}+(-1)^n
p_{k-n} s_n=0, \quad {\rm for} \quad k > n.
\end {eqnarray}
We introduce the following polynomial
\begin {eqnarray}
\Delta_n (X_1,...X_n) = \prod_{1 \leqslant i < j \leqslant n}
(X_i-X_j),
\end {eqnarray}
which can be represented as the Vandermonde determinant
\begin{eqnarray}
\Delta_n (X_1,...X_n)=
\left|
\begin{array}{cccc}
1 & 1 & ... & 1\\
X_1 & X_2 & ... & X_n\\
... & ... & ... & ...\\
X_1^{n-1} & X_2^{n-1} & ... & X_n^{n-1}
\end{array}
\right|.
\end {eqnarray}
According to the discriminant $Dis$ definition we have the
$Dis(s_1,...,s_n)$ polynomial
\begin {eqnarray}
Dis(s_1,...,s_n)=\Delta_n^2 (X_1,...X_n) = \prod_{1 \leqslant i <
j \leqslant n} (X_i-X_j)^2,
\end {eqnarray}
one can find \cite{Kostrikin_82}
\begin{eqnarray}
Dis(s_1,...s_n)=
\left|
\begin{array}{ccccc}
n & p_1 & p_2 & ... & p_{n-1}\\
p_1 & p_2 & p_3 &... & p_n\\
p_2 & p_3 & p_4 & ... & p_{n+1}\\
... & ... & ... & ...& ...\\
p_{n-1}& p_n & p_{n+1} & ... & p_{2n-2}
\end{array}
\right|.
\end {eqnarray}
Clearly, that the vanishing discriminant condition is equivalent to
an existence of multiple roots among roots $X_1,...X_n$. We apply
this technique for the quartic polynomial $R(r)$ in Eq.
(\ref{RN_D_5}). So that the symmetric $k$-power polynomials for $n =
4$ have the form
\begin {eqnarray}
p_{k}=X_{1}^{k}+X_{2}^{k}+X_{3}^{k}+X_{4}^{k}, k \geq 0.
\end {eqnarray}
The symmetric elementary polynomials for $n = 4$ have the form
\begin {eqnarray}
&& s_{1}=X_{1}+X_{2}+X_{3}+X_{4}, \nonumber\\
&& s_{2}=X_{1}X_{2}+X_{1}X_{3}+X_{1}X_{4}+X_{2}X_{3}+X_{2}X_{4}+X_{3}X_{4},\nonumber\\
&& s_{3}=X_{1}X_{2}X_{3}+X_{2}X_{3}X_{4}+X_{1}X_{3}X_{4}+X_{1}X_{2}X_{4}, \\
&& s_{4}=X_{1}X_{2}X_{3}X_{4}. \nonumber
\end {eqnarray}
We calculate the discriminant of the family $X_{1},X_{2},X_{3},X_{4}$\\
\begin {eqnarray}
Dis(s_{1},s_{2},s_{3},s_{4})=
\left|
\begin{array}{cccc}
1& 1& 1& 1\\
X_{1}& X_{2}& X_{3}& X_{4}\\
X_{1}^{2}& X_{2}^{2}& X_{3}^{2}& X_{4}^{2}\\
X_{1}^{3}& X_{2}^{3}& X_{3}^{3}& X_{4}^{3}
\end {array}
\right|^2 =\left|
\begin{array}{cccc}
4 & p_{1}& p_{2}& p_{3}\\
p_{1}& p_{2}& p_{3}& p_{4}\\
p_{2}& p_{3}& p_{4}& p_{5}\\
p_{3}& p_{4}& p_{5}& p_{6}
\end {array}
\right|
\end {eqnarray}
Expressing the polynomials $p_{k} (1\leq k \leq 6)$ in terms of the
polynomials $s_{k} (1\leq k\leq 4)$ and using Newton's equations we
calculate the polynomials and discriminant of the family
$X_{1},X_{2},X_{3},X_{4}$ in roots of
the polynomial $R(r)$; we obtain
\begin {eqnarray}
&& p_{1}=s_{1}=0, \quad p_{2}=-2s_{2}, \quad p_{3}=3s_{3},\nonumber\\
&& p_{4}=2s_{2}^{2}-4s_{4}, \quad p_{5}=-5s_{3}s_{2},\\
&& p_{6}=-2s_{2}^{3}+3s_{3}^{2}+6s_{4}s_{2},\nonumber
\end {eqnarray}
where $s_{1}=0,s_{2}=-l,s_{3}=-2l,s_{4}=-ql,$ corresponding to the
polynomial $R(r)$ in Eq. (\ref{RN_D_5}). The discriminant $Dis$ of
the polynomial $R(r)$ has the form:
\begin {eqnarray}
Dis(s_{1},s_{2},s_{3},s_{4})=
\left|
\begin{array}{cccc}
4 & 0 & 2l & -6l\\
0 & 2l & -6l & 2l(l+2q)\\
2l & -6l & 2l(l+2q) & -10l^{2}\\
-6l & 2l(l+2q) & -10l^{2} & 2l^{2}(l+6+3q)
\end{array}
\right|=\nonumber\\
=16l^{3}[l^{2}(1-q)+l(-8q^{2}+36q-27)-16q^{3}]. \label{RN_D_6}
\end {eqnarray}
The polynomial $R(r)$ thus has a multiple root if and only if
\begin {eqnarray}
l^{3}[l^{2}(1-q)+l(-8q^{2}+36q-27)-16q^{3}]=0. \label{RN_D_7}
\end {eqnarray}
Excluding the case $l=0$, which corresponds to a multiple root at
$r=0$, we find
that the polynomial $R(r)$ has a multiple root for $ r\geq r_{+}$ if and only if
\begin {eqnarray}
l^{2}(1-q)+l(-8q^{2}+36q-27)-16q^{3}=0. \label{RN_D_8}
\end {eqnarray}
If $q=0$, we obtain the well-known result for a Schwarzschild black
hole \citep{MTW,Wald84,Lightman}, $l_{\rm cr}=27$, or
$\xi_{cr}=3\sqrt{3}$ (where $l_{\rm cr}$ is the positive root of Eq.
(\ref{RN_D_8})). If $q=1$, then $l = 16$, or $\xi_{cr}=4$, which
also corresponds to numerical results given in paper \cite{Young}.
The photon capture cross section for an extreme charged
black hole turns out to be considerably smaller than the capture cross section of a
Schwarzschild black hole. The critical value of the impact parameter,
characterizing the capture cross section for a RN black hole, is determined by the equation
\begin {eqnarray}
l_{\rm cr}=\frac{(8q^{2}-36q+27)+\sqrt{D_1}}{2(1-q)}, \label{RN_D_9}
\end {eqnarray}
where
$D_1=(8q^{2}-36q+27)^{2}+64q^{3}(1-q)=-512\left(q-\dfrac{9}{8}\right)^3.$
It is clear from the last relation that there are circular unstable
photon orbits only for $q \le \dfrac{9}{8}$ (see also results in
\cite{Pugliese_11} about the same critical value). Substituting
Eq.(\ref{RN_D_9}) into the expression for the coefficients of the
polynomial $R(r)$ it is easy to calculate the radius of the unstable
circular photon orbit (which is the same as
the minimum periastron distance). The orbit of a
photon moving from infinity with the critical impact parameter, determined
in accordance with Eq.(\ref{RN_D_9}) spirals into circular orbit.
To find a radius of photon unstable orbit we will solve Eq.
(\ref{RN_D_4b}) substituting $l_{\rm cr}$ in the relation. From
trigonometric formula for roots of cubic equation we have
\begin {eqnarray}
r_{\rm crit}=2\sqrt{\frac{l_{\rm cr}}{6}} \cos{\frac{\alpha}{3}},
\label{RN_D_10}
\end {eqnarray}
where
\begin {eqnarray}
\cos \alpha={-\sqrt{\frac{27}{2 l_{\rm cr}}}}, \label{RN_D_11}
\end {eqnarray}
As it was explained in \cite{ZNDI05b} this leads to the formation of
shadows described by the critical value of $\xi_{cr}$ or, in other
words, in the spherically symmetric case, shadows are circles with
radii $\xi_{cr}$. Therefore, by measuring the shadow size, one could
evaluate the black hole charge in black hole mass units $M$. In
Fig. \ref{Fig1} a shadow radius and a radius of last unstable orbit
for photons as a function of $q$ are given as a function of charge
(including possible tidal charge with a negative $q$ and
super-extreme charge $q>1$).
\begin{figure}[th!]
\begin{center}
\includegraphics[width=10.5cm]{shadow_orbits.ps}
\end{center}
\caption{Shadow (mirage) radius (solid line) and radius of the last
circular unstable photon orbit (dot-dashed line) in $M$ units as a
function of
$q$. The critical value $q=9/8$ is shown with dashed vertical line.}
\label{Fig1}
\end{figure}
\section{Consequences}
\subsection{A disappearance of shadows for naked singularities}
In spite of the cosmic censorship hypothesis \cite{Penrose_69} that
a singularity has to be shielded by a horizon, properties of naked
singularities are a subject of intensive theoretical studies. As
usual spherical symmetrical cases are easier for analysis and RN
metrics with super extreme charge $q>1$ are investigated in a number
of papers, see, for instance \cite{Virbhadra_02} and references
therein.
So, if we assume that $q> 1$, we can see from Eq. (\ref{RN_D_9})
that for $q \leq 9/8$ we have shadows, while for $q
> 9/8$ the shadows do not exist. For these charges ($q
> 9/8$)
incoming photons always scattering by black holes for $l \neq 0$
because the polynomial $R(r)$ has no multiple roots but it has a
single positive root (it means scattering) since for great positive
$r$ we have $R(r)
> 0$ while $R(0) <0$. The degenerate case of radial trajectories of
photons ($l = 0$) can be ignored as the case with "zero measure" or
the structural unstable case using the Poincar\'e -- Pontryagin --
Andronov -- Anosov -- Arnold terminology \cite{Arnold_88}. It means
that in any small vicinity a behavior of other geodesics from the
radial ones is qualitatively different, therefore, such objects can
not be observed in nature. Therefore, shadows exist only for $q \leq
9/8$. So, $q=9/8$ is critical value which is characterized
"catastrophe" \cite{Arnold_04} or the qualitatively different
behavior of the system (the appearance and the disappearance of
shadows).
For the critical $q=9/8$ we have the smallest shadow with $l=27/2$
and a shadow size $\xi=\sqrt{13.5} \approx 3.674$ (in $M$ units) or
$37.5~\mu as$ in diameter for the black hole at the Galactic Center.
For this impact parameter we have corresponding circular unstable
orbit for photons with $r=1.5$ (in $M$ units).
\subsection{Observational constraints on a charge of the black hole
at the Galactic Center}
{ If we adopt the distance toward the Galactic Center $d_*=8.3\pm
0.4$~kpc (or $d_*=8.35\pm 0.15$~kpc \cite{Reid_14}) and mass of the
black hole $M_{BH}=(4.3\pm 0.4) \times 10^6 M_\odot$
\cite{Gillessen_09,Falcke_13} (a significant part of black hole mass
uncertainty is connected with a distance determination uncertainty
\cite{Falcke_13}), then we have the angle 10.45~$\mu as$ for the
corresponding Schwarzschild radius $R_g=2.95*
\dfrac{M_{BH}}{M_\odot}*10^5$~cm roughly with 10\% uncertainty of
black hole mass and distance estimations, so a shadow size for the
Schwarzschild black hole is around 53~$\mu as$, for a black hole
with a tidal charge ($q=-6.4$) suggested by Bin-Nun
\cite{Bin_Nun10,Bin_Nun10a,Bin_Nun10b} a shadow size is about
86.1~$\mu as$, while for the extreme charge ($q=1$) and critical
charge ($q=9/8$) the shadow sizes are 40.9~$\mu as$ and 37.5~$\mu
as$, respectively.Uncertainties of angular shadow size evaluations
are at a level around 10\% which corresponds to an
uncertainty of black hole mass evaluation.
}
\subsection{Comparison with observations}
A couple of year ago Doeleman et al. \cite{Doeleman_08} claimed that
intrinsic diameter of Sgr $A^*$ is $37^{+16}_{-10}~\mu as$ at the
$3~ \sigma$ confidence level. If we believe in GR and the central
object is a black hole, then we have to conclude that a shadow
practically coincides with the intrinsic diameter, so in spite of
the fact that a Schwarzschild black hole is marginally consistent
with observations, a Reissner -- Nordstr\"om black hole provides
much better fit of a shadow size, while a black hole with a
significant tidal charge ($q=-6.4$) is out of a more 9~$\sigma$
level interval. Later on, an accuracy of intrinsic size
measurements was significantly improved, so Fish et al.
\cite{Fish_11} gave $41.3^{+5.4}_{-4.3}~\mu as$ (at 3~$\sigma$
level) on day 95, $44.4^{+3.0}_{-3.0}~\mu as$ on day 96 and
$42.6^{+3.1}_{-2.9}~\mu as$ on day 97, so a tidal charge ($q=-6.4$)
is out of $26~\sigma$ level for day 95 and even less probable for
other observations.
The black hole in the elliptical galaxy M87 looks also perspective
to evaluate shadow size \cite{Doeleman_12} (probably even its shape
in the future to estimate a black hole spin) because the distance
toward the galaxy is $16 \pm 0.6 $~Mpc \cite{Blakeslee_09}, black
hole mass is $M_{M87}=(6.2\pm 0.4) \times 10^9 M_\odot$
\cite{Gerbhardt_11}, so that an angle $(7.3 \pm 0.5) \mu as$
corresponds to the Schwarzschild radius \cite{Doeleman_12}, so the
angle is comparable with the corresponding value considered earlier
for our Galactic Center case. Recently, it was reported that
smallest shadow size is $5.5 \pm 0.4 R_{SCH}$ with $1~\sigma$
errors (where $R_{SCH}=2 G M_{M87}/c^2$) \cite{Doeleman_12}, so
that at the moment the shadow size is consistent with the
Schwarzschild metric for the object.
\section{Conclusions}
Based on observations \cite{Doeleman_08,Fish_11} one can say that
for the Schwarzschild black hole model we have tensions between
evaluations of black hole mass done with observations of bright star
orbits near the Galactic Center and the evaluated shadow size. To
reduce tensions between estimates of the black hole mass and the
intrinsic size measurements, one can use the Reissner -- Nordstr\"om
metric with a significant charge which is comparable with the
critical one. We do not claim that the corresponding charge has an
electric origin because an interstellar environment is electrically
neutral, so the corresponding charge may be induced (like a tidal
charge induced by extra dimension) and has a non-electric origin.
Charge estimates for the Reissner -- Nordstr\"om metric given from
geodesic trajectories for orbital motions are given in
\cite{Iorio_12}.
Recent estimates of the smallest structure in the M87 published in
paper \cite{Doeleman_12} do not need an introduction of charge
(tidal or normal) to fit observational data because sizes of the
smallest spot near the black hole at the object are still consistent
with the shadow size evaluated for the Schwarzschild metric.
\begin{acknowledgments}
The author thanks D. Borka, V. Borka Jovanovi\'c, F. De Paolis, G.
Ingrosso, P. Jovanovi\'c, S. M. Kopeikin, A. A. Nucita, B. Vlahovic
for useful discussions. The authors acknowledges also A. Broderick
and C. L\"ammerzahl for conversations at GR-20/Amaldi-10 in Warsaw.
The author thanks anonymous referees for their very useful and
constructive remarks and J. Nimmo (NCCU) for his kind help to
improve English of the manuscript.
The work was supported in part by the NSF
(HRD-0833184) and NASA (NNX09AV07A) at NASA CADRE and NSF CREST
(NCCU, Durham, NC, USA) and RFBR 14-02-00754a at ICAD of RAS
(Moscow).
\end{acknowledgments}
|
\section{Introduction}\label{sec:intro}
According \cite[Theorem 1.8.1]{bcn} we denote by $\Gamma(H)$ the Hadamard graph induced by a Hadamard matrix $H$.
It is known that, for a Hadamard matrix $H$ of order $n > 1$, $\Gamma(H)$ is an antipodal bipartite distance-regular graph with $4n$ vertices and diameter $4$.
Also, \cite[page 58]{bcn} shows that every distance-regular graph of diameter $d$ induces a $d$-class association scheme whose relations are defined by
the distance function of the graph. Thus, $\Gamma(H)$ induces an association scheme $(\Omega, S)$ such that $\Omega$ is the vertex set of $\Gamma(H)$
and $S = \{ s_i \mid 0 \leq i \leq 4 \}$, where $s_i = \{ (x,y) \in \Omega \times \Omega \mid d_{\Gamma(H)}(x,y) = i \}$.
Now we consider a fission scheme of $(\Omega, S)$ such that only $s_2$ is partitioned into $\{ t_1, t_2, \dotsc, t_m \}$ and
\begin{equation}\label{A}
| xs_4 \cap yt_j | = 1 ~\text{for all}~ j=1,2, \dotsc, m,
\end{equation}
where $(y, x) \in t_j$ and $xs:=\{ y \mid (x, y) \in s \}$ for a binary relation $s$.
Since $s_0 \cup s_2 \cup s_4$ is an equivalence relation on $\Omega$ with exactly two equivalence classes, say $Y_1$ and $Y_2$, the restrictions of
$\{ s_0, s_4, t_1, t_2, \dotsc, t_m \}$ to $Y_i$ form the relation set of an association scheme for $i= 1, 2$.
In this article we deal with association schemes with the same intersection numbers as the fission scheme $(\Omega, \{t_j \mid 1 \leq j \leq m\} \cup \{s_0, s_1, s_3, s_4 \})$.
In the study of association schemes it has been one of the main topics to characterize association schemes, especially related to distance-regular
graphs by intersection numbers (see \cite[Chapter 9]{bcn} or \cite{bannai}), and many $P$- and $Q$- polynomial association schemes with large diameter are
uniquely determined by its intersection numbers: cf. \cite{bcn, ter, ega}. On the other hand, we can find quite many isomorphism classes of association schemes with the same intersection numbers as one can see in Table 3.
But, it does not guarantee that we can find such a huge number of association schemes of order $n$
with the same intersection numbers when $n$ is large enough.
In this article we prove that $(\Omega, \{t_j \mid 1 \leq j \leq m\} \cup \{s_0, s_1, s_3, s_4 \})$ is an association scheme
and give a lower bound for the number of isomorphism classes of association schemes with the same intersection numbers as $(\Omega, \{t_j \mid 1 \leq j \leq m\} \cup \{s_0, s_1, s_3, s_4 \})$.
\vskip10pt
The following are our main results which show the reason why so many isomorphism classes appear as mentioned above.
Let $(X, S)$ be an association scheme of order $n$ and $H$ a Hadamard matrix whose rows and columns are indexed by the elements of $X$.
We denote by $\mathbb{F}_2$ the finite field with two elements.
Also we denote by $x_{ab}$ an element $(x, a, b)$ of $X \times \mathbb{F}_2 \times \mathbb{F}_2$ for short.
Define
\begin{equation}\label{M}
\begin{array}{lll}
\widetilde{X}&=& \{ x_{ab} \mid x \in X, a, b \in \mathbb{F}_2 \}, \\
\widetilde{t}&=&\{(x_{ab}, x_{a(b+1)})\mid x \in X, a, b \in \mathbb{F}_2 \}, \\
\widetilde{s}&=&\{(x_{ab}, y_{ac}) \mid (x, y) \in s, a, b, c \in \mathbb{F}_2 \} \ \ \mbox{for}\ s \in S\setminus\{1_{X}\}, \\
r^{1}_H&= & \{(x_{ab}, y_{cd}) \mid x, y \in X, a, b, c, d \in \mathbb{F}_2, (1- \delta_{ac})(H^{T(a)})_{xy} = (-1)^{b+d} \}, \\
r^{-1}_H&= & \{(x_{ab}, y_{cd}) \mid x, y \in X, a, b, c, d \in \mathbb{F}_2, a\neq c \} \setminus r^{1}_H, \\
S(H)&=& \{1_{\widetilde{X}}, \widetilde{t}\}\cup \{\widetilde{s}\mid s \in S\setminus\{1_{X}\}\}\cup \{r^{\epsilon}_H\mid\epsilon=\pm 1\},
\end{array}
\end{equation}
where $H^T$ is the transpose of a matrix $H$,
\[ \delta_{ac} = \left\{
\begin{array}{ll}
1 & \hbox{if $a=c$;} \\
0 & \hbox{otherwise},
\end{array}
\right.\]
\[H^{T(a)} = \left\{
\begin{array}{ll}
H & \hbox{if $a=0$;} \\
H^T & \hbox{if $a=1$.}
\end{array}
\right.\]
Note that $(\widetilde{X}, r^{1}_H)$ is the Hadamard graph $\Gamma(H)$.
\begin{thm}\label{thm:main1}
$(\widetilde{X}, S(H))$ is an association scheme.
\end{thm}
Note that $S(H)$ is a refinement of $C_2 \wr (S \wr C_2)$, where $C_2$ is the relation set of an association scheme of order $2$
(see Section \ref{sec:pre} for the definition of wreath product).
We prepare some notations.
\begin{itemize}
\item For a finite set $X$, $Sym(X)$ is the symmetric group on $X$.
\item For a permutation $\sigma \in Sym(X)$, $P_\sigma$ is the permutation matrix with respect to $\sigma$.
\item For a permutation group $G \leq Sym(X)$, $\mathcal{P}(G) := \{ P_\sigma \mid \sigma \in G \}$.
\item $diag(\varepsilon_x \mid x \in X)$ is a diagonal matrix whose the $(x, x)$-entry is $\varepsilon_x$,
where $\varepsilon: X \rightarrow \mathbb{C}$ is a function defined by $x \mapsto \varepsilon_x$.
\item $D_x$ is a diagonal matrix such that the $(x, x)$-entry is $-1$ and the other diagonal entries are $1$.
\end{itemize}
Let $\mathrm{I} := \mathcal{P}(\mathrm{Iso}(X,S))$ and
$\mathrm{D} := \langle \{ D_x \mid x \in X \} \rangle$ (see Section \ref{sec:pre} for the definition of $\mathrm{Iso}(X,S)$).
Also let $H_1$ and $H_2$ be $n \times n$ Hadamard matrices whose rows and columns are indexed by the elements of $X$.
We say that $H_1$ is \textit{similar} to $H_2$ with respect to $(X,S)$ if there exist $(P',P), (Q',Q) \in \mathrm{D} \times \mathrm{I}$ such that $H_2 = (P'P)^{-1} H_1 Q'Q$ or $H_2^T = (P'P)^{-1} H_1 Q'Q$,
and $(P'P)^{-1} Q'Q \in \mathrm{D}\rtimes \mathcal{P}(\mathrm{Aut}(X, S))$
(see Section \ref{sec:pre} for the definition of $\mathrm{Aut}(X,S)$).
Note that the similarity is an equivalence relation.
\begin{thm}\label{thm:main2}
$(\widetilde{X}, S(H_1))$ is isomorphic to $(\widetilde{X}, S(H_2))$ if and only if
$H_1$ is similar to $H_2$ with respect to $(X,S)$.
\end{thm}
Recall that two Hadamard matrices are \textit{equivalent} if one can be transformed into the other by a series of row or column permutations
or negations.
We define a group
\[\mathrm{Aut}_{x_0}(H) = \{ (\sigma, \tau) \in Sym(X) \times Sym(X) \mid P_\sigma^{-1} H P_\tau = H, \sigma(x_0)=x_0, \tau(x_0)=x_0 \}. \]
\begin{thm}\label{thm:main3}
Let $(X, S)$ be an association scheme of order $n$, $H_0$ a Hadamard matrix whose rows and columns are indexed by the elements of $X$, and $x_0 \in X$.
Then there are at least \[\frac{(n-1)! (n-1)!}{ 2 |\mathrm{Aut}_{x_0}(H_0)| |\mathrm{Aut}(X,S)| |\mathrm{Iso}(X,S)|}\]
isomorphism classes
of association schemes in $\{ (\widetilde{X}, S(H)) \mid H ~\text{is equivalent to}~ H_0 \}$.
\end{thm}
This article is organized as follows.
In Section \ref{sec:pre}, we prepare some terminology and notations.
In Section \ref{sec:main}, we give proofs of the main theorems.
In Section \ref{sec:table}, we list tables related to the main results when $n = 4, 8$.
\section{Preliminaries}\label{sec:pre}
Based on \cite{zies,zies2}, we use the notation on association schemes.
Let $X$ be a non-empty finite set.
Let $S$ denote a partition of $X \times X$. Then the pair $(X, S)$ is
an \textit{association scheme} (or shortly \textit{scheme}) if it satisfies the following conditions:
\begin{enumerate}
\item $1_{X} := \{ (\alpha, \alpha) \mid \alpha \in X \} \in S$;
\item For each $s \in S$, $s^* := \{(\alpha, \beta) \mid (\beta, \alpha) \in s \} \in S$;
\item For all $ s, t, u \in S$, $c_{st}^u := | \alpha s \cap \beta t^* |$ is constant whenever $(\alpha, \beta) \in u$,
\end{enumerate}
where $\alpha s:=\{ \beta \in X \mid (\alpha, \beta) \in s \}$.
The numbers $\{ c_{st}^u \mid s, t, u \in S \}$ are called the \textit{intersection numbers} of $S$.
For each $ s \in S$, we abbreviate $c_{ss^*}^{1_X} $ as $n_s$, which is called the \textit{valency} of $s$.
We call $\sum_{s \in S} n_s$ the \textit{order} of $(X, S)$ which is equal to $|X|$.
The \textit{thin residue} of $S$ is the smallest subset $\Or(S)$ of $S$
such that $\bigcup_{t\in \Or(S)}t$ is an equivalence on $X$ and the factor scheme of $(X,S)$ over $\Or(S)$
is induced by a regular permutation group (see \cite[page 45]{zies2} for the definitions).
For each $s \in S$, we denote by $A_s$ the \textit{adjacency matrix} of $s$.
Namely $A_s$ is a matrix whose rows and columns are indexed by the elements of $X$ and $(A_s)_{xy} = 1$ if $(x,y) \in s$ and $(A_s)_{xy} = 0$ otherwise.
\vskip5pt
Let $(X,S)$ and $(X_1, S_1)$ be association schemes. A bijective mapping $\phi : X \cup S \rightarrow X_1 \cup S_1$ is called an
\textit{isomorphism} from $(X,S)$ to $(X_1, S_1)$ if it satisfies the following conditions:
\begin{enumerate}
\item $\phi(X) \subseteq X_1$ and $\phi(S) \subseteq S_1$;
\item For all $x, y \in X$ and $s \in S$ with $(x,y) \in s$, $(\phi(x), \phi(y)) \in \phi(s)$.
\end{enumerate}
We denote by $\mathrm{Iso}(X,S)$ the set of isomorphisms from $(X,S)$ to itself.
An isomorphism $\phi : X \cup S \rightarrow X \cup S$ is called an \textit{automorphism} of $(X, S)$ if $\phi(s) = s$ for all $s \in S$.
We denote by $\mathrm{Aut}(X,S)$ the automorphism group of $(X, S)$.
We say that two association schemes $(X,S)$ and $(X_1, S_1)$ are \textit{algebraically isomorphic} or have the \textit{same intersection numbers} if there exists a bijection $\iota : S \rightarrow S_1$ such that $c_{rs}^{t} = c_{\iota(r) \iota(s)} ^{\iota(t)}$ for all $r, s, t \in S$.
\vskip5pt
Let $(W,F)$ and $(Y,H)$ be association schemes.
For each $f \in F$ we define
\[ \overline{f}:= \{ ((w_1,y),(w_2,y)) \mid y \in Y, (w_1,w_2) \in f \} .\]
For each $h \in H \setminus \{1_Y\}$ we define
\[ \overline{h}:=\{ ((w_1,y_1),(w_2,y_2)) \mid w_1, w_2 \in W, (y_1,y_2) \in h \}. \]
Denote $F \wr H := \{ \overline{f} \mid f \in F \} \cup \{\overline{h} \mid h \in H \setminus \{1_Y\} \}$.
Then $(W \times Y, F \wr H)$ is an association scheme called the \textit{wreath product} of $(W,F)$ and $(Y,H)$.
\section{Proofs of the main theorems}\label{sec:main}
\begin{flushleft}
\textbf{Proof of Theorem~\ref{thm:main1}}
\end{flushleft}
Remark that $A_{r_H^1}$ forms the adjacency matrix of the Hadamard graph $\Gamma(H)$.
Then $A_{1_{\widetilde{X}}}$, $A_{r_H^1}$, $A_{S^\sqcup}$, $A_{r_{H}^{-1}}$, and $A_{\tilde{t}}$ are the adjacency matrices of distance $i$ graphs $\Gamma(H)_i \ (0 \leq i \leq 4)$, where $S^\sqcup=\bigcup_{s \in S \setminus \{1_X\}}\widetilde{s}$.
Since $\Gamma(H)$ is distance-regular, $(\widetilde{X}, \{1_{\widetilde{X}}, r_H^1, S^\sqcup, r_{H}^{-1}, \tilde{t}\})$ forms an association scheme.
So, it suffices to show that $A_{r_1} A_{r_2}$ is a linear combination of $\{A_{1_{\widetilde{X}}}, A_{r_H^1},A_{r_{H}^{-1}}, A_{\tilde{t}} \}\cup \{A_{\widetilde{s}}\mid s\in S\setminus \{1_X\}\}$, where at least one of $r_1$ and $r_2$ is in $\{\widetilde{s}\mid s\in S\setminus \{1_X\}\}$. From now on, let $s_0 \in S\setminus \{1_X\}\}$.
By the definition of $S(H)$, $$A_{\widetilde{s_0}}A_{1_{\widetilde{X}}}=A_{1_{\widetilde{X}}}A_{\widetilde{s_0}}=A_{\widetilde{s_0}}$$ and $$A_{\widetilde{s_0}}A_{\widetilde{t}}=A_{\widetilde{t}}A_{\widetilde{s_0}}=A_{\widetilde{s_0}}.$$
Since $(X,S)$ is an association scheme, for $s_1 \in S\setminus\{1_X\}$, $A_{s_0}A_{s_1}=\sum_{s\in S}c^s_{s_0 s_1}A_s$ and $A_{s_1}A_{s_0}=\sum_{s\in S}c^s_{s_1 s_0}A_s$, where $c_{s_0 s_1}^s$'s and $c_{s_1 s_0}^s$'s are intersection numbers of $(X,S)$.
For $x_{ab}, y_{cd}\in \widetilde{X}$, let us consider the set $$Z:=\{z_{ef}\in \widetilde{X}\mid (x_{ab}, z_{ef}) \in \widetilde{s_0} \ \mbox{and} \ (z_{ef}, y_{cd}) \in \widetilde{s_1}\}.$$
Then for $z_{ef}\in Z$, $a=e$ and $(x,z)\in s_0$ as $(x_{ab}, z_{ef}) \in \widetilde{s_0}$, and $e=c$ and $(z,y)\in s_1$ as $(z_{ef}, y_{cd}) \in \widetilde{s_1}$. And we can check the following facts:
\begin{enumerate}
\item If $(x_{ab}, y_{cd})\in 1_{\widetilde{X}}\cup \widetilde{t}$ then $x=y$ and $a=c$. So, there are $c_{s_0 s_1}^{1_X}$ choices of $z$, one choice of $e$ and two choices of $f$. Thus, $| Z | =2\cdot c_{s_0 s_1}^{1_X}$;
\item If $(x_{ab},y_{cd})\in \widetilde{s}$ for some $s\in S\setminus\{1_X\}$, then $(x,y)\in s$ and $a=c$. So, there are $c_{s_0 s_1}^{s}$ choices of $z$, one choice of $e$ and two choices of $f$. Thus, $| Z | =2\cdot c_{s_0 s_1}^{s}$;
\item If $(x_{ab}, y_{cd})\in r_H^1\cup r_H^{-1}$, then $a\neq c$. Since $a=e=c$ for any $z_{ef}\in Z$, $| Z |=0$.
\end{enumerate}
Combining (i), (ii) and (iii), we obtain $$A_{\widetilde{s_0}}A_{\widetilde{s_1}}=\sum_{s\in S\setminus \{1_X\}}2 \cdot c^{s}_{s_0 s_1}A_{\widetilde{s}}+2\cdot c_{s_0 s_1}^{1_X}(A_{1_{\widetilde{X}}}+A_{\widetilde{t}}).$$
And by symmetric argument, we can show that $$A_{\widetilde{s_1}}A_{\widetilde{s_0}}=\sum_{s\in S\setminus\{1_X\}}2 \cdot c^{s}_{s_1 s_0}A_{\widetilde{s}}+2\cdot c_{s_1 s_0}^{1_X}(A_{1_{\widetilde{X}}}+A_{\widetilde{t}}).$$
Now again, for $x_{ab}, y_{cd}\in \widetilde{X}$, let us consider the set $$Z':=\{ z_{ef}\in \widetilde{X}\mid (x_{ab}, z_{ef}) \in \widetilde{s_0} \ \mbox{and} \ (z_{ef}, y_{cd}) \in r_{H}^1\}.$$
Then $a=e$ and $(x,z)\in s_0$ as $(x_{ab}, z_{ef}) \in \widetilde{s_0}$, and $e\neq c$ and $(H^{T(e)})_{zy}=(-1)^{f+d}$ as $(z_{ef}, y_{cd}) \in r_{H}^1$. And we can check the following facts:
\begin{enumerate}
\item If $(x_{ab}, y_{cd})\in 1_{\widetilde{X}}\cup \widetilde{t}\cup S^\sqcup$, then $a=c$. Since $a=e\neq c$ for any $z_{ef}\in Z'$, $|Z'| = 0$;
\item If $(x_{ab}, y_{cd})\in r_H^1\cup r_H^{-1}$, then $a\neq c$ and so, there is one choice of $e$. Since $(x,z)\in s_0$, there are $n_{s_0}$ choices of $z$. And for fixed $d, e, y$ and $z$, there is one choice of $f$ as $(H^{T(e)})_{zy}=(-1)^{f+d}$. Thus, $|Z'| = n_{s_0}$.
\end{enumerate}
Combining (i) and (ii), we obtain $$A_{\widetilde{s_0}}A_{r_H^{1}}=n_{s_0}(A_{r_H^1} +A_{r_H^{-1}}).$$
By symmetric argument, we can show that $$A_{r_H^1}A_{\widetilde{s_0}}=n_{s_0}(A_{r_H^1} +A_{r_H^{-1}}).$$
Since $ r^{-1}_H = \{(x_{ab}, y_{cd}) \mid (1- \delta_{ac})(H^{T(a)})_{xy} = (-1)^{b+d+1}, x, y \in X, a, b, c, d \in \mathbb{F}_2 \}$,
a similar argument yields $$A_{\widetilde{s_0}}A_{r_H^{-1}}=A_{r_H^{-1}}A_{\widetilde{s_0}}=n_{s_0}(A_{r_H^1} +A_{r_H^{-1}}).$$
This completes the proof of Theorem \ref{thm:main1}.
\qed
\begin{lem}\label{lem:auto}
For a Hadamard matrix $H$ and an association scheme $(X,S)$,
let $(\widetilde{X}, S(H))$ be the association scheme constructed by $(\ref{M})$.
Then we have the following:
\begin{enumerate}
\item $\phi : x_{ab} \mapsto x_{(a+1)b}, h \mapsto \phi(h)$ is an isomorphism from $\widetilde{X} \cup S(H)$ to $\widetilde{X} \cup S(H^T)$;
\item For $y \in X$, the transposition $\alpha_y = (y_{ab}~ y_{a(b+a+1)}) \in Sym(\widetilde{X})$ induces an isomorphism $\phi$
from $\widetilde{X} \cup S(H)$ to $\widetilde{X} \cup S(H_1)$ defined by $\phi|_{\widetilde{X}}=\alpha_y , \phi|_{S(H)} : h \mapsto \phi(h)$;
\item For $y \in X$, the transposition $\beta_y = (y_{ab}~ y_{a(b+a)}) \in Sym(\widetilde{X})$ induces an isomorphism
from $\widetilde{X} \cup S(H)$ to $\widetilde{X} \cup S(H_2)$ defined by $\phi|_{\widetilde{X}}=\beta_y , \phi|_{S(H)} : h \mapsto \phi(h)$;
\item $\phi : x_{ab} \mapsto x_{a(b+a+1)}, h \mapsto \phi(h)$ is an isomorphism from $\widetilde{X} \cup S(H)$ to $\widetilde{X} \cup S(-H)$;
\item $\phi : x_{ab} \mapsto x_{a(b+a)}, h \mapsto \phi(h)$ is an isomorphism from $\widetilde{X} \cup S(H)$ to $\widetilde{X} \cup S(-H)$,
\end{enumerate}
where $\phi(h):= \{ (\phi(x), \phi(y)) \mid (x, y) \in h \}$ for $h \in S(H)$, $H_1 = D_y H$ and $H_2 = H D_y$.
\end{lem}
\begin{proof}
In the cases (i), (ii) and (iii), it is easy to see that $\phi$ is bijective on $\widetilde{X}$, $\phi(1_{\widetilde{X}}) = 1_{\widetilde{X}}$, $\phi(\widetilde{t}) = \widetilde{t}$ and $\phi(\widetilde{s}) = \widetilde{s}$ for $s \in S \setminus \{ 1_X \}$.
\begin{enumerate}
\item
Let $(x_{ab}, y_{cd}) \in r^{1}_H$.
Then $(\phi(x_{ab}), \phi(y_{cd})) = (x_{(a+1)b}, y_{(c+1)d})$.
Since $H^{T(a)} = (H^T)^{T(a+1)}$, we have $(\phi(x_{ab}), \phi(y_{cd})) \in r^{1}_{H^T}$.
This means that $\phi(r^{1}_H) = r^{1}_{H^T}$.
\item
Let $(y_{ab}, z_{cd}) \in r^{1}_H$.
When $a=0$, we have $(\alpha_y(y_{ab}), \alpha_y(z_{cd})) = (y_{0(b+1)}, z_{cd})$ and
\[ (y_{0(b+1)}, z_{cd}) \in r^{1}_{H_1} \Leftrightarrow (1 - \delta_{0c})((H_1)^{T(0)})_{yz} = (-1)^{b+d+1}.\]
Since $(1 - \delta_{0c})((H_1)^{T(0)})_{yz} = (-1)(1 - \delta_{0c})(H^{T(0)})_{yz} = (-1)(-1)^{b+d}$,
we have $(\alpha_y(y_{ab}), \alpha_y(z_{cd})) \in r^{1}_{H_1}$.
This implies that $\phi(r^{1}_H) = r^{1}_{H_1}$.
\item
Let $(y_{ab}, z_{cd}) \in r^{1}_H$.
When $a=1$, we have $(\beta_y(y_{ab}), \beta_y(z_{cd})) = (y_{1(b+1)}, z_{cd})$
Similarly to (ii), we can show $(\beta_y(y_{ab}), \beta_y(z_{cd})) \in r^{1}_{H_2}$.
This means that $\phi(r^{1}_H) = r^{1}_{H_2}$.
\item
Since $\phi$ represents $\Pi_{x \in X} \alpha_x$ as a permutation,
(ii) implies that $\phi : x_{ab} \mapsto x_{a(b+a+1)}$ is an isomorphism $\widetilde{X} \cup S(H) \rightarrow \widetilde{X} \cup S(-H)$.
\item
Since $\phi$ represents $\Pi_{x \in X} \beta_x$ as a permutation,
(iii) implies that $\phi : x_{ab} \mapsto x_{a(b+a)}$ is an isomorphism $\widetilde{X} \cup S(H) \rightarrow \widetilde{X} \cup S(-H)$.
\end{enumerate}
\end{proof}
\begin{flushleft}
\textbf{Proof of Theorem~\ref{thm:main2}}
\end{flushleft}
For convenience, we denote $\{ x_{0b} \mid b \in \mathbb{F}_2, x \in X \}$ by $X_0$
and $\{ x_{1b} \mid b \in \mathbb{F}_2, x \in X \}$ by $X_1$.
\vskip10pt
$(\Rightarrow):$
Assume $(\widetilde{X}, S(H_1)) \simeq (\widetilde{X}, S(H_2))$.
We denote by $\phi$ an isomorphism from $(\widetilde{X}, S(H_1))$ to $(\widetilde{X}, S(H_2))$.
Since $\Or(S(H_1)) = \{1_{\widetilde{X}}, \widetilde{t}\}\cup \{\widetilde{s}\mid s \in S\setminus\{1_{X}\}\}$
and $\phi(\Or(S(H_1))) = \Or(S(H_2))$, $\phi (X_0)$ is either $X_0$ or $X_1$.
By Lemma \ref{lem:auto}(i), there exists an isomorphism from $(\widetilde{X}, S(H_2))$ to $(\widetilde{X}, S(H_2^T))$.
So, we may assume $\phi (X_0) = X_0$.
Since $\phi(\Or(S(H_1))) = \Or(S(H_2))$ and $r_{H_1}^1 \in S(H_1) \setminus \Or(S(H_1))$, $\phi (r_{H_1}^1)$ is either $r_{H_2}^1$ or $r_{H_2}^{-1}$.
By Lemma \ref{lem:auto}(iv), there exists an isomorphism from $(\widetilde{X}, S(H_2))$ to $(\widetilde{X}, S(-H_2))$.
So, we also assume $\phi (r_{H_1}^1) = r_{H_2}^1 ~\text{and}~ \phi (r_{H_1}^{-1}) = r_{H_2}^{-1}$.
Using the fact that $\phi(x_{ab}) = y_{ad}$ for some $y_{ad} \in \widetilde{X}$,
we define $\sigma_a \in Sym(X)$ and $\tau_a : X \rightarrow \mathbb{F}_2$ such that
$\phi(x_{ab}) = \sigma_a(x)_{a (b + \tau_a(x))}$ for each $x_{ab} \in \widetilde{X}$.
Now we check that $\tau_a(x)$ is well defined.
It suffices to show that $\tau_a(x)$ does not depend on $b$ of $x_{ab}$.
If $\phi(x_{ab}) = y_{ad}$ and $\phi(x_{ab'}) = y_{ad'}$,
then $y_{ad} = \sigma_a(x)_{a (b + \tau_a(x))}$ and $y_{ad'} = \sigma_a(x)_{a (b' + \tau_a(x))}$.
Since $\phi$ is well defined, we have $d \neq d'$ for $b \neq b'$.
This means that $\tau_a(x) = b + d = b' + d'$.
We consider two matrices as follows.
\[ P_a := P_{\sigma_a} ~\text{and}~ Q_a := diag((-1)^{\tau_a(x)} \mid x \in X). \]
\vskip10pt
First, we claim that $(Q_0 P_0)^{-1} H_1 Q_1 P_1$ is similar to $H_1$.
For each $a \in \mathbb{F}_2$, clearly $(Q_a, P_a) \in \mathrm{D} \times \mathrm{I}$.
In order to show $(Q_0 P_0)^{-1} Q_1 P_1 \in \mathrm{D}\rtimes \mathcal{P}(\mathrm{Aut}(X, S))$,
it suffices to check $P_0^{-1} P_1 \in \mathcal{P}(\mathrm{Aut}(X, S))$, since $\mathrm{D}$ is normal in $\mathrm{D}\rtimes \mathcal{P}(\mathrm{Aut}(X, S))$.
Let $(x_{0b}, y_{0c}), (x_{1b}, y_{1c}) \in \widetilde{s}$.
Then $(\sigma_0(x)_{0(b+\tau_0(x))}, \sigma_0(y)_{0(b+\tau_0(y))})$, $(\sigma_1(x)_{1(b+\tau_1(x))},
\sigma_1(y)_{1(b+\tau_1(y))})$ $\in \phi(\widetilde{s}) = \widetilde{s_1}$ for some $s_1 \in S$.
By the definition of $s_1$, we have $(\sigma_0(x), \sigma_0(y)), (\sigma_1(x), \sigma_1(y)) \in s_1$.
This implies $P_0^{-1} P_1 \in \mathcal{P}(\mathrm{Aut}(X, S))$.
\vskip10pt
Next, we claim that $(\widetilde{X}, \phi(S(H_1))) = (\widetilde{X}, S((Q_0 P_0)^{-1} H_1 Q_1 P_1))$, i.e., $H_2 = (Q_0 P_0)^{-1} H_1 Q_1 P_1$.
It suffices to verify that $(x', y')$-entry of $H_2$ is equal to that of $(Q_0 P_0)^{-1} H_1 Q_1 P_1$.
Let $(x_{ab}, y_{cd}) \in r_{H_1}^1$.
Then $(x'_{ab'}, y'_{cd'}) \in r_{H_2}^1$, where $\phi(x_{ab})= x'_{ab'}$ and $\phi(y_{cd})= y'_{cd'}$.
If $a=0$, then $(H_1)_{xy} = (-1)^{b + d}$ and $(H_2)_{x'y'} = (-1)^{b' + d'}$.
Since
\[(P_0^{-1} Q_0^{-1} H_1 Q_1 P_1)_{x'y'} = (Q_0^{-1} H_1 Q_1)_{xy} = (-1)^{\tau_0(x)} (H_1)_{xy} (-1)^{\tau_1(y)},\]
\[b + \tau_0(x) = b' ~~\text{and}~~ d + \tau_1(y) = d',\]
we have
\begin{eqnarray*}
(P_0^{-1} Q_0^{-1} H_1 Q_1 P_1)_{x'y'} &=& (Q_0^{-1} H_1 Q_1)_{xy} \\
&=& (-1)^{\tau_0(x)} (H_1)_{xy} (-1)^{\tau_1(y)} \\
&=& (-1)^{b' + d'} = (H_2)_{x'y'}.
\end{eqnarray*}
If $a=1$, then $(H_1^{T})_{xy} = (-1)^{b + d}$ and $(H_2^{T})_{x'y'} = (-1)^{b' + d'}$.
Since
\[(P_1^{T} Q_1^{T} H_1^{T} (Q_0^{-1})^{T} (P_0^{-1})^{T} )_{x'y'} = (Q_1^{T} H_1^{T} (Q_0^{-1})^{T})_{xy} = (-1)^{\tau_0(y)} (H_1)_{yx} (-1)^{\tau_1(x)},\]
\[b + \tau_1(x) = b' ~~\text{and}~~ d + \tau_0(y) = d',\]
we have
\begin{eqnarray*}
(P_1^{T} Q_1^{T} H_1^{T} (Q_0^{-1})^{T} (P_0^{-1})^{T} )_{x'y'} &=& (P_1^{-1} Q_1^{T} H_1^{T} (Q_0^{-1})^{T} P_0)_{x'y'} \\
&=& (Q_1^{T} H_1^{T} (Q_0^{-1})^{T})_{xy} \\
&=& (-1)^{\tau_0(y)} (H_1)_{yx} (-1)^{\tau_1(x)} \\
&=& (-1)^{b' + d'} = (H_2^{T})_{x'y'}.
\end{eqnarray*}
\vskip10pt
$(\Leftarrow):$
Assume that $H_1$ is similar to $H_2$ with respect to $(X,S)$.
Since $(\widetilde{X}, S(H_2))$ is isomorphic to $(\widetilde{X}, S(H_2^T))$,
it suffices to consider the case $H_2 = P'P H_1 QQ'$, where $P, Q \in \mathrm{I}$ and $P', Q' \in \mathrm{D}$.
Then $P'P$ can be decomposed into $D_{x_1} D_{x_2} \cdots D_{x_m} P_{\sigma_0}$, where $P = P_{\sigma_0}$ and $D_{x_i} \in \mathrm{D}$ $(i=1, \dotsc, m)$.
According to the fact that $D_{x_i}$ has $-1$ at the entry corresponding to $x_i \in X$,
we define a transposition $\phi_i := ((x_i)_{0b} (x_i)_{0(b+1)}) \in Sym(X_0)$ $(i=2, \dotsc, m)$.
We define a permutation $\phi_{\sigma_0}$ in $Sym(X_0)$ by $\phi_{\sigma_0} (x_{0b}) = (\sigma_0(x))_{0b}$.
Put $\phi := \phi_1 \cdots \phi_m \phi_{\sigma_0}$.
\vskip10pt
Also, $QQ'$ can be decomposed into $Q_{\sigma_1} D_{y_1} \cdots D_{y_l}$, where $Q = Q_{\sigma_1}$ and $D_{y_i} \in \mathrm{D}$ $(i=1, \dotsc, l)$.
Similarly, we define $\psi_i$ $(i=1, \dotsc, l)$ and $\psi_{\sigma_1}$ in $Sym(X_1)$ as the above $\phi_i$ and $\phi_{\sigma_0}$.
Put $\psi := \psi_1 \cdots \psi_l \psi_{\sigma_1}$.
\vskip10pt
We claim that $\phi \cup \psi$ is an isomorphism from $(\widetilde{X}, S(H_1))$ to $(\widetilde{X}, S(H_2))$.
It is easy to see that $\phi \cup \psi$ is bijective on $\widetilde{X}$, $(\phi \cup \psi)(1_{\widetilde{X}}) = 1_{\widetilde{X}}$ and
$(\phi \cup \psi)(\tilde{t}) = \tilde{t}$.
Since $\sigma_0 \sigma_1^{-1} \in \mathrm{Aut}(X, S)$,
for each $\tilde{s} \in S(H_1) \setminus \{ 1_{\widetilde{X}}, \tilde{t}, r_{H_1}^\epsilon \}$,
we have $(\phi \cup \psi)(\tilde{s}) = (\phi_1 \cdots \phi_m \cup \psi_1 \cdots \psi_l)(\tilde{s'})$ for some $\tilde{s'} \in S(H_1)$.
Since $\phi_i$ and $\psi_j$ $(1 \leq i \leq m, 1 \leq j \leq l)$ preserve $\tilde{s'}$, we have $(\phi \cup \psi)(\tilde{s}) \in S(H_2)$.
We can check the following facts:
\begin{enumerate}
\item $\phi_i$ $(i = 1, \dotsc, m)$ and $\phi_{\sigma_0}$ correspond to multiplying $-1$ to only one row of $H_1$ and permuting rows of $H_1$, respectively;
\item $\psi_i$ $(i = 1, \dotsc, l)$ and $\psi_{\sigma_1}$ correspond to multiplying $-1$ to only one column of $H_1$ and permuting columns of $H_1$,
respectively.
\end{enumerate}
This implies $\phi (r_{H_1}^\epsilon) = r_{H_2}^\epsilon$.
Therefore, $\phi \cup \psi$ is an isomorphism from $(\widetilde{X}, S(H_1))$ to $(\widetilde{X}, S(H_2))$.
\qed
\begin{flushleft}
\textbf{Proof of Theorem~\ref{thm:main3}}
\end{flushleft}
We define an action of $GL_n(\mathbb{C}) \times GL_n(\mathbb{C})$ on $Mat_X(\mathbb{C})$ by $(H, (P, Q)) \mapsto P^{-1}HQ$.
Restricting our attention to the following subgroups of $GL_n(\mathbb{C}) \times GL_n(\mathbb{C})$ and a subset of $Mat_X(\mathbb{C})$,
we observe their orbits.
Let
\begin{equation}\label{B}
G = \{ (P, Q) \mid P, Q \in \mathrm{D}\rtimes \mathcal{P}(Sym(X)) \},
\end{equation}
\begin{equation}\label{C}
K = \{ (P, Q) \mid P, Q \in \mathrm{D}\rtimes \mathrm{I}, PQ^{-1} \in \mathrm{D}\rtimes \mathcal{P}(\mathrm{Aut}(X,S)) \},
\end{equation}
\begin{equation}\label{D}
G_{x_0} = \{ (P, Q) \mid P, Q \in \mathrm{D}\rtimes \mathcal{P}(Sym(X)_{x_0}), \}
\end{equation}
where $Sym(X)_{x_0} = \{ \sigma \in Sym(X) \mid \sigma(x_0) = x_0 \}$.
Let $\mathcal{H}$ be the set of Hadamard matrices of order $n$.
We consider an orbit of $G$ acting on $\mathcal{H}$.
Then the orbit $H_0^G$ is the set of Hadamard matrices which are equivalent to $H_0$, and
decomposed into
\[ \bigcup_{i=1}^r H_i^K,\]
where $H_i \in H_0^G$.
In particular, each orbit of $K$-action on $H_0^G$ is a subset of a similarity class with respect to $(X,S)$,
since for each $1 \leq i \leq r$, the transpose of $H_i$ may not belong to $H_i^{K}$.
By Theorem \ref{thm:main2}, the number of isomorphism classes of association schemes in $\{ (\widetilde{X}, S(H)) \mid H ~\text{is equivalent to}~ H_0 \}$ is at least $\frac{r}{2}$.
Now we give a lower bound for $r$.
Since each orbit $H_i^K$ contains a normalized Hadamard matrix,
without loss of generality, we may assume that $H_0, H_1, \dotsc, H_r$ are normalized Hadamard matrices.
By the orbit-stabilizer property (see \cite[page 57]{rotman}), we have
\[\frac{|G_{x_0}|}{|(G_{x_0})_{H_0}|} = |H_0^{G_{x_0}}| \leq |H_0^G|,\]
where $(G_{x_0})_{H_0}$ is the stabilizer of $H_0$.
\vskip10pt
\textbf{Claim 1}: \[\frac{|G_{x_0}|}{|(G_{x_0})_{H_0}|} = \frac{(n-1)! (n-1)!}{|\mathrm{Aut}_{x_0}(H_0)|} \frac{|\mathrm{D}|^2}{2}.\]
First of all, we verify $(G_{x_0})_{H_0} = \{ \pm(I_n, I_n) \} \mathcal{P}(\mathrm{Aut}_{x_0}(H_0))$.
Every element of $G_{x_0}$ is decomposed into $(P_1P_2, Q_1Q_2)$, where $P_1, Q_1 \in \mathcal{P}(Sym(X)_{x_0})$ and $P_2, Q_2 \in \mathrm{D}$.
We consider all $(P_1P_2, Q_1Q_2)$ such that
\begin{equation}\label{E}
P_2^{-1}P_1^{-1}H_0Q_1Q_2 = H_0.
\end{equation}
Since $H_0$ is normalized, $P_1^{-1}H_0Q_1$ is also normalized.
So, (\ref{E}) implies that $(P_2, Q_2)$ is either $(I_n, I_n)$ or $(-I_n, -I_n)$.
Whichever the case may be, $(P_1,Q_1)$ must satisfy $P_1^{-1}H_0Q_1 = H_0$.
This implies $(P_1, Q_1) \in \mathcal{P}(\mathrm{Aut}_{x_0}(H_0))$.
Thus, $|(G_{x_0})_{H_0}| = 2 \cdot |\mathrm{Aut}_{x_0}(H_0)|$. This completes the proof of Claim 1.
\vskip15pt
\textbf{Claim 2}: \[|K| = |\mathrm{Iso}(X, S)| \cdot |\mathrm{Aut}(X, S)| \cdot |\mathrm{D}|^2.\]
Since $\mathrm{Aut}(X, S)$ is a normal subgroup of $\mathrm{Iso}(X, S)$: cf. \cite[page 3]{mp1},
it is easy to see that $\mathrm{D}\rtimes \mathcal{P}(\mathrm{Aut}(X,S))$ is a normal subgroup of $\mathrm{D}\rtimes \mathrm{I}$.
The following is left as an exercise for the reader.
The group $K$ is a subgroup of $(\mathrm{D}\rtimes \mathrm{I}) \times (\mathrm{D}\rtimes \mathrm{I})$ and
its order is $|\mathrm{D}\rtimes \mathcal{P}(Sym(X))| \cdot |\mathrm{D}\rtimes \mathcal{P}(\mathrm{Aut}(X,S))|$.
\vskip15pt
Applying the orbit-stabilizer property for $r$ orbits of $H_0^G = \bigcup_{i=1}^r H_i^K$,
we get $|H_i^K| = \frac{|K|}{|K_{H_i}|}$ $(1 \leq i \leq r)$.
Since $\{ \pm(I_n, I_n) \}$ is a subgroup of $K_{H_i}$, we have
\[\frac{|K|}{|K_{H_i}|} \leq \frac{|K|}{2}.\]
This and Claim $1$ imply
\[\frac{(n-1)! (n-1)!}{|\mathrm{Aut}_{x_0}(H_0)|} \frac{|\mathrm{D}|^2}{2} \leq |H_0^{G_{x_0}}| \leq |H_0^G|=|\bigcup_{i=1}^r H_i^K| \leq \frac{|K|}{2}r.\]
By Claim $2$, we have
\[\frac{(n-1)! (n-1)!}{|\mathrm{Aut}_{x_0} (H_0)| |\mathrm{Aut}(X,S)| |\mathrm{Iso}(X,S)|} \leq r.\]
This completes the proof of Theorem \ref{thm:main3}
\qed
\section{Tables for isomorphism classes}\label{sec:table}
We obtain Table \ref{4} and Table \ref{16} using GAP. In these tables, $AS(n,m)$ means that the association scheme of order $n$ labeled by \verb+ #No. m + in web-site \cite{hanakimi}, $(\ast)$ means the number of similarity classes of Hadamard matrices with respect to the $(X,S)$ and $(\ast\ast)$ means the lower bound given in Theorem \ref{thm:main3}. By \cite[page 277, 1.49]{cd} and \cite{KT}, Table \ref{hm} in this paper, we deal with only one Hadamard matrix up to equivalence for the cases $n=4, 8$.
\begin{table}[h]
\begin{center}
\begin{tabular}{|c||c|c|c|c|c|c|c|c|c|c|}\hline
$n$ & 4&8&12&16&20&24&28&32&36&40 \\ \hline\hline
\verb+#+ &1&1&1&5&3&60&487&$13710027$&$>15000000$&$>366000000000$ \\ \hline
\end{tabular}
\caption{The number of equivalence classes of Hadamard matrices of order $n$}\label{hm}
\end{center}
\end{table}
\begin{flushleft}
\textbf{Isomorphism classes of association schemes induced by Hadamard matrices of order 4}
\end{flushleft}
Let $\mathcal{H}$ be the set of all Hadamard matrices of order 4. Then $|\mathcal{H}| = 768$. \\
Put $H_0:=\begin{pmatrix}
1&1&1&1 \\
1&1&-1&-1 \\
1&-1&-1&1 \\
1&-1&1&-1
\end{pmatrix}$,
$H_1:=\begin{pmatrix}
1&1&1&1 \\
1&1&-1&-1 \\
1&-1&1&-1 \\
1&-1&-1&1
\end{pmatrix}$,
$H_2:=\begin{pmatrix}
1&1&1&1 \\
1&-1&-1&1 \\
1&-1&1&-1 \\
1&1&-1&-1
\end{pmatrix}$ and
$H_3:=\begin{pmatrix}
1&1&1&1 \\
1&-1&-1&1 \\
1&1&-1&-1 \\
1&-1&1&-1
\end{pmatrix}$. \\
Then we obtain the following:
\begin{enumerate}
\item $(X,S)=AS(4,1)$ \\
$\mathcal{H}=H_0^{K_1}$ \\
$H_0^{K_1}= H_1^{K_1} = H_2^{K_1} = H_3^{K_1}$ has 768 elements;
\item $(X,S)=AS(4,2)$ \\
$\mathcal{H}=H_0^{K_2}\cupdot H_2^{K_2}$ \\
$H_0^{K_2}= H_1^{K_2}$ has 256 elements and $H_2^{K_2} = H_3^{K_2}$ has 512 elements;
\item $(X,S)=AS(4,3)$ \\
$\mathcal{H}=H_0^{K_3}\cupdot H_1^{K_3}\cupdot H_3^{K_3}$ \\
$H_0^{K_3}= H_2^{K_3}$ has 384 elements, $H_1^{K_3}$ has 128 elements and $H_3^{K_3}$ has 256 elements;
\item $(X,S)=AS(4,4)$ \\
$\mathcal{H}=H_0^{K_4}\cupdot H_2^{K_4}$ \\
$H_0^{K_4}= H_1^{K_4}$ has 256 elements and $H_2^{K_4} = H_3^{K_4}$ has 512 elements,
\end{enumerate}
where $K_i$ is the group given in (\ref{C}) defined by $AS(4,i)$.
Table \ref{4} is calculated according to Theorem \ref{thm:main3} and note that $|\mathrm{Aut}_{x_0}(H_0)|=6$ .
\begin{table}[h]
\begin{center}
\begin{tabular}{|c||c|c|c|c|c|c|}\hline
$(X,S)$&$|\mathrm{Aut}(X,S)|$&$|\mathrm{Iso}(X,S)|$&$(\ast)$& $(\ast\ast)$&$\begin{array}{c} \rm{Number \ of} \\ \rm{non-Schurian}\end{array}$& \\\hline\hline
$AS(4,1)$&24&24&1&1 &0&$AS(16,30)$ \\\hline
$AS(4,2)$&8&8&2& 1&1& $AS(16,54-55)$\\\hline
$AS(4,3)$&4&24&3& 1 &1& $AS(16,77-79)$\\\hline
$AS(4,4)$&4&8&2& 1 &2& $AS(16,89-90)$\\\hline
\end{tabular}
\caption{Isomorphism classes of association schemes induced by Hadamard matrices of order 4}\label{4}
\end{center}
\end{table}
\begin{flushleft}
\textbf{Isomorphism classes of association schemes induced by Hadamard matrices of order 8}
\end{flushleft}
Let $H$ be the Hadamard matrix which is obtained from $PG(2,2)$. Then $|\mathrm{Aut}_{x_0}(H)|=168$ (see \cite[Theorem 4]{kantor}).
Table \ref{16} is calculated according to Theorem \ref{thm:main3}.
\begin{table}[h]
\begin{center}
\begin{tabular}{|c||c|c|c|c|c|c|}\hline
$(X,S)$ & $|\mathrm{Aut}(X,S)|$&$|\mathrm{Iso}(X,S)|$ & $(\ast)$ &$(\ast\ast)$&$\begin{array}{c} \rm{Number \ of} \\ \rm{ non-Schurian}\end{array}$& \\ \hline\hline
$AS(8,1)$ &40320&40320& 1 & 1 &0&$AS(32,53)$
\\ \hline
$AS(8,2)$ &384&384& 17 & 1 &17&$AS(32,4031-4047)$
\\ \hline
$AS(8,3)$ &1152&1152& 6 & 1&6&$AS(32,4117-4122) $
\\ \hline
$AS(8,4)$ &128&128& 56 & 5 &55& $AS(32,4529-4584)$
\\ \hline
$AS(8,5)$ &48&48& 218 & 33&217& $AS(32,4646-4863) $
\\ \hline
$AS(8,6)$ &24&48& 104 & 66&104& $AS(32,5083-5186)$
\\ \hline
$AS(8,7)$ &32&192& 130 &13&129& $AS(32,5473-5602)$
\\ \hline
$AS(8,8)$ &32&64& 143 & 37&143& $AS(32,5745-5887)$
\\ \hline
$AS(8,9)$ &64&384& 37 & 4&37& $AS(32,6068-6104) $
\\ \hline
$AS(8,10)$ &16&32& 337 & 148&337&$AS(32,6105-6441)$
\\ \hline
$AS(8,11)$ &64&128& 60 & 10&59&$AS(32,6884-6943) $
\\ \hline
$AS(8,12)$ &16&32& 247 & 148&247&$AS(32,6944-7190) $
\\ \hline
$AS(8,13)$ &16&64& 377 & 74&376&$AS(32,7785-8161) $
\\ \hline
$AS(8,14)$ &16&64& 319 & 74&319&$AS(32,8598-8916)$
\\ \hline
$AS(8,15)$ &16&64& 286 & 74&286&$AS(32,9288-9573) $
\\ \hline
$AS(8,16)$ &16&64& 179 & 74&179&$AS(32,9892-10070)$
\\ \hline
$C_2 \times C_2\times C_2$&8&1344& 65 &8 &64& $AS(32,11168-11232)$
\\ \hline
$D_4$ &8&64& 441 & 148&441&$ AS(32,11305-11745) $
\\ \hline
$C_4\times C_2$&8&64& 442 &148 &442& $AS(32,12191-12632)$
\\ \hline
$Q_8$ &8&192& 138 & 50&138&$ AS(32,13083-13220) $
\\ \hline
$C_8$ &8&32& 462 & 296&462& $AS(32,13221-13682) $
\\ \hline
\end{tabular}
\caption{Isomorphim classes of association schemes induced by Hadamard matrices of order 8}\label{16}
\end{center}
\end{table}
\begin{flushleft}
\textbf{Isomorphism classes of association schemes induced by Hadamard matrices of order $2^n$ and a cyclic group of order $2^n$}
\end{flushleft}
Let $(X,S)=\mathfrak{I}(C_{2^n})$ (for the definition of $\mathfrak{I}(C_{2^n})$, see \cite{zies} page 177).
Then $\mathrm{Aut}(X,S)=R(C_{2^n})$ and $\mathrm{Iso}(X,S)=R(C_{2^n}) \rtimes \mathrm{Aut}(C_{2^n})$, where $R(C_{2^n})$ is the group
$(\{f_a:C_{2^n}\rightarrow C_{2^n}\mid a\in C_{2^n}\},\circ)$ and $f_a(x)=xa$. Therefore, $|\mathrm{Aut}(X,S)| =2^n$
and $| \mathrm{Iso}(X,S)|= 2^{2n-1}$.
Let $H$ be a Hadamard matrix induced by $PG(2, n-1)$, which is called a Sylvester matrix. Then $|\mathrm{Aut}_{x_0}(H)| = (2^n -2^0)(2^n-2^1)\cdots (2^n-2^{n-1})$.
Therefore, by Theorem \ref{thm:main3}, there are at least $\frac{(2^n-1)! (2^n-1)!}{2^{3n}(2^n -2^0)(2^n-2^1)\cdots (2^n-2^{n-1})}$ isomorphism classes of association schemes which are obtained by $\mathfrak{I}(C_{2^n})$ and $H$.
\section*{Acknowledgement}
The authors would like to thank anonymous referees for their careful reading and valuable comments.
\bibstyle{plain}
|
\section{Introduction}
\label{sec_introduction}
Random cell complexes (defined below) abound at all length scales in the physical sciences. Specifically with regard to materials science, examples include the contact graph of atoms in a metallic glass \cite{2006sheng} and the covalent bonds in an oxide glass \cite{1932zachariasen} at the atomic scale, dense dislocation networks \cite{2009motz} and the bonding of cross-linked polymers \cite{2013smallenburg} at the nanometer scale, and the grain boundary network in a polycrystal \cite{2010rowenhorst} and the cells in a metallic foam \cite{2001banhart} at the micrometer scale. One feature of all of these systems is that they defy characterization by the usual approach used in crystallography, that is, by the identification of a periodic unit and the classification of defects as deviations from periodicity. Nevertheless, some means of characterization is clearly necessary for these systems to be classified and eventually engineered.
Using metallic glass as a specific example, the arrangement of atoms appears to be homogenous and isotropic on average, and from this standpoint is quite simple. Yet, some recent experimental results \cite{2007swallen} indicate that differences in the preparation of samples with the same composition can result in measurably different mechanical properties. This suggests the presence of small variations in the atomic arrangements \cite{2013singh}, though at present the nature of these variations and the means to measure them remain unclear. This paper contends that the situation is similar to that of the characterization of crystalline materials before the advent of crystallography; the analysis of these systems would be vastly simplified by introducing language and concepts detailed enough to provide an accurate description, and yet abstract enough to apply to many different situations.
Numerous attempts have been made to introduce such a language already. The granocentric model \cite{2009clusel,2010corwin} numerically predicts the distributions of local quantities around a sphere in a polydisperse sphere packing. Shell distance is the minimum number of faces that must be crossed to go from the interior of one cell to the interior of a second cell in a cell structure \cite{1996szeto,1996asteA}. Ring statistics consider the lengths of the shortest closed paths through the network of bonds in a disordered covalent glass \cite{1993rino,2010leroux}. Homology theory is a related but more general approach that characterizes holes of arbitrary dimension \cite{2012macpherson,2013kramar,2014nakamura}. The Randi\`{c} index measures the degree of branching in organic molecules by considering the types of edges in the adjacency graph of the atoms \cite{1975randic}. Percolation theory is concerned with connected clusters of occupied vertices or edges in a graph, and particularly with the appearance of a unique infinite cluster \cite{1994stauffer,1996jacobs}. A hyperuniform distribution of points has the property that infinite-wavelength density fluctuations are absent \cite{2003torquato,2011zachary}.
While certainly not exhaustive, this selection from the available literature is intended to show that existing approaches generally have two limitations. First, they are often motivated by and formulated for a specific situation, and cannot be applied generally. Second, to the extent of our knowledge, none of them offers a complete characterization of the local topology of a physical system. That is, they do not allow the local structure to be reconstructed exactly up to a geometric deformation.
A cell complex is general in the sense that all of the physical systems described above (and many others) can be represented by means of one, and hence is used as a common point of departure in the following. A cell complex is composed of cells, where a $0$-cell corresponds to a point, a $1$-cell to a line segment, a $2$-cell to a disk, and a $3$-cell to a ball. A cell complex is constructed by placing the necessary $0$-cells, attaching the endpoints of deformed $1$-cells to the $0$-cells, attaching the boundaries of deformed $2$-cells to the $1$-cells, and attaching the surfaces of deformed $3$-cells to the $2$-cells. The cells are attached without any twisting or tearing. This allows the description of, e.g., fused quartz with the silicon atoms as $0$-cells and the oxygen atoms as $1$-cells, soap foams with junctions as $0$-cells, edges as $1$-cells, and surfaces as $2$-cells, and polycrystals with nodes as $0$-cells, triple lines as $1$-cells, boundaries as $2$-cells, and grains as $3$-cells \footnote{In the mathematical literature, the cell complexes defined here are referred to as regular cell complexes.}.
The purpose of this paper is to suggest that the swatch and the cloth serve as useful alternatives for the description of a cell complex \cite{2012masonB}. A swatch completely characterizes the local topology of a small region of the cell complex, and a cloth indicates the frequencies of the various swatch types occurring in the cell complex. The advantages of this technique are that the description is complete (any question about the local topology of the cell complex can be answered from the cloth), that the description is general (any physical system that can be represented as a cell complex is admissible), and that the description allows the construction of a distance that quantifies the similarity of the statistical topology of two different cell complexes. This distance is useful, e.g., when considering the convergence of numerical studies, when comparing simulated cell complexes with experimental ones, when quantifying the variability of cell complexes generated in a particular way, or when iteratively modifying some cell complex to reach an intended target. For instance, this paper uses the distance to define a notion of convergence to a topological steady state, and to quantify the approach of our numerical simulations to this steady state. This should be contrasted with the usual approach of following several arbitrarily chosen quantities to identify convergence.
\begin{figure}
\center
\subfloat[]{%
\label{triangle_complex}{%
\includegraphics[height=2.2cm]{%
triangle_complex.pdf}}}
\hspace{12pt}
\subfloat[]{%
\label{triangle_graph}{%
\includegraphics[height=2cm]{%
triangle_graph.pdf}}}
\caption{\label{triangle}(a) Representation of a triangle as a cell complex by three $0$-cells, three $1$-cells, and one $2$-cell. (b) Representation of a triangle as a graph with seven vertices, nine edges, and three vertex types.}
\end{figure}
The description of a physical system by a cell complex is further refined by transforming the cell complex into an equivalent adjacency graph, where a graph is composed of a set of vertices connected by edges. Every vertex of the graph corresponds to a cell of the cell complex, and every edge in the graph corresponds to two incident cells in the cell complex whose dimensions differ by one. For the purpose of illustration, the cell complex of a triangle in Figure \ref{triangle_complex} is equivalent to the adjacency graph in Figure \ref{triangle_graph}. Notice that this requires the introduction of three vertex types; circle, square and triangle vertices in the graph correspond to cells of dimension $0$, $1$, and $2$ in the cell complex. We will follow this convention throughout the paper. More generally, the representation of a cell complex by a graph allows the mathematical machinery of graph theory to be used, and the description of swatches in Section \ref{sec_swatches}, of cloths in Section \ref{sec_cloths}, and of cell complex similarity in Section \ref{sec_metric} to be applied with relatively few modifications to graphs that appear in a variety of subjects.
The utility of this approach is illustrated by means of simulations of two different physical systems. The first is a two-dimensional grain boundary network that develops by a process of normal grain growth, as discussed in Section \ref{sec_grain_growth}. This system is modeled by a set of grain boundary edges with three boundary edges meeting at every triple junction, and the same mobility and energy per unit length for every boundary edge. The Turnbull relation \cite{1951turnbull} governs the motion of the boundary edges, and is equivalent to evolution by curvature flow. The explicit form of the equations of motion depends on the properties of the triple junctions. Different formulations are derived for the cases of finite and infinite triple junction mobilities. Simulation results in Section \ref{sec_steady_state} provide evidence of a statistical steady state where all statistical quantities relating to the local topology of the grain boundary network converge, which is independent of the initial conditions and is the same for the two sets of equations. The distance on cell complexes introduced in Section \ref{sec_metric} is instrumental in making meaningful comparison of the grain boundary networks possible.
The second simulated system is a dislocation network \cite{1982hirth} in a material with no surface tractions during the process of recovery. This system is modeled as a set of dislocation edges with three edges meeting at every edge endpoint, and with a constant energy per dislocation line length; that is, all dislocation interactions are neglected and only the self-energy of the dislocations is retained. Evolution occurs by energy minimization with the same kinetics for dislocation glide and climb, resulting in curvature flow and a reduction in total line length. Our belief is that this severe simplification is justified by the need to simulate networks containing millions of edges (where the calculation of the long-range stress fields would be prohibitively expensive) to reduce the statistical error, and by our emphasis on the network topology rather than on the physics of a specific deformation process. This system is also mathematically interesting independent of the physical interpretation, and is discussed in more detail in Section \ref{sec_dislocations}. As with the grain boundary network, there appears to be a statistical steady state where all statistical quantities relating to the local topology of the dislocation network converge, and where the steady state is independent of the initial conditions. Remarkably, the long-term topology of the dislocation network seems to be invariant as to whether dislocations intersect and leave behind an adjoining edge or merely pass through one another. This suggests that in situations where only the steady state configuration is desired, substantial computational savings can be realized without loss of accuracy by not implementing one of the allowed topological events.
\section{Swatches and Local Topology}
\label{sec_swatches}
Given a cell complex, a swatch is the portion of the cell complex in a region around some specified vertex. This is motivated by the observation that random cell complexes are often compared using the frequencies of local configurations of cells. A swatch provides a definition of local configurations that completely describes the local topology and is agnostic to the details of the physical system, and therefore is a suitable basis for the current approach.
The definition of a swatch begins with the selection of a central vertex known as the root. For swatches to be directly comparable, all roots should have the same vertex type; that is, all swatches should be centered on cells of the same dimension in the underlying cell complex. The convention followed throughout this paper is for the roots to be on vertices of the graph corresponding to 0-cells of the cell complex (indicated by circles in the figures), though this is not a general requirement.
After selection of a root, an integer known as the radius of the swatch is chosen to specify the extent of the local configuration. The radius is measured using the canonical graph distance; a swatch of radius $r$ includes all of the vertices that can be reached from the root by traversing no more than $r$ edges of the adjacency graph. Notice that this allows the swatch to contain as much information about the local configuration as is necessary to measure a given property of interest, provided that the property depends only on the local topology.
\begin{figure}
\center
\subfloat[]{%
\label{free_swatch}{%
\includegraphics[height=4cm]{%
FreeSwatch_color.pdf}}}
\hspace{12pt}
\subfloat[]{%
\label{not_free_swatch}{%
\includegraphics[height=3.83cm]{%
NonFreeSwatch_color.pdf}}}
\caption{\label{swatch_examples}Swatches of radius six in a cell complex containing only $0$- and $1$-cells. The vertex color indicates distance from the root, with the root colored dark blue. (a) A free swatch where there is a single path from the root to any given vertex. (b) A swatch that contains a cycle of length four and a cycle of length two.}
\end{figure}
Figure \ref{swatch_examples} provides several examples of swatches constructed from one of the grain boundary networks described in Section \ref{sec_introduction}. The root is colored dark blue, and the color of the surrounding vertices indicates distance from the root. Grain boundaries correspond to square vertices in the graph, and triple junctions correspond to circle vertices in the graph. Physical constraints require square vertices to be of degree two and circle vertices to be of degree three, where the degree of a vertex is the number of connecting edges. The most significant difference between the swatches in Figures \ref{free_swatch} and \ref{not_free_swatch} is related to the presence of cycles, or to closed paths along the edges of the graph. The free swatch in Figure \ref{free_swatch} is distinguished by the absence of cycles, while the swatch in Figure \ref{not_free_swatch} has one cycle of length four and one cycle of length two.
A distinct advantage of defining a swatch using the language of graph theory is the compact computational representation that this affords. Let the vertices of the swatch be labeled by consecutive integers from $1$ to $n$. The vertex types can be stored in an integer array of length $n$, and the edges of the swatch can be stored in an $n \times n$ adjacency matrix where the entry in the $i$th row and $j$th column is $1$ if the $i$th and $j$th vertices share an edge, and is $0$ otherwise. While this is already sufficient to reconstruct the swatch, there remains an issue of uniqueness; rearranging the vertex labels usually results in a different (but equivalent) adjacency matrix for the swatch. This complicates the comparison of two swatches with different roots since the adjacency matrices of the swatches cannot be directly compared without simultaneously considering all permutations of the vertex labels.
This difficulty is resolved by finding a canonical labeling for every swatch. When two canonically-labeled swatches have the same vertex types and adjacency matrices, the swatches describe regions with the same local topology and are said to belong to the same swatch type. Conversely, two canonically-labeled swatches with different vertex types or adjacency matrices must describe regions with different local topologies, and hence belong to different swatch types. This reduces the classification of swatches by swatch type to the problem of finding a canonical labeling for a graph. While an algorithm that performs well in all cases is still not known, the program {\it nauty} is able to find canonical labellings quite efficiently in practice \cite{2014mckay}.
\section{Cloths and Statistical Topology}
\label{sec_cloths}
A swatch provides a complete description of the local topology around a specific root of an adjacency graph, but not of the statistical local topology of the cell complex as a whole. This suggests that a swatch of radius $r$ be constructed around every root, resulting in an ensemble of swatches. The probability that a randomly selected swatch of radius $r$ belongs to a given swatch type is known as the swatch frequency, and the set of all swatch frequencies for all values of the radius is known as the cloth. The cloth characterizes the local topology of the cell complex in the sense of determining the probability of appearance of any local configuration, as well as of prescribing all local topological properties of the cell complex (as defined toward the end of Section \ref{sec_metric}).
The idea of the cloth is related to that of local isomorphism of quasicrystals \cite{1986levine}. Two quasicrystals are considered to be locally isomorphic when they can be made to coincide exactly over an arbitrarily large region by a relative translation, or equivalently, when every atomic arrangement in one occurs somewhere in the other. This is physically significant since two quasicrystals that are locally isomorphic have the same diffraction pattern. However, this description does not allow a meaningful analysis when two quasicrystals are not locally isomorphic. By contrast, a cloth provides the information about the frequencies of every swatch type of every radius. This allows not only the occurrence but the relative rates of local bond topologies to be compared, and the extent to which the quasicrystals fails to be locally isomorphic to be measured.
The cloth is composed of levels, each containing swatch frequencies for the corresponding value of the radius. Notice that level $r$ of the cloth contains strictly more information than all levels ${s < r}$ of the cloth. This follows from the observation that a swatch of radius $r$ can be restricted to a swatch of radius $s$ with the same root (known as a subswatch) by excluding all vertices further than distance $s$ from the root. The swatch types comprising level $r$ of the cloth therefore contain strictly more information than the swatch types comprising level $s$ of the cloth, and the description of the cell complex becomes progressively more complete as the level increases.
\begin{figure}
\includegraphics[width=0.95\columnwidth]{%
TreeOfSwatches.pdf}
\caption{\label{tree_of_swatches}Levels 0 to 3 of the tree of swatch types, subject to the conditions that circle vertices be of degree three, square vertices be of degree two, and the graph be 2-connected.}
\end{figure}
The relationship between different levels of the cloth can be more clearly expressed by means of a tree of swatch types, as in Figure \ref{tree_of_swatches}. Level $r$ of the tree is composed of all swatch types of radius $r$ that are compatible with the constraints of the physical system. Whenever a swatch type of radius ${r - 1}$ is a subswatch of a swatch type of radius $r$, they are connected by an edge. Since there may be more than one swatch type of radius $r$ that satisfies this condition, the tree repeatedly branches with increasing level. Figure \ref{tree_of_swatches} is specialized to the case of the grain boundary network described in Section \ref{sec_introduction} where, apart from the restrictions on the degrees of the vertex types, the adjacency graph satisfies the further condition of being 2-connected \footnote{A graph is $2$-connected when the removal of any vertex does not disconnect the graph.}.
The information in a cloth is equivalent to an assignment of swatch frequencies to each of the swatch types of the tree, subject to the condition that the frequencies on any level sum to one. This description helps to clarify the relationship of swatch frequencies on distinct levels; supposing that $S$ is a swatch type, the frequency of $S$ is equal to the sum of the frequencies of all the descendants of $S$ on any subsequent level. Since a swatch frequency is directly proportional to the number of root vertices around which swatches of the given type occur, this is a direct consequence of $S$ being a subswatch of all of the descendants of $S$.
There are several practical concerns that arise when measuring the swatch frequencies of a finite cell complex. First, the measured swatch frequencies will converge with increasing system size only when the cell complex is statistically homogeneous. That is, any statistical feature measured within a bounded region converges to a definite limit as the region's volume increases, and the limiting value is independent of the region's position within the cell complex \cite{2012masonB}. Second, experience shows that the number of swatch types grows exceedingly quickly as a function of the level number. There may be only a single occurrence of many swatch types for radii where there are more swatch types than roots, introducing substantial sampling errors. Our efforts to reduce these errors led to the adoption of the simplified dislocation network model discussed in Section \ref{sec_dislocations} as a means to increase simulation size, given the available computational resources.
\section{Similarity and Convergence of Cell Complexes}
\label{sec_metric}
Consider a sequence of cell complexes of increasing size whose cloths become ever more similar. For example, one could generate a sequence of steady state configurations of some dynamical process. It is intuitive to think of these cell complexes as approaching an infinite, universal state. This section makes that notion rigorous by introducing a distance on cell complexes and constructing limit objects corresponding to convergent cell complex sequences. The existence of these limit objects provides strong theoretical backing for the experimental results discussed in Sections \ref{sec_steady_state}. The distance can be used for many purposes, including the comparison of cell complexes arising from different processes. Note that the mathematical results in this section require that the cell complexes in the sequence have adjacency graphs with uniformly bounded degree.
The distance on cell complexes will make use of a preliminary distance on swatches. Let the largest common subswatch of two swatches be the swatch of largest radius that is a subswatch of both. The distance between two swatches is defined as the reciprocal of the number of vertices in the largest common subswatch, or zero if the swatches are the same. For example, the largest radius for which the swatch in Figure \ref{not_free_swatch} is free is $r = 3$, and the distance to the free swatch in Figure \ref{free_swatch} is $1/13$.
Having introduced a distance on swatches, the earth mover's distance is used to define a family of distances on cell complexes. The earth mover's distance is equal to the minimum cost of transforming one probability distribution on swatch types into a different probability distribution on the same swatch types. The transformation is performed by transferring probability mass between swatch types, with the overall cost given by the sum of the costs of the individual operations. The cost of an operation is the probability mass transferred times the distance between the two swatch types \cite{1781Monge,1998Rubner}. Given two cell complexes $C_1$ and $C_2$, let $d_r\paren{C_1, C_2}$ equal the earth mover's distance between probability distributions on swatches of radius $r$ induced by the two cell complexes. Note that $d_r$ is uniformly bounded and non-decreasing in $r$, and that it stabilizes for some finite $r$ if the cell complexes are finite. The limit distance on cell complexes is defined as the limit of $d_r$ with increasing $r$, or
\begin{equation}
d\paren{C_1,C_2}=\lim_{r\rightarrow\infty}d_r\paren{C_1,C_2}. \nonumber
\end{equation}
Let $C_1, C_2, \ldots = \left\{ C_i \right\}$ be a sequence of cell complexes that it is a Cauchy sequence in the distance $d$. That is, the elements of the sequence all become arbitrarily close above some sufficiently large $i$. This condition is equivalent to the convergence of all swatch frequencies, and implies the convergence of other important properties as well. A key mathematical result of Benjamini and Schramm \cite{2001benjamini,2012lovasz} is that the sequence $\left\{C_i\right\}$ may be associated with a universal limit object $\sigma$. The limit object is not a cell complex itself, but is instead a probability distribution on the space $\mathcal{C}^\bullet$ of countably infinite, connected cell complexes with a specified root (a root must be specified because swatches are inherently rooted). Sampling from $\sigma$ may be viewed as sampling a random configuration from the universal state that the cell complex sequence approaches. Note that swatch frequencies for any radius $r$, and therefore the distance $d$, may be extended to such distributions: the frequencies are the probabilities that swatches of radius $r$ appear at the root of a random cell complex drawn from $\sigma$. This allows the distance from a finite cell complex to $\sigma$ to be computed, and makes the notion of a sequence of cell complexes converging to the probability distribution $\sigma$ sensible.
The limit distribution $\sigma$ is constructed by assigning probabilities to certain subsets of $\mathcal{C}^\bullet$ defined by swatches. Suppose that $S$ is a swatch of radius $r$, and $E_S$ is the set of all cell complexes in $\mathcal{C}^\bullet$ where $S$ appears at the root vertex. The probability of $E_S$ is then the limiting value of the swatch frequency of $S$ in the sequence $C_i$ as $i\rightarrow \infty$. It is a mathematical theorem \cite{2012lovasz} that this is sufficient to define the probability distribution $\sigma$ on the entire space $\mathcal{C}^\bullet$. By construction, $C_i$ converges to $\sigma$ in the sense that the distance $d$ between $C_i$ and $\sigma$ becomes arbitrarily small for sufficiently large $i$.
The convergence of a cell complex sequence to a limit implies the convergence of all local topological properties of that sequence as well. For example, the expected number of cycles of length four to which an edge belongs will converge. More precisely, let $H$ be the labeled adjacency graph of a square, let $G_i$ be the labeled adjacency graph of the cell complex $C_i$, and let $\text{inj}\paren{H,G_i}$ be the number of times $H$ appears in $G_i$. Although both $v\paren{G_i}$ (the number of vertices of $G_i$) and $\text{inj}\paren{H,G_i}$ will usually diverge with increasing $i$, if $\left\{G_i\right\}$ converges, the normalized quantity $\text{inj}\paren{H,G_i}/v\paren{G_i}$ will converge as well.
More generally, the normalized number of adjacency preserving maps from $H$ to $G_i$ converges for any labeled graph $H$. This may be used to find, e.g., the probability that a 0-cell is adjacent to a specified number of 1-cells (as for the number of contacts around a sphere in an random sphere packing), the probability that a 1-cell is connected to 0-cells of specified degree (as with the Randi\`{c} index \cite{1975randic} of an organic molecule), the joint probability of adjacent 2-cells being incident to specified numbers of 1-cells (as in the Aboav-Weaire relation \cite{1974weaire,1980aboav} for a 2D microstructure), or the probability of a 1-cell participating in a cycle of specified length (as for ring statistics \cite{2010leroux} in an covalent glass). In this sense, the cloth provides a complete description of the local topology of the underlying cell complex, as initially claimed in Section \ref{sec_introduction}.
Consider a dynamical process on random cell complexex, and suppose that many of the properties of the system converge as time proceeds. The steady-state hypothesis is that there is a time interval when all scale-invariant properties of the network are constant in time, though this interval will depend on the initial conditions. To connect this to the formalism established above, construct a sequence of initial conditions of increasing size and allow all of them to evolve to the steady-state condition. By the steady-state hypothesis, the cloths of the systems will be identical up to finite size effects and the cell complex sequence will converge. This implies the existence of a universal limit distribution that may be viewed as a probability distribution of swatch types for an infinite steady state.
In practice, one can track the distance from a cell complex to a reference state as the cell complex evolves. If the steady-state hypothesis holds and the reference is in the steady state condition, then the distance will decrease toward zero and stabilize for a significant interval of time. Hence, the distance provides a powerful tool to test the steady-state hypothesis. In Section \ref{sec_grain_growth}, we use this to study the convergence properties of a model grain boundary network.
Finally, we note that the subject of convergent cell complex sequences is deeper than may be apparent from this section. For instance, the analysis of global graph properties of a convergent cell complex sequence (i.e., those that can be expressed as maps from the adjacency graph $G_n$ into a graph $H$) is much more difficult than that of local topological properties, and not all of them are convergent. An example of a convergent global graph property is the number $q$-colorings of $G_n$ for sufficiently high $q$, i.e., the number of different ways that $q$ colors may be assigned to the vertices of $G_n$ such that no vertices connected by an edge have the same color. A second point is that the swatch frequencies in a cell complex are far from independent---given a swatch of radius $r$, the swatches with roots on the neighboring vertices are determined up to radius $r-1$. This fact is reflected by an important property of the limit probability distribution called involution invariance. Further exposition of these subjects is beyond the scope of this paper; the interested reader is encouraged to refer to the book by Lovasz \cite{2012lovasz}.
\section{Two Models of Grain Growth}
\label{sec_grain_growth}
Microstructure evolution in polycrystalline materials is quite complicated, with the general case involving the precipitation of solid phases, the diffusion of solute species, the formation of dislocation networks, and the interaction of stress fields with all of the above processes. Along with the scarcity of experimental values for many of the relevant material properties, this means that practical simulations of microstructure evolution often require a number of simplifying assumptions.
Perhaps the simplest system with a nontrivial evolution is a pure polycrystalline material with negligible stored strain energy. This system is represented in two dimensions by a space-filling set of grains, with two adjacent grains meeting on a grain boundary and three adjacent grain boundaries meeting on a triple junction. The disruption of the atomic bonding along the boundaries endows them with an energy per unit length, and the minimization of this energy drives the motion of the boundaries and a concomitant increase in the average area of the grains. Hence, grain growth is a result of boundary motion, and boundary motion is usually described by the Turnbull relation \cite{1951turnbull}
\begin{equation}
v_n = m_0 \exp \! \left( -\frac{Q_{gb}}{k_B T} \right) p. \nonumber
\end{equation}
Here $v_n$ is the boundary velocity in the normal direction, $m_0$ is the mobility prefactor, $Q_{gb}$ is the activation energy for boundary motion, and $p$ is the driving pressure. For a pure polycrystalline material with negligible strain energy and a constant boundary energy per unit length $\gamma$ \cite{1992taylor}, the pressure on a boundary is given by the Young--Laplace equation \cite{1805laplace}
\begin{equation}
p = \gamma \kappa,
\end{equation}
where $\kappa$ is the boundary curvature. Further assuming a constant boundary mobility $m$ allows the governing equation to be reduced to
\begin{equation}
v_n = m \gamma \kappa,
\label{eq_curvature}
\end{equation}
or the equation for curvature driven motion. This is the starting point for most simulations of grain growth in two dimensions.
\begin{figure}
\center
\subfloat[]{%
\label{fig_node_continuous}{%
\includegraphics[width=2.5cm]{%
PhysicsContinuous_2.pdf}}}
\hspace{8pt}
\subfloat[]{%
\label{fig_node}{%
\includegraphics[width=2.5cm]{%
NodePhysics_2.pdf}}}
\hspace{8pt}
\subfloat[]{%
\label{fig_vertex}{%
\includegraphics[width=2.5cm]{%
VertexPhysics_2.pdf}}}
\caption{\label{transitions} (a) A continuous grain boundary edge. (b) Node $p$ joins two discrete boundary edges. (c) Vertex $q$ joins three discrete boundary edges.}
\end{figure}
The microstructure is modeled as a network of polygonal curves in two dimensions. Grain boundaries consist of line segments that meet at nodes of degree two, and triple junctions consist of vertices of degree three. The nodes and vertices of the grain boundary network correspond to the node $p$ and the vertex $q$ of Figure \ref{transitions}, respectively. Equations of motion for the nodes are derived by considering the boundary edge in Figure \ref{fig_node_continuous}, and assuming that this edge is a planar curve of length $\Delta s$. If the length is sufficiently small, then the curvature is effectively constant and the edge can be considered as an arc of a circle. Let the angle subtended by the arc be $\Delta \theta$, the point halfway along the arc be $r$, and the normal vector at $r$ be $\nvec{n}$. The force on this edge arises from the line tension $\gamma$ being applied to the segment endpoints along the tangent vectors $\nvec{t}_1$ and $\nvec{t}_2$, and the force per unit length of boundary is $\gamma (\nvec{t}_1+\nvec{t}_2) / \Delta s$. Multiplying the force per unit length by the boundary mobility $m$ gives
\begin{align}
\frac{\Delta \vec{r}}{\Delta t} &= \frac{m \gamma (\nvec{t}_1+\nvec{t}_2)}{\Delta s} = m \gamma \frac{2 \sin(\Delta \theta / 2)}{\Delta s} \nvec{n} \nonumber \\
&\approx m \gamma \kappa \nvec{n} \nonumber
\end{align}
for the velocity of $r$, where the displacement $\Delta \vec{r}$ occurs in a time interval $\Delta t$. The second equality follows from $\nvec{t}_1$ and $\nvec{t}_2$ having equal and opposite components perpendicular to $\nvec{n}$, and components of length $\sin(\Delta \theta / 2)$ parallel to $\nvec{n}$. The approximate equality uses the small angle sine approximation and defines the curvature as $\kappa = \Delta \theta / \Delta s$. This is precisely the equation of curvature driven motion, and suggests that the equation of motion for the nodes of the discrete case can be derived in a similar manner.
The configuration in Figure \ref{fig_node} is the discrete version of the continuous boundary edge in Figure \ref{fig_node_continuous}. Two adjacent segments intersect at the node $p$, and the vectors $\vec{v}_1$ and $\vec{v}_2$ extend from $p$ to the two adjacent nodes. An equitable partition of the grain boundary network assigns half of the segments along $\vec{v}_1$ and $\vec{v}_2$ to $p$, and the remaining half to the adjacent nodes. The node $p$ is therefore associated with a boundary length of $(\norm{\vec{v}_1} + \norm{\vec{v}_2}) / 2$. The force that arises from the line tension $\gamma$ being applied along the vectors $\vec{v}_1$ and $\vec{v}_2$ is $\gamma(\nvec{v}_1+\nvec{v}_2)$, where $\nvec{v}_i$ is the unit vector along $\vec{v}_i$. As before, multiplying the force per unit length by the boundary mobility $m$ gives
\begin{equation}
\frac{\Delta \vec{p}}{ \Delta t} = \frac{2 m \gamma (\nvec{v}_1+\nvec{v}_2)}{\norm{\vec{v}_1} + \norm{\vec{v}_2}}
\label{eq_node}
\end{equation}
for the velocity of $p$, where the displacement $\Delta \vec{p}$ occurs in a time interval $\Delta t$. This is a suitable discrete approximation for curvature driven motion provided that the small angle sine approximation holds. That is, the exterior angle in Figure \ref{fig_node} should be small. Our simulations satisfy this condition by dynamically interpolating the polygonal curves to keep the exterior angle at every node below $\pi/10$.
Notice the absence of an equation of motion for triple junctions in the continuous system. The reason for this is that requiring the boundaries to move by curvature flow does not uniquely specify the behavior of the triple junctions, though any equation of motion must satisfy the following physical constraint. Let $M$ be a mobility of the triple junctions that is distinct in units and value from the mobility $m$ of the boundaries. Any equation of motion should cause the angles between boundary edges in an infinitesimal neighborhood around a triple junction to approach $2\pi/3$ in the limit of high $M.$ Conversely, decreasing $M$ should increase the deviation of the angles from $2\pi/3$ for a triple junction subject to a constant driving force.
A simple equation of motion for the triple junctions that is consistent with the above physical constraint is given by the following line of reasoning. With reference to Figure \ref{fig_vertex}, the force on the triple junction $q$ that arises from the line tension $\gamma$ being applied along the vectors $\vec{w}_1$, $\vec{w}_2$ and $\vec{w}_3$ is written as $\gamma (\nvec{w}_1 + \nvec{w}_2 + \nvec{w}_3)$, where $\nvec{w}_i$ is the unit vector along $\vec{w}_i$. Multiplying this force by the triple junction mobility $M$ gives
\begin{equation}
\frac{\Delta \vec{q}}{\Delta t} = M \gamma (\nvec{w}_1 + \nvec{w}_2 + \nvec{w}_3)
\label{eq_vert}
\end{equation}
for the velocity of $q$, where the displacement $\Delta \vec{q}$ occurs in a time interval $\Delta t$. Since Equation \ref{eq_vert} assigns a finite mobility $M$ to the vertices, Equations \ref{eq_node} and \ref{eq_vert} will be called the finite mobility equations. Note that there is some evidence of finite triple junction mobilities in the literature \cite{2000gottstein,2002gottstein}.
The finite mobility equations are not the only option. Instead, one can require that the grain boundaries meet at angles of $2 \pi / 3$ (called the Herring Angle condition), which provides a reasonable set of boundary conditions for the curvature flow along the edges.~\cite{2001kinderlehrer,2014ilmanen} This can be interpreted as giving the vertices infinite mobility, as they always move to the point which locally minimizes the lengths of the neighboring edges. The Herring Angle condition also implies that the rate of change of the area $A$ of a grain with $n$ bounding triple junctions is given by the von Neumann--Mullins relation \cite{1952vonneumann,1956mullins}
\begin{equation}
\deriv{A}{t} = m \gamma \frac{\pi}{3} ( n - 6 ).
\label{eq_mullins}
\end{equation}
That is, the rate of change of the area depends linearly on the number of bounding triple junctions. While Equation \ref{eq_mullins} is a consequence of Equation \ref{eq_curvature} and the Herring Angle condition, the quantities appearing in Equation \ref{eq_mullins} can be more reliably measured than the normal direction and curvature of a polygonal curve. This motivated the development of an alternate set of equations of motion based on the von Neumann--Mullins relation \cite{2010lazar} that will be called the von Neumann--Mullins equations.
\begin{figure}
\center
\subfloat[]{%
\label{move_1}{%
\includegraphics[height=1.25cm]{%
EdgeFlip_3.pdf}}}
\hspace{24pt}
\subfloat[]{%
\label{move_2}{%
\includegraphics[height=1.25cm]{%
DigonDeletion_2.pdf}}}
\caption{\label{moves} Topological operations that occur in the grain growth simulations include (a) edge flips and (b) digon deletions.}
\end{figure}
Two topological operations are allowed in the simulation, namely, the flip of an boundary edge and the deletion of a digon. A boundary flip occurs whenever the length of a boundary passes below a threshold value and the creation of a degree-four vertex appears imminent, as in Figure \ref{move_1}. The boundary direction is changed and connections with adjacent boundaries are shuffled to minimize the sum of the two angles opposite to the flipped boundary. A digon is deleted whenever the maximum distance between points on the two boundaries passes below a threshold value, as in Figure \ref{move_2}. One of the boundaries is deleted, and the remaining boundary is merged with the two adjacent boundaries. Note that the topological change that occurs when a grain shrinks to a point can be expressed as a combination of the two previous operations.
\begin{figure}
\center
\subfloat[]{%
\label{initial_grain}{%
\includegraphics[width=3.9cm]{%
Voronoi_2D_plain.pdf}}}
\hspace{14pt}
\subfloat[]{%
\label{perturbed_lattice}{%
\includegraphics[width=3.9cm]{%
Perturbed_2D_plain.pdf}}}
\caption{\label{grain_growth}Initial conditions for the grain boundary simulation: (a) a Voronoi tessellation of randomly distributed points and (b) a perturbed lattice.}
\end{figure}
Grain growth simulations were always performed in a square with periodic boundary conditions, with two types of initial conditions. The first is given by the Voronoi diagram of randomly distributed points, and a portion of this initial condition is shown in Figure \ref{initial_grain}. The second is a perturbed honeycomb lattice, and a portion of this is shown in \ref{perturbed_lattice}. It is generated by perturbing the vertices of the dual triangular lattice with displacements independently sampled from a two-dimensional Gaussian distribution with a standard deviation of one-fourth the lattice spacing, then computing the Voronoi diagram.
\section{A Statistical Steady State}
\label{sec_steady_state}
We used one of the cell complex distances to compare the model grain boundary networks to a reference condition throughout the simulations. The reference condition was a network with $3.1 \times 10^6$ boundaries resulting from a simulation that used the von Neumann--Mullins equations, that began from a Voronoi tessellation with $6.0 \times 10^7$ boundaries, and for which all measured scale-free properties had converged. Since the number of swatch types increases dramatically as a function of radius, the sample size required to accurately compute the cloth increases dramatically as well. Practically, $r = 7$ was the largest radius that gave reliable cloth statistics; an appreciable number of swatch types occurred only once for $r = 8$ in all of our simulations, indicating that larger samples would be needed. Nevertheless, there are so many swatch types for this radius that the cloth still offered a very detailed description, and we use the distance $d_r$ with $r = 7$ instead of the limit distance $d$ in the following.
\begin{figure}
\includegraphics[height=5.8cm]{distanceData2D_bars_new.pdf}
\caption{\label{bootstrap}Distances of several simulations to the reference condition. Error bars show the standard deviation of the distance of a steady-state configuration with the indicated number of edges to the reference condition.}
\end{figure}
\begin{figure}
\includegraphics[width=3.9cm]{SteadyState_vonNeumann_plain.pdf}
\caption{\label{mullins}The steady-state condition of the model grain boundary network.}
\end{figure}
Figure \ref{bootstrap} shows the distance to the reference condition for three simulations, one using the finite mobility equations and the other two the von Neumann--Mullins equations. For the latter two, the distance to the reference condition decreased very rapidly as they evolved, indicating convergence toward the steady state depicted in Figure \ref{mullins}. The simulation using the finite mobility equations also approached the reference condition, but not as quickly. The evolution of the systems is parametrized by the number of edges, a non-increasing function of time.
While Figure \ref{bootstrap} indicates that the distance to the reference condition initially decreases, the distance to the reference condition visibly increases for small numbers of remaining edges. This may be explained by the decrease in the sample size increasing the statistical error in the swatch frequencies and the apparent distance to the reference condition. A test of convergence to the steady state should account for this source of error. Let $R_n$ be a set of networks with $n$ edges that are already in the steady state. A measured distance may be compared to the distribution of distances from elements of $R_n$ to the reference condition; if the measured distance falls within one standard deviation of the mean of this distribution for a long time interval, then the network being considered is likely in the steady state. Practically, the set $R_n$ contains random subsamples of the reference condition. A single subsample is constructed from the vertices and edges within some radius of a randomly selected vertex, with vertices on the boundary excluded as necessary to attain the desired number of edges.
This procedure is used to evaluate the convergence of the simulations in Figure \ref{bootstrap}. The simulations began with $3.0 \times 10^7$ boundary edges. The error bars in the figure extend one standard deviation above and below the mean distance from the elements of $R_n$ to the reference condition. Note that the simulations using the von Neumann--Mullins equations (the solid green line and the evenly dashed blue line) are within one standard deviation of the subsamples throughout the interval between $100,000$ and $1000$ edges, and that the same is true for the simulation using the finite mobility equations (the dashed line) between $50,000$ and $1000$ edges. This offers strong evidence that they have converged to the steady state in the indicated intervals. Further evidence in the form of various statistical quantities suggests that the simulations using the von Neumann--Mullins equations converged with as many as $500,000$ edges remaining, though the reference condition does not contain enough edges to allow independent subsamples of that size. This data leads to a conjecture:
\begin{conj}
There exists a unique limit distribution $\sigma$ on the space of all countable, connected one-dimensional cell complexes with a root cell specified such that all generic initial conditions converge to $\sigma$ under the von Neumann--Mullins Equations.
\end{conj}
Of course, a definition of ``generic initial conditions'' is required for this to be a mathematical precise conjecture. This turns out to be a delicate question, and is addressed in a separate paper for a mathematical audience~\cite{2015Schweinhartb}.
It is not obvious \textit{a priori} that the finite mobility equations should approach the same steady state as the von Neumann--Mullins Equations. A possible explanation is that is a consequence of the fact that curvature is not a scale-free property: a circle of half the radius has twice the curvature. Thus, as the system coarsens and the average edge length increases, the average speed of an edge will decrease. In contrast, the finite vertex mobility equations are scale invariant. As a result, the vertices move faster and faster relative to the edges, in effect causing their relative mobility to increase toward infinity. It follows that the triple point angles should approach $2\pi/3,$ and therefore that the system should behave more like the von Neumann--Mullins Equations. This is corroborated by Figure \ref{fig_angles} which shows how the average deviation of the triple point angles from $2\pi/3$ changes in a simulation using the finite vertex mobility equations. The simulation is the same finite mobility system from Figure \ref{bootstrap}.
\begin{figure}
\includegraphics[height=5.8cm]{AngleDeviation.pdf}
\caption{\label{fig_angles} The average deviation of the triple point angles from $2\pi/3$ for a simulation using the finite mobility equations.}
\end{figure}
This reasoning also explains why the convergence is slower for the finite mobility equations, as they only drift toward the infinite mobility case. Still, they provide a good approximation of this behavior in the long term, which will be important for the 3D case where (as far as we know) there is no discrete, physical way to directly simulate a system with infinite vertex mobility. Let us note that a family of different universal conditions can be obtained by rescaling the mobility of the vertices as the simulation evolves to keep their speed relative to the edges constant. The resulting behavior is mathematically interesting, and is discussed in length in B. Schweinhart's thesis \cite{2015schweinhart, 2015Schweinhartb}.
\section{A Model Dislocation Network}
\label{sec_dislocations}
This section applies the concepts of a swatch and a cloth to measure the approach of a model dislocation network to a topological steady state. More specifically, the dislocation network is modeled as a network of polygonal curves in three dimensions. Line segments composing a dislocation meet at nodes of degree two, while dislocations meet at vertices of degree three. The complex calculations required to model the interactions of dislocation edges are neglected, and only the dislocation self energy is retained. That is, a dislocation edge is given a constant energy $\gamma$ per unit length and evolves following a simple line-tension model \cite{1982hirth}. There is evidence in the literature \cite{1967lothe,1967indenbom} that this approach is justified when considering the general characteristics of a complex dislocation network rather than specific dislocation reactions. Since our purpose is to study the statistical topology of a dislocation network instead of the effect of dislocations on material properties, we believe that this is a reasonable simplification. This system is also of intrinsic mathematical interest.
\begin{figure}
\center
\includegraphics[height=1.34cm]{%
Intersection_2.pdf}
\caption{\label{move_3}The third topological operation that occurs in the dislocation simulations is the edge intersection, resulting in the joining of edges.}
\end{figure}
The three topological operations that are allowed in the simulation include the two analogues of the operations in Figure \ref{moves} (an edge flip and a digon deletion) and the edge intersection, shown in Figure \ref{move_3}. An edge intersection occurs whenever two non-neighboring edges meet transversely. The edges are subdivided at the point of intersection and joined by a newly created edge. Since detecting intersections is computationally expensive, the simulation could be made substantially more efficient if this topological operation could be neglected without measurably changing the topological steady state. Using the metric on cloths introduced in Section \ref{sec_metric}, we provide evidence that this is indeed the case.
\begin{figure}
\center
\subfloat[]{%
\label{vor}{%
\includegraphics[width=3.9cm]{%
vor_200KEdges_InitialConds_graphics_v2.pdf}}}
\qquad
\subfloat[]{%
\label{rand}{%
\includegraphics[width=3.9cm]{%
random_Point5Length_100KEdges_InitialConds_graphics_v2.pdf}}}
\caption{\label{initial}Initial conditions of the dislocation network included (a) Voronoi graphs and (b) random graphs, with the constructions given in the text. Note that (b) shows a smaller volume than (a), since edges are much denser in the random graph.}
\end{figure}
The simulations were performed in a cube with periodic boundary conditions. Initial conditions for the simulations were generated from a set of random points by one of two procedures. The first was constructed from the edge set of the Voronoi diagram of the points. Since the vertices of this network had degree four, the network had to be modified to make an admissible dislocation network. Every vertex was replaced by an edge, and the four adjacent edges were assigned to the vertices to minimize the maximum angles opposite to the newly created edge. The resulting initial condition is called a Voronoi graph, and is depicted in Figure \ref{vor}. Ken Brakke's VOR3DSIM program was used to compute Voronoi tessellations \cite{BrakkeVor}. For the second procedure, points of degree less than three were randomly connected by edges with others closer than a threshold distance. If there were no points within the threshold, two cases were considered. If the point had degree two, it was considered to be a node along the edge between its two neighbors. Otherwise, it was paired with the closest possible point. This process proceeded until the creation of edges was no longer possible. The resulting initial condition is called a random graph, and is depicted in Figure \ref{rand}.
Since the model grain boundary network and the model dislocation network both satisfy the same conditions at the nodes and vertices and evolve by curvature-driven motion, Equations \ref{eq_node} and \ref{eq_vert} (the finite mobility equations) can also be used as equations of motion for the model dislocation network.
\begin{figure}
\includegraphics[height=5.8cm]{distanceData3D_2.pdf}
\caption{\label{convergence} Distances of several simulations to the reference condition.}
\end{figure}
\begin{figure}
\includegraphics[width=3.9cm]{SteadyState_200KEdges_InitialConds_graphics_v2.pdf}
\caption{\label{steadystate} A small region in the steady-state condition for the model dislocation network.}
\end{figure}
\begin{figure}
\includegraphics[height=5.8cm]{TopologicalChangeFreqLog2.pdf}
\caption{\label{intersections} Relative rates of topological changes throughout a simulation with a random graph initial condition.}
\end{figure}
To test the convergence of the simulations, we tracked their distance to a reference condition throughout their evolution. The reference condition resulted from a simulation with $\lambda=1667$ starting with a Voronoi graph with slightly more than $10^7$ edges, and has about $10^6$ edges. As discussed in the final paragraph of the previous section, the finite mobility equations are not scale invariant, so a system evolving by them can only drift toward a topological steady state. Thus, we must be careful when selecting a reference condition, and comparing other simulations to it. However, if the mobility and the number of initial edges are large enough, the system should be very close to a steady state condition toward the end of its evolution. Our measurements of several properties support this, and it appears that any deviation from the topological steady state due to the finite value of the mobility is small relative to the statistical variation due to finite size effects.
Figure \ref{convergence} shows the distance to the reference condition for three simulations. For all three, the distance to the reference condition decreased rapidly as they evolved, indicating convergence toward the state depicted in Figure \ref{steadystate}. The evolution of the systems is parametrized by the number of edges in the system, which generally decreases with time but can increase if there are frequent edge intersections. For example, the simulation for one of the random graph initial conditions (shown by the dashed red line in Figure \ref{convergence}) initially experienced many intersections, leading to a transient where the number of edges increased. As shown in Figure \ref{intersections} though, the number of intersections as a fraction of all topological changes declined as the simulation progresses, eventually decreasing to almost none in the long term. This suggests that the long-term behavior may be insensitive to the detection of edge intersections. As further evidence for this, the remaining two simulations in Figure \ref{convergence} did not detect edge intersections, and yet converged toward the same state.
\section{Conclusion}
\label{sec_conclusion}
Although random cell complexes occur throughout the physical sciences, our ability to characterize them appears to have been limited by the absence of a suitable language. This paper proposes that the topology of the cell complex be represented by a graph. A swatch (defined in Section \ref{sec_swatches}) characterizes the local topology of the cell complex around a root vertex, and provides a description of the local environment that is agnostic to the details of the physical system. A cloth (defined in Section \ref{sec_cloths}) is constructed from the probabilities of every swatch type occurring around a randomly selected root vertex, and may be used to answer any question pertaining to the statistical local topology of the cell complex. This includes, e.g., the distribution of the number of contacts around a sphere in a random sphere packing, the distribution of the number of sides in the rings in a covalent glass, and the distribution of the number of faces in the grains of a polycrystal.
For cell complexes that evolve by some dynamical process, a sequence of cell complexes can be constructed at successive points in the evolution. A distance on cell complexes is defined in Section \ref{sec_metric} such that the elements of this sequence become arbitrarily close together if the system evolves toward a steady state. This allows a precise definition of the limiting condition to be given. This was applied to a two-dimensional grain boundary network with uniform boundary energies and mobilities, and obeying one of two different sets of equations of motion. The first assumed an infinite vertex mobility, and is designed \cite{2010lazar} to accurately satisfy the von Neumann--Mullins relation \cite{1952vonneumann,1956mullins}. The second assumed a finite vertex mobility, and is derived in Section \ref{sec_grain_growth} by considering the forces on a discrete boundary edge. As described in Section \ref{sec_steady_state}, the simulations converged to steady states that do not appear to depend on the initial conditions. Furthermore, simulations with both sets of equations of motion converged to the same steady state, though the convergence is slower in the finite vertex mobility case. This shows that the distance on cloths can be used not only to measure the convergence of a simulations to a steady state, but also to quantify the extent of the differences introduced into simulation results by the use of alternative numerical implementations.
Section \ref{sec_dislocations} describes a model dislocation network where the dislocations are endowed with a constant energy per line length and the network evolves by energy minimization. As with the grain boundary network, simulations beginning from different sets of initial conditions converged toward steady states that were identical within sampling errors. Perhaps more significantly, the distance on cloths shows that the approach from the random graph initial condition to the steady state does not depend on the implementation of a separate topological operation when dislocations intersect, as in Figure \ref{move_3}. That is, the implementation of this specific model system can be significantly simplified and computational requirements can be reduced without measurably changing the statistical local topology of the steady state condition.
A distance on cell complexes is expected to be useful more generally as well. Apart from testing for the convergence of simulations and the invariance of the results to implementation details, the distance allows a meaningful comparison of simulations with experimental observations, the quantification of the variability of cell complexes generated in a particular way, and the iterative modification of a cell complex by continually reducing the distance to an intended target. We sincerely hope that this stimulates further research into statistical topology and its applications to materials science and physics.
\begin{acknowledgments}
The authors would like to thank the Institute for Advanced Study where the original idea occurred. B. Schweinhart was supported by a National Science Foundation Graduate Research Fellowship under Grant No. DGE-1148900, and the Center of Mathematical Sciences and Applications at Harvard University.
\end{acknowledgments}
|
\section{Introduction}
Vertex coloring problems (VCP) have been studied extensively since the inception of graph theory.
In classical form, problem of $k$-coloring a graph is stated like this: can we color vertices of a graph using $k$
different colors, so that no neighbouring vertices have the same color? It is known that this problem is NP-complete \cite{vcp1}.
VCPs have received much attention in the literature not only for its theoretical aspects and difficulty from the computational
point of view, but also for its real world applications, for example in: scheduling \cite{vcp2},
timetabling \cite{vcp3}, register allocation \cite{vcp4}, train platforming \cite{vcp5},
frequency assignment \cite{vcp6}, communication networks \cite{vcp7} and many other engineering fields.
In this paper we study a variation of the coloring problem. Using only two colors we want to color the vertices, so that
there is no monochromatic cycle of given length. There have been some research in solving a slightly different problem: is there
a 2-coloring such that there exists no monochromatic cycles (of any length). This problem can be viewed as partitioning a graph into
two induces forests and it is known to be NP-complete \cite{2col1} for directed graphs.
Another result worth mentioning is by Nobinon et al. \cite{2col2} where authors show that this problem is NP-complete
even for oriented graphs. They also give implementation of three exact algorithms and some inapproximability results. The motivation
to study this class of problems lies in economics -- 2-coloring without monochromatic cycles can be used in the study of rationality
of consumption behavior.
Many more papers have been written on subject of acyclic coloring (or partitioning). Papers relevant to ours include (among many others):
\cite{2col3}, \cite{2col4}, \cite{2col5}.
The rest of the paper is organized in the following way: in section 2 we define notation used in this paper, we also give
definitions of studied problems and we state the main theorem. In section 3 and 4 we prove the hardess of our coloring problem
for cycles of small length (3, and 4). Later, in section 5 we generalize the ideas used in previous sections to prove the main theorem.
We end the paper with some conclusions and we show perspectives for future work.
\section{Preliminaries}
The purpose of this section is to introduce reader to notation used in later chapters as
well as definitions of studied problems. Let $G=(V,E)$ be an undirected, unweighted graph.
The {\em cycle} in $G$ is a vertex disjoint, closed, simple path in $G$. We denote $\mathcal{C}_k$ to be a set of
all cycles in $G$ of length $k$.
Let $c: V \rightarrow \{r,b\}$ be a mapping that for each vertex in $V$ assigns one of two colors ({\em red} or {\em blue}).
We will call any such $c$: the {\em coloring} of graph $G$. Furthermore, we will say
that given coloring $c$ is {\em valid}, if a certian predicate $P(c)$ is true. Let $K_n$ be a clique of size $n$, that is: a
graph with $n$ vertices in which every vertex is connected by an edge to any other vertex.
Let \COL{k} be the decision problem of whether there exists a valid 2-coloring for given graph.
We give the validity predicate $P_k(c)$ below. It is true only if the coloring $c$ does not contain any cycles of size $k$ with
vertices of the same color.
\[
P_k(c) \equiv \forall Q \in \mathcal{C}_k \exists u,v \in Q \quad c(u) \neq c(v)
\]
Formally, our problem can be expressed as:
\[
(2,k)\text{-}\mathcal{COL} = \{G \, : \, \exists c \, P_k(c)\}
\]
We are interested in knowing how hard is the question, whether given graph $G$ belongs to \COL{k}.
In the next two sections we study the simplest variants, that is when $k=3$ and $k=4$. Cycle of size three we
call a {\em triangle}, and of size four: a {\em square}.
Let $\mathcal{SAT}$ denote the classical boolean satisfiability problem. Namely, it is the set of all boolean formulas
in CNF (conjunctive normal form)
for which there exists a truth assignment that satisfies it. It is known that this problem is NP-complete \cite{cook}.
It is also known that
a certain variation of $\mathcal{SAT}$ called $\mathcal{NAE}$-$\mathcal{SAT}$ (not-all-equal SAT) is NP-complete \cite{naesat}. In this variation we impose additional
constraint on the satisfing assignment: each clause has at least one literal that is true, and at least one that is false.
We denote $k$-$\mathcal{SAT}$ and $k$-$\mathcal{NAE}$-$\mathcal{SAT}$ (for $k \geq 3$) to be subsets of $\mathcal{SAT}$ and $\mathcal{NAE}$-$\mathcal{SAT}$ where each clause in given formula has at most $k$
literals (it's in kCNF). For $k < 3$ for both problems there exists polynomial time algorithms that solves them.
We are ready to state the main theorem:
\begin{thm}
\label{thm:main}
For any integer $k \geq 3$, \COL{k} is NP-complete.
\end{thm}
In order to prove theorem \ref{thm:main}, we will prove the following theorem:
\begin{thm}
\label{thm:red}
For any integer $k \geq 3$, there exists a computable function $f$, such that
for any boolean formula $\phi$, $\phi \in k$-$\mathcal{NAE}$-$\mathcal{SAT}$ if and only if
$f(\phi) \in$\COL{k}.
\end{thm}
\section{Two-coloring without monochromatic triangles}
In this section we prove theorem \ref{thm:red} for $k=3$. Let $\phi$ be
a boolean formula in 3CNF with $n$ variables $x_1,\ldots,x_n$ and $m$ cluses
$C_1,\ldots,C_m$. We construct desired graph $G_{\phi}$ in the following way.
Let us begin by showing an abstract form of $G_{\phi}$. The reduction consists of three gadgets: one for
each variable, one for each clause, and one for each {\em super-edge}. The super-edge $\{u,v\}$ is an edge with a property, that
any valid coloring $c$ implies that $c(u) \neq c(v)$. For starters, assume that we already have such edges at our disposal.
This is how we would construct $G_{\phi}$: a gadget for variable $x$ consists of two vertices labeled $x$ and $\neg x$ connected by
a super-edge. Gadget for clause $C=(u \vee v \vee w)$ consists of a triangle with vertices labeled $u,v$ and $w$. We connect each literal
from variable gadget to its every occurrence in clause gadgets using super-edges. Example is given in figure \ref{fig:basic_red} for
formula $\phi=(x_1 \vee \neg x_1 \vee x_2) \wedge ( \neg x_2 \vee x_3 \vee \neg x_3)$. Dashed lines represent super-edges. We prove
that this is indeed the correct reduction.
\begin{figure}[H]
\begin{center}
\includegraphics[scale=0.5]{./2col1.pdf}
\end{center}
\caption{Abstract form of graph used in reduction}
\label{fig:basic_red}
\end{figure}
\begin{lma}
For any given $\phi$, graph $G_{\phi}$ has a property, that:
\[
\phi \in 3\text{-}\mathcal{NAE}\text{-}\mathcal{SAT} \iff G_{\phi} \in \mCOL{3}
\]
\end{lma}
\begin{proof}
First we assume that $\phi \in 3$-$\mathcal{NAE}$-$\mathcal{SAT}$ and let $\hat{\sigma}(x_1,\ldots,x_n)$ be the truth assignment that certify it.
Each vertex with non-negated label $x$ in vertex gadgets is colored red if $\sigma(x)=T$ and blue otherwise. Coloring of every other
vertex is forced by super-edges. Notice that the only place where there could be any monochromatic triangle is in some clause gadget.
We cannot make that trinagle using mixture of vertices from other clause gadgets or vertex gadgets because we always have to pass
through a super-edge, hence we change a color of vertices on our path. Now if we assume on the contrary, that some clause
$C=(u \vee v \vee w)$ form a monochromatic triangle, then either $\sigma(u)=\sigma(v)=\sigma(w)=T$
or $\sigma(u)=\sigma(v)=\sigma(w)=F$, which gives a contradiction.
Now let $c$ be the valid coloring of $G_{\phi}$. Since $G_{\phi}$ has no monochromatic triangles, and from the property of super-edge
we simply assign value $T$ for all variables from vertex gadgets that have color red, and $F$ otherwise. This
gives an assignment $\hat{\sigma}(x_1,\ldots,x_n)$ that proves that $\phi \in 3$-$\mathcal{NAE}$-$\mathcal{SAT}$. To see that, observe that
every clause corresponding to clause gadget will have at least one literal that is true, and at least one that is false,
because this clause gadget does not form a monochromatic triangle, which was assumed.
\end{proof}
All we have to do now is construct a gadget for super-edge. Such gadget need to have a property, that some
selected edge $\{x,y\}$ in that gadget will always have $c(x)\neq c(y)$, for any valid coloring $c$ of that gadget
(a valid coloring also has to exist). An example of the gadget is shown in figure \ref{fig:super_edge1}. On the left picture
edge $\{x,y\}$ is pointed out. In the middle we have an example of valid coloring, and on the picture on the right we see how
coloring $\{x,y\}$ in one color gives a contradiction (vertex with a question mark cannot be colored neither red, nor blue).
The existence of this gadget completes the proof of theorem \ref{thm:red} for $k=3$ (and also theorem \ref{thm:main},
with additional observation that our reduction is polynomial with respect to size of $\phi$).
We argue, that even if a super-edge in figure \ref{fig:super_edge1} is enough to verify the
genuineness of theorem \ref{thm:red} (for $k=3$), it is not ellegant.
We give a better construction of the gadget that uses a certain coloring property
of $K_4$. Our method is also easier to generalize for $k > 3$.
The basic observation is that when we color any two vertices of $K_4$ in one selected color -- lets say red -- then
the other two vertices will have to be colored blue (otherwise there would be a monochromatic triangle).
Now if we were to {\em hook} another $K_4$ to those blue vertices (see figure \ref{fig:coloring_property}) then the two
non-colored vertices would have to be red, and so on, and so on. With this we can create {\em strings} of $K_4$-s.
\begin{figure}[H]
\begin{center}
\includegraphics[scale=0.7]{./2col4.pdf}
\end{center}
\caption{The {\em strings} created from joining 4-cliques}
\label{fig:coloring_property}
\end{figure}
The trick is to tie two ends of the string together. This will form a {\em loop}.
It is easy to verify, that loop of length 5 is the desired gadget for super-edge.
We show its properties in figure \ref{fig:super_edge2}. On the left, the edge $\{x,y\}$
is pointed out. In the middle picture we give some valid coloring, and the last picture shows how
coloring $\{x,y\}$ in a single color leads to a contradiction (vertices marked in question marks cannot be colored
without creating a monochromatic triangle). In fact, we can make an easy observation:
\begin{obs}
Any loop of odd length (for lengths greater than 3) can be used as a gadget for super-edge.
\end{obs}
It turns out that loop of length 3 cannot be used, because it is isomorphic to $K_6$ (and therefore is
not colorable). Also, coloring loops of even length would not lead to a contradiction, no matter which edge you choose for $\{x,y\}$.
We leave verification of this statements to the reader.
Our symmetric gadget is slightly bigger than the one in figure \ref{fig:super_edge1}. It's
25 edges and 10 vertices versus 21 edges and 9 vertices. It has been computed (by brute-force), that there
is no super-edge gadget that uses 8 vertices or less. We did not bother
to check if there is a gadget with number of edges less than 21. Using the symmetric
gadget we can compute number of edges and vertices used in entire $G_{\phi}$. Let $E_c$, $E_v$, $E_s$
denote a set of edges used in all clause gadgets, all vertex gadgets and all super-edge gadgets respectively.
We define $V_c$, $V_v$, $V_s$ in a similiar fashion. We have:
\begin{align*}
|E(G_{\phi})| &= |E_c| + |E_v| + |E_s| \\
&= 3m + 0 + 25(3m+n) \\
&= 78m+25n
\end{align*}
\begin{align*}
|V(G_{\phi})| &= |V_c| + |V_v| + |V_s| \\
&= 3m + 2n + (10-2)(3m+n) \\
&= 24m+10n
\end{align*}
This shows that reduction can be performed in polynomial time (with respect to $n$ and $m$) and therefore
completes (yet another) proof of theorem \ref{thm:red}. But we can improve the reduction even further and
push properties of our symmetric gadget to its limit.
We will now show what we call {\em The Necklace Reduction}. If we look at a loop of size $l$, we will spot as many as $l$
candidates for chosing the edge $\{x,y\}$. This is easily seen in figure 5. The symmetry of our gadget guarantees,
that any edge on the juncture of $K_4$-s can be considered ${x,y}$. But that leaves $l-1$ candidates unused. In nacklace reduction we
get rid of wasting so many useful edges (to some extent). We simply weave all vertex gadgets on a single loop of length $2n+1$.
Vertex gadget for variable $x_i$ (for $i=1..n$) now becomes egde on the juncture of $(2i)$-th and $(2i+1)$-th $K_4$-s (numeration
can start at any arbitrary $K_4$). We leave the rest of reduction the same as before. We have now created a beautiful necklace
of which example can be seen in figure \ref{fig:necklace} (it uses formula from previous example; some labels were omitted).
\begin{figure}[H]
\begin{center}
\includegraphics[scale=1.0]{./2col5.pdf}
\end{center}
\caption{The Necklace Reduction}
\label{fig:necklace}
\end{figure}
Number of edges and vertices drops down to:
\[
|E(G_{\phi})|= 78m + 10n + 5, \quad |V(G_{\phi})|= 27m + 4n + 2
\]
We can further improve the necklace by weaving all other super-edges, but
the construction is rather complicated. Details will be available in extended
version of this paper.
\begin{figure*}[t]
\begin{center}
\includegraphics[scale=0.6]{./2col2.pdf}
\end{center}
\caption{The super-edge gadget}
\label{fig:super_edge1}
\end{figure*}
\begin{figure*}[t]
\begin{center}
\includegraphics[scale=0.6]{./2col3.pdf}
\end{center}
\caption{The symetric super-edge gadget}
\label{fig:super_edge2}
\end{figure*}
\section{Two-coloring without monochromatic squares}
In this section we extend our reduction to cycles of length 4.
The abstract form of $G_{\phi}$ remains almost the same and the only
diffrence is that we have squares in place of triangles for clause
gadgets. In fact we use the similiar graph for higher values of $k$.
Proof of correctness is the same as before, so we leave the details
to the reader.
The most importnant part is to construct a gadget for super-edges. Now,
we want to create a graph with a selected edge $\{x,y\}$ that there exists
a valid coloring (without monochromatic squares) and that in every valid coloring
$c$: $c(x) \neq c(y)$. We use $K_6$ as a building block for the gadget and exploit
its coloring property.
\begin{figure}[H]
\begin{center}
\includegraphics[scale=0.6]{./2col6.pdf}
\end{center}
\caption{Coloring property of 6-clique}
\label{fig:coloring_property_k6}
\end{figure}
In figure \ref{fig:coloring_property_k6} on the left we see $K_6$. On the right it is the same $K_6$, but
with rearranged edges. Three arbitrarily chosen, disjoint edges have been pointed out and streched in three different directions.
Rest of the edges are less significant so we placed dotted lines in their place. Notice, that when we color vertices of top edge
in a single color -- let's say red -- then by using easy pigeon hole argument we can conclude, that exactly one of two bottom edges
will have both of its vertices colored blue.
To further simplify the $K_6$, imagine that the selected edges become nodes and that there are lines between top node
and two bottom nodes. This creates a reverse v-shaped component. The node which has two different colors associated
to it, we label as $X$ (see figure \ref{fig:coloring_property_k6_2}).
\begin{figure}[H]
\begin{center}
\includegraphics[scale=1.0]{./2col7.pdf}
\end{center}
\caption{Simplifing $K_6$}
\label{fig:coloring_property_k6_2}
\end{figure}
Now we present the trick to our gadget. We build a full binary tree of height 4, consisting of reverse v-shaped components.
It follows from coloring property of $K_6$ discussed before, that if we color root node in red, then there exist a path
from root to leaf with alternating colors (see figure \ref{fig:super_edge_sq}). Notice the analogy to the construction
of strings in previous section.
\begin{figure}[H]
\begin{center}
\includegraphics[scale=1.0]{./2col8.pdf}
\end{center}
\caption{Super-edge gadget. All leafs are connected to root.}
\label{fig:super_edge_sq}
\end{figure}
To achieve a contradition we connect all leafs to the root using two edges for each leaf in a way that they form a square.
This completes the construction. We choose root node as $\{x,y\}$.
Chosing the height 4 for $T$ is not a coincidence, as using any tree of smaller size would either not lead to contradiction
(heights 1 or 3) or would not be colorable -- for height 2 we can find a monochromatic square in any coloring. We again
leave verification to the reader.
It remains to show that our gadget has a valid coloring. We simply label all nodes by $X$. We now prove that this will
not create any monochromatic square. There are two places in our gadget that require special attention:
\begin{itemize}
\item $\mathcal{P}1$. Connections between inner nodes of the tree, and
\item $\mathcal{P}2$. Connections between leafs and root.
\end{itemize}
Both of them can be handaled in a strightforward way. For the former look at figure \ref{fig:connections_1},
where we reverse the process of
$K_6$-simplification for some subtree of $T$. We quickly verify, that there are no monochromatic squares. This is the smallest,
nontrivial subtree in which there could lurk some hidden monochromatic squares. Thanks to regular structure and symmerty of full
binary trees, any other combinations of nodes need not be checked. One could use induction for formal proof, but we will leave
it like this.
\begin{figure}[H]
\begin{center}
\includegraphics[scale=1.0]{./2col9.pdf}
\end{center}
\caption{Connections between inner nodes expanded.}
\label{fig:connections_1}
\end{figure}
$\mathcal{P}2$ couses some minor troubles. Take a look at figure \ref{fig:connections_2}.
Notice that we found a monochromatic square. This leads
to conclusion that not every coloring that labels each node by $X$ is valid. We can fix that by coloring both leafs so that
they form alternating squares with the root (the color is alternating). Now any path that passes from leaf to root have to change
the color, so there are no more threat to find a monochromatic square.
\begin{figure}[H]
\begin{center}
\includegraphics[scale=1.0]{./2col10.pdf}
\end{center}
\caption{Connections between leafs and root.}
\label{fig:connections_2}
\end{figure}
This completes the proof of theorem \ref{thm:red} for $k=4$.
We see that this is a polynomial reduction. For sake of completness lets count
number of edges and vertices in a single super-edge gadget, and then in entire graph $G_{\phi}$:
\[
\text{\#edges-in-gadget} = 243, \quad \text{\#vertices-in-gadget} = 62
\]
\begin{align*}
|E(G_{\phi})| &= |E_c| + |E_v| + |E_s| \\
&= 4m + 0 + 243(4m+n) \\
&= 976m + 243n
\end{align*}
\begin{align*}
|V(G_{\phi})| &= |V_c| + |V_v| + |V_s| \\
&= 4m + 2n + (62-2)(4m+n) \\
&= 244m + 62n
\end{align*}
\section{The general case}
In this section we finally prove theorem \ref{thm:red} for $k>4$. We do this by expanding the binary tree gadget from last section.
The tree will grow exponantially with respect to $k$, but remember that $k$ is a constant associated with the problem \COL{k}, so
our reduction will still be polynomial in size of $\phi$ (but very, very big). Our goal now is to construct a graph with a selected edge
$\{x,y\}$, that there exists a valid coloring (without monochromatic cycles of length $k$) and that in every valid coloring $c$:
$c(x) \neq c(y)$. For now assume that $k$ is even. This will simplify our reasoning.
First we construct a binary tree $T$ consisting of reverse v-shaped components introduced in previous section. Let
height of $T$ be $h=4 \lfloor \frac{k-1}{2} \rfloor$. For $i=1..\lfloor \frac{k-1}{2} \rfloor$, we will call all nodes of
depth $4i$: {\em cycle-inducing} (notice that root and leafs are also cycle-inducing).
Let $CI$ be the set of all cycle-inducing nodes in $T$. If we color root node in a single
color -- let's say red -- then there exists a path $P$ from root to some leaf, with alternating colors. Notice that all nodes in
$P \cap CI$ are now colored red. Those nodes will create a monochromatic cycle of length $k$. To achieve this, we add edges between
cycle-induced nodes in the following way.
First, we connect root and leafs just like in previous section. Next, for each cycle-induced
node of depth $4i$ ($i=1..(\lfloor \frac{k-1}{2} \rfloor-1)$) we conect it to all its descendants on depth $4(i+1)$ (they also belong
to $CI$). We add edges between them the same way we did with root and leafs. The example of how this produces monochromatic cycle is
shown on in figure \ref{fig:super_edge_general}. If we take the graph induced by $P \cap CI$, it forms a {\em donut} shown
in the right picture. We can easily identify a monochromatic cycle of length $k$.
\begin{figure}[H]
\begin{center}
\includegraphics[scale=1.0]{./2col11.pdf}
\end{center}
\caption{Super-edge gadget for general case and how to achieve contradiction.}
\label{fig:super_edge_general}
\end{figure}
Last thing to do is to prove that there exist a valid coloring of our super-edge gadget. Again we begin with labeling all
nodes in tree by $X$. We know from previous section how to handle connections between root and leafs -- we have to do the same
with all cycle-inducing nodes and their first cycle-inducing descendants. This way we will not be able to form a monochromatic cycle
that passes through two different nodes that are in $CI$. Note that at this point the gadget is correct only when value
$\lfloor \frac{k-1}{2} \rfloor$ is odd. This is true because of the way we color nodes in $CI$: the coloring of nodes on level
$4i$ force the coloring on nodes on level $4(i+1)$. This problem can be easily fixed by expanding tree another 4 levels and treating
nodes at level $h-4$ as {\em dummy nodes}.
We are left with the case when $k$ is an odd number. Note that the construction above is not working in this case, as we will not
achieve a contradiction. The fix is as follows: we change connections between leafs and root. Choose one vertex of root node and
connect all vertices in leafs to it. This creates triangles rather than squares and the {\em donut} now looks like someone has
taken a bite, but we can now find a monochromatic cycle of length $k$ for all odd numbers (if we color root node in red).
The valid coloring does not change.
For sake of completness we count the number of edges and vertices in entire reduction:
\begin{align*}
|E(G_{\phi})| &= |E_c| + |E_v| + |E_s| \\
&= km + 0 + \\
&+ (15 \cdot 2^{4\lfloor \frac{k-1}{2} \rfloor} - 2^{4\lfloor \frac{k-1}{2} \rfloor +1} \\
&+ 2 \cdot 2^{4\lfloor \frac{k-1}{2} \rfloor} + 32 \sum_{i=0}^{\lfloor \frac{k-1}{2} \rfloor}2^{4i})(km+n)
\end{align*}
\begin{align*}
|V(G_{\phi})| &= |V_c| + |V_v| + |V_s| \\
&= km + 2n + (2(2^{4\lfloor \frac{k-1}{2} \rfloor + 1})-2)(km+n)
\end{align*}
Thus, we have proved theorem \ref{thm:main}.
\section{Conclusions}
We have shown that using symmetry, one can conceive many interesting combinatorial structures and in
graph theory there is nothing more symmetric and regular than a clique. The obvious question is: can
we make the reduction smaller? We have proved that string gadget from section 3 can be used as a tool
to greatly decrease the number of edges and vertices used, but we do not know if the same can be said about
tree gadget from sections 4 and 5.
|
\section{Introduction}
The idea of feature screening came along as high dimensional data were collected in modern technology. It was aimed at dealing with the challenges of computational expediency, statistical accuracy and algorithmic stability due to high dimensionality. Fan and Lv proposed the sure independence screening (SIS) \cite{fan2008sure} and showed that the Pearson correlation ranking procedure possessed a sure screening property for linear regression with Gaussian predictors and responses. A new feature screening procedure for high dimensional data based on distance correlation \cite{szekely2007measuring}, named DC-SIS, was presented in \cite{li2012feature}. DC-SIS retained the sure screening property of the SIS, and additionally possessed new advantages of handling grouped predictors and multivariate responses by using distance correlation. Moreover, since distance correlation was applicable to arbitrary distributions, DC-SIS could also be used for screening features without specifying a regression model between the response and the predictors, and thus was robust to model mis-specification.
However, both SIS and DC-SIS relied on a user-specified model size $d$ which decided the number of predictors being selected. Let the sample size be $n$, $d$ was chosen to be multipliers of the integer part of $n/\log n$ in \cite{fan2008sure} and \cite{li2012feature} which did not depend on any other characteristics of the data. As pointed out by a referee of \cite{li2012feature}, the choice of $d$ was of great importance in practical implementation and might influence the screening results significantly. Our study is aimed at fixing this shortcoming by including an automatic stopping criteria for DC-SIS based on the property of distance covariance.
The screening procedures may fail if a feature is marginally uncorrelated, but jointly correlated with the response, or in the reverse situation where a feature is jointly uncorrelated but has higher marginal correlation than some important features. An iterative SIS was proposed in \cite{fan2008sure} to fix this problem. Current research interest involves dealing with this drawback but this work is not related to this quest.
We demonstrate our improved method through two real examples. The small round blue cell tumors (SRBCT) data were relatively easy to classify and had been studied extensively. The Cancer Genome Atlas (TCGA) ovarian cancer data, however, were much more challenging due to the large number of genes and limited sample size. The target was to identify the important genes that contribute to the sensitive or resistant status after receiving a particular chemotherapy treatment. A substantial fraction of the population was difficult to classify and a ``withholding decision" option is allowed in the support vector machine with reject option model to adapt to this fact. A multiple cross validation is used to quantify uncertainty given a humongous number of candidates, and we see a commonly observed dilemma that different variables are selected by using different subsets of the data. Comparison between the results from the original data and those from the data obtained by randomly permuting the response provide further justification on our conclusions. Furthermore, the multiple cross validation on the permuted data discloses the existence of spuriously correlated variables in high dimensional data and thus failure of variable selection and model building based on training data.
\section*{Some Preliminaries}
\subsection*{Distance correlation}
\cite{szekely2007measuring} proposed distance correlation as a measurement of dependence between two random vectors. The method has been successfully applied to various problem, see \cite{kong2012using} for example. To be specific, the authors defined the distance covariance between $X\in\mathbb{R}^p$ and $Y\in\mathbb{R}^q$ to be
\begin{displaymath}
\operatorname{V}^2(X,Y)= \frac{1}{c_p c_q}\int_{\mathbb{R}^{p+q}} \frac{\left| f_{X,Y}(s, t) - f_X(s)f_Y(t) \right|^2}{|s|_p^{1+p} |t|_q^{1+q}} dt\,ds
\end{displaymath}
where $f_{X, Y}(s, t), f_X(s)$, and $f_Y(t)$ are the characteristic functions of $(X, Y), X$, and $Y$ respectively, and $c_p$, $c_q$ are constants chosen to produce scale free and rotation invariant measure that doesn't go to zero for dependent variables. The idea is originated from the property that the joint characteristic function factorizes under independence of the two random vectors. This leads to the remarkable property that $V^2(X,Y)=0$ if and only if $X$ and $Y$ are independent.
The sample version of distance covariance and distance correlation involves pairwise distances. For a random sample $(X,Y)=\{(X_k, Y_k): k=1,...,n\}$ of $n$ i.i.d random vectors $(X,Y)$ from the joint distribution of random vectors $X$ in $\mathbb{R}^p$ and $Y$ in $\mathbb{R}^q$, the Euclidean distance matrices $(a_{ij})=(|X_i-X_j|_p)$ and $(b_{ij})=(|Y_i-Y_j|_q)$ with $i,j = 1,\ldots,n$ are computed. Define the double centering distance matrices
\begin{displaymath}
A_{ij}=a_{ij}-{\overline{a}_{i\cdot}}-{\overline{a}_{\cdot j}}+{\overline{a}_{\cdot\cdot}},\hspace{0.3cm}i,j=1,\ldots,n,
\end{displaymath}
where
\begin{displaymath}
{\overline{a}_{i\cdot}}=\frac{1}{n}\sum_{j=1}^na_{ij},\hspace{0.3cm}{\overline{a}_{\cdot j}}=\frac{1}{n}\sum_{i=1}^na_{ij},\hspace{0.3cm}{\overline{a}_{\cdot\cdot}}=\frac{1}{n^2}\sum_{i,j=1}^na_{ij},
\end{displaymath}
similarly for $B_{ij}=b_{ij}-{\overline{b}_{i\cdot}}-{\overline{b}_{\cdot j}}+{\overline{b}_{\cdot\cdot}},\hspace{0.3cm}i,j=1,\ldots,n$. Then, the sample distance covariance $V_n(X,Y)$ is defined by
\begin{displaymath}
V_n^2(X,Y)=\frac{1}{n^2}\sum_{i,j=1}^nA_{ij}B_{ij}.
\end{displaymath}
The sample distance correlation $R_n(X,Y)$ is defined by
\begin{equation}
R_n^2(X,Y)=\begin{cases}\displaystyle \frac{V_n^2(X,Y)}{\sqrt{V_n^2(X)V_n^2(Y)}}, & V_n^2(X)V_n^2(Y)>0; \\
0, &V_n^2(X)V_n^2(Y)=0,\end{cases} \notag
\end{equation}
where the sample distance variance is defined by
\begin{displaymath}
V_n^2(X)=V_n^2(X,X)=\frac{1}{n^2}\sum_{i,j=1}^nA_{ij}^2.
\end{displaymath}
\subsection*{Feature screening via distance correlation (DC-SIS)}
\cite{fan2008sure} proposed sure independence screening (SIS) procedure based on the Pearson correlation for feature selection. The distance correlation version of this technique (DC-SIS) was studied in \cite{li2012feature}. With a user-specific model size $d$, the variables whose distance correlations with the response ranking from 1st to $d$th in decreasing order were selected. The authors explored the theoretic properties of the DC-SIS and proved that the DC-SIS kept the desired sure screening property established in \cite{fan2008sure}. Moreover, due to the property of distance correlation, DC-SIS procedure was robust to model mis-specification, which was demonstrated in their simulations.
\section*{Improving DC-SIS using distance covariance}
\begin{theorem}Suppose random vectors $X,Z\in\mathbb{R}^p$ and $Y\in\mathbb{R}^q$, and assume $Z$ is independent of $(X,Y)$, then
\begin{equation}
V^2(X+Z,Y)\leq V^2(X,Y),
\end{equation}
where $V$ is the population distance variance defined in \cite{szekely2007measuring}.
\end{theorem}
\begin{proof}
\begin{align}
V^2(X+Z,Y)&= \parallel f_{X+Z,Y}(t,s) - f_{X+Z}(t)f_{Y}(s)\parallel^2 \notag\\
&=\frac{1}{c_pc_q}\int_{\mathcal{R}^{p+q}}\frac{1}{|t|_p^{1+p}|s|_q^{1+q}}|f_{X+Z,Y}(t,s) - f_{X+Z}(t)f_{Y}(s)|^2dtds. \notag
\end{align}
The following fact follows from the definition of characteristic function and independence assumption.
\begin{align}
&|f_{X+Z,Y}(t,s) - f_{X+Z}(t)f_{Y}(s)|^2 \notag\\
=&|Ee^{it^T(X+Z)+is^TY} - Ee^{it^T(X+Z)}Ee^{is^TY}|^2 \notag \\
=&|Ee^{it^TX+is^TY}Ee^{it^TZ} - Ee^{it^TX}Ee^{it^TZ}Ee^{is^TY}|^2\notag\\
=&|f_{X,Y}(t,s)f_Z(t) - f_{X}(t)f_Z(t)f_{Y}(s)|^2\notag\\
=&|f_Z(t)|^2|f_{X,Y}(t,s) - f_{X}(t)f_{Y}(s)|^2, \notag
\end{align}
Since $|f_Z(t)|\leq 1$ by the property of characteristic function\footnote{In \cite{kosorok2009discussion}, the author obtained equality here which is incorrect.}, we have
\begin{equation}
|f_{X+Z,Y}(t,s) - f_{X+Z}(t)f_{Y}(s)|^2 \leq |f_{X,Y}(t,s) - f_{X}(t)f_{Y}(s)|^2, \notag
\end{equation}
which implies
\begin{align}
V^2(X+Z,Y)&\leq\frac{1}{c_pc_q}\int_{\mathcal{R}^{p+q}}\frac{1}{|t|_p^{1+p}|s|_q^{1+q}}|f_{X,Y}(t,s) - f_{X}(t)f_{Y}(s)|^2dtds\notag\\
&=\parallel f_{X+Z,Y}(t,s) - f_{X+Z}(t)f_{Y}(s)\parallel^2 \notag\\
&=V^2(X,Y). \notag
\end{align}
\end{proof}
We know that if $E|X|_p<\infty, E|X+Z|_p<\infty$ and $E|Y|_p<\infty$, then almost surely
\begin{align}
\displaystyle \lim_{n\rightarrow\infty}V_n^2(X+Z,Y)&=V^2(X+Z,Y),\notag\\
\displaystyle \lim_{n\rightarrow\infty}V_n^2(X,Y)&=V^2(X,Y).\notag
\end{align}
Thus, for the sample distance covariance, if $n$ is large enough, we should have
\begin{equation}
V_n^2(X+Z,Y)\leq V_n^2(X,Y), \notag
\end{equation}
under the assumption of independence between $(X,Y)$ and $Z$.
In the case where $(X,Z)$ is of interest, which is the usual situation for variable selection setting, we could use the above theorem by incorporating degenerated random vectors as follows. Suppose $X\in\mathbb{R}^{p_1}$ and $Z\in\mathbb{R}^{p_2}$, then we augment $X$ and $Z$ to be $\tilde{X}=(X,0_{p_2})$ and $\tilde{Z}=(0_{p_1},Z)$ respectively. $\tilde{X}$ and $\tilde{Z}$ are therefore of the same dimension and $\tilde{X} + \tilde{Z} = (X,Z).$
We implemented the above theorem as a check for stopping for DC-SIS. For the original DC-SIS, it required a user-specified model size $d$, which was always chosen as multipliers of the integer part of $n/\log n$. For our improved screening procedure with distance correlation, we first ranked the importance of $x_i, i = 1,...,p$ using the marginal distance correlations with the response as DC-SIS did and initialized $\mathcal{S}$ as the singleton including the index of the top one variable. Instead of selecting the top $d$ variables, we kept adding variables to $x_\mathcal{S} = \{x_i:i\in\mathcal{S}\}$ according to the ordered list of variables until observing a decrease in the distance covariance between $x_\mathcal{S}$ and $y$. The procedure took the following steps and we denoted the procedure as DCOV method.
\begin{enumerate}
\item Calculate marginal distance correlations for $x_i, i = 1,...,p$ with the response.
\item Rank the variables in decreasing order of the distance correlations. Denote the ordered variables as $x_{(1)}, x_{(2)},...,x_{(p)}$. Start with $x_{\mathcal{S}} = \{x_{(1)}\}$.
\item For $i$ from $2$ to $p$, keep adding $x_{(i)}$ to $x_{\mathcal{S}}$ if $V_n^2(x_{\mathcal{S}},y)$ does not decrease. Stop otherwise.
\end{enumerate}
\section*{Real application on SRBCT data}
The small round blue cell tumors (SRBCTs) are 4 different childhood tumors named so because of their similar appearance on routine histology, which makes correct clinical diagnosis extremely challenging. However, accurate diagnosis is essential because the treatment options, responses to therapy and prognoses vary widely depending on the diagnosis. They include Ewing's family of tumors (EWS), neuroblastoma (NB), non-Hodgkin lymphoma (in our case Burkitt's lymphoma, BL) and rhabdomyosarcoma (RMS). The SRBCT data being published in \cite{khan2001classification} included the expression of 2308 genes measured on 63 samples (23 EWS, 8 BL, 12 NB and 20 RMS). This data are known as an easy-classified example and have been studied by many. \cite{lee2004multicategory} using the multicategory SVM is one of several methods that have excellent classification results on this data set. Hence, we focus more on the selected genes.
We applied our improved feature screening procedure on this dataset and compared our selection of genes with the 96 top genes reported in \cite{khan2001classification}. This is a multicategory classification and the genes were screened in a one-versus-rest fashion. Specifically, for each of the four different types of tumors, we generated a response indicator vector taking value of 0 if the sample came from the current interested type and 1 otherwise.
This allowed us to implement our screening procedure and obtained the genes which showed high distance correlation with the current type of tumor. The four groups of selected genes were combined as a whole collection of in total 176 genes. 47 genes turned out to be in common for the DCOV selection and the top 96 genes used in \cite{khan2001classification}.
We further examined the power of these two groups of genes in differentiating the 4 types of tumors by presenting the pairwise distances of the 63 samples
(\textit{Figure 1}). As shown in the plot, the samples were arranged in the order of EWS, BL, NB and RMS. The pairwise distances resulted from the two
selections of genes were scaled to maximum of 1 respectively so that they shared the same magnitude. Both groups of genes could tell the 4 types
of tumors apart. Compared with the 96 genes from \cite{khan2001classification}, however, the 176 DCOV selected genes show better distinguishability and clearer contrast over the 4 classes. Moreover, the right panel almost missed the samples labeled from 57
to 62 in the class of RMS but the 176 DCOV genes could recognize them with big differences between the in and out class pairwise distances. The dataset were known to be easy for classification and both sets of genes were able to classify the testing set of 20 samples perfectly via k-nearest neighbor method with $k=3$.
\begin{figure}
\begin{center}
\includegraphics[width=18cm]{khan_no_margin.pdf}
\caption{Comparison of pairwise distances between the two selections of genes. Left and right panel present the pairwise distances of the 63 samples over the improved DC-SIS selection of 176 genes and the 96 reported genes in \cite{khan2001classification} respectively.}
\end{center}
\end{figure}
\section*{Real application on TCGA ovarian cancer data}
\subsection*{Data description}
Ovarian cancer is the fifth-leading cause of cancer death among women in the United States; 22,240 new cases and 14,030 deaths were estimated to have occurred in 2013\cite{website,bell2011integrated}. The standard treatment for high-grade serous ovarian cancer is aggressive surgery followed by chemotherapy. Despite treatment, a vast majority of ovarian cancer patients eventually relapse and die of their disease with a major cause of chemotheraphy resistance \cite{selvanayagam2004prediction}. Identification and prediction of patients with chemoresistant thus become important for improving the outcome of ovarian cancer.
The Cancer Genome Atlas (TCGA) collected high-quality, high-dimensional, and multi-modal genetic data from women with ovarian cancer. There were 279 samples with explicit chemostatus and gene expression (Affymetrix HT-HGU133a) data in the public set. among which 191 subjects were sensitive to chemotherapy and 88 were chemoresistant. Expression data for 12042 genes after log transformation are used for analysis. The issue is to explore whether there are genes whose expression pattern is strongly correlated with the response indicating chemotherapy status.
\subsection*{DCOV gene selection results based on all the observations}
Our feature screening procedure on the gene expression data for the 279 patients selected 82 genes, among which 5 were reported to be related to ovarian cancer in the literature. IGFBP5 ranked as the 5th is one of the six members of insulin-like growth factor-binding protein (IGFBP) family and is known to be important for cell growth control, induction of apoptosis and other IGF-stimulated signaling pathways. IGFBP5 expression is shown to prevent tumor growth and inhibited tumor vascularity in a xenograft model of human ovarian cancer and is suggested that IGFBP5 plays a role as tumor suppressor by inhibiting angiogenesis \cite{rho2008insulin}. GPR3, the 7th, is a member of a family of G-protein couple receptors whose activation of PKA and subsequent increase of cyclic AMP level promotes meiotic arrest in the oocyte \cite{mehlmann2004gs}. Mice deficient in GPR3 display premature ovarian aging and loss of fertility \cite{ledent2005premature}. MAPK4, the 18th, is a member of MAPK signaling pathway. MAPK signal transduction cascade is dysregulated in a majority of human tumors \cite{basu2009nanoparticle}. It is suggested playing an important role in molecular diagnostics and molecular therapeutics for lowgrade ovarian cancer \cite{bast2010personalizing}. FZD5 ranked as the 22th encodes Frizzled-5 protein, which is believed to be the receptor for the Wnt5A ligand \cite{thiele2011expression}. The Wnt5A ligand plays a context-dependent role in human cancers. It has been demonstrated that Wnt5a is expressed at significantly lower levels in human Epithelial ovarian cancer (EOC) cell lines and in primary human EOCs compared with either normal ovarian surface epithelium or fallopian tube epithelium \cite{bitler2011wnt5a}. FGF22, the 56th, is a member of Fibroblast Growth Factors (FGFs) family, whose members possess broad mitogenic and cell survival activities, and are involved in a variety of biological processes, including embryonic development, cell growth, morphogenesis, tissue repair, tumor growth and invasion. The inhibition of FGFR2, which is a member of this family, has been found to increase cisplatin sensitivity in ovarian cancer \cite{cole2010inhibition}.
39 pathways were found to be associated with the 82 genes, among which 3 pathways are known to be important for ovarian cancer. MAPK signaling pathway is suggested playing an important role in molecular diagnostics and molecular therapeutics for lowgrade ovarian cancer \cite{bast2010personalizing}. Wnt signaling pathway is best known for its role in tumorigenesis. \cite{bast2010personalizing} demonstrated the difference in Wnt signaling pathway between normal ovarian and cancer cell lines and between benign tissue and ovarian cancer. They also pointed out that those differences implicate that Wnt signaling leads to ovarian cancer development despite the fact that gene mutations are uncommon. \cite{yin2011genetic} suggested that genetic variants in TGF-$\beta$ signaling pathway are associated with ovarian cancer risk and may facilitate the identification of high-risk subgroups in the general population.
\subsection*{Support vector machine with reject option}
We estimated the probabilities of being chemosensitive or chemoresistant by a penalized Bernoulli likelihood main effect spline model using the \texttt{R} package \texttt{gss}\cite{gu2007gss}. Aside from the additive expression effects of the selected 82 genes, we also included two more covariates, namely the cancer grade and cancer stage of the subjects. Cancer grade is an indicator for grade 2 and grade 3. Cancer stage indicates whether the subject is in stages IIIC and IV or not. As shown in \textit{Figure 2}, the estimated probabilities have high density around small and large values for sensitive and resistant patients respectively, with overlapping in the middle values. This suggested that we were less confident about the chemostatus for the patients in the middle range and so we sought a principled approach which withholds decision for such cases.
\begin{figure}
\begin{center}
\includegraphics[width=14cm]{boxplot_1.pdf}
\caption{Fitted probabilities by penalized Bernoulli likelihood model with the 82 genes.}
\end{center}
\end{figure}
\cite{wegkamp2011support,bartlett2008classification} investigated the support vector machines with a built-in reject option for binary classification where the results of classification could be $-1, +1$ or withhold decision. Given a discriminant function $f: \mathcal{X}\rightarrow \mathbb{R}$, the method only reports $sgn(f(x))\in\{-1,1\}$ if $|f(x)| > \delta$ and withholds decision if $|f(x)|\leq \delta$. Suppose that the cost of making a wrong decision is 1 and that of rejecting to make a decision is $d\in[0,\frac{1}{2}]$, then an proper risk function is
\begin{displaymath}
L_{d,\delta}(f)=El_{d,\delta}(Yf(X)) = P\{Yf(X)<-\delta\}+dP(|Yf(X)|\leq\delta)
\end{displaymath}
with the discontinuous loss function
\begin{equation}
l_{d,\delta}(z) =
\begin{cases}
1, & \hbox{if $z<-\delta$;} \\
d, & \hbox{if $|z|\leq \delta$;} \\
0, & \hbox{otherwise.}\notag
\end{cases}
\end{equation}
The classifier associated with the discriminant function
\begin{equation}
f^\ast_d(x) =
\begin{cases}
-1, & \hbox{if $\eta(x)<d$;} \\
0, & \hbox{if $d\leq \eta(x) \leq 1-d$;} \\
+1, & \hbox{if $\eta(x)>1-d$,}\notag
\end{cases}
\end{equation}
with $\eta(x) = P\{Y=+1 | X=x\}$ minimizes the risk $L_{d,\delta}(f)$ with
\begin{displaymath}
L^\ast_d=L_{d,\delta}(f^\ast_d)=E\min\{\eta(X),1-\eta(X),d\}.
\end{displaymath}
To avoid working with the discontinuous loss $l_{d,\delta}$, \cite{wegkamp2011support,bartlett2008classification} proposed a convex surrogate loss, which is the generalized hinge loss,
\begin{equation}
\phi_d(z) =
\begin{cases}
1-az, & \hbox{if $z<0$;} \\
1-z & \hbox{if $0\leq z < 1$;} \\
0, & \hbox{otherwise,}\notag
\end{cases}
\end{equation}
where $a = (1-d)/d\geq 1$. It followed that $f^\ast_d$ also minimizes the risk associated with $\phi_d$ over all measurable $f:\mathcal{X}\rightarrow\mathbb{R}$.
The discriminant functions $f$ took the form $f_{\lambda}(x)=\sum_{j=1}^M\lambda_jf_j(x)$ based on a set of known functions $f_j:\mathcal{X}\rightarrow\mathbb{R}$ and coefficients $\lambda_j\in\mathbb{R},1\leq j \leq M$. The coefficients were chosen to minimize the empirical risk
\begin{displaymath}
\hat{R}_{\phi}(f_\lambda) = \frac{1}{n}\sum_{i=1}^n\phi(Y_if_\lambda(X_i)).
\end{displaymath}
To reflect the preference for sparse solutions, which is desirable when $M$ is large compared to the sample size $n$, an $l_1$ type restriction $\|\lambda\|_1 =\sum_{j=1}^M|\lambda_j|$ was incorporated in \cite{wegkamp2011support} and $f_\lambda$ is estimated by $f_{\hat{\lambda}_r}$, where
\begin{equation}
\hat{\lambda}(r) = \textrm{arg}\min_{\lambda\in\mathbb{R}^M}\hat{R}_{\phi}(f_\lambda)+r\|\lambda\|_1
\end{equation}
and $r>0$ is a tuning parameter. We followed \cite{wegkamp2011support} to call this model support vector machines with reject option(SVM-R).
The authors in \cite{bartlett2008classification} also showed that the choice of $\delta = 1/2$ gives the optimal bound established by the excess risk of $\phi_d$ on the excess risk $L_{d,\delta}-L^\ast_d$ for any fixed $d\in [0,1/2)$ and measurable function $f$. For this reason, we fixed $\delta = 1/2$ for our practical use of the method. Furthermore, we took the set of known functions $f_j:\mathcal{X}\rightarrow\mathbb{R}$ to be linear functions of the log transformation on the 12024 genes. The optimization problem $(2)$ was formulated into a linear programming task and solved using \textsc{MATLAB}.
\textit{Figure 3} presents the 82 genes for the 279 subjects in groups according to the SVM-R classification results. The results correspond to the particular choice of $d = 1/4$ and $r = 4$ to illustrate the benefits of SVM-R. As shown in the plot, there is a big difference in the gene
expression between the subjects assigned to be resistant and sensitive. The behavior of the 82 genes for those without a certain decision tends to be in-between.
\begin{center}
\begin{figure}[h]
\includegraphics[width=18cm]{gene_1.pdf}
\caption{Gene expression data for the 82 selected genes and 279 subjects with SVM-R classification for $d = 1/4$ and $r = 4$. The subjects are grouped according to their assigned decisions by the SVM with a reject option. The left group involves 15 patients (1 sensitive and 14 resistant) classified to be resistant. The middle group is assigned to be sensitive and contains 123 sensitive and 8 resistant subjects. 67 sensitive patients and 66 resistant patients with a withhold decision are shown in the right group.}
\end{figure}
\end{center}
\subsection*{Five fold cross validation}
In order to choose the tuning parameter in SVM-R, we need to hold aside a tuning set before selecting the genes. Leaving out different observations leads to different gene selection results. Here we applied a five fold cross validation analysis to examine the variations of selections of genes and SVM-R model fitting results across different partitions of the dataset. The implementation followed the steps below.
\begin{enumerate}
\item Randomly partition the 279 subjects into 5 non-overlapping folds.
\item Select genes from the 12024 genes based on 4 folds as the training set.
\item Build SVM-R model with the selected genes and the two cancer status variables based on the training set.
\item Use the leaving-out fold as the tuning set to choose the tuning
parameter for SVM-R with mean $l-$loss, defined below, as the criteria.
\item Repeat $2.-4.$ for the 5 folds.
\end{enumerate}
The $l-$loss for a subject is $1$ if a misclassification occurs, $d$
if a withholding decision is made and $0$ otherwise. The mean $l-$loss
is the average over the $l-$losses for all the subjects in a given set
of data. We looked for tuning parameter values minimizing the mean $l-$loss.
The above procedure produced 5 selections of genes before SVM-R,
namely $S_1,\ldots,S_5$. In addition, we also have the 82 genes
selected from all the subjects previously. Table \ref{t1}
presents the pairwise intersections of these 6 sets with each
other. The union of the 5 selections includes 211 genes, which covers
77 genes in the 82 genes. 73 out of 211 genes have frequency more than
1 where 63 of them appear in the 82 genes. After implementing SVM-R,
the union of genes is reduced to 98 genes. \textit{Figure 4} displays
the histogram of these 211 genes colored by the frequency after SVM-R
runs for $d=1/5$. The pink region denotes the parts further eliminated by SVM-R, which is consistent with the DCOV selection in that SVM-R further rules out the genes with low frequency in the union. \begin{table}
\centering
\begin{tabular}{c|ccccc}
\toprule
& $S_1$ & $S_2$ & $S_3$ & $S_4$ & $S_5$ \\
\hline
$S_1$ & 53 & & & & \\
$S_2$ & 16 & 77 & & & \\
$S_3$ & 23 & 21 & 87 & & \\
$S_4$ & 18 & 16 & 15 & 33 & \\
$S_5$ & 27 & 30 & 31 & 21 & 94 \\
\hline
82 genes & 38 & 38 & 44 & 28 & 50 \\
\bottomrule
\end{tabular}
\caption{Pairwise intersections of $S_1,\ldots,S_5$ and the 82 genes. The diagonal numbers are the numbers of selected genes in each $S_i$.}
\label{t1}
\end{table}
\begin{figure}
\begin{center}
\includegraphics[width=10cm]{gene_5_loss.pdf}
\caption{Frequency for the union of $S1,\cdots, S5$, colored by
frequencies after SVM-R for $d=1/5$.}
\end{center}
\end{figure}
\subsection*{Multiple cross validation}
In order to consider uncertainty in variable selection and model
building due to different partitions of the dataset, we further
extended the five fold cross validation to multiple cross
validation (MCV) and assessed the prediction power through the
following procedure. The results were summarized in the upper part of
Table \ref{t2}.
\begin{enumerate}
\item Randomly partition 279 samples into a 1/5 tuning set, a $2/3\times 4/5 = 8/15$ training set and a $1/3\times 4/5 = 4/15$ testing set.
\item Select genes from the 12024 genes using the proposed method on the training set.
\item Build SVM-R model using the selected genes and the two cancer status variables based on the training set.
\item Use the tuning set to choose the tuning parameter for SVM-R with
mean $l-$loss as the criteria.
\item Use the model with chosen parameter to predict labels for the testing set.
\item Repeat $1.-5.$ for 50 times.
\end{enumerate}
\begin{table}
\centering
\begin{tabular}{l | c | c c c c c }
\toprule
& & num of reps & mean training & mean testing & mean
num training & mean num test \\
& $d$ & with decision & accuracy(std) & accuracy(std) &with
decision(std) & with decision(std) \\
\hline
\multirow{3}{*}{original}
&$1/3$& 50 & 0.8319(0.0914) & 0.6943(0.0544) & 101.9400(27.8043) & 49.1800(15.4295) \\
&$1/4$& 43 & 0.9371(0.0336) & 0.7807(0.1250) & 43.0698(27.0338) & 20.1860(12.9638)\\
&$1/5$& 37 & 0.9420(0.0358) & 0.8215(0.1460) & 24.4595(21.3874) & 11.4865(10.0626)\\
\hline
\multirow{3}{*}{permute}
&$1/3$& 49 & 0.7984(0.0078) & 0.6910(0.0426) & 112.1837(26.4352) & 55.5918(13.9954) \\
&$1/4$& 28 & 0.9225(0.0023) & 0.6867(0.0810) & 56.9643(21.4346) & 25.2857(10.4132)\\
&$1/5$& 9 & 0.9686(0.0015) & 0.7071(0.1322) & 53.4444(28.3333) & 24.5556(13.0682)\\
\bottomrule
\end{tabular}
\caption{Results for the 50 individual replications for $d = 1/3, 1/4$
and $1/5$. The upper and lower part are results for the original and
permuted data respectively. The third column shows the number of
replications out of 50 with at least one definite decision made on
the testing set. The fourth and fifth columns of the
table conclude the mean training and testing accuracies with
standard deviation in the parenthesis respectively
restricted to the repetitions with decision made. The last two
columns display the mean and standard deviation for the number of
patients assigned decisions for the training and testing sets
respectively given the replications with decision made.}
\label{t2}
\end{table}
To understand more about the 50 models, we further explored the prediction
results for $d = 1/5$. The prediction labels from the 50 models were
aggregated, following the idea of ensemble methods. The result for
each individual was recorded in a vector of three frequencies, namely
the frequency of being classified as sensitive subjects, the frequency
of obtaining a withholding and the frequency of being assigned to be
resistant out of the 50 models. Let $(s_i, w_i, r_i)$ be the vector for the $i$th patient.
A finer analysis was conducted by looking at the strength of being
sensitive or resistant according to $(s_i, w_i, r_i).$ A voting score
$v_i$ was defined as $(s_i - r_i) / w_i$. Hence, a positive $v_i$
indicated a tendency of being sensitive whereas a negative $v_i$
suggested more possibility of being resistant.
Table \ref{t3} (upper part) partitions the voting scores
into 5 intervals and describes the distribution of $v_i$'s as well as
the proportion of sensitive subjects within each range of $v_i$,
compared to the overall proportion of sensitive patients,
i.e. $191/279$, in the population. It turned out that the trend of
being sensitive weakened monotonically as the voting score decreased. The
stratification specified a subgroup of 15 patients, who possessed the
greatest voting scores, with very high accuracy to be
chemosensitive. The next highest voting score subgroup of 47 subjects
also showed strong confidence of being sensitive compared to the
sample proportion. The conclusion from partitioning the voting scores
was conservative but led to more convincing and steady classification results.
\begin{table}[h]
\centering
\begin{tabular}{l c c c c c c c}
\toprule
& & &\multicolumn{4}{c}{voting score}\\
\cmidrule(l){4-8}
& & &$(-0.1,0]$ & $(0,0.1]$ & $(0.1,0.2]$ & $(0.2,0.4]$ & $(0.4,1.5]$ \\
\hline
\multirow{2}{*}{original}
& \vline &frequency& 76 & 74 & 67 & 47 & 15\\
&\vline &proportion& 0.5658 & 0.6486 & 0.7164 & 0.8085 & 0.9333 \\
\hline
\multirow{2}{*}{permuted}
&\vline &frequency& 145 & 67 & 43 & 24 & 0\\
&\vline &proportion& 0.6759 & 0.6866 & 0.7209 & 0.6667 & NA \\
\bottomrule
\end{tabular}
\caption{
Frequency of voting score $v_i$'s and proportion of sensitive subjects in each subinterval for $d = 1/5$. The upper and lower parts correspond to the original and permuted data respectively.}
\label{t3}
\end{table}
Each replication of the 50 multiple cross validations gave rise to a different collection of selected genes. This issue is common to selecting
variables from a humongous number of candidates, in the not-low-hanging-fruit situation. The union of the 50 gene selections before SVM-R consisted of 1245 genes
and included all the 82 genes discussed previously. 34 out of 1245
genes were chosen at least 10 times, where 33 of them appeared in the
82 genes, but very few appeared in more than 25 runs. The $l_1$
penalty provided additional elimination, and for $d = 1/5$, 498 out of 1245 genes remained after SVM-R runs. \textit{Figure 5} displays the histogram for the 1245 genes before SVM-R. We distinguished their frequency after SVM-R by different colors. It is shown that a large number of genes with low frequency are further deleted by SVM-R model, i.e. pink color.
\begin{figure}
\begin{center}
\includegraphics[width=12cm]{gene_50_loss.pdf}
\caption{Frequency for 1245 genes being selected by DCOV method,
colored by frequencies after SVM-R for $d=1/5$.}
\end{center}
\end{figure}
\subsection*{Permutation of the response}
Our method involved several components, including variable selection,
SVM-R, MCV and the voting score, which were interacting with each
other, and led to 15 patients with over $93\%$ accuracy to be sensitive for $d = 1/5$. To further understand the mechanism and to demonstrate that the outcomes were not produced by noises, we randomly permuted the response and went through the whole procedure to compare the results with those for the original data.
It followed that the DCOV method selected genes spuriously correlated
with the permuted response based on the training data in each
replication of the 50 MCVs. The maximum distance correlation value of
the selected genes in each repetition was very close to that for the
original data. The highly correlated genes appeared due to the high
dimensionality of over 12000 genes and less than 200 training
samples.
However, the MCV step played the role of a safeguard against
the fake signals. As Table \ref{t2} depicted, the mean
training accuracies for the original and permuted data showed similar
behavior for the original and permuted data, meaning that the selected
genes were indeed important for the training data. Thus, the chosen
genes in the permutation set should provide little prediction power for the tuning and
testing data. Hence, the validation sets
selected large tuning parameter values driving all the patients with
no decision for many of the 50 replications for $d=1/4$ and
$1/5$. This did not happen for $d=1/3$ since the sample ratio
$191/279$ is slightly greater than $1/3$. For the rest of the replications
with decision making, the mean testing accuracies for the permuted data
remained at the level of the sample proportion of sensitive subjects
for all three values of $d$, which deviated much from the increasing
pattern in the original data.
These suggested that the MCV procedure was able to provide double
fail-secure for fake signals. On one hand, SVM-R placed a cap on the
conditional probability of misclassfication and eliminated the
replications where the selected genes could not produce results
achieving the specified confidence on the validation set. On the other hand, the
mean testing accuracies on the replications passing through the
safeguard of tuning sets would be no better than assigning everyone to
the sensitive class when there was no real signal.
Furthermore, the poor prediction performance of the 50 individual
models ended up with unsurprisingly disappointing voting score results
for the permuted data, as shown in the lower part of Table
\ref{t3}. Many of the patients obtained a relatively small
value of the voting score and nobody got a score in the range where
the original data had the highest accuracy, meaning that the confidence was quite
low. Moreover, the stratified ratios of sensitive subjects for different
ranges of the voting scores did not show anything insightful other than
being around the sample proportion.
\section*{Discussion}
The paper introduced a new variable selection procedure based on the property of distance covariance and demonstrated the application through two examples. The small round blue cell tumors data played a role of a toy example to show that the performance of the proposed method worked well in easy cases. The TCGA ovarian cancer data, however, were much more challenging to deal with due to the humongous number of variables and very limited sample size. The uncertainty of variable selection was discussed through gene selection results using random subsets of the data. The support vector machine with reject option was used to withhold decision for subjects who were difficult to classify. An ensemble method of combining models built on random subsets of the data was implemented to assess the prediction performance. Although we had applied these tools (DCOV, SVM-R, MCV) to biomedical data in the paper, we argue that they are quite portable across
disciplines.
As shown in Table \ref{t3}, a small portion of the model building population got classified for $d = 1/5$. Is it worthwhile to attempt the classification in such cases? It depends on the application, for example differential costs of two types of misclassification, and subjective considerations including quality of life influenced by the treatment, therapy expense and expected survival time.
Both the analysis of five fold cross validation and multiple cross validation showed the uncertainty of gene selection results based on different subsets of the data. The large number of variables that appeared only in a small number of runs suggested noises in the data and the difficulty caused by limited training sample size in the high dimensional scenario. It could also suggest the conundrum that the ``true'' model consists of a large number of variables with modest effects of which different subsets gives rise to roughly equal prediction ability. Options for further study in this and other difficult problems include allowing the DCOV stopping criteria to be modified by some amount $\epsilon$, and allowing the greedy variable selection algorithm to be doubly greedy by testing the next best $m$ of the remaining variables rather than just the next variable. It remains to obtain theoretical results to guide exploration in alternate scenarios.
The analysis of random permutation on the response served as both a
validation of our results and a discussion of what one is likely to
obtain without any true signal. If someone started with an entirely
different data set having the same proportions for the two classes with
that in the original data but no real signal at all, as what one might
get from scrambling, and went through every step, and finally obtained
a subgroup of patients with large voting score values, the result was no better than just guessing that everyone was sensitive. This experiment was also a cautionary tale that if one had not held out validation sets, the analyst could be easily fooled by spurious correlated variables and perfect training accuracy. Our proposed multiple cross validation and analysis through the voting scores provided protection against finding fake signals.
|
\section{Introduction}
\begin{comment}
In his seminal paper, Hopf (1948\nocite{Hopf1948}) conjuctered a
qualitatively mathematical picture of hydrodynamic turbulence: ``....
the corresponding phase flow in phase space $\Omega$ thus possesses
an extremely simple structure. The laminar solution represents a single
point in $\Omega$ invariant under the phase flow ..... To the flows
observed in the long run after the influence of the initial conditions
has died down there correspond certain solutions of the Navier-Stokes
equations. These solutions constitute a certain manifold $\mathcal{M}=\mathcal{M}(\mu)$
in phase space invariant under the phase flow. Presumably owing to
viscosity $\mathcal{M}$ has a finite number $N=N(\mu)$ of dimensions.''
\end{comment}
In the last decade, Navier-Stokes turbulence in channel flows has
been studied as a chaotic dynamics in the state space of a high-dimensional
system at moderate Reynolds numbers (see, for example, \cite{GibsonetalJFMCouette2008,WillisCvitanovic2013}).
Here, turbulence is viewed as an effective random walk in state space
through a repertoire of invariant solutions of the governing equations
(\cite{CvitanovicJFM2013_clockwork} and references therein). In state
space, turbulent trajectories visit the neighbourhoods of equilibria,
travelling waves or periodic orbits, switching from one saddle to
the other through their stable and unstable manifolds (\cite{CvitanovicPOT_1991},
see also \cite{ChaosBook}). Recent studies on the geometry of the
state space of Kolmogorov flows (\cite{Chandler_Kerswell2013_Kolm})
and barotropic atmospheric models (\cite{Gritsun2011,Gritsun2013})
give evidence that unstable periodic orbits provide the skeleton underpinning
the chaotic dynamics of fluid turbulence.
In pipe flows, the intrinsic continuous streamwise translation symmetry
and azimuthal discrete symmetry make difficult identifying invariant
flow structures, such as traveling waves or relative equilibria (\cite{FaisstEckhardt,WedinKerswell2004})
and relative periodic orbits (\cite{VISWANATH2007}), embedded in
turbulence. These travel downstream with their own mean velocity and
there is no unique comoving frame that can simultaneously reduce all
relative periodic orbits to periodic orbits and all travelling waves
to equilibria. Recently, this issue has been addressed by \cite{WillisCvitanovic2013}
using the method of slices (\cite{Siminos2011,Froehlich}, see also
\cite{RowleyMarsden2000,RowleyMarsden2003}) to quotient group symmetries
and reveal the geometry of the state space of pipe flows at moderate
Reynolds numbers. Further, \cite{Budanur2014} exploit the 'first
Fourier mode slice' to reduce the $SO(2)$-symmetry in spatially extended
systems. In particular, they separate the dynamics of the Kuramoto-Shivasinsky
equation into a `shape-changing' dynamics within a quotient or symmetry-reduced
space (base manifold) and a one-dimensional (1-D) transverse space
(fiber) associated with the group symmetry. This is the geometric
structure of a principal fiber bundle (e.g. \cite{Husemoller}), the
main topic of our work. In particular, we propose a generalization
of Hopf fibrations (\cite{Hopf1931}, see also \cite{Steenrod}) for
dynamical systems with translation symmetries, and apply it to symmetry-reduce
the evolution of passive scalars of turbulent pipe flows.
The paper is organized as follows. We first discuss the limitations
of the method of comoving frames, also referred to as the method of
connections (e.g. \cite{RowleyMarsden2000}). Then, we provide an
overview of principal fiber bundles and present the symmetry reduction
scheme via Hopf fibrations. As an application, two-dimensional (2-D)
Laser-Induced-Fluorescence (LIF) techniques are exploited to capture
planar fluorescent dye concentration fields tracing turbulent pipe
flow patterns at Reynolds number $\mathsf{Re}=2U_{b}R/\nu=3200$,
where $U_{b}$ is the bulk velocity, $R$ is the radius and $\nu$
is the kinematic viscosity of water. Symmetry reduction of the acquired
experimental data is then presented and discussed.
\section{Comoving frame velocities and Taylor's hypothesis}
Consider a 2-D passive scalar field $C(x,y,t)$ advected by a velocity
field $\mathbf{u}(x,y,t)=(U,V)$ and that evolves according to
\begin{equation}
\partial_{t}C+\mathbf{u}\cdot\nabla C=d\nabla^{2}C+f-s,\label{1}
\end{equation}
where $d$ is the diffusion coefficient, $f$ and $s$ are sources
and sinks, and ($x,y$) are the horizontal streamwise and vertical
cross-stream directions, respectively. The generalization to three-dimensional
geometries is straightforward, and it will not be discussed here.
Assume that (\ref{1}) admits streamwise translation symmetry, that
is if $C(x,y,t)$ is a solution so is $C(x-x_{d},y,t)$ for any drift
$x_{d}$. This can be time-varying and related to a comoving frame
velocity $U_{d}=\frac{dx_{d}}{dt}$, for which the material derivative
\begin{equation}
\frac{DC}{Dt}=\partial_{t}C+U_{d}\partial_{x}C
\end{equation}
is, in average, the smallest possible, namely
\begin{equation}
\left\langle \left(\partial_{t}C+U_{d}\partial_{x}C\right)^{2}\right\rangle _{x,y}
\end{equation}
is minimal if
\begin{equation}
U_{d}(t)=-\frac{\left\langle \partial_{t}C\partial_{x}C\right\rangle _{x,y}}{\left\langle \left(\partial_{x}C\right)^{2}\right\rangle _{x,y}},\label{Uc}
\end{equation}
where the brackets $\left\langle \cdotp\right\rangle {}_{x,y}$ denote
space average in $x$ and $y$. In the comoving frame $\left(x-U_{d}t,t\right)$,
the passive scalar appears to flow 'calmly', but still slowly drifting
(see, for example, \cite{KreilosJFM2014} for a study of Couette flows).
Only when $\frac{DC}{Dt}=0$, namely diffusion, sources and sinks
balance, the flow is steady in the comoving frame (\cite{KrogstadPoF}),
as for travelling waves (\cite{FaisstEckhardt,WedinKerswell2004}).
From (\ref{1}), (\ref{Uc}) can be written as
\begin{equation}
U_{d}(t)=\frac{\left\langle U\left(\partial_{x}C\right)^{2}+V\partial_{x}C\partial_{y}C-(f-s)\partial_{x}C\right\rangle _{x,y}}{\left\langle \left(\partial_{x}C\right)^{2}\right\rangle _{x,y}},
\end{equation}
which reveals that the comoving frame velocity is a weighted average
of the local flow velocities, sources and sinks. Although diffusion
processes are invariant under translation, they indirectly affect
$U_{d}$ through the concentration gradients. From (\ref{Uc}), averaging
along the $x$ direction only yields the vertical comoving frame velocity
profile
\begin{equation}
U_{d}(y,t)=-\frac{\left\langle \partial_{t}C\partial_{x}C\right\rangle _{x}}{\left\langle \left(\partial_{x}C\right)^{2}\right\rangle _{x}}.\label{Ucpr}
\end{equation}
The comoving frame speed $\widehat{U}_{d}$ of a Fourier mode $\widehat{C}(k,y,t)e^{ikx}$
then follows as
\begin{equation}
\widehat{U}_{d}(k,y,t)=-\mathrm{Im}\frac{\partial_{t}\widehat{C}(k,y,t)\overline{\widehat{C}}(k,y,t)}{k\left|\widehat{C}(k,y,t)\right|^{2}},\label{Uck}
\end{equation}
where $\overline{\widehat{C}}$ is the complex conjugate of $\widehat{C}$
and $\mathrm{Im}(a)$ denotes the imaginary part of $a$. Note that
$\widehat{U}_{d}$ is the same as the convective turbulent velocity
formulated by \cite{Alamo_JImenez2009} in the context of Taylor\textquoteright s
(1938) \nocite{Taylor1938} abstraction of turbulent flows as fields
of frozen eddies advected by the mean flow. If the turbulent fluctuation
$u$ is small compared to the mean flow speed $U_{m}$, then the temporal
response at frequency $\omega$ at a fixed point in space can be viewed
as the result of an unchanging spatial pattern of wavelength $2\pi/k$
convecting uniformly past the point at velocity $U_{m}$, viz. $U_{m}=\omega/k$.
This is the Taylor's hypothesis that relates the spatial and temporal
characteristics of turbulence. However, eddies can deform and decay
as they are advected by the mean flow and their speed may differ significantly
from $U_{d}$.
In this regard, \cite{Alamo_JImenez2009} concluded that the comoving
frame velocity $U_{d}$ of the largest-scale motion is close to the
local mean speed $U_{m}$, whereas $U_{d}$ drops significantly as
the scales are reduced (\cite{KrogstadPoF}). Hence, there is no unique
convection velocity, which insteads depends upon the state of evolution
of the flow. For example, it is well known that the turbulent motion
in channel flows is organized in connected regions of the near wall
flow that decelerate and then erupt away from the wall as \textquotedbl{}ejections\textquotedbl{}.
These decelerated motions are followed by larger scale connected motions
toward the wall from above as \textquotedbl{}sweeps\textquotedbl{}.
\cite{KrogstadPoF} found that the convection velocity for ejections
is distinctly lower than that for sweeps.
To gain more insights into the physical meaning of comoving frame
velocities, we have performed experiments to trace turbulent pipe
flow patterns exploiting non-intrusive LIF techniques (\cite{TianRoberts2003})
as discussed below.
\subsection{LIF measurements }
The experiments were performed in the Environmental Fluid Mechanics
Laboratory at the Georgia Institute of Technology. The LIF configuration
is illustrated in Fig. (\ref{FIGURE1}) and a detailed description
of the system is given in \cite{TianRoberts2003}. The tank has glass
walls $6.10$ m long \texttimes{} $0.91$ m wide \texttimes{} $0.61$
m deep. The front wall consists of two three-meter long glass panels
to enable long unobstructed view. The $5.5$ meter long pipe located
on the tank floor and tank was filled with water that was filtered
and dechlorinated. The pipe was transparent Perspex tube with radius
$R=2.5\mbox{ cm}$. The tank was filed up to well above the pipe to
avoid reflections of the laser sheet on the pipe wall. The water first
pomp into a damping chamber to get calm. Then, after passing a rigid
polyester air filter, it flowed into the pipe. A volume of fluorescent
dye solution continuously injected into the flow through a small hole
in the pipe wall upstream of the capture zone of length of $20R$.
The solution, a mixture of water and fluorescent dye, is supplied
from a reservoir by a rotary pump at a flowrate measured by a precision
rotameter. The flow was begun and, after waiting few minutes for the
flow to establish, laser scanning started to record the experiment.
To acquire high resolution data, we captured planar fluorescent dye
concentration fields $C(x,y,t)$ tracing turbulent pipe flow patterns
at Reynolds number $\mathsf{Re}=2U_{b}R/\nu=3200$ (bulk velocity
$U_{b}=6.42\mbox{ cm/s}$). As shown in Fig. (\ref{FIGURE1}), a laser
sheet was located at the center of the pipe to focus on flow properties
in the central plane ($y=0$ is the pipe centerline). Images of the
capture zone ($2R\times20R=5\times50\:\mbox{c\ensuremath{m^{2}}}$)
were acquired at $50$ Hz for a duration of $240$ seconds (see Fig.
(\ref{FIGURE1a})). Their vertical and horizontal resolutions are
of $65\times622$ pixels and $0.0794$ cm/pixel.
\begin{figure}[H]
\centering \includegraphics[width=0.99\textwidth]{LIFapparatus3.eps}
\protect\caption{Schematic of the LIF system of the Georgia Tech Environmental Fluid
Mechanics Laboratory (\cite{TianRoberts2003}). }
\label{FIGURE1}
\end{figure}
\begin{figure}[H]
\centering \includegraphics[width=1\textwidth]{logCsnapshot.eps} \protect\caption{LIF experiments: snapshot of the planar fluorescent dye concentration
field $C(x,y,t)$ (log scale) tracing turbulent pipe flow patterns
at Reynolds number $\mathsf{Re}=3200$ (bulk velocity $U_{b}=6.42\mbox{ cm/s}$,
flow from right to left). }
\label{FIGURE1a}
\end{figure}
\begin{figure}[H]
\centering \includegraphics[width=1\textwidth]{FIGURE_LARGESCALE_ALLSCALES_meanprofile_dynamical_velocity_spectrum.eps}
\protect\caption{Comoving frame velocity profile $U_{d}(y)$ estimated from (a) all
space scales, $U_{d,max}$$=6.32$ cm/s, (b) small scales (wavelengths
$L_{x}<0.2R$, $L_{y}<0.2R$), $U_{d,max}$$=2.76$ cm/s, and (c)
large scales ($L_{x}>2R$, $L_{y}>0.4R$), $U_{d,max}$$=8.52$ cm/s,
and pipe radius $R=2.5$ cm; (d) frequency spectrum of the large-scale
comoving frame velocity $U_{d}$ {[}see Eq. (\ref{Uc}){]}.}
\label{FIGURE2}
\end{figure}
\begin{figure}[H]
\centering \includegraphics[width=0.8\textwidth]{2Dspectrumcenterline.eps}
\protect\caption{Observed frequency-wavenumber spectrum of the fluorescent dye concentration
$C(x,y=0,t)$ at the pipe centerline. Estimated mean velocity $U_{m}=\omega/k\sim8.78$
cm/s (dashed line). $U_{m}/U_{b}=1.37$ and bulk velocity $U_{b}=6.42$
cm/s.}
\label{FIGURE3}
\end{figure}
\begin{figure}[H]
\centering \includegraphics[width=0.99\textwidth]{FIGURE_slowdown2.eps}
\protect\caption{LIF experiments: space-time evolution of the dye concentration $C(x,y=0,t)$
at the pipe centerline in the (left) lab frame ($x,t)$ and (center)
comoving frame $\mbox{\ensuremath{(x-U_{d}t,t)}}$; (right) normalized
concentration peak intensity $C/C_{max}$ tracked from the initial
time $t/T_{d}=0$ ($\Circle$) as function of the observed peak speed
$u/U_{d}$, with $C_{max}$ denoting the observed maximum value of
$C$. $U_{d}\approx$6.34 m/s and $T_{d}=U_{d}/R$.}
\label{FIGURE4}
\end{figure}
\begin{figure}[H]
\centering\includegraphics[width=0.75\columnwidth]{fiberbundle2.eps}
\protect\caption{Principal fiber bundle: a relative periodic orbit AB reduces to a
periodic orbit in the base manifold $\mathcal{M}$ by properly phase-shifting
the trajectory along the fiber $\mathcal{F}$ (or Lie-group space).
The shift is composed by a dynamical and geometric phases. The shift
induced by the dynamical phase yields the comoving trajectory A'C'B',
which is locally transversal to the fibers (parallel transport through
the fiber bundle), but it is not a closed trajectory. A further shift
by the geometric phase reduces A'C'B' to a periodic orbit on the base
manifold $\mathcal{M}$. }
\label{FIGURE5}
\end{figure}
\subsection{Data analysis}
The vertical comoving frame velocity profile can be estimated from
the measured fluorescent dye concentration field $C(x,y,t)$ using
Eq. (\ref{Ucpr}). For example, Fig. (\ref{FIGURE2}) shows $U_{d}(y)$
computed accounting for (Panel $a$) all space scales , (Panel \textbf{$\mathit{b}$})
small scales (wavelengths $L_{x}<0.2R$, $L_{y}<0.2R$) and (Panel
$c$) large scales of $C$ ($L_{x}>2R$, $L_{y}>0.4R$). Clearly,
small scales advect at lower speed than larger scales, in agreement
with \cite{KrogstadPoF}. Moreover, the maximum convective velocity
of large scales ($=8.52\mbox{ cm/s}$) is close to the centerline
mean flow speed (=$8.78$ cm/s) estimated from the frequency-wavenumber
spectrum of $C(x,y=0,t)$ {[}see Figure (\ref{FIGURE3}){]}. The frequency
spectrum of $U_{d}(t)$ estimated from Eq. (\ref{Uc}) accounting
for large scales only is also shown in Panel $d$ of Fig. (\ref{FIGURE2}).
It decays approximately as $f^{-5/3}$ as an indication that the Taylor's
hypothesis is approximately valid, possibly due to the non-dispersive
behavior of large scale motions.
In the lab frame $(x,t)$, the space-time evolution of the measured
fluorescent dye concentration $C(x,y=0,t)$ at the pipe centerline
is shown in the left panel of Figure (\ref{FIGURE4}). The associated
evolution in the comoving frame $\left(x-U_{d}t,t\right)$ is also
shown in the center panel. Here, $U_{d}$ is estimated from Eq. (\ref{Uc})
accounting for all space scales of $C$. Note the pattern-changing
dynamics of the passive scalar structures, which still experience
a drift in the comoving frame. Moreover, a generic slowdown or decelerated
motion is observed as the dye concentration bursts, possibly related
to the abovementioned turbulent flow ejections. This is clearly seen
in the the right panel of the same Figure, which reports the normalized
concentration peak intensity $C/C_{max}$ as function of the observed
peak speed $u/U_{d}$, with $C_{max}$ denoting the maximum value
of $C$. Further, $u$ is approximately 40-50\% larger than the comoving
frame velocity $U_{d}$, which is also roughly constant during the
event ($U_{d}=6.32\pm0.22$ cm/s). Note that in oceanic wave groups,
large focusing crests tend to slow down as they evolve within the
group, as a result of the natural wave dispersion of unsteady wave
trains (\cite{Banner_PRL2014,JFMFedele2014,FedeleEPL2014}). Thus,
we argue that the observed slowdown of the passive scalar bursts may
be due to the wave dispersive nature of small-scale turbulent structures.
In the following, drawing from differential geometry we will study
the fiber bundle structure of the state space associated with the
fluorescent dye concentration field evolution, which allows to explain
the observed excess speed $u-U_{d}$ of concentration bursts in terms
of geometric phases of the orbits in the bundle.
\section{Principal fiber bundles}
The geometric structure of the state space $\mathcal{P}\in\mathbb{R}^{N}$
of a dynamical system with a continuous Lie-group symmetry $G_{\alpha}$
and parameter $\alpha\in\mathbb{\mathcal{\mathbb{R}}}$, is that of
a principal fiber bundle: a base manifold $\mathcal{M}$ of dimension
$N-1$ (quotient space) and one-dimensional (1-D) fibers attached
to any point of $\mathcal{M}$ (e.g. \cite{Steenrod,Husemoller}).
The fiber $\mathcal{F}$ is the subspace of the Lie group orbit $G_{\alpha}(z)$.
One can think of the group as being an action, which pushes points
in the bundle around the bundle along the fibers (see Fig. \ref{FIGURE5}).
For example, the Euclidean space $\mathbb{R}^{3}$ can be seen as
a fiber bundle of parallel straight lines. The base manifold is a
plane cutting the whole set of parallel lines. This is a trivial fibration
since the total space $\mathcal{P}$ is both locally and globally
the direct product of the base $\mathcal{M}$ and the fiber $\mathcal{F}$,
that is $\mathbb{R}^{3}=\mbox{\ensuremath{\mathcal{M}\text{\texttimes}}\ensuremath{\mathbb{\mathcal{F}=}}}\mathbb{R}^{2}\times\mathbb{R}$.
A famous non trivial fiber bundle is the Hopf fibration of $S^{3}$
spheres by great circles $S^{1}$ and base space $S^{2}$ (\cite{Hopf1931},
see also \cite{Steenrod}). The Hopf fibration, like any fiber bundle,
is locally a product space, i.e. $S^{3}=S^{2}\times S^{1}$, but not
globally. There are numerous generalizations of it. For example, the
unit sphere $S^{2n+1}$ in the complex space $\mathbb{C}^{n+1}$ fibers
naturally over the complex projective space $\mathbb{C}P^{n}$ with
circles as fibers (see appendix).
A principal fiber bundle is denoted with the quadruplet $(\mathcal{P},\mathcal{M},G_{\alpha},\Pi)$
with total space $\mathcal{P}$ over the base manifold $\mathcal{M}$,
and a Lie group $G_{\alpha}$. The map $\Pi:\mathcal{P}\overset{}{\rightarrow}\mathcal{M}$
projects an element $z$ of the state space $\mathcal{P}$ and all
the elements of the group orbit $G_{\alpha}(z)$ into the same point
$\Pi(z)$ of the base manifold $\mathcal{M}$, viz. $\Pi(z)=\Pi(G_{\alpha}z)$,
with $\alpha\in\mathbb{R}$. In $\mathcal{P}$, a trajectory or orbit
$z(t)$ can be observed in a special comoving frame, within which
the motion is an horizontal transport through the fiber bundle, that
is the comoving orbit $y=G_{\alpha_{d}}z$ is locally transversal
to the fibers (see Fig. \ref{FIGURE5}). The proper shift $\alpha_{d}$
along the fibers to bring the motion in the comoving frame is called
dynamical phase. This increases with the time spent by the trajectory
to wander around $\mathcal{P}$ and system\textquoteright s answer
to: \textquotedblright{} how long did your trip take? \textquotedblright{}
(\cite{Berry08031984}). For example, the translational shift induced
by the constant speed of traveling waves, or relative equilibria,
is the dynamical phase. They reduce to equilibria within the base
manifold $\mathcal{M}$, whereas relative periodic orbits reduce to
periodic orbits (see Fig. \ref{FIGURE5}). In this case, the shift
along the fibers includes also a geometric phase $\alpha_{g}$, induced
by the projected motion within the $\mathcal{M}$ (\cite{Pancharatnam,Simon,Aharonov_Anandan,Garrison_GeomPhases}).
This phase is independent of time and it depends only upon the curvature
of $\mathcal{M}$, and system\textquoteright s answer to:\textquotedblright{}
where have you been? \textquotedblright{} (\cite{Berry08031984}).
Geometric phases arise due to anholonomy, that is global change without
local change. The classical example is the parallel transport of a
vector on a sphere. The change in the vector direction is equal to
the solid angle of the closed path spanned by the vector and it can
be described by Hannay's angles (\cite{Hannay}). The rotation of
Foucault\textquoteright s pendulum can be explained by means of such
a anholonomy. \cite{Pancharatnam} discovered this effect for polarized
light, and later on \cite{Berry08031984} rediscovered it for quantum-mechanical
systems. In fluid mechanics, \cite{Shapere1} exploited geometric
phases to explain self-propulsion at low Reynolds numbers. Note that
non-periodic orbits in $\mathcal{M}$ induce also a geometric phase,
that is the motion does not have to be periodic to have geometric
drift (\cite{Anandan_Nature}).
In summary, the total phase associated with any orbital path on the
base manifold is a measure of the motion induced within the fibers
of the bundle by the path. In the following, we will propose an Hopf
bundle for dynamical systems with translation symmetries.
\section{Symmetry reduction via Hopf fibrations}
For the sake of simplicity, consider a 1-D space-periodic passive
scalar field $C(x,t)$, which evolves according to
\begin{equation}
\partial_{t}C=\mathcal{N}(C),\label{ceq}
\end{equation}
where $\mathcal{N\mathrm{(\mbox{\ensuremath{C}})}}$ is a differential
operator of its argument. The extension to higher dimensions is straightforward.
The dynamics admits a continuous translation symmetry, that is if
$C(x,t)$ is a solution, so is $C(x+x_{s},t)$ for any drift $x_{s}$.
\begin{comment}
The group tangent space at $c$ is
\begin{equation}
T_{x_{0}}(c)=(G_{x_{0}}^{-1}\partial_{x_{0}}G)c=\partial_{x}c.
\end{equation}
\end{comment}
It is convenient to express $C$ as the Fourier series
\begin{equation}
\begin{array}[t]{c}
C(x,t)=C_{0}(t)+\frac{1}{2}\sum_{m=1}^{N}z_{m}(t)\exp\left(imk_{0}x\right)+c.c.=\\
\\
C_{0}(t)+\sum_{m=1}^{N}\left|z_{m}\right|\cos\left(imk_{0}x+\theta_{m}\right),
\end{array}\label{cf}
\end{equation}
where $C_{0}(t)$ is the mean, $z_{m}=\left|z_{m}\right|\exp(i\theta_{m})$
are complex Fourier amplitudes with phases $\theta_{m}$ and $k_{0}=2\pi/L_{0}$,
with $L_{0}$ denoting the domain length. The mean $C_{0}$ is invariant
under the group action and it can be decoupled from the vector $z(t)=\{z_{m}\}=\left(z_{1},...z_{N}\right)$.
This satisfies the dynamical system
\begin{equation}
\frac{dz}{dt}=\mathcal{\widehat{N}}(z,C_{0}),\label{ODE}
\end{equation}
coupled with
\begin{equation}
\frac{dC_{0}}{dt}=\left\langle \mathcal{N}(C_{0},z)\right\rangle _{x},\label{ceq-1}
\end{equation}
the space average of (\ref{ceq}). The vector
\begin{equation}
\mathcal{\widehat{N}}(z,C_{0})=\left\{ \widehat{\mathcal{N}}_{m}(z,C_{0})\right\} =\left(\widehat{\mathcal{N}}_{1}(z,C_{0}),...,\widehat{\mathcal{N}}_{N}(z,C_{0})\right),\label{ODE-1-1}
\end{equation}
follows from the discrete Fourier transform of $\mathcal{N}$ in Eq.
(\ref{ceq}).
The orbit $z(t)$ wanders in the state space $\mathcal{P}\in\mathbb{C^{\mathrm{\mathit{N}}}}$,
and the symmetry group $G_{x_{s}}$ of $z$ is the subspace
\begin{equation}
G_{x_{s}}(z)=\left\{ w\in\mathbb{C^{\mathit{N}}}:w=\{z_{m}\exp(imk_{0}x_{s})\},\:\forall x_{s}\in\mathbb{R}\right\} .
\end{equation}
The state space $\mathcal{P}$ is geometrically a principal fiber
bundle: a base manifold $\mathcal{M}$ of dimension $2N-1$ (quotient
space) and one dimensional (1-D) fibers $G_{x_{s}}(Z)$ attached to
any point $Z\in\mathcal{M}$ (e.g. \cite{Steenrod,Husemoller}). The
bundle is described by the quadruplet $(\mathcal{P},\mathcal{M},G_{x_{s}},\Pi_{j})$
and the map $\Pi_{j}$ is given as follows. For a non-vanishing Fourier
mode $z_{j}$, the desymmetrized orbit $z_{D}(t)$ within the base
manifold $\mathcal{M}\in\mathbb{C^{\mathit{N}}}$ is given by the
map $\Pi_{j}:\mathit{\mbox{\mbox{\ensuremath{\mathcal{\mathrm{\mathcal{P}}}}}}}\rightarrow\mathcal{M}$
\begin{equation}
z_{D}=\Pi_{j}(z)=Z=(Z_{1},...Z_{m},...Z_{N}),\label{ZD}
\end{equation}
where the complex amplitudes
\begin{equation}
Z_{m}=\left|z_{m}(t)\right|\exp(i\phi_{m}(t)),
\end{equation}
and phases
\begin{equation}
\phi_{m}=\theta_{m}-\frac{m\theta_{j}}{j}.
\end{equation}
Note that $Z_{j}$ is real and $\mathcal{M}$ is a $2N-1$ dimensional
manifold of $\mathbb{C}^{N}$. For $j>1$, the phase $\theta_{j}$
needs to be unwrapped in order to avoid introducing spurious discrete
symmetries. For $j=1$, the fibration reduces to the 'first Fourier
mode slice' proposed in \cite{Budanur2014}. The map $\Pi_{j}$ projects
an element $z=\{z_{m}\}$ of the state space $\mathcal{P}$ and all
the elements of its group orbit $G_{x_{d}}(z)$ into the same point
$z_{D}=\Pi_{j}(z)$ in $\mathcal{M}$, namely
\begin{equation}
\Pi_{j}(z)=\Pi_{j}(G_{x_{s}}(z))\mathrm{.}
\end{equation}
The scalar field $C_{D}$ in the symmetry-reduced frame follows from
(\ref{cf}) as
\begin{equation}
C_{D}(x,t)=C_{0}(t)+\sum_{m=1}^{N}\left|Z_{m}\right|\cos\left(mk_{0}x+\phi_{m}\right).\label{CD}
\end{equation}
Note that $\Pi_{j}$ can be written in the equivalent form
\begin{equation}
\begin{array}[t]{c}
z_{D}=\Pi_{j}(z)=\ensuremath{\left\{ z_{m}\left(\frac{\left|z_{j}\right|}{z_{j}}\right)^{m/j}\right\} }_{m=1,...N}=\\
\\
\left(z_{1}\left(\frac{\left|z_{j}\right|}{z_{j}}\right)^{1/j},...z_{j-1}\left(\frac{\left|z_{j}\right|}{z_{j}}\right)^{(j-1)/j}\mbox{\ensuremath{,}\ensuremath{\left|z_{j}\right|,z_{j+1}}\text{\ensuremath{\left(\frac{\left|z_{j}\right|}{z_{j}}\right)^{(j+1)/j}}\ensuremath{,...}}\ensuremath{z_{N}\left(\frac{\left|z_{j}\right|}{z_{j}}\right)^{N/j}}}\right),
\end{array}\label{map-1}
\end{equation}
which reveals the geometric structure of a Hopf bundle (e.g. \cite{Steenrod}).
In the following, it is shown that the associated base manifold $\mathcal{M}$,
hereafter referred to as $\mathbb{C}T^{n}$, is a generalization of
$\mathbb{C}P^{n}$ projective spaces.
\subsection{$\mathbb{C}T^{n}$ projective spaces }
The complex projective space $\mathbb{C}P^{n}$ is the quotient space
of the unit $S^{2n+1}$ hypersphere
\begin{equation}
S^{2n+1}=\left\{ z=(z_{1},...z_{n+1})\in\mathbb{C}^{n+1}:\sum_{j=1}^{n+1}\left|z_{j}\right|^{2}=1\right\}
\end{equation}
with circles $S^{1}$ as fibers under the action of the $U(1)$ group
(see also appendix). The projection $\pi_{H}:S^{2n+1}\rightarrow\mathbb{C}P^{n}$
is called the Hopf map. Since $\mathbb{C}P^{1}=S^{2}$ is the Riemann
sphere, we obtain the classical Hopf bundle $S^{3}\rightarrow S^{2}$
with fiber $S^{1}$ (\cite{Hopf1931}, see also \cite{Steenrod}).
Hereafter, we generalize $\mathbb{C}P^{n}$ spaces to complex manifolds
defined as quotient spaces
\begin{equation}
\mathbb{C}T^{n}=E^{2n+1}/T(1)
\end{equation}
of the $E^{2n+1}$ hypersurface in $\mathbb{C}^{n+1}$
\begin{equation}
E^{2n+1}=\left\{ z=(z_{1},...z_{n+1})\in\mathbb{C}^{n+1}:\sum_{j=1}^{n+1}\left|z_{j}\right|^{\frac{2(n+1)}{j}}=1\right\}
\end{equation}
under the action of the translation group $T(1)$
\begin{equation}
T_{\lambda}(z)=\left\{ Z=(Z_{1},...Z_{n+1})\in\mathbb{C}^{n+1}:Z_{k}=z_{k}\lambda^{k},\quad\left|\lambda\right|=1\right\} .
\end{equation}
An element $z$ of $\mathbb{C}T^{n}$ is identified with the equivalence
class
\[
[z]=[z_{1},...z_{n+1}]\sim(\lambda z_{1},...\lambda^{n+1}z_{n+1}),\qquad\lambda\neq0.
\]
Two points of $\mathbb{C}^{n+1}\left\backslash \left\{ 0\right\} \right.$
are equivalent if they lie on the same subspace or group orbit $T_{\lambda}(z)$.
For $n=1$, the group orbit is the space of complex parabolas $z_{2}=\lambda z_{1}^{2}$
($z_{1}\neq0$) and, for $n>1$, that of complex hypercurves $z_{n}=\lambda z_{j}^{n/j}$
($z_{j}\neq0$). All points of $T_{\lambda}(z)$ reduce to the same
point of $\mathbb{C}T^{n}$ and the projection map $\pi:\mathbb{C}^{n+1}\rightarrow\mathbb{C}T^{n}$
can be defined as follows.
The $\mathbb{C}T^{n}$ manifold can be covered by an atlas of charts
or slices $U_{j}=\left\{ [z]\left|z_{j}\neq0\right.\right\} \subset\mathbb{C}T^{n}$,
where the complex hyperplane $B_{j}=\left\{ z_{j}=0\right\} $ is
the ``border'' of $U_{j}$ (\cite{Siminos2011}). Then, the projection
onto a chart or slice $U_{j}$ is given by the map $\pi_{j}:\mathbb{C}^{n+1}\rightarrow U_{j}$
\begin{equation}
\pi_{j}([z])=\pi_{j}([z_{1},...z_{n+1}])=\left\{ z_{k}\left(\frac{\left|z_{j}\right|}{z_{j}}\right)^{k/j}\right\} _{k=1,...n+1},
\end{equation}
which is the same map as in Eq. (\ref{map-1}) for $n=N-1$. Thus,
the fiber bundle formulated in the previous section is generalization
of a Hopf fibration for dynamical systems with $T(1)$ symmetries.
The associated state space fibrates over a $\mathbb{C}T^{n}$ manifold
with fibers as the hypercurves $T_{\lambda}(z)$. fibrates over a
$\mathbb{C}T^{n}$ manifold with fibers as the hypercurves $T_{\lambda}(z)$.
Within the charts or slices, relative equilibria reduce to equilibria
and relative periodic orbits reduce to periodic orbits. Clearly, when
$z_{j}$ approaches zero, but likely never completely vanishes, the
trajectory in $\mathbb{C}T^{n}$ wanders nearby the border $B_{j}$
of the slice $U_{j}$ and the associated geometric phase tends to
become singular (see, for example, \cite{Budanur2014}). As described
in \cite{Siminos2011}, a different chart can then be chosen and the
charts' borders can be glued together via ridges into an atlas that
spans the state space region of interest.
\begin{comment}
`For the Porter-Knobloch system we can use the $TP^{1}$ space, which
is the quotient space of the $E^{3}$ hypersurface
\[
E^{3}=\left\{ (z_{1},z_{2})\in C^{2}:\left|z_{1}\right|^{4}+\left|z_{2}\right|^{2}=1\right\}
\]
under the $T(1)$ translation group. The maps $\pi_{j}$ are
\[
\pi_{1}([z_{1},z_{2}])=\frac{z_{2}}{z_{1}^{2}},\qquad\pi_{2}([z_{1},z_{2}])=\frac{z_{1}^{2}}{z_{2}}.
\]
\end{comment}
\subsection{Dynamical and geometric phases}
The desymmetrized orbit $z_{D}(t)=Z(t)=\left\{ Z_{m}(t)\right\} $
within the base manifold $\mathcal{M}$ is determined by properly
shifting $z(t)=\left\{ z_{m}(t)\right\} $ along the fibers by $x_{s}(t)$,
that is $Z=G_{-x_{s}}(z)$ and components
\begin{equation}
z_{m}=Z_{m}\exp(imk_{0}x_{s}),\qquad m=1,...N,\label{zn}
\end{equation}
To find $x_{s}$, we impose the condition of transversality to the
fiber or group orbit
\begin{equation}
\overline{T_{x_{s}}(Z)}\frac{dZ}{dt}=0,\label{orth}
\end{equation}
namely the tangent $\frac{dZ}{dt}$ at $Z$ must be orthogonal to
the group tangent space
\begin{equation}
T_{x_{s}}(Z)=(G_{x_{s}}^{-1}\partial_{x_{s}}G)Z=\{imk_{0}Z_{m}\}.
\end{equation}
From (\ref{ODE}) and (\ref{zn}), (\ref{orth}) yields
\[
\overline{T_{x_{s}}(Z)}\left(\frac{dZ}{dt}+\frac{dx_{s}}{dt}T_{x_{s}}(Z)-\widehat{\mathcal{N}}(Z,C_{0})\right)=0,
\]
from which
\begin{equation}
\frac{dx_{s}}{dt}=\frac{dx_{d}}{dt}+\frac{dx_{g}}{dt},
\end{equation}
where
\begin{equation}
\frac{dx_{d}}{dt}=U_{d}=\frac{\mathrm{Re}\left[\overline{T_{x_{s}}(Z)}\widehat{\mathcal{N}}(Z,C_{0})\right]}{\left|T_{x_{s}}(Z)\right|^{2}}=-\frac{\mathrm{Im}\sum mk_{0}\overline{Z_{m}}\widehat{\mathcal{N}}_{m}(Z,C_{0})}{\sum m^{2}k_{0}^{2}\left|Z_{m}\right|^{2}}\label{dynphase}
\end{equation}
is the velocity associated with the dynamical phase $x_{d}$, and
\begin{equation}
\frac{dx_{g}}{dt}=U_{g}=-\frac{\mathrm{Re}\left(\overline{T_{x_{s}}(Z)}\frac{dZ}{dt}\right)}{\left|T_{x_{s}}(Z)\right|^{2}}=\frac{\mathrm{Im}\sum mk_{0}\overline{Z_{m}}\frac{dZ_{m}}{dt}}{\sum m^{2}k_{0}^{2}\left|Z_{m}\right|^{2}}\label{geomphase}
\end{equation}
is that associated with the geometric phase $x_{g}$. Here, the sum
is over $m=1,...N$ and $\mathrm{\mathrm{Re}\mbox{\ensuremath{(\mbox{\ensuremath{a}})}},Im}(a)$
denote the real and imaginary parts of $a$. Note that $\overline{T_{x_{s}}(Z)}\widehat{\mathcal{N}}(Z,C_{0})$
and $\left|T_{x_{s}}(Z)\right|^{2}$ in (\ref{dynphase}) are invariant
under the $T(1)$ group action, and $x_{d}$ can also be determined
replacing $Z$ with the orbit $z$ in $\mathcal{P}$, which is usually
known or observable in applications. The total shift or drift
\begin{equation}
x_{s}=x_{d}+x_{g},\label{drift}
\end{equation}
where
\begin{equation}
x_{d}(t)=\int_{0}^{t}V_{d}d\tau,\qquad x_{g}(t)=\int_{0}^{t}V_{g}d\tau.\label{xdxg}
\end{equation}
For the Hopf bundle given by the map (\ref{ZD}) or equivalently (\ref{map-1}),
the total drift
\[
x_{s}=-\frac{\theta_{j}}{k_{0}j},\qquad j\geq1.
\]
The dynamical phase $x_{d}$ increases with the time spent by the
trajectory $z(t)$ to wander around $\mathcal{P}$, whereas the geometric
phase $x_{g}$ depends upon the path $\gamma$ associated with the
motion of $Z(t)$ in the base manifold $\mathcal{M}$. Indeed, from
(\ref{geomphase}) $x_{g}$ can be rewritten as the path integral
\begin{equation}
x_{g}=-\int_{\gamma}\frac{\mathrm{Re}\left(\overline{T_{x_{s}}(Z)}dZ\right)}{\left|T_{x_{s}}(Z)\right|^{2}}=\int_{\gamma}\frac{\mathrm{Im}\sum mk_{0}\overline{Z_{m}}dZ_{m}}{\sum m^{2}k_{0}^{2}\left|Z_{m}\right|^{2}}.\label{xg}
\end{equation}
Clearly, the geometric phase is independent of time and it depends
only upon the reduced motion within $\mathcal{M}$.
It is straightforward to recognize that $U_{d}$ in (\ref{dynphase})
is the comoving frame velocity introduced in (\ref{Uc}) and specialized
for the dynamical system (\ref{ODE}). In particular, if we shift
the orbit $z(t)$ along the fibers by $x_{d}(t)$, we obtain the comoving
frame orbit
\[
Z_{d}(t)=G_{-x_{d}}(z)=\left\{ z_{m}(t)\exp(-imk_{0}x_{d}(t)\right\}
\]
that moves through the fiber bundle locally transversal to the fibers.
This motion is referred to as an ``horizontal transport'' through
the fiber bundle. In physical space, this corresponds to an evolution
in the comoving frame $\left(x-U_{d}t,t\right)$ (see Fig. (\ref{FIGURE4})).
In general, $Z_{d}$ still experiences a shift, the geometric $x_{g}$,
if any. Indeed, the projected path $z_{D}=Z$ within the base manifold
$\mathcal{M}$ is obtained by further shifting $Z_{d}$ along the
fibers by $x_{g}$ (see Fig. \ref{FIGURE5}), namely
\[
Z(t)=G_{-x_{g}}(Z_{d})=G_{-x_{d}-x_{g}}(z)=\left\{ z_{m}(t)\exp\left[-imk_{0}\left(x_{d}+x_{g}\right)\right]\right\} .
\]
In physical space, this corresponds to a pattern-changing dynamics
in the symmetry-reduced frame $\left(x-(U_{d}+U_{g})t,t\right)$ as
discussed later on. Only relative equilibria or traveling waves have
null geometric phase, since their pattern is not dynamically changing
in the base manifold as they reduce to equilibria. In this case the
comoving and symmetry-reduced frames are the same.
\section{Symmetry reduction of LIF data}
In this section, we apply Hopf fibrations to symmetry-reduce LIF measurements
of turbulent pipe flows at $\mathsf{Re}=3200$ (see section 2.1).
Figure (\ref{FIGURE6}) illustrates the space-time evolution of a
passive scalar burst event. The left panel shows the fluorescent dye
concentration $C(x,y=0,t)$ at the pipe centerline in the lab frame
$(x,t)$. A drift in the streamwise direction $x$ is observed. The
corresponding orbit $z(t)$ in the subspace \{$\mbox{Re}(z_{11}),\mbox{Im}(z_{13}),\mbox{Re}(z_{15})$\}
of the state space $\mathcal{P}$ of dimensions $N=65\times622=40430$
is shown in the left panels of Fig. (\ref{FIGURE7}). Note that the
excursion of the orbit while the concentration $c$ lingers above
the threshold $0.8C_{max}$ is complicated (bold line) since it wanders
around its group orbit $G_{x_{s}}(z)$ as a result of the drift induced
by the translation symmetry. The center panel of Fig. (\ref{FIGURE6})
shows the space-time evolution in the comoving frame $\left(x-U_{d}t,t\right)$.
Note that the dye concentration patterns still experience a significant
drift, the geometric phase $x_{g}$ at velocity $U_{g}$ {[}see Eqs.
(\ref{geomphase}),(\ref{xg}){]}. As a result, the associated orbit
in state space still wanders around the group orbit $G_{x_{s}}(z)$.
The drift disappears in the symmetry-reduced frame $\left(x-(U_{d}+U_{g})t,t\right)$
associated with the dynamics in the chart or slice $U_{36}$ of the
base manifold $\mathcal{M}$, as shown in the right panel of Figure
(\ref{FIGURE6}). Here, one slice is sufficient to symmetry-reduce
the orbit $z$ over the analyzed time span as its component $z_{36}$
never lingers closely near zero. A pattern-changing dynamics of the
passive scalar structures is revealed as clearly seen in Figure (\ref{FIGURE8}).
The corresponding symmetry-reduced orbit $z_{D}(t)=Z(t)$, computed
from Eq. (\ref{ZD}), is shown in the right panels of Fig. (\ref{FIGURE7})
and plotted in the subspace \{$\mbox{Re}(Z_{11}),\mbox{Im}(Z_{13}),\mbox{Re}(Z_{15})$\}
of $\mathcal{M}$. Here, the excursion of the orbit while the concentration
$C$ is high (bold line) indicates an homoclinic type behavior.
\begin{figure}[H]
\centering \includegraphics[width=0.99\textwidth]{FIGURE_desymmetrizedspace_singleevent4.eps}
\protect\caption{Symmetry reduction of LIF data: space-time evolution of a passive
scalar burst event; measured concentration $C(x,y=0,t)$ at the pipe
centerline in the (left) lab frame $(x,t)$, (center) comoving frame
$\left(x-U_{d}t,t\right)$ and (right) symmetry-reduced frame $\left(x-(U_{d}+U_{g})t,t\right)$;
time average $U_{d}\approx6.74$ cm/s, $U_{g}\approx0.4U_{d}$ and
$T_{d}=U_{d}/R$.}
\label{FIGURE6}
\end{figure}
\begin{figure}[H]
\centering \includegraphics[width=0.99\textwidth]{FIGURE_statespace_singleevent4.eps}
\protect\caption{Symmetry reduction of LIF data: (left) Orbit trajectories in the subspace
($\mbox{Re}(z_{11}),\mbox{Im}(z_{13}),\mbox{Re}(z_{15})$) of the
state space $\mathcal{P}$ associated with the passive scalar dynamics
in the lab frame of Fig. 7; (right) corresponding symmetry-reduced
orbit within the chart or slice $U_{36}$ of the base manifold $\mathcal{M}$
(subspace ($\mbox{Re}(Z_{11}),\mbox{Im}(Z_{13}),\mbox{Re}(Z_{25})$).
The bold line indicates the excursion of the orbit while the concentration
$C$ lingers above the threshold $0.8C_{max}$ ($\Circle$ =initial
time, $\times$=final time).}
\label{FIGURE7}
\end{figure}
\begin{figure}[h]
\centering \includegraphics[width=1\textwidth]{patternchanging_singleevent4.eps}
\protect\caption{Symmetry reduction of LIF data: pattern-changing profiles at increasing
instants of time (from bottom to top) of the measured concentration
$C(x,y=0,t)$ at the pipe centerline in the symmetry-reduced frame
$\left(x-(U_{d}+U_{g})t,t\right)$. Associated 2-D pattern is shown
in the right panel of Fig. (\ref{FIGURE6}).}
\label{FIGURE8}
\end{figure}
\begin{figure}[H]
\centering \includegraphics[width=0.7\textwidth]{Phasespeedssingleevent4.eps}
\protect\caption{Symmetry reduction of LIF data: dynamical and geometric velocities
$U_{d}$ and $U_{g}$ associated with the orbit in state space of
Fig. (\ref{FIGURE7}).}
\label{FIGURE9}
\end{figure}
\begin{figure}[h]
\centering \includegraphics[width=1\textwidth]{FIGURE_PDF_MEANFLOW_new.eps}
\protect\caption{LIF experiments: (left) observed normalized dye concentration peak
speed $u/U_{d}$ as function of the amplitude peak $C/C_{max}$, and
(right) associated probability density function. }
\label{FIGURE10}
\end{figure}
Note that the speed $u\approx U_{d}+U_{g}$ of the dye concentration
burst is approximately 40\% larger than the comoving frame velocity
$U_{d}$, which slighlty changes during the event, whereas the geometric
phase velocity $U_{g}\approx0.4U_{d}$ as seen in Figure (\ref{FIGURE9}).
This appears a generic trend of the flow as clearly seen in Fig. \ref{FIGURE10},
which shows the observed normalized speed $u/U_{d}$ of dye concentration
peaks tracked in space as function of their amplitude $C/C_{max}$,
and the associated probability density function. As the peak amplitude
of concentration bursts increases, their speed $u$ tend to be close
to $1.43U_{d}$. Thus, the space-time evolution in the comoving frame
cannot explain the excess speed $\delta u=u-U_{d}$ (see center panel
of Fig. \ref{FIGURE9}). Indeed, $\delta u$ is roughly the geometric
phase velocity $U_{g}$ induced by the motion in the symmetry-reduced
state space $\mathcal{M}$. In physical space, this results in the
pattern-changing dynamics of the passive scalar structures revealed
in the symmetry-reduced frame (see Figure (\ref{FIGURE8})). Only
when geometric phases are small, viz. $U_{g}\ll U_{d}$, Taylor's
approximation is valid and the flow structures slightly deform as
they are advected at the comoving frame speed $U_{d}$.
\section{Conclusions}
We have presented a theoretical study on the fiber bundle structure
of the state space of dynamical systems with continuous translation
symmetries. A generalization of Hopf fibrations is proposed to quotient
such symmetries in the complex manifold $\mathbb{C}T^{n}$, which
generalizes $\mathbb{C}P^{n}$ projective spaces. As an application,
we have exploited LIF techniques to capture planar fluorescent dye
concentration fields tracing a turbulent pipe flow at Reynolds number
$\mathsf{Re}=3200$. The symmetry reduction of LIF measurements unveils
that the motion of the passive scalar structures is associated with
the dynamical and geometric phases of the corresponding orbits in
the Hopf bundle. In particular, the observed speed $u\approx U_{g}+U_{d}\approx1.43U_{d}$
of dye concentration bursts exceeds the comoving or convective velocity
$U_{d}$. The excess speed $u-U_{d}$ can be explained in terms of
the pattern-changing dynamics in the symmetry-reduced frame and associated
geometric phase velocity $U_{g}\approx0.43U_{d}.$
Hopf fibrations are promising for symmetry reduction of three-dimensional
LIF and PIV measurements as well as simulated flows of pipe turbulence,
in order to unveil the skeleton of its chaotic dynamics in state space.
Further studies along these directions are desirable.
\section{Acknowledgments}
FF acknowledges the Georgia Tech graduate courses ``Classical Mechanics
II'' taught by Prof. Jean Bellissard in Spring 2013 and ``Nonlinear
dynamics: Chaos, and what to do about it?{}`` taught by Prof. Predrag
Cvitanovic in Spring 2012. FF also thanks Prof. Alfred Shapere for
discussions on geometric phases, and Profs. Predrag Cvitanovic, Bruno
Eckhardt and Dr. Evangelos Siminos for discussions on symmetry reduction.
\section{Appendix: $\mathbb{C}P^{n}$ complex projective spaces }
A $\mathbb{C}P^{n}$ space is the set of 1-dimensional complex-linear
subspaces, or lines of the complex space $\mathbb{C}^{n+1}$. In particular,
$\mathbb{C}P^{n}$ is the quotient of $\mathbb{C}^{n+1}\left\backslash \left\{ 0\right\} \right.$
by the equivalent relation
\begin{equation}
z\sim w\Longleftrightarrow\exists\lambda\in\mathbb{C}\left\backslash \left\{ 0\right\} \right.:\: w=\lambda z,\qquad\left|\lambda\right|=1.
\end{equation}
Two points of $\mathbb{C}^{n+1}\left\backslash \left\{ 0\right\} \right.$are
equivalent iff they lie on the same line, viz. they are complex linear
independent. We denote the equivalence class of $z$ by $[z]$, which
is an element of the $U(1)$ group.
The projection map $\pi:\mathbb{C}^{n+1}\rightarrow\mathbb{C}P^{n}$
can be defined as follows. Consider
\begin{equation}
z=(z_{1},...z_{n+1})\in\mathbb{C}^{n+1},\qquad z\neq0,
\end{equation}
then $\mathbb{C}P^{n}$ is a complex manifold, which can be covered
by charts $U_{j}=\left\{ [z]\left|z_{j}\neq0\right.\right\} \subset\mathbb{C}P^{n}$,
i.e. the space of all lines not contained in the complex hyperplane
or border $B_{j}=\left\{ z_{j}=0\right\} $. Then, the maps $\pi_{j}:\mathbb{C}^{n+1}\rightarrow U_{j}$
are defined as
\begin{equation}
\begin{array}[t]{c}
\pi_{j}([z])=\pi_{j}([z_{1},...,z_{n+1}])=\left(z_{k}\left(\frac{\left|z_{j}\right|}{z_{j}}\right)\right)_{k=1,...n+1}\\
\\
\left(z_{1}\left(\frac{\left|z_{j}\right|}{z_{j}}\right),...z_{j-1}\left(\frac{\left|z_{j}\right|}{z_{j}}\right),\left|z_{j}\right|,z_{j+1}\left(\frac{\left|z_{j}\right|}{z_{j}}\right),...z_{n+1}\left(\frac{\left|z_{j}\right|}{z_{j}}\right)\right).
\end{array}
\end{equation}
It is straightforward to see that $\mathbb{C}P^{n}$ is the quotient
space of a unit $S^{2n+1}$ sphere under the action of the $U(1)$
group, namely
\begin{equation}
\mathbb{C}P^{n}=S^{2n+1}/U(1).
\end{equation}
Indeed, every line in $\mathbb{C}^{n+1}$ intersects the unit $S^{2n+1}$
sphere in a circle $S^{1}$, and we obtain a point of $\mathbb{C}P^{n}$
defined by this line by identifying all points on $S^{1}$. %
\begin{comment}
The projection $\pi:S^{2n+1}\rightarrow CP^{n}$ is called the Hopf
map. Since $CP^{1}=S^{2}$ is the Riemann sphere, we obtain the classical
Hopf bundle $S^{3}\rightarrow S^{2}$ with fiber $S^{1}$ \cite{Hopf1931}.
\end{comment}
\begin{comment}
\section{Appendix B}
For a 2-D passive scalar
\begin{equation}
\begin{array}[t]{c}
C(x,y,t)=C_{0}(t)+\frac{1}{2}\sum_{p,q=1}^{N_{x},N_{y}}z_{p,q}(t)\exp\left(ipk_{x}x+iqk_{y}y\right)+c.c.=\\
\\
C_{0}(t)+\frac{1}{2}\sum_{p,q=1}^{N_{x},N_{y}}\left|z_{p,q}\right|\mbox{cos}\left(pk_{x}x+qk_{y}y+\theta_{p,q}\right),
\end{array}\label{cf-1}
\end{equation}
where $C_{0}(t)$ is the mean, $z_{p,q}=\left|z_{p,q}\right|\exp(i\theta_{p,q})$
are complex Fourier amplitudes with phases $\theta_{p,q}$ and $k_{x,y}=2\pi/L_{x,y}$
with $L_{x,y}$ denoting the domain lengths along $x$ and $y$. Now,
reorder the double array $\{z_{p,q}$\} as a single vector $z={z_{m}}_{m=1,...N}$
with $N=N_{x}N_{y}$. Then, the evolution of $z_{m}$ is governed
by the dynamical system
\begin{equation}
\frac{dz_{m}}{dt}=\mathcal{\widehat{N}}_{m}(z_{1},z_{2},....z_{N}),\label{ODE-1}
\end{equation}
\end{comment}
\bibliographystyle{jfm}
|
\section{INTRODUCTION}
\label{sec:intro}
From the first theories on the origin of our solar system in the 18th century
(by Emanuel Swedenborg, Immanuel Kant, and Pierre-Simon Laplace)
up to the late 1990's, our understanding of the planet formation process has been
guided solely by the characteristics observed in our own planetary system.
This resulted in the classical ``core-accretion theory'',
where the planets are assembled already at their final location through dust coagulation
and agglomeration processes (e.g.\ Pollack et al.\cite{pol96}).
The discovery of the first extrasolar planet around a main-sequence star
by Mayor \& Queloz\cite{may95} and the findings of the subsequent exoplanet surveys
changed this view dramatically and revealed a surprising diversity in planetary
system architecture. Many of the detected systems exhibit Jupiter-mass planets
that orbit their host star at separations of less than 0.1\,astronomical units (``Hot Jupiter'')
or planets with masses of $\sim 10$ Earth masses (``Super-Earth'').
Neither of these planet populations is observed in our solar system, which raises
the important question of what determines the diversity of these systems
and the properties of the assembled planets.
In order to account for this menagerie of planetary systems, various planet migration mechanisms
and alternative formation scenarios (such as the gravitational instability model) have been proposed.
Accordingly, in state-of-the-art population synthesis models the final system architecture
depends on a plethora of parameters, describing the initial conditions of the protoplanetary disk,
the planetesimal formation mode, the interaction between the planet and disk (type I/II migration),
the presence of migration traps (deadzones, disk truncation, ...),
planet-planet scattering events (resonances, planet ejection, ...),
environmental factors affecting the disk evolution, and
finally scattering caused by the planetesimal disk.
These new ideas make planet formation one of the most vibrant and
active research fields, but reveal also the complexity that results from the large
number of involved processes.
It is becoming increasingly clear that new observational constraints are needed,
both to determine the initial conditions of the planet formation process and to identify the
dominant mechanisms that govern the assembly and orbital evolution of planetary systems.
With the Planet Formation Imager (PFI) project, we argue that these urgently needed
observational constraints could be obtained with a new high-angular resolution imaging facility
that would be optimized to probe these processes on the natural spatial scale where
planet formation is taking place. This natural spatial scale is the ``Hill Sphere'', which
defines the gravitational sphere of influence of the forming planet.
The Hill Sphere of a Jupiter-mass planet at the location of Jupiter in our solar system
is 0.35\,astronomical units (au, for $r=5.2$\,au) and 0.07\,au for a Jupiter-mass planet at $r=1$\,au.
From these ambitious goals we launch our efforts and have started to bring together a
science working group (SWG) that will be charged with answering fundamental questions, like:
What wavelengths and spatial scales will be key for PFI?
What different aspects of planet formation can we learn from scattered light,
thermal IR, or mm-wave imaging? What are the optimal diagnostic lines to determine
the physical conditions and kinematics in the circumplanetary accretion disk?
Which nearby star-forming regions are the most important to survey?
Can PFI also detect the panoply of warm exoplanets expected to be around most young stars?
The SWG will include observers and theoreticians to address these and other open questions.
We will conduct detailed simulations and interact with the
technical working group (TWG) in order to develop a roadmap for implementing PFI
at the budget of a mid-scale international facility project.
In the following chapters we review some of the recent advancements in high-angular resolution
observations on planet-forming disks (Sect.~\ref{sec:context}) and
introduce the PFI project (Sect.~\ref{sec:pfi}) as well as our working plan
(Sect.~\ref{sec:swg}). We conclude in Sect.~\ref{sec:conclusions}.
Further details about our organizational structure and technological considerations
are given in accompanying articles by Monnier et al., Ireland et al., and Buscher et al.\ in this volume.
\section{STATE-OF-THE-ART IN HIGH ANGULAR RESOLUTION
STUDIES ON PLANET FORMATION}
\label{sec:context}
Planet formation studies are experiencing a tremendous
level of activity, driven both by theoretical advancements
and the fundamentally new observational capabilities that are
provided by the new high-angular resolution imaging capabilities of ALMA
and adaptive optics systems like Gemini/PFI and VLT/SPHERE.
An important discovery of the last decade has been the class of
transitional and pre-transitional disks (Calvet et al.\cite{cal02}, Espaillat et al.\cite{esp07}).
These objects exhibit a strong mid-infrared (MIR) excess, but have a
significantly reduced near-infrared (NIR) excess compared to T\,Tauri or
Herbig\,Ae/Be disks (Espaillat et al.\cite{esp08}). This reduced NIR excess emission
indicates that the innermost disk regions contain only optically thin gas and dust
(transitional disks) or exhibit an extended gap, which separates the optically thick
inner disk from the outer disk (pre-transitional disks).
Coronagraphic observations using the Subaru/HiCIAO and VLT/NACO instruments revealed
intriguing spiral arm-like structures in the outer regions of several transitional and
pre-transitional disks (e.g.\ Muto et al.\cite{mut12}, Quanz et al.\cite{qua13b})
that might be triggered by embedded planets.
Structures on smaller spatial scales were probed using interferometry at multiple wavelengths.
At sub-millimeter wavelengths the SMA, CARMA, ALMA, and the
Plateau de Bure Interferometer detected central density depressions on scales of tens of au,
consistent with an extended inner disk hole or a gapped disk structure
(e.g.\ Isella et al.\cite{ise10}, Andrews et al.\cite{and11}, Casassus et al.\cite{cas13}).
Near- and mid-Infrared interferometry with the VLTI/AMBER, MIDI, and PIONIER instruments
resolved the distribution of material on (sub-)au scales and
characterized the conditions inside the gap.
Observations on HD\,100546, T\,Cha, and V1247\,Ori showed that the
near-infrared (1-2\,$\mu$m) emission is dominated by the emission of
a narrow inner disk component located near the dust sublimation radius and
smaller contributions from scattered light (Oloffson et al.\cite{olo13}) and
optically thin dust emission from within the gap (Kraus et al.\cite{kra13}).
The mid-infrared regime (N-band) is sensitive to
a wider range of dust temperatures and stellocentric radii and
includes emission contributions from the inner disk, the disk gap, and the outer disk.
The gaps of HD\,100546 and T\,Cha were found to be
highly depleted of (sub-)$\mu$m-sized dust grains,
with no significant near-/mid-infrared emission
(Benisty et al.\cite{ben10}, Oloffson et al.\cite{olo13}),
while the disk gap of V1247\,Ori is filled with optically thin
dust material (Kraus et al.\cite{kra13}).
TW\,Hya appears to be in a particularly late stage
of disk clearing, with a very extended dust-depleted
inner hole that extends out from $\sim 0.3$\,au
(Menu et al.\cite{men14}) and contains large, settled dust grains.
A particularly intriguing finding obtained with interferometry
on transitional disks is the detection of non-zero phase signals,
which indicates the presence of strong asymmetries in the
inner, au-scale disk regions.
Some detections are consistent with low-mass companions,
such as the possibly planetary-mass body around the
transitional disk of LkCa\,15 (Kraus \& Ireland\cite{kra12}).
In other cases, the asymmetries can be attributed to emission contributions
from a vertically extended disk seen under intermediate inclination angle
(e.g.\ T\,Cha and FL\,Cha; Olofsson et al.\cite{olo13} and Cieza et al.\cite{cie13}).
Another interesting case is V1247\,Ori, where
the asymmetries seem to trace complex, radially extended
disk structures (Kraus et al.\cite{kra13}) that might be caused by
the dynamical interaction of the (yet undiscovererd) gap-opening body/bodies
with the disk material.
\section{THE PFI PROJECT}
\label{sec:pfi}
The latest observational and theoretical studies suggest that planet formation
is a highly complex and multi-facetted process, where planet-induced
dynamical processes are accompanied by complex changes in the dust mineralogy.
High-angular resolution instruments both at infrared and sub-milllimeter
wavelengths are just starting to obtain a first glimpse of this complexity,
but are not able to properly resolve and characterize these intruiging structures.
Further advancements in our understanding of the planet formation process
can be expected in the coming years, for instance from the fully-operational ALMA array,
the VLTI 4-telescope mid-infrared interferometric instrument MATISSE
(Lopez et al.\cite{lop06}), and the upcoming generation of extremely large telescopes (ELTs).
However, in the relevant infrared/sub-mm wavelength regime, these facilities will
still be limited to angular scales
of {$\sim 0.01$\hbox{$^{\prime\prime}$}} or 1.5\,au for the most nearby star forming regions,
which is $\sim 30\times$ larger than the
Hill Sphere of a Jupiter-like object and $\sim 100\times$
larger than the circumplanetary accretion disk.
A facility with even higher resolution will be essential to
probe the processes on the natural scale, where planet formation is happening.
As a consequence of this realisation, the PFI project was born in
late 2013 and an international consortium has rapidly emerged
to develop realistic project goals, with participants drawn mainly from
the high resolution astronomical imaging community.
Scientists from more than a dozen different institutes in six countries
are presently on the project launch committee.
The project executives have been elected in February 2014,
including the Project Director John Monnier (University of Michigan, USA),
Project Scientist Stefan Kraus (University of Exeter, UK), and
Project Architect David Buscher (University of Cambridge, UK).
During the next 1-2 years, we will develop and prioritize the science goals
and consider all technologies and facility architectures that might be
capable of achieving our science objectives, including visible,
thermal infrared, and mm-wave imaging from the ground and from space.
\section{THE PFI SCIENCE WORKING GROUP}
\label{sec:swg}
The SWG will be lead by the Project Scientist and will be
reponsible for developing and maintaining the top-level science requirements of PFI.
It will be charged with investigating the signatures of planet formation
at different stages of disk evolution, assessing the observability of the
protoplanetary bodies, and to determine how PFI could significantly advance
our understanding of the architecture and the potential habitability of planetary systems.
Our results will be published in white papers and in
refereed journals when appropriate.
Scientists from the star formation, planet formation, planetary science,
and exoplanet community are kindly invited to contribute to our efforts.
In the following sections, we list the main topics that we
plan to investigate within our SWG sub-groups.
\subsection{Protoplanetary Disk Structure \& Disk Physics}
Protoplanetary disks set the initial conditions of planet formation and
determine the later dynamical evolution of the forming planetary system.
Fundamental aspects about disk structure and disk evolution are still poorly understood,
reflecting the fact that most of our knowledge about protoplanetary disk structure
has been derived by fitting spatially unresolved constraints,
such as line profiles or the spectral energy distribution.
Once it is in full operation, ALMA will be able to provide further insights
on the density structure of protoplanetary disks. However, with an angular resolution
of $\ge 5$\,milliarcsecond, ALMA will not be able to probe the disk regions in
the inner-most au, where substantially different processes are believed to be at work
than in the more extended disk. For instance, the disk is believed to be truncated
at the co-rotation radius at a few stellar radii and it has been proposed that this
truncation might be responsible for the pile-up of the Hot Jupiter population (Lin et al.\cite{lin96}).
Designing PFI to operate at mid-infrared wavelengths would
allow us to make use of complementary aspects with ALMA.
The sub-millimeter regime probed by ALMA is most sensitive to the thermal
emission of cold, mm-sized dust grains located in the outer disk and in the disk midplane.
The mid-infrared emission, on the other hand, traces small $\mu$m-sized dust grains
and the warm dust located in the disk surface layer.
Hydrodynamic simulations predict that the dust distribution can differ significantly
for different dust populations. For instance, theories of dust filtration predict that
large (mm-sized) dust grains are held back in the outer disk and accumulate in a narrow ring
outside of the gap, while small ($\mu$m-sized) grains filter through the gap (Zhu et al.\cite{zhu12}).
Spatial variations in the distribution of dust grain populations were also found by the recent
study on the transitional disk Oph\,IRS48 (van der Marel et al.\cite{van13}), where
the small dust grains are distributed rather homogeneously throughout the disks,
while the large dust grains are confined towards one part of the disk.
These differences in the distribution of large and small grains were modelled with
a dust trap that might be triggered by an undetected planetary-mass companion.
Being one of the most nearby (120 pc) transitional disks with a very extended
inner hole ($\sim 40$\,au), it was possible to resolve the distribution of
the $\mu$m-sized grains for Oph IRS48 with conventional VLT/VISIR imaging
observations. However, it is clear that higher resolution will be needed to study these
spatial dust variations in details, in particular for lower-luminosity objects
and in earlier evolutionary stages.
Interferometric observations in the mid-infrared wavelength regime (e.g.\ N-band)
also give access to various strong spectral features such as the 10\,$\mu$m Silicate feature
and several hydrocarbon-related (PAH) features.
Spatially and spectrally resolved investigations with the VLTI/MIDI instrument
revealed radial gradients in the dust minerology and found that the crystallinity
is higher in the inner few au of the disk, supporting theories of grain growth
(van Boekel et al.\cite{van04}).
The hydrocarbon-related emission is located in the gap region
(V1247\,Ori: Kraus et al.\cite{kra14}) and in the outer disk (Olofsson et al.\cite{olo13}).
\subsection{Planet Formation Signatures in Pre-Main-Sequence Disks}
\label{sec:pfsignatures}
\begin{figure}[p]
\centering
\includegraphics[height=21cm]{simulations-eps-converted-to.pdf}
\caption{
Radiative transfer simulation of planet-forming disks with
a single planet (left; simulation from Harries) and 4 planetary bodies (right; simulation from Dong, Whitney \& Zhu), respectively.
For details about the simulation set up, we refer the reader to
Sect.~\ref{sec:pfsignatures}.
}
\label{fig:sim}
\end{figure}
Once planets have formed in the disks, they dynamically sculpt their environment,
for instance by opening tidally-cleared gaps or triggering spiral arms and disk warps.
The aim of this sub-group is to make quantitative predictions for the
structures that we can expect to detect with PFI, both in the dust continuum emission
and in spectral lines. Below, we show radiative transfer simulations
of structures that we might expect to detect with PFI. All simulations were computed
for an inclination angle of $30^{\circ}$ (i.e.\ closer to a face-on viewing angle)
and a central star with an effective temperature of 4000\,K, a mass of 0.5\,M$_{\hbox{$\odot$}}$,
a stellar radius of 2\,R$_{\hbox{$\odot$}}$, and solar abundance.\newline
\noindent{\it 1-Planet Radiative Transfer Simulation:}
Our first simulation represents a relatively simple case of a protoplanetary disk
with a single Jupiter-mass planet located at
a separation of 5\,au from the central star.
This scenario could be observed during the T\,Tauri phase at an age around $0.5$\,Myr.
The radiative transfer computation was performed using the TORUS radiative transfer
code by Tim Harries\cite{har00}.
The planet cleared a gap with a width of 1\,au and triggered
a spiral wake that was parameterised using the analytical description
by Ogilvie and Lubow\cite{ogi02}, closely matching the results of
hydrodynamic simulations.
The density perturbation is $10^4\,(M_{p}/M_{\star})$ (where
$M_{p}$ is the mass of the planet and $M_{\star}$ the stellar mass),
or a factor 10 in density over the unperturbed disk density.
We computed the images for five wavelengths that might be of interest
for PFI, including the near-infrared (K-band around 2\,$\mu$m),
mid-infrared (N- and Q-band around 10 and 24\,$\mu$m),
far-infrared (100\,$\mu$m), and sub-mm regime (400\,$\mu$m).
Containing only a single planet, the images exhibit the minimum amount
of complexity that we should expect to detect with PFI.\newline
\noindent{\it 4-Planet Radiative Transfer Simulation:}
Our second set of simulations represents a later stage
of disk evolution, resembling the conditions that are expected during the
pre-transitional disk phase, where an extended gap region has
been cleared by multiple planets. The simulated images cover a
field-of-view of 80\,au (side-to-side) at a pixel size of 0.1\,au.\newline
The simulation has been conducted using the HOCHUNK3D radiative transfer
code by Barbara Whitney et al.\cite{whi13}.
The density profile for these simulations was computed using a
2-D hydrodynamic simulation by Zhaohuan Zhu\cite{zhu11},
which incorporates four Jupiter-mass planets located at
separations of 5, 7.5, 12.5, and 20\,au.
The vertical disk scale height was computed based on hydrostatic equilibrium,
where we assume two dust populations:
A small dust population (with grain sizes up to $\sim 1\,\mu$m) that contributes
10\% of the total dust mass and is well mixed with the gas component.
A bigger dust grain population (with sizes extending to 1\,mm) is collapsed to
20\% of the gas scale height, mimicking the effect of dust grain settling (see Dong et al.\cite{don12}).
A qualitative inspection of these continuum images suggests that the mid-infrared regime
(e.g.\ L-/M-/N-band between 3\,$\mu$m and 13\,$\mu$m)
might constitute a good wavelength regime to image the planet-related disk signatures
in the inner few au.
This wavelength regime is sensitive to the thermal emission of dust grains in the $\sim 1000-300$\,K
range, matching the temperature of hot dust in the disk surface layer in the terrestrial planet-forming region
at a few au, as well as the expected temperature of the circumplanetary accretion disk (see below).
Shorter wavelengths (e.g.\ K-band around 2\,$\mu$m) contain significant contributions
from scattered light, while longer wavelengths (e.g.\ FIR at sub-mm) are dominated
by the emission of cold material in the outer disk.
Further work by the SWG and TWG will need to confirm this assessement
and identify the wavelength band(s) that provide the optimal balance between
the brightness of the expected features and the technically achievable image contrast/imaging fidelity.
Beside the continuum tracers, we will also simulate the expected signatures
in different line tracers. Some interesting line tracers in the mid-infrared part of the spectrum
could include the CO fundamental lines (4.7\,$\mu$m), CH$_3$OH ices (3.5\,$\mu$m),
NH$_3$ ice (3.0\,$\mu$m), and C-H nanodiamonds (3.4-3.5\,$\mu$m).
Recording spectrally dispersed data in the line tracers will allow us to
construct maps in different velocity channels, constraining the kinematics of the gas
in the accretion streams and the circumplanetary accretion disk.
By imaging the disks at multiple epochs, we will also be able to study the
temporal evolution of the planet-induced disk structures and to trace the
orbital motion of the embedded protoplanets.
At its' unprecedented, sub-milliarcsecond resolution, PFI will likely be
able to detect such structural changes on timescales of just a few months,
directly revealing the ongoing, highly dynamical planet formation processes.
\subsection{Protoplanet Detection}
\begin{figure}[t]
\centering
\includegraphics[height=10cm,angle=270]{contrast-eps-converted-to.pdf}
\vspace{3mm}
\caption{
Predicted planet/star contrast for a $4M_{\rm Jup}$ protoplanet,
plotted as function of age.
The computation is based on the atmosphere models by Baraffe et al.\cite{bar98,bar02}
and on the ``Hot Start'' evolutionary tracks by Fortney et al.\cite{for08}
and has been conducted for five spectral bands.
}
\label{fig:planetcontrast}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[height=10cm]{sketch-eps-converted-to.pdf}
\caption{
Mosaic of hydrodynamics simulations of a protoplanetary disk with four embedded planets
(background), the gap that is opened by one of the planets (top-left panel),
and of the circumplanetary accretion disk (bottom-left panel), illustrating the need
to probe a wide range of spatial scales in order to study the mechanisms that
are at work during planet formation.
The background image is the 10\,$\mu$m model image of the four-planet simulation
by Dong, Whitney \& Zhu described in Sect.~\ref{sec:pfsignatures}, while
the insets show surface density profiles from simulations by Ayliffe \& Bate\cite{ayl09}.
Besides the physical scales (in au) we give the angular scale on the sky
for a distance of 140\,pc (in mas), which corresponds to the distance of the most
nearby star forming regions (e.g.\ Taurus).
}
\label{fig:sketch}
\end{figure}
The simulations discussed above show the accretion streams towards the
forming planets, but do not include the emission from the protoplanet itself.
For young planets, the brightness of the viscously heated circumplanetary disk
can significantly exceed the thermal emission of the surrounding dust.
In the age range between $10^{6}$ and $10^{7}$ years, the brightness ratio between the
central star and the protoplanet is predicted to reach $10^{3...4}$
in the mid-infrared wavelength regime (e.g.\ L-band, Figure~\ref{fig:planetcontrast}),
which is much more favourable than during the main-sequence phase, when the
corresponding contrast is typically $10^{7...8}$.
Some individual protoplanet candidates have already also been detected,
including T\,Cha (Huelamo et al.\cite{hue11}) and LkCa\,15 (Kraus \& Ireland\cite{kra12}).
These two detections have been achieved with the aperture masking interferometry technique,
which offers an efficient way to reach a good phase accuracy ($\sim 0.1^{\circ}$)
near the diffraction-limit of large single-dish telescope.
However, the detections have been made near the resolution limit and face
challenges in separating the protoplanet detection from the emission of
asymmetric disk features.
For instance, it has been argued that the phase signature on T\,Cha might
trace disk emission from the inner edge of the outer disk (at around 12\,au),
instead of a planetary-mass companion (Olofsson et al.\cite{olo13}).
Observations on the pre-transitional disks V1247\,Ori also revealed
the presence of complex disk asymmetries in the inner regions of the disk
around this object, whose phase signatures can easily be mistaken as
close companions (Kraus et al.\cite{kra13}).
The $\sim 100$-times higher angular resolution provided by PFI will be needed
to separate the protoplanets from possible disk contributions and to resolve the
circumplanetary accretion disk on scales of 0.2 Hill radii (0.1\,mas for a Jupiter-mass
planet at 1\,au; e.g.\ Ayliffe \& Bate\cite{ayl09}), enabling us to probe the
planet formation process over a wide range of spatial scales (Figure~\ref{fig:sketch}).
A major task of this working group will be to identify the optimal line tracers
that are suitable for tracing the circumplanetary accretion disk.
Intruigingly, recent adaptive optics imaging observations
at visual wavelengths were able to detect the accretion signatures of a low-mass
companion in the pre-transitional disk of HD\,142527 using the H$\alpha$-line tracer (Close et al.\cite{clo14}).
This demonstrates the feasibility for detecting the accretion signatures of low-mass
objects in pre-main-sequence (PMS) disks. For PFI we aim to identify further line tracers that
are suitable for tracing embedded planets.
\subsection{Exoplanetary System Architecture}
Planet population synthesis models aim to reproduce the observed exoplanet
system populations by linking planet formation models
with models about the dynamical processes of planet-disk interaction,
planet-planet interaction, dust evolution, and accretion physics.
These models depend on a proper knowledge of the relevant processes
and on the adjustment of a large number of free parameters, which introduces
major uncertainties.
PFI will change the situation fundamentally, both by providing robust
information about the initial conditions of planet formation and
by probing the protoplanet distribution during the evolutionary phase
that is most critical for shaping the architecture of exoplanetary systems,
namely the first $\sim 100$\,Myr.
Observing planetary systems during this time interval will allow PFI to
determine where in the disk planets form and how they migrate through
interaction with the gas-rich disk, providing insights into the mechanisms
that halt migration, for instance at deadzones, disk truncation points, or
during the disk dissipation phase (typically at $\sim 10$\,Myr).
Achieving this goal will require detecting planets in a statistically meaningful
sample of systems at different evolutionary phases, e.g.\ of the order of
100 systems in the classical T\,Tauri/Herbig~Ae/Be phase ($\sim 0.5$\,Myr),
transitional disk ($\sim 5$\,Myr),
and early debris disk phase ($\sim 50$\,Myr).
These observations will allow us to construct planet population diagrams
for these different evolutionary phases and to compare them with the
population diagrams for mature exoplanet systems.
Based on state-of-the-art population synthesis models we expect dramatic
changes during these phases, such as the inward-migration of the
Hot Jupiters during the first 1-2\,Myr and the ejection of planets
due to dynamical instabilities (e.g.\ Raymond et al.\cite{ray06}).
Another important objective of PFI will be to determine the location of the
``snow line'' for important molecules like water (H$_2$O),
marking the location where these molecules condense to form ice grains.
At the snow line, the density of solid particles in the disk increases abruptly.
This allows planets to form more efficiently beyond the snow line and enables
them to accrete gas from the disk before it dissipates, favouring the
formation of gas giant planets.
The location of the snow line is difficult to predict for different molecules
and changes during the disk evolution. Therefore, it will be crucial to
measure the location of the snow line directly, for instance using the
mid-infrared H$_2$O 3.1\,$\mu$m line.
Measuring the distribution of the snow line will allow us
to understand how water is delivered to terrestrial planets like Earth,
where it has been proposed that water can either been produced
through oxidation of atmospheric hydrogen (Ikoma et al.\cite{iko06}) or is
delivered through planetesimals from beyond the snow line (Morbidelli et al.\cite{mor00}).
Given their favourable brightness contrast, PFI will likely focus on giant planets.
However, the giant planet population is the most critical component in shaping
the architecture of planetary systems and has also a major impact on the
formation of terrestrial planets (e.g.\ Morbidelli et al.\cite{mor14}).
\subsection{Late Stages of Planetary System Formation}
The gas- and dust-rich protoplanetary disk dissipates on timescales of 5-10\,Myr
(Hernandez et al.\cite{her07}), which is also expected to halt type I+II
migration of the embedded protoplanets.
Smaller amounts of optically thin dust are still observed at the later
debris disk phase. The dust in these disks is believed to be replenished by the
collision of planetesimals.
Some debris disks show intruiging ring-like structures that have been
attributed to the dynamical influence of embedded planets (Kalas et al.\cite{kal05}),
although it has been suggested that these structures might also be shaped by
hydrodynamical instabilities without planets (Lyra \& Kuchner\cite{lyr13}).
The planetary system architecture is still subject to changes during the
debris disk phase, for instance through planet-planet interaction processes
or through dynamical interaction with the planetesimal disk.
In fact, it is now believed that our solar system also underwent a drastic
reconfiguration at an age of $\sim 700$\,Myr, owing to the migration of
the giant planets due to interaction with the planesimal disk
and a crossing of Jupiter and Saturn's 1:2 resonance (Tsiganis et al.\cite{tsi05}, Walsh et al.\cite{wal11}).
This reconfiguration might also have triggered the Late Heavy Bombardment
period of the terrestrial planets (Gomez et al.\cite{gom05}).
PFI will be able to image the disk structure and to trace the protoplanets
during the early debris disk phase, while the cooling planets are still relatively bright
at mid-infrared wavelengths. Further work will be needed in order to determine
which planet mass and which age range are accessible with PFI.
\subsection{Planet Formation in Multiple Systems}
Another unique application of PFI will be in studying planet formation
in multiple systems. Several circumbinary planets have already been
detected in mature systems (e.g.\ Kepler 16, Doyle et al.\cite{doy11}) and
it has been estimated that more than 1\% of all close binary stars have
giant planets in nearly coplanar orbits (Welsh et al.\cite{wel12}).
PFI will measure the disk truncation effects that are induced by the companion star
and determine where planet formation is possible in this environment.
The disk measurements will inform us how the tidally cleared gap affect
the migration processes.
\subsection{Star Forming Regions / Target Selection}
This working group will identify the star forming regions that are most suitable
to cover the wide range of evolutionary stages and environments
that we anticipate to probe with PFI.
For this purpose, we will consider not only typical low-mass star forming regions
like Taurus and Ophiuchus, but also high-mass star forming regions like Orion.
Comparing the disk structure in these different environments will inform us
how the planet formation process depends on the mass of the central star
and on environmental factors such as the ambient UV ionization field.
The availability of suitable target star populations on the Northern and Southern Hemisphere
will provide an important input for the TWG, both to identify potential sites
and to formulate the sensitivity requirements for the facility.
\subsection{Secondary Science Cases}
Besides investigating the primary science case, PFI has the potential to revolutionize
a wide range of other science areas from galactic and extragalactic astronomy.
The unprecedented resolution of $\sim 0.1$\,mas will enable a detailed characterization of
the fundamental parameters of exoplanet host stars (sizes, limb profile variations)
and to correct for stellar activity features like spots, which limits current exoplanet radial velocity
and transit studies.
It will enable direct spectroscopy of Hot Jupiters and the determination of their astrometric orbits.
In galactic astrophysics, some obvious applications could include imaging of photospheric structures,
such as spot patterns on main-sequence stars, the weather patterns on white dwarf stars,
shock waves on post-AGB stars, and the mass-loss processes around evolved stars.
PFI will be able to study the accretion of matter onto the black hole in X-ray binaries and
to study the circumstellar envelopes and pulsation modes in Cepheids,
which is essential for a proper calibration of the cosmological distance ladder.
Some unique science goals could also be achieved in extragalactic science,
for instance by performing detailed imaging of AGN dust tori, by performing spatially
resolved AGN reverberation mapping, by determining dynamical black hole masses
(through measurement of the rotation profile of the circumnuclear accretion disk),
and by spatially resolving the multiple images of gravitational micro-lensing events.
In the initial project phase, these secondary science cases will not drive our design decisions of PFI,
but they will be taken into account for later refinements of the technical specifications
and for adding further capabilities that do not compromise the primary science mission.
\section{CONCLUSIONS}
\label{sec:conclusions}
The PFI project aims to identify compelling science goals and a
realistic technical roadmap for taking the next step in exploring the
universe at high-angular resolution.
We argue that improved high-angular resolution imaging capabilities will
be inevitable to advance our understanding of the planet formation process.
In this contribution, we outlined our initial thoughts on the key science drivers of PFI,
where our objective is to stimulate the discussion in our science working group
and in the wider community.
Our top priorities for the next 12-24 months are to define the top-level science requirements
and to present them in a series of white papers.
This work will provide essential input for the TWG, which will formulate the instrument concept
and determine feasible facility architectures for meeting the science goals.
Further information is available on our project website
\url{http://www.planetformationimager.org}.
\acknowledgments
We thank the large number of colleagues that volunteered to contribute to our
efforts and that bring our science working group to life.
|
\section{Introduction}
There are two interesting maximally supersymmetric quantum field theories in six dimensions: the $(2,0)$ superconformal field theory and the $(1,1)$ super-Yang-Mills theory. They are believed to be the low energy limits of $(2,0)$ or $(1,1)$ little string theories (LST) \cite{Berkooz:1997cq,Seiberg:1997zk,Aharony:1998ub,Giveon:1999zm,Aharony:1999ks,Kutasov:lecturenote}. While the LST is a strongly coupled string theory that is difficult to get a handle on, it admits a 1-parameter deformation known as the double scaled little string theory (DSLST) \cite{Giveon:1999px,Giveon:1999tq}, the parameter being an effective string coupling. In the weak coupling limit, DSLST can be studied perturbatively. In particular, one can compute the spectrum and in principle perturbative scattering amplitudes to all orders in $\alpha'$.
In this paper, we will focus on the $(1,1)$ theories. In this case, the LST reduces to 6D SYM at low energies. More precisely, the interactions of massless modes of LST are expected to be described by an effective theory that is 6D SYM deformed by higher dimensional operators (such operators are highly constrained by supersymmetry and we will return to this point later). $\alpha'$ will be mapped to the 6D Yang-Mills coupling via the relation
\ie\label{idaa}
{1\over 2\pi\alpha'} = {8\pi^2\over g_{YM}^2}.
\fe
This can be understood by identifying the little string with instanton strings in the SYM. The DSLST, on the other hand, is related to 6D SYM on its Coulomb branch. The relation between the string coupling and the Coloumb branch parameter takes the form
\ie\label{idbb}
g_s \sim {1\over g_{YM} m_W}.
\fe
Thus one may anticipate a relation between scattering amplitudes of massless modes in the DSLST and those of the Cartan gluons in the 6D SYM of the schematic form
\ie
{\cal A}_{DSLST}(g_s, \A', E) = {\cal A}_{SYM}(g_{YM}, m_W, E),
\fe
provided the identifications (\ref{idaa}) and (\ref{idbb}).
Note however that the weak coupling limit of DSLST corresponds to the regime far from the origin on the Coloumb branch. In the 6D gauge theory, a priori one would expect higher dimensional operators to become important in this limit, and it is not a priori clear whether the gluon scattering amplitudes in the SYM, expanded in $1/m_W^2$, should agree with that of the perturbative amplitudes of massless modes of DSLST.
In fact, the 6D SYM diverges at three-loops, and one might be led to think that such an agreement is impossible with just the SYM perturbative amplitudes (without including any higher dimensional operators). However, a more careful inspection indicates that the scattering amplitude of gluons in the Cartan $U(1)^{k-1}$ on the Coulomb branch of $SU(k)$ gauge theory is free of UV divergence at 3-loops. This is because there are only two candidate counter terms that are dimensional 10 operators allowed by sixteen supersymmetries; one of them is a double trace operator known to be 1/4 BPS and finite in 6D SYM according to \cite{Bossard:2010pk}, the other is a single trace operator that is non-BPS but vanishes when restricted to the Cartan subalgebra of the $SU(k)$. (We will explicitly verify the absence of this UV divergence in Appendix \ref{3ldiv}.) This suggests that a certain non-renormalization theorem is at play, and one could hope that the 6D SYM by itself already captures some of the dynamics of massless gluons on the Coulomb branch of the $(1,1)$ little string theory.
Remarkably, {\bf we find agreement} between up to two-loop amplitudes of the massless gluons on the Coulomb branch of the 6D SYM, expanded to leading order in $1/m_W^2$, and the tree level amplitude of the corresponding massless string modes in DSLST, expanded up to second order in $\A'$. This is achieved through explicit computation of amplitudes in the $SU(k)$ SYM, based on previously known two-loop results derived from unitarity cut methods \cite{Dixon:1996wi,Dixon:2010gz,Bern:1996je,Brandhuber:2010mm,Bern:1997nh,Bern:2010qa}, and in the DSLST by a worldsheet computation that involves expressing $SL(2)/U(1)$ coset CFT correlators in terms of Liouville correlators, which are then expressed as integrals of Virasoro conformal blocks. The computations on the two sides utilize completely different techniques and the results are rather involved. Nonetheless, we will evaluate the results numerically for $k$ up to 5, and the answers on the two sides strikingly agree.
While our result may be viewed as a strong check of the duality between DSLST and 6D SYM, it also hints that the higher dimensional operators correcting the 6D SYM are under control. Put differently, the perturbative 6D SYM seems to know a lot more about its UV completion and particularly the instanton strings than one might have naively expected!
The paper is organized as the following. After reviewing the construction of DSLST and its spectrum, we will study the tree level four-point amplitude of massless RR vertex operators in six dimensions. In computing such amplitudes, we will need the singular limits of certain correlators in the $SL(2)/U(1)$ coset (``cigar") CFT, which are further related to Liouville correlators by Ribault and Teschner's relations \cite{Ribault:2005wp}. This will allow us to express the DSLST amplitudes as integrals of Liouville/Virasoro conformal blocks.\footnote{The integration of conformal blocks was circumvented in \cite{Aharony:2003vk} as the authors of \cite{Aharony:2003vk} were only concerned with the leading order answer in the $\A'\to 0$ limit. We find it necessary to compute the full conformal blocks in order to extract the next order terms in the $\A'$ expansion, and numerical integration of conformal blocks will be carried out.} Order by order in the $\alpha'$ expansion, these integrals can be computed numerically, thereby producing an explicit $\alpha'$ expansion of the massless amplitudes in DSLST. The DSLST results are listed in Table \ref{tab:lstresults} and the 6D SYM results are summarized in Table \ref{tab:2loop}.
These amplitudes are then compared with gluon scattering amplitudes of 6D SYM on its Coulomb branch, expanded in $1/m_W^2$ to the leading order (if these amplitudes were the full story, higher order terms in the $1/m_W^2$ expansion would be mapped to higher genus double scaled little string amplitudes).
The 1-loop and 2-loop amplitudes are perfectly finite and can be obtained using unitarity cut method \cite{Bern:1996je,Brandhuber:2010mm,Bern:1997nh}, and the 3-loop amplitudes have been reduced to scalar integrals as well \cite{Bern:2007hh} (as already asserted, they are free of UV divergences when the external lines are restricted to the Cartan subalgebra).
An agreement of the 1-loop amplitude at order $1/m_W^2$ with the low energy limit of four gluon amplitude in DSLST was found by \cite{Aharony:2003vk}.
We will carry out the $\alpha'$ expansion of DSLST amplitude to the next order in $\alpha'$ and compare it with the 2-loop $SU(k)$ SYM amplitudes (which involve highly nontrivial group theory factors). While the expressions on both sides are quite complicated, we will evaluate them numerically for $k=2,3,4,5$, and remarkably, they agree. (In the $k=5$ case, for instance, the four-point functions of nontrivial primaries in the three-state Potts model went into the DSLST amplitude computation!)
We conlude with a discussion on the implication of these results. Details on the numerical integration of conformal blocks, based on Zamolodchikov's recursion relations, are described in Appendix \ref{numerics}.
\section{Double scaled little string theory}
In Section~2.1, we review holographic descriptions of DSLST. In Sections~2.2-2.6, we construct the normalizable vertex operators of the supersymmetric $( SL(2)_k/U(1) \times SU(2)_k/U(1) ) / \bZ_k$ theory. We will find that these vertex operators $V^{sl (\eta, \bar\eta)}_{j,m,\bar m} V^{su (\eta',\bar\eta')}_{j',m',\bar m'}$ have quantum numbers satisfying (2.56) subject to the identification (2.62) and (2.64).
Our main result is in Section~2.7, where we identity the massless bosonic vertex operators of type IIA string theory on $\bR^{1,5} \times ( SL(2)_k/U(1) \times SU(2)_k/U(1) ) / \bZ_k$, which is the T-dual description of IIB (1,1) DSLST. These massless string states correspond to the scalars and gluons in the 6D SYM.
\subsection{The holographic description}
The physical definition of double scaled little string theory is the decoupled theory on $k$ NS5-branes in type IIA or IIB string theory, spread out on a circle of radius $r_0$ in a transverse plane $\mathbb{R}^2\subset \mathbb{R}^4$, in the double scaling limit $r_0, g_s\to 0$, with $r_0/g_s$ held finite in string units. Here one may consider an energy scale of interactions comparable to the string scale (without taking a low energy limit). The configuration or the decoupled theory has a residual global symmetry $\mathbb{Z}_k\times U(1)$.
In the case of IIB NS5-branes, D1-branes extending between NS5-branes correspond to W-bosons in the six-dimensional $SU(k)$ gauge theory. They become tensionless in the $r_0\rightarrow 0$ limit, where we recover the strongly coupled (1,1) LST.
On the other hand, one may directly take the decoupling limit of the string worldsheet theory in the NS5-brane background. Recall that (1,1) LST is described by IIB string theory on the linear dilaton background from the NS5-brane near horizon geometry {\cite{Callan:1991at,Aharony:1998ub}. Its deformed cousin (1,1) DSLST is described by $\cN=2$ Liouville theory in IIB which is T-dual to the ${SL(2)_k \over U(1)}$ cigar CFT in IIA, similarly for (2,0) DSLST \cite{Giveon:1999px,Ooguri:1995wj, Kutasov:1995te} (see also \cite{Sfetsos:1998xd}).
This leads to the ``holographic" description of DSLST, as IIA (in the $(1,1)$ case) or type IIB (in the $(2,0)$ case) string theory with the target space given by
\ie
{\mathbb R}^{5,1} \times \left( {SL(2)_k\over U(1)}\times {SU(2)_k\over U(1)}\right)/\bZ_k
\fe
in the NSR formalism. The $\bZ_k$ orbifolding acts simultaneously on the two (supersymmetric) coset models.
The supersymmetric coset model $SL(2)_k/U(1)$ can be constructed from the bosonic $SL(2)_{k+2}$ plus three free fermions $\lambda^a$, $a=1,2,3$, and gauging the $U(1)$ supercurrent that contains $\lambda^3$ and the total bosonic $U(1)$ current at level $k$ (combining a $U(1)$ from the bosonic $SL(2)_{k+2}$ and a $U(1)$ current from the fermions at level $-2$). The supersymmetric $SU(2)_k/U(1)$, constructed in a similar manner from the bosonic $SU(2)_{k-2}$ and three fermions, is the same as the $k$-th ${\cal N}=2$ minimal model. The total worldsheet matter central charge is
\ie
c^m = 9+{3(k+2)\over k}+{3(k-2)\over k} = 15,
\fe
as required for critical string theory.
\subsection{$SL(2)_k/U(1)$}
The ${\cal N}=2$ supersymmetric $SL(2)_k/U(1)$ coset model can be constructed as follows \cite{Giveon:2003wn}. One starts with the bosonic $SL(2)_{k+2}$ WZW model, governed by the current algebra
\ie
j^a(z) j^b(0) \sim {(k+2)\eta^{ab} \over 2z^2} + i\epsilon^{abc} {j_c(0)\over z},
\fe
where $\eta_{ab}={\rm diag}(1,1,-1)$. We will lower and raise the index $a$ by $\eta_{ab}$ and its inverse $\eta^{ab}$. The anti-holomorphic currents will be denoted as $\bar j^a$. The extension to supersymmetric $SL(2)_k$ WZW model simply requires adding three fermions $\lambda^a$, $a=1,2,3$, that obey a slightly non-standard reality condition. Namely, the OPEs of the $\lambda$'s are given by
\ie
\lambda^a(z) \lambda^b(0) \sim {\eta^{ab}\over z}.
\fe
The current $- {i\over 2} \epsilon_{abc}\lambda^b\lambda^c$ gives rise to an $SL(2)$ current algebra at level $-2$. Altogether, we have a level $k$ $SL(2)$ current
\ie
J_a = j_a - {i\over 2}\epsilon_{abc} \lambda^b \lambda^c\label{totalcurrent}
\fe
for the supersymmetric $SL(2)_k$ WZW model.
This theory has ${\cal N}=1$ superconformal symmetry, with the supercurrent
\ie
G = \sqrt{2\over k} (\eta_{ab} \lambda^a j^b - {i\over 6}\epsilon_{abc} \lambda^a\lambda^b\lambda^c).
\fe
The three fermions $\lambda^a$ are superconformal primaries with respect to $G$, while the $J^a$'s are superconformal descendants of $\lambda^a$.
The supersymmetric $SL(2)_k/U(1)$ coset model is defined by gauging the $U(1)$ ${\cal N}=1$ supermultiplet that contains the primary $\lambda^3$ and $J_3$. The coset model $SL(2)_k/U(1)$ has $\mathcal{N}=2$ superconformal symmetry with superconformal currents $G^\pm$ and $R$-current $J_R$ given by
\ie
& G^\pm = \sqrt{2\over k} j^\mp \lambda^\pm,
\\
& J_R = {k+2\over k}\lambda^+\lambda^-+ {2\over k} j^3 = \lambda^+\lambda^- + {2\over k} J^3,\label{Rcurrent}
\fe
where we have defined
\ie
\lambda^\pm \equiv {\lambda^1\pm i\lambda^2\over \sqrt{2}},~~~~
j^\pm \equiv j^1\pm i j^2.
\fe
Let us summarize the relations between the various currents in this construction. We have the bosonic $SL(2)_{k+2}$ currents $j^a$ and the fermonic currents $-{i\over2}\epsilon_{abc}\lambda^b\lambda^c$, with their sum being the total $SL(2)_k$ current $J^a$ for the supersymmetric $SL(2)_k$ WZW model. The coset model $SL(2)_k/U(1)$, as an $\mathcal{N}=2$ superconformal field theory, has the $R$-current $J_R$ defined above. Let $X$, $x$, $X_R$, and $H$ be the bosonzations of $J^3$, $j^3$, $J_R$, and $\LL^+\LL^-$ respectively,
\ie
& J^3 = - \sqrt{k\over 2} \partial X,\\
&j^3 = - \sqrt{k+2\over2} \partial x,\\
&J_R = i \sqrt{ k+2\over k} \partial X_R,\\
&\LL^+ \LL^- = i \partial H.\label{slbos}
\fe
The (chiral) bosons are normalized with the standard OPE $X(z) X(0)\sim - \log z$, etc.
It follows from \eqref{totalcurrent} and \eqref{Rcurrent} that the following relations hold among the bosonization scalars:
\ie
&i H = \sqrt{2\over k} X + i \sqrt{k+2\over k} X_R,\\
& x= \sqrt{k+2\over k}X + i \sqrt{2\over k} X_R.
\fe
Note that $H(z)x(0)\sim 0$ by definition, and it follows that $X(z)X_R(0)\sim 0$. This is necessary for $J_R$ to be well defined in the coset theory.
The NS superconformal primaries of the coset model can be constructed starting from the primaries of the supersymmetric $SL(2)_k$ WZW model, which can be taken to be the primaries of the bosonic $SL(2)_{k+2}$ WZW model, $\Phi^{sl}_{j,m,\bar m}$, with conformal weight
\ie
-{j(j+1)\over k}
\fe
and charge $m$ and $\bar m$ with respect to $j^3$ and $\bar j^3$. The range of $j$ and $m,\bar m$ will be discussed later. The $SL(2)/U(1)$ primary $V_{j,m,\bar m}^{sl}$ is then obtained by factoring out a $U(1)$ primary\footnote{As we are focusing on the massless string modes in this paper, we only need to take into account those coset primaries that come directly from current primaries.},
\ie
\Phi_{j,m,\bar m}^{sl} = V_{j,m,\bar m}^{sl} e^{\sqrt{2\over k} (m X-\bar m \bar X)},
\fe
where $X$ and $\bar X$ are related to the $U(1)$ currents by the bosonization dictionary $J^3=-\sqrt{k\over 2}\partial X$, $\bar J^3=+\sqrt{k\over 2}\bar\partial\bar X$. The plus sign convention for $\bar X$ in $\bar J^3$ is chosen for later convenience. The coset model primary $V_{j,m,\bar m}^{sl}$ has conformal weights
\ie
\Delta_{j,m} = {-j(j+1)+m^2\over k}, ~~~\bar \Delta_{j,\bar m} = {-j(j+1)+\bar m^2\over k}.
\fe
The ${\cal N}=2$ superconformal $R$-charge can be determined as follows. First recall that $J_R = i\partial H + {2\over k}J^3$ and $H$ does not have singular OPE with $\Phi_{j,m,\bar m}^{sl}$. It follows that $R(\Phi_{j,m,\bar m}^{sl}) = 2m/k$. Next, since $X_R(z)X(0)\sim 0$, the operator $e^{ m\sqrt{2\over k} X}$ is uncharged with respect to $J_R$. Combining the above two observations, we obtain the $R$-charges for the coset primary $V_{j,m,\bar m}^{sl}$,
\ie
R(V_{j,m,\bar m}^{sl}) = {2m\over k},~~~\bar R(V_{j,m,\bar m}^{sl}) = {2\bar m\over k}.
\fe
\subsubsection{Non-normalizable, delta function normalizable, and normalizable primaries}
Depending on the value of $j$, the primary operator $V_{j,m,\bar m}^{sl}$ can either be non-normalizable, delta function normalizable, or normalizable along the radial direction of the cigar. The non-normalizable primaries correspond to generic real $j$, whereas the delta function normalizable primaries are given by $j\in -{1\over 2} + i\mathbb{R}$ (the imaginary part of $j$ being the asymptotic momentum along the radial direction of the cigar). Note that the conformal weight formula is invariant under $j\to -j-1$. In fact, the operators $V_{j,m,\bar m}^{sl}$ and $V_{-j-1,m,\bar m}^{sl}$ are related by \cite{Aharony:2004xn}
\ie
& V_{j,m,\bar m}^{sl} = R(j,m,\bar m;k) V_{-j-1,m,\bar m}^{sl},
\\
& R(j,m,\bar m;k) = \nu(k)^{2j+1} {\Gamma(1-{2j+1\over k}) \Gamma(j+m+1) \Gamma(j-\bar m+1)\Gamma(-2j-1)\over \Gamma(1+{2j+1\over k}) \Gamma(m-j) \Gamma(-\bar m-j) \Gamma(2j+1) },
\\
& \nu(k) = {1\over \pi} {\Gamma(1+{1\over k}) \over \Gamma(1-{1\over k}) }.
\fe
In the above formula $m\geq \bar m$ is assumed, i.e. the momentum $n=m-\bar m$ (defined below) is non-negative. If $n$ is negative we simply exchange the role of $m$ and $\bar m$ in the formula for the reflection coefficient.
Using this reflection relation we can restrict the non-normalizable operators to real $j>-{1\over 2}$. In constructing string vertex operators, we must also impose an upper bound $j<{k-1\over 2}$ to ensure the two point function of $V_{j,m,\bar m}^{sl}$ is nonsingular \cite{Giveon:1999px,Giveon:1999tq}.\footnote{This bound is slightly stronger than that imposed by the no-ghost theorem in string theory on $SL(2)$ \cite{Evans:1998wq,Maldacena:2000hw}.}
The $j^3$, $\bar j^3$ charges $m$ and $\bar m$ are related to momentum and winding on the cigar (along the circle direction),
\ie
m - \bar m = n\in \mathbb{Z},~~~~ m+\bar m=kw\in k\mathbb{Z},
\fe
or
\ie
m = {n+wk\over 2},~~~~\bar m = {-n+wk\over 2}.
\fe
So far, for either the non-normalizable or the delta function normalizable primary operators, there are no constraining relations between $m,\bar m$ and $j$.
Importantly, there are also {\it normalizable} primary operators at special values of real $j$ \cite{Aharony:2003vk,Aharony:2004xn}, corresponding to principal discrete series of $SL(2)$ \cite{Maldacena:2000hw}, namely
\ie
j= j_* = m_0-1, m_0-2,\cdots,~~~j_*>-{1\over 2}
\fe
where $m_0$ is given by
\ie\label{m0102}
m_0 =
\begin{cases}
{\rm min}\{ |m|,|\bar m|\}, & m,\bar m<-{1\over 2}. \\
{\rm min}\{ m,\bar m\}, & m,\bar m>{1\over 2}.
\end{cases}
\fe
Only delta function normalizable primaries and normalizable primaries are being used to construct the vertex operators of DSLST. The delta function normalizable primaries of the $SL(2)/U(1)$ will lead to a continuum of string modes that propagate down the cigar, whereas the normalizable primaries will give rise to string modes localized at the tip of the cigar, thus effectively living in six dimensions \cite{Giveon:1999px,Aharony:2004xn}. The scattering amplitudes of the latter is the subject of this paper.
\subsection{$SU(2)_k/U(1)$ and ${\cal N}=2$ minimal model}
The supersymmetric coset model $SU(2)_k/U(1)$ can be constructed similarly to the $SL(2)_k/U(1)$ model. We will denote the primary operators and charges in the $SU(2)_k/U(1)$ model with primes in order to distinguish them from those in the $SL(2)_k/U(1)$ model. Let $j'_i$, $i=1,2,3$, be bosonic $SU(2)_{k-2}$ currents and let $\lambda'_i$ be three free fermions. The OPEs are (in this case there is no distinction between upper and lower indices)
\ie
&j'_i(z)j_j'(0)\sim{(k-2)\delta_{ij}\over 2z^2}+i\epsilon_{ijk}{{j'}^k(0)\over z},\\
&\lambda'_i(z)\lambda'_j(z)\sim{\delta_{ij}\over z}.
\fe
The overall $SU(2)_k$ current of the supersymmetric WZW model is given by
\ie
J_i'=j_i'-{i\over 2}\epsilon_{ijk}{\lambda'}^j{\lambda'}^k.\label{totalcurrentp}
\fe
The supersymmetric $SU(2)_k/U(1)$ coset has $\mathcal{N}=2$ superconformal currents ${G'}^\pm$ and $U(1)$ $R$-current $J_R'$,
\ie
{G'}^\pm = \sqrt{2\over k} {j'}^\mp {\lambda'}^\pm,~~~J_R' = - {2\over k} j_3' + {k-2\over k}{\lambda'}^+ {\lambda'}^-\label{Rcurrentp}.
\fe
Let $X'$, $x'$, $X_R'$, $H'$ be the bosonization of $J_3'$, $j_3'$, $J_R'$, and ${\lambda'}^+{\lambda'}^-$, respectively:
\ie
&J_3' = i \sqrt{k\over 2}\partial X',\\
&j_3' = i \sqrt{ k-2\over 2} \partial x',\\
&J_R' = i \sqrt{k-2\over k}\partial X_R',\\
&{\lambda'}^+ {\lambda'}^- = i \partial H'.
\fe
From \eqref{totalcurrentp} and \eqref{Rcurrentp}, we can read off the relations between the bosonization scalars
\ie
&X' = \sqrt{k-2\over k} x' + \sqrt{2\over k} H',\\
&X_R '= - \sqrt{2\over k} x'+\sqrt{k-2\over k} H'.\label{relation}
\fe
In particular, it follows that $X'(z)X_R'(0)\sim 0$, as is required for $J_R'$ to survive the coset construction.
The NS superconformal primaries of the $SU(2)_k/U(1)$ coset model (which is the same as the $k$-th ${\cal N}=2$ minimal model), denoted by $V^{su}_{j',m',\bar m'}$, can be constructed starting from the bosonic $SU(2)_{k-2}$ primary $\Phi^{su}_{j',m',\bar m'}$ and factoring out the $U(1)$ part. Their conformal weights and $R$-charges are
\ie
\Delta_{ j', m'} = { j'(j'+1) - m'^2\over k},~~~R = - {2m'\over k},
\fe
where $j'$ is half-integer valued, in the range $j'=0,{1\over 2},1,\cdots,{k\over 2}-1$, and $m'=-j',-j'+1,\cdots, j'$.
For later application we will recall below two examples of supersymmetric $SU(2)_k/U(1)$ coset models where the primary operators and correlators can be easily written down. The $k=2$ case has zero central charge and is a trivial theory. The $k=3$ model will be described as follows.
\subsubsection{Supersymmetric $SU(2)_3/U(1)$ and the compact boson}\label{compact}
In the $k=3$ case, the supersymmetric coset theory $SU(2)_3/U(1)$ is a compact boson\footnote{We add a prime to the field to distinguish from the linear dilaton $\phi$.} $\phi'$ of radius $1/\sqrt{3}$ \cite{Lerche:1989uy}. This compact boson CFT enjoys the $\mathcal{N}=2$ superconformal symmetry with superconformal currents
\ie
&G'^\pm (z) = \sqrt{2\over3} \exp \left[ \pm i \sqrt{3} \phi'(z)\right],\\
&\bar G'^\pm (\bar z) = \sqrt{2\over3} \exp \left[ \mp i \sqrt{3} \bar\phi'(\bar z)\right],
\fe
with dimension ${3\over 2}$. The coefficient $\sqrt{2\over3}$ is fixed by the OPE for $G'^+(z)G'^-(0)= {2c\over 3 z^3} +\cdots$.
The $R$-current can be determined by looking at the OPE for $
G'^+ (z) G'^-(0)$,
\ie
G'^+(z) G'^-(0) &\sim {2c\over 3z^3 } + {2\over z^2 } J_R'(0) + {2\over z} T'(z) +{1\over z} \partial J_R'(0) \\
& \sim {2\over3 z^3}
+ { 2 \over z^2} {i\over \sqrt{3}}\partial \phi'(0) + \cdots,
\fe
hence
\ie
J_R'(z) = {i\over \sqrt{3} } \partial \phi'(z).
\fe
The antiholomorphic $R$-current $\bar J_R'(\bar z) =-{i\over \sqrt{3} } \partial \bar\phi'(\bar z)$ has an extra sign because of our convention for $\bar G'^\pm(\bar z)$.
The Virasoro primary operators are determined as usual
\ie\label{virasoro}
\exp\left[ i \left( { \sqrt{3} n} + {w\over2\sqrt{3}}\right) \phi'(z)\right]
\exp\left[ i \left( { \sqrt{3}n } - {w\over2\sqrt{3}}\right) \bar \phi'(\bar z)\right],~~~~n,w\in \bZ.
\fe
In the following we will determine the superconformal primaries. The $\cN=2$ superconformal primary states $|\Phi\ra$ are Virasoro primaries which in addition satisfy
\ie\label{superprimary}
&G'^\pm_{r} | \Phi\ra= \oint {dz \over 2\pi i} ~z^{r+{1\over2} } G'^\pm (z) \Phi(0)= 0,~~~r>0,\\
&J'_{R,n} | \Phi \ra = \oint {dz\over 2\pi i} z^n J'_R(z) \Phi(0)=0,~~n>0.
\fe
The second condition implies that the OPE for $J'_R(z)\Phi(0)$ can't be more singular than $1/z$, which is always true for the Virasoro primaries \eqref{virasoro}.
In the NS-sector, this implies that the OPE for $G'^\pm (z)\Phi(0)$ can't be more singular than $1/z$. This imposes constraints on the momentum $n$ and the winding number $w$,
\ie
& -1 \le 3n+ {w\over 2} \le 1,\\
& -1 \le 3n-{w\over 2} \le 1,
\fe
and $3n\pm{w\over 2}$ being integers.
It follows that $(n,w)=(0, \pm2)$ or $(n,w)=(0,0)$. The latter is the identity operator. The former ones are
\ie
\text{NS}:~&\exp\left[ i { 1 \over \sqrt{3} } \phi'(z) \right] \exp\left[ - i { 1 \over \sqrt{3} } \bar \phi'(\bar z)\right],~~\Delta=\bar \Delta ={1\over 6},~~R = \bar R= {1\over 3},\\
&\exp\left[ -i { 1 \over \sqrt{3} } \phi'(z) \right] \exp\left[ i { 1 \over \sqrt{3} } \bar\phi'(\bar z)\right],~~\Delta=\bar \Delta ={1\over 6},~~R = \bar R=-{1\over 3}.
\fe
Moving on to the R-sector, the superconformal primary condition \eqref{superprimary} implies that the OPE for $G'^\pm (z)\Phi(0)$ can't be more singular than $1/z^{3/2}$. This imposes constraints on $n,w$ in \eqref{virasoro}:
\ie
& -{3\over 2} \le 3n+{w\over 2} \le {3\over2},\\
& -{3\over 2} \le 3n-{w\over 2} \le {3\over2},
\fe
and $3n\pm{w\over 2}$ being half integers. The possible solutions are $
(n,w) = (0 ,\pm1 ),~(0,\pm3)$, corresponding to the R-sector primary operators
\ie
\text{R}:~~&\exp\left[ \pm i { 1 \over 2\sqrt{3} } \phi'(z)\right] \exp\left[ \mp i { 1 \over 2\sqrt{3} } \bar \phi'(\bar z)\right] ,~~~\Delta=\bar\Delta = {1\over 24},~~R=\bar R = \pm {1\over 6},\\
&\exp\left[ \pm i { 3 \over 2\sqrt{3} } \phi'(z)\right] \exp\left[ \mp i { 3 \over 2\sqrt{3} } \bar\phi'(\bar z)\right] ,~~~\Delta=\bar\Delta = {3\over 8},~~R=\bar R = \pm {1\over 2}.
\fe
\subsection{Spectral flow}
In an ${\cal N}=2$ superconformal theory of central charge $c$, the spectral flow automorphism \cite{Schwimmer:1986mf,Lerche:1989uy}, labeled by a real parameter $\eta$, takes an operator $\cal O$ of weight $\Delta$ and $R$-charge $R$ to another operator ${\cal O}^\eta$ of weight $\Delta^\eta$ and $R$-charge $R^\eta$, related by
\ie
\Delta^\eta=\Delta+\eta R+\eta^2{c\over 6},~~~R^\eta=R+\eta{c\over 3}.
\fe
In the $SL(2)_k/U(1)$ superconformal coset theory, the spectral flowed operator $V^{sl,\eta}_{j,m}$ has weight and $R$-charge
\ie
\Delta={-j(j+1)+(m+\eta)^2\over k}+{\eta^2\over 2},~~~R={2(m+\eta)\over k}+\eta.\label{sfsl}
\fe
On the other hand, in the $SU(2)_k/U(1)$ superconformal coset, the spectral flowed operator $V^{su,\eta'}_{j',m'}$ has
\ie
\Delta={j'(j'+1)-(m'+\eta')^2\over k}+{\eta'^2\over 2},~~~R=-{2(m'+\eta')\over k}+\eta'.\label{sfsu}
\fe
In particular, when $\eta,\eta'$ are $\pm {1\over 2}$, $V^{sl,\pm{1\over 2}}_{j,m}$ and $V^{su,\pm{1\over 2}}_{j',m'}$ are R-sector vertex operators.
\subsection{$\bZ_k$ orbifold}
As already mentioned, the worldsheet CFT in the holographic description of DSLST (either type IIA or type IIB case) is
\ie
\bR^{1,5}\times\left({SL(2)_k\over U(1)}\times {SU(2)_k\over U(1)}\right)/\bZ_k,
\fe
The $\bZ_k$ orbifold is inherited from the holographic dual of the (non-doubly-scaled) LST, with worldsheet description
\ie
\bR^{1,5}\times\bR_\phi\times SU(2)_k.
\fe
Here $\bR_\phi$ is a linear dilaton direction, coming from the radial direction transverse to the stack of NS5-branes.
It is a standard fact that the supersymmetric $SU(2)_k$ WZW model can be written as the $\bZ_k$ orbifold of the product of a supersymmetric $U(1)_k$ WZW model and a supersymmetric coset model $SU(2)_k/U(1)$ \cite{DiFrancesco:1988xz},
\ie
SU(2)_k=\left(U(1)_k\times {SU(2)_k\over U(1)}\right)/\bZ_k. \label{rewriting}
\fe
Before proceeding to the $\bZ_k$ orbifolding in DSLST, let us recall how \eqref{rewriting} works.
\subsubsection{$SU(2)_k$ as a $\bZ_k$ orbifold of $U(1)_k \times {SU(2)_k\over U(1)} $}
We will write down primary operators of the supersymmetric $SU(2)_k$ WZW model in the language of $\left( U(1)_k \times { SU(2)_k\over U(1)} \right)/\bZ_k$. By comparing with the general primary operators in the unorbifolded $ U(1)_k \times { SU(2)_k\over U(1)}$, we will be able to identify the action of the $\bZ_k$ orbifold.
Let $\Phi_{j',m'}$ be a primary of the bosonic $SU(2)_{k-2}$ current algebra. In the supersymmetric $SU(2)_k$ WZW model, we can adjoin $\Phi_{j',m'}$ with $e^{ i \eta' H'}$ (where $i\partial H' = {\lambda'}^+{\lambda'}^-$). We will see shortly that $\eta'$ may be identified as the spectral flow parameter.
Now factor out the $J_3'$ charge,
\ie
\Phi_{j',m'}e^{i\eta'H'}=e^{i\sqrt{2\over k}(m'+\eta')X'}V^{su,\eta'}_{j',m'},\label{idsu2}
\fe
where recall that $X'$ is the bosonization of the $J_3'$ current via $J_3'= i\sqrt{k\over 2} \partial X'$. In the language of $\left( U(1)_k \times {SU(2)_k\over U(1)}\right)/\bZ_k$, $X'$ is the (holomorphic part of the) compact boson at radius $\sqrt{2k}$, that represents the $U(1)_k$. The operator $V^{su,\eta'}_{j',m'}$ is indeed the spectral flowed operator \eqref{sfsu} in the coset model $SU(2)_k/U(1)$, as can be checked by comparing the weights and $R$-charges on both sides.
To complete the identification of vertex operators on the two sides of \eqref{rewriting}, we need to include the antiholomorphic part as well. The vertex operators
coming from (\ref{idsu2}) are
\ie\label{primc}
e^{i\sqrt{2\over k}(m'+\eta')X'-i\sqrt{2\over k}(\bar m'+\bar\eta')\bar X'}V^{su,(\eta',\bar \eta')}_{j',m',\bar m'}.
\fe
On the other hand, if we were considering the {unorbifolded} $U(1)_k\times {SU(2)_k\over U(1)}$, the vertex operators would take the form
\ie\label{prima}
e^{i \sqrt{2\over k} M X' -i \sqrt{2\over k} \bar M \bar X' } V^{su ,(\eta',\bar \eta')}_{j' , m',\bar m'},
\fe
where the quantum numbers $M$ and $\bar M$ satisfy $M-\bar M \in \bZ, ~M+\bar M \in k\bZ$. Under the $\mathbb{Z}_k$ action, the $U(1)$ and $SU(2)/U(1)$ parts of (\ref{prima}) carry charge
$-{M-\bar M\over k}$ and ${m'+\eta' - \bar m'-\bar \eta'\over k}$ (mod 1) respectively.
(\ref{primc}) would be reproduced from (\ref{prima}) with the identification
\ie\label{ida}
M= m' +\eta' ,~~\bar M = \bar m' +\bar \eta'.
\fe
However, the quantization condition on $M$ and $\bar M$ in the $U(1)_k$ are different from that on $m', \bar m'$ of vertex operators in the supersymmetric $SU(2)_k$ via this identification.
The condition $M-\bar M\in \bZ$ translates into
\ie
m' + \eta' - \bar m ' -\bar\eta' \in \bZ. \label{U(1)1}
\fe
The other condition $M+\bar M\in k\bZ$, however, needs to be relaxed. This is achieved by taking the $\mathbb{Z}_k$ orbifold of $U(1)_k\times {SU(2)_k\over U(1)}$. Including twisted sectors now allows for $M+\bar M \in \bZ$, or
\ie
m'+\eta' +\bar m' +\bar\eta'\in \bZ,\label{U(1)2}
\fe
as desired. The special case $m'+\eta' +\bar m' +\bar\eta'\in k\bZ$ gives operators in the untwisted sector.
Note that the orbifold projection demands that the total $\mathbb{Z}_k$ charge vanishes, and this is in particular obeyed by (\ref{ida})
In LST, the $\bZ_k$ orbifold does not act on the linear dilaton $\bR_\phi$. After the deformation to DSLST, the linear dilaton $\bR_\phi$ combined with the $U(1)_k$ will be deformed to the $SL(2)_k/U(1)$ coset theory, on which the $\bZ_k$ orbifolding acts nontrivially. In preparation for the deformed case, let us introduce some notations in the linear dilaton theory.
Let $\phi$ be the linear dilation with background charge $Q^{LST}=\sqrt{2\over k}$ with the standard OPE $\phi(z)\phi(0)\sim-\log z$. Let $\psi_\phi$ be the supersymmetric partner to $\phi$. The linear dilaton fermion $\psi_\phi$ and the third component $\lambda_3'$ of the fermion in the supersymmetric $SU(2)_k$ WZW model pair up to give the current $\psi\psi^*$, where
\ie
\psi = {1\over \sqrt{2} } (\psi_\phi + i \lambda'_3), ~~~\psi^* = {1\over \sqrt{2} } (\psi_\phi - i \lambda'_3).
\fe
After the deformation to DSLST, the current $\psi\psi^*$ turns into $\lambda^+\lambda^-$ in \eqref{slbos} of the $SL(2)_k/U(1)$ coset model. Hence we will use the same symbol $H$ for the bosonization of $\psi\psi^*$,
\ie
\psi\psi^* = i\partial H.
\fe
The total superconformal $R$-current for the internal CFT $\bR_\phi\times \left( U(1)_k \times{ SU(2)_k\over U(1)}\right)/\bZ_k$ can be written in the $\mathbb{R}_\phi\times SU(2)_k$ language as
\ie
J^{tot}_R = i\partial H + i \partial H',\label{totalR}
\fe
where we recall that $i\partial H'=\lambda'^+\lambda'^-$ in the $SU(2)_k$.
Including the linear dilaton $\phi$ and the current $i\partial H = \psi \psi^\ast$, we may consider the more general vertex operator $e^{\sqrt{2\over k} j\phi}e^{i\eta H+i\eta'H'}\Phi_{j',m'}$ in $\bR_\phi\times SU(2)_k$. As before, we can factor out its $J_3'$ charge and write it in the language of $\bR_\phi\times \left( U(1)_k \times{ SU(2)_k\over U(1)}\right)/\bZ_k$ as
\ie
e^{\sqrt{2\over k} j\phi}e^{i\eta H+i\eta 'H'}\Phi_{j',m'}&=e^{\sqrt{2\over k} j\phi}e^{i\eta H}e^{i\sqrt{2\over k}(m'+\eta')X'}V^{su,\eta'}_{j',m'}.\label{defoperator}
\fe
\subsubsection{$\left( {SL(2)_k\over U(1)}\times {SU(2)_k\over U(1)}\right)/\bZ_k$}
Now we will consider the deformation from (the internal part of) the worldsheet theory of LST to that of DSLST \cite{Giveon:1999px,Giveon:1999tq}, namely
\ie
\bR_\phi \times \left(U(1)_k\times {SU(2)_k\over U(1)}\right)/\bZ_k\rightarrow
\left({SL(2)_k\over U(1)}\times {SU(2)_k\over U(1)}\right)/\bZ_k.
\fe
We would like to see how the vertex operator \eqref{defoperator} is deformed.
The weight and the $R$-charge (determined from \eqref{totalR}) of the vertex operator \eqref{defoperator} are
\ie
&\Delta= - {j(j+1)\over k} + {\eta^2+\eta'^2\over2} + {j'(j'+1)\over k},\\
&R = \eta+\eta'.
\fe
In the deformed theory, \eqref{defoperator} maps to the following spectral flowed operator in $\left({SL(2)_k\over U(1)}\times {SU(2)_k\over U(1)}\right)/\bZ_k$
\ie
V^{sl,\eta}_{ j,m'+\eta'-\eta}V^{su,\eta'}_{j',m'}\label{VV},
\fe
which indeed has the same weight and $R$-charge (by \eqref{sfsl} and \eqref{sfsu}).
In other words, the vertex operators $V^{sl,\eta}_{ j,m}V^{su,\eta'}_{j',m'}$ in ${SL(2)_k\over U(1)}\times {SU(2)_k\over U(1)}$ that obey
\ie
m+\eta=m'+\eta',
\fe
as well as the constraints \eqref{U(1)1} and \eqref{U(1)2}, will survive the orbifold.
Combining the holomorphic and antiholomorphic part, we see that a class of vertex operators in the deformed theory $\left({SL(2)_k\over U(1)}\times {SU(2)_k\over U(1)}\right)/\bZ_k$ can be written as
\ie
V^{sl,(\eta,\bar\eta)}_{ j,m,\bar m}V^{su,(\eta',\bar \eta')}_{j',m',\bar m'},
\fe
with the quantum numbers satisfying
\ie
&m+\eta=m'+\eta',~~ \bar m+\bar\eta=\bar m'+\bar \eta',\\
&m'+ \eta' - \bar m' -\bar \eta'\in \bZ,\\
&m'+\eta'+\bar m'+\bar\eta'\in\bZ.\label{Zk}
\fe
The vertex operators with $m'+\eta'+\bar m'+\bar\eta'\in k\bZ$ are in the untwisted sector.
\subsection{Identifications among vertex operators in the coset theories}
There are nontrivial identifications between vertex operators $V^{\eta}_{j,m}$ with different $\eta,~j,~m$ in both the $SL(2)/U(1)$ and $SU(2)/U(1)$ coset theories. These identifications can be traced back to the ones in the bosonic coset theories. Let us start with the bosonic $SU(2)_{k_{bos}}/U(1)$ at level $k_{bos}$, with primary operators $V^{bos}_{j',m'}$. The quantum number $j'$ lies in the range $0\le j' \le {k_{bos}\over 2}$, whereas $m'$, unlike in the bosonic $SU(2)_{k_{bos}}$ WZW model, is \textit{a priori} unconstrained. There is the following identification among primaries labeled by different quantum numbers \cite{Fateev:1985mm,Aharony:2003vk}\footnote{Note that the conformal weight formula $h_{j',m'}={j'(j'+1)\over k_{bos}+2}-{m'^2\over k_{bos}}$ only applies for $|m'|\leq j'$ which labels the primaries obtained directly from factoring out $U(1)$. Within this range, the identification is only nontrivial for $m'=j'$. Otherwise it may be regarded as a way to extend the definition of $V^{bos}_{j',m'}$.}:
\ie
&V^{bos}_{j',m'}=V^{bos}_{{k_{bos}\over 2}-j',-{k_{bos}\over 2}+m'}.
\label{id}
\fe
The identification in particular implies $V^{bos}_{j',m'}=V^{bos}_{j',m'+k_{bos}}$. This is consistent with the statement that ${m'-\bar m'\over k_{bos}}$ (mod 1) is the charge with respect to the residual $\mathbb{Z}_{k_{bos}}$ action on the coset theory.
We now describe the identifications among the vertex operators $V^{su,\eta'}_{j',m'}$ in the supersymmetric $SU(2)_k/U(1)$.
Note the following relation between the primary operator $\Phi_{j',m'}$ for the bosonic $SU(2)_{k-2}$ WZW model, the vertex operator $V^{su,\eta'}_{j',m'}$ for the supersymmetric coset $SU(2)_k/U(1)$, and the primary operator $V^{bos}_{j',m'}$ for the bosonic coset $SU(2)_{k-2}/U(1)$,
\ie
\Phi_{j',m'} e^{i{\eta'} H'} = V_{j',m'}^{su,{\eta'} } e^{ i \sqrt{2\over k } ( {\eta'} +m') X'}
=V^{bos}_{j',m'} e^{ i \sqrt{ 2\over k-2} m' x'} e^{i{\eta'} H'}.
\fe
Using (\ref{relation}), we have
\ie
V^{bos}_{j',m'} = V^{su,{\eta'} } _{j',m'} e^{ - i {\eta'} \sqrt{k-2\over k} X_R '+ i m' {2\over\sqrt{ k(k-2)} }X_R'}
\fe
From (\ref{id}) we have $V_{j',m'}^{bos} = V_{{k-2\over 2} -j', m' -{k-2\over 2}}^{bos}$ (recall that $k_{bos}= k-2$), hence
\ie
V^{su,{\eta'} }_{j',m'} = V^{su,{\eta'} }_{ {k-2\over 2} -j ', m'- {k-2\over 2}} e^{ -i \sqrt{k-2\over k}X_R'} = V^{su,{\eta'} -1}_{{k-2\over2}-j',m'-{k-2\over2}}.
\fe
Note that even though $V^{bos}_{j',m'} = V^{bos}_{j',m' + (k-2)}$, the vertex operator for the supersymmetric $SU(2)_k/U(1)$ is not invariant under a shift on $m'$ alone. Rather, we have
\ie
V^{su,\eta'}_{j',m'
+(k-2)} = V^{su,\eta'-2}_{j',m'}.
\fe
This is consistent with the interpretation of ${m'+\eta'-\bar m'-\bar\eta'\over k}$ (mod 1) as the charge with respect
to the $\mathbb{Z}_k$ symmetry. We can write the above identifications in a more compact form,
\ie
V^{su,{\eta'} }_{j',m'} = V^{su,{\eta'} -1}_{{k-2\over2}-j',m'-{k-2\over2}}= V^{su,{\eta'} +1}_{{k-2\over2}-j',m'+{k-2\over2}}.\label{id1}
\fe
Similarly, primary operators in the bosonic $SL(2)_{k+2}/U(1)$ are subject the following identification\footnote{We are considering here coset primaries that come directly from the lowest weight (resp. highest weight) principal discrete representations of $SL(2)$ \cite{Maldacena:2000hw}, i.e $ D^+_j=\{|j;m\ra:m\in j+{\mathbb N}\}$ (resp. $ D^-_j=\{|j;m\ra:m\in -j-{\mathbb N}\}$) with $-{1\over 2}<j<{k-2\over 2}$, to which the conformal weight formula $h_{j,m}=-{j(j+1)\over k_{bos}-2}+{m^2\over k_{bos}}$ applies. When restricted to this subset of primaries, the identification (2.63) is meaningful only for $j=-m-1$, i.e. it maps the highest weight state of $D^-_j$ to the lowest weight state of $D^+_{{k_{bos}\over 2}-j-2}$. The rest of the relations may be thought of as extending the definition of $V^{bos}_{j,m}$ (see \cite{Parnachev:2001gw} for an explanation of the origin of this identification).} \cite{Parnachev:2001gw,Aharony:2003vk}
\ie\label{idsl}
V^{bos}_{j,m} = V^{bos}_{ {k-2\over2} - j, {k+2\over 2}+m},
\fe
from which we obtain the identification for the supersymmetric $SL(2)_k/U(1)$ theory
\ie
V^{sl,{\eta } }_{j,m} =V^{sl,{\eta } +1}_{{k-2\over2}-j,m-{k+2\over2}}= V^{sl,{\eta } -1}_{{k-2\over2}-j,m+{k+2\over2}}.\label{id2}
\fe
Note the sign difference in $\eta\pm1$ and $m\mp {k+2\over2}$ when compared with $SU(2)_k/U(1)$. Once again, this is consistent with the interpretation of $-{m+\eta-\bar m-\bar\eta\over k}$ (mod 1) as the charge of $V^{sl,(\eta,\bar\eta)}_{j,m,\bar m}$ with respect to the $\mathbb{Z}_k$ symmetry.
\subsection{Massless string states}
Now we discuss the construction of physical vertex operators in type IIA string theory on $\mathbb{R}^{1,5}\times \left({SL(2)_k\over U(1)}\times {SU(2)_k\over U(1)}\right)/\bZ_k$, which is the T-dual description of IIB $(1,1)$ DSLST \cite{Giveon:1999px,Giveon:1999tq}. We will focus on the explicit description of massless string modes in the $\mathbb{R}^{1,5}$, localized at the tip of the cigar. We will further restrict our attention to bosonic string modes, and discuss the (NS,NS) sector and (R,R) sector separately.
\subsubsection{(NS,NS)-sector}
Consider the (NS,NS)-sector vertex operators of the form \cite{Aharony:2003vk,Aharony:2004xn}
\ie\label{vns}
{\cal V}_{NS} = e^{-\varphi-\tilde\varphi} e^{ip_\mu X^\mu} V_{j,m,\bar m}^{sl,{(\eta,\bar\eta) } } V_{j',m',\bar m'}^{su,{(\eta',\bar\eta')} }
\fe
where $\varphi,\tilde\varphi$ are the bosonized superconformal ghosts, $V^{sl,(\eta,\bar\eta)} _{j,m,\bar m}$ and $V^{su,(\eta',\bar\eta')} _{j',m',\bar m'}$ are vertex operators of $SL(2)_k/U(1)$ and $SU(2)_k/U(1)$ respectively, as described earlier. $X^\mu$ with $\mu=0,\cdots,5$ are the bosons for $\bR^{1,5}$ (not to be confused with the bosonization of $J^3 = - \sqrt{k\over2} \partial X$ in the $SL(2)$). The spectral flow parameters $\eta,\bar\eta,\eta',\bar\eta'$ are integer valued in the (NS,NS)-sector. Since $\cV_{NS}$ has no $\bR^{1,5}$ spacetime index, it should be the vertex operator for the six-dimensional scalar fields.
The mass shell condition is
\ie
{1\over 2}p^2+ {(m+{\eta} )^2-j(j+1)\over k}+{{\eta } ^2\over 2} + {j'(j'+1)-(m'+{\eta'} )^2\over k} +{{\eta'} ^2\over 2}= {1\over 2}.
\fe
We will focus on the massless case $p^2=0$. We also demand the quantum numbers to obey the $\bZ_k$ orbifold condition \eqref{Zk}. The on-shell condition for massless states then reduces to
\ie
{ j' (j'+1)- j (j+1)\over k } + {{\eta} ^2\over 2} + {{\eta'} ^2\over 2} ={1\over2}.\label{onshell}
\fe
Next let us examine the chiral GSO projection condition,
\ie
F_L + R \in 2\bZ+1,
\fe
where $F_L$ is the holomorphic worldsheet fermion number (and similarly in the anti-holomorphic sector).
The total $R$-charge can be computed using spectral flow \eqref{sfsl} and \eqref{sfsu},
\ie
R = {2(m+\eta)\over k } + {\eta }
-{2(m'+\eta')\over k} + {\eta'} = \eta + {\eta'},
\fe
where we have used the $\bZ_k$ orbifold condition \eqref{Zk} $m+\eta= m'+\eta'$.
Altogether, we need $F_L+{\eta } +{\eta'} \in 2\bZ+1$ for the GSO condition to be met.
It is straightforward to verify that in the case $F_L=0$, a {\it normalizable} (NS,NS) massless vertex operator of the form (\ref{vns}) that obeys GSO projection condition must satisfy (see Appendix~\ref{n1})
\ie
\eta^2+{\eta'} ^2=1.
\fe
In other words, one of $\eta$ and $\eta'$ must be zero and the other equal to $\pm1$.
Recall that a normalizable vertex operator must have $j=m_0-u$ with $u\in \mathbb{N}$, where $m_0={\rm min}\{|m|, |\bar m|\}$.
Consider first the case $\eta=1,~{\eta'} =0$, and so $m=m'-1$. In this case, the normalizable vertex operators must have $j=j'$.
We observe that $|m+1| = |m'| \le j' = j \leq |m|-u$. Hence we must require $m\le -1$ and $j=|m|-1$. The analysis for $\eta=0, ~{\eta'} =-1$ is identical. The allowed values of the quantum numbers for the normalizable massless vertex operators in these two cases are
\ie
&j={\ell\over 2},~~m = - {\ell+2\over 2},~~~~j' = {\ell\over 2} ,~~m' = - {\ell\over 2},\\
&\text{for}~~~\ell = 0 ,1,\cdots, k-2,
~~~~~(\eta=1,~{\eta'} =0~~{\rm or}~~\eta=0,~{\eta'} =-1).
\fe
The upper bound on $\ell$ comes from the constraint $j' \le {k-2\over 2}$. Note that $m,m'$ are both negative in these cases.
Similarly, for $\eta=0,~{\eta'} =1$, we have $j=j'$ and $m=m'+1$, which implies that $|m-1|\le |m|-u$ for some $u\in \mathbb{N}$. The analysis for $\eta=-1,~{\eta'} =0$ is identical. We end up with the solutions
\ie
&j={\ell\over 2} ,~~ m = {\ell+2\over 2},~~~~j'={\ell\over2},~~ m'={\ell\over2},\\
&\text{for}~~~\ell=0,1,\cdots,k-2,~~~(\eta=0,~{\eta'} =1~~{\rm or}~~\eta=-1,~{\eta'} =0).
\fe
Note that $m,m'$ are both positive in these cases.
To summarize, the normalizable vertex operators from the internal CFT have, in their holomorphic part,
\ie
V_{{\ell\over 2},-{\ell+2\over 2}}^{sl,1} V_{{\ell\over 2},-{\ell\over 2}}^{su,0},~~ V_{{\ell\over 2},{\ell+2\over 2}}^{sl,0} V_{{\ell\over 2},{\ell\over 2}}^{su,1},~~ V_{{\ell\over 2},-{\ell+2\over 2}}^{sl,0} V_{{\ell\over 2},-{\ell\over 2}}^{su,-1},~~ V_{{\ell\over 2},{\ell+2\over 2}}^{sl,-1} V_{{\ell\over 2},{\ell\over 2}}^{su,0},~~~\ell=0,1,\dots,k-2.\label{nsvertex0}
\fe
The operators in (\ref{nsvertex0}) are not all independent, however. Recall that we have the identifications \eqref{id1} and \eqref{id2}, and therefore,
\ie
V_{{\ell\over 2},-{\ell+2\over 2}}^{sl,1} V_{{\ell\over 2},-{\ell\over 2}}^{su,0}&=V_{{k-2-\ell\over 2},{k -\ell\over 2}}^{sl,0} V_{{k-2-\ell\over 2},{k-2-\ell\over 2}}^{su,1},
\\
V_{{\ell\over 2},-{\ell+2\over 2}}^{sl,0} V_{{\ell\over 2},-{\ell\over 2}}^{su,-1}&=V_{{k-2-\ell\over 2},{k -\ell\over 2}}^{sl,-1} V_{{k-2-\ell\over 2},{k-2-\ell\over 2}}^{su,0}.\label{idns}
\fe
In particular, the identification flips the sign of $m$ and $m'$. This will be important when we include the anti-holomorphic part of the vertex operators.
Combining with the anti-holomorphic part, normalizability requires in addition that either $m,\bar m<-{1\over 2}$ or $m,\bar m>{1\over 2}$ (see \eqref{m0102}). This constrains the possible pairings between $\eta,\eta'$ and $\bar\eta,\bar \eta'$. For example, $\eta=1,~\eta'=0$ cannot pair up with $\bar\eta=0, ~\bar \eta'=1$ since in this case $m<0$ but $\bar m>0$. On the other hand, from the identifications \eqref{idns}, some of the pairings are the same as others. For example, $\eta=1,~\eta'=0,~\bar\eta=1,~\bar\eta'=0$ would be identified as $\eta=0,~\eta'=1,~\bar\eta=0,~\bar\eta'=1$.
In the end, there are four inequivalent pairings between the holomorphic and anti-holomorphic quantum numbers that are allowed in the massless vertex operator $V^{sl,(\eta,\bar\eta)}_{j,m,\bar m}V^{su,(\eta',\bar\eta')}_{j',m',\bar m'}$
\ie
\left.\begin{array}{|c|c|c|c|}\hline \eta & \bar\eta & \eta' & \bar \eta' \\\hline ~1~ &~ 1~ & 0 & 0 \\\hline 0 & 0 & -1 & -1 \\\hline ~1~ & 0 & 0 & -1 \\\hline 0 & ~1~ & -1 & 0 \\\hline \end{array}\right.
\fe
In fact, here the normalizability condition relating $j$ to $m$ and $\bar m$ implies that $m=\bar m$, and the orbifold projection condition further implies $m'=\bar m'$. Note that even though $m,\bar m,m',\bar m'$ are all negative for the vertex operators pairing this way (see \eqref{nsvertex0}), we can use the identifications (\ref{idns}) to have them all positive by changing $\eta,\bar \eta,\eta',\bar\eta'$.
The explicit forms of the four sets of normalizable vertex operators are
\ie
&\cV^-_{NS1,\ell}=e^{-\varphi-\tilde\varphi} e^{ip_\mu X^\mu} V_{{\ell\over 2},-{\ell+2\over 2},-{\ell+2\over 2}}^{sl,(1,1)}
V_{{\ell\over 2},-{\ell\over 2},-{\ell\over 2}}^{su,(0,0)}
,
~~\cV^-_{NS2,\ell}=e^{-\varphi-\tilde\varphi} e^{ip_\mu X^\mu}
V_{{\ell\over 2},-{\ell+2\over 2},-{\ell+2\over 2}}^{sl,(0,0)} V_{{\ell\over 2},-{\ell\over 2},-{\ell\over 2}}^{su,(-1,-1)},
\\
&\cV^-_{NS3,\ell}=e^{-\varphi-\tilde\varphi} e^{ip_\mu X^\mu} V_{{\ell\over 2},-{\ell+2\over 2},-{\ell+2\over 2}}^{sl,(1,0)}
V_{{\ell\over 2},-{\ell\over 2},-{\ell\over 2}}^{su,(0,-1)}
,
~\cV^-_{NS4,\ell}=e^{-\varphi-\tilde\varphi} e^{ip_\mu X^\mu}
V_{{\ell\over 2},-{\ell+2\over 2},-{\ell+2\over 2}}^{sl,(0,1)} V_{{\ell\over 2},-{\ell\over 2},-{\ell\over 2}}^{su,(-1,0)},
\fe
or, equivalently, using the identifications \eqref{idns}, we can rewrite them as
\ie
&\cV^+_{NS1,\ell}=e^{-\varphi-\tilde\varphi} e^{ip_\mu X^\mu} V_{{\ell\over 2},{\ell+2\over 2},{\ell+2\over 2}}^{sl,(0,0)}
V_{{\ell\over 2},{\ell\over 2},{\ell\over 2}}^{su,(1,1)}
,
~~\cV^+_{NS2,\ell}=e^{-\varphi-\tilde\varphi} e^{ip_\mu X^\mu}
V_{{\ell\over 2},{\ell+2\over 2},{\ell+2\over 2}}^{sl,(-1,-1)} V_{{\ell\over 2},{\ell\over 2},{\ell\over 2}}^{su,(0,0)},
\\
&\cV^+_{NS3,\ell}=e^{-\varphi-\tilde\varphi} e^{ip_\mu X^\mu} V_{{\ell\over 2},{\ell+2\over 2},{\ell+2\over 2}}^{sl,(0,-1)}
V_{{\ell\over 2},{\ell\over 2},{\ell\over 2}}^{su,(1,0)}
,
~\cV^+_{NS4,\ell}=e^{-\varphi-\tilde\varphi} e^{ip_\mu X^\mu}
V_{{\ell\over 2},{\ell+2\over 2},{\ell+2\over 2}}^{sl,(-1,0)} V_{{\ell\over 2},{\ell\over 2},{\ell\over 2}}^{su,(0,1)},
\fe
with $\ell=0,1,\cdots,k-2$. $\cV^-_{NSi,\ell}$ is related to $\cV^+_{NSi,\ell}$ by \eqref{idns},
\ie
\cV^-_{NSi, \,\ell} =\cV^+_{NSi, \,k-2-\ell},~~~~i=1,2,3,4.
\fe
These $4(k-1)$ vertex operators correspond to $2(k-1)$ complex scalars in the low energy supersymmetric Yang-Mills theory. Only two of the $4(k-1)$ normalizable vertex operators $\cV^-_{NS1,0},\cV^-_{NS2,k-2},$ or equivalently $\cV^+_{NS1,k-2},\cV^+_{NS2,0}$, are in the untwisted sector, i.e. they satisfy the condition $m+\bar m+\eta+\bar\eta \in k\bZ$. The rest are in the twisted sectors.
Let us compare this with the NS5-brane or six dimensional gauge theory description. We are at the point in the Coulomb branch moduli space where the $k$ NS5-branes are spread on a circle in $\mathbb{R}^2\subset \mathbb{R}^4$, with a $\mathbb{Z}_k$ symmetry that permutes the NS5-branes cyclically and at the same time rotates the circle of spread. The center of mass mode decouples, and the relative motion of the NS5-branes gives rise to $4(k-1)$ massless real scalars. We can denote them by complex scalars $ Z_0, Z_1,\cdots, Z_{k-2}$ and $\widetilde Z_1,\widetilde Z_2,\cdots,\widetilde Z_{k-1}$. Here $ Z_i$ are scalars that are linear combinations of collective coordinates of the NS5-brane in the $\mathbb{R}^2$ that contains the circle of spread, while $\widetilde Z_j$ are scalars associated with the remaining traverse $\mathbb{R}^2$. The $\mathbb{Z}_k$ symmetry rotates $ Z_m$ and $\widetilde Z_m$ by the phase $e^{2\pi i m/k}$. Note that $\widetilde Z_0$ is a center of mass mode and decouples, thus absent from the DSLST spectrum. The center of mass mode in the $\phi$ direction, on the other hand, transforms with phase $e^{-2\pi i/k}$ under the $\mathbb{Z}_k$ (the sign in the exponent is a convention). This corresponds to $ Z_{k-1}$ which is absent from the spectrum. So indeed there is only a single massless {\it complex} scalar $ Z_0$ that is uncharged under the $\bZ_k$ symmetry.
\subsubsection{(R,R)-sector}
In the (R,R)-sector, we consider the vertex operators \cite{Aharony:2003vk,Aharony:2004xn}
\ie\label{RVO}
{\cal V}_R=\xi_{a,\dot a}e^{-{\varphi\over 2}-{\tilde\varphi\over 2}}e^{ip_\mu X^\mu}S_{a}\widetilde S_{\dot a}V^{sl,(\eta,\bar\eta)}_{j,m,\bar m}V^{su,(\eta',\bar\eta')}_{j',m',\bar m'},
\fe
where $S_a,~\widetilde S_{\dot a}$ are the spin fields in the $\mathbb{R}^{1,5}$, and $a$ and $\dot a$ are the indices in the $\bf 4$ and $\bf \bar 4$ of $SO(1,5)$ respectively. $\xi_{a,\dot a}$ is the polarization for a six-dimensinoal two-form field strength in the $\bf 15$. In the massless case, $\cV_R$ will give the vertex operators for the field strength of the $U(1)^{k-1}$ gauge bosons.
The on-shell condition for the vertex operator \eqref{RVO} is
\ie
{1\over 2}p^2+ {(m+{\eta} )^2-j(j+1)\over k}+{{\eta } ^2\over 2} + {j'(j'+1)-(m'+{\eta'} )^2\over k} +{{\eta'} ^2\over 2}= {1\over 4}.
\fe
We will focus on the massless case $p^2=0$, and impose the $\bZ_k$ orbifold condition $m+\eta=m'+\eta'$ \eqref{Zk}. The on-shell condition for massless states is then
\ie
{ j' (j'+1)- j (j+1)\over k } + {{\eta} ^2\over 2} + {{\eta'} ^2\over 2} ={1\over4}.
\fe
It's straightforward to derive that (see Appendix \ref{n2}) physical vertex operators surviving the GSO projection $F_L+R\in 2\bZ+{1\over 2}$ must have the following combinations of spectral flow parameters
\ie
\eta=\pm{1\over 2},~~\eta'=\mp{1\over 2}.
\fe
Let us first consider the case $\eta={1\over 2},~\eta'= -{1\over2}$. We must have $j=j'$ and $m=m'-1$. The solutions for the normalizable states are
\ie
&j={\ell\over 2},~~m = - {\ell+2\over 2},~~~~j' = {\ell\over 2} ,~~m' = - {\ell\over 2},\\
&\text{for}~~~~\ell = 0 ,1,\cdots, k-2
~~~~~(\eta={1\over2},~{\eta'} =-{1\over2}).
\fe
Note that $m,m'$ are both negative in this case.
Similarly, in the case $\eta=-{1\over 2},~\eta' = {1\over 2}$, we must have $j=j'$ and $m=m'+1$. The solutions for the normalizable states are
\ie
&j={\ell\over 2} ,~~ m = {\ell+2\over 2},~~~~j'={\ell\over2},~~ m'={\ell\over2},\\
&\text{for}~~~~\ell=0,1,\cdots,k-2,~~~(\eta=-{1\over2},~{\eta'} ={1\over 2}).
\fe
Note that $m,m'$ are both positive in this case.
However, these two sets of vertex operators with $\eta={1\over2},~\eta'=-{1\over2}$ and $\eta=-{1\over2},~\eta'= {1\over2}$ are in fact identified by \eqref{id1} and \eqref{id2},
\ie
V^{sl, 1/2}_{{\ell\over2} , - {\ell+2\over2} } V^{su, -1/2}_{ {\ell\over2} , - {\ell\over2}}
=V^{sl, -1/2}_{{k-2-\ell\over2} , {k -\ell\over2} } V^{su, 1/2}_{ {k-2-\ell\over2} , {k-2-\ell\over2}}.\label{idr}
\fe
Combining with the anti-holomorphic part, as before, normalizability of the vertex operator demands either $m,\bar m< -{1\over2}$ or $m,\bar m>{1\over2}$. The fact that $m$ and $\bar m$ must take the same sign, for instance, rules out the pairing between $\eta={1\over2},~\eta'=-{1\over2}$ in the holomorphic sector with $\bar\eta=-{1\over2},~\bar\eta'={1\over2}$ in the anti-holomorphic sector.
In the end, the gauge boson vertex operators are
\ie
{\cal V}^-_{R,\ell}=\xi_{a,\dot a}e^{-{\varphi\over 2}-{\tilde\varphi\over 2}}e^{ip_\mu X^\mu}S_{a}\widetilde S_{\dot a}
V^{sl,(1/2,1/2)}_{{\ell\over2},-{\ell+2\over 2},-{\ell+2\over 2}}V^{su,(-1/2,-1/2)}_{{\ell\over2},-{\ell\over2},-{\ell\over2}},~~\ell=0,1,\cdots,k-2,\label{gaugevertex-}
\fe
or, equivalently,
\ie
{\cal V}^+_{R,\ell}=\xi_{a,\dot a}e^{-{\varphi\over 2}-{\tilde\varphi\over 2}}e^{ip_\mu X^\mu}S_{a}\widetilde S_{\dot a}
V^{sl,(-1/2,-1/2)}_{{\ell\over2},{\ell+2\over 2},{\ell+2\over 2}}V^{su,(1/2,1/2)}_{{\ell\over2},{\ell\over2},{\ell\over2}},~~\ell=0,1,\cdots,k-2.\label{gaugevertex+}
\fe
The two are related by the reflection \eqref{idr},
\ie
\cV^-_{R,\,\ell} = \cV^+_{R,\, k-2-\ell}.
\label{reflection}
\fe
These $k-1$ vertex operators $\cV^-_{R,\ell}$, $\ell=0,1,\cdots,k-2$, correspond to the $U(1)^{k-1}$ field strengths on the Coulomb branch of the six-dimensional gauge theory. Note that there are no untwisted sector massless RR vertex operators because $m'+\eta'+\bar m' + \bar \eta'= - \ell -1$ is never a multiple of $k$. This is consistent with the fact that there are no $\bZ_k$ invariant gauge bosons at this point on the Coulomb branch of the $SU(k)$ gauge theory. In fact, under the $\mathbb{Z}_k$ action that cyclically permutes the NS5-branes, $\cV^\pm_{R,\,\ell}$ rotates with the phase $e^{\pm {2\pi i\over k}(\ell+1)}$. Namely, the $\mathbb{Z}_k$ momentum for $\cV^\pm_{R,\,\ell}$ is $\pm(\ell+1)$.
\section{Correlators and amplitudes}
The goal of this section is to compute the string tree level amplitude of four gauge bosons in DSLST. This amplitude is expressed in terms of correlators in $\bR^{1,5}$, $SL(2)_k/U(1)$, and $SU(2)_k/U(1)$. The nontrivial part of the $SL(2)_k/U(1)$ correlator is the $SL(2)$ four-point function, which can be related to a correlator in Liouville theory via Ribault and Teschner's dictionary \cite{Ribault:2005wp}. The $SU(2)_k/U(1)$ correlator, while unknown for general $k$, can be written in terms of correlators of a free boson and parafermions for $k = 2, 3, 4, 5$. We write the final scattering amplitude as a series expansion in $\A'/2$:\footnote{The $1/2$ is conventional.}
\ie
\cA_{DSLST} = \cA^{(1)}_{DSLST} + {\A' \over 2} \cA^{(2)}_{DSLST}+ \cdots.
\fe
The main results of this section are the ratios ${\A'\over2}\cA^{(2)}_{DSLST} / \cA^{(1)}_{DSLST}$ (we do not fix the overall normalization of $\cA_{DSLST}$, so only the ratio is unambiguously computed) for different $k$'s presented in Table~\ref{tab:lstresults}. Remarkably, these ratios agree with the ratios between \textit{loop} amplitudes $g_{YM}^2\cA^{2-loop}/\cA^{1-loop}$ in the 6D $SU(k)$ SYM computed in the next section.
\subsection{Winding number conserving correlators}
To begin with, consider the four-point CFT correlator of RR vertex operators ${\cal V}_{R,\ell}^\pm$ on the sphere, of the form
\ie
&\vev{{\cal V}^+_{R,\ell_1}{\cal V}^+_{R,\ell_2}{\cal V}^+_{R,\ell_3}{\cal V}^+_{R,\ell_4} },~~~\vev{{\cal V}^+_{R,\ell_1}{\cal V}^+_{R,\ell_2}{\cal V}^+_{R,\ell_3}{\cal V}^-_{R,\ell_4} },~~~\vev{{\cal V}^+_{R,\ell_1}{\cal V}^+_{R,\ell_2}{\cal V}^-_{R,\ell_3}{\cal V}^-_{R,\ell_4} },~~~{\rm etc}.
\fe
Let us examine the restrictions on the quantum numbers $\ell_i$ and the numbers of $\cV^+$'s versus $\cV^-$'s in a nontrivial correlator due to conservation laws. Recall that the $SL(2)_k/U(1)$ part $V^{sl, \pm 1/2}_{j,m}$ of the R-sector vertex operator \eqref{gaugevertex-} and \eqref{gaugevertex+} is related to the primary operator $\Phi_{j,m}$ for the bosonic $SL(2)_{k+2}$ WZW model by
\ie
\Phi_{j,m}e^{\pm i{1\over 2}H}=V^{sl,\pm1/2}_{j,m}e^{\sqrt{2\over k}(m\pm{1\over 2})X},
\fe
where recall that $i\partial H=\lambda^+\lambda^-$ is the $U(1)$ current constructed out of the free fermions in the supersymmetric $SL(2)_k$, and $J^3=-\sqrt{k\over 2}\partial X$ is a component of the overall $SL(2)$ current. We will restrict ourselves to correlators with conserved $m+\eta$ and $\bar m+\bar\eta$ quantum numbers (loosely referred to as ``winding numbers"), since such correlators in the coset theory can be computed straightforwardly from correlators of primaries of the $SL(2)$ WZW model by factoring out the $U(1)$ part. In particular we will consider winding number conserving correlators of the form $\vev{{\cal V}^+_{R,\ell_1}{\cal V}^+_{R,\ell_2}{\cal V}^-_{R,\ell_3}{\cal V}^-_{R,\ell_4} }$, so that the $\ell_i$'s are subject to the constraint
\ie
\ell_1+\ell_2-\ell_3-\ell_4=0.
\fe
Recall that $\ell_i = 2j_i$ for the RR vertex operators \eqref{gaugevertex-} and \eqref{gaugevertex+}.
For the explicit computation below, we will focus on the special case $\ell_1 = \ell_2 = \ell_3= \ell_4\equiv \ell= 0,1,\cdots, k-2$, corresponding to the assignment of $SL(2)$ quantum numbers
\ie
&j_1= j_2 =j_3 = j_4 = {\ell\over 2},\\
&m_1= m_2 =- m_3 = - m_4 = {\ell+2\over 2},~~\bar m_i =m_i.\label{jonshell}
\fe
In fact, such a correlator would be well defined for the non-normalizable vertex operators with generic values of $j_i$'s as well. The correlator $\la \cV_{R,\ell_1}^+ \cV_{R,\ell_2}^+\cV_{R,\ell_3}^- \cV_{R,\ell_4}^-\ra$ of non-normalizable vertex operators has poles in the $j_i$'s at the values corresponding to normalizable vertex operators. The correlator of normalizable vertex operators is, after integration over the worldsheet, the scattering amplitude $\cA_{DSLST}$ for the corresponding states. It is extracted from the residue of the four-point function of non-normalizable vertex operators, schematically in an LSZ form
\ie\label{vvvv}
\int_{\bC}d^2z\la \cV_{R,\ell_1}^+ \cV_{R,\ell_2}^+\cV_{R,\ell_3}^- \cV_{R,\ell_4}^-\ra \xrightarrow{ j_i \rightarrow {\ell\over 2} }
{ \cA_{DSLST}\over \prod_{i=1}^4( j_i - {\ell\over 2}) }.
\fe
Note that the DSLST amplitude has $\ell\rightarrow k-2-\ell$ reflection symmetry. This is obvious for the particular quantum number assignment due to the identification \eqref{reflection}. More generally, this is a consequence of flipping $\bZ_k$ momenta of the scattering states.
The nontrivial part of (\ref{vvvv}) is the $SL(2)_k/U(1)$ coset CFT correlator
\ie
\left\langle V^{sl, - 1/2}_{j_1, m_1}(z_1)
V^{sl, - 1/2}_{j_2 ,m_2}(z_2)
V^{sl, 1/2}_{j_3,m_3}(z_3)
V^{sl, 1/2}_{j_4,m_4}(z_4)
\right\rangle.
\fe
This is related to the correlator of bosonic $SL(2)$ primaries $\Phi_{j,m,\bar m}$ by
\ie\label{winding}
&\left\la \prod_{i=1}^4 \Phi_{j_i,m_i,\bar m_i}(z_i)\right\ra
\left\la e^{- i {1\over 2} H(z_1)}e^{- i {1\over 2} H(z_2)}e^{ i {1\over 2} H(z_3)}e^{i {1\over 2} H(z_4)}\right\ra\\
& =\left\langle V^{sl, - 1/2}_{j_1, m_1}(z_1)
V^{sl, - 1/2}_{j_2 ,m_2}(z_2)
V^{sl, 1/2}_{j_3,m_3}(z_3)
V^{sl, 1/2}_{j_4,m_4}(z_4)
\right\rangle\\
&~~~~\times
\left\la e^{ \sqrt{2\over k}{\ell+1 \over2} X(z_1)}e^{ \sqrt{2\over k}{\ell+1\over2} X(z_2)}e^{ -\sqrt{2\over k}{\ell+1\over2} X(z_3)}e^{ -\sqrt{2\over k}{\ell+1\over2} X(z_4)}\right\ra.
\fe
The problem is thus reduced to computing the sphere four-point function of $SL(2)$ primaries,
$
\left\la \prod_{i=1}^4 \Phi_{j_i,m_i,\bar m_i}(z_i)\right\ra
$.
As stressed above, this four-point function has poles in $j_i$ as $j_i\rightarrow {\ell\over 2}$. We are only interested in extracting the residue of $\la \prod_{i=1}^4 \Phi_{j_i,m_i,\bar m_i}(z_i)\ra$. We will see in the next subsection that the pole structure is manifest after we rewrite the bosonic $SL(2)_{k+2}$ correlators in terms of Liouville correlators.
\subsection{Bosonic $SL(2)_{k+2}$ correlators and Liouville correlators}
In \cite{Ribault:2005wp}, a relation was established between an $n$-point function of primaries in the bosonic $SL(2)_{k+2}$ WZW model on the sphere and a $(2n-2)$-point function in Liouville theory on the sphere. The Liouville background charge $Q=b+1/b$ is related to the $SL(2)$ level $k$ by
\ie
b^2 = {1\over k}.
\fe
The Liouville cosmological constant $\mu$ is chosen to be $\mu={1\over \pi^2k}$ \cite{Ribault:2005wp}.
Before describing this relation between the correlators, we need to specify the normalization convention on the operators of question. Let $\Phi_{j,m,\bar m}$ be the $SL(2)$ WZW primaries, and $V_\A = e^{2\A\phi}$ be Liouville primaries of conformal weight $\Delta_\A=\A(Q-\A)$. In the convention of \cite{Ribault:2005wp}, the two-point function of Liouville primaries take the form
\ie
\langle V_{\A_2} (z_2) V_{\A_1} (z_1) \rangle = 2\pi \left[ \delta(Q-\A_1-\A_2) + R^L(\A_1) \delta(\A_2-\A_1) \right] \cdot |z_{12}|^{-4\Delta_{\A_1}},
\fe
where $R^L(\A)$ is a reflection coefficient. The $SL(2)$ primaries $\Phi_{j,m,\bar m}(z)$ on the other hand are often conveniently packaged in terms of $\Phi_j(x|z)$, where $x$ is a complex auxiliary variable, such that
\ie
\Phi_{j,m,\bar m}(z) = \int d^2x \, x^{j+m} \bar x^{j+\bar m} \Phi_j(x|z).
\fe
$\Phi_j(x|z)$ are normalized such that their two-point functions take the form\footnote{It was argued in \cite{Aharony:2003vk} that the corresponding string theory two-point function has a slightly different normalization,
\ie
\langle \Phi_j \Phi_j \rangle_{string} = {1\over 2\pi^2} {2j+1\over k}\cdot {R^H(j)\over \pi \gamma(-2j-1)} .
\fe
}
\ie
\langle \Phi_{j_2}(x_2|z_2) \Phi_{j_1}(x_1|z_1) \rangle = \left[ \delta^2(x_{12}) \delta(j_1+j_2+1) + {R^H(j_1)\over \pi \gamma(-2j_1-1)} |x_{12}|^{-4j_1-4} \delta(j_1-j_2) \right] \cdot |z_{12}|^{-4\Delta_j},
\fe
where $\gamma(x)\equiv \Gamma(x)/\Gamma(1-x)$. The function $R^H(j)$ and $R^L(\A)$ coincide under the identification
\ie
\A = -b j + {1\over 2b}.
\fe
Now we can state Ribault and Teschner's relation connecting the $SL(2)$ and Liouville correlators. It takes the form\footnote{We use a different convention for $j$ than in \cite{Ribault:2005wp}: $j_{there} = - ( j_{here}+1)$. Our convention is consistent with that of \cite{Aharony:2003vk}.}
\ie
& \left\langle \prod_{i=1}^n \Phi_{j_i,m_i,\bar m_i}(z_i,\bar z_i) \right\rangle \\
&= {\pi^{2-2n} b\over (n-2)!} {\rm vol}\cdot\delta_{\sum m_i}\delta_{\sum \bar m_i} \prod_{i=1}^n {\Gamma(j_i+m_i+1)\over \Gamma(-j_i-\bar m_i)}
\prod_{1\leq r<s\leq n} z_{rs}^{m_r+m_s + {k\over 2}+1} \bar z_{rs}^{\bar m_r+\bar m_s + {k\over 2}+1}
\\
&~~~~\times
\int \prod_{a=1}^{n-2} d^2 y_a \prod_{1\leq a<b\leq n-2} |y_{ab}|^{k+2} \prod_{r=1}^n \prod_{a=1}^{n-2} (z_r-y_a)^{-m_r-{k\over 2}-1} (\bar z_r-\bar y_a)^{-\bar m_r-{k\over 2}-1}
\\
&~~~~\times \left\langle \prod_{i=1}^n V_{\A_i}(z_i,\bar z_i) \prod_{a=1}^{n-2} V_{-{1\over 2b}}(y_a,\bar y_a) \right\rangle_{Liouville},\label{connection}
\fe
where $z_{rs} = z_r -z_s$ and $y_{ab}= y_a-y_b$;
vol is the volume factor coming from the integration
\ie
{1\over 2\pi} \int {d^2x\over |x|^2}.
\fe
So strictly speaking the correlators of $\Phi_{j,m,\bar m}$'s in the $SL(2)$ WZW model are divergent. The divergent volume factor will cancel against the correlator of the $U(1)$ part of the vertex operators, and the corresponding (winding number conserving) correlator in the $SL(2)/U(1)$ coset theory will be finite.
The Liouville momenta $\A_i$ of $V_{\A_i}$ are mapped to the $SL(2)$ spins $j_i$ via
\ie\label{alphaj}
\A_i = -bj_i+{1\over 2b}.
\fe
Let us now study the pole structure in \eqref{connection} for the case of interest, namely $n=4$, $j_i\rightarrow {\ell\over 2}$, and $m_1=m_2=-m_3= -m_4= {\ell+2\over 2}$. Already the prefactors $\Gamma(j_i + m_i +1)$ give rise to poles in $j_3$, $j_4$ as
\ie
& j_3, \, j_4 \rightarrow {\ell\over 2}
\fe
while $m_3$ and $m_4$ are kept fixed at $m_3=m_4 = - {\ell+2\over 2}$.
The poles in $j_1$ and $j_2$ with positive (and fixed) $m_1$ and $m_2$, on the other hand, are expected to come from the $y_1,y_2$ integral, of the form
\ie\label{yint}
&\int d^2y_1 d^2 y_2 | y_{12}|^{k+2} \prod_{r,a=1}^2 |z_r-y_a|^{- (\ell +k+4)}
\prod_{r=3}^4 \prod_{a=1}^2 |z_r- y_a|^{\ell- k }\\
&
\times\left\la
V_{{k-\ell\over2} b +b\epsilon_1}(z_1)V_{{k-\ell\over2} b +b\epsilon_2}(z_2)
V_{{k-\ell\over2} b}(z_3)
V_{{k-\ell\over2} b }(z_4)
V_{-{1\over 2b}} (y_1)
V_{-{1\over 2b}}(y_2)
\right\ra_{Liouville}
\fe
Here we have already taken the limit $j_3, j_4\to {\ell\over 2} $, while writing
\ie
j_i={\ell\over 2}-{\epsilon_i}~~ \text{for}~~ i=1,2.
\fe
Evidently, the poles in $\epsilon_1, \epsilon_2$ must come from the integration over $y_1$, $y_2$ approaching either $z_1$ or $z_2$. Note that there is no singular contribution in the limits $y_1, y_2\to z_3, z_4$, due to the structure of the OPEs with $V_{-{1\over 2b}}$.
In Liouville theory, degenerate primaries have weights
\ie
\Delta_{m,n} = {Q^{2}\over 4} - {(n'b+m'/b)^2\over 4},~~~~n',m'\geq 1.
\fe
The first null descendant state occurs at level $n'm'$.
The primaries are given by $V_\A = e^{2\A\phi}$ with
\ie
& 2\A = {nb+{m/ b}} = b(n+mk),
\\
& (n,m) = (n'+1,m'+1)~~{\rm or}~~(-n'+1,-m'+1),
\fe
where we need either $n,m\leq 0$, or $n,m\geq 2$.
In particular, the operator $V_{-{1\over 2b}}$ is the degenerate primary labeled by $(n,m)=(0,-1)$.
It has the following OPE with $V_{b+\epsilon b} (z)$,
\ie\label{degOPE}
V_{-{1\over 2b}} (y) V_{ {k-\ell\over2} b +b\epsilon}(z)
&\sim |y-z|^{k-\ell +2 \epsilon}V_{- b{\ell\over 2} + \epsilon b}(z)\\&
+C_-({k-\ell\over2} b +b\epsilon) |y-z|^{k+\ell +2 -2\epsilon} V_{- b{\ell\over 2} +{1\over b} +\epsilon b}(z).
\fe
where the coefficient $C_-(\A)$ is given by
\ie\label{opec}
C_-(\A) = \widetilde\mu^k k^{k+1} {\gamma( {2\A\over b} -1-k) \over \gamma({2\A\over b}) },
\fe
where $\widetilde \mu \equiv \pi \mu \gamma(b^2) b^{2-2b^2}$.
By looking at the limit $y_1,y_2 \rightarrow z_1,z_2$ and using the OPE \eqref{degOPE}, the six-point function of interest in (\ref{yint}) reduces to four-point functions.
Let us first consider the contribution to the six-point function in this limit from the second term in the OPE \eqref{degOPE}, of the form
\ie
C_-({ k-\ell\over2} b+\epsilon_1 b)C_-({k-\ell\over2}b+\epsilon_2 b) \left\langle V_{-b{\ell\over2}+{1\over b} +\epsilon_1 b} (z_1)V_{-b{\ell\over2}+{1\over b} +\epsilon_2 b} (z_2)V_{ {k-\ell\over2} b}(z_3)V_{ {k-\ell\over2} b}(z_4)\right\rangle\label{CC}.
\fe
We have previously asserted that the poles in $\epsilon_1,\epsilon_2$ will come from the integral in $y_1$ and $y_2$. One may worry about further poles coming from the OPE coefficients (\ref{opec}). Indeed, the coefficient $C_-({ k-\ell\over2} b+\epsilon b)$ has a pole in $\epsilon$ at $\epsilon=0$. However, this pole will be canceled by a zero in the four-point function $\left\langle V_{-b{\ell\over2}+{1\over b} +\epsilon_1 b} (z_1)V_{-b{\ell\over2}+{1\over b} +\epsilon_2 b} (z_2)V_{ {k-\ell\over2} b}(z_3)V_{ {k-\ell\over2} b}(z_4)\right\rangle$ at $\epsilon=0$. To see this, we make use of the reflection relation in the Liouville theory
\ie
V_\A = R^L (\A) V_{Q-\A},
\fe
where
\ie
R^L(\A) = -\left[ \pi \mu \gamma(b^2) \right]^{{Q-2\A\over b}}
{\Gamma\left( 1+b(2\A-Q) \right)\Gamma\left( 1+b^{-1}(2\A-Q) \right)\over \Gamma\left( 1-b(2\A-Q) \right)\Gamma\left( 1-b^{-1}(2\A-Q) \right)},
\fe
together with the relation between $C_-$ and reflection coefficients,
\ie
C_-(\A)=R^L (\A) R^L(Q-\A-{1\over 2b}).
\fe
Applying the reflection relation on $V_{-b{\ell\over 2}+{1\over b}+\epsilon b}$, $\epsilon=\epsilon_1, \epsilon_2$, (\ref{CC}) turns into a Liouville four-point function multiplied by two factors of the form
\ie
&C_-({ k -\ell\over 2} b +\epsilon b) R^L( -b{\ell\over 2} +b^{-1} +\epsilon b) \\
&= -[\pi \mu \gamma(b^2)]^{b^{-2} -1} b^{-4}
{ \Gamma(1-k+\ell - 2\epsilon)\Gamma (- \ell-1 +2\epsilon) \over \Gamma(k-\ell) \Gamma(\ell+2) }\times
{\Gamma(2- 2b^2\ell -b^2) \Gamma( k-\ell)\over
\Gamma( b^2\ell +b^2) \Gamma(2-k+\ell-2\epsilon)}\\
&= [\pi \mu \gamma(b^2)]^{b^{-2} -1} b^{-4}
{ \Gamma (- \ell-1 +2\epsilon) \over \Gamma(\ell+2) }\times
{\Gamma(2- 2b^2\ell -b^2) \over
\Gamma( b^2\ell +b^2)} {1\over 1-k+\ell}.
\fe
In these manipulations, we have treated $k=1/b^2$ as a generic real number (as opposed to an integer), so that we do not have to worry about potential poles coming from $\Gamma(-\ell-1+2\epsilon)$. In the end when we take $k$ to be an integer, the pole from $\Gamma(-\ell-1+2\epsilon)$ will cancel against the zero from $1/\Gamma(-j_i-\bar m_i) = 1/\Gamma(-\ell-1+\epsilon_i)$ in (\ref{connection}) for $i=1,2$. On the other hand, the pole in $C_-({ k -\ell\over 2} b +\epsilon b) $ is canceled by the reflection coefficient $R( -b{\ell\over 2} +b^{-1} +\epsilon b)$ for general value of $b$. We are then left with a finite Liouville four-point function $\left\la V_{{\ell+2\over2}b}(z_1)V_{{\ell+2\over2}b}(z_2)V_{{k-\ell\over2} b}(z_3)V_{ { k-\ell\over2} b}(z_4) \right\ra_{Liouville}$.
Putting everything together, the residue of the $SL(2)$ correlator in $\epsilon_i$ with $i=1,\cdots,4$ is computed by
\ie
&\left\langle \prod_{i=1}^4 \Phi_{j_i,m_i,\bar m_i}(z_i) \right\rangle\Big|_{\epsilon_i\rightarrow 0}
= {\pi^{-6} b}\, {\rm vol}\cdot
{\Gamma(-1-\epsilon_3)\Gamma(-1-\epsilon_4)\Gamma(\ell+2)^2\over \Gamma(-\ell-1+\epsilon_1)\Gamma(-\ell-1+\epsilon_2)}\\
&
~~~~\times|z_{12}|^{2\ell +k +6}|z_{34}|^{-2\ell + k -2}|z_{13}|^{k+2}|z_{14}|^{k+2}|z_{23}|^{k+2}|z_{24}|^{k+2}
\\
&~~~~\times
\int d^2y_1 d^2y_2 |y_{12}|^{k+2} |z_1-y_1|^{-2-2\epsilon_1} |z_1-y_2|^{- (\ell+ k +4)} |z_1-y_3|^{\ell-k}|z_1-y_4|^{\ell-k} \\
&~~~~~\times
|z_2-y_1|^{-(\ell+k+4)} |z_2-y_2|^{-2-2\epsilon_2} |z_2-y_3|^{\ell-k} |z_2-y_4|^{\ell-k}\\
&~~~~\times
C_-({k-\ell\over 2} b + \epsilon_1 b)R(- b{\ell\over2}+b^{-1}+\epsilon_1 b)C_-({k-\ell\over 2} b + \epsilon_2 b)R(- b{\ell\over2}+b^{-1}+\epsilon_2 b)\\
&~~~~\times\la V_{{\ell+2 \over 2}b}(z_1)V_{{\ell+2 \over 2}b}(z_2)V_{{k-\ell \over 2}b}(z_3)V_{{k-\ell \over 2}b}(z_4) \ra_{Liouville}
\fe
In the limit $\epsilon_{1,2}\to 0$, there will be two identical contributions to the pole coming from the limiting regions $y_1\rightarrow z_1,y_2\rightarrow z_2$ and $y_1\rightarrow z_2,y_2\rightarrow z_1$,
\ie
&\int d^2y_1 d^2y_2 |y_{12}|^{k+2} |z_1-y_1|^{-2-2\epsilon_1} |z_1-y_2|^{- (\ell+ k +4)} |z_1-y_3|^{\ell-k}|z_1-y_4|^{\ell-k} \\
&~~~~~\times
|z_2-y_1|^{-(\ell+k+4)} |z_2-y_2|^{-2-2\epsilon_2} |z_2-y_3|^{\ell-k} |z_2-y_4|^{\ell-k}\\
&\sim { \pi^2\over \epsilon_1 \epsilon_2} |z_{12}|^{-2\ell - k -6}|z_{13}|^{\ell-k} |z_{14}|^{\ell-k}|z_{23}|^{\ell-k} |z_{24}|^{\ell-k}.
\fe
We then arrive at the relation
\ie\label{relsl}
\left\langle \prod_{i=1}^4 \Phi_{j_i,m_i,\bar m_i}(z_i) \right\rangle\Big|_{\epsilon_i\rightarrow 0}
&\sim {1\over \epsilon_1\epsilon_2\epsilon_3\epsilon_4} \times|z_{34}|^{-2\ell+k-2} |z_{13}|^{\ell+2} |z_{14}|^{\ell+2} |z_{23}|^{\ell+2} |z_{24}|^{\ell+2}\\
&\times\la V_{{\ell+2 \over 2}b}(z_1)V_{{\ell+2 \over 2}b}(z_2)V_{{k-\ell \over 2}b}(z_3)V_{{k-\ell \over 2}b}(z_4) \ra_{Liouville}
\fe
Here we have dropped an immaterial overall constant factor, as well as the vol factor which will cancel against the $U(1)$ correlator in passing to the coset $SL(2)/U(1)$.
Recall the decomposition of the four-point function in terms of conformal blocks in Liouville theory, \footnote{These $\A_i$ are general and are not to be confused with the $\A_i$ identified in \eqref{alphaj}.}
\ie
&\left\langle V_{\A_1}(z_1) V_{\A_2}(z_2) V_{\A_3}(z_3) V_{\A_4}(z_4) \right\rangle \\
&~~~= |z_{14}|^{-4\Delta_1} |z_{24}|^{2(\Delta_1 - \Delta_2 + \Delta_3 - \Delta_4)} |z_{34}|^{2(\Delta_1+\Delta_2-\Delta_3-\Delta_4)} |z_{23}|^{2(\Delta_4-\Delta_1-\Delta_2-\Delta_3)} \\
&~~~\times
\int_{0}^\infty {dP\over 2\pi} C(\A_1,\A_2,{Q\over 2}+iP) C(\A_3,\A_4,{Q\over 2}-iP) |F(\Delta_1,\Delta_2,\Delta_3,\Delta_4; \Delta_P; z)|^2,
\fe
where $z={z_{12}z_{34}\over z_{14}z_{32}}$ is a conformally invariant cross ratio and $\Delta_i = \A_i(Q-\A_i)$ is the conformal weight of $V_{\A_i}$. Here $\Delta_P = {Q^2\over 4} +P^2$. Importantly, in applying this formula, $\A_i$ are assumed to lie on the line ${Q\over 2} + i\mathbb{R}$. In order to go to real values of $\A_i$, for the application to $SL(2)/U(1)$ correlators, one performs an analytic continuation in $\A_i$'s, in which extra residue contribution need to be included whenever a pole of the three-point function coefficient in $P$ crosses the contour as we deform the $\A_i$'s. This significance of this phenomenon was explained in \cite{Maldacena:2001km}.
The Liouville three-point function coefficient $C(\A_1,\A_2,\A_3)$ are given by \cite{Zamolodchikov:1995aa,Teschner:2001gi}
\ie\label{structureconstant}
C(\A_1,\A_2,\A_3) = \widetilde\mu^{Q-\sum\A_i\over b} {\Upsilon_0 \prod_{i=1}^3\Upsilon(2\A_i)\over \Upsilon(\sum\A_i - Q) \Upsilon(\A_1+\A_2-\A_3)\Upsilon(\A_2+\A_3-\A_1)\Upsilon(\A_3+\A_1-\A_2)},
\fe
where $\Upsilon_0 \equiv \Upsilon'(0)$. The function $\Upsilon(x)$ is defined by
\ie
\Upsilon(-bj) = {b^{-{j(j+1)\over k}-j} \over k^{(j+1)(k-j)\over 2k} \Gamma_2(-j-1|1,k) \Gamma_2(k+j+2|1,k) },
\fe
where $\Gamma_2(x|1,\omega)$ is the Barnes double Gamma function.
In particular $\Upsilon(x)$ has zeroes at $x=-nb-m/b$ and $x=(n+1)b+(m+1)/b$, for integers $n,m\geq 0$.
In our case of interest, $\A_1=\A_2={\ell+2\over2}b$ and $\A_3=\A_4={k-\ell\over2}b$, hence
\ie\label{alphadelta}
&\Delta_1 = \Delta_2= {1\over 4k} (\ell+2)(2k-\ell),\\
&\Delta_3=\Delta_4 = {1\over 4k} (k+\ell+2)(k-\ell).
\fe
We therefore obtain from \eqref{relsl}
\ie\label{phi4decomp}
&\left\langle \prod_{i=1}^4 \Phi_{j_i,m_i,\bar m_i}(z_i) \right\rangle\Big|_{\epsilon_i\rightarrow 0}
\sim {1\over \epsilon_1\epsilon_2\epsilon_3\epsilon_4}
{\Gamma(2-2b^2\ell -b^2)^2 \over \Gamma(b^2\ell+b^2)^2}{1\over (1-k+\ell)^2} \\
&~~
\times
|z_{23}|^{ (\ell+2) ({\ell\over k}-1) } |z_{24}|^{\ell+2} |z_{13}|^{\ell+2} |z_{14}|^{ (\ell+2) ({\ell\over k}-1)}\\
&~~ \times\int_{0}^\infty {dP\over 2\pi}C(\A_1,\A_2,{Q\over 2}+iP) C(\A_3,\A_4,{Q\over 2}-iP)
|F(\Delta_1,\Delta_2,\Delta_3,\Delta_4; \Delta_P; z)|^2.
\fe
Once again, we have dropped an overall constant normalization factor that depends on $k$ only.
Note that since the crossing ratio $z$ is invariant under $(z_1,z_2)\leftrightarrow (z_3, z_4)$, the above expression is invariant under the exchange $(1,2)\leftrightarrow (3,4)$. Indeed since we have taken all $j_i$ to be equal to ${\ell\over 2}$, and $m_1=m_2=-m_3=-m_4$, this exchange simply flips the signs of all $m_i$'s, and the correlator remains invariant.
\subsection{Scattering amplitude in the double scaled little string theory}
Now let us put everything together including the $\bR^{1,5}$, $SL(2)_k/U(1)$, and $SU(2)_k/U(1)$ parts, and also the $bc$ and $\beta\gamma$ ghosts, to obtain the scattering amplitudes in the double scaled little string theory. We will expand the amplitudes in power series of the Mandelstam variables $s_{12}$ and $s_{13}$ and evaluate them numerically for $k=2,3,4,5$.
Let us take $z_1=z,~z_2=0, ~z_3=1,~z_4=\infty$. The winding number conserving $SL(2)/U(1)$ correlators are related to the correlators in the bosonic $SL(2)_{k+2}$ WZW model by \eqref{winding}. The correlators involving $H$ and $X$ in \eqref{winding} can be evaluated straightforwardly,
\ie
&\la e^{ - i {1\over 2}H(z)} e^{ - i {1\over 2}H(0)} e^{ i {1\over 2}H(1)} e^{ i {1\over 2}H(\infty)}\ra
= z^{1\over 4}(1-z)^{-{1\over 4}},\\
&\la e^{ \sqrt{2\over k}{\ell+1 \over2} X(z)}e^{ \sqrt{2\over k}{\ell+1\over2} X(0)}e^{ -\sqrt{2\over k}{\ell+1\over2} X(1)}e^{ -\sqrt{2\over k}{\ell+1\over2} X(\infty)}\ra
= {\rm vol}~z^{-{ (\ell+1)^2\over 2k}}(1-z)^{{ (\ell+1)^2\over 2k}}.
\fe
We then have
\ie
& \left\langle V^{sl, - 1/2}_{j_1, m_1}(z_1)
V^{sl, - 1/2}_{j_2 ,m_2}(z_2)
V^{sl, -1/2}_{j_3,m_3}(z_3)
V^{sl, -1/2}_{j_4,m_4}(z_4)
\right\rangle \\
& ={1\over{\rm vol }} \la \prod_{i=1}^4 \Phi_{j_i,m_i} (z_i)\ra \times
z^{ {(\ell+1)^2\over 2k } +{1\over 4} }
(1- z)^{ -{(\ell+1)^2\over 2k } -{1\over 4} }.
\fe
An explicit expression for the bosonic $SL(2)_{k+2}$ four-point function $ \la \prod_{i=1}^4 \Phi_{j_i,m_i} (z_i)\ra$ was given in (\ref{phi4decomp}) in the last subsection.
For the $\bR^{1,5}$ part, the four-point functions are (up to some $z$-independent overall factors)
\ie
&\la c\tilde c e^{ip_1\cdot X}(z,\bar z) e^{ip_2\cdot X}(0,0) c\tilde c e^{ip_3\cdot X}(1,1) c\tilde c e^{ip_4\cdot X}(\infty,\infty)\ra \sim\delta(\sum_{i=1}^4 p_i) |z|^{-{\A' \over 2} s_{12}}|1-z|^{-{\A' \over 2} s_{13}},\\
&\la e^{-\varphi(z)/2}
e^{-\varphi(0)/2}
e^{-\varphi(1)/2}
e^{-\varphi(\infty)/2}\ra = z^{-{1\over4}}(1-z)^{-{1\over4}} ,\\
&\la S_a(z)S_b(0)S_c(1)S_d(\infty)\ra = \epsilon_{abcd}z^{-{1\over4}}(1-z)^{-{1\over4}},
\fe
where $s_{ij} = -2 p_i \cdot p_j$.
The final expression for the four-gluon scattering amplitude with $\mathbb{Z}_k$ momenta $\ell+1,\,\ell+1,\,-\ell-1,\,-\ell-1$ is (up to an immaterial overall factor, the polarization factor, and the momentum conservation delta function $ \delta(\sum_{i=1}^4 p_i)$)
\ie\label{lstamplitude}
\cA_{DSLST}=&\underset{ j_i \rightarrow {\ell\over 2}}{\text{Res}} \, \int_\bC d^2z\ \la \mathcal{V}^+_{R,\ell}(z,\bar z)\mathcal{V}^+_{R,\ell}(0,0)\mathcal{V}^-_{R,\ell}(1,1)\mathcal{V}^-_{R,\ell}(\infty,\infty)\ra \\
& \sim \int_\bC d^2z\ \, |z|^{{(\ell+1)^2\over k}-s_{12} -{1\over2} } |1-z|^{\ell -{(\ell+1)^2\over k} -s_{13} +{1\over2}} \\
~~~~~~~~& \quad \times\,
\la V^{su,(1/2,1/2)}_{ {\ell\over2} ,{\ell\over2},{\ell\over 2}}(z,\bar z)
V^{su,(1/2,1/2)}_{ {\ell\over2} ,{\ell\over2},{\ell\over 2}}(0,0)
V^{su,(-1/2,-1/2)}_{ {\ell\over2} ,-{\ell\over2},-{\ell\over 2}}(1,1)
V^{su,(-1/2,-1/2)}_{ {\ell\over2} ,-{\ell\over2},-{\ell\over 2}}(\infty,\infty)\ra
\\
~~~~~~~~& \quad \times\int_{0}^\infty {dP\over 2\pi}C(\A_1,\A_2,{Q\over 2}+iP) C(\A_3,\A_4,{Q\over 2}-iP)
|F(\Delta_1,\Delta_2,\Delta_3,\Delta_4; \Delta_P; z)|^2,
\fe
where $C(\A_1,\A_2,\A_3)$ and $F(\Delta_1,\Delta_2,\Delta_3,\Delta_4)$ are the structure constant \eqref{structureconstant} and the conformal block of the Liouville theory. Here $Q=b+1/b$ and $b^2=1/k$. $\A_i$ and $\Delta_i$ are given in \eqref{alphadelta}. The Liouville conformal block will be computed using Zamolodchikov's recurrence formula as reviewed in Appendix \ref{zrec}. Once the four-point function $\la V^{su}V^{su}V^{su}V^{su}\ra$ in the supersymmetric $SU(2)_k/U(1)$ coset model is obtained, the integral in \eqref{lstamplitude} can be computed numerically (see Appendix \ref{numerics}).
For the supersymmetric $SU(2)_k/U(1)$ coset model, the general four-point functions are known but rather complicated. We will instead look into a few lower-level examples, where correlators in the supersymmetric $SU(2)_k/U(1)$ theory can be computed from correlators of free bosons and parafermions.
\subsubsection{$k=2$}
The supersymmetric $SU(2)_2/U(1)$ has zero central charge and is a trivial theory. The $\bZ_k$ momentum $\ell+1$ can only be $1$ in the $k=2$ case, i.e., $\ell=0$. The complete scattering amplitude for the double scaled little string theory is (up to some immaterial overall factor and the polarization factor)
\ie\label{ampxa}
\cA_{DSLST}
&\sim
\int_\bC d^2z\ \,
\, |z|^{-{\A' \over 2} s_{12}} |1-z|^{-{\A' \over 2} s_{13}}\\
&
\quad \times \int_{0}^\infty {dP\over 2\pi}C({1\over \sqrt{2}},{1\over \sqrt{2}},{Q\over 2}+iP) C({1\over \sqrt{2}},{1\over \sqrt{2}},{Q\over 2}-iP)
|F(\Delta_1,\Delta_2,\Delta_3,\Delta_4; \Delta_P; z)|^2.
\fe
Here $Q=b+b^{-1}={3\sqrt {2}\over 2}.$
The amplitude can be computed numerically (Appendix \ref{numerics}) and we find
\ie
k=2, ~\ell=0:~~~\cA_{DSLST} \sim 27.92 +28.18 \left( {\A' \over 2} \right)^2 (s_{12}^2+s_{13}^2+s_{14}^2)+\mathcal{O} (\A'^3s^3),
\fe
where we have included the explicit $\A'$-dependence.
The linear term in the Mandelstam variables is absent in this case due to the symmetry in $s_{12}$, $s_{13}$, and $s_{14}$, so ${\cA^{(2)}_{DSLST}}$ is zero.
\subsubsection{$k=3$}
As we have seen in Section \ref{compact}, the supersymmetric $SU(2)_3/U(1)$ coset model is the compact boson $\phi'$ CFT with radius $R=1/\sqrt{3}$.
The relevant primary operators for the gauge boson vertex operators \eqref{gaugevertex-} and \eqref{gaugevertex+} are
\ie
&V^{su, 1/2,1/2}_{ 0, 0 ,0 } = V^{su,-1/2,-1/2}_{{1\over2}, -{1\over 2} , -{1\over 2}},~~\Delta =\bar \Delta= {1\over 24} ,~~R= \bar R ={1\over 6},\\
&V^{su, 1/2,1/2}_{ {1\over 2}, {1\over 2} ,{1\over2} } = V^{su,-1/2,-1/2}_{ 0 , 0,0},~~\Delta =\bar \Delta= {1\over 24} ,~~R= \bar R =-{1\over 6}.
\fe
By comparing the dimensions and the $R$-charges, the above vertex operators can be written in terms of the compact boson $\phi'$ as
\ie
&V^{su, 1/2,1/2}_{ 0, 0 ,0 }(z,\bar z) = V^{su,-1/2,-1/2}_{{1\over2}, -{1\over 2} , -{1\over 2} }(z,\bar z) =
\exp\left[ i { 1 \over 2\sqrt{3} } \phi'(z)\right] \exp\left[ - i { 1 \over 2\sqrt{3} } \bar \phi'(\bar z)\right] ,\\
&V^{su, 1/2,1/2}_{ {1\over 2}, {1\over 2} ,{1\over2} }(z,\bar z) = V^{su,-1/2,-1/2}_{ 0 , 0,0} (z,\bar z)= \exp\left[ - i { 1 \over 2\sqrt{3} } \phi'(z)\right] \exp\left[ i { 1 \over 2\sqrt{3} } \bar \phi'(\bar z)\right] .
\fe
The four-point functions of the supersymmetric $SU(2)_3/U(1)$ coset model can then be computed straightforwardly for each value of the $\bZ_k$ momentum $\ell = 1, 2$,
\ie
& \la V^{su, 1/2,1/2}_{ 0, 0 ,0 } (z,\bar z) V^{su, 1/2,1/2}_{ 0, 0 ,0 } (0,0) V^{su, -1/2,-1/2}_{ 0, 0 ,0 } (1,1) V^{su, -1/2,-1/2}_{ 0, 0 ,0 } (\infty,\infty) \ra \\
&= \la V^{su, 1/2,1/2}_{ {1\over 2}, {1\over 2} ,{1\over 2} } (0,0) V^{su, 1/2,1/2}_{ {1\over 2}, {1\over 2} ,{1\over 2} } (z,\bar z) V^{su, -1/2,-1/2}_{ {1\over 2},-{1\over 2} ,-{1\over 2} } (1,1) V^{su, -1/2,-1/2}_{ {1\over 2}, -{1\over 2} ,-{1\over 2} } (\infty,\infty) \ra\\
&= | z| ^{ {1\over 6} } |1-z|^{ -{1\over 6}}.
\fe
Combining with the correlators in the $\bR^{1,5}$ and $SL(2)_3/U(1)$ parts, we obtain the following scattering amplitude,
\ie\label{ampxb}
\cA_{DSLST}&\sim \int_\bC d^2z\ |z|^{ {(\ell+1)^2 \over 3 } -{\A' \over 2} s_{12} -{1\over3} }
|1-z|^{\ell - {( \ell+1)^2\over 3} - {\A' \over 2} s_{13} +{1\over3} } \\
& \times \int_{0}^\infty {dP\over 2\pi}C(\A_1,\A_2,{Q\over 2}+iP) C(\A_3,\A_4,{Q\over 2}-iP)
|F(\Delta_1,\Delta_2,\Delta_3,\Delta_4; \Delta_P; z)|^2,
\fe
where
\ie
&\A_1 =\A_2 = {\ell+2\over 2\sqrt{3}},~~~\A_3 = \A_4= {3-\ell\over 2\sqrt{3}},
\fe
and $Q = b+b^{-1} = 1/\sqrt{3} + \sqrt{3}$.
The amplitudes are the same for $\ell=0$ and $\ell=1$ as expected. They are\ie
k=3,~\ell=0,1:~~~\cA_{DSLST} \sim 51.28 - 60.11\, \times{\A'\over2} s_{12}+\mathcal{O} (\A'^2s^2),
\fe
The ratio between the subleading and the leading order terms in the $\A'$ expansion is
\ie
k=3,~\ell=0,1: ~~~{\cA^{(2)}_{DSLST} \over s_{12} \cA^{(1)}_{DSLST}}=-1.172.
\fe
\subsubsection{$k=4$}
The relevant operators in the supersymmetric $SU(2)_4/U(1)$ coset model for the gauge boson vertex operators \eqref{gaugevertex-} and \eqref{gaugevertex+} are
\ie
&V^{su, 1/2,1/2}_{ 0, 0 ,0 } = V^{su,-1/2,-1/2}_{{1}, -{1} , -{1}},~~\Delta =\bar \Delta= {1\over 16} ,~~R= \bar R ={1\over 4},\\
&V^{su, 1/2,1/2}_{ {1\over2}, {1\over2} ,{1\over2} } = V^{su,-1/2,-1/2}_{ {1\over2} , -{1\over2},-{1\over2}},~~\Delta =\bar \Delta= {1\over 16} ,~~R= \bar R =0,\\
&V^{su, 1/2,1/2}_{ {1}, {1} ,{1} } = V^{su,-1/2,-1/2}_{ 0 , 0,0},~~\Delta =\bar \Delta= {1\over 16} ,~~R= \bar R =-{1\over 4}.
\fe
These three operators with dimension 1/16 can be realized as the disorder field $\sigma$ in the Ising model and a free boson $\varphi'$ \cite{Mussardo:1988av}:
\ie
&V^{su, 1/2,1/2}_{ 0, 0 ,0 }(z,\bar z) = V^{su,-1/2,-1/2}_{{1}, -{1} , -{1}}(z,\bar z)
= \exp \left[ i {1\over 2\sqrt{2}} \varphi ' (z)\right]
\exp \left[- i {1\over 2\sqrt{2}} \bar\varphi '(\bar z) \right],\\
&V^{su, 1/2,1/2}_{ {1\over2}, {1\over2} ,{1\over2} } (z,\bar z)= V^{su,-1/2,-1/2}_{ {1\over2} , -{1\over2},-{1\over2}}(z,\bar z)=\sigma(z,\bar z),\\
&V^{su, 1/2,1/2}_{ {1}, {1} ,{1} }(z,\bar z) = V^{su,-1/2,-1/2}_{ 0 , 0,0}(z,\bar z)= \exp \left[ -i {1\over 2\sqrt{2}} \varphi '(z) \right]
\exp \left[i {1\over 2\sqrt{2}} \bar\varphi ' (\bar z)\right].
\fe
The four-point functions of the $SU(2)_4/U(1)$ coset model can then be computed straightforwardly \cite{Ginsparg:1988ui} for each value of $\ell$, which takes three possible values, $0,1,2$:
\ie
\ell=0, 2 \quad &\la V^{su, 1/2,1/2}_{ 0, 0 ,0 } (z,\bar z) V^{su, 1/2,1/2}_{ 0, 0 ,0 } (0,0) V^{su, -1/2,-1/2}_{ 0, 0 ,0 } (1,1) V^{su, -1/2,-1/2}_{ 0, 0 ,0 } (\infty,\infty) \ra \\
&= \la V^{su, 1/2,1/2}_{ {1}, {1} ,{1} } (0,0) V^{su, 1/2,1/2}_{ {1}, {1} ,{1} } (z,\bar z) V^{su, -1/2,-1/2}_{ {1},-{1} ,-{1} } (1,1) V^{su, -1/2,-1/2}_{ {1}, -{1} ,-{1} } (\infty,\infty) \ra\\
&= | z| ^{ {1\over 4} } |1-z|^{ -{1\over 4}}\\
\ell=1,~~&\la V^{su, 1/2,1/2}_{ {1\over2}, {1\over2} ,{1\over2} } (z,\bar z) V^{su, 1/2,1/2}_{ {1\over2}, {1\over2} ,{1\over2} } (0,0) V^{su, -1/2,-1/2}_{ {1\over2},-{1\over2} ,-{1\over2} } (1,1) V^{su, -1/2,-1/2}_{ {1\over2}, -{1\over2} ,-{1\over2} } (\infty,\infty) \ra\\
& = {1\over2} |z|^{-{1\over4}} |1-z|^{-{1\over4}} \left(
\Big| 1+ \sqrt{1-z} \Big| + \Big| 1-\sqrt{1-z} \Big|
\right),
\fe
Combining with the correlators in the $\bR^{1,5}$ and $SL(2)_4/U(1)$ parts, we obtain the following scattering amplitude,
\ie\label{ampxc}
\cA_{DSLST}&\sim \int_\bC d^2z\ |z|^{ {(\ell+1)^2 \over 4 } -{\A' \over 2}s_{12} -{1\over4} }
|1-z|^{\ell - {( \ell+1)^2\over 4} -{\A' \over 2}s_{13} +{1\over4} } \\
&
\quad \times \left[ 1-\delta_{\ell,1} +{ \delta_{\ell,1} \over2} |z|^{-{1 \over 2}} \left(
\Big| 1+ \sqrt{1-z} \Big| + \Big| 1-\sqrt{1-z} \Big|
\right) \right]
\\
& \quad \times \int_{0}^\infty {dP\over 2\pi}C(\A_1,\A_2,{Q\over 2}+iP) C(\A_3,\A_4,{Q\over 2}-iP)
|F(\Delta_1,\Delta_2,\Delta_3,\Delta_4; \Delta_P; z)|^2,
\fe
where
\ie
&\A_1 =\A_2 = {\ell+2\over 4}, \quad \A_3 = \A_4= {4-\ell\over 4},
\fe
and $Q = b+b^{-1} = 5/2$.
The amplitudes are the same for $\ell=0$ and $\ell=2$. They are
\ie
k=4,~\ell=0,2:~~~\cA_{DSLST} \sim 78.96 - 144.6\, \times{\A'\over2} s_{12}+\mathcal{O} (\A'^2s^2).
\fe
The ratio between the subleading and the leading order terms in the $\A'$ expansion is
\ie
k=4,~\ell=0,2: ~~~{\cA^{(2)}_{DSLST} \over s_{12} \cA^{(1)}_{DSLST}}=-1.831.
\fe
As for the $\ell=1$ case, the linear term in the Mandelstam variables is absent in this case due to the symmetry in $s_{12}$, $s_{13}$, and $s_{14}$, so ${\cA^{(2)}_{DSLST}}$ is zero.
\subsubsection{$k=5$}
The relevant operators in the supersymmetric $SU(2)_5/U(1)$ coset model for the gauge boson vertex operators \eqref{gaugevertex-} and \eqref{gaugevertex+} are
\ie
&V^{su, 1/2,1/2}_{ 0, 0 ,0 }(z,\bar z) = V^{su,-1/2,-1/2}_{{3\over 2}, -{3\over 2} , -{3\over 2}}(z,\bar z),~~\Delta =\bar \Delta= {3\over 40} ,~~R= \bar R ={3\over 10},\\
&V^{su, 1/2,1/2}_{ {1\over2}, {1\over2} ,{1\over2} } (z,\bar z)= V^{su,-1/2,-1/2}_{ {1} , -{1},-{1}}(z,\bar z),~~\Delta =\bar \Delta= {3\over 40} ,~~R= \bar R ={1\over 10}\\
&V^{su, 1/2,1/2}_{ {1}, {1} ,{1} }(z,\bar z) = V^{su,-1/2,-1/2}_{{1\over2}, -{1\over2} ,-{1\over2}}(z,\bar z),~~\Delta =\bar \Delta= {3\over 40} ,~~R= \bar R =-{1\over 10}\\
&V^{su, 1/2,1/2}_{ {3\over 2}, {3\over 2} ,{3\over 2} }(z,\bar z) = V^{su,-1/2,-1/2}_{ 0 , 0,0}(z,\bar z)~~\Delta =\bar \Delta= {3\over 40} ,~~R= \bar R =-{3\over 10}.
\fe
These four operators with dimension 3/40 can be realized as a free boson $\varphi'$ and the complex order parameter $\sigma$ of dimension 1/15 in the three-state Potts model\footnote{The three-state Potts model is reviewed in Appendix~\ref{Three-state Potts Model}.} and also its complex conjugate $\sigma^*$,
\ie
&V^{su, 1/2,1/2}_{ 0, 0 ,0 }(z,\bar z) = V^{su,-1/2,-1/2}_{{3\over 2}, -{3\over 2} , -{3\over 2}}(z,\bar z)
= \exp \left[ i \sqrt{3\over 20} \varphi ' (z)\right]
\exp \left[- i \sqrt{3\over 20} \bar\varphi '(\bar z) \right],\\
&V^{su, 1/2,1/2}_{ {1\over2}, {1\over2} ,{1\over2} } (z,\bar z)= V^{su,-1/2,-1/2}_{ {1} , -{1},-{1}}(z,\bar z)=\sigma(z,\bar z)\exp \left[ i {1\over \sqrt{60}} \varphi '(z) \right]
\exp \left[-i {1\over \sqrt{60}} \bar\varphi ' (\bar z)\right],\\
&V^{su, 1/2,1/2}_{ {1}, {1} ,{1} }(z,\bar z) = V^{su,-1/2,-1/2}_{{1\over2}, -{1\over2} ,-{1\over2}}(z,\bar z)=\sigma^*(z,\bar z) \exp \left[ -i {1\over \sqrt{60}} \varphi '(z) \right]
\exp \left[i {1\over \sqrt{60}} \bar\varphi ' (\bar z)\right],\\
&V^{su, 1/2,1/2}_{ {3\over 2}, {3\over 2} ,{3\over 2} }(z,\bar z) = V^{su,-1/2,-1/2}_{ 0 , 0,0}(z,\bar z)= \exp \left[ -i \sqrt{3\over 20} \varphi '(z) \right]
\exp \left[i\sqrt{3\over 20} \bar\varphi ' (\bar z)\right].
\fe
The four-point functions of the $SU(2)_5/U(1)$ coset model for each value of $\ell$, which takes three possible values, $0,1,2,3$, are then
\ie
\ell=0, 3 \quad &\la V^{su, 1/2,1/2}_{ 0, 0 ,0 } (z,\bar z) V^{su, 1/2,1/2}_{ 0, 0 ,0 } (0,0) V^{su, -1/2,-1/2}_{ 0, 0 ,0 } (1,1) V^{su, -1/2,-1/2}_{ 0, 0 ,0 } (\infty,\infty) \ra \\
&= \la V^{su, 1/2,1/2}_{ {3 \over 2}, {3 \over 2} ,{3 \over 2} } (z, \bar z) V^{su, 1/2,1/2}_{ {3 \over 2}, {3 \over 2} ,{3 \over 2} } (0,0) V^{su, -1/2,-1/2}_{ {3 \over 2}, -{3 \over 2} ,-{3 \over 2} } (1,1) V^{su, -1/2,-1/2}_{ {3 \over 2}, -{3 \over 2} ,-{3 \over 2} } (\infty,\infty) \ra\\
&= | z| ^{ {3\over 10} } |1-z|^{ -{3 \over 10}}\\
\ell=1,2 \quad &\la V^{su, 1/2,1/2}_{ {1\over2}, {1\over2} ,{1\over2} } (z,\bar z) V^{su, 1/2,1/2}_{ {1\over2}, {1\over2} ,{1\over2} } (0,0) V^{su, -1/2,-1/2}_{ {1\over2},-{1\over2} ,-{1\over2} } (1,1) V^{su, -1/2,-1/2}_{ {1\over2}, -{1\over2} ,-{1\over2} } (\infty,\infty) \ra\\
&= \la V^{su, 1/2,1/2}_{ {1}, {1} ,{1} } (0,0) V^{su, 1/2,1/2}_{ {1}, {1} ,{1} } (z,\bar z) V^{su, -1/2,-1/2}_{ {1},-{1} ,-{1} } (1,1) V^{su, -1/2,-1/2}_{ {1}, -{1} ,-{1} } (\infty,\infty) \ra\\
& = | z| ^{ {1\over 30} } |1-z|^{ -{1\over 30}} G^{Potts}_{\sigma\sigma\sigma^*\sigma^*}(z, \bar z).
\fe
For $\ell = 0,3$, the scattering amplitude is
\ie\label{ampxe}
\cA_{DSLST}&\sim \int_\bC d^2z\ |z|^{ {(\ell+1)^2 \over 5}-{\A' \over 2}s_{12} - {1 \over 5} }
|1-z|^{ \ell - {(\ell+1)^2 \over 5} -{\A' \over 2}s_{13} +{1 \over 5}} \\
& \times \int_{0}^\infty {dP\over 2\pi}C(\A_1,\A_2,{Q\over 2}+iP) C(\A_3,\A_4,{Q\over 2}-iP)
|F(\Delta_1,\Delta_2,\Delta_3,\Delta_4; \Delta_P; z)|^2,
\fe
where
\ie
\A_1 = \A_2 = {1 \over \sqrt 5}, ~~~\A_3 = \A_4 = {5 \over 2 \sqrt 5}, \\
\fe
and $Q = b+b^{-1} = 1/\sqrt5 + \sqrt5$. The amplitude is numerically computed to be
\ie
k=5,~\ell=0,3:~ \cA_{DSLST} \sim 110.4-264.5 \times {\A'\over 2} s_{12} +\mathcal{O}(\A'^2s^2).
\fe
The ratio between the second order and the first order $\A'$ expansion is
\ie
k=5,~\ell=0,3: ~~~{\cA^{(2)}_{DSLST} \over s_{12} \cA^{(1)}_{DSLST}}=-2.397.
\fe
For $\ell = 1,2$, the scattering amplitude is
\ie\label{ampxd}
\cA_{DSLST}&\sim \int_\bC d^2z\ |z|^{ {(\ell+1)^2 \over 5 } -{\A' \over 2}s_{12} -{7\over15} }
|1-z|^{\ell - {( \ell+1)^2\over 5} -{\A' \over 2}s_{13} +{7\over15} } \times G_{\sigma\sigma\sigma^*\sigma^*}^{Potts}(z,\bar z)
\\
\times &\int_{0}^\infty {dP\over 2\pi}C(\A_1,\A_2,{Q\over 2}+iP) C(\A_3,\A_4,{Q\over 2}-iP)
| F(\Delta_1, \Delta_2, \Delta_3, \Delta_4; \Delta_P | z) |^2,
\fe
where
\ie
\A_1 = \A_2 = {3 \over 2\sqrt 5}, ~~~\A_3 = \A_4 = {2 \over \sqrt 5}, \\
\fe
and $G_{\sigma\sigma\sigma^*\sigma^*}^{Potts}(z,\bar z)$ is the four-point function in the three-state Potts model reviewed in Appendix \ref{Three-state Potts Model}. The amplitude is numerically computed to be
\ie
k=5,~\ell=1,2:~ \cA_{DSLST} \sim 220.7 - 304.6 \times {\A'\over 2} s_{12} +\mathcal{O}(\A'^2s^2).
\fe
The ratio between the subleading and the leading order terms in the $\A'$ expansion is
\ie
k=5,~\ell=1,2: ~~~{\cA^{(2)}_{DSLST} \over s_{12} \cA^{(1)}_{DSLST}}=-1.380.
\fe
We summarize the DSLST amplitudes to second order in the $\A'$ expansion in Table~\ref{tab:lstresults}. The overall normalization of the amplitudes (which depends on $k$ and $\ell$) are fixed here, so only the relative coefficients at different orders in the $\A'$ expansion, or equivalently, the expansion in the Mandelstam variables, of a particular amplitude are computed unambiguously. These exactly match with the 1-loop and 2-loop 6D $SU(k)$ SYM amplitudes listed in Table~\ref{tab:2loop}.
It is straightforward to carry out the $\A'$-expansion of the DSLST amplitudes to higher order, and they should be compared to higher than 2-loop amplitudes on the SYM side. We will comment on this later.
\begin{table}[h!]
\centering
\begin{tabular}{c|c|c|c}
$k$ & $\ell$ & ${\cal A}_{DSLST}$ & ${\cal A}_{DSLST}^{(2)} / s_{12}{\cal A}_{DSLST}^{(1)}$ \\\hline\hline
2 & 0 & $27.92 + 0\times s_{12}$ & 0\\\hline
3 & 0, 1 & $51.28 - 60.11 s_{12}$ & $-1.172$ \\\hline
4 & 0, 2 & $78.96-144.6 s_{12}$ & $-1.831$ \\
& 1 & 157.9 + $0\times s_{12}$& 0 \\\hline
5 & 0, 3 & $110.4-264.5s_{12}$ & $-2.397$ \\
& 1, 2 & $220.7-304.6s_{12}$ & $-1.380$ \\
\end{tabular}
\caption{DSLST amplitudes for different $k$ (number of NS5-branes) and $\ell+1$ ($\bZ_k$ momentum) to subleading order in the $\alpha'$ expansion (up to a $k$- and $\ell$-dependent overall normalization factor). The last column records the ratios between the subleading and leading order terms in the $\A'$ expansion, i.e. the coefficient of $s_{12}$ divided by the $s_{12}$-independent term in the full amplitudes. These ratios exactly match with the ratios of the 6D $SU(k)$ SYM amplitudes listed in the last column of Table~\ref{tab:2loop}. Here we set $\A'=2$.}
\label{tab:lstresults}
\end{table}
\section{Comparison to 6D super-Yang-Mills amplitudes}
The strong coupling limit of DSLST, namely LST, reduces to 6D $(1,1)$ $SU(k)$ SYM in the low energy limit.
It is conceivable that the full dynamics of the massless degrees of freedom at the origin of the Coulomb branch is described by a Wilsonian effective action, which is that of $SU(k)$ SYM deformed by an infinite set of higher dimensional operators, with a floating cutoff $\Lambda$.\footnote{A fully supersymmetric regulator is assumed.} It is reasonable to assume that the coefficient of the higher dimensional operators involve non-negative powers of $g_{YM}$ (and arbitrary functions of $\Lambda$).
The DSLST corresponds to a point away from the origin on the Coulomb branch of this theory, at which a $\mathbb{Z}_k\times U(1)$ subgroup of the $SO(4)$ R-symmetry is preserved. In the gauge theory description, scalar vevs of order $\langle\phi\rangle\sim m_W/g_{YM}$ are turned on.
A perturbative scattering amplitude of gluons in the Cartan $U(1)^{k-1}$ at energy $E$ schematically takes the form
\ie
{\cal A}(E) = \sum_{L=0}^\infty (g_{YM} E)^{2L+2} {\cal A}_L({m_W/E}).
\fe
A priori, ${\cal A}_L$ is not quite the same as the $L$-loop amplitude. As we expand the scalars around their vevs, we obtain couplings that involve positive powers of $g_{YM}$ and non-negative powers of $\langle\phi\rangle$. It seems reasonable to assume that the Wilsonian effective Lagrangian on the Coulomb branch at fixed $m_W$ is non-singular at $g_{YM}=0$, thus there should always be more powers of $g_{YM}$ than $\langle\phi\rangle$ in the coefficients of the operators in the Lagrangian.
With such vertices, ${\cal A}_L$ generally receives contributions from diagrams of {\it no more than} $L$ loops. By construction, the $\Lambda$ dependence drops out of ${\cal A}(E)$.
The perturbative amplitude in DSLST, on the other hand, has the structure (after identifying $g_s$ and $\A'$ with gauge theory parameters, as explained in the introduction)
\ie
{\cal A}(E) = \sum_{h=0}^\infty (g_{YM} m_W)^{-2h-2} {\cal A}^{lst}_h \quad \left(\A'E^2={g_{YM}^2E^2\over 16\pi^3}\right)
\fe
where $h$ labels the genus (which is entirely unrelated to the loop order in the SYM theory).
While ${\cal A}_L(m_W/g_{YM})$ is naturally defined by an analytic expansion in $m_W/g_{YM}$ (as the theory is free of infrared divergences), the duality with DSLST suggests that ${\cal A}_L(m_W/g_{YM})$ also has an analytic expansion in $1/m_W^2$ at large $m_W$, namely
\ie
{\cal A}_L(m_W/E) = \sum_{n=1}^\infty \left( {E\over m_W} \right)^{2n} {\cal A}_L^{(n)},
\fe
and in particular, we expect the tree level DSLST amplitude to agree with the $1/m_W^2$ part of the gauge theory amplitude,
\ie
{\cal A}_0^{lst}\left(\A'E^2={g_{YM}^2E^2\over 16\pi^3}\right) = \sum_{L=0}^\infty (g_{YM} E)^{2L+4} {\cal A}_L^{(1)} .
\fe
If we take the cutoff $\Lambda$ to infinity, the $L$-loop amplitude in the undeformed 6D SYM generally diverges for $L\geq 3$. However, as already argued (and verified explicitly in Appendix \ref{3ldiv}), the 3-loop divergence is absent when the external legs are restricted to the Cartan subalgebra. It turns out that the first UV divergence of the Cartan gluon four-point amplitude arises at four-loop.\footnote{We thank the authors of \cite{Bern:2012uf} for pointing this out. The relevant four-loop divergence can be extracted from \cite{Bern:2012uf}.} In any case, one can ask whether the perturbative $L$-loop amplitude in SYM captures (a part of) ${\cal A}_L(m_W/E)$. At $L=1$ the answer is known to be yes, and in fact the finite 1-loop amplitude in SYM, when expanded to first order in $1/m_W^2$, precisely agrees with the leading low energy term in the tree level DSLST amplitude \cite{Aharony:2003vk}. Since the 6D SYM by itself is UV finite at 2-loop as well, one might suspect that ${\cal A}_2(m_W/E)$ is also given entirely by the 2-loop 6D SYM amplitude (with $\Lambda\to \infty$ and no contribution from higher dimensional operators). Remarkably, we will find that this is indeed the case.
\subsection{Structure of perturbative amplitudes}
We will consider four-point amplitudes in $SU(k)$ maximally supersymmetric Yang-Mills theory obtained from unitarity cut methods \cite{Dixon:1996wi,Dixon:2010gz,Bern:1996je,Brandhuber:2010mm,Bern:1997nh,Bern:2010qa}. While such amplitudes are mostly studied in four-dimensional gauge theories, where the $L$-loop result can be expressed in terms of the tree level amplitude together with scalar loop integrals, such formulae admit straightforward generalizations to higher dimensions. It is known that up to 3-loop order in any spacetime dimension $D$, and at 4-loop for $D\leq 6$, the formula derived in four dimensions can be extended to $D$ dimensions by simply replacing the relevant scalar loop integrals by the $D$-dimensional loop integrals~\cite{Dixon:1996wi,Dixon:2010gz}.
We will express the amplitudes in 6D SYM in terms of 6-dimensional spinor helicity variables $\lambda_i^{Aa}$ and $\tilde \lambda_{iB\dot b}$ \cite{Cheung:2009dc,Dennen:2009vk,Bern:2010qa}, where $i=1,2,3,4$ label the external particle legs, $a,\dot b$ are $SU(2)\times SU(2)$ little group indices, and $A,B$ are spinor indices of $SO(6)$ or $SO(5,1)$ Lorentz group. The amplitudes involving various particles in a supermultiplet will be contracted with Grassmann polarization variables $\eta_{ia}$ and $\tilde \eta_i^{\dot b}$ to form a superamplitude.
It is convenient to express the latter in terms of the supermomenta
\ie
q_i^A = \lambda_i^{Aa}\eta_{ia},~~~~ \tilde q_{iB} = \tilde\lambda_{iB\dot b} \tilde\eta_i^{\dot b}.
\fe
The {\it color-ordered} four-point tree-level superamplitude can be written as
\ie
{\cal A}^{tree}(1,2,3,4) = -{i\over s_{12}s_{23}} \delta^8(\sum_{i=1}^4 {\bf q}_i).
\fe
Here $s_{ij}\equiv -(p_i+p_j)^2$ are the Mandelstam variables, and the color factor has been stripped off.
The delta function is the one in Grassmann variables. Explicitly, it can be expanded as
\ie
&\delta^8(\sum_{i=1}^4 {\bf q}_i) = \delta^4(\sum_{i=1}^4 { q}_i^A)\delta^4(\sum_{i=1}^4 {\tilde q}_{iB})
\\
&= {1\over (4!)^2} \sum_{i,j,k,\ell,m,n,r,s=1}^4 \epsilon_{ABCD} q_i^A q_j^B q_k^C q_\ell^D \epsilon^{EFGH} \tilde q_{mE} \tilde q_{nF} \tilde q_{rG} \tilde q_{sH}.
\fe
Suppose we are interested in the 4-scalar amplitude. Then each set of $(\eta_i,\tilde \eta_i)$ should appear in the superamplitude in the combination $1,\, \eta_i^2( = \epsilon^{ab}\eta_{ia}\eta_{ib}),\,\tilde\eta_i^2, \,\eta_i^2\tilde\eta_i^2$. With respect to the $SU(2)\times SU(2)$ R-symmetry (not to be confused with the little group), or rather, $U(1)\times U(1)$ Cartan generators of the R-symmetry group, these scalars have charges $(-,-)$, $(+,-)$, $(-,+)$, $(+,+)$. In the 5-brane description, the two rotations in the transverse $\mathbb{R}^2\times \mathbb{R}^2$ of the 5-brane are linear combinations of these two Cartan generators. So the scalar labeled by $1,\eta^2\tilde\eta^2$ are the collective coordinates in one $\mathbb{R}^2$ (denote by $\phi^{1,2}$), while $\eta^2$ and $\tilde\eta^2$ give scalars parameterizing the other $\mathbb{R}^2$ (denote by $\phi^{3,4}$). For instance, there are terms in the tree superamplitude proportional to $\eta_1^2 \eta_2^2 \tilde\eta_1^2\tilde\eta_2^2$, or $\eta_1^2 \eta_2^2 \tilde\eta_3^2\tilde\eta_4^2$. The former describes a four-point scattering amplitude of the scalars $\phi^1$ and $\phi^2$, while the latter describes a four-point amplitude of $\phi^3$ and $\phi^4$.
In this paper we are interested in the scattering amplitudes of the massless gluons in the Cartan $U(1)^{k-1}$ on the Coulomb branch of $SU(k)$ gauge theory, where the remaining $k^2-k$ $W$-bosons are massive. In SYM there is no tree level amplitude of the Cartan gluons due to the vanishing color factor. The nonvanishing loop amplitudes of the Cartan gluons contain $W$-bosons in the loops (as well as possibly massless gluon propagators at two loops and higher).
\begin{figure}[htb]
\begin{center}
\includegraphics[scale=1.1]{6dym1loop.pdf}
\end{center}
\caption{The 1-loop scalar integral $I_4^{1-loop}(s_{12},s_{14})$.}\label{fig:1loop}
\end{figure}
The {\it full} 1-loop amplitude in SYM is given by\footnote{While ${\cal A}^{tree}(1,2,3,4)$ refers to the color-ordered partial amplitude (the full amplitude is obtained by summing over $s_{12}, s_{13}, s_{14}$ channels), ${\cal A}^{1-loop}(1,2,3,4)$ and ${\cal A}^{2-loop}(1,2,3,4)$ are full amplitudes.} \cite{Bern:1996je} (see also \cite{Brandhuber:2010mm})
\ie
{\cal A}^{1-loop}(1,2,3,4) = - s_{12} s_{23} {\cal A}^{tree}(1,2,3,4) \left[ C_{1234} I_4^{1-loop}(s_{12},s_{23}) + (2\leftrightarrow 3) + (3\leftrightarrow 4) \right],
\fe
where $I_4^{1-loop}(s_{12},s_{14})$ in Figure \ref{fig:1loop} is the 1-loop massless scalar box integral. $C_{1234}$ is the color factor associated with the box diagram. This relation holds in any $D$, which we now take to be $D=6$, and is easily generalized to the case of a $W$-boson loop, where we simply need to replace $I_4^{1-loop}(s_{12},s_{14})$ by\footnote{The generalization to massive propagators in the loop is justified by consideration of unitarity cuts.}
\ie
I_4^{1-loop}&(s_{12},s_{14})\\
&= \int {d^6\ell\over (2\pi)^6} {1\over (\ell^2+m_W^2)((\ell+p_1)^2+m_W^2)((\ell+p_1+p_2)^2+m_W^2)((\ell-p_4)^2+m_W^2)}
\\
&= {1\over m_W^2} \int {d^6\ell\over (2\pi)^6} {1\over (\ell^2+1)^4} + {\cal O}({1\over m_W^4}) .
\fe
In the last line we have expanded the result in $1/m_W^2$. As explained, it is the order $1/m_W^2$ result of the SYM amplitude that will be compared with the {\it genus zero} amplitude of double scaled little string theory.
In the end, the 1-loop amplitude of Cartan gluons on the Coulomb branch can be written in the form
\ie
{\cal A}^{1-loop}(1,2,3,4) = -i \delta^8(\sum_{i=1}^4 {\bf q}_i) \left[ {C_{1234}+C_{1324}+C_{1243}\over m_W^2} {1\over 384\pi^3} + {\cal O}(m_W^{-4}) \right],
\fe
summed over the species of $W$-bosons if $k > 2$.
The {\it full} 2-loop amplitude is given by the tree-level amplitude multiplied by 2-loop scalar integrals \cite{Bern:1997nh}
\ie
\label{2loopamp}
&{\cal A}^{2-loop}(1,2,3,4) = -s_{12} s_{23} {\cal A}^{tree}(1,2,3,4) \\
&~~~~~~\times\bigg[ s_{12} ( {\cal A}_{1234}^{2-loop,P} + {\cal A}_{1234}^{2-loop,NP} + {\cal A}_{3421}^{2-loop,P} + {\cal A}_{3421}^{2-loop,NP})
+({\rm cyclic~in~2,3,4})\bigg],
\fe
Here ${\cal A}^{2-loop,P}_{abcd}$ and ${\cal A}^{2-loop,NP}_{abcd}$ are the color-weighted 2-loop scalar integrals given in Figure~\ref{fig:2loop}. Once again, the propagators in the loops will be replaced by the appropriate massive $W$-boson or massless gluon propagators in the amplitudes on the Coulomb branch of the theory.
\begin{figure}[htb]
\begin{center}
\includegraphics[scale=0.9]{6dym2loop.pdf}
\end{center}
\caption{In (a), the planar 2-loop scalar integral. In (b), the non-planar 2-loop scalar integral.}\label{fig:2loop}
\end{figure}
The 3- and higher-loop amplitudes generally contain logarithmic divergences. It is likely that they still contain nontrivial information that captures the DSLST amplitudes expanded to the corresponding order in $\A'$, but this is beyond the scope of the current paper.
\subsection{Evaluation of color factors and box integrals}
We will label the $W$-bosons in the loops by a pair of gauge indices $(ij)$ ($i,j=1,\cdots,k$, and $i\not=j$). There can also be massless Cartan gluons in the loops, labeled by $(ii)$ (we will not need to impose the traceless condition by hand in this case, as the overall $U(1)$ decouples due to the interaction vertices). The external massless gluons will be labeled by vectors $\vec v_1,\cdots,\vec v_4$ in the Cartan subalgebra of $su(k)$. The mass of the $(ij)$-$W$-boson is
\ie
m_{ij} = r_0 |\omega^i -\omega^j| = 2r_0 \left| \sin {\pi (i-j)\over k} \right|.
\fe
Here $\omega=e^{2\pi i/k}$ is the primitive $k$-th root of unity, $r_0$ is a (radial) Coulomb branch parameter that will be related to the inverse string coupling of DSLST.
Expanding around the point in Coulomb branch with $\mathbb{Z}_k$ symmetry, corresponding to the NS5-branes spreading out on the circle in a transverse $\mathbb{R}^2$, it is convenient to take $\vec v_a$ to be $\mathbb{Z}_k$ charge eigenstates,
\ie
v_a^j = \omega^{(j-1) n_a},~~~j=1,\cdots,k,
\fe
where
the $\mathbb{Z}_k$ momentum $n_a$ is an integer ranging from 1 to $k-1$. We also need $\mathbb{Z}_k$ charge conservation,
\ie
\sum_{a=1}^4 n_a \equiv 0 {\rm~mod~} k.
\fe
As discussed before, the gluon vertex operator $\cV^\pm_{R,\ell}$ in DSLST has $\mathbb{Z}_k$ momentum $\pm (\ell+1)$. Therefore, we see that in order to compare with the DSLST scattering amplitude $\la \cV^+_{R,\ell} (z_1,\bar z_1)\cV^+_{R,\ell} (z_2,\bar z_2)\cV^-_{R,\ell} (z_3,\bar z_3)\cV^-_{R,\ell} (z_4,\bar z_4)\ra$ computed in the last section, the $\mathbb{Z}_k$ charges $n_a$ of interest are\footnote{We shift $n_3$ and $n_4$ by $k$ for later convenience.}
\ie
n_1=n_2 = \ell+1,~~n_3=n_4=k -(\ell+1),
\fe
for $\ell=0,1,\cdots,k-2$.
The 1-loop amplitude, expanded to order $1/m_W^2$, is of the for
\ie
{\cal A}^{1-loop} =
\sum_{i\not=j} {\prod_{a=1}^4 (v_a^i-v_a^j)\over m_{ij}^2} {\cal A}(s_{12},s_{14}) + {\cal O}(m_W^{-4}),
\fe
where
\ie
{\cal A}(s_{12}, s_{14}) =
- { s_{12} s_{13} {\cal A}^{tree}(s_{12},s_{14}) \over 128 \pi^3}.
\fe
Plugging in the explicit expression for $v_a^i$, we can further write
\ie
{\cal A}^{1-loop} = {4k\over r_0^2} {\cal A}(s_{12},s_{14}) \sum_{\ell=1}^{k-1} {\prod_{a=1}^4 \sin{\pi n_a \ell\over k}\over \sin^2({\pi \ell\over k})} + {\cal O}(r_0^{-4})
\fe
As was shown in \cite{Aharony:2003vk}, the sum collapses into a curiously simple answer,
\ie
{\cal A}^{1-loop} = {2k^2\over r_0^2} {\cal A}(s_{12},s_{14}) {\rm min}\{n_a, k-n_a\} + {\cal O}(r_0^{-4}).
\fe
Now consider the 2-loop amplitude. In the planar case, let us label the $W$-boson running through vertices 1,2 by $(ij)$, the $W$-boson running through 3,4 by $(\ell m)$, and the $W$-boson in the middle line $(nr)$. Then ${\cal A}_{1234}^{2-loop,P}$ of (\ref{2loopamp}) is given by
\ie
& \sum_{i,j,\ell,m,n,r} I_4^{2-loop, P}(m_{ij},m_{\ell m}, m_{nr}) \big( \delta_{jn}\delta_{r\ell}\delta_{m i} - \delta_{j\ell}\delta_{mn}\delta_{r i} \big)^2 \prod_{a=1,2} (v_a^i-v_a^j) \prod_{a=3,4} (v_a^\ell-v_a^m) \\
&= 2 \sum_{i\not=j, i\not=\ell} I_4^{2-loop, P} (m_{ij},m_{\ell i}, m_{j\ell }) \prod_{a=1,2} (v_a^i-v_a^j) \prod_{a=3,4} (v_a^\ell-v_a^i),
\fe
where the scalar loop integral is
\ie
& I_4^{2-loop, P} = \int {d^6\ell_1\over (2\pi)^6} {d^6\ell_2\over (2\pi)^6} {1\over (\ell_1^2+m_{ij}^2)
((\ell_1+p_2)^2+m_{ij}^2)((\ell_1+p_1+p_2)^2+m_{ij}^2)}
\\
&~~~~ \times {1\over (\ell_2^2+m_{\ell i}^2)((\ell_2+p_4)^2+m_{\ell i}^2)((\ell_2-p_1-p_2)^2+m_{\ell i}^2) ((\ell_1+\ell_2)^2 + m_{j\ell}^2)}
\\
& = {1\over 4r_0^2} \int {d^6\ell_1\over (2\pi)^6}{d^6\ell_2\over (2\pi)^6} {1\over (\ell_1^2+\sin^2{\pi(i-j)\over k})^3(\ell_2^2+\sin^2{\pi(i-\ell)\over k})^3 ((\ell_1+\ell_2)^2+\sin^2{\pi(j-\ell)\over k})} + {\cal O}(r_0^{-4})
\fe
This gives the color-weighted planar amplitude
\ie
& {\cal A}_{1234}^{2-loop,P} \\
&
= {8 \over r_0^2} \sum_{i\not=j, i\not=\ell} \int {d^6\ell_1\over (2\pi)^6}{d^6\ell_2\over (2\pi)^6} {\prod_{a=1,2}e^{{\pi i n_a\over k}(j-\ell)} \sin({\pi n_a(i-j)\over k}) \prod_{a=3,4}\sin({\pi n_a(i-\ell)\over k}) \over (\ell_1^2+\sin^2{\pi(i-j)\over k})^3(\ell_2^2+\sin^2{\pi(i-\ell)\over k})^3 ((\ell_1+\ell_2)^2+\sin^2{\pi(j-\ell)\over k})} + {\cal O}(r_0^{-4}) \\
&
= {8k \over r_0^2} \sum_{m,r=0}^{k-1} \int {d^6\ell_1\over (2\pi)^6}{d^6\ell_2\over (2\pi)^6} {\prod_{a=1,2}e^{{\pi i n_a\over k}(r-m)} \sin({\pi n_a m \over k}) \prod_{a=3,4}\sin({\pi n_a r\over k}) \over (\ell_1^2+\sin^2{\pi m\over k})^3(\ell_2^2+\sin^2{\pi r\over k})^3 ((\ell_1+\ell_2)^2+\sin^2{\pi(m-r)\over k})} + {\cal O}(r_0^{-4}).
\fe
Similarly,
\ie
& {\cal A}_{1234}^{2-loop,NP} \\
&
=-{8 \over r_0^2} \sum_{i\not=j\not=\ell} \int {d^6\ell_1\over (2\pi)^6}{d^6\ell_2\over (2\pi)^6} {e^{{\pi i n_1\over k}(j-\ell) + {\pi i n_2\over k}(j-i)} \sin({\pi n_1(i-j)\over k}) \sin({\pi n_2(j-\ell)\over k}) \prod_{a=3,4}\sin({\pi n_a(i-\ell)\over k}) \over (\ell_1^2+\sin^2{\pi(i-j)\over k})^2 (\ell_2^2+\sin^2{\pi(i-\ell)\over k})^3 ((\ell_1+\ell_2)^2+\sin^2{\pi(j-\ell)\over k})^2} + {\cal O}(r_0^{-4}) \\
&
= -{8k \over r_0^2} \sum_{0\leq m\not=-r\leq k-1} \int {d^6\ell_1\over (2\pi)^6}{d^6\ell_2\over (2\pi)^6} {e^{{\pi i n_1\over k}r - {\pi i n_2\over k}m} \sin({\pi n_1 m\over k}) \sin({\pi n_2 r\over k})\prod_{a=3,4}\sin({\pi n_a (m+r)\over k}) \over (\ell_1^2+\sin^2{\pi m\over k})^2 (\ell_2^2+\sin^2{\pi r\over k})^2 ((\ell_1+\ell_2)^2+\sin^2{\pi (m+r)\over k})^3 } + {\cal O}(r_0^{-4}).
\fe
These convergent integrals and sums over color factors can be computed numerically. The results for the planar and non-planar contributions to the two-loop amplitude are given in Table~\ref{tab:PvsNP}, and the full two-loop amplitudes, whose expression is given by (\ref{2loopamp}), are listed in Table~\ref{tab:2loop}. We see that the ratios listed in the last column remarkably match with the ratios computed from DSLST that are listed in Table~\ref{tab:lstresults}, to the numerical precision of the conformal block integration. (The matching of ${\cal A}^{1-loop}$ with DSLST amplitudes, where the overall normalization is needed, was already demonstrated in \cite{Aharony:2003vk}.)
\begin{table}[h!]
\centering
\begin{tabular}{c|c|c|c}
$k$ & $\ell$ & $ { g_{YM}^2} {\cal A}^{2-loop}_P / s_{12}$ & ${ g_{YM}^2} {\cal A}^{2-loop}_{NP} / s_{12}$ \\\hline\hline
2 & 0 & 0 & 0 \\\hline
3 & 0, 1 & $-6.048$ & $-4.500$ \\\hline
4 & 0, 2 & $-16.876$ & $-12.435$ \\
& 1 & 0 & 0 \\\hline
5 & 0, 3 & $-34.594$ & $-25.327$ \\
& 1, 2 & $-39.883$ & $-29.136$ \\
\end{tabular}
\caption{The planar and non-planar contributions to the 2-loop amplitudes in leading orders of $1/m_W^2$ expansion for four-gluon scattering in 6D $SU(k)$ SYM. Here we choose the $\mathbb{Z}_k$ charges for the external gluons to be $ n_1 = n_2 =\ell+1$ and $n_3 = n_4=k-(\ell+1)$ with $\ell=0,1,\cdots,k-2$. The numbers are in units of $-s_{12} s_{23} {\cal A}^{tree} / 64 \pi^3 r_0^2$. In order to compare with the DSLST results in Table~\ref{tab:lstresults}, here we set $g_{YM}^2 = 32 \pi^3 $ (see \eqref{idaa} with $\alpha'=2$).}
\label{tab:PvsNP}
\end{table}
\begin{table}[h!]
\centering
\begin{tabular}{c|c|c|c|c}
$k$ & $\ell$ & ${\cal A}^{1-loop}$ & ${ g_{YM}^2 } {\cal A}^{2-loop} / s_{12}$ & ${ g_{YM}^2 } {\cal A}^{2-loop} / s_{12} {\cal A}^{1-loop}$ \\\hline\hline
2 & 0 & 4 & 0 & 0 \\\hline
3 & 0, 1 & 9 & $-10.548$ & $-1.1720$ \\\hline
4 & 0, 2 & 16 & $-29.311$ & $-1.8319$ \\
& 1 & 32 & 0 & 0 \\\hline
5 & 0, 3 & 25 & $-59.922$ & $-2.3969$ \\
& 1, 2 & 50 & $-69.019$ & $-1.3804$ \\
\end{tabular}
\caption{The 1 and 2-loop amplitudes in leading orders of $1/m_W^2$ expansion and their ratios for four-gluon scattering in 6D $SU(k)$ SYM. Here we choose the $\mathbb{Z}_k$ charges for the external gluons to be $ n_1 = n_2 =\ell+1$ and $n_3 = n_4=k-(\ell+1)$ with $\ell=0,1,\cdots,k-2$. ${\cal A}^{1-loop}$ and ${\cal A}^{2-loop}$ are both in units of $-s_{12} s_{23} {\cal A}^{tree} / 64 \pi^3 r_0^2$. In order to compare with the DSLST results in Table~\ref{tab:lstresults}, here we set $g_{YM}^2 = 32 \pi^3 $ (see \eqref{idaa} with $\alpha'=2$). The loop amplitudes ratios exactly match the ratios between different $\A'$ expansion orders of the DSLST amplitudes in Table~\ref{tab:lstresults}. }
\label{tab:2loop}
\end{table}
\section{Discussion}
The tree DSLST amplitudes provides all order results $g_{YM}^2$ and first order in $1/m_W^2$ of the UV completed 6D gauge theory on its Coulomb branch. While the agreement of 6D SYM amplitudes at 2-loop with DSLST at next to leading order in $\A'$ is already striking, a burning question is whether the SYM 3-loop amplitude, which as discussed is finite when the external lines are restricted to the Cartan subalgebra, agrees with the DSLST at next-to-next-to leading order in $\A'$. As the relevant 3-loop superamplitudes have already been reduced to scalar integrals, it is merely a matter of evaluating these scalar integrals to answer the question. We hope to report on the result in the near future.
One may also try to carry out the DSLST amplitude computation to higher genus, and compare with the higher order terms in the $1/m_W^2$ expansion of the SYM amplitude at each loop order. This is not easy as the relevant genus one four-point function in the cigar coset CFT is not yet known, but would nonetheless be interesting.
From the point of view of the Abelian effective action on the Coulomb branch, the 2-loop amplitude of order $1/m_W^2$ comes from the $1/4$ BPS dimension 10 operator of the form $m_W^{-2}D^2F^4+\cdots$. Presumably, our finding suggests a non-renormalization theorem of this term in the Coulomb effective action, with respect to higher dimensional non-BPS operator corrections to the non-Abelian SYM theory. If so, then the 3-loop test will be particularly important, and an agreement with the DSLST tree amplitude at the next order in $\A'$ would be more surprising.\footnote{We thank Ofer Aharony for pointing this out.}
In any case, the big question here is, to what extent will the agreement between the massless amplitudes of pure 6D SYM on the Coulomb branch and DSLST hold, and why do they agree? It so happens that the Cartan gluon amplitude becomes divergent at four-loop \cite{Bern:2012uf}. Therefore we will definitely see some nontrivial disagreement with the LST amplitude at ${\mathcal O}(\A'^3)$. At five-loop, the scalar integrand for the four-point amplitude is known but the UV divergence with external legs in the Cartan subalgebra has yet to be extracted \cite{Bern:2012uc}.\footnote{
We thank the authors of \cite{Bern:2012uf} for explaining to us the results of \cite{Bern:2012uf,Bern:2012uc,Bern:2010tq}.
}
A priori, there could be all sorts of higher dimensional operators that enter the Wilsonian effective action of the 6D gauge theory and correct the amplitudes of the SYM theory itself. After all, we do expect the presence of the dimension 10 non-BPS operator (see for instance \cite{Movshev:2009ba}) as the counter term that cancels the general 3-loop divergence, even though this operator vanishes when the fields are restricted to the Cartan subalgebra. A systematic investigation of the higher dimensional counter terms and their effect on the Cartan gluon scattering amplitude is left to future work.
Finally, let us mention that the W-bosons in the 6D SYM are dual to D1-branes stretched between the NS5-branes. The scatterings of strings with the D1-branes correspond to the scatterings of the Cartan gluons with the W-bosons, and also the scatterings of the D1-branes with themselves are dual to the scatterings of W-bosons. Some aspects of open strings and D-branes in DSLST are studied in \cite{Israel:2005fn} (also see \cite{Israel:2004jt} for the D-branes in $\cN=2$ Liouville theory). It would be interesting to extend their analysis, for example, to the closed string two-point amplitudes on a disc ending on stretched D1-branes, and compare with the scattering amplitudes of two Cartan gluons and two W-bosons.
\bigskip
\section*{Acknowledgments}
We would like to thank Ofer Aharony, Clay C\'ordova, Lance Dixon, Thomas Dumitrescu, Yu-tin Huang, Daniel Jafferis, Ingo Kirsch, Soo-Jong Rey, David Simmons-Duffin, and Andy Strominger for helpful conversations and correspondences at various stages of this project. We would like to thank the Kavli Institute for Theoretical Physics and Aspen Center for Physics during the course of this work.
C.M.C. has been supported in part by a KITP Graduate Fellowship. C.M.C. would like to thank the Physics Department of National Taiwan University for hospitality during the final stage of the work.
S.H.S. is supported by the Kao Fellowship and the An Wang Fellowship at Harvard University.
X.Y. is supported by a Sloan Fellowship and a Simons Investigator Award from the Simons Foundation. This work is also supported by NSF Award PHY-0847457, and by the Fundamental Laws Initiative
Fund at Harvard University.
|
\section{Introduction\label{intro}}
Being a subject of practical importance, elasto-plastic deformation of dense granular media has been under the focus of engineering research for many decades if not
centuries~\cite{schofield,nedderman,wood1990,kolymbas1,kolymbas2,gudehus2010,goddard2013}. The state of geotechnical theories, however, is confusing, at least for physicists: Innumerable
continuum mechanical models compete, employing strikingly different
expressions. In a recent
book on soil mechanics, phrases such as {\em morass of equations} and {\em
jungle of data} were employed as metaphors~\cite{gudehus2010}.
Moreover, this competition is among theories applicable only to the slow shear rates of elasto-plastic
deformation, while rapid dense flow (such as heap flow or
avalanches) is taken to obey yet rather different equations~\cite{hutter2007}. All this renders a unified theory capable of accounting for granular phenomena at different rates seemingly illusory.
This is the reason we adopted a different approach, focusing on the physics and leaving the rich and subtle granular phenomenology aside while constructing the theory. Our hope was to arrive at a set of equations that is firmly based in physics, broadly applicable, and affords a well founded, correlated understanding of granular media.
The formalism we employ is called the {\em hydrodynamic theory} (which physicists take to be the long-wave-length, continuum-mechanical theory of condensed systems, in contrast to its more widespread usage, as a synonym for the Navier-Stokes' equations).
The hydrodynamic formalism was pioneered by Landau~\cite{LL6} and Khalatnikov~\cite{Khal} in the context of superfluid helium, and
introduced to complex fluids by de Gennes~\cite{deGennes}. Its two crucial points are:
The input in physics that specifies the complete set of state variables, and the simultaneous consideration of energy and momentum conservation. As a result, there are many more constraints, and far less liberty, than the usual approach of constitutive relations. Moreover, being derived from physics rather than a subset of experimental data, if the theory renders some phenomena correctly, chances are that the rest is also adequately accounted for.~\footnote{We note there are also constitutive approaches which starts successfully from physics, more specifically from micromechanical properties of granular ensembles~\cite{microMech}.}
Hydrodynamic theories~\cite{hydro-1,hydro-2} have been derived for many condensed
systems, including liquid crystals
~\cite{liqCryst-1,liqCryst-2,liqCryst-3,liqCryst-4,liqCryst-5,liqCryst-6,liqCryst-7},
superfluid $^3$He~\cite{he3-1,he3-2,he3-3,he3-4,he3-5,he3-6},
superconductors~\cite{SC-1,SC-2,SC-3}, macroscopic
electro-magnetism~\cite{hymax-1,hymax-2,hymax-3,hymax-4},
ferrofluids~\cite{FF-1,FF-2,FF-3,FF-4,FF-5,FF-6,FF-7,FF-8,FF-9,FF-10}, and
polymers~\cite{polymer-1,polymer-2,polymer-3,polymer-4}.
We contend that a hydrodynamic theory is also useful and possible for granular media: Useful, because it should help to illuminate and order their
complex behavior; possible, because a significant portion is already accomplished. We call it ``granular solid hydrodynamics,'' abbreviated as {\sc gsh}.
The structure of {\sc gsh} is, as far as we can see, adequate and complete. Starting from two basic notions, {\it two-stage irreversibility} and {\it variable transient elasticity}, we have set up the theory in~\cite{granR2,granR3,granRgudehus,granR4}. In this paper, we focus on
applying these equations to varying circumstances, a large collection of
experiments. In fact, no other continuum mechanical theory comes even close. ({\sc gsh} is summarized in Sec.\ref{GSH}. It is not a derivation, only meant to keep this paper self-contained.)
There are two aspects of {\sc gsh} that we need to communicate: the ideology of its approach and the number of experiments it accounts for. Some of our starting points, such as energy conservation or the validity of thermodynamics, are not generally accepted in the granular community.
We have detailed our reasons why we believe our postulates are appropriate in~\cite{granR2,granR3,granRgudehus,granR4}, and shall not repeat them here. One of our hopes for the present paper is that the second aspect of {\sc gsh}, impressive and easily accessible, is also quietly convincing -- or at least thought-provoking, for those who still have doubts about the basic approach of {\sc gsh}.
\section{A Brief Presentation of GSH \label{GSH} }
As any hydrodynamic theory, {\sc gsh} has two parts, structure and parameters. The first is derived from general principles, but the second -- values and functional dependence of the energy and transport coefficients -- are inputs, obtained either from a microscopic theory (a tall order in any dense systems), or in a trial-and-error iteration, in which the ramifications of postulated dependences are compared to experiments and simulations.
Many details of granular phenomena depend on these parameters, and we are still in the midst of the iteration evaluating them. More specifically, we have an energy expression that is both simple and realistic, but the transport coefficients are in a less satisfactory state: Their dependence on $T_g$, obtained from more general considerations, seems quite universal, but the density dependence is not. Varying with details possibly including rigidity, shape and friction of the grains, they are material-specific and hard to arrive at in the absence of more systematic data. These need to be given by a complete range of experiments in uniform geometries employing only {\it one kind of grains}. Nevertheless, in spite of the tentative character of the density dependence assumed below, our results do show at least qualitative agreement with experiments and realistic constitutive models
\subsection{The State Variables}
A {\it complete set of state variables} is one that
uniquely determines a macroscopic state of the system. If it is given, there is no room
for ambiguity or ``history-dependence." Conversely, any such
dependences indicate that the set is incomplete. In the hydrodynamic
theory, a physical quantity is a state variable if the
energy density $w$ depends on it. In {\sc gsh}, the state variables are, in
addition to the usual ones (the density $\rho$, the momentum density $\rho
v_i$, the true entropy $s$): the granular entropy $s_g$ and the elastic strain
$u_{ij}$. Entropy $s_g$, along with $T_g\equiv\partial w/\partial s_g$,
quantifies granular jiggling and is closely associated with the averaged velocity
fluctuation $\delta\bar v\equiv\sqrt{\langle v_i^2\rangle -\langle
v_i\rangle^2}$. (It would be wrong to take $T_g\sim\delta\bar v^2$, because
any kinetic theory fails for $T_g\to0$, when enduring contacts dominate,
see~\cite{GGas}, also~\cite{granR2,granR3,granR4}.)
The elastic
strain $u_{ij}$ is associated with the deformation of the grains (or in DEM-jargon: their {\it overlap}). We do not consider the true entropy $s$ below, although it is undoubtedly a state variable, because effects such as thermal expansion are not at present under our focus.
Fabric anisotropy $f_{ij}$, the number of average contacts in different
directions, is a useful microscopic characterization of granular states. But there is insufficient evidence that it is macroscopically {\it independent}. To keep {\sc
gsh} as simple as possible, our working hypothesis is that it is not. In~\cite{luding2011}, Magnanimo and Luding employ $f_{ij}$ to account for
the anisotropic velocity of elastic waves, because their theory uses linear elasticity and does not have stressed-induced anisotropy. {\sc gsh} does and
yields velocities very close to the measured ones, without $f_{ij}$, see~\cite{ge4}. We note that anisotropy of elastic waves that persists for
isotropic stress and $u_{ij}$ would be a sign that $f_{ij}$ is an independent variable.
Denoting the (rest-frame or internal) energy density as $w=w(\rho, s_g, u_{ij})$,
we define the conjugate variables as:
\begin{equation}\label{2-2} \mu\equiv\frac{\partial w}{\partial\rho},\quad T_g\equiv\frac{\partial
w}{\partial s_g},\quad \pi_{ij}\equiv-\frac{\partial w}{\partial u_{ij}},
\end{equation}
calling $\mu$ is the chemical potential, $T_g$ the
granular temperature, and $\pi_{ij}$ the elastic stress. These are given once the energy $w$ is. (See~\cite{granR2,granR3,granRgudehus,granR4} for a a treatment including the true entropy $s$ and temperature $T\equiv{\partial
w}/{\partial s}$.)
There are three spatial scales in any granular media: (a)~the macroscopic, (b)~the mesoscopic or inter-granular, and (c)~the microscopic or inner granular. Dividing all degrees of freedom (DoF) into these three categories, we treat those of (a) differently from (b,c). Macroscopic
DoF, such as the slowly varying stress, flow and density fields, are
employed as state variables, but inter- and inner granular DoF are treated summarily:
Only their contributions to the energy are considered and taken,
respectively, as granular and true heat. So we do not account for the motion of a jiggling grain, only include its fluctuating kinetic and elastic energy as contributions to the granular
heat, $\int T_g{\rm d}s_g$.
Similarly, phonons are part of true heat, $\int T{\rm d}s$. There are a handful of macroscopic DoF~(a), many inter-granular ones (b), and innumerable inner granular ones (c). So the statistical tendency to equally distribute the energy among all DoF implies an energy decay: (a) $\to$ (b,c) and (b) $\to$ (c).
(In kinetic theories, assuming $T_g\gg T$ holds, the (b) $\to$ (c) decay is replaced by a constant restitution coefficient~\cite{granR4}.)
This is what we call {\em two-stage irreversibility}
With $v_{ij}\equiv\frac12(\nabla_iv_j+\nabla_jv_i)$,
$v^*_{ij}$ its traceless part, $v_s^2\equiv v^*_{ij}v^*_{ij}$, the balance equation for $s_g$ (closely related to the energy balance in the kinetic theory~\cite{luding2009}) is
\begin{equation}\label{2c-4}
\partial_ts_g+\nabla_i(s_gv_i-\kappa\nabla_iT_g)=
(\eta_g v_s^2+\zeta_gv^2_{\ell\ell}-\gamma T_g^2)/T_g.
\end{equation}
Here, $s_gv_i$ is the convective, and $-\kappa\nabla_iT_g$ the diffusive, flux. $\eta_g v_s^2$ accounts for viscous heating, for the increase of $T_g$ because a
macroscopic shear rate jiggles the grains. A compressional rate
$\zeta_gv^2_{\ell\ell}$ does the same, though not as effectively~\cite{granL3}. The term $-\gamma T_g^2$ accounts for the relaxation of $T_g$, the (b) $\to$ (c) decay of energy.
Our second notion, {\em variable transient elasticity}, addresses the interplay between elaticity and plasticity.
The free surface of a granular system at rest is frequently tilted. When
perturbed, when the grains jiggle and $T_g\not=0$, the tilted surface will
decay and become horizontal. The stronger the grains jiggle
and slide, the faster the decay is. We take this as indicative of a system
that is elastic for $T_g=0$, transiently elastic for $T_g\not=0$,
with a stress relaxation rate $\propto T_g$.
A relaxing stress is typical of any viscous-elastic system such as polymers~\cite{polymer-1}.
The unique circumstance here is that the relaxation rate is not a material
constant, but a function of the state variable $T_g$. As we shall see, it
is this {\em variable transient elasticity} -- a simple fact at heart -- that underlies the complex behavior of granular plasticity. This is an insight that yields a most economic way to capture granular rheology.
Employing the strain rather than stress as a state variable yields a simpler description, because the former is a geometric quantity, the latter a physical one (that includes material constants such as the stiffness). Yet one cannot use the standard strain $\varepsilon_{ij}$, because the relation between stress and $\varepsilon_{ij}$ lacks uniqueness when the system is plastic.
Engineering theories frequently divide the strain into two fields, elastic $u_{ij}$ and plastic
$\varepsilon^{p}_{ij}$, with the first accounting for
the reversible and second for the irreversible part. They then employ $\varepsilon_{ij}$ and $\varepsilon^{p}_{ij}$ as two independent variables to account for the elasto-plastic motion~\cite{Houlsby,Houlsby2}.
We believe that, on the contrary, the elastic strain $u_{ij}$
is the sole state variable. As convincingly argued by Rubin~\cite{rubin}, there is a unique relation between $u_{ij}$ and the elastic stress $\pi_{ij}$. We take $u_{ij}$ as the portion of the strain that deforms the grains, changes the elastic energy $w=w(u_{ij})$, and builds up an elastic stress $\pi_{ij}$. Employing $u_{ij}$ preserves useful features of elasticity, especially the relation, $\pi_{ij}=-\partial w(u_{ij})/\partial u_{ij}$, cf.\cite{granR2}.
This is easy to understand via
an simple analogy. The wheels of a car driving up a snowy hill will grip the ground part of the time, slipping otherwise. When the wheels grip, the car moves and its gravitational energy $w$ is increased (same as only $u_{ij}$ increases the elastic energy). Dividing the wheel's rotation $\theta$ into a gripping $\theta^{(e)}$ and a slipping $\theta^{(p)}$ portion, we may compute the torque on the wheel as $\partial w/\partial\theta^{(e)}$ [same as $\pi_{ij}=-\partial w(u_{ij})/\partial u_{ij}$]. How much the wheel turns or slips, how large $\theta$ or $\theta^{(p)}$ are, is irrelevant for the torque.
The equation for $u_{ij}$ is
\begin{equation}\label{2c-6a}
\partial_tu_{ij}-v_{ij}+\alpha_{ijk\ell}v_{k\ell}=-(\lambda_{ijk\ell}T_g)\,u_{k\ell},
\end{equation}
cf.\cite{granR2} for the general expression including the objective derivative. (In contrast to the total strain, the change in the elastic one $u_{ij}$ remains small, rendering the additional terms irrelevant -- unless one wants to describe, say, a rotating sand pile.) If
$T_g$ is finite, grains jiggle and briefly
lose or loosen contact with one another, during which their deformation is
partially lost. Macroscopically, this shows up as a relaxation of
$u_{ij}$, with a rate that grows with $T_g$, and vanishes for
$T_g=0$, with the lowest order term in a $T_g$-expansion being
$\lambda_{ijk\ell}T_g$. Within its range of stability, the energy $w$ is convex, and $-\pi_{ij}\equiv\partial w/\partial u_{ij}$ is a monotonic function of $u_{ij}$. So $-\pi_{ij}, u_{ij}$
decrease and relax at the same time, in accordance to Eq~(\ref{2c-6a}).
Conservation of momentum, $\partial_t(\rho v_i)+\nabla_j(\sigma_{ij}+\rho v_iv_j)=g_i\rho$ and mass, $\partial_t\rho=-\nabla_i(\rho v_i)$, close the set of equations. The Cauchy stress $\sigma_{ij}$ is (see~\cite{granR2,granR3,granRgudehus,granR4}):
\begin{align}\label{ST}
\sigma_{ij}=\pi_{ij}-\alpha_{k\ell ij}\pi_{k\ell}+ (P_T-\zeta_g v_{\ell\ell})\delta_{ij}-\eta_gv^*_{ij},\\
P_T\equiv-\partial(w/\rho)/\partial (1/\rho)=Ts+T_gs_g+\mu\rho-w,
\end{align}
where $P_T$ (that will turn out to be the kinetic pressure) and $\pi_{ij}$ are given by Eqs.(\ref{2-2}). The total stress $\sigma_{ij}$, though generally valid, is explicit only if $w$ is given. The terms $\propto\zeta_g,\eta_g$ are the viscous stress; the tensor $\alpha_{ijk\ell}$ is an off-diagonal Onsager coefficient that couples the stress components and softens them. The above expressions yield the structure of {\sc gsh}.
Next, we specify the energy and transport coefficients.
\subsection{The Energy\label{granEn}}
Due to a lack of interaction among the grains, the energy density $w$ vanishes when the grains are neither deformed nor jiggling. Assuming $w=w_T(\rho,s_g)+w_\Delta(\rho,u_{ij})$, we have $w_T\to0$ for $s_g\to0$, and $w_\Delta\to0$ for $u_{ij}\to0$. So, considering slightly excited, stiff grains (such that the lowest order terms in $u_{ij},s_g$ suffice), we take
\begin{align
\label{2b-2} w_T={s_g^2}/{(2\rho b)},\quad
w_\Delta=\sqrt{\Delta }(2 {\mathcal B} \Delta^2/5+ {\mathcal A}u_s^2),
\\\label{2b-2a}
\pi_{ij}=\sqrt\Delta({\cal B}\Delta+{\cal A}
{u_s^2}/{2\Delta})\delta _{ij}-2{\cal A}\sqrt\Delta\, u_{ij}^*,
\\\label{2b-2b}
P_\Delta=\sqrt\Delta({\cal B}\Delta+{\cal A}
{u_s^2}/{2\Delta}),\quad \pi_s=-2{\cal A}\sqrt\Delta\, u_s,
\\\label{2b-1}
{4P_\Delta}/{|\pi_s|}=2({\cal B}/{\cal
A})(\Delta/u_s)+{u_s}/{\Delta},\qquad\qquad \end{align}
where $\Delta\equiv -u_{\ell\ell}$, $P_\Delta\equiv\pi_{\ell\ell}/3$, $u_s^2\equiv
u^*_{ij}u^*_{ij}$, $\pi_s^2\equiv \pi^*_{ij}\pi^*_{ij}$, with
$u^*_{ij},\pi^*_{ij}$ the respective traceless tensors.
$w_T$ is an expansion in $s_g$. The quadratic term is the lowest order one
because $s_g\sim T_g=0$ is an energy minimum. (As we shall soon see, the $s_g^2$-term is in fact sufficient to account for fast dense flow and the gaseous state.)
Calling something a temperature, we also give it the dimension kelvin or energy. Taking $[s_g]=$ 1/vol, $[T_g]=$ energy, implies $[1/\rho b]=$ energy $\times$ vol. But we note the following point: Equilibration, or equality of temperatures, is usually a ubiquitous process, and what requires all temperatures to possess the same dimension. However, granular media in ``thermal contacts" do not usually equilibrate -- in the sense that the energy distribution is independent of details, and the energy flux vanishes. Given two different granular systems, 1 and 2, with only 1 being driven, there are, in the steady state, four temperatures: $T^1,T_g^1,T^2,T_g^2$, with an ongoing energy transfer: $T_g^2\to T^2$ and $T^1_g\to T^1,T_g^2$, such that none of the temperatures is equal to another. The differences depend on details such as the contact area and the respective restitution coefficients. Only when the driving stops, will they eventually become equal, but this is well approximated by $T_g^1=T_g^2=0$. Therefore, there is no harm in giving $s_g$ or $T_g$ any dimension -- as long as $T_gs_g$ is an energy density.
Given Eq.(\ref{2b-2}) with $b=b(\rho)$, there is quite generally a pressure contribution $P_T$,
\begin{equation}
-P_T\equiv\left.\frac{\partial (w_T/\rho)}{\partial 1/\rho}\right|_{s_g}=\left.\frac{\partial [(w_T-T_gs_g)/\rho]}{\partial 1/\rho}\right|_{T_g}=\frac{T_g^2\rho^2}{2}\frac{\partial b}{\partial\rho}.
\end{equation}
We choose $b=b(\rho)$ such that it yields the kinetic pressure $\propto w_T$ for the rarefied limit $\rho\to0$, and the usual form $\propto
w_T/(\rho_{cp}-\rho)$ in the dense limit $\rho\to\rho_{cp}$, see~\cite{Bocquet,luding2009},
\begin{equation}\label{2b-5}
b=b_1\rho^{a_1}+b_0\left[1-\frac{\rho}{\rho_{cp}}\right]^a,\quad
P_T=\frac{w_T}{b}\left[\frac{ab\cdot\rho/\rho_{cp}}{1-\rho/\rho_{cp}}-a_1b_1\rho^{a_1}\right]\equiv g_p(\rho)T_g^2, \end{equation}
with $a\approx0.1$ a small positive number, and $-a_1=2/3, 1$ for two and three dimensions, respectively. For $\rho\to0$, we have $b\approx b_1$, $P_T\approx-a_1w_T$, with $w_T=\frac12\rho\delta\bar v^2=\frac32T_k\rho/m$ in three dimensions (where $\delta\bar v^2\equiv{\langle v_i^2\rangle -\langle v_i\rangle^2}$, $T_k$ denotes the temperature of the kinetic theory, and $P_T=T_k\rho/m$ the usual kinetic pressure). In the dense limit, the first term in $P_T$ dominates, and the pressure is as desired $\propto w_T/(\rho_{cp}-\rho)$. (The term $\propto b_1$ is new, and not in \cite{granR2,granR3}.)
Without equilibration, there is no thermometers that measures $T_g$. It is
therefore useful to relate $T_g$ to $\delta\bar v$, a quantity that is directly
measurable, at least in simulations. This is easily done for two limits, because
$w_T=\frac12\rho\delta\bar v^2$ or $\delta\bar v=T_g\sqrt{b}$ in the rarefied
one; and $w=\rho\delta\bar v^2$ or $\delta\bar v=T_g\sqrt{b/2}$ in the dense
one. (For $\rho\to\rho_{cp}$, granular jiggling occurs in a network of linear oscillators, which oscillate weakly around the static stress. So there is on
average as much potential energy as kinetic one.) We note that, for given $\delta\bar v$, the energy $w_T$ remains finite in both limits, although $b$
diverges and $T_g$ vanishes for $\rho\to0$.
The expression for $w_\Delta$, with ${\cal A,B}>0$, is the elastic contribution. Given by the energy of linear elasticity multiplied by $\sqrt\Delta$, the form is clearly inspired by the Hertzian contact, though its connection to granular elasticity goes beyond that, and includes both {\it stress-induced anisotropy} and the {\it convexity transition} (see below). The elastic stress $\pi_{ij}$ has been validated for the following circumstances, achieving at least semi-quantitative agreement:
\begin{itemize}
\item Static stress distribution in three classic geometries: silo, sand pile,
point load on a granular sheet, calculated employing $\nabla_i\pi_{ij}=\rho g_i$, see~\cite{ge1,ge2,granR1}.
\item Incremental stress-strain relation, starting from varying static stresses~\cite{kuwano2002,ge3}.
\item Propagation of anisotropic elastic waves at varying static stresses~\cite{jia2009,ge4}.
\end{itemize}
{\it Stress-induced anisotropy}: In linear elasticity, $w\propto u_s^2$, the velocity of an elastic wave $\propto\sqrt{\partial^2w/\partial u_s^2}\,$ does not depend on $u_s$, or equivalently, the stress. For any exponent other than 2, the velocity depends on the stress, and is anisotropic if the stress is.
We note that $u_{ij}$ and $\pi_{ij}$ from the expression of Eq.(\ref{2b-2a}) are colinear, in the sense that $u_{ij}^*/u_s=\pi_{ij}^*/\pi_s$ holds (but not $\varepsilon_{ij}$). They also have the same principal axes. More recently, we have employed a slightly more complicated $w_\Delta$ that includes the third strain invariant~\cite{3inv}. Here, colinearity is lost, but strain and stress still share the same principle axis.
{\it Convexity Transition}:
In a space spanned by stress components and the density, there is a surface that
divides two regions in any granular media, one in which the grains are necessarily agitated,
another in which they may be in a static, non-dissipating state. The most obvious such surface exists with respect to the density -- when it is too small,
grains loose contact with one another and cannot stay static. Same holds if the shear stress is too larger for given pressure, say when the slope of a sand pile is too steep.
Note the collapse occurs in a completely static system. This is qualitatively different from the critical state, because the latter, and the approach to it, takes place in a dissipating system, at given rate and $T_g$. These two require different descriptions, static versus dynamic. We consider the static description here, and shall return to the critical state in Sec~\ref{critical state}.
In Eq.(\ref{2b-4}), we introduce two material parameters, $\rho_{\ell p}$ and $\rho_{cp}$. Calling the first {\it the random-loose density}, we take it to be the lowest density at which any elastic state may be maintained, where elastic solutions are stable. The second, termed {\it random-close density}, is taken as the highest one at which grains may remain uncompressed. For lack of space, grains cannot rearrange at $\rho_{cp}$, and do not execute any plastic motion.
The divide between two regions, one in which elastic solutions are stable, and another in which they are not, in which infinitesimal perturbations suffice to destroy the solution, is the surface where the second derivative of the elastic energy changes its sign, where it turns from convex to concave.
The elastic energy of Eq~(\ref{2b-2}) is convex only for
\begin{equation}\label{2b-3} u_s/\Delta\le\sqrt{2{\cal B}/{\cal A}} \quad
\text{or}\quad \pi_s/P_\Delta\le\sqrt{2{\cal A}/{\cal B}},
\end{equation}
turning concave if the condition is violated. (The second condition may be derived by considering Eq~(\ref{2b-1}),
showing $P_\Delta/\pi_s=\sqrt{{\cal B}/2{\cal A}}$ is minimal for
$u_s/\Delta=\sqrt{2{\cal B}/{\cal A}}$.)
Assuming ${\cal B}/{\cal A}$ is density-independent (typically 5/3), denoting $\bar\rho\equiv(20\rho_{\ell p}-11\rho_{cp})/9$,
we take
\begin{equation}\label{2b-4} {\cal B}={\cal B}_0
[(\rho-\bar\rho)/(\rho_{cp}-\rho)]^{0.15},
\end{equation}
with ${\cal B}_0>0$ a constant. This expression accounts for three granular characteristics:
\begin{itemize}
\item The energy is concave for any density smaller than $\rho_{\ell p}$.
\item The energy is convex between $\rho_{\ell p}$ and $\rho_{cp}$, ensuring the stability of any elastic solutions in this region. In addition, the density dependence of sound
velocities (as measured by Harding and Richart~\cite{hardin}) is well
rendered by $\sqrt{{\cal B}(\rho)}$.
\item The elastic energy diverges, slowly, at
$\rho_{cp}$, approximating the observation that the system becomes an order of magnitude stiffer there.
\end{itemize}
One may be bothered by the small exponent of 0.15, questioning whether we imply an accuracy over a few orders of magnitude. We do not: Since $\cal B$ loses its convexity at $\rho_{\ell p}$, the density is never close to $\bar\rho$ (note $\bar\rho<\rho_{\ell p}<\rho_{cp}$, with $\rho_{cp}-\rho_{\ell p}\approx\rho_{\ell p} -\bar\rho$). And although $\rho$ may in principle be close to $\rho_{cp}$, it is very difficult to reach, and the slow divergence is not really relevant. Given $ {\cal B}(\rho)$, there is also a contribution $\propto\Delta^{2.5}$ to $P_T$ from $w_\Delta$. It is neglected because it is (for small $\Delta$) much smaller than the elastic one, $P_\Delta\propto\Delta^{1.5}$.
\subsection{The Dynamics \label{dynamics}}
Dividing $u_{ij}$ into its trace $\Delta\equiv-u_{\ell\ell}$ and traceless part $u_{ij}^*$, and specifying the matrices $\alpha_{ijk\ell},\lambda_{ijk\ell}$ with two elements each, $\alpha,\alpha_1,\lambda,\lambda_1$, the equation of motion~(\ref{2c-6a}) is written as
\begin{align}
\label{2c-7}
\partial_t\Delta+(1-\alpha )v_{\ell\ell} -\alpha_1u^*_{ij}v^*_{ij}
=-\lambda_1T_g\Delta,
\\\label{2c-8}
\partial_tu^*_{ij}-(1-\alpha )v^*_{ij}
= -\lambda T_gu^*_{ij},
\\\label{2c-9}
\partial_tu_s-(1-\alpha )v_s= -\lambda T_gu_s.
\end{align}
The third equation is valid only if strain and rate are colinear, $u^*_{ij}/|u_s|=v^*_{ij}/|v_s|$. This is frequently the case for a steady rate, because any
non co-linear component of $u_{ij}$ relaxes to zero quickly.
The coefficient $\alpha$ describes softening (if $0<\alpha <1$), or more precisely a reduced gear ratio: The same shear rate yields a smaller deformation, $\partial_tu_{ij}=(1-\alpha)v_{ij}+\cdots$, but
acts also at a smaller stress, $\sigma_{ij}=(1-\alpha)\pi_{ij}\cdots$, see Eqs.(\ref{2c-2},\ref{2c-2a}). $\alpha_1$ accounts for the fact that shearing granular media will change the compression $\Delta$, implying {\em dilatancy} and {\em contractancy}. (More Onsager coefficients are permitted by symmetry, but excluded here to keep the equations simple.)
The Cauchy or total stress is now
\begin{align} \label{2c-2}
P\equiv\sigma_{\ell\ell}/3=(1-\alpha )P_\Delta+P_T-\zeta_gv_{\ell\ell},
\\\label{2c-2a}
\sigma^*_{ij}=(1-\alpha)\pi_{ij}^*-\alpha_1u^*_{ij}P_\Delta -\eta_gv^*_{ij},
\\ \label{2c-2b}
\sigma_{s}=(1-\alpha )\pi_s+\alpha_1u_sP_\Delta
+\eta_gv_s.
\end{align}
Again, the third equation (with $\sigma_s^2\equiv\sigma_{ij}^*\sigma_{ij}^*$) is
valid only if $\pi_{ij}^*$, $u_{ij}^*$ and $v^*_{ij}$ are colinear,
$\pi_{ij}^*/|\pi_s|=-u_{ij}^*/|u_s|=-v^*_{ij}/|v_s|$, often the case in steady state.
The pressure $P$ and shear stress $\sigma_s$ contain elastic contributions
$\propto\pi_s,P_\Delta$ from Eq~(\ref{2b-2b}), and seismic (ie. $T_g$-dependent) ones: $P_T\propto T_g^2$ from Eq~(\ref{2b-5}), and the viscous stress $\propto\eta_g,\zeta_g$. The coefficients $\alpha ,\alpha_1$ soften and mix the stress components. The term preceded by $\alpha_1$ is smaller by an order in the elastic strain, and may be neglected, as we shall do in this paper, if $\alpha_1$ is not too large.
The transport coefficients $\alpha,\alpha_1,\eta_g,\zeta_g$ are functions of the state variables, $u_{ij}$, $T_g$ and $\rho$. As explained above,
they are to be obtained from experiments, in a trial-and-error iteration. And the specification below is what we at present believe to be the appropriate ones. Generally speaking, we find strain dependence weak -- plausibly so because the elastic strain is a small
quantity. One expand in it, keeping only the constant terms. We also expand in $T_g$, but mostly eliminate the constant terms, as we take granular media to be fully elastic for $T_g\to0$, so the force balance $\nabla_j\sigma_{ij}=\rho{\rm g}_i$ reduces to its elastic form, $\nabla_j\pi_{ij}=\rho{\rm g}_i$. This implies
$\alpha,\alpha_1,\eta_g,\zeta_g,\kappa_g\to0$ for $T_g\to0$. In addition, we
take $\alpha,\alpha_1$ to saturate at an elevated $T_g$, such that rate-independence is established. Hence
\begin{align}\label{2c-3}
\eta_g=\eta_1T_g,\,\, \zeta_g=\zeta_1T_g,\,\,
\kappa=\kappa_1T_g,
\\\nonumber
\alpha/\bar\alpha =
\alpha_1/\bar\alpha_1={T_g}/({T_\alpha+T_g}),
\end{align}
with $\bar\alpha,\bar\alpha_1,\eta_1,\zeta_1,\kappa_1,T_\alpha$ functions of $\rho$ only. Expanding $\gamma$ in $T_g$ yields $\gamma=\gamma_0+\gamma_1 T_g$.
We keep $\gamma_0$, because the reason leading to Eqs~(\ref{2c-3}) does not apply, and because $\gamma_0\not=0$ ensures a smooth transition from the hypoplastic to the quasi-elastic regime, see Eq~(\ref{TgVs2}) below. For lack of better information, we take $T_\alpha$ and $\gamma_0/\gamma_1$ to be of the same magnitude.
Since granular media are elastic at $\rho_{cp}$, we have $\bar\alpha, \bar\alpha_1, \lambda, \lambda_1 \to 0$ for $\rho\to\rho_{cp}$,
such that Eqs.(\ref{2c-7},\ref{2c-8}) assume the elastic form, while $\gamma_1$, the relaxation rate for $T_g$, and $\eta_1$, the viscosity, diverge. Accordingly, we take (with $a_1,a_2,a_3,a_4,a_5>0$):
\begin{align}\label{DD}
r\equiv1-\rho/\rho_{cp},\quad \bar\alpha=\bar\alpha_0 r^{a_1},\quad \bar\alpha_1=\bar\alpha_{10}r^{a_2},\qquad \\ \lambda/\lambda_0=\lambda_1/\lambda_{10}=r^{a_3},\quad
\eta_1=\eta_{10} r^{-a_4},\quad \gamma_1=\gamma_{10} r^{-a_5}. \nonumber
\end{align}
(Close to $\rho_{cp}$, the dependence on $\rho_{cp}-\rho$ is the sensitive one, and we ignore any weaker ones on $\rho$ directly.)
We stand behind the temperature dependence with much more confidence than that of the density, for two reasons: First, $\rho$ is not a small quantity that one may expand in, and we lack the general arguments employed to extract the $T_g-$dependence. Second, not coincidentally, the $\rho$ dependence does not appear universal: $a_4=a_5=1$ seems to fit glass beads data, while $a_4=0.5$, $a_5=1.5$ appear more suitable for polystyrene beads~\cite{denseFlow}. For the rest of the paper, when discussing the density dependence qualitatively, we shall use what we call {\it the exemplary values}:
$a_1=a_2=a_3=a_4=a_5=1$.
At given shear rates $v_s$, the stationary state of Eq~(\ref{2c-4}) -- with viscous heating balancing $T_g$-relaxation and $\partial_ts_g=0$ -- is quickly
arrived at ($\lesssim10^{-3}$ s), implying
\begin{align}\label{2c-6}
{\gamma_1} \,h^2\, T_g^2=v_s^2\,{\eta_1}+v^2_{\ell\ell}\,{\zeta_1},
\\
\nonumber\text{where}\,\,\,
h^2\equiv1+\gamma_0/(\gamma_1T_g).
\end{align}
If the density is either constant or changing slowly, implying $v^2_{\ell\ell}\approx0$, we have a quadratic regime for small $T_g$ and low $v_s$, and a linear one at elevated $T_g,v_s$:
\begin{align}
\label{TgVs}
T_g=|v_s|\sqrt{\eta_1/\gamma_1}\quad\,\,\text{for}\quad \gamma_1T_g\gg\gamma_0,
\\\label{TgVs2}
T_g=v_s^2\,\,({\eta_1/\gamma_0})\quad\text{for}\quad \gamma_1T_g\ll\gamma_0.
\end{align}
As discussed in detail in the next section, the linear regime is the {\it hypoplastic} one, in which the system displays elasto-plastic behavior and the hypoplastic model holds. In the quadratic regime, because $T_g\propto v_s^2\approx0$ is quadratically small and negligible, the behavior is {\it quasi-elastic}, with slow, consecutive visit of static stress distributions. Note $h=1$ in the hypoplastic regime, $h\to\infty$ in the quasi-elastic one.
We revisit Eq.(\ref{2c-4}), implementing the following simplifications: (1)~$\nabla_iT_g$ is assumed to be small and linearized in; so terms such as $(\nabla_iT_g)^2$ are eliminated. (2)~$T_g$'s convective term is taken to be negligible, as is $v_{\ell\ell}\approx0$, because density change is typically both small and slow.
(3)~An extra source term $\gamma_1 h^2T_a^2$ is added to account for an ambient temperature $T_a$, which are external perturbations such as given by tapping or a sound field. Eq.(\ref{2c-4}) then reads
\begin{equation}
\label{Tg1}
b\rho\partial_tT_g-\kappa_1T_g\nabla^2T_g=
\eta_1 v_s^2-\gamma_1 h^2(T_g^2-T_a^2).
\end{equation}
Generally speaking, any source contributing to $T_g$ is already included. For instance, given a sound field and its compressional rate $v_{\ell\ell}^{s}$, there is the term on the right hand side
of Eq~(\ref{2c-4}), $\zeta_1 (v_{\ell\ell}^{s})^2$. Coarse-graining it, we may set
$\langle\zeta_1 (v_{\ell\ell}^{s})^2\rangle \equiv\gamma_1h^2T_a^2$. So adding such a term is simply a convenient way to account for a non-specific source.
Finally, we rewrite Eqs.(\ref{2c-7},\ref{2c-8},\ref{2c-9},\ref{Tg1}) as coupled relaxation equations, dimensionally streamlined with 3 time and 1 length scales,
\begin{align}\label{Tg2}
\partial_tT_g=-R_T[T_g(1-\xi_T^2\nabla^2)T_g-T_c^2-T_a^2],
\\\label{Tcf}T_c\equiv f|v_s|,\, f^2\equiv\frac1{h^2}{\frac{\eta_1}{\gamma_1}},\,
R_T\equiv\frac{\gamma_1h^2}{b\rho},\,
\xi_T^2\equiv\frac{\kappa_1}{\gamma_1h^2};
\\\label{eqD}
\partial_t\Delta+(1-\alpha)v_{\ell\ell}=-\lambda_1T_g[\Delta-(T_c|u_s|/T_gu_c)\Delta_c],
\\\label{eqU}
\partial_t u_{ij}^*=-\lambda T_g[u_{ij}^*-(T_c/T_g)u_{ij}^*|_c\,],
\\\label{eqUs}
\partial_t u_s=-\lambda T_g[u_s-(T_c/T_g)u_c],
\\\label{critV}
u_c\equiv\frac{1-\alpha}{\lambda f},\,\,
\frac{u_{ij}^*|_c} {u_c}\equiv\frac{v_{ij}^*}{|v_s|},\,\,
\frac{\Delta_c}{u_c}\equiv \frac{\alpha_1}{\lambda_1f}\frac{u_{ij}^*}{|u_s|}\frac{v_{ij}^*}{|v_s|}.
\end{align}
For constant shear rate and $T_a,v_{\ell\ell}=0$,
we have $T_g=T_c$, $\Delta=\Delta_c$, $u_s=u_c$, $u_{ij}^*=u_{ij}^*|_c$, with $\Delta_c, u_s,u_{ij}^*|_c$ rate-independent.
It is customary in soil mechanics to refer to this steady state as {\it critical}, though it is unrelated to critical phenomena in physics.
The relaxation rate $R_T$ in dense media has an inverse time scale of order ms or less. In comparison, the rates $\lambda T_g,\lambda_1 T_g\propto v_s$ are small for the shear rates typical of soil-mechanical experiments, $\lambda T_g=1$/s for $v_s=10^{-2}$/s. The length scale $\xi_T$ is a few granular diameters. Rate-independence derives from $T_g\propto T_c\equiv f|v_s|$, and is destroyed by any $T_a\ne0$.
[We note that $u_c,T_c>0$, but $u_s,v_s$ may be negative. Eq.(\ref{eqUs}) is obtained by multiplying Eq.(\ref{eqU}) with $u_{ij}^*/|u_s|$ and assuming $u_s>0$, $u_{ij}^*/|u_s|=v_{ij}^*/|v_s|=$ const, which is eg. not right in the load/unload experiment, as $u_{ij}^*/|u_s|=-v_{ij}^*/|v_s|$ right after a rate reversal, see Sec.~\ref{Load and Unload}.]
With the differential equations derived, the energy density and transport coefficients in large part specified, {\sc gsh} is a well-defined theory. It contains clear ramifications and provides little
leeway for retrospective adaptation to observations.
As we shall see in the following sections, a wide range of granular phenomena is encoded in these equations.
\subsection{Three Rate Regimes\label{sum}}
Depending on the interaction between particles, granular experiments are divided into three regimes: In the first, the particles are static and elastically deformed; in the second, they move slowly, rearranging by overcoming frictional forces; in the third, they interact by collisions. Although this interaction, of mesoscopic nature, is not manifest in a macroscopic theory, {\sc gsh} does have three regimes echoing its variation, and the control parameter is how strongly the grains jiggle -- quantified as the granular temperature $T_g$:
\begin{itemize}
\item At vanishing shear rates, grains do not jiggle, $T_g\to0$. The stress stems from deformed grains and is elastic in origin. Static stress distribution and the incremental stress-strain relation are phenomena of this regime.
Deviations from full elasticity, $u_{ij}=\varepsilon_{ij}$ and $\sigma_{ij}=\pi_{ij}$, being quadratically small, $\alpha,\alpha_1,\eta_g,\zeta_g,\kappa_g\propto T_g\propto v_s^2$, are frequently negligible. This is what we call the {\em quasi-elastic regime}.
\item At slow rates, $T_g\gg\gamma_0/\gamma_1$ is somewhat elevated, see Eq.(\ref{TgVs}). The elastic
stress may now relax, implying plasticity: When the grains jiggle and briefly loosen contact with one another, the grains' deformation and the associated stress will get partially lost, irreversibly. We call this regime {\em hypoplastic}, because this is where the hypoplastic model~\cite{kolymbas1} and other rate-independent constitutive relations are valid. Typical phenomena are the {critical state}~\cite{schofield}, and the different loading/unloading curves. Friction is a result in {\sc gsh}, not an input, and it derives from the combined effect of elastic deformation and stress relaxation. (In spite of our borrowed usage of {\it hypoplasticity}, the reversible part of the stress is derived from an energy potential.)
In the hypoplastic regime, we have $T_g=T_c\equiv f|v_s|$, $\alpha=\bar\alpha$, $\alpha_1=\bar\alpha_1$. The equations~(\ref{eqD},\ref{eqU}) for the elastic strain are explicitly rate-independent, and the stress, generally given by Eqs.(\ref{2c-2},\ref{2c-2a},\ref{2c-2b}), is simplified, because the kinetic pressure $P_T\propto T_g^2$ and the viscous stress $\eta_1T_g v_s$, both quadratic in the rate, are negligibly small. The stress is $\sigma_{ij}=(1-\alpha)\pi_{ij}$, where the factoris typically between 0.2 and 0.3. The complex
elasto-plastic motions, observed mainly in triaxial apparatus, take place in this regime.
\item At high shear rates, large $T_g$ and low densities, we are in the regime of {\it rapid dense flow}. The jiggling is so strong that it gives rise to a kinetic pressure and viscous shear stress. They compete with the elastic one as rendered by the $\mu$-rheology~\cite{midi}.
We still have $T_g\propto v_s$ at higher rates, but it is no longer small. Therefore, the kinetic pressure $P_T$ and the viscous stress become significant and compete with the elastic contribution. Both the total pressure and the shear stress may now be written as $e_1+e_2v_s^2$, with $e_1,e_2$ functions of the density.
The {\it Bagnold regime} is given for $e_2v_s^2\gg e_1$,
where all stress components depend quadratically on the rate. Typically, since $e_1\gg e_2v_s^2$ for any realistic $v_s$, it is not easy to go continuously from the rate-indepedent to the Bagnold regime at given density. However, a discontinuous transition is possible at given pressure, because $\rho$ decreases with $v_s$, eventually going below $\rho_{\ell p}$. There is then no elastic solution, $\pi_{ij}\equiv0$, or $e_1=0$. And the system is in a pure Bagnold regime.
\end{itemize}
For reasons discussed in detail in~\cite{granR4}, it is difficult to observe the transition from the hypoplastic regime to the quasi-elastic regime. And it has in fact not yet been done systematically. This is probably why soil mechanics textbooks take the hypoplastic regime to be the lowest rate one, referring to it as quasi-static. This is, we believe, conceptually inappropriate, because motions in the hypoplastic regime are irreversible and strongly dissipative, not consecutive visits of neighboring static states with vanishing dissipation. Therefore, experiments at the very low end of shear rates are highly desirable. (When pressed, we need to guess. And we expect
the {\em quasi-elastic regime} to start somewhere below $10^{-5}$/s, with the rate-independent {\em hypoplastic regime} above $10^{-3}$/s.)
Next, we employ the equations presented above to account for granular phenomena, first in the hypoplastic regime, in which the complexity of granular behavior is most developed and best documented. Then we consider dense flow, including the $\mu$-rheology and the Bagnold scaling. This is followed by the nonuniform phenomena of elastic waves, shear band and compaction. Finally, the quasi-elastic regime of vanishing rates is considered, exploring why it is hard to observe, and how best to overcome the difficulties.
\section{The Hypoplastic Regime\label{hypoplastic motion}}
Granular behavior in the hypoplastic regime are taken to generally possess rate-indepen\-dence -- meaning for given strain rates, the increase in the stress $\Delta\sigma_{ij}$ depends only on the increase in the strain, $\Delta\varepsilon_{ij}=\int v_{ij}{\rm d}t$, not the rate. As a result, engineering theories typically have rate-independence built in from the beginning. We note that it is not at all a robust feature of granular behavior. For instance, it is lost when the system is subject to an ambient temperature $T_a$ [such as given by a sound field, see the discussion around Eq.(\ref{Tg1})]: The critical stress then becomes strongly rate-dependent, vanishing for large $T_a$. And it does not extend into the higher rates of dense flow. Therefore, rate-independence is a phenomenon that cries out for an explanation, an understanding.
Moreover, it is crucial to distinguish between rate- and stress-controlled experiments. When the rate is given, $T_g$ quickly settles into its steady state value $T_c$, see Eq.(\ref{Tg2}), then the relaxation of $u_s$, accounting for the approach to the critical state, is independent of the rate, see Eq.(\ref{eqUs}). A rather different experiment is to hold the shear stress $\sigma_s$ fixed, starting with an elevated $T_g$. This $T_g$ will relax until it is zero, and the system static. There is also a rate in this case, referred to as {\it creep} sometimes -- the one that compensates the stress relaxation at a finite $T_g$. being proportional to $T_g$, this rate relaxes toward zero at the same time. Rate-independence is therefore a misplaced concept here.
Stress-controlled experiments cannot be performed in triaxial apparatus with stiff steel walls, because the correcting rates employed by the feedback loop to keep the stress constant are of
hypoplastic magnitudes. As a result, much $T_g$ is excited that distorts its relaxation, and the situation is one of consecutive constant rates, not of constant stress. Instead, one may employ a soft spring to couple the granular system with its driving device, to enable small-amplitude stress corrections without exciting much $T_g$. We consider rate-controlled experiments in Sec.\ref{critical state}, \ref{Load and Unload}, and \ref{Constitutive Relations}, stress-controlled ones in Sec.\ref{aging}, and experiments subject to an ambient temperatures $T_a$ in Sec.\ref{visElaBeh}.
\subsection{The Critical State\label{critical state}}
Grains with enduring
contacts are deformed, which gives rise to
an elastic stress. The deformation is
slowly lost when grains rattle and jiggle, because
they lose or loosen contact with one another. As a
consequence, a constant shear rate not only increases
the deformation, as in any elastic medium, but also
decreases it, because grains jiggle when being sheared
past one another. A steady state exists in which both
processes balance, such that the deformation remains
constant over time -- as does the stress. This is the critical state.
Moreover, the increase in deformation is $\propto v_s$, the relaxation is $\propto T_g$. As $T_g\propto v_s$ for elevated granular temperature, the steady-state, especially the critical stress, are rate-independent. In this section, we show how {\sc gsh} mathematically codify this physics.
\subsubsection{Stationary Elastic Solutions\label{Stationary Elastic Solution}}
The critical state is given by the stationary solution $T_g=T_c, \Delta=\Delta_c, u_s=u_c$, with
\begin{equation}\label{3b-3a} u_c=\frac{1-\alpha}{\lambda}\frac{v_s}{T_g}=\frac{1-\alpha}{\lambda f},
\quad \frac{\Delta_c}{u_s}=\frac{\alpha_1}{\lambda_1}\frac{v_s}{T_g}=\frac{\alpha_1}{\lambda_1f},
\end{equation}
see Eqs.(\ref{Tg2},\ref{eqD},\ref{eqU}). Because further shearing does not lead to any stress increase, this state is frequently referred to as {\it ideally plastic}~\cite{critState}. Note $u_c,\Delta_c$ are rate-independent (for $\alpha=\bar\alpha$, $\alpha=\bar\alpha$, $T_a=0$) and functions of the density. Same holds for the critical stress, cf Eqs.(\ref{2b-2a},\ref{2b-2b},\ref{2b-1}),
\begin{align}
\label{3b-4a}
P_c=(1-\bar\alpha)P^c_\Delta,\quad
\sigma_c=(1-\bar\alpha)\pi_c,
\\\label{3b-3c}
P^c_\Delta\equiv P_\Delta(\Delta_c,
u_c),\,\,
\pi_c\equiv\pi_s(\Delta_c, u_c), \\\label{3b-3cA}
{P^c_\Delta}/{\pi_c}=({{\cal B}}/{2{\cal
A}}){\Delta_c}/{u_c}+{u_c}/{4\Delta_c}.
\end{align}
The loci of the critical states thus calculated~\cite{GSH&Barodesy} (though employing the slightly more general energy of~\cite{3inv}) greatly resembles those calculated using either hypoplasticity or barodesy~\cite{barodesy}
The critical ratio $\sigma_c/P_c$ -- same as the Coulomb yield of Eq~(\ref{2b-3}) -- is also frequently associated with a friction angle. Since one is relevant for vanishing $T_g$, while the other requires an elevated $T_g\propto |v_s|$, it is appropriate to
identify one as the static friction angle, and the other as the
dynamic one. The latter is smaller than
the former, because the critical state is elastic,
and must stay below Coulomb yield,
$\lambda_1f/\bar\alpha_1<\sqrt{2{\cal B/A}}$, if it is viable.
Textbooks on soil mechanics state that the friction angle is
independent of the density -- although they do not, as a rule,
distinguish between the dynamic and the static one. We assume, for lack of better information, that both are, or $2(a_3-a_2)=a_5-a_4$, see Eq~(\ref{DD}).
Separately, both $\Delta_c$ and $u_c$ should increase with $\rho\to\rho_{cp}$, same holds for $P_c$ and $\sigma_c$.
\subsubsection{Approach to the Critical State at Constant Density\label{approach critical}}
Solving Eqs~(\ref{3b-2},\ref{3b-3}) for $u_s,\Delta$, at constant
$\rho, v_s$, with $h=\alpha/\bar\alpha=\alpha_1/\bar\alpha_1=1$, and the initial conditions: $\Delta=\Delta_0, u_s=0$, the relaxation into the critical state is given as
\begin{align}\label{3b-6}
u_s(t)=u_c(1-e^{-\lambda f\varepsilon_s}),\quad \varepsilon_s\equiv v_st,
\\\nonumber \Delta(t)=\Delta_c(1+f_1\,e^{-\lambda f
\varepsilon_s}+f_2e^{-\lambda_1f\varepsilon_s}), \\\nonumber
f_1\equiv\frac{\lambda_1}{\lambda-\lambda_1},\quad
f_2\equiv\frac{\Delta_0}{\Delta_c}-\frac{\lambda}{\lambda-\lambda_1}.\end{align}
Clearly, this is an exponential decay for $u_s$, and a sum of two decays for $\Delta$. It is useful, and quite demystifying, that a simple, analytical solution in terms of the elastic strain exists. Because $\lambda\approx3.3\lambda_1$, the decay of $u_s$ and $f_1$ are
faster than that of $f_2$. Note $f_2$ may be
negative, and $\Delta(t)$ is then not monotonic. The associated pressure
and shear stress are those of Eqs~(\ref{3b-4a},\ref{3b-3c},\ref{3b-3cA}). For a negative $f_2$, neither the pressure nor the shear stress is monotonic.
For the system to complete the approach to the critical state, the yield surface [such as given by Eq.(\ref{2b-3})] must not be breached during the non-monotonic course. If it happens, there is an instability, and the most probable result are shear bands, see Sec~\ref{nsb}, \ref{sb} below. Then the uniform critical state will not be reached.
\subsubsection{Approach to the Critical State at Constant Pressure\label{pressure approach}}
\begin{figure}[t]
\includegraphics[scale=.8]{fig3
\caption{Three
approaches to the critical state: These are the results of
{\sc gsh} calculations employing the parameter sets I,II,III as specified in the text. Shear stress $q\equiv(\sigma_3-\sigma_1)/\sigma_1$ and void ratio
$e\equiv\rho_g/\rho-1$ (with $\rho_g$ the grain's density)
versus the strain $\varepsilon_3$ in triaxial tests
(cylinder axis along 3), at given $\sigma_1$ and
strain rate $\varepsilon_3/t$, for an initially dense
and loose sample. }\label{fig3}
\end{figure}
Frequently, the critical state is not approached at constant density, but at constant pressure $P$ (or a stress eigenvalue $\sigma_i$). The circumstances are then more complicated. As $\Delta,u_s$ approach
$\Delta_c,u_c$, the density compensates to keep
$P(\rho,\Delta,u_s)=$ const. Along with $\rho$, the coefficients
$\alpha,\alpha_1,\lambda,\lambda_1,f$ (all functions of $\rho$), also change
with time. In addition, with $\rho$ changing, the compressional flow
$v_{\ell\ell}=-\partial_t\rho/\rho$ no longer vanishes (though it is still small). Analytic solutions
do not seem feasible now, but numerical ones are, see Fig~\ref{fig3}, which compares three sets of parameters by plotting the deviatory stress versus axial strain at given $\sigma_1$. Clearly, any could serve as a textbook illustration of the approach to the critical state. The parameters, see Eqs.(\ref{DD}), labeled as I, II, III, are:
\begin{itemize}
\item ${\cal B}_0=2, 0.22, 0.05$ GPa,\quad ${\cal B/A}=5/3, 8, 5/3$,
\quad $\bar\rho/\rho_{cp}=0.615, 0.650, 0.667$,
\item $\bar\alpha_0=1.04, 0.85, 16.25$,\quad $\bar\alpha_{10}=400, 30, 719$,
\item $\lambda_0\sqrt{\eta_{10}/\gamma_{10}}=272, 250, 2375$,\quad $\lambda/\lambda_1=3.8, 3.8, 3$,
\item $a_1=0.15, 0.15, 1.6$,\,\, $a_2=1,0. 15, 1.6$,\,\, $a_3=0.6, 0.53, 1.6$,
\,\, $a_4=a_5=0,0,-1$.
\end{itemize}
Fig \ref{fig3a} compares I to the (drained monotonic triaxial) experiment by Wichtmann~\cite{wichtmann}, II to the simulation by Thornton and Antony~\cite{thornton}, both in the plots as originally given. The comparison of III to the barodesy model~\cite{barodesy} may be found in~\cite{GSH&Barodesy}.
\begin{figure}[]
\includegraphics[scale=.8]{fig3a
\caption{A {\sc gsh} calculation employing I for comparing to the Wichtmann's experiment, and II to the simulation by Thornton and Antony, in the plots as originally given in~\cite{wichtmann,thornton}.} \label{fig3a}
\end{figure}
Generally speaking, we have three scalar state variables: $\rho,u_s,\Delta$, each with an equation of motion that depends on the rates $v_s,v_{\ell\ell}$ and the variables themselves. In addition,
$P,\sigma_s$ are functions of $\rho,u_s,\Delta$. In the last section, both rates were given,
$v_{\ell\ell}=0$, $v_s=$ const. As a result, we have $\rho=$ const, while $\Delta(t)$ and $u_s(t)$ were calculated taking the coefficients $\alpha(\rho),\alpha_1(\rho),\lambda(\rho),
\lambda_1(\rho),f(\rho)$ as constant. The stress components were then obtained as dependent
functions. A pressure-controlled experiment means that only the shear rate $v_s$ is given. Holding $P(\rho,u_s,\Delta)=$ const (or analogously $\sigma_1$) implies
the density $\rho$ (and with it also $v_{\ell\ell}=-\partial_t\rho/\rho$) is a dependent function, $\rho=\rho(P,u_s,\Delta)$. Now, in the equations of motion for $u_s$ and $\Delta$, one first eliminates
$v_{\ell\ell}$ employing $v_{\ell\ell}=-\partial_t\rho/\rho$, then eliminates both $\partial_t\rho/\rho$ and the $\rho$-dependence
of $\alpha(\rho),\alpha_1(\rho),\lambda(\rho),\lambda_1(\rho),f(\rho)$ employing
$\rho=\rho(P,u_s,\Delta)$. This changes the differential equations -- which are then solved numerically.
Many well-known features of Fig~\ref{fig3} can be understood assuming the solutions of Eq~(\ref{3b-6}) remain valid, say because the initial density is close to
the critical one, hence it does not change much during the approach to the critical state. As a
result, we may approximate $\alpha(\rho),\alpha_1(\rho),\lambda(\rho),\lambda_1(\rho),f(\rho)$ as constant, and take $v_{\ell\ell}\approx0$. In addition, we assume, for simplicity, $\lambda\gg\lambda_1$, or ${\lambda}/({\lambda-\lambda_1})\approx1$ (instead of $\,\approx1.5$). Then $f_2$ has
the same sign as $\Delta_0-\Delta_c$. The initial values are $\rho_0,\Delta_0$ and $u_s=0$, implying $P\propto{\cal B}(\rho_0)\Delta_0^{1.5}, \sigma_s=0$. For $P$ given and ${\cal B}(\rho)$ monotonically increasing with $\rho$, the pair
$\Delta_0-\Delta_c$ and $\rho_0-\rho_c$ have reversed signs. Therefore, we have a monotonic change of density for $f_2>0$, $\Delta_0>\Delta_c$, $\rho_0<\rho_c$, and non-monotonic change otherwise. At
the beginning, the faster relaxation of $f_1$ dominates, so $\Delta$ always
decreases, and $\rho$ always increases, irrespective of $\rho_0$. After
$f_1$ has run its course, $\rho$ goes on increasing for $\rho<\rho_0$ ({\em contractancy}) but switches to decreasing for $\rho>\rho_0$ ({\em dilatancy}), until the critical state is reached.
The shear stress $\sigma_s\propto \sigma_1-\sigma_2$ always increases first
with $u_s$, until $u_s$ is close to $u_c$. The subsequent behavior depends on
what $\Delta$ does. With $P\propto{\cal B}(\rho_0)\Delta_0^{1.5}$ given, $\sigma_s\propto{\cal B}\Delta^{0.5}\propto P/\Delta$ keeps growing if $\Delta$ decreases [loose
case, $f_2>0$], but becomes smaller again, displaying a peak, if $\Delta$ grows [dense
case, $f_2<0$].
\subsubsection{Shear Jamming}
A {\it jammed state} is one that can stably sustain a finite stress, especially an anisotropic one. It is therefore characterized by values for $\Delta, u_s$ that satisfy the stability conditions $u_s/\Delta\le\sqrt{2{\cal B}/{\cal A}}$, or Eq.(\ref{2b-3}).
An unjammed state violates either this or another stability conditions, such as $\phi_{\ell p}<\phi<\phi_{cp}$ (where $\phi\equiv\rho/\rho_g$, with $\rho_g$ the bulk density, is the packing fraction). Typically, the critical state is approached starting from an isotropic stress, $\Delta=\Delta_0, u_s=0$. But the approach solution Eq.(\ref{3b-6}) is also valid if the initial elastic shear strain is finite, $u_s\ne0$. Writing the solution to first order in the shear strain $\varepsilon_s\equiv v_st$,
\begin{equation}
u_s(t)=u_c\lambda f\varepsilon_s,\quad \Delta(t)=\Delta_0(1-\lambda_1f\varepsilon_s),
\end{equation}
we see a growing $u_s$ and a decreasing $\Delta$ for the initial stage. This is the reason that, if $\Delta_0$ is sufficiently small, the system will become unstable first, before it re-enters the stable region, converging eventually onto the critical state.
Shear-jamming at constant density, as observed in~\cite{SJ1} and simulated in~\cite{SJ2}, is exactly this process, starting from the initial value $\Delta, u_s=0$, or equivalently, from vanishing elastic pressure and shear stress, $P_\Delta,\pi_s=0$. So the system is unstable at the beginning, until $\Delta$ is sufficiently large to satisfy Eq.(\ref{2b-3}), and the system is securely jammed. Further steady shearing then pushes the system into the critical state.
\subsubsection{The Critical State with External Perturbations \label{external perturbation}}
If one perturbs the system, say by exposing it to weak vibrations, or by tapping it periodically, such as in a recent experiment~\cite{vHecke2011}, the critical state is modified, and a rate-dependence of the critical shear stress is observed. The stress decreases with the shaking amplitude, and increases with the shear rate, such that the decrease is compensated at
higher rates. Clearly, engineering theories with built-in rate-independence cannot account for this observation. {\sc gsh}, on the other hand, if it indeed
provides a wide-range description of granular behavior, should be able to.
The consideration of the critical state in the previous three sections takes any granular temperature $T_g$ to be a
result of the given shear rate, hence $T_g=T_c\equiv|v_s|f$. This is no longer the case here, as sound field or tapping also contributes to $T_g$. And we have~Eq.(\ref{Tg2}),
\begin{equation}\label{3b-7}
T_g^2=T_c^2+T_a^2,
\end{equation}
This is the reason the steady state values are reduced to $\bar u_c\equiv(T_c/T_g)u_c$, $\bar \Delta_c\equiv(T_c/T_g)^2\Delta_c$, see Eqs.(\ref{eqD},\ref{eqU},\ref{eqUs}), with
\begin{equation}\label{3b-8}\frac{\bar u_c^2}{u_c^2}=
{\frac{\bar \Delta_c}{\Delta_c}}={\frac{\bar \sigma_c}{\sigma_c}}=\frac1{1+{ T_a^2}/{T_c^2}}.
\end{equation}
If there is no tapping, $T_a=0$, we retrieve the unperturbed values, $\bar u_c=u_c$, $\bar\Delta_c=\Delta_c$, $\bar\sigma_c=\sigma_c$. With tapping, $\bar u_c, \bar\Delta_c, \bar\sigma_c$ decrease for increasing $T_a$, and increase with increasing shear rate $T_c\equiv f|v_s|$. see Fig~\ref{fig3b}.
\begin{figure}[t]
\includegraphics[scale=.8]{fig3b}
\caption{Suppression of the critical shear stress $\sigma^c_s$ by vibration as given by Eq.(\ref{3b-8}), assuming $\Gamma=\alpha T_a,\,\, \Omega=\beta v_s^3$ (see text for details). Inset is the experimental curve of~\cite{vHecke2011}, with the torque $\tau$ denoted as $T$, as in~\cite{vHecke2011}. (The stress dip at large $\Omega$, neglected here, is explained in~\cite{StressDip}.) }\label{fig3b}
\end{figure}
(Note we have only considered the critical state at given shear rate, not the approach to it. So the result holds both at given density and pressure.)
The above consideration is the basic physics of the observation reported in~\cite{vHecke2011}. It helps to put rate-independence, frequently deemed a fundamental property of granular media, into the proper context.
A more detailed comparison is unfortunately made difficult by the highly nonuniform experimental geometry.
Nevertheless, some comparison, even if unabashedly qualitative, may still be useful.
In~\cite{vHecke2011}, the torque $\tau$ on the disk on top of a split-bottom shear cell is related to its rotation velocity $\Omega$ and the shaking acceleration $\Gamma$. Now, $\tau$ and $\sigma_c$, $\Omega$ and $v_s$, $\Gamma$ and $T_a$, are clearly related pairs, see also Sec~\ref{compaction}.
Assuming the lowest order terms suffice in an expansion, we take $\sigma_c\propto\tau$ and $\Gamma=c_1T_a$ with $c_1\sqrt{\eta_1/\gamma_1}=$20~s (noting $T_g$ is dimensionless with an appropriate $b$). If $v_s$ were uniform, $\Omega\propto v_s$ would also hold. Since it is not, $\Omega\propto v_s^n$ with $n>1$ seems plausible, because with additional degrees of freedom such as position and width of the shear band, the system has for given $\Omega$ more possibilities to decrease its strain rate $v_s$. We take $\Omega=c_2v_s^3$ with $c_2=1{\rm rs}^2$ [implying a replacement of $T_a/v_s$ with $\Gamma/\sqrt[3]\Omega$ in Eq.(\ref{3b-8})] for the fit of Fig \ref{fig3}, but emphasize that qualitative agreement exists irrespective of $n$'s value.
In~\cite{vHecke2011}, a stress dip was in addition observed at higher rates, see Fig \ref{fig3}. This is also accounted for by {\sc gsh}, see Sec~\ref{udf} and~\cite{StressDip}.
\subsection{Load and Unload\label{Load and Unload}}
The simple reason for the difference between load and unload is that the stationary values $\Delta_c,u_{ij^*}|_c$ of Eqs.(\ref{eqD},\ref{eqU}) are altered when the shear rate $v_{ij}^*$ is reversed. The relaxation then proceed towards these new values, see the final paragraph of Sec.\ref{dynamics}. It is simple and deterministic and not in anyway {\it history-dependent}. We insert $T_g= f|v_s|$ into Eqs~(\ref{2c-7}, \ref{2c-9}),
\begin{align}\label{3b-2}
\partial_t\Delta=v_s\,\alpha_1u_s-|v_s|\,\lambda_1f\Delta,
\\\label{3b-3}
\partial_tu_s=v_s\,(1-\alpha )-|v_s|\,\lambda fu_s,
\end{align}
to see that loading ($v_s=|v_s|>0$) and unloading ($v_s=- |v_s|<0$) have different slopes: $\partial_tu_s/v_s=(1-\alpha)\mp(\lambda fu_s/h)$.
Referred to as {\em
incremental nonlinearity} in soil mechanics, this phenomenon is the reason why no backtracing takes place under reversal of shear rate: Starting from isotropic stress, $u_s=0$, see Fig~\ref{fig2}, the gradient is at first $(1-\alpha)$, becoming smaller as $u_s$ grows, until it
is zero, in the stationary case $\partial_tu_s/v_s=0$. Unloading now, the
slope is $(1-\alpha)+(\lambda fu_s/h)$, steeper than it has ever been. It is again
$(1-\alpha)$ for $u_s=0$, and vanishes for $u_s$ sufficiently negative. Same scenario holds for $\partial_t\Delta/v_s$.
\begin{figure}[t] \begin{center}
\includegraphics[scale=0.33]{fig2}
\end{center}\vspace{-2cm}
\caption{\label{fig2}The hysteretic change of the shear stress $\propto u_{s}$ with the strain, as given by Eq~(\ref{3b-3}). The sign of $v_{s}(t)$, $\varepsilon _{s}\equiv\int_{0}^{t}v_{s}(t^{\prime })dt^{\prime }$, and $u_{s}(t)$ are given respectively in (b), (c) and (d).}
\end{figure}
The stress components $P,\sigma_s$ are calculated employing Eqs~(\ref{2b-2b},\ref{2c-2},\ref{2c-2b}) for given $\Delta, u_s$. This consideration holds only for given density, it is more complicated if the pressure is given instead, same as in Sec~\ref{critical state}), but the basic physics remains the same.
In systematic studies employing discrete numerical simulation, Roux and coworkers have accumulated great knowledge about granular physics, see eg.~\cite{roux}. They distinguish between two types of strain, I and II, identifying two regimes in which either dominates. This result agrees well with the above consideration, as the relaxation term in Eq.(\ref{3b-3}), being $\propto u_s$ is small if $u_s\propto\sigma_s$ is. Slow relaxation means the system is less plastic, more elastic and the difference between load and unload is small.
\subsection{Constitutive Relations\label{Constitutive Relations}}
Granular dynamics is frequently modeled employing the strategy of
{\em rational mechanics}, by postulating a function $\mathfrak{C}_{ij}$ -- of the stress $\sigma _{ij}$, strain rate $v_{k\ell}$, and density $\rho $ -- such that the constitutive relation, ${\partial}_{t}\sigma _{ij}=\mathfrak{C}_{ij}(\sigma_{ij}, v_{k\ell}, \rho)$ holds. (More generally, ${\partial}_{t}$ is to be replaced by an appropriate objective derivative.) It forms, together with the continuity equation $\partial _{t}\rho +\nabla _{i}\rho v_{i}=0$, momentum conservation, $\partial _{t}(\rho v_{i})+\nabla_{j}(\sigma _{ij}+\rho v_iv_j)=0$, a closed set of equations for $\sigma _{ij}$, the velocity $v_{i}$, and the density $\rho $ (or void ratio $e$). Both hypoplasticity and barodesy considered below belong to this category.
(We do not consider elasto-plastic theories, but do note that, as shown by Einav~\cite{h^2}, they all form a special limit of the hypoplastic ones)
These models yield, in their range of validity, a realistic account of the complex elasto-plastic motion, providing us with highly condensed and intelligently organized empirical data. This enables us to validate {\sc gsh} and reduce the latitude in specifying the energy and transport coefficients.
The drawbacks are, first of all, the apparent freedom in fixing ${\mathfrak C}_{ij}$ -- constrained only by the data one considers, not by energy conservation or entropy production (that were crucial in deriving {\sc gsh}). This is probably the reason why there are many competing engineering models. And this liberty explodes when one includes gradient terms, hence most models refrain from the attempt to account for nonuniform situations, say elastic waves.
Second, dispensing with the the variables $T_g$ and $u_{ij}$, one reduces the model's range of validity. For instance, they hold only for $T_g=T_c\equiv f|v_s|$ and not for a $T_g$ that is either too small or oscillates too fast. Also, as the analytical solution of the approach to the critical state shows, considering $u_{ij}$ is a highly simplifying intermediate step. The case for $u_{ij}$ is even stronger when considering proportional paths and the barodesy model, see below.
\subsubsection{The Hypoplastic Model\label{Hypoplasticity}}
The {\em hypoplastic model} starts from the rate-independent
constitutive relation,
\begin{equation}\label{3b-1}
\partial_t\sigma_{ij}=H_{ijk\ell}v_{k\ell}+
\Lambda_{ij}\sqrt{v_s^2+\epsilon v_{\ell\ell}^2}, \end{equation} postulated by Kolymbas~\cite{kolymbas1}, where
$H_{ijk\ell},\Lambda_{ij},\epsilon$ are functions of the stress and void ratio. The simulated granular response is realistic for deformations at constant or slowly changing rates.
Taking $h=1$, $\alpha =\bar \alpha$, $\alpha_1=\bar\alpha_1$, $P_T, \eta_1T_g
v^0_{ij}\to0$, {\sc gsh} easily reduces to the hypoplastic model. This is because $\sigma_{ij}$ of Eqs~(\ref{2c-2},\ref{2c-2a}) is then, same as $\pi_{ij}$, a function of $u_{ij}, \rho$, and we may write $\partial_t\sigma_{mn}=({\partial\sigma_{mn}}/{\partial u_{ij}})\partial_tu_{ij}+
({\partial\sigma_{mn}}/{\partial\rho})\partial_t\rho$.
Replacing $\partial_t\rho$ with $-\rho v_{\ell\ell}$,
$\partial_tu_{ij}$ with Eq~(\ref{2c-8}), using Eq~(\ref{TgVs}) to eliminate $T_g$, we arrive at an equation with the same structure as Eq~(\ref{3b-1}). Our derived expressions for $H_{ijk\ell},\Lambda_{ij}$ is different from the postulated ones, and somewhat simpler, but they yield very
similar results, especially {\em response ellipses}~\cite{granL3}. (Response ellipses are
the strain increments as the response of the system, given unit stress
increments in all directions starting from an arbitrary point in the stress
space, or vice versa, stress increments as the response for unit strain increments.)
\subsubsection{Proportional Paths and Barodesy}
Barodesy is a recent model, again proposed by Kolymbas~\cite{barodesy}. It is more modular and better organized than hypoplasticity, with different parts
in $\mathfrak C_{ij}$ taking care of specific aspects of granular deformation, especially that of {\em proportional paths}. We take {\sc p}$\varepsilon${\sc p} and {\sc
p}$\sigma${\sc p} to denote, respectively, proportional strain and stress paths. Their behavior is summed up by the Goldscheider rule ({\sc gr}):
\begin{itemize}
\item A {\sc p}$\varepsilon${\sc p} starting from the stress $\sigma_{ij}=0$ is associated with a {\sc p}$\sigma${\sc p}. (The initial value $\sigma_{ij}=0$ is a mathematical idealization, neither easily realized nor part of the empirical data. We take it {\em cum grano salis}.)
\item A {\sc p}$\varepsilon${\sc p} starting from $\sigma_{ij}\not=0$
leads asymptotically to the same {\sc p}$\sigma${\sc p} obtained when starting at $\sigma_{ij}=0$.
\end{itemize}
Any constant strain rate $v_{ij}$ is a {\sc p}$\varepsilon${\sc p}. In the principal strain axes $(\varepsilon_1,\varepsilon_2,\varepsilon_3)$, a constant $v_{ij}$
means the system moves with a constant rate along its direction, with $\varepsilon_1/\varepsilon_2=v_1/v_2,\, \varepsilon_2/\varepsilon_3=v_2/v_3$
independent of time. {\sc gr} states there is an associated stress path that is also a straight line in the principal stress space, that there are
pairs of strain and stress path. And if the initial stress value is not on the right line, it will converge onto it.
Again, if {\sc gsh} is indeed a broad-ranged theory on granular behavior, we should be able to understand {\sc gr} with it, which is indeed the case. But we need to generalize the stationary solution as given by Eq.(\ref{critV}) to include $v_{\ell\ell}\not=0$ (using $^{ni}$ to imply non-isochoric),
\begin{equation}
\label{eq74}
u_c=\frac{1-\alpha}{\lambda f},\quad \frac{\Delta_c^{ni}}{u_c}=\frac{\alpha _1}{\lambda _{1}f}+\frac{1-\alpha}{u_c\lambda_1f}\frac{v_{\ell\ell}}{v_s},
\end{equation}
with ${\sigma_{ij}^{\ast }}/{\sigma_s}={u_{ij}^{\ast }|_c}/{u_c}={v_{ij}^{\ast }}/{v_s}$.
If the strain path is isochoric, $v_{\ell\ell}=0, \rho=$ const, both the deviatoric strain and stress are dots that remain stationary-- these are the critical state considered in
Sec~\ref{critical state}. If however $v_{\ell\ell}\not=0$, with the density $\rho[t]$ changing accordingly, ${u_{ij}^{\ast }|_c}={u_c}(\rho)\,{v_{ij}^{\ast }}/{v_s}$ and ${\sigma_{ij}^{\ast }}= {\sigma_s}(\rho)\,{v_{ij}^{\ast }}/{v_s}$ will walk down a straight line along ${v_{ij}^{\ast }}/{v_s}$, with a velocity determined, respectively, by $u_c(\rho[t])$ and $\sigma_s(\rho[t])$.
Given an initial strain deviating from that prescribed by Eq~(\ref{eq74}), $u_0\not=u_c,\Delta_0\not=\Delta_c^{ni}$, Eqs~(\ref{eqD},\ref{eqU}) clearly state that the deviation will relax, implying the strain and the associated stress will converge onto the prescribed line. This is all very well, but {\sc gr} states that it is the total stress that possesses a {\sc p}$\sigma${\sc p}. With
$\pi_{ij}=P_\Delta(\rho)[\delta_{ij}+(\pi_s/P_\Delta)v_{ij}^*/v_s]$,
this fact clearly hinges on $(\pi_s/P_\Delta)$ -- a function of $\Delta/u_s$, see Eq~(\ref{2b-1}) -- not depending on the density. As long as $v_{\ell\ell}\ll v_s$, we have $\Delta_c^{ni}/u_c\approx{\alpha _1}/{\lambda _{1}f}$, which we did assume in Eq~(\ref{DD}) is density-independent, to render the dynamic friction angle (that of the critical state) independent of $\rho$.
When looking at $\mathfrak C_{ij}$, it is easy to grasp that the construction of a constitutive relation requires vast experience in handling granular media. That we could substitute this deep knowledge with the equations of {\sc gsh} that are just as capable of accounting for elasto-plastic motion, is eye-opening. It suggests that sand, in its qualitative behavior, may be, after all, neither overly complicated, nor such a rebel against general principles.
In~\cite{GSH&Barodesy,P&G2009}, the results of {\sc gsh} are compared to that of barodesy and hypoplasticity, with frequently quantitative agreement, Some typical curves as produced by {\sc gsh} are given here, see Fig~\ref{butterfly1} and \ref{triaxial}, and the two papers for more details and the values for the parameters.
\begin{figure}[t]\begin{center}
\includegraphics[scale=2]{butterfly}
\end{center}
\caption{Upper row: radial stress $\sigma
_{1}$ versus axial stress $\sigma _{3}$, rescaled by
$B_{0}\kappa ^{-3/2}$ (with $\kappa\equiv\sqrt{\zeta
_{1}\gamma_{1}}/\rho b$). Middle row: radial strain
$\varepsilon _{1}=\int v_{xx}dt$ versus axial strain
$\varepsilon _{3}=\int v_{zz}dt$. Lower row: $e-e_0$
(with $e_0$ the initial void ratio) versus shear
strain $\varepsilon_{q}=\int(v_{zz}-v_{xx})dt$,
rescaled by $\nu _{1}\kappa $. The stress loads are
isobaric for (a,c), and nearly (or quasi-) isobaric
for (b,d); the cyclic amplitude is small for (a,b) and
large for (c,d). The associated strain loci and void
ratio are: sawtooth-like for (a), coil-like for (b),
butterfly-like (or double-looped) for (c,d).}
\label{butterfly1}\end{figure}
\begin{figure}[tbh]
\hspace{-1cm}
\includegraphics[scale=0.7]{triaxial}
\caption{In the geometry of triaxial tests, various quantities are
computed employing {\sc gsh}, as functions of the strain $\varepsilon_{xx}$, holding $\sigma_{xx}=\sigma_{yy}$ constant. (The axial direction is $z$. The case with an initially higher density is rendered in solid lines, the looser one in dashed lines.) These are: (a)~deviatoric stress $q\equiv\sigma_{zz}-\sigma_{xx}$; (b)~void ratio $e$; (c)~volumetric strain $\varepsilon_v$; (d) the friction ange, $\sin \phi _{m}\equiv q/\left( 2\sigma_{xx}+q\right) $. We chose: $\alpha ,\alpha _{1},\lambda \sim \left( 1-\rho /\rho
_{cp}\right) ^{1.6}$ and $\eta_1,\gamma_1 \sim \left( 1-\rho /\rho _{cp}\right)
^{-1}$.}
\label{triaxial}
\end{figure}
\subsection{Stress-Controlled Experiments\label{aging}}
Only rate-controlled experiments have been considered up to now. Employing Eqs~(\ref{Tg2}, \ref{eqD}, \ref{eqU}), we found that
the granular temperature quickly becomes a dependent quantity, $T_g=T_c\equiv f|v_s|$, essentially reducing {\sc gsh} to the hypoplastic model,
with the exponential relaxation of $\Delta,u_s$ reproducing the approach to the critical state. In this section, we examine what happens if we instead hold
the shear stress $\sigma_s=(1-\bar\alpha)\pi_s$ constant. (As discussed in the introductory sentences at the beginning of Sec.\ref{hypoplastic motion}, rate-independence is a misplaced concept here.)
Typical examples of experiments of given shear stresses includes relaxation of $T_g\propto v_s$ and shallow flows on an inclined plane or in rotating drums. In the second case, there is a delay between jamming ({\it angle of repose} $\varphi_{re}$) and fluidization ({\it angle of stability} $\varphi_{st}$), with $\varphi_{st}$ larger by a few degrees. All these are considered below.
\subsubsection{Diverging Strain and Long-Lived Temperature\label{TgRelaxation}}
If $T_g=0$, the system stays static, $\sigma_s=$ const, and there
is no dynamics at all. If $T_g$ is initially elevated, $u_s$ relaxes, and with it also the stress $\sigma_s$.
Maintaining a constant $\sigma_s$ (or similarly, a constant $u_s$) therefore requires a compensating shear
rate $v_s$.
As long as $T_g$ is finite, $v_s(t)$ will accumulate, resulting in a growing shear strain
$\varepsilon_s(t)=\int v_s{\rm d}t$. As we shall see, for $u_s$ close to its critical value $u_c$, the characteristic time of $T_g$ is $\propto(1-u^2_s/u^2_c)^{-1}$ and long. Adding in the fact that the relaxation of $T_g$ is algebraically slow rather than exponentially fast, the accumulated shear strain can be expected to be rather large.
In a recent experiment, Nguyen et al.~\cite{aging} pushed the system
to a certain shear stress at a given and fairly fast rate, producing an elevated $T_g$. Then, switching to maintaining the shear stress, they observed the accumulation of a large total strain
$\varepsilon_s(t)$ that appears to diverge logarithmically. The authors referred to this phenomenon
as {\it creeping}, and took it to be a {compelling evidence} that in spite of the very slow motion, their experiment contains a dynamics and was not quasi-static. we note that this conclusion sits well with a basic contention of {\sc
gsh}, that what is usually taken as quasi-static motion is in fact hypoplastic, with an elevated $T_g$, as discussed above, see also Sec~\ref{quasi elastic motion} below.
This experiment may in principle be accounted for by the equations of {\sc gsh}, though due to the highly nonuniform stress distribution, this would require solving a set of nonlinear partial differential equations with coefficients as yet uncertainly known. Hence we only consider a shear-stress controlled experiment in the hypoplastic regime with uniform variables. Also, we first assume that it is the elastic shear strain $u_s$ that is being kept constant, not the shear stress $\sigma_s\propto\sqrt\Delta\,u_s$, as both cases will turn out to be rather similar.
The relevant equations are still Eqs~(\ref{Tg2},\ref{eqD},\ref{eqU}).
At the beginning, as the strain is being ramped
up to $u_s$ employing a constant rate $v_1$, the granular temperature acquires the elevated initial value $T_0=fv_1$. Starting at $t=0$, $u_s$ is being held constant. From Eq~(\ref{eqU}), we therefore conclude
\begin{equation}\label{3b-7a}
f|v_s|/T_g\equiv T_c/T_g={u_s}/{u_c},
\end{equation}
with $v_s$ the rate needed to compensate the stress relaxation.
Inserting this into Eqs~(\ref{Tg2},\ref{eqD}),
\begin{align}\label{3b-8a}
\partial_t\Delta=-\lambda_1T_g[\Delta-(u_s/u_c)^2\Delta_c],
\\\label{3b-9}
\partial_t T_g= -r_T\,T_g^2,\,\,\,r_T\equiv R_T[1-u_s^2/u_c^2],
\end{align}
we find the $T_g$-relaxation rate reduced from $R_T$ to $r_T$. Both equations may be solved analytically, if the coefficients are constant, which they are if the density is. The pressure $P(t)$ will then change with time, same as $\Delta(t)$. This is what we consider here. (Keeping the pressure constant implies time-dependence of density and coefficients. Then, as with the critical state considered in Sec~\ref{pressure approach}, a general solution is possible only by numerical methods.)
The first equation accounts for the relaxation of $\Delta$, from both below and above $(u_s/u_c)\Delta_c$. The relaxation is faster the more elevated $T_g$ is.
Employing the initial condition $T_g=T_0$ at $t=0$, and setting $h=1$, the solution to the second equation is
\begin{equation}\label{3b-10}
T_g={T_0}/({1+{r_T}{T_0}t}).
\end{equation}
Because of Eq~(\ref{3b-7a}), the solution holds also for the shear rate,
$v_s={v_0}/(1+ r_vv_0 t)$, with $v_0\equiv T_0/f$ and $r_v\equiv(fu_c/u_s){r_T}$.
This implies a slowly growing total shear strain
\begin{equation}
\varepsilon_s-\varepsilon_0\equiv\int v_s{\rm d} t=\ln(1+r_vv_0t)/r_v.
\end{equation}
However, $\varepsilon_s$ does not diverge,
because as $T_g$ diminishes, it eventually enters the quasi-elastic regime, $\gamma_1h^2T_g^2\to \gamma_0T_g$,
where its relaxation is exponential. More specifically, writing Eq.(\ref{3b-9}) as $\partial_t T_g=-(r_0+r_TT_g)T_g$, with $r_0/\gamma_0=r_T/\gamma_1$, we have the general solution
\begin{equation}
T_g=r_0[(r_T+r_0/T_0)\exp(r_0t)-r_T]^{-1}.
\end{equation}
Assuming a large $T_0$ (implying large rate to ramp up the stress), $\Delta$ is quickly relaxed, $\Delta=(u_s/u_c)^2\Delta_c$. Fixing
$u_s$ is then equal to fixing the shear stress, $\sigma_s\propto\pi_s\propto u_s\sqrt{\Delta} =(u_s^2/u_c)\sqrt{\Delta_c} $. With
$\pi_c\propto\sqrt{\Delta_c}\,u_c$, one may rewrite the factor in $r_T$ as
\begin{equation}\label{3b-12}
1-u_s^2/u_c^2=1-\pi_s/\pi_c\approx1-\sigma_s/\sigma_c.
\end{equation}
The $T_g$-relaxation is slower the closer $\pi_s$ is to $\pi_c$, infinitely so for $\pi_s=\pi_c$. Then we have $u_s=u_c$, $\Delta=\Delta_c$, with $T_g(t)=T_0$ a constant, see Eqs.(\ref{3b-7a},\ref{3b-8a},\ref{3b-10}).
This is indistinguishable from the rate-controlled critical state, which may be maintained clearly also at given stress.
If one chooses to keep $\sigma_s$ constant from the
beginning, irrespective how far $\Delta$ has relaxed,
one needs to require
$\partial_tu_s=(u_s/2\Delta)\partial_t\Delta$, resulting in a different
proportionality $v_s\propto T_g$ to be inserted into the equations
of motion. The results are similar.
Next, we keep both the pressure and shear stress constant from the beginning. Though the general consideration does not appear analytically viable, one solution of a realistic situation exists: Keeping $\Delta, u_s=$ const in Eqs~(\ref{eqD},\ref{eqUs}), we have
\begin{equation}\label{3b-15}
\frac{u_s}{u_c}=\frac{T_c}{T_g},\quad \frac{\Delta}{\Delta_c}=\frac{T_c^2}{T_g^2}-\frac{v_{\ell\ell}}{T_g}\frac{1-\alpha}{\lambda_1\Delta_c}.
\end{equation}
For given $\Delta$, taking $u_s$ such that ${\Delta}/{\Delta_c}={u_s^2}/{u_c^2}$, we have $v_{\ell\ell}=0$ and a constant density. Inserting ${u_s}/{u_c}={T_c}/{T_g}$ into
the balance equation for $T_g$, Eq~(\ref{Tg2}), we again obtain Eq~(\ref{3b-9})
with (\ref{3b-12}). The only difference is that there is now a clear prescription for the
experiment, because constant $\Delta,u_s,\rho$ means that pressure $P$ and shear stress $\sigma_s$ are kept constant. So one proceeds by applying an arbitrary pressure, then varying the shear stress until the density no longer changes.
$T_g, v_s$ will then be as calculated.
Comparable calculation and analysis were carried out in~\cite{aging}, using two scalar equations
that may roughly be mapped to the present ones. The quantities: granular temperature $T_g$, its relaxation and production rate, $R_T$ and
$R_Tf^2$, were referred to as {\em fluidity, aging} and {\em rejuvenation parameter}.
The above consideration is therefore not new, but does provide a tensorial treatment that is embedded in {\sc gsh}, rendering it transparent, unified, and more realistic, also affording a better founded understanding.
We also not that temporary, localized regions of strong deformation (called {\it hot spots}) were observed, with the fluidity (the averaged value of which is $T_g$) identified as their rate of occurrence.
As the stress distribution in the experiments of~\cite{aging} is rather nonuniform, there will
always be areas with a shear stress close to $\sigma_c$. And the system will tend to cave in there, resulting in a larger strain accumulation than what the average value for $\sigma_s$
would predict.
In the experiment, a very soft spring was used to couple
the fan and the motor. This we believe is essential why this
experiment turned out as observed. Usually, triaxial apparatus with stiff walls are used. And the correcting rates employed by the feedback loop to keep the stress constant are of
hypoplastic magnitudes. As a result, much $T_g$ is excited, and we have the situation of consecutive constant rates, not that of constant stress. The soft spring, as discussed above, and in greater detail in Sec~\ref{soft springs}, enables quasi-static stress correction without exciting much $T_g$.
With an ambient temperature $T_a$, the $T_g$ relaxes as $\partial_tT_g=r_T(T_g-\eta T_a)$, with $\eta\equiv1/(1-u_s^2/u_c^2)$, see Eq~(\ref{Tg2}). This means, the values $T_g$ and $v_s$ respectively relax to, ${\eta T_a}$ and ${\eta v_a}$, get strongly amplified close to $u_s=u_c$. This is a large effect.
\subsubsection{Stability above the Critical Shear Stress\label{scs}}
From the consideration of the last two sections we see that a granular assembly is, for an elevated $T_g$, mechanically stable only up to the critical value for the elastic stress $\pi_c$. For $\pi_s<\pi_c$, $T_g$ grows, since $r_T$ is negative. (As we shall see in Sec.\ref{nsb}, shear bands are formed as a result of this instability.) On the other hand, for $T_g=0$, the system is stable at any static shear stresses exceeding $\pi_c$, as long as Eq.(\ref{sb12}) is not breached.
Now, since an infinitesimal $T_g$ is ubiquitous, and if it always grows, there is no stability for static shear stresses exceeding $\pi_c$. It does not always grow: Only an initial $T_g$ of hypoplastic strength will explode, not an infinitesimal one, of quasi-elastic strength. This is because $h$ diverges for $T_g\to0$, and the critical stress diverges with $h$: Since $f\propto 1/h$, we have $u_c\propto h,\Delta_c\propto h^2$, and $\sigma_c\propto h^2$, see Eqs.(\ref{critV}). Therefore, $r_T$ is always positive for very small $T_g$. In fact, what we have for strain values above $u_c$ is a metastability, a stability that may be destroyed only by granular jiggling of sufficient strength. This fact is associated with familiar phenomena: A house on a cliff collapsing due to elastic waves from a distant earth quake, or a pneumatic hammer close by; a gun shot initiating an avalanche.
The elastic strain instability for $u_s>u_c$ holds only for stress-controlled experiments, not rate-controlled ones, though this distinction is not always clear-cut in experiments. For instance, if a step motor is used for a strain-controlled experiment, and one has a strain versus time curve such as given by Fig~\ref{StepMotor} below, than the stress is being hold constant at the plateaus, rendering the stability of the uniform system precarious. This may well be the reason why shear band formation is so frequently observed in the cases where the initial density is high and the non-monotonic stress trajectory exceeds $u_c$, see Fig~\ref{fig3}.
Finally, we stress that these aspects of granular behavior are natural results of {\sc gsh}, not preconceived features planted in while constructing it. They stem from the interplay between yield and the critical state, or more precisely, between the instability of the elastic energy and the stationary solution of the elastic strain.
\subsubsection{Angle of Stability and Angle of Repose \label{sapc}}
Aranson and Tsimring were the first to construct a theory for these two angles~\cite{aranson,Aranson1}. Taking the stress $\sigma_{ij}$ as the sum of two parts, one solid, the other fluid-like, they define an
order parameter $\hat\varrho$ that is 1 for solid, and 0 for dense flow. They then postulate a free energy $f(\hat\varrho)$ such that it is stable with $\hat\varrho=1$ only for $\varphi< \varphi_{st}$, with $\hat\varrho=0$ only for $\varphi>\varphi_{re}$, and $\varphi_{st}>\varphi>\varphi_{re}$ as the bi-stable region. The solid stress is taken as an input, assumed understood from some other theory.
In comparison, the consideration below, given within the context of {\sc gsh}, is somewhat more complete and less {\em ad hoc}.
Fluidization, the collapse that occurs when one slowly tilts a plate supporting a layer of grains, is a process that happens at $T_g=0$, with no granular jiggling. Therefore, the Cauchy stress is given by the elastic one, $\sigma_{ij}=\pi_{ij}$.
On a plane inclined by the angle $\varphi$, with $y$ denoting the depth of the granular layer on the plane, and $x$ along the slope, we take the stress to be $\pi_{xx},\pi_{yy},\pi_{zz}=P_\Delta$, $\pi_{xy}=\pi_s/\sqrt2$, $\pi_{yz},\pi_{xz}=0$. Integrating $\nabla_j\pi_{ij}=g_i\rho$ assuming a variation only along $y$, we find $\pi_{xy}=g\sin\varphi\int\rho(y)dy$ and $\pi_{yy}= \pi_{xy}/\tan\varphi$. The angle of stability $\varphi_{st}$ is reached when the energetic instability of Eqs.(\ref{2b-3}) is breached. With $\pi_s^{yield}\equiv P\sqrt{2{\cal A}/{\cal B}}$ denoting the yield shear stress, it is
\begin{equation}\label{sb12}
\tan\varphi_{st}=\pi_s^{yield}/\sqrt2 P=\sqrt{{\cal A}/{\cal B}}.
\end{equation}
Effects derived from proximity to the wall or floor are considered in Sec.\ref{clogging}.
The angle of repose $\varphi_{re}$ is related to the calculation of the last two sections. As long as the shear stress is held below the critical one, $\sigma_s<\sigma_c$, the $T_g$-relaxation will run its course, and the system is in a static, mechanically stable state afterwards. At $\sigma_s=\sigma_c$, however, the system becomes critical, and no longer comes to a standstill. Therefore, $\varphi_{re}$ is given by ${\sigma_c}$,
\begin{equation}
\tan\varphi_{re}={\sigma_c}/\sqrt2\,P_c, \quad\text{with\,\,}\varphi_{re}<\varphi_{st}.
\end{equation}
The inequality holds because the critical state is an elastic solution, while $\varphi_{st}$ is the angle at which all elastic solutions become unstable.
That $\varphi_{re}$ and $\varphi_{st}$, material parameters, differ only slightly, is related to the microscopic fact that both account for the clearance with the profile of the underlying layer -- though one with granular jiggling and hence a little easier.
\subsection{The Visco-Elastic Behavior of Granular Media\label{visElaBeh} }
All visco-elastic systems (such as polymer solutions) have a characteristic time $\tau$ that separates two frequency ranges: fluid-like behavior for $\omega\tau\ll1$, and solid-like one for $\omega\tau\gg1$. Like granular media, polymers are transiently elastic, though the transiency is constant and not variable, because $\tau$ is. The hydrodynamic theory of polymers,
with a very similar elastic strain $u_{ij}$ that obeys the equation
$\partial_tu^*_{ij}-v^*_{ij}=-u_{ij}^*/\tau_{ve}$,
is capable of accounting for many visco-elastic phenomena, including shear-thinning/thicken\-ing, elongational viscosity, the Cox-Merz rule, and the rod-climbing (or Wei\ss enberg) effect~\cite{polymer-1,polymer-2,polymer-3,polymer-4}.
The main difference of the granular analogue, Eq~(\ref{2c-8}), is the fact that the relaxation time varies as $\tau\propto 1/T_g$ -- a granular system is fully elastic for $T_g\to0$, capable of sustaining a static shear stress. Moreover, rate-independence, a granular characteristics not observed in viscous elastic systems with a constant $\tau$, stems from the relation $1/\tau=\lambda T_g\propto v_s$.
However, when there is an ambient temperature in granular media, much larger than the temperature produced by the imposed shear rate, $T_a\gg T_c\equiv f|v_s|$, polymers and granular media are very similar in their behavior, because $T_a$ is also a given quantity that does not depend on the local shear rate. The ambient temperature $T_a$ may be maintained by a standing sound wave, periodic tapping, or by diffusion from a region of great granular agitation. In all cases, the resultant $T_a$ enables the relaxation of the elastic strain and stress, implying no static stress may be maintained, and the yield stress vanishes.
\subsubsection{The Creep Motion\label{creep motion}}
In granular media, one frequently observes shear bands, which borders on a non-shearing, solid part. Careful experiments reveal that the shear rate is in fact continuous, with an exponentially decaying creep motion taking place in the solid, see Komatsu et al~\cite{komatsu}, Crassous et al~\cite{crassous}.
We show here that this is a result of $T_g$ from the fluid region diffusing into the solid one, being present there
as an ambient, spatially decaying temperature $T_a$ that enables stress relaxation. If the stress is to be maintained, there must be a compensating shear rate that also decays in space, along with $T_a$, and the velocity obtained from integrating the shear rate is the observed creep motion.
Consider a ``liquid-solid boundary" at $x=0$, with the shear rate being concentrated on one side, for $x>0$. (We shall return to consider the liquid side in Sec.\ref{sb}. Here, we only take the fluid values at $x=0$ to provide the boundary conditions for $v_s,T_g$ in the solid part.) For a one-dimensional geometry, the pressure $P$, shear stress $\sigma_s$, the shear rate $v_s$ and $T_g$ are uniform, but $\rho$ need not be. We take $\rho$ to be discontinuous at $x=0$, but constant otherwise, with $v_{\ell\ell}=0$, and $T_g,v$ varying perpendicular to the boundary, along $\hat x$.
The circumstances are then quite similar to that of Sec~\ref{aging}, though variation is in space rather than time. First, with stationarity of Eqs~(\ref{eqD},\ref{eqUs}) [see also Eq.(\ref{3b-15})], we have
\begin{equation}\label{props}
\frac\Delta{\Delta_c}=\frac{u_s^2}{u_c^2}=\frac{T_c^2}{T_g^2}=\frac{\pi_s}{\pi_c}=\frac{\sigma_s}{\sigma_c},\,\,\frac\Delta{u_s}=\frac{\Delta_c}{u_c}\frac{T_c}{T_g}.
\end{equation}
With $\Delta,u_s$ fixed, so are $P,\sigma_s$, where especially $P=P_c$ if $\sigma_s=\sigma_c$. Note also that since the stable branch of $P/\sigma_s=P_\Delta/\pi_s\equiv1/\mu$ increases monotonically with $\Delta/{u_s}$, see Eq.(\ref{2b-1}), the last above equation implies that the friction $\mu$ decreases for increasing $T_c/T_g$.
The balance equation for $T_g$ [with $\partial_tT_g=0$ but including the diffusive current, see Eqs.(\ref{Tg2},\ref{3b-9})] reads
\begin{align}
\label{cr1}
\nabla^2T_g=T_g/\xi^2_{cr},\quad \xi_{cr}^2\equiv\xi_T^2/[1-\pi_s/\pi_c]
\\\label{cr2}
\text{implying}\quad v_s/v_s^{0}= T_g/T_g^{0}= \exp(-x/\xi_{cr}),
\end{align}
where $v_s^{0},T_g^{0}$ are the fluid values at $x=0$.
That the decay length $\xi_{cr}\equiv \xi_T/\sqrt{1-\sigma_s/\sigma_c}$ diverges for $\sigma_s=\sigma_c$ is not surprising, because the solid region, turning critical, ceases to exist then. Although subcritical, $\sigma_s<\sigma_c$, the solid region sustains a finite rate $v_s\not=0$, because $T_g$ is being continually diffused from the fluid region.
Note $\sigma_s$ is a uniform quantity across the boundary, yet we necessarily have $\sigma_s<\sigma_c(\rho)$ on the solid side, $\sigma_s\geq\sigma_c(\rho)$ on the fluid side, implying a lower fluid density.
Finally, the above exponential decay with the constant length $\xi_{cr}$ holds only in the hypoplastic regime. Once $T_g$ is sufficiently small, we have $h\to\infty$, and $\xi_T\propto h^{-1}$ vanishing quickly.
In two recent papers~\cite{kamrin,kamrin2}, Kamrin et al propose a nonlocal constitutive relation (KCR) well capable of accounting for steady flows in the split-bottom cell~\cite{fenistein}. A key ingredient is the fluidity $g\equiv v_s/\mu$. With $\mu\equiv\sigma_s/P$, $\mu_s\equiv\sigma_c/P_c$, it is taken to obey
\begin{equation}
\xi^2_{cr}\nabla^2g=g-g_{loc},\quad \xi_{cr}\propto1/\sqrt{|\mu-\mu_s|}.
\end{equation}
Because $g_{loc}=0$ for $\mu<\mu_s$, this relations is rather similar to Eq.(\ref{cr1}), with $g$ assuming the role of $T_g$, and the two decay lengths diverging at the same stress values
For $\mu\geq\mu_s$, the system is fluid, and $g=g_{loc}$ essentially constant. With $g_{loc}\propto\sqrt P(1-\mu_s/\mu)$, KCR is consistent with a first-order expansion of the MiDi relation, Eq.(\ref{poul}), in the inertial number.
GSH is compared to MiDi in Sec.\ref{rdf}, showing broad agreement and some relevant disagreements.
Here, we only discuss the additional differences of {\sc gsh} to KCR.
First, KCR does not take the density as a variable, leading to inconsistencies: The stress is continuous at the solid-fluid interface and strictly constant in a one-dimensional geometry. As discussed below Eq.(\ref{cr2}), {we necessarily have $\sigma_s<\sigma_c(\rho)$ on the solid side, $\sigma_s\geq\sigma_c(\rho)$ on the fluid side, implying a lower fluid density.} Without the density, the same two conditions imply a discontinuity in $\sigma_s$ or $\mu_s$ which violates momentum conservation. Another drawback is the fact that granular behavior depends sensitively
on whether density or pressure is being held constant see Sec.~\ref{pressure approach} above and \ref{udf} below. This cannot be reproduced employing KCR.
Second, being {defined} as $v_s/\mu$, the fluidity $g$ is not an independent variable like $T_g$, though it does possess a postulated, independent dynamics. If one eliminates $g$, rewrites its equation as $\mu\xi^2_{cr}\nabla^2(v_s/\mu)=v_s-v_s^{loc}$, a problem arises: This equation (in conjunction with $v_{\ell\ell}=0$) and the momentum conservation may both be used to calculate the velocity field for given density and stress. The results will in general be contradictory.
\subsubsection{Nonlocal Fluidization\label{nonlocal fluidization}}
{\em Non-local fluidization} is an observation made (and named) by Nichol et al.~\cite{nichol2010}, see also Reddy et al.~\cite{reddy2011}. In a vessel of grains, after a shear band is turned on, the medium everywhere, even further away from the band, looses its yield stress, and the Archimedes law holds: A ball stuck at whatever height without the shear band starts to sink or elevate, until its density is equal to the surrounding one.
{\sc gsh}'s explanation for this behavior is quite simple: First, $T_g$ generated by the shear band diffuses through the solid phase, as accounted for by Eq.(\ref{cr1}), permeating the medium as a spatially decaying ambient temperature $T_a$.
Second, a medium such ``fluidized" obeys, as observed earlier~\cite{huerta2005,caballero2009}), the Archimedes law, because the
ball getting stuck in the sand deforms the grains around itself and builds up an elastic shear stress. Without an ambient temperature, $T_a=0$, this stress holds up the ball's weight if it is not too large, and the ball is stationary. With $T_a\not=0$, the stress relaxes, requiring a compensating shear rate $v_s$ to maintain the stress balance, implying a
moving ball. We note that $T_a\not=0$ does not imply the grains need to jiggle violently. If the ball's descent takes an hour, a barely perceptible slip every minute would be quite sufficient. And $T_a$ is the spacial and temporal average of the changing energy contained in these slips.
More quantitatively, a solid object being dragged by a constant force $F^{ext}_i$ through a granular medium will quickly settle into a motion of constant velocity $v_\infty$, implying a stationary stress and velocity distribution in the medium, in the rest frame of the object. So Eqs.(\ref{3b-15}) holds. This is remarkable, because the elastic stress $\pi_{ij}(\Delta,u_s)$ transforms, under the replacement $\Delta,u_s\to T_c\equiv f|v_s|,v_{\ell\ell}$, into a viscous stress. And this enables one to perform a calculation similar to that needed to arrive at the Stokes' law.
The Stokes' law $F^{drag}_i=6\pi R\eta v$ is derived assuming an incompressible (and infinitely extended) medium, with $v_{\ell\ell}=0$. The resulting velocity field, scaling with $v_\infty$, is a pure geometric quantity that does not depend on any parameters, especially not the applied force $F_{ext}$~\cite{LL6}. In contrast, granular media possess sound velocities one to three times that of air and are rather compressible. As a result, both the velocity field and all parameters (that are functions of the density) will depend on $F_{ext}$.
In fact, that the viscosity seemingly depends on the mass of the steel ball ($\propto$ the gravitational force) was observed in~\cite{nichol2010}.
Inserting Eqs.(\ref{3b-15}) into
Eq~(\ref{2b-2}), we have, with $\sigma_{ij}=(1-\alpha)(P_\Delta\delta_{ij}+\pi_sv_{ij}^*/v_s)$,
\begin{equation}\label{viscStress}
P_\Delta=\frac{{\cal A}u_c^2}{2\sqrt{\Delta_c}}\frac{T_c}{T_g}
\quad\pi_s=-2{\cal A}u_c\sqrt{\Delta_c}\frac{T_c^2}{T_g^2},
\end{equation}
where $P_\Delta$ contains only the of lowest order term in $v_s, v_{\ell\ell}$, while $\pi_s$ is valid assuming $v_{\ell\ell}=0$ (and appropriate for the steel plate below). Note
$T_g=\sqrt{T_c^2+T_a^2}$ has two contributions, $T_a$ from the remote shear band, and $T_c\equiv f|v_s|$ from the nonuniform local shear rate. For $T_a=0$, $P_\Delta$ and $\pi_s$ are rate-independent, and the system is in a (nonuniform) critical state. For $T_a\gg T_c$,
the system is viscous, and one may define two effective viscosities, $P=\eta^{eff}_1v_s$, $\sigma_s=-\eta^{eff}_2v_s^2$, with $\eta^{eff}_1\propto1/T_g$, $\eta^{eff}_2\propto1/T_g^2$. In~\cite{huerta2005}, faster ascent and a smaller viscosity were observed in regions of larger granular agitation (and attributed to ``pressure screening'').
Given the form for the stress, one can calculate the velocity field depending on the geometry of the object. The drag force is then obtained by inserting the field into $\sigma_{ij}$, and integrating it over the surface of the object,
$F^{drag}_i=\oint \sigma_{ij}{\rm d}a_j$.
The simplest case is that of a {\bf steel plate}, say perpendicular to $\hat x$ and being dragged along $\hat y$. The shear rate is a constant, $v_s=\frac12\nabla_xv_y$, with $v_{\ell\ell}=0$,
and the force $F^{drag}$ per unit surface of the plate is $2\sigma_{xy}\propto v_s^2/T_a^2$.
The velocity field for a {\bf ball of radius $\boldsymbol R$} is not as easily calculated, though it is clear that, for $T_c/T_a$ small, the drag force stems from the pressure and is linear (and not quadratic as with the plate): $F^{drag}_i=\oint P{\rm d}a_i\propto v_s/T_g\propto v_\infty/T_g$, as observed in~\cite{caballero2009}.
Assuming incompressibility (as one does deriving the Stokes' law though inappropriately here), one finds
$F^{drag}_i=\oint P{\rm d}a_i=(9\pi^2/16)({\cal A}u_c^2f/\sqrt{2\Delta_c})\,(R v_\infty/T_a)$.
Any hydrodynamic theory starts from the basic assumption that its resolution is small compared to the system size, but much larger than any microscopic lengths -- in the present case, especially the grain diameter $d$. In~\cite{reddy2011}, the diameter of the probing rod, a system size, is only $2d$. Although averaging over time and runs usually retrieves the macroscopic behavior, this may not work quantitatively when the two scales are essentially the same.
Summarizing, the dichotomy of the elastic stress and a $T_g$-dependent viscosity is the basic {\sc gsh}-explanation for granular visco-elasticity. In this more general picture, creep motion may equally well be understood as the viscous motion under constant moment of inertia.
\subsection{Narrow Shear Bands\label{nsb}}
Typical constitutive models such as hypoplasticicty or barodesy do not properly account for shear bands, and the reason is the lacks of a length scale. There are
various approaches to overcome this short coming, by introducing gradient terms~\cite{wu2} or adding state variables to account for the couple stress and the Crosserat rotation~\cite{wu1}.
Especially the Crosserat method works well, but it leads to a far more complex theory, constructed for the sole purpose of solving the shear band problem. Moreover, it throws up the question about the underlying physics: If couple stress and rotational motion are important in the shear band, because it is fluid, why then are they not important in the uniformly fluid and gaseous state of granular media, see Sec.\ref{rdf}, or more generally, in nematic liquid crystals~\cite{deGennes}?
The purpose of this section is to point out that {\sc gsh} is well capable of accounting for the shear band without any modification. We consider a system of uniform density and stress, with all variables stationary, such that Eqs.(\ref{props}) hold.
The balance equation (\ref{3b-9}) for $T_g$, accounting for $T_g$'s relaxation to 0 if $\pi_s<\pi_c$, implies $T_g\equiv0$ is the uniform stationary solution, see Sec.\ref{TgRelaxation}. For $\pi_s=\pi_c$, the system is in the critical state, $T_g$ does not relax and the strain rate is indeterminate. For $\pi_s>\pi_c$, no uniform solution is stable, but a localized one is, with $T_g\equiv0$ for $x\le0$ or $x\ge\xi_{sb}$, and
\begin{align}\nonumber
\nabla^2 T_g=-T_g/\xi^2_{sb},\quad{\xi}^2_{sb}\equiv\xi_T^2/[\pi_s/\pi_c-1],
\\
v_s/v_s^0=T_g/T_g^0=\sin(\pi x/\xi_{sb})\label{nsb2}
\end{align}
in between. [Note the similarity to Eq.(\ref{cr1}). Allowing $\rho$ to vary will render $T_g$ differentiable at $0,\xi_{sb}$.] The velocity difference from 0 to $\xi_{sb}$ is $\Delta v=\int v_s{\rm d}x=\int (T_g/f)\sqrt{\pi_s/\pi_c}{\rm d}x$, hence
\begin{equation}
T_g^0/f=\sqrt{\pi_c/\pi_s}\, v_s^0 =\sqrt{\pi_c/\pi_s}\,\Delta v/(2\xi_{sb}).
\end{equation}
The critical state and the narrow shear band are the same rate-independent solution, behaving differently depending on how large $\pi_s$ is. That the correlation length $\xi_{sb}$ diverges for $\pi_s=\pi_c$ gives a retrospective justification of the term {\it critical.}
For increasing $\Delta v$, the variables $v_s,u_s,\Delta/u_s$ also grow, and the system will eventually leave the rate-independent, hypoplastic regime. Shear bands become wider then, and have to be treated as in Sec.\ref{sb}.
The above is an idealized and simplified consideration of narrow shear bands, assuming uniform density and stress, and employing {\sc gsh} expressions that have been linearized and simplified. (Neither did we invoke the higher order strains terms of Sec.\ref{clogging}, implying in essence $\xi_{sb}\gg \theta,\theta_1$.) The qualitative and structurally stable part of the results is a localized shear band solution of {\sc gsh}, for overcritical stress values, with a characteristic length that decreases with increasing $\pi_s$. When approaching the critical state non-monotonically, starting from a dense initial state, with $\pi_s>\pi_c$ for part of the path, there is a high probability for the $T_g$-instability discussed in Sec.\ref{scs} to occur and shear bands to form.
Details such as the spacial distribution of $T_g(x)\propto v_s$, or that the friction angle $\sigma_s/P=\pi_s/P_\Delta$ decreases with increasing $T_c/T_g\propto\Delta{u_s}$, however, should be taken with a grain of salt, as these depend on the details and may change with the starting assumptions and {\sc gsh} expressions. For instance, the original $T_g$ equation and the associated solution are
\begin{equation}
\label{Tg3}
\nabla_i (T_g\nabla_i T_g)=-T_g^2/\xi^2_{cb},\quad T_g=T_0\sqrt{\sin(\sqrt2\,x/\xi_{sb}}),
\end{equation}
see the discussion preceding Eq.(\ref{Tg1}), leading to the neglect of the nonlinear term $(\nabla_iT_g)^2$, small for slow variations, in both Eq.(\ref{Tg1}) and (\ref{nsb2}).
Same holds for Eq.(\ref{cr1}).
Finally, if the density is nonuniform, say due to an aggregation of macropores, we will have $\pi_s>\pi_c(\rho)$ only in some regions. $T_g$ will be larger there, diffusing away, making the situation less clear-cut.
\subsection{Clogging and the Proximity Effect\label{clogging}}
The phenomenon of clogging implies that a free surface, if several grain diameter wide, may be stable even when facing downward, implying an angle of stability of $180^\circ$,
much larger than than the usual $30^\circ$ or $40^\circ$, as discussed around Eqs.(\ref{sb12}), valid only if the surface area is sufficiently large.
Although {\sc gsh} in its present form, as given in Sec.\ref{GSH}, does not account for clogging, there is a tried and proven method of amending it. One example is the Ginzburg-Landau description of the superfluid transition~\cite{Khal}, which includes gradients of the order parameter's magnitude in the energy. In the present case, we need to include gradients of the
elastic strain that express the extra energetic cost of a nonuniform strain field. Without these terms, unclogging occurs accompanied by a discontinuity in $\Delta,u_s$.
With them, divergent gradients are forbidden by the infinite energy. A length scale on which elastic strains will change is thus introduced.
With $w_\Delta=w_\Delta(u_{ij},\nabla_ku_{ij})$, $-\pi_{ij}\equiv\partial w_\Delta/\partial u_{ij}$, $\phi_{ijk}\equiv\partial w_\Delta/\partial\nabla_ku_{ij}$, the elastic and total stress are, respectively
\begin{equation}\label{GSE3}
\hat\pi_i\equiv\pi_{ij}+\nabla_k\phi_{ijk}, \quad \sigma_{ij}=[1-\alpha(T_g)]\hat\pi_i.
\end{equation}
Denoting the two characteristic lengths as $\theta,\theta_1$, a simple example for such an energy is
\begin{align}\label{GSE8}
w_\Delta=\sqrt{\Delta }[2{\mathcal B}\Delta ^2/5+
{\mathcal A}u_s^2] +{\mathcal A}(\theta\nabla_k
u_s)^2+{\mathcal B}(\theta_1\nabla_k
\Delta)^2],
\\\label{GSE9a}\text{implying}\quad
\hat P=P-2{\cal B}\theta_1^2\nabla_k^2\Delta,\quad
\hat\pi_s=\pi_s+2{\cal A}\theta^2\nabla_k^2u_s,
\end{align}
with $P,\pi_s$ the uniform contributions, assuming $u^*_{ij}/|u_s|=$ const. Note that with this energy, the convexity transition, $u_s/\Delta\le\sqrt{2{\cal B}/{\cal A}}$ of Eq.(\ref{2b-3}) is unchanged (though $\pi_s/P_\Delta\le\sqrt{2{\cal A}/{\cal B}}$ does change), because with $w=w_1(a)+w_2(\nabla a)$ and
\begin{align*}
\delta^2w=\delta(\delta w)=\delta\left(\frac{\partial w}{\partial a}-\nabla\frac{\partial w}{\partial \nabla a}\right)\delta a=\delta\left(\frac{\partial w_1}{\partial a}-\nabla\frac{\partial w_2}{\partial \nabla a}\right)\delta a\qquad
\\
=\left(\frac{\partial^2 w_1}{\partial a^2}\delta a-\nabla\frac{\partial^2 w_2}{\partial (\nabla a)^2}\delta\nabla a\right)\delta a=\left(\frac{\partial^2 w_1}{\partial a^2}+\frac12\nabla^2\frac{\partial^2 w_2}{\partial (\nabla a)^2}\right)(\delta a)^2,
\end{align*}
$a$ standing for $\Delta$ or $u_s$, we have $\delta^2w/\delta a^2={\partial^2 w_1}/{\partial a^2}$. (Note $\int\nabla[{\partial^2 w_2}/{\partial (\nabla a)^2}]\delta\nabla a\delta a =
-\int\nabla^2 [{\partial^2 w_2}/{\partial (\nabla a)^2}]\delta a^2-\int\nabla[{\partial^2 w_2}/{\partial (\nabla a)^2}]\delta a\delta\nabla a$ if the surface integral vanishes.)
We employ this result and the model energy Eq.(\ref{GSE8}) to consider, qualitatively, clogging and the proximity effect. More quantitative treatment will be provided in a
separate work. First the effect that the angle of stability $\varphi_{st}$ is, for a few layers of grains, much larger than given in Eq.(\ref{sb12}). Simpler, that $\pi_s/P_\Delta$ can
be larger than $\sqrt{2{\cal A}/{\cal B}}$ in a one-dimensional, simple shear geometry of the width $L$.
We consider the strain fields for $-L<y<L$: $\Delta=\Delta_0$, $u_s=u_0+\alpha y^2/3L^2$ [ie. displasement $U_x=u_0y+\alpha y^3/12L^2$], with $\Delta_0,u_0=$ const, $u_0/\Delta_0\le\sqrt{2{\cal B}/{\cal A}}$ , and
$\alpha\ll u_0$ such that the direct contribution to $\pi_s$ is negligible. Then $\hat P=P$, $\hat\pi_s=\pi_s+{\cal A}\alpha\theta^2/L^2$, and the uniform correction is considerable for $\theta\gg L$. The fact that crushing is most efficient when the shearing walls are only a few grain diameters apart is clearly related to the above consideration that elastic solutions remains stable at large $\pi_s$. The grains remain static until they are crushed in narrow geometries, while transitioning into sliding, rotating, critical states in wider ones.
Next, a crude model for clogging. Since the stress vanishes for any free surfaces, we examine the 1D-situation in which it is zero for $-L<x<L$, but finite at $x=\pm L$ and beyond. Taking
$\hat P_\Delta,\hat \pi_s=0$ as the differential equations, we solve them for $-L<x<L$ subject to the boundary conditions $\Delta=\Delta_0, u_s=u_0$ for $x=\pm L$. Assuming for simplicity that $\theta_1\ll L$, we take $\Delta\equiv\Delta_0$, implying $(1-\bar\theta^2\nabla_k^2)u_s=0$ with $\bar\theta={\theta}/{\sqrt[4]{\Delta_0}}$, or
\begin{equation}
\frac{u_s(x)}{u_0}\left[1+\exp\left(\frac{2L}{-\bar\theta}\right)\right]=\exp\left[\frac{x+L}{-\bar\theta}\right]+\exp\left[\frac{x-L}{\bar\theta}\right].
\end{equation}
If $u_0$ satisfies the stability condition, $u_0/\Delta_0\le\sqrt{2{\cal B}/{\cal A}}$, the solution $u_s(x)$ also does, and therefore represents a stable elastic situation.
\section{Rapid Dense Flow\label{rdf}}
\subsection{The $\mu-$Rheology versus GSH\label{udf}}
When considering hypoplastic motion in the last section, \ref{hypoplastic motion}, we neglected the kinetic pressure $P_T$ and the viscous shear stress $\propto\eta_g$, see Eqs.(\ref{2c-2},\ref{2c-2a},\ref{2c-2b}). Here, we consider faster flows in which they are important, some times even dominant. Including them, we are leaving the rate-independent, hypoplastic regime.
Being quadratic in the shear rate, the correction come on slowly, leaving a large rate regime in which rate-independence holds.
How the stress of a system, in it stationary state, depends on the density $\rho$ and shear rate $v_s$, is called its {\it rheology}. Probing it over a wide range of shear rates is
a useful inquiry for coming to terms with complex fluids including granular media. Granular rheology has many facets, and typically, the shear rate $v_s$ is given.
If it is low, the system executes complex elasto-plastic motion with a rate-independent stress, converging onto the critical state at constant rates, with a universal shear stress $\sigma_c$ that depends only on the density, not the rate or the initial stress, as considered in Sec~\ref{hypoplastic motion}.
If $v_s$ is high and the density sufficiently low, the system is in the Bagnold regime, with all components of the stress proportional to shear rate squared~\cite{Bagnold}. We consider the whole regime below.
If the shear stress is given instead of $v_s$, circumstances are yet different. Examples are flows on an inclined plane or in a rotating drum, with a delay between jamming ({\it angle of repose} $\varphi_{re}$) and liquefaction ({\it angle of stability} $\varphi_{st}$), see Sec~\ref{sapc}. Part of the results of this section is in~\cite{P&G2013}.
\subsubsection{The $\mu-$Rheology\label{mu-rheo}} Sixty years ago, Bagnold examined how a granular system behaves at high rates and low densities, finding the pressure $P$ and shear stress $\sigma_s$ given as
\begin{equation}\label{intro1}
P=e_p(\rho)v_s^2, \quad \sigma_s=e_s(\rho)v_s^2,
\end{equation}
with $\mu_2\equiv\sigma_s/P=e_s/e_p$ a constant~\cite{Bagnold}. This result has been variously verified employing the kinetic theory to consider binary collisions among rarefied, dissipative grains~\cite{kin1,kin2,kin3,kin4}.
A decade later, granular rheology at low rates and high densities was studied. Again, a surprisingly universal so-called {\it critical state} was observed~\cite{schofield,wood1990,nedderman,gudehus2010}.
Starting from any initial stress, the system will, at constant densities and shear rates, acquire values for the pressure and shear stress that depend on $\rho$ but not the rate, with the friction $\mu_1\equiv\sigma_s^c/P^c=$ const.
Faced with these results, many find it plausible to account for the intermediate behavior by interpolating between the two rate- and density-independent plateaus~\cite{interpolate2,interpolate3,interpolate4,interpolate5},
\begin{equation}\label{intro2}
P=P^c+e_p(\rho)v_s^2, \,\, \sigma_s=\mu_1P^c+\mu_2e_pv_s^2,
\,\, \mu\equiv\sigma_s/P,
\end{equation}
implying $\mu\to\mu_1$ for $v_s\to0$ and $\mu\to\mu_2$ for $v_s\to\infty$.
Embarking on an approach independent from the above and stressing first principles, the French research group GDR MiDi consider infinitely rigid grains~\cite{midi, pouliquen1}, and point out that its rheology has only three independent numbers: the friction $\mu$, the packing fraction $\phi\equiv \rho/\rho_g$ and the inertial number $I\equiv d \sqrt{\rho_g}( v_s/\sqrt P)$, with $\rho_g$ the bulk density, $d$ the granular diameter.
Taking two as functions of the third, $\mu=\mu(I)$, $\phi=\phi(I)$, Forterre and Pouliquen~\cite{pouliquen2} take granular rheology to be accounted for by
\begin{equation}\label{poul}
\mu=\mu_1+({\mu_2-\mu_1})I/({I + I_0}),
\end{equation}
with $\mu_1\approx\sqrt2\tan 21^\circ$, $\mu_2\approx\sqrt2\tan 33^\circ$, $I_0\approx0.3$. Containing two plateaus, same as Eqs.(\ref{intro2}), this formula is shown capable of accommodating many experiments and simulations, and has recently also been successfully applied to dense suspensions~\cite{pouliquen4}.
However, there is a fundamental problem. The relations $\mu=\mu(I)$, $\phi=\phi(I)$ are (irrespective of their functional dependences) equivalent to Eqs.(\ref{intro1}), implying $P^c, \sigma_s^c=0$: First, $\phi=f(I)$ is clearly equivalent to $P=v_s^2/f^{-1}(\phi)$; second, $\mu(I)=\mu[f^{-1}(\phi)]$ is a function of $\phi$ alone, and the two plateaus are for large and small packing fractions, respectively, impßlying $P,\sigma_s\propto v_s^2$. However, this contradicts half a century worth of research in soil mechanics, unambiguously showing {\it rate-independent stresses} for elasto-plastic motion, $v_s\to0$.
The validity of Eqs.(\ref{intro1}) for infinitely rigid grains has been rigorously proven by Lois et al.~\cite{lois}, for any rates and densities, not only where the kinetic theory holds.
Yet the speed of elastic waves in glass beads is between 350 and 800 m/s~\cite{jia2009}, which in comparison to air, water, bulk glass (with velocities of 300, 1500, 4000~m/s, respectively) indicates a very soft medium. The difference between glass beads and bulk glass stems from the geometry of the Hertz contact. When considering binary collisions, assuming infinitely rigid grains reduces the collision time to zero, but does not change the physics qualitatively. Assuming incompressibility in dense media eliminates elastic waves and the critical state.
When arguing that one may treat grains as infinitely rigid, the authors of~\cite{lois}, citing a paper by Campbell~\cite{campbell}, assert that as long as $M\equiv dv_s/c_s$ (with $c_s$ the sound velocity) is small (typically for $ v_s\ll10^3$/s), grains behave as if they were perfectly stiff. This is oddly reversed, because quasi-static
deformations occur at small rates, and are disrupted at higher ones. Indeed,
perusing~\cite{campbell}, one finds Campbell stating clearly: (1)~It is the {\it inertially induced} contact deformation that vanishes with $M$. (2)~Stresses are generated by elastic deformations in the rate-independent, ``elastic-quasi-static'' regime.
\subsubsection{The Dense Flow Results of GSH\label{vsGSH}} Treating dense granular media as compressible, {\sc gsh} shows the appropriateness of Eqs.(\ref{intro2}). Starting from Eqs.(\ref{2c-2},\ref{2c-2a}), we substitute the elastic contributions with the critical state expressions, Eqs~(\ref{3b-4a}), appropriate for constant shear rates, while noting $P_T=g_pT_g^2=g_pf^2v_s^2\equiv e_p v_s^2$, see Eq.(\ref{2b-5},\ref{Tcf}), also $\eta_1T_gv_s=\eta_1f v_s^2\equiv e_sv_s^2$, to obtain
\begin{equation}\label{3b-17}
P=P_c+e_p{v_s^2},
\,\,\sigma_{s}=\sigma_c+e_sv_s^2.
\end{equation}
The observed density-independence of $\mu_1,\mu_2$ implies the constancy of $P_c(\rho)/\sigma_c(\rho)=\mu_1$ and $e_s(\rho)/e_p(\rho)=\mu_2$, an experimental input.
(A viscous stress linear in $v_s$, as observed in~\cite{campbell} at high densities, has not been included above but is a possibility, see~\cite{granR2,granR3}. It appears if a macroscopic shear flow not only heats up $T_g$, but $T$ as well. This is the case for instance when the sand is saturated with water. The pressure $P$ would receive a term liner in $T_g$ if one modify the energy, $w=w_T+w_\Delta+w_x$, by adding a cross term such as $w_x=cs_g\Delta^{1.5}$. The total pressure $P$ and $T_g$ then obtain the respective additive term, $cs_g\sqrt\Delta$ and $c\Delta^{1.5}$, implying that the linear term $\propto T_g$ exists only for $\rho>\rho_{\ell p}$ and $\Delta\ne0$.)
If the density is low, $\rho<\rho_{\ell p}$, the grains lack enduring contacts and no elastic solution is stable, $P_\Delta,\pi_s=0$, see Eq.(\ref{2b-4}). Then Eqs.(\ref{intro1}) are the appropriate formulas.
When studying granular rheology by varying the shear rate $ v_s$, one can either keep the density or the pressure constant. For any realistic shear rates, we have $P^c\gg e_p(\rho) v_s^2$, $\sigma_s^c\gg e_s(\rho) v_s^2$, and it is hard to arrive at the $\mu_2$-limit for given $\rho$. Not so for given pressure, because the density decreases for increasing $v_s$. A discontinuous transition from Eqs.(\ref{intro2}) to (\ref{intro1}) takes place at $\rho=\rho_{\ell p}$, when $\mu$ jumps from $\mu_1$ to $\mu_2$, while $P, \sigma_s$ decrease dramatically, by around three orders of magnitude~\cite{campbell}.
Two important points remain to be discussed, first when and why there is, as observed~\cite{Brodsky1,Brodsky2}, a minimum in the shear stress as a function of the rate; and second, why the MiDi relation is, in spite of its shortcomings, so successful.
Keeping the density constant, $P^c(\rho)$, $\sigma_s^c(\rho)$, $e_p(\rho)$, $e_s(\rho)$ also are, implying $P,\sigma_s$ increase monotonically with $v_s^2$. Keeping $P=$ const, the circumstances, though still given by Eqs.(\ref{intro2}), are different.
In the hypoplastic regime, the shear stress at given pressure, $\sigma_s^c=\mu_1P$, is simply a constant, similarly in the Bagnold regime, $\sigma_s=\mu_2P$.
In between, both $\rho,\sigma_s$ are rate- and density-dependent, given by $\sigma_s=\mu_1P^c(\rho)+\mu_2e_p(\rho) v_s^2$, with $\rho(P, v_s)$ from $P=P^c(\rho)+e_p(\rho) v_s^2$ plugged in. Then there is no reason for $\sigma_s(P, v_s)$ to be monotonic, see~\cite{StressDip} for more details.
Eqs.(\ref{intro2}) are algebraic relations that hold for uniform systems. To account for nonuniform ones, gradient terms from {\sc gsh} become important, and the large set of
nonlinear, partial differential equations that {\sc gsh} is needs to be solved. Even disregarding this, there are still complications that one needs to heed.
For instance, enforcing a constant total volume does not prevent the local density to vary,
and a stress dip may still occur.
To understand this better, consider two uniform volumes $V_1,V_2$, with $V_1+V_2=$ const. Being in contact via a flexible membrane, they may serve as a simple model for the continuous non-uniformity of a constant volume
experiment. Initially, the
total system is uniform, with both densities equal,
$\rho_1=\rho_2$, and both shear rates vanishing,
$\dot\gamma_1, \dot\gamma_2=0$. Now, if $\dot\gamma_2$ is
cranked up, but $ \dot\gamma_1$ remains zero, because of
pressure equality,
$P_1(\rho_1, \dot\gamma_1)=P_2(\rho_2, \dot\gamma_2)$, the
density must change and the membrane will stretch, with $\rho_2$ decreasing and
$\rho_1$ increasing. If system~1 is much larger than 2,
the stretching of the membrane will not change $\rho_1$
much, and $P_1(\rho_1, \dot\gamma_1)$ will remain
essentially constant as a result. So will $P_2=P_1$, and
the pressure-controlled limit holds in system 2. Otherwise, we have an
intermediate case between the pressure- and
density-controlled limits. In both cases, a stress dip may appear.
Finally, we give three reasons for the undeniable success of the MiDi relation: First, $\phi=\phi(I)$ is correct for $\rho<\rho_{\ell p}$, while $\mu=\mu(I)$ as given by Eq.(\ref{poul}) is right for $\rho>\rho_{\ell p}$. Very few papers span both limits and employ both relations simultaneously.
Second, many experiments are nonuniform, lying between the density- and pressure-controlled limits. An unreflective comparison of the relation to a subset of data such as the average density or stress is then neither accurate nor discriminating. In fact, by employing Eqs.(\ref{intro1},\ref{intro2}), Berzi et al.~\cite{interpolate5} were able to achieve quantitative agreement with both the simulation on simple shear in~\cite{daCruz} and the experiment on incline flows in~\cite{Opouli}. Both were deemed strong support for the MiDi relation.
Third, the frequently observed collapse of different curves, when $\mu$ is plotted as a function of $I$, may be understood because $\mu$ depends on $\hat I\equiv e_p(\rho) v_s^2/P$ alone, and $\hat I$ is close to $I^2$. [One writes $\mu=({\sigma_c}/{P})({P-e_p v_s^2})/{P_c}+{e_s v_s^2}/{P}$ $=\mu_1(1-e_pv_s^2/P)+\mu_2e_p/P$ $=\mu_1+(\mu_2-\mu_1)\hat I$.
Generally speaking, granular rheology is given by $P,\sigma_s=f(\rho,v_s)$. One may switch to $\rho,\phi=f(P,v_s)$ or $\mu,\phi=f(I,v_s)$, two variables remain and there is no collapse. $\mu=f(\hat I)$, $\hat I\stackrel{v_s\to\infty}{\longrightarrow}1$ is an exception.] Note depending whether $\rho$ or $P$ is being held constant, one must take, respectively, $\hat I=e_p(P,v_s^2)v_s^2/P$ and $\hat I=e_p(\rho)v_s^2/P(\rho,v_s^2)$.
\subsection{Wide Shear Bands\label{sb}}
The narrow shear band has already been considered in Sec.\ref{nsb}. Here, we consider a wide shear band, which is in essence the coexistence of static granular solid and uniform dense flow. In the first, the grains are deformed and at rest, $T_g=0$, with all energy being elastic. In the
second, the grains jiggle, rattle, move macroscopic distances, with $T_g\propto v_s$ and a portion of the energy in $T_g$.
Increasing the shear rate, the transition from the rate-independent critical state to the
Bagnold regime of dense flow is, as discussed in Sec~\ref{udf}, continuous at given density and discontinuous at given pressure, but always uniform. Here, we consider a nonuniform path, a narrow shear band that suddenly appears, as the result of an instability, see Sec.\ref{nsb}, then continuously widens as the externally applied velocity difference increases, until the band covers the whole system, and uniformity is restored.
Approaching the critical state with a high initial density, the evolution of the shear stress
$\sigma_s$ is non-monotonic, assuming overcritical values part of the path. This is
where the system has a high probability of breaching an instability, either of the elastic energy at a point on the yield surface, as discussed around Eq.(\ref{2b-3}), or that of $T_g$, as discussed in Sec~\ref{scs}.
The transition is difficult to account for, but the stable shear band is again simple.
As we have seen, the narrow shear band of low shearing velocity $v$ has a rate-independent width. If $v$ is higher, the system's behavior depends on the setup. For given pressure, the width $\ell$ grows linearly with $v$, implying a constant rate $v/\ell$ in the liquid phase. As a result, the shear stress, a function of the rate, remains independent of $v$. This {\em faux rate-independence} goes on until the band covers the whole system, at which point the quadratic rate dependence of uniform dense flow sets in.
For given volume, the band width remains independent of $v$, but the shear stress grows quadratically with it. The transition to uniform dense flow is for given volume discontinuous. It happens when the shear stress exceeds the critical value of the solid density, at which point the solid phase is no longer stable.
To understand wide shear bands, we study the simple case of uniform fluid and solid regions connected via a flat surface. (Separately, they are already understood.) Denoting the solid and fluid parts with the superscripts $^S$ and $^F$, respectively, these two regions have equal pressure, shear stress, and chemical potential,
\begin{equation}\label{3b-21}
P^S=P^F,\quad \sigma_s^S=\sigma_s^F, \quad\mu^S=\mu^F.
\end{equation}
[The chemical potential is defined as $\mu\equiv\partial w/\partial\rho$, see Eq~(\ref{2-2}). The equality holds because otherwise a particle current would flow across the phase boundary.] All three fields have an elastic and a seismic contribution: With $P=({1-\alpha})P_\Delta+P_T$, ${\sigma_s}=({1-\alpha})\pi_s+\eta_1T_g v_s$, see Eqs~(\ref{2c-2},\ref{2c-2a}), and $\mu=\mu_\Delta+\mu_T$, where
\begin{align}
\label{5-7}
\mu_T\equiv T_g^2\,\frac {b_0\rho}2\left[1-\frac{\rho}{\rho_{cp}}\right]^a\frac{(1+a)\rho-\rho_{cp}}{\rho_{cp}-\rho},
\\\mu_\Delta\equiv{0.15w_\Delta
}{(\rho_{cp}-\bar\rho)}/[{(\rho_{cp}-\rho)}{(\rho-\bar\rho)}].\end{align}
Denoting the width of the shear band as $\ell$, and the velocity difference across the shear band as $v$, we take
\begin{flalign
\text{in fluid:}&\quad v_s=v/\ell\propto T_g,\,\, \Delta^F=\Delta_c,\,\, u_s^F=u_c,
\\
\text{in solid:}&\quad \alpha, T_g, v_s=0.
\end{flalign}
In other words, the elastic strain $\Delta$ and $u_s$ have critical values in the $F$-phase, and appropriate static values in the $S$-phase. Strictly speaking, the discontinuities at the $S-F$ boundary are in $\rho,\Delta, u_s$, but not in $T_g$ and $v_s$, as both diffuse into the solid, decaying exponentially there, see Sec~\ref{creep motion}. We neglect this detail, approximating the decay with a discontinuity to keep the formulas simple, and to work at the qualitative understanding first. The price we pay is a slightly fuzzy $\ell$ that includes the two decay zones in the solid.
\subsubsection{The Fluid Region}
The elastic contribution $\mu_\Delta$ is a very small quantity: In $P_\Delta\propto{\cal B}\Delta^{1.5}$, a large ${\cal B}$ compensates a small $\Delta^{1.5}$, such that $P_\Delta$ is either much larger than, or comparable to, $P_T\propto T_g^2$. Now, $\mu_T$ is of the order of $P_T/\rho$, but $\mu_\Delta\propto{\cal B}\Delta^{2.5}\propto \Delta P_\Delta$ is smaller by the factor $\Delta$, around $10^{-3}-10^{-4}$. Therefore, as long as $P_T\gg\Delta P_\Delta$, we have $\mu_T\gg\mu_\Delta$, and $\mu^S=\mu^F$ reduces to $\mu_T=0$, implying the density in the shear band is (in dry sand) fixed as
\begin{equation}\label{5-8}
\rho^{F}=\rho_{cp}/(1+a).
\end{equation}
Measuring $\rho^{F}$ therefore yields the value of $a$, see Eq~(\ref{2b-5}). In what follows, we need to assume a sufficiently small $a$, such that $\rho^F>\rho_{\ell p}$. Because $\Delta^F=\Delta_c (\rho^F), u_s^F=u_c(\rho^F)$, the elastic pressure $P_\Delta(\rho,\Delta,u_s)$ in the fluid is also known.
{\bf Given Pressure\quad}
Next, we consider the case of given velocity difference $v$ across the shear band, and given external pressure, $P^{ex}=P^S=P^F$,
\begin{align}\label{sb-1}
P=P_c(\rho^F)+\frac{T_g^2}{2}\,\frac{(\rho^F)^2\,a\,b/\rho_{cp}}{(1-\rho^F/\rho_{cp})},
\\\label{sb-2}
\sigma_s=\sigma_c(\rho^F) -\eta_1T_g\,v/\ell.
\end{align}
Since $P^F,\rho^F$ fix $T_g,v_s=T_g/f$, and for given $v$, the width of the shear band $\ell=v/v_s$ is also fixed, we have all there is to know about the fluid region.
Remarkably, the system now displays a faux rate-independence: $\ell$ adjusts itself such that $T_g\propto v/\ell$ remains constant for given pressure,
independent what $v$ is. The parabola of Fig~\ref{fig4} depicts $\sigma_s$. The offset gives the elastic contributions, $\sigma_c$. The horizontal line is a result of $\ell$ adjusting.
\begin{figure}[t]
\begin{center}
\includegraphics[scale=.45]{fig4}
\end{center}\caption{Faux rate-independence: Shear stress $\sigma_s$ as a function of the velocity difference $v$, or of the apparent shear rate $v_s\equiv v/L$, for given pressure, in a simple-shear geometry. The offset gives the elastic contribution, $\sigma_c(\rho_F)$; the parabola is the case without a shear band. The thick horizontal line depicts the situation with a shear band, of width $\ell$, which is smaller towards left, and equal to the system's width $L$ at the right end. The rate-independence of $\sigma_s$ derived from $\ell$ adjusting itself such that $T_g\propto v/\ell$ remains constant for given pressure. \label{fig4}}
\end{figure}
Increasing the velocity $v$ at given pressure alters the width $\ell$, as long as it is smaller than the width of the total system $L$. For larger velocities, the system is again uniform, without a solid region. And the consideration of Sec~\ref{udf} holds. Until this point, the stress is rate-independent, much longer than without a shear band.
Given the solid density $\rho^S$ (see Sec.\ref{sr}), and the mass per unit surface $M$, mass conservation $\rho^S(L-\ell)+\rho^F\ell=M$ determines the total width $L$ for given pressure $P$.
{\bf Given Total Volume\quad}
At given total volume $L$, because of mass conservation
and because $\rho^S,\rho^F$ are given in addition to $L$, the band width $\ell$ is fixed, irrespective what the velocity $v$ is. As a result, both the shear stress and pressure grow as $(v/\ell)^2\propto v^2$, not at all rate-independent. The transition to uniform dense flow happens discontinuously, when Eq~(\ref{94}) is violated, for $\sigma_c(\rho^S)=\sigma^S$.
\subsubsection{The Solid Region\label{sr}}
Because we have terms of such different magnitudes in the connecting condition $\mu^S=\mu^F$, it fixes $\rho^F$ instead of
giving a relation between $\rho^F$ and $\rho^S$. Therefore, the condition is always satisfied,
irrespective what value $\rho^S$ assumes. So $\rho^S$ can only be a result of the dynamics: When an instability is breached, the density is changed until it gets stuck at
some value for $\rho^S$, at which the system is again stable. Then of course, $\Delta^S, u_s^S$ may be determined for given pressure and shear stress.
Nevertheless, we do know
\begin{equation}\label{92}
\rho^F<\rho^S \quad\text{and}\quad \rho^F\le\rho_c(P)
\end{equation}
must hold. The first inequality can be seen from
\begin{equation}\label{94}
\sigma_c(\rho^S)>\sigma^S=\sigma_c(\rho^F)+\eta_1T_gv_s\ge\sigma_c(\rho^F),
\end{equation}
where the first greater sign is related to the discussion in Sec~\ref{scs}; the equal sign is a connecting condition; and the second greater sign is a result of $\eta_1T_gv_s$ being positive, in addition to the fact that $\sigma_c$ is a monotonically increasing function of the density, cf. the discussion in Sec.\ref{Stationary Elastic Solution}. The second inequality, $\rho^F\le\rho_c(P)$, holds because of two reasons: First,
in the critical state, there is only one free parameter. Once $\rho$ is given, $\Delta_c,u_c,P_c,
\sigma_c$ also are. Alternatively, one may fix the external pressure $P$, then $\rho_c(P)$ is a dependent quantity. Second, given $P=P_c(\rho^F)+P_T$, we have $\rho_F=\rho_c$ for $P_T=0$. $\rho^F$ may be smaller, but if it were larger, shear band will not exist, and the flow is uniform.
For $\rho^F<\rho_{\ell p}$, there is no elastic contribution in the shear band, $P_c,\sigma_c=0$ in Eqs.(\ref{sb-1},\ref{sb-2}), and Eq.(\ref{92}) holds trivially. All other conclusions remain valid, also Fig.\ref{fig4}, though without the offset $\sigma_c$.
When the velocity $v$ decreases, the above consideration stops to be valid at some point. For instance, $\rho^F$ is no longer given if $P_T\gg\Delta P_\Delta$ does not hold. But before this happens, the narrow band solution should already have taken over.
\section{Velocity and Damping of Elastic Waves\label{elastic waves}}
That elastic waves propagate in granular media~\cite{jia1999,jia2004} is an important fact, because it is an unambiguous proof that granular media possess an
elastic regime. In this section, we consider elastic waves and propose to employ them as a tool to detect the elastic to plastic transition.
There is a wide-spread believe in the granular community that small, quasi-static increments from any
equilibrium stress state is elastic, but large ones are plastic. As discussed in
Sec~\ref{quasi elastic motion}, this assumption appears illogical, because any large increment can always be taken as the sum of small ones. In {\sc gsh}, the parameter that sets the
boundary between elastic and plastic regime is the granular temperature $T_g$. We have quasi-elastic regime for vanishing $T_g\propto v_s^2$, and the
hypoplastic one for elevated $T_g\propto v_s$.
A perturbation in the elastic strain or stress propagate as a wave only in the quasi-elastic regime, while it diffuses in the hypoplastic one.
More specifically, we derive a telegraph equation from {\sc gsh}, with a quantity $\propto T_g$ taking on the role of the electric resistance~\cite{zhang2012}. It defines a
characteristic frequency $\omega_0=\lambda T_g$, such that elastic perturbations of the frequency $\omega$ diffuse for $\omega\ll\omega_0$, and propagate for
$\omega\gg\omega_0$. We have $\omega_0\to0$ in the quasi-elastic regime, so all perturbations propagate. In the hypoplastic regime, when $T_g$ is elevated, so is
$\omega_0$, pushing the propagating range to ever higher frequencies. Eventually, the associated wave length become comparable to the granular diameter, exceeding {\sc gsh}'s range of validity.
To derive the telegraph equation, we start with two basic equations of {\sc gsh}, Eqs~(\ref{2c-8},\ref{2c-2a}),
\begin{align} \label{ew1}
\rho\partial_tv_i-(1-\alpha)\nabla_mK_{imkl}u^*_{kl}=0, \\
\partial_tu^*_{ij}-(1-\alpha)v^0_{ij}=-\lambda T_gu^*_{ij}, \label{ew2}
\end{align}
with $K_{imkl}\equiv\partial^2w/\partial u_{im}\partial u_{kl}$. (For simplicity, we concentrate on shear waves, assuming $v_{\ell\ell}\equiv0 $.)
For $T_g\to0$, the plastic terms $\lambda T_gu^*_{ij}$ and $\alpha\propto T_g$ are negligibly small, and these two equations represent conventional elasticity theory.
The wave velocity $c$ [given by the eigenvalues of $K_{imnj}q_mq_n/(\rho q^2)$ with $q_m$ the wave vector], as a function of stress and density, is then easily calculated. The results~\cite{ge4} agree well with observations~\cite{jia2009}.
There are two ways to crank up $T_g$ and the plasticity, either by introducing external perturbations $T_a$, or by increasing the amplitude
of the wave mode, because its own shear rate also creates $T_g$. The characteristic time of $T_g$ is $1/R_T\lesssim10^{-3}$~s in dense media, see Eq~(\ref{Tg2}). Therefore, we assume that the wave mode's frequency is much larger than $R_T$ , such that $T_g$ and $\alpha(T_g)$ are essentially constant, or
\begin{align}\label{ew3}
2(\partial^2_t+\lambda T_g\partial_t)\,u^*_{ij}=(1-\alpha)^2\times\qquad
\\
\nabla_m[K_{imkl}\nabla_ju^*_{kl}+K_{jmkl}\nabla_iu^*_{kl}].
\nonumber
\end{align}
Concentrating on one wave mode along $x$, with $c_{\rm qs}$ the quasi-elastic, $c\equiv(1-\alpha)c_{\rm qs}$ the actual velocity, and $\bar u\propto e^{iqx-i\omega t}$ the eigenvector's amplitude, we have the telegraph equation,
\begin{equation}\label{ew4}
(\partial^2_t+\lambda T_g\partial_t)\,\bar u= (1-\alpha)^2c_{\rm qs}^2\nabla_x^2\,\bar u
\equiv c^2\nabla_x^2\,\bar u.
\end{equation}
The coefficient $(1-\alpha)^{2}$, accounting for granular contacts softening and the effective elastic stiffness decreasing, is, in the language of electromagnetism, the inverse dielectric permeability.
Inserting $\bar u\propto e^{iqx-i\omega t}$ into Eq~(\ref{ew4}), we find
$c^2 q^2={\omega^2+i\omega\lambda T_g}$,
implying diffusion for the low frequency limit, $\omega\ll\lambda T_g$,
\begin{equation}
q\approx\pm\frac{\sqrt{\omega\lambda T_g}}c \, \frac{1+i}{\sqrt2},
\end{equation}
and propagation for the high-frequency limit, $\omega\gg\lambda T_g$,
\begin{align}
cq\approx\pm\omega \left(1+i\,{\lambda T_g}/{2\omega}
\right),
\\
\bar u\propto\exp{\left[-i\omega\left(t\mp x/c\right) \mp x({\lambda T_g}/{2 c})\right]}.
\end{align}
The first term in the square bracket accounts for wave propagation, the second a decay length $2c/\lambda T_g$, independent of the frequency if $T_g=T_a$ is an ambient temperature. It is strongly frequency and amplitude dependent if $T_g=f|v_s|\propto\omega q\bar u\propto\omega^2\bar u$ is produced by the elastic wave itself, because the inverse length varies with $T_g$, going from $T_g\propto v_s^2$ to $T_g\propto v_s$.
A brief wave pulse, arbitrarily strong, can always propagate through granular media if its duration is too brief to excite sufficient $T_g$ for the system to enter the hypoplastic regime. The duration must be much smaller than $T_g$'s characteristic time $1/R_T$.
\section{Compaction\label{compaction}}
The present understanding of {\it compaction under tapping} takes it to be a rather insular phenomenon, in need of an special entropy not useful for any of the other granular phenomena. We shall return to the so-called {\it Edwards entropy} in Sec.\ref{tapping}, after having pondered whether tapping may be related to a ubiquitous variety of compaction that has been known to engineers for a long time, the slow increase of the density at given pressure under shear, or in the presence of an ambient temperature $T_a$. This more typical phenomenon is easily understood to be a result of the fact that $\Delta$ relaxes, as accounted for by Eq.(\ref{eqD}). Keeping the pressure $P_\Delta={\cal B}(\rho)\Delta^{1.5}$ constant, the density increases to compensate. (Note that Approaching the critical state under a constant shear, the circumstances are more general, because $\Delta$ relaxes and is being increased by a shear rate at the same time. It may increase, leading to dilation, or decrease, to contraction, as considered in Sec~\ref{pressure approach}.)
\subsection{Reversible and Irreversible Compaction}
Consider the pressure $P=(1-\alpha)P_\Delta+P_T$ assuming vanishing shear strain and rate, $u_s,v_s=0$, with $P_\Delta$ the elastic, and $P_T$ the seismic, contribution, see Eqs~(\ref{2b-2b},\ref{2b-5},\ref{2b-4}),
\begin{equation}\label{gc5}
P_\Delta={}{\cal B}(\rho)\Delta^{1.5},\quad P_T= g_p(\rho)T_g^2
\end{equation}
where both $\cal B$ and $g_p$ are, for dense media, monotonically increasing functions of $\rho$.
At small $T_g$, the seismic pressure $P_T$ may be neglected, so $\rho$ must increase when $\Delta$ relaxes, for $P=P_\Delta=$ const. The increase is irreversible because the relaxation is. This is the limit most soil mechanical experiments are in. Only irreversible compaction is observed.
For $T_g$ larger, the seismic pressure $P_T$ needs to be included. Because the density change in $g_p$ is faster than in $\cal B$, the relaxation of $\Delta$ increases $P_T$ and decreases $P_\Delta$, with $P_\Delta+P_T=$ const. After the relaxation has run its course, $\Delta,P_\Delta\to0$, if one modifies
$T_g$ (ie. the amplitude of the perturbation) but maintains $P=P_T$, the density will change in response, in both direction and reversibly. Since $P_T(\rho,T_g)\equiv\partial(w/\rho)/\partial(1/\rho)$ is a thermodynamic derivative, the change is also thermodynamic.
\subsection{History Dependence versus Hidden Variables\label{hdvhv}}
Changing $T_g$ midway at constant $P$, with $\Delta$ still finite,
will mainly lead to a change in $\Delta$, because the density responds much more slowly.
It disrupts the relaxation of $\Delta$, in essence resetting its initial
condition. This phenomenon was observed in~\cite{mem} and interpreted as a
memory effect. Generally speaking, ``memory" is
usually a result of hidden variables: When the system
behaves differently in two cases, although all state
variables appear to have the same values, we speak of memory-, or
history-dependence. But an overlooked variable that
has different values for the two cases will naturally
explain the difference. In the case of compaction, the manifest and hidden
variables are $\rho$ and $\Delta$, respectively.
\subsection{Tapping and the Edwards Entropy\label{tapping}}
Numerous experiments have shown that tapping leads to reversible and irreversible compaction, see the review article~\cite{1nico}. It is usually accounted for by the
specifically tailored {\it granular statistical mechanics}~\cite{Edw} and the Edwards
entropy $S_{Ed}$, or some generalization of it. Substituting the volume
$V$ for the energy $E$, and compactivity $X$ for the
temperature $T$, this theory employs ${\rm d}V=X{\rm
d}S_{Ed}$ as the basic thermodynamic relation for a
{\it``mechanically stable agglomerate of infinitely
rigid grains at rest"}~\cite{Edw}. The entropy
$S_{Ed}$ is obtained by counting the
possibilities to package grains stably for a given volume,
equating it to $e^{S_{Ed}}$.
Two reasons prompt us to doubt its appropriateness.
First, the number of possibilities to arrange
grains concerns inter-granular degrees of freedom. These are vastly overwhelmed by the much more numerous configurations of the inner-granular degrees of freedom. In other words, the Edwards entropy $S_{Ed}$ is a special case of the granular entropy $S_g$, and as discussed in the introduction, we always have $S_g\ll S$.
One would be able to neglect $S$ and concentrate on $S_g$ if these two were only weakly coupled, if the energy decay from $S_g$ to $S$ were exceedingly slow. This is not the case, the relaxation of $T_g\propto s_g$ is fast.
Second, even assuming a weak coupling,
$S_{Ed}$ would still be a overwhelmed measure. The starting point of the Edwards entropy is the fact that the energy $E$ is always zero for infinitely rigid, non-interacting grains at rest, however they are packaged. Taking $S_g$ generally as a function of energy and volume, $S_g(E,V)$, we have,
\[{\rm d}S_g=\frac{\partial S_g}{\partial E}{\rm d}E+\frac{\partial
S_g}{\partial V}{\rm d}V \equiv\frac1{T_g}{\rm d}E+\frac P{T_g}{\rm d}V.\]
Usually, one keeps the volume constant, and consider ${\rm d}S_g=(1/T){\rm d}E$. Taking instead $E\equiv0$, we have ${\rm d}S_g=(P/T)
{\rm d}V$, equivalent to the Edwards expression ${\rm d}V=(T/P){\rm d}S_g\equiv X{\rm d}S_{Ed}$.
This derivation ignores three essential points: First, perturbing the system, allowing it to explore the phase space, introduces kinetic energy that one must include. But then $E\not\equiv0$. Second, because of the Hertz-like contact
between grains, little material is deformed at first contact, and the compressibility diverges at
vanishing compression. This is a geometric fact independent of how rigid the bulk material is.
Therefore, infinite rigidity is never a realistic limit in granular media, and there is always considerable
elastic energy stored among grains in mechanically stable agglomerates. Third, $S_{Ed}$ as defined is the granular entropy at vanishing granular motion and compression. Its phase space is therefore severely constrained.
Generally speaking, each classical particle has states in a 6D space, three for the position and three for the velocity. The Edwards entropy only includes states in a 3D space. So $\exp(S)$ is the number of states times the Loschmidt's number; $\exp(S_g)$ is the number of states in 6D space times the number of grains, and $\exp(S_{Ed})$ is the number of states in 3D space (no velocities) times the number of grains. Therefore
\begin{equation}
S_{Ed}\ll S_g\ll S.
\end{equation}
Going toward equilibrium, a system searches for the greatest number of states to equally redistribute its energy. One bears the burden of proof for the claim that it is sensible for the system to neglect $S,S_g$ and concentrate on $S_{Ed}$. In contrast, {\sc gsh} identifies compaction as a process taking place at finite $T_g$ and compares the true entropy $S$ of macrostates at that $T_g$. It also accounts for entropy increase, by detailing how macroscopic energy decays into granular heat, and how this is converted to true heat.
More specifically, maximizing the true entropy $S$, {\sc gsh} obtains two sets of equilibrium conditions, one for the solid and another for the fluid state~\cite{granR2,granR3},
\begin{align}
\nabla_i\pi_{ij}=\rho\, {g}_i,\quad T_g=0;\\
\pi_{ij}=0, \,\,\,
\nabla_iP_T=\rho\, {g}_i.
\end{align}
The first is the result of $T_g$ vanishing quickly, leaving a jammed, elastically deformed system. The second (implying $\Delta, P_\Delta=0$) holds, when $T_g=T_a$ is being maintained externally. This is the limit of reversible, thermodynamic compaction, for $\Delta=0$.
Reversible and irreversible compaction as accounted for by {\sc gsh} is a universal granular
phenomenon. It occurs at given pressure and $T_a$, however $T_a$ is created. This corresponds well to the observation that tapping, though especially efficient, is but one way to achieve compaction, leading to results vary similar to that of many other methods~\cite{1nico}. So it is natural to take the consideration of the last section to hold for tapping as well. This rings true for gentle tapping, but stronger one warrants further scrutiny.
Gentle tapping leads to granular jiggling and a small $T_g$,
though one that fluctuates in time, with periodic flare-ups. As long as $P_T$ may be neglected,
$\Delta$ will relax according to the momentary value of $T_g$, haltingly but monotonically.
Since the relaxation is a slow process, one could average over many taps to yield a coarse-grained
account. Given a granular column with a free upper surface in the
gravitational field, because a given layer is subject to a constant pressure, the density will increase to compensate for the diminishing $\Delta$. The characteristic time of $\Delta$-relaxation diverges towards the end, and is not a constant, see~\cite{compaction}.
Stronger tapping leads to a higher $T_g$, with $\Delta$ relaxing more quickly.
$P_T$ must now be included. Periodically, when all grains are at rest, $P_T$ vanishes, and $\Delta$ is necessarily increased to maintain the given pressure. This introduces a non-monotonicity into $\Delta(t)$, and raises the question, whether the system, when being tapped again to arrive at an elevated $T_g$, will pick up the relaxation of $\Delta$ where it was left when the system last crushed to a stop, and also why it should do so. If it does, we can again take tapping as coarse-grainable, intermittent compaction. Then {\sc gsh} indeed provides a complete picture for compaction, with an
understanding that is transparent, conventional and demystified.
\section{The Quasi-Elastic Regime\label{quasi elastic motion}}
Textbooks on soil mechanics take granular motion in the hypoplastic regime -- say the approach to the critical state -- to be quasi-static, because it is slow and rate-independent. Yet since it is also strongly dissipative and irreversible, we do not believe this is right: Quasi-static motion is not dissipative.
Consider sound propagation in any system including Newtonian fluid, elastic medium or liquid crystals. The sound velocity is always an order in the frequency lower than the damping.
This is a general feature: Changing a state variable $A$ slowly, dissipation is $\propto\partial_tA$. For $\omega\to0$, the motion is free of dissipation and rate-independent. One calls it {\em quasi-static} because the system is at this frequency visiting static states consecutively.
In the hypoplastic regime, reactive and dissipative terms in {\sc gsh} are of the same order in the frequency, and comparable in size -- they are exactly equal in the critical state -- and elastic waves are over-damped.
So there must be a true quasi-static regime at even lower frequencies.
A more convoluted explanation, popular in the geotechnical community, is to assume that a small incremental strain is elastic and free of dissipation, but a large one is elasto-plastic and dissipative. Unfortunately, this is incompatible with the basic notion of quasi-static motions: Starting from a static state of given stress, and applying a small incremental strain that is elastic, the system is again in a static state and an equally valid starting point. The next small increment must therefore again be elastic. Many consecutive small increments yield a large change in strain, and if the small ones are not dissipative, neither can their sum be. This cannot go on for ever, and the limit is
the elastic convexity transition of Eq.(\ref{2b-3}), at which no elastic state is stable. So in the quasi-elastic regime, granular media behave in accordance to the simplest elasto-plastic theory: completely elastic for small shear stresses, and ideally plastic when the yield stress is breached.
Together, these reasons let us believe that it is $T_g$, rather than strain amplitude that decides whether the system is elastic or elasto-plastic. Of course, small strain increments achieved with a high but short lasting shear rate will indeed provoke elastic responses, if $T_g$ does not have time to get large and produce plastic responses.
To be specific, we quote a few numbers, well aware that these are at best educated guesses for the case of dry sand:
The Bagnold regime starts at rates of one or two hundred Hz, the hypoplastic regime is say between $10^{-3}-1$Hz, and quasi-elastic regime lies possibly below $10^{-5}$Hz.
{\sc gsh} accounts well both for static stress distribution and the hypoplastic regime.
Its prediction of what should happen in between, in the quasi-elastic regime, derives from a continuous connection of these two behavior, and is not yet verified
experimentally. Granular media are taken to be completely elastic, with the elastic energy given by Eq.(\ref{2b-2}) for the static case of identically vanishing shear rate,
$v_s\equiv0$. Many known static stress distributions have thus been successfully reproduced, including silos, sand piles and point load on a granular sheet,
see~\cite{ge1,ge2,granR1}. Also, Incremental stress-strain relation starting from varying static stress points~\cite{ge3}, and the propagation of anisotropic elastic waves at varying static stresses~\cite{ge4} are well accounted for.
The elasto-plastic motion that are on display for hypoplastic shear rates and elevated $T_g$ is also in full agreement with experiments and state-of-the-art engineering theories such as hypoplasticity and barodesy, see Sec.\ref{Constitutive Relations}.
Given the two limits, there is only little leeway of how to connect both. {\sc gsh} employs $h$ of Eq.(\ref{2c-6}) as the switch, such that $h=1$ and $T_g\propto v_s$ in the rate-independent hypoplastic regime, while $h\to\infty$ and $T_g\propto v_s^2$ quadratically small in the quasi-static one. Since deviations from elasticity of all expressions vanish with $T_g\to0$, the transition is smooth.
For experiments at given shear rates, the key difference between the hypoplastic and quasi-elastic regime lies in whether the system retrace the stress-strain curve when the rate is reversed, see next section. For experiments at given shear stresses (employing a soft spring, see Sec.\ref{soft springs} below) in the hypoplastic regime, an initially elevated $T_g$ will relax sufficiently slowly to give rise to an apparently diverging creep, see Sec.\ref{aging}. This does not happen in the quasi-elastic regime. The first was observed in~\cite{aging}, and the authors concluded reasonably that the system harbors a slow dynamics and is not quasi-static.
\subsection{The Steep Stress-Strain Trajectory\label{A Steep Stress Trajectory}}
As discussed above, in the quadratic regime of very slow shear rates, $T_g\propto |v_s|^2\to0$, the granular temperature is so small that the system is essentially elastic, moving from one elastic equilibrium state to a slightly different elastic one. This is the reason we call it {\em quasi-elastic}. Because $\sigma_s=\pi_s$ and $\partial_t u_s=\partial_t\varepsilon_s=v_s$, the change of the the shear stress $\sigma_s$ is well approximated by the (hyper-) elastic relation,
\begin{equation}\label{3a-1}
\partial_t\sigma_s=\frac{\partial\sigma_s}{\partial
u_s}\partial_t u_s =\frac{\partial\pi_s}{\partial
u_s}\partial_t\varepsilon_s=-\frac{\partial^2 w}{\partial
u_s^2}v_s.
\end{equation}
Shearing a granular medium at quasi-elastic rates, the result will be a trajectory $\sigma_s(\varepsilon_s)$ that is much steeper than in experiments at hypoplastic rates, such as observed during an approaching to the critical state. The gradient is given directly by the stiffness constant ${\partial^2 w}/{\partial u^2_s}$, and possibly three to four times as large as the average between loading and unloading at hypoplastic rates [because Eq~(\ref{2c-8}) lacks the factor of $(1-\alpha)$]. This goes on until the system reaches a yield surface of the elastic energy, say Eq.(\ref{2b-3}). We expect the system to form shear bands at this point, see Sec~\ref{nsb},\ref{sb}. The critical state will not be reached. Reversing the shear rate in between will retrace the function $\sigma_s(\varepsilon_s)$.
\subsection{Soft Springs versus Step Motors\label{soft springs}}
\begin{figure}[t]
\hspace{-1.5cm}
\includegraphics[scale=0.5]{stepmotor}
\caption{Why it is hard to observe the quasi-elastic regime if step motors are used, see text.\label{StepMotor}}
\end{figure}
Quasi-elastic behavior has not been observed in triaxial apparatus, even at the lowest rates. This maybe because they are simply not slow enough. But we suspect other
reasons: First, triaxial experiments are frequently performed with sand saturated in water, and squeezing water through the narrow gaps between grains is an efficient mean
of producing $T_g$. This may push the transition from the elastic to hypoplastic regime to much lower rates than in dry grains. second, the wide usage of step motors in
the triaxial appliances may have contributed to a wrong perception. Plotting the shear rate versus time, $ v_s(t)$, different shear rates are approximately given as depicted by
the two curves of Fig~\ref{StepMotor}. Although the curves have different average rates $\langle v_s\rangle$, the time-resolved, maximal rates $ v_s^{Max}$ are identical. And if
the time span of $ v_s^{Max}$ is long enough for $T_g$ to respond, and $ v_s^{Max}$ is high enough for the system to be in the linear regime, $T_g\propto v_s^{Max}$, the
system will display consecutive hypoplastic behavior in both cases, irrespective of the average rate $\langle v_s\rangle$.
We suggest two ways here to enter the quasi-elastic regime. Since a given slow stress rate has a high shear rate at elevated $T_g$ and a low one at vanishing $T_g$, the idea is to find the latter. One method is to slowly tilt an inclined plane supporting a layer of grains. In such a situation, the shear rate remains very small, and the system starts flowing only when a yield surface is breached. In contrast, employing a feedback loop in a triaxial apparatus to maintain a stress rate would not work well, because the correcting motion itself typically has strain rates that are too high.
A second method is to insert a very {soft spring}, even a rubber band, between the
granular medium and the device moving at a given velocity $v$ to deform it. If the spring is softer by a large factor $a$ than the granular medium, it will absorb most of the displacement, leaving the granular medium deforming at a rate smaller by
the same factor $a$ than without the spring. In other words, the soft spring serves as a ``stress reservoir'' for the granular medium. The same physics applies when the feedback loop is connected via a soft spring. Little $T_g$ is then excited, as in the experiment~\cite{aging}, see Sec~\ref{aging}.
\section{Conclusions}
This paper represents half a decade worth of attempts to come to terms, at least qualitatively, with the many observations of granular dynamics, by employing {\sc gsh} as the description and unifying framework. We are happy to report that it has not failed us once, although the outcome was rarely obvious when we started to examine a new experiment. Retrospectively, of course, circumstances appear much clearer and naturally systematic, and this is also how we present them above. The range of phenomena considered is clearly considerable, much wider than any macro-theory to date. Necessarily, a number of corollary predictions have also been made, especially in the context of wide shear bands and the quasi-elastic regime. They cry out for verification. Also, an observation of the difference between yield (or elastic instability) and the critical state would be highly desirable.
|
\section{INTRODUCTION}
\label{sec:intro}
The optical and infrared interferometry community has organized a series of meetings over the past decade on the topic of ``The Future of Optical Interferometry.'' At least 7 meetings held around the world (Li\`ege, Tucson, Lisbon, Socorro, Flagstaff, Turku) brought together astronomy topic experts and instrumentalists to imagine how a future long-baseline interferometer might be used to solve critical problems in stellar astronomy and extragalactic science. These efforts were buoyed by strong endorsements by Astro2010 Decadal Survey (in the USA) and Astronet (in Europe) of relevant key astronomy areas such as star and planet formation, exoplanets, active galactic nuclei and certain areas of fundamental stellar properties. These topics specifically will be revolutionized by the high angular resolution possible only from long-baseline interferometry. In a post-ELT and post-JWST era, interferometry will be essential to continue our current pace of discovery in astronomy.
While much debate at these meetings centered on how many telescopes at what size aperture would be needed for various science cases, it was clear that funding for any ambitious new facility would be difficult to obtain in the current budget climate. At these meetings, no clear path forwards emerged except to continue to make science-driven upgrades to current facilities and to better develop the interferometer observer community along with fresh science cases.
Last year, the European Interferometry Initiative (EII) ``Future of European Interferometry" subgroup helped plan another workshop titled ``Improving the performances of current optical interferometers and future design,'' held in September 2013 at the Observatoire de Haute Provence. Through a series of talks, rich dialogue between participants, and a final round table discussion, a new consensus emerged amongst the attendees\cite{surdej2014}. The community should focus on a {\em single compelling science case} and aggressively investigate the technical requirements needed to build a capable facility. This approach should lead to a more unified and coherent effort throughout the community to develop the technologies needed to make a next-generation facility feasible. With a clear set of top-level science requirements, a well-defined and justified technical roadmap can be created to guide priorities.
Many exciting science topics were discussed -- astrometry and characterization of exoplanets, imaging of active galactic nuclei, spectro-interferometry of mass-losing stars -- but the one that resonated most was the concept of directly imaging the key stages and processes involved in planet formation.
\section{Key Science Case for Planet Formation Imager (PFI)}
The goal of Planet Formation Imager (PFI) is to image all the key processes guiding planet formation around young stars and to follow early dynamical interactions between giant planets within the first 100 million years. Planet formation is one of the most important fields of astronomy, connecting star formation with exoplanets. A robust understanding of planet formation is needed to understand the demographics we see of exoplanet architectures around mature stars based on data from conventional planet finding efforts, such as radial velocity and transit surveys. Numerical hydrodynamicists as well as analytic theorists have taken up the challenge to develop a general theory of planet formation and the subsequent disk migration phenomena, efforts that include chemistry, dust formation and growth, radiative transfer, dust dynamics, magnetic field evolution in disks, viscous accretion disk theory, dynamics, and more. New high angular resolution facilities will accelerate discoveries, most especially through mm-wave imaging (ALMA), scattered light polarimetric imaging (GPI/SPHERE/SEEDS), high contrast coronagraphy (GPI/SPHERE/Subaru), and mid-IR interferometric imaging with MATISSE/VLTI. Furthermore, JWST and the ELTs will push the angular resolution of single-aperture techniques by another factor of 3 or so. Current trends point to a robust future for the fields of planet formation and exoplanet studies.
\begin{figure}
\begin{center}
\begin{tabular}{c}
\includegraphics[height=7cm]{almanaco.png}
\end{tabular}
\end{center}
\caption[example]
{ \label{fig:almanaco}
(left panel) This ALMA image of thermal emission from large grains around SAO~206462 shows evidence for a lopsided ring (from Perez et al. 2014; Ref. \citenum{perez2014}). (right panel) NACO/VLT imaging in polarized emission reveals even more details, including spiral structures extending into the inner disk (from Garufi et al. 2013; Ref. \citenum{garufi2013}). There is a clear need for angular resolution beyond what ALMA and even 30-m telescopes can bring. The Planet Formation Imager (PFI) will image disks with $<$1 milliarcsecond angular resolution. } \end{figure}
Figure~\ref{fig:almanaco} shows recent ALMA imaging and VLT/NACO scattered-light polarimetric imaging of Herbig Ae star SAO~206462. Even the short-baseline ALMA images\cite{perez2014} begin to resolve structure in the dust emission from this object, including a deficit of emission in the center and an asymmetric, elongated ring structure. VLT/NACO imaging\cite{garufi2013} with $\sim$50~milliarcsecond angular resolution clearly finds spiral arm structure in the polarized scattered-light emission from small grains in the upper layers of the disk. These disk structures are not straightforward to understand with current theories of planet formation. ALMA will achieve more than 10$\times$ better angular resolution soon, which will give us a new view into these questions.
Unfortunately, neither the full ALMA nor scattered light imaging with even the 30-m ELTs will suffice to resolve the important planet formation physical length scales for lower-mass T Tauri stars. Figure~\ref{fig:diskscales} shows the results of a radiative transfer simulation of a hydrodynamic calculation of multi-planet formation\cite{zhu2011}. We see that while current methods can probe the overall disk structure and can resolve large gaps or internal dust cavities, the accretion streams and circumplanetary accretion disks lie on spatial scales orders of magnitude smaller than ever achievable with ALMA or large single-aperture telescopes. We must look to a new project to probe planet formation on the actual scales that regulate planet formation.
The PFI Science Case is broader than only planet formation studies through imaging dust emission. Please see the accompanying SPIE paper by Stefan Kraus (in this volume) for further details on the science cases for PFI, including how PFI will image the dynamical relaxation of giant planet orbital configurations over the first 100 Myrs.
\begin{figure}
\begin{center}
\begin{tabular}{c}
\includegraphics[height=7cm]{DiskScales.png}
\end{tabular}
\end{center}
\caption[example]
{ \label{fig:diskscales}
Here we see a result from a radiative transfer calculation (Robin Dong, private communication) of a disk undergoing vigorous planet formation (as described by Zhu et al. 2011; Ref. \citenum{zhu2011}). For nearby star forming regions (d$\sim$100~pc), thermal infrared disk emission might come from regions of size $\sim$80~AU$\sim$ 0.8'', with sub-structures on a range of fine scales. We might expect gaps of size from 5~AU~$\sim$50~milliarcseconds with circumplanetary accretion disks with size close to the planetary Hill radius $\sim$0.03~AU$\sim$0.2 milliarcseconds. In order to cover all relevant scales of planet formation, PFI should reach down to sub-milliarcsecond resolution over about a 0.5" field of view with priority given to the thermal infrared. See additional discussion in Kraus et al. (this volume). } \end{figure}
\section{PFI Project Organization}
\label{sec:title}
The Planet Formation Imager (PFI) project was initiated in November 2013 by John Monnier, Jean-Philippe Berger, and Stefan Kraus, in conjunction with a Kick-off committee consisting of Mike Ireland, Lucas Labadie, Sylvestre Lacour, Jorg-Uwe Pott, Steve Ridgway, Jean Surdej, Theo ten Brummelaar, Peter Tuthill, and Gerard Van Belle (with recent additions of Romain Petrov and Chris Haniff). The goal of the PFI Project is currently to:
\begin{enumerate}
\item Formulate the science requirements and identify the key enabling technologies;
\item Build support in the broad astronomical community as well as the interferometry community;
\item Start lobbying with decision makers (ASTRONET, ESO, Decadal Review);
\item Prepare for upcoming funding opportunities (OPTICON, MSIPS)
\end{enumerate}
The PFI project has established a set of 3 executive officers and held internal elections. John Monnier was elected as Project Director, Stefan Kraus as Project Scientist, and David Buscher as Project Architect. The Project Director is responsible for overall organization of PFI and is the main point of contact for the Project. S/he will hold regular meetings with the advisory board (PAC, Project Advisory Committee), will lead regular telecom meetings, and be responsible for public outreach through emails, websites, and social media. The Project Scientist will lead the Science Working Group (SWG) to develop and prioritize key achievable science cases, including producing and maintaining the ``Top Level Science Requirements (TLSR)''. The Project Architect will lead the Technical Working Group (TWG) to conduct concept studies in order to identify key technologies and to develop a technology roadmap consistent with the TLSRs.
In May 2014, we published a public ``Call for Participation in PFI Concept Studies." It read as follows:
\begin{quote}
We are happy to announce the First Call for Participation in PFI Concept Studies. The ambitious goal of PFI is to image planet-forming disks in nearby star-forming regions with high enough spatial resolution to resolve the key physical processes at work, to witness planet formation live as it happens with $\sim$ 0.1AU resolution or better. Scientists from more than a dozen different institutes in six countries have begun planning for initial Concept Studies, an effort led by Project Director John Monnier (U. Michigan), Project Scientist Stefan Kraus (U. Exeter), and Project Architect David Buscher (U. Cambridge). Our top priorities for the next 12-24 months will be to define the most compelling areas of science to drive the instrument concept and at the same time determine feasible architectures for meeting the science goals. We seek contributions from the international astronomical community and invite participants to join the PFI Science Working Group or the Technical Working Group. For more information and to sign up to participate, please see http://www.planetformationimager.org (initial deadline to join working groups is June 16, 2014).
\end{quote}
At the time of this writing, over 60 volunteers have expressed interest in joining the TWG and a similar number for the SWG. We are currently in the process of organizing subgroup leaders and will begin Project work during the Summer 2014.
One open issue is the timescale for building PFI. We hope to clarify this issue within 12-24 months following the work of the Science and Technical Working Groups. Realistically, most of us involved in PFI recognize that while the ambitious goals of PFI are definitely achievable, they will not be easy to meet with current technologies at a modest price point. However, we are confident that with a clear and strong design reference mission and a well-defined technology roadmap, a feasible timescale for PFI will be one that envisions deployment sometime after the Extremely Large Telescope (ELT) projects are fully operational in the late 2020s.
\section{Technical Considerations}
Alongside this article is a PFI Science Case paper by Kraus et al. that gives an overview of the priority science cases and the chief considerations guiding the top-level science requirements. Here, we will focus on technical considerations for PFI, including initial thoughts on the core science requirements, possible facility architectures, site considerations, and some crucial technologies.
\subsection{Top-Level Science Requirements}
In order to guide our architecture discussion, we must adopt a preliminary set of top-level science requirements. Based on initial work by the PFI Project, we adopt the following TLSRs for the purposes of this article:
\begin{itemize}
\item The PFI angular resolution is set to resolve the region around an embedded giant exoplanet in the disk where the planet dominates the gravitational influences, the so-called ``Hill Sphere.'' For a one Jupiter-mass planet at 1 AU around a sun-like star, we would require multiple resolution elements across 0.1~AU. We adopt a requirement of 0.03~AU resolution for distances up to 140pc (distance to Taurus star forming region), thus PFI requires 0.2 milli-arcsecond resolution.
\item PFI should be sensitive to thermal dust emission over a range of temperatures to probe the region where giant planets are thought to form and to eventually migrate. Here, we focus on the mid-infrared ($\lambda\sim10\mu$m) where dust grains at 300K are peaking in their thermal emission, giving us good sensitivity to the terrestrial planet forming zone.
\item In order to have 0.2 mas resolution at 10$\mu$m, PFI will require 10km baselines.
\item PFI should be sensitive enough to observe a range of structures in the disk. For instance (for d$=$100pc):
\begin{itemize}
\item T Tauri star photosphere: N(mag)$=$7.5
\item Best case circumplanetary disk (a la LkCa15 object\cite{kraus2012}): N(mag)$=$11
\item 10~Myr old Jupiter: N(mag)$=$15.7
\item 100~MYr old Jupiter: N(mag)$=$18.5
\item For reference, the current VLTI/MIDI with fringe tracking on the UTs can observe N(mag)$=$7.0
\end{itemize}
\item Lastly, PFI will need excellent imaging capabilities since we expect disk emission to be complicated (as shown by ALMA and scattered light imaging for Herbig Ae/Be stars). If we aim for images with 400$\times$400 pixel imaging, then we must have similar imaging to the radio VLA. PFI will likely need 20--30 telescopes in the array.
\end{itemize}
\subsection{Architecture Overview}
We consider a number of architectures below that could satisfy the above TLSRs. Note that many of the conclusions below are highly preliminary and will be critically scrutinized by the SWG and TWG over the coming 12-24 months using facility simulations. It is beyond the scope of this brief overview paper to justify all the claims below and we include specific numbers here just to give a snapshot of the PFI Project's current thinking before the detailed studies begin in earnest later in 2014.
\begin{enumerate}
\item {\bf Conventional ground-based interferometer array.} This system will focus on mid-infrared wavelengths (5--13$\mu$m) with 10 km baselines. This will require multi-km vacuum pipes with diameter $>$0.5m to minimize fresnel diffraction. For conventional sites and using $D<10$m telescopes, N-band fringe tracking on the $N=7.5$ central star will not be possible. We must use a near-infrared fringe tracker to allow coherent integration in the N-band, similar to what is now done for VLTI/UTs\cite{muller2014}. Natural guide star adaptive optics will be sufficient on $\sim$4-m class telescopes to allow both fringe tracking and mid-IR sensitivity, assuming $\sim$30~sec coherent integrations are possible. For imaging complex scenes and to have sufficient short baselines for bootstrapping, we will need $N>20$ telescopes. The $\sim$30~sec coherent integrations will allow the star to be detected at 10$\mu$m, but much longer bispectrum integrations will be needed to reach $N=15$ -- approximately 20 hours for twenty 4-m telescopes. This will allow individual giant planets to be detected around stars up to 10 Myrs. Note that this architecture could allow near-infrared imaging with $\sim$50~{\em micro}-arcsecond angular resolution down to H(mag)$=$14 in principle, supporting robust ``ancillary'' science cases in stellar astrophysics, AGN, microlensing, and more. Future detailed simulations will determine the minimum-sized aperture required for our core science -- we emphasize the above SNR calculation for 4-m apertures is merely illustrative.
\item {\bf Heterodyne array.} This system would be a mid-infrared heterodyne system similar to an expanded Infrared Spatial Interferometer (ISI)\cite{hale2000}. While the ISI was limited to very bright sources [N(mag)$<$0], new technological advances\cite{lecoarer2014} could drastically improve sensitivity as a possible PFI architecture. Ireland et al. (this volume) discuss how advances in frequency combs could allow the entire mid-IR bandwidth to be used, an increase over the ISI bandwidth of $>$1000. The heterodyne detection process at 10$\mu$m is equivalent to observing with a thermal background of 1400K, a factor many times higher than the 273K thermal background that a direct detection scheme suffers. However, the heterodyne detection process can be done closer to the telescopes with far fewer reflections, thus obtaining a transmission bonus. In addition, the heterodyne system allows amplifications of the RF signal following detection without introducing additional noise, thus all telescope combinations could be correlated without loss of SNR -- an advantage over conventional ``direct detection'' schemes. The heterodyne array will still require near-infrared fringe tracking using AO-corrected 4-m class telescopes in order to allow coherent integration in the mid-IR. While the mid-IR delays can be done in a computer following frequency down conversion (without expensive large vacuum pipe transport), conventional delay lines will still be needed for the near-IR fringe tracking channel (although perhaps only for relatively short bootstrapping baselines). The balance of advantages and disadvantages will have to be carefully considered by the Technical Working Group. One exciting possibility is that fiber optics could be used to transport near-infrared light from the telescopes to the beam combining lab which would obviate the need for long vacuum pipes connecting the telescopes to the lab. Please see the contribution by Ireland et al. (this volume) for detailed considerations on the intriguing possibility of a mid-IR Heterodyne interferometry revival.
\item {\bf Space interferometer.} The mid-infrared thermal background in space is about {\em 26 million times} lower than from the ground. It is impressive to consider that a 1mm aperture cooled space telescope is {\em in principle} as sensitive as a 8m ground-based telescope for point source detection in the thermal IR. A space interferometer will need 10km baselines just as the other designs, thus requiring long-distance formation flying capabilities not dissimilar from those being developed for the gravity wave interferometer LISA and with some technology connections with the formation flying maneuvers over short distances that are now routine for remote docking of cargo ships to the International Space Station (see also Ref. \citenum{lecoroller2012}). One huge advantage of a space platform is that PFI would have access to imaging across the thermal IR, from the mid- to far-IR.
This would open up new science cases unthinkable from the ground, including studies of ices, mineralogy, debris disks, etc. It is encouraging that the NASA's 2013 Strategic
Astrophysics Roadmap highlighted space interferometry
as an essential technology in the ``Visionary Era'' in optical, IR, and even X-ray astronomy.
While much work has been done studying infrared space interferometry in the context of detecting terrestrial planets (TPFI\cite{lay2007}, Darwin\cite{leger2007}), the sensitivity and nulling requirements for PFI are more relaxed than for these earlier planning efforts. Indeed, while imaging disks will be easier in some ways compared to imaging Earth-like planets, space-PFI will require many more telescopes and reconfigurations than TPF/Darwin to allow high fidelity imaging of complex scenes. Space-PFI may have more in common with the ``Stellar Imager\cite{carpenter2003}'' project than Terrestrial Planet Finder. Thinking ahead, space-PFI will be a natural follow-on to JWST and will afford a robust extragalactic science case in addition to the Planet Formation Imaging science. We note that the far-IR space community is already looking into detailed studies of space interferometers (FISICA, SPIRIT/SPECS\cite{leisawitz2008}), although generally with far fewer telescopes and over modest $B<100$m baselines. We share many technology requirements and hope to coordinate our efforts to invest in space interferometry technology studies.
\item {\bf ALMA with longer baselines.} ALMA is already beginning to do breakthrough imaging of dust continuum and molecules in disks. Eventually ALMA will have 10-20$\times$ greater angular resolution when the 350$\mu$m band observations are extended to the full 16km baselines, yielding an angular resolution of $\sim$5 milliarcseconds. This will allow AU-sized structures to be imaged in nearby star-forming regions, an incredibly exciting prospect for learning about planet formation. While there is no more space on the high plateau site, it might be possible to extend ALMA baselines by connecting with new telescopes built on nearby plateaus to boost resolution to the sub-milliarcsecond level. There is some effort to identify possible locations (e.g., the Long Latin America Millimeter Array LLAMA\cite{arnal2011}). The advantage of extending an existing successful facility is manifest although the science potential may not be as great as for a shorter wavelength facility. For instance, it might not be possible to detect exoplanets themselves in the mm/sub-mm and also the thermal emission from small grains will be largely invisible in the sub-mm. It is clear that a mid-IR PFI would be a powerful complement to the mm-wave ALMA and the Science Working Group will look into scientific trade-offs of observing at different wavelengths.
\item{\bf Ground-based coronagraph.} The ``Planet Formation Imager'' moniker was first attached to a 2nd generation Thirty-Meter Telescope instrument concept\cite{vasisht2006}. Indeed, a visible-light extreme AO on a 30-m telescope will have $\sim$4~milliarcsecond resolution, similar to sub-mm ALMA. However, for the same reasons as above, 4~mas is not sufficient to resolve a 1-Hill-Sphere gap in a young disk or to isolate a circumplanetary disk from its surrounding material. That said, an ELT-PFI will be a powerful complement to mid-IR PFI by probing scattered light that is sensitive to small dust particles in the upper layers of the disk. As for ALMA, the SWG will scrutinize the trade-offs between the baseline PFI concept of a mid-IR, sub-mas resolution array and a 30-m coronagraph instrument. Furthermore, a 3--5$\mu$m imaging system on an ELT might be sufficient to detect circumplanetary accretion disks and young exoplanets, either complementing PFI or perhaps moving PFI away from this science toward the more challenging (and rewarding) thermal IR dust imaging core science case.
\item{\bf Space occulter.} Advances in starshade technology in the context of exoplanet discovery\cite{kasdin2012}
suggest new possibilities for space-based imaging systems. Angular resolution can be obtained through occultations of astronomical bodies by large shades in a way similar to classic lunar occultation work\cite{ridgway1977}. Fresnel interference sets the resolution to be proportional to $\theta\propto\sqrt{\frac{\lambda}{D}}$, which unfortunately means that a 30AU separation is required between a 10km occulting body and the distant collecting telescope to obtain milli-arcsecond resolution at $\lambda=$10$\mu$m. This does not seem a feasible prospect, even using asteroids as natural occulters\cite{dunham2013}.
\end{enumerate}
\subsection{Sites}
Optical and infrared interferometers typically require the very best sites although adaptive optics and fringe tracking on relative bright targets might relax these conditions. In addition to considering conventional mid-latitude sites, the PFI project will seriously investigate the unusual attributes of the Antarctic High Plateau Sites (Dome A, C, F) which offer immense promise for both thermal infrared observing (low background) and excellent atmospheric stability\cite{lawrence2004} that allow for much longer coherence times and might allow for novel off-axis fringe tracking. Some aspects of the PFI core science might be done in Antarctica with a specialized few-telescope system, for instance searching for hot exoplanets and accreting circumplanetary disks in the 3--5$\mu$m range -- similar to the ELT-PFI case discussed in the last section.
The core science of PFI requires a large number of young stars with disks for study. This limits us to considerations of the nearest young associations with active planet formation. Any site selection will need to account for the accessibility to the well-known Taurus, Orion, and Ophiuchus regions. In addition, PFI would like to study the evolving characteristics of cooling gas giant planets in stars found in young clusters over a range of ages ($<$100~Myrs). The SWG will collect a comprehensive database of nearby star-forming regions to guide these deliberations.
\subsection{Technology Spotlights}
The Technical Working Group will also identify key technologies that could prove decisive to making PFI feasible.
Although we still have to confirm that a 10-km array of 20--30 telescopes with 4 to 8\,m apertures will give the required sensitivity, angular resolution, dynamic range, and sensitivity to dynamics, there are no known technology showstoppers to PFI. Many believe we could build a conventional ground-based interferometer version of PFI with today's technology (e.g., Buscher et al, these proceedings). If PFI can not be built until after the ELTs, then there is time to invest in new technologies that might significantly reduce the cost and risk of a PFI facility -- these efforts may prove crucial.
Here, we highlight a few technologies that could have a large influence on the direction of PFI moving forward.
\begin{itemize}
\item {\bf Inexpensive 4-m to 8-m class telescopes.} While there is some skepticism within our Project that major advances are possible in this area, we believe the community should focus on developing breakthroughs in this area. PFI will greatly benefit from large numbers of telescopes and thus unit costs should be minimized. There have been advances in carbon fiber reinforced polymer mirrors\cite{andrews2010} that, when coupled with inexpensive low-order adaptive optics, could radically change our telescope designs. For PFI, the mirror specifications can be degraded compared to typical optical systems -- PFI will focus on the infrared only, the telescopes will naturally incorporate adaptive optics, and they also will require only a narrow field-of-view. Scattered light off high-order surface imperfections might reduce throughput but will not ruin the imaging ability for a single-mode system like PFI. There were new ideas presented in 3d printing, active optics, and more on display at this meeting in Montreal. Current interferometer groups might consider partnering with industry to field-test these technologies over the coming decade.
\item {\bf Frequency combs.} The heterodyne concept will require mid-IR frequency combs that do not yet exist.
\item{\bf Fiber optics.} The use of single mode fibers for beam transport might be crucial for the heterodyne architecture. Much work has been done in this area for the OHANA\cite{perrin2003} project and new lab results were recently presented at this meeting (Anderson et al.).
\item {\bf Mid-IR photonics.} For the direct detection scheme, we will need efficient mid-IR combiners that can combine large numbers of baselines simultaneously. This work is just beginning with new materials and some progress was reported at this meeting (see, e.g., Cvetojevic et al., Martiarena et al., Minardi et al. )
\item{\bf Space technology.} Any progress towards space interferometry technology demonstrations is welcome and sorely needed. There is renewed interest in nano- or micro-sat flight opportunities to test ideas in formation flying, fiber nulling combination (Lacour et al., this meeting), and more.
\end{itemize}
Other interesting technologies include inexpensive laser guide stars, realistic 10km delay lines (super low-loss multiple reflections from LIGO legacy, ``ultra-flat'' zero dispersion fiber optics), fringe tracking (demonstrate 30 second coherencing, perform multi-hour bispectrum integrations), novel combiners for large numbers (N$>$20) of beams (both near-IR and mid-IR, densified pupil coronagraph), 40$\mu$m heterodyne detectors for dome A (could avoid near-IR fringe tracking possibly), new operations models (take advantage of robotic telescopes and large redundancy to slash operating costs).
\subsection{Intermediate Milestones on the Roadmap}
There are possible intermediate facilities that might demonstrate PFI technologies while advancing PFI science goals.
\begin{itemize}
\item L/M band imaging using J/H/K band coherencing: VLTI-MATISSE at 5--10$\mu$m (this project is funded with results expected by 2020), L/M band modes for CHARA/MROI (not funded at the moment)
\item Imaging NIR scattered-light disks in polarized emission with ``ALMA resolution'' (5~mas): NIR polarization+nulling instrument at VLTI -- a kind of high-resolution GPI/SPHERE.
\item Antarctica High Plateau interferometer: 1--5$\mu$m optimized for exoplanets.
\end{itemize}
\section{Conclusions}
While some argue that ``interferometry is inevitable'' (see Rinehart paper, these proceedings), there is no guarantee that the prodigious pace of discovery we currently enjoy in Astronomy is sustainable. The PFI Project hopes to identify and specify compelling science goals to guide our thinking into the post-ELT era. This is the first SPIE meeting for PFI and we hope to see maturing of the concept over the next few years. PFI is a long-term goal requiring coherent efforts over many years, possibly decades, to reach fruition. We welcome new participation and encourage interested readers to visit {\bf planetformationimager.org} to get involved in one of the working groups.
\acknowledgments
The PFI Project wishes to thank the organizers of the 2013 workshop at the Observatoire de Haute Provence (OHP), where PFI was born. JDM especially wishes to think EII for travel support that allowed him to attend the OHP meeting.
|
\section{Introduction}
Over the past decade the cosmological standard model based on the $\Lambda$CDM paradigm has been firmly established and tested using a variety of data.
One of the fundamental parameters of the $\Lambda$CDM model is the present expansion rate of the universe, $H_0$, also known as the Hubble constant.
The value of this parameter can either be inferred indirectly from e.g.\ measurements of the cosmic microwave background at high redshift or from large scale
structure surveys at intermediate or low redshift, or it can be measured directly by measuring velocities and distances to standard candles in the nearby universe.
Both methods have been used successfully in the past, and the precision of both methods is now at a level where the agreement between the two types of measurements is a constraining consistency test of the underlying model. Both methods have advantages and disadvantages: The direct measurement is very model independent, but prone to systematics related to local flows and the standard candle assumption. On the other hand, the indirect method is very robust and precise, but relies completely on the underlying model to be correct.
Any disagreement between the two types of measurements could in principle point to a problem with the underlying $\Lambda$CDM model.
The recently released data from the Planck satellite seems to point to a relatively low value of the Hubble constant, while direct measurements relying mainly on Type Ia supernovae clearly favour a higher value, with the two being inconsistent at the $2.4\sigma$ level \cite{Riess:2011, Planck:2013}.
There are several possible explanations for the disagreement between the two estimates of the expansion rate. It could be caused by a problem with the assumed cosmological model, or it could be that one (or both) of the estimates is either inaccurate or biased. The second possibility has recently been considered by Estathiou in \cite{Efstathiou:2013} and by Clarkson et al. in \cite{Clarkson:2014}. Using an improved distance calibration, Estathiou has re-analysed the data from \cite{Riess:2011} and found a lower value for the local Hubble parameter, which decreases the tension between the two estimates. Ref.~\cite{Clarkson:2014} consider relativistic corrections to the distance to the CMB last scattering surface, and show that second order lensing corrections can possibly increase this distance by several percent, which in turn causes an increase in the best fit value of the Hubble parameter from the Planck data.
Yet another possibility - the one that is tested in this paper - is that it is merely a result of the spatial variation in the expansion rate of our universe.\\
\\
\noindent
For a given cosmological model, the expansion rates estimated by observers at different locations are expected to vary according to some underlying distribution. To gain knowledge of the spread of this distribution, we perform N-body simulations of the model in question. Subsequently, structures are identified using a halo finder that generates a catalogue containing the masses, positions, velocities and substructures of all halos found in each simulation. From the halo catalogues, lists of observers are selected, under the assumption that the standard observer reside in a halo similar to the Local Group. Each observer is assumed to estimate the local Hubble parameter in his neighbourhood by measuring the distances and velocities of halos located in a sphere around him, since this is where we would expect supernovae to be found. Finally, confidence intervals are constructed from the local Hubble parameter calculated by each of these observers. This is done in order to answer the question: Could the discrepancy between the inferred and the measured Hubble constants be due to cosmic variance?
The same question has been addressed in \cite{Marra:2013} and \cite{Wojtak:2013}. In \cite{Marra:2013}, Marra et al. calculate the variation of local Hubble parameters based on the power spectrum which expresses the variations of the underlying density field, and this in turn affects the velocity field and hence the local Hubble parameters. In \cite{Wojtak:2013}, Wojtak et al. use the same approach as in this paper, i.e. N-body simulations, to estimate the spread in the local Hubble parameters. In their paper, they focus primarily on how the local Hubble parameters are affected by observer positions. While this is also discussed in the present paper, we furthermore investigate how the apparent variation is affected by only observing a fraction of the sky, and whether the cosmic evolution that takes place between the time that light is emitted by a supernova and observed by us can have a significant effect.
We finally note that measurements of variations in the Hubble parameter using type Ia supernovae is equivalent to measurements of the local velocity field using the same tracers (see e.g.\ \cite{Haugboelle:2006uc,Ma:2013oja,Feindt:2013pma,Turnbull:2011ty,Davis:2010jq,Weyant:2011hs}).
\\
\noindent
The paper is organized as follows: In section 2, we describe the N-body-simulations that our analysis is based upon and the halo finder used to identify structure. In section 3, the calculation of the local Hubble parameters is outlined, and the selection of observers and of observed halos is motivated and described. In section 4, we present the outcome of the analysis and finally we discuss the results in section 5.
\section{Method}
Below the N-body simulations and identification of halos are described in detail.
\subsection{Simulations}
The best fit cosmological parameters determined by the Planck collaboration, as specified in the last column of table 5 in \cite{Planck:2013}, are used as the basis for N-body simulations of the universe. The N-body simulations are performed using a modified version of the GADGET-2 code \cite{Springel:2005}, with initial conditions generated using a code written by J.~Brandbyge \cite{Brandbyge:2006} based on transfer functions computed using CAMB \cite{Lewis:2002}. Specifically, the transfer functions are calculated with $(\Omega_b,\Omega_{CDM})=(0.048,0.26)$, whereas only cold dark matter is used in the N-body simulations. A flat universe is assumed, and $(h,\sigma_8)=(0.68,0.84)$. The simulations are run from a redshift of $z=50$ until $z=0$.
The reference simulation is done in a box of sidelength $512 \di{Mpc/h}$ with $512^3$ particles. We perform two simulations of this size and resolution, but with different seeds for the random number generator that is used to construct the initial conditions, in order to determine how much the results can be expected to vary between different occurrences of a universe with the same parameters. In order to check for numerical effects due to the simulation resolution and the simulation volume, we also perform two simulations with $1024^3$ dark matter particles in boxes of $512 \di{Mpc/h}$ and $1024 \di{Mpc/h}$, respectively. By comparing the results obtained from the standard simulation with the results obtained using a bigger box and a greater resolution, we conclude that this does not significantly affect the results, and that the chosen box size and resolution are sufficient to make the desired inquiries. Still, all the simulations are of much smaller volume than the ones that are the basis for the analysis performed by Wojtak et al. in \cite{Wojtak:2013}, which are of several $\di{Gpc/h}$, while the mass resolution is similar. The sample on which they base their analysis of the Hubble flow is therefore substantially larger than the ones used in this paper, and as a result the analyses presented here will have a greater sample noise and a greater correlation between the observers. The latter is to some extend mitigated by using several boxes with independent random seeds. We identify 600 observers in each simulation, and use these to estimate of the mean and spread of the Hubble constant.
The observers in the simulations are spread across the entire simulation volume. In order for each of them to be able to observe in all directions, the box is taken to be periodic. To avoid an observer considering the same halo twice in his calculation of the Hubble constant, the maximal observed distance should be half a box length, that is $256\di{Mpc/h}$, corresponding to a redshift of $z = 0.087$. This corresponds approximately to the greatest observed distance of $z \sim 0.1$ used in supernova surveys, \cite{Riess:2011}. As in \cite{Riess:2011}, in order to reduce the effect of the local, coherent flow, no halos closer than $30\di{Mpc/h}$, corresponding to $z = 0.01$, are used.
\subsection{Identifying halos}
To identify bound structures in the simulations described above, we make use of the halo finder ROCKSTAR \cite{Behroozi:2011,Manual:Rockstar} which identifies halos using a variant of the Friends-of-Friends (FOF) algorithm. At first, particles that are close together in space are grouped together using the FOF algorithm, which assembles particles that are within a specified distance of each other. Within each FOF-group, a measure of the phase-space distance between two particles $p_1$ and $p_2$ (with positions $\vec{x_1}$, $\vec{x_2}$ and velocities $\vec{v_1}$, $\vec{v_2}$, respectively) is defined as $d(p_1,p_2) = \sqrt{\frac{\fabs{\vec{x_1}-\vec{x_2}}^2}{\sigma_x^2}+\frac{\fabs{\vec{v_1}-\vec{v_2}}^2}{\sigma_v^2}}$, where $\sigma_x^2$ and $\sigma_v^2$ are the variances of the particle positions and velocities within the group. This metric combines distance in position and velocity into a single measure and, using this, locations in which the mean phase-space-distance between particles is low are identified as local maxima of phase-space density. These maxima are used as seed halos, and all particles in the original FOF-group are assigned to the seed that is closest in phase-space. At last, halos located in more massive hosts are categorized as subhalos, unbound particles are removed, and halo masses are calculated as spherical overdensities. For each halo, the position of the halo centre is calculated as the mean of the positions of the central particles, and the halo velocity is calculated as the mean of the velocities of the innermost 10\% of the halo particles.
\section{Analysis}
As mentioned, it is assumed that each observer estimates the local Hubble parameter, $H_{loc}$, by measuring velocities and positions of halos in the close vicinity. The apparent velocities of observed halos contain two components: one from the expansion of space, and one from peculiar motion. Assuming a pure Hubble flow, corresponding to no peculiar motion, the radial velocities would be given by Hubble's law:
\begin{align}
v = H_0 r,
\end{align}
where $r$ is the radial distance to the halo and $H_0$ is the global Hubble constant. Assuming that the observers cannot determine how large the peculiar component of the radial velocity is, they fit the distances and velocities of the halos around them to Hubble's law, using a least squares estimate, hereby getting a value for the local Hubble constant:
\begin{align}
H_{loc}=\frac{\bar{r}\bar{v}}{\bar{r^2}},
\end{align}
where a bar denotes the mean and $v$ is the absolute velocity of the halo away from the observer, meaning that each observer has transformed to the CMB rest frame.
In actual observations, the position and velocity information of distant halos is primarily obtained from supernovae. In \citep{Riess:2011}, 240 supernovae are used in estimating the local Hubble parameter, and for this reason we choose 240 observed halos for each of the observers in our mock survey. This is done by assuming that the probability of a supernova occurring in a given halo is proportional to the halo mass, and therefore making a mass-weighted selection of observed halos (this is in contrast to what is done in \cite{Wojtak:2013}, where every halo within a given distance is used). This procedure results in a redshift distribution of the observed halos which is peaked at a higher value than the one used in \cite{Riess:2011}, as shown in figure \ref{fig:redshiftdistribution}. In consequence we expect to slightly underestimate the variance of the local Hubble parameter at large distances.
There are some indications \cite{Perrett:2012,Rodney:2014} that the rate of type Ia supernovae not only depends on the stellar mass of a galaxy -- or more precisely the star formation history -- but there also exist a prompt component in the type Ia distribution correlated with the instantaneous star formation rate of the galaxy. Given the current uncertainty in the fraction of prompt type Ia supernovae, this has not been taken into account in the selection of halos. Instead we assume a direct proportionality between the type Ia supernovae rate, stellar mass, and halo mass.
When 240 observed halos have been picked for a given observer, they are distributed into bins according to their distance, with each bin characterized by the maximum distance of the observed halos it contains. The bin distances, which will henceforth be denoted by $r_{max}$, ranges from $67\di{Mpc/h}$ to $256\di{Mpc/h}$. The local Hubble constants are then estimated first using only the innermost bin, then the two innermost bins and so forth, until all the bins are included.
\begin{figure}
\center
\includegraphics[width = 0.49\textwidth]{redshiftdistribution.pdf}
\includegraphics[width = 0.49\textwidth]{hubblediagram.pdf}
\caption{Left: Comparison between the redshift distribution measured by a typical observer in the mock catalogues (blue) and the redshift distribution of supernovae from Hicken et al., \cite{Hicken:2009a} (green), which closely resembles the one used by Riess et al. in \cite{Riess:2011}. Right: Hubble diagram obtained by the same mock observer.}
\label{fig:redshiftdistribution}
\end{figure}
\subsection{Objectives}
In order to determine how the variance of the local Hubble constants depend on various aspects of the observations, the observed halos are selected and handled in different manners, as is the choice of observer positions.
Furthermore, effects of simulation seed, resolution and volume are checked for, and we additionally study how the number of observed halos affect the variance.
Below we describe each of these analyses in more detail. In figure \ref{fig:refsheet} a few of the analyses are sketched for visualisation, and an overview is given in table \ref{tab:parameters}.
\begin{table}
\center
\begin{tabular}{cccccc}
Name & Nsim & Box [Mpc/h] & Survey geometry & Observers & Past lightcones \\
\hline
A.0 & 512 & 512 & Full sky & Random in space & No \\
A.1 & 512 & 512 & Full sky & Random in halos & No \\
A.2 & 512 & 512 & Full sky & Local Group like & No \\
A.3 & 512 & 512 & One cone & Local Group like & No \\
A.4 & 512 & 512 & Two cones & Local Group like & No \\
A.5 & 512 & 512 & Full sky & Local Group like & Yes \\
B & 512 & 512 & Full sky & Local Group like & No \\
C & 1024 & 512 & Full sky & Local Group like & No \\
D & 1024 & 1024 & Full sky & Local Group like & No \\
\end{tabular}
\caption{Four different simulations have been run, named A-D. The number of particles in each simulation is Nsim$^3$. For the standard simulation (A), six different analyses have been performed, named A.0-A.5. Simulation B has the same characteristics as A, but with a different seed for the random number generator. All simulations are performed with the best fit parameters from the Planck satellite: $(\Omega_b,\Omega_{CDM},h,\sigma_8)=(0.048,0.26,0.68,0.84)$.}
\label{tab:parameters}
\end{table}
\paragraph{Choice of observers:}
The standard observer is chosen from the halo catalogues as a subhalo of mass $10^{12}-10^{13} M_{\odot}/\di{h}$ in a host halo with mass $5\cdot 10^{14}-5\cdot 10^{15} M_{\odot}/\di{h}$, since this approximately corresponds to our position in the Local Group galaxy cluster that resides in the Virgo Super Cluster. The significance of the observer positions is checked by using two alternative selections: One in which the observer positions consist of positions chosen randomly in the simulation volume, and one in which the observer positions are chosen randomly among all the halos in the simulation.
\begin{figure}
\center
\includegraphics{haloplots.png}
\caption{Positions of halos in a slap of thickness $10\di{Mpc/h}$, with illustrations of the observation performed by each observer (green star), with values of $r_{max}$ shown as green circles (or arcs). For clarity, only every 5th value of $r_{max}$ is shown. }
\label{fig:refsheet}
\end{figure}
\paragraph{Survey geometry:}
The actual supernovae used in estimating the expansion rate of the local universe are not distributed across the entire sky. In some directions, our sight is blocked, for example in the plane of the Milky Way. To account for this in the analysis, at first only a small patch of the sky is observed. This patch is subsequently enlarged until it covers all of the sky. The number of supernovae is the same no matter if the whole sky or only a fraction is observed. We perform this analysis both with one cone, and with two cones pointing in opposite directions.
\paragraph{Cosmic evolution:}
Observing out to a redshift of $z\sim 0.1$, corresponds to observing 1.3 billion years back in time. When doing the same "observation" in the output of our N-body simulations, the time it takes light to travel from the distant halos is ignored, and therefore so are the changes in the growth and structure of the halos. We investigate the significance of this effect by using a locally developed plugin to GADGET-2 that during the course of the simulation reconstructs the past lightcone of any observer in the box, dumping it to a snapshot on disk. Such snapshots are created for the positions of our chosen observers. Then the halo finder is applied to these past lightcone snapshots, using a functionality in ROCKSTAR in which the change in cosmological parameters as one looks backwards in time is taken into account in the analysis. Constructing the mock observations using past lightcones takes into account the evolution in the cosmological model, and more accurately reproduces the actual observations.
\FloatBarrier
\section{Results}
In Figures \ref{fig:Planck512}-\ref{fig:Planck512_lightcone} we show how the mean of $H_{loc}/H_0$ among the chosen observers depend on the number of observed halos and the maximum observed distances for different choices of observers and observed halos. Confidence intervals are shown as coloured bands around the mean (solid line). The chosen confidence levels are 68.3, 95.4 and 99.7 per cent, all calculated symmetrically so that equal fractions of the locally measured Hubble constants fall below and above the interval. For comparison, the results from the standard analysis (on the right in figure \ref{fig:Planck512}) are shown as punctuated lines in all other plots.\\
\\
\noindent
The first plot in figure \ref{fig:Planck512_varSN} shows $H_{loc}/H_0$ as a function of the number of observed halos. In figure \ref{fig:Planck1024}, the analysis is repeated for two simulations with $1024^3$ particles, in boxes of sidelength $512\di{Mpc/h}$ and $1024\di{Mpc/h}$, respectively. This is done order to check that the simulation resolution and simulation volume have no significant effect on the results.
The set of parameters used in the N-body simulations are the ones published by the Planck collaboration:
$(\Omega_b,\Omega_{CDM},h,\sigma_8)=(0.048,0.26,0.68,0.84)$.
In figure \ref{fig:Planck512_RandPos} we show how the observer positions affect the measured values of $H_{loc}$ (on the left). On the right, the observers are chosen randomly among all the halos in the simulation. In figure \ref{fig:Planck512_lightcone}, the observations are performed using past lightcones of the Local Group like observers, so that cosmic evolution from the time that light was emitted from an observed halo is taken into account. In figure \ref{fig:Planck512_skyfraction} we show how the width of the 68.3 per cent confidence interval varies as a function of both the distance $r_{max}$ and the covered percentage of the sky, and in figure \ref{fig:Planck512_skyfraction_256} the width of the confidence interval is plottet as a function of the observed skyfraction at $r_{max}=256\di{Mpc/h}$. The results are summarized in table \ref{tab:results}.
\begin{figure}
\center
\includegraphics[width = 0.49\textwidth]{Planck512_varSN.pdf}
\includegraphics[width = 0.49\textwidth]{Planck512.pdf}
\caption{Left: Mean and scatter of the local Hubble parameter as a function of the number of observed halos. The largest observed distance is held constant at $256\di{Mpc/h}$, while the number of observed halos is varied from 2 to 200. Right: Mean and scatter of the local Hubble parameter as a function of the maximum observed distance. The chosen observers are Local Group like halos, each estimating the local Hubble constant by measuring the distances and radial velocities of 240 supernovae in nearby halos. The simulation consists of $512^3$ particles in a periodic box of sidelength $512\di{Mpc/h}$.}
\label{fig:Planck512}
\label{fig:Planck512_varSN}
\end{figure}
\begin{figure}[htb!]
\center
\includegraphics[width = 0.49\textwidth]{Planck1024.pdf}
\includegraphics[width = 0.49\textwidth]{Planck1024_box1024.pdf}
\caption{Left: Mean and scatter of the local Hubble parameter as a function of the maximum observed distance, calculated for a simulation with $1024^3$ particles in a box of $512 \di{Mpc/h}$. Right: Calculated for a simulation with $1024^3$ particles in a box of $1024 \di{Mpc/h}$. Punctuated lines indicate the result from the standard analysis.}
\label{fig:Planck1024}
\label{fig:Planck1024_box1024}
\end{figure}
\begin{figure}[htb!]
\center
\includegraphics[width = 0.49\textwidth]{Planck512_RandPos.pdf}
\includegraphics[width = 0.49\textwidth]{Planck512_RandHalos.pdf}
\caption{Left: Mean and scatter of the local Hubble parameter as a function of the maximum observed distance, when the chosen observers are distributed randomly in the simulation volume. Right: Observers chosen randomly among all the halos in the simulation. Punctuated lines indicate the result from the standard analysis.}
\label{fig:Planck512_RandPos}
\label{fig:Planck512_RandHalos}
\end{figure}
\begin{figure}[htb!]
\center
\includegraphics[width = \textwidth]{Planck512_lightcone.pdf}
\caption{Mean and scatter of the local Hubble parameter as a function of the maximum observed distance calculated for Local Group like halos, with the effect of cosmic evolution taken into account by using lightcones. Punctuated lines indicate the result from the standard analysis.}
\label{fig:Planck512_lightcone}
\end{figure}
\begin{figure}[htb!]
\center
\includegraphics[width = \textwidth]{Planck512_skyfraction.pdf}
\caption{Width of 68.3\% confidence interval for $H_{loc}/H_0$ when observing only a fraction of the sky. Left: One cone. Right: Two cones, pointing in opposite directions of the sky.}
\label{fig:Planck512_skyfraction}
\end{figure}
\begin{figure}[htb!]
\center
\includegraphics[width = \textwidth]{Planck512_skyfraction_256.pdf}
\caption{Width of 68.3\% confidence interval for $H_{loc}/H_0$ as a function of the observed fraction of the sky at $r_{max}=256\di{Mpc/h}$. The full line corresponds to observations in one cone, the punctuated line to observations in two cones pointing in opposite directions.}
\label{fig:Planck512_skyfraction_256}
\end{figure}
\begin{table}
\center
\begin{tabular}{cccccccc}
Name & Description & $\mu_{67}[\%]$ & $\mu_{150}[\%]$ & $\mu_{256}[\%]$ & $\sigma_{67}[\%]$ & $\sigma_{150}[\%]$ & $\sigma_{256}[\%]$ \\
\hline
• & \textbf{Observers} & • & • & • & • & • & • \\
A.0 & Random positions in space & 0.1 & 0.0 & 0.0 & 4.6 & 0.9 & 0.3 \\
A.1 & Random positions in halos & 0.1 & -0.1 & 0.0 & 4.7 & 0.9 & 0.3 \\
A.2 & Local Group-like halos & -2.0 & -0.3 & 0.0 & 4.5 & 0.9 & 0.2 \\
\hline
• & \textbf{10\% sky coverage}& • & • & • & • & • & • \\
A.3 & One cone & -2.6 & -0.2 & -0.0 & 4.8 & 1.2 & 0.4 \\
A.4 & Two cones & -2.6 & -0.3 & -0.0 & 4.2 & 1.1 & 0.4 \\
\hline
• & \textbf{Cosmic evolution}& • & • & • & • & • & • \\
A.5 & Lightcone & -2.5 & -0.3 & 0.0 & 4.4 & 0.9 & 0.2 \\
\hline
• & \textbf{Simulation}& • & • & • & • & • & • \\
B & Different seed & -1.7 & -0.2 & 0.0 & 4.6 & 0.9 & 0.4 \\
C & Box=512, Nsim=1024 & -2.1 & -0.4 & -0.1 & 4.6 & 1.0 & 0.3 \\
D & Box=1024, Nsim=1024 & -1.7 & -0.2 & 0.0 & 5.0 & 1.1 & 0.4 \\
\end{tabular}
\caption{Mean ($\mu$) and variance ($\sigma$) of $H_{loc}/H_0$ at selected distances. The subscripts specify the distance in $\di{Mpc}/h$. All results are given in per cent.}
\label{tab:results}
\end{table}
\FloatBarrier
\section{Discussion}
\begin{figure}[htb!]
\center
\includegraphics[width = \textwidth]{Planck512_band.pdf}
\caption{The value of the local Hubble parameter found by Riess et al. (green line), is plotted together with the value and spread of $H_{loc}/H_0$ found in the standard analysis. The green band indicates the $1\sigma$ uncertainty in the result from Riess et al.}
\label{fig:Planck512_band}
\end{figure}
\noindent
Figure \ref{fig:Planck512_band} shows the mean and variance of the standard simulation (as in figure \ref{fig:Planck512}) together with the value of the Hubble parameter found by Riess et al. in \cite{Riess:2011} of $73.8\pm 2.4$, plotted as a green line, with the uncertainty of the measured value shown as a green band. Only for very small values of $r_{max}$ is there any overlap between the measured range of $H_0$ and the confidence bands obtained from a Planck like universe. At a distance of $r_{max} = 256\di{Mpc}/h$, less than $0.3\%$ of the Local Group like observers would observe a value as high as the one we see. We therefore conclude that the variance of the expansion field does not lift the discrepancy between the Hubble constant determined from measurements of the CMB and that obtained by direct measurements of recessional velocities in our local universe. This has also been concluded by Marra et al. in \cite{Marra:2013} and by Wojtak et al. in \cite{Wojtak:2013}. At an intermediate distance of $150\di{Mpc/h}$, we find the width of the 63.8\% confidence interval for the local Hubble constant to be $0.9\%-1.1\%$ (depending on the simulation), which is in good agreement with the value of $0.9\%$ found by Wojtak et al. Using equation 5 from \cite{Marra:2013} and the redshift distribution from \cite{Hicken:2009a}, we compute the standard deviation of $H_{loc}/H$ in the redshift range $0.023 < z < 0.1$. Averaging over all observers, we get 1.1\% for the standard simulation. This is 38\% bigger than the value of 0.8\% found by Wojtak et al., and 8\% smaller than the value of 1.2\% found by Marra et al. from analytical calculations based on the power spectrum.
We observe a tendency for the local Hubble parameter as measured by observers in halos to be systematically lower than the overall expansion rate, whereas observers distributed randomly in space tends to overestimate $H_0$. This can be explained by noting that observers in halos are positioned in in-fall regions as a consequence of ongoing structure formation, whereas observers positioned at random will have a tendency to be located in regions less dense than average, because these take up a greater fraction of the simulation volume than the overdense regions. Both effects are smoothed out when observing halos at large distances. This is in good agreement with the result obtained in \cite{Wojtak:2013}. Observers positioned randomly in halos will in mean observe a Hubble constant very close the actual value.
\section{Conclusion}
We have carefully studied how local measurements of $H_0$ can be influenced by a variety of different parameters related to survey geometry, depth, and size, as well as observer position in space.
We find that variations in the local expansion field {\it cannot} explain the difference in the Hubble parameter obtained indirectly by measurements of the CMB by the Planck collaboration and that obtained by direct measurements. This result has been found to be insensitive to the percentage of the sky observed in direct measurements, and to whether or not cosmic evolution is taken into consideration. At small distances, observers positioned in Local Group like halos will have a tendency to measure a Hubble constant that is lower than the true value, whereas observers positioned at random in space have a tendency to measure a higher value. However, these effects become negligible when the largest observed distance exceed a few hundred $\di{Mpc/h}$, and therefore have no significant effect on the scale used in the local measurement of the Hubble parameter.
Our conclusion is that the discrepancy between the value of $H_0$ inferred from Planck and the value found from direct measurements must be ascribed to other sources than a variation in the local velocity field.
\acknowledgments
We acknowledge computing resources from Center for Scientific Computing Aarhus. Research at Centre for Star and Planet Formation is funded by the Danish National Research Foundation. TH is supported by a Sapere Aude Starting Grant from The Danish Council for Independent Research.
\bibliographystyle{utcaps}
\providecommand{\href}[2]{#2}\begingroup\raggedright |
\section{Introduction}
Given a class of representable matroids,
the following are two basic
questions about the class.
Over which fields are the
members of the class representable?
Are there efficient algorithms to
construct representations
for every member of the class?
Here an algorithm is efficient
if its running time is
polynomial in the size of the ground set.
For instance, every transversal matroid is
representable over all sufficiently large
fields~\cite[Corollary~12.2.17]{Oxl92},
but it is not known exactly over which
fields they are representable, and the
existence of efficient algorithms to
construct representations is an
open problem too.
The interest for these problems has been mainly motivated by
their connections to coding theory and cryptology,
mainly to secret sharing.
Determining over which fields
the uniform matroids are representable
is equivalent to solving the
Main Conjecture for Maximum Distance Separable Codes.
For more details, and a proof of this conjecture in the prime case,
see~\cite{Ball2011}, and for further information
on when the conjecture is known to hold, see~\cite[Section~3]{HS2001}.
As a consequence of the results by
Brickell~\cite{Bri89}, every representation
of a matroid $M$ over a finite field
provides ideal linear secret sharing schemes for
the access structures that are ports of the matroid $M$.
Because of that, the representability of certain classes
of matroids is closely connected to the search for efficient
constructions of secret sharing schemes
for certain classes of access structures.
The reader is referred to~\cite{MaPa10}
for more information about
secret sharing and its connections to matroid theory.
Several constructions of ideal linear
secret sharing schemes
for families of relatively simple
access structures with interesting
properties for the applications have been
proposed~\cite{BTW08,Bri89,FaPa10,FPXY11,HeSa06,Ng06,PaSa00,Sim88,Tas07,TaDy06}.
They are basic and natural generalizations
of Shamir's~\cite{Sha79}
threshold secret sharing scheme.
A unified approach to all those proposals
was presented in~\cite{FMP07}.
As a consequence, the open questions about the existence
of such secret sharing schemes for some sizes of the secret value
and the possibility of constructing them efficiently
are equivalent to determining the representability
of some classes of multi-uniform matroids.
See~\cite{FaPa11,FPXY11}
for more information on this line of work.
In this paper, we analyze the representability
of the bi-uniform matroids.
They were introduced by Ng and Walker~\cite{NgWa01},
but ideal secret sharing schemes for
the access structures that are determined
by them were previously presented in~\cite{PaSa00}.
Bi-uniform matroids are defined in terms of their symmetry properties,
specifically the number of clonal classes,
a concept introduced in~\cite{GOVW98}.
Two elements in the ground set of a matroid
are said to be \emph{clones\/} if the map that interchanges them
and fixes all other elements is an automorphism of the matroid.
Being clones is clearly an equivalence relation,
and its equivalence classes are called the \emph{clonal classes\/} of the matroid.
Uniform matroids are precisely those having only one clonal class.
A matroid is said to be \emph{bi-uniform\/} if
it has at most two clonal classes.
Of course, this definition can be generalized to
\emph{$m$-uniform matroids\/} for every positive integer $m$.
A bi-uniform matroid is determined by
its rank, the number of elements in each clonal class,
and the ranks of the two clonal classes,
which are called the \emph{sub-ranks\/} of the bi-uniform matroid.
It is not difficult to check
that every bi-uniform matroid is a transversal matroid,
and hence it is representable over all
sufficiently large fields.
Moreover, as a consequence of the results in~\cite{FMP07},
every bi-uniform matroid is representable
over all fields with at least ${N \choose k}$ elements,
where $N$ is the size of the ground set
and $k$ is the rank.
The same result applies to tri-uniform
matroids~\cite{FMP07},
but it does not apply to $4$-uniform matroids
because the Vamos matroid is not
representable~\cite[Proposition 6.1.10]{Oxl92}.
Even though the proof in~\cite{FMP07} is constructive,
no efficient method to find representations
for the bi-uniform matroids can be derived from it.
A method to construct a representation
for every bi-uniform matroid was presented
by Ng~\cite{Ng03}, but it was not proved to be efficient.
Efficient methods to find representations
for the bi-uniform matroids in which
one of the sub-ranks is equal to the rank
can be derived from the
constructions of ideal hierarchical
secret sharing schemes
by Brickell~\cite{Bri89} and by Tassa~\cite{Tas07}.
These constructions are analyzed
in Section~\ref{pt:related}.
In this work, we prove for the first time
that there exist efficient algorithms
to find representations for \emph{all\/}
bi-uniform matroids.
In addition, our constructions provide representations
over finite fields that are in many cases
smaller than the ones used in~\cite{Bri89,Ng03,Tas07}.
A detailed comparison is given in Section~\ref{pt:related}.
More specifically, we present three different representations of bi-uniform matroids.
All of them can be obtained in time polynomial in the size of the ground set.
An important parameter in our discussions is
$d = m + \ell -k$,
where $k$ is the rank of the matroid while
$m$ and $\ell$ are its sub-ranks.
The cases $d = 0$ and $d = 1$ are reduced to the
representability of the uniform matroid.
Our first construction (Theorem~\ref{st:rep4d2}) corresponds to the case $d = 2$,
and we prove that every such bi-uniform matroid
is representable over $\mathbb{F}_q$ if $q$ is odd
and every clonal class has at most $(q-1)/2$ elements.
The other two constructions apply to the general case,
and they are both based on a family of linear evaluation codes.
Our second construction (Theorem~\ref{st:repextfield})
provides a representation of the bi-uniform
matroid over ${\mathbb F}_{q_0^s}$,
where $s > d(d-1)/2$ and $q_0$ is a prime power
larger than the size of each clonal class.
Finally, we present a third construction in Theorem~\ref{st:repprimefield}.
In this case, if $m \ge \ell$,
a representation of the bi-uniform matroid
is obtained over every prime field $\mathbb{F}_p$
with $p > K^h$, where
$K$ is larger than half the number of elements in
each clonal class and
$h = md(1 + d(d-1)/2)$.
\section{Related work}
\label{pt:related}
The existence of ideal secret sharing schemes
for the so-called bipartite and tripartite access structures
was proved in~\cite{PaSa00} and in~\cite{FMP07}, respectively.
These proofs are constructive and, in particular, they
provide a method to find representations
for all bi-uniform matroids.
Such a representation can be found
over every field with at least
${N \choose k}$ elements,
where $N$ is the size of the ground set
and $k$ is the rank.
This method is not efficient
because exponentially many
determinants have to be computed
to find a valid representation.
This problem is avoided in the
method proposed by Ng~\cite{Ng03},
which provides a representation
for every given bi-uniform matroid.
Specifically, Ng gives a representation for
the bi-uniform matroid
with rank $k$ and sub-ranks $m,\ell$
over every finite field of the form
$\mathbb{F}_{q_0^s}$, where $q_0 > 14$,
each clonal class has at most $q_0$ elements,
and $s$ is at least $k$
and co-prime with $d = m + \ell - k$.
This method may be efficient,
but this fact is not proved in~\cite{Ng03}.
In addition, the degree $s$ of the extension field
depends on the rank $k$, while in our
efficient construction
in Theorem~\ref{st:repextfield}, this degree depends
only on $d$.
Therefore, if $d$ is small compared to $k$,
our construction works over smaller fields.
Efficient methods to construct
ideal hierarchical secret sharing schemes were
given by Brickell~\cite{Bri89} and by Tassa~\cite{Tas07}.
When applied to some particular cases, these
methods provide representations for bi-uniform
matroids in which one of the sub-ranks is equal to the rank.
Brickell's construction provides a representation
for every such bi-uniform matroid
over fields of the form $\mathbb{F}_{q_0^s}$,
where $q_0$ is a prime power larger
than the size of each clonal class
and $s$ is at least the square of the rank of the matroid.
An irreducible polynomial
of degree $s$ over $\mathbb{F}_{q_0}$
has to be found,
but this can be done in time polynomial in $q_0$ and $s$
by using the algorithm given by Shoup~\cite{Sho90}.
Therefore, a representation can be found in
time polynomial in the size of the ground set.
Clearly, the size of the field is much smaller in the
representations that are obtained by the method described
in Theorem~\ref{st:repextfield}.
Representations for those bi-uniform matroids
are efficiently obtained from Tassa's construction over
prime fields $\mathbb{F}_p$
with $p$ larger than
$N^{(k-1)(k-2)/2}$, where $N$
is the number of elements in the ground set.
If $d$ is small compared to $n$, the size of the field
in our construction (Theorem~\ref{st:repprimefield})
is smaller.
Representations for bi-uniform matroids in which
one of the sub-ranks is equal to the rank of
the matroid and the other one is equal to $2$
are obtained from the constructions of
ideal hierarchical secret sharing schemes in~\cite{BeWe93}.
These are representations over $\mathbb{F}_q$,
where the size of the ground set
is at most $q+1$ and the size of each clonal class is around $q/2$.
These parameters are similar
to the ones in Theorem~\ref{st:rep4d2},
but our construction is more general.
\section{The bi-uniform matroid}
\label{pt:defs}
A {\em matroid} $M=(E,F)$ is a pair in which $E$
is a finite set, called the \emph{ground set\/}, and
$F$ is a nonempty set of subsets of $E$,
called {\em independent sets}, such that
\begin{enumerate}
\item
every subset of an independent set is an independent subset, and
\item
for all $A \subseteq E$, all maximal independent
subsets of $A$ have the same cardinality,
called the {\em rank of $A$} and denoted $r(A)$.
\end{enumerate}
A {\em basis} $B$ of $M$ is a maximal independent set.
Obviously all bases have the same cardinality, which is called
the \emph{rank of $M$}.
If $E$ can be mapped to a subset of vectors of a vector space over a field ${\mathbb K}$
so that $I \subseteq E$ is an independent set if and only if the vectors assigned
to the elements in $I$ are linearly independent, then the matroid is said to be
{\em representable over ${\mathbb K}$}.
The independent sets of the {\em uniform matroid} of rank $k$
are all the subsets $B$ of the set $E$ with the property that $|B| \leq k$.
If the uniform matroid is representable over a field ${\mathbb K}$ then there is a map
$$
f \ : \ E \rightarrow \ {\mathbb K}^k
$$
such that $f(E)$ is a set of vectors with the property that every subset
of $f(E)$ of size $k$ is a basis of ${\mathbb K}^k$.
For positive integers $k,m,\ell$ with
$1 \le m, \ell \le k$ and $m + \ell \geq k$,
and a partition $E = E_1 \cup E_2$ of the ground set with
$|E_1| \geq m$ and $|E_2| \geq \ell$,
the independent sets of the {\em bi-uniform matroid of rank $k$ and sub-ranks $m$, $\ell$}
are all the subsets $B$ of the ground set with the property that
$|B | \leq k$, $|B \cap E_1| \leq m$ and $|B \cap E_2| \leq \ell$.
Since the maximal independent subsets of $E_1$
have $m$ elements, $r(E_1) = m$.
Similarly, $r(E_2) = \ell$.
If the bi-uniform matroid is representable over a field ${\mathbb K}$ then there is a map
$$
f \ : \ E \rightarrow \ {\mathbb K}^k
$$
such that $f(E)$ is a set of vectors with the property that every subset
$D$ of $f(E)$ of size $k$ with
$|D \cap f(E_1)| \leq m$ and $|D \cap f(E_2)| \leq \ell$
is a basis of ${\mathbb K}^k$.
The dimensions of
$\langle f(E_1) \rangle$
and $\langle f(E_2) \rangle$ are
$m = r(E_1)$ and $\ell = r(E_2)$, respectively.
Thus, if the bi-uniform matroid is representable over ${\mathbb K}$ then we can construct a set
$S \cup T$ of vectors of ${\mathbb K}^k$ such that
$\dim(\langle S \rangle )=m$ and $\dim(\langle T \rangle )=\ell$,
with the property that every subset $B$ of $S \cup T$ of size $k$
with $|B \cap S| \leq m$ and $|B \cap T| \leq \ell$ is a basis.
\section{Necessary conditions}
\label{pt:neccond}
We present here some necessary conditions for a
bi-uniform matroid to be representable over
a finite field ${\mathbb F}_q$.
The following lemma implies that restricting a representation of the bi-uniform matroid
on $E = E_1 \cup E_2$, one gets a representation of the
uniform matroid on $E_1$ of rank $m$ and the uniform matroid on $E_2$ of rank $\ell$.
Therefore, the known necessary conditions for the
representability of the uniform matroid over ${\mathbb F}_q$
can be applied to the bi-uniform matroid.
\begin{lemma}
If $f$ is a map from $E$ to ${\mathbb K}^k$ which gives a representation of the bi-uniform matroid
of rank $k$ and sub-ranks $m$ and $\ell$ then $f(E_1)$ has the property that every
subset of $f(E_1)$ of size $m$ is a basis of $\langle f(E_1) \rangle$.
Similarly, $f(E_2)$ has the property that every subset of $f(E_2)$ of size $\ell$
is a basis of $\langle f(E_2) \rangle$.
\end{lemma}
\begin{proof}
If $L'$ is a set of $m$ vectors of $f(E_1)$ which are linearly dependent then
$L' \cup L$, where $L$ is a set of $k-m$ vectors of $f(E_2)$,
is a set of $k$ vectors of $f(E)$ which do not form a basis of ${\mathbb K}^k$.
\end{proof}
The \emph{dual\/} of a matroid $M$ is the matroid $M^*$
on the same ground set such that its bases
are the complements of the bases of $M$.
Given a representation of $M$ over $\mathbb{K}$,
simple linear algebra operations provide
a representation of $M^*$ over the same
field~\cite[Section~2.2]{Oxl92}.
In particular, if $\mathbb{K}$ is finite,
a representation of $M^*$ can be
efficiently obtained from a representation of $M$.
By the following proposition,
the dual of a bi-uniform matroid is a bi-uniform matroid
with the same partition of the ground set.
\begin{proposition}
The dual of the bi-uniform matroid of rank $k$ and sub-ranks $m$ and $\ell$
on the ground set $E = E_1 \cup E_2$ is the
bi-uniform matroid of rank $k^* = |E_1|+|E_2|-k$
and sub-ranks $m^* = |E_1| + \ell - k$ and $\ell^* = |E_2|+ m -k$.
\end{proposition}
\begin{proof}
Clearly, a matroid and its dual have the same automorphism group.
This implies that the dual of a bi-uniform matroid is bi-uniform
for the same partition of the ground set.
The values for the rank and the sub-ranks of $M^*$
are derived from the formula that relates the
rank function $r$ of matroid $M$
to the rank function $r^*$ of its dual $M^*$.
Namely,
\(
r^*(A) = |A| - r(E) + r(E \setminus A)
\)
for every $A \subseteq E$~\cite[Proposition~2.1.9]{Oxl92}.
\end{proof}
Clearly, $k = m = \ell$ if and only if
$m^* = |E_1|$ and $\ell^* = |E_2|$,
and in this case both $M$ and $M^*$
are uniform matroids.
We assume from now on that
$m < k$ or $\ell < k$
and that
$m < |E_1|$ or $\ell < |E_2|$,
The results in this paper indicate that
the value $d = m + \ell - k$, which is equal to
the dimension of $\langle S \rangle \cap \langle T \rangle$,
is maybe the most influential parameter
when studying the representability
of the bi-uniform matroid over finite fields.
Observe that the value of this parameter is the same
for a bi-uniform matroid $M$ and for its dual $M^*$.
If $d = 0$, then the problem reduces to the representability of the uniform matroid.
Similarly, if $d = 1$ then, by adding to $S \cup T$ a nonzero vector
in the one-dimensional intersection of $\langle S \rangle$ and $\langle T \rangle$,
the problem again reduces to the representability of the uniform matroid.
From now on, we assume that $d = m + \ell - k \geq 2$.
\begin{proposition}
If $k \leq m+\ell-2$ and the bi-uniform matroid of rank $k$
and sub-ranks $m, \ell$ is
representable over ${\mathbb F}_q$, then $|E| \leq q+k-1$.
\end{proposition}
\begin{proof}
Take a subset $A$ of $S$ of size $k-\ell$.
Then $\langle A \rangle \cap \langle T \rangle = \{ 0 \}$
because $A \cup C$ is a basis for every
subset $C$ of $T$ of size $\ell$.
Since $k - \ell \le m - 2$, we can project
the points of $S \setminus A$ onto
$\langle S \rangle \cap \langle T \rangle$,
by defining $A'$ to be a set of $|S|-(k-\ell)$ vectors,
each a representative of a distinct $1$-dimensional subspace
$\langle x,A \rangle\cap (\langle S \rangle \cap \langle T \rangle)$
for some $x \in (S \setminus A)$.
Let $B$ be a subset of $T$ of size $\ell-2$.
For all $x \in A'$, if $\langle B,x \rangle$ contains $\ell-1$ points of $T$
then $\langle A,B,x \rangle$ is a hyperplane of ${\mathbb F}_q^k$ containing $k$ points of $S \cup T$,
at most $m-1$ points of $S$ and $\ell-1$ points of $T$.
This cannot occur since such a set must be a basis, by hypothesis.
Thus, each of the $q+1$ hyperplanes containing $\langle B \rangle$
contains at most one vector of $A' \cup (T\setminus B)$. This gives
\(
|T|-(\ell-2)+|S|-(k-\ell) \leq q+1
\),
which gives the desired bound, since $E=S \cup T$.
\end{proof}
\begin{proposition}
If $q \leq k \leq m+\ell-2$, then the bi-uniform matroid is not representable over~${\mathbb F}_q$.
\end{proposition}
\begin{proof}
Assume that the bi-uniform matroid is representable and we have the sets of vectors $S$ and $T$ as before.
Let $e_1,\ldots,e_m$ be vectors of $S$. These vectors form a basis for $\langle S\rangle$
and we can extend them with $k - m$ vectors $e_{m+1},\ldots,e_k$ of $T$ to a basis of $\langle S,T\rangle$.
For every vector in $T$ that is not in the basis $\{e_1,\ldots, e_k \}$,
all its coordinates in this basis are non-zero.
Indeed, if there is such a vector with a zero coordinate in the $i\geq m+1$ coordinate
then the hyperplane $X_i=0$ contains $m$ vectors of $S$ and $k-m$ vectors of $T$, which does not occur.
Similarly, if the zero coordinate is in the $i\leq m$ coordinate then the hyperplane $X_i=0$
contains $m-1$ vectors of $S$ and $k-m+1$ vectors of $T$, which also does not occur.
Thus, by multiplying the vectors in the basis
by some nonzero scalars, we can assume that
$e_1 + \cdots + e_k$ is a vector of $T$
and all the coordinates of the other vectors in
$T \setminus \{e_{m+1},\ldots,e_k\}$ are non-zero.
Since $\ell \ge k - m + 2$,
there is a vector $z \in T\setminus \{e_{m+1},\ldots,e_k, e_1+ \cdots +e_k \}$.
Since $k\geq q$ there are coordinates $i$ and $j$ such that $z_i=z_j$.
If $1 \leq i \leq m$ and $1 \leq j \leq m$ then the hyperplane $X_i=X_j$
contains $m-2$ vectors of $S$ and $k-m+2 \leq \ell$ vectors of $T$, which cannot occur.
If $1 \leq i \leq m$ and $m+1 \leq j \leq k$ then the hyperplane $X_i=X_j$
contains $m-1$ vectors of $S$ and $k-m+1$ vectors of $T$, which also cannot occur.
Finally, if $m+1 \leq i \leq k$ and $m+1 \leq j \leq k$ then the hyperplane $X_i=X_j$
contains $m$ vectors of $S$ and $k-m$ vectors of $T$, which cannot occur, a contradiction.
\end{proof}
\section{Representations of the bi-uniform matroid}
\label{pt:reps}
\begin{theorem}
\label{st:rep4d2}
The bi-uniform matroid of rank $k$ and sub-ranks
$m$ and $\ell$ with $d = m + \ell - k = 2$
is representable over ${\mathbb F}_q$ if
$q$ is odd and $\max\{|E_1|,|E_2|\} \leq (q - 1)/2$.
\end{theorem}
\begin{proof}
Let $L$ denote the set of non-zero squares of ${\mathbb F}_q$ and
$(-1)^{\ell+m}\eta $ a fixed non-square of ${\mathbb F}_q$.
Consider the subsets of ${\mathbb F}_q^k$
$$
S = \{ (t,t^2,\ldots,t^{m-2},1,t^{m-1},0,\ldots,0) \ | \ t \in L \}
$$
and
$$
T = \{ (0,\ldots,0,\eta,t^{\ell-1},t^{\ell-2},\ldots,t) \ | \ t \in L \},
$$
where the coordinates are with respect to the basis $\{e_1,\ldots,e_k\}$.
We prove in the following that any
injective map which maps the elements of $E_1$
to a subset of $S$ and the elements of $E_2$ to a subset of $T$
is a representation of the bi-uniform matroid.
Since every set of $S \cup \{e_{m-1},e_{m}\}$
of size $m$ is a basis of $\langle S \rangle$,
every set formed by $m - 2$ vectors in $S$
and $\ell$ vectors in $T$ is a basis.
Symmetrically, the same holds for every
$m$ vectors in $S$ and $\ell - 2$ vectors in $T$.
The proof is concluded by showing that
there is no hyperplane $H$ of ${\mathbb F}_q^k$ containing $m - 1$ points of $S$
and $\ell - 1$ points of $T$.
Suppose that, on the contrary, such a hyperplane $H$ exists.
Since $S \cup A$ span ${\mathbb F}_q^k$ for every
$A \subseteq T$ of size $\ell - 2$,
the hyperplane $H$ intersects $\langle S \rangle$
in an $(m-1)$-dimensional subspace.
Symmetrically, $H \cap \langle T \rangle$ has dimension $\ell-1$.
Therefore, $H$ intersects
$\langle e_{m-1},e_m \rangle = \langle S \rangle \cap \langle T \rangle$
in a one-dimensional subspace.
Take elements $a_1$ and $a_2$ of ${\mathbb F}_q$,
not both zero, with $a_1 e_{m-1} + a_2 e_m \in H$.
The $m-1$ vectors of $H \cap S$ together with $a_1 e_{m-1}+a_2 e_m$ are linearly dependent.
Thus, there are $m-1$ different elements $t_1,\ldots,t_{m-1}$ of $L$ such that
$$
\det\left(\sum_{i=1}^{m-2} t_1^i e_i + e_{m-1} + t_1^{m-1} e_m, \ldots,
\sum_{i=1}^{m-2} t_{m-1}^i e_i + e_{m-1} + t_{m-1}^{m-1} e_m, a_1 e_{m-1}+a_2 e_m \right) = 0.
$$
Expanding this determinant by the last column gives
$$
a_2 (-1)^m V(t_1,\ldots,t_{m-1})=
a_1 V(t_1,\ldots, t_{m-1}) \prod_{i=1}^{m-1} t_i,
$$
where $V(t_1,\ldots,t_{m-1})$ is the determinant of the Vandermonde matrix.
Since $a_1=0$ implies $a_2=0$, we can assume that $a_1 \neq 0$ and so $a_2a_1^{-1}(-1)^{m} \in L$.
Analogously, the $\ell-1$ vectors of $H \cap T$
together with $a_1 e_{m-1} + a_2 e_m$ are linearly dependent, and hence
there are $\ell-1$ elements $u_1,\ldots,u_{\ell-1}$ of $L$ such that
$$
\det \left(\eta e_{m-1}+\sum_{i=1}^{\ell-1} u_1^i e_{k+1-i},\ldots,
\eta e_{m-1}+\sum_{i=1}^{\ell-1} u_{\ell-1}^i e_{k+1-i},a_1 e_{m-1}+a_2 e_m \right)=0.
$$
Expanding this determinant by the last column gives
$$
\eta a_2(-1)^{\ell} V(u_1,\ldots,u_{\ell-1})=
a_1 V(u_1,\ldots,u_{\ell-1}) \prod_{i=1}^{\ell-1} u_i.
$$
Since $a_1=0$ implies $a_2=0$, we can assume that $a_1 \neq 0$ and so $\eta a_2a_1^{-1}(-1)^{\ell} \in L$,
and since $a_2a_1^{-1}(-1)^{m} \in L$, this gives $\eta (-1)^{\ell+m} \in L$.
However, $\eta$ was chosen so that this is not the case.
\end{proof}
We describe in the following a family of linear evaluation codes
that will provide different representations of
the bi-uniform matroid for all possible values of the
rank $k$ and the sub-ranks $m, \ell$.
Take $\beta \in {\mathbb F}_q$ and the subspace $V$ of
${\mathbb F}_q[x] \times {\mathbb F}_q[y]$ defined by
$$
V= \{ (f(x),g(y)) \ | \ f(x)=f_1(x)+x^{m-d}g_1(\beta x),\ g(y)=g_1(y)+y^{d}g_2(y),
$$
$$
\mathrm{deg}(f_1) \leq m-d-1,\ \mathrm{deg}(g_1) \leq d-1,\ \mathrm{deg}(g_2) \leq \ell-d-1\},
$$
where $d = m + \ell - k$.
Let $F_1=\{ x_1,\ldots,x_{N_1} \}$ and
$F_2=\{ y_1,\ldots,y_{N_2} \}$ be subsets of ${\mathbb F}_q \setminus \{0\}$,
where $N_1 = |E_1|$ and $N_2 = |E_2|$.
Define $C = C(F_1,F_2,\beta)$ to be the linear evaluation code
$$
C = \{((f(x_1), \ldots, f(x_{N_1}),
g(y_1), \ldots, g(y_{N_2})) \ | \ (f,g) \in V \}.
$$
Note that $\mathrm{dim}C=\mathrm{dim}V=m-d+\ell-d+d=k$.
Every linear code determines a matroid,
namely the one that is represented
by the columns of a generator matrix $G$,
which is the same for all generator matrices of the code.
We analyze now under which conditions
the code $C = C(F_1, F_2,\beta)$ provides a
representation over ${\mathbb F}_q$ of the bi-uniform matroid
by identifying $E_1$ and $E_2$ to $F_1$ and $F_2$, respectively
(that is, to the first $N_1$ columns and the last $N_2$ columns of $G$, respectively).
Clearly, for every $A \subseteq E$
with $|A \cap E_1| > m$ or $|A \cap E_2| > \ell$,
the corresponding columns of $G$ are linearly dependent.
Let $B$ be a basis of the bi-uniform matroid with
$|B \cap E_1| = m - t_1$ and
$|B \cap E_2| = \ell - t_2$,
where $0 \le t_i \le d$ and $t_1 + t_2 = d$.
We can assume that $B \cap E_1$ is mapped to
$\{x_1, \ldots, x_{m - t_1} \} \subseteq F_1$
and $B \cap E_2$ is mapped to
$\{y_1, \ldots, y_{\ell - t_2} \} \subseteq F_2$.
The corresponding columns of $G$ are linearly independent
if and only if $(f,g) = (0,0)$ is the only element in $V$ satisfying
\begin{equation}
\label{eq:kernel}
(f(x_1), \ldots, f(x_{m - t_1}),
g(y_1), \ldots, g(y_{\ell - t_2})) = 0.
\end{equation}
Let
$$
r(x) = (x - x_1) \cdots (x- x_{m - t_1}) =
\sum_{i=0}^{m-t_1}r_i x^i
$$
and
$$
s(y) = (y - y_1) \cdots (y - y_{\ell - t_2}) =
\sum_{i=0}^{\ell-t_2}s_i y^i.
$$
Then $(f,g) \in V$ satisfy~(\ref{eq:kernel}) if and only if
$f(x) = a(x)r(x)$ for some polynomial
$a(x) = \sum_{i=0}^{t_1-1} a_i x^i$ and
$g(y) = b(y)s(y)$ for some polynomial
$b(y)=\sum_{i=0}^{t_2-1} b_i y^i$.
Since $f(x) = a(x)r(x) = f_1(x)+x^{m-d}g_1(\beta x)$,
$$
g_1(\beta x) = \sum_{i=0}^{t_1-1} a_i
\left(\sum_{j= 0}^{d-t_1+i} r_{m-d+j-i}\, x^j\right),
$$
where $r_j = 0$ if $j < 0$.
On the other hand, $g(y) = b(y)s(y) = g_1(y)+y^d g_2(y)$ and so
$$
g_1(y) = \sum_{i=0}^{t_2-1} b_i \left(\sum_{j=i}^{d-1} s_{j-i}\, y^j\right),
$$
where $s_j = 0$ if $j > \ell - t_2$. Hence,
\begin{equation}
\label{eq:lincomb}
\sum_{i=0}^{t_1-1} a_i \left(\sum_{j=0}^{d-t_1+i} r_{m-d+j-i}\, x^j \right)
=\sum_{i=0}^{t_2-1} b_i\left(\sum_{j=i}^{d-1} s_{j-i} (\beta x)^j \right).
\end{equation}
If $(f,g) \neq 0$ then either $a$ or $b$ is nonzero and so there
is a linear dependence between the~$d$ polynomials in Equation~(\ref{eq:lincomb}).
Therefore, the determinant of the $d \times d$ matrix
\begin{equation}
\label{eq:bigmatrix}
\left(
\begin{array}{ccccccccc}
r_{m-d} & r_{m-d+1} & \cdots & \cdots & \cdots &r_{m-t_1} & 0 & \cdots & 0 \\
r_{m-d-1} & r_{m-d} & \cdots & \cdots & \cdots & \cdots & r_{m-t_1} &\cdots & 0 \\
\vdots & \vdots & & & & & & \ddots & \vdots \\
& & \cdots & \cdots & \cdots & \cdots & \cdots & \cdots & r_{m-t_1} \\
\hline
s_0 & s_1 \beta & \cdots & \cdots & \cdots & \cdots & \cdots & \cdots & \\
0 & s_0 \beta & s_1 \beta^2 & \cdots & \cdots & \cdots & \cdots & \cdots & \\
\vdots & \ddots & \ddots & \ddots & & & & & \vdots \\
0 & \cdots & 0 & s_0 \beta^{t_2-2} & s_1 \beta^{t_2-1} & \cdots & \cdots & \cdots & s_{t_1 + 1}\beta^{d-1}\\
0 & \cdots & \cdots & 0 & s_0 \beta^{t_2-1} & s_1 \beta^{t_2} & \cdots & \cdots & s_{t_1}\beta^{d-1}
\end{array}
\right)
\end{equation}
is zero.
In conclusion, the code $C(F_1,F_2,\beta)$ provides a
representation over $\mathbb{F}_q$ of the bi-uniform matroid
if and only if the determinant of the matrix~(\ref{eq:bigmatrix})
is nonzero for every choice of $m - t_1$ elements in $F_1$ and
$\ell - t_2$ elements in $F_2$ with $0 \le t_i \le d$ and $t_1 + t_2 = d$.
Clearly, this is always the case if $t_1 = 0$ or $t_2 = 0$.
Otherwise, that determinant can be expressed
as an ${\mathbb F}_{q}$-polynomial on $\beta$.
The degree of this polynomial $\varphi(\beta)$ is at most $d(d-1)/2$.
In addition, $\varphi(\beta)$ is not identically zero because
the term with the minimum power of $\beta$ is
equal to $1 \beta \cdots \beta^{t_2 - 1} s_0^{t_2} r_{m-t_1}^{t_1}$,
and $r_{m-t_1} = 1$ and $s_0 \neq 0$.
In the next two theorems we present two different ways to select
$F_1, F_2, \beta$ with that property.
\begin{theorem}
\label{st:repextfield}
The bi-uniform matroid of rank $k$ and sub-ranks $m$ and $\ell$
with $d = m + \ell - k \ge 2$ is representable over ${\mathbb F}_q$
if $q = q_0^s$ for some $s > d(d-1)/2$ and
some prime power $q_0 > \max\{|E_1|,|E_2|\}$.
Moreover, such a representation can be obtained
in time polynomial in the size of the ground set.
\end{theorem}
\begin{proof}
Take $F_1$ and $F_2$ from
${\mathbb F}_{q_0} \setminus \{0\}$
and take $\beta \in {\mathbb F}_q$ such that its
minimal polynomial over ${\mathbb F}_{q_0}$ is of degree $s$.
The algorithm by Shoup~\cite{Sho90} finds
such a value $\beta$ in time polynomial in $q_0$ and $s$.
Then the code $C(F_1,F_2,\beta)$ gives a representation
over ${\mathbb F}_{q}$ of the bi-uniform matroid.
Indeed, all the entries in the matrix~(\ref{eq:bigmatrix}), except the
powers of $\beta$, are in ${\mathbb F}_{q_0}$.
Therefore, $\varphi(\beta)$ is a nonzero ${\mathbb F}_{q_0}$-polynomial
on $\beta$ with degree smaller than $s$.
\end{proof}
Our second construction of a code $C(F_1,F_2,\beta)$
representing the bi-uniform matroid is done over a
prime field $\mathbb{F}_p$.
We need the following well known bound
on the roots of a real polynomial.
\begin{lemma}
\label{st:bound}
The absolute value of every root
of the real polynomial $c_0 + c_1 x + \cdots + c_n x^n$
is at most
\(
1 + \max_{0 \le i \le n-1} |c_i|/ |c_n|.
\)
\end{lemma}
\begin{theorem}
\label{st:repprimefield}
Let $M$ be the bi-uniform matroid of of rank
$k$ and sub-ranks $m$ and $\ell$ with $d = m + \ell - k \ge 2$
and $m \ge \ell$.
Take $N = \max\{|E_1|, |E_2|\}$ and $K = \lceil N/2 \rceil + 1$.
Then $M$ is representable over ${\mathbb F}_p$
for every prime $p > K^h$,
where $h = md(1 + d(d-1)/2)$.
Moreover, such a representation can be obtained
in time polynomial in the size of the ground set.
\end{theorem}
\begin{proof}
First, we select the value $\beta$ and the sets $F_1, F_2$
among the integers in such a way that the determinant
of the real matrix~(\ref{eq:bigmatrix}) is always nonzero.
Then we find an upper bound on the absolute
value of this determinant.
The code $C(F_1,F_2,\beta)$ will represent
the bi-uniform matroid over $\mathbb{F}_p$
if $p$ is larger than that bound.
Consider two sets of nonzero integer numbers
$F_1, F_2$ with $|F_i| = |E_i|$ in the interval $[-(K-1),K-1]$.
Take $m - t_1$ values in $F_1$ and $\ell - t_2$ values in $F_2$,
where $1 \le t_i \le d-1$ and $t_1 + t_2 = d$.
Then the values $r_i$ appearing in the
matrix~(\ref{eq:bigmatrix}) satisfy
\[
|r_{m - t_1 - i}| \le {m - t_1 \choose i} (K-1)^i
\]
for every $i = 0, \ldots, m - t_1$,
and hence
\(
\sum_{i = 0}^{m-t_1} |r_i| \le K^{m-t_1}.
\)
Analogously,
\(
\sum_{i = 0}^{\ell-t_2} |s_i| \le K^{\ell-t_2}.
\)
Since $r_{m-t_1} = s_{\ell - t_2} = 1$ and $m \ge \ell$,
all values $|r_i|$, $|s_j|$ are less than or equal to $K^m - 1$.
Then $\varphi(\beta)$ is a real
polynomial on $\beta$ with degree at most $d(d-1)/2$ such that the absolute value
of every coefficient is at most $(K^m - 1)^d < K^{md} - 1$.
Take $\beta = K^{md}$.
By Lemma~\ref{st:bound}, $\varphi(\beta) \ne 0$.
Moreover,
\[
|\varphi (\beta)| \le (K^{m} - 1)^{d} \;
\frac{\beta^{d(d-1)/2+1} - 1}{\beta -1} < K^h.
\]
Finally, consider a prime $p > K^{h}$
and reduce $\beta = K^{md}$ and the elements in $F_1$ and $F_2$ modulo~$p$.
The code $C(F_1,F_2,\beta)$ represents
the bi-uniform matroid $M$ over $\mathbb{F}_p$.
Observe that the number of bits that are needed
to represent the elements in $\mathbb{F}_p$
is polynomial in the size of the ground set.
\end{proof}
|
\section{Introduction}
Let $L$ be a locally compact topological group acting continuously on a locally Hausdorff topological space $M$. This action is called {\it \textbf{proper}} if for every compact subset $C \subset M$ the set
$$L(C):=\{ g\in L \ | \ g\cdot C \cap C \neq \emptyset \}$$
is compact. In this paper, our main concern is the following question posed by T. Kobayashi \cite{kob3}
\begin{center}
How ``large'' subgroups of $G$ can act properly
\end{center}
\begin{center}
on a homogeneous space $G/H$? \textbf{(Q1)}
\end{center}
We restrict our attention to the case where $M=G/H$ is a homogeneous space of reductive type and always assume that $G$ is a linear connected reductive real Lie group with the Lie algebra $\mathfrak{g}.$ Let $H\subset G$ be a closed subgroup of $G$ with finitely many connected components and $\mathfrak{h}$ be the Lie algebra of $H.$
\begin{definition}
The subgroup $H$ is reductive in $G$ if $\mathfrak{h}$ is reductive in $\mathfrak{g},$ that is, there exists a Cartan involution $\theta $ for which $\theta (\mathfrak{h}) = \mathfrak{h}.$
The space $G/H$ is called the homogeneous space of reductive type.
\label{def1}
\end{definition}
\noindent
Note that if $\mathfrak{h}$ is reductive in $\mathfrak{g}$ then $\mathfrak{h}$ is a reductive Lie algebra.
It is natural to ask when does a closed subgroup of $G$ act properly on a space of reductive type $G/H.$ This problem was treated, inter alia, in \cite{ben}, \cite{bt}, \cite{kas}, \cite{kob2}, \cite{kob4}, \cite{kob1}, \cite{kul} and \cite{ok}. In \cite{kob2} one can find a very important criterion for a proper action of a subgroup $L$ reductive in $G.$ To state this criterion we need to introduce some additional notation. Let $\mathfrak{l}$ be the Lie algebra of $L.$ Take a Cartan involution $\theta$ of $\mathfrak{g}.$ We obtain the Cartan decomposition
\begin{equation}
\mathfrak{g}=\mathfrak{k} + \mathfrak{p}.
\label{eq1}
\end{equation}
Choose a maximal abelian subspace $\mathfrak{a}$ in $\mathfrak{p}.$ The subspace $\mathfrak{a}$ is called the \textbf{\textit{maximally split abelian subspace}} of $\mathfrak{p}$ and $\text{rank}_{\mathbb{R}}(\mathfrak{g}) := \text{dim} (\mathfrak{a})$ is called the \textbf{\textit{real rank}} of $\mathfrak{g}.$ It follows from Definition \ref{def1} that $\mathfrak{h}$ and $\mathfrak{l}$ admit Cartan decompositions
$$\mathfrak{h}=\mathfrak{k}_{1} + \mathfrak{p}_{1} \ \text{and} \ \mathfrak{l}=\mathfrak{k}_{2} + \mathfrak{p}_{2},$$
given by Cartan involutions $\theta_{1}, \ \theta_{2}$ of $\mathfrak{g}$ such that $\theta_{1} (\mathfrak{h})= \mathfrak{h}$ and $\theta_{2} (\mathfrak{l})= \mathfrak{l}.$ Let $\mathfrak{a}_{1} \subset \mathfrak{p}_{1}$ and $\mathfrak{a}_{2} \subset \mathfrak{p}_{2}$ be maximally split abelian subspaces of $\mathfrak{p}_{1}$ and $\mathfrak{p}_{2},$ respectively. One can show that there exist $a,b \in G$ such that $\mathfrak{a}_{\mathfrak{h}} := \text{\rm Ad}_{a}\mathfrak{a}_{1} \subset \mathfrak{a}$ and $\mathfrak{a}_{\mathfrak{l}} := \text{\rm Ad}_{b}\mathfrak{a}_{2} \subset \mathfrak{a}.$ Denote by $W_{\mathfrak{g}}$ the Weyl group of $\mathfrak{g}.$ In this setting the following holds
\begin{theorem}[Theorem 4.1 in \cite{kob2}] The following three conditions are equivalent
\begin{enumerate}
\item $L$ acts on $G/H$ properly.
\item $H$ acts on $G/L$ properly.
\item For any $w \in W_{\mathfrak{g}},$ $w\cdot \mathfrak{a}_{\mathfrak{l}} \cap \mathfrak{a}_{\mathfrak{h}} =\{ 0 \}.$
\end{enumerate}
\label{twkob}
\end{theorem}
\noindent
Note that the criterion 3. in Theorem \ref{twkob} depends on how $L$ and $H$ are embedded in $G$ up to inner-automorphisms. Theorem \ref{twkob} provides a partial answer to Q1.
\begin{corollary}[Corollary 4.2 in \cite{kob2}] If $L$ acts properly on $G/H$ then
$$\text{\rm rank}_{\mathbb{R}}(\mathfrak{l}) + \text{\rm rank}_{\mathbb{R}}(\mathfrak{h}) \leq \text{\rm rank}_{\mathbb{R}} (\mathfrak{g}).$$
\label{coko}
\end{corollary}
\noindent
Hence the real rank of $L$ is bounded by a constant which depends on $G/H,$ no matter how $H$ and $L$ are embedded in $G.$ In this paper we find a similar, stronger restriction for Lie groups $G,H,L$ by means of a certain tool which we call the a-hyperbolic rank (see Section 2, Definition \ref{dd2} and Table \ref{tab1}). In more detail we prove the following
\begin{theorem}
If $L$ acts properly on $G/H$ then
$$\mathop{\mathrm{rank}}\nolimits_{\text{\rm a-hyp}}(\mathfrak{l}) + \mathop{\mathrm{rank}}\nolimits_{\text{\rm a-hyp}}(\mathfrak{h}) \leq \mathop{\mathrm{rank}}\nolimits_{\text{\rm a-hyp}} (\mathfrak{g}).$$
\label{twgl}
\end{theorem}
Recall that a homogeneous space $G/H$ of reductive type admits a \textbf{\textit{compact Clifford-Klein form}} if there exists a discrete subgroup $\Gamma \subset G$ such that $\Gamma$ acts properly on $G/H$ and $\Gamma \backslash G/H$ is compact. The space $G/H$ admits a \textbf{\textit{standard compact Clifford-Klein form}} in the sense of Kassel-Kobayashi \cite{kako} if there exists a subgroup $L$ reductive in $G$ such that $L$ acts properly on $G/H$ and $L \backslash G/H$ is compact. In the latter case, for any discrete cocompact subgroup $\Gamma ' \subset L,$ the space $\Gamma ' \backslash G/H$ is a compact Clifford-Klein form. Therefore it follows from Borel's theorem (see \cite{bor}) that any homogeneous space of reductive type admitting a standard compact Clifford-Klein form also admits a compact Clifford-Klein form.
\newline
It is not known if the converse statement holds, but all known reductive homogeneous spaces $G/H$ admitting compact Clifford-Klein forms also admit standard compact Clifford-Klein forms.
As a corollary to Theorem \ref{twgl}, we get examples of the semisimple symmetric spaces without standard compact Clifford-Klein forms. In particular, we cannot find the first example in the existing literature:
\begin{corollary}
The homogeneous spaces $G/H=SL(2k+1, \mathbb{R})/SO(k-1,k+2)$ \linebreak and $G/H=SL(2k+1, \mathbb{R})/Sp(k-1,\mathbb{R})$ for $k \geq 5$ do not admit standard compact Clifford-Klein forms.
\label{co1}
\end{corollary}
\begin{remark}
Let us mention the following results, related to the above corollary.
\begin{itemize}
\item T. Kobayashi proved in \cite{kobadm} that $SL(2k,\mathbb{R})/SO(k,k)$ for $k\geq 1$ and $SL(n,\mathbb{R})/Sp(l,\mathbb{R})$ \linebreak for $0<2l \leq n-2$ do not admit compact Clifford-Klein forms.
\item Y. Benoist proved in \cite{ben} that $SL(2k+1,\mathbb{R})/SO(k,k+1)$ for $k\geq 1$ does not admit compact Clifford-Klein forms.
\item Y. Morita proved recently in \cite{mor} that $SL(p+q,\mathbb{R})/SO(p,q)$ does not admit compact Clifford-Klein forms if $p$ and $q$ are both odd.
\end{itemize}
Note that these works are devoted to the problem of existence of compact Clifford-Klein forms on a given homogeneous space (not only standard compact Clifford-Klein forms).
\end{remark}
\section{The a-hyperbolic rank and antipodal hyperbolic orbits}
Let $\Sigma_{\mathfrak{g}}$ be a system of restricted roots for $\mathfrak{g}$ with respect to $\mathfrak{a}.$ Choose a system of positive roots $\Sigma^{+}_{\mathfrak{g}}$ for $\Sigma_{\mathfrak{g}}.$ Then a fundamental domain of the action of $W_{\mathfrak{g}}$ on $\mathfrak{a}$ can be define as
$$\mathfrak{a}^{+} := \{ X\in \mathfrak{a} \ | \ \alpha (X) \geq 0 \ \text{\rm for any} \ \alpha \in \Sigma^{+}_{\mathfrak{g}} \}.$$
Note that
$$sX+tY \in \mathfrak{a}^{+},$$
for any $s,t \geq 0$ and $X,Y\in \mathfrak{a}^{+}$. Therefore $\mathfrak{a}^{+}$ is a convex cone in the linear space $\mathfrak{a}.$
Let $w_{0} \in W_{\mathfrak{g}}$ be the longest element. One can show that
$$-w_{0}: \mathfrak{a} \rightarrow \mathfrak{a}, \ \ X \mapsto -(w_{0} \cdot X)$$
is an involutive automorphism of $\mathfrak{a}$ preserving $\mathfrak{a}^{+}.$ Let $\mathfrak{b} \subset \mathfrak{a}$ be the subspace of all fixed points of $-w_{0}$ and put
$$\mathfrak{b}^{+} := \mathfrak{b} \cap \mathfrak{a}^{+}.$$
Thus $\mathfrak{b}^{+}$ is a convex cone in $\mathfrak{a}.$ We also have $\mathfrak{b} = \text{Span} (\mathfrak{b}^{+}).$
\begin{definition}
The dimension of $\mathfrak{b}$ is called the a-hyperbolic rank of $\mathfrak{g}$ and is denoted by
$$\mathop{\mathrm{rank}}\nolimits_{\text{a-hyp}} (\mathfrak{g}).$$
\label{dd2}
\end{definition}
\noindent
The a-hyperbolic ranks of the simple real Lie algebras can be deduce from Table \ref{tab1}.
\begin{center}
\begin{table}[h]
\centering
{\footnotesize
\begin{tabular}{| c | c | c |}
\hline
\multicolumn{3}{|c|}{ \textbf{\textit{A-HYPERBOLIC RANK}}} \\
\hline
$\mathfrak{g}$ & $\mathop{\mathrm{rank}}\nolimits_{\text{a-hyp}} (\mathfrak{g})$ & $\text{rank}_{\mathbb{R}} (\mathfrak{g})$ \\
\hline
$\mathfrak{sl}(2k,\mathbb{R})$ & $k$ & $2k-1$ \\
{\scriptsize $k\geq 2$} & & \\
\hline
$\mathfrak{sl}(2k+1,\mathbb{R})$ & $k$ & $2k$ \\
{\scriptsize $k\geq 1$} & & \\
\hline
$\mathfrak{su}^{\ast}(4k)$ & $k$ & $2k-1$ \\
{\scriptsize $k\geq 2$} & & \\
\hline
$\mathfrak{su}^{\ast}(4k+2)$ & $k$ & $2k$ \\
{\scriptsize $k\geq 1$} & & \\
\hline
$\mathfrak{so}(2k+1,2k+1)$ & $2k$ & $2k+1$ \\
{\scriptsize $k\geq 2$} & & \\
\hline
$\mathfrak{e}_{6}^{\text{I}}$ & 4 & 6 \\
\hline
$\mathfrak{e}_{6}^{\text{IV}}$ & 1 & 2 \\
\hline
\end{tabular}
}
\caption{
This table contains all real forms of simple Lie algebras $\mathfrak{g}^{\mathbb{C}}$ for which $\text{rank}_{\mathbb{R}}(\mathfrak{g}) \neq \mathop{\mathrm{rank}}\nolimits_{\text{a-hyp}}(\mathfrak{g}).$ The notation is close to \cite{ov2}, Table 9, pages 312-316.
}
\label{tab1}
\end{table}
\end{center}
\noindent
A method of calculation of the a-hyperbolic rank of a simple Lie algebra can be found in \cite{bt}. The a-hyperbolic rank of a semisimple Lie algebra equals the sum of a-hyperbolic ranks of all its simple parts. For a reductive Lie algebra $\mathfrak{g}$ we put
$$\mathop{\mathrm{rank}}\nolimits_{\text{a-hyp}} (\mathfrak{g}) := \mathop{\mathrm{rank}}\nolimits_{\text{a-hyp}}([\mathfrak{g},\mathfrak{g}]).$$
There is a close relation between $\mathfrak{b}^{+}$ and the set of antipodal hyperbolic orbits in $\mathfrak{g}.$ We say that an element $X \in \mathfrak{g}$ is {\it \textbf{hyperbolic}}, if $X$ is semisimple (that is, $\mathrm{ad}_{X}$ is diagonalizable) and all eigenvalues of $\mathrm{ad}_{X}$ are real.
\begin{definition}
An adjoint orbit $O_{X}:=\mathop{\mathrm{Ad}}\nolimits (G)(X)$ is said to be hyperbolic if $X$ (and therefore every element of $O_{X}$) is hyperbolic. An adjoint orbit $O_{Y}$ is antipodal if $-Y\in O_{Y}$ (and therefore for \linebreak every $Z\in O_{Y},$ $-Z\in O_{Y}$).
\end{definition}
\begin{lemma}[c.f. Fact 5.1 and Lemma 5.3 in \cite{ok}]
There is a bijective correspondence between antipodal hyperbolic orbits $O_{X}$ in $\mathfrak{g}$ and elements $Y \in \mathfrak{b}^{+}.$ This correspondence is given by
$$\mathfrak{b}^{+}\ni Y \mapsto O_{Y}.$$
Furthermore, for every hyperbolic orbit $O_{X}$ in $\mathfrak{g}$ the set $O_{X} \cap \mathfrak{a}$ is a single $W_{\mathfrak{g}}$ orbit in $\mathfrak{a}$.
\label{lma}
\end{lemma}
\section{The main result}
We need two basic facts from linear algebra.
\begin{lemma}
Let $V_{1},V_{2}$ be vector subspaces of a real linear space $V$ of finite dimension. Then
$$\text{\rm dim} (V_{1}+V_{2})= \text{\rm dim} (V_{1}) + \text{\rm dim} (V_{2}) - \text{\rm dim} (V_{1}\cap V_{2}).$$
\label{lma1}
\end{lemma}
\begin{lemma}
Let $V_{1},...,V_{n}$ be a collection of vector subspaces of a real linear space $V$ of a finite dimension and let $A^{+} \subset V$ be a convex cone. Assume that
$$A^{+} \subset \bigcup_{k=1}^{n} V_{k}.$$
Then there exists a number $k,$ such that $A^{+} \subset V_{k}.$
\label{lma2}
\end{lemma}
We also need the following, technical lemma. Choose a subalgebra $\mathfrak{h}$ reductive in $\mathfrak{g}$ which corresponds to a Lie group $H \subset G.$ Let $\mathfrak{b}_{[\mathfrak{h},\mathfrak{h}]}^{+} \subset \mathfrak{a}_{\mathfrak{h}}$ be the convex cone constructed according to the procedure described in the previous subsection (for $[\mathfrak{h},\mathfrak{h}]$).
\begin{lemma}
Let $X\in \mathfrak{b}_{[\mathfrak{h},\mathfrak{h}]}^{+}.$ The orbit $O_{X}:=\mathop{\mathrm{Ad}}\nolimits (G)(X)$ is an antipodal hyperbolic orbit in $\mathfrak{g}.$
\label{lma3}
\end{lemma}
\begin{proof}
By Lemma \ref{lma} the vector $X$ defines an antipodal hyperbolic orbit in $\mathfrak{h}.$ Therefore we can find $h \in H \subset G$ such that
$\text{\rm Ad}_{h}(X) = - X$. Since a maximally split abelian subspace $\mathfrak{a} \subset \mathfrak{g}$ consists of vectors for which $\mathrm{ad}$ is diagonalizable with real values and
$$X \in \mathfrak{b}_{[\mathfrak{h},\mathfrak{h}]}^{+} \subset \mathfrak{a}_{\mathfrak{h}} \subset \mathfrak{a},$$
thus the vector $X$ is hyperbolic in $\mathfrak{g}.$ It follows that $\mathop{\mathrm{Ad}}\nolimits (G)(X)$ is a hyperbolic orbit in $\mathfrak{g}$ \linebreak and $-X \in \mathop{\mathrm{Ad}}\nolimits (G)(X).$
\end{proof}
Now we are ready to give a proof of Theorem \ref{twgl}.
\begin{proof}
Assume that $\mathop{\mathrm{rank}}\nolimits_{\text{a-hyp}}(\mathfrak{l}) + \mathop{\mathrm{rank}}\nolimits_{\text{a-hyp}}(\mathfrak{h}) > \mathop{\mathrm{rank}}\nolimits_{\text{a-hyp}} (\mathfrak{g})$ and let $\mathfrak{b}_{[\mathfrak{h},\mathfrak{h}]}^{+},$ $\mathfrak{b}_{[\mathfrak{l},\mathfrak{l}]}^{+},$ $\mathfrak{b}^{+}$ be appropriate convex cones. If $X \in \mathfrak{b}_{[\mathfrak{h},\mathfrak{h}]}^{+}$ then $O_{X}^{H} := \mathop{\mathrm{Ad}}\nolimits (G)(X)$ is an antipodal hyperbolic orbit in $\mathfrak{h}.$ By Lemma \ref{lma3} the orbit $O_{X}^{G} := \mathop{\mathrm{Ad}}\nolimits (G)(X)$ is an antipodal hyperbolic orbit in $\mathfrak{g}.$ By Lemma \ref{lma} there exists $Y \in \mathfrak{b}^{+}$ such that
$$O_{X}^{G} = O_{Y}^{G} = \mathop{\mathrm{Ad}}\nolimits (G)(Y).$$
Since $\mathfrak{b}_{[\mathfrak{h},\mathfrak{h}]}^{+} \subset \mathfrak{a}_{\mathfrak{h}} \subset \mathfrak{a}$ and $\mathfrak{b}^{+} \subset \mathfrak{a}$ thus (according to Lemma \ref{lma}) we get $X=w_{1} \cdot Y$ for a certain $w_{1} \in W_{\mathfrak{g}}.$ Therefore
$$\mathfrak{b}_{[\mathfrak{h},\mathfrak{h}]}^{+} \subset W_{\mathfrak{g}} \cdot \mathfrak{b}^{+} = \bigcup_{w\in W_{\mathfrak{g}}} w \cdot \mathfrak{b}^{+} \subset \bigcup_{w\in W_{\mathfrak{g}}} w \cdot \mathfrak{b}.$$
Analogously
$$\mathfrak{b}_{[\mathfrak{l},\mathfrak{l}]}^{+} \subset \bigcup_{w\in W_{\mathfrak{g}}} w \cdot \mathfrak{b}^{+} \subset \bigcup_{w\in W_{\mathfrak{g}}} w \cdot \mathfrak{b}.$$
By Lemma \ref{lma2} there exist $w_{\mathfrak{h}},w_{\mathfrak{l}} \in W_{\mathfrak{g}}$ such that
$$\mathfrak{b}_{[\mathfrak{h},\mathfrak{h}]}^{+} \subset w_{\mathfrak{h}}^{-1} \cdot \mathfrak{b} \ \ \text{\rm and} \ \ \mathfrak{b}_{[\mathfrak{l},\mathfrak{l}]}^{+} \subset w_{\mathfrak{l}}^{-1} \cdot \mathfrak{b},$$
because $W_{\mathfrak{g}}$ acts on $\mathfrak{a}$ by linear transformations. Therefore
$$\mathfrak{b}_{[\mathfrak{h},\mathfrak{h}]} \subset w_{\mathfrak{h}}^{-1} \cdot \mathfrak{b} \ \ \text{\rm and} \ \ \mathfrak{b}_{[\mathfrak{l},\mathfrak{l}]} \subset w_{\mathfrak{l}}^{-1} \cdot \mathfrak{b}$$
where $\mathfrak{b}_{[\mathfrak{h},\mathfrak{h}]} := \text{Span}(\mathfrak{b}_{[\mathfrak{h},\mathfrak{h}]}^{+})$ and $\mathfrak{b}_{[\mathfrak{l},\mathfrak{l}]} := \text{Span} (\mathfrak{b}_{[\mathfrak{l},\mathfrak{l}]}^{+}).$ We obtain
$$w_{\mathfrak{h}} \cdot \mathfrak{b}_{[\mathfrak{h},\mathfrak{h}]} \subset \mathfrak{b} , \ w_{\mathfrak{l}} \cdot \mathfrak{b}_{[\mathfrak{l},\mathfrak{l}]} \subset \mathfrak{b}.$$
By the assumption and Lemma \ref{lma1}
$$\text{\rm dim}(w_{\mathfrak{h}} \cdot \mathfrak{b}_{[\mathfrak{h},\mathfrak{h}]} \cap w_{\mathfrak{l}} \cdot \mathfrak{b}_{[\mathfrak{l},\mathfrak{l}]}) >0.$$
Choose $0 \neq Y \in w_{\mathfrak{h}} \cdot \mathfrak{b}_{[\mathfrak{h},\mathfrak{h}]} \cap w_{\mathfrak{l}} \cdot \mathfrak{b}_{[\mathfrak{l},\mathfrak{l}]}.$ Then
$$w_{\mathfrak{l}} \cdot X_{\mathfrak{l}}=Y= w_{\mathfrak{h}} \cdot X_{\mathfrak{h}} \ \text{\rm for some} \ X_{\mathfrak{h}} \in \mathfrak{b}_{[\mathfrak{h},\mathfrak{h}]}\backslash \{ 0 \} \ \text{\rm and} \ X_{\mathfrak{l}} \in \mathfrak{b}_{[\mathfrak{l},\mathfrak{l}]}\backslash \{ 0 \}. $$
Take $w_{2} := w_{\mathfrak{h}}^{-1}w_{\mathfrak{l}} \in W_{\mathfrak{g}},$ we have $X_{\mathfrak{h}}= w_{2} \cdot X_{\mathfrak{l}}$ and $X_{\mathfrak{h}} \in \mathfrak{a}_{\mathfrak{h}}, \ X_{\mathfrak{l}} \in \mathfrak{a}_{\mathfrak{l}}.$ Thus $0 \neq X_{\mathfrak{h}} \in w_{2} \cdot \mathfrak{a}_{\mathfrak{l}} \cap \mathfrak{a}_{\mathfrak{h}}.$ The assertion follows from Theorem \ref{twkob}.
\end{proof}
We can proceed to a proof of Corollary \ref{co1}. For a reductive Lie group $D$ with a Lie algebra $\mathfrak{d}$ with a Cartan decomposition
$$\mathfrak{d} = \mathfrak{k}_{\mathfrak{d}} + \mathfrak{p}_{\mathfrak{d}}$$
we define $d(G) := \text{dim} (\mathfrak{p}_{\mathfrak{d}}).$ We will need the following properties
\begin{theorem}[Theorem 4.7 in \cite{kob2}]
Let $L$ be a subgroup reductive in $G$ acting properly on $G/H.$ The space $L \backslash G /H$ is compact if and only if
$$d(L)+d(H)=d(G).$$
\label{twkk}
\end{theorem}
\begin{theorem}[\cite{yo}]
If $J \subset G$ is a semisimple subgroup then it is reductive in $G.$
\label{twy}
\end{theorem}
\begin{proposition}
Let $L \subset G$ be a semisimple Lie group acting properly on \linebreak $G/H=SL(2k+1, \mathbb{R})/SO(k-1,k+2)$ or $G/H=SL(2k+1, \mathbb{R})/Sp(k-1,\mathbb{R}).$ Then
$$\text{\rm rank}_{\mathbb{R}}(\mathfrak{l}) \leq 2.$$
\label{p1}
\end{proposition}
\begin{proof}
Because $\mathop{\mathrm{rank}}\nolimits_{\text{a-hyp}}(\mathfrak{g}) = 1+ \mathop{\mathrm{rank}}\nolimits_{\text{a-hyp}}(\mathfrak{h})$ thus it follows from Table \ref{tab1} and Theorem \ref{twgl} that if $L$ is simple then $\text{rank}_{\mathbb{R}}(\mathfrak{l}) \leq 2.$ On the other hand if $L$ is semisimple then each (non-compact) simple part of $\mathfrak{l}$ adds at least $1$ to a-hyperbolic rank of $\mathfrak{l}.$ Thus we also have $\text{rank}_{\mathbb{R}}(\mathfrak{l}) \leq 2.$
\end{proof}
\textit{Proof of Corollary \ref{co1}.} Assume now that $L$ is reductive in $G.$ Since the Lie algebra $\mathfrak{l}$ is reductive therefore
$$\mathfrak{l} = \mathfrak{c}_{\mathfrak{l}} + [\mathfrak{l},\mathfrak{l}],$$
where $\mathfrak{c}_{\mathfrak{l}}$ denotes the center of $\mathfrak{l}.$ It follows from Corollary \ref{coko} that
\begin{equation}
\text{\rm rank}_{\mathbb{R}}(\mathfrak{l}) \leq k+1,
\label{eq2}
\end{equation}
and by Proposition \ref{p1} we have $\text{rank}_{\mathbb{R}}([\mathfrak{l},\mathfrak{l}]) \leq 2.$ Note that
$$d(G)-d(H)\geq k^{2} +2k +2.$$
We will show that if $L$ acts properly on $G/H$ and $k\geq 5$ then
\begin{equation}
d(L) < k^{2} + 2k +2.
\label{eq4}
\end{equation}
Let $[\mathfrak{l},\mathfrak{l}] = \mathfrak{k}_{0} + \mathfrak{p}_{0}$ be a Cartan decomposition. From (\ref{eq2})
\begin{equation}
d(L) \leq \text{\rm dim} (\mathfrak{c}_{\mathfrak{l}}) + \text{\rm dim} (\mathfrak{p}_{0}) \leq k+1 + \text{\rm dim} (\mathfrak{p}_{0}).
\label{eq7}
\end{equation}
Also, if $\text{rank}_{\mathbb{R}}([\mathfrak{l},\mathfrak{l}]) =2$ then it follows from Table \ref{tab1} that (the only) non-compact simple part of $[\mathfrak{l},\mathfrak{l}]$ is isomorphic to $\mathfrak{sl}(3,\mathbb{R}),$ $\mathfrak{su}^{\ast}(6),$ $\mathfrak{e}_{6}^{\text{IV}}$ or $\mathfrak{sl}(3,\mathbb{C})$ (treated as a simple real Lie algebra). In such case
\begin{equation}
\text{\rm dim} (\mathfrak{p}_{0}) < 27.
\label{eq5}
\end{equation}
Therefore assume that $\text{rank}_{\mathbb{R}} ([\mathfrak{l},\mathfrak{l}])=1$ and let $\mathfrak{s} \subset [\mathfrak{l},\mathfrak{l}]$ be (the only) simple part of a non-compact type. We have
\begin{equation}
\text{\rm rank}_{\mathbb{R}} (\mathfrak{s}) =1.
\label{eq3}
\end{equation}
It follows from Theorem \ref{twy} that $\mathfrak{s}$ is reductive in $\mathfrak{g}.$ Therefore $\mathfrak{s}$ admits a Cartan decomposition
$$\mathfrak{s} = \mathfrak{k}_{\mathfrak{s}} + \mathfrak{p}_{\mathfrak{s}}$$
compatible with $\mathfrak{g}= \mathfrak{k} + \mathfrak{p},$ that is $\mathfrak{k}_{\mathfrak{s}} \subset \mathfrak{k}.$ We also have $\text{dim}(\mathfrak{p}_{s})=\text{dim}(\mathfrak{p}_{0}).$ Since $\mathfrak{k} = \mathfrak{so}(2k+1)$ we obtain
$$\text{\rm rank} (\mathfrak{k}_{s}) \leq \text{\rm rank} (\mathfrak{k}) = k.$$
Using the above condition together with (\ref{eq3}) we can check (by a case-by-case study of simple Lie algebras) that
\begin{equation}
\text{\rm dim} (\mathfrak{p}_{\mathfrak{s}}) < 4k.
\label{eq6}
\end{equation}
Now (\ref{eq7}), (\ref{eq5}) and (\ref{eq6}) imply that
$$d(L) < 5k+1$$
for $k \geq 6,$ and $d(L)<33$ for $k=5.$ Thus we have showed (\ref{eq4}). The assertion follows from Theorem \ref{twkk}.
|
\section{Introduction}
\label{intro}
Collisional ring galaxies are scarce, yet very interesting objects. Their most prominent feature is the lack
of a typical spiral structure, replaced by a narrow, ring--shaped accumulation of gas and stars. Such a peculiar distribution
of matter is believed to form during collision of a spiral, gas--rich galaxy with a small, early type one (\citealt{LT};
\citealt{TS}).
Galaxy pair consisting of NGC\,2444 and 2445 is one of the few systems of ring galaxies that have been
included in the Arp's Catalogue, denoted Arp\,143 \citep{arp}. However, it differs significantly from the usual image of
a ring system, as the ring structure in NGC\,2445 is distorted, similar in appearance to a trapezoid with rounded vertices. The
whole visible structure is dominated by local maxima of optical emission that are regions of intensive star formation
\citep{burridge, tail, multiv}. All of them are relatively young, and the central region is suspected to be undergoing starburst
activity \citep{multiv}. Moreover, the pair is known to be "traling smoke", as the {\rm H}{\sc i} morphology \citep{tail}
reveals a 150\,kpc long tail of the neutral gas emission extending towards north. All this is suggestive for a collision of
NGC\,2445 with a compact galaxy -- possibly its companion, NGC\,2444 \citep{multiv,beirao}. This collision disrupted the spiral disk
and resulted in morphological distortions as well as in intensification of the star formation in particular regions of the collisional ring.\\
Not much is known about the radio emission from Arp\,143. In fact, the only work that aimed at revealing the morphology of
the radio-emitting medium of this object was that of \citet{burmil}, who used the Westerbork Synthesis Radio Telescope to study
several interacting galaxies. These authors identified a source of radio emission within NGC\,2445, associating it with one of the
{\rm H}{\sc ii} regions. However, they excluded the possibility of a purely thermal origin of the emission, as the luminosity was far
too high to be caused by thermal processes only. Apart from that paper, the only available radio data was that from large surveys (e.g.
\citealt{sulentic} or \citealt{davis}). \citet{colrings} mention a study of Arp\,143 (as well as of ten other ring systems) made with the
Very Large Array (VLA) at the frequency of 8440\,MHz, but no detailed information is provided. Using Arp\,10 as an example, these Authors
suggest that the spectral index of the emission from ring galaxies might steepen inwards from the ring. Such phenomenon could be a result of
a change from thermal to non-thermal emission and/or of ageing of the highest energy electrons, which leads to their absence in the inner
region of the ring. However, almost nothing is known about the magnetic field and its properties. Also, cross--identifications between
radio-emitting structures and their counterparts in other domains of the electromagnetic spectrum have not yet been made.\\
In this paper we present results from our recent observing project at the Giant Metrewave Radio Telescope (GMRT), in which Arp\,143 has
been studied at 234 and 612\,MHz. The new observations are analysed together with the previously unpublished, archive VLA data at 1490,
4860, 8440, and 14940\,MHz to produce high resolution maps of the radio emission and to study the magnetic field of this galaxy pair.
\begin{figure*}
\resizebox{\hsize}{!}{\includegraphics{rgb.ps}}
\caption{
GMRT map of the TP emission from Arp\,143 at 612\,MHz overlaid upon an RGB image,
with locations of the star-forming regions and intergalactic ridge indicated.
The contour levels are $3,5,10,20,30,50,100 \times$ 0.1\,mJy/beam
(r.m.s. noise level). The angular resolution is 10\,arcsec.
The beam is represented by a circle in the lower left corner of the image.
Details of the maps used to produce this RGB composite can be found in the text
(Sect.~\ref{result}).
}
\label{RGB}
\end{figure*}
\section{Observations and data reduction}
\label{observ}
\subsection{GMRT data}
\label{obsgmrt}
The GMRT near Pune, India was used to observe Arp\,143 at 234 and 612\,MHz. The observations were carried out in a dual
frequency mode as a part of our project ``Magnetic field and galaxy interactions -- from loose
groups to mergers'' (project code 23\_025). The observations were carried out in February, 2013.
The total observing time was 8\,h and the bandwidths were 16 and 32\,MHz at 234 and 612\,MHz, respectively. The (u,v) data
were reduced using the Astronomical Image Processing System (\textsc{aips}), including calibration and RFI--flagging. After
obtaining initial images at both frequencies, they were processed by a self-calibration pipeline in order to correct
the phase information. The final images have been (u,v)-tapered to obtain circular beams and then corrected for the primary
beam shape. Two images have been made: one with resolution of 16\,arcsec at 234\,MHz, and another at 612\,MHz, with
resolution of 10\,arcsec. Original beam sizes and noise levels of these maps can be found in Table~\ref{beamsizes}.
\begin{table*}
\caption{\label{beamsizes}Basic information on the observational datasets used in this study}
\begin{center}
\begin{tabular}{rrrrrrrl}
\hline
\hline
Freq. [MHz] & Telescope & Proj. code & Date & TOS [h] & Org. beam [arcsec] & Fin. beam [arcsec] &
Noise [mJy\slash beam]\\
\hline
234 & GMRT & 23\_025 & 10.02.2013 & 8.0 & 14x11 & 16 & 0.4 \\
612 & GMRT & 23\_025 & 10.02.2013 & 8.0 & 7.5x5 & 10 & 0.1 \\
1490 & VLA B & AD 182 & 05.09.1986 & 1.5 & 4x4 & 10 & 0.08\\
4860 & VLA C & AD 182 & 14.11.1986 & 2.7 & 4x4 & 10 & 0.02\\
8440 & VLA D & AA 146 & 25.08.1992 & 0.5 & 9x7 & 10 & 0.06\\
14940 & VLA C & AJ 105 & 29.05.1984 & 2.8 & 1.3x1.1 & 1.3x1.1 & 0.08\\
\hline
\end{tabular}
\end{center}
\end{table*}
\subsection{Archive VLA data}
\label{obsvla}
In order to construct the radio continuum spectrum, and to derive the spectral age as well as to provide the magnetic field
strength estimates for Arp\,143, we searched through the archive of the VLA of the National Radio Astronomy Observatory (NRAO)
\footnote{NRAO is a facility of National Science
Foundation operated under cooperative agreement by Associated Universities, Inc.}. We have obtained archived
data at 1490, 4860, 8440, and 14940\,MHz. Details on these observations can be found in Table~\ref{beamsizes}. All
these sets were calibrated, flagged and imaged using the \textsc{aips}. The 4860\,MHz data allowed to run self-calibration.
Except for the highest frequency, all the data sets were (u,v)--tapered to obtain a circular beam of 10\arcsec.
The 14940\,MHz image has not been tapered, as it was intended to be used only to derive the size of the galactic core of
NGC\,2445 and its flux. Additionally, the size of the primary beam at this frequency is nearly the same as the size of NGC\,2445,
resulting in non-reliable flux values besides the very center of this galaxy.
To include the calibration uncertainties, we have assumed a 5 per cent error for each integrated flux value for all the radio maps
except the 234\,MHz map, for which we adopt 8 per cent error.
\section{Results}
\label{result}
Fig.~\ref{RGB} shows a composite RGB image built from \textit{u, g, r} bands of the Sloan Digital Sky Survey (SDSS) data. An additional,
H$\alpha$ component (map from \citealt{romano}, taken from the NASA Extragalactic Database) was added to the red channel. Before merging,
the H$\alpha$ map was rescaled to have signal values significantly higher than the median value of the \textit{r} band emission in order to make the
regions of molecular emission easily distinguishable. The colour image was then overlaid with the contours of
radio emission at 612\,MHz. Designations of different structures described in this paper -- namely southern (S), northwestern (NW) and eastern
(E) star-forming regions, intergalactic ridge and both galaxies -- have also been marked. Measured fluxes of these structures are presented
in Table~\ref{fluxess}.
\begin{table*}
\caption{Flux densities (with errors) obtained for the selected regions of radio emission. All values given in mJy. ND means no detection.}
\begin{center}
\begin{tabular}[]{lccccccc}
\hline
\hline
Region & 234 MHz & 612 MHz & 1420 MHz & 4860 MHz & 8440 MHz & 14960 MHz \\
\hline
Core & 15.58 $\pm$ 1.40 & 11.14 $\pm$ 0.57 & 6.92 $\pm$ 0.36 & 3.59 $\pm$ 0.26 & 1.86 $\pm$ 0.11 & 0.81 $\pm$ 0.09 \\
NW & 30.97 $\pm$ 2.61 & 20.07 $\pm$ 1.01 & 9.33 $\pm$ 0.49 & 3.24 $\pm$ 0.16 & 1.39 $\pm$ 0.12 & ND \\
S & ND & 3.57 $\pm$ 0.25 & ND & 0.57 $\pm$ 0.04 & ND & ND \\
E & ND & 0.66 $\pm$ 0.11 & ND & 0.08 $\pm$ 0.02 & ND & ND \\
Ridge & 3.30 $\pm$ 0.69 & 1.28 $\pm$ 0.10 & ND & 0.16 $\pm$ 0.03 & ND & ND \\
\hline
\end{tabular}
\end{center}
\label{fluxess}
\end{table*}
The lowest frequency used in our study is 234\,MHz and the map at this frequency is presented in Fig.~\ref{pband}. It has the lowest
resolution (circular beam of 16 arcseconds in diameter). The emission concentrates in the central-northern region of NGC\,2445 and extends
northwest, vanishing in the outer regions of NGC\,2444 (which is otherwise a non-radio-emitting galaxy). The central part consists of the core
and the NW region. Because of the beam size, the emission coming from these two entities is not separated. The radio emission coincides almost
perfectly with the optically--emitting material in the northern part of the ring, but diminishes abruptly about 10 arcseconds south from the core
of NGC\,2445. Only an isolated patch of emission coincident with the southern star-forming region was detected. This is caused by the lower sensitivity
to faint, extended structures at 234\,MHz compared to 612\,MHz.
\begin{figure*}
\begin{subfigure}[p]{0.4\textwidth}
\includegraphics[width=\textwidth]{pbandN.ps}
\captionsetup{width=\textwidth}
\caption{
GMRT map at 234\,MHz.
The r.m.s. noise level is\\ 0.4\,mJy/beam.
The angular resolution is 16\,arcsec.
}
\label{pband}
\end{subfigure}%
\begin{subfigure}[p]{0.4\textwidth}
\includegraphics[width=\textwidth]{gbandN.ps}
\captionsetup{width=0.95\textwidth}
\caption{
GMRT map at 612\,MHz.
The r.m.s. noise level is 0.1\,mJy/beam.
The angular resolution is 10\,arcsec.
}
\label{gband}
\end{subfigure}
\begin{subfigure}[p]{0.4\textwidth}
\includegraphics[width=\textwidth]{lbandN.ps}
\captionsetup{width=\textwidth}
\caption{
VLA map at 1490\,MHz.
The r.m.s. noise level is\\ 0.08\,mJy/beam.
The angular resolution is 10\,arcsec.
}
\label{lband}
\end{subfigure}%
\begin{subfigure}[p]{0.4\textwidth}
\includegraphics[width=\textwidth]{cbandN.ps}
\captionsetup{width=0.95\textwidth}
\caption{
VLA map at 4860\,MHz.
The r.m.s. noise level is 0.02\,mJy/beam.
The angular resolution is 10\,arcsec.
}
\label{cband}
\end{subfigure}
\begin{subfigure}[p]{0.4\textwidth}
\includegraphics[width=\textwidth]{xbandN.ps}
\captionsetup{width=\textwidth}
\caption{
VLA map at 8440\,MHz.
The r.m.s. noise level is\\ 0.06\,mJy/beam.
The angular resolution is 10\,arcsec.
}
\label{xband}
\end{subfigure}%
\begin{subfigure}[p]{0.4\textwidth}
\includegraphics[width=\textwidth]{ubandN.ps}
\captionsetup{width=0.95\textwidth}
\caption{
VLA map of the core at 14940\,MHz.
The r.m.s. noise level is 0.08\,mJy/beam.
The angular resolution is 1.28x1.13\,arcsec.
}
\label{uband}
\end{subfigure}
\caption{
Maps of the TP emission from Arp\,143 at various frequencies overlaid upon an SDSS {\it g}-band image.
The contour levels are $-3$ (dashed), $3,5,10,20,50,100\times$ r.m.s. noise level.
The beam is represented by an ellipse in the lower left corner of the image.
}
\end{figure*}
The map at 612\,MHz (Fig.~\ref{gband}) has a modest resolution of 10 arcsec and a long integration time, resulting in a deep and detailed radio map.
Despite higher resolution and frequency, this map shows more extended emission than that at 234\,MHz. Apart from the core and the NW region, two
other star-forming areas can be easily identified. Both in the eastern and the southern part of the ring, local maxima of the visible light
distribution are coincident with peaks of the radio emission; the eastern structure is the weakest one. The northern extension that forms a
ridge--like structure between NGC\,2444 and NGC\,2445 is also visible; it extends even further than that at 234\,MHz, reaching NGC\,2444.
The map at 1490\,MHz (made from archive data) is presented in Fig.~\ref{lband}. Short integration time and lack of the shortest baselines (due to the
wide B-configuration of the VLA) resulted in the absence of most of the extended emission as well as in higher noise level. Only the core and the NW region
are visible. There is no trace of emission from other parts of the collisional ring, nor from the aforementioned intergalactic ridge.
It should be noted that the archive VLA data used in this study are of worse (u,v) coverage than the GMRT ones, resulting in lower detectability of the weak, extended
emission. This mostly applies to the 1490 and 8440\,MHz data, for which the (u,v) plane sampling is poor due to the integration times lower than
2 hours.
Considerably better (u,v) coverage and lack of strong RFI allowed to obtain a detailed map of the extended emission at 4860\,MHz (Fig.~\ref{cband}).
Similarly to the 612\,MHz map, not only the emission from the NW star formation region
and the core can be seen, but the whole collisional ring is emitting at this frequency. From its structure one can distinguish local maximum corresponding
to the southern star forming region. Surprisingly, the eastern star forming region is indistinguishable from the extended emission. A patch of emission
north from the ring, spatially coincident with the ridge detected both at 234 at 612\,MHz is easily visible. It should be noted here that, unlike at 612\,MHz,
the inner part of the galaxy -- an area between the ring and the core -- is not visible in emission at 4860\,MHz; this effect is more pronounced in the
spectral index map (Fig.~\ref{spixGC}), indicating a significantly steeper spectrum there.
At 8440\,MHz the NW region is still visible, albeit significantly less prominent than at lower frequencies. Most of the extended emission has not been detected,
because of the very short integration time of these archival data. Despite that, there is some minor extension of the NW region into the northern part of the
collisional ring, but it barely exceeds the 3\,$\sigma$ level (Fig.~\ref{xband}). Throughout the external parts of the image several isolated patches of
noise--level radio emission, boosted by the primary beam correction are visible.
The primary beam of the observations at the highest frequency (14940\,MHz) is very small, therefore only the very central part of NGC\,2445 is visible.
Only the core emerges from the noise (Fig.~\ref{uband}). With the high resolution of this map, we can estimate the angular size of the radio-emitting region,
which is approximately 3x2\,arcseconds (at the 3\,$\sigma$ level). This (assuming the distance estimate of $54.9 \pm 3.8$\,Mpc from \citealt{devac}) yields
the linear size of approximately 0.8 by 0.5\,kpc. Similarly to the 8440\,MHz image (Fig.~\ref{xband}), loose patches of emission resulting from the primary
beam correction are visible in the outer parts of the image.
Large primary beams in our low-frequency observations allowed us to search for the continuum counterpart of the giant {\rm H}{\sc i} tail extending north
from the ring galaxy \citep{tail}. However, nothing was detected above the noise level in all the maps. Some remarks on the upper limit of the magnetic
field strength in this area can be found in Sect.~\ref{magfield}.
\section{Discussion}
\label{discus}
\subsection{Spectral index}
\label{spixtext}
The spectral index map has been calculated between 612 and 4860\,MHz and has a resolution of 10 arcsec. The input maps have been clipped
at the 5$\sigma$ level to ensure that the noise fluctuations would not be taken into account. The choice of this two maps -- instead of
the ones at the lowest and highest frequencies -- was dictated by their resolution and quality. The 234\,MHz map (Fig.~\ref{pband})
has a rather low resolution of 16 arcsec, which does not allow to clearly distinguish different emitting regions. Moreover, despite large
beamsize, it does not show larger extent of emission than the 612\,MHz map. The 8440\,MHz map (Fig.~\ref{xband}) has a modest resolution
of 10 arcsec, but it suffers from the short integration time, resulting in significant losses of the extended flux. This is not the case of
the 4860\,MHz map (Fig.~\ref{cband}), which clearly shows the ring structure as well as a patch spatially connected to the intergalactic
ridge. Therefore, the spectral index map -- as well as all spectral index values used and presented in this study -- have been calculated
between the best-quality 612 and 4860\,MHz maps. Throughout the paper we are using the $S_{\nu} \propto \nu^{-\alpha}$ definition of the
spectral index $\alpha$.
The emission from the core of NGC\,2445 has a steeper spectrum (a mean spectral index of $\approx 0.64 \pm 0.08$) than expected if it were
a purely thermal source. This indicates that it has a synchrotron origin, as we would expect the thermal emission to manifest with a
significantly flatter spectrum, which is not the case here (see \citealt{pacholczyk} for details).
The spectrum of the star forming regions is steeper, with the mean spectral indices typical for an ageing population of relativistic electrons.
The values for the particular regions are $\alpha \approx 0.81 \pm 0.07$ for the northwestern, $\alpha \approx 1.03 \pm 0.19$ for the eastern
and $\alpha \approx 0.89 \pm 0.06$ for the southern region. This means that the synchrotron emission dominates everywhere over the thermal component,
even at the higher of these frequencies.
This is not unusual; \citet{niklas28} have shown that the median thermal fraction at 1\,GHz is $ 0.08 \pm 0.01$. The median non-thermal
spectral index given by these authors is $0.83 \pm 0.02$, very close to the values derived for the star-forming regions in NGC\,2445.
This means domination of the high--energy, relativistic electrons supplied by the supernovae. \citet{colrings} suggest that
this process of electron supply would start some $10^6$ years after the collision. The estimated age of the density wave that gave birth to
the NW region is $\approx$ 85\,Myr \citep{beirao}, indicating that such scenario is possible. The inner part of the ring has not
been detected at 4860\,MHz, indicating a very steep spectrum there. Assuming the 3-r.m.s. level as a constraint on the emission at 4860\,MHz, we have
estimated the spectral index to be $> 1.8$ between 612 and 4860\,MHz. Steepening of the spectral index inwards from the ring was mentioned
by \citet{colrings}, who provided two possible explanations for such phenomenon: change from the mostly thermal to non-thermal radiation and/or
ageing of the synchrotron electrons. Our findings suggest that in case of Arp\,143 the second scenario is more probable.
The intergalactic ridge has a steep spectrum -- characteristic for an ageing population of electrons -- with a mean spectral index
$\alpha \approx 1.01 \pm 0.12$. This is consistent with its identification as an intergalactic structure presented in Sect.~\ref{result}. Also, it is
steeper than that of the star-forming regions NW and S, suggesting a higher spectral age, as it is far away from the star-forming
areas.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{spix_ryphlame.ps}}
\caption{
Map of the spectral index calculated between 612 and 4860\,MHz
with contours of the radio emission at 612\,MHz overlaid.
The contour levels are $-3$ (dashed), $3,5,10,20,30,50\times$ 0.02\,mJy/beam
(r.m.s. noise level). The angular resolution is 10\,arcsec.
}
\label{spixGC}
\end{figure}
\subsection{Magnetic field}
\label{magfield}
Both the strength of the magnetic field $B_{\mathrm{TOT}}$ and its energy density $E_{\mathrm{B}}$ in the selected areas were calculated
assuming the energy equipartition between the cosmic rays and the magnetic field, following the formulae presented in \citet {bfeld}.
We applied the \textsc{bfeld} code, which uses following parameters: total pathlength through the source $D$, proton--to--electron energy density ratio
$K_0$, spectral index $\alpha$, and the mean synchrotron surface brightness of the chosen region to estimate its total field strength as well as the
magnetic energy density. The mean brightness has been obtained from the region's integrated flux density $S$ by dividing it by the square root of the
number of the beams in the integration area. $K_0$ was fixed as 100; such a value is suggested also for starburst galaxies \citep{beckstarburst}. Values
of all parameters used, together with the results of the estimation are presented in Table~\ref{params}. The spectral index was calculated between the
maps at 612 and 4860\,MHz, which have the best quality.
According to \citet{niklas28}, thermal fraction at 1490\,MHz for majority of galaxies does not exceed 10 per cent; even at 10000\,MHz its mean value is about
25 per cent. This consistently yields thermal fraction at 4860\,MHz significantly less than some 15-20 per cent. Except for the nucleus, a low thermal fraction is
implied by steep radio spectra of discussed regions. Even in case of maximum value of 20 per cent, its neglecting leads to an overestimate of magnetic field
strength by no more than 5-7 percent. Therefore, we decided not to take thermal emission into account.
\begin{table*}
\caption{Parameters used for the estimation of the magnetic field properties and resulting values}
\begin{center}
\begin{tabular}[]{lccccccc}
\hline
\hline
Region & {\it D}\,[kpc]& $\alpha$ & $S_{4.86}$\,[mJy] &$B_{\mathrm{TOT}}$ [$\mu$G]&$E_{\mathrm{B}}$ [$\mathrm{erg}\,\mathrm{cm^{-3}}$] \\
\hline
Core & 0.65 $\pm$ 0.15 & 0.64 $\pm$ 0.08 & 3.24 $\pm$ 0.16 & 38.8 $\pm$ 1.7 & 60 $\pm$ 5 $\times 10^{-12}$ \\
NW & 5.5 $\pm$ 0.5 & 0.81 $\pm$ 0.07 & 3.59 $\pm$ 0.13 & 12.0 $\pm$ 0.4 & 5.8 $\pm$ 0.5 $\times 10^{-12}$ \\
S & 5.5 $\pm$ 0.5 & 0.89 $\pm$ 0.06 & 0.57 $\pm$ 0.04 & 9.9 $\pm$ 0.3 & 3.9 $\pm$ 0.3 $\times 10^{-12}$ \\
E & 5.5 $\pm$ 0.5 & 1.03 $\pm$ 0.19 & 0.08 $\pm$ 0.02 & 8.7 $\pm$ 0.6 & 3.0 $\pm$ 0.4 $\times 10^{-12}$ \\
Ridge & 5.5 $\pm$ 0.5 & 1.01 $\pm$ 0.12 & 0.16 $\pm$ 0.03 & 9.2 $\pm$ 0.6 & 3.3 $\pm$ 0.3 $\times 10^{-12}$ \\
\hline
\end{tabular}
\end{center}
\label{params}
\end{table*}
For the NW part of the ring, we have used a pathlength of 5500\,pc -- derived from the angular size of the emitting region under an
assumption of a cylindrical symmetry.
The same size has been also adopted for all the other regions within the disk of NGC\,2445 as well as for the ridge between it and
NGC\,2444. The estimated strength of the magnetic field in the NW region is therefore $12.0 \pm 0.4\,\mu$G,
and the energy density is $5.8 \pm 0.5 \times 10^{-12} \mathrm{erg}\,\mathrm{cm^{-3}}$. This means that the magnetic field in this
region is somewhat stronger than the average found in normal, spiral galaxies ($9 \pm 1.3\,\mu$G -- \citealt{niklas}). It is also
comparable to that found in the shock region between the galaxies forming the Stephan's Quintet ($11.0 \pm 2.2\,\mu$G -- \citealt{SQ}).
Values derived for the other star--forming regions are very similar (see Table~\ref{params}).
Estimates for the intergalactic ridge do not differ much from that for the star--forming regions. As its spectrum is steeper
than that of the star--forming regions (except for the eastern region, which has a comparable $\alpha$), the strength of the magnetic field
is nearly the same despite lower surface brightness. For an intergalactic structure, value of $\approx 9 \mu$G is a rather high strength.
There is no clear counterpart to this entity in any other spectral domain. Most possibly, the magnetic field --
enhanced in the star--forming NW region -- is dragged with the intergalactic gas during the tidal interaction with the companion galaxy.
Such a structure can result from a partially ordered B--field, dragged and stretched during the interaction.
The total magnetic field is, in terms of its strength and energy density, similar to that of the iconic colliding pair of galaxies, the
Taffies \citep{taffy}. Unfortunately, the setup of the low-frequency observations (simultaneous dual-frequency mode 234/612\,MHz)
did not allow to perform full Stokes observations, and we have not detected polarisation in any of the VLA datasets . Polarisation data is
necessary to confirm or reject this scenario.
As the core of the collisional ring galaxy exhibits a spectrum that suggests a non-thermal origin of the emission ($\alpha \approx 0.64 \pm 0.08$),
we derived estimates for the magnetic field of the core, too. We used a pathlength of 650\,pc -- a mean of its linear dimensions (see Sect.~\ref{result}).
The resulting field is strong, as it reaches $38.8 \pm 1.7\,\mu$G. Its energy density is equal to $6.0 \pm 0.5 \times 10^{-11} \mathrm{erg}\,\mathrm{cm^{-3}}$.
The core is supposed to undergo starburst activity \citep{multiv}, and for a compact region of efficient electron supply such a
number is not surprising.
The radio continuum counterpart for the tidal tail has not been detected, but we could obtain the upper limit for the magnetic field strength and
energy in its area. We have assumed that the tail has a cylindrical symmetry with a diameter of 20\,kpc \citep{tail} and a steep
spectral index of 1.0. Once again, we used the \textsc{bfeld} code, calculating the magnetic field strength using information at
234, 612 and 4860\,MHz. The upper limit for the strength was estimated as $< 4.6 \mu$G. The corresponding limit for the magnetic field
energy is $< 1.1\times 10^{-12} \mathrm{erg}\,\mathrm{cm^{-3}}$.
However, these values are of high uncertainty, as the estimates of the tail parameters are rather rough. Assuming the radio-emitting medium to be
more shallow (which is possible, as there are hints of narrowing in certain regions of the tail), the limit for the magnetic field strength
would rise up -- to approximately 8.5--12\,$\mu$G (in case of the depth of 5 and 1\, kpc, respectively) and its energy density would reach
2.8 -- 6.3 $\times 10^{-12} \mathrm{erg}\,\mathrm{cm^{-3}}$. This indicates that a relatively strong magnetic field could remain undetected.
\subsection{Age of the structures}
\label{age}
Good frequency coverage allows to estimate the spectral age of the core and the NW region. The amount of time elapsed
since the last acceleration of the particles in a given structure can be calculated under assumption that the observed steepening of the radio
spectrum is caused by the synchrotron and/or inverse-Compton processes. We decided to use the \textsc{synage} package \citep{murgia},
which has an implementation of the Jaffe--Perola (\citealt{JP}, JP), Kardashev--Pacholczyk (\citealt{kardashev};~\citealt{pacholczyk}, KP), and
continous injection (\citealt{pacholczyk};~\citealt{myears};~\citealt{carilli}, CI) models of electron energy losses. The JP model assumes that the particles get
isotropised in the pitch angle with the time-scale of isotropisation much smaller than the radiative lifetime. The KP model assumes that each electron
maintains its original pitch angle. The CI model includes the continuous injection of a power-law distributions of relativistic electrons. The observed
spectrum in this case is the sum of the emission from the various electron populations at different synchrotron ages, ranging from zero to the age of the
source. The flux density at frequencies below the synchrotron break rises with time, since new particles are being added. All these models assume
a constant magnetic field. \textsc{synage} uses the flux values at different wavelengths (spectral energy distribution, SED) to determine the spectral
index of the injected electron population $\alpha_{\mathrm{inj}}$, and the spectral break frequency $\nu_{\mathrm{break}}$ (frequency above which the
observed spectrum steepens from the initial one, in GHz). Using the magnetic field strength determined in Sect~\ref{magfield},
now expressed in nT, the spectral age can be calculated as:\\
\begin{equation}
\tau = 50.3 \frac{B^{1\slash 2}}{B^2 + B^{2}_{\mathrm{IC}}} \times (\nu_{\mathrm{break}}(1 + z))^{-1\slash 2} \mathrm{[Myr]}
\end{equation}
where $B_{\mathrm{IC}} = 0.338(1 + z)^{2}$ is the CMB magnetic field equivalent \citep{CMB}.
Estimated values of the injection spectral index $\alpha_{\mathrm{inj}}$, break frequency $\nu_{\rm{break}}$, and the spectral age are summarised in Table\,
\ref{fits}.
Both standard JP and KP models give a good representation of the data, but the fit derived using the JP model is slightly better (judging on the $\chi^{2}$
value -- see Fig.~\ref{agecore}). The fit given by the CI model is worse, as the goodness-of-fit parameter $\chi^{2}$ is of an order of magnitude higher than
for JP and KP. No matter the chosen model, it is clearly visible that the spectral age of the core is very low, ranging from 1.6 to 5.2\, Myrs. Bearing in mind
the possible starburst occurrence \citep{multiv} this is a reasonable value; starburst activity would result in an efficient supply of young electrons,
producing a flat spectrum. The derived spectral index of the injected electron population (significantly below 0.5 for JP and KP, and $0.46\pm 0.05$ for CI)
indeed suggests a very flat, albeit still reliable spectrum; \citet{SNR} give $\alpha=0.3$ as the flattest spectrum available for the supernova remnants (SNR),
and supernovae are the main sources of the non-thermal emitting electrons in the star-forming galaxies. Moreover, \citet{hummel} lists several galaxies for which the
injection spectra have $\alpha$ in range of 0.3--0.5, similar to our estimates, both for the core and the NW region.
The spectral fits for the NW region can be seen in Fig.~\ref{ageshock}. Similarly to the case of the core, the JP and KP fits give better representation than
the CI model. The age estimates range from 9.0 to 39.1\,Myrs. Even in case of the lowest value, the result is higher than the estimates of the age of the
star-forming knots given by \citet{beirao}, who identifies two very young (2.5 and 3.5\,Myrs) structures in the area coincident with our NW region. This means
that the high energy electrons associated with the magnetic field are older than those involved in the recent star--forming processes. This is consistent with
the steep, non-thermal spectrum of this area (Sect.~\ref{spixtext}). Most likely the magnetic field has been enhanced during the propagation of the density wave
that formed the collisional ring (see \citealt{colrings}), and the recent star--forming activity has weaker effect on its properties (e.g. flattens the spectrum,
but as for it is a non-dominant component, the overall spectrum is still relatively steep). The injection spectral index of around 0.5 is reasonable for an
extended region of synchrotron radio emission.
Finally, we note that the definite rejection of the CI model on the
basis of available data may be premature. While for the core the
mentioned fitted spectrum (Fig.~\ref{agecore}) has generally an overmuch small
curvature, for the NW region the CI model differs from the data and
other fits only at the high and low ends of the analysed spectrum
(Fig.~\ref{ageshock}). We keep in mind that the flux measurements at
low frequencies may be affected by free-free absorption caused by
ionised gas. This would decrease the flux at 234\,MHz. At the high-frequency
end of the radio spectrum, a small $\lambda$/D ratio and a short
integration time (hence a poor (u,v) plane coverage) of interferometric
observations may cause a substantial flux loss of faint extended emission.
With the currently available data both effects cannot be evaluated quantitatively.
We estimated possible effects of hypothetical free-free absorption and
of loss of extended structures at lowest and highest frequencies,
respectively upon the results of CI model fits. We performed an
experiment by increasing arbitrarily the measured flux densities of the
NW region at extreme frequencies by values needed to improve the
$\chi^2$ for the CI model. It turns out that achieving similar
goodness-of-fit as for the JP and KP models for the NW region requires
approximately 10 per cent higher flux at 234\,MHz and 30 per cent
higher flux at 8440\,MHz. We found that the break frequency, and thus
the age estimate (noted as NW-SIM in Table~\ref{fits}), appeared to be
similar to those obtained from the original CI fit. In the light of
these estimates and of the current data status we cannot simply reject
the CI model. Though it implies (regardless possible flux losses) a
substantially higher limits to the spectral age than JP and KP spectral
fits, it should be still kept in mind as an acceptable solution for the
NW region.
\begin{figure*}
{\includegraphics{core_new.eps}}
\caption{
Three different models of the electron energy losses models fitted to the SED of the core of NGC\,2445: Jaffe-Perola (JP -- red), Kardashev-Pacholczyk
(KP -- green) and continous injection (CI -- blue). Details of the fitting scheme and characteristic fit parameters are provided in Sect.~\ref{age}.
}
\label{agecore}
\end{figure*}
\begin{figure*}
{\includegraphics{shock_new.eps}}
\caption{
Three different models of the electron energy losses models fitted to the SED of the NW region of NGC\,2445: Jaffe-Perola (JP -- red), Kardashev-Pacholczyk
(KP -- green) and continous injection (CI -- blue). Details of the fitting scheme and characteristic fit parameter are provided in Sect.~\ref{age}.
}
\label{ageshock}
\end{figure*}
\begin{table*}
\caption{\label{fits}Parametres of the spectral fits}
\begin{center}
\begin{tabular}{rrrrrc}
\hline
\hline
Region & Model & $\chi^{2}$ & $\alpha_{\mathrm{inj}}$ & $\nu_{\mathrm{break}} [GHz]$ & Spectral age [Myrs]\\
\hline
Core & JP & 0.25 & $0.38^{+0.04}_{-0.01}$ & $12.1^{+2.7}_{-1.0}$ & 1.6 -- 2.1\\
Core & KP & 0.48 & $0.38^{+0.03}_{-0.01}$ & $ 6.2^{+1.3}_{-0.5}$ & 2.2 -- 2.9\\
Core & CI & 12.18 & $ 0.46\pm 0.05 $ & $ 2.8^{+1.6}_{-1.0}$ & 2.9 -- 5.2\\
NW & JP & 5.77 & $ 0.49\pm 0.07 $ & $ 9.3^{+4.8}_{-2.5}$ & 9.0 -- 14.1\\
NW & KP & 6.34 & $0.49^{+0.08}_{-0.07}$ & $ 5.0^{+3.2}_{-1.4}$ & 11.8 -- 19.5\\
NW & CI & 10.81 & $0.54^{+0.09}_{-0.10}$ & $ 2.3^{+2.7}_{-1.4}$ & 15.1 -- 39.1\\
NW-SIM& CI & 2.77 & $0.52^{+0.11}_{-0.14}$ & $ 2.8^{+5.0}_{-2.4}$ & 12.1 -- 58.6\\
\hline
\end{tabular}
\end{center}
\end{table*}
\subsection{Morphology of the radio emission distribution}
As mentioned in Sect.~\ref{intro}, the collisional ring of NGC\,2445 is far from being circular/elliptical. So is the radio emission distribution
in this galaxy: its SE side is less prominent than the NW one (not only in the radio regime). The emission is weaker -- not only at 234\,MHz (Fig.~\ref{pband}), but also in better
quality maps at 612 (Fig.~\ref{gband}) and 4860\,MHz (Fig.~\ref{cband}). Star-forming regions in the SE part have also somewhat weaker magnetic fields
and are less extended. Additionally, the angular distance from the core to the ring is lower in the NW part than in the SE part. The asymmetry may result
from an ongoing interaction with NGC\,2444, resulting in the compression of the NW part of the ring, which leads to the amplification of the magnetic field.
Such interaction has been suggested multiple times (e.g. \citealt{multiv}, \citealt{beirao}).
The distribution of the radio-emitting medium is very similar to that of the molecular emission (\citealt{beirao}),
including lack of the PAH radio emission in the southernmost knot (labeled {\it D} in \citealt{beirao}) and weak emission in the northernmost knot ({\it G} in
\citealt{beirao}). Both these knots have the lowest far-UV star-forming rates. This suggests that the radio emission deficiency might be result of faster
CR diffusion when compared to the synchrtotron losses. Weaker magnetic field -- especially in the southern part -- allows CR electrons to dffuse regardless of
their energy.
In general, the overall radio distribution seems to confirm the scenario of an off-axis collision and later, on-going interaction with NGC\,2444.
The radio ring follows the distribution of the neutral and ionised gas (\citealt{higdon}, \citealt{beirao}), which is suspected to have formed as a
result of such collision. The asymmetry of the emission together with the intergalactic ridge provides an
evidence of an interaction with the companion galaxy. However, further details of the interaction history would need examination of the regular magnetic
field properties, i.e. analysis of the polarisation data (like that was done for NGC\,3627 by \citealt{3627}, or the Virgo cluster spirals, by \citealt{virgo}),
\section{Conclusions}
\label{conclusions}
We observed the tight galaxy pair Arp\,143 with the GMRT at 234 and 612\,MHz. Radio emission maps were analysed together with the archive
VLA data at 1490, 4860, 8440, and 14940\,MHz to study the spectral age of the features seen in the system, morphology of the radio emitting medium
and the magnetic field strength and energy within the pair. The results obtained are summarised below:\\
-- The radio emission from NGC\,2445 concentrates in the northwestern part of the optical ring, extending to the outer parts of NGC\,2444.
Several distinct radio emitting regions can be identified: the core, three star-forming regions, and an intergalactic ridge.\\
-- The galactic core is a not very compact (0.8 on 0.5\,kpc), synchrotron--emitting source. It possesses a strong magnetic field of
$ 38.8 \pm 1.7 \mu$G. It has a relatively flat spectrum ($\alpha = 0.64 \pm 0.08$) and the age estimate yields $1.6-5.2$\,Myrs (depending
on the model fitted). This is consistent with its identification as a starburst region.\\
-- The northwestern region of radio emission, coincident with the star--forming knots, possesses magnetic field with strength of $12.0 \pm 0.4 \mu$G
-- somewhat stronger than the typical galactic fields. Its spectrum is typical for an ageing, yet not very old electron population, with a mean
spectral index of 0.81 $\pm$ 0.07. The spectral age estimate of $9.0-39.1$\,Myrs is higher than that of the star-forming knots given by
\citet{beirao}, indicating that most of the radio emission is associated with an older electron population, that was injected while a density
wave was propagating through the intergalactic medium.\\
-- Two other -- southern and eastern -- regions of star-formation have been detected. Their magnetic fields ($ 9.9 \pm 0.3$ and
$ 8.7 \pm 0.6 \mu$G, respectively) are somewhat weaker than that of the northwestern one.\\
-- Distribution of the radio emission from NGC\,2445 shows significant asymmetry, with the northwestern part being more luminous than the
southeastern one. This supports the scenario of an ongoing interaction with NGC\,2444.\\
-- NGC\,2444 is in general a non-radio-emitting galaxy, apart from the intergalactic radio ridge connecting it with NGC\,2445. The magnetic
field in the ridge is comparable to those found in the star-forming regions. It is also similar to that found in the intergalactic
bridge of the Taffy galaxies \citep{taffy}.\\
-- There are no signs of radio emission from the {\rm H}{\sc i} tail reported by \citet{tail}.
Depending on the -- yet unknown -- depth of the radio emitting medium, the upper limit for the magnetic field in the tidal tail is $< 4.5-12\,\mu$G, yielding
$< 1.1-6.3\times 10^{-12} \mathrm{erg}\,\mathrm{cm^{-3}}$ for its energy density.
\section*{acknowledgements}
We wish to thank the anonymous refferee, whose comments and suggestions allowed us to
significantly improve this article.
We thank the staff of the GMRT that made these observations possible. GMRT
is run by the National Centre for Radio Astrophysics of the Tata Institute
of Fundamental Research.
BNW is indebted to Eric Greisen, NRAO, for his help in preparing the RGB composite
maps using the \textsc{aips}.
This research has been supported
by the scientific grant from the National Science Centre (NCN), dec.
No.\,2011/03/B/ST9/01859.
This research has made use of
the NASA/IPAC Extragalactic Database (NED) which is operated by the Jet Propulsion
Laboratory,
California Institute of Technology, under contract with the National Aeronautics and
Space Administration.
This research has made use of NASA's Astrophysics Data System.
Funding for the SDSS and SDSS-II has been provided by the Alfred P. Sloan
Foundation, the Participating Institutions,
the National Science Foundation, the U.S. Department of Energy, the National
Aeronautics and Space Administration,
the Japanese Monbukagakusho, the Max Planck Society, and the Higher Education
Funding Council for England.
The SDSS Web Site is http://www.sdss.org/.
|
\section{Problem Definition}
Consider a reaction network with $d$ species and $K$ reactions whose \emph{stoichiometric} vectors are given by $\zeta_1,\dots,\zeta_K$. We assume that propensities of various reactions depend on a vector of $p$ parameters $\theta = (\theta_1,\dots,\theta_p)$, which may represent systems's parameters such as, reaction rate constants, hill coefficients, cell volume etc. When the state is $x$, the $k$-th reaction fires at rate $\lambda_k(x,\theta)$ and it moves the state to $(x+\zeta_k)$. In the stochastic setting, the reaction dynamics can be represented by a Markov process whose generator is
\begin{align*}
\mathbb{A}_{\theta} f(x) = \sum_{k=1}^K \lambda_k(x,\theta) \Delta_{\zeta_k} f(x),
\end{align*}
where $\Delta_{\zeta_k} f(x) = f(x+\zeta_k)-f(x)$.
Let $(X_\theta(t))_{t \geq 0}$ be a process with generator $\mathbb{A}_{\theta}$. Then its random time change representation (see \cite{EK}) is given by
\begin{align*}
X_\theta(t) = X_\theta(0) + \sum_{k=1}^K Y_k\left( \int_{0}^{t} \lambda_k( X_\theta(s),\theta )ds \right) \zeta_k,
\end{align*}
where $\{ Y_k : k =1,\dots,K\}$ is a family of independent unit rate Poisson processes. For any function $f : \mathbb{N}^d_0 \to \mathbb{R}$ that expresses an output of interest, define
\begin{align*}
\Psi_\theta(x,f,t) = \mathbb{E}\left( f( X_\theta(t) ) \vert X_\theta(0) =x \right).
\end{align*}
Then for any $i=1,\dots,p$, the \emph{first-order} sensitivity
\begin{align}
\label{def:firstsens}
S^{(i)}_{\theta} (x, f, t) = \frac{ \partial \Psi_\theta(x,f,t) }{ \partial \theta_i }
\end{align}
measures how sensitive the expected value of the output at time $t$, $\mathbb{E}( f( X_\theta(t) )$, is to small changes in parameter $\theta_i$.
Many methods exist in the literature to estimate $S^{(i)}_{\theta} (x, f, t) $ (see \cite{KSR1,KSR2,DA,Gupta,Gupta2,Gir}). In this paper we deal with the problem of estimating the \emph{second-order} sensitivity
\begin{align*}
S^{(i,j)}_{\theta} (x, f, t) = \frac{ \partial^2 \Psi_\theta(x,f,t) }{ \partial \theta_i \partial \theta_j },
\end{align*}
for any $i,j \in \{1,\dots,p\}$. Such a quantity measures the local curvature of the mapping $\theta \mapsto \mathbb{E} ( f(X_\theta(t)) )$ and this information is useful in implementing optimization schemes such as the Newton Raphson method.
In \cite{Beth}, the authors estimate $S^{(i,j)}_{\theta} (x, f, t) $ using a finite-difference approximation of the form
\begin{align}
\label{daz_fd}
S^{(i,j)}_{\theta} (x, f, t) \approx \mathbb{E}\left( \frac{ f( X_{\theta_1^\epsilon}(t) ) - f( X_{\theta_2^\epsilon}(t) ) - f( X_{\theta_3^\epsilon}(t) ) + f( X_{\theta_4^\epsilon}(t) ) }{\epsilon^2} \right),
\end{align}
for a small $\epsilon$, where each $X_{\theta_l^\epsilon}$ is a process with generator $\mathbb{A}_{\theta_l^\epsilon} $ and initial state $x$, and $\theta_l^\epsilon$-s denote perturbations of the parameter $\theta$ defined as follows:\footnote{Here $e_i$ denotes the vector $(0,\dots,0,1,0,\dots) \in \mathbb{R}^q$ where the $1$ is at the $i$-th location.}
\begin{align*}
\theta_1^\epsilon = \theta +(e_i+e_j) \epsilon, \quad \theta_2^\epsilon = \theta +e_i \epsilon, \quad \theta_3^\epsilon = \theta +e_j \epsilon \quad \textnormal{and} \quad \theta_1^\epsilon = \theta.
\end{align*}
Moreover the processes $X_{\theta_1^\epsilon},X_{\theta_2^\epsilon},X_{\theta_3^\epsilon}$ and $X_{\theta_4^\epsilon}$, are intelligently coupled to lower the variance of the associated estimator. The main drawback of finite-difference approximation is that it introduces a \emph{bias} in the estimate and generally the size or even the sign of this bias is unknown, which can cause problems in certain applications.
Interestingly, the coupling described in \cite{Beth} can be used to derive an exact formula for $S^{(i,j)}_{\theta} (x, f, t)$ by extending the ideas presented in \cite{Gupta} in the context of first-order sensitivity. The advantage of such a formula is that it allows one to construct an efficient unbiased estimator for $S^{(i,j)}_{\theta} (x, f, t)$, in the same way as the formula for the first-order sensitivity $S^{(i)}_{\theta} (x, f, t)$ in \cite{Gupta} was used in \cite{Gupta2} to devise an unbiased estimator for this quantity. In Section \ref{mainresult} we present our main result that expresses $S^{(i,j)}_{\theta} (x, f, t)$ as the expectation of a certain random variable. In Section \ref{algo} we describe how this result can be used for obtaining unbiased estimates of $S^{(i,j)}_{\theta} (x, f, t)$ in an efficient way.
\section{Main Result} \label{mainresult}
From now on, let $\lambda_0(x,\theta)$ be the function denoting the sum of propensities
\begin{align*}
\lambda_0(x,\theta) = \sum_{k=1}^K \lambda_k(x,\theta).
\end{align*}
\begin{theorem}
\label{th:mainresult}
Suppose $(X_\theta(t))_{t \geq 0}$ is the Markov process with generator $\mathbb{A}_\theta$ and initial state $x_0$. Let $\sigma_l$ be the $l$-th jump time of the process for $l=0,1,\dots$. Then for any function $f : \mathbb{N}^d_0 \to \mathbb{R}$ and $t \geq 0$, $S^{(i,j)}_{\theta} (x_0, f, t) = \mathbb{E}\left( s^{(i,j)}_\theta (x_0,f,t) \right)$ where
\begin{align}
\label{defnsij}
& s^{(i,j)}_\theta (x_0,f,t)\\ &= \sum_{k=1}^K \left[ \int_{0}^t \frac{ \partial^2 \lambda_k ( X_\theta(s) ,\theta ) }{ \partial \theta_i \partial \theta_j } \Delta_{\zeta_k} f ( X_\theta(s) ) ds \right. \notag\\
& \left. + \sum_{l=0, \sigma_l<t}^\infty \frac{ \partial \lambda_k( X_\theta(\sigma_l) ,\theta ) }{ \partial \theta_i } \left( \int_{0}^{t - \sigma_l}
(S^{(j)}_{\theta} (X_\theta(\sigma_l) + \zeta_k, f, t - \sigma_l -s) - S^{(j)}_{\theta} (X_\theta(\sigma_l), f, t - \sigma_l - s) ) e^{-\lambda_0( X_\theta(\sigma_l) ,\theta ) s } ds \right) \right. \notag \\ & \left. + \sum_{l=0, \sigma_l<t}^\infty \frac{ \partial \lambda_k( X_\theta(\sigma_l) ,\theta ) }{ \partial \theta_j } \left( \int_{0}^{t - \sigma_l}
(S^{(i)}_{\theta} (X_\theta(\sigma_l) + \zeta_k, f, t -\sigma_l -s) - S^{(i)}_{\theta} (X_\theta(\sigma_l), f, t-\sigma_l -s) ) e^{-\lambda_0( X_\theta(\sigma_l) ,\theta ) s } ds \right) \right. \notag \\ & \left. + \sum_{l=0, \sigma_l<t}^\infty\frac{ \partial^2 \lambda_k ( X_\theta(\sigma_l) ,\theta ) }{ \partial \theta_i \partial \theta_j } \left( \int_{0}^{t - \sigma_l}
( \Psi_\theta (X_\theta(\sigma_l) + \zeta_k, f, t - \sigma_l -s) - \Psi_\theta(X_\theta(\sigma_l), f, t - \sigma_l-s) \right. \right. \notag \\ & \left. \left. \qquad \qquad \qquad \qquad \qquad \qquad -\Delta_{\zeta_k} f(X_\theta(\sigma_l) ) ) e^{-\lambda_0( X_\theta(\sigma_l) ,\theta ) s } ds \right) \right], \notag
\end{align}
\end{theorem}
\begin{proof}
The proof follows by a simple extension of the ideas presented in \cite{Gupta}. We start with the finite-difference approximation of the form \eqref{daz_fd}, where the processes $X_{\theta_1^\epsilon},X_{\theta_2^\epsilon},X_{\theta_3^\epsilon}$ and $X_{\theta_4^\epsilon}$ are coupled in the same way as described in \cite{Beth}. Using Dynkin's formula and exploiting the coupling, we can pass to the limit $\epsilon \to 0$ and prove the relation given above. The details of the proof shall be provided elsewhere.
\end{proof}
Observe that Theorem \ref{th:mainresult} expresses the second-order sensitivity as the expectation of a random variable which only involves first-order sensitivities and expectations of the underlying Markov process. Since many efficient methods exist to estimate such first-order sensitivities and expectations, one can hope to use Theorem \ref{th:mainresult} to construct an efficient estimator for the second-order sensitivity. However the main difficulty is that one has to estimate a ``new" quantity (first-order sensitivity and/or expectation) at each jump time in the observation time period $[0,t]$. This can be very cumbersome as the number of jumps can be very high. However using the ideas in \cite{Gupta2} we can get around this problem and only estimate these ``new" quantities at a \emph{small} number of jump times, and still achieve an unbiased estimate for the second-order sensitivity. We describe this approach in the next section.
\section{Algorithm} \label{algo}
We use the same notation as in Theorem \ref{th:mainresult}. Let $(X_\theta(t))_{t \geq 0}$ be the Markov process with generator $\mathbb{A}_\theta$, initial state $x_0$, and let $\sigma_l$ be the $l$-th jump time of the process for $l=0,1,\dots$. The total number of jumps until time $t$ is given by the random variable
\begin{align}
\label{defn_eta}
\eta_t = \max \{ i \geq 0 : \sigma_i < t \}.
\end{align}
For simplicity we assume that there are no absorbing states (that is, $\lambda_0(x,\theta) >0$ for all $x \in \mathbb{N}^d_0$). For each $l=0,\dots,\eta_l$ let $\gamma_l$ be an independent exponentially distributed random variable with rate $\lambda_0( X_\theta( \sigma_l ),\theta )$ and define
\begin{align}
\label{defn_Gammal}
\Gamma_l = \left\{
\begin{array}{cc}
1 & \textnormal{ if } \gamma_l < (t - \sigma_l) \\
0 & \textnormal{ otherwise}
\end{array}
\right\}.
\end{align}
For each $k\in \{1,\dots,K\}$ and $q \in \{ i,j\}$ let $ \beta^{(q)}_{kl} $ be given by
\begin{align*}
\beta^{(q)}_{kl} = \textnormal{Sign}\left( \frac{ \partial \lambda_k( X_\theta( \sigma_l) ,\theta ) }{ \partial \theta_q } \right) \ \textnormal{ where } \ \textnormal{Sign}(x)= \left\{
\begin{array}{cc}
1 & \textnormal{ if } x > 0 \\
-1 & \textnormal{ if } x < 0 \\
0 & \textnormal{ if } x= 0
\end{array}
\right\}.
\end{align*}
Similarly let
\begin{align*}
\beta^{(i,j)}_{kl} = \textnormal{Sign}\left( \frac{ \partial^2 \lambda_k( X_\theta( \sigma_l) ,\theta ) }{ \partial \theta_i \theta_j } \right) .
\end{align*}
Now we choose a normalizing constant $c > 0$, which specifies the ``density" of jump times at which we estimate a new quantity of the form $ S^{(i)}_{\theta}(\cdot), S^{(j)}_{\theta}(\cdot)$ or $\Psi_\theta(\cdot)$. More details on the role of $c$ and how it can be chosen can be found in \cite{Gupta2}. Please note that the estimator we construct will remain unbiased for any choice of $c$, but its variance may vary. For each $q \in \{ i,j\}$, if $ \beta^{(q)}_{kl} \neq 0$ and $\Gamma_l = 1$, then let $\rho^{(q)}_{kl}$ be an independent $\mathbb{N}_0$-valued random variable whose distribution is Poisson with parameter
\begin{align}
\label{rateofpoisoon}
\frac{c}{ \lambda_0( X_\theta(\sigma_l) ,\theta ) } \left| \frac{ \partial \lambda_k (X_\theta(\sigma_l) ,\theta ) }{ \partial \theta_q } \right|.
\end{align}
Similarly if $ \beta^{(i,j)}_{kl} \neq 0$ and $\Gamma_l = 1$, then let $\rho^{(i,j)}_{kl}$ be an independent $\mathbb{N}_0$-valued random variable whose distribution is Poisson with parameter
\begin{align}
\label{rateofpoisoon}
\frac{c}{ \lambda_0( X_\theta(\sigma_l) ,\theta ) } \left| \frac{ \partial^2 \lambda_k (X_\theta(\sigma_l) ,\theta ) }{ \partial \theta_i \partial \theta_j } \right|.
\end{align}
Let
\begin{align*}
\Delta t_l = \left\{
\begin{array}{cc}
(\sigma_{l+1} - \sigma_l) & \textrm{ for } l = 0,\dots,\eta_t-1 \\
(T - \sigma_\eta) & \textrm{ for } l = \eta_t \\
\end{array}\right\}
\end{align*}
and define
\begin{align}
\label{expr:sthetahat}
\widehat{s}^{(i,j)}_\theta (x_0,f,t) &= \sum_{k=1}^K \sum_{l = 0}^{\eta_t } \left[
\frac{ \partial^2 \lambda_k ( X_\theta(\sigma_l) ,\theta ) }{ \partial \theta_i \partial \theta_j } \Delta_{\zeta_k}f( X_\theta(\sigma_l) )\left( \Delta t_l - \frac{ \Gamma_l }{ \lambda_0( X_\theta(\sigma_l) ,\theta ) } \right) \right. \\ & \left.
+
\frac{1}{c} \Gamma_l \left( \beta^{(i)}_{kl} \rho^{(i)}_{kl} \widehat{S}^{(i)}_{kl} + \beta^{(j)}_{kl} \rho^{(j)}_{kl} \widehat{S}^{(j)}_{kl} + \beta^{(i,j)}_{kl} \rho^{(i,j)}_{kl} \widehat{D}_{kl} \right) \right], \notag
\end{align}
where the construction of random variables $\widehat{S}^{(i)}_{kl}$, $\widehat{S}^{(j)}_{kl}$ and $\widehat{D}_{kl}$ is described below.
Let $(Z_1(t))_{t \geq 0}$ and $(Z_2(t))_{t \geq 0}$ be two processes with random time change representations given by:
\begin{align*}
Z_1(t) &=( X_\theta(\sigma_l)+ \zeta_k) + \sum_{k = 1}^K \widehat{Y}_k\left( \int_{0}^{t} \lambda_k( Z_1(s) ,\theta) \wedge \lambda_k( Z_2(s) ,\theta) ds \right)\zeta_k \\
&+ \sum_{k = 1}^K \widehat{Y}^{(1)}_k\left( \int_{0}^{t} \left(\lambda_k( Z_1(s) ,\theta) - \lambda_k( Z_1(s) ,\theta) \wedge \lambda_k( Z_2(s) ,\theta) \right) ds \right) \zeta_k \\ \textnormal{ and } \
Z_2(t) &=X_\theta(\sigma_l)+ \sum_{k = 1}^K \widehat{Y}_k\left( \int_{0}^{t} \lambda_k( Z_1(s) ,\theta) \wedge \lambda_k( Z_2(s) ,\theta) ds \right)\zeta_k \\
&+ \sum_{k = 1}^K \widehat{Y}^{(2)}_k\left( \int_{0}^{t} \left(\lambda_k( Z_2(s) ,\theta) - \lambda_k( Z_1(s) ,\theta) \wedge \lambda_k( Z_2(s) ,\theta) \right) ds \right) \zeta_k,
\end{align*}
where $\{\widehat{Y}_k, \widehat{Y}^{(1)}_k,\widehat{Y}^{(2)}_k : k =1,\dots,K\}$ is an independent family of unit rate Poisson processes. Let
\begin{align}
\label{defn_dkl}
\widehat{D}_{kl} = f( Z_1( t - \sigma_l - \gamma_l ) ) - f( Z_2( t - \sigma_l - \gamma_l ) ),
\end{align}
then we must have
\begin{align}
\label{unbpr1}
\mathbb{E}\left( \widehat{D}_{kl} \vert \mathcal{F}_t\right) = \Psi_{\theta} (X_\theta(\sigma_l) + \zeta_k, f, t - \sigma_l - \gamma_l) - \Psi_{\theta} (X_\theta(\sigma_l), f, t - \sigma_l - \gamma_l),
\end{align}
where $\{ \mathcal{F}_s\}_{s \geq 0}$ is the filtration generated by the process $(X_\theta(t))_{t \geq 0}$.
We now construct $\widehat{S}^{(q)}_{kl}$ for each $q \in \{i,j\}$. Using the \emph{Poisson Path Algorithm}(PPA) given in \cite{Gupta2}, with the underlying Markov process as $Z_1$, we can generate one realization of a random variable $\widehat{s}^{(q)}_1$ whose expectation is
\begin{align*}
\mathbb{E}\left( \widehat{s}^{(q)}_1 \right) = S^{(q)}_{\theta} (X_\theta(\sigma_l) + \zeta_k, f, t - \sigma_l - \gamma_l).
\end{align*}
Similarly using PPA with the underlying Markov process as $Z_2$, we can generate one realization of another random variable $\widehat{s}^{(q)}_2$ whose expectation is
\begin{align*}
\mathbb{E}\left( \widehat{s}^{(q)}_2 \right) = S^{(q)}_{\theta} (X_\theta(\sigma_l), f, t - \sigma_l - \gamma_l).
\end{align*}
Defining
\begin{align}
\label{defn_sqkl}
\widehat{S}^{(q)}_{kl} = \widehat{s}^{(q)}_1 - \widehat{s}^{(q)}_2,
\end{align}
we must have that
\begin{align}
\label{unbpr2}
\mathbb{E}\left( \widehat{S}^{(q)}_{kl} \vert \mathcal{F}_t\right) = S^{(q)}_{\theta} (X_\theta(\sigma_l) + \zeta_k, f, t - \sigma_l - \gamma_l) - S^{(q)}_{\theta} (X_\theta(\sigma_l), f, t - \sigma_l - \gamma_l).
\end{align}
Using relations \eqref{unbpr1} and \eqref{unbpr2} one can show using a simple conditioning argument that
\begin{align*}
\mathbb{E}\left( s^{(i,j)}_\theta (x_0,f,t ) \right) = \mathbb{E}\left( \widehat{s}^{(i,j)}_\theta (x_0,f,t) \right),
\end{align*}
where $ s^{(i,j)}_\theta (x_0,f,t ) $ and $\widehat{s}^{(i,j)}_\theta (x_0,f,t) $ are defined by \eqref{defnsij} and \eqref{expr:sthetahat} respectively. Hence Theorem \ref{th:mainresult} guarantees that by generating realizations of the random variable $\widehat{s}^{(i,j)}_\theta (x_0,f,t)$, we can obtain an unbiased estimate for second-order parameter sensitivity $S^{(i,j)}_{\theta} (x_0, f, t) $.
A single realization of the random variable $\widehat{s}^{(i,j)}_\theta (x_0,f,t)$ (given by \eqref{expr:sthetahat}) can be computed using \\$\Call{GenerateSample}{x_0,t,c}$ (Algorithm \ref{gensensvalue}). This method simulates the process $X_{\theta}$ according to SSA and at each state $x$ and jump time $s=\sigma_l$, the following happens:
\begin{itemize}
\item The exponential random variable $\gamma$ (where $\gamma = \gamma_l$ in \eqref{expr:sthetahat}) is generated and the corresponding $\Gamma_l$ (see \eqref{defn_Gammal}) is calculated.
\item If $\gamma < (t-s)$ then for each $k=1,\dots,K$ such that either $\partial \lambda_k(x,\theta)/ \partial \theta_i , \partial \lambda_k(x,\theta)/ \partial \theta_j $ or $\partial^2 \lambda_k(x,\theta)/ \partial \theta_i \partial \theta_j $ is non-zero, we generate the appropriate Poisson random variable ($\rho^{(i)}_{kl},\rho^{(j)}_{kl}$ or $\rho^{(i,j)}_{kl}$), and if this random variable is positive then the appropriate quantity ($\widehat{S}^{(i)}_{kl},\widehat{S}^{(j)}_{kl}$ or $\widehat{D}_{kl}$) is calculated and the sample value is updated according to \eqref{expr:sthetahat}.
\end{itemize}
In this algorithm, we assume that the function \emph{rand()} returns independent samples from the uniform distribution on $[0,1]$. Moreover $n \sim \Call{Poisson } { r}$ implies that $n$ is an independently generated random variable having Poisson distribution with parameter $r$. When the state of the process is $x$, the next time increment ($\Delta s$) and reaction index ($k$), as prescribed by Gillespie's SSA \cite{GP}, can be calculated using function $\Call{SSA}{x}$ (see Algorithm \ref{SSA}).
\begin{algorithm}[h]
\caption{Computes the next time increment ($\Delta s$) and reaction index $(k)$ for Gillespie's SSA}
\label{SSA}
\begin{algorithmic}[1]
\Function{SSA}{$x$}
\State Set $r_1 = \mathrm{rand}()$ , $r_2 = \mathrm{rand}()$ and $k = 0$
\State Calculate $\Delta s= -\log(r_1)/\lambda_0(x,\theta)$
\State Set $S = 0$
\While {$S < r_2$}
\State Update $k \gets k + 1$
\State Update $S \gets S + \lambda_k(x,\theta)/ \lambda_0(x,\theta) $
\EndWhile
\State \Return $(\Delta s, k)$
\EndFunction
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[h]
\caption{Generates one realization of $\widehat{s}^{(i,j)}_\theta (x_0,f,t) $ according to \eqref{expr:sthetahat} }
\label{gensensvalue}
\begin{algorithmic}[1]
\Function{GenerateSample}{$x_0,t,c$}
\State Set $x = x_0$, $s = 0$ and $S = 0$
\While {$ s < t $}
\State Calculate $(\Delta s, k_0 ) =$ SSA$(x)$
\State Update $\Delta s \gets \min\{\Delta s, t -s\}$ and set $\gamma = -\frac{ \log(rand())}{ \lambda_0(x,\theta) }$
\If{$\gamma \geq (t-s)$ }
\State Update $S \gets S + \left( \frac{ \partial \lambda^2_k(x,\theta) }{ \partial \theta_i \partial \theta_j } \right) (f(x+\zeta_k) -f(x) ) \Delta s$
\Else
\State Update $S \gets S + \left( \frac{ \partial \lambda^2_k(x,\theta) }{ \partial \theta_i \partial \theta_j } \right) (f(x+\zeta_k) -f(x) ) \left( \Delta s - \frac{ 1 }{ \lambda_0(x,\theta) } \right)$
\For {$k = 1$ to $K$}
\State Set $r^{(i)} = \left| \frac{ \partial \lambda_k(x,\theta) }{ \partial \theta_i } \right|$, $r^{(j)} = \left| \frac{ \partial \lambda_k(x,\theta) }{ \partial \theta_j } \right|$ and $r^{(i,j)} = \left| \frac{ \partial^2 \lambda_k(x,\theta) }{ \partial \theta_i \partial \theta_j } \right|$
\State Set $\beta^{(i)} =\textnormal{Sign}\left( \frac{ \partial \lambda_k(x,\theta) }{ \partial \theta_i } \right)$, $\beta^{(j)} = \textnormal{Sign}\left( \frac{ \partial \lambda_k(x,\theta) }{ \partial \theta_j } \right)$ and $\beta^{(i,j)} = \textnormal{Sign}\left( \frac{ \partial^2 \lambda_k(x,\theta) }{ \partial \theta_i \partial \theta_j } \right)$
\If {$r^{(i)} > 0$}
\State Set $n \sim \Call{Poisson } { \frac {r^{(i)} c }{ \lambda_0(x,\theta) } }$
\If {$n > 0$}
\State Calculate $\widehat{S}^{(i)}_{kl}$ according to \eqref{defn_sqkl} with $q=i, \sigma_l =s, \gamma_l = \gamma$ and $X_\theta(\sigma_l)=x$.
\State Update $S \gets S +\left( \frac{\beta^{(i)} n}{c} \right) \widehat{S}^{(i)}_{kl}$
\EndIf
\EndIf
\If {$r^{(j)} > 0$}
\State Set $n \sim \Call{Poisson } { \frac {r^{(j)} c }{ \lambda_0(x,\theta) } }$
\If {$n > 0$}
\State Calculate $\widehat{S}^{(j)}_{kl}$ according to \eqref{defn_sqkl} with $q=j,\sigma_l =s, \gamma_l = \gamma$ and $X_\theta(\sigma_l)=x$.
\State Update $S \gets S +\left( \frac{\beta^{(j)} n}{c} \right) \widehat{S}^{(j)}_{kl}$
\EndIf
\EndIf
\If {$r^{(i,j)} > 0$}
\State Set $n \sim \Call{Poisson } { \frac {r^{(i,j)} c }{ \lambda_0(x,\theta) } }$
\If {$n > 0$}
\State Calculate $\widehat{D}_{kl}$ according to \eqref{defn_dkl} with $\sigma_l =s, \gamma_l = \gamma$ and $X_\theta(\sigma_l)=x$.
\State Update $S \gets S +\left( \frac{\beta^{(i,j)} n}{c} \right) \widehat{D}_{kl}$
\EndIf
\EndIf
\EndFor
\EndIf
\State Update $s \gets s +\Delta s$ and $x \gets x+\zeta_{k_0}$
\EndWhile
\State \Return $S$
\EndFunction
\end{algorithmic}
\end{algorithm}
|
\section{Introduction}
\label{sec:introduction}
Despite being very rare, massive stars such as luminous blue variable (LBVs),
red super giants (RSGs), and Wolf-Rayet stars (WRs) play a pivotal role in
enriching the interstellar medium (ISM) through mass loss, and
they are an important source of heavier
elements contributing to the chemical enrichment of galaxies
\citep[e.g.,][]{ref:Maeder_1981}.
The deaths of these massive stars are associated with some of the highest energy
phenomena in the universe such as core-collapse supernovae
\citep[ccSNe,][]{ref:Smartt_2009}, long-duration gamma-ray
bursts \citep[e.g.,][]{ref:Stanek_2003},
neutrino bursts \citep[e.g.,][]{ref:Bionta_1987} and gravitational
wave bursts \citep[e.g.,][]{ref:Ott_2009}.
The physical mechanism, energetics and
observed properties of these events depend on the structure and
terminal mass of the evolved stars at core-collapse, which in turn are
determined by stellar mass loss \citep[see, e.g., review by][]{ref:Smith_2014}.
In addition, there is also evidence that some supernova (SN) progenitors undergo
major mass ejection events shortly before exploding
\citep[e.g.,][]{ref:GalYam_2007,ref:Smith_2008,ref:Ofek_2013a}, further altering
the properties of the explosion and implying a connection between some eruptive
mass-loss events and death. It is generally agreed that the effects of winds
are metallicity dependent~\citep[e.g.,][]{ref:Meynet_1994,ref:Heger_2003} and the SNe
requiring a very dense circumstellar medium
\citep[e.g.,][]{ref:Schlegel_1990,ref:Filippenko_1997} predominantly
occur in lower metallicity galaxies
\citep[e.g.,][]{ref:Stoll_2011}. This strongly suggests that the
nature and distribution of stars undergoing impulsive mass loss will also be
metallicity dependent and a full understanding requires exploring galaxies
beyond the Milky Way.
Understanding the evolution of massive (M$\gtrsim30 $\,M$_\odot$) stars is
challenging even when mass loss is restricted to continuous winds
\citep[e.g.,][]{ref:Fullerton_2006}. However, shorter, episodic eruptions,
rather than steady winds, may be the dominant mass loss mechanism in the
tumultuous evolutionary stages toward the end of the lives of the most massive
stars \citep[e.g.,][]{ref:Humphreys_1984,ref:Smith_2006} as they undergo
periods of photospheric instabilities leading to stellar transients
($M_V\lesssim-13$) followed by rapid ($\dot{M}\gtrsim 10^{-4}\,M_\odot$year)
mass-loss in the last stages of their evolution
\citep[see][]{ref:Kochanek_2012a,ref:Smith_2014}. Deciphering the
rate of these eruptions and their consequences is challenging because
no true analog of $\eta$\,Car in mass, luminosity, energetics, mass lost and age has
been found (see \citealp{ref:Smith_2011,ref:Kochanek_2012a}),
and the associated transients are significantly fainter than supernova explosions
and are easily missed. These phases are as difficult to model theoretically as
they are to simulate computationally.
Dense winds tend to form dust,
although for hot stars the wind must be dense enough to form a pseudo-photosphere
in the wind~\citep{ref:Davidson_1987} that shields the dust formation region from the UV emission of the star
\citep{ref:Kochanek_2011c}. The star will then be heavily obscured by dust for an extended
period after the eruption (see, e.g., \citealp{ref:Humphreys_1994}).
The Great Eruption of $\eta$\,Car between 1840 and 1860 is the most
studied case of a stellar outburst (see, e.g., \citealp{ref:Humphreys_2012}).
The $\sim10 M_{\odot}$ ejecta are now seen as a dusty nebula around
the star absorbing and then reradiating $\sim90\%$ of the light in the mid-IR.
This means that dusty ejecta are a powerful and long-lived signature of eruption.
The emission from these dusty envelopes peaks
in the mid-IR with a characteristic red color and a rising or flat
spectral energy distribution (SED) in the \textit{Spitzer} IRAC~\citep{ref:Fazio_2004} bands.
In the Galaxy, stars with resolved shells of dust emission are easily found at
24\,$\micron$~\citep{ref:Wachter_2010,ref:Gvaramadze_2010}.
The advantage of the 24\,$\micron$ band is that it can be used to identify dusty ejecta
up to $10^3 - 10^4$\,years after formation. A minority of these objects are very luminous stars
(L\,$\gtrsim10^{5.5}\,$\,L$_\odot$) with massive ($\sim0.1-10\,$\,M$_\odot$) shells (see summaries by
\citealp{ref:Humphreys_1994,ref:Humphreys_1999,ref:Smith_2006,ref:Smith_2009,ref:Vink_2009}).
These include AG\,Car~\citep{ref:Voors_2000},
the Pistol Star~\citep{ref:Figer_1999}, G79.29$+$0.46~\citep{ref:Higgs_1994},
Wray\,17$-$96~\citep{ref:Egan_2002}, and IRAS\,18576$+$0341~\citep{ref:Ueta_2001}.
These systems are significantly older ($10^3-10^4$\,years) than $\eta$\,Car, which makes it difficult
to use the ejecta to probe the rate or mechanism of mass-loss.
Still, the abundance of Galactic shells implies that the rate of
$\eta$\,Car-like eruptions is on the order of a modest fraction of the ccSN rate~\citep{ref:Kochanek_2011c}.
Their emission peaks
in the shorter IRAC bands when they are relatively young ($\sim10-100$\,years)
because the dust becomes cooler and the emission shifts to longer wavelengths
as the ejected material expands \citep{ref:Kochanek_2012a}.
It is difficult to quantify searches for such objects in our Galaxy because it is
hard to determine the distances and the survey
volume because we have to look through the crowded and dusty disk of the Galaxy.
Surveys of nearby galaxies are both better defined and can be used to build larger samples
of younger systems whose evolution can be studied to better understand the mechanism.
We previously demonstrated in \citet{ref:Khan_2010,ref:Khan_2011} that it is
possible to identify post-eruptive massive stars in galaxies beyond the Local
Group using the mid-IR excess created by warm circumstellar dust despite
the crowding problems created by the limited spatial resolution of \textit{Spitzer}
at greater distances.
In \citet{ref:Khan_2013} (``Paper\,I'' hereafter) we used archival \textit{Spitzer} IRAC
images of seven $\lesssim4$\,Mpc galaxies (closest to farthest: NGC\,6822, M\,33, NGC\,300,
NGC\,2403, M\,81, NGC\,0247, NGC\,7793) in a pilot study to search for
extragalactic analogs of $\eta$\,Car. We found 34 candidates with flat or
rising mid-IR spectral energy distributions (SEDs) and total mid-IR luminosity
$L_{mIR}\gtrsim 5\times10^5 L_\odot$.
Here, in Paper\,II, we characterize
these sources and quantify the rate of episodic
mass loss from massive stars in the last stages of evolution.
First, we construct extended optical through far-IR SEDs
using archival HST, 2MASS, and \textit{Herschel} data as well as
ground based data (Section\,\ref{sec:data}). Then, we classify the sources as either stellar or non-stellar
based on properties of the extended SEDs and model the SEDs
to infer the properties of the underlying star and the obscuring circumstellar
medium (Section\,\ref{sec:analysis}). Next, we relate these properties to the observed
ccSN rate of the targeted galaxies to quantify the rate of episodic
mass loss in the last stages of massive star evolution
(Section\,\ref{sec:discussion3}). Finally, we
consider the implications of our
findings for theories and observations of massive star evolution and their fates
(Section\,\ref{sec:conclusions}).
\section{Additional Wavelength Coverage}
\label{sec:data}
In this Section, we describe the details of how we obtained the photometric
measurements at various wavelengths to determine the properties of the
candidates from Paper\,I.
The optical through far-IR photometry are reported in
Table\,\ref{tab:photo}, and the extended SEDs are shown in Figures
\ref{fig:accept} and \ref{fig:reject}.
We utilized VizieR\footnote{http://vizier.u-strasbg.fr/}
\citep{ref:Vizier} to search for other observations of the candidates, in
particular for WISE (\citealt{ref:Wright_2010}, $12\micron$), 2MASS
(\citealt{ref:Cutri_2003}, $JHK_s$), SDSS (\citealt{ref:Adelman_2009}, $ugriz$)
and X-ray detections. For M\,33, we used the $UBVRI$ images from the
\citet{ref:Massey_2006} optical survey, and archival HST images of
NGC\,300, NGC\,2403, M\,81, NGC\,247 and NGC\,7793. Finally, we used
\textit{Herschel} PACS data to supplement the \textit{Spitzer} measurements.
For the \textit{Spitzer} IRAC 3.6, 4.5, 5.8 and $8\micron$ as well as MIPS
\citep{ref:Reike_2004} 24, 70, and $160\micron$ data, we use the
measurements reported in Paper\,I. For M\,33, our measurements were
based on IRAC data from \citet{ref:McQuinn_2007}
and MIPS data from the
\textit{Spitzer Heritage Archive}\footnote{http://sha.ipac.caltech.edu/applications/Spitzer/SHA/}.
Data from the LVL survey \citep{ref:Dale_2009} were used for NGC\,300 and
NGC\,247, and data from the SINGS survey \citep{ref:Kennicutt_2003} for
NGC\,6822, NGC\,2403, and M\,81.
We used the \textit{Herschel} PACS \citep{ref:Poglitsch_2010} 70, 100, and
$160\micron$ images available from the public
\textit{Herschel Science Archive}\footnote{http://herschel.esac.esa.int/Science\_Archive.shtml}.
Although both MIPS and PACS cover the same far-IR wavelength range
($70-160\micron$), Herschel has significantly higher resolution
(see Figure\,\ref{fig:mips_pacs}). All three PACS
band data were available for M\,33 and NGC\,7793, 70 and 160$\micron$ data
were available for NGC\,2403 and M\,81, and 100 and 160$\micron$ data were
available for NGC\,300. There are no publicly available PACS images of the
candidates in NGC\,247. We used aperture photometry (IRAF\footnote{IRAF is
distributed by the National Optical Astronomy Observatory, which is operated by
the Association of Universities for Research in Astronomy (AURA) under
cooperative agreement with the National Science Foundation.} ApPhot/Phot) with
the extraction apertures and aperture corrections from \citet{ref:Balog_2013}
and given in Table\,\ref{tab:pacs}. As with our treatment of the MIPS 70 and 160$\micron$
measurements in Paper\,I, we treat the measurements obtained in the PACS bands
as upper limits because the spatial resolution of these bands requires
increasingly large apertures at longer wavelengths. For similar
reasons, we also treat the WISE $12\micron$ fluxes, where available, as upper
limits.
For the optical photometry of the candidates in M\,33, we used the Local Group
Galaxies Survey $UBVRI$ images \citep{ref:Massey_2006}. First we verified that the coordinates
match with the IRAC images to within few$\times0\farcs1$ and then used $1\farcs0$
radius extraction apertures centered on the IRAC source locations. We
transformed the aperture fluxes to Vega-calibrated magnitudes using zero point
offsets determined from the difference between our aperture magnitudes and
calibrated magnitudes for bright stars in the \cite{ref:Massey_2006} catalog
of M\,33.
For the candidates in NGC\,300, M\,81, NGC\,2403, and NGC\,247, we
searched the ACS Nearby Galaxy Survey
\citep[ANGST,][]{ref:Dalcanton_2009} $B$, $V$ and (where available) $I$ band
point source catalogs derived
using DOLPHOT \citep{ref:Dolphin_2000}.
We verified that the IRAC and HST astrometry of the NGC\,300, NGC\,2403 and NGC\,247 images agree
within (mostly) $\lesssim0\farcs1$ to (in a few cases) $0\farcs3$.
We corrected the astrometry of the M\,81 HST images using the LBT images described later in this section
to achieve similar astrometric accuracy.
We also used the HST $I$-band
photometry of M\,81 from HST program GO-10250 (P.I. J.\,Huchra).
We retrieved all publicly available archival HST images of NGC\,7793 overlapping
the IRAC source locations along with the associated photometry tables from the Hubble Legacy
Archive\footnote{http://hla.stsci.edu/}. The HST and \textit{Spitzer}
images have a significant (few$\times 1\farcs0$) astrometric mis-match, and there are too few
reference stars in the HST images to adequately improve the astrometry. Therefore, we utilized the
IRAF GEOXYMAP and GEOXYTRAN tasks to locally match the overlapping HST and \textit{Spitzer}
images of NGC\,7793 within uncertainties of $0\farcs1 \sim 0\farcs3$.
We have variability data for the galaxies M\,81 and NGC\,2403
from a Large Binocular Telescope survey in the $U B V R$ bands
that is searching for failed supernovae
\citep{ref:Kochanek_2008}, and studying supernova progenitors
and impostors \citep{ref:Szczygiel_2012}, and
Cepheid variables \citep{ref:Gerke_2011}. We analyzed 27 epochs of data for
M\,81 and 28 epochs of data for NGC\,2403, spanning a 5\,year period.
The images were analyzed with the ISIS image subtraction
package \citep{ref:Alard_1998,ref:Alard_2000} to produce light
curves (see Figure\,\ref{fig:lc}).
\section{Characterizing the Candidates}
\label{sec:analysis}
In this section, we first discuss how we classify the candidates
based on their SEDs. Next, we describe the non-stellar and stellar
sources. Finally, we model the SEDs of the stellar sources
to determine their physical properties. Figure\,\ref{fig:accept}
shows the SEDs of the stellar sources with the best fit SED models
over plotted and
Figure\,\ref{fig:reject} shows the SEDs of the non-stellar sources.
\subsection{Source Classification}
We classify the candidates either as
stellar or non-stellar based on their photometric properties.
We focus on identifying two tell-tale signatures of the SED of
a luminous star obscured by warm circumstellar dust ---
low optical fluxes or flux-limits
compared to the mid-IR luminosities and signs of the SEDs turning
over between $8\,\micron$ and $24\,\micron$.
Towards longer
wavelengths, emission from warm circumstellar dust should peak
between the IRAC $8\,\micron$ and MIPS $24\,\micron$ bands.
It is almost impossible for mass lost from a single star to
to both have a significant optical depth and a dust temperature
cold enough to peak at wavelengths longer than $\sim24\,\micron$.
Such systems are almost certainly star clusters with significant
amounts of cold dust.
Therefore, any SED that appears
to have a steep slope between $8$ and $24\,\micron$ is considered
to be a likely cluster, rather than a single dust obscured star.
Frequently, these sources are also too luminous to be a single
star. At the shorter wavelengths, we
expect a dusty star to have relatively lower luminosity
compared to its mid-IR luminosity and redder optical
colors.
We examine the HST $B-V$/$V$ and the $V-I$/$V$ color magnitude diagrams
(CMDs) for each source for which HST data is available.
The presence of a very red optical
counterpart or the absence of a luminous star supports the
existence of significant dust obscuration. On the other hand, the presence of a blue
or bright optical counterpart makes it likely that the
source is a star cluster, a background galaxy/AGN, or a foreground star.
We first search
for bright and/or red optical sources
within the $0\farcs3$ matching radius that can be the obvious counterpart of the bright and
red IRAC source. Next, if multiple bright and/or red optical matches are found,
we identify the best astrometric match to the IRAC location. Finally, if no
reasonable match is found, we adopt the flux of the brightest of the
nearby sources as a conservative upper limit on the optical luminosity of the
candidate.
To demonstrate these, we discuss the case of M\,81-12 in detail.
M\,81-12 has a steeply rising optical and mid-IR SED
(Figure\,\ref{fig:Eta_sed}) with two distinct peaks --- one in the near-IR,
between the $R$-band and 3.6\,\micron, and another in the mid-IR between
8 and 24\,\micron.
Figure\,\ref{fig:cmd} shows the HST optical CMD for sources
near the location of M\,81-12. Besides the sources within the $0\farcs3$
matching radius, it also shows all sources within $0\farcs3 - 2\farcs0$ of the candidate
using a different symbol to emphasize the absence of any other unusual
nearby sources.
We detect a very red ($B=23.95$, $V=21.98$, $I=19.07$,
$B-V\simeq2$, $V-I\simeq2.9$) HST counterpart with an excellent astrometric
match ($<0\farcs1$, Figure\,\ref{fig:m81-12_all}) to the IRAC position.
This source is the brightest,
red HST point source within $2\farcs0$ of the IRAC location
(Figure\,\ref{fig:cmd}) and so we define it to be the counterpart used in the SED.
The LBT $V$ and $R$ band light curves show
a variable source with the correlated irregular variability ($\sim$0.4\,mag,
Figure\,\ref{fig:lc}) typical of many evolved massive stars \citep[e.g.,][]{ref:Kourniotis_2013}. Based on the
SED shape and the unambiguous detection of a red, variable optical
counterpart, we conclude that M\,81-12 is a massive, dust-obscured, single star.
In addition to Object\,X (M\,33-1), we identified
17 additional dust obscured stars and
classified 16 others as non-stellar. We left one source (N\,7793-12)
unclassified due to a lack of sufficient optical data
(it falls on an HST/ACS chip gap). It could well be a dusty star, but we
do not discuss it further.
\subsection{The 18 Stars and 16 Non-stellar Sources}
\label{sec:stars}
We identify 18 (including Object\,X/M\,33-1) sources as dusty stars. Of these, four
are in M\,33 (1, 3, 4, 7), one is in NGC\,300 (N\,300-1), four are in NGC\,2403
(2, 3, 4, 5), five are in M\,81 (5, 6, 11, 12, 14), and four are in NGC\,7793
(3, 9, 10, 13). None are in NGC\,6822 (a low-mass, low SFR galaxy) or
NGC\,247 (all three candidates turned out to be non-stellar).
These stars have low optical fluxes or flux limits
and their SEDs turn over between $8\,\micron$ and $24\,\micron$.
Moreover, M\,81-11, M\,81-12 and N\,2403-2
are detected as optically variable sources in the LBT monitoring data. N\,2403-3 is a saturated source
in the HST images, and we use the LBT flux measurements as upper limits on its
optical flux. N\,2403-3 and N\,2403-5 are not variable in the LBT data.
M\,81-5 is 0\farcs56 from a variable X-ray source
with maximum luminosity of $2\times10^{38}$\,ergs\,s$^{-1}$ \citep{ref:Liu_2011},
which is consistent with the source being an X-ray binary~\citep{ref:Remillard_2006}.
N\,7793-3 is also a X-ray source \citep{ref:Liu_2011}, with a maximum X-ray luminosity
of $3.9\times10^{37}$erg$s^{-1}$ and is classified as an HMXB by \cite{ref:Mineo_2012}.
There are 16 candidates whose SEDs indicate that they are not self-obscured stars.
Five sources in M\,33 (2, 5, 6, 8, 9) have SEDs that nearly monotonically rise
from the optical to $24\,\micron$, unambiguously indicating the presence of cold
dust associated with star clusters. As
we discussed in Paper\,I, it is unlikely for an ultra-compact
star cluster to host both evolved massive stars and significant
amounts of intra-cluster dust.
Eight sources cannot be dust obscured stars given their
very high optical luminosities: N\,2403-1 (likely a foreground star),
M\,81-7, M\,81-10, N\,247-3 (likely a foreground star, optical magnitudes
from \citealp{ref:STSci_2001}), N\,7793-4,
N\,7793-8, N\,7793-11, and N\,7793-14.
The three observed with the LBT, M\,81-7, M\,81-10 and N\,2403-1, are not variable.
We consider three more sources as most likely
non-stellar due to reasons that are unique in each case~---
\begin{itemize}
\item{N\,247-1} is located far from the plane of its edge-on host
and is unlikely to be associated with the host.
\item{N\,7793-1} is located at the edge of its host galaxy
and the PACS far-IR
flux limits are significantly lower than those of the sources
that we classified as obscured stars, indicating an absence of
the diffuse emission commonly associated with star forming regions.
\item{N\,7793-6} has an SED that can conceivably be produced by a hot star
with significant circumstellar material, although the near-IR peak seems
too narrow. However, a close inspection of the
HST image shows that this source is in a dense star-forming region with
significant diffuse light indicating the presence of intra-cluster dust.
None of the sources in the HST image are a good astrometric match to the
IRAC location.
It is more likely, in this case, that warm intercluster dust is producing
the mid-IR flux excess. The optical fluxes adopted here are those of the
most luminous HST source within a larger matching radius of $0\farcs5$.
\end{itemize}
In Paper\,I, we anticipated that further analysis would show that most, if not all, of the
candidates are in fact non-stellar sources. Based on the expected surface density
of extragalactic contaminants, of the 46 initial candidates
we estimated that all but $6\pm6$ are background galaxies/AGN
with 11 already being identified as such. Here we find that
$18$ (including Object\,X) of the candidates are dusty massive stars and
very few of the other sources are background galaxies. We do not
presently have an explanation for the fewer than expected background
sources in the targeted fields.
\subsection{SED Modeling}
We fit the SEDs of the 18 self-obscured stars using DUSTY
\citep{ref:Ivezic_1997,ref:Ivezic_1999,ref:Elitzur_2001} to model radiation
transfer through a spherical dusty medium surrounding a star and
Figure\,\ref{fig:accept} shows the best fit models.
We estimate the properties of a black-body source obscured by a surrounding dusty shell
that would produce the best fit to the observed SED (see
Figure\,\ref{fig:Dusty2} for an example). We considered models with either
graphitic or silicate \citep{ref:Draine_1984} dust. We distributed the
dust in a shell with a $\rho \propto 1/r^2$ density distribution.
The models are defined by the stellar luminosity ($L_*$),
stellar temperature ($T_*$), the total (absorption plus scattering)
$V$-band optical depth ($\tau_V$), the dust
temperature at the inner edge of the dust distribution ($T_d$), and the shell
thickness $\zeta=R_{out}/R_{in}$.
The exact value of $\zeta$ has little effect on the results, and
after a series of experiments with $1<\zeta<10$, we fixed $\zeta=4$ for
the final results.
We embedded DUSTY inside a Markov
Chain Monte Carlo (MCMC) driver to fit each SED by varying
$T_*$, $\tau_V$, and $T_d$. We limit $T_*$ to a maximum value of 30,000\,K
to exclude unrealistic temperature regimes.
The parameters of the best fit model determine the radius of the inner edge of
the dust distribution ($R_{in}$).
The mass of the shell is
\begin{equation}
\label{eqn:mejecta}
M_e = \frac{4 \pi R_{in}^2 \tau_V}{\kappa_V}
\end{equation}
where we simply scale the mass for a $V$ band dust opacity of
$\kappa_V=100\,\kappa_{100}$~cm$^2~$g$^{-1}$ and the result can be
rescaled for other choices as $M_e \propto \kappa_V^{-1}$.
Despite using a finite width shell, we focus on $R_{in}$ because it is
well-constrained while $R_{out}$ (or $\zeta$) is not. We can also
estimate an age for the shell as
\begin{equation}
\label{eqn:age}
t_e = \frac{R_{in}}{v_e}
\end{equation}
where we scale the results to $v_e=100\,v_{e100}$\,km\,s$^{-1}$.
For a comparison sample, we followed the same procedures
for the SEDs of three well-studied dust obscured stars:
$\eta$\,Car \citep{ref:Humphreys_1994};
the Galactic OH/IR star IRC+10420 \citep{ref:Jones_1993,ref:Humphreys_1997,ref:Tiffany_2010};
and M\,33's Variable\,A, which had a brief period of high mass loss leading to dust
obscuration over the last $\sim50$\,years
\citep{ref:Hubble_1953,ref:Humphreys_1987,ref:Humphreys_2006}. We use the same SEDs for these stars as in
\citet{ref:Khan_2013}.
In Table\,\ref{tab:mcmc}, we report $\chi ^ 2$, $\tau_V$, $T_d$, $T_*$, $R_{in}$, $L_*$,
$M_e$ (Equation\,\ref{eqn:mejecta}), and $t_e$ (Equation\,\ref{eqn:age})
for the best fit models for these three sources as well as the newly
identified stars. The stellar
luminosities required for both dust types are mutually consistent
because the optically thick dust shell acts as a calorimeter.
However, because the stars are heavily obscured and we have limited
optical/near-IR SEDs, the stellar temperatures generally are not well constrained.
In some cases, for different dust types,
equally good models can be obtained for either a hot ($>25000\,K$, such as a LBV in quiescence) or a
relatively cooler ($<10000\,K$, such as a LBV in outburst) star.
Indeed, for many of our 18 sources, the best fit
is near the fixed upper limit of $T_* = 30000\,K$.
To address this issue, we also tabulated the models on a grid of three fixed
stellar temperatures, $T_* = 5000\,K, 7500\,K, 20000\,K$, for each dust type.
The resulting best fit parameters are reported in Tables \ref{tab:graphitic}
and \ref{tab:silicate}.
Figure\,\ref{fig:HR_like} shows the integrated luminosities
of the newly identified self-obscured stars described in Section\,\ref{sec:stars}
as a function of $M_e$
for the best fit graphitic models of each source.
Object\,X, IRC$+10420$, M\,33\,Var\,A, and $\eta$\,Car are shown for comparison.
Figure\,\ref{fig:Dusty} shows the same quantities, but for various dust
models and temperature assumptions.
It is apparent from Figure\,\ref{fig:Dusty} and Tables
\ref{tab:mcmc}, \ref{tab:graphitic} and \ref{tab:silicate} that the integrated luminosity
and ejecta mass estimates are robust to these uncertainties.
The exceptions are N\,2403-4 and N\,7793-3.
Without any optical or near-IR data, many of the models of N\,7793-3
are unstable so we simply drop it.
The only models having a luminosity in significant excess of
$10^6\,L_\odot$ are some of the fixed temperature models of N\,2403-4.
These models have a poor goodness of fit and can be ignored.
One check on our selection methods is to examine the distribution of shell radii.
Crudely, we can see a shell until
it either becomes optically thin or too cold, so the probability
distribution of a shell's radius assuming a constant expansion velocity is
\begin{equation}
\frac{dN}{dR_{in}} = \frac{1}{R_{max}} = \hbox{constant}
\end{equation}
for $R_{in}<R_{max}$. An ensemble of shells with similar $R_{max}$ should
then show this distribution.
Figure\,\ref{fig:complete} shows the cumulative histogram (excluding N\,7793-3)
of the inner shell radii ($R_{in}$).
The curves show the expected distribution where we simply normalized to the
point where $F\left(<R_{in}\right)\simeq0.5$. The agreement shows that our sample
should be relatively complete up to $R_{max}\simeq10^{16.5}$-$10^{17}$\,cm
which corresponds to a maximum age of
\begin{equation}
\label{eqn:rad}
t_{max} \simeq 300\,{v_{e100}^{\,\,-1}}\,\,\,\hbox{years}.
\end{equation}
Figure\,\ref{fig:tv} shows the age ($t_e=R_{in}/{{v_{e100}^{-1}}}$) of the shells
as a function of $M_e$.
We also show lines corresponding to optical depths of
$\tau_V=1,10,100$. As expected, we see no sources with very low or high
optical depths, as we should have trouble finding sources with $\tau_V<1$ due to a
lack of mid-IR emission and $\tau_V\gtrsim100$ due to
the dust photosphere being too cold (peak emission in the far-IR).
Indeed, most of the dusty stars have $ 1 < \tau_V < 10 $ and
none has $ \tau_V > 100$. The large $t_e$ estimate for $\eta$\,Car when
scaled by $v_{e100}$ is due to its unusually
large ejecta velocities ($\sim600$km\,s$^{-1}$ along the
long axis \citep{ref:Cox_1995,ref:Smith_2006a} compared to
typical LBV shells ($\sim50-100$km\,s$^{-1}$, \citealp{ref:Tiffany_2010}).
\section{Implications}
\label{sec:discussion3}
The advantage of surveying external galaxies with a significant supernova
rate is that we can translate our results into estimates of abundances and
rates. We scale our rates using the observed supernova rate of
$R_{SN}=0.15$~year$^{-1}$ ($0.05 < R_{SN} < 0.35$ at 90\% confidence).
As we discussed in Paper\,I, this is significantly higher than standard star
formation rate estimates for these galaxies, but the SN rate is directly
proportional to the massive star formation rate rather than an indirect
indicator, and similar discrepancies, although not as dramatic, have been
noted in other contexts \citep[e.g.,][]{ref:Horiuchi_2011}.
In this section we first outline how we will
estimate rates, and then we discuss the constraints on analogs of $\eta$\,Car
and the implications of our sample of luminous dusty stars.
We are comparing a sample of $N_{SN}=3$ supernovae observed over $t_{SN}=20$~years
to a sample of $N_c$ candidate stars which are detectable by our selection procedures
for a time $t_d$. In Paper\,I we used DUSTY to model the detection of expanding
dusty shells and found that a good estimate for the detection time period was
\begin{equation}
t_d = t_w + 66 \left( { 100~\hbox{km~s}^{-1} \over v_e } \right)
\left( { L_* \over 10^6 L_\odot } \right)^{0.82}
\left( { M_e \over M_\odot } \right)^{0.043}~\hbox{years}
\end{equation}
for shells with masses in the range $-1 \leq \log M_e/M_\odot \leq 1$ around
stars of luminosity $5.5 \leq \log L_*/L_\odot \leq 6.5$ where $t_w$ is the
duration of the ``wind'' phase and the second term is an estimate of how long
the shell will be detected after the heavy mass loss phase ends. The principle
uncertainty lies in the choice of the velocity, $v_e$. If the rate of events
in the sample is $R_e$, then we expect to find $N_e = R_e t_d$ candidates.
The transient rate in a sample of galaxies is less interesting than comparing the rate
to the supernova rate. Let $f_e$ be the fraction of massive ($M_{ZAMS}>8\,M_\odot$) stars that
create the transients, where $f_e = (M_C/8M_\odot)^{-1.35}$ if we
assume a Salpeter IMF \citep{ref:Kennicutt_1998b}, that all stars more massive than $8M_\odot$ become
supernovae and that all stars more massive than $M_C$ cause the transients.
If each star undergoes an average of $N_e$ eruptions, then the rate of transients is related
to the rate of supernovae by $ R_e = N_e f_e R_{SN} = F_e R_{SN}$. The
interesting quantity to constrain is $F_e = N_e f_e$ rather than $R_e$. Poisson
statistics provide constraints on the rates,
where $P(D|R) \propto (R t)^N \exp(-R t)$ for $N$ events observed over a time
period $t$. This means that the probability of the rates given the data is
\begin{equation}
P(R_{SN},R_e|D) \propto P(R_{SN}) P(R_e) (R_{SN} t_{SN})^3 (R_e t_d)^{N_c}
\exp(-R_{SN}t_{SN}-R_e t_d)
\end{equation}
where $P(R_{SN})$ and $P(R_e)$ are priors on the rates which we will assume to
be uniform and we have set $N_{SN}=3$. If we now change variables to compute
$F_e$ and marginalize over the unknown supernova rate, we find that the
probability distribution for the ratio of the rates is
\begin{equation}
P(F_e |D) \propto F_e^{N_c} \left( F_e t_d + t_{SN}\right)^{-5-N_c}
\end{equation}
with the standard normalization that $\int P(F_e|D) dF_e \equiv 1$.
For our estimates of $F_e$ we present either 90\% confidence upper limits
or the value corresponding to the median probability and symmetric 90\% probability
confidence regions. Note that the probability distribution really just depends
on the product $F_e t_d$, so the results for any given estimate of $t_d$
are easily rescaled.
\subsection{No $\eta$\,Car Analog Is Found}
It is immediately obvious from Figure\,\ref{fig:HR_like} that none of the sources we identified closely
resemble $\eta$\,Car. Their typical luminosities of $10^{5.7\pm0.2}L_\odot$ correspond
to $\sim 40M_\odot$ stars
\citep{ref:Maeder_1981,ref:Maeder_1987,ref:Maeder_1988,ref:Stothers_1996,ref:Meynet_1994}
rather than the higher masses usually associated with
LBV outbursts. Since we identify a significant population of fainter stars, this is
unlikely to be a selection effect, and we conclude that these galaxies contain no analogs
of $\eta$\,Car.
There are two ways we can interpret the result. First, we
can ignore the existence of $\eta$\,Car, and set $N_c=0$. Alternatively, we can
acknowledge the existence of $\eta$\,Car, in which case $N_c=1$, since $\eta$\,Car
passes our selection criterion and mid-IR surveys of our Galaxy for objects as luminous
as $\eta$\,Car are probably complete.
For the first case, the 90\% confidence upper limit is
$F_e < 0.077 t_{d200}^{-1}$ where the period over which such
systems can be detected is scaled to $t_d = 200 t_{d200}$~years.
For the second case, where we include $\eta$\,Car, we find that
$F_e = 0.046 t_{d200}^{-1}$ with $0.0083 < F_e t_{d200} < 0.19$ at 90\% confidence.
In either case, the rate of transients comparable to $\eta$\,Car is a small
fraction of the supernova rate.
Stars as massive as $\eta$\,Car are also rare, representing only $f_e=0.02$
to $0.04$ of all massive stars for a mass range from $70$/$100M_\odot$ to
$200M_\odot$. If every sufficiently massive star had one eruption, the
results including $\eta$\,Car correspond to a minimum mass of $M_C = 65M_\odot$
($26 M_\odot < M_C < 138 M_\odot$). If every star has an
average of two eruptions, the mass limits rise to $M_C = 94M_\odot$
($42 M_\odot < M_C < 162 M_\odot$). Similarly the upper limit from
ignoring the existence of $\eta$\,Car corresponds to $M_C > 48M_\odot$
for an average of one eruption or $M_C > 72 M_\odot$ for an average
of two. \cite{ref:Kochanek_2011c} estimated that the abundance of lower optical
depth shells found at $24\mu$m around massive stars in the Galaxy was
roughly consistent with all stars more massive than $M_C = 40M_\odot$
having an average of two eruptions, corresponding to $F_e \simeq 0.2$,
which is consistent with the present results, but close to the upper limits.
\subsection{An Emerging Class of Dust Obscured Stars}
All the newly identified stars have luminosities within a narrow range of
$\log L/L_\odot \simeq 5.5$-$6.0$ (see Figure\,\ref{fig:Dusty}), which roughly corresponds
to initial stellar masses of $M_{ZAMS}\simeq 25$-$60 M_\odot$
\citep[see Section 4 of][and references therein]{ref:deJager_1998}.
Local examples of evolved stars in this luminosity range are the
Yellow Hypergiants (YHGs) such as IRC$+$10420, $\rho$\,Cas and
HR\,8752 \citep{ref:deJager_1997,ref:Smith_2006},
many of which are also partially obscured by dust ejecta. There is
no means of cleanly surveying the Galaxy for these objects and they
are so rare that samples in the Galaxy and the Magellanic Clouds
do not provide good statistics for their abundances, life times or
total mass loss. Our well-defined sample of likely extragalactic
analogs provides a means of addressing some of these questions.
If we assume these objects are similar to stars like IRC$+$10420, their
expansion velocities will be more like $50$~km/s than the $100$~km/s
of the typical LBV shell. Hence, it seems more appropriate to scale
the results to $t_d = 500 t_{d500}$~years. This also matches the
estimated age of the phase of dusty mass loss by IRC$+$10420 \citep{ref:Tiffany_2010}.
With 18 candidates, this detection period
then leads to a median estimate that $F_e = 0.20 t_{d500}^{-1}$ with
$0.086 < F_e t_{d500} < 0.55$. If we associate these with the mass
range from $25$ to $60M_\odot$, they represent a fraction of
$f_e \simeq 0.15$ of massive stars, so the average number of
episodes per star, $N_e=F_e/f_e \simeq 1.3 t_{d500}^{-1}$ with
a possible range of $0.58 < N_e t_{d500} < 3.7$, although this
does not include the uncertainties in $f_e$
Figure\,\ref{fig:HR_like} shows that the median mass causing the obscuration is
$M_e \sim 0.5 M_\odot$. The total mass lost in all the eruptions is then of order
$N_e M_e $, which would be of order $0.3$-$1.9 t_{d500}^{-1} M_\odot$.
This implies that the periods of optically thick (dusty) mass loss cannot
dominate the overall
mass loss of the star. To make the mass lost in these phases dominate
either requires that we have grossly overestimated $t_d$, or that
the mass range of the stars is much narrower.
A related point is that these phases represent a negligible fraction
of the post-main-sequence life times of the stars, at most lasting
a few thousand years.
\section{Conclusions}
\label{sec:conclusions}
In our survey, we have found no true analogs of $\eta$\,Car.
This implies that the rate of Great Eruption-like
events is of order $F_e = 0.046 t_{d200}^{-1}$ ($0.0083 < F_e t_{d200} < 0.19$)
of the ccSN rate, which is roughly consistent with each $M \gtrsim 70M_\odot$
star undergoing 1 or 2 such outbursts in its lifetime. This is scaled by
an estimated detection period of order $t_d =200 t_{d200}$\,years.
We do identify a significant population of lower luminosity dusty stars that
that are likely similar to IRC$+10420$.
Stars enter this phase at the rate $F_e = 0.20 t_{d500}^{-1}$
($0.086 < F_e t_{d500} < 0.55$) compared to the ccSN rate and for a
detection period of $t_d = 500 t_{d500}$\,years. Here the
detection period is assumed longer because the expansion velocities
are likely slower. This rate is comparable to
having all stars with $25 < M < 60M_\odot$ undergoing such a phase
once or twice.
If the estimated detection periods and mass ranges are roughly correct,
and our completeness is relatively high, there are two interesting
implications for both populations. First, these high optical depth
phases represent a negligible fraction of the post-main sequence
lifetimes of these stars, at most lasting a few thousand years.
This implies that these events have to be associated with special
periods in the evolution of the stars. The number of such events a
star experiences is also small, one or two, not ten or twenty.
Second, while a significant amount of mass is lost in the eruptions,
they cannot be a dominant contribution to mass loss. For these
high mass stars, standard models
\citep[e.g.,][]{ref:Maeder_1981,ref:Maeder_1987,ref:Maeder_1988,ref:Stothers_1996,ref:Meynet_1994}
typically strip the stars of their hydrogen envelopes and beyond,
implying total mass losses of all but the last $5$-$10M_\odot$.
The median mass loss in Figure\,\ref{fig:HR_like} is $M_e \sim 0.5 M_\odot$ and
if every star underwent two eruptions, the typical total would
be $N_e M_e \sim M_\odot$. Clearly there are some examples that
require significantly larger $M_e$, but we simply do not find
enough heavily obscured stars for this phase to represent more
than a modest fraction of the total mass loss ($\sim 10\%$ not $\sim 50\%$).
For the stars similar to IRC$+$10420, this is consistent with the
picture that the photospheres of blue-ward evolving Red Super Giants (RSGs) with
$\log(L_*/L_\odot)= 5.6\sim6.0$ and $T_{star} \simeq 7000$-$12500$~K,
become moderately unstable, leading to periods of
lower effective temperature and enhanced mass loss as the stars try
to evolve into a ``prohibited'' region of the HR diagram that
\citet{ref:deJager_1997},
\citet{ref:deJager_1998} and \citet{ref:Nieuwenhuijzen_2000}
termed the ``yellow void''.
In this phase, the stars lose enough mass to evolve into a hotter,
less massive star on the blue side of the HR diagram. This is also
the luminosity regime of the ``bistability jump'' in wind speeds
driven by opacity changes which \citet{ref:Smith_2004} hypothesizes
can explain the absence of LBVs and the existence of YHGs with high
mass loss rates and dust formation (\citealt{ref:Vink_2009}) in this luminosity range.
In fact, \citet{ref:Humphreys_2002} propose that IRC$+10420$,
which is identified by our selection criterion, is such a star.
While these arguments supply a unique, short-lived evolutionary
phase, there may be problems with the absolute scale of the mass
loss, since estimates are that IRC$+$10420 started with a mass
of $\sim 40 M_\odot$ and has lost all but $6\sim15 M_\odot$
\citep{ref:Nieuwenhuijzen_2000}.
The only other similarly unique phase in the lives of these stars is the final
post-carbon ignition phase. There are now many examples of stars
which have had outbursts shortly before exploding as supernovae
\citep[e.g.,][]{ref:Pastorello_2007,ref:Mauerhan_2012,ref:Prieto_2012,ref:Pastorello_2012,ref:Ofek_2013a}
and superluminous supernovae that are most
easily explained by surrounding the star with a large amount of
previously ejected mass
\citep{ref:Smith_2007b,ref:GalYam_2007,ref:Smith_2008,ref:Kozlowski_2010,ref:Ofek_2013a}.
Powering these supernovae requires mass ejected in
the last years to decades of the stellar life
\citep[e.g.,][]{ref:Chevalier_1994,ref:Chugai_2003,ref:Smith_2009,ref:Moriya_2014}.
and it seems natural to associate these events with the mass ejections of LBVs like
$\eta$\,Car \citep[e.g.,][]{ref:Smith_2007b,ref:GalYam_2009}. The
statistical properties and masses of either of the classes of dusty
stars we discuss are well-matched to the statistical requirements for
explaining these interaction powered supernovae if the instability is
associated with the onset of carbon burning \citep[see][]{ref:Kochanek_2011c}.
If there is only one eruption mechanism, it must be associated with a relatively
long period like carbon burning (thousands of years) rather than
the shorter, later nuclear burning phases, because we observe many
systems like $\eta$\,Car that have survived far longer than these
final phases last. If the mechanism for producing the ejecta around
the superluminous supernovae is associated with nuclear burning phases
beyond carbon, then we must have second eruption mechanism to explain
$\eta$\,Car or other still older LBVs surrounded by massive dusty
shells. If there indeed are two mass loss mechanisms --- one commencing
$\gtrsim 10^3$\,years from core-collapse and the other occurring in
the $\sim 1$\,year prior to
core-collapse --- then the self-obscured stars identified in this
work may very well be experiencing the earlier of these two mechanisms.
Otherwise, in a larger sample to $\sim100$ such stars, one should be
exploding as a ccSN every $\sim10$ years.
The dusty stars can be further characterized by their variability,
which will help to follow the evolution of the dust. For the
optically brighter examples, it may be possible to spectroscopically
determine the stellar temperatures, although detailed study may
only become possible with the James Webb Space Telescope (\textit{JWST}).
It is relatively easy to expand our survey to additional galaxies.
For very luminous sources like $\eta$\,Car analogs this is probably
feasible to distance of $10$\,Mpc, while for the lower luminosity
IRC$+$10420 analogs this is likely only feasible at the distances
of the most distant galaxies in our sample ($\sim4$\,Mpc). Larger galaxy
samples are needed not only to increase the sample of dusty luminous
stars (and hopefully find a true $\eta$ Car analog!), but also to
have a sample with a larger number of supernovae, or equivalently
a higher star formation rate. Our estimate of the abundance of
IRC$+$10420 analogs is limited by the small number of ccSN
($3$) in our sample more than by the number of dusty stars ($18$) identified.
Finally, while we have shown that surveys for the stars are feasible
using archival Spitzer data, JWST will be a far more powerful probe
of these stars. The HST-like resolution \citep{ref:Gardner_2006}
will be enormously useful
to either greatly reduce the problem of confusion or to greatly expand
the survey volume. Far more important will be the ability to carry
out the survey at $24\mu$m, which will increase the time over which
dusty shells can be identified from hundreds of years to thousands
of years, greatly improving the statistics and our ability to survey
the long term evolution of these systems and the relationship between
stellar eruptions and supernovae.
\acknowledgments
We thank Hendrik Linz for helping us analyze the Herschel PACS data and John Beacom for
numerous productive discussions.
We extend our gratitude to the SINGS Legacy Survey and LVL Survey for making
their data publicly available. This research has made use of
observations made with the NASA/ESA Hubble Space Telescope,
and obtained from the Hubble Legacy Archive, which is a collaboration
between the Space Telescope Science Institute (STScI/NASA), the Space
Telescope European Coordinating Facility (ST-ECF/ESA) and the
Canadian Astronomy Data Centre (CADC/NRC/CSA).
RK and KZS are supported in part by NSF grant AST-1108687.
|
\section{Introduction}
\subsection{Overview}
Given an open bounded connected Lipschitz set $\Omega\subset\mathbb{R}^N$ and an exponent $1<p<\infty$, we consider $\lambda_p(\Omega)$ and $\mu_p(\Omega)$ the first nontrivial eigenvalues of the $p-$Laplace operator with Dirichlet and Neumann boundary conditions, respectively. We remind that these can be variationally characterized by
\[
\lambda_p(\Omega)=\min_{v\in W^{1,p}_0(\Omega)}\left\{\int_\Omega |\nabla v|^p\, dx\,:\, \int_\Omega |v|^p\, dx=1\right\},
\]
and
\[
\mu_{p}(\Omega)=\min_{v\in W^{1,p}(\Omega)}\left\{\int_\Omega |\nabla v|^p\, dx\, :\, \int_\Omega |v|^p\, dx=1 \mbox{ and }\int_\Omega |v|^{p-2}\, v\, dx=0\right\}.
\]
Since exact values of such quantities are known only for specific values of $p$ and special domains $\Omega$, it is important to give (sharp) estimates for these quantities
in terms of (simple) geometric quantities such as measure, perimeter, diameter, relative isoperimetric constants and so on. In this direction, the reader could consult
for instance \cite{A, AM, BCT, BCT2, FNT} and the references therein.
\par
With this respect the most celebrated example is the {\it Faber--Krahn inequality} (see \cite[Chapter 3]{He} for example) which asserts that
the following minimization problem
\begin{equation}
\label{FKmin}
\inf\{\lambda_{p}(\Omega)\, :\, |\Omega|\le c\},
\end{equation}
is (uniquely) solved by
$N-$dimensional balls of measure $c$. By taking advantage of the homegeneity properties of the functional $\Omega\mapsto \lambda_p(\Omega)$, the previous can be summarized as
\begin{equation}
\label{FK}
\lambda_p(\Omega)\ge \lambda_p(B)\,\left(\frac{|B|}{|\Omega|}\right)^\frac{p}{N},
\end{equation}
where $B$ is now any $N-$dimensional ball. Then $\lambda_p(\Omega)$ can be bounded from below in a sharp way just in terms the measure of the set $\Omega$. We point out that by using {\it isoperimetric inequality} or {\it isodiametric inequality}, from \eqref{FK} we can infer similar lower bounds for $\lambda_p$ in terms of the perimeter or the diameter of $\Omega$.
\par
Observe that problem \eqref{FKmin} becomes trivial, when we replace $\lambda_p(\Omega)$ with $\mu_p(\Omega)$. Indeed, the latter is actually zero each time $\Omega$ is disconnected.
It turns out that the natural counterpart for $\mu_{p}$ is rather the maximization problem, i.e.
\begin{equation}
\label{SW}
\sup\{\mu_{p}(\Omega)\, :\, |\Omega|\ge c\}.
\end{equation}
Again, this is generally expected to be solved by $N-$dimensional balls of volume $c$. Unfortunately so far this problem has resisted all the attempts to be attacked with the unique exception of the case $p=2$ (and partially of the limiting cases $p=1$ and $p=\infty$, see \cite{EFKNT} and \cite{EKNT, RS} respectively). The {\it Szeg\H{o}--Weinberger inequality} \cite {Sz, We} states in fact that for $p=2$ problem \eqref{SW} is (uniquely) solved by $N-$dimensional balls of measure $c$.
As before, the result can be rewritten in scaling invariant form as \begin{equation}
\label{SWscaling}
\mu_{2}(\Omega)\le \mu_{2}(B) \left(\frac{|B|}{|\Omega|}\right)^{\frac{2}{N}},
\end{equation}
with equality holding if and only if $\Omega$ is an $N-$dimensional ball. We recall that the proof of \eqref{SW} for $p=2$ crucially exploits some pecularities of the Laplacian, like linearity and the knowledge of the explicit form of eigenfunctions on balls.
\vskip.2cm
A couple of comments on the Szeg\H{o}-Weinberger result are in order. First of all, inequality \eqref{SWscaling} says that $\mu_2(\Omega)$ can be estimated {\it from above} just in terms of the measure of $\Omega$. But differently from the case $\lambda_p$, now {\it we can not directly infer} similar upper bounds for $\mu_2(\Omega)$ in terms of perimeter or diameter. Then one may wonder whether such a kind of estimates hold true or not for every $1<p<\infty$, at least for some particular classes of sets.
\par
Secondly, we notice that
if $B$ is any $N-$dimensional ball, using the fact that $\mu_{2}(B) < \lambda_{2}(B)$ (see \cite{He} or Proposition \ref{prop:nodal} below), from \eqref{SWscaling} we can infer
\begin{equation}
\label{mild}
\mu_{2}(\Omega)\le \lambda_{2}(B) \left(\frac{|B|}{|\Omega|}\right)^{\frac{2}{N}},
\end{equation}
which can be see as a weak version of the Szeg\H{o}-Weinberger inequality. Again, a natural question is whether inequality \eqref{mild} can be extended to the case of $p\not=2$ or not.
\par
The last two questions are the starting point of our analysis. In this paper we prove indeed sharp upper bounds on $\mu_p(\Omega)$ in terms of diameter, as well as generalizations of \eqref{mild} for $p \neq 2$, under the additional constraint that $\Omega$ is a convex set.
\subsection{A sharp upper bound}
Then our main scope is to investigate the following shape optimization problem with convexity and diameter constraints
\begin{equation}
\label{mild2}
\mu^*:=\sup\Big\{\mu_{p}(\Omega)\, :\, \Omega\subset\mathbb{R}^N \mbox{ convex},\, \mathrm{diam}(\Omega)\ge 1\Big\}.
\end{equation}
Of course, by homogeneity of the quantities involved the value $1$ has no bearing and could be replaced by any constant $c>0$.
In Theorem \ref{p_cop} we show that the previous upper bound is finite, then we compute it and show at the same time that {\it this problem does not admit a solution}. Notably, we show that for every admissible set $\Omega$ there holds
\[
\mu_{p}(\Omega)<\mu^*,
\]
and we find a sequence $\{\Omega_n\}_{n\in\mathbb{N}}\subset\mathbb{R}^N$ of convex sets degenerating to a segment such that
\[
\lim_{n\to\infty} \mu_{p}(\Omega_n)=\mu^*.
\]
We refer to Section \ref{sec:3} for more details.
As we will show, the previous result can be summarized by the following scaling invariant sharp inequality
\begin{equation}
\label{maestro}
\mu_{p}(\Omega)<\lambda_{p}(B)\,\left(\frac{\mathrm{diam}(B)}{\mathrm{diam}(\Omega)}\right)^{p},
\end{equation}
where $B$ is any $N-$dimensional ball and inequality sign is strict. The proof of \eqref{maestro} is based on a clever choice of a special test function which reminds ideas exploited in \cite{Ch}.
By joining \eqref{maestro} and the {\it isodiametric inequality}, we immediately get (see Corollary \ref{coro:weaksw} below)
\[
\mu_{p}(\Omega)<\lambda_{p}(B)\,\left(\frac{|B|}{|\Omega|}\right)^{\frac{p}{N}},
\]
which generalizes \eqref{mild} to $p\not=2$ for convex sets, as announced above. By keeping in mind the way such an estimate was proved for $p=2$, the previous can be seen as {\it the trace of a potentially existing Szeg\H{o}-Weinberger inequality for the $p-$Laplacian}.
\par
For ease of completeness and in order to neatly motivate some of the studies performed in this paper, it is useful to recall at this point that the {\it minimization} problem
\begin{equation}
\label{mild3}
\inf\Big\{\mu_{p}(\Omega)\, :\, \Omega\subset\mathbb{R}^N \mbox{ convex},\, \mathrm{diam}(\Omega)\le 1\Big\},
\end{equation}
highlights the same features as problem \eqref{mild2}. For example, here as well the infimum can be computed and is not attained. More interestingly, a minimizing sequence is again given by a family of convex sets collapsing on a segment. For $p=2$ this is a celebrated result by Payne and Weinberger (see \cite{PW}), recently generalized in \cite{ENT,FNT} and \cite{Va} to $p\not=2$. The result can be summarized by the sharp inequality
\begin{equation}
\label{mild4}
\mu_{p}(\Omega)> \left(\frac{\pi_p}{\mathrm{diam}(\Omega)}\right)^p,
\end{equation}
where the constant $\pi_p$ is defined by
\begin{equation}
\label{pip}
\pi_p=2\,\int_0^{(p-1)^\frac{1}{p}} \left(1-\frac{t^p}{p-1}\right)^{-\frac{1}{p}}\, dt=2\,(p-1)^\frac{1}{p}\,\frac{\pi/p}{\sin(\pi/p)}.
\end{equation}
\subsection{Generalized eigenvalues} It is quite natural to wonder if similar conclusions can be drawn also in the case of the following generalized notion of first eigenvalues
\[
\lambda_{p,q}(\Omega)=\min_{v\in W^{1,p}_0(\Omega)}\left\{\int_\Omega |\nabla v|^p\, dx\, :\, \int_\Omega |v|^q\, dx=1\right\},
\]
and
\[
\mu_{p,q}(\Omega)=\min_{v\in W^{1,p}(\Omega)}\left\{\int_\Omega |\nabla v|^p\, dx\, :\, \int_\Omega |v|^q\, dx=1\,\mbox{ and }\,\int_\Omega |v|^{q-2}\, v\, dx=0\right\}.
\]
where $q\not =p$.
Quite interestingly, it turns out that for $q>p$ one has the following picture:
\begin{itemize}
\item one can prove the analogue of \eqref{maestro};
\vskip.2cm
\item this estimate {\it is not} sharp;
\vskip.2cm
\item the maximization problem
\[
\sup\Big\{\mu_{p,q}(\Omega)\, :\, \Omega\subset\mathbb{R}^N \mbox{ convex},\, \mathrm{diam}(\Omega)\ge 1\Big\}
\]
now {\it admits a solution};
\vskip.2cm
\item a lower bound like \eqref{mild4} is not possible (and the infimum in \eqref{mild3} is $0$);
\end{itemize}
On the contrary, for $q<p$ all the previous statements have to be reverted. In particular, we have
\[
\sup\Big\{\mu_{p,q}(\Omega)\, :\, \Omega\subset\mathbb{R}^N \mbox{ convex},\ \mathrm{diam}(\Omega)\ge 1\Big\}=+\infty,
\]
and it is rather the minimization problem for $\mu_{p,q}$ which is now well-posed (see Section \ref{sec:4} for more details).
\subsection{Plan of the paper}
In Section \ref{sec:2} we prove some basic results concerning properties of $\mu_{p,q}(\Omega)$ and $\lambda_{p,q}(\Omega)$. Section \ref{sec:3} is devoted to the investigation of problem
\eqref{mild2}. In Section \ref{sec:4} we consider the case $p\not = q$.
As a consequence of some the estimates proved in the paper, in Section \ref{sec:5} we exhibit a nodal domain property for Neumann eigenfunctions. Roughly speaking, this shows that for $q\ge p$ eigenfunctions associated to $\mu_{p,q}(\Omega)$ can not have a closed nodal line. Finally, Appendix \ref{sec:1d} deals with a one-dimensional variational problem and its extremals, whose properties are crucially exploited to prove optimality of the estimate \eqref{maestro}.
\begin{ack}
The first author has been partially supported by the Gaspard Monge Program for Optimization (PGMO), by EDF and the Jacques Hadamard Mathematical Foundation, through the research contract MACRO.
Part of this work has been done during some visits of the first author to Napoli. The Departement of Mathematics of the University of Napoli and its facilities are kindly acknowledged.
\end{ack}
\section{Preliminaries}
\label{sec:2}
We fix two exponents $p$ and $q$ such that $1<p<\infty$ and
\[
1<q<p^*:=\left\{\begin{array}{lr}
\displaystyle \frac{N\, p}{N-p},&\qquad \mbox{ if } 1<p<N,\\
&\\
+\infty,& \mbox{ if }p\ge N.\\
\end{array}
\right.
\]
For every $\Omega\subset\mathbb{R}^N$ open bounded Lipschitz set, we use the standard Sobolev spaces
\[
W^{1,p}(\Omega)=\big\{u\in L^p(\Omega)\, :\, \nabla u\in L^p(\Omega;\mathbb{R}^N)\big\},
\]
and $W^{1,p}_0(\Omega)$, the latter being the completion of $C^\infty_0(\Omega)$ with respect to the norm of $W^{1,p}(\Omega)$. We then define the two quantities
\[
\mu_{p,q}(\Omega)=\inf_{v\in W^{1,p}(\Omega)\setminus\{0\}}\left\{\frac{\displaystyle\int_\Omega |\nabla v|^p\, dx}{\displaystyle \left(\int_\Omega |v|^q\, dx\right)^\frac{p}{q}}\, :\, \int_\Omega |v|^{q-2}\, v\, dx=0\right\}
\]
and
\[
\lambda_{p,q}(\Omega)=\inf_{v\in W^{1,p}_0(\Omega)\setminus\{0\}} \frac{\displaystyle\int_\Omega |\nabla v|^p\, dx}{\displaystyle\left(\int_\Omega |v|^p\, dx\right)^\frac{p}{q}}.
\]
It is useful to recall that $\mu_{p,q}(\Omega)$ can be defined through the unconstrained minimization
\begin{equation}
\label{rallai}
\mu_{p,q}(\Omega)=\inf_{v\in \widehat W^{1,p}(\Omega)}\frac{\displaystyle\int_\Omega |\nabla v|^p\, dx}{\displaystyle \min_{t\in\mathbb{R}} \left(\int_\Omega |v-t|^q\, dx\right)^\frac{p}{q}},
\end{equation}
where we set $\widehat W^{1,p}(\Omega)=\{v\in W^{1,p}(\Omega)\, :\, \int_\Omega |\nabla v|^p\, dx>0\}$.
Also, we have that if $\lambda$ is such that the equation
\[
-\Delta_p u=\lambda\, \|u\|^{p-q}_{L^q(\Omega)}\, |u|^{q-2}\, u\quad \mbox{ in }\Omega,\qquad u=0,\quad \mbox{ on }\partial\Omega,
\]
admits a nontrivial solution in $W^{1,p}_0(\Omega)$, then $\lambda\ge \lambda_{p,q}(\Omega)$.
We start with a preliminary result on the quantities $\mu_{p,q}$ and $\lambda_{p,q}$.
\begin{lm}
\label{lm:comparison}
Let $1<p<\infty$ and $1<s<q<p^*$, then we have
\[
\lambda_{p,q}(\Omega)\le |\Omega|^{\frac{p}{s}-\frac{p}{q}}\, \lambda_{p,s}(\Omega)\qquad \mbox{ and }\qquad \mu_{p,q}(\Omega)\le |\Omega|^{\frac{p}{s}-\frac{p}{q}}\, \mu_{p,s}(\Omega).
\]
\end{lm}
\begin{proof}
The result is a plain consequence of H\"older inequality. Let us prove for example the second inequality: we pick $u_s\in \widehat W^{1,p}(\Omega)$ a function minimizing the Rayleigh quotient \eqref{rallai} which defines $\mu_{p,s}(\Omega)$.
We then define $t_q$ as the minimizer of
\[
t\mapsto \int_\Omega |u_s-t|^q\, dx,
\]
thus we get
\[
\begin{split}
\mu_{p,q}(\Omega)&=\min_{v\in \widehat W^{1,p}(\Omega)} \frac{\displaystyle\int_\Omega |\nabla v|^p\, dx}{\min\limits_{t\in\mathbb{R}}\displaystyle \left(\int_\Omega |v-t|^q\, dx\right)^\frac{p}{q}}\le \frac{\displaystyle\int_\Omega |\nabla u_s|^p\, dx}{\displaystyle \left(\int_\Omega |u_s-t_q|^q\, dx\right)^\frac{p}{q}}\\
&\le |\Omega|^{\frac{p}{s}-\frac{p}{q}}\, \frac{\displaystyle\int_\Omega |\nabla u_s|^p\, dx}{\displaystyle \left(\int_\Omega |u_s-t_q|^s\, dx\right)^\frac{p}{s}}\le |\Omega|^{\frac{p}{s}-\frac{p}{q}}\, \frac{\displaystyle\int_\Omega |\nabla u_s|^p\, dx}{\min\limits_{t\in\mathbb{R}}\displaystyle \left(\int_\Omega |u_s-t|^s\, dx\right)^\frac{p}{s}}
\end{split}
\]
which in turn gives the desired inequality, by minimality of $u_s$.
\end{proof}
\begin{comment}
The following simple continuity result will be useful.
\begin{lm}
\label{lm:continuo}
Let $1<p<\infty$ and $1<q<p^*$, then we have
\[
\lim_{s\to q} \lambda_{p,s}(\Omega)=\lambda_{p,q}(\Omega)\qquad\mbox{ and }\qquad \lim_{s\to q} \mu_{p,s}(\Omega)=\mu_{p,q}(\Omega).
\]
\end{lm}
\begin{proof}
We just prove the second equality, the first one can be proved along the same lines. Let $u_q\in W^{1,p}(\Omega)\setminus\{0\}$ be a minimizer for the variational problem defining $\mu_{p,q}(\Omega)$, i.e. such that
\[
\frac{\displaystyle\int_\Omega |\nabla u_q|^p\, dx}{\displaystyle \left(\int_\Omega |u_q|^{q}\, dx\right)^\frac{p}{q}}=\mu_{p,q}(\Omega),\qquad \mbox{ and }\qquad \int_\Omega |u_q|^{q-2}\, u_q\ dx=0.
\]
Observe that since $u_q\not =0$, the second condition above implies that $u_q$ is not constant.
We then define $t_{s}$ to be the minimizer of
\[
t\mapsto \int_\Omega |u_q-t|^s\, dx,\qquad t\in\mathbb{R},
\]
thus we obtain
\[
\begin{split}
\limsup_{s\to q}\mu_{p,s}(\Omega)&\le \limsup_{s\to q}\frac{\displaystyle\int_\Omega |\nabla u_q|^p\, dx}{\min\limits_{t\in\mathbb{R}}\displaystyle \left(\int_\Omega |u_q-t|^{s}\, dx\right)^\frac{p}{s}}\\
&=\limsup_{s\to q}\frac{\displaystyle\int_\Omega |\nabla u_q|^p\, dx}{\displaystyle \left(\int_\Omega |u_q-t_s|^{s}\, dx\right)^\frac{p}{s}}=\frac{\displaystyle\int_\Omega |\nabla u_q|^p\, dx}{\displaystyle \left(\int_\Omega |u_q|^{q}\, dx\right)^\frac{p}{q}}=\mu_{p,q}(\Omega).
\end{split}
\]
where we used that\footnote{The convex functions
\[
t\mapsto \int_\Omega |u-t|^s\, dx,
\]
are converging locally uniformly to the function
\[
t\mapsto \int_\Omega |u-t|^q\, dx.
\]
Moreover, for $|s-q|$ sufficiently small, there exists $m>0$ such that we have $|t_s|<m$. Then by uniform convergence and uniqueness of the minimizer $t_q$ we can infer $\lim_{s\to q} t_s=t_q=0$.} $t_s$ goes to $0$ as $s$ goes to $q$.
In order to prove that
\[
\liminf_{s\to q} \mu_{p,s}(\Omega)\ge \mu_{p,q}(\Omega),
\]
as well, we have to distinguish two cases. If $s\nearrow q$, then we can simply apply Lemma \ref{lm:comparison}, thus we obtain
\[
\liminf_{s\nearrow q} \mu_{p,s}(\Omega)=\liminf_{s\nearrow q} |\Omega|^{\frac{p}{s}-\frac{p}{q}}\, \mu_{p,s}(\Omega)\ge \mu_{p,q}(\Omega).
\]
On the contrary, for $s\searrow q$, we pick $u_s\in W^{1,p}(\Omega)$ such that
\[
\int_\Omega |\nabla u_s|^p\, dx=\mu_{p,s}(\Omega),\qquad \int_\Omega |u_s|^s\, dx=1,\qquad \int_\Omega |u_s|^{s-2}\, u_s=0,
\]
then in particular $\{u_s\}$ is a bounded sequence in $W^{1,p}(\Omega)$, thus there exists a subsequence (not relabeled) which strongly converges when $s\searrow q$ in $L^q(\Omega)$. If we call $u$ this limit function, we have
\[
\int_\Omega |u|^p=\lim_{s\searrow q} \int_\Omega |u_s|^s\, dx=1\qquad \mbox{ and }\qquad \int_\Omega |u|^{q-2}\, u\, dx=\lim_{s\searrow q} \int_\Omega |u_s|^{s-2}\, u_s\, dx=0
\]
and of course
\[
\mu_{p,q}(\Omega)\le\int_\Omega |\nabla u|^p\, dx\le \liminf_{s\searrow q} \int_\Omega |\nabla u_s|^p\, dx=\liminf_{s\searrow q} \mu_{p,s}(\Omega).
\]
This concludes the proof.
\end{proof}
\end{comment}
We will also need the following very simple geometric result for convex sets.
\begin{lm}
\label{lm:normali}
Let $\Omega\subset\mathbb{R}^N$ be an open convex set, and let $x_0\in\partial\Omega$. Then
\[
\langle x-x_0,\nu_\Omega(x)\rangle\ge 0,\qquad\mbox{ for $\mathcal{H}^{N-1}-$a.e. }x\in\partial\Omega,
\]
where $\nu_\Omega(x)$ denotes the outer unit normal to $\partial \Omega$ at the point $x$.
\end{lm}
\begin{proof}
Since $\Omega$ is convex, given $x\in\partial\Omega$ we have that
\[
\overline\Omega\subset \{y\in\mathbb{R}^N\, :\, \langle y-x,\nu_\Omega(x)\rangle\le 0\},
\]
i.e. the hyperplane orthogonal to $\nu_\Omega(x)$ and passing from $x$ is a supporting hyperplane for $\Omega$. In particular, since $x_0\in\overline\Omega$ we get $\langle x_0-x,\nu_\Omega(x)\rangle\le 0$, which concludes the proof.
\end{proof}
\section{A Szeg\H{o}-Weinberger inequality for convex sets}
\label{sec:3}
The following is the main result of the paper. This shows that the nonlinear spectral optimization problem
\[
\sup\Big\{\mu_{p,p}(\Omega)\, :\, \Omega\subset\mathbb{R}^N \mbox{ convex},\, \mathrm{diam}(\Omega)\ge 1\Big\},
\]
does not admit a solution, but a maximizing sequence is given by a family of convex sets suitably degenerating to a segment. Of course, the value $1$ for the diameter constraint plays no special role and could be replaced by any constant $c>0$.
\begin{teo}
\label{p_cop}
Let $\Omega\subset\mathbb{R}^N$ be an open bounded convex set and $1<p<\infty$. Then we have
\begin{equation}
\label{coppolicchio}
\mu_{p,p}(\Omega)<\lambda_{p,p}(B)\,\left(\frac{\mathrm{diam}(B)}{\mathrm{diam}(\Omega)}\right)^{p},
\end{equation}
where $B$ is any $N-$dimensional ball.
\par
Equality sign in \eqref{coppolicchio} is never achieved but the inequality is sharp. More precisely, there exists a sequence $\{\Omega_k\}_{k\in\mathbb{N}}\subset\mathbb{R}^N$ of convex sets such that:
\begin{itemize}
\item $\mathrm{diam}(\Omega_k)=d>0$, for every $k\in\mathbb{N}$;
\item $\Omega_k$ converges to a segment of length $d$ in the Hausdorff topology;
\item we have
\begin{equation}
\label{figurine}
\lim_{k\to\infty}\mu_{p,p}(\Omega_k)=\lambda_{p,p}(B_{d/2}),
\end{equation}
where $B_{d/2}$ is an $N-$dimensional ball having radius $d/2$.
\end{itemize}
\end{teo}
\begin{proof}
We split the proof into two parts: at first we prove \eqref{coppolicchio}, then we construct the sequence $\{\Omega_k\}_{k\in\mathbb{N}}$ verifying \eqref{figurine}.
\vskip.2cm\noindent
\emph{Proof of \eqref{coppolicchio}.} First of all, we observe that inequality \eqref{coppolicchio} is in scaling invariant form. Then without loss of generality, we can confine ourselves to prove that
\[
\mu_{p,p}(\Omega)<\lambda_{p,p}(B),
\]
where $B$ is the ball centered at the origin such that $\mathrm{diam}(\Omega)=\mathrm{diam}(B)$. Let us take $u\in C^{1,\alpha}(\overline B)\cap C^\infty(B\setminus\{0\})$ the first Dirichlet eigenfunction for the ball $B$, normalized by the conditions
\[
\|u\|_{L^p(B)}=1\qquad \mbox{ and }\qquad u>0.
\]
This is radially symmetric and solves
\begin{equation}
\label{palla}
-\Delta_p u=\lambda_{p,p}(B)\, u^{p-1}\qquad \mbox{ and }\qquad u=0\quad \mbox{ on } \partial B.
\end{equation}
We then take two points $x_0,x_1\in\partial\Omega$ such that $|x_0-x_1|=\mathrm{diam}(\Omega)$,
and we define the two caps
\[
\Omega_i=\left\{x\, :\, |x-x_i|<\frac{\mathrm{diam}(\Omega)}{2}\right\}\cap \Omega,\qquad i=0,1,
\]
which are mutually disjoint (see Figure \ref{fig:fig3}).
\begin{figure}
\includegraphics[scale=.25]{fig3}
\caption{The construction of the two caps $\Omega_0$ and $\Omega_1$.}
\label{fig:fig3}
\end{figure}
We then take the function
\[
\varphi(x)=u(x-x_0)\cdot 1_{\Omega_0}(x)-c\, u(x-x_1)\cdot 1_{\Omega_1}(x)\in W^{1,p}(\Omega),
\]
where the costant $c\in\mathbb{R}$ is given by
\[
c=\frac{\displaystyle\int_{\Omega_0} u(x-x_0)^{p-1}\, dx}{\displaystyle\int_{\Omega_1} u(x-x_1)^{p-1}\, dx},\qquad \mbox{ so that }\qquad \int_{\Omega} |\varphi|^{p-2}\, \varphi\, dx=0.
\]
By using this function $\varphi$ in the Rayleigh quotient defining $\mu_{p,p}(\Omega)$, we get
\[
\mu_{p,p}(\Omega)<
\frac{\displaystyle\int_{\Omega_0} |\nabla u(x-x_0)|^p\, dx+c^p\,\int_{\Omega_1} |\nabla u(x-x_1)|^p\, dx}{\displaystyle \int_{\Omega_0} |u(x-x_0)|^p\, dx+c^p \,\int_{\Omega_1} |u(x-x_1)|^q\, dx},
\]
where the strict inequality holds since $\varphi$ {\it can not be an eigenfunction}\footnote{Observe that if the Rayleigh quotient of $\varphi$ achieves the minimal value $\mu_{p,q}(\Omega)$, then $\varphi$ would solve
\[
-\Delta_p \varphi=\mu_{p,q}(\Omega)\, |\varphi|^{q-2}\, \varphi,\qquad \mbox{ in }\Omega,
\]
in a weak sense. Let us take $y_0\in \partial\Omega_0\cap \Omega$, by picking a ball $B_\varrho(y_0)$ with radius $\varrho$ sufficiently small so that $B_\varrho(y_0)\subset\Omega\setminus\Omega_1$, we would obtain that $\varphi$ is a nonnegative solution of the equation above in $B_\varrho(y_0)$. Then by Harnack's inequality (see \cite[Theorem 1.1]{Tr}) one obtains
\[
0<\max_{B_\varrho(y_0)} \varphi\le C\, \min_{B_\varrho(y_0)} \varphi=0,
\]
thus getting a contradiction. We point out that {\it we are not using any unique continuation argument}.}.
By performing an integration by parts in the integrals at the numerator, we obtain\footnote{Observe that we have $u(x-x_0)=0$ on $\partial\Omega_0\cap\Omega$.}
\[
\begin{split}
\int_{\Omega_0} |\nabla u(x-x_0)|^p\, dx&=\int_{\partial\Omega\cap \partial\Omega_0} |\nabla u(x-x_0)|^{p-2}\, \frac{\partial u}{\partial\nu_\Omega}(x-x_0)\, u(x-x_0)\, d\mathcal{H}^{N-1}(x)\\
&-\int_{\Omega_0} \Delta_p u(x-x_0)\, u(x-x_0)\, dx\\
&=\int_{\partial\Omega\cap \partial\Omega_0} |\nabla u(x-x_0)|^{p-2}\, \frac{\partial u}{\partial\nu_\Omega}(x-x_0)\, u(x-x_0)\, d\mathcal{H}^{N-1}(x)\\
&+\lambda_{p,p}(B)\,\int_{\Omega_0} |u(x-x_0)|^p\, dx,
\end{split}
\]
where we used the equation \eqref{palla} solved by $u$. Observe that the first integral in the right-hand side is non-positive. Indeed $u$ is a radially decreasing function, then (with a small abuse of notation) we have
\[
\langle\nabla u(x-x_0),\nu_\Omega)(x)\rangle=u'(|x-x_0|)\, \left\langle \frac{x-x_0}{|x-x_0|},\nu_\Omega(x)\right\rangle,
\]
and the claim follows from Lemma \ref{lm:normali}, since $u'\le 0$. The same computations apply to the other terms appearing in the numerator, thus obtaining
\[
\mu_{p,p}(\Omega)< \lambda_{p,p}(B)\,\frac{\displaystyle \int_{\Omega_0} |u(x-x_0)|^q\, dx+c^p \,\int_{\Omega_1} |u(x-x_1)|^p\, dx}{\displaystyle\int_{\Omega_0} |u(x-x_0)|^p\, dx+c^p\,\int_{\Omega_1} |u(x-x_1)|^q\, dx}=\lambda_{p,p}(B),
\]
which concludes the proof of \eqref{coppolicchio}.
\vskip.2cm\noindent
\emph{Optimality of \eqref{coppolicchio}}. Let $B$ be a ball of diameter $d$. We now prove optimality of \eqref{coppolicchio}: for this we need to construct a sequence of convex sets $\{\Omega_k\}_{k\in\mathbb{N}}$, all sharing the same diameter $d$, and such that
\begin{equation}
\label{optimality}
\lambda_{p,p}(B)\le \liminf_{k\to\infty} \mu_{p,p}(\Omega_k).
\end{equation}
For all $s\in\mathbb{R}$ and $k\in \mathbb{N}\setminus\{0\}$ let us denote by
\[
\mathcal{C}^-_k(s)=\left\{ (x_1,x')\in\mathbb{R}\times\mathbb{R}^{N-1}\, :\, (x_1-s)_->k\,|x'| \right\}
\]
and
\[
\mathcal{C}^+_k(s)=\left\{ (x_1,x')\in\mathbb{R}\times\mathbb{R}^{N-1}\, :\, (x_1-s)_+>k\,|x'| \right\}
\]
the left and right circular infinite cone in $\mathbb{R}^N$ whose axis is the $x_1-$axis, having vertex in $(s,0)\in\mathbb{R}\times\mathbb{R}^{N-1}$, and whose opening angle is $\alpha=2\,\arctan(1/k)$. We set
\[
\Omega_k=\mathcal{C}^-_k\left(\frac{d}{2}\right)\cap\mathcal{C}^+_k\left(-\frac{d}{2}\right).
\]
In dimension $N=2$, $\Omega_k$ is nothing but a rhombus of diagonals $d$ and $d/k$. In higher dimension $\Omega_k$ is obtained by gluing together the basis of two right circular cones of height $d/2$ and radii $d/(2\,k)$ (see Figure \ref{fig:fig2}).
\begin{figure}
\includegraphics[scale=.3]{fig2}
\caption{The maximizing sequence $\Omega_k$ of Theorem \ref{p_cop}.}
\label{fig:fig2}
\end{figure}
\par
We claim that for this family inequality \eqref{optimality} holds true.
We start observing that whenever $u\in W^{1,p}(\Omega_k)$ then the rescaled function $v(x_1,x')=u\left(x_1,x'/k\right)$ belongs to $W^{1,p}(\Omega_1)$ and we have
\[
\int_{\Omega_1} \left(|\partial_{x_1}v|^2+k^2|\nabla_{x'}v|^2\right)^\frac p2\, dx=k^{N-1}\,\int_{\Omega_k} |\nabla u|^p\, dx,\qquad \int_{\Omega_1} |v|^p\, dx =k^{N-1}\, \int_{\Omega_k} |u|^p\, dx,
\]
and
\[
\int_{\Omega_1} |v|^{p-2}\, v\, dx=k^{N-1}\, \int_{\Omega_k} |u|^{p-2}\, u\, dx=0.
\]
Thus we obtain
\[
\begin{split}
\mu_{p,p}&(\Omega_k)=\min_{u\in W^{1,p}(\Omega_k)\setminus\{0\}}\left\{\frac{\displaystyle\int_{\Omega_k} |\nabla u|^p\, dx}{\displaystyle \int_{\Omega_k} |u|^p\, dx}\, :\, \int_{\Omega_k} |u|^{p-2}\, u\, dx=0\right\},\\
&=\min_{v\in W^{1,p}(\Omega_1)\setminus\{0\}}\left\{\frac{\displaystyle\int_{\Omega_1} \left( |\partial_{x_1}v|^2+k^2\,|\nabla_{x'}v|^2\right)^\frac p2\, dx}{\displaystyle \int_{\Omega_1} |v|^p\, dx}\, :\, \int_{\Omega_1} |v|^{p-2}\, v\, dx=0\right\}=:\gamma_k(\Omega_1).
\end{split}
\]
Now we denote by $u_k$ a function which minimizes the Rayleigh quotient defining $\mu_{p,p}(\Omega_k)$ and by $v_k(x_1,x')=u_k\left(x_1,x'/k\right)$ the corresponding function which minimizes the functional defining $\gamma_k(\Omega_1)$. Without loss of generality we can assume that $\|v_k\|_{L^p(\Omega_1)}=1$. Inequality \eqref{coppolicchio} implies that
\begin{equation}
\label{kappa}
\int_{\Omega_1} \left( |\partial_{x_1}v_k|^2+k^2\,|\nabla_{x'}v_k|^2\right)^\frac{p}{2}\, dx\le C_{N,p,d},\qquad \mbox{ for all } k\in\mathbb{N}\setminus\{0\},
\end{equation}
then there exists $w\in W^{1,p}(\Omega_1)\setminus\{0\}$ so that $v_k\to w$ weakly in $W^{1,p}(\Omega_1)$ and strongly in $L^p(\Omega_1)$. Moreover we also have\footnote{The bound \eqref{kappa} implies that for every given $k_0\in\mathbb{N}\setminus\{0\}$, we have
\[
k_0^p\,\int_{\Omega_1} |\nabla_{x'} w|^p\, dx\le\int_{\Omega_1} \left( |\nabla_{x_1} w|^2+k_0^2\,|\nabla_{x'}w|^2\right)^\frac p2\, dx\le\liminf_{k\to\infty}\int_\Omega \left( |\nabla_{x_1}v_k|^2+k_0^2\,|\nabla_{x'}v_k|^2\right)^\frac{p}{2}\, dx\le C,
\]
which in turn gives $\nabla_{x'} w\equiv 0$ by the arbitrariness of $k_0$.}
\[
\begin{split}
\nabla_{x'} w\equiv 0,\qquad \mbox{ and }\qquad\int_{\Omega_1} |w|^{p-2}\, w\, dx=0.
\end{split}
\]
Thus $w$ does not depend on the $x'$ variable and we will write for simplicity $w=w(x_1)$ with a slight abuse of notation. For all $s\in[-d/2,d/2]$ we denote by $\Gamma_s$ the section of $\Omega_1$ which is orthogonal to the $x_1-$axis at $x_1=s$ and set $g(s)=\mathcal{H}^{N-1}(\Gamma_s)$. Then we get
\[
\begin{split}
\liminf_{k\to\infty} \gamma_k(\Omega_1)&=\liminf_{k\to\infty} \frac{\displaystyle\int_{\Omega_1} \left( |\partial_{x_1}v_k|^2+k^2|\nabla_{x'}v_k|^2\right)^\frac{p}{2}\, dx}{\displaystyle \int_{\Omega_1} |v_k|^p\, dx}\ge\liminf_{k\to\infty} \frac{\displaystyle\int_{\Omega_1} |\partial_{x_1}v_k|^p\, dx}{\displaystyle \int_{\Omega_1} |v_k|^p\, dx}\\
&\ge \frac{\displaystyle\int_{\Omega_1} |w'|^p\, dx}{\displaystyle \int_{\Omega_1} |w|^p\, dx}=\frac{\displaystyle\int_{-d/2}^{d/2} |w'|^p \,g\, ds}{\displaystyle \int_{-d/2}^{d/2} |w|^p \,g\, ds}\\
&\ge\min_{\phi\in W^{1,p}\left(\left(-d/2,d/2\right)\right)\setminus\{0\}}
\left\{\frac{\displaystyle\int_{-d/2}^{d/2} |\phi'|^p\,g\,\,ds}{\displaystyle \int_{-d/2}^{d/2} |\phi|^p\,g\,ds}\, :\, \int_{-d/2}^{d/2} |\phi|^{p-2}\,\phi\,g\, \,ds=0\right\}
\end{split}
\]
Let us denote by $\eta$ the previous minimal value, then by Lemma \ref{lm:apres} a minimizer $f$ does exist and is a solution to the following boundary value problem
\[
\left\{\begin{array}{lc}
-\big(g\,|f'|^{p-2}\,f'\big)'=\eta \,g\,|f|^{p-2}\,f,& \mbox{ in } (-d/2,d/2),\\ &\\
f'(-d/2)=f'(d/2)=0.&
\end{array}
\right.
\]
Still by Lemma \ref{lm:apres} we have that $f(0)=0$ and hence $f$ solves
\[
\left\{\begin{array}{lc}
-\big(g\,|f'|^{p-2}\,f'\big)'=\eta \,g\,|f|^{p-2}\,f,& \mbox{ in } (0,d/2),\\ &\\
f(0)=f'(d/2)=0.&
\end{array}
\right.
\]
Finally, by reminding that $g(s)=\omega_{N-1}(d/2-s)^{N-1}$ for $0\le s\le d/2$, if we set $h(r)=f(d/2-r)$ then this solves
\[
\left\{
\begin{array}{lc}
-\big(r^{N-1}\,|h'|^{p-2}\,h'\big)'=\eta \, r^{N-1}\,|h|^{p-2}\, h, & \mbox{ in } (0,d/2), \\&\\
h'(0)=h(d/2)=0,&
\end{array}
\right.
\]
which means that the radial function $H(x)=h(|x|)$ is a Dirichlet eigenfunction of $-\Delta_p$ of a $N$-dimensional ball of radius $d/2$, namely $B$. Hence $\eta\ge \lambda_{p,p}(B)$ and we get
\[
\liminf_{k\to\infty} \mu_{p,p}(\Omega_k)=\liminf_{k\to\infty}\gamma_k(\Omega_1)\ge\lambda_{p,p}(B).
\]
This concludes the proof.
\end{proof}
From Theorem \ref{p_cop} and the isodiametric inequality
\begin{equation}
\label{isodiametric}
\frac{\mathrm{diam}(B)}{\mathrm{diam}(\Omega)}\le \left(\frac{|B|}{|\Omega|}\right)^\frac{1}{N},
\end{equation}
we can infer the following upper bound on $\mu_{p,p}$, in terms of the $N-$dimensional measure. We recall that for $p=2$ this is indeed a consequence of Szeg\H{o}-Weinberger inequality.
\begin{coro}
\label{coro:weaksw}
Let $\Omega\subset\mathbb{R}^N$ be an open bounded convex set and $1<p<\infty$. Then we have
\[
\mu_{p,p}(\Omega)<\lambda_{p,p}(B)\,\left(\frac{|B|}{|\Omega|}\right)^{\frac{p}{N}},
\]
where $B$ is any $N-$dimensional ball.
\end{coro}
\section{The case $p\not = q$}
\label{sec:4}
In this section we discuss variants and extensions of Theorem \ref{p_cop} for the quantity $\mu_{p,q}$ when $p\not =q$.
\subsection{The case $p<q$} Actually, with the very same proof of Theorem \ref{p_cop} we can prove the following upper bound. This time, the resulting inequality {\it is not sharp} (see Remark \ref{oss:nonva!} below). For this reason, though the argument is the same, we prefer to give a separate statement.
\begin{teo}
\label{p_cop2}
Let $\Omega\subset\mathbb{R}^N$ be an open bounded convex set and $1<p<q<p^*$. Then we have
\begin{equation}
\label{coppolicchio2}
\mu_{p,q}(\Omega)<\lambda_{p,q}(B)\,\left(\frac{\mathrm{diam}(B)}{\mathrm{diam}(\Omega)}\right)^{p+\frac{N\,p}{q}-N},
\end{equation}
where $B$ is any $N-$dimensional ball.
\end{teo}
\begin{proof}
We use the same notation as in the proof of Theorem \ref{p_cop}. With $u\in C^{1,\alpha}(\overline B)\cap C^\infty(B\setminus\{0\})$ we now indicate the function achieving $\lambda_{p,q}(B)$, normalized by the conditions
\begin{equation}
\label{normalizzato}
\|u\|_{L^q(B)}=1\qquad \mbox{ and }\qquad u>0.
\end{equation}
By optimality it solves $-\Delta_p u=\lambda_{p,q}(B)\, u^{q-1}$ in $B$, with homogeneous Dirichlet boundary conditions.
As before, we consider the two caps $\Omega_0$ and $\Omega_1$ and take
\[
\varphi(x)=u(x-x_0)\cdot 1_{\Omega_0}(x)-c\, u(x-x_1)\cdot 1_{\Omega_1}(x)\in W^{1,p}(\Omega),
\]
where $c\in\mathbb{R}$ is chosen so to guarantee $\int_\Omega |\varphi|^{q-2}\,\varphi\,dx=0$.
By inserting $\varphi$ in the Rayleigh quotient defining $\mu_{p,q}(\Omega)$ and proceeding as in Theorem \ref{p_cop} we now end up with
\[
\mu_{p,q}(\Omega)< \lambda_{p,q}(B)\,\frac{\displaystyle \int_{\Omega_0} |u(x-x_0)|^q\, dx+c^p \,\int_{\Omega_1} |u(x-x_1)|^q\, dx}{\displaystyle\left(\int_{\Omega_0} |u(x-x_0)|^q\, dx+c^q\,\int_{\Omega_1} |u(x-x_1)|^q\, dx\right)^\frac{p}{q}}.
\]
The term on the right-hand side is of the form
\[
\frac{A+t^\frac{p}{q}\, B}{(A+t\,B)^\frac{p}{q}}.
\]
For $p<q$ the previous expression is maximal for $t=1$. Such a maximal value is given by $(A+B)^{1-p/q}$, we thus get
\begin{equation}
\label{restinooo}
\mu_{p,q}(\Omega)<\lambda_{p,q}(B)\,\left[\int_{\Omega_0} |u(x-x_0)|^q\, dx+\int_{\Omega_1} |u(x-x_1)|^q\, dx\right]^{1-\frac{p}{q}}.
\end{equation}
We have $1-p/q>0$ and the sum of the two terms into square brackets is less than $1$ by \eqref{normalizzato}, thus we can finally infer \eqref{coppolicchio2}.
\end{proof}
As in the case $p=q$, Theorem \ref{p_cop2} implies the following generalization of Corollary \ref{coro:weaksw}.
\begin{coro}
Let $\Omega\subset\mathbb{R}^N$ be an open bounded convex set and $1<p\le q<p^*$. Then we have
\[
\mu_{p,q}(\Omega)<\lambda_{p,q}(B)\,\left(\frac{|B|}{|\Omega|}\right)^{\frac{p}{N}+\frac{p}{q}-1},
\]
where $B$ is any $N-$dimensional ball.
\end{coro}
\begin{oss}[About sharpness]
\label{oss:nonva!}
This time, the estimate \eqref{coppolicchio2} is not sharp. We keep the same notation as in the proof of Theorem \ref{p_cop2} and still consider $q>p$. By adding and subtracting the term $\lambda_{p,q}(B)$
on the right-hand side of \eqref{restinooo}, recalling \eqref{normalizzato} and using the concavity of $t\mapsto t^{1-p/q}$, we get
\[
\begin{split}
\mu_{p,q}(\Omega)&<\lambda_{p,q}(B)+\lambda_{p,q}(B)\, \left[\left(\int_{\Omega_0\cup \Omega_1} |u|^q\, dx\right)^{1-\frac{p}{q}}-\left(\int_B |u|^q\, dx\right)^{1-\frac{p}{q}}\right]\\
&\le \lambda_{p,q}(B)+\frac{q-p}{q}\, \lambda_{p,q}(B)\, \left[\int_{\Omega_0\cup \Omega_1} |u|^q\, dx-\int_{B} |u|^q \, dx\right]\\
\end{split}
\]
Since $u$ is radially decreasing, a simple rearrangement argument finally gives
\begin{equation}
\label{restino0}
\mu_{p,q}(\Omega)<\lambda_{p,q}(B)-\frac{q-p}{q}\, \lambda_{p,q}(B)\,\int_{B\setminus T_\Omega} |u|^q\, dx
\end{equation}
where $T_\Omega$ is the ball centered at the origin, such that $|T_\Omega|=|\Omega_0\cup\Omega_1|$. Now observe that by using the {\it quantitative isodiametric inequality} (see \cite[Theorem 1]{MPP})
\begin{equation}
\label{isodiam}
\begin{split}
|B\setminus T_\Omega|=|B|-|\Omega_0\cup \Omega_1|&=\Big(|B|-|\Omega|\Big)+\Big(|\Omega|-|\Omega_0\cup \Omega_1|\Big)\\
&\ge \frac{|\Omega|}{C_N}\, \mathcal{A}(\Omega)^2+\Big||\Omega|-|\Omega_0\cup\Omega_1|\Big|,
\end{split}
\end{equation}
where $C_N>0$ is a dimensional constant and $\mathcal{A}(\Omega)$ is the {\it Fraenkel asymmetry} of $\Omega$, defined by
\[
\mathcal{A}(\Omega)=\inf\left\{\frac{2\,|\Omega\setminus \Omega_\#|}{|\Omega_\#|}\, :\, \Omega_\# \mbox{ ball with } |\Omega_\#|=|\Omega|\right\}.
\]
Now suppose that there exists a sequence of convex sets $\{\Omega_n\}_{n\in\mathbb{N}}\subset\mathbb{R}^N$ such that
\[
\mathrm{diam}(\Omega_n)=\mathrm{diam}(B)\qquad \mbox{ and }\qquad \lim_{n\to\infty}\mu_{p,q}(\Omega_n)=\lambda_{p,q}(B),
\]
for $q>p$. Then from \eqref{restino0} one would obtain
\[
\int_{B\setminus T_{\Omega_n}} |u|^q\, dx=0.
\]
Since $u>0$ in $B$, this would imply $|B\setminus T_{\Omega_n}|\to 0$ and thus from \eqref{isodiam}
\begin{equation}
\label{restino}
\lim_{n\to\infty}\mathcal{A}(\Omega_n)=0 \qquad \mbox{ and }\qquad \lim_{n\to\infty} \Big||\Omega_n|-|\Omega_{n,0}\cup\Omega_{n,1}|\Big|=0.
\end{equation}
The first condition in \eqref{restino} implies that $\Omega_n$ converges\footnote{In the $L^1$ sense, i.e. the characteristic functions $\{1_{\Omega_n}\}_{n\in\mathbb{N}}$ converge in $L^1(\mathbb{R}^N)$ to $1_B$.} to a ball, in contrast with the fact that $|\Omega|>|\Omega_0\cup\Omega_1|$ for a ball (see Figure \ref{fig:fig1}).
\begin{figure}
\includegraphics[scale=.25]{fig1}
\caption{The two caps $\Omega_0$ and $\Omega_1$ can not cover the whole ball.}
\label{fig:fig1}
\end{figure}
\end{oss}
As in the case $p=q$, we can then ask whether the following shape optimization problem
\begin{equation}
\label{shape}
\sup\Big\{\mu_{p,q}(\Omega)\, :\, \Omega \mbox{ open and bounded convex set},\, \mathrm{diam}(\Omega)\ge c\Big\},
\end{equation}
admits a solution or not. Quite surprisingly, this time we can infer existence of an optimal shape.
\begin{teo}[Existence of a maximizer]
\label{teo:esistenza}
Let $1<p<q<p^*$, for every $c>0$ problem \eqref{shape} admits a solution, i.e. there exists an open and bounded convex set $\mathcal{K}\subset\mathbb{R}^N$ such that
\[
\mu_{p,q}(\Omega)\, \Big(\mathrm{diam}(\Omega)\Big)^{p+\frac{N\,p}{q}-N}\le \mu_{p,q}(\mathcal{K})\, \Big(\mathrm{diam}(\mathcal{K})\Big)^{p+\frac{N\,p}{q}-N},
\]
for every $\Omega \subset\mathbb{R}^N$ open and bounded convex set.
\end{teo}
\begin{proof}
By Theorem \ref{p_cop2} we already know that the suprem \eqref{shape} is finite.
Let us call it $N_c$ and take a maximizing sequence of admissible sets $\{\Omega\}_{k\in\mathbb{N}}\subset\mathbb{R}^N$. Of course we can assume
\begin{equation}
\label{dalbasso}
\mu_{p,q}(\Omega_k)\ge \frac{N_c}{2}>0,\qquad \mbox{ for every }k\in\mathbb{N}.
\end{equation}
Since $\mu_{p,q}$ scales like a length to a negative power, we can also assume that
\[
\mathrm{diam}(\Omega_k)=c,\qquad \mbox{ for every }k\in\mathbb{N}.
\]
Finally, we can suppose that there exists a uniform constant $\delta>0$ such that
\begin{equation}
\label{salva}
|\Omega_k|\ge \delta,\qquad \mbox{ for every }k\in\mathbb{N},
\end{equation}
since otherwise we would have that $\mu_{p,q}(\Omega_k)$ goes to zero (see Remark \ref{oss:azzero} below).
\par
Thanks to the bound on the diameters, we can assume that the whole sequence $\{\Omega_k\}_{k\in\mathbb{N}}$ is contained in a common compact set $D\subset\mathbb{R}^N$. Thus the sequence is relatively compact for the complementary Hausdorff topology in $D$: more precisely, there exists an open set $\Omega\subset D$ such that $\Omega_k$ (up to a subsequence) converges in the Hausdorff complementary distance to $\Omega$ (see \cite[Corollaire 2.2.24]{HP}). Moreover, $\Omega$ is still convex and its diameter equals $c$ (see \cite[Section 2.2.3]{HP}). We also observe that the characteristic functions $\{1_{\Omega_k}\}_{k\in\mathbb{N}}$ converges to $1_\Omega$ strongly\footnote{By convexity, a uniform bound on $\mathrm{diam}(\Omega_k)$ implies a uniform bound on their perimeters and measures. Then it is sufficient to use the compact embedding $BV(D)\hookrightarrow L^1(D)$, where $BV(D)$ is the space of functions with bounded variation.} in $L^1(D)$ and $\ast-$weakly in $L^\infty(D)$.
\par
Without loss of generality, we can assume that $\Omega$ contains the origin, since $\mu_{p,q}(\Omega)$ is not affected by translations.
We are now going to prove that
\begin{equation}
\label{scs}
\limsup_{k\to\infty} \mu_{p,q}(\Omega_k)\le \mu_{p,q}(\Omega).
\end{equation}
At this aim, let us take $u\in W^{1,p}(\Omega)$ a function attaining the infimum in the definition of $\mu_{p,q}(\Omega)>0$. Since $\Omega$ contains the origin, for every $\varepsilon>0$ the set $\Omega^\varepsilon:=(1+\varepsilon)\,\Omega$ is such that
\[
\Omega\Subset \Omega^\varepsilon.
\]
Then by Hausdorff convergence for every $\varepsilon>0$ there exists $k_\varepsilon\in\mathbb{N}$ such that
\[
\Omega_k\subset \Omega^\varepsilon,\qquad \mbox{ for every }k\ge k_\varepsilon.
\]
We also set
\[
u_\varepsilon(x)=u\left(\frac{x}{1+\varepsilon}\right),\qquad x\in \Omega^\varepsilon,
\]
then for every $0<\varepsilon<1$ and every $k\ge k_\varepsilon$, we take $t_{\varepsilon,k}\in\mathbb{R}$ such that
\[
\int_{\Omega_k} |u_\varepsilon-t_{\varepsilon,k}|^q\, dx=\min_{t\in\mathbb{R}}\int_{\Omega_k} |u_\varepsilon-t|^q\, dx,
\]
We claim that the sequence $\{t_{\varepsilon,k}\}_{k\in\mathbb{N}}$ in bounded uniformly in $k$ and $0<\varepsilon<1$, i.e. there exists $C>0$ such that
\begin{equation}
\label{tt}
|t_{\varepsilon,k}|\le C,\qquad \mbox{ for every }\quad 0<\varepsilon<1\quad \mbox{ and }\quad k\ge k_\varepsilon.
\end{equation}
Indeed, observe that by convexity of the map $\tau\mapsto \tau^q$ and \eqref{salva}, we have
\[
\begin{split}
\int_{\Omega_k} |u_\varepsilon-t_{\varepsilon,k}|^q\, dx&\ge \frac{1}{2^{q-1}}\, |\Omega_k|\, |t_{\varepsilon,k}|^q-\int_{\Omega_k} |u_\varepsilon|^q\, dx\\
&\ge \frac{\delta}{2^{q-1}}\, |t_{\varepsilon,k}|^q-\int_{\Omega_\varepsilon} |u_\varepsilon|^q\, dx\\
&=\frac{\delta}{2^{q-1}}\, |t_{\varepsilon,k}|^q\,-(1+\varepsilon)^{N}\, \int_\Omega |u|^q\, dx,
\end{split}
\]
and on the other hand
\[
\begin{split}
\int_{\Omega_k} |u_\varepsilon-t_{\varepsilon,k}|^q\, dx&\le \frac{\displaystyle\int_{\Omega_k} |\nabla u_\varepsilon|^p\, dx}{\mu_{p,q}(\Omega_k)}\le 2\,\frac{(1+\varepsilon)^{N-p}}{N_c}\, \int_\Omega |\nabla u|^p\, dx,
\end{split}
\]
where we used \eqref{dalbasso} and the very definition fo $u_\varepsilon$.
By keeping the two estimates together, we finally get \eqref{tt}.
\par
Thus we can suppose that $t_{\varepsilon,k}$ converges (up to a subsequence) to $t_\varepsilon\in\mathbb{R}$ as $k$ goes to $\infty$, and $t_\varepsilon$ is in turn uniformly bounded. Then we get
\[
\limsup_{k\to\infty} \mu_{p,q}(\Omega_k)\le \limsup_{k\to\infty}\frac{\displaystyle\int_{\Omega_k} |\nabla u_\varepsilon|^p\, dx}{\displaystyle\left(\int_{\Omega_k} |u_\varepsilon-t_{\varepsilon,k}|^q\,dx \right)^{p/q}}\le \frac{\displaystyle\int_{\Omega} |\nabla u_\varepsilon|^p\, dx}{\displaystyle\left(\int_{\Omega} |u_\varepsilon-t_{\varepsilon}|^q\,dx \right)^{p/q}}
\]
for every $0<\varepsilon<1$, where we also used the $\ast-$weak convergence of the characteristic functions, recalled above. We now observe that
\[
\lim_{\varepsilon\to 0} \int_\Omega |u_\varepsilon-t_\varepsilon|^q\, dx=\int_\Omega |u-\widetilde t|^q\, dx\ge \min_{t\in\mathbb{R}} \int_\Omega |u-t|^q\, dx,
\]
where $\widetilde t\in\mathbb{R}$ is an accumulation point of the net $\{t_\varepsilon\}_{\varepsilon>0}$, and also
\[
\lim_{\varepsilon\to 0} \|\nabla u_\varepsilon-\nabla u\|_{L^p(\Omega)}=0.
\]
Thus it is now sufficient to take the limit as $\varepsilon$ goes to $0$ in order to get \eqref{scs}, by arbitrariness of $u$. This finally gives that $\Omega$ is a solution of \eqref{shape}.
\end{proof}
\begin{oss}[Lower bounds and minimization]
\label{oss:azzero}
For $p<q$ the quantity $\mu_{p,q}(\Omega)$ can not be bounded {\it from below} in terms of $\mathrm{diam}(\Omega)$ only. In other words, for $q>p$ we have
\[
\inf\Big\{\mu_{p,q}(\Omega)\, :\, \Omega\subset\mathbb{R}^N \mbox{ convex},\ \mathrm{diam}(\Omega)\le c\Big\}=0.
\]
A minimizing sequence is given by any family of convex sets $\{\Omega_k\}_{k\in\mathbb{N}}\subset\mathbb{R}^N$ such that
\begin{equation}
\label{vanishing}
\lim_{k\to\infty} |\Omega_k|=0\qquad \mbox{ and }\qquad \mathrm{diam}(\Omega_k)=c.
\end{equation}
Indeed, by applying Lemma \ref{lm:comparison} we get
\[
\mu_{p,q}(\Omega)\le |\Omega|^{1-\frac{p}{q}}\, \mu_{p,p}(\Omega).
\]
If we now apply Theorem \ref{p_cop} to the right-hand side, we gain
\[
\mu_{p,q}(\Omega)<\lambda_{p,p}(B)\,|\Omega|^{1-\frac{p}{q}}\,\left(\frac{\mathrm{diam}(B)}{\mathrm{diam}(\Omega)}\right)^{p}.
\]
Thus for a sequence of convex sets verifying \eqref{vanishing}, we get that $\mu_{p,q}(\Omega_k)$ converges to $0$.
\end{oss}
\subsection{The case $p>q$}
In this case, we can show that an upper bound on $\mu_{p,q}$ like that of \eqref{coppolicchio2} can not hold true and actually we have
\[
\sup\{\mu_{p,q}(\Omega)\, :\, \Omega\subset\mathbb{R}^N \mbox{ convex},\ \mathrm{diam}(\Omega)\ge c\}=+\infty.
\]
Indeed, a maximizing sequence is given by any family of open convex sets $\{\Omega_k\}_{n\in\mathbb{N}}\subset\mathbb{R}^N$ such that
\[
\mathrm{diam}(\Omega_k)=c>0 \qquad \mbox{ and }\qquad \lim_{n\to\infty} |\Omega_k|=0.
\]
Actually, this is a consequence of estimate \eqref{op!} below.
\begin{prop}
Let $1<q<p$ and $\Omega\subset\mathbb{R}^N$ be an open and bounded convex set. Then we have
\begin{equation}
\label{op!}
\mu_{p,q}(\Omega)\ge \left(\frac{\pi_p}{\mathrm{diam}(\Omega)}\right)^p\, |\Omega|^{\frac{q}{p}-1},
\end{equation}
and
\begin{equation}
\label{controcoppolicchio}
\mu_{p,q}(\Omega)\ge \left(\frac{\pi_p}{|B|^{\frac{1}{q}-\frac{1}{p}}\,\mathrm{diam}(B)}\right)^p\,\left(\frac{\mathrm{diam}(B)}{\mathrm{diam}(\Omega)}\right)^{p+\frac{N\,p}{q}-N},
\end{equation}
where the constant $\pi_p$ is given by \eqref{pip}.
\end{prop}
\begin{proof}
Again by Lemma \ref{lm:comparison} with $s=p>q$, we get
\[
\mu_{p,p}(\Omega)\le |\Omega|^{1-\frac{q}{p}}\, \mu_{p,q}(\Omega).
\]
By using the lower bound \eqref{mild4}, we can obtain \eqref{op!}.
\vskip.2cm\noindent
Estimate \eqref{controcoppolicchio} is obtained by combining \eqref{op!} with the isodiametric inequality \eqref{isodiametric}.
\end{proof}
The estimate \eqref{controcoppolicchio} is the counterpart of Theorem \ref{p_cop2} for the case $q<p$. Thus this time it is the minimum problem
\[
\inf\Big\{\mu_{p,q}(\Omega)\, :\, \Omega\subset\mathbb{R}^N \mbox{ open and convex},\ \mathrm{diam}(\Omega)\le c\Big\},
\]
that actually makes sense. By suitably adapting the proof of Theorem \ref{teo:esistenza}, one can see that the previous problem admits indeed a solution. We leave the details to the interested reader.
\section{A nodal domain property}
\label{sec:5}
If $u$ is a function achieving the infimum in the problem defining $\mu_{p,q}(\Omega)$, then by {\it nodal domain} we mean every connected component of the (open) sets
\[
\{x\in\Omega\, :\, u(x)>0\}\qquad \mbox{ and }\qquad \{x\in\Omega\,: \, u(x)<0\}.
\]
As a consequence of Theorems \ref{p_cop} and \ref{p_cop2}, in the case $p\ge q$ we have the following result.
\begin{prop}
\label{prop:nodal}
Let $\Omega\subset\mathbb{R}^N$ be an open and bounded convex set and $1<p\le q<p^*$. Then
\begin{equation}
\label{debole}
\mu_{p,q}(\Omega)<\lambda_{p,q}(\Omega).
\end{equation}
Moreover, every nodal domain of a function achieving $\mu_{p,q}(\Omega)$ has to intersect $\partial\Omega$.
\end{prop}
\begin{proof}
The proof of \eqref{debole} immediately follows by combining \eqref{coppolicchio}, the Faber-Krahn inequality
\[
|B|^{\frac{p}{q}+\frac{p}{N}-1}\, \lambda_{p,q}(B)\le |\Omega|^{\frac{p}{q}+\frac{p}{N}-1}\, \lambda_{p,q}(\Omega)
\]
and the isodiametric inequality.
\par
To prove the second assertion, let us argue by contradiction. We take $v$ achieving $\mu_{p,q}(\Omega)$ and we assume that the open set $\{x\in\Omega\, :\, v>0\}$ has a connected component $\omega\Subset \Omega$. We can further suppose that $\|v\|_{L^q(\Omega)}=1$,
then $v\in W^{1,p}_0(\omega)$ and it solves
\[
-\Delta_p v=\mu_{p,q}(\Omega)\, v^{q-1},\qquad \mbox{ in }\omega,
\]
so that
\[
\int_{\omega} |\nabla v|^p=\mu_{p,q}(\Omega)\, \int_\omega |v|^q\, dx\le \mu_{p,q}(\Omega)\, \left(\int_{\omega} |v|^{q}\, dx\right)^\frac{p}{q},
\]
thanks to the fact that $1=\|v\|_{L^q(\Omega)}\ge \|v\|_{L^q(\omega)}$ and $p/q\le 1$. This yields $\lambda_{p,q}(\omega)\le \mu_{p,q}(\Omega)$.
By using the strict monotonicity of $\lambda_{p,q}(\Omega)$ with respect to set inclusion and \eqref{debole}, we then get
\[
\lambda_{p,q}(\Omega)<\lambda_{p,q}(\omega)\le \mu_{p,q}(\Omega)<\lambda_{p,q}(\Omega),
\]
which gives the desired contradiction.
\end{proof}
\begin{oss}
When $p=q=2$, the previous argument to infer that first nontrivial Neumann eigenfunctions can not have a closed nodal line was originally due to Pleijel (see \cite{Pl}). For the Laplacian, inequality \eqref{debole} was conjectured by Kornhauser and Stakgold (see \cite{KS}) and can be obtained (again) as a byproduct of the Szeg\H{o}--Weinberger inequality \eqref{SWscaling}.
\end{oss}
|
\section{Introduction}
It was shown by Ba\~{n}ados, Silk and West (hereafter, BSW) \cite{ban} that
near-horizon collision of two particles moving towards the extremal Kerr
black hole can result in the indefinite growth of their energy $E_{c.m.}$ in
the center of mass frame. Soon after this observation, several arguments
were pushed forward against the possibility of physical realization of this
effect. The first one consisted in that realistic astrophysical black holes
cannot be exactly extremal \cite{ted}. However, it was refuted since the BSW
effect was extended to the nonextremal horizons of Kerr \cite{gp} and other
stationary axially symmetric black holes \cite{prd}. Another objection was
based on the role of gravitational radiation, which was assumed to bound the
BSW effect \cite{berti}. However, recent studies showed that under rather
general and weak assumptions, the BSW effect survives even if a force
(modeling the effect of radiation, backreaction, etc.) acts on the
particles \cite{rad}. This was obtained for the extremal horizons. Now our
goal is to consider the possibility of the BSW effect near horizons of
nonextremal black holes when particles move under the action of some force.
It is worth mentioning that there are two kinds of potential limitations on
the BSW effect. The first one concerns the possibility to get unbounded $E_{c.m.}$, which involves only processes in the immediate vicinity of the horizon. The second kind is related to the issue of astrophysical relevance and potential observational significance of the BSW effect. In this regard the behavior of debris after collision in the asymptotically flat region is also important. The observable energy and mass at infinity for the extremal Kerr metric were found to be restricted by some upper limits in \cite{p} (built on \cite{pir3}), and a similar result was obtained in \cite{j}. Extension to more general ``dirty'' black holes was done in \cite{z}. A
separate question is whether fluxes at infinity can exceed the sensitivity of
modern devices. In general, the situation remains controversial -- \cite{mc},
\cite{com}. Here we discuss only the first kind of limitation, having
obvious theoretical value, and put aside the second kind, which is important
but needs separate further treatment. Up to date, there are already many
other different aspects of the BSW effect that remain beyond the scope of
the present paper.
In this paper we discuss the BSW effect under the action of a force of a
rather generic character. When the corresponding results are applied to the
question regarding gravitational self-force, important reservations are in
order. The true gravitational self-force differs from simple external
force and depends on the particle's position, velocity, etc. in a highly
nonlinear way. The full analysis also needs to
consider the motion of the particle in nonstationary and nonaxisymmetric background \cite{bar}. In this sense, the present paper, as well as our previous one \cite{rad}, should be considered the first step to understanding BSW effect under the action of gravitational self-force,
which we model with the help of a``usual'' force.
\section{General formulas}
Let us consider the axially symmetric stationary black hole metric%
\begin{equation}
ds^{2}=-N^{2}dt^{2}+g_{\phi }(d\phi -\omega dt)^{2} +\frac{dr^{2}}{A}%
+g_{z}dz^{2}.
\end{equation}%
All metric coefficients do not depend on $t$ and $\phi $. The lapse function
$N$ turns to zero at the horizon, where $N^2 \sim A$.
Let us consider particle's motion in the equatorial plane and, for
simplicity, put the mass of each particle equal to unity $m_{1,2}=1$.
Hereafter we will use two frames, which are convenient for description of
the processes near horizon -- the ``OO frame'', which is attached to an
observer orbiting a black hole with the zero angular momentum \cite{72} and
the ``FO frame'', which is attached to an observer falling into the black
hole.
It is convenient to parametrize a particle's for-velocity $u^\mu$ by its
energy $E$ and angular momentum (here $\mu=t,\phi,r,z$)
\begin{equation}
u_\mu =(-E, L, u_r , 0). \label{EL}
\end{equation}
Then the normalization condition for the four-velocity $u^\mu u_\mu =-1$ can
be presented as
\begin{equation}
\frac{1}{A}(u^{r})^{2}=\frac{X^{2}}{N^{2}}-\frac{L^{2}}{g_{\phi }}-1,
\label{norm}
\end{equation}
where
\begin{equation}
X=E-\omega L.
\end{equation}
Equation (\ref{norm}) can be rewritten again to give
\begin{align}
u^{r}&=\pm \frac{\sqrt{A}}{N}\;Z, \label{pr} \\
Z^{2}&=X^{2}-N^{2}\Big(\frac{L^{2}}{g_{\phi }}+1\Big) . \label{z}
\end{align}
For free motion Eqs. (\ref{norm}) and (\ref{pr}) can be obtained, as usual,
as the first integrals of the geodesic equations. If the motion is not
geodesic, the equations remain valid, but $E$ and $L$ are not, in general,
integrals of motion anymore, and should be treated as useful notation only.
Let us denote the components of acceleration in the OO frame by $a_{o}^{(\mu
)}$. For simplicity, we assume hereafter that $A=N^{2}$ (if this is not so,
one can redefine the radial coordinate for motion in the equatorial plane to
achieve this). Then, using Eqs. (112), (116) and (117) of \cite{rad}, we
have
\begin{align}
a_{o}^{(\phi )}& =-\frac{Z}{\sqrt{g_{\phi }}}L^{\prime }; \label{EQao-phi}
\\
a_{o}^{(t)}& =-\frac{Z}{N}(X^{\prime }+L\omega ^{\prime }) =-\frac{%
Z(E^{\prime }-\omega L^{\prime })}{N}; \label{EQao-t} \\
a_{o}^{(r)}& =-\frac{X}{Z}a_{o}^{(t)}-N\frac{LL^{\prime }}{g_{\phi }}.
\label{EQao-r}
\end{align}
By definition, the energy in the center of mass frame of colliding particles
having unit masses is equal to
\begin{equation}
E_{c.m.}^{2}=-(u_{1}^{\mu }+u_{2}^{\mu })(u_{1\mu }+u_{2\mu })
=2+2\gamma_{c.m.}, \label{cm}
\end{equation}
where
\begin{equation}
\gamma _{c.m.}=-u_{1\mu }u_{2}^{\mu }
\end{equation}
is the relative Lorentz factor. Assuming that both particles move towards
the black hole, so that in (\ref{pr}) we should take the minus sign, the
direct calculation, using (\ref{EL}) and (\ref{norm}), gives
\begin{equation}
\gamma _{c.m.} =\frac{X_{1}X_{2}-Z_{1}Z_{2}}{N^{2}} -\frac{L_{1}L_{2}}{%
g_{\phi}}. \label{ga}
\end{equation}
\section{Critical and near-critical particles}
In the context of the BSW effect a particle is called ``usual'' if $%
X_{H}\neq 0$ and ``critical'' if $X_{H}=0$ (subscript ``H" here denotes
quantities calculated on the horizon). The effect near extremal horizons is
realized in collision of one usual and one critical particles. Let us
suppose now that there are two particles colliding near a nonextremal
horizon, where
\begin{equation}
N^{2}\sim \xi,\qquad \xi\equiv \frac{r-r_H}{r_H} . \label{nx}
\end{equation}
For the critical particle in the horizon limit, according to (\ref{z}), we
would have $Z^2 <0$, which means that it cannot actually reach the horizon
\cite{prd}: exactly critical particles do not exist in this case. However,
let us then consider a usual particle, for which the expansion
\begin{equation}
X=X_{H}+x_{1}\xi +x_{2}\xi ^{2}+\cdots
\end{equation}
starts from the first nonvanishing term $X_H \neq 0$. Such a particle
evidently can reach the horizon, and we can choose its point of collision
with another particle close to the horizon $\xi_c \ll 1$ (subscript ``c''
denotes the point of collision). Additionally, we can choose $X_{H}$ to be
small to the same order
\begin{equation}
X_{H}\sim N_c\sim \sqrt{\xi _{c}}. \label{xn}
\end{equation}
Such a particle is called ``near critical'', and
\begin{equation}
Z_H\sim N_c \sim \sqrt{\xi _{c}}. \label{zn}
\end{equation}
Then, in case the other particle is usual, in accordance with \cite{gp,prd}, their relative Lorentz factor at the point of collision is
\begin{equation*}
\gamma_{c.m.}\sim N^{-1}_c ,
\end{equation*}
which can be made arbitrarily large by choosing the point of collision $%
\xi_c $ sufficiently close to the horizon and tuning $X_H$ accordingly.
\section{Behavior of the force}
As seen in the previous section, the formal description of the BSW effect on
the kinematic level does not change with the introduction of force \cite%
{gp,prd}. However, in order to understand whether the effect is actually
preserved, we should check i) if it is compatible with the force acting on
particles being \emph{finite} (an obvious physical requirement) and ii)
whether it is possible, given some reasonably arbitrary finite force, to
find/tune the near-critical particle, which is needed for the effect. Here
we will take advantage of the results of analysis performed in \cite{rad}
for extremal horizons to show that the main conclusions remain valid for
nonextremal ones.
The force acting on a particle, which must be bounded, is the one calculated
in that particle's frame. We are interested only in usual particles here, as
critical ones do not exist near nonextremal horizons (see above). The FO
frame \cite{rad} is constructed so that a usual particle's Lorentz factor in
this frame is finite, thus the force acting on the particle in the FO frame
must be finite as well. The OO frame is related to the FO frame through the
Lorentz boost, which is singular in the horizon limit, with $\gamma \sim 1/N
\to \infty$, so the components of the force in the OO frame can diverge.
Let us denote the tetrad components of a particle's acceleration in the FO
frame, which must be finite, as $a_{f}^{(i)}$. The components of
acceleration in the OO frame $a_{o}^{(i)}$ are related with them via the
singular Lorentz boost, and the explicit relations between their asymptotics
are given by Eqs. (68)--(71) of \cite{rad}:
\begin{align}
a_{f}^{(t)}& =(a_{f}^{(t)})_{0}+(a_{f}^{(t)})_{1}N+O(N^{2}); \label{af-t} \\
a_{f}^{(r)}& =(a_{f}^{(r)})_{0}+(a_{f}^{(r)})_{1}N+O(N^{2}); \label{af-r} \\
a_{o}^{(t)}& =+\frac{(a_{f}^{(t)})_{0}-(a_{f}^{(r)})_{0}}{N}+\big[%
(a_{f}^{(t)})_{1}-(a_{f}^{(r)})_{1}\big]+O(N); \label{ao-t} \\
a_{o}^{(r)}& =-\frac{(a_{f}^{(t)})_{0}-(a_{f}^{(r)})_{0}}{N}-\big[%
(a_{f}^{(t)})_{1}-(a_{f}^{(r)})_{1}\big]+O(N). \label{ao-r}
\end{align}
Thus $a_{o}^{(t)}$ and $a_{o}^{(r)}$ can diverge as $1/N$. The $\phi $ and $%
z $ components are the same in the two frames and must be bounded, so
according to Eqs. (72) and (73) of \cite{rad},
\begin{equation}
a_{f}^{(\phi )}=a_{o}^{(\phi )}=O(1). \label{o1}
\end{equation}
This behavior is insensitive to the type of the horizon, extremal or not.
Let us see if this asymptotic behavior is compatible with ``equations of
motion'' (\ref{EQao-phi})--(\ref{EQao-r}) for near-critical particles. Using
the asymptotes (\ref{xn}) and (\ref{zn}) and assuming that $L$, $E$, $L^{\prime}
$ and $E^{\prime}$ are finite, we get
\begin{align}
& a_o^{(\phi)}\sim \sqrt{\xi}; \label{phi} \\
& a_o^{(t)},a_o^{(r)}\sim 1. \label{atr}
\end{align}
Here we have omitted the subscript ``c''. We see that the kinematic
restrictions (\ref{ao-t})--(\ref{o1}) are satisfied, and the dynamic
constraints (\ref{phi}) and (\ref{atr}) are even stronger than the kinematic
ones. This means that it is the eqs. (\ref{phi}), (\ref{atr}) that
constitute the actual constraints on the behavior of the force near the
horizon, where collision occurs, for near-critical particles to exist.
In the same way, one can check that for usual particles with $X_H , Z_H \neq
0 $ the dynamic constraints, which follow from (\ref{EQao-phi}) and (\ref{EQao-r}), coincide with the kinematic ones (\ref{ao-t})--(\ref{o1}):
\begin{equation}
a_o^{(\phi)}\sim 1,\quad a_o^{(r)},a_o^{(t)}\sim 1/N .
\end{equation}
\section{Example: Reissner-Nordstr\"{o}m metric}
For the purely radial motion in the Reissner-Nordstr\"{o}m metric the
equations of motion of a particle with mass $m$ and charge $q$ read
\begin{align}
ma_{o}^{(t)} =-&\frac{qQ}{r^{2}}\,\frac{Z}{mN}, \\
ma_{o}^{(r)} =+&\frac{qQ}{r^{2}}\,\frac{X}{mN}, \\
& m^{2}a^{2}=\Big(\frac{qQ}{r^{2}}\Big)^{2}.
\end{align}
For near-critical particles with $X\sim Z\sim N$ [(\ref{xn}) and (\ref{zn})], we
get
\begin{equation*}
a_o^{(t)},a_o^{(r)}\sim 1 ,
\end{equation*}
analogously to (\ref{atr}). For usual particles, $X_H , Z_{H}\neq 0$, so we
have
\begin{equation*}
a_o^{(t)},a_o^{(r)}\sim 1/N ,
\end{equation*}
which is still allowed, according to Eqs. (\ref{ao-t}), (\ref{ao-r}).
\section{Near-critical particles and effect of dissipation}
The second question that may not be quite clear \textit{a priori} is whether it is
always possible to fine-tune a near-critical particle. Let us suppose that
dissipation is neglected. Then, the solution $x^{\mu }(n)$ is specified by
initial data and for each set of data there exists a single solution.
Instead of fixing conditions at the initial moment of time, however, we can
fix them at the moment when the near-critical particle reaches the horizon: $%
r(0)=r_{H}$, $\dot{r}(0)=X_{H}$ [see (\ref{pr})].
As, by assumption, dissipation is neglected, the system is time symmetric
with respect to time inversion $t\mapsto -t$. Therefore, by integrating
equations of motion back in time, we can recover the trajectory that leads
to near-horizon collision with the unbound $E_{c.m.}$ This is achieved by
taking arbitrarily small $X_{H}$ from the very beginning. Thus the BSW
effect survives in spite of the presence of the force.
In practice, however, a particle can experience the influence of dissipative
forces, such as gravitational radiation reaction. Either dissipation arises
due to terms proportional to velocity (or its higher odd powers) or it
cannot be described in terms of forces at all. What is important,
dissipation violates the symmetry between the two directions of time, which
devaluates the above reasoning. However, if dissipation is small enough, the
presented arguments retain their validity.
Dissipation is small if the time of relaxation $\tau_{diss}$ is much greater
than the characteristic dynamic time scales. In the context of gravity, we
should compare proper time intervals. Let us consider motion of the
near-critical particle. Such a particle moves between the horizon and the
turning point $r=r_{0}$, so collision occurs somewhere within this interval,
which shrinks to zero when $X_{H}\rightarrow 0$ (see for details \cite{gp42} and \cite{circ}).
The proper time of movement until collision is less than the time of
movement from the horizon to the turning point $r=r_{0}$, so its upper
estimate is
\begin{equation}
\tau _{dyn}=\int\limits_{r_{H}}^{r_{0}}\frac{dr}{|u^{r}|}. \label{dist}
\end{equation}%
Then, for small $r-r_{H}$, we have
\begin{equation}
N^{2}\approx 2\kappa (r-r_{H})=2\kappa r_{H}\xi , \qquad 0 \leq \xi_c \leq
\xi \leq \xi _{0},
\end{equation}%
where $\xi _{0}$ corresponds to the turning point. In accordance with (\ref%
{xn}), we have $X_{H}^{2}=\xi _{0}b$, with $b=2\kappa r_{H}(\frac{L^{2}}{%
g_{\phi }}+1)=O(1)$.
Now, it follows from Eqs. (\ref{pr}) and (\ref{z}) that
\begin{equation}
\tau \approx 2\frac{r_{H}}{\sqrt{b}}\sqrt{\xi _{0}}.
\end{equation}%
Thus the effect of dissipation is small as long as
\begin{equation}
\frac{\tau _{diss}}{r_{H}}\gg \xi _{c}, \label{dis}
\end{equation}%
which holds automatically for sufficiently small $\xi _{c}$. If a particle
is usual, the effect of dissipation is irrelevant in the context of the BSW
effect at all since it simply transforms a usual trajectory into another
usual one. Thus for near-horizon collision the dissipation effects can be
neglected and cannot restrict the BSW effect, so that the energy in the
center of mass frame can be made arbitrarily large.
It is worth noting that the above discussion does not apply directly to the
case of extremal horizon, since the proper time of reaching the extremal
horizon for a critical particle is infinite. In that case, however, the
existence of the BSW effect is confirmed via different reasoning, based on
the direct analysis of near-horizon trajectories \cite{rad}.
\section{Explicit procedure of tuning}
In this section we demonstrate explicitly, how the procedure of tuning can
be realized for near-critical particles. As the particle is not exactly
critical, tuning should be understood in the approximate sense (small but
nonzero $X_{H}$ on the horizon). As an example, we consider the case of the
azimuthal force, when $a_{o}^{(r)}=0$. Then, it follows from (\ref{EQao-phi}%
) that
\begin{equation}
g_{\phi }X(X^{\prime }+L\omega ^{\prime })=N^{2}LL^{\prime } \label{xg}
\end{equation}%
which corresponds to Eq. (134) of \cite{rad}.
Now, we fix small $X_{H}$ and seek the solution in the form of series%
\begin{align}
N^{2}&=2\kappa \xi +\nu _{2}\xi ^{2}+\nu _{3}\xi ^{3}+\cdots \\
\omega &=\omega _{H}-\omega _{1}\xi +\omega _{2}\xi ^{2}+\cdots , \\
g_{\phi }&=g_{H}+g_{1}\xi +g_{2}\xi ^{2}+\cdots , \\
X&=X_{H}+x_{1}\xi +x_{2}\xi ^{2}+\cdots , \label{xe} \\
L&=l_{H}+l_{1}\xi +l_{2}\xi ^{2}+\cdots .
\end{align}
Equating the terms of the zeroth order by $\xi $, we obtain from (\ref{xg})
that%
\begin{equation}
x_{1}=\omega _{1}l_{H}.
\end{equation}
The terms of the first order entail
\begin{equation}
l_{1}=\frac{2g_{H}X_{H}(x_{2}+\omega _{2}l_{H})}{2\kappa
l_{H}+g_{H}X_{H}\omega _{1}}.
\end{equation}%
Repeating the procedure iteratively, we get $l_{2}=l_{2}(X_{H},x_{1},x_{2})$%
, etc. Substituting (\ref{xe}) into (\ref{z}), we find%
\begin{align}
Z^{2}&=X_{H}^{2}-z_{1}\xi , \\
z_{1}&=2\kappa \Big(\frac{l_{H}^{2}}{g_{H}}+1\Big)
\end{align}
where we neglected the term of the order $X_{H}$ in $z_{1}$.
Then, it follows from (\ref{ga}) that%
\begin{equation}
\gamma _{c.m.}\sim \frac{X_{H}-\sqrt{X_{H}^{2}-z_{1}\xi }}{\xi }\text{.}
\end{equation}%
In the region between the horizon and the turning point $0\leq \xi \leq \xi
_{0}=X_{H}^{2}/ z_{1} $ this factor has the order $X_{H}^{-1}$ and can be
made as large as one likes.
\section{Summary}
We have shown that the BSW effect near nonextremal horizons retains its
validity even if the particle experiences the action of forces, provided
some rather weak and reasonable restrictions are imposed on these forces.
For the near-critical particle that plays the crucial role in the effect,
the corresponding conditions are described by Eqs. (\ref{phi}) and (\ref{atr}). In combination with the previous similar results for extremal horizons
\cite{rad}, this means that the BSW effect turns out to be rather viable and
shows properties of universality. In application of the obtained results
to the issue of gravitation self-force, this should be considered
as the model approach and first approximation only, so the full analysis
requires further study.
|
\section{Introduction}
In this paper we use the Jackson $q$-integral to define an operation of integration with respect to a Markov process which first arose in non-commutative probability and which we shall call the $q$-Brownian motion.
The non-commutative $q$-Brownian motion was introduced in \cite{BS91} who put the formal approach of \cite{frisch2003parastochastics} on firm mathematical footing. Stochastic integration with respect to the non-commutative $q$-Brownian was developed in Ref. \cite{donati2003stochastic}.
According to \cite[Corollary 4.5]{BKS-97}, there exists a unique classical Markov process $(B^{(q)})_{t\in[0,\infty)}$ with the same univariate distributions and the same transition operators as the non-commutative $q$-Brownian motion. Markov process $(B^{(q)})_{t\geq 0}$ is a martingale and which converges in distribution to the Brownian motion as $q\to 1$.
With some abuse of terminology we call $(B^{(q)}_t)$ the $q$-Brownian motion and we review its basic properties in Section \ref{Sec-qB}.
Our definition of integration with respect to $(B^{(q)}_t)$ uses the orthogonal martingale polynomials and mimics the well-known properties of the Ito integral.
The integral of an instantaneous function $f(B^{(q)}_s,s)$ of the process is denoted by $$\int_0^t f(B^{(q)}_s,s)\mathbf{\partial} B^{(q)},$$
where to avoid confusion with the Ito integral with respect to the martingale $(B^{(q)}_t)$ we use $\mathbf{\partial}$ instead of $d$.
We define this integral for $0<q<1$ and for polynomials $f(x,t)$ in variable $x$ with coefficients that are bounded for small enough $t$.
Our definition uses the Jackson integral \cite{Jackson-1910} which is reviewed in Section \ref{Sec-Jackson}. This approach naturally leads to the $q$-analog of the $L_2$-isometry
\begin{equation}\label{EQ-iso}
E\left(\left|\int_0^t f(B^{(q)}_s,s)\mathbf{\partial} B^{(q)}\right|^2\right)=\int_0^tE\left(\left|f(B^{(q)}_s,s)\right|^2\right) d_qs,
\end{equation}
with the Jackson $q$-integral appearing on the right hand side. In Theorem \ref{Thm1} we establish this formula for polynomial $f(x,t)$ and in Corollary \ref{Cor} we use it to extend the integral to more general functions. Proposition \ref{P4.5} shows that the integral is well defined for analytic functions $f(x)$ that do not depend on $t$. In Section \ref{Sec-SDE} we use Corollary \ref{Cor} to exhibit a solution of the "linear $q$-equation" $\mathbf{\partial} Z = a Z \ddB^{(q)}$.
Our second main result is a version of Ito's formula which takes the form
\begin{multline}\label{EQ-ito}
f(B^{(q)}_t,t)-f(0,0)\\ =\int_0^t (\nabla_{x}^{(s)} f)(B^{(q)}_s,s)\ddB^{(q)}+\int_0^t (\mathcal{D}_{q,s} f) (B^{(q)}_{qs},s)d_qs +\int_0^t (\Delta_{x}^{(s)} f) (B^{(q)}_{qs},s)d_qs.
\end{multline}
In Theorem \ref{Thm2}, this formula is
established for polynomial $f(x,t)$ and $q\in(0,1)$.
In Remark \ref{Rem-Ito} we point out that we expect this formula to hold in more generality.
As in the Ito formula, the operators $\mathcal{D}_{q,s}$, $\nabla_x^{(s)}$ and $\Delta_x^{(s)}$ should be interpreted as acting on the appropriate variable of the function $f(x,s)$, which is then evaluated at $x=B^{(q)}_s$ or at $x=B^{(q)}_{qs}$.
The time-variable operator $\mathcal{D}_{q,s} f$ is the $q$-derivative with respect to variable $s$:
\begin{equation}\label{Def:Dq}
(\mathcal{D}_{q,s} f)(x,s)=\frac{f(x,s)-f(x,qs)}{(1-q)s}.
\end{equation}
This expression is well defined for $s>0$ which is all that is needed in \eqref{EQ-ito} when $q\in(0,1)$.
The operators $\nabla_x^{(s)}$ and $\Delta_x^{(s)}$ act only on the ``space variable" $x$ but they depend on the time variable $s$, and are defined as the ``singular integrals" with respect to two time-dependent probability kernels:
\begin{eqnarray}
\label{EQ:nabla} \\ \nonumber
(\nabla_{x}^{(s)} f)(x,s)&=&\int_\mathbb{R} \frac{f(y,s)-f(x,s)}{y-x}\nu_{x,s}(dy) \\
\label{EQ:Delta} \\ \nonumber
(\Delta_{x}^{(s)} f)(x,s)&=&
\iint_{\mathbb{R}^2} \frac{(y-x)f(z,s)+(x-z)f(y,s)+(z-y)f(x,s)}{(x-y)(y-z)(z-x)}\mu_{x,s}(dy,dz).
\end{eqnarray}
Probability measures $\nu_{x,s}(dy)$ and $\mu_{x,s}(dy,dz)$ can be expressed in terms of the transition probabilities $P_{s,t}(x,dy)$ of the Markov process $(B^{(q)}_t)_{t\in[0,\infty)}$ as follows:
$$\nu_{x,s}(dy)= P_{sq^2,s}(qx,dy)$$ and $$\mu_{x,s}(dy,dz)=P_{qs,s}(x,dy)\nu_{y,s}(dz)=P_{qs,s}(x,dy)P_{q^2s,s}(qy,dz).$$ Transition probabilities $P_{s,t}(x,dy)$ appear in \eqref{EQ:Pst} below. The same probability kernel $\nu_{x,s}(dy)$ appears also in \cite[Proposition 21]{anshelevich2013generators}.
As $q\to 1$ probability measures $ \nu_{x,s}(dy)$ and $\mu_{x,s}(dy,dz)$ converge to degenerated measures. Using Taylor expansion one can check that for smooth enough functions $ (\nabla_{x}^{(s)} f)(x,s)$ converges to $\partial f/\partial x$ and
$ (\Delta_{x}^{(s)} f)(x,s)$ converges to $\tfrac12\partial^2 f/\partial x^2$ as $q\to1$., so \eqref{EQ-ito} is a $q$-analog of the Ito formula.
We note that there seems to be some interest in $q$-analogs of the Ito formula. In particular, Ref. \cite{haven2009quantum,haven2011ito} discuss formal $q$-versions of the Ito formula and its
applications to financial mathematics.
\arxiv{
\subsection{Limits as $q\to1$}
We conclude this introduction with the verification that
as $q\to 1$, formulas \eqref{EQ-iso} and \eqref{EQ-ito} coincide with the usual formulas of Ito calculus, at least on polynomials.
Formula \eqref{EQ-iso} becomes the familiar $$E\left(\int_0^t f(W_s,s)dW\right)^2=\int_0^t E(f(W_s,s))^2ds,$$ as
the Jackson $q$-integral becomes the Riemann integral in the limit. As $q\to 1$ formula \eqref{EQ-ito} becomes
$$
f(W_t,t)=\int_0^t \left(\frac{\partial}{\partial x}f\right)(W_s,s)dW+\int_0^t \left(\frac{\partial}{\partial s}f\right)(W_s,s)ds+\frac12 \int_0^t \left(\frac{\partial^2}{\partial x^2}f\right)(W_s,s)ds,
$$
but this is less obvious. Clearly, $\mathcal{D}_{q,s} f$ becomes $\frac{\partial}{\partial s}$ in the limit.
Since $P_{q^2s,s}(qx,dy)\xrightarrow{{\mathcal{D}}} \delta_x(dy)$ in distribution as $q\to 1$ and $\frac{f(y)-f(x)}{y-x}=f'(x)+\frac12 f''(x)(y-x)+\dots$, we see that $\nabla_{x}^{(s)}$ becomes $\frac{\partial}{\partial x}$, when acting on smooth enough functions of variable $x$. Next, we expand the numerator of the integrand in \eqref{EQ:nabla} into the Taylor series at $x$ (here we take $f$ to be a polynomial in variable $x$). We get
\begin{multline}\label{1.6}
(y-x)f(z)+ (x-z)f(y)+(z-y)f(x) \\=(y-x)\left[f(x)+f'(x)(z-x)+\frac12 f''(x)(z-x)^2\right]
\\+(x-z)\left[f(x)+f'(x)(y-x)+\frac12 f''(x)(y-x)^2\right]+(z-y)f(x)\\ +R(x,y,x)
= \frac12(x-y)(y-z)(z-x)f''(x)+R(x,y,z).
\end{multline}
}\arxiv{
After inserting \eqref{1.6} into \eqref{EQ:Delta}, the first term gives $\tfrac12\tfrac{\partial ^2}{\partial x^2}$, since $\mu_{x,s}(dy,dz)$ is a probability measure.
The reminder term $R(x,y,z)$ consists of the higher order terms. The coefficient at $f^{(k)}(x)/k!$ in $R$ is
\begin{multline*}
(y-x)(z-x)^k+(x-z)(y-x)^k=
(y-x)(z-x)\left((z-x)^{k-1} - (y-x)^{k-1}\right)\\=
(x-y)(y-z)(z-x) \sum_{j=0}^{k-2} (z-x)^j(y-x)^{k-2-j},
\end{multline*} and it is clear that for $k\geq 3$ we have $\lim_{q\to 1^-}\int (z-x)^j(y-x)^{k-2-j} \mu_{x,s}(dy,dz)= 0$, since $\mu_{x,s}(dy,dz)\xrightarrow{{\mathcal{D}}} \delta_x(dy)\delta_x(dz)$ in distribution as $q\to1$.
}
\section{$q$-Brownian motion}\label{Sec-qB}
With $B^{(q)}_0=0$, the univariate distribution of $B^{(q)}_t$ is a $q$-Gaussian distribution
\begin{multline}
\label{EQ:q-gauss} \gamma_{t;q}(dy) \\= \frac{\sqrt{1-q}}{2\pi \sqrt{4t-(1-q)y^{2}}}\prod_{k=0}^{\infty
}\left( (1+q^{k})^{2}-(1-q)\frac{y^{2}}{t}q^{k}\right) \prod_{k=0}^{\infty
}(1-q^{k+1}) dy,
\end{multline}
supported on the interval $|y|\leq 2\sqrt{t}/\sqrt{1-q}$. Here, parameter $t$ is the variance. (We note that there are several other distributions that are also called $q$-Gaussian, see the introduction to \cite{diaz2009gaussian}.)
The transition probabilities $P_{s,t}(x,dy)$ are non-homogeneous and, as noted in \cite{Bryc-Wesolowski-2013-gener}, are well defined for all $x\in\mathbb{R}$.
The explicit form will not be needed in this paper, but for completeness we note that for $|x|\leq 2\sqrt{s}/\sqrt{1-q}$ and $s<t$ transition probability $P_{s,t}(x,dy)$ has density supported on the interval $|y|\leq 2\sqrt{t}/\sqrt{1-q}$ which can be written as an infinite product
\begin{multline}
\label{EQ:Pst}
P_{s,t}(x,dy)\\ = \frac{\sqrt{1-q}}{2\pi \sqrt{4t-(1-q)y^{2}}} \prod_{k=0}^{\infty }\frac{
(t-sq^{k})\left( 1-q^{k+1}\right) \left(
t(1+q^{k})^{2}-(1-q)y^{2}q^{k}\right) }{(t-sq^{2k})^{2}-(1-q)
q^{k}(t+sq^{2k})xy +(1-q)(sy^{2}+tx^{2})q^{2k}}dy.
\end{multline}
With $B^{(q)}_0=0$, we have $\gamma_{t;q}(dy)=P_{0,t}(0,dy)$ and the separable version of the process satisfies $|B^{(q)}_t|\leq 2\sqrt{t}/\sqrt{1-q}$ for all $t\geq 0$ almost surely.
Formulas \eqref{EQ:q-gauss} and \eqref{EQ:Pst} are taken from \cite[Section 4.1]{Bryc-Wesolowski-03}, but they are well known, see \cite[Theorem 4.6]{BKS-97} and
\cite{anshelevich2013generators,Bryc-Matysiak-Szablowski,szabl2012q}. We note that for $|x|>2\sqrt{s}/\sqrt{1-q}$ transition probability $P_{s,t}(x,dy)$ has
an additional discrete component; for $q=0$ this discrete component is explicitly written out in \cite[Section 5.3]{Biane98}.
As $q\to1$, the transition probabilities $P_{s,t}(x;dy)$ converge in variation norm to the (Gaussian) transition probabilities of the Wiener process. This can be deduced from the convergence of densities established in the appendix of \cite{ismail1988askey}.
Our definition of the integral relies on the so called continuous $q$-Hermite polynomials $\{h_n(x;t):n=0,1,\dots\}$. These are monic polynomials in variable $x$ that are defined by the three step recurrence
\begin{equation} \label{EQ:rec-q-hermite}
xh_n(x;t)=h_{n+1}(x;t)+t [n] h_{n-1}(x;t),
\end{equation}
where $[n]=[n]_q=1+q+\dots+q^{n-1}$, $h_0(x;t)=1$, $h_1(x;t)=x$. (These are renormalized monic versions of the ``standard" continuous $q$-Hermite polynomials as given in \cite[Section 13.1]{Ismail-05} or in \cite[Section 3.26]{Koekoek-Swarttouw}.)
Polynomials $\{h_n(x;t)\}$ play a special role because they are orthogonal martingale polynomials for $(B^{(q)})$. The orthogonality is
\begin{equation}\label{EQ:ortho}
\int h_n(x;t)h_m(x;t)\gamma_{t;q}(dx)=\delta_{m,n} [n]! t^n,
\end{equation}
where $[n]!=[n]_q!=[1]_q[2]_q\dots[n]_q$. The martingale property
\begin{equation}\label{EQ:mart}
h_n(x;s)=\int h_n(y;t)P_{s,t}(x,dy)
\end{equation}
holds for all real $x$ and $s<t$.
When $q=1$ recursion \eqref{EQ:rec-q-hermite} becomes the recursion for the Hermite polynomials, which are the martingale orthogonal polynomials for the Brownian motion \cite{Schoutens00}.
It is easy to check from \eqref{EQ:rec-q-hermite} that
\begin{equation}
\label{EQ:h(t)2h(1)}
h_n(x;t)=t^{n/2} h_n(x/\sqrt{t};1).
\end{equation}
From \eqref{EQ:h(t)2h(1)} it is clear that if $|x|\leq 2\sqrt{t}/\sqrt{1-q}$ then
\begin{equation}\label{EQ:h-growth}
|h_n(x;t)|\leq C_n t^{n/2}
\end{equation}
for some constant $C_n$.
Explicit sharp bound with $C_n=(1-q)^{-n/2}\sum_{k=0}^n[^n_k]$ can be deduced from \cite[(13.1.10)]{Ismail-05}.
\section{The Jackson $q$-integral}\label{Sec-Jackson}
For reader's convenience we review basic facts about the Jackson integral \cite{Jackson-1910}. In addition to introducing the notation, we explicitly spell out the ``technical" assumptions that suffice for our purposes. In particular, since we will integrate functions defined on $[0,\infty)$ only, we assume that $q\in(0,1)$. We follow \cite{kac2002quantum} quite closely, but essentially the same material can be found in numerous other sources such as \cite[Section 11.4]{Ismail-05}.
Suppose $a:[0,\infty)\to \mathbb{R}$ is bounded in a neighborhood of $0$, and let $b:[0,\infty)\to\mathbb{R}$ be such that
\begin{equation}
\label{EQ:bound}
|b(t)-b(0)|\leq C t^\delta
\end{equation} for some $C<\infty$ and $\delta>0$ in a neighborhood of $0$.
\begin{definition} For $t>0$ and $q\in(0,1)$, the Jackson $q$-integral of $a(s)$ with respect to $b(s)$ is defined as
\begin{equation}
\label{J-int}
\int_0^t a(s)d_q b(s):=\sum_{k=0}^\infty a(q^k t)\left(b(q^k t)-b(q^{k+1}t)\right).
\end{equation}
\end{definition}
When $b(t)=t$, formula \eqref{J-int} takes the following form that goes back to \cite{Jackson-1910}
\begin{equation}
\label{J-ds-int}
\int_0^t a(s)d_q s=(1-q)t\sum_{k=0}^\infty q^k a(q^k t).
\end{equation}
Recalling \eqref{Def:Dq}, it is clear that \eqref{J-int} is $\int_0^t a(s) (\mathcal{D}_{q,s} b)(s) d_q s$.
We will also need the $q$-integration by parts formula \cite[\S5]{Jackson-1910}. This formula is derived in \cite[page 83]{kac2002quantum} under the assumption of differentiability of functions $a(t), b(t)$. Ref. \cite[Theorem 11.4.1]{Ismail-05} gives a $q$-integration by parts formula under more general assumptions, but since the conclusion is stated in a different form than what we need, we give a short proof.
\begin{proposition}
For any pair of functions $a,b$ that satisfy bound \eqref{EQ:bound} near $t=0$ we have
\begin{equation}\label{EQ:by-parts}
\int_0^t a(s)d_q b(s)= a(t)b(t)-a(0)b(0)-\int_0^t b(q s) d_q a(s).
\end{equation}
\end{proposition}
\begin{proof}
Under our assumptions on $a,b$, the series defining $\int_0^t a(s)d_q b(s)$ and $\int_0^t b(q s) d_q a(s)$ converge, so we can pass to the limit as $N\to\infty$ in the telescoping sum identity
\begin{multline*}
\sum_{n=0}^N a(q^n t)\left(b(q^n t)-b(q^{n+1}t)\right)+\sum_{n=0}^N b(q^{n+1}t)\left(a(q^n t)-a(q^{n+1}t)\right) \\ =a(t)b(t)-a(q^{N+1}t)b(q^{N+1}t).
\end{multline*}
\end{proof}
We also note the $q$-anti-differentiation formula, which is a special case $a(s)=1$ of formula \eqref{EQ:by-parts}
\begin{equation}
\label{EQ:antideriv}
b(t)-b(0)=\int_0^t (\mathcal{D}_{q,s} b)(s)d_qs.
\end{equation}
\section{Integration of polynomials with respect to $B^{(q)}_t$}\label{Sec-SI}
Our definition is designed to imply the relation
\begin{equation}\label{WdW}
\int_0^t h_n(B^{(q)}_s,s)\ddB^{(q)}= \frac{1}{[n+1]_q}h_{n+1}(B^{(q)}_t;t)
\end{equation}
and relies on expansion into polynomials $\{h_n(x;t)\}$.
Let $q\in(0,1)$ and suppose that $f(x,t)=\sum_{m=0}^d a_m(t)x^m$ is a polynomial of degree $d$ in variable $x$ with coefficients $a_m(t)$ that depend on $t\geq 0$.
Then, for $t>0$, we can expand $f$ into the $q$-Hermite polynomials \eqref{EQ:rec-q-hermite},
\begin{equation}\label{EQ:f2h}
f(x,t)=\sum_{m=0}^d \frac{b_m(t)}{[m]!}h_m(x;t).
\end{equation}
\begin{lemma}\label{Lem:1} If all the coefficients $a_m(t)$ of the polynomial $f(x,t)$ are bounded in a neighborhood of $0$ then the coefficients $b_m(t)$ in expansion \eqref{EQ:f2h} are bounded in the same neighborhood of $0$. If all the coefficients $a_m(t)$ satisfy estimate \eqref{EQ:bound} then
coefficients $b_m(t)$ satisfy estimate \eqref{EQ:bound}.
\end{lemma}\begin{proof}
Since $h_0,\dots,h_{m-1}$ are monic, they span the polynomials of degree ${m-1}$ so that by orthogonality \eqref{EQ:ortho} we have $\int x^j h_m(x,t)\gamma_{t;q}(dx)=0$ for $j<m$. For the same reason, $\int x^m h_m(x,t)\gamma_{t;q}(dx) =\int h_m^2(x,t)\gamma_{t;q}(dx)=t^m[m]!$. Thus the coefficients are
\begin{multline*}
b_m(t)=\frac{1}{t^{m}} \int f(x,t) h_m(x;t)\gamma_{t;q}(dx) =\frac{1}{t^{m}}\sum_{j=m}^d a_j(t) \int x^j h_m(x;t)\gamma_{t;q}(dx)
\\= \frac{1}{t^{m/2}}\sum_{j=m}^d a_j(t) \int x^j h_m(x/\sqrt{t};1)\gamma_{t;q}(dx),
\end{multline*}
where in the last line we used \eqref{EQ:h(t)2h(1)}. Since random variable $B^{(q)}_t/\sqrt{t}$ has the same distribution as $ B^{(q)}_1$, changing the variable of integration to $y=x/\sqrt{t}$ we get
\begin{multline*}
b_m(t) =\sum_{j=m}^d a_j(t) t^{(j-m)/2}\int y^j h_m(y;1)\gamma_{1;q}(dy) \\= [m]! a_m(t)+\sum_{j=m+1}^d a_j(t) t^{(j-m)/2}\int y^j h_m(y;1)\gamma_{1;q}(dy).
\end{multline*}
Thus $b_m(0)=[m!]a_m(0)$ and for $0<t<1$ we get $|b_m(t)-b_m(0)|\leq [m]! |a_m(t)-a_m(0)|+C \sqrt{t}$.
\end{proof}
\begin{definition}
Let $f(x,t)$ be a polynomial in variable $x$ with coefficients bounded in a neighborhood of $t=0$.
We define the $\mathbf{\partial} B^{(q)}$ integral as the sum of the Jackson $q$-integrals with respect to random functions $h_{m+1}(B^{(q)}_{s};s)$ as follows:
\begin{equation}
\label{EQ:q-Int-1}
\int_0^t f(B^{(q)}_s,s)\ddB^{(q)}= \sum_{m=0}^d \frac{1}{[m+1]!}\int_0^t b_m(s) d_q h_{m+1}(B^{(q)}_{s};s).
\end{equation}
\end{definition}
Since $|B^{(q)}_t|\leq 2\sqrt{t}/\sqrt{1-q}$ for all $t\geq 0$ almost surely, from Lemma \ref{Lem:1} and inequality \eqref{EQ:h-growth}
we see that the Jackson $q$-integrals on the right hand side of \eqref{EQ:q-Int-1} are well defined.
As a special case we get the following formula for the integrals of the deterministic functions.
\begin{example}\label{Ex:4.1} For non-random integrands, we have
$$\int_0^t f(s)\ddB^{(q)}=\int_0^t f(s)d_qB^{(q)}_s.$$
Indeed, since $h_0=1$ the expansion \eqref{EQ:f2h} is $f(s)h_0(x;s)$ and since $h_1(x;s)=x$, we get $ d_q h_{1}(B^{(q)}_{s};s) = d_qB^{(q)}_s$.
By computing the fourth moments one can check that the distribution of random variable $\int_0^t f(s)d_qB^{(q)}_s$ in general is not $\gamma_{\sigma^2,q}$.
\end{example}
When the coefficients of $f(x,t)$ satisfy estimate \eqref{EQ:bound},
we can use \eqref{EQ:by-parts} to write the $\mathbf{\partial} B^{(q)}$ integral as
\begin{multline}
\label{EQ:q-Int-2}
\int_0^t f(B^{(q)}_s,s)\ddB^{(q)}\\=\sum_{m=0}^d \frac{b_m(t)}{[m+1]!}h_{m+1}(B^{(q)}_t;t)-\sum_{m=0}^d \frac{1}{[m+1]!}\int_0^t h_{m+1}(B^{(q)}_{qs};qs)d_q b_m(s).
\end{multline}
Taking $b_m(t)=0$ or $1$ we get formula \eqref{WdW}.
Formula \eqref{EQ:q-Int-2} leads to a couple of explicit examples.
\begin{example} \label{Ex:4.2}
Noting that $x=h_1$, $x^2=h_2(x;t)+t$, and $h_3(x;t)=x^3-(2+q)xt$, we get the following $q$-analogs of the formulas that are well known for the Ito integrals.
$$
\int_0^t B^{(q)}_s\ddB^{(q)} =\frac{1}{1+q} \left((B^{(q)}_t)^2-t\right),
$$
$$
\int_0^t (B^{(q)}_s)^2\ddB^{(q)} =\frac{1}{(1+q)(1+q+q^2)}(B^{(q)}_t)^3-\frac{2+q}{(1+q)(1+q+q^2)}t B^{(q)}_t + \int_0^t s d_qB^{(q)}_s.
$$
\end{example}
We remark that a similar definition could be introduced for a wider class of the $q$-Meixner processes from \cite{Bryc-Wesolowski-2013-gener,Bryc-Wesolowski-03}, compare \cite[Theorem 7]{Schoutens:998lr}.
\subsection{The $L_2$-isometry}
Our first main result is the $q$-analog of the Ito isometry.
\begin{theorem}\label{Thm1}
If $0<q<1$ and $f(x,t)$ is a polynomial in $x$ with coefficients bounded in a neighborhood of $t=0$ then \eqref{EQ-iso} holds.
\end{theorem}
\begin{proof}
With \eqref{EQ:f2h}, by orthogonality \eqref{EQ:ortho} for the $q$-Hermite polynomials, we have
$$
E \left(\left(f(B^{(q)}_s,s)\right)^2\right)=\sum_{m=0}^d \frac{b_m^2(s) s^m}{[m]!}.
$$
So the right hand side of \eqref{EQ-iso} is
\begin{equation}
\label{EQ:RHS}
\sum_{m=0}^d \frac{1}{[m]!} \int_0^tb_m^2(s) s^m d_q(s)=\sum_{m=0}^d \frac{1}{[m+1]!} \int_0^tb_m^2(s) d_q(s^{m+1}).
\end{equation}
On the other hand, expanding the right hand side of \eqref{EQ:q-Int-1} we get
\begin{multline*}
E\left(\left(\int_0^t f(B^{(q)}_s,s)\mathbf{\partial} B^{(q)}\right)^2\right)
=\sum_{m,n=0}^d
\frac{1}{[m+1]![n+1]!}\sum_{k,j=0}^\infty b_m(q^k t)b_n(q^jt) \\
\times E\left(\left(h_{n+1}(B^{(q)}_{tq^k};tq^k)-h_{n+1}(B^{(q)}_{tq^{k+1}};tq^{k+1}) \right) \left(h_{m+1}(B^{(q)}_{tq^j};tq^j)-h_{m+1}(B^{(q)}_{tq^{j+1}};tq^{j+1})\right)\right).
\end{multline*}
Since $h_n(B^{(q)}_t;t)$ is a martingale in the natural filtration,
for $j<k$ we have
$$
E\left(h_{m+1}(B^{(q)}_{tq^j};tq^j)-h_{m+1}(B^{(q)}_{tq^{j+1}};tq^{j+1})|B^{(q)}_{tq^k}, B^{(q)}_{tq^{k+1}} \right)=0.
$$
Similarly,
$$E\left(h_{n+1}(B^{(q)}_{tq^k};tq^k)-h_{n+1}(B^{(q)}_{tq^{k+1}};tq^{k+1})|B^{(q)}_{tq^j}, B^{(q)}_{tq^{j+1}} \right)=0$$
for $j>k$. So the only contributing terms are $j=k$ and
\begin{multline*}
E\left(\left(\int_0^t f(B^{(q)}_s,s)\mathbf{\partial} B^{(q)}\right)^2\right)
=\sum_{m,n=0}^d
\frac{1}{[m+1]![n+1]!}\sum_{k=0}^\infty b_m(q^k t)b_n(q^kt) \\ \times
E\left(\left(h_{n+1}(B^{(q)}_{tq^k};tq^k)-h_{n+1}(B^{(q)}_{tq^{k+1}};tq^{k+1}) \right) \left(h_{m+1}(B^{(q)}_{tq^k};tq^k)-h_{m+1}(B^{(q)}_{tq^{k+1}};tq^{k+1})\right)\right).
\end{multline*}
Noting that
$$ E\left(\left(h_{n+1}(B^{(q)}_{tq^k};tq^k)-h_{n+1}(B^{(q)}_{tq^{k+1}};tq^{k+1}) \right) \left(h_{m+1}(B^{(q)}_{tq^k};tq^k)-h_{m+1}(B^{(q)}_{tq^{k+1}};tq^{k+1})\right)\right)=0$$ for $m\ne n$, we see that
\begin{multline*}
E\left(\left(\int_0^t f(B^{(q)}_s,s)\mathbf{\partial} B^{(q)}\right)^2\right)\\
=\sum_{m=0}^d
\frac{1}{[m+1]!^2}\sum_{k=0}^\infty b_n^2(q^k t)
E\left(\left(h_{m+1}(B^{(q)}_{tq^k};tq^k)-h_{m+1}(B^{(q)}_{tq^{k+1}};tq^{k+1})\right)^2 \right)
\\=
\sum_{m=0}^d
\frac{1}{[m+1]!}\sum_{k=0}^\infty b_n^2(q^k t)
t^{m+1}\left(q^{k(m+1)}- q^{(k+1)(m+1)}\right).
\end{multline*}
Here we used the fact that for $s<t$, by another application of the martingale property and \eqref{EQ:ortho}, we have
$$E\left(\left(h_k(B^{(q)}_{t};t)-h_k(B^{(q)}_{s};s)\right)^2\right)=E \left(h_k^2(B^{(q)}_{t};t)\right)-E\left(h_k^2(B^{(q)}_{s};s)\right)=[k]!(t^k-s^k).$$
This shows that the right hand side of \eqref{EQ:q-Int-1} matches \eqref{EQ:RHS}.
\end{proof}
Suppose that for every $t>0$ the function $f(x,t)$ as a function of $x$ is square integrable with respect to $\gamma_{t;q}(dx)$ so that we have an $L_2(\gamma_{t;q}(dx))$-convergent expansion
\begin{equation}\label{EQ-L2}
f(x,t)=\sum_{n=0}^\infty \frac{b_n(t)}{[n]!}h_n(x;t).
\end{equation}
\begin{corollary}\label{Cor} If each of the coefficients $b_n(t)$ in \eqref{EQ-L2} is bounded in a neighborhood of $0$ and the series
$$\sum_{n=0}^\infty \frac{1}{[n]!}\int_0^t b_n^2(s)s^n d_qs$$ converges, then the sequence of random variables
\begin{equation}
\label{EQ:L2-def}
\int_0^t \sum_{k=0}^n \frac{b_k(s)}{[k]!}h_k(B^{(q)}_s;s)\mathbf{\partial} B^{(q)}
\end{equation}
converges in mean square as $n\to\infty$.
\end{corollary}
When the mean-square limit \eqref{EQ:L2-def} exists, it is natural to denote it as
\begin{equation}\label{EQ:l.i.m}
\int_0^t f(B^{(q)}_s,s)\ddB^{(q)} .
\end{equation}
We remark that the verification of the assumptions of Corollary \ref{Cor} may not be easy. In Section \ref{Sec-SDE} we consider function
$$f(x,t)=\prod_{k=0}^\infty \left(1-(1-q) q^k x +t q^{2k}\right)^{-1}$$
and we rely on explicit expansion \eqref{EQ-L2} to show that the integral \eqref{EQ:l.i.m} is well defined for $t<1/(1-q)$. The following result deals with analytic functions that do not depend on variable $t$.
\begin{proposition}\label{P4.5}
Suppose $f(x)=\sum_{k=0}^\infty a_k x^k$ with infinite radius of convergence. Then $ \int_0^t f(B^{(q)}_s)\ddB^{(q)}$ is well defined for all $t>0$.
\end{proposition}
\begin{proof}
Since the support of $\gamma_{t;q}(dx)$ is compact and $f$ is bounded on compacts, it is clear that we can expand $f(x)$ into \eqref{EQ-L2}.
We will show that the coefficients $b_n(t)$ of the expansion are bounded in a neighborhood of $0$. (The proof shows also that $b_n(t)$ satisfy estimate \eqref{EQ:bound};
the later is not needed here, but this property is assumed in Theorem \ref{Thm2} below.)
As in the proof of Lemma \ref{Lem:1}, we have
\begin{multline*}
b_n(t)=t^{-n}\int f(x) h_n(x;t)\gamma_{t;q}(dx)
= t^{-n/2} \int h_n(y;1) \sum_{k=0}^\infty a_k t^{k/2} y^k\gamma_{1;q}(dy).
\end{multline*}
Since on the support of $\gamma_{1;q}(dy)$ the series in the integrand converges absolutely and is dominated by a constant that does not depend on $y$, we can switch the order of summation and integration so that
\begin{multline*}
b_n(t) = t^{-n/2} \sum_{k=0}^\infty a_k t^{k/2} \int h_n(y;1)y^k\gamma_{1;q}(dy)\\
= t^{-n/2} \sum_{k=n}^\infty a_k t^{k/2} \int h_n(y;1)y^k\gamma_{1;q}(dy),
\end{multline*}
where in the last line we used the orthogonality of $h_n(y;1)$ and $y^k$ for $k<n$.
Choose $R>0$ such that $R^2>4t/(1-q)$. Expressing the coefficients $a_k$ by the contour integrals we get
\begin{multline*}
b_n(t) = \frac{1}{2\pi i} t^{-n/2} \sum_{k=n}^\infty \oint _{|z|=R}\frac{f(z)}{z^{k+1} }dz \, t^{k/2} \int h_n(y;1)y^k\gamma_{1;q}(dy)
\\ =\frac{1}{2\pi i} \oint _{|z|=R} f(z)\sum_{k=n}^\infty \frac{t^{k/2-n/2}}{z^{k+1} } t^{k/2} \int h_n(y;1)y^k\gamma_{1;q}(dy) dz.
\end{multline*}
Since $|y|\leq 2/\sqrt{1-q}$ on the support of $\gamma_{1;q}(dy)$, from \eqref{EQ:ortho} we get
$$
\left| \int h_n(y;1)y^k\gamma_{1;q}(dy)\right|\leq \left(\frac2{\sqrt{1-q}}\right)^k \int |h_n(y;1)|\gamma_{1;q}(dy)\leq \left(\frac2{\sqrt{1-q}}\right)^k \sqrt{[n]!}.
$$
Summing the resulting geometric series under the $dz$-integral, we get
\begin{equation}
\label{b-bound}
|b_n(t)|\leq \max_{|z|=R}|f(z)| \sqrt{[n]!}\left(\frac{2}{R\sqrt{1-q}}\right)^n \frac{R\sqrt{1-q}}{R\sqrt{1-q}-2\sqrt{t}}.
\end{equation}
This shows that $b_n(t)$ is bounded in any neighborhood of $t=0$.
Next, we use \eqref{J-ds-int} and inequality \eqref{b-bound} to prove the convergence of the series.
\begin{multline*}
\sum_{n=0}^\infty \frac{1}{[n]!}\int_0^t b_n^2(s)s^n d_qs
=\sum_{n=0}^\infty \frac{(1-q)t^{n+1}}{[n]!}\sum_{k=0}^n b_n^2(q^{k} t) q^{(n+1)k}
\\
\leq C \sum_{n=0}^\infty \left(\frac{4t}{R^2(1-q)}\right)^n (1-q)\sum_{k=0}^n q^{k(n+1)}
= C \sum_{n=0}^\infty \left(\frac{4t}{R^2(1-q)}\right)^n \frac{1-q}{1-q^{n+1}} \\ \leq C \sum_{n=0}^\infty \left(\frac{4t}{R^2(1-q)}\right)^n <\infty,
\end{multline*}
with constant $C=tR^2(1-q)\max_{|z|=R}|f(z)|^2/(R\sqrt{1-q}-2\sqrt{t})^2 $.
We now apply Corollary \ref{Cor} to infer that $ \int_0^t f(B^{(q)}_s)\ddB^{(q)}$ is well defined. Since $R$ was arbitrary, the conclusion holds for all $t>0$.
\end{proof}
\subsection{The $q$-Ito formula}
Our second main result is the $q$-version of the Ito formula.
\begin{theorem}\label{Thm2}
Let $0<q<1$. Suppose $f(x,t)$ is a polynomial in $x$ with coefficients that satisfy estimate \eqref{EQ:bound} in a neighborhood of $t=0$.
Then \eqref{EQ-ito} holds.
\end{theorem}
\subsection{Proof of Theorem \ref{Thm2}}
By linearity, it is enough to consider a single term in \eqref{EQ:q-Int-2}. We first consider the constant term, $f(x,t)=b(t)$. Then
$\nabla_x f=0$ as the expression $f(x,s)-f(y,s)$ under the integral in the definition \eqref{EQ:nabla} vanishes.
Similarly, $\Delta_x f=0$, as the expression $(z-x)f(y,s)+(x-y)f(z,s)+(y-z)f(x,s) $ in the definition \eqref{EQ:Delta} vanishes.
So \eqref{EQ-ito} in this case reduces to the $q$-integral identity
$b(t)-b(0)=\int_0^t (\mathcal{D}_{q,s} b)(s)d_qs$ which was already recalled in \eqref{EQ:antideriv}.
Now we consider a non-constant term of degree $m+1$. We write \eqref{EQ:q-Int-2} as
\begin{multline}
\label{EQ-one-term-1}
b(t)h_{m+1}(B^{(q)}_t;t)=[m+1]_q\int_0^t b(s) h_m(B^{(q)}_s;s)\mathbf{\partial} B^{(q)} \\ +\int_0^t h_{m+1}(B^{(q)}_{qs};qs)(\mathcal{D}_{q,s} b)(s)d_qs.
\end{multline}
We first note the following identity.
\begin{lemma} \label{L:BW} For a fixed $s>0$, we have
$\nabla_x^{(s)} \left(h_{m+1}(x;s)\right)=[m+1]h_m(x;s)$.
\end{lemma}
\begin{proof}
This is a special case of \cite[Lemma 2.3]{Bryc-Wesolowski-2013-gener}, applied to $\tau=\theta=0$.
\end{proof}
Thus, noting that $\nabla_x^{(s)}$ acts only on variable $x$, we have
\begin{multline}
\label{EQ-one-term-2}
b(t)h_{m+1}(B^{(q)}_t;t) \\= \int_0^t \nabla_x^{(s)} \left(b(s) h_{m+1}(x;s)\right)\Big|_{x=B^{(q)}_s}\mathbf{\partial} B^{(q)}+\int_0^t h_{m+1}(B^{(q)}_{qs};qs)(\mathcal{D}_{q,s} b)(s)d_qs.
\end{multline}
Next, we consider the second term on the right hand side of \eqref{EQ-one-term-1}. Using the $q$-product identity
\begin{equation}
\label{EQ-q-prod}
\mathcal{D}_{q,s}(b(s)h(s))=b(s)\mathcal{D}_{q,s}(h(s))+h(qs)\mathcal{D}_{q,s}(b(s)),
\end{equation}
we write
$$h_{m+1}(x;qs)(\mathcal{D}_{q,s} b)(s)=\mathcal{D}_{q,s}\left(b(s)h_{m+1}(x,s)\right) - b(s)\mathcal{D}_{q,s}(h_{m+1}(x,s)).$$
This recovers the second term in \eqref{EQ-ito}: \eqref{EQ-one-term-2} becomes
\begin{multline}
\label{EQ-one-term-3}
b(t)h_{m+1}(B^{(q)}_t;t)= \int_0^t \nabla_x^{(s)} \left(b(s) h_{m+1}(x;s)\right)\Big|_{x=B^{(q)}_s}\mathbf{\partial} B^{(q)} \\ +\int_0^t
\mathcal{D}_{q,s}\left(b(s) h_{m+1}(x;s)\right)\Big|_{x=B^{(q)}_{qs}}d_qs +\int_0^t D_x^{(s)}\left(b(s)h_{m+1}(x;s)\right) \Big|_{x=B^{(q)}_{qs}}d_qs,
\end{multline}
where $D_x^{(s)}$ is a linear operator on polynomials in variable $x$, whose action on monomials $x^n=\sum_{j=0}^n a_j(s)h_j(x;s)$ is defined by
$$
D_x^{(s)}(x^n)= - \sum_{j=0}^n a_j(s)\mathcal{D}_{q,s}\left(h_j(x;s)\right).
$$
It remains to identify $D_x^{(s)}$ with $\Delta_x^{(s)}$. To do so we use the approach from \cite{Bryc-Wesolowski-2013-gener}. We fix $s>0$ and introduce an auxiliary linear operator $A$ that acts on polynomials $p(x)$ in variable $x$ by the formula
\begin{equation}
\label{D2P}
A(p(x))=D_x^{(s)}\left(x p(x)\right)-x D_x^{(s)}\left(p(x)\right).
\end{equation}
\begin{lemma}\label{L:B}
\begin{equation}
\label{EQ:B2} A\left(h_{m}(x;s)\right)=[m]_qh_{m-1}(x;qs).
\end{equation}
\end{lemma}
\begin{proof} By definition, $A\left(h_{m}(x;s)\right)=D_x^{(s)}\left(x h_{m}(x;s)\right)-x D_x^{(s)}\left(h_{m}(x;s) \right)$, and we evaluate each term separately.
From recursion \eqref{EQ:rec-q-hermite} we get
\begin{equation}\label{EQ:Dx}
D_x^{(s)}\left(x h_{m}(x;s)\right)
=-\mathcal{D}_{q,s}\left(h_{m+1}(x;s)\right)-s[m]_q\mathcal{D}_{q,s}\left(h_{m-1}(x;s)\right).
\end{equation}
Next, we have
\begin{multline*}
x D_x^{(s)}\left(h_{m}(x;s)\right)=-x \mathcal{D}_{q,s}\left(h_{m}(x;s)\right)=- \mathcal{D}_{q,s}\left(xh_{m}(x;s)\right) \\ =
- \mathcal{D}_{q,s}\left(h_{m+1}(x;s)+s [m] h_{m-1}(x;s) \right)\\=
- \mathcal{D}_{q,s}\left(h_{m+1}(x;s)\right)- [m]_q \mathcal{D}_{q,s}\left(s h_{m-1}(x;s) \right).
\end{multline*}
Since $\mathcal{D}_{q,s}(s)=1$, applying \eqref{EQ-q-prod} we get
$$xD_x^{(s)}\left(h_{m}(x;s)\right)
=- \mathcal{D}_{q,s}\left(h_{m+1}(x;s)\right) -s[m]_q \mathcal{D}_{q,s}\left(h_{m-1}(x;s)\right) - [m]_q h_{m-1}(x;qs).$$
To end the proof, we subtract this from \eqref{EQ:Dx}.
\end{proof}
Using martingale property \eqref{EQ:mart} and Lemma \ref{L:BW}, we rewrite \eqref{EQ:B2} as
\begin{multline*}
A\left(h_{m}(x;s)\right)=\int [m]_q h_{m-1}(y;s) P_{qs,s}(x,dy)\\ =\iint \frac{h_{m}(z;s)-h_m(y;s)}{z-y}\nu_{s,y}(dz)P_{qs,s}(x,dy).
\end{multline*}
So by linearity, for any polynomial $p$ in variable $x$ we have
$$
(A p)(x)=\iint \frac{p(z)-p(y)}{z-y}\mu_{x,s}(dy,dz).
$$
Since \eqref{D2P} gives $D_x^{(s)}(x^n)=A(x^{n-1})+x D_x^{(s)}(x^{n-1})$ and $D_x^{(s)}(x)=-\mathcal{D}_{q,s}(h_1(x;s))=0$ this determines $D_x^{(s)}$ on monomials:
$$
D_x^{(s)}(x^n)=\sum_{k=0}^{n-1}x^k A(x^{n-k-1})=\iint \left(\sum_{k=0}^{n-1} x^k\frac{z^{n-k-1}-y^{n-k-1}}{z-y}\right)\mu_{x,s}(dy,dz).
$$
By the geometric sum formula, the integrand is
$$\sum_{k=0}^{n-1} x^k\frac{z^{n-k-1}-y^{n-k-1}}{z-y}=\frac{ (x-z)y^n+(y-x)z^n + (z-y)x^n}{(z-x)(x-y) (y-z)}.$$
This shows that $ D_x^{(s)}=\Delta_x^{(s)}$ on polynomials. With this identification, \eqref{EQ-one-term-3} concludes the proof of Theorem \ref{Thm2}.
\begin{remark}\label{Rem-Ito} Operators $\nabla_x^{(s)}$ and $\Delta_x^{(s)}$ are well defined on functions with bounded second derivatives in variable $x$, so we expect that \eqref{EQ-ito} holds in more generality.
For an analytic function $f(x)$ one can use contour integration to verify that each term in \eqref{EQ-ito} can be approximated uniformly on compacts by the same expression applied to a polynomial.
Thus formula \eqref{EQ-ito} for polynomials yields
$$
f(B^{(q)}_t)=f(0)+\int_0^t (\nabla_x^{(s)} f)(B^{(q)}_s,s)\ddB^{(q)}+\int_0^t (\Delta_x^{(s)} f)(B^{(q)}_{qs},s) d_q s.
$$
\end{remark}
\arxiv{
\begin{proof}[Proof of Remark \ref{Rem-Ito}]
Using Cauchy formula with $R$ large enough, we have
\begin{equation}\label{CF1}
g(x,t):=(\nabla_x^{(t)}f)(x,t)=\frac{1}{2\pi i} \oint_{|\zeta|=R} \frac{f(\zeta)}{(\zeta-x)(\zeta-y)}d\zeta d\nu_{x,t}(dy).
\end{equation}
With $p_n(x)=\sum_{k=0}^n a_k x^k\to f(x)$, we see that
\begin{multline*}
g(x,t)=\frac{1}{2\pi i} \oint_{|\zeta|=R} \frac{p_n(\zeta)}{(\zeta-x)(\zeta-y)}d\zeta d\nu_{x,t}(dy)\\+\frac{1}{2\pi i} \oint_{|\zeta|=R} \frac{f(\zeta)-p_n(\zeta)}{(\zeta-x)(\zeta-y)}d\zeta d\nu_{x,t}(dy)
\end{multline*}
can be approximated uniformly in $x$ from the support of $\gamma_{t;q}$ by polynomials. In particular, expanding $p_n(x)$ into the $q$-Hermite polynomials, from Lemma \ref{L:BW} we deduce that
$$
g(x,t)=\sum_{k=0}^\infty \frac{b_n(t)}{[n]!}\nabla_x^{(t)}(h_n)(x,t) =
\sum_{k=0}^\infty \frac{b_{n+1}(t)}{[n]!} h_n(x;t)
$$
and the convergence is uniform over $x$ from the support of $\gamma_t$.
The estimates from proof of Proposition \ref{P4.5} show that $\int_0^t (\nabla_x^{(s)} f)(B^{(q)}_s,s)\ddB^{(q)}$ is well defined.
Next we note that
$$(\Delta_x^{(s)}f)(x,t)=\frac{1}{2\pi i} \oint_{|\zeta|=R} \frac{f(\zeta)}{(\zeta-x)(\zeta-y)(\zeta-z)}d\zeta d\mu_{x,s}(dy,dz),$$
so $(\Delta_x^{(s)}f)(x,t)$ is also a limit of $(\Delta_x^{(s)}p_n)(x,t)$, and the convergence is uniform over $x$ from the support of $\gamma_t$ and over all $ s\in[0,t]$.
By Theorem \ref{Thm2}, formula \eqref{EQ-ito} is valid for polynomials $p_n(x)$ and the contour integral identities show that we can pass to the limit as $n\to\infty$.
\end{proof}
}
\section{A $q$-analogue of the stochastic exponential} \label{Sec-SDE}
In this section we consider a more general function $f(x,t)$ for which the mean-square limit \eqref{EQ:l.i.m} is applicable.
Our goal is to exhibit a solution to the ``differential equation" $\mathbf{\partial} Z=a Z \ddB^{(q)}$ which we interpret in integral form as
\begin{equation}
\label{ODE}
Z_t=c+ a \int_0^t Z_s\mathbf{\partial} B^{(q)}
\end{equation}
with deterministic parameters $a,c\in\mathbb{R}$. In view of the fact that we can integrate only instantaneous functions of $B^{(q)}_t$, we seek a solution in this form, and we seek the series expansion.
The solution parallels the development in Ito theory and relies on the identity \eqref{WdW}
which is a special case of our definition \eqref{EQ:q-Int-2}.
Thus, as a mean-square expansion the solution of \eqref{ODE} is
\begin{equation}
\label{EQ:Z-series}
Z_t=c \sum_{n=0}^\infty \frac{a^n h_n(B^{(q)}_t;t)}{[n]_q!}=c\prod_{k=0}^\infty \left(1-(1-q)a q^k B^{(q)}_t +a^2t q^{2k}\right)^{-1}.
\end{equation}
Formal calculations give
$$c+a\int_0^t Z_s \ddB^{(q)}= c+c a \sum_{n=0}^\infty \frac{a^n h_{n+1}(B^{(q)}_t;t)}{[n+1]_q!}=Z_t,$$
as $h_0(x;t)=1$.
Note that the product expression on the right hand side of \eqref{EQ:Z-series} is well defined for all $t$, as with probability one $|B^{(q)}_t|< 2\sqrt{t}/\sqrt{1-q}$.
However, the solution of \eqref{ODE} ``lives" only on the finite interval $0\leq t < a^{-2}(1-q)^{-1}$.
This is because for $0<q<1$ the product $\prod_{j=1}^{\infty}(1-q^j)$ converges, so the series
$$
\sum_n \frac{a^{2n}}{[n]!} \int_0^t s^n d_qs=\sum_n \frac{a^{2n}t^{n+1}}{[n+1]!}=\sum_n \frac{(1-q)^{n}a^{2n}t^{n+1}}{\prod_{j=2}^{n+1}(1-q^j)}
$$
converges when $(1-q)a^2t<1$ and hence Corollary \ref{Cor} can be applied only to this range of $t$. This is the same range of $t$ where $(Z_t)$ is a martingale \cite[Corollary 4]{szabl2012q}.
\arxiv{
\section{Deterministic integrands}\label{Sec-Determ} In the Ito-Wiener theory, the distribution of the integral of a deterministic function $b(s)$ is Gaussian with the variance $\int_0^t b^2(s)ds$.
In this section we show that for $0<q<1$ the distribution of the random variable $\int_0^1 b(s)\mathbf{\partial} B^{(q)}$ depends on $b(s)$ and $q$ in a more complicated way.
This gives rise to a plethora of $q$-Gaussian laws that all converge to the Gaussian law as $q\to 1$.
To illustrate this phenomenon, we consider
$$Z=\int_0^1 s^r \mathbf{\partial} B^{(q)}=\sum_{k=0}^\infty q^{kr}\left(B^{(q)}_{q^k}-B^{(q)}_{q^{k+1}}\right)
$$ for $r\geq 0$. Then $E(Z)=0$ and from \eqref{EQ-iso} we see that
\begin{equation}
\label{EZ^2}
E(Z^2)=\int_0^1 s^{2r}d_qs=\frac{1-q}{1-q^{2r+1}}=\frac{1}{[2r+1]_q}.
\end{equation}
To show that the distribution of $Z$ depends on $r$ and $q$ in a complicated way, we compute the fourth moment. Recall that the $4$-th moment of $\gamma_{t;q}$ is $t^2(2+q)$, where $t$ is the variance. We will show that this pattern fails for $Z$.
}
\arxiv{
Noting that by martingale property $E(B^{(q)}_{u}-B^{(q)}_t|B^{(q)}_s)=0$ for $s<t<u$, to compute $E(Z^4)$ we need only the formulas
\begin{equation}
\label{EQ:DW^4}
E\left((B^{(q)}_{t}-B^{(q)}_{s})^4\right)=(t-s) ((q+2) t-3 q s)
\end{equation}
for $t<s$ and
\begin{equation}
\label{EQ:DW2DW2}
E\left((B^{(q)}_{t_2}-B^{(q)}_{t_1})^2(B^{(q)}_{u_2}-B^{(q)}_{u_1})^2\right)=(u_2-u_1)(t_2-t_1)
\end{equation}
\begin{equation}
\label{EQ:DW1DW3}
E\left((B^{(q)}_{t_2}-B^{(q)}_{t_1})(B^{(q)}_{u_2}-B^{(q)}_{u_1})^3\right)=-(1-q)(u_2-u_1)(t_2-t_1)
\end{equation}
for $t_1<t_2\leq u_1<u_2$.
}
\arxiv{
The calculations use the first four martingale polynomials: the first three were already listed in Example \ref{Ex:4.2} and the fourth one is
$h_4(x;t)=x^4-\left(q^2+2
q+3\right) t x^2 +\left(q^2+q+1\right) t^2$.
We use them to compute recurrently conditional moments. This gives martingale property that we already used, the quadratic martingale property $E((B^{(q)}_t)^2|B^{(q)}_s)=(B^{(q)}_s)^2+t-s$, and
a bit more cumbersome formulas
\begin{equation}
\label{m[3]}
E((B^{(q)}_t)^3|B^{(q)}_s)=(B^{(q)}_s)^3+(t-s)(2+q)B^{(q)}_s
\end{equation}
\begin{equation}
\label{m[4]}
E((B^{(q)}_t)^4|B^{(q)}_s)=(B^{(q)}_s)^4+(t-s)(3+2q+q^2)(B^{(q)}_s)^2+(t-s)((2+q)t-(1+q+q^2)s)
\end{equation}
After a calculation these formulas give
$$E\big((B^{(q)}_t-B^{(q)}_s)^4\big|B^{(q)}_s\big)= (t-s)\left((1-q)^2 (B^{(q)}_s)^2+(q+2) t- \left(q^2+q+1\right) s\right).$$ To derive \eqref{EQ:DW^4} we take the expected value of this expression.
To derive \eqref{EQ:DW2DW2}, we use martingale property and quadratic martingale property to compute
$$E\big((B^{(q)}_{u_2}-B^{(q)}_{u_1})^2\big|B^{(q)}_{t_2}\big)= E\big((B^{(q)}_{u_2}-B^{(q)}_{t_2})^2\big|B^{(q)}_{t_2}\big)-E\big((B^{(q)}_{u_1}-B^{(q)}_{t_2})^2\big|B^{(q)}_{t_2}\big)=u_2-u_1.$$
To derive \eqref{EQ:DW1DW3} we use again the formulas for conditional moments to compute $E\big((B^{(q)}_{u_2}-B^{(q)}_{u_1})^3\big|B^{(q)}_{u_1}\big)=-(1 - q) (u_2 - u_1) B^{(q)}_{u_1}$. This gives
$$E\left((B^{(q)}_{t_2}-B^{(q)}_{t_1})(B^{(q)}_{u_2}-B^{(q)}_{u_1})^3\right)=-(1-q)(u_2-u_1)\left(E(B^{(q)}_{t_2}B^{(q)}_{u_1})- E(B^{(q)}_{t_1}B^{(q)}_{u_1})\right).$$
}
\arxiv{We now compute $E(Z^4)$. Denoting $\xi_k=B^{(q)}_{q^{k}}-B^{(q)}_{q^{k+1}}$, we see that
$$E(Z^4)= \sum_{k=0}^\infty q^{4rk} E\xi_k^4+ 6 \sum_{k=1}^\infty\sum_{j=0}^{k-1} q^{2r(j+k)}E(\xi_j^2\xi_k^2)+
4 \sum_{k=1}^\infty\sum_{j=0}^{k-1} q^{r(3j+ k)}E(\xi_k\xi_j^3),
$$
as all other terms vanish.
From \eqref{EQ:DW^4} we see that $E((B^{(q)}_{t q^{k}}-B^{(q)}_{t q^{k+1}})^4)=(1-q)^2 t^2 q^{2 k}(2+3q)$, so
$$
\sum_{k=0}^\infty q^{4rk} E\xi_k^4=\frac{(1-q)^2 (2+3 q)}{1-q^{4 r+2}}.
$$
From \eqref{EQ:DW2DW2} we get
$$
\sum_{k=1}^\infty\sum_{j=0}^{k-1} q^{2r(j+k)}E(\xi_j^2\xi_k^2)=(1-q)^2\sum_{k=1}^\infty\sum_{j=0}^{k-1} q^{(2r+1)(j+k)} =\frac{(1-q)^2 q^{2 r+1}}{\left(1-q^{2 r+1}\right)^2
\left(1+q^{2 r+1}\right)}.
$$
From \eqref{EQ:DW1DW3} we get
$$
\sum_{k=1}^\infty\sum_{j=0}^{k-1} q^{r(3j+ k)}E(\xi_k\xi_j^3)= -(1-q)^3 \sum_{j=0}^{k-1} q^{(3r+1)j+(r+1)k} =-\frac{(1-q)^3 q^{r+1}}{\left(1-q^{r+1}\right) \left(1-q^{4
r+2}\right)}.
$$
This gives
$$E(Z^4)=\frac{(1-q)^2 \left( 2+3 q -6 q^{r+1}+q^{r+2}+4 q^{2 r+1}-3 q^{2 r+2}-q^{3
r+3}\right)}{\left(1-q^{r+1}\right) \left(1-q^{2
r+1}\right)^2 \left(1+q^{2 r+1}\right)}.$$
So using \eqref{EZ^2} we get
$$\frac{E(Z^4)}{(E(Z^2))^2}=\frac{2+3q-6 q^{r+1}+q^{r+2}+4 q^{2 r+1}-3 q^{2 r+2}-q^{3 r+3} }{\left(1-q^{r+1}\right) \left(1+q^{2 r+1}\right)}.$$
}
\subsection*{Acknowledgements}
I would like to thank Magda Peligrad for positive feedback and encouragement. Referee's comments helped to improve the presentation.
\bibliographystyle{plain}
|
\section{Introduction}
The \acl{sze} \cite[\acsu{sze}]{sunyaev70,sunyaev72}, is a spectral distortion of the \ac{cmb} arising from interactions between CMB photons and hot, ionised gas. Surveys of galaxy clusters using the \ac{sze} have opened a new window on the Universe by providing samples of hundreds of massive galaxy clusters with well-understood selection over a broad redshift range. Both space- and ground-based instruments, including the {\it Planck}\ satellite \citep{tauber10}, the \acl{spt} \acused{spt}\citep[\acs{spt};][]{carlstrom11}, and the Atacama Cosmology Telescope \citep[ACT;][]{fowler07}, have released catalogs of their \ac{sze} selected clusters. The cluster samples have provided new cosmological constraints \cite[]{reichardt13, hasselfield13,planck13-20} and have enabled important evolution studies of cluster galaxies and the intracluster medium over a broad range of redshift \cite[e.g.,][]{zenteno11, semler12, mcdonald13}.
Understanding the relationship between the \ac{sze} observable and cluster mass is important for both cosmological applications and astrophysical studies. Among observables, the integrated Comptonization from the \ac{sze} has been shown by numerical simulations \cite[]{motl05,nagai07} to be a good mass proxy with low intrinsic scatter. Cluster mass estimates derived from X-ray observations of \ac{sze} selected clusters have largely confirmed this expectation \cite[]{andersson11,planckearlyXI}. A related quantity, the \ac{spt} signal-to-noise $\xi$, is linked to the underlying virial mass of the cluster by a power law with log-normal scatter at the $\sim 20$~per cent level \cite[][hereafter B13]{benson13}.
Probing the \ac{sze} signature of low mass clusters and groups is also important, although it is much more challenging with the current generation of experiments. These low mass clusters and groups are far more numerous and are presumably important environments for the transformation of galaxies from the field to the cluster. Studies of their baryonic content show that low mass clusters and groups are not simply scaled-down versions of the more massive clusters \cite[e.g.,][]{mohr99, sun09, lagana13}. This breaking of self-similarity in moving from the cluster to the group mass scale is likely due to processes such as star formation and \ac{agn} feedback.
The {\it Planck}\ team has recently studied this low mass population by stacking the {\it Planck}\ maps around samples of X-ray selected clusters in the nearby universe \cite[][hereafter P11]{planck11-10}. They show that the \ac{sze} signal is consistent with the self-similar scaling relation based on the X-ray luminosity over a mass range spanning 1.4 orders of magnitude.
Here we pursue a study of the \ac{sze} signatures of low mass clusters extending over a broad range of redshift. We use the \ac{sptsz} data with the \ac{xbcs} over $6~\sdeg$ from which a sample of 46 X-ray groups and clusters have been selected \cite[][hereafter S12]{suhada12}. The \ac{sptsz} data enable us to extract cluster \ac{sze} signal with high angular resolution and low instrument noise, making the most of this small sample.
The paper is organised as follows. In \Fref{sec:xbcs-data-description}, we describe the data used from the \ac{xbcs} and the extraction of the \ac{sze} signature from the \ac{sptsz} maps. In \Fref{sec:xbcs-method}, we introduce the calibration method for the mass-observable scaling relation, and we apply it to the cluster sample in \Fref{sec:xbcs-result}. We also discuss possible systematic effects and present a discussion of the point source population associated with our sample. We conclude in \Fref{sec:xbcs-conclusions} with a prediction of the improvement based on future surveys.
The cosmological model parameters adopted in this paper are the same as the ones used for the X-ray measurement from the \ac{xbcs} project (S12): $(\Omega_{\mathrm M}, \Omega_{\Lambda},H_{0})=(0.3,0.7,70~\text{km s}^{-1}\text{Mpc}^{-1})$. The amplitude of the matter power spectrum, which is needed to estimate bias corrections in the analysis, is fixed to $\sigma_{8}=0.8$.
\section{Data Description and Observables}\label{sec:xbcs-data-description}
In this analysis, we adopt an X-ray selected sample of clusters, described in \Fref{sec:xbcs-xmmbcs}, together with published $L_{\text X}$-mass scaling relations to examine the corresponding \acsu{sptsz} significance- and $Y_\text{500}$-mass relations. The \ac{sptsz} observable $\xi$ is measured by a matched filter approach, which we discuss in \Fref{sec:xbcs-spt} and \Fref{sec:xbcs-spt-xi}. The estimation of $Y_\text{500}$ is described in \Fref{sec:xbcs-yintegrated}.
\subsection{X-ray Catalog}\label{sec:xbcs-xmmbcs}
The \ac{xbcs} project consists of an X-ray survey mapping $14~\sdeg$ area of the southern hemisphere sky that overlaps the {\it griz} bands \acl{bcs} \cite[\acsu{bcs}]{desai12} and the mm-wavelength \ac{sptsz} survey \cite[]{carlstrom11}. S12 analyse the initial $6~\sdeg$ core area, construct a catalog of 46 galaxy clusters and present a simple selection function. Here we present a brief summary of the characteristics of that sample. The cluster physical parameters from Table 2 (S12) are repeated in \Fref{tab:xbcs-cat} with the same IDs.
The initial cluster sample was selected via a source detection pipeline in the 0.5--2~keV band. The spatial extent of the clusters leads to the need to have more counts to reach a certain detection threshold than are needed for point sources. S12 modelled the extended source sensitivity as an offset from the point source limit; the cluster sample is approximately a flux-limited sample with $f_\text{min}=1\times10^{-14} \text{erg s}^{-1}\text{cm}^{-2}$.
The X-ray luminosity $L_{\text X}$ was measured in the detection band (0.5\ -\ 2.0~keV) within a radius of $R_\text{500c}$, which is iteratively determined using mass estimates from the $L_{\text X}$-mass relation and is defined such that the interior density is 500~times the critical density of the Universe at the corresponding redshift. This luminosity was converted to a bolometric luminosity and to a 0.1\ -\ 2.4~keV band luminosity using the characteristic temperature for a cluster with this 0.5\ -\ 2.0~keV luminosity and redshift (see equation~3 in S12). The core radius, $R_\text{c}$, of the beta model is calculated using (see equation~1 in S12):
\begin{equation}
R_\text{c} = 0.07\times R_{500} \Big(\frac{T}{\text{1~keV}}\Big)^{0.63},
\end{equation}
where $T$ is X-ray temperature determined through the $L_{\text X}-T$ relation.
The redshifts of the sample are primarily photometric redshifts extracted using the \ac{bcs} optical imaging data. The optical data and their processing and calibration are described in detail elsewhere \cite[]{desai12}. The photometric redshift estimator has been demonstrated on clusters with spectroscopic redshifts and on simulations \cite[]{song12a} and has been used for redshift estimation within the \ac{sptsz} collaboration \cite[]{song12b}. The typical photometric redshift uncertainty in this \ac{xbcs} sample is $\avg{\Delta z/(1+z)}=0.023$, which is determined using a subsample of 12 clusters ($z<0.4$) with spectroscopic redshifts. This value is consistent with the uncertainty $\avg{\Delta z/(1+z)}=0.017$ we obtained on the more massive main sample \ac{sptsz} clusters.
The X-ray luminosities and photometric redshifts of the sample are shown in \Fref{fig:xbcs-lxza} in black squares and the approximate flux limit of the sample is shown as a red curve. For comparison, we also include a high mass \ac{sptsz} cluster sample (blue triangles) with published X-ray properties \cite[]{andersson11}.
In the analysis that follows we use the X-ray luminosity as the primary mass estimator for each cluster. We adopt the $L_{\text X}$-mass scaling relation used in S12, which is based on the hydrostatic mass measurements in an ensemble of 31 nearby clusters observed with {\it XMM-Newton} \cite[{REXCESS},][]{pratt09}:
\begin{equation} \label{eq:xbcs-lm}
L_{\text X}=L_{0}\Big(\frac{M_\text{500c}}{2\times10^{14}\msun}
\Big)^{\alpha_\text{LM}}E(z)^{7/3},
\end{equation}
where $H(z)=H_0 E(z)$. The intrinsic scatter in $L_{\text X}$ at fixed mass is modelled as lognormal distributions with widths $\sigma_{L_{\text X}}$, and the observational scatter is given in S12.
This scaling relation includes corrections for Malmquist and Eddington biases. Both biases are affected by the intrinsic scatter and the skewness of the underlying sample distribution. In general, the bias on the true mass is $\Delta \ln M\ \propto\ \gamma \sigma_{\ln M}^2$, where $\mr{d}n(M)/\mr{d}\ln M \propto M^\gamma$ is the slope of the mass distribution and $\sigma_{\ln M}$ is the scatter in mass at fixed observable \citep[for more discussion, we refer the reader to][]{stanek06,vikhlinin09b,mortonson11}. Typically $\gamma$ is negative, and the result is that mass inferred from an observable must be corrected to a lower value than that suggested by naive application of the scaling relation.
The scaling relation parameters for different X-ray bands are listed in \Fref{tab:xbcs-lxm}. We find the choice of luminosity bands has negligible impact on the parameter estimation given the current constraint precision. In addition, we investigate using the $L_{\text X}$-mass scaling relations from {\it Chandra} observations \citep{vikhlinin09b,mantz10b}. These studies draw upon higher mass
cluster samples than the {REXCESS} sample, and therefore we adopt the \citet{pratt09} relation for our primary analysis. We discuss the impact of changing the $L_{\text X}$-mass scaling relation in \Fref{sec:xbcs-xim}.
\begin{table} \begin{center} \caption{$L_{\text X}$-mass relations with
different luminosity bands
(\Fref{eq:xbcs-lm}).} \label{tab:xbcs-lxm}
\begin{tabular}[h]{|c|c|c|c|} \hline Type &
$L_{0}[10^{44}\text{erg\ s}^{-1}]$& $\alpha_\text{LM}$ & $\sigma_{\ln L_{\text X}}$\\ \hline
0.5--2.0 keV & $0.48\pm0.04$ & $1.83\pm0.14$ & $0.412\pm0.071$ \\
0.1--2.4 keV & $0.78\pm0.07$ & $1.83\pm0.14$ & $0.414\pm0.071$ \\
Bolometric & $1.38\pm0.12$ & $2.08\pm0.13$ &
$0.383\pm0.061$ \\ \hline \end{tabular}
\end{center} \end{table}
\begin{figure}
\includegraphics[width=3.5in]{xbcs_andersson_lx_z.pdf}
\caption[The luminosity-redshift distribution of the XMM-BCS
clusters from \protect\citet{suhada12} and the \ac{sptsz} clusters from
\protect\citet{andersson11}]{The luminosity-redshift distribution of the \ac{xbcs}
clusters from S12 (black dots) and the \ac{sptsz} clusters from
\protect\citet[blue triangles]{andersson11}. The X-ray sample is
selected with a flux cut that varies somewhat across the field.
The red line is the corresponding luminosity sensitivity determined by the median flux limit in the 0.5 - 2.0~keV band. The \ac{sptsz} sample is more massive and approximately mass limited.}
\label{fig:xbcs-lxza}
\end{figure}
\subsection{SPT Observations} \label{sec:xbcs-spt}
The \ac{spt} \citep{carlstrom11} is a 10-metre diameter, millimetre-wavelength, wide field telescope that was deployed in 2007 and has been used since then to make arcminute-resolution observations of the CMB over large areas of the sky. The high angular resolution is crucial to detecting the \ac{sze} signal from high-redshift clusters. The \ac{sptsz} survey \citep[e.g.,][]{story13}, completed in 2011, covers a 2500 deg$^2$ region of contiguous sky area in three bands -- centred at 95, 150, and 220 GHz -- at a typical noise level of $< 18 \mu$K per one-arcminute pixel in the 150 GHz band.
The details of the \ac{sptsz} observation strategy, data processing and mapmaking are documented in \cite{schaffer11}; we briefly summarise them here. The \ac{sptsz} survey data were taken primarily in a raster pattern with azimuth scans at discrete elevation steps. A high-pass filter was applied to the time-ordered data to remove low-frequency atmospheric and instrumental noise. The beams, or angular response functions, were measured using observations of planets and bright \acp{agn} in the field. The main lobe of the beam for a field observation is well-approximated as a Gaussian with a \ac{fwhm} of 1.6, 1.2, and 1.0~arcmin at 95, 150, and 220~GHz, respectively. The final temperature map was calibrated by the Galactic H\,{\sevensize\sc II} regions RCW38 and MAT5a \cite[c.f.][]{vanderlinde10}. The \ac{sptsz} maps used in this work are from a 100 deg$^2$ field centred at $(\alpha, \delta)$ = $(23^\circ\ 30', -55^\circ)$ and consist of observations from the 2008 and 2010 \ac{sptsz} observing seasons. The characteristic depths are 37, 12 and $35~\mu\text{K-arcmin}$ at 95, 150 and 220~GHz, respectively.
\subsection{SPT-SZ Cluster Significance}\label{sec:xbcs-spt-xi}
The process of determining the \ac{sptsz} significance for our X-ray sample is very similar to the process of finding clusters in \ac{sptsz} maps, but there are certain key differences, which we highlight below. Clusters of galaxies are extracted from \ac{sptsz} maps through their distinct angular scale- and frequency-dependent imprint on the \ac{cmb}. We adopt the multi-frequency matched filter approach \cite[]{melin06} to extract the cluster signal. The matched filter is designed to maximise the given signal profile while suppressing all noise sources. A detailed description appears elsewhere \citep{vanderlinde10, williamson11}. Here we provide a summary. The \ac{sze} introduces a spectral distortion of the \ac{cmb} at given
frequency $\nu$ as:
\begin{equation}
\label{eq:xbcs-cmby}
\Delta T_\text{CMB}(\bm{\theta},\nu) = y(\bm{\theta}) g(\nu) T_\text{CMB},
\end{equation}
where $g(\nu)$ is the frequency dependency and the Compton-y parameter $y(\bm{\theta})$ is the \ac{sze} signature at direction $\bm{\theta}$, which is linearly related to the integrated pressure along the line-of-sight. To model the \ac{sze} signal $y(\bm{\theta})$, two common templates are adopted: the circular $\beta$ model \cite[]{cavaliere76} and the Arnaud profile \cite[]{arnaud10}. The cluster profiles are convolved with the \ac{spt} beams to get the expected signal profiles. The map noise assumed in constructing the filter includes the measured instrumental and atmospheric noise and sources of astrophysical noise, including the primary \ac{cmb}. Point sources are identified in a similar manner within each band independently, using only the instrument beams as the source profile \cite[][]{vieira10}.
Once \ac{sptsz} maps have been convolved with the multi-frequency matched filter, clusters are extracted with a simple peak-finding algorithm, with the primary observable $\xi$ defined as the maximum signal-to-noise of a given peak across a range of filter scales. The \ac{sptsz} significance $\xi$ is a biased estimator that links to the underlying $\zeta$ as $\avg{\xi} = \sqrt{\zeta^2+3}$, because it is the maximum value identified through a search in sky position and filter angular scale \citep{vanderlinde10}. The observational scatter of $\xi$ around $\zeta$ is a unit-width Gaussian distribution corresponding to the underlying RMS noise of the \ac{sptsz} filtered maps.
In this work, we use the 90~GHz and 150 GHz maps and employ the method described above to define an \ac{sptsz} significance for each X-ray selected cluster, but with two important differences: 1) We measure the \ac{sptsz} significance at the X-ray location, and 2) we use a cluster profile shape informed from the X-ray data. We define this \ac{sptsz} significance as $\xi_{\text X}$, which is related to the unbiased \ac{sptsz} significance $\zeta$ as:
\begin{equation}\label{eq:xizeta}
\zeta=\avg{\xi_{\text X}},
\end{equation}
where the angle brackets denote the average over many realizations of the experiment. The observational scatter of $\xi_{\text X}$ around $\zeta$ is also a unit-width Gaussian distribution. Therefore $\xi_{\text X}$
is an unbiased estimator of $\zeta$, under the assumption that the true X-ray position and profile are identical to the true \ac{sze} position and profile -- a reasonable assumption -- given that both the X-ray and the \ac{sze} signatures are reflecting the intracluster medium properties of the clusters. Note, however, that in the midst of a major merger the different density weighting of the X-ray and SZE signatures can lead to offsets \citep[][]{molnar12}.
We model the relationship between $\zeta$ and the cluster mass through
\begin{equation}\label{eq:zetam}
\zeta = A_{\text{SZ}}^\text{SPT}\Big(\frac{M_\text{500c}}{4.3\times10^{14} \msun} \Big)^{B_{\text{SZ}}} \Big[ \frac{E(z)}{E(0.6)}\Big]^{C_{\text{SZ}}},
\end{equation}
where the intrinsic scatter on $\zeta$ is described by a log-normal distribution of width $D_{\text{SZ}}$
\cite[B13;][]{reichardt13}. We use $A_{\text{SZ}}^\text{SPT}$ to denote the amplitude of the original \ac{sptsz} scaling relation. The differences in the depths of the \ac{sptsz} fields results in a re-scaling of the \ac{sptsz} cluster significance in spatially filtered maps. For the field we study here, the relation requires a factor of 1.38 larger normalisation compared to the value in \cite{reichardt13}.
For the massive \ac{sptsz} clusters (with $\xi>4.5$), the $\zeta$-mass relation is best parametrized as shown in \Fref{tab:xbcs-constraint} with $C_{\text{SZ}}=0.83\pm 0.30$ and $D_{\text{SZ}}=0.21\pm0.09$ (B13). In our analysis, we examine the characteristics of the lower mass clusters within the \ac{sptsz} survey. To avoid a degeneracy between the scaling relation amplitude and slope, we shift the pivot mass to $1.5\times 10^{14}~\msun$, near the median mass of our sample and term the associated amplitude $A_{\text{SZ}}$. At this pivot mass, with the normalisation factor mentioned previously, the equivalent amplitude parameter for the main \ac{sptsz} sample corresponds to $A_{\text{SZ}}=1.50$. In \Fref{tab:xbcs-constraint} we also note the priors we adopt in our analysis of the low mass sample. For our primary analysis we adopt flat priors on the amplitude and slope parameters and fix the redshift evolution and scatter at the values obtained by B13.
\subsection{Integrated \texorpdfstring{$Y_\text{500}$}{Y500}}\label{sec:xbcs-yintegrated}
To facilitate the comparison of our sample with cluster physical properties reported in the literature, we also convert the $\xi_{\text X}$ to $Y_\text{500}$, which is the integration of the Compton-y parameter within a spherical volume with radius $R_\text{500c}$. The central $y_{0}$ is
linearly linked to $\xi_{\text X}$ in the matched filter approach \citep{melin06}, with the corresponding Arnaud profile or $\beta$
profile as the cluster template. The characteristic radii ($R_\text{500c}$ and $R_\text{c}$) are based on the X-ray measurements (S12), because the \ac{sze} observations are too noisy to constrain the profile accurately.
The projected circular $\beta$ profile for the filter is:
\begin{equation}
\label{eq:betaprofile}
y^\mr{(\beta)}_\text{cyl}(r) \propto (1-r^2/R_\text{c}^2)^{-(3\beta-1)/2},
\end{equation}
where $\beta$ is fixed to 1, consistent with higher signal to noise cluster studies \citep{plagge10}. And the spherical $Y_\text{500}$ within the $R_\text{500c}$ is
\begin{equation}
\label{eq:betay500}
Y_\text{500}^\mr{(\beta)} = y_0 \times \pi R_\text{c}^{2} \ln(1+R_\text{500c}^{2}/R_\text{c}^{2}) \times f(R_\text{500c}/R_\text{c}),
\end{equation}
where $f(x)$ corrects the cylindrical result to the spherical value for the $\beta$ profile as:
\begin{equation}
\label{eq:beta-deproject}
f(x) = 2\frac{\ln(x+\sqrt{1+x^2})-x/\sqrt{1+x^2}}{\ln(1+x^2)}.
\end{equation}
The $Y_\text{500}^\text{(A)}$ for the Arnaud profile is calculated similarly except that the projected profile is calculated numerically within $5R_\text{500c}$ along the line-of-sight direction:
\begin{equation}\label{eq:arnaud-y-cyl}
y^\text{(A)}_\text{cyl}(r) \propto
\int_{-5R_\text{500c}}^{5R_\text{500c}}
P\Big(\frac{\sqrt{r^2+z^2}}{R_\text{500c}}\Big) \mathrm{d}z,
\end{equation}
where the pressure profile has the form
\begin{equation}
P(x) \propto
(c_\text{500}x)^{-\gamma_\text{A}} [1+(c_\text{500}x)^{\alpha_\text{A}}]^{(\gamma_\text{A}-\beta_\text{A})/\alpha_\text{A}},
\end{equation}
with $[c_\text{500},\gamma_\text{A},\alpha_\text{A},\beta_\text{A}]=[1.177,0.3081,1.0510,5.4905]$ \cite[]{arnaud10}. The integration up to $5R_\text{500c}$ includes more than 99~per cent of the total pressure contribution. The spherical $Y_\text{500}$ for the
Arnaud profile is:
\begin{equation}
\label{eq:xbcs-yszproj-sph}
Y_\text{500}^\text{(A)} = 2\pi y_{0} \int_0^{R_\text{500c}} y^\text{(A)}_\text{cyl}(r)
r\mr{d}r /1.203,
\end{equation}
where the numerical factor 1.203 is the ratio between cylindrical integration and spherical integration for the adopted Arnaud profile parameters.
Measurements of $Y_\text{500}$ are sensitive to the assumed profile. The Arnaud profile depends only on $R_\text{500c}$, while the $\beta$ profile depends on both $R_\text{500c}$ and $R_\text{c}$ and therefore $Y_\text{500}$ is sensitive to the ratio $R_\text{c}/R_\text{500c}$. We find that with $R_\text{c}/R_\text{500c}=0.2$ the $\beta$ and Arnaud profiles provide $Y_\text{500}$ measurements in good agreement; this ratio is consistent with the previous \ac{sze} profile study using high mass clusters \citep{plagge10}. Interestingly, the X-ray data indicate a characteristic ratio of $0.11\pm0.03$ for our sample, and a shift in the $R_\text{c}/R_\text{500c}$ ratio from 0.2 to 0.1 leads to a $\sim$40~per cent decrease in $Y_\text{500}$. Given that the {\it Planck}\ analysis to which we compare is carried out using the Arnaud profile, we adopt that profile for the analysis in Section~\ref{sec:xbcs-yszm} below.
The $Y_\text{500}$-mass scaling relation has been modelled using a representative local X-ray cluster sample \cite[]{arnaud10} and further studied in the \ac{sze} \cite[][P11]{andersson11} as
\begin{equation}\label{eq:xbcs-ym}
Y_\text{500} = \ay \Big(\frac{M_{500}}{1.5\times10^{14} M_{\odot}}\Big)^{\by}E(z)^{2/3}\Big[\frac{D_\text{A}(z)}{500\text{Mpc}}\Big]^{-2},
\end{equation}
where $D_\text{A}(z)$ is the angular-diameter distance and the intrinsic scatter on $Y_\text{500}$ is described by a log-normal distribution of width $\sigma_\mr{\ln Y}=0.21$. The observational scatter of $Y_\text{500}$ is propagated from the scatter of $\xi_{\text X}$.
In \Fref{sec:xbcs-result}, we fit this relation to the observations.
\section{Method}\label{sec:xbcs-method}
In this section, we describe the method we developed to fit the \ac{sze}-mass scaling relations of the low mass cluster population selected through the \ac{xbcs} and observed by the \ac{spt}. In principle, we could use our cluster sample observed in X-ray and \ac{sze} to simultaneously constrain the cosmology and the scaling relations, in the so-called self-calibration approach \cite[]{majumdar04}. However, self-calibration requires a large sample. Without this, we take advantage of strong, existing cosmology constraints \citep[e.g.,][]{planck13-16,bocquet14} and knowledge of the $L_{\text X}$-mass scaling relation \citep[e.g.][]{pratt09}. We focus only on the \ac{sze}-mass scaling relations, exploring the \ac{sze} characteristics of low mass galaxy clusters and groups. In \Fref{sec:xbcs-lik} we present the method and in \Fref{sec:xbcs-submocks} we validate it using mock catalogs.
\begin{figure*}[htb]
\centering
\begin{subfigure}{0.3\textwidth}
\hskip-0.75\textwidth \includegraphics[width=3.5in]{xr_mock08_3d.pdf}
\end{subfigure}
~
\begin{subfigure}{0.3\textwidth}
\includegraphics[width=3.5in]{xr_mock08_3d_Dsz.pdf}
\end{subfigure}
\caption[Constraints on the $\zeta$-mass relation from an analysis of the mock catalog]{Constraints on the $\zeta$-mass relation from an analysis of the mock catalog. The left panel constrains $A_{\text{SZ}}$, $B_{\text{SZ}}$, and $C_{\text{SZ}}$ with fixed $D_{\text{SZ}}$. And the right panel shows the result by fixing $C_{\text{SZ}}$ instead of $D_{\text{SZ}}$. The red lines and stars denote the input values of the scaling relation parameters of the mock catalog. Histograms in each case show the recovered projected likelihood distribution for each parameter. Joint constraints for different pairs of parameters are shown in blue with different shades indicating the 1, 2, and 3$\,\sigma$ levels. }
\label{fig:xbcs-mock-test}
\end{figure*}
\subsection{Description of the Method}\label{sec:xbcs-lik}
The selection biases on scaling relations include the Malmquist bias and the Eddington bias,
which are manifestations of scatter and population variations associated with the selection observable. Several methods have previously been developed (e.g., \citealt{vikhlinin09b, mantz10a, allen11}; B13; \citealt{bocquet14}) to account for the sampling biases when fitting scaling relation and cosmological parameters simultaneously. In this analysis, we use a likelihood function that can be derived from the one presented in B13. For a detailed discussion we refer the reader to Appendix~\ref{sec:likelihood}; here we present an overview of the key elements of this likelihood function.
The likelihood function $\mathcal{L}(\bm{r}_\text{SZ})$ we use to constrain the \ac{sze}-mass relations is the product of the individual conditional probabilities to observe each cluster with \ac{sze} observable $Y_i$ (e.g., \ac{sptsz} significance $\xi_{\text X}$ or $Y_\text{500}$), given the cluster has been observed to have an X-ray observable $L_i$ and redshift $z_i$:
\begin{equation}\label{eq:xbcs-fullexp}
\mathcal{L}(\bm{r}_\text{SZ})=\Pi_{i}\ P(Y_{i}|L_{i},z_{i},\bm{c}, \bm{r}_\text{X}, \bm{r}_\text{SZ}, \Theta_\text{X}),
\end{equation}
where $i$ runs over the cluster sample, $\bm{r}_\text{SZ}$ contains the parameters describing the \ac{sze} mass-observable scaling relation that we wish to study, $\bm c$ contains the cosmological parameters, $\bm{r}_\text{X}$ contains the parameters describing the X-ray mass-observable scaling relation, and the survey selection in X-ray is encoded within $\Theta_\text{X}$. Note that the redshifts are assumed to be accurate such that the X-ray luminosity ($L_{\text X}$) is used instead of the true survey selection observable, which is the X-ray flux.
As noted above, given the size of our dataset we adopt fixed cosmology $\bm c$ and X-ray scaling relation parameters $\bm{r}_\text{X}$ to focus on the \ac{sze}-mass scaling relation. In \Fref{sec:xbcs-result} we examine the sensitivity of our results to the current uncertainties in cosmology and the X-ray scaling relation and find them to be unimportant for our analysis.
Within this context, the conditional probability density function for cluster $i$ can be written as the ratio of the expected number of clusters $\mr{d}N$ with observables $Y_i$, $L_i$ and $z_i$ within infinitesimal volumes $\mr{d}Y$, $\mr{d}L$ and $\mr{d}z$:
\begin{equation}\label{eq:xbcs-single-like}
P(Y_{i}|L_{i},z_{i},\bm{r}_\text{SZ},\Theta_\text{X}) =\frac{\mr{d}N(Y_{i}, L_{i}, z_{i}|\bm{r}_\text{SZ},\Theta_\text{X})}{\mr{d}N(L_{i}, z_{i}|\Theta_\text{X})},
\end{equation}
where we have dropped the cosmology $\bm{c}$ and X-ray scaling relation parameters $\bm{r}_\text{X}$ because they are held constant. Typically, the survey selection $\Theta_\text{X}$ is a complex function of the redshift and X-ray flux, but in the above expression it is simply the probability that a cluster with X-ray luminosity $L_i$ and redshift $z_i$ is observed
(i.e. $\mr{d}N\left(Y_{i},L_{i},z_{i}|\bm{r}_\text{SZ},\Theta_\text{X}\right)=\Theta_\text{X}(L_{i},z_{i})\mr{d}N(Y_{i},L_{i},z_{i}|\bm{r}_\text{SZ})$);
in \Fref{eq:xbcs-single-like} this same factor appears in both the numerator and denominator, and therefore it cancels out. Thus, studying the \ac{sze} properties of an X-ray selected sample does not require detailed modelling of the selection. If the selection were based on both $L$ and $Y$, then there would be no cancellation, because the selection probability in the numerator would be just $\Theta(L_{i},Y_{i},z_{i})$ while in the denominator it would have to be marginalised over the unobserved $Y$ as $\int \Theta(Y,L_{i},z_{i})\mr{d}Y$ (see \Fref{eq:xbcs-selection-yl}).
With knowledge of the cosmologically dependent mass function $n(M,z)\equiv\mr{d}N(M,z|\bm{c})/\mr{d}M\mr{d}z$ \cite[]{tinker08}, the ratio of the expected number of clusters can be written as:
\begin{equation}\label{eq:xbcs-simplify-likelihood}
P(Y_{i}|L_{i},z_{i},\bm{r}_\text{SZ}) =\frac{\int \mr{d}M P(Y_{i},L_{i}|M,z_{i},\bm{r}_\text{SZ})\,n(M,z_{i})}{\int \mr{d}M P(L_{i}|M,z_{i})\,n(M,z_{i})}.
\end{equation}
We emphasise that there is a residual dependence on the X-ray selection in our analysis in the sense that we can only study the \ac{sze} properties of the clusters that have sufficient X-ray luminosity to have made it into the sample. This effectively limits the mass range over which we can use the X-ray selected sample to study the \ac{sze} properties of the clusters.
To constrain the scaling relation in the presence of both observational uncertainties and intrinsic scatter, we further expand the conditional probability density functions in Equation~(\ref{eq:xbcs-simplify-likelihood}):
\begin{align}
P(Y_{i},L_{i}|M,z_{i},\bm{r}_\text{SZ}) = \iint&\mr{d}Y_\mr{t} \mr{d}L_\mr{t}\ P(Y_{i}, L_{i}|Y_\mr{t},L_\mr{t}) \nonumber\\
&\times P(Y_\mr{t},L_\mr{t} |M,z_{i},\bm{r}_\text{SZ}) , \label{eq:xbcs-ymlcore} \\
P(L_{i}|M,z_{i}) = \int&\mr{d}L_\mr{t}\ P(L|L_\mr{t}) P(L_\mr{t}|M,z_{i}),\label{eq:xbcs-mlcore}
\end{align}
where, as above, $Y_{i}$ and $L_{i}$ are the observed values, and $Y_\mr{t}$ and $L_\mr{t}$ are the true underlying observables related to mass through scaling relations that have intrinsic scatter. The first factor in each integral represents the measurement error, and the second factor describes the relationship between the pristine observables and the halo mass. Improved data quality affects the first factor, but cluster physics dictates the form of the second. These second factors are fully described by the power law mass-observable relations in Equations (\ref{eq:xbcs-lm}), (\ref{eq:zetam}), and (\ref{eq:xbcs-ym}) together with the adopted log-normal scatter.
We use this likelihood function under the assumption that there is no correlated scatter in the observables; in \Fref{sec:xbcs-submocks} we use mock samples that include correlated scatter to examine the impact on our results.
\subsection{Validation with Mock Cluster Catalogs}\label{sec:xbcs-submocks}
We use mock samples of clusters to validate our likelihood and fitting approach and to explore our ability to constrain different parameters. Specifically, we generate ten larger mock surveys of 60~$\sdeg$, with a similar flux limit of $1\times 10^{-14} \text{erg s}^{-1}\text{cm}^{2}$ and $z>0.2$. Each mock catalog contains $\sim400$ clusters, or approximately eight times as many as in the observed sample. The $\xi_{\text X}$ of the sample spans $-2.2\le\xi_{\text X}\le7.8$ with a median value of $1.4$. We include both the intrinsic scatter and observational uncertainties for both the $L_{\text X}$ and the $\xi_{\text X}$ in the mock catalog. The intrinsic scatter is lognormal distributed with values given as $\sigma_{\lnL_{\text X}}$ ($D_{\text{SZ}}$). The observational uncertainties in $L_{\text X}$ and $\xi_{\text X}$ are modelled as normal distributions. The standard deviation used for $L_{\text X}$ is proportional to $\sqrt{L_{\text X}}$ to mimic the Poisson distribution of photon counts, while the standard deviation for $\xi_{\text X}$ is 1.
Here we focus on recovering the four \ac{sptsz} $\zeta$-mass relation parameters from the mock catalog; the fiducial values for these parameters are the B13 best-fitting values. We scan through the parameter space using a fixed grid. The following results contain 41 bins in each parameter direction. Given the limited constraining power, we validate the parameters using two different sets of priors. In the first set we adopt flat priors on $A_{\text{SZ}}$, $B_{\text{SZ}}$, and $C_{\text{SZ}}$ with fixed $D_{\text{SZ}}$. In the second set we adopt flat positive priors on $A_{\text{SZ}}$, $B_{\text{SZ}}$, and $D_{\text{SZ}}$ with fixed $C_{\text{SZ}}$. All other relevant parameters are fixed, including the $L_{\text X}$-mass scaling and the cosmological model.
Our tests show good performance of the method. Using ten mock samples that are each ten times larger than our observed sample, and fitting for 3 parameters in each mock, we recover the parameters to within the marginalised $1\,\sigma$ statistical uncertainty 70~per cent of the time and to within $2\,\sigma$ for the rest. \Fref{fig:xbcs-mock-test} illustrates our $\zeta$-mass parameter constraints from one mock sample. Note that the constraints on $C_{\text{SZ}}$ and $D_{\text{SZ}}$ are both weak and exhibit no significant degeneracy with the other two \ac{sptsz} scaling parameters. We take this as motivation to fix $C_{\text{SZ}}$ and $D_{\text{SZ}}$ and focus on the amplitude $A_{\text{SZ}}$ and slope $B_{\text{SZ}}$ in the analysis of the observed sample. We have repeated this testing in the case of the $Y_\text{500}$-mass relation, and we see no difference in behavior.
We also investigate the sensitivity of our method when a correlation between intrinsic scatter in the X-ray and $\xi_{\text X}$ is included. Cluster observables can be correlated through an analysis approach. For example, if one uses the $L_{\text X}$ as a virial mass estimate, then when $L_{\text X}$ scatters up by 40~per cent, it leads to a 5~per cent increase in radius, and 8~per cent increase in $Y_\text{500}$ if the underlying SZE brightness distribution is described by the \cite{arnaud10} profile. In comparison, the intrinsic scatter of $Y_\text{500}$ about mass is about 20~per cent, which in this example would still dominate over the correlated component of the scatter. Correlated scatter in different observable-mass relations can also reflect underlying physical properties of the cluster that impact the two observables in a similar manner.
We find that even with a correlation coefficient $\rho=0.5$ between the intrinsic scatter of the two observables, the change in constraints extracted using a no correlation assumption is small. Thus, our approximation does not lead to significant bias in the analysis of this sample. This result is also consistent with the fact that by extending Equations~(\ref{eq:xbcs-ymlcore}) and (\ref{eq:xbcs-mlcore}) to include multi-dimensional log-normal scatter distributions, we find the constraint on correlated scatter in the mock catalog to be very weak. We therefore do not include the possibility of correlated scatter when studying the real sample.
\section{Results}\label{sec:xbcs-result}
In this section, we present the observed relationship between the \ac{sze} significance $\xi_{\text X}$ at the position of the X-ray selected cluster and the predicted value given the measured X-ray luminosity of the system. Thereafter, we test -- and rule out -- the null hypothesis that the \ac{sze} signal at the locations of the X-ray selected clusters is consistent with noise. We then present constraints on the \ac{sptsz} $\zeta$-mass and $Y_\text{500}$-mass relations. We end with a discussion of possible systematics and a presentation of the point source population for this X-ray selected group and cluster sample.
\subsection{SPT Significance Extraction}
We extract the $\xi_{\text X}$ from the \ac{sptsz} multi-frequency-filtered map at the location of each \ac{xbcs} selected cluster as described in \Fref{sec:xbcs-data-description}. In the primary analysis, we adopt
three matched-filtered maps from the \ac{sptsz} data, one each for $\beta$-model profiles with $R_\text{c} =$ 0.25, 0.5, and 0.75 arcmin, and we extract the value of $\xi_{\text X}$ for each cluster from the map that most closely matches the X-ray-derived $R_\text{c}$ value for that cluster. The $\xi_{\text X}$ is extracted at the X-ray-derived cluster position. The measured $\xi_{\text X}$ values are presented in \Fref{tab:xbcs-cat}. We have also tried extracting \ac{sptsz} significance by making a matched-filtered map for every cluster, using a filter with the exact X-ray-derived value of $R_\text{c}$, and the change in the results is negligible.
We have also investigated the dependence of $\xi_{\text X}$ on the assumed cluster profile. We repeated the analysis described above using the Arnaud profile and a $\beta$ profile with $\beta=2/3$. The resulting changes in the extracted values of $\xi_{\text X}$ are less than 3~per cent of the measurement uncertainty on the individual $\xi_{\text X}$ values. A similar lack of sensitivity to the assumed cluster profile is seen in the $\xi > 5$ \ac{sptsz} derived cluster samples.
The cluster with the strongest detection in the \ac{sptsz} maps is illustrated in \Fref{fig:xbcs-clu044}, which contains a pseudo-colour optical image with %
\ac{sptsz} signal-to-noise contours in white. The \ac{sptsz} significance, $\xi$, of this cluster is 6.23 corresponding to maximum signal-to-noise in the filtered map (SPT-CLJ2316-5453, Bleem et al. in prep.), whereas the $\xi_{\text X}$ is 4.58 at the X-ray position with $R_\text{c}$ of 0.367~arcmin. This reduction in signal to noise is expected because there is noise in the \ac{sze} map, and the \ac{sptsz} cluster is selected to lie at the peak $\xi$.
\begin{figure}
\hskip-0.0in\includegraphics[width=3.3in]{clu044.pdf}
\caption[BCS optical image of cluster 044.]
{\ac{bcs} optical pseudo-colour image of cluster 044 in {\it gri} bands. The yellow circle (1.5~arcmin diameter) centred at the X-ray peak indicates the rough size of the \ac{spt} beam (1.2~arcmin FWHM in 150~GHz and 1.6~arcmin in 95~GHz). The \ac{sptsz} filtered map is overlaid with white contours, which are marked with the significance levels. The offset between the X-ray centre and the \ac{sze} peak is $0.75$\,arcmin, and the BCG for this system lies near those two centres.}
\label{fig:xbcs-clu044}
\end{figure}
\begin{table*}
\begin{center}
\caption[SPT significance of the XMM-BCS sample.]{\ac{sptsz} $\xi_{\text X}$ of \ac{xbcs} sample.}\label{tab:xbcs-cat}
\begin{tabular}{|c|r|r|r|r|r|r|c|c|}
\hline
ID & \begin{tabular}{@{}c@{}} $L_{\mathrm{X,500,bol}}$ \\
$[10^{42}~\text{erg/s}]$\end{tabular} &
\begin{tabular}{@{}c@{}} $\Delta L_{\mathrm{X,500,bol}}$ \\
$[10^{42}~\text{erg/s}]$ \end{tabular} &
Redshift &\begin{tabular}{@{}c@{}} Redshift \\
uncertainty \end{tabular} & \begin{tabular}{@{}c@{}} $R_\text{c}$ \\
$[\text{arcmin}]$ \end{tabular} &
$\xi_{\text X}$ &\begin{tabular}{@{}c@{}} SPT point source \\ separation [arcmin] and SN\end{tabular} &
\begin{tabular}{@{}c@{}} SUMSS point source \\ separation
[arcmin] \end{tabular} \\ \hline
011 & $ 345.2$&$ 51.6$ & $ 0.97$&$ 0.10$ & 0.185 & 0.99& - & - \\
018 & $ 66.3$&$ 6.5$ & $ 0.39$&$ 0.04$ & 0.239 & 1.90& - & 0.92 \\
032 & $ 684.0$&$ 56.8$ & $ 0.83$&$ 0.07$ & 0.272 & 3.04& - & 1.70, 2.30, 3.97 \\
033 & $ 209.0$&$ 17.6$ & $ 0.79$&$ 0.05$ & 0.189 & 2.34& - & - \\
034 & $ 16.0$&$ 2.5$ & $ 0.28$&$ 0.02$ & 0.197 & -0.38& - & - \\
035 & $ 91.0$&$ 14.3$ & $ 0.67$&$ 0.05$ & 0.164 & 2.78& - & 0.10, 1.56 \\
038 & $ 16.3$&$ 2.5$ & $ 0.39$&$ 0.05$ & 0.147 & -0.20& - & 1.85 \\
039 & $ 19.4$&$ 1.2$ & $ 0.18$&$ 0.04$ & 0.315 & -0.34& - & 2.91 \\
044 & $ 310.5$&$ 20.5$ & $ 0.44$&$ 0.02$ & 0.367 & 4.58& 3.87\quad 4.84 & 0.22 \\
069 & $ 124.9$&$ 21.5$ & $ 0.75$&$ 0.07$ & 0.165 & 1.38& 3.40\quad 6.34& 3.42 \\
070 & $ 137.9$&$ 2.8$ & $ 0.152$&$ 0.001$ & 0.726 & 1.80& - & - \\
081 & $ 93.1$&$ 15.4$ & $ 0.85$&$ 0.12$ & 0.133 & -1.56& - & - \\
082 & $ 53.6$&$ 9.2$ & $ 0.63$&$ 0.05$ & 0.144 & 0.55& - & - \\
088 & $ 122.1$&$ 16.7$ & $ 0.43$&$ 0.04$ & 0.271 & -0.10& - & 2.96 \\
090 & $ 25.4$&$ 5.8$ & $ 0.58$&$ 0.02$ & 0.120 & 0.30& - & - \\
094 & $ 26.3$&$ 2.9$ & $ 0.269$&$ 0.001$ & 0.243 & 2.20& - & 1.48 \\
109 & $ 196.9$&$ 28.8$ & $ 1.02$&$ 0.09$ & 0.145 & 1.09& - & 0.19 \\
110 & $ 68.8$&$ 9.3$ & $ 0.47$&$ 0.06$ & 0.205 & -1.07& - & 0.10 \\
126 & $ 82.0$&$ 6.1$ & $ 0.42$&$ 0.02$ & 0.240 & 0.03& - & 1.22 \\
127 & $ 8.4$&$ 1.0$ & $ 0.207$&$ 0.001$ & 0.207 & 1.28& - & - \\
132 & $ 319.3$&$ 35.7$ & $ 0.96$&$ 0.17$ & 0.182 & 1.74& - & - \\
136 & $ 86.8$&$ 7.3$ & $ 0.36$&$ 0.02$ & 0.282 & -3.58& 1.11\quad 5.84 & 1.00 \\
139 & $ 8.7$&$ 1.2$ & $ 0.169$&$ 0.001$ & 0.252 & -0.17& - & 0.44 \\
150 & $ 37.7$&$ 1.8$ & $ 0.176$&$ 0.001$ & 0.403 & -3.34& 0.13\quad 4.23 & 0.05, 2.29 \\
152 & $ 3.4$&$ 0.6$ & $ 0.139$&$ 0.001$ & 0.219 & -0.45& - & - \\
156 & $ 166.0$&$ 11.7$ & $ 0.67$&$ 0.06$ & 0.202 & 3.01& - & - \\
158 & $ 104.2$&$ 15.6$ & $ 0.55$&$ 0.03$ & 0.205 & 1.94& - & - \\
210 & $ 45.0$&$ 9.0$ & $ 0.83$&$ 0.09$ & 0.105 & 0.18& - & - \\
227 & $ 14.5$&$ 1.8$ & $ 0.346$&$ 0.001$ & 0.157 & -1.03& - & 0.06\\
245 & $ 38.1$&$ 7.1$ & $ 0.62$&$ 0.03$ & 0.130 & 0.24& - & 1.38 \\
275 & $ 17.8$&$ 2.7$ & $ 0.29$&$ 0.03$ & 0.198 & -0.46& - & 2.12 \\
287 & $ 31.1$&$ 11.0$ & $ 0.57$&$ 0.04$ & 0.131 & -0.02& - & - \\
288 & $ 89.0$&$ 17.4$ & $ 0.60$&$ 0.04$ & 0.180 & -0.25& - & 0.62 \\
357 & $ 66.3$&$ 8.3$ & $ 0.48$&$ 0.06$ & 0.198 & -0.97& - & - \\
386 & $ 17.7$&$ 4.8$ & $ 0.53$&$ 0.05$ & 0.115 & 0.83& 0.417\quad 4.53$^*$ & - \\
430 & $ 4.5$&$ 0.9$ & $ 0.206$&$ 0.001$ & 0.167 & -0.67& - & - \\
444 & $ 69.1$&$ 13.8$ & $ 0.71$&$ 0.05$ & 0.141 & -0.13& - & - \\
457 & $ 1.1$&$ 0.3$ & $ 0.100$&$ 0.001$ & 0.201 & -1.24& - & - \\
476 & $ 6.2$&$ 0.7$ & $ 0.101$&$ 0.001$ & 0.365 & -0.12& - & 1.03 \\
502 & $ 47.2$&$ 4.2$ & $ 0.55$&$ 0.05$ & 0.156 & -0.30& - & - \\
511 & $ 23.4$&$ 3.7$ & $ 0.269$&$ 0.001$ & 0.233 & 0.11& - & 0.15, 2.37 \\
527 & $ 160.8$&$ 26.2$ & $ 0.79$&$ 0.06$ & 0.172 & 0.83& - & 3.96 \\
528 & $ 6.4$&$ 2.1$ & $ 0.35$&$ 0.02$ & 0.117 & 0.57& - & - \\
538 & $ 5.1$&$ 2.1$ & $ 0.20$&$ 0.02$ & 0.179 & 0.30& - & - \\
543 & $ 134.5$&$ 29.6$ & $ 0.57$&$ 0.03$ & 0.217 & 1.10& - & - \\
547 & $ 4.1$&$ 1.3$ & $ 0.241$&$ 0.001$ & 0.140 & -6.45& 0.20\quad 6.75 & 0.12, 2.89 \\
\hline
\end{tabular}
\end{center}
\raggedright
$^*$Detected in 220~GHz.
\end{table*}%
\subsection{Testing the Null Hypothesis}
To gain a sense of the strength of the \ac{sze} detection of the ensemble of \ac{xbcs} clusters, we test the measured significance around \ac{sze} null positions. A single null catalog consists of the same number of clusters as the \ac{xbcs} sample where the X-ray luminosities and redshifts are maintained, but the \ac{sptsz} significances $\xi_{\text X}$ are measured at random positions. We then carry out a likelihood analysis of three null catalogs. When fixing the slope $B_{\text{SZ}}$ of the scaling relation, we find that the normalisation factor $A_{\text{SZ}}$ is constrained to be $<0.56$ at 99~per cent confidence level for all three null samples we tested. Because this constraint on the amplitude is small compared to the expected normalisation for the \ac{xbcs} sample, we have essentially shown that there should be sufficient signal to noise to detect the \ac{sze} signature of the cluster ensemble.
\subsection{SPT \texorpdfstring{$\zeta$}{zeta}-mass Relation}
\label{sec:xbcs-xim}
We explore the \ac{sze} signature of low mass clusters by constraining the $A_{\text{SZ}}$ and $B_{\text{SZ}}$ parameters with the approach described and tested above. The X-ray luminosity-mass scaling relation, \Fref{eq:xbcs-lm}, is directly adopted with the additional observational uncertainties of each cluster that are listed in \Fref{tab:xbcs-cat} (bolometric luminosities presented in S12).
We present results for four different subsets of our sample: 1) the full sample without removal of any cluster; 2) the sample excluding any cluster with a point source detected at $>$4\,$\sigma$ in any \ac{spt} observing band within a 4~arcmin radius of the X-ray cluster (see \Fref{tab:xbcs-cat}), hereafter SPT-NPS sample; 3) the SPT-NPS clusters with redshift larger than 0.3, hereafter SPT-NPS($z>0.3$), which is the best match to the selection of the \ac{sptsz} high mass sample in B13 and 4) the sample without any \acl{sumss} \cite[\acsu{sumss},][]{bock99,mauch03} point sources in 4~arcmin radius. We discuss further the astrophysical nature and impact of point sources in \Fref{sec:xbcs-PScontamination}.
In \Fref{fig:xbcs-corrbias}, we illustrate the $\zeta$-mass relation obtained by plotting the observed $\xi_{\text X}$ versus the expected $\avg{\zeta(L_{\text X},z)}$, estimated using \Fref{eq:xbcs-fullexp}. Here we use the best fit scaling relation from the SPT-NPS (black points only). Note that the typical bias correction on the mass is about 10 percent at the high mass end.
We explore the likelihood as a function of $A_{\text{SZ}}$ and $B_{\text{SZ}}$ and show the parameter constraints for the three samples in \Fref{tab:xbcs-params}, and we show the likelihood distribution of the SPT-NPS sample in \Fref{fig:xbcs-AszBsz}. We also show marginalised single parameter probability distributions, which we use to calculate the 68~per cent confidence region for each parameter. This confidence region along with the modal value is reported in \Fref{tab:xbcs-params}. For comparison, the constraints from the B13 analysis are shown in red.
\begin{figure}
\hskip-0.2in\includegraphics[width=3.5in]{correctbias.pdf}
\caption[The measured significance $\xi_{\text X}$ versus the expected \ac{sptsz} $\avg{\zeta(L_{\text X},z)}$ from XMM-BCS sample]{The measured significance $\xi_{\text X}$ versus the expected \ac{sptsz} $\avg{\zeta(L_{\text X},z)}$, where the best-fitting relation from the SPT-NPS sample and sampling bias corrections are applied. Overplotted is the line of equality. Clusters close to \ac{spt} point sources are marked with red diamonds.}
\label{fig:xbcs-corrbias}
\end{figure}
\begin{table}
\begin{center}
\caption[Constraints on the SPT zeta-mass relation
parameters.]{Constraints on the SZE $\zeta$-mass relation
parameters.\label{tab:xbcs-params}}
\begin{tabular}{|c|c|c|}
\hline
& $A_{\text{SZ}}$ & $B_{\text{SZ}}$ \\ \hline
SPT High Mass (B13) & $1.50\pm0.34$& $1.40\pm0.16$\\
Prior & $[0.1-5]$ & $[0.1-6]$ \\
Full sample & $1.38^{+0.46}_{-0.36}$ & $2.80^{+0.66}_{-0.63}$\\
SPT-NPS & $1.37^{+0.48}_{-0.38}$&$2.14^{+0.86}_{-0.66}$ \\
SPT-NPS ($z>0.3$) &$1.37^{+0.60}_{-0.46}$ & $2.31^{+1.31}_{-0.86}$\\
SPT-No-SUMSS & $1.42^{+0.58}_{-0.43}$ & $2.14^{+0.91}_{-0.71}$ \\
\hline
\end{tabular}
\label{tab:xbcs-constraint}
\end{center}
\end{table}
All three low mass subsamples show similar normalisation to the extrapolated high mass \ac{sptsz} sample, but there is a preference for larger slopes. The SPT-NPS sample is the best for comparison to the \ac{sptsz} high mass sample used in B13; this is because the \ac{spt} point sources have been removed to mimic the \ac{spt} cluster catalog selection and because there is no measurable difference between the SPT-NPS samples with or without the redshift cut.
The fact that we find consistent results with or without a low-redshift cut may at first be surprising, given that analyses of the high-mass \ac{sptsz} cut all clusters below z=0.3. In the \ac{sptsz} high mass sample, the low redshift clusters are cut because the angular scales of these clusters begin to overlap the scales where there is significant \ac{cmb} primary anisotropy, making extraction with the matched filter approach using two frequencies difficult. However the \ac{xbcs} clusters are low mass systems with corresponding $R_\text{c}$ less than 1~arcmin even at low redshift. So we are able to recover the same scaling relation with or without the low redshift clusters.
The fully marginalised posterior probability distributions for $B_{\text{SZ}}$ can be used to quantify consistency between the two datasets. We do this for any pair of the distributions $P_i\left(\theta\right)$ by first calculating the probability density distribution of the difference $\Delta \theta$:
\begin{equation}
P(\Delta \theta) = \int \mr{d}\theta P_{1}(\theta)P_{2}(\theta-\Delta\theta).
\label{eq:xbcs-1dp2exceed}
\end{equation}
We then calculate the likelihood $p$ that the origin ($\Delta\theta=0$) lies within this distribution as
\begin{equation}
p = \int_S \mr{d}\Delta\theta\ P(\Delta\theta)
\end{equation}
where $S$ is the space where $P(\Delta\theta)<P(\Delta\theta=0)$. We then convert this $p$ value to an equivalent $N$-$\sigma$ significance within a normal distribution.
\begin{figure}
%
\hskip-0.3in\includegraphics[width=3.5in]{lik2d_2.pdf}
\caption[Constraints on the SPT $\zeta$-mass relation parameters $A_{\text{SZ}}$ and $B_{\text{SZ}}$ for the non-point source sample (SPT-NPS)]{Constraints on the \ac{sptsz} $\zeta$-mass relation parameters $A_{\text{SZ}}$ and $B_{\text{SZ}}$ for the non-point source sample (SPT-NPS). The different shading indicates 1, 2, and 3$\sigma$ confidence regions. The constraints from the \ac{sptsz} high mass clusters (B13) are shown in red with 68~per cent confidence regions marked with dashed lines. The amplitudes for low and high mass clusters are compatible, but the slope is higher for low mass systems by about 1.4$\,\sigma$.}
\label{fig:xbcs-AszBsz}
\end{figure}
Overall, there is no strong statistical evidence that the low mass clusters behave differently than expected by simply extrapolating the high mass scaling relation to low mass; the slope parameter $B_{\text{SZ}}$ of the \ac{sptsz} high mass and SPT-NPS samples differs by only $1.4\,\sigma$ (\Fref{tab:xbcs-constraint}). The full sample has a $2.6\,\sigma$ higher $B_{\text{SZ}}$ than the \ac{sptsz} high mass sample \citep{benson13}. This steeper slope is presumably due to the contaminating effects of the \ac{spt} point sources. We find three outliers below the $L_{\text X}\text{-}\xi_{\text X}$ distribution (\Fref{fig:xbcs-corrbias}) that are all contaminated by \ac{spt} point sources. We list the separation between the cluster centres and the nearest \ac{spt} point source in \Fref{tab:xbcs-cat}.
It is clear from \Fref{fig:xbcs-corrbias} and from the results for the full sample
that including X-ray-selected clusters that are associated with point
sources that are independently detected in \ac{sptsz} data can bias the
derived \ac{sze}-mass relation. In these cases, the affected clusters can
be removed from the sample, and this particular bias can be easily avoided.
Point sources that are not detected in the \ac{sptsz} data but which could be
significantly affecting the measured \ac{sze} signal -- particularly in low-mass
clusters and groups -- do remain a potential issue. We discuss this and the effect
of point sources on our results more generally in \Fref{sec:xbcs-PScontamination}.
In addition to the X-ray bolometric luminosities, we test the luminosities based on two other bands (0.5--2.0~keV and 0.1--2.4~keV) as predictors of the cluster mass. After applying the appropriate $L_{\text X}$-mass relations listed in \Fref{tab:xbcs-lxm} we find that the changes to the parameter estimates are small. The largest change is on the slope of the \ac{sptsz} $\zeta$-mass relation, but the difference is less than $0.2\,\sigma$. Thus, the choice of X-ray luminosity band is not important to our analysis.
Our results show some dependence on the assumed $L_{\text X}$-mass scaling relation. Adopting the \citet{vikhlinin09b} scaling relation has no significant impact on our results. However, with the \citet{mantz10b} $L_{\text X}$-mass relation, the slope decreases to $B_{\text{SZ}}\sim1.57$ from 2.14, which makes the SPT-NPS sample almost a perfect match to the high mass \ac{sptsz} scaling relation. This shift is not surprising, because the \citet{mantz10b} $L_{\text X}$-mass relation has a very different slope from \citet{pratt09} (1.63 vs. 2.08, respectively). This causes clusters with a $L_{\text X}<1\times10^{44}~\mathrm{erg\ s}^{-1}$to have significantly lower estimated masses when assuming the \citet{mantz10b} relation (20~per cent on average and $\sim40$~per cent at the low mass end). We expect the \citet{pratt09} relation to be more appropriate for our analysis, because the \citet{mantz10b} relation was calibrated from higher mass clusters, using only clusters with $L_{\text X} >2.5\times10^{44}~\mathrm{erg\ s}^{-1}$, above the majority of \ac{xbcs} clusters. Also we note the change of $\xi_{\text X}$ caused by the updated $R_\text{500c}(L_{\text X})$ is negligible, which has been shown also in \citet{saliwanchik13}.
\begin{figure}
\hskip-0.3in\includegraphics[width=3.5in]{yszm_2.pdf}
\caption[Constraints on the $Y_\text{500}$-mass relation parameters $\ay$ and $\by$ for the non-point source sample (SPT-NPS)]{Constraints on the $Y_\text{500}$-mass relation parameters $\ay$ and $\by$ for the non-point source sample (SPT-NPS). The SPT-NPS constraints are shown in blue and different shades show the 1, 2, and 3\,$\sigma$ levels. The red is for the \ac{sptsz} result \protect\citep{andersson11}, and the green is the best fit from the {\it Planck}\ analysis (P11). Marginalised constraints for each parameter are shown in blue with best fit and $1\,\sigma$ confidence regions marked by solid and dashed lines, respectively.}
\label{fig:xbcs-ym}
\end{figure}
\subsection{SZE \texorpdfstring{$Y_\text{500}$}{Y500}-mass
Relation}\label{sec:xbcs-yszm}
We measure the $Y_\text{500}$-mass relation, using the SPT-NPS sample. A similar fitting approach is used to account for the selection bias and with the same shifted pivot mass in \Fref{eq:xbcs-ym} of $1.5\times10^{14}~\msun$. The best fit parameters and uncertainties are presented in \Fref{tab:xbcs-yszm} along with the results from \cite{andersson11} and P11, which are adjusted to use our lower pivot mass. The $Y_\text{500}$ is based on the Arnaud profile and the $L_{\text X}$ is based on the X-ray luminosity measured within the 0.1--2.4~keV band, which facilitates the comparison with the P11\ result. The impact from different profiles is discussed later in this section.
\Fref{fig:xbcs-ym} shows the joint parameter and fully marginalised constraints for $\ay$ and $\by$. The shaded regions denote the 1, 2, and 3\,$\sigma$ confidence regions as in \Fref{fig:xbcs-AszBsz} with blue for the SPT-NPS, red for the \ac{sptsz} sample \cite[]{andersson11}, and green for the {\it Planck}\ sample (P11). This figure shows that the low mass SPT-NPS sample has rather weak constraints that are shifted with respect to the high mass \ac{sptsz} sample and the {\it Planck}\ sample.
We estimate the significance of the difference using the method described in \Fref{sec:xbcs-xim}. We quantify the consistency between any pair of the two-parameter distributions $P_i\left(\bm\theta\right)$ by calculating a $p$ value in a manner similar to that in \Fref{eq:xbcs-1dp2exceed} with the null hypothesis $\Delta \bm \theta =0$. Using this approach, we calculate that the SPT-NPS sample is roughly consistent with the high mass \ac{sptsz} sample (a $1.4\,\sigma$ difference) but is in tension with the {\it Planck}\ result (a $2.8\,\sigma$ difference).
Also shown in \Fref{fig:xbcs-ym} are the fully marginalised single parameter constraints. These distributions indicate that the normalisation differs by $0.8\sigma$ ($1.6\,\sigma$), and the slope parameter differs by $1.7\,\sigma$ ($1.7\,\sigma$) for the \ac{sptsz} ({\it Planck}) sample. Alternatively, we fix $\by=1.67$ ($1.78$) to limit the impact of the large uncertainty on the slope on the constraint of the normalisation. In this case, we find $\ay=1.33^{+0.34}_{-0.31}$ ($1.37^{+0.36}_{-0.32}$) and the discrepancy on $\ay$ is $1.5\,\sigma$ ($3.1\,\sigma$) for the \ac{sptsz} ({\it Planck}) sample. As in the $\zeta$-mass relation, there is no strong statistical evidence that the \ac{sptsz} clusters at low mass behave differently than those at high mass. Tighter constraints on the high mass \ac{sptsz} scaling relation will be helpful to understand the tension.
The tension with the {\it Planck}\ sample is intriguing; here we discuss several possible issues that could contribute. One difference is in the mass ranges probed in the two studies. In P11, the {\it Planck}\ team studies the relation between X-ray and \ac{sze} properties of 1600 clusters from the Meta-Catalogue of X-ray detected Clusters of galaxies \cite[{MCXC},][]{piffaretti11} that span two decades in luminosity ($10^{43}~\mathrm{erg\ s^{-1}} \lesssim L_{500,[0.1\ -\ 2.4~\mathrm{keV}]} E(z)^{-7/3}\lesssim 2\times10^{45}~\mathrm{erg\ s^{-1}}$). In contrast, our sample spans the range $10^{42}\mathrm{erg\ s^{-1}} \lesssim L_{500,[0.1\ -\ 2.4~\mathrm{keV}]} E(z)^{-7/3}\lesssim 10^{44}~\mathrm{erg\ s^{-1}}$ extending into the galaxy group regime. Thus, it is interesting to probe for any mass trends in the discrepancy. In \Fref{fig:xbcs-cmpplanck}, we show our measurements along with the {\it Planck}\ relation with fixed slope and redshift evolution as listed in Table~4 in P11\ (solid black line). At the luminous (massive) end, our sample matches well with the {\it Planck}\ result (cyan points are taken from Figure~4 in P11). Beyond the {\it Planck}\ sample at the faint end, we find the preference for lower $Y_\text{500}$ relative to the {\it Planck}\ relation.
In the {\it Planck}\ analysis, an $L_{\text X}$-mass relation without Malmquist bias correction is used \citep{pratt09}. They argue that based on the similarity between the {REXCESS} and {MCXC} samples, there is no bias correction needed. In our analysis, we use the Malmquist bias-corrected relation and our likelihood corrects for selection bias. Using the non-corrected relation \citep{pratt09} has very little impact. Interestingly, if we adopt the \citet{mantz10b} relation, the tension between our result and the {\it Planck}\ result disappears mainly due to the lower masses predicted by the relation as discussed in \Fref{sec:xbcs-xim}. However, given that the {\it Planck}\ analysis adopted the \citet{pratt09} relation, it is with this same relation that the most meaningful comparisons can be made.
\begin{table}
\caption{Constraints on the $Y_\text{500}$-mass relation.}
\label{tab:xbcs-yszm}
\centering
\begin{tabular}{|crr|}
\hline
Parameter & $\ay[10^{-4} \mathrm{arcmin^{2}]}$& $\by$ \\
\hline
SPT-NPS & $1.59^{+0.63}_{-0.48}$ & $2.94^{+0.77}_{-0.74}$ \\
SPT-No-SUMSS & $1.72^{+1.01}_{-0.66}$& $3.29^{+0.84}_{-0.96}$\\
SPT & $2.19\pm0.63$ & $1.67\pm0.29$ \\
{\it Planck} & $2.57\pm0.11$ & $1.78\pm0.05$ \\
\hline
\end{tabular}
\end{table}
\begin{figure}
\hskip-0.15in\includegraphics[width=3.5in]{cmp_planck2.pdf}
\caption[Comparison with the {\it Planck}\ $Y_\text{500}-L_{\text X}$ relation]{Comparison with the {\it Planck}\ $Y_\text{500}-L_{\text X}$ relation. The green dots are XMM-BCS\ clusters with $1\,\sigma$ uncertainty on $\xi_{\text X}$ and measured uncertainties on $L_{\text X}$ converted from the 0.5--2~keV band. Blue points are inverse variance weighted means of ensembles of the XMM-BCS\ sample. The black line is the {\it Planck}\ \ac{sze} relation from table 4 in P11\ with the last four binned data points from figure~4 (P11) in cyan. The red line is the best-fit relation from the SPT sample. The correction of selection bias leads to a higher-than-measured $Y_\text{500}$\ at high mass (luminous) end as the mass function is steep.) Consistent with our parameter constraints in \Fref{fig:xbcs-ym}, our measurements prefer a lower value than the {\it Planck}\ relation. Clusters close to \ac{spt} point sources are marked with red diamonds.}
\label{fig:xbcs-cmpplanck}
\end{figure}
Second, the {\it Planck}\ relation is dominated by the high mass clusters, and their measurements at the low luminosity end (marked by cyan points in \Fref{fig:xbcs-cmpplanck}) also tend to fall below their best fit relation. The lowest luminosity {\it Planck}\ point has a $Y_\text{500}$\ that is 68~per cent (2\,$\sigma$ offset) of the value of the best fit model at the same X-ray luminosity. Interestingly, the best fit normalisation of the SPT-NPS sample is 53~per cent of the {\it Planck}\ model normalisation. In this sense, the tension between the two low mass samples is less than the tension between our sample and the best-fitting {\it Planck}\ relation.
Third, we note the redshift dependence of $Y_\text{500}$-mass relation could lead to a different normalisation because the SPT-XBCS sample is on average at higher redshift than the {\it Planck}\ sample. In P11, they show a weak redshift evolution of $Y_\text{500}$, where the index of $E(z)$ term is $-0.007\pm 0.518$. When they fit with the redshift evolution fixed to the self-similar expectation ($2/3$), it changes the $Y_\text{500}$ normalisation by $-5$~per cent ($0.451/0.476$), because $E(z)$ is larger than 1 for $z>0$. In comparison, if we assume an index of 0 for $E(z)$ it will increase our $Y_\text{500}$ normalisation by 19~per cent compared to the $E(z)^{2/3}$ case (\ac{xbcs} sample has a mean redshift of $0.48$). In this sense, there is some systematic uncertainty in the tension between the two samples that depends on the true redshift evolution of the $Y_\text{500}$-mass relation. If the samples evolve self-similarly, then the {\it Planck}\ normalisation should be reduced by 5~per cent.
Finally, the comparison to {\it Planck}\ is complicated because of differences between the \ac{spt} and {\it Planck}\ instruments and datasets and also differences between the analyses. Our analysis of \ac{sptsz} data calculates the \ac{sze} signal exclusively at frequencies below the \ac{sze} null (95 GHz and 150 GHz), where the \ac{sze} signal is negative, while {\it Planck}\ also includes information from frequencies above the 220 GHz \ac{sze} null, where the signal is positive. Thus, contamination from sources like radio galaxies with steeply falling spectra, which primarily affect the lowest-frequency bands in both instruments, would tend to bias both the {\it Planck}\ and \ac{sptsz} relations in the same way. But there are other possible sources of contamination such as dusty star-forming galaxies that are much brighter at higher frequencies. A population of star-forming galaxies associated with clusters could artificially increase the {\it Planck}\ measured $Y_\text{500}$, but could only negatively bias the \ac{sptsz} measurements. In their paper, the Planck team shows that at the low mass end (stellar mass smaller than $10^{11.25}\msun$), the $Y_\text{500}$ estimated by six high frequency bands directly is higher than the $Y_\text{500}$ estimated when using a thermal dust model and the $Y_\text{500}$ estimated just using the three low frequencies (100, 143, 217~GHz) \citep{planck13-11}. However, at higher masses where we are seeing the discrepany between the SPT and {\it Planck}\ signals there is no clear evidence for a dust related systematic in the cross-checks carried out by the {\it Planck}\ team. We present 2.8$\sigma$ significant evidence for dusty galaxy flux in our cluster ensemble in \Fref{sec:xbcs-PScontamination} below. If present, this flux is likely contributing to some degree to the discrepancy we find between the SPT and {\it Planck}\ SZE signatures on these mass scales.
In summary, there are several potential contributing factors to the 2.8$\sigma$ tension between the two results. None of them provide a convincing explanation for the offset on their own, but there are indications that differential sensitivity to dusty galaxy flux in SPT and {\it Planck}\ could be playing a role. What is needed next is a larger sample with higher quality data to probe this tension and -- if the tension persists -- to provide insights into the underlying causes of the discrepancy.
\subsection{Potential Systematics}\label{sec:xbcs-sensitivity}
In the likelihood approach, we fix the cosmological parameters and assume no redshift uncertainty to improve the efficiency of the calculation. We test both of these assumptions and find that neither significantly impacts the analysis. Specifically, the mass function used for correcting the sampling bias is adopted from a fixed cosmology $(\Omega_{\mathrm M}, \Omega_{\Lambda},H_{0})=(0.3,0.7,70~\text{km s}^{-1}\text{Mpc}^{-1})$. When we alter these to the recent WMAP results for $\Lambda$CDM \cite[]{komatsu11}, we find a negligible impact.
We test the importance of possible photometric redshift biases by shifting the redshifts of all clusters up (or down) by $1\,\sigma$. We update $L_{\text X}$ appropriately for the new redshifts, and we find a small ($0.5\,\sigma$) shift in the normalisation and no change to the slope. Therefore, redshift biases at this level would not significantly bias the analysis.
\subsection{Point Source Population}\label{sec:xbcs-PScontamination}
As already noted (see Section~\ref{sec:xbcs-xim}), there is a tendency for the systems with the most negative $\xi_{\text X}$ to be those with nearby \ac{spt} point sources (see \Fref{fig:xbcs-corrbias}). In this section, we explore this association in more detail, testing whether it is biasing our constraints on the \ac{sze} mass--observable relations. For the purposes of our analysis, an object is identified as an \ac{spt} point source if it appears as a 4$\,\sigma$ detection in a single frequency point-source filtered \ac{sptsz} map in any of the three bands (95, 150, or 220~GHz). An area within a 4~arcmin radius around each point source is defined, and all X-ray selected clusters within that region are flagged. There are six clusters flagged in our sample, and these are denoted with red diamonds in the figures presented above. Given the number densities of the \ac{spt} point sources (6~deg$^{-2}$ in this field) and the X-ray selected clusters together with the association radius, we estimate a 36~per cent chance that these point sources are random associations with the clusters.
If we consider a smaller 2~arcmin association radius between the X-ray centre and the \ac{spt} point source location, we still find four associations: three of which correspond to the most negative $\xi_{\text X}$ in \Fref{fig:xbcs-corrbias}, and the fourth is detected only at 220~GHz by \ac{spt} (and therefore is likely a dusty galaxy). With the smaller association radius the probability of a random association drops to 7~per cent, providing $\sim2\,\sigma$ evidence that these point sources are physically associated with the X-ray selected groups.
To further study the point source issue, we cross-match our cluster sample with radio sources detected at 843~MHz by the \ac{sumss}. The survey covers the whole sky at $\delta\leq-30^{\circ}$ with $|b|>10^{\circ}$ down to limiting source brightness of 6~mJy~beam$^{-1}$. For the cross-matching, we utilise the latest version 2.1 of the catalog\footnote{\url{http://www.physics.usyd.edu.au/sifa/Main/SUMSS}} and a similar matching radius of 2~arcmin. This threshold is much larger than the \ac{sumss} positional uncertainty, which has a median value of $\sim2.3$~arcsec.
Within 2~arcmin of the X-ray centres, we find a total of 19 \ac{sumss} point sources matching 18 clusters from our sample. In comparison, given the number density of \ac{sumss} sources \cite[$31.6$~deg$^{-2}$,][]{mauch03}, the number density of our clusters, and our association radius, we would expect to find $\sim 5$ clusters randomly overlapping with point sources in the $6~\sdeg$ survey; there is a $3\times10^{-4}$~per cent chance of explaining the associations as random superpositions.
Thus, our small sample provides clear evidence of physical associations between low frequency radio point sources and X-ray selected groups and clusters; this is consistent with previous findings \cite[]{best05,lin07} that low frequency radio sources are associated with cluster galaxies in both optically and X-ray selected cluster samples. As expected, given the tendency for radio galaxies to have steeply falling spectra as a function of frequency, only a small fraction (3 out of 19) of these low frequency radio galaxies are detectable at \ac{spt} frequencies.
We use the \ac{bcs} data \citep{desai12} to examine the optical counterparts of the six \ac{spt} point sources that lie within 4~arcmin of our X-ray selected group and cluster sample. We do this by first associating the \ac{spt} point sources with a \ac{sumss} source, which in general is only possible for the radio galaxies and not the dusty galaxies \citep{vieira10}. For our sample, three of the \ac{spt} point sources within 4~arcmin of the X-ray selected groups and clusters have \ac{sumss} counterparts. All three of these have strongly negative $\xi_{\text X}$ (see \Fref{fig:xbcs-corrbias}). For two of the three point sources, the optical counterpart is the group \acs{bcg}. In the third case the \ac{spt} point source corresponds to a quasar candidate \cite[MRC~2319-550;][]{wright90} and does not appear to be a cluster member. The three remaining \ac{spt} point sources do not have \ac{sumss} counterparts and are likely dusty galaxies; the \ac{sze} signatures $\xi_{\text X}$ of those systems are not obviously impacted. Thus we confirm that in two of our 46 low mass systems there are associated radio galaxies bright enough to be detected at \ac{spt} frequencies.
Based on the prediction from \cite{lin09}, we would have expected that radio sources completely fill in the $Y_\mathrm{SZ}$ signal (100~per cent contamination) at a redshift of 0.1 (or a redshift of 0.6) in approximately 2.5 (or 0.5) percent of clusters with similar mass ($M_{200}=10^{14}\msun$). For our 46 cluster sample, we would have expected this to happen for 1.15 (or 0.23) clusters, consistent with the two clusters we find associated with radio galaxies detected as point sources by \ac{sptsz}. We also expect a 20~per cent level $Y_\text{SZ}$ contamination on 9 (2)~per cent of the sample. This predicted contamination is significantly smaller than our current uncertainties on the $Y_\mathrm{SZ}$ normalisation, and therefore cannot be tested in this analysis.
We repeat the \ac{sze}-mass relation analysis while excluding the half of the clusters with \ac{sumss} point source associations. We find that the results are qualitatively similar using either the SPT-NPS or SPT-No-SUMSS sample (see Tables~\ref{tab:xbcs-constraint} and \ref{tab:xbcs-yszm}), although the uncertainties increase; this is consistent with the expectation that the level of the effect is too small to be measured with our sample. As already shown in Tables~\ref{tab:xbcs-constraint} and \ref{tab:xbcs-yszm}, our analysis shows no statistically significant difference in the \ac{sze}-mass relations when excluding or including the systems with nearby \ac{spt} point sources.
As pointed out in \Fref{sec:xbcs-yszm}, the dusty star-forming galaxies would have a net negative biasing impact on the \ac{sptsz} measurement. We examine the contamination from the dusty galaxies, which are not bright enough to be directly detectable in the 150~GHz and 95~GHz bands. To do this we measure the specific intensities at 220~GHz in a single frequency adaptive filter that uses cluster profiles at the locations of our X-ray selected cluster sample. In the SPT-NPS sample, the evidence for dusty galaxies is significant at the 2.8\,$\sigma$ level. We then convert the 220~GHz intensities to temperature fluctuations at 150~GHz and 95~GHz by assuming the intensity follows $I\propto \nu^{3.6}$ for dusty sources \cite[]{shirokoff11}. These are then converted to the corresponding values of $Y_\text{500}$. Dividing then by the expected $Y_\text{500}$ for a cluster of this redshift and X-ray luminosity, we then estimate the inverse variance weighted mean contamination to be $32\pm18$~per cent and $7\pm4$~per cent at 150~GHz and 95~GHz, respectively. Together, this contamination would lead the \ac{sptsz} observed $Y_\text{500}$ signature to be biased low by $\sim(17\pm9)$~per cent. This fractional contamination depends on the mass and redshift of the cluster together with the typical star formation activity. In particular, as a function of mass the SZE signature grows as $Y_\text{500}\propto M^{5/3}$, whereas the blue or star forming component of the galaxy population falls \citep[e.g.][]{weinmann06}; thus, contamination would fall with mass. As one pushes to even higher redshift than this sample (i.e. $z>1$) where star formation is more prevalent, the contamination would be expected to increase.
This level of contamination is consistent with a recent study of $\sim 550$ galaxy clusters selected via optical red-sequence techniques. Using \emph{Herschel} and \ac{spt} mm-wave data to jointly fit an \ac{sze}+dust spectral model, \cite{bleem13} finds the contamination at 150~GHz to be $40 \pm 30$~per cent for low-richness optical groups ($M_{\textrm{200}} \sim 1 \times 10^{14}\msun$). The fractional contamination declines as a function of optical richness and is measured to be $5 \pm 5$~per cent for the richest 3 per cent of clusters in the sample sample ($M_{\textrm{200}} \sim 3 \text{--} 6 \times 10^{14}\msun$). A larger sample size combined with deeper mm-wave data will improve our ability to estimate the contamination from dusty galaxies in clusters and groups.
In summary, this small sample of 46 X-ray selected groups and low mass clusters provides high significance evidence of having physically associated low frequency SUMSS radio galaxies. For the \ac{spt} point source sample within 2~arcmin, there is less than $2\,\sigma$ statistical evidence of physical association, but two of the sources have optical counterparts that are in the groups. Although we would expect physically associated high frequency radio galaxies to bias the \ac{sze} mass-observable relation, our analysis provides no evidence of this impact. We use the 220~GHz \ac{sptsz} data in this sample to estimate that the $Y_\text{500}$ measured by the \ac{spt} is biased $\sim 17\pm 9$~per cent low. A larger sample from a broader survey (through XMM-XXL or eROSITA, for example) or a deeper \ac{sze} survey would both help to improve our understanding of the impact of point sources.
\section{Conclusions}\label{sec:xbcs-conclusions}
Using data from the \ac{sptsz} survey, we have explored the \ac{sze} signatures of low mass clusters and groups selected from a uniform {\it XMM-Newton} X-ray survey. The cluster and group sample from the \ac{xbcs} has a well understood selection, and previously published calibrations of the $L_{\text X}$-mass relation allow us to estimate the masses of each of these systems. Although these systems have masses that are too low for them to have been individually detected within the \ac{sptsz} survey, we are able to use the ensemble to constrain the underlying relationship between the halo mass and the \ac{sze} signature for low mass systems.
Our method corrects for the Eddington bias and shows that there is no Malmquist like bias effect on the \ac{sze} mass-observable relation within this X-ray selected sample. We test our likelihood using a large mock sample, and we show with the current sample size we can at most extract constraints from two scaling relation parameters: the power law amplitude $A_{\text{SZ}}$ and slope $B_{\text{SZ}}$ (see Equations~\ref{eq:zetam} and \ref{eq:xbcs-ym}).
We separate the sample of 46 groups and clusters into three subsamples: (1) the full sample, (2) the point source-free sample, for which we exclude systems with point sources detected at significance $>$4 at either 95, 150, or 220~GHz in the \ac{sptsz} data within 4~arcmin radius of the X-ray centre, and (3) the point source-free sample, with clusters at $z<0.3$ excluded. We find that, due to the point source contamination in three of the lowest $\xi_{\text X}$ groups, the full sample exhibits a steep slope ($B_{\text{SZ}}=2.80^{+0.66}_{-0.63}$) that is in tension at 2.6$\,\sigma$ with the high mass \ac{spt} sample ($B_{\text{SZ}}=1.40\pm0.16$). The point source free subsample has a slope ($B_{\text{SZ}}=2.14^{+0.86}_{-0.66}$) that is in rough agreement with the slope of the high mass \ac{spt} sample ($1.4\sigma$ difference). We find no evidence that the low redshift clusters deviate from the scaling relation of the point source free sample.
We also measure the $Y_\text{500}$-mass relation for our sample and compare it to the results from the \ac{sptsz} high mass clusters and the {\it Planck}\ sample. Our low mass sample exhibits a preference for lower normalisation and steeper slope than the other two samples, but the uncertainties are large (see \Fref{fig:xbcs-ym} and \Fref{tab:xbcs-yszm}). Within the \ac{spt} samples, there is no statistically significant evidence for differences in the scaling relation as one moves from high to low masses. On the other hand, the {\it Planck}\ sample exhibits a 2.8\,$\sigma$ significant tension with our sample. As shown in \Fref{fig:xbcs-cmpplanck}, the lowest X-ray luminosity portion of our sample has lower $Y_\text{500}$ than expected from the {\it Planck}\ relation. We discuss a range of possible explanations for this tension (Section~\ref{sec:xbcs-yszm}), in particular contamination from dusty sources. Given the significance level of the tension the appropriate next step is to enlarge the sample to better quantify the differences in the \ac{sze} signatures of low and high mass clusters and the possible differences between {\it Planck}\ and \ac{spt}.
We examine radio point source contamination. Cross-matching our X-ray selected groups and clusters with the \ac{sumss} catalog, we find that 18 of 46 members have associated 843~MHz \ac{sumss} point sources within 2~arcmin. This represents highly significant evidence of physical association between our sample and low frequency point sources. At higher frequencies, we find four systems with associated \ac{spt} detected point sources; three of these also have SUMSS counterparts. Two of these three point sources have optical counterparts that lie within the X-ray group, and the third is a quasar candidate that is likely unassociated with the group. Having two out of 46 groups or clusters with physically associated bright, high frequency point sources is consistent with the expectations from \citet{lin09}. The predicted contamination from undetected radio point sources \citep[]{lin07,lin09} in the remainder of the sample is significantly smaller than our measurement uncertainty on the $Y_\text{500}$ normalisation, and so we cannot test these predictions here.
We also examine the impact of undetected dusty galaxies. Using the \ac{sptsz} 220~GHz band, we find 2.8$\,\sigma$ significant evidence of a flux excess due to dusty galaxies. Extrapolating to lower frequencies, we estimate that the measured $Y_\text{500}$ signature is biased low by $\sim(17\pm9)$~per cent in this ensemble of low mass clusters and groups. Given the different frequency coverage of {\it Planck}\ and \ac{spt}, it is not clear that the {\it Planck}\ bias due to dusty galaxy flux would be the same. If flux from dusty galaxies would induce a smaller negative bias or even a positive bias in {\it Planck}\ $Y_\text{500}$ measurements, then that would reduce the tension between the {\it Planck}\ $Y_\text{500}$--mass relation and ours.
We point out that these contamination levels are for this X-ray selected low mass sample with a median mass of $10^{14}\msun$, which is significantly below the typical mass of \ac{spt} selected clusters. Given the increasing rarity of blue and star forming galaxies as one moves from groups to high mass clusters \citep[e.g][]{weinmann06}, any contamination in the \ac{spt} selected sample would be much lower.
Finally, the receiver on the \ac{spt} was upgraded in 2012. The SPTpol camera provides sensitivity to \ac{cmb} polarization and, more importantly for \ac{sze} work, increased sensitivity to CMB temperature fluctuations. The final SPTpol maps are expected to cover 500 square degrees of sky to noise levels of $\sim 5$ and $\sim 9 \mr{\mu K-arcmin}$ at 150 and 95~GHz \citep{austermann12}. Meanwhile, the XXL survey \cite[]{pierre11} has increased the survey area that has a characteristic 10~ks {\it XMM-Newton} exposure from $6~\sdeg$ to $25~\sdeg$. This should enable an interesting new insight into possible differences in the \ac{sze} signatures of low and high mass clusters. We make a forecast with a mock catalog that consists of 144 clusters within redshift range 0.2--1.2 and a bolometric flux limit of $1\times10^{-14}~\text{erg s}^{-1}\text{cm}^{-2}$. Analysing this sample with the appropriate SPTpol increase in depth indicates that with the future sample we can tighten the fractional error on $A_{\text{SZ}}$ to 6~per cent compared to our current result of 30~per cent. On $B_{\text{SZ}}$ the uncertainty shrinks from 34 to 8~per cent. These improvements should enable a more revealing comparison of the \ac{sze} signatures of low and high mass clusters and perhaps also enable a detailed study of potential contamination of the \ac{sze} signal by associated radio or dusty galaxies.
%
\section*{Acknowledgments}
We acknowledge the support of the DFG through TR33 ``The Dark Universe'' and the Cluster of Excellence ``Origin and Structure of the Universe''. Some calculations have been carried out on the computing facilities of the Computational Center for Particle and Astrophysics (C2PAP). The South Pole Telescope is supported by the National Science Foundation through grant PLR-1248097. Partial support is also provided by the NSF Physics Frontier Center grant PHY-1125897 to the Kavli Institute of Cosmological Physics at the University of Chicago, the Kavli Foundation and the Gordon and Betty Moore Foundation grant GBMF 947. This work is also supported by the U.S. Department of Energy. Galaxy cluster research at Harvard is supported by NSF grants AST-1009012 and DGE-1144152. Galaxy cluster research at SAO is supported in part by NSF grants AST-1009649 and MRI-0723073. The McGill group acknowledges funding from the National Sciences and Engineering Research Council of Canada, Canada Research Chairs program, and the Canadian Institute for Advanced Research.
\bibliographystyle{mn2e} |
\section{Introduction}
Graphene is a two dimensional hexagonal lattice of carbon atoms and is one
of the most important topics in solid state physics due to the vast
application in nano-electronics, opto-electronics, superconductivity and
Josephson junctions (\cite{novo},\cite{intro1},\cite{intro2},\cite{B} and
\cite{BBBB}). The band structure shows that the conduction and valence band
touch at the Dirac point and the dispersion relation is approximately linear
and isotropic \cite{A}. This linear dispersion near the symmetry points have
striking similarities with those of massless relativistic Dirac fermions
\cite{B}. This leads to a number of fascinating phenomena such as the
half-quantized Hall effect (\cite{C},\cite{D}) and minimum quantum
conductivity in the limit of vanishing concentration of charge carriers \cit
{novo}. Although this is an outstanding experimental result, there is no
consensus about the theoretical value computed through different theoretical
methods (see \cite{Ziegler}), neither the physical reason for such minimum
value (see \cite{yuriv}), where the minimum is due to the impurity resonance
and is not related to the Dirac point.
In particular, one of the theoretical methods used to compute response
functions within the linear response theory is the Kubo formalism \cite{kubo
. Deviations of charge and current densities from their equilibrium values
are described by density and current response functions through Kubo
formulas using the same two-particle Green function. Although, generally is
unable to obtain exact relations between these response functions, several
approximations can be obtained taking into account the dimensionality and
dispersion relation of the system (see \cite{Janis}). But these approximate
relations are based on the continuity equation and Ward identities and is
not clear if these assumptions are valid for linear dispersion relations and
spinor wave functions.
In turn, impurities in graphene can be considered in various type of forms:
substitutional, where the site energy is different from those of carbon
atoms, which originates resonances \cite{mahan} and as adsorbates, that can
be placed on various points in graphene; sixfold hollow site of a honeycomb
lattice, twofold bridge site of two neighboring carbons or top site of a
carbon atom \cite{roten}. Theoretical as well as experimental studies have
indicated that substitutional doping of carbon materials can be used to
tailor their physical and/or chemical properties (\cite{strobel}, \cit
{jiang}, \cite{sankaran}). In particular, nitrogen or boron dopants can be
added to pristine graphene (\cite{11},\cite{12},\cite{13},\cite{14}).
The detection and absorption of low levels of hydrogen becomes very
important for sensor gas and hydrogen energy. Different methods of hydrogen
detection are not entirely selective or it have a high cost of manufacture
due to their complexity. Pd-doped reduced graphene have a clear response to
hydrogen and are very selective (\cite{pandey}, \cite{corral}). In the other
side, the decoration of carbon support by transition metals can also be
independently used to enhance the hydrogen storage of the specimens.
Transition metals eliminate the hydrogen dissociation barrier altogether
\cite{adams}.
In this sense, the density and current response function of doped-graphene
with low concentration of subtitutional impurities is of major importance
for the consequences in the sensor effect (\cite{sch}, \cite{bar}). In
particular, the simplest graphene-based sensor detects the conductivity
change upon adsorption of analyte molecules. The change of conductivity
could be attributed to the changes of charge carrier concentration in the
graphene induced by adsorbed gas molecules. It has been proposed that such
device may be capable of detecting individual molecule \cite{kong}. These
reactions release captured electrons in the interaction zone between the gas
and the sensor, and increase their concentration in the conductivity zone.
But the conductivity of electrons are based on the diffusion phenomena of
charge carriers through the sample. Electrons moving in randomly distributed
scatterers has a diffusive character, which is described at long distances
by a diffusion equation. It has been shown that it is possible to supress
diffusion (see \cite{Anderson}), giving rise to a localization phenomena,
which will affect the sensor characteristics of the material. In turn, a
dynamical generalization of the diffusion constant from the electron-hole
correlation function cannot be linked to the frequency dependent
conductivity (see eq.(3.18) and eq.(3.19) of \cite{Janis}). In this sense,
the aim of this work is two-fold: to introduce a generalization of the
linear response theory for linear dispersion relation and spinor wave
functions, to apply it to graphene, and the subsequent computation of
minimal conductivity and dynamical diffusion, to analize the general
behavior of the system under local perturbations and the implications for
sensor gas.
This work will be organized as follow: In section II, the impurity averaged
Green function will be computed. In section III and IV, a generalization of
the conductivity tensor and response function will be computed using the
current definition for relativistic Dirac fermions. In section V, different
limit behavior of the current and density response functions are computed.
The Boltzmann limit is introducedshowing the minimal conductivity value. In
section VI, the dynamical diffusion will be computed through the relaxation
function, showing how the obtain the full renormalization group equation for
the Fermi velocity. Finally, the conclusion are presented. Appendix A and B
are introduced for self-contained lecture.
\section{Impurity averaged Green function}
The Hamiltonian of clean graphene in the $K$ point in the Brillouin zone and
in the long wavelength approximation reads (see \cite{B}
\begin{equation}
H=v_{F}\left(
\begin{array}{cc}
0 & p_{x}+ip_{y} \\
p_{x}-ip_{y} &
\end{array
\right) \label{h1}
\end{equation
where $v_{F}\sim 10^{6}m/s$ is the Fermi velocity. The eigenfunctions of
this Hamiltonian read
\begin{equation}
\psi _{\mathbf{k}}(\mathbf{r})=\frac{1}{\sqrt{2}}\left(
\begin{array}{c}
1 \\
\lambda e^{i\varphi _{\mathbf{k}}
\end{array
\right) e^{i\mathbf{k}\cdot \mathbf{r}} \label{h2}
\end{equation
wher
\begin{equation}
\varphi _{\mathbf{k}}=\arctan (\frac{k_{y}}{k_{x}}) \label{h3}
\end{equation
In turn, the eigenvalues read
\begin{equation}
E_{\lambda }(\mathbf{k})=\lambda v_{f}\hbar k \label{h4}
\end{equation
where $k=\left\vert \mathbf{k}\right\vert $ and where $\lambda =1$ are
positive energy states (conduction band) and $\lambda =-1$ are negative
energy states (valence band).With the eigenfunctions of eq.(\ref{h2}) we can
compute the retarded and advanced Green function for conduction electrons
(\lambda =1)$ in momentum space\footnote
We assume that the valence-band states do not contribute to low temperature
conductivity.}
\begin{equation}
G_{0}^{R(A)}(\mathbf{q},E)=\frac{1}{E-v_{f}\hbar q\mp is}\left(
\begin{array}{cc}
1 & e^{i\varphi _{\mathbf{q}}} \\
e^{-i\varphi _{\mathbf{q}}} &
\end{array
\right) \label{h7}
\end{equation
where the minus sign correspond to the retarded Green function and the plus
sign to the advanced Green function. The contribution to second order in the
perturbation expansion in the impurity potential reads (see \cite{Rammer},
eq.(3.31), page 136
\begin{equation}
\left\langle G_{2}^{R}(\mathbf{k},\mathbf{k}^{\prime },E)\right\rangle
^{i=j}=\delta _{\mathbf{k,k}^{\prime }}n_{i}[G_{0}^{R}(\mathbf{k
,E)]^{2}\int \frac{d^{2}\mathbf{k}^{\prime }}{(2\pi )^{2}}\left\vert V_{imp}
\mathbf{k}-\mathbf{k}^{\prime })\right\vert ^{2}IG_{0}^{R}(\mathbf{k
^{\prime },E) \label{h8}
\end{equation
where $n_{i}$ is the impurity concentration, $\left\vert V_{imp}(\mathbf{k}
\mathbf{k}^{\prime })\right\vert ^{2}I$ is a diagonal matrix
\begin{equation}
\left\vert V_{imp}(\mathbf{k}-\mathbf{k}^{\prime })\right\vert ^{2}I=\left(
\begin{array}{cc}
\left\vert V_{imp}(\mathbf{k}-\mathbf{k}^{\prime })\right\vert ^{2} & 0 \\
0 & \left\vert V_{imp}(\mathbf{k}-\mathbf{k}^{\prime })\right\vert ^{2
\end{array
\right) \label{h9}
\end{equation
and the angle brackets represent the configurational averaging that can be
computed a
\begin{equation}
\left\langle A\right\rangle =\int \overset{N}{\underset{i=1}{\prod }}
\mathbf{r_{i}}A(\mathbf{r_{1},r_{2},...,r_{N}})P(\mathbf
r_{1},r_{2},...,r_{N}}) \label{h9.1}
\end{equation
where $P(\mathbf{r_{1},r_{2},...,r_{N}})=P(\mathbf{r_{1}})P(\mathbf{r_{2}
)...P(\mathbf{r_{N}})$ and $P(\mathbf{r_{i}})$ is the probability density
for having the impurity located around point $\mathbf{r_{i}}$.\footnote
In this case we are assuming that the positions of the impurities are
distributed independently.}\ In eq.(\ref{h8}), the Fourier transform of
G_{2}^{R}(\mathbf{r},\mathbf{r}^{\prime },E)~$has been taken first.
Replacing last equation and eq.(\ref{h7}) in eq.(\ref{h8}) the diagonal part
of the averaged Green function read
\begin{gather}
\int \frac{d^{2}\mathbf{k}^{\prime }}{(2\pi )^{2}}\frac{\left\vert V_{imp}
\mathbf{k}-\mathbf{k}^{\prime })\right\vert ^{2}}{E-v_{f}\hbar k^{\prime }-i
}= \label{h11} \\
\int \frac{d^{2}\mathbf{k}^{\prime }}{(2\pi )^{2}}\left\vert V_{imp}(\mathbf
k}-\mathbf{k}^{\prime })\right\vert ^{2}\left( \frac{E-v_{f}\hbar k^{\prime
}{(E-v_{f}\hbar k^{\prime })^{2}+s^{2}}+i\frac{s}{(E-v_{f}\hbar k^{\prime
})^{2}+s^{2}}\right) \notag
\end{gather
If we consider for simplicity that the impurity potential is a Dirac delta
potential, then\footnote
In this case, the disorder introduced by the delta Dirac impurity potential
is an on-site diagonal disorder.
\begin{equation}
V_{imp}(\mathbf{k})=\int d^{2}\mathbf{r}e^{-i\mathbf{r}\cdot \mathbf{q
}V_{imp}(\mathbf{r})=\int d^{2}\mathbf{r}e^{-i\mathbf{r}\cdot \mathbf{q
}V_{0}\delta (\mathbf{r})=V_{0} \label{h11.1}
\end{equation
Using the last result, the integral of eq.(\ref{h11}) read
\begin{gather}
V_{0}^{2}\underset{s\rightarrow 0}{\lim }\int \frac{d^{2}\mathbf{k}^{\prime
}{(2\pi )^{2}}\left( \frac{E-v_{f}\hbar k^{\prime }}{(E-v_{f}\hbar k^{\prime
})^{2}+s^{2}}+i\frac{s}{(E-v_{f}\hbar k^{\prime })^{2}+s^{2}}\right) =
\label{h13} \\
i\pi V_{0}^{2}\int \frac{d^{2}\mathbf{k}^{\prime }}{(2\pi )^{2}}\delta
(E-v_{f}\hbar k^{\prime })=i\pi V_{0}^{2}n(E_{F}) \notag
\end{gather
where we have used that $\delta (x)=\frac{1}{\pi }\underset{s\rightarrow 0}
\lim }\frac{s}{x^{2}+s^{2}}$ and $n(E)$ is the density of states at the
Fermi energy.\footnote
In last equation the real part of is strictly not zero, but is a constant
that do not depends on the momentum. In this sense, this value is arbitrary
and has no observable consequences. For this we can assume that is zero or
redefine the reference for measuring energy.} At this point is important to
notice that in clean graphene, the density of states $n(E)$ at the Fermi
energy is $n(E_{F})=0$ (see \cite{Peres}, eq.(33)). Nevertheless, when
impurities are introduced, the density of states at the Fermi energy is not
zero (see \cite{Peres}, figure 3), which implies that disorder introduce an
imaginary term to the self-energy.
The non-diagonal term read
\begin{gather}
\int \frac{d^{2}\mathbf{k}^{\prime }}{(2\pi )^{2}}\frac{V_{0}^{2}}
E-v_{f}\hbar k^{\prime }-is}\frac{k_{x}^{\prime }+ik_{y}^{\prime }}
k^{\prime }} \label{h14} \\
\int \frac{d^{2}\mathbf{k}^{\prime }}{(2\pi )^{2}}\frac{V_{0}^{2}}{k^{\prime
}}\left( \frac{k_{x}^{\prime }(E-v_{f}\hbar k^{\prime })-k_{y}^{\prime }s}
(E-v_{f}\hbar k^{\prime })^{2}+s^{2}}+i\frac{sk_{x}^{\prime }+k_{y}^{\prime
}(E-v_{f}\hbar k^{\prime })}{(E-v_{f}\hbar k^{\prime })^{2}+s^{2}}\right)
\notag
\end{gather
Introducing polar coordinates in the wave vector $\mathbf{k}^{\prime }$,
k_{x}^{\prime }=k^{\prime }\cos \lambda $ and $k_{y}^{\prime }=k^{\prime
}\sin \lambda $, is not difficult to show the last integral is zero due to
the cosine and sine functions, which are integrated between $0$ and $2\pi $.
Then, the averaged Green function at second order in the perturbation
expansion in the impurity potential reads
\begin{equation}
\left\langle G_{2}^{R}(\mathbf{k},\mathbf{k}^{\prime },E)\right\rangle
^{i=j}=\delta _{\mathbf{k,k}^{\prime }}[G_{0}^{R}(\mathbf{k},E)]^{2}\left(
\begin{array}{cc}
i\eta & 0 \\
0 & i\eta
\end{array
\right) \label{h16}
\end{equation
where $\eta =\pi n_{i}V_{0}^{2}n(E)$. By introducing the one-particle
irreducible propagator, which correspond to all the diagrams which cannot be
cut in two by cutting an internal line, the impurity averaged propagator can
be written as a geometric series in terms of the self-energy (see \cit
{Rammer}, page 141
\begin{equation}
G=G_{0}+G_{0}\Sigma G_{0}+...=G_{0}\sum\limits_{n=0}^{+\infty }(\Sigma
G_{0})^{n}=G_{0}\left( I-\Sigma G_{0}\right) ^{-1} \label{h16.2}
\end{equation
where $\Sigma =\Sigma _{1}^{R}+\Sigma _{2}^{R}+...$ contains the
contributions at different orders in the perturbation expansion of the
impurity concentration. With the computation done in eq.(\ref{h16}) we
finally obtai
\begin{equation}
G^{R(A)}(\mathbf{q})=G_{0}\left( I-\Sigma G_{0}\right) ^{-1}=\frac{1}
E-v_{f}\hbar q\mp is-i\eta }\left(
\begin{array}{cc}
1 & e^{i\varphi _{\mathbf{q}}} \\
e^{-i\varphi _{\mathbf{q}}} &
\end{array
\right) \label{h16.5}
\end{equation
This last result is the impurity averaged Green function which take into
account the first contribution of the self-energy by comparing last equation
with eq.(\ref{h7}). This is known as the full Born approximation, which
include electronic scattering from a single impurity. The diagonal part
contains the shifted pole due to the imaginary part of the self energy. The
non-diagonal part contains the same contribution multiplied by a phase
factor. The last result will be used in the following sections.
\section{Current response function}
In this section, a generalization of the conducitivity tensor for Dirac
fermion systems, that is, linear dispersion relation and spinors wave
functions, will be introduced. To do it we will follow the development
introduced in \cite{Rammer} and by taking into account the differences
introduced by Dirac systems. The Hamiltonian of Bloch electrons in the long
wavelength approximation in a electric field and random impurities read
\begin{equation}
H=v_{f}\mathbf{\sigma }\cdot (\mathbf{p}-e\mathbf{A})+V_{imp}(\mathbf{r})
\label{con1}
\end{equation
where $\mathbf{A}(\mathbf{r})$ is the vector potential that is related to
the electric field a
\begin{equation}
\mathbf{E}=-\frac{\partial \mathbf{A}}{\partial t} \label{con2}
\end{equation
and where $V_{imp}(\mathbf{r})$ is the impurity field. We can compute the
current density to linear order in the external electric field (see
eq.(7.84) of \cite{Rammer})
\begin{equation}
\mathbf{j}(\mathbf{r},t)=Tr(\rho _{0}(t)\mathbf{j})-\frac{i}{\hbar
\int_{t_{i}}^{t}d\overline{t}Tr(\rho _{0}(t_{r})[\mathbf{j}_{p}(\mathbf{r
,t),H_{A}(\overline{t})]+O(E^{2}) \label{con9}
\end{equation
where the charge current density operator can be written a
\begin{equation}
\mathbf{j}_{p}=v_{F}\left\vert \mathbf{r}\right\rangle \mathbf{\sigma
\left\langle \mathbf{r}\right\vert \label{con5.1}
\end{equation
which is the usual definition of current in relativistic Dirac system, where
$v_{F}$ plays the role of velocity of light and
\begin{equation}
H_{A}(t)=ev_{f}\mathbf{\sigma }\cdot \mathbf{A} \label{con5.2}
\end{equation
Taking into account the direction of the current in index notation and to
linear order in the electric field we obtai
\begin{equation}
j_{\alpha }(\mathbf{r},t)=\left\langle j_{\alpha }(\mathbf{r
,t)\right\rangle _{0}+\sum\limits_{\beta }^{{}}\int d\mathbf{r}^{\prime
}\int_{t}^{+\infty }Q_{\alpha \beta }(\mathbf{r},t;\mathbf{r}^{\prime
},t^{\prime })A_{\beta }(\mathbf{r}^{\prime },t^{\prime }) \label{con10}
\end{equation
where $Q_{\alpha \beta }$ is the current response function. Taking into
accout that in linear response, each frequency contributes additively, only
is necesary to study what happens at one driving frequenc
\begin{equation}
\mathbf{A}(\mathbf{r},t)=\mathbf{A}(\mathbf{r},\omega )e^{-i\omega t}
\label{con13}
\end{equation
Then, the Fourier transform of the current read
\begin{equation}
j_{\alpha }(\mathbf{r},\omega )=\left\langle j_{\alpha }(\mathbf{r},\omega
)\right\rangle _{0}+\sum\limits_{\beta }^{{}}\int d\mathbf{r}^{\prime
}Q_{\alpha \beta }(\mathbf{r},\mathbf{r}^{\prime },\omega )A_{\beta }
\mathbf{r}^{\prime },\omega ) \label{con14}
\end{equation
wher
\begin{equation}
Q_{\alpha \beta }(\mathbf{r},\mathbf{r}^{\prime },\omega )=K_{\alpha \beta }
\mathbf{r},\mathbf{r}^{\prime },\omega )-K_{\alpha \beta }(\mathbf{r}
\mathbf{r}^{\prime },0) \label{con15}
\end{equation
an
\begin{equation}
K_{\alpha \beta }(\mathbf{r},\mathbf{r}^{\prime },\omega
)=\sum\limits_{\lambda \lambda ^{\prime }}^{{}}\frac{\rho _{\lambda }-\rho
_{\lambda ^{\prime }}}{\epsilon _{\lambda }-\epsilon _{\lambda ^{\prime
}}+\hbar \omega +is}\left\langle \lambda \left\vert j_{\alpha }^{p}(\mathbf{
})\right\vert \lambda ^{\prime }\right\rangle \left\langle \lambda ^{\prime
}\left\vert j_{\beta }^{p}(\mathbf{r}^{\prime })\right\vert \lambda
\right\rangle \label{con16}
\end{equation
where $\left\vert \lambda \right\rangle $ are eigenstates of unperturbed
Hamiltonian and $\rho _{\lambda }$ is the mean ocuppation number for a
energy level $\epsilon _{\lambda }$. At this point, if we use the usual
definition of current in non-relativistic quantum mechanic
\begin{equation}
\mathbf{j}=\frac{e}{2m}\{\mathbf{P},\left\vert \mathbf{r}\right\rangle
\left\langle \mathbf{r}\right\vert \} \label{con17}
\end{equation
then
\begin{equation}
\left\langle \lambda \left\vert j_{\alpha }^{p}(\mathbf{r})\right\vert
\lambda ^{\prime }\right\rangle =-\frac{i\hbar e}{2m}[\psi _{\lambda
^{\prime }}^{\ast }(\mathbf{r})\overrightarrow{\nabla }\psi _{\lambda }
\mathbf{r})-\psi _{\lambda }(\mathbf{r})\overrightarrow{\nabla }\psi
_{\lambda ^{\prime }}^{\ast }(\mathbf{r})] \label{con18}
\end{equation
In the same line of thought, we can use the definition of relativistic Dirac
current, the
\begin{equation}
v_{F}\left\langle \lambda \mid \mathbf{r}\right\rangle \mathbf{\sigma
\left\langle \mathbf{r}\mid \lambda ^{\prime }\right\rangle =v_{F}\psi
_{\lambda }^{\dag }(\mathbf{r})\mathbf{\sigma }\psi _{\lambda ^{\prime }}
\mathbf{r}) \label{con19}
\end{equation
and in the same wa
\begin{equation}
\left\langle \lambda ^{\prime }\left\vert j_{\beta }^{p}(\mathbf{r
)\right\vert \lambda \right\rangle =v_{F}\psi _{\lambda ^{\prime }}^{\dag }
\mathbf{r}^{\prime })\mathbf{\sigma }\psi _{\lambda }(\mathbf{r}^{\prime })
\label{con20}
\end{equation
Introducing eq.(\ref{con19}) and eq.(\ref{con20}) into eq.(\ref{con16}) and
writing in index notation which allows to move the functions $\psi $ and the
Pauli matrices we hav
\begin{gather}
K_{\alpha \beta }(\mathbf{r},\mathbf{r}^{\prime },\omega
)=e^{2}v_{F}^{2}\int \frac{dE}{2\pi }\int \frac{dE^{\prime }}{2\pi }\frac
\rho (E)-\rho (E^{\prime })}{E-E^{\prime }+\hbar \omega +is}\psi
_{l}^{\lambda }(\mathbf{r}^{\prime })\psi _{i}^{\lambda \ast }(\mathbf{r
)\sigma _{ij}^{\alpha }\sigma _{kl}^{\beta }\psi _{j}^{\lambda ^{\prime }}
\mathbf{r})\psi _{k}^{\lambda ^{\prime }\ast }(\mathbf{r}^{\prime })=
\label{con20.2} \\
e^{2}v_{F}^{2}\int \frac{dE}{2\pi }\int \frac{dE^{\prime }}{2\pi }\frac{\rho
(E)-\rho (E^{\prime })}{E-E^{\prime }+\hbar \omega +is}A_{li}(\mathbf{r
^{\prime },\mathbf{r},E)\sigma _{ij}^{\alpha }\sigma _{kl}^{\beta }A_{jk}
\mathbf{r},\mathbf{r}^{\prime },E^{\prime }) \notag
\end{gather
where we have used the relation between the spectral weigth $A(\mathbf{r}
\mathbf{r}^{\prime },E)$ and the wave functions (see Appendix A, eq.(\re
{con21}) and eq.(\ref{con22})). Finally, applying the relation between the
spectral weight and the Green function we obtai
\begin{gather}
K_{\alpha \beta }(\mathbf{r},\mathbf{r}^{\prime },\omega
)=-e^{2}v_{F}^{2}\int \frac{dE}{2\pi }\int \frac{dE^{\prime }}{2\pi }\frac
\rho (E)-\rho (E^{\prime })}{E^{\prime }-E+\hbar \omega +is}\times
\label{con20.3} \\
\left[ G_{li}^{R}(\mathbf{r}^{\prime },\mathbf{r},E)-G_{li}^{A}(\mathbf{r
^{\prime },\mathbf{r},E)\right] \sigma _{ij}^{\alpha }\sigma _{kl}^{\beta
\left[ G_{jk}^{R}(\mathbf{r},\mathbf{r}^{\prime },E^{\prime })-G_{jk}^{A}
\mathbf{r},\mathbf{r}^{\prime },E^{\prime })\right] \notag
\end{gather
which is the desired generalization of the current response function for
Dirac fermion systems. In this case, the Pauli matrices play the rol of
momentum in eq.(7.96) of \cite{Rammer}. In the momentum space, the
current-current response function read
\begin{gather}
K_{\alpha \beta }(\mathbf{q},\mathbf{q}^{\prime },\omega
)=e^{2}v_{F}^{2}\int \frac{d^{2}\mathbf{k}}{(2\pi )^{2}}\int \frac{d^{2
\mathbf{k}^{\prime }}{(2\pi )^{2}}\int \frac{dE}{2\pi }\int \frac{dE^{\prime
}}{2\pi }\frac{\rho (E)-\rho (E^{\prime })}{E^{\prime }-E+\hbar \omega +is
\times \label{con20.6.1.0} \\
Tr(\left[ G^{R}(\mathbf{k}+\mathbf{q}^{\prime },\mathbf{k}^{\prime }+\mathbf
q},E)-G^{A}(\mathbf{k}+\mathbf{q}^{\prime },\mathbf{k}^{\prime }+\mathbf{q
,E)\right] \sigma ^{\alpha }\left[ G^{R}(\mathbf{k}^{\prime }-\mathbf{q}
\mathbf{k}-\mathbf{q}^{\prime },E^{\prime })-G^{A}(\mathbf{k}^{\prime }
\mathbf{q},\mathbf{k}-\mathbf{q}^{\prime },E^{\prime })\right] \sigma
^{\beta }) \notag
\end{gather
Introducing the impurity averaging of two Green function (see \cite{Rammer},
eq.(8.3)
\begin{gather}
\left\langle G^{R(A)}(\mathbf{k}+\mathbf{q}^{\prime },\mathbf{k}^{\prime }
\mathbf{q},E)G^{A(R)}(\mathbf{k}^{\prime }-\mathbf{q},\mathbf{k}-\mathbf{q
^{\prime },E^{\prime })\right\rangle = \label{con20.6.1.2} \\
\delta _{\mathbf{q},\mathbf{q}^{\prime }}\left\langle G^{R(A)}(\mathbf{k}
\mathbf{q},\mathbf{k}^{\prime }+\mathbf{q},E)G^{A(R)}(\mathbf{k}^{\prime }
\mathbf{q},\mathbf{k}-\mathbf{q},E^{\prime })\right\rangle \notag
\end{gather
computing one of the energy integration and exploting the analytical
properties of the averaged Green function we obtain (see \cite{Rammer}, page
283
\begin{equation}
K_{\alpha \beta }(\mathbf{q},\omega )=K_{\alpha \beta }^{AR}(\mathbf{q
,\omega )+K_{\alpha \beta }^{AA}(\mathbf{q},\omega )+K_{\alpha \beta }^{RR}
\mathbf{q},\omega ) \label{con20.6.1.3}
\end{equation
wher
\begin{equation}
K_{\alpha \beta }^{RA}(\mathbf{q},\omega )=e^{2}v_{F}^{2}\sigma
_{ij}^{\alpha }\sigma _{kl}^{\beta }\int \frac{dE}{2\pi i}[\rho (E+\hbar
\omega )-\rho (E)]\Phi _{lijk}^{AR}(E+\hbar \omega ,E,\mathbf{q})
\label{con20.6.1.4}
\end{equation
an
\begin{equation}
K_{\alpha \beta }^{AA}(\mathbf{q},\omega )=-e^{2}v_{F}^{2}\sigma
_{ij}^{\alpha }\sigma _{kl}^{\beta }\int \frac{dE}{2\pi i}\rho (E+\hbar
\omega )\Phi _{lijk}^{AA}(E+\hbar \omega ,E,\mathbf{q}) \label{con20.6.1.5}
\end{equation
\begin{equation}
K_{\alpha \beta }^{RR}(\mathbf{q},\omega )=e^{2}v_{F}^{2}\sigma
_{ij}^{\alpha }\sigma _{kl}^{\beta }\int \frac{dE}{2\pi i}\rho (E)\Phi
_{lijk}^{RR}(E+\hbar \omega ,E,\mathbf{q}) \label{con20.6.1.6}
\end{equation
where the electron (hole)-electron (hole) correlation function $\Phi
_{lijk}^{ab}$ read
\begin{equation}
\Phi _{lijk}^{ab}(E+\hbar \omega ,E,\mathbf{q})=\int \frac{d^{2}\mathbf{k}}
(2\pi )^{2}}\int \frac{d^{2}\mathbf{k}^{\prime }}{(2\pi )^{2}}\left\langle
G_{li}^{a}(\mathbf{k}+\mathbf{q},\mathbf{k}^{\prime }+\mathbf{q},E+\hbar
\omega )G_{jk}^{b}(\mathbf{k}^{\prime }-\mathbf{q},\mathbf{k}-\mathbf{q
,E)\right\rangle \label{con20.6.1.7}
\end{equation
The final conductivity tensor can be written in terms of the current
response function $K_{\alpha \beta }(\mathbf{q},\omega )$ (see \cite{Rammer
, eq.(8.51)) using the Kramer-Kronig relatio
\begin{equation}
\sigma _{\alpha \beta }(\mathbf{q},\omega )=\frac{K_{\alpha \beta }(\mathbf{
},\omega )-K_{\alpha \beta }(\mathbf{q},0)}{i\omega } \label{con20.8}
\end{equation
As we can see in eq.(\ref{con20.6.1.0}), we have the multiplication of one
Green matrix functions with the Pauli matrix in the $\alpha $ direction and
the other with the Pauli matrix in the $\beta $ directio
\begin{equation}
G_{li}^{R(A)}(\mathbf{k}+\mathbf{q},\mathbf{k}^{\prime }+\mathbf{q},E+\hbar
\omega )\sigma _{ij}^{\alpha }=M_{lj}^{R(A)}(\mathbf{k}+\mathbf{q},\mathbf{k
^{\prime }+\mathbf{q},E+\hbar \omega ) \label{con20.9}
\end{equation
\begin{equation}
G_{jk}^{R(A)}(\mathbf{k}^{\prime }-\mathbf{q},\mathbf{k}-\mathbf{q},E)\sigma
_{kl}^{\beta }=M_{jl}^{R(A)}(\mathbf{k}^{\prime }-\mathbf{q},\mathbf{k}
\mathbf{q},E) \label{con20.10}
\end{equation
The two possible Pauli matrices are $\sigma ^{x}$ and $\sigma ^{y}$ and in
particular if we choose the direction of the Pauli matrix in such a way that
$\sigma ^{\alpha }$ is $\sigma ^{\alpha }=\sigma ^{x}\cos \phi e_{x}+\sigma
^{y}\sin \phi e_{y}$ and $\sigma ^{\beta }=\sigma ^{x}\cos \theta
e_{x}+\sigma ^{y}\sin \theta e_{y}$, where $\phi $ and $\theta $ are angles
in real space, the
\begin{equation}
\sigma ^{\alpha }=\left(
\begin{array}{cc}
0 & e^{-i\phi } \\
e^{i\phi } &
\end{array
\right) \text{ \ \ \ \ \ \ \ }\sigma ^{\beta }=\left(
\begin{array}{cc}
0 & e^{-i\theta } \\
e^{i\theta } &
\end{array
\right) \label{con20.13}
\end{equation
At this point, we have to used the perturbation expansion of the product of
two matrix Green functions in the impurity concentration which has been
computed in last section.
\section{Density response function}
In a similar way, we can generalize the density response function $\chi
\mathbf{r},\mathbf{r}^{\prime },\omega )$ for linear dispersion and spinor
wave functions, which is defined as (see eq.(7.23) of \cite{Rammer}
\begin{equation}
\chi (\mathbf{r},\mathbf{r}^{\prime },\omega )=-\sum\limits_{\lambda \lambda
^{\prime }}^{{}}\frac{\rho _{\lambda }-\rho _{\lambda ^{\prime }}}{\epsilon
_{\lambda }-\epsilon _{\lambda ^{\prime }}+\hbar \omega +is}\left\langle
\lambda \mid \mathbf{r}\right\rangle \left\langle \mathbf{r}\mid \lambda
^{\prime }\right\rangle \left\langle \lambda ^{\prime }\mid \mathbf{r
^{\prime }\right\rangle \left\langle \mathbf{r}^{\prime }\mid \lambda
\right\rangle \label{dif1}
\end{equation
Using eq.(\ref{con22}) and computing the Fourier transform we obtain for the
density response functio
\begin{gather}
\chi (\mathbf{q},\mathbf{q}^{\prime },\omega )=\int \frac{d^{2}\mathbf{k}}
(2\pi )^{2}}\int \frac{d^{2}\mathbf{k}^{\prime }}{(2\pi )^{2}}\int \frac{dE}
2\pi }\int \frac{dE^{\prime }}{2\pi }\frac{\rho (E)-\rho (E^{\prime })}
E^{\prime }-E+\hbar \omega +is}\times \label{dif4} \\
\left[ G_{ij}^{R}(\mathbf{k}+\mathbf{q}^{\prime },\mathbf{k}^{\prime }
\mathbf{q},E)-G_{ij}^{A}(\mathbf{k}+\mathbf{q}^{\prime },\mathbf{k}^{\prime
}+\mathbf{q},E)\right] \left[ G_{ji}^{R}(\mathbf{k}^{\prime }-\mathbf{q}
\mathbf{k}-\mathbf{q}^{\prime },E^{\prime })-G_{ji}^{A}(\mathbf{k}^{\prime }
\mathbf{q},\mathbf{k}-\mathbf{q}^{\prime },E^{\prime })\right] \notag
\end{gather
Introducing the impurity averaging of two Green function
\begin{gather}
\chi (\mathbf{q},\omega )=\int \frac{d^{2}\mathbf{k}}{(2\pi )^{2}}\int \frac
d^{2}\mathbf{k}^{\prime }}{(2\pi )^{2}}\int \frac{dE}{2\pi }\int \frac
dE^{\prime }}{2\pi }\frac{\rho (E)-\rho (E^{\prime })}{E^{\prime }-E+\hbar
\omega +is}\times \label{dif4.0.1} \\
Tr(A(\mathbf{k}+\mathbf{q},\mathbf{k}^{\prime }+\mathbf{q},E)A(\mathbf{k
^{\prime }-\mathbf{q},\mathbf{k}-\mathbf{q},E^{\prime })) \notag
\end{gather
and computing one of the energy integration by exploting the analytical
properties of the averaged Green function we obtai
\begin{equation}
\chi (\mathbf{q},\omega )=\chi ^{AR}(\mathbf{q},\omega )+\chi ^{AA}(\mathbf{
},\omega )+\chi ^{RR}(\mathbf{q},\omega ) \label{dif4.1}
\end{equation
wher
\begin{equation}
\chi ^{RA}(\mathbf{q},\omega )=\int \frac{dE}{2\pi i}\left[ \rho (E+\hbar
\omega )-\rho (E)\right] \Phi _{lijk}^{AR}(E+\hbar \omega ,E,q)
\label{dif4.2}
\end{equation
an
\begin{equation}
\chi ^{AA}(\mathbf{q},\omega )=-\int \frac{dE}{2\pi i}\rho (E+\hbar \omega
)\Phi _{lijk}^{AA}(E+\hbar \omega ,E,\mathbf{q}) \label{dif4.3}
\end{equation
\begin{equation}
\chi ^{RR}(\mathbf{q},\omega )=\int \frac{dE}{2\pi i}\rho (E)\Phi
_{lijk}^{RR}(E+\hbar \omega ,E,\mathbf{q}) \label{dif4.4}
\end{equation
With this result, the generalization of the density response function for
linear dispersion relation and spinor wave function is obtained.
\section{Current and density response relations}
With the generalization of the current and density response functions for
linear dispersion relations and spinors, we can proceed to obtain a relation
between those functions. Introducing a Kronecker delta product $\delta
_{ij}\delta _{kl}$ in $\chi (\mathbf{q},\omega )$ and using eq.(\ref{con20.8
) we can writ
\begin{equation}
i\omega \sigma _{\alpha \beta }(\mathbf{q},\omega )+e^{2}v_{F}^{2}\chi
\mathbf{q},\omega )=\xi (\mathbf{q},\omega )-K_{\alpha \beta }(\mathbf{q},0)
\label{cu2.1}
\end{equation
wher
\begin{equation}
\xi (\mathbf{q},\omega )=Tr(\Lambda \Gamma )=\Lambda _{ijkl}^{\alpha \beta
}\Gamma _{ijkl}(\mathbf{q},\omega ) \label{cu4}
\end{equation
wher
\begin{equation}
\Lambda _{ijkl}^{\alpha \beta }=e^{2}v_{F}^{2}(\sigma _{ij}^{\alpha }\sigma
_{kl}^{\beta }+\delta _{ij}\delta _{kl}) \label{cu5}
\end{equation
an
\begin{gather}
\Gamma _{ijkl}(\mathbf{q},\omega )=\int \frac{d^{2}\mathbf{k}}{(2\pi )^{2}
\int \frac{d^{2}\mathbf{k}^{\prime }}{(2\pi )^{2}}\int \frac{dE}{2\pi }\int
\frac{dE^{\prime }}{2\pi }\frac{\rho (E)-\rho (E^{\prime })}{E^{\prime
}-E+\hbar \omega +is}\times \label{cu6} \\
\left\langle \left[ G_{li}^{R}(\mathbf{k}+\mathbf{q},\mathbf{k}^{\prime }
\mathbf{q},E)-G_{li}^{A}(\mathbf{k}+\mathbf{q},\mathbf{k}^{\prime }+\mathbf{
},E)\right] \left[ G_{jk}^{R}(\mathbf{k}^{\prime }-\mathbf{q},\mathbf{k}
\mathbf{q},E^{\prime })-G_{jk}^{A}(\mathbf{k}^{\prime }-\mathbf{q},\mathbf{k
-\mathbf{q},E^{\prime })\right] \right\rangle \notag
\end{gather
Relation eq.(\ref{cu2.1}) is analogous to the relation introduced in \cit
{Janis}, eq.(38), but in the former case, the relation obtained is not a
definition as it occurs in \cite{Janis}. The main difference is that in
graphene and in general for spinor systems with linear dispersion relation,
the space derivate is replaced by the Pauli matrix, then the Fourier
transform do not introduce any momentum $\mathbf{p}$. The non-appearance of
the momentum in the current response function implies a different functional
behavior, but the same diagrammatic expansion. In the other side, whenever
there is a continuity equation, which expresses the charge conservation, it
is possible to obtain a direct relation between isotropic conductivity and
density response function
\begin{equation}
\sigma (\mathbf{q},\omega )=-\frac{ie^{2}\omega }{q^{2}}\chi (\mathbf{q
,\omega ) \label{cu7}
\end{equation
In the particular case of graphene, a continuity equation can be obtained,
which is identical to continuity equation for quantum relativistic systems.
Combining last equation and eq.(\ref{cu2.1}) we obtain for the conductivity
tensor $\sigma _{\alpha \alpha }(q,\omega )=\sigma (q,\omega )
\begin{equation}
\sigma _{\alpha \beta }(\mathbf{q},\omega )=\frac{\omega \left[ \xi (\mathbf
q},\omega )-K_{\alpha \beta }(\mathbf{q},0)\right] }{i\omega
^{2}+iv_{F}^{2}q^{2}} \label{cu8}
\end{equation
The static limit of the homogeneous conductivity can be computed as (see
eq.(40a) of \cite{Janis}
\begin{equation}
\sigma =\underset{\omega \rightarrow 0}{\lim }\underset{\mathbf{q
\rightarrow 0}{\lim }\sigma (\mathbf{q},\omega )=\underset{\omega
\rightarrow 0}{\lim }\frac{1}{i\omega }\xi (0,\omega )-\frac{1}{i\omega
\underset{\mathbf{q}\rightarrow 0}{\lim }\underset{\omega \rightarrow 0}
\lim }K_{\alpha \beta }(\mathbf{q},\omega )=e^{2}D\frac{\partial n}{\partial
\mu } \label{cu9}
\end{equation
that relates the diffusion constant $D$ with the static conductivity, known
as Einstein relation. At zero temperature, $\frac{\partial n}{\partial \mu
=n_{F}$, where $n_{F}$ is the density of states at the Fermi energy. In the
other side, replacing eq.(\ref{cu7}) in eq.(\ref{cu2.1}) we obtain for the
density response functio
\begin{equation}
\chi (\mathbf{q},\omega )=\frac{q^{2}\left[ \xi (\mathbf{q},\omega
)-K_{\alpha \alpha }(\mathbf{q},0)\right] }{e^{2}\omega
^{2}+e^{2}v_{F}^{2}q^{2}} \label{cu10}
\end{equation
which is similar to eq.(42) of \cite{Janis}, but in this case, this relation
is exact.
\subsection{Current and density limits}
To compute the two limits $\omega \rightarrow 0$ and $\mathbf{q}\rightarrow
0 $ for the response functions it is only necessary to study the tensor
\Gamma _{ijkl}(\mathbf{q},\omega )$ that can be separated a
\begin{equation}
\Gamma _{ijkl}(\mathbf{q},\omega )=\Gamma _{ijkl}^{RA}(\mathbf{q},\omega
)+\Gamma _{ijkl}^{AA}(\mathbf{q},\omega )+\Gamma _{ijkl}^{RR}(\mathbf{q
,\omega ) \label{dc1}
\end{equation
In particular, the $\omega \rightarrow 0$ limit read
\begin{equation}
\underset{\omega \rightarrow 0}{\lim }\Gamma _{ijkl}(\mathbf{q},\omega
)=\int \frac{dE}{2\pi i}\rho (E)\left[ \Phi _{lijk}^{RR}(E,E,\mathbf{q
)-\Phi _{lijk}^{AA}(E,E,\mathbf{q})\right] \label{dc2}
\end{equation
using that $G^{A}=[G^{R}]^{\ast }$we hav
\begin{equation}
\Phi _{lijk}^{RR}(E,E,\mathbf{q})-\Phi _{lijk}^{AA}(E,E,\mathbf{q})=-2i\Im
\left[ \Phi _{lijk}^{RR}(E,E,\mathbf{q})\right] \label{dc3}
\end{equation
Finally, taking the $\mathbf{q}\rightarrow 0$ limit and using the Ward
identity (see \cite{janis2}) we obtai
\begin{equation}
\underset{\mathbf{q}\rightarrow 0}{\lim }\underset{\omega \rightarrow 0}
\lim }\Gamma _{ijkl}(\mathbf{q},\omega )=\int \frac{dE}{\pi }\rho (E)\int
\frac{d^{2}\mathbf{k}}{(2\pi )^{2}}\Im \left[ \frac{\partial g_{R}(\mathbf{k
,E)}{\partial E}\right] f_{li}(\mathbf{k})f_{jk}(\mathbf{k}) \label{dc4}
\end{equation
wher
\begin{equation}
g_{R}(\mathbf{k},E)=\frac{1}{E-\hbar v_{F}k-i\eta -is} \label{dc5}
\end{equation
an
\begin{equation}
f(\mathbf{k})=\left(
\begin{array}{cc}
1 & e^{i\varphi _{\mathbf{k}}} \\
e^{-i\varphi _{\mathbf{k}}} &
\end{array
\right) =\left(
\begin{array}{cc}
1 & e^{i\lambda } \\
e^{-i\lambda } &
\end{array
\right) \label{dc6}
\end{equation
where $\lambda $ is the polar angle of the wave vector. Applying the chain
rule in the derivat
\begin{equation}
\underset{\mathbf{q}\rightarrow 0}{\lim }\underset{\omega \rightarrow 0}
\lim }\Gamma _{ijkl}(\mathbf{q},\omega )=-\int \frac{dE}{\pi }\frac{\partial
\rho }{\partial \mu }\int \frac{d^{2}\mathbf{k}}{(2\pi )^{2}}\Im \left[
g_{R}(\mathbf{k},E)\right] f_{li}(\mathbf{k})f_{jk}(\mathbf{k}) \label{dc7}
\end{equation
Those integral matrix elements that contains $e^{\pm i\lambda }$ will not
contribute because the angular integration vanish. Multypling last result
with $e^{2}v_{F}^{2}\sigma _{ij}^{\alpha }\sigma _{kl}^{\alpha }$ and taking
the $T\rightarrow 0$ limit we obtain
\begin{equation}
\underset{\mathbf{q}\rightarrow 0}{\lim }\underset{\omega \rightarrow 0}
\lim }\left( K_{\alpha \alpha }^{AA}(\mathbf{q},\omega )+K_{\alpha \alpha
}^{RR}(\mathbf{q},\omega )\right) =-\underset{\mathbf{q}\rightarrow 0}{\lim
\underset{\omega \rightarrow 0}{\lim }K_{\alpha \alpha }(\mathbf{q},\omega
)=-e^{2}v_{F}^{2}n_{F}(\eta ) \label{dc8}
\end{equation
The longitudinal conductivity depends only in the electron-hole correlation
function as it is expected
\begin{equation}
\underset{\mathbf{q}\rightarrow 0}{\lim }\underset{\omega \rightarrow 0}
\lim }\sigma (\mathbf{q},\omega )=\underset{\omega \rightarrow 0}{\lim
\frac{K_{\alpha \alpha }^{RA}(0,\omega )}{i\omega } \label{dc8.1}
\end{equation
In the other side, taking the $\mathbf{q}\rightarrow 0$ limit and using the
Ward identity we obtain
\begin{gather}
\underset{\mathbf{q}\rightarrow 0}{\lim }\Gamma _{ijkl}(\mathbf{q},\omega
)=\int \frac{dE}{2\pi i}\int \frac{d^{2}\mathbf{k}}{(2\pi )^{2}}[\rho
(E+\hbar \omega )\left( g_{A}(\mathbf{k},E+\hbar \omega )-g_{R}(\mathbf{k
,E+\hbar \omega )\right) - \label{dc8.2} \\
\rho (E)\left( g_{A}(\mathbf{k},E)-g_{R}(\mathbf{k},E)\right) ]f_{li}
\mathbf{k})f_{jk}(\mathbf{k})=0 \notag
\end{gather
Because both contributions gives the density of states at the Fermi level
when the tensor $f_{li}(\mathbf{k})f_{jk}(\mathbf{k})$ is contracted with
\Lambda _{ijkl}^{\alpha \beta }$. Last equation and the result of eq.(\re
{dc8}) implies that the tensor $\Gamma _{ijkl}(\mathbf{q},\omega )$ is not
analytical in the $\mathbf{q}\rightarrow 0$ and $\omega \rightarrow 0$ limit
as it occurs in conventional systems.
\subsection{Boltzmann limit and minimum conductivity}
The Boltzmann limit can be introduced by making the following approximation
(see \cite{Rammer}
\begin{gather}
\left\langle G_{li}^{R(A)}(\mathbf{k}+\mathbf{q},\mathbf{k}^{\prime }
\mathbf{q},E)G_{jk}^{R(A)}(\mathbf{k}^{\prime }-\mathbf{q},\mathbf{k}
\mathbf{q},E^{\prime })\right\rangle \sim \label{au1} \\
\left\langle G_{li}^{R(A)}(\mathbf{k}+\mathbf{q},\mathbf{k}^{\prime }
\mathbf{q},E)\right\rangle \left\langle G_{jk}^{R(A)}(\mathbf{k}^{\prime }
\mathbf{q},\mathbf{k}-\mathbf{q},E^{\prime })\right\rangle \notag \\
=\delta _{k,k^{\prime }}G_{li}^{R(A)}(\mathbf{k}+\mathbf{q},E)G_{jk}^{R(A)}
\mathbf{k}-\mathbf{q},E^{\prime }) \notag
\end{gather
where $G_{li}^{R(A)}(\mathbf{k}+\mathbf{q},E)$ is the impurity averaged
Green function computed in Section I. Because in the $\omega \rightarrow 0$
limit, the conductivity will depends on the electron-hole correlation
function $\Phi _{lijk}^{RA}$, we will compute $\Gamma _{ijkl}^{RA}(\mathbf{q
,\omega )$. Introducing a shift $E\rightarrow E-\frac{\hbar \omega }{2}$ we
hav
\begin{equation}
\Gamma _{ijkl}^{RA}(\mathbf{q},\omega ,\eta )=-i\int \frac{dE}{2\pi }\left[
\rho (E+\frac{\hbar \omega }{2})-\rho (E-\frac{\hbar \omega }{2})\right]
\int \frac{d^{2}\mathbf{k}}{(2\pi )^{2}}G_{li}^{R}(\mathbf{k-q},E+\frac
\hbar \omega }{2})G_{jk}^{A}(\mathbf{k}-\mathbf{q},E-\frac{\hbar \omega }{2})
\label{au2}
\end{equation
Because we have to compute the trace $\Lambda _{ijkl}^{\alpha \alpha }\Gamma
_{ijkl}(\mathbf{q},\omega ,\eta )$, the only tensor elements that are not
zero read
\begin{gather}
\xi (\mathbf{q},\omega )=e^{2}v_{F}^{2}(\Lambda _{1212}^{\alpha \alpha
}\Gamma _{1212}^{RA}+\Lambda _{1221}^{\alpha \alpha }\Gamma
_{1221}^{RA}+\Lambda _{2112}^{\alpha \alpha }\Gamma _{2112}^{RA}+\Lambda
_{2121}^{\alpha \alpha }\Gamma _{2121}^{RA}+ \label{au3} \\
\Lambda _{1111}^{\alpha \alpha }\Gamma _{1111}^{RA}+\Lambda _{1122}^{\alpha
\alpha }\Gamma _{1122}^{RA}+\Lambda _{2211}^{\alpha \alpha }\Gamma
_{2211}^{RA}+\Lambda _{2222}^{\alpha \alpha }\Gamma _{2222}^{RA}) \notag
\end{gather
wher
\begin{equation}
\Lambda _{1212}^{\alpha \alpha }=\sigma _{12}^{\alpha }\sigma _{12}^{\alpha
}=e^{2}v_{F}^{2}e^{2i\phi }\text{ \ \ \ \ \ \ \ \ \ \ \ \ \ }\Lambda
_{2121}^{\alpha \alpha }=\sigma _{21}^{\alpha }\sigma _{21}^{\alpha
}=e^{2}v_{F}^{2}e^{-2i\phi } \label{au4}
\end{equation
\begin{equation}
\Lambda _{1221}^{\alpha \alpha }=\Lambda _{2112}^{\alpha \alpha }=\Lambda
_{1111}^{\alpha \alpha }=\Lambda _{1122}^{\alpha \alpha }=\Lambda
_{2211}^{\alpha \alpha }=\Lambda _{2222}^{\alpha \alpha }=e^{2}v_{F}^{2}
\label{au5}
\end{equation
In tur
\begin{equation}
\Gamma _{1221}^{RA}=\Gamma _{2112}^{RA}=\Gamma _{1111}^{RA}=\Gamma
_{2222}^{RA}=g^{RA}(k,q,E,\omega ) \label{au7}
\end{equation
an
\begin{eqnarray}
\Gamma _{2121}^{RA} &=&g^{RA}(\mathbf{k},\mathbf{q},E,\omega )e^{i(\varphi _
\mathbf{k}+\mathbf{q}}+\varphi _{\mathbf{k}-\mathbf{q}})}\text{ \ \ \ \ \ \
\ \ \ \ \ }\Gamma _{1122}^{RA}=g^{RA}(\mathbf{k},\mathbf{q},E,\omega
)e^{i(\varphi _{\mathbf{k}-\mathbf{q}}-\varphi _{\mathbf{k}+\mathbf{q}})}
\label{au8} \\
\Gamma _{2211}^{RA} &=&g^{RA}(\mathbf{k},\mathbf{q},E,\omega )e^{i(\varphi _
\mathbf{k}+\mathbf{q}}-\varphi _{\mathbf{k}-\mathbf{q}})}\text{ \ \ \ \ \ \
\ \ \ \ \ }\Gamma _{1212}^{RA}=g^{RA}(\mathbf{k},\mathbf{q},E,\omega
)e^{-i(\varphi _{\mathbf{k}+\mathbf{q}}+\varphi _{\mathbf{k}-\mathbf{q}})}
\notag
\end{eqnarray
wher
\begin{gather}
g^{RA}(\mathbf{k},\mathbf{q},E,\omega )=g_{R}(\mathbf{k}+\mathbf{q},E+\frac
\hbar \omega }{2})g_{A}(\mathbf{k}-\mathbf{q},E-\frac{\hbar \omega }{2})=
\label{au12} \\
\frac{1}{\left( E+\frac{\hbar \omega }{2}-\hbar v_{F}\left\vert \mathbf{k}
\mathbf{q}\right\vert -i\eta -is\right) \left( E-\frac{\hbar \omega }{2
-\hbar v_{F}\left\vert \mathbf{k}-\mathbf{q}\right\vert -i\eta +is\right) }
\notag
\end{gather
For $\mathbf{q}=0
\begin{gather}
\xi ^{RA}(0,\omega )=\Lambda _{ijkl}^{\alpha \alpha }\Gamma
_{ijkl}^{RA}(0,\omega )=-e^{2}v_{F}^{2}\int \frac{dE}{2\pi }\left[ \rho (E
\frac{\hbar \omega }{2})-\rho (E-\frac{\hbar \omega }{2})\right]
\label{au13} \\
\int \frac{d^{2}\mathbf{k}}{(2\pi )^{2}}\frac{e^{2i\phi }e^{-2i\varphi _
\mathbf{k}}}+e^{-2i\phi }e^{2i\varphi _{\mathbf{k}}}+2}{\left( E+\frac{\hbar
\omega }{2}-\hbar v_{F}k-i\eta -is\right) \left( E-\frac{\hbar \omega }{2
-\hbar v_{F}k-i\eta +is\right) } \notag
\end{gather
writin
\begin{equation}
e^{\pm 2i\varphi \mathbf{_{\mathbf{k}}}}=\cos (2arctg(\frac{k_{y}}{k_{x}
))+i\sin (2arctg(\frac{k_{y}}{k_{x}}))=\frac{1}{k^{2}}\left(
k_{x}^{2}-k_{y}^{2}\pm 2ik_{x}k_{y}\right) \label{au14}
\end{equation
using that $k_{x}=k\cos \lambda $ and $k_{y}=k\sin \lambda $ and computing
the angular integration, we obtai
\begin{gather}
\xi ^{RA}(0,\omega )=\Lambda _{ijkl}^{\alpha \alpha }\Gamma
_{ijkl}^{RA}(0,\omega )=-2e^{2}v_{F}^{2}\int \frac{dE}{2\pi }\left[ \rho (E
\frac{\hbar \omega }{2})-\rho (E-\frac{\hbar \omega }{2})\right]
\label{au15} \\
\int \frac{d^{2}\mathbf{k}}{(2\pi )^{2}}\frac{1}{\left( E+\frac{\hbar \omega
}{2}-\hbar v_{F}k-i\eta -is\right) \left( E-\frac{\hbar \omega }{2}-\hbar
v_{F}k-i\eta +is\right) } \notag
\end{gather
The small parameter $is$ can be disregard because the self-energy $i\eta $
moves the poles of $\xi ^{RA}(0,\omega )$ away from the real line. Using a
simplified version of the Ward identity, we can compute the integral in $k$
as follows
\begin{gather}
\frac{1}{2\pi }\int_{0}^{1/a}\frac{kdk}{(E+\frac{\hbar \omega }{2
-v_{f}\hbar k-i\eta )(E-\frac{\hbar \omega }{2}-v_{f}\hbar k-i\eta )}=
\label{au16} \\
\frac{1}{2\pi \hbar \omega }\int_{0}^{1/a}dk\left( \frac{k}{E+\frac{\hbar
\omega }{2}-v_{f}\hbar k-i\eta }-\frac{k}{E-\frac{\hbar \omega }{2
-v_{f}\hbar k-i\eta }\right) \notag
\end{gather
A special feature about graphene is the no disorder limit $\eta \rightarrow
0 $. In this case, the density of states at the Fermi level is zero
n(E_{F})=0 $ which implies that there is no charge carriers. Nevertheless, a
minimal conductivity value can be found as follows:\ last equation can be
separated in a real and imaginary part, but the principal part will not
contribute to the conducticity because it vanishes since $\frac{k(E+\frac
\hbar \omega }{2}-v_{f}\hbar k)}{(E+\frac{\hbar \omega }{2}-v_{f}\hbar
k)^{2}+\eta ^{2}}$ is an odd function of $E\pm \frac{\hbar \omega }{2
-v_{f}\hbar k$, the width $\eta $ is small and $k$ is a slow varying
function from~$0$ to $1/a$, the
\begin{equation}
\underset{\eta \rightarrow 0}{\lim }\frac{i}{2\pi \hbar \omega
\int_{0}^{1/a}dk\left[ \frac{k\eta }{(E+\frac{\hbar \omega }{2}-v_{f}\hbar
k)^{2}+\eta ^{2}}-\frac{k\eta }{(E-\frac{\hbar \omega }{2}-v_{f}\hbar
k)^{2}+\eta ^{2}}\right] =-\frac{i}{2v_{F}^{2}\hbar ^{2}} \label{au17}
\end{equation
Using last result in eq.(\ref{au15}) and taking into account tha
\begin{equation}
\rho (E+\frac{\hbar \omega }{2})-\rho (E-\frac{\hbar \omega }{2})=-\frac
\sinh (\frac{\beta \hbar \omega }{2})}{\cosh (\frac{\beta \hbar \omega }{2
)+\cosh (\beta E)} \label{au19}
\end{equation
which behaves at low temperatures as $\rho (E+\frac{\hbar \omega }{2})-\rho
(E-\frac{\hbar \omega }{2})\sim 1\,$between $-\frac{\hbar \omega }{2}$ and
\frac{\hbar \omega }{2}$ and zero in the remaining energy values, the
\begin{equation}
\int \frac{dE}{2\pi }\left[ \rho (E+\frac{\hbar \omega }{2})-\rho (E-\frac
\hbar \omega }{2})\right] b(E)=\int_{-\frac{\hbar \omega }{2}}^{\frac{\hbar
\omega }{2}}b(E)dE \label{au20}
\end{equation
where $b(E)$ is any function. Eq.(\ref{au15}) finally read
\begin{equation}
\xi ^{RA}(0,\omega )=\Lambda _{ijkl}^{\alpha \alpha }\Gamma
_{ijkl}^{RA}(0,\omega ,\eta )=\frac{ie^{2}\hbar \omega }{2\pi \hbar ^{2}}
\label{au21}
\end{equation
by applying eq.(\ref{cu9}
\begin{equation}
\sigma _{0}=\frac{1}{i\omega }\frac{ie^{2}\hbar \omega }{2\pi \hbar ^{2}}
\frac{e^{2}}{2\pi \hbar } \label{au22}
\end{equation
Altough there is no disorder ($\eta \rightarrow 0~$limit) and in consequence
no density of states at the Fermi energy, is unusual to obtain a minimal
conductivity. This result is agreement with the result found in (\cite{Peres
, eq.(2.53)), but in disagreement with other results (see \cite{Ziegler})
\footnote
The conductivity of eq.(\ref{au22}) must be multiplied by the degeneracy
given by spin and valley $K$ and $K^{\prime }$. Then, the value would be
4\sigma _{0}$.} As we point before, we are using the Born approximation to
treat impurity effects in graphene, which is valid only in the weak
scattering regime. This impose conditions on the possible value of the the
impurity potential $V_{0}$, in particular, it should be less than the
bandwidth because we are in the linear dispersion regime. In turn, this
approximation ommit scatterings on pairs and larger groups of impurities,
then it is expected to remain valid provided cluster effects are
insignificant. In the other side, when impurity concentration is gradually
increased, individual impurity states begin to overlap and the contribution
from these states to the self--energy is becoming more pronounced in the
vicinity of the impurity state energy and a spectrum rearrangement appears
for a critical concentration (see \cite{yuriv}). This impose several
restrictions to the possible values for the the concentration of impurities
and the potential $V_{0}$ value (see \cite{yuriv2} and \cite{yuriv3}), which
in turn impose several restrictions to the approximation used in this work,
because it cannot be applied in a close vicinity of the Dirac point in the
spectrum due to the increase in cluster scattering. Nevertheless, in \cit
{loktev1} and \cite{loktev2}, a $E_{F}\rightarrow 0$ limit is taken on the
average Green function and by using the Ioffe-Regel criterion (see \cit
{ioffe}), one of the solutions of this limit implies that the
self-consistent method is not applicable near the nodal point, which is
equivalent to the conditions found in \cite{yuriv2}, but another low energy
asymptotics solution exists, which impose more suitable conditions for the
applicability of the Born approximation (see eq.(9) of \cite{loktev2}). This
point deserves more attention, because the low energy limit in the graphene
Green function and correlation functions raise up a non-analytical behavior
which produces different results (see \cite{Ziegler}). Another important
point is to compute minimum conductivity by taking into account the Velic
\'{y}-Ward identity, which introduce a two-particle irreducible vertex
consistent with the coherent-potential approximation for the self-energy
(see \cite{veli}, \cite{baym} and \cite{baym2}). In particular, a Cooper
pole could be computed in the two-particle irreducible vertex due to
backscattering, which will dominate the low-energy behavior of the
conductivity and this could give some insight for the minimum conductivity
puzzle.
\section{Dynamical diffusion}
A dynamical generalization of the diffusion constant from the electron-hole
correlation function cannot be linked to the frequency dependent
conductivity (see eq.(3.18) and eq.(3.19) of \cite{Janis}). For this, is
necessary to obtain a dynamical diffusion from a different procedure. The
relaxation of a non-equilibrium particle density distribution can be studied
through the diffusion equation
\begin{equation}
\frac{\partial \delta n}{\partial t}-D\nabla ^{2}\delta n=0 \label{diff1}
\end{equation
where the Fourier transformed solution read
\begin{equation}
\delta n(\mathbf{q},\omega )=\frac{\delta n(t=0,\mathbf{q})}{i\omega -Dq^{2}}
\label{diff2}
\end{equation
The induced non-equilibrium density variation that arose as a response to a
weak inhomogeneous electric field, where this perturbation is first slowly
switched on during the time interval $(-\infty ,0)$ and then suddenly turned
off at $t=0$ reads (see \cite{belitz}
\begin{equation}
\delta n(\mathbf{q},t)=eV(\mathbf{q})\theta (t)\int_{-\infty }^{0}dt^{\prime
}e^{\epsilon t^{\prime }}\chi (\mathbf{q},t-t^{\prime })=eV(\mathbf{q})\phi
\mathbf{q},t) \label{diff3}
\end{equation
where $V(\mathbf{q})$ is the Fourier transform of the scalar potential and
\phi (\mathbf{q},t)$ is the relaxation function. The Fourier transform of
last equation gives a relation between $\delta n(\mathbf{q},\omega )$ and
\phi (\mathbf{q},\omega )$, the
\begin{equation}
\phi (\mathbf{q},\omega )=\frac{\frac{\partial n}{\partial \mu }}{-i\omega
+D(\omega )q^{2}} \label{diff4}
\end{equation
wher
\begin{equation}
\frac{\partial n}{\partial \mu }=\frac{\delta n(t=0,\mathbf{q})}{eV(\mathbf{
})} \label{diff5}
\end{equation
From eq.(\ref{diff4}) we can obtain the dynamical diffusion
\begin{equation}
2\frac{\partial n}{\partial \mu }D(\omega )=\omega ^{2}\frac{\partial
^{2}\phi }{\partial \mathbf{q}^{2}}\mid _{\mathbf{q}=0} \label{diff6}
\end{equation
In turn, from eq.(\ref{diff3}) we obtain a relation between the relaxation
function $\phi (\omega ,q)$ and the response function $\chi (\omega ,q)$
\begin{equation}
i\omega \phi (\mathbf{q},\omega )=\chi (\mathbf{q},\omega )-\chi (\mathbf{q
,0) \label{diff7}
\end{equation
Using eq.(\ref{cu10}), we can obtain the dynamical diffusion in terms of
\xi (\mathbf{q},\omega )$ without taking the $\eta \rightarrow 0$ limi
\begin{equation}
2\frac{\partial n}{\partial \mu }D(\omega ,\eta )=\frac{2}{ie^{2}\omega
(\xi (0,\omega )-\underset{\mathbf{q}\rightarrow 0}{\lim }\underset{\omega
\rightarrow 0}{\lim }K_{\alpha \alpha }(\mathbf{q},\omega ))+\frac{i\omega }
e^{2}v_{F}^{2}}(\frac{\partial ^{2}K_{\alpha \alpha }(\mathbf{q},0)}
\partial \mathbf{q}^{2}}\mid _{\mathbf{q}=0}-\frac{\partial ^{2}\xi (\mathbf
q},0)}{\partial \mathbf{q}^{2}}\mid _{\mathbf{q}=0}) \label{diff8}
\end{equation
The dynamical diffusion will contain two contributions at order $O(\omega )
. The first one contains the diffusion pole $1/\omega $ of the relaxation
function and will not depends on disorder. The second term will be
proportional to $\omega $ and the factor will be a $\eta $ dependent
function. From last section, we found that $\xi (0,\omega )=0$ and that
\underset{\mathbf{q}\rightarrow 0}{\lim }\underset{\omega \rightarrow 0}
\lim }K_{\alpha \alpha }(\mathbf{q},\omega )=e^{2}v_{F}^{2}n_{F}$, the
\begin{equation}
2\frac{\partial n}{\partial \mu }D(\omega )=-\frac{2v_{F}^{2}n_{F}}{i\omega
-i\omega \frac{\partial ^{2}\Gamma _{iill}(\mathbf{q},0)}{\partial \mathbf{q
^{2}}\mid _{\mathbf{q}=0} \label{diff8.1}
\end{equation
where we have used eq.(\ref{cu2.1}) to eq.(\ref{cu5}). The $\eta $ dependent
factor will depends on the electron-hole correlation function, but in this
case, we have take into account the $\mathbf{q}$ dependence. Using eq.(\re
{dc4}), last term of the r.h.s. of eq.(\ref{dc2}) can be written a
\begin{equation}
\frac{\partial ^{2}\Gamma _{iill}(\mathbf{q},0)}{\partial \mathbf{q}^{2}
\mid _{\mathbf{q}=0}=-\int \frac{dE}{\pi }\rho (E)\frac{\partial ^{2}\Im
\left[ \Phi _{liil}^{RR}(E,E,\mathbf{q})\right] }{\partial \mathbf{q}^{2}
\mid _{\mathbf{q}=0} \label{diff8.2}
\end{equation
Writing
\begin{equation}
\frac{\partial ^{2}\Im \left[ \Phi _{liil}^{RR}(E,E,\mathbf{q})\right] }
\partial \mathbf{q}^{2}}\mid _{\mathbf{q}=0}=\frac{dS}{dE} \label{diff8.3}
\end{equation
Eq.(\ref{diff8.2}) can be written a
\begin{equation}
\frac{\partial ^{2}\Gamma _{iill}(\mathbf{q},0)}{\partial \mathbf{q}^{2}
\mid _{\mathbf{q}=0}=-\int \frac{dE}{\pi }\rho (E)\frac{\partial ^{2}\Im
\left[ \Phi _{liil}^{RR}(E,E,\mathbf{q})\right] }{\partial \mathbf{q}^{2}
\mid _{\mathbf{q}=0}=\frac{1}{\pi }S(-\infty )+\frac{S(0)}{\pi }
\label{diff8.4}
\end{equation
where we have integrate by parts and used that $\frac{\partial \rho }
\partial E}=\delta (E)$ in the $T\rightarrow 0$ limit. \ Taking into account
eq.(\ref{con20.6.1.7}) in the Boltzmann limit, the electron-electron
correlation function $\Phi _{liil}^{RR}(E,E,\mathbf{q})$ can be written a
\begin{gather}
\Phi _{liil}^{RR}(E,E,\mathbf{q})= \label{diff8.5} \\
\int_{0}^{1/a}\int_{0}^{2\pi }\frac{dkkd\lambda }{(2\pi )^{2}}\frac{1+\frac
k^{2}-q^{2}}{\sqrt{(k^{2}-2qk\cos \lambda +q^{2})(k^{2}+2qk\cos \lambda
+q^{2})}}}{(E-\hbar v_{F}\sqrt{k^{2}+2kq\cos \lambda +q^{2}}-i\eta )(E-\hbar
v_{F}\sqrt{k^{2}-2kq\cos \lambda +q^{2}}-i\eta )} \notag
\end{gather
where we have put the $\mathbf{q}$ direction in the same direction as $k_{x}
, then $\mathbf{q\cdot k}=qk\cos \lambda $ where $\lambda $ is the polar
angle of $\mathbf{k}$. In appendix C we have computed $\frac{\partial
^{2}\Im \left[ \Phi _{liil}^{RR}(E,E,q)\right] }{\partial \mathbf{q}^{2}
\mid _{\mathbf{q}=0}$, where the result read
\begin{equation}
\frac{\partial ^{2}\left[ \Phi _{liil}^{RR}(E,E,\mathbf{q})\right] }
\partial \mathbf{q}^{2}}\mid _{\mathbf{q}=0}=\frac{1}{\pi }\int_{0}^{1/a}d
\left[ -\hbar v_{F}g_{R}^{3}(k,E)+k\hbar ^{2}v_{F}^{2}g_{R}^{4}(k,E)-\frac
g_{R}^{2}(k,E)}{k}\right] \label{diff8.5.1}
\end{equation
Last integral will contain give a divergent result in the limit
k\rightarrow 0$, which is an infrared divergence due to the masless behavior
of electrons. To isolate the divergence, we can expand the integral in
powers of $k$ before introducing the integral limit
\begin{gather}
\frac{\partial ^{2}\left[ \Phi _{liil}^{RR}(E,E,\mathbf{q})\right] }
\partial \mathbf{q}^{2}}\mid _{\mathbf{q}=0}=-\frac{10+6iarctg(\frac{\eta }{
})}{6\pi (E-i\eta )^{2}}+\frac{1}{2\pi (E-i\eta )^{2}}\ln (\frac{E^{2}+\eta
^{2}}{\hbar ^{2}v_{F}^{2}k^{2}}) \label{diff8.6} \\
+\frac{1}{\pi }\underset{j=1}{\overset{+\infty }{\sum }}b_{j}\frac{(\hbar
v_{F}k)^{j}}{(E-i\eta )^{j+2}} \notag
\end{gather
where
\begin{equation}
b_{j}=\frac{j^{3}-3j^{2}-10j-6}{6j} \label{diff8.6.1}
\end{equation
Introducing a lower cutoff $\Lambda $, integral of eq.(\ref{diff8.5.1}) read
\begin{equation}
\frac{\partial ^{2}\left[ \Phi _{liil}^{RR}(E,E,\mathbf{q})\right] }
\partial \mathbf{q}^{2}}\mid _{\mathbf{q}=0}=\frac{\ln (a\Lambda )}{\pi
(E-i\eta )^{2}}+\frac{1}{\pi }\underset{j=1}{\overset{+\infty }{\sum }}b_{j
\frac{(\hbar v_{F})^{j}}{(E-i\eta )^{j+2}}(\frac{1}{a^{j}}-\Lambda ^{j})
\label{diff8.7}
\end{equation
Taking the imaginary part of eq.(\ref{diff8.7}
\begin{gather}
\Im \frac{\partial ^{2}\left[ \Phi _{liil}^{RR}(E,E,\mathbf{q})\right] }
\partial \mathbf{q}^{2}}\mid _{\mathbf{q}=0}=\frac{2E\eta \ln (a\Lambda )}
\pi (E^{2}+\eta ^{2})^{2}} \label{diff8.9} \\
+\frac{1}{\pi }\underset{j=1}{\overset{+\infty }{\sum }}b_{j}(\hbar
v_{F})^{j}(E^{2}+\eta ^{2})^{-\frac{1}{2}(j+2)}\sin ((j+2)arctg(\frac{\eta }
E}))(\frac{1}{a^{j}}-\Lambda ^{j}) \notag
\end{gather
Integrating in $E$ and taking the two limits of eq.(\ref{diff8.4}
\begin{equation}
\frac{\partial ^{2}\Gamma _{iill}(\mathbf{q},0)}{\partial \mathbf{q}^{2}
\mid _{\mathbf{q}=0}=-\frac{\ln (a\Lambda )}{\pi \eta }-\frac{1}{\pi
\underset{j=1}{\overset{+\infty }{\sum }}b_{j}\frac{\sin (1+j)\frac{\pi }{2}
}{j+1}\frac{(\hbar v_{F})^{j}}{\eta ^{j+1}}(\frac{1}{a^{j}}-\Lambda ^{j})
\label{diff8.10}
\end{equation
Last equation depends on the lower cutoff $\Lambda $, which is not desired.
A correct procedure can be applied by assuming that the Fermi velocity
v_{F} $ will change with $\Lambda $.\footnote
The Fermi velocity is one of the parameters of the Hamiltonian.} A
renormalization group equation can be obtained by assuming that the
dynamical diffusion do not depends on $\Lambda $. The
\begin{gather}
\frac{4n_{F}v_{F}}{i\omega }\frac{dv_{F}}{d\Lambda }+i\omega \frac{\partial
}{\partial v_{F}}\left( \frac{\partial ^{2}\Gamma _{iill}(\mathbf{q},0,\eta
,\Lambda ,v_{F}(\Lambda ))}{\partial \mathbf{q}^{2}}\mid _{\mathbf{q
=0}\right) \frac{dv_{F}}{d\Lambda }+ \label{diff8.11} \\
i\omega \frac{\partial }{\partial \Lambda }\left( \frac{\partial ^{2}\Gamma
_{iill}(\mathbf{q},0,\eta ,\Lambda ,v_{F}(\Lambda ))}{\partial \mathbf{q}^{2
}\mid _{\mathbf{q}=0}\right) =0 \notag
\end{gather
Eq.(\ref{diff8.10}) is suitable to compute differents orders of $\hbar $ to
the renormalization group equation for $v_{F}$. At order $\hbar ^{0}$ we
obtain
\begin{figure}[tbp]
\centering
\includegraphics[width=105mm,height=70mm]{disorder.eps}
\caption{Dynamical diffusion as a function of disorder for different values
of $\protect\omega $ in arbitrary units. From red to black solid lines,
\protect\omega $ increase. Dashed lines for dynamical diffusion at order
O(\hbar ^{2})$. }
\label{dynamical1}
\end{figure}
\begin{equation}
-\frac{4v_{F}n_{F}}{i\omega }\frac{dv_{F}}{d\Lambda }+\frac{i\omega }{\pi
\eta }\ln (a\Lambda )=0 \label{diff8.12}
\end{equation
and the solution reads\footnote
In eq.(\ref{diff8.13}) $\Lambda _{0}=1$ as a low limit of the cutoff has
been used.
\begin{equation}
v_{F}(\Lambda )=\sqrt{v_{F_{0}}^{2}-\frac{\omega ^{2}}{2\pi \eta n_{F}}\ln
(\Lambda )} \label{diff8.13}
\end{equation
because we are in the approximation $\omega \rightarrow 0$, last equation
read
\begin{equation}
v_{F}(\Lambda )=v_{F_{0}}-\frac{\omega ^{2}}{4\pi n_{F}\eta v_{F}^{(0)}}\ln
(\Lambda ) \label{diff8.13.1}
\end{equation
which is similar to the results found in (\cite{atta}, \cite{song}) and
shows a singular behavior with the impurity factor $\eta $ that is similar
to the singular behavior of the Fermi velocity with impurities found in \cit
{yuriv4}.\footnote
If we introduce a upper cutoff $k_{\infty }=\frac{1}{\Lambda }$, the result
of eq.(\ref{diff8.13}) follows the same behavior as other results.} Using
eq.(\ref{diff8.13}), the dynamical diffusion at order $\hbar ^{0}$ read
\begin{equation}
D(\omega )=\frac{iv_{F_{0}}^{2}}{\omega }+\frac{i\omega }{2\pi \eta n_{F}
\ln (a) \label{diff8.14}
\end{equation
\begin{figure}[tbp]
\centering
\includegraphics[width=105mm,height=70mm]{frequency.eps}
\caption{Dynamical diffusion as a function of $\protect\omega $ for
different values of $\protect\eta $ in arbitrary units. From red to black
solid lines, $\protect\eta $ increase. Dashed lines for dynamical diffusion
at order $O(\hbar ^{2})$. }
\label{dynamical2}
\end{figure}
From last equation, there is no real value for $\omega $ where $D(\omega )=0
, which is expected because suppression of diffusion can be achieved by
taking into account maximally crossed diagrams in the perturbation
expansion. Nevertheless, we can plot $D(\omega ,\eta )$ as a function of
\omega $ for different values of $\eta $ and $D(\omega ,\eta )$ as a
function of $\eta $ for different values of $\omega $. Both figures show the
diffusion pole at $\omega =0$ \ (see \ref{dynamical1} and \ref{dynamical2}).
Dynamical diffusion tends to $\frac{v_{F_{0}}^{2}}{2i\omega }$ for $\eta
\rightarrow \infty $. In turn, dynamical diffusion shows a minimum which
corresponds to the following frequenc
\begin{equation}
\omega =\sqrt{\frac{2\pi v_{F_{0}}^{2}\eta n_{F}}{\ln (a)}} \label{diff8.15}
\end{equation
which is proportional to the impurity potential $V_{0}$. This implies that
at low resonance frequencies, a decrease in the diffusion can be expected.
The full renormalization group equation can be computed in the no-disorder
limit $\eta \rightarrow 0$, which gives the following differential equation
at order $O(\hbar ^{j})$
\begin{equation}
\frac{dv_{F}}{v_{F}}=\frac{\Lambda ^{j-1}d\Lambda }{(\frac{1}{a^{j}}-\Lambda
^{j})} \label{diff8.16}
\end{equation
where the solution read
\begin{equation}
v_{F}(\Lambda )=v_{F_{0}}(\frac{1-a^{-j}}{\Lambda ^{j}-a^{-j}})^{1/j}
\label{diff8.17}
\end{equation
In the cut off limit $\Lambda \rightarrow 0$, the Fermi velocity change a
\begin{equation}
\frac{v_{F}}{v_{F_{0}}}=a(a^{-j}-1)^{1/j} \label{diff8.18}
\end{equation
which decreases with increasing order of $\hbar $. In turn, Fermi velocity
at higher orders of $\hbar $ and in the no disorder limit do not depends on
the frequency of the external perturbation. High impurity concentrations in
graphene can lead to a diffusion suppresion which would leave without effect
the high performance of the material sample as a gas sensor. Weak
localization of electrons in doped graphene implies to take into account
higher orders in the diagrammatic perturbation expansion of the current
response function. Some theoretical computations has been done (see \cit
{wl0}, \cite{wl1}). For conventional impurities, the correction becomes
positive and it leads to the fact that anti-localization is realized, which
would enhance the gas sensor peformance. In contrast, negative corrections
for short-range impurities are expected from symmetry consideration. This
suggest that the high sensitivity of graphene to detect individual dopants
is highly dependent on the quantum corrections to the conductivity. Finally,
taking into account the first quantum correction to the renormalization
group equation for $v_{F}$ we obtai
\begin{equation}
v_{F}dv_{F}=\frac{5\omega \hbar ^{2}v_{F}^{2}}{3\pi \eta ^{3}}\frac{d\Lambda
}{L(\Lambda ,\omega ,\eta ,a)}-\frac{1}{\pi \eta }\frac{d\Lambda }{\Lambda
^{2}L(\Lambda ,\omega ,\eta ,a)} \label{diff9}
\end{equation
wher
\begin{equation}
L(\Lambda ,\omega ,\eta ,a)=\frac{1}{\Lambda }\left( \frac{4n_{F}}{\omega
^{2}}+\frac{5}{3\pi }\frac{\hbar ^{2}}{\eta ^{3}}(\frac{1}{a^{2}}-\Lambda
^{2})\right) \label{diff10}
\end{equation
Solution of eq.(\ref{diff10}) in a integral form read
\begin{equation}
v_{F}^{2}(\Lambda )=v_{F_{0}}^{2}-\frac{2}{\pi \eta }\int_{1}^{\Lambda
\frac{d\Lambda ^{\prime }}{\Lambda ^{\prime 2}L(\Lambda ^{\prime },\omega
,\eta ,a)}+\frac{10\omega \hbar ^{2}}{3\pi \eta ^{3}}\int_{1}^{\Lambda
\frac{v_{F}^{2}(\Lambda ^{\prime })d\Lambda ^{\prime }}{L(\Lambda ^{\prime
},\omega ,\eta ,a)} \label{diff11}
\end{equation
at order $O(\hbar ^{2})$ we need to compute the second term of r.h.s. of
last equatio
\begin{equation}
\int_{1}^{\Lambda }\frac{d\Lambda ^{\prime }}{\Lambda ^{\prime 2}L(\Lambda
^{\prime },\omega ,\eta ,a)}=\frac{3a^{2}\pi \omega ^{2}\eta ^{3}}{10\hbar
^{2}\omega ^{2}+24a^{2}n_{F}\pi \eta ^{3}}\ln \left[ \frac{\left( 5\hbar
^{2}\omega ^{2}(a^{2}-1)-12a^{2}n_{F}\pi \eta ^{3}\right) \Lambda ^{2}}
5\hbar ^{2}\omega ^{2}(a^{2}\Lambda ^{2}-1)-12a^{2}n_{F}\pi \eta ^{3}}\right]
\label{diff12}
\end{equation
Introducing $v_{F}^{2}(\Lambda )$ in eq.(\ref{diff11}) inside the integral
of the r.h.s. in the same equation and using eq.(\ref{diff12}), the
dynamical diffusion reads at order $O(\hbar ^{2})
\begin{equation}
D(\omega )=\frac{v_{F_{0}}^{2}}{\omega }+\frac{i\omega \ln (a)}{2\pi \eta
n_{F}}+\frac{5\omega }{12\pi }\frac{v_{F_{0}}^{2}\hbar ^{2}}{\eta
^{3}n_{F}a^{2}} \label{diff13}
\end{equation
The correction introduced at order $O(\hbar ^{2})$ can be seen in both
figures as dashed lines. In the case of dynamical diffusion in terms of
frequency, the correction is small and only is appreciable for low values of
$\eta $. In this sense, quantum corrections to the diffusion do not alter
the behavior under local perturbations at linear order in $\omega $.
\section{Conclusion}
In this work a generalization of linear response theory with Kubo formula
has been introduced for linear dispersion relations and spinor wave
functions. A minimal conductivity can be found in the no disorder limit and
the result is in discordance by a factor of $2$ with other theoretical
results, although there is no consensus of the physical reason of such
value. Using the generalization introduced in the first sections, an exact
relation between current and density response functions can be obtained.
Combining this result with the relation obtained with the continuity
equation, an exact functional form of response functions are obtained, where
in particular, a singular behavior appears at $\omega \rightarrow 0$ and
q\rightarrow 0$ limit. Finally, dynamical diffusion is computed through the
relaxation function at low order in $\omega $. A regularization is
introduced to avoid infrared divergences, which introduce a renormalization
group equation for the Fermi velocity. Different contributions to this
equation can be analyzed at different order in $\hbar $. Different results
are obtained which are of importance for local pertubations of graphene
sample.
\section{Acknowledgment}
This paper was partially supported by grants of CONICET (Argentina National
Research Council) and Universidad Nacional del Sur (UNS) and by ANPCyT
through PICT 1770, and PIP-CONICET Nos. 114-200901-00272 and
114-200901-00068 research grants, as well as by SGCyT-UNS., E.A.G. and
P.V.J. are members of CONICET. P.B. and J. S.A. are fellow researchers at
this institution.
The authors are extremely grateful to the referee, whose relevant
observations have greatly improved the final version of this paper.
\section{Appendix}
\subsection{Spectral weight}
The Green function for Dirac fermion systems read
\begin{equation}
G_{ij}^{R(A)}(\mathbf{r},\mathbf{r}^{\prime },E)=\int \frac{d^{2}\mathbf{k}}
(2\pi )^{2}}\frac{\psi _{i}^{(k)}(\mathbf{r})\psi _{j}^{(k)\dag }(\mathbf{r
^{\prime })}{E-v_{f}\hbar k(\mp )is} \label{con21}
\end{equation
we can define the spectral weight a
\begin{equation}
A_{ij}(\mathbf{r},\mathbf{r}^{\prime },E)=i[G_{ij}^{R}(\mathbf{r},\mathbf{r
^{\prime },E)-G_{ij}^{A}(\mathbf{r},\mathbf{r}^{\prime },E)] \label{con22}
\end{equation
If we integrate the spectral weight in the volume we obtain the density of
state
\begin{equation}
\int d^{2}\mathbf{r}A_{ij}(\mathbf{r},\mathbf{r},E)=2\pi \delta _{ij}\int
\frac{d^{2}\mathbf{k}}{(2\pi )^{2}}\delta (E-v_{f}\hbar k)=2\pi n(E)\delta
_{ij} \label{con24}
\end{equation
where $n(E)$ is the density of states.
\subsection{Electron-electron correlation function}
The electron-electron hole correlation function read
\begin{equation}
\Phi _{liil}^{RR}(E,E,\mathbf{q})=\int_{0}^{1/a}\int_{0}^{2\pi }\frac
dkkd\lambda }{(2\pi )^{2}}\varrho (\mathbf{k},\mathbf{q},E) \label{ee1}
\end{equation
where $\varrho (\mathbf{k},\mathbf{q},E)=g_{R}(\left\vert \mathbf{k}+\mathbf
q}\right\vert ,E)g_{R}(\left\vert \mathbf{k}-\mathbf{q}\right\vert ,E)\alpha
(\mathbf{k},\mathbf{q})
\begin{equation}
g_{R}(\left\vert \mathbf{k}\pm \mathbf{q}\right\vert ,E)=\frac{1}{E-\hbar
v_{F}\left\vert \mathbf{k}\pm \mathbf{q}\right\vert -i\eta } \label{ee2}
\end{equation
an
\begin{equation}
\alpha (\mathbf{k},\mathbf{q})=1+\frac{k^{2}-q^{2}}{\left\vert \mathbf{k}
\mathbf{q}\right\vert \left\vert \mathbf{k}-\mathbf{q}\right\vert }
\label{ee3}
\end{equation
We can take the $q$ derivate inside the integral in $k$. Taking into account
tha
\begin{equation}
\frac{\partial g_{R}(\left\vert \mathbf{k}\pm \mathbf{q}\right\vert ,E)}
\partial \mathbf{q}}=-\frac{\hbar v_{F}\frac{\partial \left\vert \mathbf{k
\pm \mathbf{q}\right\vert }{\partial \mathbf{q}}}{(E-\hbar v_{F}\left\vert
\mathbf{k}\pm \mathbf{q}\right\vert -i\eta )^{2}}=-\hbar v_{F}\frac{\partial
\left\vert \mathbf{k}\pm \mathbf{q}\right\vert }{\partial \mathbf{q}
g_{R}^{2}(\left\vert \mathbf{k}\pm \mathbf{q}\right\vert ,E) \label{ee4}
\end{equation
then
\begin{equation}
\frac{\partial \varrho (\mathbf{k},\mathbf{q},E)}{\partial \mathbf{q}
=\varrho (\mathbf{k},\mathbf{q},E)B(\mathbf{k},\mathbf{q},E) \label{ee5}
\end{equation
wher
\begin{equation}
B(\mathbf{k},\mathbf{q},E)=-\hbar v_{F}\frac{\partial \left\vert \mathbf{k}
\mathbf{q}\right\vert }{\partial \mathbf{q}}g_{R}(\left\vert \mathbf{k}
\mathbf{q}\right\vert ,E)-\hbar v_{F}\frac{\partial \left\vert \mathbf{k}
\mathbf{q}\right\vert }{\partial \mathbf{q}}g_{R}(\left\vert \mathbf{k}
\mathbf{q}\right\vert ,E)+\frac{\partial \alpha (\mathbf{k},\mathbf{q})}
\partial \mathbf{q}}\frac{1}{\alpha (\mathbf{k},\mathbf{q})} \label{ee6}
\end{equation
The second derivate read
\begin{equation}
\frac{\partial ^{2}\varrho (\mathbf{k},\mathbf{q},E)}{\partial \mathbf{q}^{2
}=\varrho (\mathbf{k},\mathbf{q},E)\left[ B^{2}(\mathbf{k},\mathbf{q},E)
\frac{\partial B(\mathbf{k},\mathbf{q},E)}{\partial \mathbf{q}}\right]
\label{ee7}
\end{equation
wher
\begin{gather}
\frac{\partial B(\mathbf{k},\mathbf{q},E)}{\partial \mathbf{q}
=g_{R}(\left\vert \mathbf{k}+\mathbf{q}\right\vert ,E)\left[ -\hbar v_{F
\frac{\partial ^{2}\left\vert \mathbf{k}+\mathbf{q}\right\vert }{\partial
\mathbf{q}^{2}}+\hbar ^{2}v_{F}^{2}\left( \frac{\partial \left\vert \mathbf{
}+\mathbf{q}\right\vert }{\partial \mathbf{q}}\right) ^{2}g_{R}(\left\vert
\mathbf{k}+\mathbf{q}\right\vert ,E)\right] + \label{ee8} \\
g_{R}(\left\vert \mathbf{k}-\mathbf{q}\right\vert ,E)\left[ -\hbar v_{F
\frac{\partial ^{2}\left\vert \mathbf{k}-\mathbf{q}\right\vert }{\partial
\mathbf{q}^{2}}+\hbar ^{2}v_{F}^{2}\left( \frac{\partial \left\vert \mathbf{
}-\mathbf{q}\right\vert }{\partial \mathbf{q}}\right) ^{2}g_{R}(\left\vert
\mathbf{k}-\mathbf{q}\right\vert ,E)\right] + \notag \\
\frac{1}{\alpha (\mathbf{k},\mathbf{q})}\left[ \frac{\partial ^{2}\alpha
\mathbf{k},\mathbf{q})}{\partial \mathbf{q}^{2}}-\frac{1}{\alpha (\mathbf{k}
\mathbf{q})}\left( \frac{\partial \alpha (\mathbf{k},\mathbf{q})}{\partial
\mathbf{q}}\right) ^{2}\right] \notag
\end{gather
Finally using that
\begin{equation}
\frac{\partial \left\vert \mathbf{k}\pm \mathbf{q}\right\vert }{\partial
\mathbf{q}}=\frac{q\pm k\cos \lambda }{\left\vert \mathbf{k}\pm \mathbf{q
\right\vert } \label{ee9}
\end{equation
and that the second derivate read
\begin{equation}
\frac{\partial ^{2}\left\vert \mathbf{k}\pm \mathbf{q}\right\vert }{\partial
\mathbf{q}^{2}}=\frac{1}{\left\vert \mathbf{k}\pm \mathbf{q}\right\vert }
\frac{(q\pm k\cos \lambda )^{2}}{\left\vert \mathbf{k}\pm \mathbf{q
\right\vert ^{3}} \label{ee10}
\end{equation
Putting $\mathbf{q}=0$ in eq.(\ref{ee7})
\begin{equation}
\varrho (k,0,E)=2g_{R}(\mathbf{k},E)g_{R}(\mathbf{k},E) \label{ee11}
\end{equation
and using that $\frac{\partial \alpha (k,\mathbf{q})}{\partial \mathbf{q}
\mid _{\mathbf{q}=0}=0$,,
\begin{equation}
B^{2}(k,0,E)=0 \label{ee12}
\end{equation
In turn,
\begin{equation}
\frac{\partial B(\mathbf{k},\mathbf{q},E)}{\partial \mathbf{q}}=2g_{R}
\mathbf{k},E)\left[ -\hbar v_{F}\frac{\sin ^{2}\lambda }{k}+\hbar
^{2}v_{F}^{2}\cos ^{2}\lambda g_{R}(k,E)\right] -\frac{2\sin ^{2}\lambda }
k^{2}} \label{ee13}
\end{equation
where we have used tha
\begin{equation}
\frac{\partial ^{2}\alpha (\mathbf{k},\mathbf{q})}{\partial \mathbf{q}^{2}
\mid _{\mathbf{q}=0}=-\frac{4\sin ^{2}\lambda }{k^{2}} \label{ee14}
\end{equation
Finally the second derivate of $\varrho (\mathbf{k},\mathbf{q},E)$ at
\mathbf{q}=0$ read
\begin{equation}
\frac{\partial ^{2}\varrho (\mathbf{k},\mathbf{q},E)}{\partial \mathbf{q}^{2
}\mid _{\mathbf{q}=0}=4g_{R}^{2}(\mathbf{k},E)\left[ g_{R}(\mathbf{k},E
\left[ -\hbar v_{F}\frac{\sin ^{2}\lambda }{k}+\hbar ^{2}v_{F}^{2}\cos
^{2}\lambda g_{R}(\mathbf{k},E)\right] -\frac{\sin ^{2}\lambda }{k^{2}
\right] \label{ee15}
\end{equation
which is the desired result which will be used in Section VI.
|
\section{Introduction}
\label{sec:Introduction}
It is expected that investigation of the hadron physics in extreme conditions will give a clue for our understanding of QCD (Quantum Chromodynamics).
In particular studying asymmetric nuclear matter is also important to derive the equation of state inside neutron stars~\cite{Lattimer:2006xb}, which will give a clue to understand the recently found very heavy neutron star~\cite{Demorest:2010bx,Antoniadis:2013pzd}.
We often draw the QCD phase diagram on the plane of temperature $T$ and the baryon chemical potential $\mu_B$~\cite{Fukushima:2010bq,Fukushima:2013rx}. It is expected that various phases exist on the plane $(T, \mu_B)$ of the phase diagram: e.g. the quark-gluon plasma phase and the color superconducting phase. Similarly,
finite isospin chemical potential $\mu_I$ provide a rich phase structure
which includes the pion condensation phase.
There are many works studying the phase diagram at $\mu_I\neq0$.
In particular, the pion condensation phase transition on the plane $(\mu_B, \mu_I)$ are studied by introducing $\mu_I$ together with $\mu_B$ in the Nambu-Jona-Lasinio (NJL) model~\cite{Sasaki:2010jz,Zhang:2006gu,Toublan:2003tt,
Barducci:2004tt,He:2005nk,Mu:2010zz,He:2006tn,He:2005tf,Zhang:2013oia,Andersen:2007qv} and holographic QCD models~\cite{Lee:2013oya,Parnachev:2007bc}
and so on~\cite{Klein:2003fy,Nishida:2003fb,Toublan:2004ks,Abuki:2013vwa}.
As a first step to study the rich phase structure, it is interesting to study the
phase transition from the normal hadron phase to
the pion condensation phase
together with the equation of state
in the pion condensation phase on the plane $(\mu_B, \mu_I)$.
References~\cite{Sasaki:2010jz,Zhang:2006gu,He:2005nk,Barducci:2004tt,Toublan:2003tt,Mu:2010zz} show the
$T$ \textendash $\mu_B$ \textendash $\mu_I$ phase diagram via the NJL model, in which the dependence of the isospin density on the isospin chemical potential is shown only for $T=\mu_B=0$.
On the other hand, by using holographic QCD models~\cite{Parnachev:2007bc,Lee:2013oya}, the pion condensation phase transition is discussed.
In Ref.~\cite{Parnachev:2007bc}, they draw the phase diagram on the plane $(\mu_B, \mu_I)$ in the Sakai-Sugimoto model.
Reference~\cite{Lee:2013oya} also studies stability of the normal hadron phase at finite isospin density by introducing the baryon charge as the Reissner-Nordstr\"{o}m (RN) blackhole charge in a hard wall holographic QCD model.
However, the equation of state in the pion condensation phase is not discussed in these works.
In the previous work~\cite{Nishihara:2014nva}, we studied the pion condensation in the pure mesonic matter using a holographic QCD model by introducing the isospin chemical potential as a UV boundary value of the gauge field.
We showed that the phase transition from the normal hadron phase to the pion condensation phase is of the second order and the critical value of the isospin chemical potential is equal to the pion mass, consistently with the chiral Lagrangian analysis~\cite{Son:2000xc}.
In Ref.~\cite{Nishihara:2014nva}, we studied the $\mu_I$-dependence of the chiral condensate defined by
$\tilde{\sigma} \equiv \sqrt{ \langle \sigma \rangle^2 + \langle \pi^a \rangle^2 }$, and showed that, although the ``$\sigma$"-condensate decreases rapidly with the isospin chemical potential in the pion condensation phase, the $\pi$-condensate increases more rapidly. As a result the chiral condensate $\tilde{\sigma}$ keeps increasing, which implies the enhancement of the chiral symmetry breaking in the pion condensation phase.
The symmetry structure for this is understood in the following way:
When the isospin chemical potential is introduced, the chiral symmetry $\mbox{SU}(2)_R \times \mbox{SU}(2)_L$ is explicitly broken to $\mbox{U}(1)_R^{(3)} \times \mbox{U}(1)_L^{(3)} = \mbox{U}(1)_V^{(3)} \times \mbox{U}(1)_A^{(3)}$, where the superscript $^{(3)}$ implies that the generator $T_3$ of SU(2) is used for the U(1) as $\exp[ i \theta_V T_3 ] \in \mbox{U}(1)_V^{(3)}$.
In the normal hadron phase the $\mbox{U}(1)_A^{(3)}$ is broken by the ``$\sigma$''-condensate spontaneously and the quark mass explicitly.
In the pion condensation phase, on the other hand, the $\mbox{U}(1)_V^{(3)}$ symmetry is spontaneously broken by the $\pi$-condensate, which generates a massless Nambu-Goldstone boson.
Since both $\mbox{U}(1)_A^{(3)}$ and $\mbox{U}(1)_V^{(3)}$ are subgroups of the chiral $\mbox{SU}(2)_R \times \mbox{SU}(2)_L$ symmetry, the above structure implies that the chiral symmetry is never restored in the mesonic matter with the isospin chemical potential, and actually the breaking is enhanced in the pion condensation phase.
We note that the above properties are obtained in the pure mesonic matter, so that it is interesting to ask whether they are changed by the existence of the nucleon in the matter.
In this paper, we adopt a simple way for introducing the baryonic sources: We include a point-like nucleon source at the IR boundary coupling to the iso-triplet vector meson in the hard wall holographic QCD model as in Ref.~\cite{Kim:2007zm}, and studied the pion condensation in the asymmetric nuclear matter.
We will show that the phase transition from the normal hadron phase to the pion condensation phase is delayed in the asymmetric nuclear matter compared with the pure mesonic matter.
In other words, the critical chemical potential is larger than the pion mass.
On the other hand, the enhancement of the chiral symmetry breaking still occurs since the chiral condensate $\tilde{\sigma}$ keeps increasing with the isospin chemical potential.
This paper is organized as follows:
In section~\ref{sec:model}, we briefly review the holographic QCD model used in our analysis, and introduce the baryonic charge following Ref.~\cite{Kim:2007zm}.
Section~\ref{sec:delay} is devoted to the study of the symmetry energy and the pion mass in the normal hadron phase.
In section~\ref{sec:pion condensation},
we study the pion condensation phase and obtain the relation between the isospin chemical potential and the isospin number density as well as the chiral condensate.
In section~\ref{sec:HLS}, we make an analysis of the pion mass in the normal hadron phase using the four dimensional chiral model based on the hidden local symmetry~\cite{Bando:1987br,Harada:2003jx}.
We give a summary and discussions in section~\ref{sec:summary}.
We also show the equations of motion in appendix~\ref{app:EOM}.
\section{Model}
\label{sec:model}
In the present analysis, we employ
a holographic QCD model given in Refs.~\cite{Erlich:2005qh,Da Rold:2005zs,DaRold:2005vr} for the mesonic part.
Then the mesonic action in the five dimensional space is given by
\bee{
S_5=S_X
+S^\m{BD}
}
where
\bee{
S_X=&\int d^4x \int_\epsilon^{z_m} dz
\nn\\&
\sqrt{g} \mathrm{Tr} \left\{ |DX|^2 -m_5^2|X|^2 -\frac{1}{4g_5^2}\left(F^2_L+F^2_R\right)\right\}
\ ,
\\
S^\m{BD}=&
-\int d^4x \int_\epsilon^{z_m} dz
\nn\\
&\sqrt{g} \ \mathrm{Tr} \left\{ \lambda z_m |X|^4-m^2 z_m |X|^2 \right\}\delta\left(z-z_m\right)
}
with $m_5^2=-3$.
The metric is written as
\bee{
ds^2&=a^2(z) \left(\eta_{\mu\nu}dx^\mu dx^\nu -dz^2\right) =g_{MN}dx^M dx^N
}
with
\bee{
a(z)=\frac{1}{z}
\ ,
\label{metric}
}
where $z_m$ and $\epsilon$ are the IR-cutoff and UV-cutoff.
Here $N$ and $M$ run over 0,1,2,3,5 and $\eta_{\mu\nu}$ is the defined as the Mankowski metric: $\eta_{\mu\nu}={\rm diag} (1,-1,-1,-1)$.
\footnote{
Although there is a Chern-Simons term in addition, the term does not affect our result since we assume the rotational invariance in the present analysis.}
The model has the chiral symmetry U$(2)_L\times $U$(2)_R$ (=U$(1)_L\times $U$(1)_R\times$SU$(2)_L\times $SU$(2)_R$), under which
the fields transform in the following form:
\begin{eqnarray}
X &\rightarrow & X'= g_L X g_R^\dagger
\ ,
\label{X trans}\\
L_M &\rightarrow& {L'}_M = g_L L_M g_L^\dagger +i g_L \partial_M g_L^\dagger
\ ,
\\
R_M &\rightarrow& {R'}_M = g_R R_M g_R^\dagger +i g_R \partial_M g_R^\dagger
\label{L trans}
\end{eqnarray}
with $g_R \in \mbox{U}(2)_R$ and $g_L \in \mbox{U}(2)_L$.
The covariant derivative and the field strength are defined as
\begin{eqnarray}
D_M X&=&\partial_M X-iL_M X + i X R_M
\ ,
\\
F^L_{MN}&=&\partial_M L_N -\partial_N L_M - i\left[ L_M , L_N \right]
\label{strength}
\end{eqnarray}
and similar for $F^R_{MN}$.
These fields are
parametrized as
\begin{eqnarray}
&&
L^I_M=\mathrm{Tr}\left[L_M\sigma^I\right]
\ ,
\quad
R^I_M=\mathrm{Tr}\left[R_M\sigma^I\right]
\ ,
\label{parameterize2}
\\
&&
V^I_M=\frac{R^I_M+L^I_M}{2}
\ ,
\quad
A^I_M=\frac{R^I_M-L_M^I}{2}
\ ,
\label{parameterize4}
\\
&&
X=\frac{1}{2}\left(S^0 \sigma^0 +S^a \sigma^a \right)e^{i\pi^b \sigma^b+i \eta}
\label{parameterize5}
\end{eqnarray}
where
$\sigma^I=(\sigma^0 ,\sigma^a)=(1 ,\sigma^a)$ and $\sigma^a$ are the Pauli matrices.
In the following analysis we adopt the gauge $L_5=R_5=0$ and the IR boundary condition
$\left. F^L_{5\mu}\right|_{z_m}=\left.F^R_{5\mu}\right|_{z_m}=0$.
Now, let us include the effects of the nucleon into the model.
Here we introduce baryonic sources for the quark number density $n_q$ and the baryonic contribution to the isospin number density, denoted by $n_I^{\rm {Baryon}}$, through the following term
\footnote{The sign of this term is uniquely determined from the definition of the chemical potential
introduced
in \eqref{isospin chemical}.}:
\bee{
S_\m{int}=\int d^4x \int_\epsilon^{z_m} dz~
\left[ V_0^0 n_q + V_0^3 n_I^{\rm {Baryon}} \right]
\delta\left(z-z_m+\delta z\right)
\label{Sint}
}
where $\delta z$ ($>0$) is an infinitesimal length and
$V_0^0$ and $V_0^3$ are the gauge fields corresponding to the quark number density and isospin number density.
The baryon number density $n_B$ is defined as $n_B=n_q/N_c$.
We introduced the baryonic sources by the $\delta$-function having a peak near the IR boundary
\cite{Kim:2007zm}, which doesn't modify the IR boundary conditions.
In the present analysis,
we assume that the proton (neutron) does not appear as long as the proton (neutron) chemical potential $\mu_p$ ($\mu_n$) is smaller than the mass of a nucleon, denoted by $m_N$.
Therefore
our analysis will be done for the following three cases separately:
\bee{
&{\rm (i)}~~ -m_N \leq \mu_p < m_N \ , -m_N \leq \mu_n < m_N \ ,
\nn\\
&{\rm (ii)}~~~~~ m_N \leq \mu_p \ , ~~~~~~~~-m_N \leq \mu_n < m_N \ ,
\nn\\
&{\rm (iii)}~~~~m_N \leq \mu_p \ , ~~~~~~~~~~~~m_N \leq \mu_n \ .
}
The proton and neutron chemical potential $\mu_p$ and $\mu_n$ are
related with the isospin chemical potential $\mu_I$ and the baryon chemical potential $\mu_B$ through $\mu_p=\mu_B+\mu_I/2$ and $\mu_n=\mu_B-\mu_I/2$.
The assumption implies $2n_I^{\rm Baryon}=n_B=0$ in Case-(i) and $2n_I^{\rm Baryon}=n_B=n_p$ in Case-(ii) because $n_I^{\rm Baryon}$ and $n_B$ are expressed as the difference between the proton density $n_p$ and the neutron density $n_n$ and the sum of them, respectively: $n_I^{\rm Baryon}=\frac{n_p-n_n}{2}\ , n_B=n_p+n_n$.
Case-(i) corresponds to the pure mesonic case which is studied in Ref.~\cite{Nishihara:2014nva}.
On the other hand, $n_I^{\rm Baryon}$ and $n_B$ are independent of the each other in Case-(iii).
We will show the results of our analysis in
the Case-(ii) and the Case-(iii)
to compare with the pure mesonic case.
We note that the
four-dimensional part of the gauge symmetry is fixed when
$S_{\rm int}$ is introduced.
In other words,
$S_{\rm int}$ is not invariant under the four-dimensional gauge transformation.
Then, we introduce the
quark number chemical potential $\mu_q$ and the isospin chemical potential $\mu_I$
as the UV boundary values of the time components of the gauge fields
as
\footnote{
The baryon number chemical potential $\mu_B$ is related to $\mu_q$ as $\mu_B\equiv N_c \mu_q$.
}
\bee{
\left. V_0^0\right|_\epsilon=\mu_q-c_{(0)}
\ , \quad
\left. V_0^3\right|_\epsilon=\mu_I-c_{(3)}
\ ,
\label{isospin chemical}
}
where
the constants $c_{(0)}$ and $c_{(3)}$ are corresponding to the degree of freedom of the gauge transformation.
In the next section,
we will determine the values of $c_{(0)}$ and $c_{(3)}$ by the physical requirements
for the pion mass and the equation of state between the chemical potential and the density.
This holographic QCD model involves the following five parameters,
\bee{
g_5^2
\ ,~~
z_m
\ ,~~
m_q
\ ,~~
\lambda
\ ,~~
m^2
\ .
}
To match this model with QCD, the parameter $g_5^2$ is adjusted as~\cite{Erlich:2005qh}
\bee{
\frac{1}{g_5^2}=\frac{N_c}{12\pi^2}
\ .
\label{g_5}
}
For the physical inputs to determine the parameters, we use
the pion mass $m_\pi=139.6$MeV, the pion decay constant $f_\pi=92.4$MeV, the $\rho$ meson mass $m_\rho=775.8$MeV, and the $a_0$ meson mass.
As in Ref.~\cite{Nishihara:2014nva}, we use the $a_0$ meson mass $m_{a_0} = 980\,$MeV as a reference value,
and see the dependence of our results on the scalar meson mass.
The values of the parameters corresponding to $m_{a_0} = 980\,$MeV are determined as
\bee{
z_m=1/(323 {\rm MeV})
\ ,~~
m_q=2.29 {\rm MeV}
\ ,~~
\nn\\
\lambda=4.4
\ ,~~
m^2=5.39
\ .
\label{parameters}
}
As in Ref.~\cite{Nishihara:2014nva}, we assume that the pion condensation phase has the rotational symmetry, $L_i=R_i=0$
\footnote{Note that we also take expectation values of operators made by $L_i$ and $R_i$ such as $\sum_i R_i R_i$, which are invariant under the rotational symmetry, vanish since our present analysis is of the leading order in the large $N_c$ expansion and the hadronic loop contributions are suppressed.
},
and the iso-triplet scalars do not condense, $S^a=0$.
Furthermore, we take $V_0^1=V_0^2=0$ and $A_0^3=\pi^3=0$ which form a set of solutions of the equation of motion (EOM) for these fields.
Similarly, the set of the $\eta=0$ and $A_0^0=0$ satisfies the EOM and we take this solution.
The grand potential density $\Omega$ is given from the Lagrangian $\mc{L}$:
\bee{
\Omega =&-\int_\epsilon^{z_m} dz \mc{L}
\label{omega}
}
where the explicit form of the Lagrangian $\mc{L}$ is shown in \eqref{Lagrangian}.
One can derive the EOM for the vector field $V^0_0$ and $V^3_0$ from \eqref{omega},
\bee{
{} \pd_5\frac{a}{g_5^2}\pd_5 V_0^0
&=n_q \delta (z-z_m+\delta z)
\nn\ ,\\
{} \pd_5\frac{a}{g_5^2}\pd_5 V_0^3
&-{} \frac{a^3\left(S^0\right)^2}{2}\left[2\sin^2 b ~V_0^3
+\theta\sin 2b \sin \zeta \right]
\nn \\
&~~~~~~~~~~~~~~~~~~~
=n_I^{\rm {Baryon}} \delta (z-z_m
+\delta z)
\ .
\label{EOM for v}
}
By parameterizing $V_0^0$ and $V_0^3$ as
\bee{
V_0^0(z)=&{\mu}_q-c_{(0)}+\varphi^0 (z)+g_5^2n_q \int_\epsilon^z d\tilde{z} \tilde{z} \theta (\tilde{z}-z_m+\delta z)
\ ,\nn\\
V_0^3(z)=&\mu_I-c_{(3)}+\varphi^3 (z)+g_5^2n_I^{\rm {Baryon}} \int_\epsilon^z d\tilde{z} \tilde{z} \theta (\tilde{z}-z_m+\delta z)
}
where $\theta$ is a step function,
\eqref{EOM for v} is rewritten as
\bee{
{} \pd_5\frac{a}{g_5^2}\pd_5 \varphi^0
=&0
\ ,
\nn\\
{} \pd_5\frac{a}{g_5^2}\pd_5 \varphi^3
=& \frac{a^3\left(S^0\right)^2}{2}\left[2\sin^2 b ~\braa{\mu_I-c_{(3)}+\varphi^3}\right.
\nn\\&
\left.
~~~~~~~~~~~~~~~~~~~
+\theta\sin 2b \sin \zeta \right]
\ .
\label{eom for phi}
}
Here the boundary conditions are given by
\bee{
\left.V_0^{(0,3)} \right|_\epsilon=\mu_{(q,I)}-c_{(0,3)}
&\ ,~~
\left.\partial_5V_0^{(0,3)} \right|_{z_m}=0
\nn\\
\ra~~
\left.\varphi^{(0,3)} \right|_\epsilon=0
&\ ,~~
\left\{
\mt{
\left.\partial_5\varphi^{0} \right|_{z_m}=&-g_5^2z_m n_{q}
\\
\left.\partial_5\varphi^{3} \right|_{z_m}=&-g_5^2z_m n_I^{{\rm {Baryon}}}
}
\right.
\ .
\label{BD for V}
}
\section{Symmetry energy and delay of the phase transition}
\label{sec:delay}
In this section we first
study the dependence of the pion mass on the isospin chemical potential $\mu_I$ in the normal hadron phase to show the delay of the pion condensation compared with the pure mesonic matter studied in Ref.~\cite{Nishihara:2014nva}.
Next, we investigate the symmetry energy to check whether the present way to introduce the baryonic matter works well in the normal hadron phase by comparing our result with its empirical value.
For studying the hadron phase we set $b=0$ and $\theta=0$ in the equations of motion in Eqs.~(\ref{EOM for v}) and (\ref{eom for phi}).
In Case-(iii), we first derive the relation between the chemical potential $\mu_I$ and the isospin number density $n_I$.
For $b=\theta=0$, it is easy to solve the equation of motion (\ref{eom for phi}) with the boundary conditions in \eqref{BD for V} to have
\bee{
\varphi^3=&-\frac{g_5^2 n_I^{\rm {Baryon}}}{2} z^2
\ .
}
Substituting this solution into \eqref{omega},
we obtain
\bee{
\Omega
\supset&
{} -\int_\epsilon^{z_m} dz~\frac{a}{2g_5^2}\left(\partial_5 V_0^3 \right)^2
-n_I^{\rm {Baryon}} \left.V_0^3\right|_{z_m-\delta z}
\nn\\
=&
{} \frac{g_5^2 z_m^2}{4} \braa{n_I^{\rm {Baryon}}}^2
-\braa{\mu_I-c_{(3)}}n_I^{\rm {Baryon}}
}
where $a=\frac{1}{z}$. Minimizing the $\Omega$
in terms of
the $n_I^{\rm {Baryon}}$ for a given value of the isospin chemical potential $\mu_I$ yields the relation between the isospin chemical potential $\mu_I$ and the isospin number
density of the asymmetric matter $n_I^{\rm {Baryon}}$:
\bee{
n_I^{\rm {Baryon}}=\frac{2}{g_5^2z_m^2}\braa{\mu_I-c_{(3)}}
\ .
\label{nI.vs.mu-hadron_phase}
}
Here in the normal hadron phase the isospin density $n_I$ equals to the density $n_I^{\rm {Baryon}}$ because mesons carrying isospin charge do not condense.
Similarly, for the quark number density we also have
\bee{
n_q=\frac{2}{g_5^2z_m^2}\braa{\mu_q-c_{(0)}}
\ .
}
The baryon number density, $n_B=n_q/N_c$, appears when the baryon chemical potential $\mu_N$ is larger than the mass of nucleon $m_N=939$MeV.
This implies that the $c_{(0)}$ is determined as $c_{(0)}= m_N/N_c $,
as in Ref.~\cite{Kim:2007zm}.
This argument yields the following relation:
\bee{
n_B=\frac{2}{g_5^2z_m^2N_c^2}\braa{\mu_B-m_N}
\ .
}
Let us next study the $\mu_I$-dependence of the pion mass.
The equations of motion for the pion fluctuation up till the quadratic order in the momentum space are given by
\bee{
&
-\frac{1}{a \braa{aS^0}^2}\partial_5 \brac{a\braa{aS^0}^2\partial_5 \pi^\pm}
=
E^{\pm}(\mu_I)
\left[ A_0^\pm +\pi^\pm E^{\pm}(\mu_I)\right]
\ ,
\nn \\ &
\frac{1}{a(aS^0)^2}\partial_5(\frac{a}{g_5^2}\partial_5 A_0^\pm)
=\left[ A_0^\pm+ \pi^\pm E^{\pm} (\mu_I)\right]
\label{eom for pion}
}
where the fields are parameterized as
\bee{
\pi^{\pm}=\frac{\pi^1\mp i\pi^2}{\sqrt{2}}~,~~A_0^{\pm}=i\frac{A_0^1 \mp i A_0^2}{\sqrt{2}}
}
and
\bee{
E^{\pm}(\mu_I)=M\pm\mu_I\left(1-\frac{z^2}{z_m^2}\right)\mp c_{(3)} \frac{z^2}{z_m^2}
\label{eigen value}
\ .
}
$S^0$ is the solution of \eqref{S^0} and $M$ is the energy of the static pion.
Equation~(\ref{eom for pion}) together with the boundary conditions,
$\left.\pi\right|_\epsilon=\left.\pd_5\pi\right|_{z_m}=0$, yield
the value of the $M$ as the eigenvalue.
The lowest value of the eigenvalue $M$ is identified with the pion mass, $m_\pi^{*}$.
Here the parameter $c_{(3)}$ is determined as zero by assuming that $\pi^+$ and $\pi^-$ are degenerating at $\mu_I=0$:
$E^{\pm}(\mu_I)=M\pm\mu_I\braa{1-z^2/z_m^2}$.
Figure~\ref{fig1} shows the $\mu_I$ dependence of the pion mass in the normal hadron phase.
The $\pi^-$ mass drawn by the red curve increases with the isospin chemical potential.
The $\pi^+$ mass by the green curve, on the other hand, decreases and reaches zero
at
$\mu_I=235$MeV,
which implies that the $\pi^+$ condenses and that the transition to the pion condensation phase occurs.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=70mm]{fig1_v5.eps}
\end{center}
\caption[]{
$\mu_I$ dependence of the pion masses.
The red and green curves show the masses of $\pi^-$ and $\pi^+$, respectively.
We also show the $\mu_I$ dependence of the $\pi^+$ mass in the pure mesonic matter obtained in Ref.~\cite{Nishihara:2014nva} by the blue curve.
}
\label{fig1}
\end{figure}
We would like to stress that the $\pi^+$ mass here decreases more slowly than the one obtained
in the pure mesonic matter
shown by the blue curve.
One can easily see that the critical value of the isospin chemical potential for the phase transition is larger than the pion mass for the pure mesonic case.
This is due to the existence of the baryons in the matter, which can also be understood by an analysis of the chiral Lagrangian based on the Hidden Local Symmetry as shown in section~\ref{sec:HLS}.
The energy density of the system is defined as $\mc{E}=\Omega+n_q \mu_q+n_I \mu_I $ at zero temperature and
given by
\bee{
\mc{E}=&
{} \frac{g_5^2 z_m^2}{4} \braa{n_I}^2
{}+ \frac{g_5^2 z_m^2N_c^2}{4} \braa{n_B}^2
\ .
}
Then, the symmetry energy is obtained as
\footnote{Note that this definition of the symmetry energy is different from the one used in Ref.~\cite{Park:2011zp}.}
\bee{
E_{\rm sym}\braa{n_B}\equiv\left.\frac{\pd \braa{\mc{E}/n_B}}{\pd \alpha^2}\right|_{\alpha=0}
=\frac{g_5^2 z_m^2}{16} n_B
}
where $\alpha\equiv\frac{2n_I}{n_B}$.
At the saturation density $n_0=0.16 {\rm fm}^{-3}$, we can estimate $E_{\rm sym}\braa{n_0}=29 {\rm MeV}$ by using \eqref{g_5} and \eqref{parameters}, which is comparable to the empirical value of $32. 3 \pm 1.0$\,MeV \cite{Zhang:2013wna}.
In Refs.~\cite{Chen:2004si,Shetty:2007zg}, the value of the parameter $\gamma$ defined as $E_{\rm sym}(n_B)=E_{\rm sym}(\rho_0)\braa{\frac{n_B}{n_0}}^\gamma$ is estimated as $\gamma=0.55$ - $0.69$, which is different from the result of the present analysis, $\gamma=1$.
The slope parameter of the symmetry energy is give by
\bee{
L\equiv
3n_0\left.\frac{\pd E_{\rm sym} (n_B)}{\pd n_B}\right|_{n_B=n_0}
=3n_0\frac{g_5^2 z_m^2}{16}
}
and its value is estimated as $L=87$MeV, where its empirical value is known as $45.2 \pm 10.0$\,MeV \cite{Zhang:2013wna}.
We may understand that the deviations of values of $\gamma$ and $L$ are caused by the next leading order in the large $N_c$ expansion.
In Case-(ii), as we stated
in section~\ref{sec:model}, $2n_I^{\rm Baryon}$ and $n_B$ are equal to $n_p$, which
leads to
the following solutions of \eqref{eom for phi} in the normal hadron phase, $b=\theta=0$:
\bee{
\varphi^3=&-\frac{g_5^2 n_I^{\rm {Baryon}}}{2} z^2=-2\frac{g_5^2 n_p}{2} z^2
\ ,\nn\\
\varphi^0=&-\frac{g_5^2 n_q}{2} z^2=-N_c\frac{g_5^2 n_p}{2} z^2
\ .
}
Now,
the grand potential density is given by
\bee{
\Omega
\supset&
{} -\int_\epsilon^{z_m} dz~\frac{a}{2g_5^2}\brac{\left(\partial_5 V_0^0 \right)^2 +\left(\partial_5 V_0^3 \right)^2 }
\nn\\
&-n_I^{\rm {Baryon}} \left.V_0^3\right|_{z_m-\delta z}
-n_q \left.V_0^0\right|_{z_m-\delta z}
\nn\\
=&
{}\frac{g_5^2}{4}\braa{{N_c^2}+\frac{1}{4}}z_m^2{n_p^2}
-\braa{ \mu_p-N_c c_{(0)}-\frac{c_{(3)}}{2}}n_p
\ ,
}
where we used $2n_I^{\rm {Baryon}}=n_B=n_p$ and $\mu_p= \mu_B+\frac{\mu_I}{2}$.
Minimizing this in terms of $n_p$,
we have
\bee{
n_p=\frac{2}{g_5^2z_m^2}\frac{4}{1+4N_c^2}\braa{\mu_p-c_{(0)}-\frac{c_{(3)}}{2}}
}
Thus, we
set $c_{(0)}+\frac{c_{(3)}}{2}=m_N$
because the proton density $n_p$ must vanish as $\mu_p \ra m_N$.
In Case-(ii), the pion mass depends on not only $\mu_I$ but also $\mu_B$ through
\bee{
E^{\pm}(\mu_I)=&M\pm\mu_I\braa{1-\frac{1}{1+4N_c^2}\frac{z^2}{ z_m^2}
}
\nn\\&
\mp2\frac{1}{1+4N_c^2}\braa{\mu_B-m_N}
\frac{z^2}{ z_m^2}
\mp c_{(3)}\frac{z^2}{ z_m^2}
}
in \eqref{eigen value}.
In the limit $\mu_I\ra 0$ at $\mu_B=m_N$
\footnote{
We can not take the limit $\mu_I\ra 0$ at $\mu_B\neq m_N$ in Case-(ii).
}
the degeneration of the charged pions gives us $c_{(3)}=0$.
The $\mu_I$ dependence of the pion mass is the almost same
as that in
the pure mesonic case (Case-(i)) because
the difference is suppressed by
the factor $\frac{1}{1+4N_c^2}=\frac{1}{37}$.
\section{Pion Condensation Phase}
\label{sec:pion condensation}
Next, we study the equation of state
in the asymmetric nuclear matter.
First, we will perform the following analysis in Case-(ii) and in Case-(iii),
separately, and show the results of in Case-(iii).
In the last of this section, our results on the ($\mu_I$, $\mu_B$) plane will be shown, which are given by combining the result of each Case.
\newpage
\begin{widetext}
{}From the Lagrangian
\eqref{Lagrangian}, the equations of motion are obtained as
\bee{
\partial_5 \left( -a^3 \partial_5 S^0 \right) + a^3 S^0 \left(\partial_5 b\right)^2
-3a^5S^0
-a^3S^0\left[\sin^2 b \; ( \varphi^3 +\mu_I )^2
+\theta\sin 2b \;\sin \zeta \; (\varphi^3 +\mu_I)
+\theta^2 - \theta^2 \sin^2 b \; \sin^2 \zeta \right] =&0
\ ,
\nn \\
\partial_5 \left( - a^3 \left(S^0\right)^2 \partial_5 b \right)
-\frac{a^3\left(S^0\right)^2}{2}\left[\sin 2b \; \left\{( \varphi^3 +\mu_I )^2
- \theta^2 \sin^2 \zeta\right\}+2\theta \cos 2b \;\sin \zeta
~( \varphi^3 +\mu_I )\right]=&0
\ ,
\nn \\
\partial_5\left(\frac{a}{g_5^2}\partial_5 \theta \right)
-\frac{a}{g_5^2}\theta \left( \partial_5 \zeta \right)^2
-\frac{a^3\left(S^0\right)^2}{2}\left[\sin 2b \; \sin \zeta
~( \varphi^3 +\mu_I )+ 2\theta \left\{ 1 - \sin^2 b \;\sin^2 \zeta \right\}\right]=&0
\ ,
\nn \\
\partial_5\left(\frac{a}{g_5^2}\theta^2\partial_5 \zeta \right)
-\frac{a^3\left(S^0\right)^2}{2}\left[\theta \sin 2b \;\cos \zeta \;\;( \varphi^3 +\mu_I )
- \theta^2 \sin^2 b\sin 2 \zeta\right]=&0
\ ,
\nn \\
\partial_5 \left( \frac{a}{g^2_5} \partial_5 \varphi^3 \right)
-\frac{a^3\left(S^0\right)^2}{2}\left[2\sin^2 b
\;( \varphi^3 +\mu_I )+\theta\sin 2b \;\sin \zeta \right]=&0
\ .
\label{EOM}
}
\end{widetext}
These differential equations are solved with the boundary conditions listed in Table \ref{Boundary conditions}.
\begin{table}[h]
\begin{center}
\begin{tabular}{ccc}\hline\hline
Variables & UV & IR
\\\hline
$S^0$ & $\frac{S^0}{z} |_\epsilon=m_q$ &
$ \partial_5 S^0 |_{z_m}=
\left. -\frac{S^0}{2z_m} \left(\lambda \left(S^0\right)^2 - 2 m^2 \right) \right\vert_{z_m}$
\\
$b$ & $b |_\epsilon=0$ & $\partial_5 b |_{z_m}=0$
\\
$\theta$ & $\theta |_\epsilon=0$ & $\partial_5 \theta |_{z_m}=0$
\\
$\zeta$ & $\zeta |_\epsilon=\frac{\pi}{2}$ & $\partial_5 \zeta |_{z_m}=0$
\\
$\varphi^3$ & $\varphi^3 |_\epsilon=0$ &
\begin{tabular}{ll}
$\varphi^3 |_{z_m}=-\mu_I$ & in Case-(iii) \\
(IR condition) & in Case-(ii) \\
\end{tabular}
\\
\hline\hline
\end{tabular}
\end{center}
\caption{Boundary conditions for the relevant wave functions.
(IR condition) implies that $N_c^2\left.\pd_5 \varphi^3\right|_{z_m} +\frac{1}{2}\left.\varphi^3\right|_{z_m}+\mu_B+\frac{\mu_I}{2}-m_N=0$ is satisfied.
}
\label{Boundary conditions}
\end{table}
We note that, in Case-(iii), the IR boundary condition for $\varphi^3$ and the value of $n_I^{\rm Baryon}$ are determined by
the minimization condition for the grand potential
\bee{
0 =
\frac{\pd \Omega}{\pd n_I^{\rm {Baryon}}}=&-\int dzV_0^3
\delta\left(z-z_m+\delta z\right)
\nn\\
=& -\mu_I-\left.\varphi^3 \right|_{z_m-\delta z}
\ ,
}
and the condition in \eqref{BD for V}
\bee{
n_I^{\rm {Baryon}}=\left.-\frac{1}{g_5^2}\frac{\partial_5 \varphi^3 }{z}\right|_{z_m}
\ .
}
In Case-(ii),
on the other hand,
the condition
$\pd \Omega / \pd n_p=0$ together with
the solution of $\varphi^0$ and the condition in \eqref{BD for V}
leads to
\begin{equation}
N_c^2\left.\pd_5 \varphi^3\right|_{z_m} +\frac{1}{2}\left.\varphi^3\right|_{z_m}+\mu_B+\frac{\mu_I}{2}-m_N=0
\ .
\end{equation}
It should be noticed that the condition provides the $\mu_B$ dependence of the isospin number density $n_I$ in Case-(ii), although the coupled equations of motion in Eq.~(\ref{EOM}) do not include $\mu_B$.
On the other hand, in Case-(iii), neither the boundary conditions nor the equations of motion have $\mu_B$ dependence, which implies that the isospin number density is independent of the baryon number chemical potential.
Now, let us study the isospin number density in Case-(iii), which
is defined by
\bee{
n_I=-\frac{\pd \Omega}{\pd \mu_I}=n_I^{\rm {Meson}}+n_I^{\rm {Baryon}}
}
where the $n_I^{\rm {Meson}}$ expresses the mesonic contribution to the isospin number density given by
\bee{
n_I^{\rm {Meson}}=\int dz \frac{a^3\left(S^0\right)^2}{2}\left[2\sin^2 b ( \varphi +\mu_I )+\theta\sin 2b \sin \zeta \right]
\ .
}
In Fig.~\ref{fig2}, we show the resultant equation of state between the isospin density and the isospin chemical potential obtained by solving \eqref{EOM}.
For $\mu_I < 235$\,MeV there is no pion condensation, so that the isospin number density increases linearly with the chemical potential following \eqref{nI.vs.mu-hadron_phase} as drawn by the red curve in Fig.~\ref{fig2}. At $\mu_I^c = 235$MeV the phase transition occurs from the normal hadron phase to the pion condensation phase.
This critical chemical potential $\mu_I^c = 235$\,MeV is consistent with the one determined from the pion mass shown in \figref{fig1}, but the value
is larger than the critical value for the pure mesonic matter, for which $\mu_I^c = m_\pi$~\cite{Nishihara:2014nva} as seen by the green curve in Fig.~\ref{fig2}.
This delay of the phase transition is due to the existence of the baryons, which can also be understood by an analysis of the four dimensional chiral Lagrangian as shown in the next section.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=70mm]{fig2_v4.eps}
\end{center}
\caption[]{
Equation of state
between the isospin chemical potential $\mu_I$ and the isospin number density $n_I$ drawn by the red curve.
The blue and green curves show the $\mu_I$ dependences of the mesonic contribution $n_I^{\rm {Meson}}$ and the baryonic contribution $n_I^{\rm {Baryon}}$, respectively.
The pink
dots show the result shown in Ref.~\cite{Nishihara:2014nva} for pure mesonic matter.}
\label{fig2}
\end{figure}
In the pion condensation phase, the pion contribution to the isospin number density increases monotonically with the chemical potential as shown by the pink curve in Fig.~\ref{fig2}, while the baryonic contribution by the blue curve is almost constant: $n_I^{\rm {Baryon}} \sim 0.2\ {\rm fm}^{-3}$.
As a result the mesonic contribution dominates the isospin number density.
This implies that the energy provided by the isospin chemical potential is mostly used for generating the pion condensation rather than converting the neutron into proton.
Figure \ref{fig7} shows the dependence of the equation of state on the scalar meson mass.
The value of parameter $\lambda$ is determined from the mass of the $a_0$ meson,
where $\lambda=1.0$, $4.4$ and $100$ correspond to the $m_{a_0}=610, 980$ and $1210$ MeV, respectively.
We find that the critical value of the isospin chemical potential is independent of $\lambda$ and the behavior of the equation of state is not sensitive to the value of $\lambda$.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=70mm]{fig7_v2.eps}
\end{center}
\caption[]{
Dependence of the equation of state on the value of $\lambda$.
The green, red and blue curves are for $\lambda=1.0, 4.4$ and $100$, respectively.
}
\label{fig7}
\end{figure}
As we stated in the introduction, the existence of the isospin chemical potential $\mu_I$ explicitly breaks the chiral symmetry group SU$(2)_R\times$ SU$(2)_L$ down to its
subgroup $\mbox{U}(1)_R^{(3)} \times \mbox{U}(1)_L^{(3)} = \mbox{U}(1)_V^{(3)} \times \mbox{U}(1)_A^{(3)}$, where the superscript $^{(3)}$ implies that the generator $T_3$ of SU(2) is used for the U(1) as $\exp[ i \theta_V T_3 ] \in \mbox{U}(1)_V^{(3)}$.
For studying the order parameters for the phase transition, we define the following $\pi$-condensate and the ``$\sigma$''-condensate~\cite{Nishihara:2014nva}:
\begin{eqnarray}
\langle\pi^a\rangle
&\equiv&
\frac{1}{2}\mathrm{Tr} \left[i\sigma^a a\left(\partial_5 \frac{X}{z}\right)
+{\rm h.c.}\right]_\epsilon
=\langle \bar{q} \gamma_5 \sigma^a q \rangle
\ ,
\nonumber \\
\langle\sigma\rangle
&\equiv&
\frac{1}{2}\mathrm{Tr} \left[ a\left(\partial_5 \frac{X}{z}\right)
+{\rm h.c.}\right]_\epsilon
= \langle \bar{q} q \rangle
\ .
\end{eqnarray}
We plot the ``$\sigma$''-condensate and the $\pi$-condensate obtained by the present analysis in Fig.~\ref{fig3}, together with those condensates for the pure mesonic matter.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=70mm]{fig3_v5.eps}
\end{center}
\caption{
Dependence of the ``$\sigma$''-condensate $\langle \sigma \rangle$ and the $\pi$-condensate $\langle \pi \rangle$ on the isospin chemical potential $\mu_I$.
The condensates are scaled by the ``$\sigma$''-condensate at the vacuum indicated by $\langle \sigma \rangle_0$.
The blue curves are the dependence of $\langle \sigma \rangle$ and $\langle \pi \rangle$ shown in Ref.~\cite{Nishihara:2014nva} for the pure mesonic matter.
}
\label{fig3}
\end{figure}
This figure shows that the present behavior is quite similar to the previous one except the difference of the phase transition point:
In the normal hadron phase the ``$\sigma$''-condensate exists, which leads to the break down of
the $\mbox{U}(1)_A^{(3)}$ symmetry, but $\pi$-condensate is zero.
At the phase transition point, the $\pi$-condensate appears, which spontaneously breaks the $\mbox{U}(1)_V^{(3)}$ symmetry, while the ``$\sigma$''-condensate starts to decrease very rapidly.
For large $\mu_I$, the ``$\sigma$''-condensate is almost zero while the $\pi$-condensate keeps increasing.
We next show the chiral circle in Fig.~\ref{fig4}.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=70mm]{fig4_v4.eps}
\end{center}
\caption{
Chiral circle shown by red curve.
The black curve is an unit circle and the green curve is the chiral circle for the pure mesonic matter.
}
\label{fig4}
\end{figure}
The red solid curve shows that the behavior for the nuclear matter is quite similar to the one for the pure mesonic matter shown by the green dotted line:
Although the ``$\sigma$''-condensate decreases and the $\pi$-condensate increases, the chiral condensate defined by
\begin{equation}
\tilde{\sigma} \equiv \sqrt{ \langle \sigma \rangle^2 + \langle \pi^a \rangle^2 }
\end{equation}
stays constant until about 150\,MeV above the critical chemical potential.
In the large $\mu_I$ region, the chiral condensate $\tilde{\sigma}$ grows very rapidly.
This implies that the enhancement of the chiral symmetry breaking occurs in the asymmetric nuclear matter, similarly to the one in the pure mesonic matter as shown in Ref.~\cite{Nishihara:2014nva}.
Next, we
study the equation of state and the condensates in Case-(ii) as well as those in Case-(i) in a similar way.
The resultant
$\langle\pi^1\rangle$, $\langle\sigma\rangle$ and $n_I$
in the entire ($\mu_B$, $\mu_I$) plane
are shown in Fig.~\ref{fig8},~\ref{fig9} and \ref{fig10}, respectively.
The green lines in these figures show the boundary between Case-(i) and Case-(ii) and that between Case-(ii) and Case-(iii), which are corresponding to $\mu_p=m_N$ and $\mu_n=m_N$.
Figure~\ref{fig8} shows that there is the first order transition on the boundary between the pion condensation phase in Case-(ii) and the normal hadron phase in Case-(iii).
In Fig.~\ref{fig9}, we see that $\langle\sigma\rangle$ decreases discontinuously at the first order transition line in response to sudden increase of $\langle\pi^1\rangle$ in Fig.~\ref{fig8}.
Figure~\ref{fig10} shows the equation of state on the ($\mu_B$, $\mu_I$) plane.
In these figures, the values of $\langle\pi^1\rangle$, $\langle\sigma\rangle$ and $n_I$ drastically change on $\mu_n=m_N$ which is the boundary between Case-(ii) and Case-(iii).
\begin{figure}[ht]
\begin{center}
\includegraphics[width=70mm]{fig8_v7.eps}
\end{center}
\vspace{0.2cm}
\caption{
$\frac{\langle\pi^1\rangle}{\langle\sigma\rangle_0}$ vs. $\mu_B$ vs. $\mu_I$.
The green lines are the boundaries between Case-(i) and Case-(ii) and between Case-(ii) and Case-(iii).
}
\label{fig8}
\end{figure}
\begin{figure}[ht]
\begin{center}
\includegraphics[width=70mm]{fig9_v7.eps}
\vspace{0.2cm}
\end{center}
\caption{
$\frac{\langle\sigma\rangle}{\langle\sigma\rangle_0}$ vs. $\mu_B$ vs. $\mu_I$.
The green lines are the boundaries between Case-(i) and Case-(ii) and between Case-(ii) and Case-(iii).
}
\label{fig9}
\end{figure}
\begin{figure}[ht]
\begin{center}
\includegraphics[width=70mm]{fig10_v6.eps}
\end{center}
\caption{
$n_I$ vs. $\mu_B$ vs. $\mu_I$.
The green lines are the boundaries between Case-(i) and Case-(ii) and between Case-(ii) and Case-(iii).
}
\label{fig10}
\end{figure}
\section{An analysis by the chiral Lagrangian based on the Hidden Local Symmetry}
\label{sec:HLS}
In this section, we show that the delay of the phase transition to the pion condensation phase is understood as the baryonic matter effect in the framework of the four dimensional chiral Lagrangian including the $\rho$ meson based on the hidden local symmetry (HLS)~\cite{Bando:1987br,Harada:2003jx}.
The mesonic part of the HLS Lagrangian is given by
\bee{
\mathcal{L}=&F_\pi^2 \tr\brac{\hat{\alpha}_{\perp \mu}\hat{\alpha}_\perp^\mu}
+aF_\pi^2 \tr\brac{\hat{\alpha}_{\parallel \mu}\hat{\alpha}_\parallel^\mu}
\nn\\
&+\frac{F_\pi^2}{4}\tr\brac{\xi_L \chi \xi_R^\dagger+\xi_R \chi^\dagger \xi_L^\dagger}
-\frac{1}{2g^2}\tr\brac{V_{\mu\nu}V^{\mu\nu}}
}
where $\chi$ is an external field which has the expectation value corresponding to the pion mass, $\brae{\chi}=m_\pi^2 {\bf 1}$.
The $\hat{\alpha}_{\perp \mu}$ and $\hat{\alpha}_{\parallel \mu}$ are defined as
\bee{
\hat{\alpha}_{\perp, \parallel \mu}=&\frac{D_\mu\xi_{L} \cdot \xi_{L}^\dagger\pm D_\mu\xi_{R} \cdot\xi_{R}^\dagger}{2i}
}
where $\xi_{L,R}$ are the fields including pions, $V_\mu$ is the gauge field including the rho and omega mesons and the covariant derivative of these fields are
\bee{
D_\mu\xi_{L}=&\partial_\mu \xi_{L}-iV_\mu\xi_{L}+i\xi_{L}\mathcal{L}_\mu
\ ,\nn\\
D_\mu\xi_{R}=&\partial_\mu \xi_{R}-iV_\mu\xi_{R}+i\xi_{R}\mathcal{R}_\mu
\ .
}
The baryon and isospin chemical potentials, $\mu_B$ and $\mu_I$, are introduced as the expectation value of the time component of the external gauge fields: $\brae{\mc{L}_0}=\brae{\mc{R}_0}=\frac{\mu_B}{2}\sigma^0+\frac{\mu_I}{2}\sigma^3$.
Here we introduce the following terms including the baryons explicitly:
\bee{
\bar{N} i \gamma^\mu D_\mu N+G\bar{N}\gamma^\mu \hat{\alpha}_{\mu \parallel}N
}
where $N$ is the baryon field and $D_\mu$ is a covariant derivative defined as $D_\mu N = \left( \partial_\mu - i V_\mu \right) N$.
We replace the bilinear baryon fields by the mean field as
\begin{equation}
\braa{V_0^{3} +G\hat{\alpha}_{\parallel 0}^3}n_I^{\rm {Baryon}}
+\braa{V_0^{0} +G\hat{\alpha}_{\parallel 0}^0}n_B
\end{equation}
where $\hat{\alpha}_{\parallel 0}^{(0,3)}=\tr\braa{\hat{\alpha}_{\parallel 0}\sigma^{(0,3)} }$, $V_0^{3}$ is the time component of the neutral rho meson and $V_0^{0}$ is the time component of the omega meson.
Taking the unitary gauge of the HLS and integrating out the rho and omega mesons and assuming the rotational symmetry, we obtain the following effective Lagrangian for the pion coupling to the baryonic sources:
\bee{
\mathcal{L}=&F_\pi^2 \tr\brac{\hat{\alpha}_{\perp 0}\hat{\alpha}_\perp^0}
+\frac{F_\pi^2}{4}\tr\brac{\xi_L \chi \xi_R^\dagger+\xi_R \chi^\dagger \xi_L^\dagger}
\nn\\&
-\frac{1}{2a'F_\pi^2}n_B^2-\frac{1}{2a'F_\pi^2}\braa{n_I^{\rm {Baryon}}}^2
+\mu_Bn_B
+{\alpha}_{\parallel 0}^3n_I^{\rm {Baryon}}
\label{the effective Lagrangian based on the HLS}
}
where $a'\equiv\frac{a}{\braa{1-G}^2} $ \footnote{Since the value of the parameter $a$ is known as about two in the HLS~\cite{Harada:2003jx}, $a'=\frac{a}{\braa{1-G}^2} $ is larger than zero.} and ${\alpha}_\parallel^\mu=\hat{\alpha}_\parallel^\mu+V^\mu$.
Existence of the terms in the last line of \eqref{the effective Lagrangian based on the HLS} causes the deviation from the result obtained by the $\mc{O}(p^2)$ chiral Lagrangian without the baryonic sources, which delays the transition to the pion condensation comparing to of the pure mesonic analysis.
Figure \ref{fig6} shows the relation between the pion mass and the isospin chemical potential for $a'=0.7$ (green), $0.5$ (blue), $0.3$ (pink) and of the holographic QCD model (the red curve). The dotted black line corresponds to the case for the pure pion matter, $a'=0$.
This figure shows that the point at which the curve reaches zero depends on the value of $a'$.
The critical value of the isospin chemical potential for $0 < a' < 1$
is larger than the pion mass, which implies that
delay of the transition is understood by using a model based on the HLS with the baryonic sources.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=70mm]{fig6_v5.eps}
\end{center}
\caption{
The $\mu_I$ dependence of the $\pi^+$ mass obtained from the chiral Lagrangian.
The green, blue and pink curves are the results for $a'=0.7, 0.5$ and $0.3$ respectively. The $\pi^+$ mass given from the analysis of the holographic QCD model is indicated by the red curve.
We also show the $\mu_I$ dependence of the $\pi^+$ mass in the pure mesonic matter obtained in Ref.~\cite{Nishihara:2014nva} by the dotted black line.
}
\label{fig6}
\end{figure}
\section{A summary and discussions}
\label{sec:summary}
We introduced a baryonic source at the IR boundary coupling to the iso-triplet vector meson in the hard wall holographic QCD mode, and studied the pion condensation in the asymmetric nuclear matter.
We showed that the phase transition from the normal matter to the pion condensation phase is delayed in the asymmetric nuclear matter compared with the pure mesonic matter.
Furthermore, our result shows that the meson contribution to the isospin number density increases with the chemical potential, while the baryon contribution stays constant.
We would like to stress
that the chiral symmetry breaking is enhanced in the asymmetric nuclear matter as in the pure mesonic matter.
We show the phase diagram obtained from the present analysis in \figref{fig5}, where the blue and red area express the hadron phase and the pion condensation phase respectively.
The phase transition is of the second order except on the yellow line expressing the first order.
In Case-(i), the phase transition to the pion condensation occurs at which the isospin chemical potential is equal to the pion mass as shown in Ref.~\cite{Nishihara:2014nva}.
On the other hand, in Case-(iii), done by the present analysis, the critical point of the transition is delayed compared
with in Case-(i).
A similar delay also occurs in Case-(ii), although the effect is very tiny and it is hard to see in Fig.~\ref{fig5}.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=60mm]{fig5_v14.eps}
\end{center}
\caption{
Phase diagram: $\mu_B$ vs. $\mu_I$.
The blue and red area express the hadron phase and the pion condensation phase respectively.
}
\label{fig5}
\end{figure}
The model which we used in section \ref{sec:HLS}
explicitly includes
the rho and omega mesons.
The existence of the rho meson is essential for the delay of the phase transition.
This indicate that the phase transition point in the NJL model may be changed
by including
the following vector 4-Fermi interaction~\cite{Asakawa:1989bq,Kitazawa:2002bc,Fukushima:2008is,Zhang:2009mk,Bratovic:2012qs,Zhang:2013oia}:
\bee{
g_v\brac{\braa{\bar{\psi}\sigma^a\gamma_\mu \psi }^2+\braa{\bar{\psi}\sigma^a\gamma_\mu \gamma_5\psi }^2}
}
where $g_v$ is a positive coupling constant, $\psi$ is a quark field and $\sigma^a$ are Pauli matrices in the flavor space.
In the present analysis, we put the baryonic charge at the IR boundary. In more general case, the charge is spread into the bulk by the gauge interaction. Furthermore, the coupling of the baryon to the scalar mesons is not included.
Such effects could be included by the holographic mean field approach \cite{Harada:2011aa,He:2013gta}, which is left for future publication.
References \cite{Lee:2013oya,Park:2011zp,Jo:2009xr} studied
the asymmetric matter in the hard wall holographic QCD model.
Our results for the meson mass splitting and the symmetry energy are comparable to their results.
\section*{ACKNOWLEDGEMENTS}
The authors would like to thank Kenji Fukushima for stimulating discussion on the symmetry structure.
This work was supported in part by Grant-in-Aid for Scientific Research
on Innovative Areas (No. 2104) ``Quest on New Hadrons with Variety of Flavors'' from MEXT, and
the JSPS Grant-in-Aid for Scientific Research
(S) No.~22224003, (c) No.~24540266.
|
\section{Introduction}
It is an elementary exercise to show that if a series converges, then the sequence formed by its individual terms must tend to zero. The converse, however, does not follow; the archetype here is the harmonic series
\begin{equation} \label{harmonicseries}
1+\frac{1}{2} + \frac{1}{3} + \frac{1}{4} + \cdots
\end{equation}
which diverges to infinity. In 1914, Kempner \cite{kempner} proved that if one omits from the above series all terms whose denominator contains the digit 9, the resulting series will converge to a finite value. Kempner went on to show that this value was less than 80, and Baillie \cite{baillie} later gave the result to 20 decimal places; the sum is approximately $22.92067$.
There is nothing special about the digit 9. One can, in fact, fix any finite string of digits and omit all terms from (\ref{harmonicseries}) whose denominator contains this string to get a convergent series. A proof of this can be found in Section 9.9 of Hardy and Wright's well-known text \cite{hardywright}. It should also be noted that Schmelzer and Baillie \cite{schmelzerbaillie} have provided an algorithm to compute the value of these series to high precision. It follows, from the convergence of such series, that we have the somewhat paradoxical statement that almost all integers contain any given string of digits. This is meant in the following way: if we fix a string $S$ and let $R(x)$ count the positive integers up to $x$ which do not contain the string $S$, then
$$\lim_{x \rightarrow \infty} \frac{R(x)}{x} = 0.$$
The purpose of this paper is to consider the consequences of this result in the setting of prime numbers. It is known that the series of the reciprocals of the primes diverges and, as this is a subseries of the harmonic series, it follows that there are infinitely many primes which contain any given string (see Behforooz \cite{behforooz}). Our first result gives an explicit upper bound on the least prime containing some fixed string.
\begin{thm} \label{primebound}
Let $S$ be a string of $l$ digits. Then there exists a prime number $p$ containing $S$ within its decimal representation and satisfying the bound
\begin{equation} \label{thmeqn}
\frac{\log p}{\log \log p} \leq 5.7 l^2 \times 10^l.
\end{equation}
\end{thm}
Actually, the proof of this theorem gives something stronger, for it does not rely on the composition of the string but only its length. As such, for \textit{any} string $S$ of length $l$ there exists a prime $p \leq N$ such that $p$ contains $S$ in its decimal representation and
\begin{equation} \label{forn}
\frac{\log N}{\log \log N} = 5.7 l^2 \times 10^l.
\end{equation}
This bound is considerably weak when contrasted with some calculations. Using text files containing the first two million primes (see Caldwell's tables \cite{caldwell}), we wrote a program in Java which, given a positive integer $l$, would return the least number $M$ such that all strings of length $l$ were contained in primes at most $M$. Table 1 compares these computations with the approximate values for $N$ we get by solving (\ref{forn}).
\begin{table}[ht]
\caption{Experimentally determined values for $M$ against values for $\log N$ given by Theorem 1.}
\label{tab:pixs}
\centering
\begin{tabular}{c c c }
\\ \hline\hline
$l$ & $M$ & $\log N$ \\[0.5 ex] \hline
$1$ & 83 & 330.7 \\
$2$ & 1847 & 22887.4 \\
$3$ & 50411 & 689676\\
$4$ & 793343 & $1.51 \cdot 10^7$ \\
$5$ & 9810001 & $2.77 \cdot 10^8$ \\
\hline\hline
\end{tabular}
\end{table}
Theorem \ref{primebound} thus seems an extremely weak result. There are some potential reasons for this; we see no grounds why certain strings should be favoured by primes nor others by composites, except for at the basic level where even-integer strings should not appear at the end of primes. There are also well-known divisibility rules which depend on decimal digits, though we see no way in which one would implement these.
To this end, the proof of Theorem \ref{primebound} relied on a simple argument. If we consider the integers in the interval $[1,x]$ and let $\pi(x)$ denote the number of primes in this interval and $R(x)$ denote those integers which avoid some string $S$, then we can solve the inequality
$$\pi(x) > R(x)$$
using known estimates for $\pi(x)$.
We can speculate somewhat on the correct order of Theorem \ref{primebound}, using an entertainingly loose argument based on the solution to the coupon collector's problem (see, for example, Isaac \cite{prob}):
\begin{center}
\textit{Suppose that there is an urn of n different coupons, from which coupons are being collected, equally likely, with replacement. What is the expected number of collections required to guarantee that all different coupons have been collected?}
\end{center}
It is known that the expected number is asymptotic to $n \log n$. Our speculation on the correct order of Theorem \ref{primebound} will be questionable for we suppose that, for a given prime with at least $l$ digits, it is equally likely that it will contain any string of length $l$ i.e. there is no bias amongst strings. We must also suppose that there is at most one string in each prime, which is risible to say the least.
Then, consider that for each prime following $10^{l-1}$, we wish to draw from it a string of length $l$. There are $9 \cdot 10^{l-1}$ such strings, and so we expect to have to go through the first
$$9 \cdot 10^{l-1} \log(9 \cdot 10^{l-1})$$
primes which exceed $10^{l-1}$ in order to collect all strings of the given length. By the prime number theorem, there are approximately
$$\frac{10^{l-1}}{(l-1) \log 10}$$
primes which do not exceed $10^{l-1}$. Thus, we should have drawn all strings of length $l$ by $N$, where
$$\pi(N) \approx \frac{10^{l-1}}{(l-1) \log 10} + 9 \cdot 10^{l-1} \log(9 \cdot 10^{l-1}).$$
It is straightforward enough to take $l$ to be large so that we may neglect some terms and then invert the formula to get that
$$N \sim k l^2 10^l$$
for some constant $k$. One may compare this with (\ref{thmeqn}). Though there is not an ample amount of computational data, the values for $M$ in Table 1 do not strongly contradict this.
In Section 3, we give a short proof for the following theorem, which turns out to be a corollary of Theorem 1.2 of Green and Tao's \cite{greentao} celebrated paper and our proof of Theorem \ref{primebound}.
\begin{thm} \label{greentaocorollary}
Let $S$ be a string of length $l$. Then there are arbitrarily long arithmetic progressions of prime numbers whose decimals contain the string $S$.
\end{thm}
\section{Proof of Theorem \ref{primebound}}
Following Hardy and Wright \cite{hardywright}, we will consider integers with decimal representations in base $r$. This will allow for simplicity, for a string of length $l$ will correspond to a single digit if we choose $r = 10^l$. We denote by $v$ an integer whose decimal representation avoids the digit $b$. Then the number of such integers for which $r^{l-1} \leq v < r^l$ is $(r-1)^{l}$ in the case that $b=0$ and $(r-2)(r-1)^{l-1}$ if $b \neq 0$. In either case, the count does not exceed $(r-1)^l$. Thus, if
$$r^{k-1} \leq x < r^k,$$
then the number $R(x)$ of $v$ up to $x$ does not exceed
$$(r-1)+(r-1)^2 + \cdots + (r-1)^k = \frac{(r-1)^{k+1}-(r-1)}{r-2} < \frac{(r-1)^{k+1}}{r-2}.$$
Now, when there are more primes than there are numbers whose decimal representations avoid $b$, we have that there must be some prime whose representation contains $b$. Therefore, we wish to solve the inequality
$$\pi(x) > R(x)$$
for $x$, where $\pi(x)$ denotes the number of primes not exceeding $x$. We can do this by continuously strengthening the inequality to get a simpler inequality involving $x$. We can employ Corollary 1 of Rosser and Schoenfeld \cite{rosserschoenfeld1962}, namely that
\begin{equation} \label{piestimate}
\pi(x) > \frac{x}{\log x}
\end{equation}
for all $ x \geq 17$, with our upper bound for $R(x)$, so that we may instead solve the stronger inequality
\begin{equation} \label{one}
\frac{x}{\log x} > \frac{(r-1)^{k+1}}{r-2}
\end{equation}
for $x \geq 17$. The reader should note that there are sharper estimates than (\ref{piestimate}) available (see Dusart \cite{dusart} for instance) but these do not lead to a significant improvement of the main result; an improvement would only be made on the error terms which we have chosen to neglect.
Using the fact that $r^{k-1} \leq x$, we can rearrange (\ref{one}) to get
$$\log x < \frac{r-2}{r(r-1)} \Big( \frac{r}{r-1} \Big)^k.$$
Taking the logarithm of both sides and using
$$\frac{\log x}{\log r} < k$$
we can get
$$\log \log x < \log\Big( \frac{r-2}{r(r-1)} \Big) + \log x \Big( 1- \frac{\log(r-1)}{\log r} \Big).$$
It remains to rearrange and use the bound
$$\log r - \log(r-1) < \frac{1}{r}$$
along with the trivial bound $\log \log x > 1$ when $x > \exp(e)$ to get that
$$\frac{\log x}{\log \log x} > r \log^2 r \Big(1+\frac{1 + \log(\frac{r-1}{r-2})}{\log r}\Big).$$
The above inequality implies that $\pi(x) > R(x)$, and so there must exist some prime $p$ containing the digit $b$ which satisfies
$$\frac{\log p}{\log \log p} \leq r \log^2 r \Big(1+\frac{1 + \log(\frac{r-1}{r-2})}{\log r}\Big).$$
Now note that a string of length $l$ corresponds to a single digit in base $r=10^l$, so we have that for any string of length $l$ there exists a prime $p$ whose decimal contains the string and satisfies
$$\frac{\log p}{\log \log p} \leq 5.7 l^2 \times 10^l$$
so long as $l \geq 6$. Table 1 confirms this bound for the remaining values of $l$, and this completes the proof of Theorem \ref{primebound}.
\section{Proof of Theorem \ref{greentaocorollary}}
We require the following lemma, which is Theorem 1.2 of Green and Tao \cite{greentao}.
\begin{lem}
Let $A$ be any subset of the prime numbers of positive relative upper density, thus $\limsup_{N \rightarrow \infty} \pi(N)^{-1} | A \cap [1,N]|>0$. Then $A$ contains infinitely many arithmetic progressions of length $k$ for all $k$.
\end{lem}
Theorem 2 clearly follows from showing that for any string $S$, the set $A$ of primes whose decimal contains the string $S$ has positive relative upper density. The set $A$ actually has a density of 1. We let $B$ denote the set of primes whose decimal does not contain the string $S$, and we write $A(N) = A \cap [1,N]$ and $B(N) = B \cap [1,N].$
From the proof of Theorem \ref{primebound}, we have that
$$\frac{R(N)}{N} < \frac{r (r-1)}{r-2} \Big( \frac{r-1}{r} \Big)^k.$$
for $r^{k-1} \leq N < r^k$. Clearly,
\begin{eqnarray*}
\frac{ | A \cap [1,N]|}{\pi(N)} & = & 1 - \frac{|B(N)|}{\pi(N)} \\
& = & 1 + O\bigg( \frac{R(N)}{\pi(N)} \bigg)\\
& = & 1 + O\bigg( \frac{R(N) \log N}{N} \bigg).
\end{eqnarray*}
It is straightforward to check that
$$R(N) = o(N / \log N)$$
as $N \rightarrow \infty$ and so the error term tends to zero. This completes the proof of Theorem \ref{greentaocorollary}.
\newpage
\bibliographystyle{plain}
|
\section{Introduction}
Stokes profiles are used to characterize the properties of partially polarized light: the Stokes $I$ profile gives information about the intensity (total number of photons), $Q$ and $U$ about the linear polarization, and $V$ about the circular polarization \citep{Rees1989,Landi1992,delToro2003}. In an atmosphere at rest, harboring a vertically homogeneous magnetic field, under the assumption of local thermodynamic equilibrium (LTE), and within the Zeeman regime, Stokes $V$ is expected to be strictly antisymmetric with respect to the zero-crossing wavelength, and remains unshifted with respect to the position of the Stokes $I$ line core. Under these assumptions the so-called area asymmetry (or $\delta A$), that is, the wavelength integral of the signed Stokes $V$ signal normalized to its total area, is zero. A non-vanishing $\delta A$ is a necessary and sufficient condition for the existence of vertical variations in the line of sight component of the velocity and/or magnetic field vector \citep{Landolfi1996}. Recent observations at high spatial resolution have revealed Stokes $V$ profiles with distinct asymmetries in active regions \citep{Balasu1997}, plages \citep{MartinezPillet1997,Sigwarth1999} and the quiet Sun \citep{Vitichie2011,Sainz2012}. In the case of extreme asymmetries, one of the lobes of Stokes $V$ is missing, and therefore the $\delta A=1$. These are commonly referred to as single-lobed Stokes $V$ profiles.
Several theoretical models featuring large gradients in velocity and magnetic field have been suggested to explain the existence of single-lobed Stokes $V$ profiles. \cite{Steiner2000} proposed that these profiles can be produced when the line of sight (LOS) passes through a magnetopause, i.e., a separatrix layer dividing the atmosphere in two regions with different magnetic fields. Magnetopauses occur frequently in the solar atmosphere, with examples being the canopies created by expanding flux tubes in the network \citep{Grossmann2000} and the interfaces between different magnetic components in sunspot penumbrae \citep{Schlichenmaier1998}.
In the present work, we focus on the analysis of highly asymmetric ($\delta A \rightarrow 1$) and highly Doppler-shifted Stokes $V$ signals in the quiet Sun detected with the spectropolarimeter SP \citep{Lites2013_SP} on board the Hinode spacecraft \citep{Kosugi2007,Tsuneta2008}. Although these profiles are probably included in the extensive analysis carried out by \cite{Sainz2012}, these authors do not focus on those profiles, which besides having $\delta A \rightarrow 1$, are also strongly Doppler-shifted. The origin of Stokes $V$ profiles featuring both properties still remains under debate.
\cite{BellotRubio2001}, \cite{Shimizu2008}, \cite{Nagata2008} and \cite{Fischer2009} concluded that the cause of the strong red-shifted signals they found is convective collapse inside a magnetic flux tube. This process, which was first proposed on theoretical grounds by \cite{Parker1978}, \cite{Webb1978}, and \cite{Spruit1979}, makes the plasma inside a magnetic flux tube unstable, cool down, and fall along the field lines, sometimes at supersonic speeds. Meanwhile, the flux tube narrows and the magnetic field inside intensifies. This happens until a critical magnetic field (about $1-2$ kG), capable of suppressing the instability, is reached. The downflowing plasma can hit denser layers beneath it and rebound upward. This upflow could explain the extremely blue-shifted signals observed by \cite{SocasNavarro2005}.
Another mechanism that could produce strong up- and downflows is magnetic reconnection \citep{Parker1963}. The change in the topology of the magnetic field lines, favored by the low electrical conductivity of the photospheric layers, can produce two opposite jets that could reach supersonic velocities. Sweet-Parker reconnection, as modeled by \cite{Chae2002} and observed by \cite{Litvinenko2007}, could produce jets with outflowing speeds in the range of 3-10 km/s. This mechanism was proposed by \cite{QuinteroNoda2014} to explain the detected strong downflows embedded in a heated environment where the magnetic field intensity decreases during the process. \cite{Borrero2013} also invoked this scenario to explain the physical parameters inferred through the inversion of the strong Doppler-shifted profiles reported in \cite{Borrero2010}.
Last but not least, the siphon flow mechanism is also capable of producing highly Doppler-shifted signals. This phenomenon arises in a flux tube that connects magnetic elements of opposite polarity provided that there is a difference in the magnetic field between both footpoints. The difference in the magnetic field creates a gas pressure imbalance that accelerates the plasma from the high-pressure point toward the low-pressure footpoint. Under this mechanism, the velocities can also reach supersonic values \citep[][and references therein]{Montesinos1993,Rueedi1992}.
In the following sections we will study the polarimetric properties of two Fe {\sc i} lines observed by Hinode/SP in those regions where the Stokes $V$ profiles are both highly blue-shifted and present only one lobe (i.e., area asymmetry $\delta A \rightarrow 1$). Then, we will follow their vertical trace through the analysis of the chromospheric Ca {\sc ii h} spectral band. Finally, we will infer the atmospheric stratification of the physical parameters using the Stokes Inversion based on Response functions (SIR) \citep{RuizCobo1992} code. We aim to reveal the nature of these strongly asymmetric and highly Doppler-shifted signals as well as to reveal their underlying physical mechanism.
\section{Observations and data analysis}
\label{observations}
The polarimetric data used in this paper were acquired with the spectropolarimeter \citep[SP ; ][]{Lites2013_SP} on board the Hinode spacecraft \citep{Kosugi2007}. We selected a data set with a field of view of $328^{\prime\prime}\times154^{\prime\prime}$ recorded at disk center on March 10$^{\rm th}$, 2007. The SP instrument measures the Stokes vector of the Fe~{\sc i} line-pair at 630 nm with a spectral and spatial sampling of 2.15 pm pixel$^{-1}$ and 0.16$^{\prime\prime}$ ({\it normal map}), respectively. The exposure time is 4.8 seconds per slit position, making it possible to achieve a noise level of $1.1\times10^{-3}~I_{c}$ in Stokes $V$ and $1.2\times10^{-3}~I_{c}$ in Stokes $Q$ and $U$. Here $I_{c}$ refers to the mean continuum intensity in the granulation.
The second data set is a time series observation, based on a raster scan of 18 slits with 1.6 seconds of exposure time per slit position. It spans three hours of observation on September 25$^{\rm th}$, 2007 ($\mu$=1). In this case, the field of view is $2.9^{\prime\prime}\times38^{\prime\prime}$ and the final time cadence is 36 seconds. This data set belongs to the Hinode Operation Plan 14, entitled Hinode/Canary Islands campaign, and it allows us to analyze the time evolution of small magnetic structures. Although the noise level is slightly higher ($\approx 2\times 10^{-3}~I_{c}$) than in the first data set, it is still good enough to study polarization signals above the $10^{-2}~I_{c}$ level.
We also have at our disposal co-aligned CN and Ca~{\sc ii~h} broadband images recorded by the Broadband Filter Imager (BFI) instrument on Hinode \citep{Tsuneta2008}. The BFI images were acquired simultaneously with the second dataset (time series). The CN filter is located at 383.3 nanometers (filter width of 0.52 nm FWHM) and conveys information about the photosphere. The Ca~{\sc ii~h} filter is located at 396.85 nm (filter width of 0.22 nm FWHM) and contains information from both the photosphere and the chromosphere \citep[see][]{pietarila2009}. The exposure times were 0.1 s and 0.3 s, respectively.
Data from CN and Ca~{\sc ii~h} spectral bands consist of a series of images of the same field of view, taken every 30 seconds. Because of the difference between the cadence of the acquisition of these images and the cadence of each Hinode/SP raster (36 seconds), we choose the broadband images that are closest in time to the ninth slit position of the raster scan performed by the SP instrument. With this, we guarantee that the time delay between any SP raster and the BFI images is never larger than 18 seconds.
\begin{figure*}
\centering
\includegraphics[width=18cm]{spa1t.eps}
\vspace{+20pt}
\caption{\footnotesize Typical spatial distribution of the circular polarization profiles (Stokes $V$) around the events detected on the {\it normal} map. Each panel displays adjacent profiles from the two iron lines located at 6301.5 and 6302.5 \AA \ measured by Hinode/SP. The vertical scale ranges from -4\% to +2\% of $I_c$. Blue denotes the detected single-lobed blue-shifted profiles. Green shows pixels with significant linear polarization signals ($\ge 0.01~I_{c}$). Finally, the red lines indicate pixels with single-lobed red-shifted Stokes $V$ profiles. Vertical lines indicate the rest center wavelength of each line.}
\label{spa1t}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=18cm]{spa2.eps}
\vspace{+20pt}
\caption{Same as Figure \ref{spa1t}, except showing an example of a blue-shifted case that appears unrelated to any red-shifted signal. The vertical scale goes from -3\% to +8\% of $I_c$.}
\label{spa2}
\end{figure*}
The events of interest in this work are those presenting strong Doppler-shifted Stokes $V$ profiles with a single lobe. We have followed the method described in \cite{MartinezPillet2011a} to find them, that is, examining the Stokes $V$ signal at $\pm 272$~m\AA \ from the center of the Fe {\sc i} 6302.5 \AA \ line (i.e., on the continuum) and generating Stokes $V$ \textit{continuum} magnetograms from the red and blue sides of that line. Then, we analyzed these magnetograms looking for pixels with Stokes $V$ signals above 0.5\% of the mean continuum intensity of the map ($\ge 0.05~I_{c}$). As the detected events were grouped in patches of different sizes, we decided to define a minimum area of detection of two pixels to prevent isolated pixels from entering the analysis. Isolated spikes can be produced by cosmic rays or data compression issues.
\begin{figure*}
\centering
\includegraphics[width=29.5cm]{Magn.eps}
\caption{Spatial distribution of the blue and red cases over the Fe {\sc i} 6302.5~\AA \ magnetogram from the Hinode/SP \textit{normal} map. Blue squares indicate the location of the blue events, while red squares designate the position of the red events.}
\label{magn}
\end{figure*}
\section{General properties of the single-lobed Stokes $V$ profiles}
\subsection{Polarimetric description}\label{profiles_sec}
With the selection criteria described in Section~\ref{observations}, we have found 398 strongly blue-shifted Stokes $V$ profiles belonging to 64 different events and 158 highly red-shifted Stokes $V$ profiles corresponding to 46 events. Hereafter, we will refer to these events as {\it blue} and {\it red} cases, respectively. Although not demanded by our selection criteria, all the selected circular polarization profiles happen to be highly asymmetric, $\delta A \rightarrow 1$, featuring one single lobe. In addition, the intensity profiles (Stokes $I$) also possess some common features. For instance, Stokes $I$ in the blue cases, present a mean continuum signal above one, indicating that they are located over bright regions on the continuum map (granular edges). The red cases, however, show the opposite behavior and display a mean continuum intensity below one, thus appearing preferably over dark regions on the continuum map (intergranular lanes). We have also found that the outer-blue wing of Stokes $I$ is strongly pushed toward shorter wavelengths in the blue cases. Likewise, the outer red wing of Stokes $I$, in the red cases, presents a strong shift toward larger wavelengths. This produces strongly asymmetric intensity profiles that are indicative of large depth variations in the LOS component of the velocity.
In addition, we have searched for blue- or red-shifted circular polarization signals inside a $2\times2\arcsec$ region centered in every detected event, and we found that 73\% of the blue cases are related to a red event. The rest of the blue cases, 27\% of the total, seem to be isolated events. Interestingly, red cases never appear isolated. The distance between blue and the associated red cases varies but, whenever both events are close enough, they are connected by pixels that present strong linear polarization. In those in-between regions with large levels of linear polarization, Stokes $V$ is also highly asymmetric ($\delta A \rightarrow 1$), but in this case the profiles are not single lobed. Instead, they present two lobes of the same sign. Figure~\ref{spa1t} shows a typical example of the spatial distribution of the Stokes $V$ profiles whenever a blue event appears related to a red event. This figure is composed of $6\times6$ panels representing adjacent pixels. Each panel displays the Stokes $V$ profiles of both iron lines, Fe {\sc i} 6301.5 and 6302.5~\AA. Dotted vertical lines represent the rest center wavelength. In every panel, the vertical scale goes from -4\% to +2\% of I$_c$. We plotted in blue the Stokes $V$ profiles that we classify as belonging to a blue event, while red indicates the profiles with Stokes $V$ belonging to a nearby red event. Green profiles in between the blue and red profiles correspond to the Stokes $V$ profiles where the linear polarization is above 1\% of I$_c$. We note that the closest pixels to the green profiles that display Stokes $V$ blue- or red-shifted signals, also show a non-negligible amount of linear polarization, albeit below the 1\% level. The rest of the pixels in blue or red (e.g., panels on the top leftmost, $[1,1]$, and bottom rightmost, $[6,6]$, part in Fig.~\ref{spa1t}) do not possess a significant amount of linear polarization. We emphasize that the polarization signals in these observations, with circular and linear polarization levels higher than $0.03~I_{c}$ and $0.01~I_{c}$, respectively, are much stronger than those in the quiet-Sun internetwork \citep[cf. Fig.~4 in][]{borrero2013clv}.
An interesting property of the events in this configuration (Fig.~\ref{spa1t}) is that the sign of the single lobe in Stokes $V$ is the same for the red and blue cases: if one of the patches displays negative single-lobed profiles, then the associated event will also show negative Stokes $V$ profiles but with the opposite wavelength shift, and vice-versa (e.g., compare panels $[2,2]$ and $[5,5]$). We also see that the missing lobe is always the one closer to the rest wavelength (vertical dotted lines in each panel). These two facts seem to indicate that the circular polarization profiles in the blue and red cases are produced by magnetic fields of opposite polarity that harbor plasma flows along opposite directions. This can be realized by picturing circular polarization profiles where the missing lobe, in both red and blue cases, would be present close to the rest wavelength (vertical dotted lines). In addition, we observe that the Stokes $V$ profiles in the region that connects the red and blue events (green profiles in Fig.~\ref{spa1t}) are also highly asymmetric. Unlike the blue and red Stokes $V$ profiles, these profiles are not single-lobed, but rather they possess two lobes of the same sign (i.e., panel $[3,3]$).
The properties described above point to a configuration of the magnetic field in the form of a magnetic loop, where the footpoints possess a magnetic field of opposite polarities and velocities of opposite signs. The detection of linear polarization signals between footpoints further supports this picture. Interestingly, this scenario could also explain the observed asymmetries in Stokes $V$. The circular polarization profiles at the footpoints would have single-lobed shapes because the observer sees two different atmospheres along the line of sight: one with magnetic field and the other without it \citep{Viticchie2012}. On the other hand, the Stokes $V$ profiles in the region with large linear polarization signals (green colors in Fig.~\ref{spa1t}) can be ascribed to the limited spatial resolution of the telescope, either produced by the contamination of the point spread function the telescope introduces in each pixel \citep{vanNoort2012} or by the presence of two magnetic fields with opposite polarities on the same resolution element \citep{Sigwarth2001}. For instance, Stokes $V$ in panel $[3,3]$ can be thought as being produced by a mixture of the Stokes $V$ profiles from panels $[2,2]$ and $[4,4]$. In spite of all the information gathered by the visual inspection of the profiles, it is so far not possible to establish whether the magnetic loop corresponds to a $\Omega$-loop or a U-loop. We will answer this question in the following sections.
Finally, Figure~\ref{spa2} illustrates an example of the remaining 27\% of detected blue-shifted events, where no red-shifted counterpart could be identified within a $2\times2\arcsec$ region around it. It was not possible to detect any significant linear polarization signals there either. In this figure, the vertical scale ranges from -3\% to +8\% of I$_c$. The blue-shifted single-lobed Stokes $V$ profiles in Figure~\ref{spa2} are very similar to those displayed in Figure~\ref{spa1t}. The only discernible difference is that the amplitude of the blue-shifted Stokes $V$ lobe in the isolated events is generally larger than the amplitude of the blue-shifted Stokes $V$ lobe when the blue event is accompanied by a red event.
\subsection{Spatial distribution and rate of occurrence}
\label{spatial_dist}
To analyze the spatial distribution of the detected events we have plotted in Figure \ref{magn} the locations of all blue and red cases on a magnetogram constructed by integrating the signed circular polarization signals across the Fe {\sc i} 6302.5 {\AA} spectral line. Blue squares mark the position of the blue cases, while red squares correspond to the red cases. In order to make them visible, the size of each square has been magnified with respect to the real area occupied by each event. The locations of these events seem to be related to regions where the magnetic field is concentrated. They are close to network areas and rarely appear on the internetwork.
These events appear in patches of different sizes over the entire map. The mean areas are 6.2$\pm$3.1 and 3.5$\pm$2.1 pixels, for blue and red cases respectively. The smaller area of the latter could explain why we detect fewer of them, 46 red events compared to 64 detected blue events. To obtain the rate of occurrence of these events we divide the number of detected cases, 64 blue cases and 46 red cases, by the total area of the map ($328^{\prime\prime}\times154^{\prime\prime}$). This provides a rate of occurrence of $1.2\times10^{-3}$ blue cases per arcsec$^{2}$ and $9.1\times10^{-4}$ red cases per arcsec$^{2}$. These values indicate that we would find nearly 12 different blue events, or nine red cases, in a FOV of $100\times100$ arcsec$^{2}$.
\subsection{Analysis of the time series}
\label{time_evol}
We complement the information presented in Sections~\ref{profiles_sec}, and ~\ref{spatial_dist} using the time series from Hinode/SP described in Section~\ref{observations}. In the three hours of observations available we found five cases that are analogous to those found in the {\it normal} map (see Fig.~\ref{magn})\footnote{Considering the total area spanned by the time series and the rate of occurrence determined in Sect.~\ref{spatial_dist} a total of about 40 events should have been detected. The fact that we see ten times fewer might indicate that the slit of the raster scan probably sit on an internetwork region, where many fewer events are usually seen (see Fig.~\ref{magn})}. Because of the small FOV ($2.9^{\prime\prime}\times38^{\prime\prime}$) it is difficult to detect any blue case, along with its associated red event, that falls into the small scanned region during the entire lifetime of the event. Out of five detected events, only one of them remains within the field of view throughout its evolution. In Figure~\ref{chromos}, we analyze the time evolution of this particular event using Hinode/SP scans as well as CN and Ca {\sc ii~h} broadband images from Hinode/BFI. The event lasts nearly ten minutes. From top to bottom each row displays the continuum intensity from Hinode/SP, CN broadband images, Fe {\sc i} 6302.5 \AA \ magnetogram (obtained in the same way as Fig.~\ref{magn}), and Ca {\sc ii h} broadband images. Each of the four columns correspond to a different time step in the event's evolution: $\Delta t=0$, 72, 144 and 216 seconds. Following the same convention used in Fig.~\ref{spa1t}, we have also marked in blue the pixels harboring blue-shifted single-lobed Stokes $V$, in red the pixels displaying red-shifted single-lobed Stokes $V$, and in green the pixels with a linear polarization above 1\% of $I_{c}$.
\begin{figure*}
\centering
\includegraphics[width=15.cm]{chromos2.eps}
\caption{Evolution of a single event detected with the Hinode/SP time series. Each row corresponds to a different time in its evolution. The first column displays the continuum signal from Hinode/SP, the second column shows the CN broadband image from Hinode/BFI, the third column displays the Fe {\sc i} 6302.5 \AA \ magnetogram from Hinode/SP, and the fourth column shows the Ca {\sc ii h} broadband images from Hinode/BFI. We have also highlighted the pixels with blue- and red-shifted single-lobed Stokes $V$ profiles with blue and red crosses, respectively. Pixels with linear polarization signals above $0.01~I_{c}$ are indicated with green crosses.}
\label{chromos}
\end{figure*}
The left most column in Fig.~\ref{chromos} corresponds to the initial detection of the event. At this time, the event displays only a patch of large linear polarization signals and almost no blue- or red-shifted Stokes $V$ profiles. The pixels where the linear polarization is above 1\% of $I_{c}$ exhibit $Q$-like circular polarization profiles characterized by two lobes of the same sign (similar to those in Fig.~\ref{magn}; panel $[3,3]$). After 72 seconds (second column) two patches of blue- and red-shifted single-lobed Stokes $V$ profiles developed and they appear connected through the region of enhanced linear polarization. The third and fourth columns show that the patches of blue- and red-shifted single-lobed circular polarization profiles separate, while the pixels with high linear polarization signals vanish. The footpoints keep moving away from each other for the remaining of the event's lifetime (10 min). According to the Fe {\sc i} continuum and CN broadband images (first and second rows), the blue- and red-shifted patches lie over bright and dark regions, respectively. The Fe {\sc i} 6302.5 \AA \ magnetogram (third column in Fig.~\ref{chromos}) further demonstrates (see Sect.~\ref{profiles_sec}) that these two patches appear over magnetic fields of opposite polarities. Finally, we do not find any traces of these events higher up in the atmosphere (Ca {\sc ii h} images on the bottom row in Fig.~\ref{chromos}) at any time in their evolution.
As already mentioned in Sect.~\ref{profiles_sec} (where we studied instantaneous images of these events) the observed configuration, as seen from its temporal evolution, is also compatible with a magnetic loop where the linear polarization connects two footpoints with opposite velocities and magnetic fields of opposite polarity. Since the linear polarization precedes the appearance of the footpoints of the loops, the time evolution could correspond to either an emerging $\Omega$-loop or to a submerging U-loop. These kinds of events have been previously reported in \cite{MartinezGonzalez2007,Centeno2007,MartinezGonzalez2009,Ishikawa2010,Viticchie2012}. All of them describe the structure as two footpoints of opposite Stokes $V$ signals connected by linear polarization, but only \citep{Ishikawa2010} found opposite velocities, and only \cite{Viticchie2012} detected single-lobed Stokes $V$ profiles.
\begin{figure*}
\centering
\includegraphics[width=18.4cm]{Mean_1.eps}
\caption{Mean atmospheric stratifications obtained from the inversion of the selected pixels in the Hinode/SP \textit{normal} map (Fig.~\ref{magn}). The first and second rows represent the results for blue and red events, respectively. Color coding follows this designation. The first column displays the temperature $T$, second column the LOS velocity ${\rm v}_{\rm LOS}$ with positive values representing downflows, third column the magnetic field intensity $B$, and fourth column the inclination of the magnetic field $\gamma$. Gray areas indicate the deviation from the mean stratifications obtained from the 358 pixels belonging to blue events and 158 pixels belonging to red events (see Sect.~\ref{profiles_sec}).}
\label{mean}
\end{figure*}
\section{Stokes profiles inversions}
\subsection{Configuration}\label{conf}
The SIR code \citep[Stokes inversion based on Response Functions;][]{RuizCobo1992} allows us to infer the optical-depth dependence of the atmospheric parameters at each pixel through the inversion of the observed Stokes profiles. In our inversions we employ a single magnetic component parametrized by: seven nodes in the temperature $T(\tau_{500})$\footnote{The parameter $\tau_{500}$ refers to the optical depth evaluated at a wavelength where there are no spectral lines (continuum). In our case this wavelength is 500 nm.}, three in the line of sight (LOS) component of the velocity ${\rm v}_{\rm LOS}(\tau_{500})$, three in the magnetic intensity $B(\tau_{500})$, two for the inclination of the magnetic field with respect to the observer's LOS $\gamma(\tau_{500})$, and one for the azimuthal angle of the magnetic field in the plane perpendicular to the observer's LOS $\phi(\tau_{500})$. Variables such as micro- and macro-turbulence are fixed to zero and not inverted. At each iteration step the synthetic profiles are convolved with the spectral transmission profile of Hinode/SP \citep{Lites2013_SP}. Since each node corresponds to a free parameter during the inversion, our model includes a total of 16 free parameters. The total number of data points is 448 distributed across 112 wavelength positions for each of the four Stokes profiles.
The nodes in $B(\tau_{500})$, $\gamma(\tau_{500})$, and ${\rm v}_{\rm LOS}(\tau_{500})$ are necessary to reproduce the extreme area asymmetries ($\delta A \rightarrow 1$) in the observed (see Figs.~\ref{spa1t} and~\ref{spa2}) circular polarization profiles \citep{Landolfi1996}. Additionally, the inferred physical parameters could be reliant on the initial atmosphere. To avoid this dependence, we have decided to invert each individual pixel with 25 different initial atmospheric models. These models were determined by randomly perturbing the temperature stratification of the Harvard-Smithonian Reference Atmosphere (HSRA) model \citep{Gingerich1971}. The rest of the physical parameters of the initial model ($B$, $\gamma$ and ${\rm v}_{\rm LOS}$) were determined via a uniform probability distribution and considered to be independent of the optical-depth. Out of the 25 solutions obtained for a given pixel, we retain the one that yields the smallest value of $\chi^2$.
\begin{figure*}
\centering
\includegraphics[width=16cm]{tau0.eps}
\caption{Results from the inversion of the observed Stokes profiles in the same region as in Figure~\ref{chromos}. Each panel displays the spatial distribution of the physical parameters at an optical depth $\log\tau_{500}=0$ (i.e. continuum). {\it From top to bottom}: temperature $T$, LOS velocity ${\rm v}_{\rm LOS}$, and LOS magnetic field $B_{\rm LOS}=B \cos\gamma$. {\it From left to right}: different time steps during the evolution of the event, $\Delta t=$ 0, 72, 144, and 216 seconds.}
\label{77}
\end{figure*}
\subsection{Statistical properties}\label{stats}
We have inverted all 556 selected pixels in Hinode/SP \textit{normal} map (398 blue- and 158 red-shifted; Sect.~\ref{profiles_sec}) employing the configuration described in Sect.~\ref{conf}. Figure~\ref{mean} presents the averaged physical parameters as a function of the optical-depth evaluated at a wavelength of 500 nm ($\log\tau_{500}$) for the blue (upper row) and red cases (bottom row). The shaded-grey areas in this figure indicate the standard deviation obtained from the ensemble of pixels belonging to each of the two cases.
The mean temperature stratification (left most panels) is very similar in both blue and red cases. In the deep photosphere, $\log\tau_{500} \ge 0$, both kinds of events follow the stratification from the HSRA model (dashed lines) very closely. Above this region, $\log\tau_{500} \le 0$, the inferred temperature in both cases is larger than that of the HSRA model. However, there are two particular layers ($\log\tau_{500} \approx -3.5$ and $\log\tau_{500} \approx -2$) where the temperature in these events is the same as in the aforementioned model. The only minor difference between the inferred temperature stratifications appears close to the continuum, $\log\tau_{500} = 0$, where we obtain that the blue cases are slightly hotter (on average) than HSRA, whereas the red cases are slightly cooler. This was foreseeable because the former tend to appear at bright regions as seen on the continuum intensity, while the latter occur at dark regions (see Sect.~\ref{profiles_sec} and upper row in Fig.~\ref{chromos}).
\begin{figure*}
\centering
\includegraphics[width=16cm]{tau2.eps}
\caption{Same as Figure \ref{77}, except showing the physical parameters at $\log\tau_{500}=-2$.}
\label{83}
\end{figure*}
Unlike the temperature, the inferred LOS velocities (second column in Fig.~\ref{mean}) are very different in the blue and red cases. In the former, the LOS velocity changes from large blue-shifts in the deep photosphere, ${\rm v}_{\rm LOS} = -5$ km~s$^{-1}$ at $\log\tau_{500} = 0$, to red-shifts in the higher photospheric layers: ${\rm v}_{\rm LOS} = 3$ km~s$^{-1}$ at $\log\tau_{500} = -3$. In the red cases, however, the velocity stratification is exactly opposite, with large red-shifts in the deep photosphere, ${\rm v}_{\rm LOS} = 5$ km~s$^{-1}$ at $\log\tau_{500} = 0$, which become blue-shifts at higher photospheric layers: ${\rm v}_{\rm LOS} = -3$ km~s$^{-1}$ at $\log\tau_{500} = -3$. The LOS velocities inferred from the inversion close to $\log\tau_{500} = 0$ correspond to what should have been expected from the visual inspection of the profiles (see Fig.~\ref{spa1t}), since the blue cases featured highly blue-shifted Stokes $V$ profiles and enhanced absorption on the blue wing of Stokes $I$, and vice versa for the red cases (see Sect.~\ref{profiles_sec}).
As far as the magnetic properties of these events are concerned, the inversion returns a magnetic field strength $B$ that, despite being rather large (about 1 kG) in the deep photosphere, quickly decreases toward higher photospheric layers and becomes negligible above $\log\tau_{500} = -2$. This behavior is observed in both blue and red cases (see third column in Fig.~\ref{mean}). Meanwhile, the inclination of the magnetic field indicates that the field lines are nearly vertical close to $\log\tau_{500} = 0$, and they gradually become horizontal as we move upward in the photosphere. The value of the inclination, $\gamma$, at $\log\tau_{500} = 0$ follows the polarity of the magnetic field already seen in Fig.~\ref{chromos} (third row) and is the opposite for blue and red cases. The inferred inclination for $\log\tau_{500} \le -2$ carries no significance because of the large scatter (see shaded areas) and because the magnetic field strength vanishes above this layer ($B \rightarrow 0$ when $\log\tau_{500} \le -2$).
\subsection{Evolution}\label{evolution}
Besides the data from the Hinode/SP {\it normal map} (Sect.~\ref{stats}), we have also inverted the Stokes profiles from the time series described in Sect.~\ref{time_evol}. This allows us to follow the evolution of the physical parameters in the example presented in Fig.~\ref{chromos}. Inferring the magnetic field vector $B$, $\gamma$, and $\phi$ from noisy data is a difficult task \citep{borrero2011} and therefore, we have decided to invert only for the temperature $T(\tau_{500})$, and LOS velocity ${\rm v}_{\rm LOS}(\tau_{500})$ in those pixels in Fig.~\ref{chromos} where the total polarization is below 2 \% of $I_{c}$. In this case, we use three nodes for the temperature and the LOS velocity. Pixels with polarization values above this threshold were treated in the way described in Sect.~\ref{stats}. The inversion of all pixels in Figure~\ref{chromos} yields the physical parameters ($T$, $B$, $\gamma$, $\phi$, and ${\rm v}_{\rm LOS}$) as a function of $(x,y,\log\tau_{500},t)$. To visualize the results, we present in Figures~\ref{77} and~\ref{83} maps of the physical parameters at two different optical depths in the photosphere: $\log\tau_{500}=0,-2$, respectively. Each of these two figures display, from the top to bottom, the temperature $T(x,y)$, LOS velocity ${\rm v}_{\rm LOS}(x,y)$, and finally the LOS magnetic field $B_{\rm LOS}(x,y)$ obtained as the product between $B$ and $\cos\gamma$. Each column presents a different time in the event's evolution: $\Delta t=0,72,144,216$ s.
We note that the temperature $T(x,y)$ at $\log\tau_{500}=0$ (upper panels in Fig.~\ref{77}) closely follows the intensity pattern (i.e., granulation) outside the strong magnetic field regions (see first and second rows in Fig.~\ref{chromos}). The Wilson depression inside the magnetic element causes the $\tau_{500}=1$ level to shift with respect to its surroundings. This effect, in combination with the existence of vertical gradients in the temperature, produces the apparent discontinuities in the temperature between the magnetic and non magnetic region. The color palette for ${\rm v}_{\rm LOS}(x,y)$ in the second row in Figures~\ref{77} and~\ref{83} has been chosen such that positive velocities are in red (red-shifts or downflows) and negative velocities are in blue (blue-shifts or upflows). In addition, $B_{\rm LOS}(x,y)$ (third row in Figs.~\ref{77} and~\ref{83}) shows positive polarity magnetic fields (pointing toward the observer) in red, while negative polarity magnetic fields (pointing away from the observer) are indicated in blue.
Figure~\ref{77} demonstrates that, in the two regions with strong magnetic fields, the polarities and LOS velocities are inverted. As mentioned in Sects.~\ref{profiles_sec} and~\ref{time_evol}, these two regions correspond to the footpoints of the magnetic loop that separate from each other as time passes. Higher up in the photosphere, $\log\tau_{500}=-2$ (see Figure~\ref{83}), we can see that the temperature in the regions with strong magnetic fields are close to those of the HSRA model, where T$_{\rm HSRA}$(log $\tau_{500}$=-2)=4660 K. This is in agreement with the previous section (see also Fig.~\ref{mean}). We also observe that the LOS magnetic field and velocity above the location of the footpoints have vanished (see also the third column in Fig.~\ref{mean}). The fact that ${\rm v}_{\rm LOS}$ and $B_{\rm LOS}$ are only present at the bottom of the photosphere lead us to identify this structure as an $\Omega$-loop structure rising through the atmosphere. A more detailed analysis of the loop structure will be performed in the next section.
\section{Side view of the loop}
\label{sideview}
\subsection{Method}
\label{sideview_method}
In this section, we will analyze the vertical structure of the loop in Figures~\ref{77} and~\ref{83} as seen from the side. To that end we employ the results from the inversion presented in Section~\ref{evolution}. Here we focus only on the time step $\Delta t=72$ seconds. At this time the loop is completely developed. In addition, the footpoints and the linear polarization signal that connects them are clearly discernible. To determine the physical properties of the loop as seen from the side we average, at each optical depth, the physical quantities temperature, LOS velocity, and magnetic field vector along the direction perpendicular to the line that connects both footpoints. This connecting line, or {\it bisector}, is denoted by the inclined solid black line in Figure~\ref{bisec}. In this fashion, we can transform all relevant physical parameters from the $(x,y,\tau_{500})$ coordinate system into $(L,\tau_{500})$, where $L$ refers to the direction along the bisector in Figure~\ref{bisec}.
Once in this new reference frame $(L,\tau_{500})$, we need to translate the physical parameters into a frame where the vertical axis corresponds to the geometrical height $z$ instead of the optical depth scale $\tau_{500}$. To achieve this, we employ a simplified version of the method described in \citet{klaus2010proc,klaus2010apj}. The idea is to determine a $z(\tau_{500})$ conversion for each pixel along the $L$ direction. This can be done by integrating the relation $dz=-d\tau_{500}/[\rho(\tau_{500}) \kappa(\tau_{500})]$. Here, $\kappa$ and $\rho$ correspond to the opacity and density, respectively, of the averaged ($L$-direction) atmosphere, with the latter being obtained through the assumption of hydrostatic equilibrium along the vertical $z$-axis. The solution to this equation requires an integration constant. In our case, this constant is taken as the Wilson depression $z_w=z(\tau_{500}=1)$. With this procedure, a different $z(\tau_{500})$ conversion can be established for every point $L$ along the loop. In order to establish a common $z$-scale for all pixels, we allow the aforementioned integration constant to vary along the bisector, $z_w(L)$, such that the divergence of the magnetic field, $\nabla \cdot {\rm \bf B}$, is minimized along the loop.
Before this minimization is performed, we need to solve the 180$^{\circ}$ ambiguity in the azimuth ($\phi$) of the magnetic field. This ambiguity makes it impossible to distinguish between the following two solutions: ${\rm \bf B}=(B_x,B_y,B_z)$ and ${\rm \bf B}^{\dagger}=(-B_x,-B_y,B_z)$. If left unattended, the terms corresponding to the horizontal derivatives in $\nabla \cdot {\rm \bf B}$ ($\partial B_x/\partial x$ and $\partial B_y/\partial y$) will dominate over the vertical derivative ($\partial B_z/\partial z$), thereby hindering the minimization of the divergence of the magnetic field by changing the boundary condition $z_w(L)$. The ambiguity is solved by choosing, at each pixel along the bisector in Fig.~\ref{bisec}, either ${\rm \bf B}$ or ${\rm \bf B}^{\dagger}$, such that the magnetic field in the $XY$-plane always points along the same quadrant. The red headless arrows in Fig.~\ref{bisec} show the direction of the magnetic field on the horizontal plane after the correction for the 180$^{\circ}$-ambiguity, illustrating that the field lines connect both footpoints of the structure.
\begin{figure}[t]
\centering
\includegraphics[width=8.0cm]{field_bisector.eps}
\caption{Magnetic field intensity bisector reference, black line, for the magnetic loop analyzed using the time series. The direction along the bisector is $L$. It forms an angle of $-57^{\circ}$ with the horizontal axis. The selected time instant, $\Delta t=72$ $s$, corresponds to the second column in Figures \ref{chromos}, \ref{77} and \ref{83}. Red headless arrows display the behavior of the horizontal component of the magnetic field after solving the 180$^{\circ}$-ambiguity.}
\label{bisec}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=15.0cm]{loop_80.eps}
\caption{Side view of the observed magnetic loop. {\it Top left panel}: magnetic field $B(L,z)$. {\it Top right panel}: inclination of the magnetic field $\gamma(L,z)$. {\it Bottom left panel}: logarithm of the gas pressure $P_g(L,z)$. {\it Bottom right panel}: LOS component of the velocity ${\rm v}_{\rm LOS}(L,z)$. White/black lines trace the magnetic field lines.}
\label{loop80}
\end{figure*}
\subsection{Results}
\label{sideview_results}
After a common $z$ scale has been established for all pixels along the bisector in Fig.~\ref{bisec}, we can study the behavior of the physical parameters as a function of the common vertical $z$-coordinate and the distance $L$ along the loop. This is equivalent to looking at the structure from the side, hence the title of Section~\ref{sideview}. Figure \ref{loop80} displays: magnetic field $B(L,z)$ (upper left panel), inclination of the magnetic field $\gamma(L,z)$ (upper right panel), gas pressure $P_g(L,z)$ (lower left panel; logarithmic scale), and LOS velocity ${\rm v}_{\rm LOS}(L,z)$ (lower right panel). Each panel indicates the magnetic field lines of the loop displayed in white or black. We can see that the minimization of $\nabla \cdot {\rm \bf B}$ (which shifts up or down the vertical scale at each pixel along the bisector; see Sect.~\ref{sideview_method}) forces the central sections of the loop to be located higher in the photosphere than the footpoints. This behavior produces a clear $\Omega$-loop shape: the footpoints (formed deep in the photosphere) are characterized by somewhat vertical fields with opposite polarities: $\gamma \approx 45^{\circ}$ on the left footpoint ($L \approx 700$ km), but $\gamma \approx 120^{\circ}$ on the right footpoint ($L \approx 1600$ km). Meanwhile, the apex of the loop ($L \approx 1200$ km) possesses a more horizontal magnetic field ($\gamma \approx 90^{\circ}$). In addition, the LOS velocity ${\rm v}_{\rm LOS}$ at each footpoint has the opposite sign (bottom-right panel), thus indicating that the plasma flows from the right footpoint (upflow; ${\rm v}_{\rm LOS} < 0$) toward the left footpoint (downflow; ${\rm v}_{\rm LOS} > 0$). We emphasize that the values of ${\rm v}_{\rm LOS}$ at the footpoints are of the order of $3-4$ km~s$^{-1}$, but if we project these velocities along the loop's axis, assuming that the velocity and magnetic field vectors are parallel, we obtain absolute values of $5-6$ km~s$^{-1}$ at the footpoints. These velocities are significantly faster than the typical convective velocities seen in the granulation.
The bottom left panel in Fig.~\ref{loop80} allows us to detect gas pressure imbalances $\Delta P_g$ between two points along the loop at a fixed height. This panel should be treated with caution however, because of the exponential dependence of $P_g$ with height. This dependence means that small errors in the determination of the Wilson depression at each point along the loop, $z_w(L)$, greatly increase the uncertainty in the gas pressure difference between any two given points.
During the inversion, only one node was given to the azimuthal angle of the magnetic field $\phi$ (see Sect.~\ref{conf}). Because of this, $\phi$ does not depend on $z$ (or $\tau$). However, the azimuth angle does feature an interesting behavior along the loop. This behavior, $\phi(L)$, is indicated in Figure~\ref{azimuth}, where we observe that the average magnetic field is not completely aligned with the structure. In the upflowing footpoint (${\rm v}_{\rm LOS} < 0$; $L \approx 1600$ km), the azimuth of the field closely follows the inclination of the bisector ($-57^{\circ}$ in Fig.~\ref{bisec}), but it gradually deviates from this value as we approach the downflowing footpoint (${\rm v}_{\rm LOS} > 0$; $L \approx 700$ km). We conclude, therefore, that the magnetic field shows a certain degree of twist along the loop. This can be produced by a structure with magnetic helicity. The torsion of the magnetic field lines is also visible in the relative orientation between the bisector and the red headless arrows in Figure \ref{bisec}.
\begin{figure}
\centering
\includegraphics[width=8.0cm]{azimuth.eps}
\caption{Magnetic field azimuth. Squares mark the mean value of the inverted pixels and the dotted line designates the orientation of the black line plotted in Figure \ref{bisec}, i.e., the bisector.}
\label{azimuth}
\end{figure}
\section{Discussion}
The analysis of the data from the Hinode/SP time-series revealed two footpoints with opposite polarities connected by a region that harbors clear linear polarization signals ($> 0.01~I_{c}$; see Sects~\ref{profiles_sec}, ~\ref{time_evol}, and Fig.~\ref{chromos}). Because of this, it is tempting to ascribe the inferred configuration of the magnetic field to a magnetic loop. The analysis of the horizontal component of the magnetic field shows that the field lines go from one footpoint to the other (Fig.~\ref{bisec}), albeit with some degree of twist (Fig.~\ref{azimuth}). Furthermore, by imposing the $\nabla \cdot {\rm \bf B} = 0$ condition, we have been able to determine (see Sect.~\ref{sideview}) that the linear polarization signals (corresponding to horizontal magnetic fields) come from higher photospheric layers than those responsible for the circular polarization signals (associated with more vertical magnetic fields) of opposite polarity at the footpoints. All these results lend support to a configuration of the magnetic field in the form of an $\Omega$-loop. This kind of topology has already been reported by \cite{MartinezGonzalez2007,Centeno2007,MartinezGonzalez2009,Ishikawa2010}. A similar type of configuration, including the presence of single-lobed Stokes $V$ profiles, has also been described in \cite{Viticchie2012}.
Once we have established the configuration of the magnetic field, we turn our attention to the inferred velocities. The fact that the LOS velocities are also opposite at the loop footpoints (see the second column in Fig.~\ref{mean} and the bottom right panel in Fig.~\ref{loop80}) indicates that the magnetized plasma flows from one footpoint towards the other. In spite of not finding any clear gas pressure imbalance between the footpoints (see Sect.~\ref{sideview_results}), we believe that the siphon flow mechanism \citep[][ and references therein]{Montesinos1993} is the only physical mechanism (of those listed in Sect.~1) that explains the observed flow along this magnetic $\Omega$-loop (i.e., the observed kinematic and magnetic properties).
There are, however, some features of the inferred physical parameters that do not entirely match the above scenario. We refer in particular to the strong increase in the temperature in the mid-photosphere at the footpoints of the loop (see leftmost panels in Fig.~\ref{mean}), and the change in the sign (also in the mid-photosphere) of the LOS component of the velocity ${\rm v}_{\rm LOS}$ where the magnetic field drops to zero (second and third columns in Fig.~\ref{mean}). It is possible to argue, however, that these two features appear as a consequence of the limitations in our modeling of the solar atmosphere. In the following, we will elaborate on this particular point.
\begin{figure*}
\centering
\includegraphics[width=17cm]{Sint.eps}
\caption{Synthetic Stokes profiles using a temperature stratification without the inferred increased at high layers, first row. The same study but changing to zero the LOS velocity at high layers is presented in the second row. Black designates the original profiles obtained using the mean atmosphere showed in blue in Figure \ref{mean} while blue lines mark the newer results.}
\label{sint}
\end{figure*}
In Section~\ref{conf} we mentioned that, according to \cite{Landolfi1996}, several nodes in the magnetic field and LOS velocity were needed to reproduce the large area asymmetries in the observed Stokes $V$ profiles (see Fig.~\ref{spa1t}). As a result, we have inferred an optical depth dependence of the LOS velocity where ${\rm v}_{\rm LOS}(\tau_{500})$ changes from large blue-shifts in the deep photosphere ($\log\tau_{500} \approx 0$) to large red-shifts in the higher layers ($\log\tau_{500} \approx -3$) at the upflowing footpoint, and vice versa at the downflowing footpoint. The question is therefore whether this change in the sign of the velocity in the high photosphere (and also the temperature enhancement) is needed to reproduce the asymmetries in the observed circular polarization profiles. To investigate this we have performed two different experiments.
In the first experiment,, we took the optical depth dependence of the physical parameters inferred from the inversion of the blue cases (upper row in Fig.~\ref{mean}) and proceeded to calculate the emerging Stokes profiles from this mean atmospheric model using the synthesis module of SIR. The results are represented by solid black lines in the upper panels of Figure~\ref{sint}. As expected, the circular polarization profiles are highly asymmetric ($\delta A \rightarrow 1$) and very similar to the blue Stokes profiles in Figs.~\ref{spa1t} and~\ref{spa2} (as they should because the atmospheric model was actually obtained from the inversion of those profiles). We then perform a synthesis using those very same physical parameters, but substituting the temperature enhancement above $\log\tau_{500} < -2$ for the temperature stratification of the HSRA model \citep{Gingerich1971}. The new atmosphere and resulting Stokes profiles are indicated in blue in the upper row of Fig.~\ref{sint}. As can be seen, Stokes $V$ has changed very little after modifying the model's temperature stratification. Stokes $I$, however, does show some differences.
The second experiment is identical to the first, except now we maintain the temperature stratification $T(\tau_{500})$ of the original model (see leftmost bottom row in Fig.~\ref{sint} and upper left panel in Fig.~\ref{mean}). This time, however, we neglect the change of sign in the LOS velocity in the upper layers, and, instead, we set ${\rm v}_{\rm LOS}=0$ above $\log\tau_{500} < -2$ (see second column bottom panel in Fig.~\ref{sint}). Then we perform a synthesis to calculate the emergent Stokes profiles from this modified atmosphere. The results are indicated in blue in the bottom row of Fig.~\ref{sint}. Again, while Stokes $I$ shows some differences, Stokes $V$ is almost the same as in the original model.
The same two experiments except using, as an original atmosphere, the red case (bottom row in Fig.~\ref{mean}) instead of the blue case (upper row in Fig.~\ref{mean}), lead to identical results. From these tests we can conclude that the neither the temperature enhancement nor the velocity change in the higher photospheric layers are needed to reproduce the asymmetry of the observer circular polarization profiles. Interestingly, they are needed to reproduce Stokes $I$ instead. Once the effect of these gradients above $\log\tau_{500} < -2$ on the asymmetries of Stokes $V$ have been ruled out, there is no particular reason as to why the atmosphere in this particular photospheric layer must be really there. In fact, at this point, this scenario is indistinguishable from one in which we extract the physical parameters above $\log\tau_{500} < -2$ in Fig.~\ref{mean}, and place them in a spatially unresolved second component located at the same height as the footpoint. This second component would be unmagnetized and carry a downflow associated with the upflowing footpoint (blue pixels in Fig.~\ref{spa1t}). However, next to the downflowing footpoint (red pixels in Fig.~\ref{spa1t}), this second component would carry an unmagnetized upflow. Its role would be limited to fitting, as we just described above, Stokes $I$, in particular its wings (see Sect.~\ref{profiles_sec}).
Independent of which of the two aforementioned scenarios is correct (non magnetic atmosphere above the loop's footpoints, or at the same height), it does not seem coincidental that this additional component harbors opposite LOS velocities than those of the footpoints. Unfortunately, with the current observations it is not possible to distinguish between these two possibilities, nor to study the nature of this second component and whether it interacts with the magnetic $\Omega$-loop.
\section{Conclusions}
The inspection of circular polarization signals at $\Delta\lambda=\pm272$~m{\AA} reveals the presence of pixels with highly blue-shifted single-lobed Stokes $V$ profiles grouped in patches of about 6.2$\pm$3 pixels. Very often (in 73 \% of the detected cases), these patches are accompanied by highly red-shifted single-lobed Stokes $V$ profiles that occupy a smaller, 3.5$\pm$2.1 pixels, area. In the remaining cases, 27 \% of the total, the patches of highly blue-shifted single-lobed Stokes $V$ profiles are isolated from other magnetic structures. We have referred to these patches as \emph{blue} and \emph{red} cases. The polarization profiles observed in these events are such that Stokes $I$ features a strongly blended blue or red wing in the blue or red cases, respectively. Although the amplitude of the circular polarization profiles is always higher in the blue cases than in the red cases, the sign of their only existing lobe is the same. These two kinds of events appear within 2\arcsec of each other. Interestingly, whenever they are very close, the pixels in between them display enhanced linear polarization signals. In this region of enhanced linear polarization, Stokes $V$ has a shape that closely resembles that of Stokes $Q$.
The study of the continuum intensity map $I_{c}$ shows that blue cases appear over bright regions, usually located at the granular edges, while red cases are placed in dark intergranular lanes. Overlaying the events to a magnetogram reveals that most of them occur around magnetic field concentrations (i.e., network), and almost never in the quiet-Sun internetwork.
The inversion of the Stokes profiles, detected as blue or red cases in the \emph{normal} map, allowed us to retrieve the optical depth stratification of the physical parameters through the solar photosphere: $T(\tau_{500})$, $B(\tau_{500})$, $\gamma(\tau_{500})$, ${\rm v}_{\rm LOS}(\tau_{500})$, etc. We have found that the temperature at $\log\tau_{500} = -1$ and $-3$ is larger than in the HSRA model (average quiet-Sun model), but lower than that model at $\log\tau_{500} = -2$. This happens in both blue and red cases, and the only difference between them at $\log\tau_{500} =0$, is that blue/red cases are slightly hotter/cooler than HSRA model. This last feature is just a consequence of the blue cases having larger continuum intensities than the red cases. The magnetic field strength $B$ in both types of events also behaves very similarly, with values of up to 1 kG in the deep photosphere ($\log\tau_{500} = 0$), which quickly decrease upward until vanishing at $\log\tau_{500}=-2$. The inclination of the magnetic field $\gamma$, on the other hand, indicates that the magnetic field has opposite polarities in these events, that is to say, if $\gamma < 90^{\circ}$ for blue case, then $\gamma > 90^{\circ}$ for the associated red case (or vice versa). The velocities also display a rather different behavior. In the blue cases, the LOS velocity ${\rm v}_{\rm LOS}$ changes from large blue-shifts in the deep photosphere ($\log\tau_{500} = 0$) to large red-shifts in the upper photosphere ($\log\tau_{500} = -3$). In the red cases, we observe exactly the opposite, that is, large red-shifts in the deep photosphere that turn into large blue-shifts higher up.
The inversions of the Stokes profiles, detected as blue or red cases in the time series, revealed the same properties of the physical parameters as in the \emph{normal} map. With the time series, however, we have been able to analyze the evolution of these events and discovered that their magnetic field configuration is that of an $\Omega$-loop, where the footpoints move further from each other as time passes. The physical process that best explains the observed configuration of the magnetic field while, at the same time, accounting for the high velocities at the footpoints, is a siphon flow. However, we were not able to detect the required gas pressure or magnetic field imbalance between footpoints. Moreover, although we have argued that it might be an artefact from the geometrical model we have defined prior to our inversions, the large velocity gradients at each of the footpoints, are difficult to reconcile with the picture of a siphon flow along a magnetic arch. Thus, at this point we are not able to fully establish the origin or physical mechanism responsible for these events.
Data with better spatial resolution, or including spectral lines that convey more reliable information on the higher layers of the solar photosphere, could help in the interpretation of the phenomena described in this paper. Hopefully, this will be possible with future ground-based facilities such as EST \citep{Collados_EST} and ATST \citep{Keil_ATST}, or from satellites like Solar-C \citep{Suematsu_SolarC}. It would be also interesting to search for these events into realistic MHD simulations to find out whether there is some kind of relation between their emergence and what occurs in the surrounding medium \cite[see][]{danilovic2014}.
\begin{acknowledgement}
C.Q.N. thanks to the researchers of the Kiepenheuer-Insitut f$\ddot{\rm u}$r Sonnenphysik (Freiburg, Germany) for their insightful comments that helped to develop this work. C.Q.N. also thanks to the director Oskar von der L$\ddot{\rm u}$he for the opportunity to have a wonderful stay in the KIS Institute, the same stay that seeds this paper. This work has been funded by the Spanish MINECO through Projects No. AYA2009-14105-C06-03 and AYA2011-29833-C06. The data used here were acquired in the framework of Hinode Operation Plan 14 (Hinode-Canary Islands joint campaign). \textit{Hinode} is a Japanese mission developed and launched by ISAS/JAXA, with NAOJ as a domestic partner, and NASA and STFC (UK) as international partners. It is operated by these agencies in cooperation with ESA and NSC (Norway). Extensive use of the NASA Astrophysical Data System has been made.
\end{acknowledgement}
\bibliographystyle{aa}
|
\section{Introduction}
\label{sec:1}
The study of transport processes on networks of varying types has attracted much recent interest \cite{duch, holme}, and is important from the point of view of applications. Earlier studies of transport processes on networks has been carried out for important classes like scale free and random networks \cite{tadic, ohira, zhao}. However, an important class of networks, viz., that of branching hierarchical networks \cite{pre11, pre12}, has not yet been extensively explored. Networks of the branching, hierarchical type are common in both real and engineering contexts. Examples of such networks include river networks \cite{river}, models of granular media \cite{copper}, and voter models \cite{voter,liggett}, the Domany Kinzel cellular-automata model \cite{domany} and the branching hierarchical model of the lung inflation process \cite{suki, suki2}. In this paper, we explore transport processes such as avalanche dynamics, and percolation on typical branching hierarchical lattices, and on the V lattice, a special realization of such lattices. We show that the behavior of transport on the V lattice is very distinct from that on typical realizations, and the V lattice is thus the critical realization of the branching hierarchical lattices. We also show that this special behavior is soon lost when the lattice is perturbed, indicating its critical nature. We also discuss the implications of our results.
\section{The Model}
\label{sec:2}
The base model for the load bearing hierarchical network \cite{pre11, janaki} discussed here is a regular $2D$ lattice of sites, where each site is connected at random to exactly one of its two neighbors in the layer below. The choice of the connection between the left and right neighbors to a given site $i^D$ at any $D$th layer is made with some probability $p$, where ($0<p<1$), for a connection to the left neighbor ${\it i_l^{D-1}}$, and probability $(1-p)$ to the right neighbor ${\it i^{D-1}_r}$.
Each site has the capacity to bear unit weight if it is not connected to either of its neighbors in the layer above, and can bear weight $w+1$ if it is connected to site(s) whose capacities add up to $w$, in the layer above. Thus, the capacity $w(i^D)$ of the $i^{th}$ site in the $D^{th}$ layer is given by the equation
\begin{equation}
w(i^D)=l(i^{D-1}_l,i^D)w(i^{D-1}_l)+l(i^{D-1}_r,i^D)w(i^{D-1}_r)+1
\end{equation}
The quantity $l(i^{D-1}_l,i^D)=1$ if a connection exists between ${\it i^{D-1}_l}$ and ${\it i^{D}}$, and $0$ if otherwise. The network consists of clusters, where a cluster is a set of sites connected with each other. The trunk is defined as the set of connected sites in the largest cluster with the highest weight bearing capacity in each layer. The sum of the weight bearing capacities of the sites along the trunk is defined as the trunk capacity $W_T$ of the given realization of the network. This model has similarity with the critical case $q(0,1)$ of model of granular media \cite{copper}, models of river networks \cite{river}, and the Takayasu model of the aggregation process with injection \cite{takayasu}. Other models analogous to our model are voter models \cite{voter,liggett}, the Domany-Kinzel cellular-automata model \cite{domany}, and the branching hierarchical model of the lung inflation process \cite{suki, suki2}.
The base model has a very specific and unique realization which incorporates the largest possible V shaped cluster that the base model could have. The V-cluster includes all the sites in the topmost layer, and ($M-D+1$) sites in the $D$th layer. One of the arms of the V constitutes the trunk, and all other connections run parallel to the arm of the V that is opposite to the trunk. The V-cluster is the largest possible cluster the base model could have. The trunk of this V-cluster bears the largest possible trunk capacity of the base model. We call this special lattice the V lattice. Figure.\ref{fig:vlattice} shows the V lattice configuration. Structures similar to the V lattice can be seen in riverine deltas \cite{riverdelta}, in Martian gullies \cite{marsgully}, and in granular flows \cite{shinbrot}, if the channels of maximal flow capacity are considered\footnote{In the natural structures, the channels of high capacity can have some channels of lower capacity joining them, leading to an overall symmetry between left and right connections, however one direction is favored by the channels of high capacity, due to some feature like the nature of the geographic terrain, leading to variants of the V structure for the high capacity channels alone.}.
\begin{figure}[htb]
\begin{center}
\includegraphics[height=5.5cm,width=10cm]{vlattice}
\vspace{0.5cm}
\caption{\label{fig:vlattice} The V lattice network of $D=8$ layers with $8$ sites per layers. The number in bracket beside each site denotes its weight bearing capacity. Solid lines show connection between the sites. The beaded line is the trunk of the V-cluster $C_2$.}
\end{center}
\end{figure}
\section{The V lattice: The critical case for transport}
\label{sec:2}
In this section we will discuss the critical nature of the V lattice, a special realization of the base lattice. The V lattice exhibits critical behavior for avalanche phenomena \cite{pramana}, site percolation phenomena \cite{pre12}, and capacity distribution and failure rate \cite{pre11} quite distinct from the behavior seen for the base lattice model, manifesting its criticality. Here, we focus only on the avalanche and percolation properties of the V lattice.
\subsection{Avalanche times distribution of the V lattice}
\label{subsec:2}
The behavior of avalanche processes for a given network topology provides an interesting example of the interplay between the nature of a transport process and the topology of the substrate. Here we aim to study the avalanche process of weight transmission on the V lattice. The avalanche is defined here in terms of weight transmission on the network.
To initiate the avalanche process on the network, some test weight $W_{test}$ is deposited on a randomly chosen site in the first layer of the network. Since our base network is a directed network, the flow of weight transmission takes place in the downward direction. The site in the first layer absorbs weight equal to its capacity and transmits the excess weight to its neighboring site it is connected to in the layer below. This process of weight transmission continues till there is no excess weight left, and the transmission is successful. If the $W_{test}$ is sufficiently large that it reaches the last layer without being fully absorbed, which completes one cycle of weight transmission, the excess weight is then deposited randomly on a site in the first layer and the second cycle of weight transmission starts here. If in the next cycle, the receiving site happens to be the one which has already saturated its capacity in the previous cycle, and now it does not have spare capacity to absorb any weight, the transmission is then said to be failed at that site. Otherwise the weight transmission continues till all excess weight is absorbed, and the transmission is said to be successful in this case. This process of the cycling of weight transmission through the lattice network is defined as an avalanche. The duration of an avalanche, or avalanche time, is defined to be the total number of layers traversed during all cycles of successful weight transmission by the test weight in the network.
We discuss the distribution of avalanche times of the V lattice. The avalanche times distribution $P(t)$ is, in fact, the distribution of the number of layers traversed during all cycles of successful avalanche transmissions by a test weight placed at a random site in the first layer for any lattice.
\begin{figure}[htb]
\begin{center}
\begin{tabular}{cccc}
(a)&
\hspace{-0.5cm}
\includegraphics[height=4.5cm,width=5.5cm]{aval_vlat_prob_100ls_05wt_bar}&
(b)&
\hspace{-0.5cm}
\includegraphics[height=4.5cm,width=5.5cm]{aval_vlat_prob_100ls_2wt_bar}\\
\end{tabular}{}
\caption{The avalanche time distribution $P(t)$ corresponding to $1000$ realizations for the V lattice network of side $M=100$ when tested for weights equal to (a) $0.05W_T$, and (b) $0.2W_T$. Small regimes for $0.05W_T$ ( as shown in inset of (a) ) and $0.2W_T$ ( as shown in inset of (b) ) display power law \mbox{$i.e.$} $P(t)\sim t^{-\alpha}$ behavior with exponent $\alpha=1.18$ and $\chi^2=0.0009$, and $\alpha=2.96$ and $\chi^2=52.197$ respectively.}
\label{fig:vlat}
\end{center}
\end{figure}
\begin{figure}[t]
\sidecaption[t]
\includegraphics[scale=0.55]{aval_vlat_prob_100ls_59wt_bar}
\caption{The avalanche times distribution $P(t)$, corresponding to $1000$ realizations, for the V lattice network of side $M=100$ when tested for weights equal to (a) $0.5W_T$, and (b) $0.9W_T$. No power law regime is seen in the distributions corresponding to $0.5W_T$, and $0.9W_T$.}
\label{fig:vlat_59wt}
\end{figure}
The avalanche times distribution $P(t)$ of the V lattice has been studied by Kachhvah and Gupte \cite{pramana} for test weights which are fractions of the trunk capacity of the V lattice. It has been observed that for this lattice, the avalanche times distribution $P(t)$ displays a power law regime, i.e., $P(t)\sim t^{-\alpha}$, which gradually starts disappearing as the test weight starts approaching the trunk capacity of the V lattice.
Figs. \ref{fig:vlat}, and \ref{fig:vlat_59wt} display the avalanche times distributions, corresponding to $1000$ realizations, for the V lattice. The power law distribution is seen in the V lattice as weight transmissions on the V lattice can achieve success at any one of the layers. This behavior of the existence and subsequent disappearance of the power law regime in the distribution has been one of the indications that the V lattice is a critical case of the base lattice. It is to be noted that the avalanche times distribution averaged over many realizations of the base lattice shows Gaussian behavior Fig. \ref{fig:base_aval} \cite{pramana}, and is quite distinct from the power-law behavior seen here.
\begin{figure}[t]
\sidecaption[t]
\includegraphics[scale=.55]{aval_durt_100ls_12wt_hist}
\caption{The avalanche times distributions $P(t)$ for the base lattice, corresponding to $1000$ realizations of networks of side $M=100$, exhibits Gaussian behavior when tested for weights equal to $0.1W_T$ and $0.2W_T$, where $W_T$ is the trunk capacity of base lattice.}
\label{fig:base_aval}
\end{figure}
\subsection{Explosive percolation on the V lattice}
\label{subsec:2}
The problem of site percolation has frequently been studied on networks due to its relevance for the problem of information transfer on the network.
Therefore, it is important to study site percolation on the hierarchical networks. The site percolation problem is set up as follows:
Typically, the transmission of information is modeled by packets of information hopping on the sites of the substrate networks. These packets could be the data packets of information in the Internet, which links computers of heterogeneous and high capacities capable of transmitting packets at high rates. For our hierarchical networks, the packets are deposited at a randomly chosen site on the topmost layer of the network. Each site retains the number of packets which saturates its capacity and the remaining packets are transmitted further. A packet at a given site sees the nearest neighbor sites linked to itself. If the targeted neighboring site is not fully occupied (i.e., it has not saturated its capacity), the packet moves there and looks for the next site that has spare capacity. If the target site is fully occupied, then the packet stops on the site which it occupies, and the transmission of the packet ends at that site. In this fashion, all packets hop from one site to another according to the vacancy available on the neighboring sites, and this process continues till all the packets come to rest at some site.
When the packet transmission has come to rest, at that time some of the sites would be occupied while the other would remain unoccupied or free. In this scenario, the network is composed of two sub-networks one of the unoccupied or free one and other of the occupied one, with the size of each network being a measure of the saturated or available capacity on the network.
Hence, we study the transition to percolation in the occupied or unoccupied sub-networks. For the free sub-network, one anticipate a transition from the percolating to the non-percolating state when the packet density $\mu$ increases, where the packet density is the ratio of total number of packets to total number of sites in the network.
In order to analyze the transition, we simulate the stochastic dynamics of packet transmission in the V lattice for different values of the packet density $\mu$ \cite{pre12}.
To study the percolation transition, the order parameter is defined to be the percolation strength $S=S_m/L$, where $S_m$ indicates the number of sites belonging to the largest connected cluster of the unoccupied sub-network and $L$ denotes the total number of sites in the network. The complement of the percolation strength $S$ is defined as $S_1=(1-S)$, which is a measure of the fraction of occupied sites or of the size of the occupied sub-network. We note that $S$, the size of the connected cluster of unoccupied sites is a measure of the capacity available for transport for a given packet density. This goes to zero at the jamming transition where no more capacity is available, and its complement $S_1$ goes to the size of the lattice. Here $S_1$ is compared with the size of the percolating cluster which spans the size of the lattice at the percolation transition.
\begin{figure}[htb]
\begin{center}
\begin{tabular}{cccc}
(a)&
\hspace{-0.6cm}
\includegraphics[height=4.5cm,width=5.5cm]{vlat_unoccupied_perc_sd_100ls}&
(b)&
\hspace{-0.6cm}
\includegraphics[height=4.5cm,width=5.5cm]{vlat_perc_spdev_100ls}\\
\end{tabular}{}
\caption{\label{fig:perco} (a) The order parameter $S$ and (b) its complement $S_1$ as a function of packet density $\mu$, averaged over $500$ realizations, for the V lattice networks of $(100\times100)$ sites.}
\end{center}
\end{figure}
The numerical study by \cite{pre12} shows that the percolation transition in the V lattice is a discontinuous or explosive one, which can be observed from the Figure. \ref{fig:perco}, and laws of the finite size scaling for continuous percolation transition does not hold for the V lattice. The percolation transition in the V lattice is contrary to second order continuous transition seen in the base lattice Fig. \ref{fig:base_percoS} \cite{pre12}, with associated second order percolation exponents. Here, again the V lattice shows critical behavior.
\begin{figure}[t]
\sidecaption[t]
\includegraphics[height=4.5cm,width=6.5cm]{orig_unoccupied_perc_sd_100ls_ps_fin}
\caption{The order parameter $S$ as a function of packet density $\mu$, averaged over $500$ realizations, for the free sub-network in the base lattice network of $(100\times100)$ sites.}
\label{fig:base_percoS}
\end{figure}
\section{Perturbed V lattices: annihilation of criticality}
\label{sec:3}
We have seen that the V lattice is the critical case of the base model as it displays behaviors for the avalanche times distribution and site percolation quite distinct from those seen for the base lattice. Here we are interested in exploring whether a slight perturbation introduced in the structure of the V lattice destroys the critical nature of the V lattice for avalanche and site percolation phenomena.
For this, perturbed V lattices are generated by switching the direction of the connections running parallel to arm opposite to the trunk in the V-cluster, with some probability $0<p<0.5$, but leaving the trunk or backbone untouched.
\subsection{Avalanche time distributions for the perturbed V lattice}
\label{subsec:3}
To demonstrate that the V lattice is the only realization of the base lattice which displays power law behavior for the avalanche time distribution, we studied these distributions for the perturbed V lattices, for 1000 realizations and test weights equal to $0.05W_T$ as shown in Fig. \ref{fig:vlat_perturbed}. It is observed that even a small perturbation ($p=0.1$) in the topology of the V-cluster of the V lattice, destroys the power law behavior seen for the V lattice. Thus, the V lattice is found to be the only case of the base lattice which displays power law behavior.
\begin{figure}[t]
\sidecaption[t]
\includegraphics[height=5.0cm,width=7.0cm]{vlat_aval_times_100ls_05wt_p}
\caption{The distributions of avalanche times of the V lattice and the perturbed V lattices obtained by switching connections with probability $p=0.1,\ldotp\ldotp\ldotp0.4$, corresponding to $1000$ realizations of a network of side $M=100$ when tested for weights equal to $0.05W_T$.}
\label{fig:vlat_perturbed}
\end{figure}
\subsection{The percolation transition for the perturbed V lattice}
\label{subsec:3}
Again, in order to demonstrate that the V lattice is the only realization of the base lattice which displays explosive percolation transition, we explore the percolation transition on the perturbed V lattices. The order parameter $S_1$ is computed as a function of the packet density $\mu$ (see Fig. \ref{fig:vlat_jump} (a)) for the V lattice perturbed by different values of $p$, \mbox{$i.e.$}, $0<p<0.5$. From Fig. \ref{fig:vlat_jump}(a), one can notice that as the perturbation in the V lattice is increased (\mbox{$i.e.$}, as $p$ increases), the size of the jump $\Delta S_1$ also reduces, where $\Delta S_1$ is the difference between the values of $S_1$ before and after the transition. In Fig. \ref{fig:vlat_jump}(b), the size of the largest jump $\Delta S_1$ is plotted, corresponding to different $p$, for different lattice sizes $L$. It is apparent from Fig. \ref{fig:vlat_jump}(b) that the largest jump size $\Delta S_1$, scales as a power law with system size $L$, defined as:
\begin{equation}
\Delta S_1 \sim L^{-\phi}.
\end{equation}
The above relation is similar to the relation between the largest jump size and the system size in \cite{nagler} to show the discontinuity of the percolation transition. We find that $\phi>0$ for all processes of the V lattice corresponding to different $p$. However, the values of $\phi$ are quite small. If, in the thermodynamic limit of infinite system size, we have
\begin{equation}
lim_{L\rightarrow \infty}\Delta S_1=0.
\end{equation}
i.e. if, in the limit of infinite system size, the size of the largest jump $\Delta S_1$ for the giant cluster goes to zero, then transitions are said to be weakly discontinuous.
The percolation transitions in the perturbed V lattices are weakly discontinuous. We conclude that even a slight perturbation introduced in the V lattice destroys the nature of the percolation transition in the V lattice.
\begin {figure}[htb]
\begin{center}
\begin{tabular}{cccc}
(a)&
\hspace{-0.5cm}
\includegraphics[height=4.5cm,width=5.5cm]{vlat_perc_ps_100ls_dev_occup}&
(b)&
\hspace{-0.5cm}
\includegraphics[height=4.5cm,width=5.5cm]{vlat_perc_ps_jump_ls_occup}\\
\end{tabular}{}
\caption{\label{fig:vlat_jump} Plot of (a) the order parameter $S_1$ of the occupied sub-lattice as a function of critical packet density $\mu$ for different values of $0<p<0.5$. (b) The largest jump size $\Delta S_1$ of the V lattice, scales as a power law as function of the lattice size $L$ \mbox{$i.e.$}, $\Delta S_1 \sim L^{-\phi}$, where $\phi$ is $0.0019$, $0.0023$, $0.0029$, $0.0059$, $0.0079$ and $0.011$ for $p$ equal to $0.05$, $0.15$, $0.25$, $0.35$, $0.45$, $0.50$, respectively.}
\end{center}
\end{figure}
We thus conclude that the V lattice is indeed a critical case of the base lattice, as even a slight distortion in the structure of the V lattice destroys its critical nature.
\section{Conclusions}
To summarize, we have observed that in the case of branching hierarchical structure, the lattice with V-shaped clusters shows special behavior for transport processes which use this lattice as the substrate. Avalanche processes on this lattice show power law behavior, and percolation behavior
on this lattice belongs to the explosive percolation class. This behavior sharply contrasts with the behavior seen for transport on typical realizations of the hierarchical networks, where avalanche transmissions are Gaussian, and the transition to percolation is of the usual second order type. Small perturbations to the V-cluster, rapidly destroy the special behavior, indicating the critical nature of the lattice. We note that this is one of the few examples where the nature of the substrate topology has led to the identification of a transition of the explosive percolation class. We hope our results go some way towards the identification of special topologies where critical behavior is observed for transport properties. Such an identification maybe of utility in practical systems like power grids and computer and communication networks.
\input{referenc}
\end{document}
|
\section{Introduction}
Quantum spin liquids are exotic phases of matter beyond
Landau's paradigm of symmetry-breaking\cite{wen}. In contrast to other familiar ground states of quantum magnets (such as antiferromagnets or ferromagnets) the quantum spin liquid ground state has a non-local entanglement between its local degrees of freedom. Similar `long range entanglement' also appears in the ground state of some other states of matter, for instance in the fractional quantum Hall states, and in Fermi/non-Fermi liquid metals. Since the original conception of the possibility of the quantum spin liquid, there has been tremendous progress in describing them theoretically. Many different kinds of quantum spin liquids are known to be theoretically possible. In the last decade a number of experimental candidates have also appeared. Interestingly all the existing experimental candidates seem to have gapless excitations which are not related to Goldstone modes of any broken symmetry. The theory of such gapless quantum spin liquids is however much less developed than the theory of gapped
quantum spin liquid states.
The currently known experimental candidate spin liquid materials may be conveniently grouped into two broad categories. The first - dubbed ``weak Mott insulators" - are close to the Mott transition and have significant virtual charge fluctuations. Both the layered organics
$\kappa-(ET)_2Cu_2(CN)_3$ and $EtMe_3 Sb[Pd(dmit)_2]_2$, and the three dimensional hyperkagome iridate $Na_4Ir_3O_8$ are all Mott insulating at ambient pressure but can be driven\cite{kanoda1,kato,takagi} through the Mott transition with application of moderate pressure. Indeed quantum spin liquid behavior may well be a common fate of {\em weak} Mott insulators.
The second category - dubbed ``strong Mott insulators" - have large charge gaps that are well separated from their exchange scales. These two classes of spin liquids likely require different theoretical approaches.
In weak Mott insulators gapless spin excitations are perhaps expected. At short length/time scales such insulators look roughly the same as a metal. As confirmed by various theoretical calculations\cite{sfs1,sfs2,matthewnumerics}, it is then reasonable that at longer length scales even though the charge localizes the spin continues to be carried by itinerant neutral fermions (the spinons).
Remarkably gapless excitations are found even in candidate spin liquids which are strong Mott insulators. Striking examples are the Kagome systems\cite{yslee,hiroi} $ZnCu_3(OH)_6Cl_2$
(Herbertsmithite) and $Cu_3V_2O_7(OH)_2.2H_2O$ (Volborthite). Similarly the recently reported spin-1 spin liquid\cite{balicas} $Ba_3NiSb_2O_9$ is also a strong Mott insulator.
Recent progress in Density Matrix Renormalization Group calculations of the isotropic spin-$1/2$ Kagome magnet\cite{dmrg} reveal a large spin gap ($0.14 J)$ which is not seen in the experiments on Herbertsmithite\cite{han}. The real model for this material is more complicated and must include Dzyaloshinski-Moriya as well as other anisotropies. Further there are significant impurity effects attributed to excess Cu spins sitting in between the Kagome planes. Other complications may exist in other materials. Nevertheless the surprisingly common occurance of gapless spin liquids in strongly Mott insulating materials leads to some fundamental questions in the theory of spin liquids.
In what theoretical framework should we discuss these gapless spin liquids?
Currently one framework that is known is to start with a slave particle description of the
physical spin operator in terms of fermionic neutral spin-$1/2$
spinons. The resulting spinon Hamiltonian is then first treated at
a mean field level. At this level of treatment the spinon spectrum
may well be gapless (with Fermi points or even a Fermi surface).
Going beyond mean field requires including fluctuations. The
resulting theory typically includes a fluctuating gauge field.
Thus in this approach gapless spin liquids are described by an
effective theory that involves gapless fermionic spinons coupled
to a fluctuating gauge field. If this theory is stable then this
is a legitimate description of a possible gapless spin liquid
phase.
The slave particle approach described above is deservedly popular and it certainly enables description of a class of quantum spin liquids.
However while this seems natural for weak Mott insulators (as is confirmed by many existing calculations) it is hardly obvious that this is the way forward in dealing with gapless spin liquids in, say, the Kagome magnets, or in the spin-$1$ magnet. As currently no other methods are known, fermionic spinon based approaches are the ``knee-jerk" reaction of theorists to the announcement of any experimental candidate gapless spin liquid.
A big open question in the field is whether there are other approaches that enables access to a different class of gapless spin liquids. More specifically do gapless spin liquids exist that are beyond the existing fermionic spinon (+ gauge field) paradigm? If so what is their phenomenology? Could they be more natural candidates for some of the
spin liquids that are reported?
In this paper we introduce a new theoretical route to gapless quantum spin liquids in spin systems with XY symmetry which appears to be distinct from the conventional fermionic spinon route. We utilize a dual description of such a spin system in terms of vortices in the XY spin. We show that quantum vortex liquid phases exist where there are gapless fermionic fields that carry the vorticity. We access these gapless vortex liquid phases through a `dual' parton approach where we fractionalize the fundamental vortex field into fermionic half-vortices. These may then form a gapless state. We describe an example of this construction for spin-1 quantum XY models on a honeycomb lattice. The dual parton approach is complementary to the standard one which fractionalizes the physical spin itself. Indeed it is likely that the phases we access have no simple and intuitive description within the standard approach. In Sec. \ref{parton} we give a construction of the phases in the usual parton language, but the
construction is quite complicated and requires auxiliary degrees of freedom on the lattice. This makes it much more natural to think in terms of fractionalizing vortices rather than spins.
An interesting earlier attempt with a motivation similar to ours was made in Ref. \onlinecite{avl1,avl2,avl3}, where the vortices were ``fermionized'' through a Chern-Simons flux-attachment. The fermions can then be put into a gapless band, and the resulting state becomes gapless and $U(1)$-symmetric.
However, this construction has problems with implementing time-reversal symmetry. In the simplest context in which such a fermionized vortex duality was attempted, it was shown in Ref. \onlinecite{topo} that such an approach would require an extra topological term into the original (un-dualized) description. Moreover, it was found recently\cite{avts12,hmodl} that such states realize time-reversal symmetry anomalously, and could
appear only on the surface of certain bosonic topological insulators. Therefore such a state realized in two dimensions will break time-reversal symmetry (an example was discussed in Ref.~\onlinecite{2dvrtxmtl}), and hence is not suitable to describe symmetric quantum spin liquids.
Closer to our approach is Ref. \onlinecite{hermele} which also employed a dual fermionic parton decomposition of the fundamental vortex field. The goal however was different from ours and that work did not attempt to find stable quantum spin liquid phases through the dual parton approach.
\section{Duality, vortices, and fractionalization}
\label{duality}
A spin system with $XY$ symmetry can be fruitfully viewed as a system of interacting bosons (with $S^z$ playing the role of boson number and $S^+$ the role of the boson creation operator $b^\dagger$). For a bosonic system with global~$U(1)$ symmetry, it is known that one can make a
duality mapping and describe the system in terms of vortices\cite{dual}. Specifically, one can write the conserved $U(1)$ current as the flux of a
non-compact $U(1)$ gauge field
\begin{equation}
j^{\mu}=\frac{\epsilon^{\mu\nu\lambda}}{2\pi}\partial_{\nu}a_{\lambda},
\end{equation}
The gauge field $a_{\mu}$ couples to a formally bosonic field $\Phi$ that corresponds to vortices in the order parameter of the global $U(1)$ symmetry.
If the vortices are gapped, we get a superfluid/ordered magnet with the global $U(1)$ symmetry broken, in which the gapless photons of the $a_{\mu}$ gauge field
corresponds to the Goldstone mode. But if the vortices are condensed instead, the whole system will be gapped due to the Higgs mechanism and we get a trivial Mott insulator/paramagnet. One can then ask the following question:
is it possible for the vortices to be in a stable gapless phase, so that the whole system is gapless while the global $U(1)$ symmetry is still preserved?
The route we will take is to fractionalize the vortex into two fermions, schemetically we have
\begin{equation}
\label{vfrac}
\Phi\sim\psi_1\psi_2,
\end{equation}
where $\Phi$ represents the vortex field rather than the physical spin as in usual parton construction, and $\psi_{1,2}$ are fermions representing ``fractionalized'' vortices. Such a ``dual'' parton construction can easily be made time-reversal invariant.
As in the usual parton construction the dual parton representation introduces an $SU(2)$ gauge redundancy. In this paper we will restrict ourselves to states where this $SU(2)$ gauge structure is broken down to $Z_2$. This will already be enough to produce a number of interesting states of the spin/boson system.
Before describing the gapless states we are interested in let us briefly describe some conventional states that will help build intuition about these fractionalized vortices. Consider the simplest such fractionalized vortex state, in which the fermionic fractional vortices $\psi$ is gapped, and couple to $a_\mu$ with gauge charge $1/2$. Then we may integrate them out to get a Maxwell action for the $a_\mu$. The gauge field fluctuations are thus gapless.
Physically this is a superfluid phase of the original bosons. However the presence of the gapped fractional vortex means that it is a {\em paired} superfluid where boson pairs $b^2$ are condensed without condensation of individual bosons $b$. (In spin language this is a `spin nematic' phase). The excitation spectrum of such a paired superfluid is well known. There is the usual gapless superfluid sound mode which in the dual description is identified with the propagating photon. The single boson survives as a gapped `Bogoliubov' quasiparticle, and may be described as an Ising spin $s$. In addition there is a half-vortex excitation where the phase of $b^2$ winds by $2\pi$. The Ising spin $s$ in turn acquires a phase $\pi$ upon encircling this vortex. Thus the Ising spin and the half-vortex are mutual semions. If we assign bose statistics to the half-vortex, its bound state with the Ising spin $s$ yields an excitation that is a fermion and also carries half-vorticity. Clearly we identify this with the
$\psi$ particles in the dual parton description.
Since we have assumed a state that has broken the dual $SU(2)$ gauge structure to $Z_2$, the $\psi$ carry a $Z_2$ gauge charge (in addition to the $U(1)$ gauge charge representing their vorticity). Correspondingly there is a $Z_2$ gauge vortex (the vision) which clearly must be identified with the $s$ particle, {\em i.e} the unpaired boson in the paired superfluid.
The original physical boson is the composite of a vison $s$ and
a $2\pi$-flux of the U(1) gauge field. Condensing the original boson means condensing the vison $s$, which confines the half-vortices, in agreement
with the usual description.
One can also consider a different phase in which the $\psi$ fermions are paired $\langle \psi\psi\rangle \neq0$. In such a phase $a_{\mu}$ is gapped, and we get a fractionalized liquid
with $Z_2$ topological order. The pair condensation quantizes the magnetic flux of $a_\mu$ in units of $2\pi$, which corresponds to an excitation $b_v$ with physical charge $1$ and boson statistics. This $b_v$ is however not to be identified with the physical boson $b$. Indeed the unpaired $\psi$ fermion survives as a Bogoliubov quasiparticle which is a mutual semion with the $b_v$. This is in contrast with the physical boson $b$ which is local with respect to all excitations. The state obtained this way has the topological order of a $Z_2$ quantum spin liquid but with symmetry realized in an unfractionalized manner.
The most interesting situation - which we explore in this paper - is when we put the $\psi$ fermions into a gapless band structure, such as a massless Dirac band. The gapless fermions will then couple to the
gauge field $a_{\mu}$ strongly, and form a gapless state which is not ordered. This is a gapless quantum spin liquid state which is potentially not accessible within the standard fermionic spinon-gauge field paradigm.
\section{Construction with frustrated quantum $XY$ model}
We now illustrate the construction of an example of such a gapless fractionalized quantum vortex liquid.
Consider a quantum $XY$ antiferromagnet on a two-dimensional triangular lattice. The Hamiltonian can be written as a rotor model ($b\sim e^{i\phi}$) in a
background static gauge field $\mathcal{A}^0$:
\begin{equation}
\label{rotor}
H=-J\sum_{\langle ij\rangle }\cos\left(\phi_i-\phi_j+\mathcal{A}^0_{ij} \right) +U\sum_i n_i^2+...
\end{equation}
where $\mathcal{A}^0$ gives a $\pi$ flux on each triangular plaquette (corresponding to antiferromagnetic exchange). We can think of the $\pi$ flux as requiring that there be an average vortex filling of $1/2$ per site on the dual honeycomb lattice. Going then to the vortex picture, we get a theory
of hard-core bosons (the vortices) at half-filling on the honeycomb lattice, coupled with a non-compact U(1) gauge field\cite{avl1}:
\begin{equation}
\tilde{H}=-2t\sum_{\langle ij\rangle }e^{ia_{ij}}\Phi_i^{\dagger}\Phi_j+h.c.+H_{Maxwell}+...
\end{equation}
where one may also have short range vortex interaction terms in general. For spin-half antiferromagnets ({\em i.e} where the original rotor number is $1/2$ per site on average), the vortices will themselves see a background $\pi$-flux
on each plaquette.
This system of hard-core bosonic vortices at half-filling could be fractionalized. To explore this
possibility, we fractionalize the vortex operator $\Phi$ into two fermions using the slave-particle formulation:
\begin{equation}
\Phi_i=\frac{1}{2}\epsilon^{\alpha\beta}\psi_{i,\alpha}\psi_{i,\beta}, \hspace{20pt} N_i=\frac{1}{2}\psi^{\dagger}_{\alpha}\psi_{\alpha},
\end{equation}
where $N$ denotes the vortex density, and $\alpha, \beta=1,2$ are the pseudo-spin indices, which transform under the internal $SU(2)$ gauge symmetry as $\psi_{\alpha}\to U_{\alpha\beta}\psi_{\beta}$.
The lattice symmetries act on $\psi_{i,\alpha}$ in the same ways as on $\Phi$ (up to an $SU(2)$ gauge transform). For a spin model, time
reversal acts on vortices as $\mathcal{T}: \Phi_i\to \Phi_i $, we have $\mathcal{T}: \psi_{i,\alpha}\to\psi_{i,\alpha}$ (again up to a gauge rotation).
The particle-hole symmetry (coming from $\pi$ rotation of spins around $x$ axis) transformation acting on the vortex is non-trivial:
$\mathcal{C}: \Phi_i\to \Phi_i^{\dag} $, which leads to
$\mathcal{C}: \psi_{i,\alpha}\to W_{i,\alpha\beta}\psi^{\dag}_{i,\beta}$ where $W$ is unitary with ${\rm{det}}(W)=-1$.
Our goal is to explore phases in which the fermions $\psi_{1,2}$ are deconfined and gapless. The gaplessness of the fermions should be stable in the sense that it is protected by symmetries.
It is instructive to reinterpret the ``fermionized vortex'' theory of Refs. \onlinecite{avl1,avl2,avl3} using this dual parton construction. It corresponds to putting $\psi_1$ in a Chern-insulator and $\psi_2$ in a gapless Dirac band. However, since time-reversal is broken in such a phase, the gaplessness is unprotected.
Now consider a particular mean field ansatz that meets our need:
\begin{equation}
\label{ansatz}
H_{mean}=-\sum_{ij}\left( \psi_{i\alpha}^{\dag}u^{\alpha\beta}_{ij}\psi_{j\beta}+h.c.\right),
\end{equation}
with the hopping matrices $u_{ij}$ given by
\begin{eqnarray}
\label{ans}
u_{i,i+\vec{a}_1}&=&u_{i,i+\vec{a}_2}=u_{i,i+\vec{a}_3}=\eta\tau^0+\lambda\tau^3, \\
u_{i,i+\vec{a}_1+\vec{a}_2-\vec{a}_3}&=&u_{i,i+\vec{a}_1-\vec{a}_2+\vec{a}_3}=u_{i,i-\vec{a}_1+\vec{a}_2+\vec{a}_3}=\xi\tau^1, \nonumber
\end{eqnarray}
where $\vec{a}_i$ are the three nearest-neighbor vectors on the honeycomb lattice, $\eta,\lambda,\xi$ are all real and $\tau^l$ are Pauli matrices acting on the $SU(2)$ gauge indices.
It is easy to see that $\langle\psi^{\dagger}_i\tau_{\mu}\psi_i\rangle=0$ on any site $i$ due to the particle-hole and time-reversal symmetries preserved by the mean field band structure. Therefore the mean field ansatz satisfies the gauge constraints on average and no further chemical potential term is needed.
To determine the remaining gauge structure in the phase described by Eq.~\eqref{ans}, one needs to calculate the $SU(2)$ gauge fluxes of the hopping matrices $u_{ij}$ on various loops, and
all the fluxes must be invariant under the unbroken gauge group\cite{wen}. It is then straightforward to see that only the $\mathbb{Z}_2$ gauge group $\psi_i\to(-1)^{s_i}\psi_i$ survives.
The ansatz given in Eq.~\eqref{ans} realizes all the lattice symmetries trivially, and is also manifestly time-reversal invariant.
Hence $\psi_{\alpha}$ transforms in exacly the same way as $\Phi$.
For charge conjugation $\mathcal{C}$, by inspection one can see that we should choose $\mathcal{C}: \psi_{i,\alpha}\to i(-1)^i\psi^{\dagger}_{i,\alpha}$, where $(-1)^i$ takes opposite values on different sublattices.
The fermions $\psi$ should also be coupled to the non-compact $U(1)$ gauge field $a_{\mu}$, and from the structure of the ansatz it is clear that the only way to do this consistently is to assign charge-$1/2$ to both $\psi_{1,2}$.
The virtue of the ansatz Eq.~\eqref{ans} is that it supports a gapless band structure protected by symmetries. It is straightforward to show that the band structure is described by four Dirac cones (similar to Graphene) near $\pm\vec{Q}$, and the low energy `mean field' Hamiltonian can be written as
\begin{eqnarray}
\label{eff}
H_{eff}(\vec{k})=&&\frac{\sqrt{3}}{2}\left(\eta\tau^0+\lambda\tau^3-2\xi\tau^1\right) \nonumber\\
&&\otimes\left( k_x\sigma^1\otimes v^3-k_y\sigma^2\otimes v^0\right),
\end{eqnarray}
where $\sigma^i$ acts on sub-lattice indices and $v^i$ on valley indices.
The symmetry actions on the low energy fermions in the above basis can be worked out through standard procedures:
we have the lattice translation $T_{(1,0)}: \psi\to \exp\left( i\frac{4\pi}{3}\sigma^0\otimes v^3\right)\psi$; $\pi/3$ rotation around the center of an honeycomb plaquette (a site of the original triangular lattice)
$R_{\pi/3}\psi=\sigma^2\otimes v^2e^{-i\frac{\pi}{6}\sigma^3\otimes v^3}\psi$; modified $x$-reflection
$\tilde{\mathcal{R}}_x=\mathcal{R}_x\mathcal{C}: \psi(k_x,k_y)\to\tau^0\otimes\sigma^0\otimes v^1\psi(-k_x,k_y)$ (note that a simple reflection flips vorticity); charge conjugation $\mathcal{C}:\psi(\vec{k})\to\tau^0\otimes\sigma^3\otimes v^1\psi^{\dag}(-\vec{k})$;
time reversal $\mathcal{T}$ ($\psi(k_x,k_y)\to\tau^0\otimes\sigma^0\otimes v^1\psi(-k_x,-k_y)$ and complex conjugation).
We can now analyse generally what fermion-bilinear terms are allowed by symmetries in the low-energy theory. It is then straightforward
to show that Eq.~\eqref{eff} is the most general form of symmetry-allowed low energy hamiltonian of the fermions. In particular,
a mass term that opens up a fermion gap is not allowed by symmetries. Hence the gaplessness of the fermions are symmetry-protected, at least perturbatively.
The above analysis can also be applied to a physical hard-core boson system on a honeycomb lattice at half-filling. The resulting state is a gapless $Z_2$ fractionalized liquid. The charge-$1/2$ fermions form four Dirac nodes, with a velocity anisotropy in the pseudo-spin space. As we will see below, when we view the theory instead as a vortex theory, the coupling to the $U(1)$ gauge field $a_{\mu}$ removes the velocity anisotropy at low energy.
The low energy Lagrangian with the $a_{\mu}$ field included can be written as
\begin{equation}
\label{dirac}
\mathcal{L}=\bar{\psi}\left[-i(\gamma^{\mu}+\hat{\gamma}^{\mu})(\partial_{\mu}+i\tilde{a}_{\mu})\right]\psi+\frac{1}{2e^2}f_{\mu\nu}^2.
\end{equation}
We have chosen the normalization $\tilde{a}_{\mu}=a_{\mu}/2$, $\eta=1$ and $\bar{\psi}=i\psi^{\dagger}\gamma^0$, where $\gamma^{\mu}=(\tau^0\otimes\sigma^3\otimes v^3, \tau^0\otimes\sigma^2\otimes v^0, \tau^0\otimes\sigma^1\otimes v^3)$, and $\hat{\gamma}^{\mu}=(0,(\lambda\tau^3-2\xi\tau^1)\otimes \sigma^2\otimes v^0, (\lambda\tau^3-2\xi\tau^1)\otimes\sigma^1\otimes v^3)$. This is not quite Dirac, but after including the fluctuation of the U(1) gauge field, it will renormalize
to a Dirac theory with emergent Lorentz symmetry. For small $\lambda$ amd $\xi$ and large $N_f$ (here we have $N_f=4$), we have
to first order
\begin{equation}
\frac{1}{\lambda}\frac{d\lambda}{dl}=\frac{1}{\xi}\frac{d\xi}{dl}=-\frac{64}{5\pi^2N_f}.
\end{equation}
Hence they are irrelevant to first order. The calculation is essentially identical to that in Ref. \onlinecite{asl}, where it was shown that the velocity anisotropy in real space was irrelevant (see Appendix \ref{rg} for details). Hence the low energy fixed point is simply the $QED_3$ with four flavors of Dirac fermions. It is believed that for flavor number $N_f$ not too small (greater than certain critical value $N_{f,c}$), the $QED_3$ fixed point is a CFT that is stable against spontaneous chiral symmetry breaking and fermion mass generation. The currently known\cite{tarun} upper-bound for $N_{f,c}$ is $N_{f,c}<6.6$. If the actual value of $N_{f,c}$ is less than four, our theory would describe a stable critical phase, rather than just a fine-tuned critical point.
One could also consider slightly modifying the system, by changing the flux on each plaquette in the rotor model Eq.~\eqref{rotor} from $\pi$ to $(\pi+2\pi\delta)$. This changes the vortex filling to $(1/2+\delta)$, which is also the filling fraction of the $\psi$ fermions. The same mean field ansatz Eq.~\eqref{ansatz} would then describe small fermi surfaces coupled with the gauge field $a_{\mu}$. As discussed in Ref. \onlinecite{max}, such a theory could describe a stable phase. However, we will not study this phase in detail since the modified system is harder to realize.
Since our purpose is mainly to illustrate the new formalism, the $QED_3$ fixed point theory is enough to convey the message.
The critical phase thus obtained has symmetries that are absent in the microscopic model, but emerge at low energy. These include the Lorentz invariance and the $SU(4)$ flavor symmetry. The $SU(4)$ group is generated by $\{\tau^0\otimes\sigma^0\otimes v^3, \tau^0\otimes\sigma^2\otimes v^1,\tau^0\otimes\sigma^2\otimes v^2,\tau^{1,2,3}\otimes\sigma^0\otimes v^0\}$
and their tensor products, which gives 15 generators in total, denoted by $T^a$, and by construction we have $[T^a,\gamma^{\mu}]=0$.
\section{Physical properties}
Now we look at particular features of the specific gapless vortex liquid state constructed above by considering physical observables. As a critical theory, we expect many of the physical observables will have algebraic correlation functions, and the exponents can be calculated using the CFT description.
A notable exception, however, is the in-plane spin-spin correlations. A spin-$1$ excitation $S^{\pm}$ is represented as the composite of the vison $s$ seen by the half-vortices $\psi_{\alpha}$ and a half-monopole in $a_{\mu}$. Since the vison $s$ is assumed to be gapped, we expect $S^{\pm}$ to be also gapped, and the in-plane spin-spin correlations $\langle S^+S^-\rangle $ will thus be short-ranged.
The out-of-plane spin-spin correlation functions $\langle S^zS^z\rangle $, on the contrary, decays algebraically. In fact, since $S^z$ is conserved in the CFT with the corresponding current represented as $j\sim da$, its scaling dimension must be $h_j=2$. We therefore have an interesting state with gapped $S^{\pm}$ but critical $S^z$.
In fact, the rich symmetry structure of our theory gives many other conserved currents which all have scaling dimension $h_j=2$. These include the vorticity $J_{\mu}=-i\bar{\psi}\gamma_{\mu}\psi$ and the $SU(4)$ flavor current $J^a_{\mu}=-i\bar{\psi}\gamma_{\mu}T^a\psi$.
The more interesting observables are nematic (spin-$2$) order parameters like $(S^+)^2$. In the dual picture these nematic operators are
represented as monopoles in $QED_3$. There are four flavors of Dirac fermions and each of them gives a zero-mode in the presence of $\pm 2\pi$ flux of $\tilde{a}_{\mu}$. A gauge-invariant state created by a monopole event should have half of the zero-modes filled. Hence there are six possible monopoles, obtained by filling two of the four zero-modes. We show in Appendix \ref{mono} that the monopole operators indeed transform in the same way as $(S^{\pm})^2$ at the three low energy momenta $(0,\pm\vec{Q})$.
The scaling dimension of the nematic operators is thus given by that of the monopole operators, which
can be calculated in the large-$N_f$ limit \cite{monopole,monopole2} (here we have $N_f$=4): $h_n\approx0.265N_f-0.038\approx1.02$. The relatively small scaling dimension reveals the proximity to nematically ordered phases.
To actually go to a nematic phase, the fermions $\psi_{\alpha}$ must acquire a mass gap. Since all the fermion mass terms break some global symmetries, the mass gap must be dynamically generated through spontaneous symmetry breaking, which agrees with the intuition that an ordered state on a frustrated lattice should break some symmetries other than the global $U(1)$.
Possible mass terms are the flavor $SU(4)$ adjoint $N^a=-i\bar{\psi}T^a\psi$ and scalar $M=-i\bar{\psi}\psi$. It turns out\cite{asl} that $M$ has a relatively large scaling dimension, so the primary instability comes from the $N^a$ terms.
The scaling dimensions of all the $N^a$ operators (which must be the same due to the $SU(4)$ symmetry) have been calculated\cite{asl} to leading order in $1/N_f$ which gives $h_N\approx2-64/3\pi^2N_f\approx1.46$.
In particular, these include the coplanar order parameter (spin chirality) $\kappa\sim K_z: \tau^{\mu\neq2}\otimes\sigma^3\otimes v^0$, and
the collinear order parameter (bond energy wave) $K_{\pm}: \tau^{\mu\neq2}\otimes\sigma^1\otimes (v^1\pm i v^2)$, which are expected to order in usual magnetic phases.
The large number of operators with the same relatively small scaling dimensions gives a clear manifestation of the emergent $SU(4)$ flavor symmetry. Physical observables that transform the same way with $N^a$ under microscopic symmetries will thus have the same scaling dimensions $h_N$.
It is straightforward to see that eight distinct physical operators are connected by the $SU(4)$ flavor symmetry.
We list all the physical operators in Appendix \ref{ops}.
Finally we mention some of the thermodynamic properties of this state. Clearly the low-$T$ heat capacity will be $C \propto T^2$, and the uniform spin susceptibility (for field coupling to $S^z$) will be $\chi^z \propto T$. The proportionality constants will depend on the (non-universal) Dirac velocity $v$ in a universal way such that the Wilson ratio $\frac{T\chi^z}{C}$ is a universal constant characteristic of the CFT (computable in the $\frac{1}{N_f}$ expansion).
There is another $QED_3$ fixed point for the theory in Eq.~\eqref{dirac}, by choosing the $\gamma$ matrices differently. We discuss this fixed point in Appendix \ref{afp}. We show that physical observables behave differently in this new fixed point, so it is indeed a distinct phase from the one discussed so far.
\section{Relationship to other states}
We now briefly consider how the gapless quantum vortex liquid state is related to other more familiar phases of the quantum $XY$ magnet. We have already discussed in Sec. \ref{duality} and later that if the vortex fields $\psi$ acquires a gap then the result is a phase with long range spin-nematic order ({\em i.e} where $(S^+)^2$ is ordered without ordering of $S^+$). As also discussed in Sec. \ref{duality}, if the $\psi$ pair and condense, the result is a $Z_2$ quantum spin liquid but without fractionalization of the global $U(1)$ quantum number.
Although being conceptually close to a nematic phase, the gapless vortex liquid can also be found near other conventional states in principle, via a direct phase transition.
To make a transition into a simple ordered state in which $\langle b\rangle \neq 0$, simply condense the vison $s$ seen by $\psi$, then the fermions $\psi$ will be confined and the vortex $\Phi$ will be gapped, which is nothing but an ordinary superfluid.
The trivial Mott insulator is also accessible through condensing the composite of the fermion half-vortex $\psi$ and the vison $s$ (which is a boson $v\sim\psi s$ due to the mutual semion statistics), which will confine all the fractional particles and make the system gapped.
\section{A parton construction}
\label{parton}
Here we give a parton construction of the phases we discuss. For this purpose we consider a modified system, in which a rotor $b\sim e^{i\phi}$ lives on the site of the triangular lattice, and an auxiliary rotor $\tilde{b}$ lives at the center of each plaquette of the triangular lattice. The auxiliary rotors thus form a honeycomb lattice.
We further demand the $U(1)$ rotation symmetry to act only on the $b$ rotors, but not on the auxiliary $\tilde{b}$ rotors. In other words we allow terms like $\Delta H\sim h\tilde{b}+h.c.$ in the Hamiltonian for the auxiliary sites. Now consider fractionalizing the auxiliary rotors as
\begin{equation}
\tilde{b}=\Phi_1\Phi_2,
\end{equation}
where $\Phi_1$ and $\Phi_2$ are bosons coupling to an emergent $U(1)$ gauge field $A_{\mu}$. For the $b$ rotors, we go to the dual pictur in terms of the vortex field $\Phi$ and the non-compact $U(1)$ gauge field $a_{\mu}$ whose flux is the charge of the $U(1)$ symmetry of the $b$ rotors. We then condense the following object:
\begin{equation}
\langle\Phi^{\dagger}_1\Phi\rangle\neq0.
\end{equation}
This is equivalent to putting the $b$ rotors and the $\Phi_1$ bosons into the $(001)$-hierarchical quantum hall state\cite{iqhe}. The condensate will Higgs the gauge field $A^-_{\mu}=(A_{\mu}-a_{\mu})/2$ and leaves only one gauge field $A^+_{\mu}=(A_{\mu}+a_{\mu})/2$ un-Higgsed. Since the gauge field $a_{\mu}$ in the vortex picture is non-compact, the un-Higgsed gauge field $A^+_{\mu}$ is also non-compact. Furthermore it is easy to check that $2\pi$-flux of $A^+_{\mu}$ carries $2\pi$-flux of $a_{\mu}$, which carries charge-$1$ under the $U(1)$-$XY$ symmetry.
The final effective theory is thus the same as the dual-vortex theory: the uncondensed boson $\Phi_2$ coupling to the non-compact gauge field $A^+_{\mu}$, where the flux of the gauge field carries $U(1)$ charge. We can now further fractionalize $\Phi_2$ as we did for Eq.~\eqref{vfrac}:
\begin{equation}
\Phi_2=\psi_1\psi_2,
\end{equation}
and the field theory for the phase we discussed thusfar is recovered.
We should emphasize that even though this is a construction in the usual parton language, it is much more natural to discuss our phase in the dual parton language, where the fractionalization if introduced straightforwardly with no auxiliary degrees of freedom required.
\section{Discussion}
We have described a concrete example of a gapless quantum spin liquid phase as a gapless fractionalized quantum vortex liquid. It is certainly desirable to find some concrete spin Hamiltonians to realize such phases. However this task is very challenging at this point. Instead we have focused on the more tractable phenomenological side: if these phases are indeed realized in some spin systems, what are the interesting features that could clearly distinguish them from the more familiar phases? We addressed this issue for the particular example in this work.
Clearly the dual parton approach developed here can be used to construct a variety of other gapless quantum vortex liquid states. An interesting example is a state where the fractionalized vortices form a gapless Fermi surface rather than Fermi points. The coupling of the vortices to the non-compact gauge field will lead to a low energy field theory similar to that of a spinon Fermi surface spin liquid\cite{sfs1,sfs2} or the HLR state\cite{hlr} of the half-filled Landau level. Of course as in the Dirac case discussed here the identification of physical operators in terms of the fields of the low energy theory will be different and will lead to different physical properties.
The states described in this paper should open our eyes to other new possible routes to gapless spin liquid behavior and suggest alternate possibilities for building phenomenologies of existing experimental candidates.
\textbf{Acknowledgement}: We thank F. Wang for helpful discussions. This work was supported by NSF DMR-1305741. This work was partially supported by a Simons Investigator award from the
Simons Foundation to Senthil Todadri.
|
\section{Introduction}
A cotorsion pair $\UV$, essentially equal to the notion of a torsion pair (\cite{IY}) on a triangulated category $\C$, is a unifying notion of $t$-structure (\cite{BBD}) and cluster tilting subcategory (\cite{KR}, \cite{KZ}).
Generalizing the case of $t$-structures and cluster tilting subcategories, a sequence of additive full subcategories $\Wcal\se\Hcal\se\Ccal$ is associated to any cotorsion pair, which gives an abelian category $\HW$ and a cohomological functor $H\co\C\to\HW$ (\cite{N1}, \cite{AN}). We call $\HW$ the {\it heart} of $\UV$.
In a work by Buan and Marsh (\cite{BM}), a generalization of (\cite{KZ}) has been given. From a rigid object in a triangulated category which is Krull-Schmidt $k$-linear $\Hom$-finite over a field $k$ equipped with a Serre functor, an integral preabelian category is constructed as an ideal quotient there.
This can be regarded as a {\it heart} of a pair of cotorsion pairs as in \cite{N2}.
We call a pair of cotorsion pairs $\Pcal=\STUV$ a {\it twin cotorsion pair} if it satisfies $\Exto(\Scal,\Vcal)=0$ (\cite{N2}). In fact, in the same manner as for the single cotorsion pair, a sequence of additive full subcategories $\Wcal\se\Hcal\se\Ccal$ is associated to any twin cotorsion pair, and it gives a preabelian category $\HW$ called {\it heart} and an additive functor $H\co\C\to\HW$.
Recently, Marsh and Palu (\cite{MP}) have established an equivalence of ideal quotients associated to rigid objects related by a mutation. In this article, to interpret this as an equivalence of hearts, we investigate a condition for the hearts to be equivalent.
Equivalences which we consider are the natural ones, namely, those compatible with associated functors as follows.
\begin{introdfn}
Let $\C$ and $\C\ppr$ be triangulated categories. Let $\Pcal=\STUV$ and $\Pcal\ppr=\STUVp$ be twin cotorsion pairs on $\C$ and $\C\ppr$, respectively. Let $F\co\C\ov{\simeq}{\lra}\C\ppr$ be a triangle equivalence. $\Pcal$ is said to be {\it heart-equivalent to $\Pcal\ppr$ along $F$} if there exists an equivalence of categories $E\co\HW\ov{\simeq}{\lra}\HWp$ which makes the following diagram commutative up to a natural isomorphism.
\[
\xy
(-10,7)*+{\C}="0";
(10,7)*+{\C\ppr}="2";
(-10,-7)*+{\HW}="4";
(10,-7)*+{\HWp}="6";
{\ar^{F}_{\simeq} "0";"2"};
{\ar_{H} "0";"4"};
{\ar^{H\ppr} "2";"6"};
{\ar_{E}^{\simeq} "4";"6"};
{\ar@{}|\circlearrowright "0";"6"};
\endxy
\]
\end{introdfn}
With this terminology, the problem we consider in this article is stated as follows.
\begin{introproblem}
Let $\C,\C\ppr,\Pcal,\Pcal\ppr$ and $F$ be as above.
Under which condition, does the following {\rm (I)} and {\rm (II)} become equivalent?
\begin{enumerate}
\item[{\rm (I)}] $\Pcal$ is heart-equivalent to $\Pcal\ppr$ along $F$.
\item[{\rm (II)}] The following conditions are satisfied.
\begin{itemize}
\item[{\rm (i)}] $H\ppr F(\Ucal)=H\ppr F(\Tcal)=0$.
\item[{\rm (ii)}] $HF\iv(\Ucal\ppr)=HF\iv(\Tcal\ppr)=0$.
\end{itemize}
\end{enumerate}
\end{introproblem}
It can be easily confirmed that {\rm (I)} always implies {\rm (II)}. Thus our goal is to find a condition, with which {\rm (II)} implies {\rm (I)}.
In our main theorem, this is given as follows.
\begin{introthm}
In the above situation, if the conditions
\begin{itemize}
\item[{\rm (B)}] \ $F(\Scal)\se\Scal\ppr\ast\Vcal\ppr [1]$, \ $F(\Vcal)\se\Scal\ppr [-1]\ast\Vcal\ppr$,
\item[{\rm (C)}] \ $F\iv(\Scal\ppr)\se\Scal\ast\Vcal [1]$, \ $F\iv(\Vcal\ppr)\se\Scal [-1]\ast\Vcal$,
\end{itemize}
are satisfied, then the above {\rm (I)} and {\rm (II)} become equivalent.
\end{introthm}
In section 2, we review the definitions and results used in this article, mainly from \cite{N2}. In section 3, we reduce the problem to a manageable, equivalent one. In section 4 we introduce what happens in the case of single cotorsion pairs. Following the argument by Zhou and Zhu (\cite{ZZ}), conditions {\rm (I)} and {\rm (II)} are shown to be equivalent in this case, without any extra assumption. In section 5, for general twin cotorsion pairs, we give a sufficient condition for {\rm (II)} to imply {\rm (I)}. In section 6, we demonstrate how this condition can be applied to the equivalence given in \cite{MP}.
We also would like to remark that hearts of twin cotorsion pairs on exact categories and their equivalences have been studied by Liu in \cite{L1} and \cite{L2}.
\bigskip
For any category $\K$, we write abbreviately $X\in\K$, to indicate $X$ is an object of $\K$.
For any $X,Y\in\K$, let $\K(X,Y)$ denote the set of morphisms from $X$ to $Y$.
If $\K$ is additive, then for a subcategory $\Mcal\se\K$, we define its right perpendicular category $\Mcal\ppp$ to be the full subcategory of $\K$ consisting of those $X\in\K$ satisfying $\K(\Mcal,X)=0$. Dually, ${}\ppp\Mcal$ denotes the full subcategory of those $X$ satisfying $\K(X,\Mcal)=0$.
If $\Ical\subseteq\K$ is a full additive subcategory, then $\K/\Ical$ is defined to be the ideal quotient of $\K$ by $\Ical$. Namely, $\K/\Ical$ is an additive category defined as follows.
\begin{itemize}
\item[-] Objects in $\K/\Ical$ are the same as those in $\K$.
\item[-] For any $X,Y\in\K$, the morphism set is defined by
\[ (\K/\Ical)(X,Y)=\K(X,Y)/\{ f\in\K(X,Y) \mid f\ \text{factors through some}\ I\in\Ical\}. \]
\end{itemize}
\section{Review of Definitions and results}
Throughout this article, $\C$ denotes a triangulated category with suspension functor $[1]$.
\begin{dfn}
Let $\Ucal,\Vcal\se\C$ be a full subcategories closed under isomorphisms, finite direct sums and summands.
The pair $\UV$ is a {\it cotorsion pair on} $\C$ if it satisfies the following.
\begin{itemize}
\item[{\rm (i)}] $\C=\Ucal\ast\Vcal [1]$.
\item[{\rm (ii)}] $\Exto\UV=0$, where $\Exto(X,Y)=\C(X,Y[1])$ for any $X,Y\in\C$.
\end{itemize}
Here, $\Ucal\ast\Vcal [1]$ denotes the full subcategory of $\C$ consisting of those $C\in\C$ admitting a distinguished triangle
\[ U\rightarrow C\rightarrow V[1]\rightarrow U[1] \]
with $U\in\Ucal$ and $V\in\Vcal$.
\end{dfn}
\begin{dfn}
Let $\ST,\UV$ be two cotorsion pairs on $\C$. The pair $\Pcal=\STUV$ is a {\it twin cotorsion pair on} $\C$ if it satisfies $\Exto(\Scal,\Vcal)=0$. Note that this condition is equivalent to $\Scal\se\Ucal$, and also to $\Vcal\se\Tcal$.
\end{dfn}
Remark that a cotorsion pair $\UV$ on $\C$ can always be regarded as a twin cotorsion pair $(\UV,\UV)$. This is regarded as a {\it degenerated} case of a twin cotorsion pair, as follows.
\begin{dfn}
A twin cotorsion pair $\Pcal=\STUV$ is said to be {\it degenerated to a
single cotorsion pair} if it satisfies
\[ \Scal=\Ucal\quad \text{and}\quad \Tcal=\Vcal. \]
\end{dfn}
\begin{dfn}\label{DefNotaW}
For a twin cotorsion pair $\Pcal=\STUV$ on $\C$, put
\begin{eqnarray*}
&\Wcal=\Ucal\cap\Tcal,&\\
&\Cp=\Wcal\ast\Vcal [1],\ \ \ \Cm=\Scal [-1]\ast\Wcal,&\\
&\Hcal=\Cp\cap\Cm.&
\end{eqnarray*}
We call the ideal quotient $\Hcal/\Wcal$ the {\it heart} of $\Pcal$. Remark that there are inclusions of full subcategories
\[
\xy
(-16,0)*+{\HW}="0";
(0,8)*+{\CpW}="2";
(0,-8)*+{\CmW}="4";
(16,0)*+{\CW}="6";
{\ar@{^(->} "0";"2"};
{\ar@{^(->} "0";"4"};
{\ar@{^(->} "2";"6"};
{\ar@{^(->} "4";"6"};
{\ar@{}|\circlearrowright "0";"6"};
\endxy.
\]
\end{dfn}
For any morphism $f\co A\to B$ in $\C$, its image under the quotient functor $\C\ov{\quot}{\lra}\CW$ is denoted by $\und{f}\co A\to B$.
\begin{lem}
$\Ucal\se\Cm$ and $\Tcal\se\Cp$ holds.
\end{lem}
\begin{lem}
$\Hcal\cap(\Ucal\ast\Tcal)=\Wcal$ holds.
\end{lem}
\begin{proof}
These are confirmed easily.
\end{proof}
\begin{fact}
$($Corollary 3.8, 3.9 in \cite{N2}.$)$
\begin{enumerate}
\item Inclusion $\CpW\hr\CW$ has a left adjoint $\tp\co\CW\to\CpW$.
\item Inclusion $\CmW\hr\CW$ has a right adjoint $\tm\co\CW\to\CmW$.
\end{enumerate}
\end{fact}
\begin{dfn}
Let $C\in\C$ be any object.
\begin{enumerate}
\item A diagram
\begin{equation}\label{ReflTri}
\xy
(-28,0)*+{S[-1]}="0";
(-8,0)*+{C}="2";
(8,0)*+{Z_C}="4";
(24,0)*+{S}="6";
(-18,-12)*+{U}="8";
(-18,-5)*+{_{\circlearrowright}}="10";
{\ar^{} "0";"2"};
{\ar^{z_C} "2";"4"};
{\ar_{} "4";"6"};
{\ar_{} "0";"8"};
{\ar_{} "8";"2"};
\endxy
\end{equation}
is called a {\it reflection triangle} if it satisfies the following.
\begin{itemize}
\item[{\rm (i)}] $S\in\Scal, U\in\Ucal$.
\item[{\rm (ii)}] $S[-1]\to C\ov{z_C}{\lra}Z_C\to S$ is a distinguished triangle in $\C$.
\item[{\rm (iii)}] $Z_C\in\Cp$.
\end{itemize}
\item Dually, a diagram
\begin{equation}\label{CoreflTri}
\xy
(-24,0)*+{V}="0";
(-8,0)*+{K_C}="2";
(8,0)*+{C}="4";
(28,0)*+{V[1]}="6";
(18,-12)*+{T}="8";
(18,-5)*+{_{\circlearrowright}}="10";
{\ar^{} "0";"2"};
{\ar^{k_C} "2";"4"};
{\ar^{} "4";"6"};
{\ar_{} "4";"8"};
{\ar_{} "8";"6"};
\endxy
\end{equation}
is called a {\it coreflection triangle} if it satisfies the following.
\begin{itemize}
\item[{\rm (i)}] $V\in\Vcal, T\in\Tcal$.
\item[{\rm (ii)}] $V\to K_C\ov{k_C}{\lra}C\to V[1]$ is a distinguished triangle in $\C$.
\item[{\rm (iii)}] $K_C\in\Cm$.
\end{itemize}
\end{enumerate}
\end{dfn}
\begin{fact}$($Corollary 3.8 in \cite{N2}.$)$
For any reflection triangle $(\ref{ReflTri})$, there exists a unique morphism $\ze\in(\CW)(\tp(C),Z_C)$ which is compatible with the adjunction $C\to\tp(C)$ as follows.
\[
\xy
(-10,6)*+{C}="0";
(10,6)*+{Z_C}="2";
(0,-8)*+{\tp(C)}="4";
(0,10)*+{}="5";
{\ar^{\und{z}_C} "0";"2"};
{\ar_(0.4){\mathrm{adjunction}} "0";"4"};
{\ar^(0.4){\ze} "2";"4"};
{\ar@{}|\circlearrowright "4";"5"};
\endxy
\]
This $\ze$ is an isomorphism in $\CW$.
Dually for coreflection triangles.
\end{fact}
\begin{rem}\label{RemExistRef}$($Definition 3.4 in \cite{N2}.$)$
For any $C\in\C$, a reflection triangle always exists. Indeed, by the condition $\C=\Scal\ast\Tcal [1]=\Ucal\ast\Vcal [1]$ and the octahedron axiom, we can draw a diagram
\[
\xy
(-16.2,16)*+{S_C[-1]}="0";
(-14.9,-1)*+{U_C}="2";
(-13.5,-17.5)*+{T_C}="4";
(-4,2)*+{C}="6";
(2,-5.2)*+{Z_C}="8";
(17,8.5)*+{V_C[1]}="10";
(-8.5,-7.5)*+_{_{\circlearrowright}}="12";
(-11.5,5.5)*+_{_{\circlearrowright}}="14";
(4.3,0.5)*+_{_{\circlearrowright}}="14";
{\ar_{} "0";"2"};
{\ar_{} "2";"4"};
{\ar^{} "0";"6"};
{\ar_{} "2";"6"};
{\ar_{z_C} "6";"8"};
{\ar^{} "6";"10"};
{\ar_{} "4";"8"};
{\ar_{} "8";"10"};
\endxy
\]
satisfying $S_C\in\Scal, U_C\in\Ucal, V_C\in\Vcal, T_C\in\Tcal$, in which
\begin{eqnarray*}
&S_C[-1]\to C\to Z_C\to S_C,\quad S_C[-1]\to U_C\to T_C\to S_C,&\\
&U_C\to C\to V_C[1]\to U_C[1],\quad T_C\to Z_C\to V_C[1]\to T_C[1]&
\end{eqnarray*}
are distinguished triangles. Since $T_C\in\Tcal\cap(\Ucal\ast\Scal)=\Wcal$, it follows $Z_C\in\Cp$.
Dually for the existence of coreflection triangles.
\end{rem}
Reflection triangles are \lq\lq functorial" in the following sense.
\begin{rem}\label{RemFunctRefl}
Let $A,B\in\C$ be any object, and let
\begin{equation}\label{ReflTriA}
\xy
(-28,0)*+{S_A[-1]}="0";
(-8,0)*+{A}="2";
(8,0)*+{Z_A}="4";
(24,0)*+{S_A}="6";
(-18,-12)*+{U_A}="8";
(-18,-5)*+{_{\circlearrowright}}="10";
{\ar^{} "0";"2"};
{\ar^{z_A} "2";"4"};
{\ar_{} "4";"6"};
{\ar_{} "0";"8"};
{\ar_{} "8";"2"};
\endxy
\end{equation}
\begin{equation}\label{ReflTriB}
\xy
(-28,0)*+{S_B[-1]}="0";
(-8,0)*+{B}="2";
(8,0)*+{Z_B}="4";
(24,0)*+{S_B}="6";
(-18,-12)*+{U_B}="8";
(-18,-5)*+{_{\circlearrowright}}="10";
{\ar^{} "0";"2"};
{\ar^{z_B} "2";"4"};
{\ar_{} "4";"6"};
{\ar_{} "0";"8"};
{\ar_{} "8";"2"};
\endxy
\end{equation}
be reflection triangles. For any morphism $f\in\C(A,B)$, there exists a morphism of diagrams from $(\ref{ReflTriA})$ to $(\ref{ReflTriB})$ as
\[
\xy
(-26,7)*+{S_A[-1]}="0";
(-7,7)*+{A}="2";
(8,7)*+{Z_A}="4";
(22,7)*+{S_A}="6";
(-16,6)*+{}="7";
(-16,16)*+{U_A}="8";
(-26,-7)*+{S_B[-1]}="10";
(-7,-7)*+{B}="12";
(8,-7)*+{Z_B}="14";
(22,-7)*+{S_B}="16";
(-16,-6)*+{}="17";
(-16,-16)*+{U_B}="18";
{\ar^{} "0";"2"};
{\ar^{z_A} "2";"4"};
{\ar^{} "4";"6"};
{\ar^{} "0";"8"};
{\ar^{} "8";"2"};
{\ar^{} "0";"10"};
{\ar^{f} "2";"12"};
{\ar^{g} "4";"14"};
{\ar^{} "6";"16"};
{\ar^{} "10";"12"};
{\ar_{z_B} "12";"14"};
{\ar_{} "14";"16"};
{\ar^{} "10";"18"};
{\ar^{} "18";"12"};
{\ar@{}|\circlearrowright "0";"12"};
{\ar@{}|\circlearrowright "2";"14"};
{\ar@{}|\circlearrowright "4";"16"};
{\ar@{}|\circlearrowright "7";"8"};
{\ar@{}|\circlearrowright "17";"18"};
\endxy
\]
In fact, if we take a distinguished triangle
\[ W\to Z_A\to V[1]\to W[1]\qquad(W\in\Wcal,V\in\Vcal), \]
then in the diagram
\[
\xy
(0,12)*+{S_A[-1]}="0";
(0,4)*+{A}="1";
(0,-4)*+{B}="2";
(0,-12)*+{Z_B}="3";
(-16,-12)*+{W}="4";
(16,-12)*+{V[1]}="5";
(12,0)*+{U_A}="6";
(-2,0)*+{}="7";
{\ar^{} "0";"1"};
{\ar_{f} "1";"2"};
{\ar_{z_B} "2";"3"};
{\ar^{} "4";"3"};
{\ar^{f} "3";"5"};
{\ar^{} "0";"6"};
{\ar_{} "6";"3"};
{\ar@{}|\circlearrowright "6";"7"};
\endxy,
\]
the equality $\C(U_A,V[1])=0$ implies that the composed morphism $S_A[-1]\to Z_B$ factors through $W$, and thus should be the zero morphism, by $\C(S_A[-1],W)=0$.
Also remark that $\und{g}\in(\CpW)(Z_A,Z_B)$ with the commutativity $\und{g}\ci\und{z}_A=\und{z}_B\ci\und{f}$ is unique by the adjoint property.
Dually for coreflection triangles.
\end{rem}
\begin{rem}\label{RemTauUT}$($Lemmas 3.10, 3.11 in \cite{N2}.$)$
For any $C\in\C$, the following holds.
\begin{enumerate}
\item $\tp(C)=0\ \ \LR\ \ C\in\Ucal$.
\item $\tm(C)=0\ \ \LR\ \ C\in\Tcal$.
\end{enumerate}
\end{rem}
\begin{fact}\label{FactUSTV}
$($Lemmas 2.12, 2.13 in \cite{N2}.$)$
\begin{enumerate}
\item If $U[-1]\to A\to B\to U$ is a distinguished triangle satisfying $U\in \Ucal$, then
\[ A\in\Cm\ \ \tc\ \ B\in\Cm. \]
\item If $S[-1]\to A\to B\to S$ is a distinguished triangle satisfying $S\in \Scal$, then
\[ A\in\Cm\ \ \LR\ \ B\in\Cm. \]
\item If $T\to A\to B\to T[1]$ is a distinguished triangle satisfying $T\in \Tcal$, then
\[ B\in\Cp\ \ \tc\ \ A\in\Cp. \]
\item If $V\to A\to B\to V[1]$ is a distinguished triangle satisfying $V\in \Vcal$, then
\[ B\in\Cp\ \ \LR\ \ A\in\Cp. \]
\end{enumerate}
\end{fact}
\begin{prop}
For the compositions of functors
\begin{eqnarray*}
&\tp\tm:=\big(\CW\ov{\tm}{\lra}\CmW\hr\CW\ov{\tp}{\lra}\CpW\hr\CW \big),&\\
&\tm\tp:=\big(\CW\ov{\tp}{\lra}\CpW\hr\CW\ov{\tm}{\lra}\CmW\hr\CW \big),&
\end{eqnarray*}
there exists a natural isomorphism $\tp\tm\cong\tm\tp$.
\end{prop}
\begin{proof}
For any $C\in\C$, take a reflection triangle $(\ref{ReflTri})$ and a coreflection triangle $(\ref{CoreflTri})$.
The equalities $\C(S_C[-1],T_C)=0$ and $\C(U_C,V_C[1])=0$ give morphisms $Z_C\to T_C$ and $U_C\to K_C$, which make
\[
\xy
(16,-13)*+{V_C[1]}="0";
(10,8)*+{Z_C}="2";
(0,-1)*+{C}="6";
(-10,8)*+{K_C}="8";
(-16,-13)*+{S_C[-1]}="10";
(-24,-3)*+{U_C}="12";
(24,-3)*+{T_C}="14";
(10.5,1.5)*+_{_{\circlearrowright}}="20";
(-10.5,1.5)*+_{_{\circlearrowright}}="22";
(-14,-7)*+_{_{\circlearrowright}}="24";
(14,-7)*+_{_{\circlearrowright}}="26";
{\ar^{} "6";"14"};
{\ar_{} "6";"0"};
{\ar^{z_C} "6";"2"};
{\ar^{k_C} "8";"6"};
{\ar_{} "10";"6"};
{\ar^{} "12";"6"};
{\ar_{} "10";"12"};
{\ar^{} "12";"8"};
{\ar^{} "2";"14"};
{\ar_{} "14";"0"};
\endxy
\]
commutative. Then the composed morphisms $S_C[-1]\to K_C$ and $Z_C\to V_C[1]$ make the following diagram commutative.
\[
\xy
(19,-13)*+{V_C[1]}="0";
(8,4)*+{Z_C}="2";
(0,-2)*+{C}="6";
(-8,4)*+{K_C}="8";
(-19,-13)*+{S_C[-1]}="10";
(-26,-4)*+{U_C}="12";
(26,-4)*+{T_C}="14";
(9,-3)*+_{_{\circlearrowright}}="20";
(-9,-3)*+_{_{\circlearrowright}}="22";
(-19,-5)*+_{_{\circlearrowright}}="24";
(19,-5)*+_{_{\circlearrowright}}="26";
{\ar^{} "2";"0"};
{\ar_{} "6";"0"};
{\ar_{z_C} "6";"2"};
{\ar_{k_C} "8";"6"};
{\ar_{} "10";"6"};
{\ar^{} "10";"8"};
{\ar_{} "10";"12"};
{\ar^{} "12";"8"};
{\ar^{} "2";"14"};
{\ar_{} "14";"0"};
\endxy
\]
If we complete $S_C[-1]\to K_C$ into a distinguished triangle $S_C[-1]\to K_C\to Q\to S_C$, then by the octahedron axiom, we obtain a diagram
\[
\xy
(19,-13)*+{V_C[1]}="0";
(8,4)*+{Z_C}="2";
(0,16)*+{Q}="4";
(0,-2)*+{C}="6";
(-8,4)*+{K_C}="8";
(-19,-13)*+{S_C[-1]}="10";
(0,7)*+_{_{\circlearrowright}}="12";
(9,-3)*+_{_{\circlearrowright}}="14";
(-9,-3)*+_{_{\circlearrowright}}="16";
{\ar^{} "2";"0"};
{\ar^{} "4";"2"};
{\ar_{} "6";"0"};
{\ar_{z_C} "6";"2"};
{\ar_{k_C} "8";"6"};
{\ar_{} "10";"6"};
{\ar^{} "8";"4"};
{\ar^{} "10";"8"};
\endxy
\]
in which
\begin{eqnarray*}
&S_C[-1]\to K_C\to Q\to S_C,\quad S_C[-1]\to C\to Z_C\to S_C,&\\
&K_C\to C\to V_C[1]\to K_C[1],\quad Q\to Z_C\to V_C[1]\to Q[1]&
\end{eqnarray*}
are distinguished triangles. By Fact \ref{FactUSTV},
\begin{itemize}
\item[-] $Z_C\in\Cp$ implies $Q\in\Cp$,
\item[-] $K_C\in\Cm$ implies $Q\in\Cm$.
\end{itemize}
Thus we have $Q\in\Hcal$, and
\[
\xy
(-28,0)*+{S_C[-1]}="0";
(-8,0)*+{K_C}="2";
(8,0)*+{Q}="4";
(24,0)*+{S_C}="6";
(-18,-12)*+{U_C}="8";
(-18,-5)*+{_{\circlearrowright}}="10";
{\ar^{} "0";"2"};
{\ar^{} "2";"4"};
{\ar_{} "4";"6"};
{\ar_{} "0";"8"};
{\ar_{} "8";"2"};
\endxy
\]
\[
\xy
(-24,0)*+{V_C}="0";
(-8,0)*+{Q}="2";
(8,0)*+{Z_C}="4";
(28,0)*+{V_C[1]}="6";
(18,-12)*+{T_C}="8";
(18,-5)*+{_{\circlearrowright}}="10";
{\ar^{} "0";"2"};
{\ar^{} "2";"4"};
{\ar^{} "4";"6"};
{\ar_{} "4";"8"};
{\ar_{} "8";"6"};
\endxy
\]
are reflection and coreflection triangles, respectively. This implies $\tp(K_C)\cong Q\cong\tm(Z_C)$.
\end{proof}
\begin{cor}
$\tp(\CmW)\se\HW$ and $\tm(\CpW)\se\HW$ hold.
\end{cor}
\begin{dfn}
Define functors $h$ and $H$ to be the compositions
\begin{eqnarray*}
h&=&\tp\tm\ =\ \big(\CW\ov{\tm}{\lra}\CmW\ov{\tp}{\lra}\HW\big),\\
H&=&\big(\C\ov{\quot}{\lra}\CW\ov{h}{\lra}\HW\big).
\end{eqnarray*}
Also, let $\iota$ be the inclusion functor
\[ \iota\co\HW\hr\CW. \]
We have a natural isomorphism $h\ci\iota\cong\Id_{\HW}$.
\end{dfn}
\begin{prop}\label{RemHTauU}
For any $C\in\C$, the following are equivalent.
\begin{enumerate}
\item $H(C)=0$.
\item $\tm(C)\in\Ucal/\Wcal$.
\item $\tp(C)\in\Tcal/\Wcal$.
\end{enumerate}
\end{prop}
\begin{proof}
This follows from Remark \ref{RemTauUT}.
\end{proof}
\begin{prop}
$H(\Ucal\ast\Tcal)=0$ holds. Thus $H$ factors through $\C\ov{\quot}{\lra}\C/\Ucal\ast\Tcal$ as follows.
\[
\xy
(-6,6)*+{\C}="0";
(0,-10)*+{}="1";
(-10,-6)*+{\C/\Ucal\ast\Tcal}="2";
(10,-6)*+{\HW}="4";
{\ar_{\quot} "0";"2"};
{\ar^{H} "0";"4"};
{\ar^{} "2";"4"};
{\ar@{}|\circlearrowright "0";"1"};
\endxy
\]
In particular we have $H(\Ucal)=H(\Tcal)=0$.
\end{prop}
\begin{proof}
Let $U\to C\to T\to U[1]$ be any distinguished triangle satisfying $U\in\Ucal$ and $T\in\Tcal$. If we decompose $T$ into a distinguished triangle
\[ V_T\to U_T\to T\to V_T[1]\quad (U_T\in\Ucal, V_T\in\Vcal), \]
then by the octahedron axiom,
we obtain a diagram
\[
\xy
(-20,16)*+{U}="0";
(0.5,15)*+{L}="2";
(17,14)*+{U_T}="4";
(-2,4)*+{C}="6";
(6.1,-0.6)*+{T}="8";
(-5.8,-14.2)*+{V_T[1]}="10";
(7.5,8.5)*+_{_{\circlearrowright}}="12";
(-5.5,11.5)*+_{_{\circlearrowright}}="14";
(-0.5,-4.3)*+_{_{\circlearrowright}}="14";
{\ar^{} "0";"2"};
{\ar^{} "2";"4"};
{\ar_{} "0";"6"};
{\ar^{} "2";"6"};
{\ar^{} "6";"8"};
{\ar_{} "6";"10"};
{\ar^{} "4";"8"};
{\ar^{} "8";"10"};
\endxy
\]
in which
\begin{eqnarray*}
&U\to L\to U_T\to U[1],\quad U\to C\to T\to U[1],&\\
&L\to C\to V_T[1]\to L[1],\quad U_T\to T\to V_T[1]\to U_T[1]&
\end{eqnarray*}
are distinguished triangles.
Since $U,U_T\in\Ucal$ implies $L\in\Ucal\se\Cm$, this gives a coreflection triangle
\[
\xy
(-24,0)*+{V_T}="0";
(-8,0)*+{L}="2";
(8,0)*+{C}="4";
(28,0)*+{V_T[1]}="6";
(18,-12)*+{T}="8";
(18,-5)*+{_{\circlearrowright}}="10";
{\ar^{} "0";"2"};
{\ar^{} "2";"4"};
{\ar^{} "4";"6"};
{\ar_{} "4";"8"};
{\ar_{} "8";"6"};
\endxy
\]
which yields $\tm(C)\cong L\in\Ucal/\Wcal$, and thus $H(C)=0$.
\end{proof}
\section{Problem setting}
Let $\C\ppr$ be another triangulated category,
and let $\Pcal\ppr=(\STp,(\Ucal\ppr, \Tcal\ppr))$ be a twin cotorsion pair
on $\C\ppr$.
\begin{nota}\label{NotaPrime}
We denote the associated categories as
\begin{eqnarray*}
&\Wcal\ppr=\Ucal\ppr\cap\Tcal\ppr,&\\
&\Cpp=\Wcal\ppr\ast\Vcal\ppr[1],\quad \Cmp=\Scal\ppr[-1]\ast\Wcal\ppr,&\\
&\Hcal\ppr=\Cpp\cap\Cmp,&
\end{eqnarray*}
and functors as
\begin{eqnarray*}
&h\ppr=\tpp\tmp\co\CWp\to\HWp,&\\
&H\ppr=\big(\C\ppr\ov{\quot}{\lra}\CWp\ov{h\ppr}{\lra}\HWp\big),&\\
&\iota\ppr\co\HWp\hr\CWp.&
\end{eqnarray*}
\end{nota}
\begin{dfn}\label{DefHeartEquiv}
Let $\C,\C\ppr$ and $\Pcal,\Pcal\ppr$ be as above, and let $F\co\C\ov{\simeq}{\lra}\C\ppr$ be a triangle equivalence. $\Pcal$ is said to be {\it heart-equivalent to $\Pcal\ppr$ along $F$} if there exists an equivalence of categories $E\co\HW\ov{\simeq}{\lra}\HWp$ which makes the following diagram commutative up to a natural isomorphism.
\begin{equation}\label{DiagFE}
\xy
(-10,7)*+{\C}="0";
(10,7)*+{\C\ppr}="2";
(-10,-7)*+{\HW}="4";
(10,-7)*+{\HWp}="6";
{\ar^{F}_{\simeq} "0";"2"};
{\ar_{H} "0";"4"};
{\ar^{H\ppr} "2";"6"};
{\ar_{E}^{\simeq} "4";"6"};
{\ar@{}|\circlearrowright "0";"6"};
\endxy
\end{equation}
Remark that this notion of a heart-equivalence is an equivalence relation, in an obvious sense.
\end{dfn}
\begin{prop}
Let $\C,\C\ppr,\Pcal,\Pcal\ppr$ and $F$ be as above.
Then the following {\rm (I)} implies {\rm (II)}.
\begin{enumerate}
\item[{\rm (I)}] $\Pcal$ is heart-equivalent to $\Pcal\ppr$ along $F$.
\item[{\rm (II)}] The following conditions are satisfied.
\begin{itemize}
\item[{\rm (i)}] $H\ppr F(\Ucal)=H\ppr F(\Tcal)=0$.
\item[{\rm (ii)}] $HF\iv(\Ucal\ppr)=HF\iv(\Tcal\ppr)=0$.
\end{itemize}
\end{enumerate}
\end{prop}
\begin{proof}
This immediately follows from $H(\Ucal)=H(\Tcal)=0$ and $H\ppr(\Ucal\ppr)=H\ppr(\Tcal\ppr)=0$.
\end{proof}
\smallskip
The following is our problem in this article.
\begin{problem}\label{ProbOne}
Conversely, does {\rm (I)} follow from {\rm (II)} with some extra conditions?
\end{problem}
\smallskip
We are going to reduce this problem to a more manageable one (Problem \ref{ProbTwo}).
First, remark that this is reduced to the case $F=\Id$.
\begin{rem}\label{ClaimOne}
Let $F\co\C\ov{\simeq}{\lra}\C\ppr$ be a triangle equivalence, and let $\Pcal=\STUV$ be a twin cotorsion pair on $\C$.
If we put $F(\Pcal)=((F(\Scal),F(\Tcal)),(F(\Ucal),F(\Vcal)))$, then the following holds.
\begin{enumerate}
\item $F(\Pcal)$ is a twin cotorsion pair on $\C\ppr$.
\item $\Pcal$ is heart-equivalent to $F(\Pcal)$ along $F$.
\end{enumerate}
Thus, replacing $\Pcal$ by $F(\Pcal)$, we may assume $\C=\C\ppr$ and $F=\Id_{\C}$ from the first.
\end{rem}
\medskip
From this we may assume $F=\Id_{\C}$, and $\Pcal\ppr$ is a twin cotorsion pair on $\C$. We will keep using
Notation \ref{NotaPrime} also in this case. For example, $H\ppr$ denotes a functor $H\ppr\co \C\to\HWp$.
\medskip
Second, note that the candidate for $E$ in $(\ref{DiagFE})$ is unique up to natural transformations.
\begin{rem}\label{ClaimTwo}
Let $\Pcal=\STUV$ and $\Pcal\ppr=\STUVp$ be twin cotorsion pairs on $\C$. Assume that the condition
\[ H\ppr(\Wcal)=0 \]
is satisfied. Then the following holds by the commutativity of the following diagrams.
\[
\xy
(-16,13)*+{\Hcal}="-2";
(-4,-3)*+{}="-1";
(0,8)*+{\C}="0";
(0,-13)*+{}="1";
(-16,-8)*+{\HW}="2";
(16,-8)*+{\HWp}="4";
{\ar@{^(->}^{} "-2";"0"};
{\ar_{\quot} "-2";"2"};
{\ar^{H} "0";"2"};
{\ar^{H\ppr} "0";"4"};
{\ar_{E} "2";"4"};
{\ar@{}|\circlearrowright "0";"1"};
{\ar@{}|\circlearrowright "-2";"-1"};
\endxy
\ ,\quad
\xy
(-16,13)*+{\Hcal}="-2";
(-1,-8)*+{}="-1";
(0,8)*+{\C}="0";
(9,-13)*+{}="1";
(-16,-8)*+{\HW}="2";
(0,-8)*+{\CW}="3";
(16,-8)*+{\HWp}="4";
{\ar@{^(->}^{} "-2";"0"};
{\ar_{\quot} "-2";"2"};
{\ar_{\quot} "0";"3"};
{\ar^{H\ppr} "0";"4"};
{\ar_{\iota} "2";"3"};
{\ar_(0.46){\Hbarp} "3";"4"};
{\ar@{}|\circlearrowright "0";"1"};
{\ar@{}|\circlearrowright "-2";"-1"};
\endxy
\].
\begin{enumerate}
\item $H\ppr$ induces a functor $\Hbarp\co\CW\to\HWp$ which makes the following diagram commutative up to a natural isomorphism.
\[
\xy
(-6,6)*+{\C}="0";
(0,-10)*+{}="1";
(-10,-6)*+{\CW}="2";
(10,-6)*+{\HWp}="4";
{\ar_{\quot} "0";"2"};
{\ar^{H\ppr} "0";"4"};
{\ar_(0.46){\Hbarp} "2";"4"};
{\ar@{}|\circlearrowright "0";"1"};
\endxy
\]
\item If a functor $E\co\HW\to\HWp$ makes the diagram
\[
\xy
(0,7)*+{\C}="0";
(0,-10)*+{}="1";
(-11,-6)*+{\HW}="2";
(11,-6)*+{\HWp}="4";
{\ar_{H} "0";"2"};
{\ar^{H\ppr} "0";"4"};
{\ar_{E} "2";"4"};
{\ar@{}|\circlearrowright "0";"1"};
\endxy
\]
commutative up to a natural isomorphism, then there exists a natural isomorphism
\[ E\cong\Hbarp\ci\iota, \]
where $\iota\co\HW\hr\CW$ is the inclusion.
\end{enumerate}
\end{rem}
Thus, under the conditions $H\ppr(\Wcal)=0$ and $H(\Wcal\ppr)=0$ (remark that condition {\rm (II)} in Problem \ref{ProbOne} implies these conditions), we may restrict our attention to the functors
\[ E=\Hbarp\ci\iota\co\HW\to\HWp \quad\text{and}\quad E\ppr=\Hbar\ci\iota\ppr\co\HWp\to\HW, \]
and it is enough to find a condition which induces
\begin{itemize}
\item[{\rm (a)}] $E\ci H\cong H\ppr$, $E\ppr\ci H\ppr\cong H$,
\item[{\rm (b)}] $E\ci E\ppr\cong\Id$, $E\ppr\ci E\cong\Id$.
\end{itemize}
However, {\rm (b)} follows from {\rm (a)}, as follows.
\begin{lem}\label{ClaimThree}
Let $\Pcal$ and $\Pcal\ppr$ be twin cotorsion pairs on $\C$,
and assume that $H\ppr(\Wcal)=0$ and $H(\Wcal\ppr)=0$ are satisfied. Let $\Hbarp\co\CW\to\HWp$ and $\Hbar\co\C/\Wcal\ppr\to\HW$ be the unique functors induced from $H\ppr$ and $H$ respectively, as in Remark \ref{ClaimTwo}.
Put
\begin{eqnarray*}
&E=\Hbarp\ci\iota\co\HW\to\HWp,&\\
&E\ppr=\Hbar\ci\iota\ppr\co\HWp\to\HW.&
\end{eqnarray*}
If $E$ and $E\ppr$ satisfy {\rm (a)}, then they also satisfy {\rm (b)}.
\end{lem}
\begin{proof}
$E\ppr\ci E\cong\Id$ follows from the commutativity of the following diagram up to natural isomorphisms.
\[
\xy
(-28,26)*+{\Hcal}="0";
(-28,-24)*+{\HW}="2";
(-10,10)*+{\C}="4";
(-10,-14)*+{\C/\Wcal}="6";
(10,-24)*+{\HWp}="10";
(30,10)*+{\HW}="12";
(-19,0)*+{_{\circlearrowright}}="21";
(-7.5,15)*+{_{\circlearrowright}}="22";
(-4,-10)*+{_{\circlearrowright}}="23";
(-9,-20)*+{_{\circlearrowright}}="25";
(10,-2.5)*+{_{\circlearrowright}}="28";
{\ar_{\quot} "0";"2"};
{\ar@{_(->} "0";"4"};
{\ar^{\quot} "0";"12"};
{\ar^{\iota} "2";"6"};
{\ar_{E} "2";"10"};
{\ar_{\quot} "4";"6"};
{\ar^{H\ppr} "4";"10"};
{\ar^{H} "4";"12"};
{\ar_{\Hbarp} "6";"10"};
{\ar_{E\ppr} "10";"12"};
\endxy
\]
Similarly for $E\ci E\ppr\cong\Id$.
\end{proof}
By Remarks \ref{ClaimOne}, \ref{ClaimTwo} and Lemma \ref{ClaimThree}, Problem \ref{ProbOne} has been reduced to the following.
\begin{problem}\label{ProbTwo}
Let $\Pcal=\STUV$ and $\Pcal\ppr=\STUVp$ be twin cotorsion pairs on $\C$. Suppose that the condition
\[ H\ppr(\Ucal)=H\ppr(\Tcal)=0 \]
is satisfied. Let $\Hbarp\co\CW\to\HWp$ be the functor induced from $H\ppr$, and put $E=\Hbarp\ci\iota$. With some extra condition, does there exist a natural isomorphism $E\ci H\cong H\ppr$?
\end{problem}
\section{Degenerated case}
Consider the case each of $\Pcal$ and $\Pcal\ppr$ is degenerated to a single cotorsion pair. In this case, it requires no extra condition. This is essentially due to \cite{ZZ} (see also \cite{L2}).
\begin{rem}
If $\Pcal$ is degenerated, then its heart $\HW$ becomes an abelian category (\cite{N1}), and the functor $H\co\C\to\HW$ becomes cohomological (\cite{AN}).
\end{rem}
\begin{prop}\label{PropSingle}
Let $\UV$ and $\UVp$ be cotorsion pairs on $\C$. Suppose that the condition
\[ H\ppr(\Ucal)=H\ppr(\Vcal)=0 \]
is satisfied. Let $\Hbarp\co\CW\to\HWp$ be the functor induced from $H$, and put $E=\Hbarp\ci\iota$. Then, $E\ci H\cong H\ppr$ holds.
\end{prop}
\begin{proof}
For each object $A\in\C$, choose a reflection triangle $(\ref{ReflTriA})$, and a coreflection triangle
\[
\xy
(-24,0)*+{V_A}="0";
(-8,0)*+{K_A}="2";
(8,0)*+{A}="4";
(28,0)*+{V_A[1]}="6";
(18,-12)*+{T_A}="8";
(18,-5)*+{_{\circlearrowright}}="10";
{\ar^{} "0";"2"};
{\ar^{k_A} "2";"4"};
{\ar^{} "4";"6"};
{\ar_{} "4";"8"};
{\ar_{} "8";"6"};
\endxy.
\]
By Remark \ref{RemFunctRefl}, for any morphism $f\in\C(A,B)$, we obtain a morphism of diagrams
\[
\xy
(26,7)*+{V_A[1]}="0";
(7,7)*+{A}="2";
(-8,7)*+{K_A}="4";
(-22,7)*+{V_A}="6";
(16,6)*+{}="7";
(16,16)*+{T_A}="8";
(26,-7)*+{V_B[1]}="10";
(7,-7)*+{B}="12";
(-8,-7)*+{K_B}="14";
(-22,-7)*+{V_B}="16";
(16,-6)*+{}="17";
(16,-16)*+{T_B}="18";
{\ar^{} "2";"0"};
{\ar^{k_A} "4";"2"};
{\ar^{} "6";"4"};
{\ar^{} "8";"0"};
{\ar^{} "2";"8"};
{\ar^{} "0";"10"};
{\ar_{f} "2";"12"};
{\ar_{g} "4";"14"};
{\ar^{} "6";"16"};
{\ar_{}"12";"10"};
{\ar_{k_B} "14";"12"};
{\ar_{} "16";"14"};
{\ar_{} "18";"10"};
{\ar_{} "12";"18"};
{\ar@{}|\circlearrowright "2";"14"};
{\ar@{}|\circlearrowright "4";"16"};
{\ar@{}|\circlearrowright "7";"8"};
{\ar@{}|\circlearrowright "17";"18"};
{\ar@{}|\circlearrowright "0";"12"};
\endxy
\]
and
\[
\xy
(-28,7)*+{S_{K_A}[-1]}="0";
(-7,7)*+{K_A}="2";
(8,7)*+{Z_{K_A}}="4";
(22,7)*+{S_{K_A}}="6";
(-17,6)*+{}="7";
(-17,16)*+{U_{K_A}}="8";
(-28,-7)*+{S_{K_B}[-1]}="10";
(-7,-7)*+{K_B}="12";
(8,-7)*+{Z_{K_B}}="14";
(22,-7)*+{S_{K_B}}="16";
(-17,-6)*+{}="17";
(-17,-16)*+{U_{K_B}}="18";
{\ar^{} "0";"2"};
{\ar^{z_{K_A}} "2";"4"};
{\ar^{} "4";"6"};
{\ar^{} "0";"8"};
{\ar^{} "8";"2"};
{\ar^{} "0";"10"};
{\ar^{g} "2";"12"};
{\ar^{h} "4";"14"};
{\ar^{} "6";"16"};
{\ar_{} "10";"12"};
{\ar_{z_{K_B}} "12";"14"};
{\ar_{} "14";"16"};
{\ar^{} "10";"18"};
{\ar^{} "18";"12"};
{\ar@{}|\circlearrowright "2";"14"};
{\ar@{}|\circlearrowright "4";"16"};
{\ar@{}|\circlearrowright "7";"8"};
{\ar@{}|\circlearrowright "17";"18"};
{\ar@{}|\circlearrowright "0";"12"};
\endxy,
\]
which imply $H(f)=\tp\tm(\und{f})=\tp(\und{g})=\und{h}$.
Since $H\ppr$ is cohomological and $H\ppr(\Ucal)=H\ppr(\Tcal)=0$ by assumption, we obtain morphisms of exact sequences
\[
\xy
(-27,7)*+{0}="0";
(-11,7)*+{H\ppr(K_A)}="2";
(11,7)*+{H\ppr(A)}="4";
(31,7)*+{H\ppr(V_A[1])}="6";
(50,7)*+{\exact};
(-27,-7)*+{0}="10";
(-11,-7)*+{H\ppr(K_B)}="12";
(11,-7)*+{H\ppr(B)}="14";
(31,-7)*+{H\ppr(V_B[1])}="16";
(50,-7)*+{\exact};
{\ar^{} "0";"2"};
{\ar^(0.54){H\ppr(k_A)} "2";"4"};
{\ar^(0.44){0} "4";"6"};
{\ar_{H\ppr(g)} "2";"12"};
{\ar_{H\ppr(f)} "4";"14"};
{\ar^{} "6";"16"};
{\ar_{} "10";"12"};
{\ar_(0.54){H\ppr(k_B)} "12";"14"};
{\ar_(0.44){0} "14";"16"};
{\ar@{}|\circlearrowright "2";"14"};
{\ar@{}|\circlearrowright "4";"16"};
\endxy
\]
and
\[
\xy
(-35,7)*+{H\ppr(S_{K_A}[-1])}="0";
(-11,7)*+{H\ppr(K_A)}="2";
(14,7)*+{H\ppr(Z_{K_A})}="4";
(31,7)*+{0}="6";
(42,7)*+{\exact};
(-35,-7)*+{H\ppr(S_{K_B}[-1])}="10";
(-11,-7)*+{H\ppr(K_B)}="12";
(14,-7)*+{H\ppr(Z_{K_B})}="14";
(31,-7)*+{0}="16";
(42,-7)*+{\exact};
{\ar^(0.56){0} "0";"2"};
{\ar^{H\ppr(z_{K_A})} "2";"4"};
{\ar^{} "4";"6"};
{\ar_{} "0";"10"};
{\ar^{H\ppr(g)} "2";"12"};
{\ar^{H\ppr(h)} "4";"14"};
{\ar_(0.56){0} "10";"12"};
{\ar_{H\ppr(z_{K_B})} "12";"14"};
{\ar_{} "14";"16"};
{\ar@{}|\circlearrowright "0";"12"};
{\ar@{}|\circlearrowright "2";"14"};
\endxy.
\]
These yield a commutative diagram
\[
\xy
(-37,7)*+{H\ppr(A)}="2";
(-11,7)*+{H\ppr(K_A)}="4";
(15,7)*+{H\ppr(Z_{K_A})}="6";
(30.5,7)*+{=EH(A)};
(-37,-7)*+{H\ppr(B)}="12";
(-11,-7)*+{H\ppr(K_B)}="14";
(15,-7)*+{H\ppr(Z_{K_B})}="16";
(30.5,-7)*+{=EH(B)};
{\ar_{H\ppr(k_A)}^{\cong} "4";"2"};
{\ar^{H\ppr(z_{K_A})}_{\cong} "4";"6"};
{\ar_{H\ppr(f)} "2";"12"};
{\ar^{H\ppr(g)} "4";"14"};
{\ar^{H\ppr(h)=EH(f)} "6";"16"};
{\ar^{H\ppr(k_B)}_{\cong} "14";"12"};
{\ar_{H\ppr(z_{K_B})}^{\cong} "14";"16"};
{\ar@{}|\circlearrowright "2";"14"};
{\ar@{}|\circlearrowright "4";"16"};
\endxy.
\]
Thus if we put
\[ \eta_A=H\ppr(z_{K_A})\ci H\ppr(k_A)\iv, \]
then $\eta=\{\eta_A\}_{A\in\C}$ gives a natural isomorphism $\eta\co H\ppr\ov{\cong}{\Longrightarrow}E\ci H$.
\end{proof}
In terms of the original problem (Problem \ref{ProbOne}), we obtain the following.
\begin{cor}
Let $\Pcal=\UV$ and $\Pcal\ppr=\UVp$ be cotorsion pairs on $\C$ and $\C\ppr$, respectively.
Let $F\co\C\ov{\simeq}{\lra}\C\ppr$ be a triangle equivalence. Then, the following are equivalent.
\begin{enumerate}
\item $\Pcal$ is heart-equivalent to $\Pcal\ppr$ along $F$.
\item $H\ppr F(\Ucal)=H\ppr F(\Vcal)=0$ and $HF\iv(\Ucal\ppr)=HF\iv(\Vcal\ppr)=0$ are satisfied.
\end{enumerate}
\end{cor}
\begin{proof}
This immediately follows from Proposition \ref{PropSingle}.
\end{proof}
\section{A sufficient condition in general case}
We return to the general case of twin cotorsion pairs. Let $\Pcal=\STUV$ be a twin cotorsion pair on $\C$. We start with some improvements of the results from \cite{N2}.
The following gives a partial converse to Lemma 5.1 in \cite{N2}.
\begin{prop}\label{PropCokOne}
Let $S[-1]\ov{e}{\lra}A\ov{f}{\lra}B\ov{g}{\lra}S$ be a distinguished triangle with $S\in\Scal$. Then,
\[ H(S[-1])\ov{H(e)}{\lra}H(A)\ov{H(f)}{\lra}H(B)\to 0 \]
is a cokernel sequence in $\HW$.
\end{prop}
\begin{proof}
By the adjoint property of $\tp$, it is enough to show that the sequence
\begin{equation}\label{SeqE}
0\to(\CW)(\tm(B),Y)\ov{-\ci\tm(\und{f})}{\lra}(\CW)(\tm(A),Y)\ov{-\ci\tm(\und{e})}{\lra}(\CW)(\tm(S[-1]),Y)
\end{equation}
is exact for any $Y\in\Cp$.
Take a coreflection triangle
\[
\xy
(-24,0)*+{V_B}="0";
(-8,0)*+{K_B}="2";
(8,0)*+{B}="4";
(28,0)*+{V_B[1]}="6";
(18,-12)*+{T_B}="8";
(18,-5)*+{_{\circlearrowright}}="10";
{\ar^{} "0";"2"};
{\ar^{k_B} "2";"4"};
{\ar^{v_B} "4";"6"};
{\ar_{} "4";"8"};
{\ar_{} "8";"6"};
\endxy
\]
and complete $v_B\ci f\co B\to V_B[1]$ into a distinguished triangle
\[ V_B\to L\ov{\ell}{\lra}A\ov{v_B\ci f}{\lra}V_B[1]. \]
By the octahedron axiom, we obtain a diagram
\[
\xy
(-20,16)*+{S[-1]}="0";
(0.5,15)*+{L}="2";
(17,14)*+{K_B}="4";
(-2,4)*+{A}="6";
(6.1,-0.6)*+{B}="8";
(-5.8,-14.2)*+{V_B[1]}="10";
(7.5,8.5)*+_{_{\circlearrowright}}="12";
(-5.5,11.5)*+_{_{\circlearrowright}}="14";
(-0.5,-4.3)*+_{_{\circlearrowright}}="14";
{\ar^{s} "0";"2"};
{\ar^{h} "2";"4"};
{\ar_{e} "0";"6"};
{\ar^{\ell} "2";"6"};
{\ar^{f} "6";"8"};
{\ar_{} "6";"10"};
{\ar^{k_B} "4";"8"};
{\ar^{v_B} "8";"10"};
\endxy
\]
in which
\begin{eqnarray*}
&S[-1]\to L\to K_B\to S,\quad S[-1]\to A\to B\to S,&\\
&L\to A\to V_B[1]\to L[1],\quad K_B\to B\to V_B[1]\to K_B[1]&
\end{eqnarray*}
are distinguished triangles.
Then by Fact \ref{FactUSTV}, $K_B\in\Cm$ implies $L\in\Cm$, and thus
\[
\xy
(-24,0)*+{V_B}="0";
(-8,0)*+{L}="2";
(8,0)*+{A}="4";
(28,0)*+{V_B[1]}="6";
(18,-12)*+{T_B}="8";
(18,-5)*+{_{\circlearrowright}}="10";
{\ar^{} "0";"2"};
{\ar^{\ell} "2";"4"};
{\ar^{v_B\ci f} "4";"6"};
{\ar_{} "4";"8"};
{\ar_{} "8";"6"};
\endxy
\]
becomes a coreflection triangle. Thus we may assume $\tm(\und{f})=\und{h}$. Since $S[-1]\in\Scal [-1]\se\Cm$, we may also assume $\tm(\und{e})=\und{s}$.
\medskip
{\rm (1)} Let $X\in\C$ be any object, and let $x\in\C(L,X)$ be any morphism.
If $x$ satisfies $\und{x}\ci\und{s}=0$, then $x\ci s$ factors through some $W\in\Wcal$ as follows.
\[
\xy
(-8,0)*+{S[-1]}="2";
(8,0)*+{L}="4";
(24,0)*+{K_B}="6";
(-8,-12)*+{W}="8";
(8,-12)*+{X}="10";
{\ar^(0.54){s} "2";"4"};
{\ar^(0.46){h} "4";"6"};
{\ar_{} "2";"8"};
{\ar^{x} "4";"10"};
{\ar_{} "8";"10"};
{\ar@{}|\circlearrowright "2";"10"};
\endxy
\]
Since $\C(S[-1],W)=0$ it follows $x\ci s=0$, and thus $x$ factors through $h$.
\smallskip
{\rm (2)} Let $Y\in\Cp$ be any object, and let $y\in\C(K_B,Y)$ be any morphism. Decompose $Y$ into a distinguished triangle
\[ V_Y\to W_Y\ov{w_Y}{\lra}Y\ov{v_Y}{\lra}V_Y[1] \]
satisfying $V_Y\in\Vcal, W_Y\in\Wcal$.
If $y$ satisfies $\und{y}\ci\und{h}=0$, then $y\ci h$ factors through $w_Y$ as follows.
\[
\xy
(-8,0)*+{L}="2";
(8,0)*+{K_B}="4";
(22,0)*+{S}="6";
(-8,-12)*+{W_Y}="8";
(8,-12)*+{Y}="10";
(25,-12)*+{V_Y[1]}="12";
{\ar^{h} "2";"4"};
{\ar^{} "4";"6"};
{\ar_{} "2";"8"};
{\ar^{y} "4";"10"};
{\ar_{w_Y} "8";"10"};
{\ar_(0.46){v_Y} "10";"12"};
{\ar@{}|\circlearrowright "2";"10"};
\endxy
\]
Thus $v_Y\ci y$ should factor through some morphism $S\to V_Y[1]$. However, since $\C(S,V_Y[1])=0$, this implies $v_Y\ci y=0$. Thus $y$ factors through $w_Y$, which means $\und{y}=0$.
Exactness of $(\ref{SeqE})$ follows from {\rm (1)} and {\rm (2)}.
\end{proof}
\begin{prop}\label{PropCokThree}
Let $C[-1]\to A\ov{f}{\lra}B\ov{g}{\lra}C$ be a distinguished triangle. Suppose $C$ admits a coreflection triangle
\[
\xy
(-24,0)*+{V}="0";
(-8,0)*+{S}="2";
(8,0)*+{C}="4";
(28,0)*+{V[1]}="6";
(18,-12)*+{T}="8";
(18,-5)*+{_{\circlearrowright}}="10";
{\ar^{} "0";"2"};
{\ar^{} "2";"4"};
{\ar^{} "4";"6"};
{\ar_{} "4";"8"};
{\ar_{} "8";"6"};
\endxy
\]
satisfying $S\in\Scal$. Then, $H(f)$ is epimorphic in $\HW$.
\end{prop}
\begin{proof}
Take a coreflection triangle
\[
\xy
(-24,0)*+{V_B}="0";
(-8,0)*+{K_B}="2";
(8,0)*+{B}="4";
(28,0)*+{V_B[1]}="6";
(18,-12)*+{T_B}="8";
(18,-5)*+{_{\circlearrowright}}="10";
{\ar^{} "0";"2"};
{\ar^{k_B} "2";"4"};
{\ar^{} "4";"6"};
{\ar_{} "4";"8"};
{\ar_{} "8";"6"};
\endxy.
\]
Remark that $H(k_B)$ is isomorphic.
Then, we obtain the following commutative diagram by (the dual of) Remark \ref{RemFunctRefl}.
\[
\xy
(0,14)*+{K_B}="2";
(18,14)*+{S}="4";
(-18,0)*+{A}="10";
(0,0)*+{B}="12";
(18,0)*+{C}="14";
(15,-7)*+{}="17";
(28,-7)*+{T}="18";
(0,-14)*+{V_B[1]}="22";
(18,-14)*+{V[1]}="24";
{\ar^{h} "2";"4"};
{\ar_{k_B} "2";"12"};
{\ar^{} "4";"14"};
{\ar^{f} "10";"12"};
{\ar^{g} "12";"14"};
{\ar^{} "14";"18"};
{\ar^{} "18";"24"};
{\ar^{} "12";"22"};
{\ar^{} "14";"24"};
{\ar^{} "22";"24"};
{\ar@{}|\circlearrowright "2";"14"};
{\ar@{}|\circlearrowright "12";"24"};
{\ar@{}|\circlearrowright "17";"18"};
\endxy
\]
If we complete $h$ into a distinguished triangle
\[ S[-1]\to L\ov{\ell}{\lra}K_B\ov{h}{\lra}S, \]
then we have a commutative diagram
\[
\xy
(-22,6)*+{S[-1]}="0";
(-7,6)*+{L}="2";
(6,6)*+{K_B}="4";
(20,6)*+{S}="6";
(-22,-6)*+{C[-1]}="10";
(-7,-6)*+{A}="12";
(6,-6)*+{B}="14";
(20,-6)*+{C}="16";
{\ar^{} "0";"2"};
{\ar^{\ell} "2";"4"};
{\ar^{h} "4";"6"};
{\ar^{} "0";"10"};
{\ar^{a} "2";"12"};
{\ar^{k_B} "4";"14"};
{\ar^{} "6";"16"};
{\ar_{} "10";"12"};
{\ar_{f} "12";"14"};
{\ar_{} "14";"16"};
{\ar@{}|\circlearrowright "0";"12"};
{\ar@{}|\circlearrowright "2";"14"};
{\ar@{}|\circlearrowright "4";"16"};
\endxy.
\]
By Proposition \ref{PropCokOne}, $H(\ell)$ is epimorphic. Thus $H(f)\ci H(a)=H(k_B)\ci H(\ell)$ also becomes epimorphic. This implies $H(f)$ is epimorphic.
\end{proof}
\begin{prop}\label{RemCokThree}
For any object $C\in\C$, the following are equivalent.
\begin{enumerate}
\item $H(C)=0$ and $C\in\Scal\ast\Vcal [1]$.
\item $\tm(C)\in\Ucal/\Wcal$ and $C\in\Scal\ast\Vcal [1]$.
\item $C$ admits a coreflection triangle
\[
\xy
(-24,0)*+{V}="0";
(-8,0)*+{S}="2";
(8,0)*+{C}="4";
(28,0)*+{V[1]}="6";
(18,-12)*+{T}="8";
(18,-5)*+{_{\circlearrowright}}="10";
{\ar^{} "0";"2"};
{\ar^{} "2";"4"};
{\ar^{} "4";"6"};
{\ar_{} "4";"8"};
{\ar_{} "8";"6"};
\endxy
\]
satisfying $S\in\Scal$.
\end{enumerate}
\end{prop}
\begin{proof}
The equivalence $(1)\LR(2)$ follows from Proposition \ref{RemHTauU}. Suppose {\rm (3)} holds. This implies $\tm(C)\cong S\in\Ucal/\Wcal$, and also $C\in\Scal\ast\Vcal [1]$.
Conversely, suppose {\rm (2)} holds. By $\tm(C)\in\Ucal/\Wcal$, this $C$ admits a coreflection triangle
\[
\xy
(-24,0)*+{V}="0";
(-8,0)*+{U}="2";
(8,0)*+{C}="4";
(28,0)*+{V[1]}="6";
(18,-12)*+{T}="8";
(18,-5)*+{_{\circlearrowright}}="10";
{\ar^{} "0";"2"};
{\ar^{} "2";"4"};
{\ar^{} "4";"6"};
{\ar_{} "4";"8"};
{\ar_{} "8";"6"};
\endxy
\]
satisfying $U\in\Ucal$. By $C\in\Scal\ast\Vcal [1]$, it can be also decomposed into a distinguished triangle
\[ V_0\to S\to C\to V_0[1] \]
satisfying $S\in\Scal, V_0\in\Vcal$. Then by $\C(U,V_0[1])=0$, we obtain the following commutative diagram.
\[
\xy
(25,7)*+{V[1]}="0";
(7,7)*+{C}="2";
(-8,7)*+{U}="4";
(-23,7)*+{V}="6";
(15,6)*+{}="7";
(15,16)*+{T}="8";
(25,-6)*+{V_0[1]}="10";
(7,-6)*+{C}="12";
(-8,-6)*+{S}="14";
(-23,-6)*+{V_0}="16";
{\ar^{} "2";"0"};
{\ar^{} "4";"2"};
{\ar^{} "6";"4"};
{\ar^{} "8";"0"};
{\ar^{} "2";"8"};
{\ar^{} "0";"10"};
{\ar@{=} "2";"12"};
{\ar_{} "4";"14"};
{\ar^{} "6";"16"};
{\ar_{} "12";"10"};
{\ar_{} "14";"12"};
{\ar_{} "16";"14"};
{\ar@{}|\circlearrowright "0";"12"};
{\ar@{}|\circlearrowright "2";"14"};
{\ar@{}|\circlearrowright "4";"16"};
{\ar@{}|\circlearrowright "7";"8"};
\endxy
\]
This gives the following coreflection triangle, and thus {\rm (3)} holds.
\[
\xy
(-24,0)*+{V_0}="0";
(-8,0)*+{S}="2";
(8,0)*+{C}="4";
(28,0)*+{V_0[1]}="6";
(18,-12)*+{T}="8";
(18,-5)*+{_{\circlearrowright}}="10";
{\ar^{} "0";"2"};
{\ar^{} "2";"4"};
{\ar^{} "4";"6"};
{\ar_{} "4";"8"};
{\ar_{} "8";"6"};
\endxy
\]
\end{proof}
\begin{cor}\label{CorCok}
Let $C[-1]\ov{e}{\lra}A\ov{f}{\lra}B\ov{g}{\lra}C$ be a distinguished triangle. Suppose $C$ satisfies $H(C)=0$ and $C\in\Scal\ast\Vcal [1]$.
Then,
\[ H(C[-1])\ov{H(e)}{\lra}H(A)\ov{H(f)}{\lra}H(B)\to 0 \]
becomes a cokernel sequence in $\HW$.
\end{cor}
\begin{proof}
By Proposition \ref{PropCokThree}, the morphism $H(f)$ is epimorphic.
Take a coreflection triangle
\[
\xy
(-24,0)*+{V}="0";
(-8,0)*+{S}="2";
(8,0)*+{C}="4";
(28,0)*+{V[1]}="6";
(18,-12)*+{T}="8";
(18,-5)*+{_{\circlearrowright}}="10";
{\ar^{} "0";"2"};
{\ar^{k} "2";"4"};
{\ar^{} "4";"6"};
{\ar_{} "4";"8"};
{\ar_{} "8";"6"};
\endxy
\]
satisfying $S\in\Scal$. If we complete $s=-e\ci k[-1]$ into a distinguished triangle
\[ S[-1]\ov{s}{\lra}A\ov{d}{\lra}D\to S, \]
then by the octahedron axiom, we obtain a commutative diagram
\[
\xy
(-16,16)*+{S[-1]}="0";
(-15,-1)*+{C[-1]}="2";
(-14,-19)*+{V}="4";
(-4,2)*+{A}="6";
(2.3,-5.3)*+{D}="8";
(17,8.5)*+{B}="10";
(-8.5,-7.5)*+_{_{\circlearrowright}}="12";
(-11.5,5.5)*+_{_{\circlearrowright}}="14";
(4.3,0.5)*+_{_{\circlearrowright}}="14";
{\ar_{-k[-1]} "0";"2"};
{\ar_{} "2";"4"};
{\ar^{s} "0";"6"};
{\ar_(0.56){e} "2";"6"};
{\ar_{d} "6";"8"};
{\ar^{f} "6";"10"};
{\ar_{} "4";"8"};
{\ar_{h} "8";"10"};
\endxy
\]
in which
\begin{eqnarray*}
&S[-1]\to A\to D\to S,\quad S[-1]\to C[-1]\to V\to S,&\\
&C[-1]\to A\to B\to C,\quad V\to D\to B\to V[1]&
\end{eqnarray*}
are distinguished triangles.
Applying Proposition \ref{PropCokOne} and its dual to the triangles
\begin{eqnarray*}
&S[-1]\ov{s}{\lra}A\ov{d}{\lra}D\to S,&\\
&S[-1]\ov{-k[-1]}{\lra}C[-1]\to V\to S,&\\
&V\to D\ov{h}{\lra}B\to V[1],&
\end{eqnarray*}
we obtain exact sequences
\begin{eqnarray*}
&H(S[-1])\ov{H(s)}{\lra}H(A)\ov{H(d)}{\lra}H(D)\to 0\quad \exact,&\\
&H(S[-1])\ov{-H(k[-1])}{\lra}H(C[-1])\to 0\quad\exact,&\\
&0\to H(D)\ov{H(h)}{\lra}H(B)\quad\exact.&
\end{eqnarray*}
These form a commutative diagram
\[
\xy
(-24,14)*+{H(S[-1])}="0";
(-24,0)*+{H(C[-1])}="2";
(-24,-12)*+{0}="4";
(0,-13)*+{0}="5";
(0,0)*+{H(A)}="6";
(12.5,-8)*+{H(D)}="8";
(29,0)*+{H(B)}="10";
(23.5,-14)*+{0}="12";
{\ar_{-H(k[-1])} "0";"2"};
{\ar_{} "2";"4"};
{\ar^{H(s)} "0";"6"};
{\ar_(0.56){H(e)} "2";"6"};
{\ar_{H(d)} "6";"8"};
{\ar^{H(f)} "6";"10"};
{\ar_{} "5";"8"};
{\ar_{H(h)} "8";"10"};
{\ar_{} "8";"12"};
(-16,4)*+_{_{\circlearrowright}}="24";
(13,-3)*+_{_{\circlearrowright}}="26";
\endxy,
\]
which implies the exactness of
\[ H(C[-1])\ov{H(e)}{\lra}H(A)\ov{H(f)}{\lra}H(B). \]
Together with the epimorphicity of $H(f)$, we obtain the conclusion.
\end{proof}
Corollary \ref{CorCok} gives a sufficient condition for Problem \ref{ProbTwo}.
\begin{cor}\label{CorProT}
Let $\Pcal=\STUV$ and $\Pcal\ppr=\STUVp$ be twin cotorsion pairs on $\C$. If they satisfy the conditions
\[ H\ppr(\Ucal)=H\ppr(\Tcal)=0 \]
and
\[ \Scal\se\Scal\ppr\ast\Vcal\ppr [1],\quad \Vcal\se\Scal\ppr [-1]\ast\Vcal\ppr, \]
then $E=\Hbarp\ci\iota\co\HW\to\HWp$ satisfies $E\ci H\cong H\ppr$.
\end{cor}
\begin{proof}
For each object $A\in\C$, choose a reflection triangle
\begin{equation}\label{ReflTriALas}
\xy
(-28,0)*+{S_A[-1]}="0";
(-8,0)*+{A}="2";
(8,0)*+{Z_A}="4";
(24,0)*+{S_A}="6";
(-18,-12)*+{U_A}="8";
(-18,-5)*+{_{\circlearrowright}}="10";
{\ar^{s_A} "0";"2"};
{\ar^{z_A} "2";"4"};
{\ar_{} "4";"6"};
{\ar_{} "0";"8"};
{\ar_{} "8";"2"};
\endxy
\end{equation}
and a coreflection triangle
\[
\xy
(-24,0)*+{V_A}="0";
(-8,0)*+{K_A}="2";
(8,0)*+{A}="4";
(28,0)*+{V_A[1]}="6";
(18,-12)*+{T_A}="8";
(18,-5)*+{_{\circlearrowright}}="10";
{\ar^{} "0";"2"};
{\ar^{k_A} "2";"4"};
{\ar^{} "4";"6"};
{\ar_{} "4";"8"};
{\ar_{} "8";"6"};
\endxy.
\]
Similarly as in the proof of Proposition \ref{PropSingle}, it suffices to show that $H\ppr(z_A)$ and $H\ppr(k_A)$ become isomorphisms in $\HWp$.
By assumption, $S_A$ satisfies $H\ppr(S_A)=0$ and $S_A\in\Scal\ppr\ast\Vcal\ppr [1]$. Thus, applying Corollary \ref{CorCok} to $(\ref{ReflTriALas})$, we obtain a cokernel sequence
\[ H\ppr(S_A[-1])\ov{H\ppr(s_A)}{\lra}H\ppr(A)\ov{H\ppr(z_A)}{\lra}H\ppr(Z_A)\to 0. \]
Since $H\ppr(s_A)$ factors through $H\ppr(U_A)=0$, this means $H\ppr(z_A)$ is an isomorphism.
Dually, we have a kernel sequence
\[ 0\to H\ppr(K_A)\ov{H\ppr(k_A)}{\lra}H\ppr(A)\ov{0}{\lra}H\ppr(V_A[1]), \]
which means $H\ppr(k_A)$ is an isomorphism.
\end{proof}
If we rephrase this in the setting of Problem \ref{ProbOne}, we obtain the following.
\begin{thm}\label{MainThm}
Let $\Pcal=\STUV$ and $\Pcal\ppr=\STUVp$ be twin cotorsion pairs on $\C$ and $\C\ppr$. Let $F\co\C\ov{\simeq}{\lra}\C\ppr$ be a triangle equivalence. If $\Pcal$ and $\Pcal\ppr$ satisfy the conditions
\begin{itemize}
\item[{\rm (A)}] \ $H\ppr F(\Ucal)=H\ppr F(\Tcal)=0$, \ $HF\iv(\Ucal\ppr)=HF\iv(\Tcal\ppr)=0$,
\item[{\rm (B)}] \ $F(\Scal)\se\Scal\ppr\ast\Vcal\ppr [1]$, \ $F(\Vcal)\se\Scal\ppr [-1]\ast\Vcal\ppr$,
\item[{\rm (C)}] \ $F\iv(\Scal\ppr)\se\Scal\ast\Vcal [1]$, \ $F\iv(\Vcal\ppr)\se\Scal [-1]\ast\Vcal$,
\end{itemize}
then $\Pcal$ is heart-equivalent to $\Pcal\ppr$ along $F$.
\end{thm}
\begin{proof}
This follows immediately from Corollary \ref{CorProT}.
\end{proof}
\section{An application}
In the rest, we assume the following.
\begin{assumption}
$\ $
\begin{enumerate}
\item $\C$ is $k$-linear for some field $k$, and has a Serre functor $\Sbf$.
\item $\Dcal\se\C$ is a functorially finite rigid subcategory, closed under isomorphisms, finite direct sums and summands.
\item $\UVp$ is a cotorsion pair satisfying $\Dcal\se\Ucal\ppr$.
\item $\Ucal=\mu\iv(\Ucal\ppr ;\Dcal)=(\Dcal\ast\Ucal\ppr [1])\cap {}\ppp\Dcal [1]$.
\item $\Ucal\ppr=\mu(\Ucal ;\Dcal)=(\Ucal [-1]\ast\Dcal)\cap \Dcal [-1]\ppp$.
\item $({}\ppp\Ucal [1],\Ucal)$ is a cotorsion pair.
\end{enumerate}
And, additionally,
\begin{enumerate}
\setcounter{enumi}{6}
\item $\Ucal\se\Dcal [-1]\ppp$.
\item $\Ucal\ppr\se {}\ppp\Dcal [1]$.
\end{enumerate}
Remark that {\rm (7)} and {\rm (8)} are automatically satisfied if $\C$ is 2-Calabi-Yau.
\end{assumption}
\begin{prop}\label{PropPropLast}
Put $F=\Sbf\ci [-2]\co\C\ov{\simeq}{\lra}\C$. With the above assumption, the twin cotorsion pair $(({}\ppp\Ucal [1],\Ucal),({}\ppp\Dcal [1],\Dcal))$ is heart-equivalent to $((\Dcal,\Dcal [-1]\ppp),(\Ucal\ppr,\Vcal\ppr))$ along $F$.
\end{prop}
This recovers the following result in \cite{MP}.
\begin{fact}
(Theorem 2.9 in \cite{MP})
Assume $\C$ is a Krull-Schmidt $k$-linear $\Hom$-finite triangulated category with a Serre functor. Let $T$ be a basic rigid object, and let
\[ T=\ovl{T}\oplus R,\qquad T\ppr=\ovl{T}\oplus R^{\ast} \]
be direct sums in $\C$, where $R$ and $R^{\ast}$ are related by a distinguished triangle
\[ R^{\ast}\to B\to R\to R^{\ast}[1] \]
in which $B\in\add\ovl{T}$, and $B\to R$ is a minimal right $\add\ovl{T}$-approximation. Then there is an equivalence
\[ \big((\add T)\ast(\add\ovl{T}[1])\big)/\add T\, \simeq\, \big((\add \ovl{T}[-1])\ast(\add T\ppr)\big)/\add T\ppr. \]
Indeed, this equivalence follows from Proposition \ref{PropPropLast}, if we put $\Dcal=\add\overline{T},\, \Ucal=\add T$ and $\Ucal\ppr=\add T\ppr$.
\end{fact}
\begin{proof} ({\it of Proposition \ref{PropPropLast}.})
We confirm the conditions {\rm (A),(B),(C)} in Theorem \ref{MainThm}.
Obviously, we have
\begin{eqnarray*}
&F(\Ucal)\se F({}\ppp\Dcal [1])=\Sbf({}\ppp\Dcal [-1])=\Dcal [-1]\ppp,&\\
&F\iv(\Ucal\ppr)\se F\iv(\Dcal [-1]\ppp)=\Sbf\iv(\Dcal [1]\ppp)={}\ppp\Dcal [1],&
\end{eqnarray*}
and thus {\rm (A)} is satisfied.
It remains to show the following.
\begin{itemize}
\item[{\rm (B1)}] \ $F({}\ppp\Ucal [1])\se\Dcal\ast\Vcal\ppr [1]$.
\item[{\rm (B2)}] \ $F(\Dcal)\se\Dcal [-1]\ast\Vcal\ppr$.
\item[{\rm (C1)}] \ $F\iv(\Dcal)\se {}\ppp\Ucal [1]\ast\Dcal [1]$.
\item[{\rm (C2)}] \ $F\iv(\Vcal\ppr)\se {}\ppp\Ucal\ast\Dcal$.
\end{itemize}
\medskip
\noindent {\bf [Confirmation of {\rm (B1)}]} This requires condition {\rm (8)}. Since $F({}\ppp\Ucal [1])=\Ucal [-1]\ppp$, it is equivalent to show $\Ucal [-1]\ppp\se\Dcal\ast\Vcal\ppr [1]$.
For any $X\in\Ucal [-1]\ppp$, decompose it into a distinguished triangle
\[ V_X\ppr\to U_X\ppr\to X\to V_X\ppr [1]\quad(U_X\ppr\in\Ucal\ppr,V_X\ppr\in\Vcal\ppr) \]
and then, $U_X\ppr$ into
\[ U[-1]\to U_X\ppr\to D\to U\quad(U\in\Ucal,D\in\Dcal). \]
By $X\in\Ucal [-1]\ppp$, we obtain a diagram
\[
\xy
(-16.2,16)*+{U_X\ppr}="0";
(-14.8,-1)*+{D}="2";
(-13.5,-17.5)*+{U}="4";
(-4,2)*+{X}="6";
(2,-5.5)*+{V_X\ppr [1]}="8";
(16.7,8.5)*+{P}="10";
(-8.5,-7.5)*+_{_{\circlearrowright}}="12";
(-11.5,5.5)*+_{_{\circlearrowright}}="14";
(4.3,0.5)*+_{_{\circlearrowright}}="14";
{\ar_{} "0";"2"};
{\ar_{} "2";"4"};
{\ar^{} "0";"6"};
{\ar_{} "2";"6"};
{\ar_{} "6";"8"};
{\ar^{} "6";"10"};
{\ar_{} "4";"8"};
{\ar_{} "8";"10"};
\endxy
\]
in which
\begin{eqnarray*}
&U\ppr_X\to D\to U\to U\ppr_X[1],\quad U\ppr_X\to X\to V\ppr_X[1]\to U\ppr_X[1],&\\
&U\to V\ppr_X[1]\to P\to U[1],\quad D\to X\to P\to D[1]&
\end{eqnarray*}
are distinguished triangles.
Then for any $U\ppr\in\Ucal\ppr$ and any $u\ppr\in\C(U\ppr,P)$, the diagram
\[
\xy
(8,14)*+{U\ppr}="0";
(-8,0)*+{V_X\ppr [1]}="2";
(8,0)*+{P}="4";
(28,0)*+{U[1]}="6";
(18,-12)*+{D[1]}="8";
(18,-5)*+{_{\circlearrowright}}="10";
{\ar^{u\ppr} "0";"4"};
{\ar^{} "2";"4"};
{\ar^{} "4";"6"};
{\ar_{} "4";"8"};
{\ar_{} "8";"6"};
\endxy
\]
and the condition $\Ucal\ppr\se {}\ppp\Dcal [1]$ together with $\C(U\ppr,V\ppr_X[1])=0$ show $u\ppr=0$. Namely $P$ belongs to $\Ucal^{\prime\perp}=\Vcal\ppr [1]$, and thus it follows $X\in\Dcal\ast\Vcal\ppr [1]$.
\medskip
\noindent {\bf [Confirmation of {\rm (B2)}]}
We have
\begin{eqnarray*}
\C(\Ucal\ppr [-1],F(\Dcal))&=&\C(\Ucal\ppr [-1],\Sbf(\Dcal [-2])
\ \cong\ \C(\Dcal [-2],\Ucal\ppr [-1])^{\vee}\\
&\cong&\C(\Dcal [-1],\Ucal\ppr)^{\vee}\ =\ 0.
\end{eqnarray*}
This means $F(\Dcal)\se\Vcal\ppr$. Here, ${(-)}^{\vee}$ denotes the $k$-dual.
\medskip
\noindent {\bf [Confirmation of {\rm (C1)}]}
We have
\begin{eqnarray*}
\C(F\iv(\Dcal),\Ucal [1])&\cong&\C(\Dcal [2],\Sbf(\Ucal [1]))
\ \cong\ \C(\Ucal [1],\Dcal [2])^{\vee}\\
&\cong&\C(\Ucal,\Dcal [1])^{\vee}\ =\ 0.
\end{eqnarray*}
This means $F\iv(\Dcal)\se {}\ppp\Ucal [1]$.
\medskip
\noindent {\bf [Confirmation of {\rm (C2)}]}
This requires condition {\rm (7)}. Since $F\iv(\Vcal\ppr)=F\iv(\Ucal\ppr [-1]\ppp)={}\ppp\Ucal\ppr [1]$, it is equivalent to show ${}\ppp\Ucal\ppr [1]\se {}\ppp\Ucal\ast\Dcal$.
For any $X\in\ppp\Ucal\ppr [1]$, decompose it into a distinguished triangle
\[ Q_X\to X\to U_X\to Q_X[1]\quad(Q_X\in {}\ppp\Ucal,U_X\in\Ucal) \]
and then, $U_X$ into
\[ D\to U_X\to U\ppr [1]\to D[1]\quad(D\in \Dcal,U\ppr\in\Ucal\ppr). \]
By $X\in {}\ppp\Ucal\ppr [1]$, we obtain a diagram
\[
\xy
(-20,16)*+{R}="0";
(0.5,15)*+{Q_X}="2";
(17,14)*+{U\ppr}="4";
(-2,4)*+{X}="6";
(6.1,-0.6)*+{D}="8";
(-5.8,-14.2)*+{U_X}="10";
(7.5,8.5)*+_{_{\circlearrowright}}="12";
(-5.5,11.5)*+_{_{\circlearrowright}}="14";
(-0.5,-4.3)*+_{_{\circlearrowright}}="14";
{\ar^{} "0";"2"};
{\ar^{} "2";"4"};
{\ar_{} "0";"6"};
{\ar^{} "2";"6"};
{\ar^{} "6";"8"};
{\ar_{} "6";"10"};
{\ar^{} "4";"8"};
{\ar^{} "8";"10"};
\endxy
\]
in which\begin{eqnarray*}
&R\to Q_X\to U\ppr\to R[1],\quad R\to X\to D\to R[1],&\\
&Q_X\to X\to U_X\to Q_X[1],\quad U\ppr\to D\to U_X\to U\ppr [1]&
\end{eqnarray*}
are distinguished triangles.
Then for any $U\in\Ucal$ and any $u\in\C(R,U)$ the diagram
\[
\xy
(-18,12)*+{D[-1]}="0";
(-28,0)*+{U\ppr [-1]}="2";
(-8,0)*+{R}="4";
(8,0)*+{Q_X}="6";
(-8,-14)*+{U}="8";
(-18,5)*+{_{\circlearrowright}}="10";
{\ar_{} "0";"4"};
{\ar_{} "2";"0"};
{\ar^{} "2";"4"};
{\ar^{} "4";"6"};
{\ar^{u} "4";"8"};
\endxy
\]
and the condition $\Ucal\se\Dcal [-1]\ppp$ shows $u=0$. Namely $R$ belongs to ${}\ppp\Ucal$, and thus we have $X\in {}\ppp\Ucal\ast\Dcal$.
\end{proof}
\section*{Acknowledgement}
This article has been written when the author was staying at LAMFA, l'Universit\'{e} de Picardie-Jules Verne, by the support of JSPS Postdoctoral Fellowships for Research Abroad. He wishes to thank the hospitality of Professor Serge Bouc, Professor Radu Stancu and the members of LAMFA.
|
\section{Introduction}
The discovery of a well-defined color bimodality in galaxy populations at redshifts $z = 0 - 2$
is arguably one of the most significant legacies of galaxy surveys. Extensive
discussion on the color properties in large samples of galaxies
can be found in the literature \citep[e.g.][]{bow92,str01,bal04,bel04,wei05,bra09}.
A major task for modern studies of galaxy formation is therefore to understand how and when
internal processes and environment drive the evolution of galaxies from the blue to the red cloud.
While internal mechanisms, including ejective feedback from supernovae or active galactic nuclei,
are deemed responsible for suppressing and quenching star formation at all densities,
environmental processes in the form of tidal interactions, harassment, and ram-pressure or viscous
stripping are key players in shaping the observed color bimodality within rich galaxy clusters,
especially at the faint end of the galaxy luminosity function.
The origin and effects of these processes have been the subject of scrutiny of many
studies and review articles
\citep[e.g.][]{lar80,has98,lew02,gav02,kau04,balo04,bos06,bos08,gav10,smi12,dre13}.
\begin{figure*}
\includegraphics[scale=0.58]{nicefov.eps}
\caption{{\it Left:} Archival HST/ACS image in the F475W filter (PID 11683; PI Ming Sun)
with, superposed, the MUSE field of view at the two locations targeted by our observations.
In this figure, north is up and east is to the left. The coordinate system is centered at the J2000
position of ESO137$-$001\ ($\alpha=$16:13:27.3, $\delta=$-60:45:51). At the redshift of ESO137$-$001,
$1' = 18.4$ kpc. {\it Right:} RGB color image obtained combining
images extracted from the MUSE datacube in three wavelength intervals ($\lambda = 5000-6000~$\AA\ for the
B channel, $\lambda = 6000-7000~$\AA\ for the G channel, and $\lambda = 7000-8000~$\AA\ for the R channel).
A map of the H$\alpha$\ flux is overlaid in red using a logarithmic scale, revealing the extended gas
tail that originates from the high-velocity encounter of ESO137$-$001\ with the intracluster medium.}\label{fig:fov}
\end{figure*}
\begin{figure*}
\includegraphics[scale=0.55]{haflux.eps}
\caption{{\it Left:} Map of the H$\alpha$\ surface brightness at a resolution of
$60\times60$ pc, obtained from the MUSE datacube as described in the text.
Fluxes have been corrected for foreground Galactic extinction, but not for internal
dust extinction. For visualization purposes, the map has been convolved with a median filter of $5\times 5$
pixels. {\it Right:} The $S/N$ map obtained by dividing the flux values by the associated
errors in each pixel of the filtered datacube, which we have rescaled as described in the text.
The dashed regions mark areas that are not covered by our observations. The image coordinate system
is the same as the one adopted in Figure \ref{fig:fov}, but in proper physical units.
The unprecedented depth of our MUSE observations uncover an extended primary tail, as well as
a secondary southern tail with embedded HII regions.}\label{fig:flux}
\end{figure*}
The smoking gun of ongoing environmental transformation in nearby clusters is the disturbed
gas content of member galaxies. Globally, radio surveys at 21 cm reveal that spiral galaxies in rich environments
are deficient of $\rm H\,{\sevensize I}$\ gas compared to matched samples at lower densities \citep[e.g.][]{hay84,gio85}.
A deficiency of molecular gas has also been reported \citep{fum09,bos14}.
More significantly, these gas-deficient galaxies, and particularly those residing near the cluster cores,
often exhibit truncated and disturbed gas disks \citep[e.g.][]{vol08,chu09} or, in some cases,
one-sided $\rm H\,{\sevensize I}$\ tails \citep[e.g.][]{chu07}.
Furthermore, extreme cases of disturbance are seen in ``head-tail'' radio continuum
emission or in prominent H$\alpha$\ and UV tails arising from galaxies moving towards the centers of
massive clusters, such as Coma, Virgo, and A1367
\citep[e.g.][]{gav85,gav01,yos02,cor06,ken08,smi10,yag10,fos12}.
Detailed multiwavelength studies of these archetypal galaxies which are
undergoing extreme gravitational and hydrodynamic transformations in the harsh cluster environment
become an invaluable tool for unveiling the rich physics responsible for the morphological and color
transformation of cluster galaxies. Many examples of these peculiar objects have been reported in
the recent literature \citep[e.g.][]{vol04,sun06,fum08,hes10,fum11,arr12,fos12,ken14,ebe14}.
In synergy with models and numerical simulations
\citep[e.g.][]{gun72,nul82,moo96,qui00,sch01,vol01,bek03,kro08,ton10,rus14}, these
peculiar systems are in fact ideal laboratories to constrain the efficiency with which
gas can be removed, quenching star formation. Extrapolated to the more general cluster population,
results from these studies offer a unique perspective for how the red sequence is assembled
in dense environments.
A particularly stunning example of a galaxy which is undergoing environmental
transformation is ESO137$-$001, member of the massive Norma cluster.
The Norma cluster, also known as A3627, has a dynamical mass of
$M_{\rm dyn,cl} \sim 1\times 10^{15}~$M$_\odot$\ and lies in the Great Attractor region
at a radial velocity of $v_{\rm cl}= 4871 \pm 54~$$\rm km~s^{-1}$, or redshift
$z_{\rm cl}=0.01625\pm 0.00018$ \citep{wou08,nis12}.
ESO137$-$001\ is located at a projected distance of
only $14.5'$ from the central cluster galaxy WKK6269, corresponding to $\sim 267$ kpc for the
adopted cosmology \citep[$H_0=69.7$ and $\Omega_{\rm m}=0.236$ for which $1' = 18.4$ kpc;][]{hin13}.
Observations across the entire electromagnetic spectrum, including X-ray, optical, infrared, and
millimetric data \citep{sun06,sun07,sun10,siv10,jac14}, show that ESO137$-$001\ is suffering from
an extreme ram-pressure stripping event during its first approach to the center of A3627.
Most notably, ESO137$-$001\ exhibits a double tail which extend for $\sim 80$ kpc as seen in X-ray,
pointing in the opposite direction from the cluster center \citep{sun06,sun10}. This tail, which is
also detected in H$\alpha$\ \citep{sun07} and contains both cold and warm molecular gas \citep{siv10,jac14},
is thought to originate from the interaction between the hot intracluster medium (ICM) and the
warm/cold interstellar medium (ISM) which is being ablated from the stellar disk by ram pressure
\citep[e.g.][]{sun10,jac14}.
A complete picture of how ram pressure operates in this galaxy and how the cold, warm, and hot gas
phases coexist and interact inside the extended tail requires a detailed knowledge of the gas kinematics
and of emission-line ratios. To date, these diagnostics have been limited to pointed radio observations
\citep{jac14} or spectroscopy of selected HII regions \citep{sun07,sun10}.
To dramatically improve the status of current observations, we exploit the unprecedented capabilities
the Multi Unit Spectroscopic Explorer \citep[MUSE;][]{bac10} at the UT4 Very Large Telescope.
With its unique
combination of high efficiency ($\sim 0.35$ at 6800~\AA), extended wavelength coverage
($\sim 4800-9300~$\AA\ sampled at $1.25~$\AA), and large field of view ($1'\times 1'$ in Wide Field Mode,
sampled with $0.2''$ pixels), MUSE is a unique instrument
to map at optical wavelengths the kinematics and line ratios of the disk and tail of ESO137$-$001.
In this paper, we present MUSE observations of ESO137$-$001\ which were conducted as part of the
science verification
run. After discussing the data reduction (Sect. \ref{sec:obs}) and the data analysis technique
(Sect. \ref{sec:ana}), we present a study of the kinematics of the H$\alpha$\ emitting gas and of the
stellar disk of ESO137$-$001\ (Sect. \ref{sec:disc}). A subsequent paper in this series
(Fossati et al., in prep.)
will analyze emission and absorption line ratios to characterize the physical properties of the
emitting gas and of the stellar populations in the disk and tail of ESO137$-$001.
\begin{figure*}
\includegraphics[scale=0.55]{havel.eps}
\caption{{\it Left:} Velocity map of the H$\alpha$\ line relative to the galaxy systemic redshift
at a resolution of $60\times 60$ pc, obtained from the MUSE datacube as described in the text.
For visualization purposes, the map has been convolved with a median filter of $5\times 5$ pixels.
In the inset, we show the stellar kinematics from Figure \ref{fig:stkin}, on the same color
scale adopted for the gas velocity map. To facilitate the comparison, we show the position of the
inset in the main panel (grey box), and we use crosses as a ruler with steps of 2 kpc.
{\it Right:} The error map of the fitted centroids, rescaled as described
in the text. The dashed regions mark areas which are not covered by our observations.
The image coordinate system is the same as the one adopted in Figure \ref{fig:flux}.
The gas and stellar kinematics are remarkably aligned to distances of $\sim 20~$kpc
along the tail, suggesting that ESO137$-$001\ is fast moving along a radial orbit in the plane of
the sky.}\label{fig:vel}
\end{figure*}
\begin{figure*}
\includegraphics[scale=0.55]{hasigma.eps}
\caption{{\it Left:} Velocity dispersion map of the H$\alpha$\ line at a
resolution of $60\times 60$ pc, obtained from the MUSE datacube as described in the text.
Values have been corrected for instrumental resolution, and pixels in which lines are
unresolved are masked in grey in the online edition. For visualization purposes, the map has been convolved with a
median filter of $5\times 5$ pixels. {\it Right:} The error map of the fitted line widths,
rescaled as described in the text. The dashed regions mark areas which are not covered by our observations.
The image coordinate system is the same as the one adopted in Figure \ref{fig:flux}. The transition
from laminar to turbulent flow is seen at $\sim 20~$kpc from the galaxy center.
For a model-dependent velocity of $\sim 3000~$$\rm km~s^{-1}$, turbulence arises on timescales
$\ge 6.5$Myr.}\label{fig:sig}
\end{figure*}
\section{Observations and data reduction}\label{sec:obs}
ESO137$-$001\ was observed with MUSE during the science verification run on
UT June 21, 2014, under program 60.A-9349(A) (PI Fumagalli, Hau, Slezak).
Observations were conducted while the galaxy was transiting at
airmass $\sim 1.25$, in photometric conditions and good seeing
($\sim 0.7''-0.8''$). The disk and the inner tail of ESO137$-$001\ (``Field A'' in Figure \ref{fig:fov})
were observed with three exposures of 900s each, with dithers of $13''-16''$
and a rotation of 90 degrees in the instrument position angle at each position.
A single 900s exposure (``Field B'' in Figure \ref{fig:fov}) was then acquired to cover
the extent of the primary tail. Natural seeing mode and slow guiding mode were used.
The final data product is obtained using a combination of recipes from the
early-release MUSE pipeline (v0.18.1) and a set of in-house IDL codes, which we have developed
to improve the quality of the illumination, sky subtraction, and flux calibration of the MUSE data
specifically for the problem at hand. First, using the MUSE pipeline, we construct for each
IFU a master bias, a master flat field,
and a master sky flat, together with a wavelength solution. Next, we apply these
calibrations to each science exposure and to exposures of the standard star GD153,
which was observed for 30s at the beginning of the night in photometric conditions.
After computing the instrument response curve for flux calibration and the telluric correction spectrum
within the MUSE pipeline, we apply a smoothing kernel of 80 pixel to the
obtained sensitivity function so as to remove residual small-scale fluctuations in the
instrument response, which are present in the current version of the MUSE pipeline.
The quality of photometric calibration is confirmed by aperture photometry
on the $r$-band image reconstructed from the final datacube, which is in excellent agreement with
$r$-band magnitudes in APASS, the AAVSO All-Sky Photometric Survey \citep{hen12}.
Using the MUSE pipeline, we then produce resampled datacubes for each exposure (in spaxels of
1.25\AA\ and $0.2''\times 0.2''$), also saving individual calibrated pixel tables which contain the data
and variance pixels before interpolation. Cosmic rays are also identified as 7$\sigma$ outliers. At this stage,
the MUSE pipeline could be used to compute the sky subtraction with the algorithm offered by the
pipeline recipes. However, when sky subtraction is
enabled, the reconstructed cube exhibits an evident variation in the illumination of each slice which,
as noted in the pipeline manual, is currently not fully corrected by sky flats.
Due to this variation, and perhaps to the particularly crowded nature of our field, the
sky-subtracted cubes exhibit large negative residuals which would affect our subsequent analysis.
To improve the data quality, we perform both an illumination correction and sky subtraction,
using a set of IDL procedures which we have specifically developed to handle the reduced MUSE pixel tables.
First, the illumination correction is computed by fitting eight bright sky lines in all slices
within each exposure and by normalizing them to the median flux distribution across the entire field of view.
After applying this illumination correction, we then generate a sky model in each exposure by combining pixels
of comparable wavelength and within the 20-30th flux percentiles in a master sky spectrum.
This operation is performed on the pixel table, to improve the spectral sampling of the sky model.
We then subtract the sky model from each pixel, by interpolating the master sky spectrum
with a spline function. Following these procedures, we find that the newly reconstructed datacubes present
a more uniform illumination pattern and do not have evident residuals in the background level.
As a further test of our procedure, we verify that the line flux from extragalactic sources
in the sky-subtracted cubes and the non-sky subtracted cubes agree to within the typical flux
standard deviation (see below). At each step, our IDL procedures propagate errors in the datacube
which contains the pixel variance (also known as ``stat'' datacube).
At the end of the reduction, we combine the four exposures and we reconstruct the
final datacube inside the MUSE pipeline, after having registered the astrometry
to the available HST imaging with in-house IDL codes to achieve a sub-arcsecond precision
in both the absolute and relative astrometry. Data are interpolated over a regular grid of $0.2''$
spatially and $1.25$\AA\ spectrally. Finally, we apply the heliocentric correction to the wavelengths
computed in air by the pipeline. An RGB color image produced by combining three slices of 1000\AA\
from the final MUSE datacube is shown in Figure \ref{fig:fov}. In this figure,
we also superimpose a map of the H$\alpha$\ flux, obtained as described in the next section.
The final datacube has an excellent image quality, with full width at half maximum (FWHM)
of $0.73''\pm 0.05''$ for point sources in the field. We also characterize the noise properties by
measuring in each spatial pixel the standard deviation of fluxes in the wavelength interval
$6740-6810$\AA, which is clean of bright sky line residuals. The mean surface brightness limit
at $3\sigma$ confidence level is $3.0\times 10^{-18}$ erg~s$^{-1}$~cm$^{-2}$~\AA$^{-1}$~arcsec$^{-2}$
for Field A and $4.9\times 10^{-18}$ erg~s$^{-1}$~cm$^{-2}$~\AA$^{-1}$~arcsec$^{-2}$ for Field B.
Noise properties are uniform across the field, with a spatial variation expressed in
unit of standard deviation about the mean of
$4\times 10^{-19}$ erg~s$^{-1}$~cm$^{-2}$~\AA$^{-1}$~arcsec$^{-2}$ for Field A and
$1\times 10^{-18}$ erg~s$^{-1}$~cm$^{-2}$~\AA$^{-1}$~arcsec$^{-2}$ for Field B.
For an unresolved line of $\Delta \lambda = 2.55$\AA\ at $\sim 6700$\AA, the line surface
brightness limit becomes $7.6\times 10^{-18}$ erg~s$^{-1}$~cm$^{-2}$~arcsec$^{-2}$
for Field A and $1.2\times 10^{-17}$ erg~s$^{-1}$~cm$^{-2}$~arcsec$^{-2}$ for Field B.
\section{Emission and absorption line modeling}\label{sec:ana}
To characterize the kinematics of the gas in the disk and tail, and the galaxy
stellar rotation, we extract maps of the zeroth, first and second moment of
the H$\alpha$\ ($\lambda 6563$) line, together with a map of the first moment of
the \ion{Ca}{II} triplet ($\lambda\lambda\lambda 8498,8542,8662$),
as detailed below. Throughout this analysis, for consistency with previous work
\citep{sun10,jac14}, we assume a systemic heliocentric velocity of $v_{\rm sys} = 4661 \pm 46~$$\rm km~s^{-1}$\ for
ESO137$-$001, equivalent to redshift $z_{\rm sys}=0.01555 \pm 0.00015$.
\subsection{Emission lines}
Maps of the H$\alpha$\ line flux $F_{\rm H\alpha}$ (Figure \ref{fig:flux}), radial velocity relative to systemic
$\Delta v = v - v_{\rm sys}$ (Figure \ref{fig:vel}), and of the velocity dispersion $\sigma$
(Figure \ref{fig:sig}) are obtained from the final datacube using the IDL custom code {\sc kubeviz}.
With {\sc kubeviz}, we fit a combination of 1D Gaussian functions to the H$\alpha$\
and [NII] $\lambda\lambda 6548,6583$ lines, keeping the relative velocity separation
and the flux ratio of the two [NII] lines constant. While measuring the intrinsic
line width $\sigma$, we also fix the instrumental line width $\sigma_{\rm ins}$ at each wavelength
using an interpolation function constructed with a third degree polynomial fit to the resolution curve
in the MUSE manual. At the observed wavelength of H$\alpha$\ computed for the systemic redshift of ESO137$-$001\
($\lambda =6664.87~$\AA), we find $R=2611.9$ or $\sigma_{\rm ins}=49~$$\rm km~s^{-1}$, in agreement with the
resolution measured from skylines within a resampled datacube before sky subtraction.
The spatial variation of the spectral resolution is found to be negligible for emission lines
with an intrinsic width of $\gtrsim 10~$$\rm km~s^{-1}$.
Before the fit, the datacube is median filtered in the spatial direction with a kernel
of $5\times 5$ pixels to increase the individual $S/N$ per pixel. No spectral smoothing
is performed. In each spectrum, the continuum is then evaluated inside two symmetric windows
around the H$\alpha$-[NII] line complex. Only values between the 40th and 60th percentile are modeled
with a linear polynomial to estimate the continuum under each line. The Gaussian fit is performed on
all the continuum-subtracted spaxels, assigning a minimum value of $\sigma=1~$$\rm km~s^{-1}$\ to spectrally
unresolved emission lines. During the fit, {\sc kubeviz} takes into account the noise from the ``stat''
datacube, thus optimally suppressing skyline residuals. However, the adopted variance underestimates
the real error, most notably because it does not account for correlated noise. We therefore renormalize
the final errors on the fitted parameters so as to obtain a reduced $\tilde \chi^2\sim 1$. In the
subsequent analysis, we mask spaxels where the $S/N$ of the H$\alpha$\ flux is $<3$ on the filtered data.
Further masking is applied to the spaxels for which the line centroids or the line widths
are extreme outliers compared to the median value of their distributions, or the errors on these quantities
exceeds 50 $\rm km~s^{-1}$\ (i.e. conditions indicative of a non-converged fit). Throughout this analysis, the
H$\alpha$\ flux has been corrected to account for appreciable Galactic extinction towards the Galactic plane
\citep[$A_{\rm H\alpha}=0.55$;][]{sun07}, but not for internal extinction. However, our conclusions
primarily rely on kinematics that are not severely affected by the presence of internal dust.
As a consistency check on the velocity measurements,
we also model the [SII] $\lambda\lambda 6717,6731$ doublet
following the same procedure. The derived velocities $\Delta v$ from H$\alpha$\ and [SII] are
found to be in excellent agreement, to better than 1~$\rm km~s^{-1}$\ on average over the entire cube.
The difference in the two reconstructed velocity maps is $\sim 18~$$\rm km~s^{-1}$\ for all regions where
[SII] is detected at $S/N>3$, and $\sim 7~$$\rm km~s^{-1}$\ where $S/N>10$. A similar agreement is found
when comparing the two measurements of velocity dispersion. In the following, we will
only present and analyze the velocity field measured via H$\alpha$\ emission, the
strongest transition, which we use as tracer for the bulk of the ionized hydrogen. A detailed analysis
the properties of [SII] and other emission lines will be presented in the second paper of this series.
We also note that, given the tracer used in our study, we will focus on the ram-pressure stripping of
warm disk gas, which differ from strangulation, the stripping of the hot gas in the halo.
As a final check, we compare the fluxes of HII regions in our data with the
values reported by \citet{sun07} in their table 1. We focus on
three isolated regions \citep[ID 2, 3, and 7 in figure 1 of][]{sun07}, for which we
find a flux difference of 1\%, 3\%, and 6\%, respectively.
This test confirms the excellent agreement in the flux calibrations of the two data sets.
\subsection{Absorption lines}
Following a similar procedure to the one adopted for modeling emission lines,
we extract the rotation curve of the stellar disk of ESO137$-$001\ by fitting Gaussian functions to
the \ion{Ca}{II} triplet. Specifically, we apply to the datacube
a median filter of $10\times 10$ pixels in the spatial direction, with no filtering in the spectral
direction. We then identify stars on the continuum-subtracted image in the rest-frame wavelength range
$\lambda = 8380-8950~$\AA, which we masks with circular regions of $1.2''$.
The velocity relative to systemic is then found by fitting all the spaxels
which have $S/N>5$ per spaxels as measured on the filtered continuum image, using Gaussian functions
with initial guesses at the rest-frame wavelengths of the \ion{Ca}{II} triplet lines.
We find that the best result is achieved by fitting only the strongest $\lambda8662$ line,
a choice which reduces the noise introduced by poorly converged fits on the weaker transitions
at the lowest $S/N$. A map of the reconstructed velocity field is in Figure \ref{fig:stkin}.
The use of Gaussian functions, although rather simple, is preferred to full
template fitting when handling $\sim 10000$ spaxels in order to reduce the run time
of the analysis by orders of magnitude. Nevertheless, we have conducted extensive tests
comparing results from Gaussian fits to results from full template fitting on a coarsely binned datacube,
so as to verify that our method yields accurate centroids. For this test, we rebin
the data using the Voronoi tessellation method with the code developed by \citet{cap03}.
The binning procedure is performed on a continuum image in the rest-frame
wavelength range $\lambda = 8380-8950~$\AA, requiring a final $S/N$ in the
re-binned regions of $>80$. Galactic stars are again identified and masked.
We then fit to the optimally-combined spectra in each Voronoi cell templates from \citet{cen01},
using the Penalized Pixel-Fitting method (pPXF) developed by \citet{cap04}. During the fitting procedure,
we assume an instrumental line width at relevant wavelengths of $\sigma_{\rm ins}=37~$$\rm km~s^{-1}$, computed as
described in the previous section. Next, we apply our Gaussian fitting technique
to the same binned spectra. When comparing these two independent measurements, we find a narrow
Gaussian distribution, centered at $2~$$\rm km~s^{-1}$\ and with standard deviation of $15~$$\rm km~s^{-1}$,
value which we assume as the typical uncertainty on the velocity measurement over the entire datacube.
As for the case of emission lines, higher signal-to-noise regions are most likely affected by smaller
uncertainties. Furthermore, the 2D velocity field measured in the coarser Voronoi cells is
in excellent agreement with the higher-resolution map recovered with Gaussian fits,
confirming the robustness of the measurement presented in Figure \ref{fig:stkin}.
\begin{figure}
\includegraphics[scale=0.55]{hakin.eps}
\caption{Velocity map of the stellar component in the disk of ESO137$-$001\ as traced
by the \ion{Ca}{II} triplet at a resolution of $60\times 60$ pc.
The map has been smoothed with a median filter of $10\times 10$ pixels for visualization
purposes. The dashed contours mark isophotes measured on the stellar continuum between
$\lambda = 8380-8950~$\AA, in steps of 0.4 mag~arcsec$^{-2}$. The image coordinate system
is the same as the one adopted in Figure \ref{fig:flux}, but within a smaller box centered
at the galaxy position. The ordered velocity field is aligned with
respect to the stellar isophotes implying that gravitational interactions cannot be
the dominant mechanism for gas removal.}\label{fig:stkin}
\end{figure}
\begin{figure}
\includegraphics[scale=0.34, angle=90]{haxray.eps}
\caption{Same as the left panel of Figure \ref{fig:flux}, but with logarithmic contours
from Chandra observations in steps of 0.05 dex to show the location of the X-ray
emitting gas. The X-ray map has been smoothed, but in fact the X-ray front is at the same position of the
H$\alpha$\ front \citep[see][]{sun10}. The X-ray tails and the diffuse H$\alpha$\ emission appear to
be co-spatial.}\label{fig:xray}
\end{figure}
\section{Discussion}\label{sec:disc}
Compared to traditional narrow band imaging in $80-100~$\AA\ wide filters, MUSE enables us
to extract from the datacube flux maps in very narrow wavelength windows
($\sim 3-4~$\AA), in which we optimally weight the signal from emission lines over the
background sky noise. In addition, by fitting emission lines in each spatial pixel, we can more precisely
recover the line flux also in the presence of bright unrelated sources, such as foreground stars, if the
Poisson noise introduced by these sources does not outweigh the signal. These key advantages,
combined with the large VLT aperture, allow us to reconstruct a map of the
H$\alpha$\ emission in the disk and tail of ESO137$-$001\ which is one order of magnitude deeper than
previous observations \citep{sun07}, down to surface brightness limits of
$F_{\rm H\alpha} \sim 10^{-18}$ erg~s$^{-1}$~cm$^{-2}$~arcsec$^{-2}$ (see Figure \ref{fig:flux}).
At this depth, our observations reveal that the H$\alpha$\ tail extends continuously for more
than 30 kpc, with surface brightness
$F_{\rm H\alpha} \sim 10^{-17}-10^{-15}$ erg~s$^{-1}$~cm$^{-2}$~arcsec$^{-2}$.
Figure \ref{fig:xray} also shows that the H$\alpha$\ emission is co-spatial in projection with the primary
X-ray tail. At the west edge of our field of view
($\Delta \alpha \sim 28~{\rm kpc},\Delta \delta \sim 21~{\rm kpc}$), one can see a significant surface
brightness enhancement, which is known to harbor $1.5-1.8\times 10^{8}$M$_\odot$\
of molecular gas \citep{jac14} and it is coincident with an X-ray bright region within the primary
tail \citep{sun10}. On both sides of this primary tail, both towards north and south, we also detect
a group of compact sources, some of which were previously identified as HII regions in
spectroscopic observations by \citet{sun07} and \citet{sun10}.
At the superior depth of our observations, the southern
clump complex ($\Delta \alpha \sim 8~{\rm kpc},\Delta \delta \sim -2~{\rm kpc}$)
appears to be embedded in a lower surface brightness component, similar to the diffuse
emission of the primary tail, stretching with the same projected orientation in the plane of the sky,
with a position angle of approximately $-54\deg$. This second H$\alpha$\ tail is co-aligned with the
secondary X-ray tail visible in the Chandra map, suggesting a common origin of the two gas phases,
which possibly originate from material that has been stripped from the southern part of the galaxy.
Unfortunately, our observations do not cover the full extent of this secondary tail,
as visible for instance in figure 2 of \citet{sun10}. Additional observations in the south-west
will thus be critical for a detailed investigation of the properties of the secondary tail.
Near the disk, embedded within the H$\alpha$\ emission of this secondary tail, there are few
bright H$\alpha$\ knots, which \citet{sun10} identified as HII regions.
In contrast to the more diffuse emission, these bright HII regions (and in particular the southernmost clump)
do not appear to be tightly correlated with the X-ray flux, hinting to a different physical
origin of the emission in compact regions compared to the diffuse emission in the tails.
Our observations, in agreement with previous studies, point to ram-pressure stripping
as the primary mechanism of gas removal. Indeed, we note how the stellar
velocity field in Figure \ref{fig:stkin} exhibits an ordered velocity gradient along the major axis,
as expected for a nearly edge-on rotating disk. Compared to the isophotes measured on an image
extracted from the datacube in the wavelength range $\lambda = 8380-8950$\AA, the measured velocity
field appears to be symmetric relative to the major axis and it is also well-centered at the optical
location of ESO137$-$001. This ordered motion and an extremely disturbed morphology of the ISM compared
to unperturbed isophotes support the conclusion of earlier studies that
ram-pressure stripping is responsible for the gas ablation. Indeed, ram-pressure stripping produces
only subtle transformations in the stellar disks \citep[e.g.][]{smi12b}, which largely retain memory
of the original morphology and kinematics \citep[e.g.][]{bos08,tol11}. One may however argue that
the velocity map in Figure \ref{fig:stkin} suggests the presence of small perturbations in the velocity
field in the northern part of the stellar disk. However, we are more cautious in interpreting
these features that are detected
only in regions of low surface brightness and in proximity to two bright foreground stars, which
may contaminate our measurement of the velocity fields at these locations.
Our observations, combined with the available HST imaging (see Figure \ref{fig:fov}), support a
picture in which the galaxy suffered from ram pressure in the disk outskirts in the past, and it is
still being ablated of its gas from the innermost regions.
Through models in which the efficiency of ram-pressure stripping depends on both gas density and
galactocentric radius, as shown for example in figure 8 by \citet{jac14}, we can establish a
correspondence between the spatial location of the H$\alpha$\ emitting gas and the time at which this
gas was stripped. Moreover, lower density material in the less-bound outskirts of ESO137$-$001\ is stripped
more easily at earlier times. This means that material originating from the disk outskirts traces
a more rapid stripping process which will come to an end before the stripping from the nucleus.
Our map shows that the H$\alpha$\ bright knots that are offset from the primary
tail are being stripped from the northern and southern parts of the disk, respectively.
At these large galactocentric radii, there is no evidence of H$\alpha$\ emitting gas inside the stellar disk
of ESO137$-$001\ to our sensitivity limit. Thus, ram pressure has already cleared the galaxy ISM in these
less-bound regions. Conversely, the galaxy ISM is still present in the inner $2-3$ kpc of ESO137$-$001\
\citep[see also][]{jac14}, and this material is still being stripped, feeding the primary tail.
We therefore conclude that the primary tail originates from the inner regions of ESO137$-$001\ and not
from the northern trailing arm, as suggested by the earlier analysis of H$\alpha$\ narrow-band
imaging \citep{sun10}. This outside-in stripping process offers a natural explanation for the different
morphology of the primary H$\alpha$\ tail compared to the two complexes of H$\alpha$\ bright clumps in the north
and south direction. We speculate that the diffuse gas from the outer disk was removed in the
past and it has already mixed with the hot ambient medium downstream. As a consequence, we are now
witnessing only the last episode of stripping of the densest clumps of gas, which trace the
location of the older secondary tails. This is also a plausible explanation for why the primary H$\alpha$\ tail
is narrow compared to the optical disk \citep{sun10}. Originally, there may have been a
single tail, fed by material stripped at all galactocentric radii, differently than what we see today.
A qualitative picture for how an inside-out stripping affects the morphology of the tail can be seen,
for instance, in the right panels of figure 1 in \citet{qui00}.
The analysis of the gas and stellar kinematics offers additional insight into how ram-pressure
stripping occurs in ESO137$-$001\ and how this galaxy moves as it approaches the
center of the cluster. The H$\alpha$\ map shows that ESO137$-$001\ is moving towards the center the Norma
cluster on a projected orbit with position angle of $\sim -54\deg$. The gas radial velocity,
as traced by H$\alpha$\ (Figure \ref{fig:vel}), reveals no
evident gradient along the primary tail, implying that the motion occurs in the plane of the sky.
Conversely, a clear velocity gradient is seen in the
direction parallel to the galaxy major axis, perpendicular to the tail. This velocity pattern was
already observed by \citet{sun10} who studied selected HII regions with Gemini/GMOS spectroscopy.
The unprecedented quality of the MUSE datacube, however, dramatically improves the fidelity
with which the gas kinematics of ESO137$-$001\ is traced, also within the diffuse component.
Our observations confirm that ESO137$-$001\ is approaching the cluster center on a nearly-radial orbit, with no
significant tangential velocity component. The stripped gas retains a remarkable degree of coherence
in velocity to $\sim 20$ kpc downstream,
which is originally imprinted by the rotation curve of the stellar disk.
The map of the stellar kinematics (Figure \ref{fig:stkin}) shows a clear rotational
signature in the plane
of the sky, with peak to peak velocity of $\pm 70-80~$$\rm km~s^{-1}$\ in projection out to
radii of $4-5~$kpc. These values can be compared to the rotational velocity of $\sim 110-120~$$\rm km~s^{-1}$,
which \citet{jac14} infer from the $K-$band luminosity.
Radial velocities in the range $\sim 80-120~$$\rm km~s^{-1}$\ are visible in the stripped gas primarily in the
southern secondary tail, which in our picture originates from material that detached from the outermost
part of the disk of ESO137$-$001, at radii $\gtrsim 5~$kpc where stars rotate at
comparable velocity \citep[see inset in Figure \ref{fig:vel} and cf. the rotational velocity estimates
by][]{jac14}. A hint of the same effect, but with reversed
sign, is visible in the northern part of the primary tail where the stripped gas is blue-shifted compared
to systemic velocity, in agreement with the velocity field of the stellar disk.
The infall velocity $v_{\rm inf}$ is unknown from direct observations,
but we can combine several pieces of evidence to conclude that ESO137$-$001\
is fast moving towards the center of the cluster.
Given the cluster velocity dispersion, which \citet{wou08} measured as
$\sigma_{\rm cl} \sim 949$~$\rm km~s^{-1}$, one could derive an approximate
estimate for the orbital velocity of ESO137$-$001\ as $v_{\rm inf} \sim \sqrt3 \sigma_{\rm cl} \sim 1600~$$\rm km~s^{-1}$.
However, the study of \citet{jac14} suggests that a much higher velocity $\gtrsim 3000$~$\rm km~s^{-1}$\ is
needed for an orbit that crosses the cluster virial radius.
Detailed hydrodynamic calculations are required to link the infall velocity to the
disturbances seen in ESO137$-$001, but here we simply note that such a high velocity is qualitatively
in line with cases of extreme ram pressure. For ESO137$-$001, gas is ablated
even from the most bound regions of the galaxy potential well, implying high ram pressure.
Indeed, \citet{jac14} estimates that during its interaction
with the dense ICM, with electron densities of $1.0-1.4\times 10^{-3}~\rm cm^{-3}$ \citep{sun10}, ESO137$-$001\
can suffer from ram pressure up to a peak of $\sim 2.1\times 10^{-10}~\rm dyne~cm^{-2}$ for their
assumed velocity.
As noted, a coherent radial velocity is visible on scales of at least
$L_{\rm coe}\sim 20~$kpc from the galaxy center along the tail.
Besides, the H$\alpha$\ emitting gas exhibits a modest
degree of turbulent motion on similar scales, as seen from the velocity dispersion map
(Figure \ref{fig:sig}). The bulk of the gas has
a velocity dispersion $\sigma \sim 20-40~\rm$$\rm km~s^{-1}$, in qualitative agreement with the
narrow CO line profiles shown by \citet{jac14}. A lower velocity dispersion
($\sigma \sim 15-25~\rm$$\rm km~s^{-1}$) is seen in the bright H$\alpha$\ knots which are
offset from the primary tail, as expected for a colder medium which hosts sites of ongoing star formation.
Beyond $\sim 20~$kpc from the galaxy position, a higher degree of velocity
dispersion can be seen, with typical values above $\sim 40~$$\rm km~s^{-1}$\ and peaks exceeding
$\sim 100~$$\rm km~s^{-1}$. We need to be more cautious in interpreting kinematics
within Field B, due to the lower signal-to-noise of our observations. Nevertheless,
a progressively higher velocity dispersion moving far from the ESO137$-$001\ disk along the tail
was also noted by \citet{jac14} who found larger FWHM and more complex line profiles in their
CO observations at larger distances from the disk.
Moreover, Gaussian fits to coarsely binned data with higher $S/N$
show larger velocity widths in the outer tail, consistently with the measurements presented
in Figure \ref{fig:sig}.
For an infall velocity $\gtrsim 3000$~$\rm km~s^{-1}$, gas lying at $\gtrsim 20~$kpc
from the ESO137$-$001\ disk has been stripped no less than $\sim 6.5~$Myr ago,
assuming instantaneous acceleration of the removed material to the ICM velocity.
Thus, our observations suggest that substantial turbulence due to fluid instabilities or residual rotational
velocity in the stripped gas grows on timescales of at least $\sim 6.5~$Myr. However, both results of numerical
simulations \citep[e.g.][]{roe08,ton12} and observations \citep{yos12,ken14} reveal that the stripped gas is not
accelerated to orbital velocity instantaneously, and that the denser medium does not
typically reach the infall velocity. Based on these results, the inferred timescale should
be regarded as a conservative lower limit to the growth of instabilities in the tail.
It is also interesting to note that this laminar flow persists despite the shear between the ambient
ICM and the H$\alpha$\ emitting, perhaps due to the presence of magnetic fields \citep{rus14,ton14}.
Finally, we can appeal to the observed coherence to argue that the relative velocity $v_{\rm gas}$
between the gas tail and the galaxy likely exceeds the galaxy circular velocity $v_{\rm cir}$.
Our argument starts by noting that the tail is confined by the ambient ICM.
If that were not the case, in the absence of a restoring centripetal force, material ablated from the disk
at a radius $r_{\rm det}$ where the disk circular velocity is $v_{\rm cir} = v_{\rm los} / \sin i$ would leave
on a tangential orbit, causing the tail to flare. Here, $i = 66\deg$ is the disk inclination angle and
$v_{\rm los}$ is the observed line of sight velocity. Unsurprisingly, this is not what we observe,
as the gas in the tail is confined by the ambient ICM pressure.
In presence of a confining medium, a second velocity component perpendicular to the galaxy orbital motion
contributes to the shear between the tail and the ICM. However, we noted how
the tail retains a line-of-sight velocity similar to the disk rotation, which suggests
that the tail cannot be spinning at higher velocity compared to the
velocity of the gas in the tail relative to the galaxy. In a simple geometry, a parcel of gas that leaves the
disk with velocity $v_{\rm cir}$ rotates by an angle
$\alpha_{\rm rot} \approx \frac{L_{\rm coe}}{r_{\rm det}} \times \frac{v_{\rm cir}}{v_{\rm gas}}$
by the time it has been displaced to a distance $L_{\rm coe}$ with velocity $v_{\rm gas}$.
Thus, the lack of substantial turbulence leads us to conclude that the angle $\alpha_{\rm rot}$
is likely small, which is to say that $v_{\rm cir} < v_{\rm gas}$ and there are no significant instabilities
due to the disk rotation that grow substantially at the interface between the warm gas in the tail and
the hot ICM. Given that $v_{\rm cir}\sim 80-100~\rm km~s^{-1}$ in the outer disk, and that most likely
$v_{\rm gas} < v_{\rm inf}$, we also conclude that $v_{\rm cir} \ll v_{\rm inf}$, i.e. the galaxy is
fast moving in the plane of the sky.
Clearly, this quite simple discussion should be verified with detailed
hydrodynamic simulations designed to reproduce the properties of ESO137$-$001, and that model the
interaction of the warm/hot components, the effect of hydrodynamic and thermal instabilities,
and the role of magnetic fields. All these quantities are in fact
encoded in the phenomenology revealed by our MUSE observations.
\section{Summary and future work}
In this paper we have presented MUSE science verification observations of the central disk and gaseous
tail of ESO137$-$001, a spiral galaxy infalling at high velocity on a nearly-radial orbit
towards the center of the massive Norma cluster at $z \sim 0.0162$. During its encounter with the
ICM, ESO137$-$001\ experiences extreme ram-pressure stripping, which gives rise to
an extended tail that has been detected at multiple wavelengths, from the X-ray to the radio.
The new capabilities of MUSE, with its large field of view, extended wavelength coverage, and high
throughput, have allowed us to map the H$\alpha$\ emitting gas inside ESO137$-$001\ and its tail to
surface brightness limits of $F_{\rm H\alpha} \sim 10^{-18}$ erg~s$^{-1}$~cm$^{-2}$~arcsec$^{-2}$,
and to reconstruct the stellar kinematics traced by the \ion{Ca}{II} triplet.
These new data offer an unprecedented view of ESO137$-$001\ and its tail, which is traced in H$\alpha$\
emission to beyond 30 kpc. These observations allow us to conclude that the ISM of
this galaxy suffered and it is still suffering from ram-pressure stripping at different
galactocentric radii: the less bound material has been completely stripped from the outer disk,
giving rise to clumpy and irregular tails, while the galaxy center still contains part of the original
ISM which feeds the primary H$\alpha$\ tail, aligned with the X-ray emission. The stripped gas retains
the original rotation velocity of the stellar disk, but exhibits a remarkable degree of coherence in
velocity space for $\sim 20~$kpc downstream.
Beyond $\sim 20~$kpc from the galaxy disk along the tail, a greater degree of turbulence is seen, with peak
velocity dispersion $\gtrsim 100~$$\rm km~s^{-1}$. These pieces of evidence are
indicative that the galaxy is fast moving in the plane of sky, with infall velocity greater than
the rotational velocity. For an infall velocity of $\sim 3000$~$\rm km~s^{-1}$, indicated by the orbital study
of \citet{jac14}, material at $\sim 20~$kpc has been stripped no less than $\sim 6.5~$Myr ago,
suggesting that the transition from laminar to turbulent flow occurs on comparable, or most likely,
longer timescales. Finally, the ordered stellar rotation combined with regular isophotes in the disk
rules out gravitational interactions as the main process responsible for the stripping.
This work offers only a first glimpse of the rich MUSE datacube, in which we detect
multiple transitions including [OIII], [OI], [NII], [SII], H$\beta$ in emission, as well as
absorption lines within the stellar disk, most notably \ion{Ca}{II} and H$\alpha$.
All these transitions will allow us to better constrain the physical properties of
the disk and tail of ESO137$-$001, including the properties of the bow shock and the
interaction between the stripped warm ISM and the hot X-ray emitting plasma.
Furthermore, a detailed comparison of these observations with results of numerical simulations
will be crucial to establish the hydrodynamic and thermal properties of the gas which is
subject to ram-pressure stripping, including the role of instabilities and magnetic fields.
These studies will be presented in future papers of this series.
Our work demonstrates the terrific potential that MUSE and future large field-of-view IFUs on 8-10m class
telescopes have in studying ram-pressure stripping, building on
earlier IFU studies \citep[e.g.][]{mer13}.
By targeting galaxies like ESO137$-$001\ in which ram pressure is near to its maximum, these observations
will provide key information on how gas is ablated from all galactocentric radii, how the diffuse
and dense ISM components interact and mix with the ambient ICM, and on what timescales star formation
responds to this gas removal.
Once applied to more typical cluster galaxies, results from these detailed analyses
will offer a comprehensive view of how ram pressure affects the build-up
of the observed red sequence in rich clusters. In turn, this will help to disentangle the role of
environment from the role of internal processes in shaping the color bimodality seen in the
densest regions of the Universe.
\section*{Acknowledgments}
It is a pleasure to thank Mark Swinbank for his advice on how to handle MUSE data,
together with Emanuele Farina, David Wilman, and Elke Roediger for comments on this manuscript.
We thank David J. Wilman and Joris Gerssen for the development of the KUBEVIZ code, and
Eric Slezak for his contribution to the MUSE science verification programme.
We thank the referee, Jeffrey Kenney, for insightful comments on this manuscript.
This work is based on observations made with ESO telescopes at the La Silla Paranal Observatory
under programme ID 60.A-9349(A). M. Fumagalli acknowledges support by the Science and Technology
Facilities Council [grant number ST/L00075X/1]. M Fossati acknowledges the support of the
Deutsche Forschungsgemeinschaft via Project ID 387/1-1. For access to the data used in this paper
and the IDL code to process the MUSE datacubes, please contact the authors.
|
\section{Introduction and main results} \label{section:1}
The Kardar-Parisi-Zhang (KPZ) equation
is the stochastic partial differential equation (SPDE) \begin{equation} \label{eq:KPZ}
\partial_t h = \frac12 \partial_x^2 h + \frac12 \big(\partial_x h \big)^2
+ \dot{W}(t,x), \quad x\in\R,
\end{equation}
where $\dot{W}(t,x)$ is
the space-time Gaussian white noise, in particular, it has
covariance
$$
E[\dot{W}(t,x)\dot{W}(s,y)] = \de(x-y)\de(t-s).
$$
We consider \eqref{eq:KPZ} in one dimension on the whole line,
or on finite intervals with periodic boundary conditions.
This SPDE is used in the physics literature as a general model for the
fluctuations of a growing interface and $h=h(t,x)\in \R$ describes the
height of the interface at time $t$ and position $x\in \R$.
The coefficients $\tfrac12$ are not essential, since one can change
them under scalings in time, position and values of $h$. The importance of the
KPZ equation comes from the fact that it reflects the behavior of a wide class of microscopic systems.
Unfortunately, the equation \eqref{eq:KPZ} is ill-posed and does not
make sense as written. Indeed, the linear SPDE obtained from
\eqref{eq:KPZ} by dropping the nonlinear term $\tfrac12 (\partial_x h)^2$
has a solution $h(t,x)$ which is $(\tfrac12-\e)$-H\"older continuous
for every $\e>0$ in the spatial variable $x\in \R$, and this suggests
that the term $\tfrac12 (\partial_x h)^2$ would diverge in the equation \eqref{eq:KPZ}.
To compensate for this diverging factor, one needs to
introduce a renormalization and the correct form of the KPZ
equation would be
\begin{equation} \label{eq:KPZ-ren}
\partial_t h = \frac12 \partial_x^2 h + \frac12 \left(\big(\partial_x h \big)^2
-\de_x(x)\right) + \dot{W}(t,x), \quad x\in\R.
\end{equation}
The delta function $\de_x$ evaluated at $x$ is certainly $+\infty$.
At least heuristically, one can derive \eqref{eq:KPZ-ren} from
the well-defined linear stochastic heat equation \eqref{eq:1.1} explained below
by applying Cole-Hopf transform and It\^o's formula and
$-\tfrac12 \de_x(x)$ appears as an It\^o correction term; see
\eqref{eq:1-aa} and \eqref{eq:2-a} below, in which
$\xi^\e \to \de_0(0)=\de_x(x)$, since $\eta_2^\e(x)\to\de_0(x)$,
and $\dot{W}^\e\to\dot{W}$ as $\e\downarrow 0$ at least
formally. The result is called the \emph{Cole-Hopf solution}. Recently, Hairer \cite{Hai} has shown how to
make sense of \eqref{eq:KPZ-ren} (with periodic boundary conditions) by introducing a renormalization structure
without passing through the Cole-Hopf transform. The resulting (unique) solutions
are indeed the Cole-Hopf solutions \eqref{eq:1.6}.
In this sense, and also because they arise in the weakly asymmetric limit of exclusion models \cite{BG}, the Cole-Hopf solution is
the physically correct solution of the KPZ equation.
To explain the Cole-Hopf solution of the KPZ equation, let us
consider the following one dimensional linear stochastic heat
equation (SHE) for $Z=Z(t,x)$ with a multiplicative noise:
\begin{equation} \label{eq:1.1}
\partial_t Z = \frac12 \partial_x^2 Z + Z \dot{W}(t,x), \quad x\in\R,
\end{equation}
having an initial value $Z(0,x)\ge 0$. Unlike the KPZ equation \eqref{eq:KPZ}, this is a well-posed SPDE. In fact, the solution of the SPDE
\eqref{eq:1.1} is defined in a weak sense, that is, $Z(t) = \{Z(t,x);x\in\R\}$
is called a {\it solution in generalized functions' sense} if it is adapted
to the increasing family
of $\si$-fields generated by $\{\dot{W}(s,x); 0\le s \le t, x\in \R\}$ and satisfies
\begin{equation} \label{eq:1.2-a}
\lan Z(t),\fa\ran = \lan Z(0),\fa\ran
+ \frac12 \int_0^t \lan Z(s),\partial_x^2\fa\ran ds
+ \int_0^t \int_\R Z(s,x)\fa(x) W(dsdx),
\end{equation}
for every $\fa\in C_0^\infty(\R)$, where $\lan Z,\fa\ran = \int_\R Z(x)
\fa(x)dx$ and the last term is defined as the It\^o integral with
respect to the space-time Gaussian white noise. The process
$Z(t)$ is called a {\it mild solution} if it is adapted and satisfies
the stochastic integral equation
\begin{equation} \label{eq:1.3-a}
Z(t,x) = \int_\R p(t,x,y) Z(0,y)dy +
\int_0^t\int_\R p(t-s,x,y) Z(s,y)W(dsdy),
\end{equation}
where $p(t,x,y) = \frac1{\sqrt{2\pi t}} e^{-(y-x)^2/2t}$ is the
heat kernel on $\R$. These two
notions of solutions are equivalent in the class
$\mathcal{C}_{\rm tem}$, which consists of all $Z\in
\mathcal{C}= C(\R,\R)$ satisfying
$$
\|Z\|_{C_r} = \sup_{x\in\R} e^{-r |x|} |Z(x)| < \infty,
$$
for all $r >0$, equipped with the topology induced by norms
$\{\| \cdot \|_{C_r}; r>0\}$. It is known that, if
$Z(0)\in \mathcal{C}_{\rm tem}$, the SPDE \eqref{eq:1.1} has
a unique solution, in both generalized functions' and mild
senses, such that $Z(\cdot) \in C([0,\infty),
\mathcal{C}_{\rm tem})$ a.s. Moreover, the strong comparison theorem
holds, that is, if $Z(0)$ satisfies in addition that
$Z(0) \in \bar{\mathcal{C}}_+ = C(\R,[0,\infty))$ and $Z(0,x)>0$ for
some $x\in \R$, then $Z(t) \in C((0,\infty),\mathcal{C}_+)$ a.s.,
where $\mathcal{C}_+ =C(\R,\R_+), \R_+ = (0,\infty)$,
equipped with the usual topology of uniform convergence on
each bounded interval; see \cite{Mue} and Corollary 1.4
of \cite{Shiga}.
The {\it Cole-Hopf solution} of the KPZ equation is defined from
the solution of \eqref{eq:1.1} as
\begin{equation} \label{eq:1.6}
h(t,x) := \log Z(t,x),
\end{equation}
which is well-defined since $Z(t,x)>0$.
As we mentioned above, in order to link the Cole-Hopf solution
to the KPZ equation, we need to deal with an infinite It\^o correction
term. In other words, a certain renormalization factor which balances
with this diverging term should be introduced in the KPZ equation.
The simplest approximation scheme is to consider
\begin{equation} \label{eq:1-aa}
\partial_t h = \frac12 \partial_x^2 h + \frac12 \big((\partial_x h)^2
-\xi^\e \big)+ \dot{W}^\e(t,x), \quad x \in \R,
\end{equation}
where $\dot{W}^\e(t,x)= \big(\dot{W}(t)*\eta^\e\big)(x)$ is a smeared noise
defined from a usual symmetric convolution kernel $\eta^\e$ which tends to
the $\de$-function $\de_0$ as $\e\downarrow 0$ and
$\xi^\e=\eta_2^\e(0)$ with $\eta_2^\e=\eta^\e*\eta^\e$;
see Section \ref{sec:2.1} for more details.
Under the Cole-Hopf transform \eqref{eq:1.6} or $Z(t,x) :=
e^{h(t,x)}$, by applying It\^o's formula,
this is equivalent to the SPDE
\begin{equation} \label{eq:2-a}
\partial_t Z = \frac12 \partial_x^2 Z + Z \dot{W}^\e(t,x),
\quad x \in \R,
\end{equation}
see \cite{BG}, (3.6). It is easy to see that the solution of \eqref{eq:2-a}
converges to that of \eqref{eq:1.1} as $\e\downarrow 0$, and therefore
the solution of \eqref{eq:1-aa} converges to the Cole-Hopf solution of
the KPZ equation. However, from the point of view of invariant measures,
\eqref{eq:1-aa} is not a good approximation; in fact it is an open problem to describe
the invariant measures of \eqref{eq:1-aa}.
The present paper introduces a renormalization different from
\eqref{eq:1-aa} or the one considered in \cite{Hai}, better adapted to finding the invariant measures.
We consider the following KPZ approximating equation:
\begin{equation} \label{eq:1}
\partial_t h = \frac12 \partial_x^2 h + \frac12 \big((\partial_x h)^2
-\xi^\e \big)* \eta_2^\e + \dot{W}^\e(t,x), \quad x \in \R.
\end{equation}
It is a common fact (or folklore) that the invariant measures are essentially
unchanged if we apply an operator $A$ (in our case the convolution with $\eta^\e$)
to the noise term and apply $A$ twice to the drift term at the same time;
see e.g., \cite{HH}. Here, the convolution commutes with the
second derivatives so that we don't put it in the first term.
Then, the Cole-Hopf transform $Z^\e(t,x)
= e^{h^\e(t,x)}$ applied to the solution $h=h^\e(t,x)$ of \eqref{eq:1} leads to
an SHE with a smeared noise and an extra complex nonlinear term involving
a certain renormalization structure:
\begin{equation} \label{eq:8}
\partial_t Z = \frac12 \partial_x^2 Z + \frac12 Z\left\{\left(
\frac{\partial_x Z}Z \right)^2*\eta_2^\e - \left(\frac{\partial_x Z}Z \right)^2
\right\} + Z \dot{W}^\e(t,x), \quad x\in \R.
\end{equation}
One of the main contributions of this paper is to show that this nonlinear term,
that is the middle term in the right hand side of \eqref{eq:8}, can be replaced by a simple
linear term divided by a specific constant $24$ in the limit when the corresponding
tilt process is in equilibrium; see Theorems \ref{thm:3.1} and \ref{thm:3.1-M}
below. Thus, we derive the SPDE
\begin{equation} \label{eq:8-b}
\partial_t Z = \frac12 \partial_x^2 Z + \frac1{24} Z+ Z \dot{W}(t,x), \quad x\in \R,
\end{equation}
in the limit as $\e\downarrow 0$. Or, we can rephrase it, that the solution
$h^\e(t,x)$ of \eqref{eq:1} converges to $h(t,x)+\tfrac1{24}t$, where
$h(t,x)$ is the Cole-Hopf solution defined by \eqref{eq:1.6}.
The constant $\tfrac1{24}$
frequently appears in KPZ related papers and describes the speed of
the vertical drift of the interface. The same constant
also appears in our formulation and this provides it with a probabilistic
meaning. If the convolution kernel $\eta^\e$ is asymmetric and satisfies a certain
condition, a constant different from $\frac1{24}$
appears in the limit; see Remark \ref{rem:3.3}-(2).
For technical reasons (see Lemma \ref{cor:3.10} and Proposition
\ref{prop:3.12}), in order to study the limit $\e\downarrow 0$ , we need to work with the KPZ
approximating equation \eqref{eq:1}
on finite intervals with periodic boundary conditions:
\begin{equation} \label{eq:1-M}
\partial_t h = \frac12 \partial_x^2 h + \frac12 \big((\partial_x h)^2
-\xi^\e \big)* \eta_2^\e + \dot{W}^\e(t,x), \quad x \in \SS_M,
\end{equation}
where $\SS_M=\R/M\Z (=[0,M))$, $M\ge 1$, is a continuous one-dimensional
torus of size $M$. The convolution $*\eta_2^\e$ is defined in a
periodic sense. A similar SPDE was studied in Da Prato et.\ al.\ \cite{DDT}.
The Cole-Hopf transformed process $Z^{\e,M}(t,x)
= e^{h^{\e,M}(t,x)}, x\in\SS_M$ applied to the solution $h=h^{\e,M}(t,x)$ of \eqref{eq:1-M}
satisfies the SPDE:
\begin{equation} \label{eq:8-M}
\partial_t Z = \frac12 \partial_x^2 Z + \frac12 Z\left\{\left(
\frac{\partial_x Z}Z \right)^2*\eta_2^\e - \left(\frac{\partial_x Z}Z \right)^2
\right\} + Z \dot{W}^\e(t,x), \quad x\in \SS_M.
\end{equation}
We also consider the SPDE
\begin{equation} \label{eq:8-b-M}
\partial_t Z = \frac12 \partial_x^2 Z + \frac1{24} Z+ Z \dot{W}(t,x), \quad x\in \SS_M,
\end{equation}
which will appear in the limit as $\e\downarrow 0$.
To state our first result, we introduce some notation.
Let $B^M=\{B^M(x);x\in \SS_M\}$ be the pinned Brownian motion
satisfying $B^M(0)=B^M(M)=0$. Let $\nu^M$ and $\nu^{\e,M}$ be the distributions
of $B^M$ and $\{B^M*\eta^\e(x) - B^M*\eta^\e(0);x\in\SS_M\}$ with the convolution
defined in a periodic sense on the space $\mathcal{C}_M=C(\SS_M)
(=C(\SS_M,\R))$, respectively.
For a random continuous function $h=\{h(x);x\in\SS_M\}$ on $\SS_M$, we call
$\nabla h=\{h(x)-h(y); x,y\in\R\}$ for $h$ periodically extended on $\R$
the tilt variables of $h$ and denote
$\nabla h\stackrel{\rm{law}}{=} \nu^{\e,M}$ (or $\nu^M$) if the law of
$\{h(x)-h(0);x\in\SS_M\}$ is given by $\nu^{\e,M}$ (or $\nu^M$).
In fact, $\nu^{\e,M}$ is invariant for the tilt process determined from
the SPDE \eqref{eq:1-M}; see Theorem \ref{thm:2}-(1).
Our convergence result on $\SS_M$ is now formulated as follows.
\begin{thm} \label{thm:0}
We fix $M\ge 1$ and assume that the law of $Z^{\e,M}(0,\cdot) = e^{h^{\e,M}(0,\cdot)}$
is determined by $(h^{\e,M}(0,0), \nabla h^{\e,M}(0)) \stackrel{\rm{law}}{=}
\de_{h_0}\otimes\nu^{\e,M}$ with some $h_0\in \R$. Then, for every $t\ge 0$,
the law of the solution $Z^{\e,M}(t,\cdot)$ of the SPDE \eqref{eq:8-M} on
the space $\mathcal{C}_M$ weakly converges as $\e\downarrow 0$ to that
of the solution $Z^M(t,\cdot)$ of the SPDE \eqref{eq:8-b-M} with the initial
distribution determined by $Z^M(0,\cdot) = e^{h^M(0,\cdot)}$ such that
$(h^M(0,0), \nabla h^M(0)) \stackrel{\rm{law}}{=} \de_{h_0}\otimes\nu^M$.
\end{thm}
This theorem immediately implies that the distribution of the tilt variables
of the logarithm of
the solution $Z^M(t,\cdot)$ of the SPDE \eqref{eq:8-b-M} is given by $\nu^M$
for every $t\ge 0$.
We next pass to the limit $M\to\infty$ by extending $Z^M(t,x), x\in \SS_M$
and $\nu^M$ periodically on $\R$.
We need some more notation. Let
$B=\{B(x);x\in \R\}$ be the two-sided Brownian motion satisfying $B(0)=0$,
that is, $\{B(x);x\ge 0\}$ and $\{B(-x);x\ge 0\}$ are both Brownian motions
with time parameter $x\ge 0$ and mutually independent. Let
$\nu$ and $\nu^\e$ be the distributions of $B$ and $\{B*\eta^\e(x)
- B*\eta^\e(0);x\in\R\}$ on the space $\mathcal{C}$, respectively.
The tilt variables $\nabla h$ of $h$ on $\R$ are similarly defined
as above and denote
$\nabla h\stackrel{\rm{law}}{=} \nu$ (or $\nu^\e$) if the law of
$\{h(x)-h(0);x\in\R\}$ is given by $\nu$ (or $\nu^\e$). We introduce
the weighted $L^2$-space $L_r^2(\R)$, $r>0$, which is a family of all
measurable functions $u$ on $\R$ such that
\begin{equation} \label{eq:L_r-norm}
\| u\|_{L_r^2} := \left(\int_\R u(x)^2 e ^{-2r\chi(x)} dx\right)^{1/2} < \infty,
\end{equation}
where $\chi\in C^\infty(\R)$ is a fixed function satisfying
$\chi(x)=|x|$ for $|x|\ge 1$.
Then, it is standard to show that $Z^M(t,x), x\in \R$ converges to
the solution of the SPDE \eqref{eq:8-b} weakly on $C([0,\infty), L_r^2(\R))$, $r>0$
and $\nu^M$ (periodically extended on $\R$)
converges to $\nu$ weakly on $L_r^2(\R)$, $r>1$ as $M\to\infty$;
see Proposition \ref{prop:SPDE-conv} and its consequences.
As a byproduct, though the factor $\tfrac1{24}Z$ is different in \eqref{eq:8-b},
we can investigate the invariant
measures of the SPDE \eqref{eq:1.1} on $\R$. Let $\mu^c, c\in \R$ be
the distribution of $e^{B(x)+cx}, x\in\R$ on $\mathcal{C}_+$, where
$B(x)$ is the two-sided Brownian motion such that $\mu^c(B(0)\in dx) =dx$.
In particular, $\mu^c$ are not probability measures.
Our second result is that $\mu^c$ are invariant for the process $Z(t)$,
which is a solution of the SPDE \eqref{eq:1.1}.
\begin{thm} \label{thm:1.1}
For every bounded, integrable and continuous function $G$ on $\mathcal{C}_+$,
we have that
$$
\int_{\mathcal{C}_+} G(Z(t))d\mu^c =
\int_{\mathcal{C}_+} G(Z(0))d\mu^c,
$$
for all $t\ge 0$ and $c\in \R$, where the integrals in both sides
are defined under the condition that $Z(0)$ is distributed under
$\mu^c$. More precisely, for example, the left hand side is given by
$$
\int_{\mathcal{C}_+} E_Z[G(Z(t))] \mu^c(dZ),
$$
where $E_Z[\cdot]$ stands for the expectation with respect to the
solution of the SPDE \eqref{eq:1.1} with an initial value $Z(0)=Z \in \mathcal{C}_+$.
\end{thm}
At the level of the process $\partial_x \log Z(t,x)$, where the corresponding invariant
measure is a white noise, the result was derived earlier by Bertini and Giacomin \cite{BG}
via the weakly asymmetric limit of simple exclusion processes.
Note that the measure $\mu^c$ is invariant, but not reversible
for the process $Z(t)$. Only a kind of Yaglom reversibility holds; see Remark
\ref{rem:Yaglom} below.
If $Z(t,x)$ is a solution of (\ref{eq:1.1}) and $c\in\mathbb{R}$
then $Z^c(t,x):=e^{cx + \tfrac12 c^2 t}Z(t,x+ct)$ is also a solution
(with a new space-time Gaussian white noise $\dot{W}$). Therefore,
if one can show the invariance of $\mu^0$ for $Z(t)$, since $\mu^0$ is
invariant under both shifts $Z(x)\mapsto Z(x+a)$ and $Z(x)\mapsto e^a Z(x)$,
we have that the $\mu^c$ is also invariant for the process $Z(t)$. For this
reason, in the proof of Theorem \ref{thm:1.1}, we may
assume $c=0$ without loss of generality and write $\mu$ for $\mu^0$.
We will mostly work with tilt variables $\partial_x h$ or $\partial_x \log Z$
rather than heights to avoid the difficulty caused by the non-normalizability
of the measure $\mu$. This is carried
out by introducing an equivalence relation to the state space
$\mathcal{C}_+$ of $Z(t,x)$; see Section \ref{section:3.1}.
Or, one can say that we are only
interested in the shapes of height functions
$h(t,x)=\log Z(t,x)$ by identifying its vertical translations:
$h(t,x) \sim h(t,x)+c$ for all $c\in \R$; see Remark \ref{rem:2.1-c}
below.
\begin{rem} \label{rem:1.1}
One expects $\mu^c$, $c\in\mathbb{R}$ to be all the extremal
invariant measures (except constant multipliers) for the process
$Z(t)$ as in \cite{FS}, but we will not investigate this here. \end{rem}
The paper is organized as follows. In Section 2, we consider
the KPZ approximating equations \eqref{eq:1} and \eqref{eq:1-M},
and study their invariant measures; see Theorems \ref{thm:2}
and \ref{thm:2.7}. This is accomplished by introducing
finite dimensional approximations due to the discretization in space,
and then taking limits. In Section 3, we consider its Cole-Hopf transform
and pass to the limit. We need to replace a complicated nonlinear term
by a simple linear term in the limit; see Theorems \ref{thm:3.1} and
\ref{thm:3.1-M}. This procedure has a similarity
to the so-called Boltzmann-Gibbs principle \cite{KL}, \cite{KLO},
which plays an important role in
establishing the equilibrium fluctuation limits
for large scale interacting systems. We rely
on Wiener-It\^o expansion.
\section{KPZ approximating equation}
This section studies the invariant measures of the KPZ approximating
equations \eqref{eq:1} and \eqref{eq:1-M}. For this purpose, we first consider the
associated tilt process $u=\partial_x h$, which satisfies the SPDEs
\eqref{eq:b-4.1} and \eqref{eq:b-4.1-torus}
stated below, and introduce its finite dimensional
approximations due to the spatial discretization. Indeed, in view of
finding invariant measures, it is important to choose a special
discretization scheme.
The infinitesimal invariance is shown in Lemma \ref{lem:D-IP}.
Then, we find the invariant measures of the SPDEs \eqref{eq:b-4.1}
and \eqref{eq:b-4.1-torus} by passing to the limits; see Theorem \ref{thm:2}.
This part is standard, especially because $0<\e<1$ is fixed and
the noises in \eqref{eq:b-4.1}, \eqref{eq:b-4.1-torus} or \eqref{eq:1},
\eqref{eq:1-M} are smooth. The results are the same even if
we replace $\xi^\e$ by any other constants.
One can actually check the infinitesimal invariance directly for
\eqref{eq:b-4.1} without introducing the spatial discretization; see
Remark \ref{rem:D^3}. The reason we do not do it that way is that there are not clear enough results in the infinite dimensional setting
telling us that the infinitesimal invariance implies the global invariance.
\subsection{Approximating equations and invariant measures
for tilt processes} \label{sec:2.1}
Let $\eta \in C_0^\infty(\R)$ be a function satisfying $\eta(x)
\ge 0, \eta(x) = \eta(-x)$ and $\int_\R \eta(x)dx=1$. We set
$\eta^\e(x) = \eta(x/\e)/\e$ for $\e>0$, $\eta_2(x) =
\eta*\eta(x)$, and $\eta_2^\e(x) =\eta_2(x/\e)/\e$. Note that
$\eta_2^\e(x) = \eta^\e*\eta^\e(x)$. To fix ideas, we assume
that supp$\,\eta \subset [-1,1]$, so that supp$\,\eta^\e \subset
[-\e,\e]$ and supp$\,\eta_2^\e \subset [-2\e,2\e]$.
Define the smeared noise:
\begin{equation} \label{eq:W^e}
W^\e(t,x) = \lan W(t), \eta^\e(x-\cdot) \ran,
\end{equation}
and consider the KPZ approximating equation \eqref{eq:1} on $\R$
for $h= h^\e(t,x)$. By the symmetry of $\eta$, we have that
\begin{equation} \label{eq:2.3}
\xi^\e = \int_\R \eta^\e(y)^2 dy \; (=\eta_2^\e(0)),
\end{equation}
in \eqref{eq:1}.
The solution $h$ of the SPDE \eqref{eq:1} is smooth in $x$ and
we are concerned with the associated tilt process $u=\partial_x h$,
which satisfies the stochastic Burgers' equation:
\begin{equation} \label{eq:b-4.1}
\partial_t u = \frac12 \partial_x^2 u + \frac12 \partial_x(u^2*\eta_2^\e)
+ \partial_x \dot{W}^\e(t,x), \quad x \in \R.
\end{equation}
In this respect, \eqref{eq:1} is a kind of stochastic
Hamilton-Jacobi equation.
Similarly, the tilt process $u=\partial_x h$ of the solution $h$ of
the SPDE \eqref{eq:1-M} on $\SS_M$, $M\ge 1$, satisfies
\begin{equation} \label{eq:b-4.1-torus}
\partial_t u = \frac12 \partial_x^2 u + \frac12 \partial_x(u^2*\eta_2^\e)
+ \partial_x \dot{W}^\e(t,x), \quad x \in \SS_M.
\end{equation}
Note that $\int_{\SS_M} u(t,x)dx=0$ holds for \eqref{eq:b-4.1-torus}.
Let $\nu^\e$ be the distribution of $\partial_x(B*\eta^\e(x))$ on
$\mathcal{C}$, where $B$ is the two-sided Brownian motion satisfying
$B(0)=0$; we abuse the notation for $\nu^\e$ compared with that
introduced in Section \ref{section:1}, since the meanings
are clear. Note that $\nu^\e$ is a probability measure
which is independent of the choice of the value of $B(0)$.
Similarly, $\nu^{\e,M}$ is the distribution of $\partial_x(B^M*\eta^\e(x))$
on $\mathcal{C}_{M,0}$, where $B^M$ is the pinned Brownian motion,
$\mathcal{C}_{M,0} =\{u\in \mathcal{C}_M; \int_{\SS_M}u(x)dx=0\}$
and recall $\mathcal{C}_M = C(\SS_M)$.
Then, the first main result of this section is formulated as in the
following theorem. This will be extended to the height process $h$
in Section \ref{section:2.7}; see Theorem \ref{thm:2.7}.
\begin{thm} \label{thm:2}
{\rm (1)} The probability measure $\nu^{\e,M}$ on $\mathcal{C}_{M,0}$ is invariant
under \eqref{eq:b-4.1-torus}, that is, for the tilt process $\partial_x h$ of
the solution $h$ of the SPDE \eqref{eq:1-M}.\\
{\rm (2)} The probability measure $\nu^\e$ on $\mathcal{C}$ is invariant
under \eqref{eq:b-4.1}, that is, for the tilt process determined from
the SPDE \eqref{eq:1}.
\end{thm}
\subsection{Invariant measure of KPZ approximating equation
on a discrete torus}
In this section, we introduce the KPZ approximating equation
on a discrete torus $\T_N=\{1,2, \ldots,N\}$ with periodic boundary condition.
To study its invariant measure, it is important to choose a special discretization
scheme as we will explain. Let $\a: \Z \to [0,\infty)$ be given and
satisfy the conditions $\a(i)=\a(-i)$ and $\a(i)=0$ for
$i: |i|\ge K$ with some $K\ge 1$. We naturally regard $\a$ as a function on $\T_N$
assuming that $N$ is sufficiently large compared
with the size of the support of $\a$: $N> 2K$.
For $h=(h(i))_{i\in\T_N}\in \R^{\T_N}$, we define $\De h \in \R^{\T_N}$ by
$\De h(i) = h(i+1)+h(i-1)-2h(i), i \in \T_N$ and two functions $G_1(h)
=(G_1(i,h))_{i\in\T_N}$, $G_2(h)=(G_2(i,h))_{i\in\T_N}$ by
\begin{align*}
& G_1(i,h) = (h_{i+1}-h_i)^2 + (h_i-h_{i-1})^2, \\
& G_2(i,h) = (h_{i+1}-h_i)(h_i-h_{i-1}), \quad i\in \T_N,
\end{align*}
respectively. We sometimes write $h_i$ for $h(i)$.
These are discrete analogues of $2(\partial_x h)^2$ and $(\partial_x h)^2$,
respectively. For functions $\b, \ga$ on $\T_N$, we define the convolution
$\b *\ga$ on $\T_N$ by $(\b*\ga)(i) = \sum_{k\in \T_N} \b(i-k)\ga(k), i\in \T_N$,
where $i-k$ is defined in modulo $N$.
We consider the stochastic differential equation for $h_t
=(h_t(i))_{i\in\T_N} \in \R^{\T_N}$:
\begin{equation} \label{eq:SDE-discrete}
dh_t(i) = \frac{\la_1}2\De h_t(i)dt + \la_2 \{\a_2*G_1(i,h_t)+\a_2*G_2(i,h_t)\}dt +
\la_3 dw_t^\a(i), \quad
i\in \T_N,
\end{equation}
where $\la_1, \la_2, \la_3\in\R$ are arbitrary constants,
$\a_2 = \a*\a$, $w_t^\a = \a*w_t$ and $w_t = (w_t(i))_{i\in\T_N}$
is a family of independent Brownian motions.
We consider three operators on $\R^{\T_N}$:
\begin{align*}
& {\mathcal L}_0^\a f(h) = \frac{\la_1}2 \sum_{i\in \T_N} \De h(i)
\frac{\partial f}{\partial h_i} + \frac{\la_3^2}2 \sum_{i,j\in \T_N}
\a_2(i-j) \frac{\partial^2 f}{\partial h_i\partial h_j}, \\
& {\mathcal A}_1^\a f(h) = \sum_{i\in \T_N} (\a_2*G_1)(i,h)
\frac{\partial f}{\partial h_i}, \\
& {\mathcal A}_2^\a f(h) = \sum_{i\in \T_N} (\a_2*G_2)(i,h)
\frac{\partial f}{\partial h_i},
\end{align*}
for $f\in C^2(\R^{\T_N})$.
Then, ${\mathcal L}^\a = {\mathcal L}_0^\a +\la_2 {\mathcal A}_1^\a
+ \la_2 {\mathcal A}_2^\a$
is the generator of the SDE \eqref{eq:SDE-discrete}.
Let $\a^{-1}=\a_N^{-1}$ be
the inverse matrix of $\a=\{\a(i-j)\}_{i,j\in\T_N}$. Note that
the matrix $\a$ may not be invertible in general, but we can
always make $\det \a\not=0$ by slightly perturbing $\a$
and we consider such $\a$.
Let $\mu_N(dh) = e^{- I_N^\a(h)} dh$ be an infinite measure on
$\R^{\T_N}$, where $dh=\prod_{i\in\T_N} dh(i)$ and
$$
I_N^\a(h) = \frac{\la}2 \sum_{j\in\T_N} \{\a^{-1}*h(j+1)- \a^{-1}*h(j)\}^2,
\quad \la=\frac{\la_1}{\la_3^2}.
$$
\begin{lem} \label{lem:D-IP}
For every $f, g\in C_b^2(\R^{\T_N})$, we have that
\begin{equation} \label{eq:D-IP-1}
\int g(h) {\mathcal L}_0^\a f(h) d\mu_N=\int f(h) {\mathcal L}_0^\a g(h) d\mu_N.
\end{equation}
In particular, $\int{\mathcal L}_0^\a f(h)d\mu_N=0$.
We also have that
\begin{equation} \label{eq:D-IP-2}
\int {\mathcal A}_1^\a f(h) d\mu_N= - \int {\mathcal A}_2^\a f(h) d\mu_N.
\end{equation}
Accordingly, we have that
\begin{equation} \label{eq:D-IP-3}
\int {\mathcal L}^\a f(h) d\mu_N=0.
\end{equation}
\end{lem}
\begin{proof}
We first compute derivatives of $ I_N^\a$:
\begin{align} \label{eq:derivative-I}
& \frac{\partial}{\partial h_i} I_N^\a(h) \\
& =\la \sum_{j\in\T_N} \{\a^{-1}*h(j+1)- \a^{-1}*h(j)\}
\frac{\partial}{\partial h_i}\{\a^{-1}*h(j+1)- \a^{-1}*h(j)\} \notag \\
& =\la \sum_{j\in\T_N} \sum_{k\in\T_N} \{\a^{-1}(j+1-k)- \a^{-1}(j-k)\}h(k)\cdot
\{\a^{-1}(j+1-i)- \a^{-1}(j-i)\} \notag \\
& = \la\sum_{k\in\T_N} \{2\a_2^{-1}(i-k)- \a_2^{-1}(i+1-k)- \a_2^{-1}(i-1-k)\}h(k)
\notag \\
& = - \la\sum_{k\in\T_N} \a_2^{-1}(i-k) \De h(k)
= -\la (\a_2^{-1}*\De h)(i). \notag
\end{align}
We now prove the symmetry \eqref{eq:D-IP-1} of $ {\mathcal L}_0^\a$. To this end,
\begin{align*}
\int g \frac{\partial^2 f}{\partial h_i\partial h_j} d \mu_N
& = - \int \frac{\partial}{\partial h_j} \left(g e^{-I_N^\a(h)}\right)
\frac{\partial f}{\partial h_i} dh \\
& = - \int \left( \frac{\partial g}{\partial h_j} - g
\frac{\partial I_N^\a}{\partial h_j} \right)
\frac{\partial f}{\partial h_i} d\mu_N \\
& = - \int \left( \frac{\partial g}{\partial h_j} + g \la
(\a_2^{-1}*\De h)(j) \right)
\frac{\partial f}{\partial h_i} d\mu_N,
\end{align*}
by \eqref{eq:derivative-I}. Therefore, we have that
\begin{align*}
\int g {\mathcal L}_0^\a f d\mu_N
& = -\frac{\la_3^2}2 \sum_{i,j} \int \a_2(i-j) \frac{\partial g}{\partial h_j}
\frac{\partial f}{\partial h_i} d\mu_N
+ \frac{\la_1}2 \sum_i \int g \De h(i) \frac{\partial f}{\partial h_i} d\mu_N\\
& \qquad - \frac{\la_3^2}2\cdot \la \sum_{i,j}\int g \, \a_2(i-j)
(\a_2^{-1}*\De h)(j) \frac{\partial f}{\partial h_i} d\mu_N \\
& = -\frac{\la_3^2}2 \sum_{i,j} \int \a_2(i-j) \frac{\partial g}{\partial h_j}
\frac{\partial f}{\partial h_i} d\mu_N.
\end{align*}
This shows \eqref{eq:D-IP-1}.
We next prove \eqref{eq:D-IP-2}. For $\ell =1,2$,
\begin{align*}
& \int {\mathcal A}_\ell^\a f(h) d\mu_N=
- \sum_i \int f \frac{\partial}{\partial h_i}\left\{ (\a_2*G_\ell)(i)
e^{-I_N^\a(h)} \right\} dh \\
& = - \int f \left\{ \sum_i \frac{\partial}{\partial h_i} (\a_2*G_\ell)(i)
- \sum_i (\a_2*G_\ell)(i) \frac{\partial I_N^\a}{\partial h_i}
\right\} d\mu_N.
\end{align*}
Here, noting that $\frac{\partial G_\ell}{\partial h_i}(j)=0$ if $j\not= i,
i\pm 1$, the first sum vanishes both for $\ell=1,2$:
\begin{align*}
& \sum_i \frac{\partial}{\partial h_i} (\a_2*G_\ell)(i) \\
& = \sum_i \frac{\partial}{\partial h_i} \{\a_2(-1) G_\ell(i+1) +
\a_2(0) G_\ell(i) + \a_2(1) G_\ell(i-1)\} \\
& = \left\{
\begin{aligned}
& -2 (\a_2(0)+\a_2(1)) \sum_i \De h(i) = 0, \quad \text{ for } \ell=1, \\
& \a_2(0) \sum_i \De h(i) + \a_2(1) \sum_i\{
-(h_{i+2}-h_{i+1})+(h_{i-1}-h_{i-2})\} = 0, \quad \text{ for } \ell=2.
\end{aligned}
\right.
\end{align*}
On the other hand, from \eqref{eq:derivative-I}, the second sum
can be rewritten as
\begin{align*}
- \sum_i (\a_2*G_\ell)(i) \frac{\partial I_N^\a}{\partial h_i}
= \la \sum_i (\a_2*G_\ell)(i) (\a_2^{-1}* \De h)(i)
= \la \sum_i G_\ell(i) \De h(i),
\end{align*}
for $\ell=1,2$. However,
\begin{align*}
& \sum_i G_1(i) \De h(i) \\
& = \sum_i \{ (h_{i+1}-h_i)^2 + (h_i-h_{i-1})^2\} \{ (h_{i+1}-h_i)- (h_i-h_{i-1})\} \\
& = \sum_i (h_{i+1}-h_i)^3 - \sum_i (h_{i+1}-h_i)^2(h_i-h_{i-1})
+ \sum_i (h_i-h_{i-1})^2(h_{i+1}-h_i) - \sum_i (h_i-h_{i-1})^3 \\
& = \sum_i G_2(i) \{(h_i-h_{i-1}) - (h_{i+1}-h_i)\} \\
& = - \sum_i G_2(i) \De h(i).
\end{align*}
This implies \eqref{eq:D-IP-2}. \eqref{eq:D-IP-3} is immediate
from \eqref{eq:D-IP-1} and \eqref{eq:D-IP-2}.
\end{proof}
We can apply Echeverria's result \cite{Ech} for the finite dimensional
SDE \eqref{eq:SDE-discrete} and Lemma \ref{lem:D-IP} proves the
invariance of $\mu_N$.
We define the tilt variables $u=(u(i))_{i\in\T_N}$ associated with
$h=(h(i))_{i\in\T_N}$ by $u(i) = \nabla h(i) := h(i+1)-h(i)$.
Note that $u$ always satisfies $\sum_{i\in\T_N} u(i)=0$, and
$I_N^\a(h) = \tilde I_N^\a(u) := \frac{\la}2 \sum_j \{\a^{-1}*u(j)\}^2$.
Let $\nu_N(du) = e^{-\tilde I_N^\a(u)} du/ Z_N^\a$ be
a probability measure on $\R_0^{\T_N} := \{u\in
\R^{\T_N}; \sum_{i\in\T_N} u(i)=0\}$, where $du$ is the Lebesgue
measure on this space and $Z_N^\a$ is a normalizing constant.
Then, our result for the height process
$h$ can be transformed into that for the tilt process $u$:
\begin{prop}
The probability measure $\nu_N$ on $\R_0^{\T_N}$ is invariant
for the tilt process $u=\nabla h$ of the SDE \eqref{eq:SDE-discrete}.
\end{prop}
Note that we will later consider $u^N\equiv \nabla_Nh = N\nabla h$;
see Lemma \ref{lem:2.5-a} and \eqref{eq:SDE-u} below.
\subsection{Invariant measure of KPZ approximating equation
on a continuous torus}
Under a proper scaling in space $i\mapsto x = i/N$, parameters
$\la_1, \la_2, \la_3$ and $\a(\cdot)$, one can show that
the stationary solution of \eqref{eq:SDE-discrete} converges
weakly to that of the SPDE \eqref{eq:1-M} with $M=1$, i.e.,
\eqref{eq:1-M} for $x \in \SS=\R/\Z (=[0,1))$, or, for the
corresponding tilt process, to the SPDE \eqref{eq:b-4.1-torus}
with $M=1$ for fixed $\e>0$, by showing the tightness
of the sequence of stationary solutions of the SDE \eqref{eq:SDE-discrete}.
The goal is to show the following proposition, whose proof will be
completed in Section \ref{section:2.5}. This proposition proves
Theorem \ref{thm:2}-(1).
\begin{prop}\label{prop:2.4-M}
The probability measure $\nu^{\e,1}$ on $\mathcal{C}_{1,0}$
is invariant under the SPDE \eqref{eq:b-4.1-torus} with $M=1$.
By a simple scaling argument, we see that the probability
measure $\nu^{\e,M}$ on $\mathcal{C}_{M,0}$ is invariant
under the SPDE \eqref{eq:b-4.1-torus}.
\end{prop}
We first show the convergence of the invariant measure.
For $u\in \R_0^{\T_N}$, we define $u^N=\{u^N(x);x\in\SS\}$ by a linear
interpolation of $\{u^N(\frac{i}N) := Nu(i)\}_{i\in \T_N}$, that is
\begin{align} \label{eq:interpolation}
u^N(x) & = u^N(\tfrac{i+1}N)\cdot N(x-\tfrac{i}N)+u^N(\tfrac{i}N)
\cdot N(\tfrac{i+1}N-x) \\
& = N^2 u(i+1)(x-\tfrac{i}N)+N^2u(i)(\tfrac{i+1}N-x),
\quad x \in [\tfrac{i}N,\tfrac{i+1}N). \notag
\end{align}
\begin{lem} \label{lem:2.5-a}
Consider $\nu_N$ on $\R_0^{\T_N}$ by choosing $\a: \a(i) =
\frac1N\eta^\e(\frac{i}N)$ and $\la=N$. Then, as $N\to\infty$,
the distribution of $u^N$ under $\nu_N$ weakly converges to
$\nu^{\e,1}$ on the space $\mathcal{C}_1=C(\SS)$.
\end{lem}
\begin{proof}
We first observe that the law of $\{\nabla_N(\a*B^1(\frac{\cdot}N))(i)\}$
coincides with that of $\{u^N(\frac{i}N) = Nu(i)\}_{i\in \T_N}$
under $\nu_N$, where $\{B^1(x);x\in \SS\}$ is the pinned Brownian
motion such that $B^1(0)=B^1(1)=0$. In fact, for every $f\in C_b(\R_0^{\T_N})$,
\begin{align*}
E^{\nu_N}[f(u)]
& =\frac1{Z_N^\a} \int_{\R_0^{\T_N}} f(u)
e^{-\frac{N}2\sum_j \{\a^{-1}*u(j)\}^2} du \\
& =\frac1{\tilde Z_N^\a} \int_{\R_0^{\T_N}} f(\a*\tilde u)
e^{-\frac{N}2\sum_j \tilde u(j)^2} d\tilde u \\
& = E[f(\a*\nabla B^1(\tfrac{\cdot}N))],
\end{align*}
where we have applied the change of variables:
$u= \a*\tilde u$, that is, $\tilde u = \a^{-1}*u$ and
$d\tilde u= C_Ndu$ for the second line and note that the distribution
of $\{\tilde u(j)\}$ under the probability measure
$e^{-\frac{N}2\sum_j \tilde u(j)^2} d\tilde u/\tilde Z_N^\a$
is equal to that of $\{(\nabla B^1(\frac{\cdot}N))(j)\}$ for the
third line. Since $\a*\nabla B^1(\frac{\cdot}N)
= \nabla(\a* B^1(\frac{\cdot}N))$, the above computation
implies that the law of $\{\nabla_N (\a*B^1(\frac{\cdot}N))(i)\}$
is equal to that of $\{u^N(\frac{i}N)=Nu(i)\}$ under $\nu_N$.
However, it is easy to see that the linear interpolation of
$\{\nabla_N (\a*B^1(\frac{\cdot}N))(i)\}$ converges in
$C(\SS)$ to $\{\partial_x(\eta^\e*B^1)\}$ as $N\to\infty$ a.s.,
and this completes the proof.
\end{proof}
We choose $\la_1=N^2, \la_2=\frac16 N^2, \la_3= \sqrt{N}$
and $\a(i) = \frac1N\eta^\e(\frac{i}N)$ in the SDE
\eqref{eq:SDE-discrete}, and set
$$
U_N(t) = \frac1N \sum_{i\in \T_N} \{u_t(i)^2 + (\nabla_Nu_t(i))^2\},
$$
where $u_t(i) := \nabla_Nh_t(i) = N(h_t(i+1)-h_t(i))$. Note that
$u_t=\{u_t(i)\}$ satisfies the following SDE:
\begin{align}\label{eq:SDE-u}
du_t(i) =& \frac12\De_Nu_t(i)dt + \frac16 \nabla_N[\a_2* \{
u_t(\cdot)^2+u_t(\cdot-1)^2+ u_t(\cdot)u_t(\cdot-1)\}](i) dt \\
&+ \sqrt{N} \nabla_N dw_t^\a(i), \notag
\end{align}
where $\De_Nu(i):=N^2\De u(i)$. We define
$\{u_t^N(x); x \in \SS\}$ by the linear interpolation
of $\{u_t^N(\frac{i}N) := u_t(i)\}_{i\in\T_N}$
as in \eqref{eq:interpolation}. Then
$c_1 \|u_t^N\|_{H^1(\SS)}^2 \le U_N(t) \le c_2 \|u_t^N\|_{H^1(\SS)}^2$
with some $0<c_1$ and $c_2<\infty$. We denote Sobolev spaces
of order $s\ge 0$ on $\SS$ by $H^s(\SS)$.
We consider the stationary
solution of \eqref{eq:SDE-u}, that is, the initial value
before scaling is taken as $\{u_0(i)/N\}_{i\in\T_N}
\stackrel{\rm{law}}{=} \nu_N$.
\begin{lem} \label{lem:2.6-a}
{\rm (1)} For every $T>0$, we have the uniform bound:
$$
\sup_{N\in\N}E\left[\sup_{0\le t \le T} U_N(t)
\right] <\infty.
$$
{\rm (2)} For every $T>0, \fa\in C^\infty(\SS)$ and $0\le s < t\le T$,
$$
E[\lan u_t^N-u_s^N,\fa\ran_{\SS}^4] \le C(\fa) (t-s)^2,
$$
holds with $C(\fa)=C_T(\fa)>0$,
where $\lan u^N,\fa\ran_{\SS} = \int_\SS u^N(x)\fa(x)dx$. \\
{\rm (3)} In particular, $\{u_t^N\}_{N\in\N}$ is tight on
$C([0,T],C(\SS))$ for every $T>0$.
\end{lem}
\begin{proof}
The tightness on $C([0,T],H^s(\SS))$ with $s<1$
follows from (1) and (2) noting that the embedding $H^1(\SS)
\subset H^s(\SS)$ is compact by Rellich's theorem; see, e.g.,
the proof of Proposition 3.1 in \cite{F92}. Therefore, (3) follows
by noting $H^s(\SS)\subset C(\SS)$ continuously embedded
if $s>1/2$.
To show (1), we apply It\^o's formula to see that
\begin{align*}
d U_N(t) & = \frac1N\sum_i \{2u_t(i) d u_t(i)+ (du_t(i))^2
+ 2\nabla_N u_t(i) d\nabla_N u_t(i)+ (d\nabla_N u_t(i))^2\} \\
& = \frac1N\sum_i \big[ u_t(i) \big\{\De_N u_t(i)+
\frac13 \nabla_N[\a_2* \{u_t(\cdot)^2+u_t(\cdot-1)^2+
u_t(\cdot)u_t(\cdot-1)\}](i) \big\} \\
& \qquad + \nabla_N u_t(i) \big\{\nabla_N \De_N u_t(i)+
\frac13 \nabla_N^2[\a_2* \{u_t(\cdot)^2+u_t(\cdot-1)^2+
u_t(\cdot)u_t(\cdot-1)\}](i) \big\} \\
& \qquad + (-N \nabla_N^2 \a_2(0) + N \nabla_N^4\a_2(0))\big]
\cdot dt\\
& \qquad + \frac{\sqrt{N}}N \sum_i \{2u_t(i) \nabla_N dw_t^\a(i)
+ 2\nabla_Nu_t(i) \nabla_N^2 dw_t^\a(i)\} \\
&=: b_N(t)dt+dm_N(t),
\end{align*}
where $m_N(t)$ denotes the martingale part.
Therefore, we have that
\begin{align} \label{eq:U-N}
E[\sup_{0\le t \le T} U_N(t)] \le E[U_N(0)] + \int_0^TE[|b_N(t)|]dt
+ E[\sup_{0\le t \le T} m_N(t)].
\end{align}
However, by the stationarity of $u_t$, we easily see that
\begin{align*}
& E[U_N(0)] = \frac1N\sum_i E^{\nu_N}[\bar u(i)^2+(\nabla_N\bar u(i))^2] \le C,
\end{align*}
uniformly in $N$, where $\bar u(i)=u^N(\frac{i}N)
(\stackrel{\rm{law}}{=} \nabla_N(\a*B(\frac{\cdot}N))(i))$, and
\begin{align*}
& E[|b_N(t)|] \le \frac1N\sum_i \big(
E^{\nu_N}[ |\bar u(i) \big\{\De_N \bar u(i)+
\frac13 \nabla_N[\a_2* \{\bar u(\cdot)^2+\bar u(\cdot-1)^2+
\bar u(\cdot)\bar u(\cdot-1)\}](i) \big\} \\
& \qquad\qquad\qquad
+ \nabla_N \bar u(i) \big\{\nabla_N \De_N \bar u(i)+
\frac13 \nabla_N^2[\a_2* \{\bar u(\cdot)^2+\bar u(\cdot-1)^2+
\bar u(\cdot)\bar u(\cdot-1)\}](i) \big\}| ] \\
& \qquad\qquad\qquad
+|-N \nabla_N^2 \a_2(0) + N \nabla_N^4\a_2(0)|\big) \\
& \qquad\quad\;\; \le C,
\end{align*}
since $|-N \nabla_N^2 \a_2(0) + N \nabla_N^4\a_2(0)|$ is bounded in $N$
(asymptotically converging to $|-(\eta_2^\e)''(0)+(\eta_2^\e)''''(0)|$
as $N\to\infty$), and
$E^{\nu_N}[ |\nabla_N^\ell \bar u(i)|^p], \ell =0,1,2,3, p\ge 1$
are all independent of $i$ (because of the shift invariance of
$\nu_N$) and uniformly bounded in $N$.
Moreover, by Doob's inequality and then by the stationarity of $u_t$,
\begin{align*}
E[\sup_{0\le t \le T} m_N(t)]^2
& \le E[\sup_{0\le t \le T} m_N(t)^2] \le 4E[m_N(T)^2] \\
& = 4\int_0^Tdt \frac4{N}\sum_j E\big[\big\{\sum_i(u_t(i)\nabla_N\a(i-j)
+ \nabla_Nu_t(i)\nabla_N^2\a(i-j))\big\}^2\big] \\
& \le CT.
\end{align*}
Note that $\nabla_N\a(i-j) = \eta^\e(\frac{i+1-j}N)- \eta^\e(\frac{i-j}N)$
and $\nabla_{N}^2 \a(i-j) = N\{ \eta^\e(\frac{i+2-j}N) -2 \eta^\e(\frac{i+1-j}N)
+ \eta^\e(\frac{i-j}N)\}$ are both $O(1/N)$. This proves (1).
To show (2), from the definition \eqref{eq:interpolation} of the linear
interpolation for $u_t^N(x)$,
one can rewrite $\lan u_t^N,\fa\ran_{\SS}$ as a sum in $i$. Then, applying the
summation by parts in $i$, we obtain that
\begin{equation} \label{eq:u^N_t}
\lan u_t^N,\fa\ran_{\SS} = \frac1N \sum_i u_t(i) \tilde \fa^N(i),
\end{equation}
where
$$
\tilde\fa^N(i) = N^2 \int_\SS \{1_{[\tfrac{i-1}N,\tfrac{i}N)}(x) (x-\tfrac{i-1}N)
+ 1_{[\tfrac{i}N,\tfrac{i+1}N)}(x) (\tfrac{i+1}N-x) \}\fa(x)dx.
$$
However, the Taylor expansion of $\fa(x)$ around $x=i/N$ in the right
hand side up to the third order leads to
$$
\tilde\fa^N(i) = \fa(\tfrac{i}N) + \tfrac1{12N^2} \fa''(\tfrac{i}N) +
r^N(i), \quad i\in \T_N,
$$
with remainder terms $r^N(i)$ satisfying $|r^N(i)|\le C/N^3$.
This implies that $|R^N(i)|$, $|\nabla_N R^N(i)|$, $|\De_N R^N(i)| \le C/N$
for $R^N(i) := \tilde\fa^N(i)-\fa(\tfrac{i}N)$. Then, from \eqref{eq:SDE-u}
and \eqref{eq:u^N_t}, for $0\le s \le t$, we have that
\begin{align} \label{eq: u_t-u_s}
\lan u_t^N-u_s^N,\fa\ran_{\SS}= I^{(1)}+ I^{(2)} + I^{(3)},
\end{align}
where
\begin{align*}
I^{(1)}&= \frac1{2N} \int_s^t \sum_i u_r(i) \{ (\De_N\fa(\tfrac{\cdot}N))(i)
+ \De_N R^N(i)\} dr, \\
I^{(2)}& = - \frac1{6N} \int_s^t \sum_i
\a_2* \{u_r(\cdot)^2+u_r(\cdot-1)^2+ u_r(\cdot)u_r(\cdot-1)\}(i) \\
& \qquad\qquad\qquad\qquad \times
\{ (\nabla_N\fa(\tfrac{\cdot}N))(i) + \nabla_N R^N(i)\} dr, \\
I^{(3)}&= \frac1{\sqrt{N}}\sum_i\nabla_N(w_t^\a-w_s^\a)(i)\{\fa(\tfrac{i}N)+R^N(i)\}.
\end{align*}
Noting that $E[|u_r(i)|^p] = E^{\nu_N}[|\bar u(i)|^p]$ (by stationarity) are bounded
in $N$ for $p\ge 1$, we easily see that
$
E[(I^{(1)})^4], E[(I^{(2)})^4] \le C(\fa)(t-s)^4.
$
Moreover, it is also easy to see that
$
E[(I^{(2)})^4] \le C(\fa)(t-s)^2.
$
This proves (2).
\end{proof}
\subsection{The martingale problems associated
with the SPDEs \eqref{eq:b-4.1} and \eqref{eq:b-4.1-torus}}
To complete the proofs of Proposition \ref{prop:2.4-M} and then
Theorem \ref{thm:2}, we introduce the martingale formulations for the
SPDEs \eqref{eq:b-4.1} and \eqref{eq:b-4.1-torus}. To this end,
we first introduce the (pre) generators of the processes $h(t)$
determined by \eqref{eq:1} or \eqref{eq:1-M}.
Let $\mathcal{D} = \mathcal{D}(\mathcal{C})$ be the class of
all tame functions $\Phi$ on
$\mathcal{C}=C(\R)$, that is, those of the form:
\begin{equation} \label{eq:2.tame}
\Phi(h) = f(\lan h,\fa_1\ran, \ldots, \lan h,\fa_n\ran), \quad
h\in \mathcal{C},
\end{equation}
with $n= 1,2,\ldots$, $f=f(z_1,\ldots,z_n)\in C_b^2(\R^n), \fa_1
\ldots, \fa_n \in C_0^\infty(\R)$, where $\lan h,\fa\ran =
\int_\R h(x)\fa(x)dx$. We define its functional derivatives by
\begin{align} \label{eq:2.5-D}
& D\Phi(x;h)
= \sum_{i=1}^n \partial_{z_i}f(\lan h,\fa_1\ran, \ldots, \lan h,
\fa_n\ran) \fa_i(x), \\
& D^2\Phi(x_1,x_2;h)
= \sum_{i,j=1}^n \partial_{z_i}\partial_{z_j}
f(\lan h,\fa_1\ran, \ldots, \lan h,\fa_n\ran) \fa_i(x_1)\fa_j(x_2).
\label{eq:2.6-D}
\end{align}
The class $\mathcal{D}_{\infty} = \mathcal{D}_{\infty}
(\mathcal{C})$ stands for the family
of all $\Phi \in \mathcal{D}$ determined by \eqref{eq:2.tame}
with $f\in C_\infty^2(\R^n)$ such that
$
\lim_{|z|\to\infty} \left\{|f(z)| + |\partial_{z_i}f(z)| + |\partial_{z_i}\partial_{z_j}f(z)|
\right\} =0.
$
For $\Phi\in \mathcal{D}$, define two operators $\mathcal{L}_0^\e$
and $\mathcal{A}^\e$ by
\begin{align*}
\mathcal{L}_0^\e\Phi(h) & = \frac12 \int_{\R^2}
D^2 \Phi(x_1,x_2;h) \eta_2^\e(x_1-x_2) dx_1 dx_2
+ \frac12 \int_\R \partial_x^2h(x) D\Phi(x;h) dx, \\
\mathcal{A}^\e\Phi(h) & = \frac12\int_\R\big((\partial_x h)^2 -\xi^\e \big)
*\eta_2^\e(x) D\Phi(x;h) dx.
\end{align*}
Then, $\mathcal{L}^\e := \mathcal{L}_0^\e + \mathcal{A}^\e$ is the (formal)
generator corresponding to the SPDE \eqref{eq:1}. In fact, by applying
It\^o's formula, we have that
$$
d\Phi(h_t) = \lan D\Phi(x;h_t), d h_t(x)\ran_\R
+ \frac12 \lan D^2\Phi(x_1,x_2;h_t), dW^\e(t,x_1)
dW^\e(t,x_2)\ran_{\R^2}
$$
and note that
\begin{equation} \label{eq:2.cov}
dW^\e(t,x_1)dW^\e(t,x_2) = \eta_2^\e(x_1-x_2) dt.
\end{equation}
The (formal) generator corresponding to the SPDE
\eqref{eq:b-4.1} for the tilt process $u=\partial_x h$ is given by
$\mathcal{L}^{\e,U} = \mathcal{L}_{0}^{\e,U} + \mathcal{A}^{\e,U}$,
where
\begin{align*}
\mathcal{L}_0^{\e,U}\Phi(u) & = \frac12 \int_{\R^2}
D^2 \Phi(x_1,x_2;u) \, \partial_{x_1} \partial_{x_2}\{\eta_2^\e(x_1-x_2)\} dx_1 dx_2
+ \frac12 \int_\R \partial_x^2u(x) D\Phi(x;u) dx, \\
\mathcal{A}^{\e,U}\Phi(u) & = \frac12\int_\R
\partial_x (u^2 *\eta_2^\e)(x) D\Phi(x;u) dx,
\end{align*}
for $\Phi=\Phi(u)\in \mathcal{D}$, which is given by \eqref{eq:2.tame}
with $u$ in place of $h$. Note that the derivatives $\partial_x^2$ and
$\partial_x$ in these operators can be moved to $D\Phi(x;u)$
by the integration by parts.
We similarly define $\mathcal{D}(\mathcal{C}_M)$ and
$\mathcal{D}(\mathcal{C}_{M,0})$ as the classes of all
$\Phi$ on $\mathcal{C}_M$ and $\mathcal{C}_{M,0}$, respectively,
of the forms \eqref{eq:2.tame} with $\fa_i \in C^\infty(\SS_M)$
and $\lan h,\fa\ran_{\SS_M} := \int_{\SS_M} h(x)\fa(x)dx$
in place of $\lan h,\fa\ran$.
Then, operators $\mathcal{L}_{0,M}^\e, \mathcal{A}_M^\e$ together
with $\mathcal{L}_M^\e := \mathcal{L}_{0,M}^\e + \mathcal{A}_M^\e$
on $\mathcal{D}(\mathcal{C}_M)$ and $\mathcal{L}_{0,M}^{\e,U},
\mathcal{A}_M^{\e,U}$ together with $\mathcal{L}_M^{\e,U} :=
\mathcal{L}_{0,M}^{\e,U} + \mathcal{A}_M^{\e,U}$ on
$\mathcal{D}(\mathcal{C}_{M,0})$ are defined as
$\mathcal{L}_{0}^\e, \mathcal{A}^\e, \mathcal{L}^\e$
and $\mathcal{L}_{0,M}^{\e,U}, \mathcal{A}_M^{\e,U},
\mathcal{L}_M^{\e,U}$, respectively, by replacing the integrals over
$\R^2$ and $\R$ by those over $\SS_M^2$ and $\SS_M$,
respectively. We also consider the classes of functions
$\mathcal{D}_\infty(\mathcal{C}_M)$ and
$\mathcal{D}_\infty(\mathcal{C}_{M,0})$.
\begin{rem} \label{rem:2.1-c}
We can regard $\mathcal{L}^\e$ as the generator of the tilt process
$u$ by replacing its domain. In fact, let $\mathcal{D}_{\nabla}=
\mathcal{D}_{\nabla}(\mathcal{C})$ be the class of all
$\Phi\in\mathcal{D}$ with $\fa_i$ satisfying $\int_\R\fa_i dx=0,
1\le i \le n$. This is a natural class of functions for tilt variables,
since, under the equivalence relation $h \sim h+c$ with some
$c\in\R$, we have
$\Phi(h) = \Phi(h+c)$ if $\Phi \in \mathcal{D}_{\nabla}$ so that
$\Phi$ is a function on the quotient space $\tilde{\mathcal{C}}
= \mathcal{C}/\!\!\sim$.
For the function
$\Phi \in \mathcal{D}_{\nabla}$, though we write its variable by
$h$, the height $h$ itself has no meaning. In particular, if $h$ is
differentiable, $\Phi \in \mathcal{D}_{\nabla}$ can be considered
as a function of its tilt $u:=h'\equiv \partial_x h$: if $\Phi(u) =
f(\lan u,\psi_1\ran,\ldots, \lan u,\psi_n\ran)$ with $\psi_1,\ldots,
\psi_n\in C_0^\infty(\R)$, then $\lan u,\psi_i\ran =
\lan h, \fa_i\ran$ with $\fa_i:=-\psi_i'$ and $\fa_i$ satisfies the condition
$\int_\R \fa_i dx=0$, which is the additional condition
imposed on $\Phi \in \mathcal{D}_{\nabla}$. We can also define
$\mathcal{D}_{\nabla}(\mathcal{C}_M)$ as the class of all
$\Phi\in \mathcal{D}(\mathcal{C}_M)$ with $\fa_i \in C^\infty(\SS_M)$
satisfying $\int_{\SS_M} \fa_i(x)dx =0$.
\end{rem}
We now introduce the martingale problems associated
with the SPDEs \eqref{eq:b-4.1} and \eqref{eq:b-4.1-torus}
on extended spaces. Recall that $\e>0$ is fixed so that the noise
$\dot{W}^\e(t,x)$ is smooth in $x$.
As a state spaces for the SPDE \eqref{eq:b-4.1},
we take $C(\R)\cap L_r^2(\R), r>0,$ where $L_r^2(\R)$
is the weighted $L^2$-space; recall \eqref{eq:L_r-norm}.
\begin{lem}\label{lem:2.3.1}
{\rm (1)} If the probability measure $P$ on
$C([0,\infty), C(\R)\cap L_r^2(\R))$
is a solution of the
$(\mathcal{L}^{\e,U},\mathcal{D}_{\infty})$-martingale problem,
then there exists $\dot{W}^\e(t,x)$, which is defined on this space
and a Gaussian smeared noise under $P$, such that the
coordinate function $u(t)$ is a solution of the SPDE \eqref{eq:b-4.1}
in the generalized functions' sense; i.e., \eqref{eq:b-4.1} holds multiplied by
any test function $\fa\in C_0^\infty(\R)$ and integrated over $\R$
(as in \eqref{eq:1.2-a}). \\
{\rm (2)} Similar results hold on $\SS_M$: Under the solution $P$
on $C([0,\infty),\mathcal{C}_{M,0})$ of the
$(\mathcal{L}_M^{\e,U},\mathcal{D}_{\infty}(\mathcal{C}_{M,0}))$-martingale
problem, the coordinate function $u(t)$ satisfies the SPDE
\eqref{eq:b-4.1-torus} in the generalized functions' sense
with a certain Gaussian smeared noise $\dot{W}^\e(t,x)$ on
$[0,\infty)\times \SS_M$.
\end{lem}
\begin{proof}
To prove (1), we use two types of functions $\Phi_1(u) = \lan u,\fa\ran$
and $\Phi_2(u) = \lan u,\fa_1\ran \lan u,\fa_2\ran$; more precisely, their
cut-off functions such as $\Phi_{1,N}(u) = g_N(\lan u,\fa\ran)$ with
$g_N\in C_\infty^2(\R)$ satisfying $g_N(x)=x$ for $|x|\le N$
and similarly defined functions $\Phi_{2,N}$ for $\Phi_2$. We denote by
$
b(u,\fa) = \tfrac12 \lan \partial_x^2 u + \partial_x (u^2*\eta_2^\e),\fa\ran.
$
Then,
\begin{align} \label{eq:2.l.1}
M_t(\fa) := & \lan u(t),\fa\ran - \lan u(0),\fa\ran-
\int_0^t \mathcal{L}^{\e,U} \Phi_1(u(s))ds \\
= & \lan u(t),\fa\ran -\lan u(0),\fa\ran- \int_0^t b(u(s),\fa)ds \notag
\end{align}
is a local martingale. Moreover, by noting that $\Phi_2(u(t))- \int_0^t
\mathcal{L}^{\e,U} \Phi_2(u(s))ds$ is a local martingale and
$\mathcal{L}^{\e,U} \Phi_2(u) = \lan \partial_{x_1}\partial_{x_2}
\{\eta_2^\e(\cdot-\cdot)\},\fa_1\otimes
\fa_2\ran_{\R^2} + b(u,\fa_1)\lan u,\fa_2\ran +b(u,\fa_2)\lan u,\fa_1\ran$,
and by applying It\^o's formula, we can easily see that
\begin{align} \label{eq:2.l.2}
M_t(\fa_1) M_t(\fa_2) -
t \int_{\R^2} \partial_{x_1}\partial_{x_2}
\{\eta_2^\e(x_1-x_2)\} \fa_1(x_1) \fa_2(x_2) dx_1dx_2
\end{align}
is a local martingale. This implies that the cross variation of
two local martingales $M_t(\fa_1)$ and $M_t(\fa_2)$ with
$\fa_1, \fa_2\in C_0^\infty(\R)$ is given by
\begin{align} \label{eq:2.l.3}
\lan M(\fa_1), M(\fa_2)\ran_t = t \lan V\fa_1,\fa_2\ran,
\end{align}
where the right hand side denotes the inner product in
$L^2(\R) =L^2(\R,dx)$ and
$$
V\fa(x) = \partial_x \int_{\R} \eta_2^\e(x-x_1) (-\partial_{x_1}) \fa(x_1) dx_1.
$$
Introducing operators: $R=\partial_x$ and
\begin{align*}
& Q\fa(x) = \int_{\R} \eta_2^\e(x-x_1) \fa(x_1) dx_1, \\
& Q^{\frac12}\fa(x) = \int_{\R} \eta^\e(x-x_1) \fa(x_1) dx_1,
\end{align*}
we can rewrite $V$ as $V = (R Q^{\frac12})(R Q^{\frac12})^*$ as operators
on $L^2(\R)$. Note that $Q$ and $Q^{\frac12}$ are symmetric on $L^2(\R)$,
$(Q^{\frac12})^2 =Q$, and, in particular, $Q$ is non-negative, but
$\|Q^{\frac12}\|_{\rm HS}^2 = \int_{\R^2}(\eta^\e(x_1-x_2))^2 dx_1dx_2 =\infty$
so that ${\rm Tr} \, Q =\infty$.
By the martingale representation theorem (see, e.g., \cite{DZ} Theorem 8.2,
actually stated only in case ${\rm Tr} \, Q<\infty$, and also
Remark \ref{rem:cov-Q} below), \eqref{eq:2.l.3} implies that
$$
M_t(x) = R W^Q(t,x) \equiv \partial_x W^Q(t,x),
$$
where $W^Q$ is the $Q$-Wiener process, which has the representation
\eqref{eq:W^e} with a space-time Gaussian white noise. This with
\eqref{eq:2.l.1} implies the conclusion of (1). The proof of (2)
is similar.
\end{proof}
\begin{rem} \label{rem:cov-Q}
In our case, as we pointed out, the assumption ${\rm Tr} \, Q<\infty$
is not satisfied. To overcome this, we may first define $M_t^N(x)$
by restricting $M_t(x)$ on $[-N,N]$ and periodically extending it to
$[-N-2\e,N+2\e]$. For $M_t^N$, the corresponding $Q$-operator
is given by
\begin{align*}
Q^N\fa(x) = \int_{-N-2\e}^{N+2\e} \eta_2^\e(x-x_1) \fa(x_1) dx_1, \quad x\in [-N,N],
\end{align*}
with $\fa$ defined on $[-N,N]$ but periodically extended to
$[-N-2\e,N+2\e]$ and this operator becomes of trace class. Therefore, one can
apply Theorem 8.2 of \cite{DZ} and construct $W^{Q^N}(t,x)$. Then,
by the consistency of $W^{Q^N}(t,x)$, one can extend it to the whole
line $\R$.
\end{rem}
\subsection{Proof of Proposition \ref{prop:2.4-M}} \label{section:2.5}
We may assume $M=1$ without loss of generality. Recall that
$\{u_t^N(t); x\in \SS\}$ is defined by the linear interpolation of the
stationary solution $u_t=\{u_t(i)\}$ of \eqref{eq:SDE-u} in such
a manner that $u_t^N(\tfrac{i}N) = u_t(i), i\in
\T_N$, and it is tight on $C([0,T],C(\SS))$ from Lemma
\ref{lem:2.6-a}. Therefore, by Skorohod's representation theorem,
we can realize on a proper probability space such that
$u_t^N$ converges to some $u_t$ in $C([0,T],C(\SS))$ as
$N\to\infty$ a.s.\ for every $T>0$.
We abuse the notation. Then, for every $\Phi\in
\mathcal{D}(\mathcal{C}(\SS))$, we have that
\begin{equation} \label{eq:dPhi}
d\Phi(u_t^N) = \lan D\Phi(\cdot;u_t^N), du_t^N\ran_{\SS}
+ \frac12 \lan D^2\Phi(x_1,x_2;u_t^N), du_t^N(x_1) du_t^N(x_2)\ran_{\SS^2}.
\end{equation}
However, from \eqref{eq: u_t-u_s}, we have that
\begin{align} \label{eq:du_t^N}
d \lan u_t^N,\fa\ran
= & \frac1{2N} \sum_i u_t^N(\tfrac{i}N) \{ (\De_N\fa(\tfrac{\cdot}N))(i)
+ \De_N R^N(i)\} dt \\
& - \frac1{6N} \sum_i \a_2* \{u_t(\frac{\cdot}N)^2+u_t(\tfrac{\cdot-1}N)^2
+ u_t(\tfrac{\cdot}N)u_t(\tfrac{\cdot-1}N)\}(i) \notag \\
& \qquad\qquad\qquad\qquad \times
\{ (\nabla_N\fa(\tfrac{\cdot}N))(i) + \nabla_N R^N(i)\} dt \notag \\
& - \frac1{\sqrt{N}}\sum_i \{(\nabla_N\fa(\tfrac{\cdot}N))(i)+\nabla_NR^N(i)\}
\sum_j \a(i-j) dw_t(j). \notag
\end{align}
In particular, recalling that $|\nabla_NR^N(i)| \le C/N$,
\begin{align*}
& d \lan u_t^N,\fa_1\ran d \lan u_t^N,\fa_2\ran \\
& = \frac1N \sum_{i,j}\a_2(i-j)
\{(\nabla_N\fa_1(\tfrac{\cdot}N))(i)+\nabla_NR_1^N(i)\}
\{(\nabla_N\fa_2(\tfrac{\cdot}N))(j)+\nabla_NR_2^N(j)\} dt\\
& = \frac1{N^3} \sum_{i,j,k}\eta^\e(\tfrac{i-k}N)\eta^\e(\tfrac{k-j}N)
\{(\nabla_N\fa_1(\tfrac{\cdot}N))(i)+\nabla_NR_1^N(i)\}
\{(\nabla_N\fa_2(\tfrac{\cdot}N))(j)+\nabla_NR_2^N(j)\} dt\\
& \to \int_{\SS^3}\eta^\e(x-z)\eta^\e(y-z)\fa_1'(x)\fa_2'(y)dxdydz \cdot dt
= \int_{\SS^2}\eta_2^\e(x-y)\fa_1'(x)\fa_2'(y)dxdy \cdot dt,
\end{align*}
as $N\to\infty$.
Since $u_t^N$ converges to $u_t$ in the space
$C([0,T],C(\SS))$ a.s., for the limit $u_t$, we see
from \eqref{eq:dPhi} and \eqref{eq:du_t^N} that
$$
\Phi(u_t)- \int_0^t \mathcal{L}_1^{\e,U}\Phi(u_s)ds
$$
is a martingale for every $\Phi\in \mathcal{D}(C(\SS))$ and therefore
for $\Phi\in \mathcal{D}(\mathcal{C}_{1,0})$.
This completes the proof of Proposition \ref{prop:2.4-M} with the
help of Lemma \ref{lem:2.3.1}-(2) and Lemma \ref{lem:2.5-a}.
\subsection{Invariant measure of KPZ approximating equation
on $\R$}
Let $u_t^M=\{u_t^M(x); x\in \SS_M=[0,M)\}$ be the stationary solution of
the SPDE \eqref{eq:b-4.1-torus}, that is, $u_0^M \stackrel{\rm{law}}{=}
\nu^{\e,M}$, constructed in Proposition \ref{prop:2.4-M}.
We extend $u_t^M$ periodically on $\R$.
\begin{lem} \label{lem:2.8-a}
{\rm (1)} For every $T>0$ and $r>0$, we have
$$
\sup_{M\ge 1}E\left[\sup_{0\le t \le T} \|u_t^M\|_{H_r^1(\R)}^2
\right] <\infty,
$$
where $\|u\|_{H_r^1(\R)}^2 = \|u\|_{L_r^2(\R)}^2
+ \|\partial_xu\|_{L_r^2(\R)}^2$. \\
{\rm (2)} For every $T>0, \fa\in C_0^\infty(\R)$ and $0\le s<t\le T$,
$$
E[\lan u_t^M-u_s^M,\fa\ran^4] \le C(\fa) (t-s)^2,
$$
holds with $C(\fa)=C_T(\fa)>0$. \\
{\rm (3)} In particular, $\{u_t^M\}_{M\ge 1}$ is tight on
$C([0,T],C(\R)\cap L_r^2(\R))$ for every $T, r>0$.
\end{lem}
\begin{proof}
The proof is parallel to that of Lemma \ref{lem:2.6-a}. Indeed,
(3) follows from (1) and (2) noting that the embedding $H_r^1(\R)
\subset H_{r'}^s(\R)$ is compact if $r'>r>0$ and $s<1$, and also
$H_r^s(\R) \subset C(\R)$ if $s>1/2$; see \cite{F95}, p.284 for
the weighted Sobolev spaces $H_r^s(\R)$.
To show (1), set $U^M(t) = \|u_t^M\|_{H_r^1(\R)}^2$. Then, by
It\^o's formula,
\begin{align*}
d U^M(t) & = \int_\R\{2u_t(x) d u_t(x)+ (du_t(x))^2
+ 2\partial_x u_t(x) d\partial_x u_t(x)+ (d\partial_x u_t(x))^2\}
e^{-2r\chi(x)}dx \\
& = \int_\R [ u_t(x) \{\partial_x^2 u_t(x)+\partial_x (u_t^2*\eta_2^\e)\}
+ \partial_x u_t(x) \{\partial_x^3 u_t(x)+\partial_x^2 (u_t^2*\eta_2^\e)\} \\
& \qquad + (-(\eta_2^\e)''(0)+(\eta_2^\e)''''(0)) ]
e^{-2r\chi(x)}dx \cdot dt\\
& \qquad + \int_\R \{2u_t(x) d\partial_xW^\e(t,x) +
2\partial_x u_t(x) d\partial_x^2 W^\e(t,x)\} e^{-2r\chi(x)}dx\\
&=: b^M(t)dt+dm^M(t),
\end{align*}
where $u_t=u_t^M$ and $W^\e(t,x)$ originally defined on $\SS_M$
is periodically extended on $\R$. We can bound $E[\sup_{0\le t \le T} U^M(t)]$
by the sum of three terms similarly to \eqref{eq:U-N}. However, we easily see that
\begin{align*}
& E[U^M(0)] = E^{\nu^{\e,M}}[\|\partial_x B^M*\eta^\e\|_{H_r^1(\R)}^2] \le C
\quad (\text{uniformly in }M),\\
\intertext{where $\{B^M(x);x\in \SS_M\}$ is periodically extended on $\R$, and}
& E[|b^M(t)|] \le \int_\R e^{-2r\chi(x)}dx \big(
E^{\nu^{\e,M}}[ | u(x) \{\partial_x^2 u(x)+\partial_x (u^2*\eta_2^\e)\} \\
& \qquad \qquad \qquad
+ \partial_x u(x) \{\partial_x^3 u(x)+\partial_x^2 (u^2*\eta_2^\e)\}|]
+ (-(\eta_2^\e)''(0)+(\eta_2^\e)''''(0))\big) \\
& \qquad\qquad\;\, \le C,
\end{align*}
since $E^{\nu^{\e,M}}[ |\partial_x^\ell u(x)|^p], \ell =0,1,2,3, p\ge 1$
are all independent of $x$ (because of the shift invariance of
$u(x)$ under $\nu^{\e,M}$) and uniformly bounded in $M$.
Moreover, by Doob's inequality and then by the stationarity of $u_t$,
\begin{align*}
E[\sup_{0\le t \le T} m^M(t)]^2
& \le 4\int_0^Tdt \int_{\R^2} 8e^{-2r(\chi(x)+\chi(y))}dxdy \big\{
E^{\nu^{\e,M}}[u(x)u(y)] \partial_x \partial_y\eta_{2,M}^\e(x-y) \\
& \qquad\qquad \qquad\qquad \qquad
+ E^{\nu^{\e,M}}[\partial_x u(x)\partial_y u(y)] \partial_x^2 \partial_y^2
\eta_{2,M}^\e(x-y) \big\} \\
& \le CT,
\end{align*}
where $\eta_{2,M}^\e(x-y)$ is defined in the sense of modulo $M$
in $x-y$; note that $\partial_x \partial_y \eta_{2,M}^\e$ and
$\partial_x^2 \partial_y^2\eta_{2,M}^\e$ are bounded in $M, x, y$.
We have estimated as $m^M(T)^2 \le 2(m_1^M(T)^2+m_2^M(T)^2)$
by decomposing $m^M(T)$ into the sum of two stochastic integrals
$m_1^M(T)$ and $m_2^M(T)$. This proves (1).
For (2), denoting $u_t=u_t^M$ again, we see that
\begin{align*}
\lan u_t-u_s,\fa\ran
&= \frac12 \int_s^t \{\lan u_r,\fa''\ran -
\lan u_r^2*\eta_2^\e,\fa'\ran \}dr
-\{W^\e(t,\fa')-W^\e(s,\fa')\} \\
&=: I^{(1)}+ I^{(2)}.
\end{align*}
However, we easily see that
$
E[(I^{(1)})^4] \le C(\fa)(t-s)^4
$
by the stationarity of $u_t$, and
$
E[(I^{(2)})^4] = C(\fa)(t-s)^2,
$
since $W^\e(t,\fa')$ is a Brownian motion multiplied by a certain
constant. This proves (2).
\end{proof}
Let $u=\{u(x);x\in \SS_M\}$ be a $\mathcal{C}_{M,0}$-valued random variable
distributed under $\nu^{\e,M}$ and, by periodically extending $u$
on $\R$, we can regard $\nu^{\e,M}$ as a probability distribution
on $\mathcal{C}$. Then, the following lemma is easy and the proof
is omitted.
\begin{lem} \label{lem:2.9-b}
The distribution $\nu^{\e,M}$ weakly converges to $\nu^\e$
on the space $\mathcal{C}$ as $M\to\infty$.
\end{lem}
We are now ready to give the proof of Theorem \ref{thm:2}.
\begin{proof}[Proof of Theorem \ref{thm:2}]
The assertion (1) is already shown by Proposition \ref{prop:2.4-M}.
Let us prove (2). We have shown in Lemma \ref{lem:2.8-a} that the periodically
extended stationary solution $\{u_t^M\}_{M\ge 1}$ of
the SPDE \eqref{eq:b-4.1-torus} is tight on
$C([0,T],C(\R)\cap L_r^2(\R))$ for every $r>0$.
Therefore, by Skorohod's representation theorem, we can
realize on a proper probability space that $u_t^M$ converges to
some $u_t$ in $C([0,T],C(\R)\cap L_r^2(\R))$ for every $T, r>0$
as $M\to\infty$ a.s. Then, for every
$\Phi\in \mathcal{D}(\mathcal{C})$,
$$
\Phi(u_t^M)- \int_0^t \mathcal{L}_M^{\e,U}\Phi(u_s^M)ds
$$
is a martingale. Here, in the operator $\mathcal{L}_M^{\e,U}$,
the function $\eta_2^\e$ should be understood in
the sense of modulo $M$. However, noting that the supports of the functions
$\fa_1,\ldots,\fa_n$ appearing in $\Phi$ are compact, we see that
$\mathcal{L}_M^{\e,U}\Phi(u^M)$ converges to $\mathcal{L}^{\e,U}\Phi(u)$
as $M\to\infty$ if $u^M$ converges to $u$ in $C([0,T],C(\R)\cap L_r^2(\R))$.
Thus, one can prove that, for the limit $u_t$,
\begin{equation} \label{eq:mart-d}
\Phi(u_t)- \int_0^t \mathcal{L}^{\e,U}\Phi(u_s)ds
\end{equation}
is a martingale for every $\Phi\in \mathcal{D}(\mathcal{C})$.
This completes the proof of Theorem \ref{thm:2}-(2) with the
help of Lemma \ref{lem:2.3.1}-(1) and Lemma \ref{lem:2.9-b}.
\end{proof}
As a corollary, we can prove the infinitesimal invariance of
$\mathcal{L}^{\e,U}$, the symmetry of $\mathcal{L}_0^{\e,U}$ and the
asymmetry of $\mathcal{A}^{\e,U}$ under $\nu^\e$, respectively, or integration
by parts formulas, and similar results on $\SS_M$.
\begin{cor} \label{cor:asymmetry}
{\rm (1)}
For every $\e>0$ and $\Phi\in \mathcal{D}(\mathcal{C})$, we have that
\begin{equation} \label{eq:Inv-7}
\int \mathcal{L}^{\e,U} \Phi d\nu^\e =0.
\end{equation}
The operators $\mathcal{L}_0^{\e,U}$ and $\mathcal{A}^{\e,U}$ are symmetric
and asymmetric with respect to $\nu^\e$, respectively, that is,
for every $\Phi, \Psi\in \mathcal{D}(\mathcal{C})$,
\begin{align} \label{eq:Inv-aaa}
\int \Psi\mathcal{L}_0^{\e,U} \Phi d\nu^\e
&= \int \Phi\mathcal{L}_0^{\e,U} \Psi d\nu^\e,\\
\intertext{and}
\int \Psi \mathcal{A}^{\e,U} \Phi d\nu^\e
&= -\int \Phi \mathcal{A}^{\e,U} \Psi d\nu^\e. \label{Inv-bbb}
\end{align}
{\rm (2)} Similar results hold on $\SS_M$ with $\mathcal{L}_M^{\e,U}$,
$\mathcal{L}_{0,M}^{\e,U}$, $\mathcal{A}_M^{\e,U}$ and $\nu^{\e,M}$
in place of $\mathcal{L}^{\e,U}$, $\mathcal{L}_0^{\e,U}$,
$\mathcal{A}^{\e,U}$ and $\nu^\e$, respectively.
\end{cor}
\begin{proof}
We give the proof of (1) only.
\eqref{eq:Inv-7} follows by taking the average of the martingale
\eqref{eq:mart-d} and noting that $u_t\stackrel{\rm{law}}{=} \nu^\e$.
\eqref{eq:Inv-aaa} can be shown from \eqref{eq:D-IP-1}
rewritten at the tilt level and by taking the limits twice as we did,
or it can be directly shown by noting that $\nu^\e$ is reversible
for the Ornstein-Uhlenbeck process determined by the SPDE:
\begin{equation*}
\partial_t u = \frac12 \partial_x^2 u
+ \partial_x \dot{W}^\e(t,x), \quad x \in \R.
\end{equation*}
Since $\mathcal{A}^{\e,U} = \mathcal{L}^{\e,U} - \mathcal{L}_0^{\e,U}$,
\eqref{eq:Inv-7} and \eqref{eq:Inv-aaa} with $\Psi=1$ prove that
\begin{equation} \label{eq:Inv-6}
\int \mathcal{A}^{\e,U}\Phi d\nu^\e =0,
\end{equation}
and \eqref{Inv-bbb} follows from this by noting that
$\mathcal{A}^{\e,U}(\Phi \Psi) = \Psi \mathcal{A}^{\e,U} \Phi
+\Phi \mathcal{A}^{\e,U} \Psi$.
\end{proof}
\begin{rem}\label{rem:D^3}
We can alternatively prove the infinitesimal invariance \eqref{eq:Inv-7} directly
using the Wiener-It\^o expansion, see \cite{F14}.
\end {rem}
\begin{rem} \label{rem:Yaglom}
(Yaglom reversibility)
Corollary \ref{cor:asymmetry}
suggests that the generator of the time reversed process
under $\nu^\e$ is given by $\mathcal{L}_0^{\e,U} -
\mathcal{A}^{\e,U}$. Coming back to the level of the height processes,
a simple computation shows that $\mathcal{L}_0^\e \check{\Phi}(h) =
\mathcal{L}_0^\e \Phi(\check{h})$ and $\mathcal{A}^\e \check{\Phi}(h) =
-\mathcal{A}^\e \Phi(\check{h})$ for $\Phi\in \mathcal{D}$, where
$\check{h}$ and $\check{\Phi}$ are defined by the transformations
$\check{h}(x) = -h(-x)$ and $\check{\Phi}(h) = \Phi(\check{h})$,
respectively. This means that
$\check{h}(t,x):= - h(t,-x)$ determined from the solution
$h(t,x)$ of the SPDE \eqref{eq:1} admits the (pre) generator
$\mathcal{L}_0^\e - \mathcal{A}^\e$.
\end{rem}
\subsection{Invariant measure for the height process}
\label{section:2.7}
Theorem \ref{thm:2} deals with the tilt processes only, but this can
be easily extended to the height process. Theorem \ref{thm:2.7}
will not be used later, but we state it for its own interest. Set
\begin{equation*}
X_t^\e = \frac12\int_0^t \partial_x^2 h^\e(s,0)ds
+ \frac12 \int_0^t \big((\partial_x h^\e(s))^2
-\xi^\e \big)* \eta_2^\e(0)ds + W^\e(t,0),
\end{equation*}
for the solution $h^\e(t,x)$ of \eqref{eq:1}.
The key point is that, as functions of $h^\e$, the first and
second terms of $X_t^\e$ are defined on the quotient space
$\tilde{\mathcal{C}}$ defined in Remark \ref{rem:2.1-c}.
Therefore, once $h^\e(t)\in
\tilde{\mathcal{C}}$ is determined by solving the SPDE
\eqref{eq:b-4.1}, we can recover its height at $x=0$ as
\begin{equation}\label{smoothed2}
h^\e(t,0) = h^\e(0,0) + X_t^\e.
\end{equation}
\begin{thm} \label{thm:2.7}
For any bounded, integrable and continuous function $G=G(h_0,h)$
on $\R\times\tilde{\mathcal{C}}$ and for any $\e>0$, $t\ge 0$,
we have that
\begin{equation}\label{smoothed3}
\int_{\R\times\tilde{\mathcal{C}}} G(h^\e(t,0),h^\e(t))dh_0d\nu^\e
= \int_{\R\times\tilde{\mathcal{C}}} G(h^\e(0,0),h^\e(0))
dh_0 d\nu^\e,
\end{equation}
where $dh_0$ means that $h^\e(0,0)$ is distributed under the
Lebesgue measure on $\R$. Note that
$\R\times\tilde{\mathcal{C}}$ can be identified with
$\mathcal{C}$. Similar results hold on $\SS_M$.
\end{thm}
\begin{proof}
From \eqref{smoothed2} and then by the translation-invariance
of the Lebesgue measure and performing the integral in $dh_0$
first, the left hand side of \eqref{smoothed3} is equal to
$$
\int_{\R\times\tilde{\mathcal{C}}} G(h^\e(0,0)+X_t^\e,h^\e(t))
dh_0d\nu^\e
= \int_{\R\times\tilde{\mathcal{C}}} G(h^\e(0,0),h^\e(t))
dh_0 d\nu^\e.
$$
But, this is equal to the right hand side of \eqref{smoothed3}
by the invariance of $\nu^\e$ under $h^\e(t)\in
\tilde{\mathcal{C}}$ due to Theorem \ref{thm:2}-(2).
\end{proof}
\section{Cole-Hopf transform of KPZ approximating equation
and proofs of Theorems \ref{thm:0} and \ref{thm:1.1}}
Our goal is to study the limit of the KPZ approximating equation
\eqref{eq:1} on $\R$ or \eqref{eq:1-M} on $\SS_M$
as $\e\downarrow 0$. To this end,
we move to the level of the corresponding Cole-Hopf transformed
process $Z^\e(t)$ rather than staying with \eqref{eq:1} or \eqref{eq:1-M},
and show that $Z^\e(t)$ converges to the solution $Z(t)$ of the
SPDE \eqref{eq:8-b} on $\R$ or \eqref{eq:8-b-M} on $\SS_M$
at least if the corresponding
tilt process is stationary. This implies that the solution $h^\e(t)$ of
\eqref{eq:1} or \eqref{eq:1-M} converges to $h(t) + \tfrac1{24}t$
as $\e\downarrow 0$, where $h(t)$ is the Cole-Hopf solution of
the KPZ equation defined by \eqref{eq:1.6}. We can actually
do this only for \eqref{eq:1-M}; due to a technical reason,
we do not have Proposition \ref{prop:3.12} on $\R$.
Since all arguments except this do work on $\R$,
we state the results on $\R$ in Sections
\ref{section:3.1}--\ref{section:3.3}. Then, we study the
SPDEs on $\SS_M$ in Sections \ref{section:3.4} and
\ref{subsection:3.4.3}. Finally in Section \ref{section:3.6},
letting $M\to\infty$, as a byproduct, we find an invariant measure
of the SHE \eqref{eq:1.1} on $\R$.
\subsection{The equation for $Z^\e(t)$}\label{section:3.1}
Under the transformation $h\mapsto Z$ defined by $Z=e^h$, the
KPZ approximating equation \eqref{eq:1} is transformed into the
equation \eqref{eq:8} for $Z=Z^\e(t)$.
In fact, by applying It\^o's formula and recalling \eqref{eq:2.cov}
with $x_1=x_2=x$,
\begin{align*}
dZ & = e^h dh + \frac12 e^h (dW^\e)^2 \\
& = \frac12 Z \Big(\partial_x^2 h + \big((\partial_x h)^2
-\xi^\e \big)* \eta_2^\e\Big) dt + Z dW^\e + \frac12 Z \xi^\e dt \\
& = \frac12 Z \Big(\partial_x^2 h + (\partial_x h)^2* \eta_2^\e\Big) dt
+ Z dW^\e.
\end{align*}
Thus, \eqref{eq:8} is obtained noting that $\partial_x^2 h + (\partial_x h)^2
= Z^{-1} \partial_x^2 Z$ and $\partial_x h = \partial_x Z/Z$.
The derivation of \eqref{eq:8-M} from \eqref{eq:1-M} is the same.
We define the notion of tilt variables associated with the process
$Z(t)$. This is a reformulation of those defined for $h$
above Remark \ref{rem:1.1} or in Remark \ref{rem:2.1-c}.
For $Z^1, Z^2\in \mathcal{C}_+$, we say $Z^1 \sim Z^2$
if there exists $c>0$ such that $Z^1(x)=cZ^2(x)$ for all $x\in \R$.
Then, by the linearity and uniqueness of solutions of the SPDE
\eqref{eq:1.1}, we see that $Z^1(t) \sim Z^2(t)$ holds if $Z^1(0)
\sim Z^2(0)$ for two solutions $Z^1(t), Z^2(t)$ of \eqref{eq:1.1}.
Thus, \eqref{eq:1.1} defines a stochastic evolution $\tilde{Z}(t)$
on the quotient space $\tilde{\mathcal{C}}_+ :=
\mathcal{C}_+/\!\!\sim$. The SPDE \eqref{eq:8} has the same
character, though it is nonlinear.
\subsection{Wrapped processes} \label{sec:3.2}
To avoid the complexity arising from the infiniteness of the
invariant measure of $h^\e(t,x)$, we introduce a
modified process $g^\e(t,x)$ of $h^\e(t,x)$.
Let us take $\rho\in C_0^\infty(\R)$ satisfying $\rho\ge 0$,
supp $\rho\subset [-1,1]$, supp $\rho$ is connected,
and $\int_\R\rho(x)dx=1$, and fix it in the rest of the paper
except the last step of Section \ref{subsection:3.4.3},
where we take two different $\rho$'s.
We define a wrapped process $g^\e(t,x)$ of $h^\e(t,x)$ by
$g^\e(t,x)= h^\e(t,x) + N^\e(t)$ with $N^\e(t) = - [h^\e(t,\rho)]$,
more precisely its right continuous modification, where
$[h]\in \Z$ stands for the integer part of $h\in \R$ and
$h^\e(t,\rho) = \int_\R h^\e(t,x)\rho(x)dx$.
In particular, $g^\e(x,\rho)$ defined from $g^\e(t,x)$
similarly to $h^\e(t,\rho)$ always satisfies
$g^\e(t,\rho) \in [0,1]$ a.s.\ and $g^\e(t,\rho) \equiv h^\e(t,\rho)$
modulo $1$.
In the next lemma, the initial distribution of $h^\e(0,\cdot)$ is
taken to be $\pi\otimes\nu^\e$, where $\pi$ is a uniform
measure on $[0,1]$, under the decomposition of the height:
\begin{equation} \label{eq:3.1-map}
g \mapsto (g(\rho),\{g(x)-g(\rho);x\in\R\}),
\end{equation}
into the height averaged by $\rho$ and the tilt variable.
Then, $g^\e(t)$ considered as a $[0,1]\times\tilde{\mathcal{C}}$-valued
process under the map \eqref{eq:3.1-map} is stationary in $t$:
\begin{lem} \label{lem:3.1-a}
The probability measure $\pi\otimes\nu^\e$ on $[0,1]\times
\tilde{\mathcal{C}}$ is invariant under $g^\e(t,x)$.
\end{lem}
\begin{proof}
Take a periodic and smooth function $f$ on $[0,1]$ and
set $\Psi(h) = f(h(\rho))$ for $h\in \mathcal{C} \cong
[0,1]\times\tilde{\mathcal{C}}$ under the map
\eqref{eq:3.1-map}, where $h(\rho) = \int_\R h(x)\rho(x)dx$. Then, since
$$
D\Psi(x;h) = f'(h(\rho)) \rho(x), \quad
D^2\Psi(x_1,x_2;h) = f''(h(\rho)) \rho(x_1)\rho(x_2),
$$
we have that
\begin{align*}
L_{\nabla h}^\e f(h(\rho)) :=& \mathcal{L}^\e \Psi(h)
= \frac{\xi_\rho^\e}2 f''(h(\rho)) + \frac12 b^\e(\nabla h)
f'(h(\rho)),
\end{align*}
where
\begin{align*}
& \xi_\rho^\e = \int_{\R^2} \rho(x_1)\rho(x_2) \eta_2^\e(x_1-x_2)
dx_1 dx_2, \\
& b^\e(\nabla h) = h(\rho'') + \int_\R
((\partial_x h)^2-\xi^\e)*\eta_2^\e(x) \rho(x)dx.
\end{align*}
Note that $b^\e(\nabla h)$ is a tilt variable.
Take another function $\Phi=\Phi(\nabla h)$ of tilt variables
$\nabla h = \{\partial_x h; x \in \R\}$. Then, since
\begin{align*}
\mathcal{L}^\e(\Psi\Phi)
& = \Phi\mathcal{L}^\e\Psi + \Psi\mathcal{L}^\e\Phi
+ \int_{\R^2} D\Psi(x_1;h)D\Phi(x_2;h) \eta_2^\e(x_1-x_2)
dx_1dx_2 \\
& = \Phi L_{\nabla h}^\e f + f\mathcal{L}^\e\Phi
+ f'(h(\rho)) \lan D\Phi(\cdot;h)* \eta_2^\e,\rho\ran,
\end{align*}
noting that $\mathcal{L}^\e\Phi$ and $\lan D\Phi(\cdot;h)* \eta_2^\e,\rho\ran$
are tilt variables, we have that
\begin{align} \label{eq:3.aa}
& \int_{[0,1]\times\tilde{\mathcal{C}}}\mathcal{L}^\e(\Psi\Phi)
d\pi\otimes\nu^\e
= E^{\nu^\e}\left[\Phi(\nabla h) \int_0^1 L_{\nabla h}^\e f(h(\rho))d\pi\right]\\
& \qquad \qquad \notag+ \int_0^1f(h(\rho))d\pi \,
E^{\nu^\e}[\mathcal{L}^\e\Phi]
+ \int_0^1 f'(h(\rho)) d\pi\,
E^{\nu^\e}[\lan D\Phi(\cdot;h)* \eta_2^\e,\rho\ran ].
\end{align}
However, we easily see that
$$
\int_0^1 L_{\nabla h}^\e f(h(\rho))d\pi =
\int_0^1 L_{\nabla h}^\e f(a)da =0
$$
for all fixed $\nabla h$ by the periodicity of $f$, and also
$$
\int_0^1 f'(h(\rho)) d\pi = \int_0^1 f'(a) da=0.
$$
Moreover, noting that $\mathcal{L}^\e$ acting on $\Phi=\Phi(\nabla h)$
through $h$ coincides with $\mathcal{L}^{\e,U}$ acting on $\Phi=\Phi(u)$
through $u$, Corollary \ref{cor:asymmetry} shows that the second term
in the right hand side of \eqref{eq:3.aa} vanishes, and therefore
we have that
$$
\int_{[0,1]\times\tilde{\mathcal{C}}}\mathcal{L}^\e(\Psi\Phi)
d\pi\!\otimes\!\nu^\e=0.
$$
This can be extended to linear combinations of the functions of the
form $\Psi\Phi$, and concludes the proof of the lemma.
\end{proof}
We next introduce the Cole-Hopf transform $Y^\e(t,x) =e^{g^\e(t,x)}$ of
the wrapped process $g^\e(t,x)$. The initial distribution of
$h^\e(0,\cdot)$ is taken as mentioned above
Lemma \ref{lem:3.1-a}. $Y^\e(t,x)$ is called a wrapped
process of $Z^\e(t,x) = e^{h^\e(t,x)}$ and satisfies $Y_\rho^\e(t)\in [1,e]$
a.s., where we define
\begin{equation} \label{eq:3.5-c}
Y_\rho= \exp\left\{ \int_\R \log Y(x) \rho(x)dx\right\},
\end{equation}
for $Y=\{Y(x)>0; x\in \R\}$ and $Y_\rho^\e(t) = (Y^\e(t))_\rho$.
\begin{lem} \label{lem:3.2-R}
$Y^\e(t,x)$ satisfies the following equation in generalized functions' sense:
\begin{align} \label{eq:3-Y^e-2}
Y^\e(t,x) = Y^\e(0,x) + & \frac12 \int_0^t \partial_x^2 Y^\e(s,x)ds
+\int_0^t A^\e(x,Y^\e(s)) ds \\
& + \int_0^t Y^\e(s,x) dW^\e(s,x) + N^\e(t,x), \notag
\end{align}
where
\begin{align} \label{eq:3.4}
A^\e(x,Y) & = \frac12 Y(x)\left\{\left(
\frac{\partial_x Y}Y \right)^2*\eta_2^\e(x)
- \left(\frac{\partial_x Y}Y \right)^2(x) \right\}, \\
N^\e(t,x) & =
\int_0^t \left\{ (e-1) 1_{\{ Y_\rho^\e(s-)=1\}}
+ (e^{-1}-1) 1_{\{ Y_\rho^\e(s-)=e\}} \right\}Y^\e(s-,x) N^\e(ds).
\label{eq:3.5}
\end{align}
\end{lem}
\begin{proof}
Note that $N^\e(t)=-[h^\e(t,\rho)]$ is expressed as
$$
N^\e(t) = \int_0^t \sum_{a=\pm1} a \, n^\e(ds,a)
$$
with a certain point process $n^\e(ds,a)$ on $\mathbb{X}
= \{\pm1\}$. Thus, applying It\^o's formula for $Y^\e(t,x) =
Z^\e(t,x) e^{N^\e(t)} \equiv F(Z^\e(t,x),N^\e(t))$ with $F(z,n) = ze^n,
z\in\R, n\in \Z$, we have that
\begin{align*}
Y^\e(t,x) &= Y^\e(0,x) + \int_0^t \frac{\partial F}{\partial z}
(Z^\e(s,x),N^\e(s)) dZ^\e(s,x) \\
&\qquad +\int_0^{t+} \sum_{a=\pm1} \{ F(Z^\e(s,x),N^\e(s-)+a)
- F(Z^\e(s,x),N^\e(s-))\} n^\e(ds,a) \\
& = Y^\e(0,x) + \int_0^t e^{N^\e(s)} dZ^\e(s,x)
+ N^\e(t,x),
\end{align*}
where $N^\e(t,x)$ is defined by \eqref{eq:3.5}.
The conclusion follows from \eqref{eq:8}.
\end{proof}
\subsection{Asymptotic behavior of the nonlinear term in
\eqref{eq:3-Y^e-2}} \label{section:3.3}
We need to analyze the limit of the third term in the right hand
side of \eqref{eq:3-Y^e-2} as $\e\downarrow 0$ at least in the
stationary situation. The goal of this subsection is to show the
following theorem, by which one can replace $A^\e(x,Y^\e(s))$
with a linear function $\tfrac1{24} Y^\e(s,x)$ if $h^\e(0,\cdot)$
is distributed under $\pi\otimes\nu^\e$.
\begin{thm} \label{thm:3.1}
For every $\fa\in C_0(\R)$ satisfying $\rm{supp}\,\fa \cap
\rm{supp}\,\rho = \emptyset$ (so that $\dist(\rm{supp}\,\fa,
\rm{supp}\,\rho)>0$), we have that
$$
\lim_{\e\downarrow 0} E^{\pi\otimes\nu^\e}
\left[ \sup_{0\le t \le T}\left\{ \int_0^t \hat A^\e(\fa,Y^\e(s))ds \right\}^2
\right] =0,
$$
where
\begin{align*}
\hat A^\e(\fa,Y) & = \int_\R \hat A^\e(x,Y)
\fa(x)dx, \\
\hat A^\e(x,Y) & = A^\e(x,Y) - \frac1{24} Y(x).
\end{align*}
In particular, under the time average, $A^\e(\fa,Y^\e(s))$
can be replaced by $\tfrac1{24}\int_\R Y^\e(s,x)
\fa(x)dx$ in $L^2(\Om)$ in a strong topology as $\e\downarrow 0$
under the equilibrium situation, if $\fa$ satisfies the above
conditions.
\end{thm}
\begin{rem} \label{rem:3.1}
{\rm (1)} The time average is essential to show this theorem.
At each fixed time, we never have this type of statement;
see Remark \ref{rem:3.1} below. \\
{\rm (2)} The constant $\frac1{24}$ frequently appears in KPZ computations;
see e.g.\ Theorem 2.3 of \cite{BG}, Theorem 1.1 of \cite{C} and
Proposition 5.1 of \cite{BCF}.
\end{rem}
The proof of Theorem \ref{thm:3.1} will be carried out at the
level of the height processes $g^\e(t,x)$ or $h^\e(t,x)$ not at that
of the transformed processes $Y^\e(t,x)$ or $Z^\e(t,x)$,
and in a similar way
to that of the Boltzmann-Gibbs principle, which is needed in
the study of the equilibrium fluctuation and establishes a
replacement of a certain complex term by a linear term.
In particular, we deduce an equilibrium dynamic problem into
a static problem.
To this end, we first consider the symmetric part
$\mathcal{S}^\e := \tfrac12(\mathcal{L}^\e+\mathcal{L}^{\e\ast})$
of the generator $\mathcal{L}^\e$ of the height process
and the corresponding Dirichlet
form. Since $\mathcal{L}^\e = \mathcal{L}_0^\e + \mathcal{A}^\e$,
and $\mathcal{L}_0^\e$ is symmetric and $\mathcal{A}^\e$ is asymmetric
with respect to $\pi\otimes \nu^\e$, we see that $\mathcal{S}^\e
= \mathcal{L}_0^\e$; see Corollary \ref{cor:asymmetry} (at least
at the level of tilt variables) and
arguments in the proof of Lemma \ref{lem:3.1-a}.
The corresponding Dirichlet form is given in the next lemma.
\begin{lem} \label{lem:3.4}
\begin{align*}
\|\Phi\|_{1,\e}^2 := & \lan \Phi, (-\mathcal{L}_0^\e)\Phi\ran_{\pi\otimes\nu^\e}
= \frac12 E^{\pi\otimes\nu^\e}\left[ \int_\R
\left( D\Phi(\cdot;h)*\eta^\e\right)^2(x)dx \right].
\end{align*}
\end{lem}
Before giving the proof of this lemma, we note that the limit
as $\e\downarrow 0$ of $\nu^\e$ for tilt variables
(and therefore defined on $\tilde{\mathcal{C}}$) can be identified
with the Gaussian random measure $\nu$ on $(\R,
\mathcal{B}(\R))$ determined from $dB$. More precisely, under $\nu$, random variables $\{X(A); A\in
\mathcal{B}(\R)\}$ are given and
\begin{enumerate}
\item $X(A) \stackrel{\text{law}}{=} N(0,|A|)$ with $|A|=$
the Lebesgue measure of $A$,
\item If $\{A_i \in \mathcal{B}(\R)\}_{i=1}^n$ are disjoint, then
$\{X(A_i)\}_{i=1}^n$ are independent and $X(\cup_{i=1}^nA_i) = \sum_{i=1}^n
X(A_i)$ a.s.
\end{enumerate}
Such $X(A)$ can be constructed from $X((a,b]) := B(b)-B(a)$
in terms of the two-sided Brownian motion $\{B(x);x\in \R\}$
satisfying, for instance, $B(0)=0$.
\begin{proof}[Proof of Lemma \ref{lem:3.4}]
We first note that $\mathcal{L}_0$ defined as the limit of
$\mathcal{L}_0^\e$ as $\e\downarrow 0$, that is,
$$
\mathcal{L}_0\Phi(h) = \frac12 \int_{\R}
D^2 \Phi(x,x;h) dx + \frac12 \int_\R \partial_x^2h(x) D\Phi(x;h) dx,
$$
is the generator of
the Ornstein-Uhlenbeck process determined by the SPDE
$$
\partial_t h = \frac12 \partial_{x}^2 h + \dot{W}(t,x),\quad x \in \R,
$$
and $\pi\otimes\nu$ is reversible under the wrapped process
$g(t,x)$ of $h(t,x)$ so that it is reversible under $\mathcal{L}_0$.
It is easy to see that
\begin{equation} \label{eq:Lem3-1-a}
\lan \Psi, (-\mathcal{L}_0)\Phi\ran_{\pi\otimes\nu}
= \frac12 E^{\pi\otimes\nu}\left[ \int_\R
D\Psi(x;h)D\Phi(x;h)dx \right].
\end{equation}
Now, for a given $\Phi$, we set $\tilde\Phi^\e(h) :=
\Phi(h*\eta^\e)$ and take $\tilde\Phi^\e$ and $\tilde\Psi^\e$
in place of $\Phi$ and $\Psi$, respectively, in \eqref{eq:Lem3-1-a}.
Then, noting that
\begin{align*}
D \tilde\Phi^\e(x;h) & = D\Phi(\cdot;h*\eta^\e)*\eta^\e(x) \\
D^2 \tilde\Phi^\e(x_1,x_2;h) & = D^2\Phi(\cdot,\cdot;h*\eta^\e)*(\eta^\e)^{\otimes 2}(x_1,x_2),
\end{align*}
we can show that
$
\mathcal{L}_0 \tilde\Phi^\e(h)
= \mathcal{L}_0^\e \Phi(h*\eta^\e)
$
and therefore
$
\lan \tilde\Psi^\e, (-\mathcal{L}_0) \tilde\Phi^\e\ran_{\pi\otimes\nu}
= \lan \Psi, (-\mathcal{L}_0^\e) \Phi\ran_{\pi\otimes\nu^\e}
$
by the change of variables. On the other hand, the right hand
side of \eqref{eq:Lem3-1-a} with $\tilde\Phi^\e$ and
$\tilde\Psi^\e$ in place of $\Phi$ and $\Psi$, respectively,
is rewritten as
\begin{align*}
\frac12 & E^{\pi\otimes\nu}\left[ \int_\R
D\Psi(\cdot;h*\eta^\e)*\eta^\e(x)
D\Phi(\cdot;h*\eta^\e)*\eta^\e(x) dx \right] \\
& = \frac12 E^{\pi\otimes\nu^\e}\left[ \int_\R
D\Psi(\cdot;h)*\eta^\e(x)
D\Phi(\cdot;h)*\eta^\e(x) dx \right],
\end{align*}
by the change of variables again.
This concludes the proof of the lemma.
\end{proof}
The basic tool of the proof of Theorem \ref{thm:3.1} is the
Wiener-It\^o expansion.
Recall that the multiple Wiener integral of order $n\ge 1$ with
a kernel $\fa_n \in \hat{L}^2(\R^n)$, i.e.\ $\fa_n \in L^2(\R^n)$
and symmetric in $n$-variables, is defined by
\begin{align*}
I(\fa_n) & = \frac1{n!} \int_{\R^n} \fa_n(x_1,\ldots,x_n) dB(x_1)\cdots dB(x_n) \\
& = \int_{-\infty}^\infty dB(x_1) \int_{-\infty}^{x_1} dB(x_2)
\cdots \int_{-\infty}^{x_{n-1}} \fa_n(x_1,\ldots,x_n) dB(x_n),
\end{align*}
where $B$ is the two-sided Brownian motion on $\R$ introduced
above.
Set $\mathcal{H}_n = \{I(\fa_n) \in L^2(\tilde{\mathcal{C}},\nu);
\fa_n \in \hat{L}^2(\R^n)\}$ for $n\ge 1$ and $\mathcal{H}_0
= \{\text{const}\}$. Then, the well-known
Wiener-It\^o (Wiener chaos) expansion of $\Phi \in \mathcal{H}
:= L^2(\tilde{\mathcal{C}},\nu)$ is given by
\begin{equation}\label{eq:11-b}
\Phi = \sum_{n=0}^\infty I(\fa_n) \in \bigoplus_{n=0}^\infty \mathcal{H}_n,
\end{equation}
with some $\fa_0\in\R$ and $\fa_n\in \hat{L}^2(\R^n)$,
where $I(\fa_0)=\fa_0$, and
\begin{align} \label{eq:2.6}
\|\Phi\|_{L^2(\nu)}^2
= \sum_{n=0}^\infty \| I(\fa_n) \|_{L^2(\nu)}^2
= \sum_{n=0}^\infty \frac1{n!} \|\fa_n \|_{L^2(\R^n)}^2
\end{align}
holds because of the orthogonality and then by It\^o isometry.
The expansion \eqref{eq:11-b} identifies $\Phi \in
L^2(\tilde{\mathcal{C}},\nu)$ with the element
$
\varphi = \{\varphi_n\}_{n=0}^\infty \in
\bigoplus_{n=0}^{\infty} \hat{L}^2( \mathbb{R}^n)
$
of the symmetric Fock space. The reason to do this is that it
gives an explicit representation of $D$: $D\Phi(x)$ has
representation $\{D\varphi_n\}_{n=1}^\infty$
where $D\varphi_n \in \hat{L}^2( \mathbb{R}^{n-1})$ is given by
\begin{align}\label{five}
D\varphi_n(x; x_1,\ldots,x_{n-1})
& = -\frac1n \sum_{i=1}^n \partial_i \varphi_n
( x_1,\ldots,x_{i-1}, x, x_i, \ldots, x_{n-1})\\
& = - \partial_1 \varphi_n
(x, x_1,\ldots,x_{n-1}). \notag
\end{align}
The factor $\tfrac1n$ arises when we replace $\tfrac1{n!}$
with $\tfrac1{(n-1)!}$, and the second equality is due to the
symmetry of $\fa_n$.
The next task toward the proof of Theorem \ref{thm:3.1}
is to express the norm $\|\Phi\|_{1,\e}$ of
$\Phi\in L^2(\tilde{\mathcal{C}},\nu)$ in terms of its Wiener
chaos expansion \eqref{eq:11-b}.
\begin{lem} \label{lem:3.5}
For $\Phi\in L^2(\tilde{\mathcal{C}},\nu)$,
$$
\|\Phi\|_{1,\e}^2 = \frac12 \sum_{n=0}^\infty \frac1{n!}
\int_{\R^{n+1}} \big(D\fa_{n+1}(x;x_1,\ldots,x_n)*
(\eta^\e)^{\otimes (n+1)}\big)^2 dxdx_1\cdots dx_n.
$$
\end{lem}
\begin{proof}
Lemma \ref{lem:3.4} applied for a function $\Phi$ of tilt variables
gives
\begin{align*}
\|\Phi\|_{1,\e}^2
= \frac12 \int_\R E^{\nu}\left[
\left( D\Phi(\cdot;B*\eta^\e)*\eta^\e\right)^2(x)\right] dx.
\end{align*}
Here, we see that
\begin{align*}
& \left( D\Phi(\cdot;B*\eta^\e)*\eta^\e\right)(x)\\
& = \sum_{n=1}^\infty \frac1{(n-1)!} \int_{\R^{n-1}}
D\fa_n(x;x_1,\ldots,x_{n-1})*(\eta^\e)^{\otimes n}
dB(x_1)\cdots dB(x_{n-1}).
\end{align*}
Therefore, the conclusion follows from \eqref{eq:2.6}.
\end{proof}
We are now almost ready to start the proof of Theorem \ref{thm:3.1}.
But, before starting, we slightly extend Lemma 2.4 of \cite{KLO},
p.48 stated for temporally homogeneous functions
$V(x)$ to more general temporally inhomogeneous $V(s,x)$
as follows. Recall that this lemma holds generically under the stationary situation:
$L$ is the generator of a process $X_t$, $\pi$ is its invariant probability
measure, $S=(L+L^*)/2$ and semi-norms $\|\cdot\|_{-1}$ and $\|\cdot\|_1$ are
defined based on the operator $S$: $\|f\|_1^2 = E^\pi[f\cdot (-S)f]$ and
$\|f\|_{-1}^2 = \sup_g \{2E^\pi[fg]-\|g\|_1^2\}$. This extension is actually
needed only for the proof of Lemma \ref{cor:3.10}, and not for that
of Theorem \ref{thm:3.1}.
\begin{lem} \label{lem:3.6-a}
For $V=V(s,x)$,
$$
E^\pi\left[ \sup_{0\le t \le T} \left(\int_0^tV(s,X_s)ds\right)^2\right]
\le 24 \int_0^T \|V(s,\cdot)\|_{-1}^2ds.
$$
\end{lem}
\begin{proof}
We give the outline of the proof. For $f=f(t,x)$, let $M_t$ be
the martingale
\begin{equation} \label{eq:Dynkin-1}
M_t = f(t,X_t)-f(0,X_0)-\int_0^t \left(\frac{\partial}{\partial s} +L\right) f(s,X_s)ds.
\end{equation}
Then, we have
\begin{equation} \label{eq:Dynkin-2}
E^\pi[M_t^2] = 2\int_0^t \|f(s,\cdot)\|_1^2ds.
\end{equation}
In fact, the proof of \eqref{eq:Dynkin-2} is similar to that of (2.16) in \cite{KLO}, p.47.
Note that, because of the temporal inhomogeneity in our situation,
two terms $E^\pi[f(t,X_t)^2]- E^\pi[f(0,X_0)^2]$ and
$E^\pi [\int_0^t \frac{\partial}{\partial s} f(s,X_s)^2ds] (=
\int_0^t \frac{\partial}{\partial s} E^\pi [f(s,X_s)^2]ds)$ appear, but these terms
just cancel.
For given $V(s,\cdot)$, let us take $f(s,\cdot)$ such that
$E^\pi[(Sf(s,\cdot)-V(s,\cdot))^2] \le \de$ and $\|f(s,\cdot)\|_1 \le
\|V(s,\cdot)\|_{-1} +\de$ for $\de>0$ and $s\in [0,T]$,
and define $M_t$ as in \eqref{eq:Dynkin-1} and $M_t^-$ by
$$
M_t^- = f(T-t,X_{T-t})-f(T,X_T)-\int_0^t \left(-\frac{\partial}{\partial s} +L^*\right)
f(T-s,X_{T-s})ds, \quad t \in [0,T].
$$
Then, by a simple computation, we have
\begin{align*}
M_t+M_T^--M_{T-t}^- & = - \int_0^t (L+L^*)f(s,X_s)ds \\
& =-2 \int_0^t \{V(s,X_s) - R(s,X_s)\} ds,
\end{align*}
where $R:=V-Sf$ satisfies $E^\pi[R^2(s,\cdot)]\le \de$.
This combined with \eqref{eq:Dynkin-2} implies the concluding estimate as in \cite{KLO}.
\end{proof}
Due to this lemma, we have the bound:
\begin{align} \label{eq:thm3-1-1}
& E^{\pi\otimes\nu^\e}
\left[ \sup_{0\le t \le T} \left\{ \int_0^t \hat A^\e(\fa(s,\cdot),Y^\e(s))ds \right\}^2
\right] \\
& \qquad \le 24 T \sup_{\Phi\in L^2(\pi\otimes\nu^\e)}
\left\{ 2E^{\pi\otimes\nu^\e}
\left[ \hat A^\e(\fa,Y) \Phi \right] - \|\Phi\|_{1,\e}^2\right\}.
\notag
\end{align}
In fact, the reason we introduced the wrapped process mostly
lies in applying this bound.
For $\Phi = \Phi(h(\rho),\nabla h)\in L^2(\pi\otimes\nu^\e)$,
we can rewrite
\begin{align} \label{eq:3.14-b}
2 E^{\pi\otimes\nu^\e}
\left[ \hat A^\e(\fa,Y) \Phi \right]
= E^{\pi}\left[ Y_\rho E^{\nu^\e}
\left[ B^\e(\fa,Y) \Phi(h(\rho),\nabla h) \right] \right],
\end{align}
where
\begin{align*}
B^\e(x,Y) & = \left\{\left(\frac{\partial_x Y}Y \right)^2
*\eta_2^\e(x) - \left(\frac{\partial_x Y}Y \right)^2(x)
-\frac1{12} \right\} \frac{Y(x)}{Y_\rho}
\left(= \frac{2\hat A^\e(x,Y)}{Y_\rho} \right), \\
B^\e(\fa,Y) & = \int_\R B^\e(x,Y) \fa(x) dx,
\end{align*}
and recall that $Y_\rho\equiv e^{h(\rho)}$,
$h(x)=\log Y(x)$ is defined by \eqref{eq:3.5-c}.
The integration of $\Phi(h(\rho),\nabla h)$ under $\nu^\e$ is
performed in $\nabla h$ by fixing $h(\rho)$. Note that
$B^\e(x,Y)$ is a tilt variable, though $\hat A^\e(x,Y)$ is not.
The bound \eqref{eq:thm3-1-1} reduces the equilibrium
dynamic problem into a static problem.
The key for the proof of Theorem \ref{thm:3.1} is the following static bound:
\begin{prop} \label{prop:Phi}
For $\Phi = \Phi(\nabla h) \in L^2(\tilde{\mathcal{C}},\nu)$
such that $\|\Phi\|_{1,\e}<\infty$, and $\fa\in C_0(\R)$ satisfying the
condition of Theorem \ref{thm:3.1}, we have that
\begin{equation} \label{eq:B-1}
\left| E^{\nu^\e}\left[ B^\e(\fa,Y) \Phi\right] \right|
\le C(\fa) \sqrt{\e} \|\Phi\|_{1,\e},
\end{equation}
for all $0 < \e <1 \wedge \left(\frac14 \dist(\rm{supp}\,\fa,\rm{supp}\,\rho)\right)$
with some positive constant $C(\fa)$, which depends only on
$\|\fa\|_\infty$ and the size of $\rm{supp}\,\fa$. In particular, taking $\Phi=1$,
we see $E^{\nu^\e}\left[ B^\e(\fa,Y) \right] =0$.
\end{prop}
As we pointed out, we will work with the height processes
and not with the Cole-Hopf transformed processes.
In this respect, $\left(\frac{\partial_x Y}Y \right)^2 -\xi^\e$
is transformed back to $(\partial_x h)^2 -\xi^\e$. However,
recalling that $\partial_x h=\partial_x(B*\eta^\e)$ under $\nu^\e$
in the stationary situation, by It\^o's formula, we have that
\begin{align}\label{eq:6.1}
(\partial_xh)^2 & = \left\{ \int_\R \eta^\e(x-y)dB(y)\right\}^2 \\
& = \Psi^\e(x) + \int_\R \eta^\e(x-y)^2dy = \Psi^\e(x)+\xi^\e, \notag
\end{align}
where $\Psi^\e(x)$ is a Wiener functional of second order given by
\begin{align} \label{eq:2.Psi}
\Psi^\e(x) &
= \int_{\R^2} \eta^\e(x-x_1)\eta^\e(x-x_2)dB(x_1)dB(x_2)\\
&\equiv 2 \int_{x_1<x_2} \eta^\e(x-x_1)\eta^\e(x-x_2)dB(x_1)dB(x_2). \notag
\end{align}
Therefore, $\left(\frac{\partial_x Y}Y \right)^2 -\xi^\e= \Psi^\e(x)$
and $A^\e(x,Y)$ is rewritten as
\begin{equation} \label{eq:10-b}
\frac12 Y(x) \{\Psi^\e*\eta_2^\e(x) - \Psi^\e(x)\}.
\end{equation}
\begin{rem} \label{rem:3.1}
{\rm (1)} The term $\Psi^\e*\eta_2^\e(x) - \Psi^\e(x)$
does not converge in a strong sense. In fact, if we take
$\eta(x) = \tfrac1{\sqrt{2\pi}} e^{-x^2/2}$ for simplicity, then
an explicit computation shows that
$$
E^{\nu^\e} [\{\Psi^\e*\eta_2^\e(x) - \Psi^\e(x)\}^2]
= \frac1{\pi \e^2} \left( \frac1{4\sqrt{5}}- \frac1{2\sqrt{3}}
+ \frac14 \right).
$$
{\rm (2)} We easily have that
$$
E^{\nu^\e} [\Psi^\e(x)^2] = \frac1{\e^2} \eta_2(0)^2.
$$
Comparing with {\rm (1)}, we see that taking the difference
does not improve the magnitude in $\e$.\\
{\rm (3)} The term $\Psi^\e*\eta_2^\e(x) - \Psi^\e(x)$ converges
to $0$ in a weak sense. More precisely, for every
$\Phi\in L^2(\nu)$ whose second order Wiener chaos has a
continuous kernel $\fa_2\in C(\R^2)\cap \hat{L}^2(\R^2)$,
we have that
$$
\lim_{\e\downarrow 0}
E^{\nu^\e} [\{\Psi^\e*\eta_2^\e(x) - \Psi^\e(x)\}\Phi]=0.
$$
However, this is not sufficient to analyze
the limit of \eqref{eq:10-b} because of the extra $Y(x)$. \\
{\rm (4)} If one could have a bound on $\Psi^\e*\eta_2^\e(x)
- \Psi^\e(x)$ in a Sobolev norm $\|\cdot\|_{H_r^{-\a}}$ with
possibly $\a< 1/2$,
then one might be able to control the limit of \eqref{eq:10-b}.
However, we only have that
$$
E^{\nu^\e} \left[\left| \int_\R \{\Psi^\e*\eta_2^\e(x) - \Psi^\e(x)\}
\psi(x)dx \right|\right] \le C \e \|\partial_x^2\psi\|_\infty,
$$
for every $\psi\in C_b^2(\R)$. This (with interpolation) roughly
implies
$$
E^{\nu^\e} \left[\| \Psi^\e*\eta_2^\e(x) - \Psi^\e(x)
\|_{H^{-\a}_r(\R)} \right] \le C \e^{\a-1},
$$
which is expected to converge to $0$ only if $\a>1$.
Under the multiplication of $Y^\e(x)$, which is roughly in
$C^{\frac12-\de}$ (uniformly in $\e$), the convergence of
\eqref{eq:10-b} to $0$ cannot be expected. \\
{\rm (5)} Proposition \ref{prop:Phi}
shows that $\|B^\e(\fa,Y)\|_{-1,\e} \le C(\fa)\sqrt{\e}$.
For its $L^2$-norm, we only have $\|B^\e(\fa,Y)\|_{L^2(\Om)}
\le C(\fa)\e^{-1/2}$.
Indeed, in the proof of the proposition stated below, we can estimate
$|f_{n+2}-f_{n+2}| \le |f_{n+2}| + |f_{n+2}|$ to avoid to have their derivatives,
but we lose the factor $\e$ in doing so.
\end{rem}
The results mentioned in Remark \ref{rem:3.1} are
not useful in our situation because of the extra term $Y(x)$.
Since $Y(x)$ is not a tilt variable, we need to consider
$\frac{Y(x)}{Y_\rho}$ instead as in $B^\e(x,Y)$.
Once Proposition \ref{prop:Phi} is shown, \eqref{eq:3.14-b} with $\Phi\equiv 1$
implies that $E^{\pi\otimes\nu^\e}\left[ \hat A^\e(\fa,Y) \right] =0$.
This means that we may assume $\Phi\in L^2(\pi\otimes\nu^\e)$
in the right hand side of \eqref{eq:thm3-1-1} satisfies
$E^{\pi\otimes\nu^\e}\left[ \Phi \right] =0$. For such $\Phi$,
we have that
\begin{align*}
\left| 2E^{\pi\otimes\nu^\e}
\left[ \hat A^\e(\fa,Y) \Phi \right] \right|
& \le E^\pi\left[ Y_\rho C(\fa) \sqrt{\e} \|\Phi(h(\rho),\cdot)\|_{1,\e}
\right] \\
& \le e \, C(\fa) \sqrt{\e} \int_0^1 d\pi \left\{\tfrac12
E^{\nu^\e}\left[ \int_\R
\left( D\Phi(\cdot;h(\rho),\nabla h)*\eta^\e\right)^2(x)dx
\right] \right\}^{1/2} \\
& \le e \, C(\fa) \sqrt{\e} \left\{\tfrac12
E^{\pi\otimes\nu^\e}\left[ \int_\R
\left( D\Phi(\cdot;h(\rho),\nabla h)*\eta^\e\right)^2(x)dx
\right] \right\}^{1/2} \\
& \le e\, C(\fa) \sqrt{\e} \|\Phi\|_{1,\e},
\end{align*}
where the operator $D$ acts only on the tilt variable $\nabla h$.
We have used Proposition \ref{prop:Phi}, Lemma \ref{lem:3.4}
and $Y_\rho\in [1,e]$ for the second line, Schwarz's inequality for
the third line and Lemma \ref{lem:3.6} stated below for the
fourth line. Therefore, the right hand side of \eqref{eq:thm3-1-1}
is bounded by
$$
24 T \sup_{\a\in \R} \{ e C(\fa) \sqrt{\e}\a-\a^2\}
= 24 T (\tfrac{e}2\, C(\fa) \sqrt{\e})^2,
$$
in which we write $\a = \|\Phi\|_{1,\e}$. This tends to $0$ as
$\e\downarrow 0$, and concludes the proof of Theorem
\ref{thm:3.1}.
\begin{lem} \label{lem:3.6}
For $\Phi=\Phi(h(\rho),\nabla h)$ such that
$E^{\pi\otimes\nu^\e}\left[ \Phi \right] =0$,
\begin{align*}
\|\Phi\|_{1,\e}^2 = & \frac{\xi_\rho^\e}2 E^{\pi\otimes\nu^\e}\left[
\left( \frac{\partial\Phi}{\partial h(\rho)}(h(\rho),\nabla h)\right)^2
\right] \\
& \quad + \frac12 E^{\pi\otimes\nu^\e}\left[ \int_\R
\left( D\Phi(\cdot;h(\rho),\nabla h)*\eta^\e\right)^2(x)dx \right].
\end{align*}
In the above formula, $D$ acts only on the tilt variables and
$\xi_\rho^\e$ is defined in the proof of Lemma \ref{lem:3.1-a}.
\end{lem}
\begin{proof}
Denote $D$ acting on $h=(h(\rho),\nabla h)$ by $\tilde D$ for
distinction. Then, it can be expressed as
$$
\tilde D \Phi(x;h(\rho),\nabla h) =
\frac{\partial\Phi}{\partial h(\rho)}(h(\rho),\nabla h) \rho(x)
+ D\Phi(x;h(\rho),\nabla h),
$$
so that Lemma \ref{lem:3.4} implies that
\begin{align*}
\|\Phi\|_{1,\e}^2 = \frac12 E^{\pi\otimes\nu^\e}\left[
\int_\R \left\{ \frac{\partial\Phi}{\partial h(\rho)}(h(\rho),\nabla h)
\rho*\eta^\e(x)
+ D\Phi(\cdot;h(\rho),\nabla h)*\eta^\e(x)\right\}^2dx \right].
\end{align*}
We expand the square inside the integration, then the cross term
becomes
\begin{align} \label{eq:cross}
E^{\pi\otimes\nu^\e}\left[
\frac{\partial\Phi}{\partial h(\rho)}(h(\rho),\nabla h)
\int_\R \rho*\eta^\e(x)
\big( D\Phi(\cdot;h(\rho),\nabla h)*\eta^\e\big)(x)dx \right] =0
\end{align}
and this shows the conclusion. \eqref{eq:cross} is shown,
first for $\Phi$ of the form $\Phi = \sum_{i=1}^\ell f_i(h(\rho))
\Phi_i(\nabla h)$, where we choose $\{f_i(a)=\sqrt{2}\sin
\pi i a \}_{i=1}^\infty$ which is a complete orthonormal system
of $L^2([0,1],\pi)$. Since $E^{\pi\otimes\nu^\e}\left[ \Phi \right] =0$,
we may assume $E^\pi[f_i]=0$ so that $i$ are even.
Indeed for such $\Phi$, the left hand side of
\eqref{eq:cross} can be rewritten as
\begin{align*}
\sum_{i,j=2: \text{ even}}^\ell \int_0^1 f_i' f_j d\pi \,
E^{\nu^\e}\left[ \Phi_i \int_\R \rho*\eta^\e(x)
\big( D\Phi_j(\cdot;\nabla h)*\eta^\e\big)(x)dx \right].
\end{align*}
However, we easily see that
$
\int_0^1 f_i' f_j d\pi =0
$
for all even $i, j \ge 2$ and this proves \eqref{eq:cross}. The general
$\Phi \in L^2(\pi\!\otimes\!\nu^\e)$ can be
approximated by the functions of the above form.
\end{proof}
Only the proof of Proposition \ref{prop:Phi} is left. Before
giving it, we recall the diagram formula in order to compute
the second order chaos in the products of two
Wiener functionals $\frac{Y(x)}{Y_\rho}$ and $\Phi$.
Let $n_1, \ldots, n_m \in\N$ be given. We call $\Ga$ the set of all
diagrams $\ga$, which are collections of (undirected) edges connecting
vertices in $V :=\{(j,\ell); \ell=1,\ldots, m, j=1,\ldots, n_\ell\}$,
in such a way that each edge in $\ga$ connects two vertices $(j_1,\ell_1)$
and $(j_2,\ell_2)$ only when $\ell_1 \not= \ell_2$ and each
vertex is attached to at most one edge. We denote
$N= \sum_{\ell=1}^m n_\ell$ and the number of edges in $\ga$ by
$|\ga|$. We define the function $\fa_\ga\in \hat{L}^2(\R^{N-2|\ga|})$
as follows: We first introduce a function $\fa$ of $N$ variables $\{x_{j,\ell}\}$ by
$$
\fa(x_{j,\ell}, \ell=1,\ldots, m, j=1,\ldots, n_\ell)
:= \prod_{\ell=1}^m \fa_\ell(x_{j,\ell}, j=1,\ldots, n_\ell).
$$
We call the variables $\{x_{j,\ell}, \ell=1,\ldots, m, j=1,\ldots, n_\ell\}$
simply as $\{x_1,\ldots,x_N\}$ and call $\fa$ again its symmetrization.
Then, $\fa_\ga$ is defined from $\fa$ by truncating the last
$2|\ga|$-variables:
$$
\fa(x_1,\ldots,x_{N-2|\ga|}) := \int_{\R^{|\ga|}}
\fa(x_1,\ldots,x_{N-2|\ga|}, p_1,p_1, \ldots. p_{|\ga|},p_{|\ga|})
dp_1\cdots dp_{|\ga|}.
$$
Then, we have the following diagram formula; see Major
\cite{Ma}, Section 5, under a slightly different setting.
\begin{lem}
For $\fa_1\in \hat{L}^2(\R^{n_1}), \ldots, \fa_m\in \hat{L}^2(\R^{n_m})$
with $n_1, \ldots, n_m \ge 1$, we have
$$
I(\fa_1) \cdots I(\fa_m) = \sum_{\ga\in \Ga} \frac{(N-2|\ga|)!}
{n_1! \cdots n_m!} I (\fa_\ga).
$$
\end{lem}
We are now ready to give the proof of Proposition \ref{prop:Phi}.
\begin{proof}[Proof of Proposition \ref{prop:Phi}]
We first notice that, under $\nu$, $\frac{Y(x)}{Y_\rho}$ has an expression:
$$
\frac{Y(x)}{Y_\rho} = e^{B(x)-\int_\R B(y)\rho(y)dy},
$$
and the exponent can be rewritten as
$$
B(x)-\int_\R B(y)\rho(y)dy= \int_\R \phi_x(u)dB(u),
$$
where
\begin{equation} \label{eq:3.17-x}
\phi_x(u) = 1_{(-\infty,x]}(u) + \th(u), \quad \th(u) = - \int_u^\infty\rho(y)dy.
\end{equation}
Note that, from the condition of $\rho$, $\phi_x(\cdot)$ has a compact support:
supp $\phi_x \subset [x\wedge(-1),x\vee 1]$, $|\phi_x(u)|\le 1$, has
a jump only at $u=x$ and $\th$ is smooth. Therefore, $\frac{Y(x)}{Y_\rho}$ has
the following Wiener-It\^o expansion under $\nu$:
\begin{align} \label{eq.WIexp-Y}
\frac{Y(x)}{Y_\rho}
= e^{a(x)} \left\{ 1 + \sum_{n=1}^\infty \frac1{n!}\int_{\R^n}
\phi_x^{\otimes n} (u_1,\ldots,u_n) dB(u_1) \cdots dB(u_n)\right\},
\end{align}
where $a(x) = \tfrac12 \int_\R \phi_x^2(u)du$. In fact, one can apply the well-known
result for the expansion of exponential martingales
written, e.g., in \cite{KS}, p.167 for $M_t= \int_{-\infty}^t\phi_x(u)dB(u)$,
and letting $t\to\infty$.
Since $\Psi^\e$ and $\Psi^\e*\eta^\e$ are both second order Wiener chaoses,
to compute the expectation
\begin{equation}\label{eq:Exp}
E^{\nu^\e} \left[\{\Psi^\e*\eta_2^\e(x)
- \Psi^\e(x)\} \frac{Y(x)}{Y_\rho} \Phi \right],
\end{equation}
what we need to obtain is the kernel $\bar\fa_2(x_1,x_2)$
of the second order Wiener chaos in the product $\Phi\cdot
\frac{Y(x)}{Y_\rho}$. Denoting the kernel of the $n$th order Wiener
chaos in $\frac{Y(x)}{Y_\rho}$ except the factor $e^{a(x)}$
by $\psi_n(u_1,\cdots,u_n) =
\phi_x^{\otimes n}(u_1,\cdots,u_n)$, in the expansion of the product
\begin{equation} \label{eq:3.product}
\Phi\cdot \frac{Y(x)}{Y_\rho} = e^{a(x)}\left( \sum_{m_1=0}^\infty I(\fa_{m_1})\right)
\left( \sum_{m_2=0}^\infty I(\psi_{m_2})\right)
= e^{a(x)}\sum_{m_1,m_2=0}^\infty I(\fa_{m_1}) I(\psi_{m_2}),
\end{equation}
we apply the diagram formula to get the explicit formula for
$\bar\fa_2(x_1,x_2)$:
\begin{align*}
\bar\fa_2(x_1,x_2)&= e^{a(x)} \sum_{n=0}^\infty
\begin{pmatrix} n+2 \\ 2 \end{pmatrix} n!
\times
\frac{2!}{(n+2)!n!} \int_{\R^n} \fa_{n+2}(\uu,x_1,x_2)
\psi_n(\uu) d\uu \\
&\;\; + e^{a(x)} \sum_{n=0}^\infty
\begin{pmatrix} n+2 \\ 2 \end{pmatrix} n!
\times
\frac{2!}{(n+2)!n!} \int_{\R^n} \fa_n(\uu)
\psi_{n+2}(\uu,x_1,x_2) d\uu \\
&\;\; +e^{a(x)} \sum_{n=0}^\infty (n+1)^2 n!\times
\frac{2!}{(n+1)!(n+1)!} \int_{\R^n} \fa_{n+1}(\uu,x_1)
\psi_{n+1}(\uu,x_2) d\uu \\
&\;\; +e^{a(x)} \sum_{n=0}^\infty (n+1)^2 n!\times
\frac{2!}{(n+1)!(n+1)!} \int_{\R^n} \fa_{n+1}(\uu,x_2)
\psi_{n+1}(\uu,x_1) d\uu \\
&=: \bar\fa_2^{(1)}(x_1,x_2) + \bar\fa_2^{(2)}(x_1,x_2) +
\bar\fa_2^{(3)}(x_1,x_2) + \bar\fa_2^{(4)}(x_1,x_2),
\end{align*}
where we denote $\uu = (u_1,\ldots,u_n)$ and $d\uu = du_1\cdots du_n$.
Note that, to obtain the second order term, $(m_1,m_2)$ which
we need to take care are only of the forms $\{(n+2,n), (n,n+2),
(n+1,n+1), n \ge 0\}$, in which case $N= 2n+2$, and we may only
consider $\ga$ satisfying $N-2|\ga| = 2$, i.e., $|\ga|=n$. For example,
when $(m_1,m_2) = (n+2,n)$, the prefactor in the diagram formula
becomes as above, since $|\Ga| = \begin{pmatrix} n+2 \\ 2 \end{pmatrix}n!$,
and when $(m_1,m_2)=(n+1,n+1)$, $|\Ga|=(n+1)^2n!$
In the rest, we will show that the contributions of $\bar\fa_2^{(1)}$,
$\bar\fa_2^{(3)}$, $\bar\fa_2^{(4)}$ to the expectation \eqref{eq:Exp}
are small, while that of $\bar\fa_2^{(2)}$ exactly cancels with
$\frac1{12} E^{\nu^\e} \left[\frac{Y(x)}{Y_\rho} \Phi \right]$,
when integrated with a test function $\fa=\fa(x)$.
Under $\nu^\e$, the kernels $\fa_n$ and $\psi_n$ are replaced by
$\fa_n*(\eta^\e)^{\otimes n}$ and $\psi_n*(\eta^\e)^{\otimes n}$,
respectively, where $\fa_n*(\eta^\e)^{\otimes n}$ is defined by
$$
\fa_n*(\eta^\e)^{\otimes n}(\uu)
:= \int_{\R^n} \fa_n(\underline{v}) \eta^\e(u_1-v_1)\cdots
\eta^\e(u_n-v_n) d\underline{v},
$$
where $\underline{v}= (v_1,\ldots,v_n)$ and $d\underline{v}=
dv_1\cdots dv_n$. Recall that $\Psi^\e*\eta_2^\e(x) - \Psi^\e(x)$
is a second order Wiener chaos with the kernel:
\begin{equation}\label{eq:ker-a}
2 \left\{ \int_\R \eta^\e(y-x_1) \eta^\e(y-x_2) \eta_2^\e(x-y) dy
- \eta^\e(x-x_1) \eta^\e(x-x_2) \right\}.
\end{equation}
Then, the contribution of the first term $\bar\fa_2^{(1)}$ to the
expectation \eqref{eq:Exp}
is given by, neglecting the factor $e^{a(x)} \sum_{n=0}^\infty
\frac1{n!}$ ($2$ in \eqref{eq:ker-a} cancels with $\frac12$
appearing in \eqref{eq:2.6} for $n=2$),
\begin{align*}
& \int_{\R^2}dx_1dx_2 \left\{
\int_{\R^n} \left( \fa_{n+2}*(\eta^\e)^{\otimes (n+2)} \right)
(\uu,x_1,x_2) \left(\phi_x^{\otimes n}*(\eta^\e)^{\otimes n}\right)
(\uu) d\uu \right\} \\
& \qquad \times
\left\{ \int_\R \eta^\e(y-x_1) \eta^\e(y-x_2) \eta_2^\e(x-y) dy
- \eta^\e(x-x_1) \eta^\e(x-x_2) \right\} \\
& = \int_{\R^n} \phi_x^{\otimes n}(\uu) d\uu \left\{ \int_\R \left(\fa_{n+2}*(\eta_2^\e)^{\otimes (n+2)} \right)
(\uu,y,y) \eta_2^\e(x-y)dy \right. \\
& \qquad\qquad\qquad\qquad\qquad \left. \phantom{\int_{\R^n}}
- \left(\fa_{n+2}*(\eta_2^\e)^{\otimes (n+2)} \right)(\uu,x,x)\right\}.
\end{align*}
In the above computation, we first move $\eta^\e$ in
$\phi_x^{\otimes n}*(\eta^\e)^{\otimes n}$ to $\fa_{n+2}*(\eta^\e)^{\otimes (n+2)}$, which gives $(\fa_{n+2}*(\eta_2^\e)^{\otimes n}*(\eta^\e)^{\otimes 2})(\uu,x_1,x_2)$. Then, we integrate in $x_1$ and $x_2$ and obtain the above formula.
Therefore, the contribution of $\bar\fa_2^{(1)}$ to the
left hand side of \eqref{eq:B-1} is given by
$
\sum_{n=0}^\infty \frac1{n!} I_n^{(1)},
$
where
\begin{align*}
I_n^{(1)} := & \int_\R \fa(x) e^{a(x)}dx \int_{\R^n} \phi_x^{\otimes n}
(\uu) d\uu\\
& \times \left\{ \int_R (f_{n+2}*(\eta^\e)^{\otimes (n+2)})
(\uu,y,y) \eta_2^\e(x-y)dy - (f_{n+2}*(\eta^\e)^{\otimes (n+2)})
(\uu,x,x) \right\},
\end{align*}
with $f_{n+2} := \fa_{n+2}*(\eta^\e)^{\otimes (n+2)}$.
We have rewritten as $\fa_{n+2}*(\eta_2^\e)^{\otimes (n+2)}=
f_{n+2}*(\eta^\e)^{\otimes (n+2)}$ in the above formula.
The difference in the braces in the right hand side of
$I_n^{(1)}$ can be expressed as the
expectation:
$$
E[f_{n+2}(\uu+U^\e, x+Y^\e+X_1^\e, x+Y^\e+X_2^\e) -
f_{n+2}(\uu+U^\e, x+X_1^\e, x+X_2^\e)],
$$
where $U^\e = (U_i^\e)_{i=1}^n$ is an $\R^n$-valued random
variable with independent components $U_i^\e$ distributed
under $\eta^\e(y)dy$ ($\eta^\e$ in short), $X_1^\e, X_2^\e$
are $\R$-valued random variables distributed under $\eta^\e$,
$Y^\e$ is an $\R$-valued random variable distributed under
$\eta_2^\e$, and all random variables are mutually independent.
We estimate
\begin{align*}
& | f_{n+2}(\uu+U^\e, x+Y^\e+X_1^\e, x+Y^\e+X_2^\e) -
f_{n+2}(\uu+U^\e, x+X_1^\e, x+X_2^\e) | \\
& =\left| \int_0^1 \frac{\partial}{\partial \la}
f_{n+2}(\uu+U^\e, x+\la Y^\e+X_1^\e, x+\la Y^\e+X_2^\e) d\la\right| \\
& =\left| \int_0^1 Y^\e \sum_{k=n+1}^{n+2} \partial_k
f_{n+2}(\uu+U^\e, x+\la Y^\e+X_1^\e, x+\la Y^\e+X_2^\e) d\la
\right|\\
& \le 4\e \int_0^1 \left| \partial_{n+2}
f_{n+2}(\uu+U^\e, x+\la Y^\e+X_1^\e, x+\la Y^\e+X_2^\e) \right|
d\la, \quad \text{ a.s.}.
\end{align*}
Note that supp$\,\eta \subset [-1,1]$ implies
supp$\,\eta_2^\e \subset [-2\e,2\e]$ and therefore
$|Y^\e| \le 2\e$ a.s. We also used the symmetry of $f_{n+2}$
in $(n+2)$-variables. Accordingly, since $\fa\in C_0(\R)$
implies supp $\fa\subset [-K,K]$ with some $K\ge 1$ and
supp $\phi_x \subset [x\wedge(-1),x\vee 1] \subset [-K,K]$
for $x\in [-K,K]$, we have
\begin{align*}
|I_n^{(1)}| & \le 4\e \|\fa e^{a(x)}\|_\infty \int_{-K}^K dx \int_{\R^n}
1_{[-K,K]^n} (\uu) d\uu \\
& \qquad \times \int_0^1 E[|\partial_{n+2}
f_{n+2}(\uu+U^\e, x+\la Y^\e+X_1^\e, x+\la Y^\e+X_2^\e)|] d\la\\
& \le 4\e \|\fa e^{a(x)}\|_\infty \int_{-K-3\e}^{K+3\e} dx \int_{\R^n}
1_{[-K-\e,K+\e]^n} (\uu) d\uu \\
& \qquad \times E[|\partial_{n+2}
f_{n+2}(\uu, x, x+X_2^\e-X_1^\e)|].
\end{align*}
The last line is obtained by putting the integrals in $(x,\uu)$
inside the expectation and then applying the change of variables $\uu'=\uu+U^\e$ and $x'=x+\la Y^\e+X_1^\e$. We enlarge
the domains of the integrations a little bit. Since $X_2^\e-X_1^\e$
is distributed under $\eta_2^\e$, this is equal to
\begin{align*}
4\e \|\fa e^{a(x)}\|_\infty \int_{\R^{n+2}} 1_{[-K-3\e,K+3\e]}(x)
1_{[-K-\e,K+\e]^n} (\uu)
|\partial_{n+2} f_{n+2}(\uu, x, x+y)| d\uu dx \eta_2^\e(y)dy.
\end{align*}
Apply Schwarz's inequality, and we obtain that
\begin{align*}
|I_n^{(1)}| & \le 4\e \|\fa e^{a(x)}\|_\infty
\left\{ \int_{\R^{n+2}} 1_{[-K-3\e,K+3\e]}(x)
1_{[-K-\e,K+\e]^n} (\uu) d\uu dx \eta_2^\e(y)dy \right\}^{1/2} \\
& \qquad \times \left\{ \int_{\R^{n+2}}
|\partial_{n+2} f_{n+2}(\uu, x, x+y)|^2 d\uu dx \eta_2^\e(y)dy
\right\}^{1/2} \\
& \le 4 \sqrt{\e} \|\fa e^{a(x)}\|_\infty \|\eta_2\|_\infty^{1/2}
(2K+6)^{(n+1)/2} \left\{\int_{\R^{n+2}}
|\partial_{n+2} f_{n+2}(\uu, x, y)|^2 d\uu dx dy
\right\}^{1/2},
\end{align*}
since $0<\e<1$. We have used a rough estimate: $|\eta_2^\e(y)| \le
\|\eta_2\|_\infty/\e$. Thus, we obtain that
\begin{align*}
\sum_{n=0}^\infty \frac1{n!} |I_n^{(1)}| &
\le C \sqrt{\e} \sum_{n=0}^\infty \frac{ (2K+6)^{(n+1)/2}}{n!}
\left\{\int_{\R^{n+2}}
|\partial_{n+2} f_{n+2}|^2 d\ux \right\}^{1/2} \\
& \le C \sqrt{\e} \left\{\sum_{n=0}^\infty \frac{ (n+1)^2(2K+6)^{(n+1)}}{(n+1)!}
\right\}^{1/2} \left\{\sum_{n=0}^\infty \frac1{(n+1)!}
\int_{\R^{n+2}} |\partial_{n+2} f_{n+2}|^2 d\ux \right\}^{1/2} \\
& \le C' \sqrt{\e} \|\Phi\|_{1,\e},
\end{align*}
by Schwarz's inequality and Lemma \ref{lem:3.5}.
Next, the contribution of the second term $\bar\fa_2^{(2)}$ to the
expectation \eqref{eq:Exp} is given by, again neglecting the factor $e^{a(x)} \sum_{n=0}^\infty\frac1{n!}$,
\begin{align*}
& \int_{\R^2}dx_1dx_2 \left\{
\int_{\R^n} \left( \fa_{n}*(\eta^\e)^{\otimes n} \right)
(\uu) \left(\phi_x^{\otimes (n+2)}*(\eta^\e)^{\otimes (n+2)}\right)
(\uu,x_1,x_2) d\uu \right\} \\
& \qquad\qquad \times
\left\{ \int_\R \eta^\e(y-x_1) \eta^\e(y-x_2) \eta_2^\e(x-y) dy
- \eta^\e(x-x_1) \eta^\e(x-x_2) \right\} \\
& = \int_{\R^n} \left(f_{n}*(\eta^\e)^{\otimes n} \right)
(\uu) \phi_x^{\otimes n}(\uu) d\uu \\
& \times \left\{ \int_\R
\left(\phi_x^{\otimes 2}*(\eta_2^\e)^{\otimes 2}\right)(y,y)
\eta_2^\e(x-y)dy - \left(\phi_x^{\otimes 2}*(\eta_2^\e)^{\otimes 2}\right)(x,x)
\right\}.
\end{align*}
Here, the last term in the braces can be represented by means of expectations:
\begin{align*}
& \int_\R \left(\phi_x^{\otimes 2}*(\eta_2^\e)^{\otimes 2}\right)(y,y)
\eta_2^\e(x-y)dy - \left(\phi_x^{\otimes 2}*(\eta_2^\e)^{\otimes 2}\right)(x,x)\\
&= E[\phi_x^{\otimes 2}(x+R_1+R_3, x+R_2+R_3)]
- E[\phi_x^{\otimes 2}(x+R_1, x+R_2)],
\end{align*}
where $\{R_1, R_2, R_3\}$ are independent random
variables distributed under $\eta_2^\e$. By expanding
$$
\phi_x^{\otimes 2}= 1_{(-\infty,x]}^{\otimes 2} +1_{(-\infty,x]}\otimes \th
+ \th \otimes 1_{(-\infty,x]}+ \th^{\otimes 2},
$$
the above difference of two expectations can be rewritten as
$$
J_1+2J_2^\e(x)+J_3^\e(x),
$$
where
\begin{align*}
& J_1 = E[1_{(-\infty,0]}^{\otimes 2}(R_1+R_3, R_2+R_3)]
- E[1_{(-\infty,0]}^{\otimes 2}(R_1, R_2)] \\
& J_2^\e(x) = E[1_{(-\infty,0]}(R_1+R_3) \th(x+ R_2+R_3)]
- E[1_{(-\infty,0]}(R_1)\th(x+R_2)] \\
& J_3^\e(x) = E[\th(x+R_1+R_3) \th(x+ R_2+R_3)]
- E[\th(x+R_1)\th(x+R_2)].
\end{align*}
However, we see that ${\rm supp}\, J_2^\e \subset ({\rm supp}\, \rho)^{4\e}
:= \{y\in\R; |y-x|\le 4\e \text{ for some } x\in {\rm supp}\, \rho\}$ and
${\rm supp}\, J_3^\e \subset ({\rm supp}\, \rho)^{4\e}$; recall that
${\rm supp}\, \rho$ is connected and $E[1_{(-\infty,0]}(R_1+R_3)] =
E[1_{(-\infty,0]}(R_1)] = \frac12$ by the symmetry of $R_i$ for $J_2^\e(x)$.
Therefore, from our assumption: $\dist({\rm supp}\, \fa, {\rm supp}\, \rho)>4\e$,
$\fa(x) J_2^\e(x) = \fa(x) J_3^\e(x) =0$
for all $x\in\R$. On the other hand, we have
$$
J_1=\frac1{12}.
$$
In fact, by the symmetry of $\eta_2^\e$,
\begin{align}\label{eq:1/3}
P(R_1+R_3>0, R_2+R_3>0)
& = P(R_1-R_3>0, R_2-R_3>0) \\
& = P \left(R_3 = \min_{i=1,2,3} R_i\right) = \tfrac13, \notag
\end{align}
for the first expectation and the second one is $\frac14$
as easily seen. The constant $\frac1{12}$ is
universal in the sense that it does not depend on the specific
distributions of independent random variables $\{R_1, R_2, R_3\}$
if they are symmetric (and have densities).
Summarizing these, we see that the contribution of $\bar\fa_2^{(2)}$
to the expectation \eqref{eq:Exp}, when multiplied by $\fa$, is given by
\begin{equation}\label{eq:ker-b}
\tfrac1{12}e^{a(x)} \sum_{n=0}^\infty \frac1{n!}
\int_{\R^n} \left(f_{n}*(\eta^\e)^{\otimes n} \right)(\uu)
\phi_x^{\otimes n}(\uu) d\uu.
\end{equation}
However, the series in \eqref{eq:ker-b} just cancels with the expectation
$E^{\nu^\e} \left[\frac{Y(x)}{Y_\rho} \Phi \right]$ divided by $12$. Indeed,
this expectation can be computed by picking up the $0$th order term in the product \eqref{eq:3.product}, and it is again an application of the diagram formula. Indeed, we may take care $(m_1,m_2)$ of the forms $(n,n), n\ge 0$ only, in which case $N=2n$, and may consider only
$\ga$ such that $|\ga|=n$, i.e., diagrams connecting all vertices of the forms $(j_1,\ell_1)$ and $(j_2,\ell_2)$ with $\ell_1\not= \ell_2$. The number of such $\ga$'s is given by $|\Ga|=n!$. Thus, we have
\begin{align*}
E^{\nu^\e} \left[\frac{Y(x)}{Y_\rho} \Phi \right]
& = e^{a(x)} \sum_{n=0}^\infty \frac1{n!}
\int_{\R^n} \left( \fa_{n}*(\eta^\e)^{\otimes n} \right)
(\uu) \left(\phi_x^{\otimes n}*(\eta^\e)^{\otimes n}\right)(\uu) d\uu \\
& = e^{a(x)} \sum_{n=0}^\infty \frac1{n!}
\int_{\R^n} \left(f_{n}*(\eta^\e)^{\otimes n} \right)
(\uu) \phi_x^{\otimes n}(\uu) d\uu,
\end{align*}
which is exactly the same series in \eqref{eq:ker-b} except
the factor $\frac1{12}$.
The contribution from the third term $\bar\fa_2^{(3)}$ to the
expectation \eqref{eq:Exp} is given by, neglecting the factor
$a(x) \sum_{n=0}^\infty \frac{2}{n!}$,
\begin{align*}
& \int_{\R^2}dx_1dx_2 \left\{
\int_{\R^n} \left( \fa_{n+1}*(\eta^\e)^{\otimes (n+1)} \right)
(\uu,x_1) \left( \phi_x^{\otimes (n+1)}*(\eta^\e)^{\otimes (n+1)}\right)
(\uu,x_2) d\uu \right\} \\
& \qquad\qquad \times
\left\{ \int_\R \eta^\e(y-x_1) \eta^\e(y-x_2) \eta_2^\e(x-y) dy
- \eta^\e(x-x_1) \eta^\e(x-x_2) \right\} \\
& = \int_{\R^n} \phi_x^{\otimes n}(\uu) d\uu
\left\{ \int_\R
\left(\fa_{n+1}*(\eta_2^\e)^{\otimes (n+1)} \right)
(\uu,y) \left(\phi_x*\eta_2^\e\right)(y)
\eta_2^\e(x-y)dy \right. \\
& \qquad \qquad\qquad\qquad \left. \phantom{\int_\R }
- \left(\fa_{n+1}*(\eta_2^\e)^{\otimes (n+1)} \right)
(\uu,x) \left(\phi_x*\eta_2^\e\right)(x) \right\}.
\end{align*}
However, we see that
\begin{align*}
& \int_\R \left(\phi_x*\eta_2^\e\right)(y)
\eta_2^\e(x-y)dy = E[\phi_x(x+R_1+R_2)] = \frac12+\th*\eta_4^\e(x), \\
& \phi_x*\eta_2^\e(x) = E[\phi_x(x+R_1)] = \frac12+\th*\eta_2^\e(x),
\end{align*}
where $\{R_1, R_2\}$ are independent random variables
distributed under $\eta_2^\e$, from the symmetry of $R_i$.
Therefore, the contribution of $\bar\fa_2^{(3)}$ in the
left hand side of \eqref{eq:B-1} is given by
$
\sum_{n=0}^\infty \frac2{n!} I_n^{(3)},
$
where
\begin{align*}
I_n^{(3)} := & \int_\R \fa(x) e^{a(x)}dx \int_{\R^n} \phi_x^{\otimes n}
(\uu) d\uu\\
& \; \times \int_\R \left\{ f_{n+1}*(\eta^\e)^{\otimes (n+1)}
(\uu,y) - f_{n+1}*(\eta^\e)^{\otimes (n+1)}(\uu,x) \right\}
(\phi_x*\eta_2^\e)(y) \eta_2^\e(x-y)dy \\
& + \int_\R \fa(x) e^{a(x)} \{\th*\eta_4^\e(x)- \th*\eta_2^\e(x)\}dx
\int_{\R^n} f_{n+1}*(\eta^\e)^{\otimes (n+1)}(\uu,x) \phi_x^{\otimes n}
(\uu) d\uu.
\end{align*}
The second term in $I_n^{(3)}$
vanishes from our assumption by noting that
${\rm supp}\{\th*\eta_4^\e(x)- \th*\eta_2^\e(x)\} \subset
({\rm supp} \,\rho)^{4\e}$. For the first term,
estimating $|(\phi_x*\eta_2^\e)(y)| \le 1$, the absolute value of
the last integral in $y$ can be bounded by
$$
E[|f_{n+1}(\uu+U^\e, x+Y^\e+X_1^\e) -f_{n+1}(\uu+U^\e, x+X_1^\e)|],
$$
where $U^\e, X_1^\e, Y^\e$ are the same as before, and this can be estimated further by
\begin{align*}
2\e \int_0^1 E\left[ \left| \partial_{n+1}
f_{n+1}(\uu+U^\e, x+\la Y^\e+X_1^\e) \right|\right] d\la.
\end{align*}
Thus,
\begin{align*}
|I_n^{(3)}| & \le 2\e \|\fa e^{a(x)}\|_\infty \int_{\R^{n+1}}
1_{[-K-3\e,K+3\e]}(x) 1_{[-K-\e,K+\e]^n} (\uu)
|\partial_{n+1} f_{n+1}(\uu, x)| d\uu dx \\
& \le 2\e \|\fa e^{a(x)}\|_\infty (2K+6)^{(n+1)/2}
\left\{ \int_{\R^{n+1}} |\partial_{n+1} f_{n+1}(\ux)|^2 d\ux
\right\}^{1/2},
\end{align*}
by Schwarz's inequality. Therefore,
we get that
\begin{align*}
\sum_{n=0}^\infty \frac2{n!} |I_n^{(3)}|
& \le C\e \sum_{n=0}^\infty \frac1{n!} (2K+6)^{(n+1)/2}
\left\{ \int_{\R^{n+1}} |\partial_{n+1} f_{n+1}(\ux)|^2 d\ux
\right\}^{1/2} \\
& \le C'\e \|\Phi\|_{1,\e}.
\end{align*}
The contribution of $\bar\fa_2^{(4)}$ can be estimated similarly,
and this concludes the proof of the proposition.
\end{proof}
\begin{rem} \label{rem:3.3}
{\rm (1)}
The assumption that ${\rm supp}\, \fa$ and
${\rm supp}\, \rho$ separate is needed to treat the terms
$\bar\fa_2^{(3)}$ and $\bar\fa_2^{(4)}$. For $\bar\fa_2^{(2)}$,
this assumption is unnecessary by changing $\frac1{12}$
into $\frac1{12} + 2J_2^\e(x) + J_3^\e(x)$ in the definition of
$B^\e(x,Y)$. \\
{\rm (2)} Due to the symmetry of $\eta(x)$, we obtain the constant
$\frac1{12}$. For general asymmetric $\eta$, if it satisfies
$\int_0^\infty \eta_4(y)dy= \int_0^\infty \eta_2(y)dy$
(i.e., $P(R_1+R_2>0) = P(R_1>0)$ to treat $\bar\fa_2^{(3)}$ and
$\bar\fa_2^{(4)}$), this constant $\frac1{12}$ is replaced by
$$
J_1 = P(R_1+R_3>0, R_2+R_3>0) - P(R_1>0)^2,
$$
where $\{R_1, R_2, R_3\}$ are independent random variables
distributed under $\eta_2(y)dy$. For example, if
$\text{supp}\,\eta\subset [0,\infty)$ (or $(-\infty,0]$), then
$R_1, R_2, R_3>0$ a.s.\ and we have $J_1=0$.
\end{rem}
\subsection{SPDE on $\SS_M$} \label{section:3.4}
The arguments developed in Sections
\ref{section:3.1}--\ref{section:3.3} work for the
SPDE's on $\SS_M, M\ge 2,$ instead of $\R$ in a similar
way. We outline it. Let $h^{\e,M}(t,x)$, $x\in \SS_M$ be
the solution of the SPDE \eqref{eq:1-M}.
It is periodically extended on $\R$. Then, the Cole-Hopf
transformed process $Z^{\e,M}(t,x) = e^{h^{\e,M}(t,x)}$
satisfies the SPDE \eqref{eq:8-M}.
Taking $\rho$ as in Section \ref{sec:3.2}, we consider
wrapped processes $g^{\e,M}(t,x)$ and $Y^{\e,M}(t,x)$ of
$h^{\e,M}(t,x)$ and $Z^{\e,M}(t,x)$, respectively. Note that
the integral $h^{\e,M}(t,\rho) =\int_{\SS_M}h^{\e,M}(t,x)
\rho(x)dx$ is defined in a periodic sense; that is,
$\int_{-M/2}^{M/2} h^{\e,M}(t,x)\rho(x)dx$ recalling that $M\ge 2$.
We take the initial distribution of $h^{\e,M}(0,\cdot)$ to be
$\pi\otimes\nu^{\e,M}$ under the map \eqref{eq:3.1-map}
with $\R$ replaced by $\SS_M$. Then, the probability measure
$\pi\otimes \nu^{\e,M}$ is
invariant under $g^{\e,M}(t,x)$ as in Lemma \ref{lem:3.1-a}.
Lemma \ref{lem:3.2-R} is also parallel:
\begin{lem} \label{lem:3.2-R-M}
$Y^{\e,M}(t,x)$ satisfies the following equation in generalized functions' sense:
\begin{align} \label{eq:3-Y^e-2-M}
Y^{\e,M}(t,x) = Y^{\e,M}(0,x) + & \frac12 \int_0^t \partial_x^2 Y^{\e,M}(s,x)ds
+\int_0^t A^\e(x,Y^{\e,M}(s)) ds \\
& + \int_0^t Y^{\e,M}(s,x) dW^\e(s,x) + N^{\e,M}(t,x), \quad x\in \SS_M, \notag
\end{align}
where $A^\e(x,Y)$ is defined by \eqref{eq:3.4} with convolution $*\eta_2^\e$
considered in a periodic sense,
\begin{align} \label{eq:3.5-M}
N^{\e,M}(t,x) & =
\int_0^t \left\{ (e-1) 1_{\{ Y_\rho^{\e,M}(s-)=1\}}
+ (e^{-1}-1) 1_{\{ Y_\rho^{\e,M}(s-)=e\}} \right\}\\
&\hspace{5cm} \times Y^{\e,M}(s-,x) N^{\e,M}(ds), \notag
\end{align}
and $N^{\e,M}(t) = -[h^{\e,M}(t,\rho)]$.
\end{lem}
The limit $\nu$ of $\nu^\e$ as $\e\downarrow 0$ was
identified with the distribution of Gaussian random measure
$X=\{X(A); A\in \mathcal{B}(\R)\}$. The limit $\nu^M$
of $\nu^{\e,M}$ as $\e\downarrow 0$ is nothing but
$X=\{X(A); A\in \mathcal{B}([0,M))\}$ restricted on
$[0,M)$ and conditioned to be $X([0,M))=0$. Such conditional
random variables $X^M=\{X^M(A); A\in \mathcal{B}([0,M))\}$
can be realized by $X^M((a,b]) =B^M(b)-B^M(a)$,
$0\le a<b\le M$, where $B^M = \{B^M(x);x\in [0,M)\}$ is
the pinned Brownian motion such that $B^M(0)=B^M(M)=0$.
The Wiener-It\^o expansion \eqref{eq:11-b} is modified in the following
way under $\nu^M$: Taking a usual standard Brownian motion
$B=\{B(x); x\in [0,M]\}$ such that $B(0)=0$, the pinned Brownian
motion is realized as
\begin{equation} \label{eq:3.B^M}
B^M(x) = B(x)- \frac{x}MB(M), \quad x\in [0,M],
\end{equation}
so that $B^M(0)=B^M(M)=0$. By applying It\^o's formula and
noting $B(M) = \int_0^MdB(x)$, the multiple Wiener integral
with respect to $B^M$ can be rewritten into that with respect
to $B$ as follows:
\begin{align*}
I^M(\fa_n) := & \frac1{n!} \int_{\SS_M^n} \fa_n(x_1,\ldots,x_n) dB^M(x_1)\cdots dB^M(x_n) \\
=& \frac1{n!} \int_{[0,M)^n} \fa_n^M(x_1,\ldots,x_n) dB(x_1)\cdots dB(x_n) = I(\fa_n^M),
\end{align*}
where $\fa_n^M$ is defined by
\begin{align} \label{eq:3.fa^M}
\fa_n^M(x_1,\ldots,x_n) & = \sum_{k=0}^n \frac{(-1)^k}{M^k} \sum_{1\le i_1<\cdots<i_k\le n}
\int_{[0,M)^k} \check{\fa}_{n; i_1,\ldots,i_k} dy_{i_1}\cdots dy_{i_k}, \\
\check{\fa}_{n; i_1,\ldots,i_k} &= \fa_n(x_1,\ldots,x_{i_1-1},y_{i_1},x_{i_1+1},\ldots,x_{i_k-1},y_{i_k},x_{i_k+1},
\ldots,x_n). \notag
\end{align}
Note that $\fa_n^M$ is symmetric in $\ux=(x_1,\ldots,x_n)$ and satisfies
\begin{equation} \label{eq:fa_n=0}
\int_{[0,M)} \fa_n^M(x_1,\ldots,x_n)dx_i=0,
\end{equation}
for every $1\le i \le n$. Thus, $\Phi \in L^2(\tilde{\mathcal{C}}_M,\nu^M)$
has a Wiener-It\^o expansion
\begin{equation*}
\Phi = \sum_{n=0}^\infty I(\fa_n^M),
\end{equation*}
with some $\fa_0\in\R$ and $\fa_n^M\in \hat{L}_0^2([0,M)^n)
:= \{\fa_n^M\in \hat{L}^2([0,M)^n); \fa_n^M$ satisfies the condition
\eqref{eq:fa_n=0}$\}$. The quotient space $\tilde{\mathcal{C}}_M$
is defined similarly to $\tilde{\mathcal{C}}$.
Note that, for two symmetric functions $f_n=f_n(\ux)$ and $\fa_n=\fa_n(\ux)$,
from \eqref{eq:3.fa^M}, we see that
\begin{equation} \label{eq:P^M-dual}
\int_{[0,M)^n} f_n^M(\ux)\fa_n(\ux)d\ux=
\int_{[0,M)^n} f_n(\ux)\fa_n^M(\ux)d\ux,
\end{equation}
and $(\fa_n^M)^M = \fa_n^M$, or $\fa_n^M =\fa_n$ for
$\fa_n\in \hat{L}_0^2([0,M)^n)$. Moreover, under the representation
\eqref{eq:3.B^M} of $B^M$, \eqref{eq:6.1} is modified as
\begin{align*}
(\partial_x h)^2 &= \left\{\int_{[0,M)} \eta^\e(x-y) dB^M(y)\right\}^2 \\
&= \left\{\int_{[0,M)} \left(\eta^\e(x-y) -\tfrac1M\right) dB(y)\right\}^2
= \Psi^{\e,M}(x)+\xi^{\e,M},
\end{align*}
where $\xi^{\e,M}= \xi^\e- \frac1{M}$ and
\begin{align} \label{eq:Psi^M}
\Psi^{\e,M}(x)
& = \int_{[0,M)^2} \left(\eta^\e(x-x_1)-\tfrac1M\right)
\left(\eta^\e(x-x_2)-\tfrac1M\right) dB(x_1)dB(x_2) \\
& = \int_{[0,M)^2} \Big(\eta^\e(x-x_1)\eta^\e(x-x_2)\Big)^M dB(x_1)dB(x_2). \notag
\end{align}
Theorem \ref{thm:3.1} can be reformulated as follows in the present setting:
\begin{thm} \label{thm:3.1-M}
For every $\fa\in C(\SS_M)$ satisfying $\rm{supp}\,\fa \cap
\rm{supp}\,\rho = \emptyset$ (so that $\dist(\rm{supp}\,\fa,
\rm{supp}\,\rho)>0$ in a periodic sense), we have that
$$
\lim_{\e\downarrow 0} E^{\pi\otimes\nu^{\e,M}}
\left[ \sup_{0\le t \le T}\left\{ \int_0^t \hat A^\e(\fa,Y^{\e,M}(s))ds \right\}^2
\right] =0,
$$
where $\hat A^\e(\fa,Y)$ is defined similarly to that in Theorem \ref{thm:3.1}
by taking the integral over $\SS_M$.
\end{thm}
\begin{proof}
We modify the proofs of Proposition \ref{prop:Phi} and Theorem \ref{thm:3.1}.
By \eqref{eq:Psi^M}, the expectation \eqref{eq:Exp} under $\nu^{\e,M}$
in the present setting is equal to
\begin{align} \label{eq:3.XYZ}
\int_{[0,M)^2} \bar\fa_2^\e(x_1,x_2) & \left(
\int_{\SS_M} \Big(\eta^\e(y-x_1)\eta^\e(y-x_2)\Big)^M\eta_2^\e(x-y)dy
\right.\\
& \qquad\qquad\qquad \left. - \Big(\eta^\e(x-x_1)\eta^\e(x-x_2)\Big)^M\right)dx_1dx_2,
\notag
\end{align}
where $\bar\fa_2^\e$ is defined from $\bar\fa_2$ given below \eqref{eq:3.product}
with $\R^n$ replaced by $\SS_M^n$ and $\fa_n$, $\psi_n$ replaced by
$P^M\fa_n*(\eta^\e)^{\otimes n}$, $P^M\psi_n*(\eta^\e)^{\otimes n}$,
respectively. The operator $P^M$ is defined by $P^M\fa_n=\fa_n^M$ and note
that the constant $\xi^{\e,M}$ plays no role in \eqref{eq:3.XYZ}.
By \eqref{eq:P^M-dual}, the operator
$P^M$ acting on $\psi_n, \psi_{n+1}, \psi_{n+2}$ can be moved to
$P^M\fa_{n+2}$, $P^M\fa_{n+1}$, $P^M\fa_{n}$ under the integrals in the
variable $\uu$ and to $(\eta^\e(y-x_1)\eta^\e(y-x_2))^M$ under the
integrals in the variable $(x_1,x_2)$. Note that we can separate $P^M$
in these two variables. Thus, we can remove $P^M$ from
$\psi_n, \psi_{n+1}, \psi_{n+2}$. Writing $P^M\fa_{n}, P^M\fa_{n+1}, P^M\fa_{n+2}$
simply by $\fa_{n}, \fa_{n+1}, \fa_{n+2}$ again and noting $(P^M)^2 = P^M$,
we can also drop $P^M$ from $\fa_{n}, \fa_{n+1}, \fa_{n+2}$.
Now, we expand
\begin{equation} \label{eq:eta-eta}
\Big(\eta^\e(x-x_1)\eta^\e(x-x_2)\Big)^M
= \eta^\e(x-x_1)\eta^\e(x-x_2) - \frac1M\big(\eta^\e(x-x_1)+\eta^\e(x-x_2) \big)
+ \frac1{M^2}.
\end{equation}
The contribution to \eqref{eq:3.XYZ} from the first term of \eqref{eq:eta-eta}
is exactly the same as in the proof of Proposition \ref{prop:Phi}.
Note that the expectation $\frac1{12} E^{\nu^{\e,M}}[\frac{Y(x)}{Y_\rho}\Phi]$
can be computed similarly in the present setting. The contribution of the last constant
$\frac1{M^2}$ cancels when we take the difference in \eqref{eq:3.XYZ}.
The contribution to \eqref{eq:3.XYZ} from the second term of \eqref{eq:eta-eta}
becomes
\begin{align*}
-\frac2M & \int_{[0,M)^2} \bar\fa_2^\e(x_1,x_2)
\left( \int_{\SS_M} \eta^\e(y-x_1)\eta_2^\e(x-y)dy- \eta^\e(x-x_1)\right) dx_1dx_2\\
& = -\frac2M \int_{[0,M)^2} \bar\fa_2^\e(x_1,x_2)
\left( \eta_3^\e(x-x_1)- \eta^\e(x-x_1)\right) dx_1dx_2
\end{align*}
by the symmetry of $\bar\fa_2^\e(x_1,x_2)$. However, since
$\int_{[0,M)}\fa_{n+2}(\uu,x_1,x_2)dx_2= \int_{[0,M)}\fa_{n+1}(\uu,x_2)dx_2=0$
by \eqref{eq:fa_n=0}, the contributions of $\bar\fa_2^{(1)}$ and $\bar\fa_2^{(4)}$
vanish. It is therefore enough to compute the contributions of $\bar\fa_2^{(2)}$
and $\bar\fa_2^{(3)}$ only.
The computation of the contribution of
$$
\int_{[0,M)^2} \bar\fa_2^{(2),\e}(x_1,x_2)
\left( \eta_3^\e(x-x_1)- \eta^\e(x-x_1)\right) dx_1dx_2,
$$
with $\bar\fa_2^{(2),\e}$ defined from $\bar\fa_2^{(2)}$ by replacing $\fa_n$, $\psi_{n+2}$ by
$\fa_n*(\eta^\e)^{\otimes n}$, $\psi_{n+2}*(\eta^\e)^{\otimes (n+2)}$, respectively,
can be essentially reduced to the computation of
\begin{align*}
& \int_{[0,M)^2} \phi_x^{\otimes 2}*(\eta^\e)^{\otimes 2}(x_1,x_2)
\left( \eta_3^\e(x-x_1)- \eta^\e(x-x_1)\right) dx_1dx_2\\
& = \{\phi_x*\eta_4^\e(x)-\phi_x*\eta_2^\e(x)\} \int_{[0,M)} (\phi_x*\eta^\e)(x_2)dx_2\\
& = \{\th*\eta_4^\e(x)-\th*\eta_2^\e(x)\} \int_{[0,M)} (\phi_x*\eta^\e)(x_2)dx_2.
\end{align*}
As we saw in the proof of Proposition \ref{prop:Phi}, the support of this function
is contained in $(\text{supp}\,\rho)^{4\e}$ so that it vanishes when we multiply
the test function $\fa=\fa(x)$.
The computation of the contribution of
$$
\int_{[0,M)^2} \bar\fa_2^{(3),\e}(x_1,x_2)
\left( \eta_3^\e(x-x_1)- \eta^\e(x-x_1)\right) dx_1dx_2,
$$
with $ \bar\fa_2^{(3),\e}$ defined from $ \bar\fa_2^{(3)}$ similarly as above
can be essentially reduced to the computation of
\begin{align*}
\int_{\SS_M^{n+1}} & \phi_x^{\otimes n}(\uu)(\phi_x*\eta^\e)(x_2)d\uu dx_2\\
& \times \big\{f_{n+1}*((\eta^\e)^{\otimes n}\otimes \eta_3^\e)(\uu,x)
- f_{n+1}*((\eta^\e)^{\otimes n}\otimes \eta^\e)(\uu,x)\big\}.
\end{align*}
However, the difference in the braces can be rewritten as
$$
E[f_{n+1}(\uu+U^\e,x+Y^\e+X_1^\e) - f_{n+1}(\uu+U^\e,x+X_1^\e)],
$$
where $U^\e=\{U_i^\e\}_{i=1}^n$, $U_i^\e\stackrel{\rm{law}}{=}\eta^\e$,
$X_1^\e\stackrel{\rm{law}}{=}\eta^\e$, $Y^\e\stackrel{\rm{law}}{=}\eta_2^\e$
and $\{U_i^\e, X_1^\e, Y^\e\}$ are independent. This can be estimated exactly
in the same way as $I_n^{(3)}$ in the proof of Proposition \ref{prop:Phi}.
Therefore, we have a parallel assertion to Proposition \ref{prop:Phi}
in the present setting and this completes the proof of the theorem.
\end{proof}
The proof of Theorems \ref{thm:3.1} or \ref{thm:3.1-M} can be extended
to obtain the following lemma for $h^{\e,M}(t,x)$. To prove this lemma,
we need Poincar\'e inequality so that this holds only on $\SS_M$ and
not on $\R$.
\begin{lem} \label{cor:3.10}
Assume that a measurable and bounded function $\fa=\fa(s,x)$ on
$[0,T]\times \SS_M$ is given. Then, we have
\begin{align} \label{eq:cor3.10}
\sup_{0<\e<1} & E^{\pi\otimes\nu^{\e,M}}\left[ \sup_{0\le t \le T} \left(
\int_0^t H^{\e,M}(\fa(s,\cdot),\partial h^{\e,M}(s))ds\right)^2 \right] \\
& \qquad \le 12M^2 \int_0^T \|\fa(s,\cdot)\|_{L^\infty(\SS_M)}^2 ds, \notag
\end{align}
where
\begin{equation} \label{eq:H-e}
H^{\e,M}(x,\partial h) = \frac12 \big((\partial_x h)^2 -\xi^{\e,M} \big)
*\eta_2^\e(x)
\end{equation}
and $H^{\e,M}(\fa,\partial h) = \int_{\SS_M} H^{\e,M}(x,\partial h) \fa(x)dx$.
\end{lem}
\begin{proof}
By Lemma \ref{lem:3.6-a}, noting that $H^{\e,M}(\fa,\partial h)$ is a
tilt variable, the expectation in \eqref{eq:cor3.10} is bounded by
\begin{align} \label{eq:24sup}
24\int_0^T \sup_{\Phi\in L^2(\nu^{\e,M})}
\left\{ 2E^{\nu^{\e,M}}
\left[ \tilde\Psi^{\e,M}(\fa(s,\cdot)) \Phi \right] - \|\Phi\|_{1,\e}^2\right\} ds,
\end{align}
where $\tilde\Psi^{\e,M}(\fa) = \frac12\int_{\SS_M} \Psi^{\e,M}*\eta_2^\e(x)
\fa(x)dx$, $\Psi^{\e,M}(x)$ is given by \eqref{eq:Psi^M} and
$\|\Phi\|_{1,\e}^2$ is defined on $\SS_M$. Since $\Psi^{\e,M}$
is a second order Wiener chaos, we have that
\begin{align*}
2 E^{\nu^{\e,M}}
\left[ \tilde\Psi^{\e,M}(\fa) \Phi \right]
& = \int_{\SS_M} \fa*\eta_2^\e(x)dx
\int_{\SS_M^2} f_2(x_1,x_2) \Big(\eta^\e(x-x_1)\eta^\e(x-x_2)\Big)^Mdx_1dx_2 \\
& = \int_{\SS_M^2} f_2^M(x_1,x_2)\psi(x_1,x_2)dx_1dx_2,
\end{align*}
where
$$
\psi(x_1,x_2) = \int_{\SS_M}\fa*\eta_2^\e(x)\eta^\e(x-x_1)\eta^\e(x-x_2)dx,
$$
and $f_2=\fa_2\otimes (\eta^\e)^{\otimes 2}$ with
the kernel $\fa_2\in \hat L_0^2([0,M)^2)$ of the second order Wiener chaos of $\Phi$.
Note that \eqref{eq:fa_n=0} implies $\int_{\SS_M} \fa_2(x_1,x_2)dx_i=0$,
$i=1,2$, so that
\begin{equation} \label{eq:vanish}
\int_{\SS_M} f_2(x_1,x_2)dx_i=0.
\end{equation}
This shows $f_2^M=f_2$. Moreover, Lemma \ref{lem:3.5} (similar on $\SS_M$) implies
\begin{equation} \label{eq:Sob-Poin}
\frac12 \int_{\SS_M^2} \left( \frac{\partial f_2}{\partial x_1}(x_1,x_2)\right)^2 dx_1dx_2
\le \| \Phi\|_{1,\e}^2.
\end{equation}
We now estimate
\begin{align} \label{eq:3.33}
\left| 2E^{\nu^{\e,M}}
\left[ \tilde\Psi^{\e,M}(\fa) \Phi \right] \right|
\le \int_{\SS_M} \|f_2(\cdot,x_2)\|_{L^\infty(\SS_M)} \|\psi(\cdot,x_2)\|_{L^1(\SS_M)} dx_2.
\end{align}
However, we easily see that
$$
\|\psi(\cdot,x_2)\|_{L^1(\SS_M)} \le \|\fa\|_{L^\infty(\SS_M)},
$$
for every $x_2\in \SS_M$ and, by \eqref{eq:vanish} with $i=1$,
\begin{align*}
|f_2(x_1, x_2)| & = \left| f_2(x_1, x_2) - \frac1M \int_{\SS_M} f_2(y, x_2)dy \right| \\
& = \frac1M \left| \int_{\SS_M} dy \int_y^{x_1} \frac{\partial f_2}{\partial z}(z,x_2)dz\right|
\le \int_{\SS_M} \left| \frac{\partial f_2}{\partial z}(z,x_2) \right| dz.
\end{align*}
Thus, by \eqref{eq:3.33}, Schwarz's inequality and \eqref{eq:Sob-Poin}, we obtain
\begin{align*}
\left| 2E^{\nu^{\e,M}}
\left[ \tilde\Psi^{\e,M}(\fa) \Phi \right] \right|
& \le \|\fa\|_{L^\infty(\SS_M)}
\int_{\SS_M^2} \left| \frac{\partial f_2}{\partial x_1}(x_1,x_2) \right| dx_1dx_2 \\
& \le M \|\fa\|_{L^\infty(\SS_M)}
\left( \int_{\SS_M^2} \left| \frac{\partial f_2}{\partial x_1}(x_1,x_2) \right|^2 dx_1dx_2
\right)^{1/2} \\
& \le \sqrt{2} M \|\fa\|_{L^\infty(\SS_M)} \| \Phi\|_{1,\e}.
\end{align*}
Combining this with \eqref{eq:24sup}, the conclusion is shown
similarly to the last part of the proof of Theorem \ref{thm:3.1}.
\end{proof}
\begin{rem} \label{rem:3.4-c}
{\rm (1)} For $p>1$, $L^p$-norms of $\psi$ diverge as $\e\downarrow 0$ in general.\\
{\rm (2)} On $\R$, we have the same estimate as \eqref{eq:3.33} and, if
$\text{supp}\, \fa \subset [-K,K]$, then $\|\psi(\cdot,x_2)\|_{L^1(\R)} =0$
if $|x_2|\ge K+3\e$. Therefore, Morrey's theorem (\cite{Adams}, p.97 (8)),
which implies $\|f_2(\cdot,x_2)\|_{L^\infty(\R)} \le C \|f_2(\cdot,x_2)\|_{H^1(\R)}$,
shows that
$$
\left| 2E^{\nu^{\e}}
\left[ \tilde\Psi^\e(\fa) \Phi \right] \right|
\le C \sqrt{K} \|\fa\|_{L^\infty(\R)} \{\|\partial f_2/\partial x_1\|_{L^2(\R^2)} +
\|f\|_{L^2(\R^2)}\}.
$$
However, since Poincar\'e inequality is missing, this cannot be bounded by
$\|\Phi\|_{1,\e}$, because of the last term $\|f\|_{L^2(\R^2)}$. This estimate
can be easily extended to $\fa$ such that $\sup_{x\in\R}|\fa(x)|\sqrt{1+|x|}
< \infty$.
\end{rem}
This lemma is applied to prove the following proposition, which shows that
the height average cannot move very quickly. Recall that
the initial distribution of $h^{\e,M}(t,x)$ is given by $h^{\e,M}(0,\cdot)
\stackrel{\rm{law}}{=} \pi\otimes\nu^{\e,M}$ under the map \eqref{eq:3.1-map}
defined on $\SS_M$.
Note that the wrapping is not introduced for $h^{\e,M}(t,x)$.
This result will be used to remove the wrapping from
$Y^{\e,M}(t,x)$ in the next section.
\begin{prop} \label{prop:3.12}
For every $T>0$ and $\fa\in C^2(\SS_M)$,
\begin{equation} \label{eq:H-3}
\sup_{0<\e<1} E\left[\sup_{0\le t \le T}h^{\e,M}(t,\fa)^2\right] <\infty.
\end{equation}
\end{prop}
\begin{proof}
For every $\fa\in C^2(\SS_M)$, from the SPDE \eqref{eq:1-M}, we have
\begin{align} \label{eq:H-1}
h^{\e,M}(t,\fa) = & h^{\e,M}(0,\fa) + \frac12 \int_0^t h^{\e,M}(s,\fa'')ds
+ \int_0^t H^{\e,M}(\fa,\partial h^{\e,M}(s))ds \\
&+ \frac12(\xi^{\e,M}-\xi^\e)t \lan\fa,1\ran_{\SS_M}
+ \lan W(t),\fa*\eta^\e\ran_{\SS_M}. \notag
\end{align}
By Lemma \ref{cor:3.10}, it is shown that
$$
\sup_{0<\e<1} E\left[ \sup_{0\le t \le T} \left(
\int_0^t H^{\e,M}(\fa,\partial h^{\e,M}(s))ds\right)^2 \right] <\infty.
$$
It is easy to see that $\xi^{\e,M}-\xi^\e= -\frac1M$ is finite,
$$
E\left[ \sup_{0\le t \le T} \lan W(t),\fa*\eta^\e\ran_{\SS_M}^2\right]
\le 4 E[\lan W(T),\fa*\eta^\e\ran_{\SS_M}^2]
= 4T \|\fa*\eta^\e\|_{L^2(\SS_M)}^2
\le 4T \|\fa\|_{L^2(\SS_M)}^2 <\infty,
$$
and
$$
E\left[ \sup_{0\le t \le T} \left(
\int_0^t h^{\e,M}(s,\fa'')ds\right)^2 \right] \le T \int_0^T E[h^{\e,M}(s,\fa'')^2]ds.
$$
Therefore, once we can prove
\begin{equation} \label{eq:H-2}
\sup_{0<\e<1} \sup_{0\le t \le T}E[h^{\e,M}(s,\fa)^2] <\infty,
\end{equation}
for every $\fa\in C(\SS_M)$, \eqref{eq:H-1} shows \eqref{eq:H-3}.
The next task is to give the proof of \eqref{eq:H-2}. To this end,
we rewrite the equation for $h^{\e,M}(t,x)$ into a mild form:
\begin{align*}
h^{\e,M}(t,\fa) = & \int_{\SS_M} h^{\e,M}(0,y)\psi(t,y)dy + \int_0^t
\int_{\SS_M} \psi(t-s,y) H^{\e,M}(y,\partial h^{\e,M}(s))dsdy \\
& + \frac12 (\xi^{\e,M}-\xi^\e) \int_0^t\int_{\SS_M} \psi(t-s,y) dsdy
+ \int_0^t \int_{\SS_M} \psi(t-s,y) W^\e(dsdy),
\end{align*}
where $\psi(t,y) = \int_{\SS_M} p^M(t,x,y)\fa(x)dx$ and
$p^M$ is the heat kernel on $\SS_M$. We apply Lemma \ref{cor:3.10}
by dropping $\sup_{0\le t\le T}$ and then regarding $\fa(s,y) =\psi(t-s,y)$
for fixed $t$, we have
\begin{align*}
& E^{\pi\otimes \nu^{\e,M}}\left[\left( \int_0^t
\int_{\SS_M} \psi(t-s,y) H^{\e,M}(y,\partial h^{\e,M}(s))dsdy \right)^2\right] \\
& \le 12M^2 \int_0^t \|\psi(t-s,\cdot)\|_{L^\infty(\SS_M)}^2ds
\le 12M^2 t \|\fa\|_{L^\infty(\SS_M)}^2<\infty.
\end{align*}
We also see
$$
E^{\pi\otimes \nu^{\e,M}}\left[\left( \int_0^t \int_{\SS_M}
\psi(t-s,y) W^\e(dsdy) \right)^2 \right]
= \int_0^t \int_{\SS_M} (\psi(t-s,\cdot)*\eta^\e(y))^2dsdy
\le C <\infty.
$$
Since the first term has a uniform bound recalling
$h^{\e,M}(0,\cdot) \stackrel{\rm{law}}{=} \pi\otimes\nu^{\e,M}$
and the integral in the third term is simply $t\lan\fa,1\ran_{\SS_M}$,
this completes the proof of \eqref{eq:H-2}.
\end{proof}
\begin{rem} \label{rem:3.5}
As we will see in the next subsection, $N^{\e,M}(t,\fa) = \int_{\SS_M}
N^{\e,M}(t,x)\fa(x)dx$
in \eqref{eq:3-Y^e-2-M} converges weakly in $L^2(\Om)$ as $\e\downarrow 0$,
since all other terms do. Therefore, its $L^2$-norm is automatically bounded in $\e$:
\begin{equation} \label{eq:3.4-a}
\sup_{0<\e<1} E\big[\big(N^{\e,M}(t,\fa)\big)^2\big] < \infty.
\end{equation}
Proposition \ref{prop:3.12} gives a stronger estimate with supremum in $t$
inside the expectation.
\end{rem}
\subsection{Proof of Theorem \ref{thm:0}} \label{subsection:3.4.3}
We are now at the position to complete the proof of Theorem \ref{thm:0}.
We fix $M\ge 2$ (or $M\ge 1$ by making the support of $\rho$ smaller)
and denote $h^{\e,M}(t,x)$, $Z^{\e,M}(t,x)$, $Y^{\e,M}(t,x)$
simply by $h^{\e}(t,x)$, $Z^{\e}(t,x)$, $Y^{\e}(t,x)$, respectively, in
this subsection.
\vskip 2mm
\noindent
{\it Step 1.} $\,$
Let $h^{\e}(t,x)$ be the solution of the SPDE \eqref{eq:1-M} such that
$h^{\e}(0,0)=h_0\in \R$ and $\nabla h^{\e}(0)\stackrel{\rm{law}}{=}\nu^{\e,M}$.
We may assume $h_0=0$ without loss of generality.
Then, $h^\e(0,\cdot) (\in \mathcal{C}_M) \stackrel{\rm{law}}{=}
\de_0\otimes \nu^{\e,M}$, $0<\e<1$ is tight.
\begin{lem} \label{lem:3.12-ub}
{\rm (1)} The uniform bound \eqref{eq:H-3} holds also for this $h^{\e}(t,x)$. \newline
{\rm (2)} The tilt variable of $h^\e(t,\cdot)$ under the map \eqref{eq:3.1-map}
on $\SS_M$ is $\nu^{\e,M}$-distributed for all $t\ge 0$.
\end{lem}
\begin{proof}
The bound \eqref{eq:H-3} was shown for
the solution $\tilde h^{\e}(t,x)$ of the SPDE \eqref{eq:1-M} such that
$\tilde h^{\e}(0,\cdot) = \big(\tilde h^{\e}(0,\rho), \tilde h^{\e}(0,\cdot)
-\tilde h^{\e}(0,\rho)\big) \stackrel{\rm{law}}{=}\pi\otimes\nu^{\e,M}$
under the map \eqref{eq:3.1-map} on $\SS_M$. However, two solutions have a
simple relation: $\tilde h^{\e}(t,x) = h^{\e}(t,x) - h^{\e}(0,\rho)+m$ with
a $\pi$-distributed random variable $m$ (which is independent of
tilt variables of $h^\e(0,\cdot)$). Therefore, we see that
$|h^\e(t,\fa)-\tilde h^\e(t,\fa)| \le \{1+|h^\e(0,\rho)|\}\int_{\SS_M}|\fa(x)|dx$
so that \eqref{eq:H-3} holds for the solution $h^\e(t,x)$
we are now considering. The second assertion (2) follows by
noting that the tilt variables of $h^\e(t,\cdot)$ and $\tilde h^\e(t,\cdot)$
coincide: $h^\e(t,x)-h^\e(t,\rho) = \tilde h^\e(t,x) - \tilde h^\e(t,\rho)$.
\end{proof}
Lemma \ref{lem:3.12-ub} taking $\fa=\rho$ in \eqref{eq:H-3}
shows that, for every fixed $t>0$, $h^\e(t,\cdot)$, $0<\e<1$, is tight on
$\mathcal{C}_M$.
We fix $T>0$ arbitrarily. Then, from Lemma \ref{lem:3.12-ub} again,
$(h^\e(0,\cdot), h^\e(T,\cdot), X^\e:=\sup_{0\le t \le T}|h^\e(t,\rho)|, W(t))$,
$0<\e<1$, is jointly tight on
$\mathcal{C}_M\times \mathcal{C}_M\times \R\times C([0,T], H^{-\a}(\SS_M))$
with suitably chosen $\a>0$. In particular, every subsequence $\{\e'\downarrow 0\}$ of
$\{\e\in (0,1)\}$ contains a weakly convergent subsequence (in law sense)
$(h^{\e''}(0,\cdot), h^{\e''}(T,\cdot), X^{\e''}, W(t))$. Thus, by Skorohod's
theorem, we can find a probability space $(\tilde\Om,\tilde P)$ and
$(\tilde h^{\e''}(0,\cdot), \tilde h^{\e''}(T,\cdot), \tilde X^{\e''}, \tilde W_{\e''}(t))$
defined on this space having the same law as
$(h^{\e''}(0,\cdot), h^{\e''}(T,\cdot), X^{\e''}, W(t))$,
and $(\tilde h(0,\cdot), \tilde h(T,\cdot), \tilde X, \tilde W(t))$ also
defined on this space such that
$(\tilde h^{\e''}(0,\cdot), \tilde h^{\e''}(T,\cdot), \tilde X^{\e''}, \tilde W_{\e''}(t))$
converges to
$(\tilde h(0,\cdot), \tilde h(T,\cdot), \tilde X, \tilde W(t))$
$\tilde P$-a.s.\ $\tilde \om$ as $\e'' \downarrow 0$.
From the space-time Gaussian white noise $\tilde W_{\e''}(t)$,
one can construct a smeared noise $\tilde W_{\e''}^{\e''}(t) := \tilde W_{\e''}(t)*
\eta^{\e''}$ and consider the SPDE \eqref{eq:8-M} with initial value
$\tilde Z^{\e''}(0,\cdot) = e^{\tilde h^{\e''}(0,\cdot)}$. Then, we have its solution
$\tilde Z^{\e''}(t,\cdot)$. $\tilde Z^{\e''}(T,\cdot)$ is consistent with
$\tilde h^{\e''}(T,\cdot)$ and $\tilde Z^{\e''}(t,0)$ is consistent with
$\tilde h^{\e''}(t,0)$, respectively, $\tilde P$-a.s.
For every $L\ge 1$, we define the wrapped process $\tilde Y_L^{\e''}(t,x)
\equiv \tilde Y_L^{\e''}(t,x;\tilde \om,m)$
on the extended probability space $(\tilde\Om\times [-L,L],\tilde P\otimes\pi_L)$
in such a manner that $\tilde Y_L^{\e''}(t,x) = e^m \tilde Z^{\e''}(t,x)$ modulo
$2L$ averaged by $\rho$ in an exponential sense such as \eqref{eq:3.5-c},
i.e., $\log (\tilde Y_L^{\e''}(t))_\rho \in [e^{-L},e^L]$ (instead of $[1,e]$)
and $\log (\tilde Y_L^{\e''}(t))_\rho = m+ \log (\tilde Z^{\e''}(t))_\rho$
modulo $2L$ with $\pi_L$-valued random variable $m$, where $\pi_L$
is a uniform probability measure on $[-L,L]$.
In other words, the initial value
$\log (\tilde Y_L^{\e''})_\rho(0)$ is distributed under $\pi_L$.
\noindent
{\it Step 2}. $\,$ Since $\tilde X^{\e''} = \sup_{0\le t \le T}|\tilde h^{\e''}(t,\rho)|$
converges to $\tilde X$ as $\e''\downarrow 0$
in $\R$, $\tilde P$-a.s., and $0\le \tilde X<\infty$, $\tilde P$-a.s.,
we see that $\lim_{L\to\infty} \tilde P(\tilde \Om_L)=1$ holds, where
$$
\tilde \Om_L := \left\{ \sup_{0<\e''<1} \sup_{0\le t \le T}
|\tilde h^{\e''}(t,\rho)| \le \tfrac{L}2 \right\} \; (\subset \tilde \Om).
$$
In particular, we have that
\begin{equation} \label{eq:A-1}
\tilde Y_L^{\e''}(t,x;\tilde \om,m) = e^m \tilde Z^{\e''}(t,x;\tilde \om)
\end{equation}
for all $t\in [0,T], m: |m|\le \tfrac{L}2, 0<\e''<1$ and $\tilde \om \in
\tilde \Om_L$.
We prepare the following uniform bound on $\tilde Y_L^{\e''}(t,x)$ defined on $\tilde\Om$.
\begin{lem}
We have that
\begin{equation} \label{eq:3.Y}
\tilde E[\tilde Y_L^{\e''}(t,x;\cdot,m)^{2p}] \le e^{4p^2(|x|+c)+2Lp},
\end{equation}
for every $0<\e''<1$, $t\ge 0$, $p\ge 1$, $x\in [-\tfrac{M}2,\tfrac{M}2]$ and $m\in [-L,L]$ with some
$c>0$.
\end{lem}
\begin{proof}
Since the tilt variable for $\tilde Y^{\e''}(t)$ is distributed under $\nu^{\e'',M}$,
we have that
\begin{align*}
\tilde E[\tilde Y_L^{\e''}(t,x)^{2p}]
& \le \tilde E[(\tilde Y_L^{\e''}(t))_\rho^{4p}]^{1/2}
\tilde E \left[ \frac{\tilde Y_L^{\e''}(t,x)^{4p}}{(\tilde Y_L^{\e''}(t))_\rho^{4p}}\right]^{1/2}\\
& \le e^{2Lp} E[e^{4p\big(B^M(\cdot)-\int_{\SS_M} B^M(y)\rho(y)dy\big)*\eta^{\e''}(x)}]^{1/2} \\
& = e^{2Lp} E[e^{4p\int_{\SS_M} \phi_\cdot(u)*\eta^{\e''}(x)dB^M(u)}]^{1/2} \\
& = e^{2Lp} e^{\tfrac14 (4p)^2\int_{\SS_M} \big(\phi_\cdot(u)*\eta^{\e''}(x)
- \frac1M \int_{\SS_M} \phi_\cdot(v)*\eta^{\e''}(x)dv\big)^2 du} \\
& \le e^{2Lp} e^{4p^2\int_{\SS_M} \big(\phi_\cdot(u)*\eta^{\e''}(x)\big)^2 du} \\
& \le e^{4p^2(|x|+ c)+2Lp},
\end{align*}
where $\phi_x(u)$ is defined in \eqref{eq:3.17-x}. Note that
$\phi_\cdot(u)*\eta^{\e''}(x) = 1_{[u,\infty)}*\eta^\e(x)+\th(u)$,
$1_{[u,\infty)}*\eta^\e(x) =0$ $(u\ge x+\e$), $1 (u\le x-\e$),
$\in [0,1]$ (otherwise), and $\th(u)=0 (u\ge K)$, $\th(u)=-1
(u\le -K)$, where $K>0$ is taken such that $\text{supp}\, \rho
\subset [-K,K]\subset [-1,1]$. Recall that supp$\,\eta^\e \subset [-\e,\e]$.
This shows \eqref{eq:3.Y}.
\end{proof}
Since \eqref{eq:3.Y} with $p=1$ implies the weak compactness of
$\tilde Y_L^{\e''}(\cdot,\cdot)$, by the similar argument to Krylov and
Rozovskii \cite{KR}, p.1264 and by the diagonal argument
in $L\in \N$, one can find a subsequence $\e'''\downarrow 0$ of $\e''$
such that
\begin{align} \label{eq:3.25}
\tilde Y_L^{\e'''}(t,x) \to \tilde Y_L(t,x;\tilde\om,m),
\end{align}
weakly in $L^2([0,T]\times\tilde\Om\times [-L,L] ,\bar{\mathcal{T}}_L,
dtd\tilde Pd\pi_L; L^2(\SS_M))$ with some $\tilde Y_L(t,x;\tilde\om,m)$,
where $\bar{\mathcal{T}}_L$ is the completion of the $\si$-field
of progressively measurable sets on $[0,T]\times\tilde\Om\times [-L,L]$
with respect to $dtd\tilde Pd\pi_L$ for every $L\in \N$.
This combined with \eqref{eq:A-1} shows
\begin{align} \label{eq:3.26}
\tilde Z^{\e'''}(t,x) \to \tilde Z(t,x),
\end{align}
weakly in $L^2([0,T]\times\tilde\Om_L, \bar{\mathcal{T}},
dtd\tilde P; L^2(\SS_M))$,
for all $L\in \N$ with some $\tilde Z(t,x)$, where $\bar{\mathcal{T}}$ is
defined similarly to $\bar{\mathcal{T}}_L$. Furthermore, by definition,
we see that
\begin{equation} \label{eq:A-2}
\tilde Y_L(t,x;\tilde \om,m) = e^m \tilde Z(t,x;\tilde \om)
\end{equation}
in $L^2([0,T],L^2(\SS_M))$ for a.e.\ $m: |m|\le \tfrac{L}2, L\in \N$
and $\tilde P$-a.s.\ $\tilde \om \in\tilde
\Om_L$.
The identity \eqref{eq:A-2} can be rewritten as
\begin{equation*}
\tilde Z(t,x;\tilde \om) = e^{-m}\tilde Y_L(t,x;\tilde \om,m),
\end{equation*}
for a.e.\ $m: |m|\le \tfrac{L}2, L\in \N$ and $\tilde P$-a.s.\
$\tilde \om \in\tilde \Om_L$.
\noindent
{\it Step 3}.$\,$
From \eqref{eq:3-Y^e-2-M} considered on $\tilde\Om$ and modulo $2L$
rather than modulo $1$, under a multiplication of $\fa = \fa(x)
\in C^\infty(\SS_M)$ and dropping superscripts $M$ as above, we have that
\begin{align} \label{eq:3-Y^e-3}
& \lan \tilde Y_L^{\e'''}(t),\fa\ran_{\SS_M}
= \lan \tilde Y_L^{\e'''}(0),\fa\ran_{\SS_M} \\
& \quad + \int_0^t
\lan \tilde Y_L^{\e'''}(s), \tfrac12 \partial_x^2\fa
+ \tfrac1{24}\fa\ran_{\SS_M} ds
+ \int_0^t \int_{\SS_M} \hat A^{\e'''}(x,\tilde Y_L^{\e'''}(s)) \fa(x) dsdx
\notag \\
& \quad
+ \int_0^t \int_{\SS_M} \tilde Y_L^{\e'''}(s,x) \fa(x) \tilde W^{\e'''}(dsdx)
+ \int_{\SS_M} \tilde N_L^{\e'''}(t,x) \fa(x) dx, \notag
\end{align}
where $\tilde W^{\e'''} := \tilde W_{\e'''}*\eta^{\e'''}$ and
$\tilde N_L^{\e'''}$ is defined correspondingly.
By Theorem \ref{thm:3.1-M}, for each fixed $L\in \N$,
the third term in the right hand side
of \eqref{eq:3-Y^e-3} converges to $0$ strongly in $L^2(\tilde\Om\times [-L,L],
d\tilde Pd\pi_L)$ as $\e (=\e''') \downarrow 0$, if $\text{supp}\,\fa \cap
\text{supp}\,\rho=\emptyset$.
The fourth term in the right hand side of \eqref{eq:3-Y^e-3}
involving stochastic integrals converges in the following sense:
\begin{lem}\label{lem:3.1-ab}
For every $0\le t\le T$, as $\e (=\e''') \downarrow 0$,
\begin{equation} \label{eq:3.stoch-int}
\int_0^t \int_{\SS_M} \tilde Y_L^{\e'''}(s,x)\fa(x) \tilde W^{\e'''}(dsdx)
\rightarrow \int_0^t \int_{\SS_M} \tilde Y_L(s,x)\fa(x) \tilde W(dsdx),
\end{equation}
weakly in $L^2(\tilde \Om\times [-L,L],\tilde{\mathcal F}_t,\tilde P\otimes\pi_L)$,
where $\tilde{\mathcal F}_t$ is the $\si$-field generated by
$\{\tilde{W}(s); 0\le s \le t\}$.
\end{lem}
\begin{proof}
Rewriting as
\begin{align*}
\int_0^t \int_{\SS_M} \tilde Y_L^{\e'''}(s,x)\fa(x) \tilde W^{\e'''}(dsdx)
& = \int_0^t \int_{\SS_M}
\left(\tilde Y_L^{\e'''}(s,\cdot)\fa(\cdot)\right)* \eta^{\e'''}(x) \tilde W_{\e'''}(dsdx)\\
& = \int_0^t \int_{\SS_M^2} \tilde Y_L^{\e'''}(s,y)\fa(y)
\eta^{\e'''}(x-y) \tilde W_{\e'''}(dsdx)dy,
\end{align*}
the difference of the both sides of \eqref{eq:3.stoch-int} is given by
\begin{align*}
& \int_0^t \int_{\SS_M^2} \tilde Y_L^{\e'''}(s,y) \fa(y)\eta^{\e'''}(x-y)
\{ \tilde W_{\e'''}(dsdx) - \tilde W(dsdx)\} dy \\
& \quad +
\int_0^t \int_{\SS_M^2} \{\tilde Y_L^{\e'''}(s,y)-\tilde Y_L(s,y)\}
\fa(y)\eta^{\e'''}(x-y) \tilde W(dsdx)dy \\
& \quad + \int_0^t \int_{\SS_M^2} \{\tilde Y_L(s,y)-\tilde Y_L(s,x)\}
\fa(y)\eta^{\e'''}(x-y) \tilde W(dsdx)dy \\
& \quad + \int_0^t \int_{\SS_M} \tilde Y_L(s,x)
\{ \fa*\eta^{\e'''}(x)-\fa(x)\} \tilde W(dsdx) \\
& \quad =: I^{(1,\e''')} + I^{(2,\e''')} + I^{(3,\e''')} + I^{(4,\e''')}.
\end{align*}
For the first term $I^{(1,\e''')}$, since both $\tilde W_{\e'''}$ and $\tilde W$ are
the space-time Gaussian white noises and $\tilde W_{\e'''}$ converges to $\tilde W$
a.s., we see that $\tilde W_{\e'''} - \tilde W$ is also a space-time Gaussian white noise
multiplied by a constant $c_{\e'''}$ converging to $0$ as $\e'''\downarrow 0$. Thus,
\begin{align*}
\tilde E[\{I^{(1,\e''')}\}^2]
& = c_{\e'''}^2 \int_0^t \int_{\SS_M} \tilde E \left[ \left(
\int_{\SS_M} \tilde Y_L^{\e'''}(s,y) \fa(y)\eta^{\e'''}(x-y) dy
\right)^2 \right]ds dx\\
& \le c_{\e'''}^2 \int_0^t \int_{\SS_M} \fa(y)^2 \tilde E [
\tilde Y_L^{\e'''}(s,y)^2 ]ds dy,
\end{align*}
which converges to $0$ as $\e'''\downarrow 0$ by
noting \eqref{eq:3.Y}. For the last term,
since $\|\fa*\eta^{\e'''}-\fa\|_{L^\infty(\SS_M)} \to 0$, we easily see that
$$
\tilde E[\{I^{(4,\e''')}\}^2] \to 0 \quad \text{as } \e'''\downarrow 0.
$$
For other terms, note that functions $\Phi$ of the forms
$$
\Phi = \int_0^t \int_{\SS_M} \fa(s,x,\tilde\om) \tilde W(dsdx)
$$
with bounded and continuous kernels $\fa(s,x,\tilde\om)$ form a dense
family in $L^2(\tilde\Om,\tilde{\mathcal F}_t,\tilde P)\ominus \{\text{const}\}$.
Then, we have that
$$
\tilde E[I^{(2,\e''')}\Phi] = \int_0^t ds \int_{\SS_M^2}
\tilde E[\{\tilde Y_L^{\e'''}(s,y)-\tilde Y_L(s,y)\} \fa(s,x,\tilde\om)]
\fa(y)\eta^{\e'''}(x-y) dxdy.
$$
If we can replace $\fa(s,x,\tilde\om)$ with $\fa(s,y,\tilde\om)$, then
$\eta^{\e'''}(x-y)$ disappears under the integration in $x$ and
this expectation converges to $0$ by \eqref{eq:3.25}
and applying Lebesgue's convergence theorem noting
\eqref{eq:3.Y}. The replacement of $\fa(s,x,\om)$ with
$\fa(s,y,\om)$ with a small error negligible as $\e'''\downarrow 0$
can be justified by the continuity of $\fa(\cdot,\cdot,\om)$
noting \eqref{eq:3.Y}.
Similarly for $I^{(3,\e''')}$, we have that
$$
\tilde E[I^{(3,\e''')}\Phi] = \int_0^t ds \int_{\SS_M^2}
\tilde E[\{\tilde Y_L(s,y)-\tilde Y_L(s,x)\} \fa(s,x,\tilde\om)] \fa(y)\eta^{\e'''}(x-y) dxdy.
$$
First, by the continuity of $\fa(\cdot,\cdot,\tilde\om)$ and noting
\eqref{eq:3.Y}, which implies a corresponding bound on $\tilde Y_L(s,y)$,
we can replace
this by
\begin{equation} \label{eq:3.YYY}
\int_0^t ds \int_{\SS_M^2} \tilde E[\tilde Y_L(s,y)\fa(s,y,\tilde\om)
-\tilde Y_L(s,x) \fa(s,x,\tilde\om)] \fa(y)\eta^{\e'''}(x-y) dxdy
\end{equation}
with a small error negligible as $\e'''\downarrow 0$. Then, noting the
symmetry of $\eta^{\e'''}$, we can rewrite \eqref{eq:3.YYY} as
$$
\int_0^t ds \int_{\SS_M^2} \tilde E[\tilde Y_L(s,y)\fa(s,y,\tilde\om)]
(\fa(y) -\fa(x)) \eta^{\e'''}(x-y) dxdy.
$$
However, since $\tilde E[\tilde Y_L(s,y)\fa(s,y,\tilde\om)] \in L^2([0,t]\times \SS_M)$,
this tends to $0$. This shows that $E[I^{(3,\e''')}\Phi]$ converges to $0$ as $\e'''\downarrow 0$.
\end{proof}
From Theorem \ref{thm:3.1-M} applied on $\tilde\Om$, \eqref{eq:3.25} and Lemma
\ref{lem:3.1-ab}, the last term in the right hand side of
\eqref{eq:3-Y^e-2-M} (on $\tilde\Om$ and modulo $2L$, or
\eqref{eq:3-Y^e-3} integrated with $\fa(x)$)
must converge weakly
to a certain $\tilde{N}_L(t,x) = \tilde{N}_L(t,x;\tilde\om,m)$ in
$L^2(\tilde\Om\times [-L,L], \tilde{\mathcal F}_T,\tilde P\otimes\pi_L)$;
c.f.\ Remark \ref{rem:3.5}. Thus we obtain the equation:
\begin{align} \label{eq:3-12-a}
\tilde Y_L(t,x) = \tilde Y_L(0,x) + & \frac12 \int_0^t \partial_x^2 \tilde Y_L(s,x)ds
+ \frac1{24} \int_0^t \tilde Y_L(s,x) ds \\
& + \int_0^t \tilde Y_L(s,x) d\tilde W(s,x) + \tilde{N}_L(t,x), \notag
\end{align}
for a.e.\ $x\in (\text{supp}\, \rho)^c$, and a.e.\ $(t,\tilde\om,m)$ in the limit.
Recalling the definition of $\tilde\Om_L$, we see that the limit
$\tilde{N}_L(t,x) = \tilde{N}_L(t,x;\tilde\om,m)$ of the jump part of $\tilde Y_L^{\e'''}$
vanishes on $\tilde\Om_L\times [-\frac{L}2,\frac{L}2]$, i.e., $\tilde N_L(t,x)=0$
for all $t\in [0,T]$ on $\tilde \Om_L\times [-\frac{L}2,\frac{L}2]$. Moreover,
since $\tilde P(\cup_{L\in\N}\tilde\Om_L)=1$ holds, from
\eqref{eq:3-12-a} and \eqref{eq:A-2},
we see that $\tilde Z(t,x;\tilde \om)$ satisfies the SPDE
\eqref{eq:8-b-M} on $(\text{supp}\, \rho)^c$. However, choosing another $\bar\rho$,
which satisfies the same condition as $\rho$ stated at the beginning of Section
\ref{sec:3.2}, such that $\text{supp}\, \rho \cap \text{supp}\, \bar\rho =
\emptyset$, we see that $\tilde Z(t,x;\tilde \om)$ satisfies the SPDE
\eqref{eq:8-b-M} on the whole torus $\SS_M$. Note that $\tilde Z$ is defined
independently of the choice of $\rho$.
Since the solution of \eqref{eq:8-b-M} is unique and continuous in $(t,x)$,
we find that $\tilde Z(t,x;\tilde \om)$ is continuous in $(t,x)$, $\tilde P$-a.s.
In particular, $\tilde Z(T,x;\tilde \om)$ is consistent with $\tilde h(T,\cdot)$,
which was introduced in Step 1 of this subsection. Since $\tilde Z^{\e'''}(T)$
converges to $\tilde Z(T)$ $\tilde P$-a.s.\ and the limit is uniquely
characterized by the SPDE \eqref{eq:8-b-M}, we don't need to
take the subsequences. This concludes the proof of Theorem
\ref{thm:0}.
\subsection{Proof of Theorem \ref{thm:1.1}} \label{section:3.6}
We first prove that the solution $Z^M(t,x), x\in \R$ of the SPDE
\eqref{eq:8-b-M} on $\SS_M$ periodically extended to the whole
line $\R$ weakly converges to the solution $Z(t,x), x\in \R$ of the
SPDE \eqref{eq:8-b}:
\begin{prop} \label{prop:SPDE-conv}
Assume that $Z(0,\cdot)\in L_r^2(\R), r>0,$ is given and the initial
value of $Z^M$ is determined by $Z^M(0,x) = Z(0,x)$ for $|x|\le M/2$
and periodically extended to $\R$. Then, we have the followings.\\
{\rm (1)} $\{Z^M(t,x)\}_{M\ge 1}$ is tight on $C([0,\infty),L_r^2(\R))$.\\
{\rm (2)} $Z^M(t,x)$ weakly converges to $Z(t,x)$ on
the space $C([0,\infty),L_r^2(\R))$ as $M\to\infty$.
\end{prop}
To prove this proposition, we prepare a lemma. Recall that
the fundamental solution $p^M$ of the parabolic operator
$\tfrac{\partial}{\partial t} - \tfrac12 \tfrac{\partial^2}{\partial x^2}$
on $\SS_M$ is given by
$$
p^M(t,x,y) = \sum_{n=-\infty}^\infty p(t,x,y+nM),
\quad x,y \in \SS_M = [0,M).
$$
We define a function $\chi_M(x), x\in \R$ by $\chi_M(x)=|x|$
for $|x|\le \frac{M}2$ and then by periodically extending it on $\R$.
The following lemma is an extension of Lemma 6.2 of \cite{Shiga}
to $\SS_M$; see also \cite{F91}.
\begin{lem} \label{lem:tight-Z}
{\rm (1)} For every $0<\de<1$ and $T>0$, there exists
$C=C_{\de,T}>0$ such that
$$
\int_0^{t'}\int_{\SS_M} \left( p^M(t'-s,x',y) - 1_{\{s\le t\}}p^M(t-s,x,y)
\right)^2dy \le C(|t-t'|^{1/2}+|x-x'|^{1-\de}),
$$
holds for $0\le t < t'\le T, x,x'\in \SS_M$. \\
{\rm (2)} For every $r\in \R$ and $T>0$,
$$
\sup_{M\ge 1}\sup_{0\le t \le T} \sup_{x\in \SS_M}
e^{-r\chi_M(x)}\int_{\SS_M} p^M(t,x,y)e^{r\chi_M(y)}dy<\infty.
$$
\end{lem}
\begin{proof}
To show (1), we first expand the square inside the integral in $y$
and apply Chapman-Kolmogorov's equality for the integral of
products of $p^M$. Then, (1) is shown by the following easily
shown uniform bounds on $p^M$:
$$
0<p^M(t,x,y)\le \frac{C}{\sqrt{t}}, \quad
\left|\frac{\partial p^M}{\partial x}(t,x,y)\right| \le \frac{C}t,
$$
for $0<t\le T$ with $C=C_T>0$ which is independent of $M$.
For (2), we note that
$$
\chi_M(x)-|a|\le \chi_M(x+a)\le \chi_M(x)+|a|, \quad x, a \in \R,
$$
and therefore
$$
\int_{\SS_M} p^M(t,x,y)e^{r\chi_M(y)}dy
= E\left[ e^{r\chi_M(x+B(t))}\right]
\le e^{r\chi_M(x)}E\left[ e^{|rB(t)|}\right] \le Ce^{r\chi_M(x)},
$$
where $B(t)$ is a Brownian motion starting at $0$.
\end{proof}
\begin{proof}[Proof of Proposition \ref{prop:SPDE-conv}]
For (1), we follow the proof of Theorem 2.2 in \cite{Shiga}.
The SPDE \eqref{eq:8-b-M} can be rewritten into the mild form:
\begin{align} \label{eq:Z-mild}
Z^M(t,x) = & \int_{\SS_M} Z^M(0,y) p^M(t,x,y)dy \\
& + \int_0^t ds \int_{\SS_M} e^{\frac1{24}(t-s)}p^M(t-s,x,y)
Z^M(s,y) W(dsdy). \notag
\end{align}
We first show that
$$
\sup_{M\ge 1}\sup_{0\le t \le T} \int_{\SS_M} e^{-r\chi_M(x)}
E[|Z^M(t,x)|^{2p}] dx<\infty,
$$
for $p\ge 1$ and $T>0$ by using Lemma \ref{lem:tight-Z}-(2).
Then, by Lemma \ref{lem:tight-Z}-(1), we have the bound:
$$
E[|X^M(t,x)- X^M(t',x')|^{2p}]
\le C\{|t-t'|^p+|x-x'|^{2p(1-\de)}\},
$$
for every $0\le t<t'\le T, \; x, x'\in \SS_M: |x-x'|\le 1$,
where $X^M(t,x)$ is the second term in the right hand side
of \eqref{eq:Z-mild}. Since the first term is easily treated, this
proves the tightness.
To show (2), we rely on the martingale formulation as in the
proof of Theorem \ref{thm:2}. We consider the operator
$$
\mathcal{L}^Z\Phi(Z) =\int_\R \left( \frac12 Z^2(x)
D^2\Phi(x,x;Z) +\Big(\frac12 \partial_x^2Z(x) +\frac1{24} Z(x)\Big)
D\Phi(x;Z) \right) dx,
$$
for $\Phi\in \mathcal{D}(\mathcal{C})$ and
$\mathcal{L}_M^Z\Phi$ for $\Phi\in \mathcal{D}(\mathcal{C}_M)$
by replacing the integral over $\R$ by $\SS_M$. Then, the distribution
$P^M$ of $Z^M(\cdot,\cdot)$ is a solution of
$(\mathcal{L}_M^Z, \mathcal{D}(\mathcal{C}_M))$-martingale
problem and $\{P^M\}$ is tight. Then, it is easy to see that
any weak limit of $P^M$ as $M\to\infty$ is a solution of
$(\mathcal{L}^Z, \mathcal{D}(\mathcal{C}))$-martingale
problem. Since the limit is unique, this concludes the proof of (2).
\end{proof}
Since $\nu^M$ (periodically extended on $\R$)
weakly converges to $\nu$ as $M\to\infty$ on $L_r^2(\R), r>1$
(note $E^\nu[e^{2B(x)}] = e^{2|x|}, x\in \R$), this
proposition shows that $\nu$ is invariant for the tilt variable of the
logarithm of the solution $Z(t,\cdot)$ of the SPDE \eqref{eq:8-b}.
Let $\bar Z(t)$ be the solution of the SPDE \eqref{eq:8-b} and set
$$
Z(t) := e^{-\tfrac1{24}t} \bar Z(t).
$$
Then, one can easily see that $Z(t)$ is a solution of \eqref{eq:1.1}
and $Z(t) \sim \bar Z(t)$ by the definition.
Since the tilt variable of $Z^\e(t)$ has $\tilde\mu^\e$ as
invariant probability measure, we see in the limit the tilt variable
of $\bar Z(t)$ has $\tilde\mu$ as invariant probability measure.
Since $Z(t) \sim \bar Z(t)$, this concludes that the tilt variable of
$Z(t)$ also has $\tilde\mu$ as invariant probability measure.
In other words, we obtain the invariance of the
distribution of the geometric Brownian motion for the tilt process
determined by the stochastic heat equation \eqref{eq:1.1}:
\begin{prop} \label{prop:3.7}
For any bounded and continuous function $G=G(Z)$ on
$\tilde{\mathcal{C}}_+$ and for any $\e>0$, $t\ge 0$, we have that
\begin{equation*}
\int_{\tilde{\mathcal{C}}_+} G(Z(t))d\tilde\mu
= \int_{\tilde{\mathcal{C}}_+} G(Z(0))d\tilde\mu,
\end{equation*}
where $\tilde\mu$ is a probability distribution on $\tilde{\mathcal{C}}_+$ of
$Z(\cdot)=e^{B(\cdot)}$ with $B(\cdot)\in \tilde{\mathcal{C}}$ distributed under $\nu$.
\end{prop}
In order to deduce Theorem \ref{thm:1.1} from Proposition
\ref{prop:3.7}, one needs to now recover the height at the origin $h(t,0)$, which is defined
by $h(t,x) = \log Z(t,x)$. However, since this does not really
work as in Section \ref{section:2.7},
we consider its smooth approximation.
Let $\eta^\e$ be the function introduced previously, and define a function
$h^\e(x,Z) := \log (Z*\eta^\e(x))$ of $x\in \R$ and $Z\in \mathcal{C}_+$.
Let $Z(t)$ be the solution of the SPDE \eqref{eq:1.1}. Then, by It\^o's
formula, the approximated height $h^\e(t,x) := h^\e(x,Z(t))$ satisfies
\begin{equation}\label{smoothed1}
h^\e(t,x) = h^\e(0,x) + \int_0^t b^\e(x,Z(s))ds + \int_0^t \int_\R
\si^\e(x,y,Z(s)) W(dsdy),
\end{equation}
where
\begin{align*}
& b^\e(x,Z) = \frac12 \left\{ \partial_x^2 h^\e(x,Z)
+ (\partial_x h^\e(x,Z))^2
- \frac{(Z^2*(\eta^\e)^2)(x)}{(Z*\eta^\e(x))^2} \right\}, \\
& \si^\e(x,y,Z) = \frac{\eta^\e(x-y)Z(y)}{Z*\eta^\e(x)}.
\end{align*}
The key point is that, as functions of $Z$, both $b^\e$ and $\si^\e$ are
defined on the quotient space $\tilde{\mathcal{C}}_+$. Therefore, once
$Z(t)\in\tilde{\mathcal{C}}_+$ is determined by solving the SPDE
\eqref{eq:1.1}, we can recover its height as
\begin{equation}\label{smoothed2-b}
h^\e(t,0) = h^\e(0,0) + X_t,
\end{equation}
where $X_t$ is the sum of the second and third terms in the right hand side
of \eqref{smoothed1} with $x=0$, which depends only on $Z(\cdot)$.
\begin{lem}
For any bounded, integrable and continuous function $G=G(h,Z)$ on
$\mathcal{C}_+ \equiv \R\times\tilde{\mathcal{C}}_+$ and for any $\e>0$, $t\ge 0$, we have that
\begin{equation}\label{smoothed3-b}
\int_{\R\times\tilde{\mathcal{C}}_+} G(h^\e(t,0),Z(t))dh_0d\tilde\mu
= \int_{\R\times\tilde{\mathcal{C}}_+} G(h^\e(0,0),Z(0))dh_0d\tilde\mu.
\end{equation}
\end{lem}
\begin{proof}
From \eqref{smoothed2-b} and the translation-invariance of the Lebesgue measure,
the left hand side of \eqref{smoothed3-b} is equal to
$$
\int_{\R\times\tilde{\mathcal{C}}_+} G(h^\e(0,0)+X_t,Z(t))dh_0d\tilde\mu
= \int_{\R\times\tilde{\mathcal{C}}_+} G(h^\e(0,0),Z(t))dh_0d\tilde\mu,
$$
by performing the integral in $dh_0$ first.
But, this is equal to the right hand side of \eqref{smoothed3-b}
by the invariance of $\mu$ under $Z(t)\in\tilde{\mathcal{C}}_+$;
see Proposition \ref{prop:3.7}.
\end{proof}
Letting $\e$ tend to zero in \eqref{smoothed3-b}, we obtain
\eqref{smoothed3-b} also for
$(h(t,0),Z(t))$. This implies the invariance of the product measure
$dh_0d\tilde\mu$ for this joint process with the state space
$\R\times\tilde{\mathcal{C}}_+ $.
However, since the image measure of $dh_0d\tilde\mu$ under the
map $(h_0,Z) \in \R\times\tilde{\mathcal{C}}_+ \mapsto e^{h_0}Z \in
\mathcal{C}_+$ is nothing but $\mu$, the proof of Theorem \ref{thm:1.1} is completed.
\vskip 5mm
\noindent
{\bf Acknowledgement.} $\,$
The authors thank Makiko Sasada for pointing out
a simple proof of \eqref{eq:1/3}.
|
\section{Introduction}
The origins of vortex dynamics lie in the famous work of Helmholtz of 1858 , where he introduced the concepts of vortex line, vortex filament and derived the vorticity equation for an ideal incomprenssible fluid \cite{H}. Helmholtz also introduced planar point vortices and their equations of motion in order to model a $2$-dimensional slice of columnar vortex filaments. Some years later, Kirchhoff (1876) gave a Hamiltonian formulation of Helmholtz's equations for point vortices (see \cite{K} for more details). This model has been widely used to provide finite-dimensional approximation to vorticity evolution in fluid dynamics. Kirchhoff proved that $n$ point vortices in the plane located at $z_i=(x_i, y_i) \in \mathbb{R}^{2}$ with vortex strength $\Gamma_i \neq 0 \in \mathbb{R}$ for $i=1, \dots, n$ satisfy
\begin{equation}\label{Kirchof}
\Gamma_{i} \frac{dx_{i}}{dt} = \frac{\partial H}{\partial y_{i}}, \qquad \Gamma_{i} \frac{dy_{i}}{dt} = -\frac{\partial H}{\partial x_{i}},
\end{equation}
where
\[
H = -\frac{1}{2}\sum_{i<j} \Gamma_{i}\Gamma_{j} \log [(x_{i} - x_{j})^{2} + (y_{i} - y_{j})^{2}].
\]
Computing the derivatives indicated in (\ref{Kirchof}), we can find the velocity of the $i$-th vortex:
\begin{equation}\label{Hamiltoniano}
\frac{dx_{i}}{dt} = -\sum_{j\neq i} \Gamma_{j}\frac{y_{i} - y_{j}}{r_{ij}^{2}}, \qquad \frac{dy_{i}}{dt} = \sum_{j\neq i} \Gamma_{j}\frac{x_{i} - x_{j}}{r_{ij}^{2}},
\end{equation}
where
\[
r_{ij}^{2} = (x_{i} - x_{j})^{2} + (y_{i} - y_{j})^{2}.
\]
Let $
\textbf{J} =
\left(
\begin{array}{cc}
0 & \textbf{I} \\
-\textbf{I} & 0
\end{array}
\right),
$ the standard symplectic matrix $2n \times 2n$, and let $\nabla_j$ denote the two-dimensional partial gradient with respect to $z_j$, the previous equation can be written in vectorial form as
\begin{equation}\label{em}
\textbf{G} \dot{z_{i}} = \textbf{J} \nabla_{i} H = -J \sum_{j=1}^{n}\frac{\Gamma_{i}\Gamma_{j}}{r_{ij}}(z_{i}-z_{j}), \qquad 1 \leq i \leq n,
\end{equation}
where $\textbf{G} = \; \text{diag}(\Gamma_{1}, \cdots ,\Gamma_{n})$ is the diagonal matrix.
In general the $n$-vortex system is simpler than the $n$-body problem of point masses governed by Newtonian gravity for a given $n$. For example the three-vortex system is always integrable, whereas in the Newtonian three-body problem there are chaotic regimes. Somewhat offsetting this relative simplicity is the larger set of parameter values that we must investigate (since $\Gamma_i<0$ is allowed).
In 2001, Kossin and Schubart \cite{KS} conducted numerical experiments describing the evolution of thin annular rings with large vorticity as a model for the behavior seen in the eyeball of intensifying hurricanes. In a conservative, idealized setting, they find examples of ``vortex crystals'', formations of mesovortices (namely a vertical vortex of air associated with a thunderstorm that occurs at less one kilometer from the ground) that rigidly rotate as a solid body. One particular formation of four vortices, situated very close to a rhombus configuration, is observed to last for the final 18 hours of a 24-hour simulation. Rigidly rotating polygonal configurations have also been found in the eyeballs of hurricanes in weather research and forecasting models from the Hurricane Group at the National Center for Atmospheric Research (see the website \cite{C} for some revealing simulations).
It is natural to explore these rigidly rotating configurations in a dynamical systems setting by studying relative equilibria of the planar $n$- vortex problem.
The three-vortex case was extensively studied by Gr\"{o}bli \cite{G}, Kossin and Schubart \cite{KS} and Hern\'andez-Gardu\~{n}o and Lacomba \cite{GL}. Equilateral triangles are always relative equilibria. There are also collinear equilibria and/or relative equilibria, which are determined by a cubic equation in a shape parameter with coefficients which are linear in the vortex strengths. Depending of the parameters it is possible to obtain one, two or three collinear relative equilibria. In 2008 M. Hampton and R. Moeckel proved the finiteness of relative equilibria in the four vortex problem \cite{HM}. The four vortex case with two pairs of equal vorticities was studied by Hampton, Roberts and Santoprete \cite{HS}. The linear and nonlinear stability of certain symmetric configurations of point vortices on the sphere forming relative equilibria was studied in \cite{LMR}. These configurations include, in particular, kite configurations.
The purpose of this paper is to study the planar relative equilibria of the 4-vortex problem with three equal vorticities. We set three vorticities equal to 1, and the fourth vorticity $\Gamma_{4}$ is taken as a real parameter. Our main goal is to classify, describe, and count the number and type of solutions of relative equilibria as $\Gamma_{4}$ varies. We say that a planar non-collinear central configuration of the 4-vortex problem has a kite shape if it has an axis of symmetry passing through two of the vorticities. The kite configuration is \textit{convex} if none of the bodies is located in the interior of the convex hull of the other three, otherwise and if the configuration is not collinear we say that the kite configuration is \textit{concave}. When counting solutions, we use the standard convention from celestial mechanics that solutions which are identical under scaling or rotation are considered equivalent.
When $\Gamma_{4} =1$ (all vortex strengths equal), there are 34 solutions. In this case, it is not possible to distinguish if the vorticity $\Gamma_{4}$ is at a vertex of the triangle formed by the convex hull of the four vorticities, or if it is in its interior. There are only four geometrically distinct configurations: a square, an equilateral triangle with a vortex at the center, an equilateral triangle with a vortex on a vertex of the triangle formed by the convex hull of the four vorticities, and a collinear configuration. This is different from the Newtonian case, where an additional symmetric, concave solution exists, consisting of an isosceles triangle with an interior body on the axis of symmetry. An interesting bifurcation occurs as $\Gamma_{4}$ decreases through $\Gamma_{4} =1$, the equilateral triangle with the vortex 4 at the center of the triangle goes to an isosceles triangle with the interior vortex on the line of symmetry, and the equilateral triangle with the vortex 4 at the vertex of the equilateral triangle formed by the convex hull of the other three vorticities vanish. Thus, the number of solutions decreases from 34 to 23 for the case $\Gamma_{4}>1$. If $\Gamma_{4}$ increases from $0$ through $1$, the equilateral triangle splits into three different solutions. If the vortex 4 is at the vertex of the equilateral triangle formed by the convex hull of the other three vorticities, then the solution for $\Gamma_{4} = 1$ bifurcates into two different isosceles triangles. If vortex 4 is at the center of the triangle, the equilateral solution goes to an isosceles configuration. Thus, the number of solutions decreases from 34 to 29 for the case $0<\Gamma_{4}<1$.
As $\Gamma_{4}$ flips sign, there is one bifurcation value at $\Gamma_{4} = 0$. When $\Gamma_{4} = 0$ there are 26 solutions. The equilateral triangle bifurcates in two different solutions, a isosceles triangle with $\Gamma_{4}$ in its interior, and $\Gamma_{4}$ on the vertex of an isosceles triangle. If $\Gamma_{4}$ approach to 0 there is one kite concave configuration having $\Gamma_{4}$ in the interior of an isosceles triangle. In the collinear case we have bifurcation value at $\Gamma_{4} = -1/2$.
The paper is organized as follows: In Section 2, we define a relative equilibrium and explain how to use mutual distances as variables in the 4-vortex problem. In Section 3, we describe the relevant algebraic techniques used to analyze and quantify the number of solutions. Section 4 examines the interplay between symmetry and equality of vorticities in two special cases: absolute equilibria and rigid translations. Sections 5 and 6 cover the collinear case and kite configurations.
\section{Relative Equilibria}
A motion of $n$ vortices is said to be a \textit{relative equilibrium} if, and only if there exists a real number $\lambda$, called angular velocity, such that, for every $i,j$ and for all time $t$:
\[
z_{i} - z_{j} = e^{-J\lambda t}(z_{i}(0) - z_{j}(0)).
\]
Then one of the following statements is satisfied:
\begin{itemize}
\item If the $z_{j}$ are constant, then the motion is said to be an \textit{absolute equilibrium}. In this case, we have $\lambda = 0$.
\item If there exists a velocity of translation $v \neq 0$ such that, for every $j$ and for all time $t$: $z_{j}(t) = z_{j}(0) + tv$, then the motion is said to be a \textit{rigid translation}. Again, we have $\lambda = 0$.
\item If $\lambda \neq 0$. then there exists a \textit{center of rotation} $c$ such that, for every $j$ and for all time $t$:
\[
z_{j}(t) = c + e^{-J\lambda t}(z_{j}(0) -c).
\]
If, moreover: $\sum_{j=1}^{n}\Gamma_{j}\neq 0$, the center of rotation is the center o vorticity.
\end{itemize}
In fact, looking for these motions we observe that the search of relative equilibria is equivalent to look for the configurations which generate them, as shown in the following proposition which is proved in \cite{O}:
\begin{proposition}\label{mer}
A motion of $n$ vortices is a relative equilibrium if, and only if, at a certain time, there exists a real number $\lambda$ such that for every $i, j$ we have $ \dot{z_{i}}-\dot{z_{j}} = - \lambda(z_{i} - z_{j})$.
\begin{itemize}
\item[1.] It is an absolute equilibrium if, and only if, at a certain time, for every $j$ we have $ \dot{z}_{j} = 0.$
\item[2.] It is a rigid translation if, and only if, at a certain time, there exists $v \neq 0$ such that for every $j$ we have $ \dot{z}_{j} = v$.
\item[3.] It is a relative equilibrium with $\lambda \neq 0$ if, and only if, at a certain time, there exists $c$ such that for every $j$ we have,
\end{itemize}
\begin{equation}\label{re}
\dot{z}_{j} = -\lambda(z_{j} - c).
\end{equation}
\end{proposition}
\begin{definition}
The following quantities are defined:
\begin{align*}
\text{Total vorticity} & \; & \Gamma &= \sum \Gamma_{l}\\
\text{Angular momentum} & \; & L &= \sum_{l<k}\Gamma_{l}\Gamma_{k}\\
\text{Moment of vorticity} & \; & M &= \sum \Gamma_{l}z_{l}\\
\text{Center of vorticity } & \; & c &= M/ \Gamma \;(\text{when} \; \Gamma \neq 0)\\
\text{Moment of inertia} & \; & I &= \frac{1}{2}\sum \Gamma_{l}\|z_{l}\|^{2}
\end{align*}
\end{definition}
Applying the matrix $\textbf{G}$ to both sides of equation (\ref{re} ) in Proposition (\ref{mer}), and using the equations of motion (\ref{em}) we obtain
\begin{equation}\label{inercia}
\lambda \nabla (I-I_{0}) = \nabla H,
\end{equation}
where $\nabla = (\nabla_{1}, \dots , \nabla_{n})$ and $I = I_{0}$. We observe that equation (\ref{inercia}) is Lagrange multiplier problem, with $\lambda$ as the Lagrange multiplier and any solution can be interpreted as a critical point of the Hamiltonian $H(z)$ under the condition that $I$ remains constant. Using the homogeneity of the functions $H$ and $I$, equation (\ref{inercia}) implies that the angular velocity $\lambda$ in a relative equilibria is given by
\begin{equation}\label{lambda}
\lambda = \frac{-L}{2I}.
\end{equation}
\subsection{Equations in mutual distances}
We now consider the case of $n=4$ vortices. Our presentation follows the approach of \cite{Sch} in describing the work of Dziobek \cite{Dz} for the Newtonian $n-$body problem. We want to express equation (\ref{inercia}) in terms of the mutual distance variables $r_{ij}$. In 1900, Dziobek gave the innovative idea of using the mutual distances of the bodies as unknowns in order to formulate the equations of motion. His work reduces the problem of searching for relative equilibria to the study of systems of non-linear polynomial equations that exhibit many symmetries.
Between four vortices there are six mutual distances, which are not independent if the vortices are planar; generically they describe a tetrahedron in $\mathbb{R}^{3}$ in place of a configuration in the plane. In order that they describe a planar relative equilibria we need an additional constraint, which is obtained by setting the volume of the tetrahedron equal to zero. This restriction follows from the Cayley-Menger determinant:
\[
S = \left|
\begin{array}{ccccc}
0 & 1& 1 & 1 & 1 \\
1 & 0 & r_{12}^{2} & r_{13}^{2} & r_{14}^{2} \\
1 & r_{12}^{2} & 0 & r_{23}^{2} & r_{24}^{2} \\
1 & r_{13}^{2} & r_{23}^{2} & 0 & r_{34}^{2} \\
1 & r_{14}^{2} & r_{24}^{2} & r_{34}^{2} & 0 \\
\end{array}
\right|.
\]
Hence, relative equilibria configurations are obtained as critical points of following equation
\begin{equation}\label{cp}
H - \lambda(I - I_{0}) - \frac{\mu}{32} S = 0,
\end{equation}
depending on $\lambda, \mu, r_{12}, \dots , r_{34}$, where $\lambda$ and $\mu$ are Lagrange multipliers.
\smallskip
\noindent
To find $S$ restricted to planar configurations, we use the following important formula
\begin{equation*}
\frac{\partial S}{\partial r_{ij}^{2}} = -32 A_{i}A_{j},
\end{equation*}
where $A_i$ is the oriented area of the triangle $T_i$ whose vertices are all except the $i-$th vortex. Setting the gradient of equation (\ref{cp}) equal to zero yields the equations
\[
\frac{\partial H}{\partial r_{ij}^{2}} - \lambda \frac{\partial I}{\partial r_{ij}^{2}} - \frac{\mu}{32}\frac{\partial S}{\partial r_{ij}^{2}} = 0.
\]
If $\Gamma \neq 0$, $I$ can be written in terms of the mutual distances as
\[
I = \frac{1}{2\Gamma} \sum_{i<j}\Gamma_{i}\Gamma_{j} r_{ij}^{2},
\]
respectively, so
\[
\frac{\partial I}{\partial r_{ij}^{2}} = \frac{\Gamma_{i}\Gamma_{j}}{\Gamma}.
\]
\noindent
Using this, we obtain the following equations for a four-vortex central configuration:
\[
\Gamma_{i}\Gamma_{j}(r_{ij}^{-2} + \lambda ') = \mu A_{i}A_{j},
\]
where $\lambda = \lambda ' \Gamma, I = I_{0}$ and $S = 0$.
Explicity we have:
\begin{eqnarray}\label{md}
\Gamma_{1}\Gamma_{2}(r_{12}^{-2} + \lambda') = \mu A_{1}A_{2}, & \quad \Gamma_{3}\Gamma_{4}(r_{34}^{-2} + \lambda') = \mu A_{3}A_{4},\nonumber \\
\Gamma_{1}\Gamma_{3}(r_{13}^{-2} + \lambda') = \mu A_{1}A_{3}, & \quad \Gamma_{2}\Gamma_{4}(r_{24}^{-2} + \lambda') = \mu A_{2}A_{4},\\
\Gamma_{1}\Gamma_{4}(r_{14}^{-2} + \lambda') = \mu A_{1}A_{4}, & \quad \Gamma_{2}\Gamma_{3}(r_{23}^{-2} + \lambda') = \mu A_{2}A_{3}.\nonumber
\end{eqnarray}
This yields to the Dziobek equations for vortices \cite{Dz}:
\begin{equation}\label{Dziobek}
(r_{12}^{-2} + \lambda')(r_{34}^{-2} + \lambda') = (r_{13}^{-2} + \lambda')(r_{24}^{-2} + \lambda') = (r_{14}^{-2} + \lambda')(r_{23}^{-2} + \lambda').
\end{equation}
From the above equations we find
\[
\lambda = \frac{r_{12}^{-2}r_{34}^{-2} - r_{13}^{-2}r_{24}^{-2}}{r_{12}^{-2} + r_{34}^{-2} - r_{13}^{-2} - r_{24}^{-2}} = \frac{r_{13}^{-2}r_{24}^{-2} - r_{14}^{-2}r_{23}^{-2}}{r_{13}^{-2} + r_{24}^{-2} - r_{14}^{-2} - r_{23}^{-2}} = \frac{r_{14}^{-2}r_{23}^{-2} - r_{12}^{-2}r_{34}^{-2}}{r_{14}^{-2} + r_{23}^{-2} - r_{12}^{-2} - r_{34}^{-2}}.
\]
Now, using the different ratios of two vorticities that can be found from the equations in (\ref{md}), we obtain:
\begin{eqnarray}\label{fv}
\frac{\Gamma_{1}A_{2}}{\Gamma_{2}A_{1}} = \frac{\rho_{23}+\lambda '}{\rho_{13} + \lambda '} = \frac{\rho_{24}+\lambda '}{\rho_{14} + \lambda '} = \frac{\rho_{23}-\rho_{24}}{\rho_{13}-\rho_{14}}, \nonumber \\
\frac{\Gamma_{1}A_{3}}{\Gamma_{3}A_{1}} = \frac{\rho_{23}+\lambda '}{\rho_{12} + \lambda '} = \frac{\rho_{34}+\lambda '}{\rho_{14} + \lambda '} = \frac{\rho_{23}-\rho_{34}}{\rho_{12}-\rho_{14}}, \nonumber \\
\frac{\Gamma_{1}A_{4}}{\Gamma_{4}A_{1}} = \frac{\rho_{24}+\lambda '}{\rho_{12} + \lambda '} = \frac{\rho_{34}+\lambda '}{\rho_{13} + \lambda '} = \frac{\rho_{24}-\rho_{34}}{\rho_{12}-\rho_{13}}, \nonumber \\
\frac{\Gamma_{2}A_{3}}{\Gamma_{3}A_{2}} = \frac{\rho_{13}+\lambda '}{\rho_{12} + \lambda '} = \frac{\rho_{34}+\lambda '}{\rho_{24} + \lambda '} = \frac{\rho_{13}-\rho_{34}}{\rho_{12}-\rho_{24}}, \\
\frac{\Gamma_{2}A_{4}}{\Gamma_{4}A_{2}} = \frac{\rho_{14}+\lambda '}{\rho_{12} + \lambda '} = \frac{\rho_{34}+\lambda '}{\rho_{23} + \lambda '} = \frac{\rho_{14}-\rho_{34}}{\rho_{12}-\rho_{23}}, \nonumber \\
\frac{\Gamma_{3}A_{4}}{\Gamma_{4}A_{3}} = \frac{\rho_{14}+\lambda '}{\rho_{13} + \lambda '} = \frac{\rho_{24}+\lambda '}{\rho_{23} + \lambda '} = \frac{\rho_{14}-\rho_{24}}{\rho_{13}-\rho_{23}}, \nonumber
\end{eqnarray}
where $\rho_{ij} = r_{ij}^{2}$, and $\lambda '$ is a constant.
Eliminating $\lambda'$ from equation (\ref{Dziobek}) and factoring we obtain the important relation
\begin{equation}\label{8}
(r_{13}^{2}-r_{12}^{2})(r_{23}^{2}-r_{34}^{2})(r_{24}^{2}-r_{14}^{2}) = (r_{12}^{2}-r_{14}^{2})(r_{24}^{2}-r_{34}^{2})(r_{13}^{2}-r_{23}^{2}).
\end{equation}
\noindent
Assuming that the six mutual distances determine a configuration in the plane, equations (\ref{8}) give a necessary condition for the existence of a four-vortex relative equilibrium. The corresponding vortex strengths are then found from equations (\ref{fv}).
\section{Algebraic Techniques}
In this section we describe an algebraic technique useful for analyzing solutions to our problem: elimination theory using Gr\"{o}bner bases.
\subsection{Elimination Theory and Gr\"obner Bases}
Elimination theory is the classical name for algorithmic approaches to eliminating some variables between polynomials of several variables. The linear case would be handled by \textit{Gaussian elimination}. In the same way, computational techniques for elimination can in practice be based on \textit{Gr\"obner bases} methods.
We mention some elements from elimination theory and the theory of Gr\"obner bases that will prove useful in our analysis. For more details see \cite{DJD}.
Let $K$ be a field and consider the polynomial ring $K[x_1,...,x_n]$ of polynomials in $n$ variables $x_i$ over $K$. Let $f_1,...f_l$ be $l$ polynomials in $K[x_1,...,x_n]$ and consider the ideal $I = <f_1, . . . , f_l>=<\mathcal{F}>$ generated by these polynomials.
\begin{definition}
An order is a relation on the monomials $>$ such that it is a total order (ie, given two monomials, one is always greater than the other), such that if $x^\alpha>x^\beta$, and $x^\gamma$ any monomial, we have $x^{\alpha+\gamma}>x^{\beta+\gamma}$. That is, it is preserved by multiplication of monomials. Finally, we want it to be a well-ordering. That is, every non-empty subset has a smallest element.
\end{definition}
\begin{definition}
Let $I$ an ideal in $K[x_1, \dots , x_n]$. We call the $k-$th elimination ideal the ideal $I \cap K[x_{k+1}, \dots ,x_n]$ in $K[x_{k+1}, \dots, x_n]$.
\end{definition}
\noindent
Note that if $k=0$, we just get $I$.
\smallskip
\begin{theorem}\textbf{(The Elimination Theorem)}
If $G$ is a Gr\"obner bases for $I$ with respect to lexicographic order with $x_1>x_2> \cdots >x_n$, then for all $0 \leq k \leq n$, we have
\[
G_k = G \cap K[x_k+1, \dots ,x_n]
\]
is a Gr\"obner bases for the $k-$th elimination ideal.
\end{theorem}
\section{Special Cases}
In this section we will study the \textit{equilibria} and \textit{rigid translation}.
A necessary condition on the vorticities for the existence of rigidly translating solutions is that $\Gamma = \sum \Gamma_{i} = 0$. O'Neil proved that for almost every such choice of vortex strengths, there are exactly $(n-1)!$ rigidly translating configurations \cite{O}. He showed that for almost every choice of vortex strengths satisfying the necessary conditions $L = \sum \Gamma_{i}\Gamma_{j} =0$, there are $(n-2)!$ equilibria.
\noindent
In the four- vortex case, the equilibria can be found explicitly. Since the equations are invariant under translations, set $z_{3} = (1, 0)$, and $z_{4} = (0,0)$. The solutions for $(z_{1}, z_{2})$ are
\[
z_{1} = \frac{(2\Gamma_{4} + \Gamma_{2}, \pm \Gamma_{2}\sqrt{3})}{2(\Gamma_{2} + \Gamma_{3} + \Gamma_{4})}, \quad z_{2} = \frac{(2\Gamma_{4} + \Gamma_{1}, \mp \Gamma_{1}\sqrt{3})}{2(\Gamma_{1} + \Gamma_{3} + \Gamma_{4})}.
\]
In our case, we suppose $\Gamma_{1} = \Gamma_{2} = \Gamma_{3} = 1$, $L=0$ when $\Gamma_{4} = -1$. In this case, two solutions $(z_{1}, z_{2}, z_{3}, z_{4}) $ of relative equilibria are
\[
\left(\frac{(- 1, \sqrt{3})}{2}, \frac{(- 1, -\sqrt{3})}{2}, 1, 0\right), \quad \left(\frac{(- 1, - \sqrt{3})}{2}, \frac{(-1, \sqrt{3})}{2}, 1, 0\right).
\]
\noindent
The configuration is an equilateral triangle with $\Gamma_{4}$ in the convex hull formed by the other vorticities.
\medskip
Now, for rigid translations we have that $\Gamma = 0$ when $\Gamma_{3} = -3$. Following \cite{HM}, for relative equilibria we use the equations
\[
S_{1} = S_{2} = S_{3} = S_{4} = s_{0},
\]
and
\begin{equation}\label{17}
\frac{1}{s_{12}} + \frac{1}{s_{34}} = \frac{1}{s_{13}} + \frac{1}{s_{24}} = \frac{1}{s_{14}} + \frac{1}{s_{23}},
\end{equation}
where $S_{i} - \Gamma_{j} r_{ij}^{2}+\Gamma_{k}r_{ik}^{2} +\Gamma_{l} r_{il}^{2}$ and $I = \sum_{i=1}^{n}\|z_{i}-c\|^{2} = s_{0}$. We clear denominators in the equations (\ref{17}) to get a polynomial system. There are two types of symmetric relative equilibria. The first is the equilateral triangle with vortex 4 at its center. The second type is a concave kite with the three equal vorticities on the exterior isosceles triangle.
\section{Collinear Relative Equilibria}
Collinear relative equilibria of the four-vortex problem can be studied directly from equation (\ref{re}) since in this case it reduces to
\begin{equation}\label{colineal}
\lambda(x_{i} - c) = \sum_{i\neq j} \frac{\Gamma_{i}}{x_{i} - x_{j}},
\end{equation}
where $c \in \mathbb{R}$. Clearing denominators from these equations yields a polynomial system. Rather than fixing $\lambda$ or $c$, we use the homogeneity and translation invariance of the system and set $x_{3} = -1, x_{4} = 1, \Gamma_{1} = \Gamma_{2} = \Gamma_{3} = 1$ and $\Gamma_{4}$ will be treated as a parameter.
\medskip
Using this approach we get the following results:
\subsection{Symmetric Solutions}
Given our setup, symmetric configurations correspond to solutions where $x_1=-x_2$. In this case the center of vorticity $c$ is located at the origin. Substituting these values in (\ref{colineal}) we get the following equation system:
\begin{eqnarray*}
-2\lambda x_2^4+(2\lambda + 2\Gamma_4 +3)x_2^2 + 2(1-\Gamma_4)x_2 &=& 1,\\
2\lambda x_2^4-(2\lambda + 2\Gamma_4 +3)x_2^2 + 2(1-\Gamma_4)x_2 &=& -1,\\
(\Gamma_4-2\lambda)x_2^2 &=& \Gamma_4 +4 -2\lambda,\\
(2\lambda-1)x_2^2 &=&2\lambda-5.
\end{eqnarray*}
\noindent
Solving the system we get that the only solutions are possible when all vorticities are equals
\[
-2\lambda x_2^4 + (2\lambda+2\Gamma_4+3)x_2^2-1 = 0, \quad \Gamma_4=1, \; \text{and} \; \lambda = \frac{x_2^2-5}{x_2^2-1}.
\]
Using $y=x_2^{2}$ we have four real solutions for $(x_1, x_2)=$
\[
(-\sqrt{3} \mp \sqrt{2}, \sqrt{3} \pm \sqrt{2}) \quad \text{and} \quad (\sqrt{3} \pm \sqrt{2}, -\sqrt{3} \mp \sqrt{2}).
\]
\subsection{Asymmetric solutions}
To locate any asymmetric solutions, we introduce the variables $u$ and $v$ along with the equations
\[
u(x_1+x_2)-1 \quad \text{and} \quad v(x_1-x_2)-1.
\]
\noindent
Adding these two equations to the original polynomial system obtained from (\ref{colineal}), we compute a Gr\"obner basis $G_{col}$ with respect to the lex order where $c>\lambda>u>v>x_1>x_2>\Gamma_4$. We get a 12th-degree polynomial in $x_2$ with coefficients in $\Gamma_4$.
\begin{equation*}
\begin{split}
p(x_{2}) &= (13\Gamma^{4}+32\Gamma^{3}+2\Gamma^{5}+4+20
\Gamma +37\Gamma^{2}) {x_2}^{12}+ (-32\,{\Gamma }^{4}+20\,{\Gamma }^{2}+38\,\Gamma \\
& \quad -30{\Gamma }^{3} +12-8\,{\Gamma }^{5}) { x_2}^{11}- (1836\,{\Gamma }^{2}+312+
1250\,\Gamma +1204\,{\Gamma }^{3}+338\,{\Gamma }^{4}\\
& \quad +28\,{\Gamma }^{5}) {x_2}^{10} + (664\,{\Gamma }^{4}-1346\,\Gamma +88\,
{\Gamma }^{5}+1234\,{\Gamma }^{3}+100\,{\Gamma }^{2}-740) {x_2}^{9}\\
& \quad + (3007\,{\Gamma }^{4}+13688\,{\Gamma }^{3} +26937\,{\Gamma }^{2}+254\,{\Gamma }^{5}+23290\,\Gamma +7020) {x_2}^{8}\\
& \quad + (-8636\,{\Gamma }^{3}-272\,{\Gamma }^{5}-5640\,{\Gamma }^{2}+8492\,\Gamma +8712-2656\,{\Gamma }^{4}) {x_2}^{7}\\
& \quad + (1288\,{\Gamma }^{5}+62688+156484\,\Gamma +145312\,{\Gamma }^{2} -14092\,{\Gamma }^{4}-62936\,{\Gamma }^{3}) {x_2}^{6}\\
& \quad+(-6092\,{\Gamma }^{3}+6476\,\Gamma -656\,{\Gamma }^{5}+9528\,{\Gamma }^{2}-4144\,{\Gamma }^{4} -5112) {x_2}^{5} \\
& \quad + (114080\,{\Gamma }^{3}+261207\,{\Gamma }^{2}+334552\,\Gamma +24859\,{\Gamma }^{4}+2078\,{\Gamma }^{5}+166860) {x_2}^{4}\\
& \quad+(-93138\,\Gamma +3864\,{\Gamma }^{5}-112340+91162\,{\Gamma }^{3}+32192\,{\Gamma }^{4}+78260\,{\Gamma }^{2}) {x_2}^{3}\\
& \quad + (2916\,{\Gamma }^{5}+13886\,{\Gamma }^{4}+9048+2348\,{\Gamma }^{3}-66484\,{\Gamma }^{2}-49626\,\Gamma) {x_2}^{2}\\
& \quad +(-12102\,{\Gamma }^{3} +7068+1080\,{\Gamma }^{5}+600\,{\Gamma }^{4}+9846\,\Gamma -6492\,{\Gamma }^{2}) {x_2}\\
& \quad-1148+162\,{\Gamma }^{5}+1050\,\Gamma -472\,{\Gamma }^{3}+1227\,{\Gamma }^{2}-711\,{\Gamma }^{4}.
\end{split}
\end{equation*}
\begin{theorem}
Let the 4-vortex problem with $\Gamma_{1} = \Gamma_{2} = \Gamma_{3} = 1$ and $\Gamma_{4}$ taken as a parameter. In addition, we fix $x_{3} = -1$ y $x_{4} = 1$. There are collinear relative equilibria for every $\Gamma_{4} \in (-1, +\infty)$:
\begin{itemize}
\item If $\Gamma_{4} \in (-1, -1/2)$, we have six relative equilibria configurations.
\item If $\Gamma_{4} = -1/2$, we have seven collinear solutions.
\item If $\Gamma_{4} \in (-1/2, 1)$, there are twelve collinear solutions.
\item If $\Gamma_4 = 1$, there are eight collinear solutions.
\item If $\Gamma_4 \in (1, +\infty)$, there are twelve collinear solutions.
\end{itemize}
\end{theorem}
Using \texttt{Mathematica}, we can find the $\Gamma_4$ values for which we have changes of sign in the polynomial $p(x_2)$. By Descartes' rule of signs, we can find how many real roots can take $p(x_2)$ and $p(-x_2)$. Numerically we can use \textsl{Sturm}'s theorem to count the exact number of real roots, so that we obtain: For $\Gamma \in [-\infty, -1]$, there are no real solutions other than the degenerate $x_{2} = \pm 1$ (i.e., the second vortex coincides with the fourth vortex).
For $\Gamma \in (-1, -1/2)$, there are six different roots for $p(x_{2})$. For $\Gamma = -1/2$ there are seven different roots, and when $\Gamma \in (-1/2, +\infty]$, $p(x_{2})$ has twelve roots. In Fig. \ref{collineal} we can see all the collinear relative equilibria for $\Gamma_{4} = 1/2$.
\begin{figure}[h]
\centering
\includegraphics[scale=0.4]{colineal-01.eps}
\vspace{-5 cm}
\caption{\label{collineal}{\small Set of collinear solutions for $\Gamma_{4} = 1/2$. Vortices $\Gamma_{1}=\Gamma_{2}=\Gamma_{3}=1$ are detonated by white disks and vortex $\Gamma_{4}$ by dark one. In the first line we can see the same order for the three configurations but the distance among the vortices is different. The same applies to the following three lines}}
\end{figure}
\section{Symmetric Strictly Planar Relative Equilibria}
In this section we investigate all possible symmetric planar relative equilibria in the four-vortex problem with $\Gamma_1=\Gamma_2=\Gamma_3=1$. The two possible configurations are a concave kite, and a convex kite. We use the techniques used in \cite{EJ} for the 4 body problem.
\subsection{The kite family}
We say that a planar relative equilibria has a kite shape if it has an axis of symmetry passing through two of the vorticities. The kite configuration is \textit{convex} if none of the vortices is located in the interior of the convex hull of the other three, otherwise, if the configuration is not collinear, we say that the kite configuration is \textit{concave}.
\begin{figure}[h]
\centering
\includegraphics[scale=0.3]{graficoscometa.eps}
\caption{\label{cometa}{\small Kite relative equilibria: (a) $k,l>0$ (left), \; (b) $k<0, l>0$ (center), \; (c) $k>0, l<0$ (right).}}\label{Kite}
\end{figure}
In order to study these kite central configurations we choose the axes of coordinates with origin at the center of mass of $\Gamma_3$ and $\Gamma_4$, the $y-$axis as the axis of symmetry, and the $x-$axis orthogonal to it. Taking conveniently the unity of length, we can suppose that the positions of the masses $\Gamma_1, \Gamma_2,
\Gamma_3$ and $\Gamma_4$ are $(-1, 0), (1, 0), (0, -k)$ and $(0, l)$ respectively, and that always $\Gamma_4$ is over $\Gamma_3$ on the $y-$axis; i.e. $k + l > 0$. See Figure \ref{Kite}.
Using the symmetries of the kite configuration, Dziobek's equations reduce to the following two equations
\begin{eqnarray}\label{convex}
\Gamma_{4}(k+l)[(1+l^{2})^{-1} - (k+l)^{-2}] + 2k[4^{-1} - (1+k^{2})^{-1}] & = & 0, \nonumber \\
f(k, l) = (k+l)[(1+k^{2})^{-1} - (k+l)^{-2} ] + 2l[4^{-1} - (1+l^{2})^{-1} ] & = & 0,
\end{eqnarray}
\noindent
Solutions $(k, l)$ of equations (\ref{convex}) with $k, l \in \mathbb{R}$, and $k + l > 0$ provide the planar non-collinear relative equilibria of the 4-vortex problem for the vorticities $\Gamma_4 \in \mathbb{R}$ and $\Gamma_{1} = \Gamma_{2} = \Gamma_{3} = 1$. From the first equation of (\ref{convex}), $k$ cannot be zero; and from the second one $l$ also cannot be zero.
\begin{proposition}\label{equilateral}
The equilateral triangle, with the three vorticities equal to 1
on its vertices and the vorticity $\Gamma_4$ at its barycenter, always is a relative equilibria of the planar 4-vortex problem.
\end{proposition}
\medskip
\begin{proof} Since the triangle is equilateral, we have that $r_{12} = r_{13} = r_{23} = 2$, so $k = \sqrt{3}$, substituting the $k$ value we have that $l = -1/\sqrt{3}$. \qquad
\end{proof}
The first equation of (\ref{convex}) can be written as
\[
\Gamma_{4} = \Gamma_{4}(k,l) = \frac{2k[(1+k^{2})^{-1}-4^{-1}]}{(k+l)[(1+l^{2})^{-1}-(k+l)^{-2}]} = \frac{k(3-k^{2})(1+l^{2})(k+l)}{2(1+k^{2})(k^{2}+2kl-1)}.
\]
\noindent
then we can write system (\ref{convex}) as the system
\begin{equation}\label{systems}
\left\{
\begin{array}{ll}
\Gamma_{4} = \Gamma_{4}(k, l), \\
f(k, l) = 0.
\end{array}
\right.
\end{equation}
\begin{figure}[h
\centering
\includegraphics[scale=0.5]{fyk-01.eps}
\caption{{\small Arcs of $f=0$ where $\Gamma_{4}(k,l)>0$ and $\Gamma_{4}(k, l)<0$ (black curves). The dashed curve is $l=(1-k^2)/2k$ }}\label{fyk}
\end{figure}
Using \textit{Maple} we plot the curve $f=0$, from here we see that this curve has only two branches in the region $k+l>0$; one contained in the half-plane $l>0$, and the other one contained in the fourth quadrant $\{(k,l):k>0, l<0\}$; see Figure \ref{fyk}.
Using the expression of $\Gamma_{4}(k,l)$ it follows that
\[
\text{sign}(\Gamma_{4}) = \text{sign}\left(\frac{k(3-k^{2})}{k^{2}+2kl-1}\right).
\]
We note that $k(3-k^{2})$ is positive in $(-\infty, -\sqrt{3}) \cup (0, \sqrt{3})$; is zero in $\{-\sqrt{3}, 0, \sqrt{3}\}$; and negative in the complement. The intersection of the vertical line $k=\sqrt{3}$ with the curve $f=0$ provides the three points:
\[
P_{1} =(\sqrt{3}, 1\text{.}19175),\quad P_{5} = (\sqrt{3}, -0\text{.}17633), \quad P_{6} = (\sqrt{3}, -1/\sqrt{3}).
\]
\noindent
While the intersection of $k=0$ with $f=0$ consists of only the point $P_{4} = (0, 1\text{.}2072)$. Finally, the intersection of $k = -\sqrt{3}$ with the curve $f=0$ provides a unique point $P_{3} = (-\sqrt{3}, 2\text{.}74748)$ in the region $k+l>0$.
\smallskip
The curve $l(k) = \dfrac{1-k^{2}}{2k}$ only has points with $k>0$ in $k+l>0$, this curve is monotone decreasing and has a unique branch which intersects the curve $f=0$ at the points (see Figure \ref{fyk}).
\begin{eqnarray*}
P_{2} & = & (k(l_{1}), l_{1} = 1), \\
P_{6} & = & (\sqrt{3}, -1/\sqrt{3}), \\
P_{7} & = & (k(l_{2}), l_{2} = -1 ),
\end{eqnarray*}
\noindent
Then, $\Gamma_{4}>0$ if:
\begin{itemize}
\item $l > -1/\sqrt{3}$ and ($-\sqrt{3} < k < 0$ or $-l + \sqrt{1+l^{2}} < k < \sqrt{3}$), or
\item $l < -1/\sqrt{3}$ and $\sqrt{3} < k< -l + \sqrt{1+l^{2}}$.
\end{itemize}
\noindent
And $\Gamma_{4}<0$ if:
\begin{itemize}
\item $l > -1/\sqrt{3}$ y $0 < k < -l + \sqrt{1+l^{2}}$, y
\item $l < -1/\sqrt{3}$ y $0 < k <\sqrt{3}$.
\end{itemize}
\bigskip
\begin{theorem}
The planar 4-vortex problem with three masses equal to 1 and the fourth one equal to $\Gamma_{4} \in \mathbb{R}$ has exactly one convex central configuration which is kite. Moreover, depending on the value of $\Gamma_{4}$, it has 1, 2, 3 or 4 kite concave relative equilibria.
\begin{enumerate}
\item For all $\Gamma_{4} \in \mathbb{R}$ the equilateral triangle with the three equal vorticities located on its vertices and the vorticity $\Gamma_{4}$ located at its barycentre is always a kite concave relative equilibrium. In the remainder of the statements we omit the description of this concave central configuration.
\item If $\Gamma_{4}=0$ there are exactly 2 additional kite concave relative equilibria configurations. In one of them, $\Gamma_{4}$ is on a vertex of an isosceles triangle and in the other $\Gamma_{4}$ is in its interior.
\item If $0 < \Gamma_{4} < 1$, there are 3 additional kite concave relative equilibria configurations. In two of them $\Gamma_4$ is on a vertex of an isosceles triangle and in the other $\Gamma_{4}$ is in its interior.
\item If $\Gamma_{4} = 1$, then in the convex configuration the four vorticities are located at the vertices of a square, and there is exactly 1 additional kite concave configuration where $\Gamma_{4}$ is on a vertex of an equilateral triangle
\item If $\Gamma_{4} > 1$, there is exactly 1 additional kite concave configuration having $\Gamma_{4}$ in the interior of an isosceles triangle.
\item If $\Gamma_{4} < 0$, there is exactly 1 additional kite concave configuration having $\Gamma_{4}$ in the interior of an isosceles triangle.
\end{enumerate}
\end{theorem}
\begin{proof}
After intersecting the $\Gamma_4$ function with $f$, we get the following arcs:
\begin{itemize}
\item [(i)] The open arc $\gamma_1$ going from $P_1$ to $P_2$. We observe that on this arc $k$ and $l$ are positive, so the corresponding relative equilibria associated to the points $(k,l)$ of this arc are convex. Since $\Gamma_4(k,l)$ takes the value zero on $P_2$ and takes the value $+\infty$ on $P_1$, there is at least one convex central configuration for every value of $\Gamma_4>0$.
\item [(ii)] The open arc $\gamma_2$ going from $P_2$ to $P_4$. Since on this arc $k$ and $l$ are positive, the corresponding relative equilibria associated to points $(k,l)$ of this arc are convex. Since $\Gamma_4$ takes the value zero on $P_4$ and the value $-\infty$ on $P_2$, there is at least one convex central configuration for every value of $\Gamma_4<0$.
\item [(iii)] The open arc $\gamma_3$ going from $P_3$ to $P_4$. Since on this arc $k<0$ and $l>0$, the corresponding relative equilibria associated to points $(k,l)$ of this arc are concave having the vorticity $\Gamma_4$ as a vertex of the triangle formed by the convex hull of the four vorticities. Later, we will show that the $\Gamma_4$ function has a unique critical point on this arc. Therefore, the $\Gamma_4$ value of this point is $1$. In this arc, when $k$ approaches $0$, $l$ tends to $+\infty$ . So the $\Gamma_4(k,l)$ curve cannot intercept the $f=0$ curve or cut at one or two points. As will be seen below, when $\Gamma_4=1$, these curves are located in a single point, i.e., there is an equilateral triangle concave configuration. When $\Gamma_4$ takes values greater than $1$, the two curves do not intersect, and when $\Gamma_4 <1$, the curves intersect in two points, on those points we have isosceles triangle concave configurations.
\item[(iv)] The open arc $\gamma_4$ going from $P_5$ to $P_7$. Since on this arc $k>0$ and $l<0$, the corresponding relative equilibria associated to points $(k,l)$ of this arc are concave having the vorticity $\Gamma_4$ in the interior of the triangle formed by the convex hull of the other three vorticities. As $\Gamma_4(k,l)$ takes the value $0$ on $P_5$ and takes the value $+\infty$ on $P_7$ , there is at least one concave configurations for every $\Gamma_4 \in \mathbb{R}$ value. We remark that $\Gamma_4(k,l)$ on the point $P_6$ takes the value $\Gamma_4=1$ and this point correspond to equilateral triangle of Proposition \ref{equilateral}.
\end{itemize}
We now discuss the case $\Gamma_4=0$, since Dziobek equations are valid for $\Gamma_4 \neq 0$ this case must be studied directly from the equations (\ref{re}) where the center of vorticity is fixed at $c = -k/3$. So we have that:
\begin{eqnarray*}
\lambda(-1, k/3) &=& \frac{(-3-k^{2}, 2k)}{2(1+k^{2})}, \\
\lambda(1, k/3) &=& \frac{(3+k^{2}, 2k)}{2(1+k^{2})}, \\
\lambda(0, -2k/3) &=& \frac{(0, -2k)}{1+k^{2}}, \\
\lambda(0, l+k/3) &=& \frac{(0, 3l^{2}+2kl+1)}{(1+l^{2})(k+l)}.
\end{eqnarray*}
\noindent
From there we have three independent equations
\begin{equation*}
\lambda = \frac{3+k^{2}}{2(1+k^{2})}, \quad \lambda = \frac{3}{1+k^{2}}, \quad \lambda = \frac{3(3l^{2}+2kl+1)}{(3l+k)(1+l^{2})(k+l)}.
\end{equation*}
\noindent
Solving this system, we obtain the following $(k,l)$ solutions, which satisfy $k+l>0$:
\[
(\sqrt{3}, 1\text{.}19175), \; (\sqrt{3}, -0\text{.}176327), \; (\sqrt{3}, -1/\sqrt{3}) \; \text{y} \; (-\sqrt{3}, 2\text{.}74748),
\]
and correspond to a convex configuration, one of this is an isosceles triangle with $\Gamma_4$ in the interior of the convex hull formed by the other three vorticities, the other one is the equilateral triangle given in Proposition \ref{equilateral} and the last one is an isosceles triangle configuration with $\Gamma_4$ as a vertex of the triangle. \qquad
\end{proof}
\medskip
Now we will look for the extremals of the function $\Gamma_ {4} (k, l)$ along the open arc of the curve $f (k, l) = 0$. These points correspond to the bifurcation values on the number of relative equilibria. For this, we use the method of Lagrange multipliers.
\smallskip
\begin{proposition}
The function $\Gamma_4(k,l)$ restricted to the arcs $\gamma_1 \cup \gamma_2 \cup \gamma_3 \cup \gamma_4$ has a unique critical point on the arc $\gamma_3$. The value of $\Gamma_4(k,l)$ at this point is $\Gamma_4=1$.
\end{proposition}
\smallskip
\begin{proof} Let $F=\Gamma_4(k,l) + \lambda f(k,l)$, where $\lambda$ is a Lagrange multiplier. We look for the extremals of the function $\Gamma_4(k,l)$ on the arcs $\gamma_1 \cup \gamma_2 \cup \gamma_3 \cup \gamma_4$. We first look for the extremals of this function on the curve $f(k,l)=0$, and after we must choose the extremals on these arcs. So, we must solve the system
\begin{equation}\label{gradiente}
\frac{\partial F}{\partial k} = 0, \quad \frac{\partial F}{\partial l} = 0, \quad f(k,l) = 0,
\end{equation}
of three equations in the three variables $k,l$ and $\lambda$. First, we eliminate the variable $\lambda$ using the first two equations of (\ref{gradiente}), obtaining a system of the form
\begin{equation}\label{h}
h(k,l) = 0, \quad f(k,l)=0,
\end{equation}
where
\[
h(k,l) = (1+l^{2})(k+l)^{4}h_{1}(k,l), \qquad f(k,l) = f_{1}(k,l) + 3l^{2}(l^{2}-1)-2,
\]
here $h_1$ is a polynomial in the variables $k$ and $l$ of degrees 9 and 6 respectively, and $f_1$ is a polynomial in the variables $k$ and $l$ of degrees 3 and 4 respectively,
\begin{equation*}
\begin{split}
h_{1}(k,l) & = 6+36\,{l}^{4}-24\,{k}^{2}+9\,{l}^{6}-20\,{k}^{4}+24\,{k}^{6}-18\,{k}^{8}+9\,{l}^{2}+30\,{l}^{5}{k}^{5}\\
\quad &+ 304\,{k}^{4}{l}^{4} +84\,{l}^{3}{k}^{5}-3\,{k}^{9}l-50\,{l}^{6}{k}^{4}+{k}^{9}{l}^{5}+84\,{k}^{6}{l}^{6}+32\,{l}^{7}{k}^{5}+200\,{k}^{7}{l}^{3}\\
\quad & +208\,{k}^{6}{l}^{4}+76\,{k}^{7}{l}^{5}-27\,kl+28\,{k}^{8}{l}^{4}+6\,{l}^{3}{k}^{9}+{k}^{8}{l}^{6}+12\,l{k}^{3}+104\,{l}^{3}{k}^{3}\\
\quad & +9\,k{l}^{5}+126\,{k}^{4}{l}^{2}
+36\,{l}^{6}{k}^{2}+54\,k{l}^{3}+36\,{l}^{2}{k}^{2}+49\,{k}^{8}{l}^{2}+252\,{k}^{3}{l}^{5}\\
\quad & -100\,{k}^{7}l+102\,{k}^{5}l-140\,{k}^{6}{l}^{2} ,
\end{split}
\end{equation*}
and
\begin{equation*}
f_{1}(k,l) = kl(k(l^{2}+3)(k+l)+1+5l^{2}).
\end{equation*}
In order to find the solution of system (\ref{h}), we will use the resultant of two polynomials. The resultant Res$[h_1, f_1, k]$ si
\[
6144 l(l^2-3)(l^2+1)^12(3l^2-1)^2 r(l).
\]
with $r$ a polynomial in the variable $l$ of degree 20,
\begin{eqnarray*}
r(l) &=& 36\,{l}^{20}+315\,{l}^{18}-
2457\,{l}^{16}-1776\,{l}^{14}+33264\,{l}^{12}-61986\,{l}^{10}+51534\,{
l}^{8}\\
\quad &-&18904\,{l}^{6}+324\,{l}^{4}+1455\,{l}^{2}+243.
\end{eqnarray*}
\noindent
When equalling to zero the resultant of these two polynomials, the only real roots that we get are $l=\pm \sqrt{3}$ and $\pm1/\sqrt{3}$. For each real root of $l$, we compute numerically the real roots for $f_1$ and check if the point $(k,l)$ is a solution of the system (\ref{gradiente}) with $\lambda \neq 0$. After that, we have that the only point that is solution to the system (\ref{gradiente}) on $\gamma_1 \cup \gamma_2 \cup \gamma_3 \cup \gamma_4$ with $\lambda \neq 0$ is
\[
(k, l) = (-1/\sqrt{3}, \sqrt{3}),
\]
which is equivalent to an equilateral triangle configuration, i.e., at this point we have a bifurcation since the number of relative equilibria changes. Computing the Hessian of $\Gamma_4(k,l)$ we get that this point is a minimum on the arc $\gamma_3$, and the $\Gamma_4(k, l)$ value at this point is $1$. \qquad
\end{proof}
\bigskip
Counting the different positions for the three equal vortices in configuration of relative equilibria described previously and adding the 4 classes of collinear relative equilibria we obtain the number of planar relative equilibria in kite configuration. The result is summarizing as:
\medskip
\begin{corollary}
The planar 4-vortex problem with three vorticities equal to 1 and the fourth one equal to $\Gamma_4$ has the following classes of kite relative equilibria:
\begin{itemize}
\item[1.] 26 \, if \, $\Gamma_4=0$,
\item[2.] 29 \, if \, $0<\Gamma_4<1$,
\item[3.] 34 \, if \, $\Gamma_4=1$,
\item[4.] 23 \, if \, $\Gamma_4>1$,
\item[5.] 20 \, if \, $-1/2<\Gamma_4<0$,
\item[6.] 15 \, if \, $\Gamma_4 = -1/2$,
\item[7.] 14 \, if \, $-1 <\Gamma_4<-1/2$,
\item[8.] 8 \, if \, $\Gamma_4 < -1$.
\end{itemize}
\end{corollary}
\begin{proof}
The idea is to count the number and type of solutions (equivalence classes) for different values of $\Gamma_4$.
When all vorticities are equals ($\Gamma_4 =1$) we have $6$ square configurations, $8$ configurations of equilateral triangle where $\Gamma_4$ is an interior vortex (that is $\Gamma_4$ is in the interior of the convex hull of the other three), 8 configurations of equilateral triangle where $\Gamma_4$ is an exterior vortex (it is located at one vortex of the equilateral triangle) and 12 collinear solutions.
\noindent
If $\Gamma_4 = 0$, we have 12 collinear configurations, 6 convex configurations, 2 configurations of equilateral triangle with interior vortex, 3 isosceles configurations with interior vortex and 3 isosceles configurations with exterior vortex.
\noindent
If $0 < \Gamma_4 < 1$ we have 12 collinear configurations, 6 convex configurations, 2 configurations of equilateral triangle with the different vorticity inside, 6 relative equilibria where $\Gamma_4$ is on a vertex of an isosceles triangle and 3 relative equilibria where $\Gamma_4$ is in the interior of an isosceles triangle.
\noindent
If $\Gamma_4>1$ we have 12 collinear configurations, 6 convex configurations, 2 configurations of equilateral triangle with the different vorticity inside, and 3 relative equilibria where $\Gamma_4$ is in the interior of an isosceles triangle.
\noindent
If $-1/2 < \Gamma_4 < 0$, we have 12 collinear configurations, 6 convex configurations, and 2 configurations of equilateral triangle with the different vorticity inside.
\noindent
If $\Gamma_4=-1/2$, there are 7 collinear relative equilibria, 6 convex relative equilibria, and 3 relative equilibria of equilateral triangle.
\noindent
If $-1 < \Gamma_4 < -1/2$, there are 6 collinear relative equilibria, 6 convex relative equilibria, and 3 relative equilibria of equilateral triangle. \qquad
\end{proof}
\subsubsection{Rhombus configuration}
Now, we will study a special case of convex kite configuration, the rhombus. The special thing about this configuration is that, because of the two lines of symmetry, we can find an explicit expression for the value of the sides which depend on different vorticities.
One of the properties of the rhombus is that all four sides are congruent, so $r_{13} = r_{14} = r_{23} = r_{24}$ where the diagonals satisfy the relation $4r_{13}^2 = r_{12}^2+r_{34}^2$. We fix the areas orientation as $A_1=A_2=-A_3=-A_4$ (since $A_1+A_2+A_3+A_4$). Let $x=r_{34}/r_{12}$ be the ratio between the diagonals of the rhombus. From the first and second equation of (\ref{Dziobek}) we have that $x^2=1$, i.e., the diagonal are equal, so the configuration is a square with all the vorticities equal. From the third and fifth equation in (\ref{Dziobek}) we have an expression for $\Gamma_4$ given by
\begin{equation}\label{rombo}
\Gamma_4 = \frac{x^2-3}{1-3x^2}.
\end{equation}
$\Gamma_4$ is positive when $1/\sqrt{3} < x < \sqrt{3}$, and negative when $x>\sqrt{3}$ or $0<x<1/\sqrt{3}$ (see Fig. \ref{fyg}). Solving equation (\ref{rombo}) for $x^2$, yields two different rhombus families. One when $\Gamma_4 \in (-1/3, +\infty)$, and the other one when $\Gamma_4<3$. This can be seen by inverting equation (\ref{rombo}), which yields
\begin{equation}\label{35}
x^{2} = \frac{\Gamma_{4}+3}{3\Gamma_{4} + 1}.
\end{equation}
\begin{figure}
\centering
\includegraphics[scale=0.4]{rombo.eps}
\caption{{\small Plot of $x$ vs $\Gamma_{4}$, where $x=r_{34}/r_{12}$ is the ratio between the diagonals of the rhombus}}\label{fyg}
\end{figure}
Since $\lambda = \dfrac{-L}{2I}$, we can calculate the angular velocity ,
\begin{equation}\label{36}
\lambda =- \frac{3(1+\Gamma_{4})}{r_{12}^{2}},
\end{equation}
\noindent
The numerator vanishes at $\Gamma_{4} = -1$, in this case $x^{2} < 0$, i.e., it is not possible to have absolute equilibria when the vortices are in a rhombus configuration. The value of the angular velocity $\lambda$ is negative when $\Gamma_{4} > -1$, and $\lambda$ is positive when $\Gamma_{4} < -1$. We summarize our conclusions in the following theorem:
\medskip
\begin{theorem}
There are two one-parameter families of rhombus relative equilibria with vortex strengths $\Gamma_{1} = \Gamma_{2} = \Gamma_{3} = 1$. The vortices 1 and 2 lie on opposite sides of each other, as do vortices 3 and 4. The mutual distances are given by
\begin{equation}
\left(\frac{r_{34}}{r_{12}}\right)^{2} = \frac{\Gamma_{4} + 3}{3\Gamma_{4}+1} \qquad \text{and} \qquad \left(\frac{r_{13}}{r_{12}}\right)^{2} = \Gamma_{4} + 1,
\end{equation}
describing two distinct solutions. For $\Gamma_{4} \in(-\infty, -3)$, we have a solution that has $\lambda >0$. The other solution is when $\Gamma_{4} \in (-1/3, 0)$ that has $\lambda < 0$. The case $\Gamma_{4} = 1$ reduces to square. For $\Gamma_{4} >0$, the larger vortex lies on the shorter diagonal.
\end{theorem}
\begin{proof} The formula for $(r_{13}/r_{12})^{2}$ comes from $1+x^{2} = 4(r_{13}/r_{12})^{2}$.
For the case $\Gamma_{4} > 0$ we get from equation (\ref{rombo}) that $\Gamma_{4} < 1$ if and only if $1 < x < \sqrt{3}$. Beginning with the square at $\Gamma_{4} = 1$, as $\Gamma_{4} \in (0,1)$ , the ratio of the diagonals of the rhombus increases from $1$ to $\sqrt{3}$, so the different vorticity ($\Gamma_{4}$) is located on the longer diagonal, while if $\Gamma_{4}>1$, the ratio of the diagonals of the rhombus is always less than $1$ and decreases as $\Gamma_{4}$ increases. In this case $\Gamma_{4}$ is located on the shorter diagonal.
At $\Gamma_{4} = -3$ a bifurcation occurs, and a new family is born emerging out of a binary collision between vortices $3$ and $4$. This family has the vortex $\Gamma_{4}$ located on the shorter diagonal. \qquad
\end{proof}
\section*{Acknowledgments}
This work has been partially supported by CONACYT-M\'exico, grant 128790 and by a NSERC Discovery grant.
|
\section{Introduction}
\vskip4mm Let $R$ be a ring with an identity. An element $a\in R$
is called nil-clean if there exists an idempotent $e\in R$ such
that $a-e\in R$ is a nilpotent. A ring $R$ is nil-clean provided
that every element in $R$ is nil-clean. In ~\cite[Question 3]{D},
Diesl asked: Let $R$ be a nil clean ring, and let $n$ be a
positive integer. Is $M_n(R)$ nil clean? In ~\cite[Theorem
3]{BGDT}, Breaz et al. proved that their main theorem: for a field
$K$, $M_n(K)$ is nil-clean if and only if $K\cong {\Bbb Z}_2$.
They also asked if this result could be extended to division
rings. As a main result in ~\cite{KLZ}, Ko\c{s}an et al. gave a
positive answer to this problem. They showed that the preceding
equivalence holds for any division ring.
A ring $R$ is abelian if every idempotent in $R$ is central.
Clearly, every division ring is abelian. We extend, in this
article, the main results of Breaz et al.~\cite[Theorem 3]{BGDT}
and that of Ko\c{s}an et al.~\cite[Theorem 3]{KLZ}. We shall prove
that for an abelian ring $R$, $M_n(R)$ is nil-clean if and only if
$R/J(R)$ is Boolean and $M_n(J(R))$ is nil. As a corollary, we
also prove that the converse of a result of Ko\c{s}an et al.'s is
true.
Throughout, all rings are associative with an identity. $M_n(R)$
will denote the ring of all $n\times n$ full matrices over $R$
with an identity $I_n$. $GL_n(R)$ stands for the $n$-dimensional
general liner group of $R$.
\section{The main result}
We begin with several lemmas which will be needed in our proof of
the main result.
\begin{lem}~\cite[Proposition 3.15]{D} \label{21} Let $I$ be a nil ideal of a ring $R$. Then $R$ is nil-clean if and only if
$R/I$ is nil-clean.
\end{lem}
\begin{lem}~\cite[Theorem 3]{KLZ}\label{22} Let $R$ be a division ring. Then the following are equivalent:
\begin{enumerate}
\item [(1)]{\it $R\cong {\Bbb Z}_2$.}
\item [(2)]{\it $M_n(R)$ is nil-clean for all $n\in {\Bbb N}$.}
\item [(3)]{\it $M_n(R)$ is nil-clean for some $n\in {\Bbb N}$.}
\end{enumerate}
\end{lem}
Recall that a ring $R$ is an exchange ring if for every $a\in R$
there exists an idempotent $e\in aR$ such that $1-e\in (1-a)R$.
Clearly, every nil-clean ring is an exchange ring.
\begin{lem} \label{23} Let $R$ be an
abelian exchange ring, and let $x\in R$. Then $RxR=R$ if and only
if $x\in U(R)$.\end{lem} \begin{proof} If $x\in U(R)$, then
$RxR=R$. Conversely, assume that $RxR=R$. As in the proof of
~\cite[Proposition 17.1.9]{CH}, there exists an idempotent $e\in
R$ such that $e\in xR$ such that $ReR=R$. This implies that $e=1$.
Write $xy=1$. Then $yx=y(xy)x=(yx)^2$. Hence, $yx=y(yx)x$.
Therefore $1=x(yx)y=xy(yx)xy=yx$, and so $x\in U(R)$. This
completes the proof.\end{proof}
Set
$$J^*(R)=\bigcap \{ P~|~P~\mbox{is a maximal ideal of}~R\}.$$ We
will see that $J(R)\subseteq J^*(R)$. In general, they are not the
same. For instance, $J(R)=0$ and $J^*(R)=\{ x\in
R~|~dim_F(xV)<\infty\}$, where $R=End_F(V)$ and $V$ is an
infinite-dimensional vector space over a field $F$.
\begin{lem} \label{24} Let $R$ be an
abelian exchange ring. Then $J^*(R)=J(R)$.\end{lem} \begin{proof}
Let $M$ be a maximal ideal of $R$. If $J(R)\nsubseteq M$, then
$J(R)+M=R$. Write $x+y=1$ with $x\in J(R),y\in M$. Then $y=1-x\in
U(R)$, an absurd. Hence, $J(R)\subseteq M$. This implies that
$J(R)\subseteq J^*(R)$. Let $x\in J^*(R)$, and let $r\in R$. If
$R(1-xr)R\neq R$, then we can find a maximal ideal $M$ of $R$ such
that $R(1-xr)R\subseteq M$, and so $1-xr\in M$. It follows that
$1=xr+(1-xr)\in M$, which is imposable. Therefore $R(1-xr)R=R$. In
light of Lemma~\ref{23}, $1-xr\in U(R)$, and then $x\in J(R)$.
This completes the proof.\end{proof}
A ring $R$ is local if $R$ has only maximal right ideal. As is
well know, a ring $R$ is local if and only if for every $a\in R$,
either $a$ or $1-a$ is invertible if and only $R/J(R)$ is a
division ring.
\begin{lem}\label{25} Let $R$ be a ring with no non-trivial idempotents, and let $n\in {\Bbb N}$. Then the following are equivalent:
\begin{enumerate}
\item [(1)]{\it $M_n(R)$ is nil-clean.}
\item [(2)]{\it $R/J(R)\cong {\Bbb Z}_2$ and $M_n(J(R))$ is nil.}
\end{enumerate}
\end{lem}
\begin{proof} $(1)\Rightarrow (2)$ In view of ~\cite[Proposition 3.16]{D}, $J(M_n(R))$ is
nil, and then so is $M_n(J(R))$.
Let $a\in R$. By hypothesis, $M_n(R)$ is nil-clean. If $n=1$, then
$R$ is nil-clean. Then $a\in N(R)$ or $a-1\in N(R)$. This shows
that $a\in U(R)$ or $1-a\in U(R)$, and so $R$ is local. That is,
$R/J(R)$ is a division ring. As $R/J(R)$ is nil-clean, it follows
from Lemma~\ref{22} that $R/J(R)\cong {\Bbb Z}_2$. We now assume
that $n\geq 2$. Then there exists an idempotent $E\in M_n(R)$ and
a nilpotent $W\in GL_n(R)$ such that $I_n+\left(
\begin{array}{cccc}
a&&\\
&0&\\
&&\ddots&\\
&&&0
\end{array}
\right)=E+W$. Set $U=-I_n+W$. Then $U\in GL_n(R)$. Hence,
$$U^{-1}\left(
\begin{array}{cccc}
a&&\\
&0&\\
&&\ddots&\\
&&&0
\end{array}
\right)=U^{-1}E+I_n=\big(U^{-1}EU\big)U^{-1}+I_n.$$ Set
$F=U^{-1}EU$. Then $F=F^2\in M_n(R)$, and that
$$(I_n-F)U^{-1}\left(
\begin{array}{cccc}
a&&\\
&0&\\
&&\ddots&\\
&&&0
\end{array}
\right)=I_n-F.$$ Write $I_n-F=\left(
\begin{array}{cccc}
e&0&\\
*&0&\\
\vdots&&\ddots&\\
*&0&&0
\end{array}
\right).$ As $R$ possesses no non-trivial idempotents, $e=0$ or
$1$. If $e=0$, then $I_n-F=0$, and so $E=I_n$. This shows that
$\left(
\begin{array}{cccc}
a&&\\
&0&\\
&&\ddots&\\
&&&0
\end{array}
\right)=W$ is nilpotent; hence that $a\in R$ is nilpotent. Thus,
$1-a\in U(R)$.
If $e=1$, then $F=\left(
\begin{array}{cccc}
0&0&\\
*&1&\\
\vdots&&\ddots&\\
*&0&&1
\end{array}
\right).$ Write $U^{-1}=\left(
\begin{array}{cc}
\alpha&\beta\\
\gamma&\delta
\end{array}
\right),$ where $\alpha\in R, \beta\in M_{1\times (n-1)}(R),$
$\gamma\in M_{(n-1)\times 1}(R)$ and $\delta\in M_{(n-1)\times
(n-1)}(R)$. Then $$\left(
\begin{array}{cc}
\alpha&\beta\\
\gamma&\delta
\end{array}
\right)\left(
\begin{array}{cccc}
a&&\\
&0&\\
&&\ddots&\\
&&&0
\end{array}
\right)=\left(
\begin{array}{cc}
0&0\\
x&I_{n-1}
\end{array}
\right)\left(
\begin{array}{cc}
\alpha&\beta\\
\gamma&\delta
\end{array}
\right)+I_n,$$ where $x\in M_{(n-1)\times 1}(R)$. Thus, we get
$$\begin{array}{c}
\alpha a=1, \gamma a=x\alpha+\gamma, 0=x\beta+\delta+I_{n-1}.
\end{array}$$ One easily checks that
$$\left(
\begin{array}{cc}
1&\beta\\
0&I_{n-1}
\end{array}
\right)\left(
\begin{array}{cc}
1&0\\
x&I_{n-1}
\end{array}
\right)U^{-1}\left(
\begin{array}{cc}
1&0\\
\gamma a&I_{n-1}
\end{array}
\right)=\left(
\begin{array}{cc}
\alpha+\beta\gamma a&0\\
0&-I_{n-1}
\end{array}
\right).$$ This implies that $u:=\alpha+\beta\gamma a\in U(R)$.
Hence, $\alpha=u-\beta\gamma a$. It follows from $\alpha a=1$ that
$(u-\beta\gamma a)a=1$. As $R$ is connected, we see that
$a(u-\beta\gamma a)=1$, and so $a\in U(R)$. This shows that $a\in
U(R)$ or $1-a\in U(R)$. Therefore $R$ is local, and then $R/J(R)$
is a division ring. Since $M_n(R)$ is nil-clean, we see that so is
$M_n(R/J(R))$. In light of Lemma~\ref{22}, $R/J(R)\cong {\Bbb
Z}_2$, as desired.
$(2)\Rightarrow (1)$ In view of Lemma~\ref{21}, $M_n(R/J(R))$ is
nil-clean. Since $M_n(R)/J(M_n(R))\cong M_n(R/J(R))$ and
$J\big(M_n(R)\big)=M_n(J(R))$ is nil, it follows from
Lemma~\ref{22} that $M_n(R)$ is nil-clean, as asserted.
\end{proof}
\begin{ex} \label{210} Let $K$ be a field, and let $R=K[x,y]/(x,y)^2$. Then $M_n(R)$ is nil-clean if and only if $K\cong {\Bbb Z}_2$.
Clearly, $J(R)=(x,y)/(x,y)^2$, and so $R/J(R)\cong K$. Thus, $R$ is a local ring with a nilpotent Jacobson radical. Hence, $R$ has no non-trivial idempotents.
Thus, we are done by Lemma~\ref{25}.\end{ex}
We come now to our main result.
\begin{thm}\label{26} Let $R$ be abelian, and let $n\in {\Bbb N}$. Then the following are equivalent:
\begin{enumerate}
\item [(1)]{\it $M_n(R)$ is nil-clean.}
\item [(2)]{\it $R/J(R)$ is Boolean and $M_n(J(R))$ is nil.}
\end{enumerate}
\end{thm}
\begin{proof} $(1)\Rightarrow (2)$ Clearly, $M_n(J(R))$ is nil. Let $M$ be a maximal ideal of
$R$, and let $\varphi_M: R\to R/M.$ Since $M_n(R)$ is nil-clean,
then so is $M_n(R/M)$. Hence, $R/M$ is an exchange ring with all
idempotents central. In view of ~\cite[Lemma 17.2.5]{CH}, $R/M$ is
local, and so $R/M$ is connected. In view of Lemma~\ref{25},
$R/M/J(R/M)\cong {\Bbb Z}_2$. Write $J(R/M)=K/M$. Then $K$ is a
maximal ideal of $R$, and that $M\subseteq K$. This implies that
$M=K$; hence, $R/M\cong {\Bbb Z}_2$. Construct a map $\varphi_M:
R/J^*(R)\to R/M, r+J^*(R)\mapsto r+M$. Then
$\bigcap\limits_{M}Ker\varphi_M=\bigcap\limits_{M}\{
r+J^*(R)~|~r\in M\}=0$. Therefore $R/J^*(R)$ is isomorphic to a
subdirect product of some ${\Bbb Z}_2$. Hence, $R/J^*(R)$ is
Boolean. In light of Lemma~\ref{24}, $R/J(R)$ is Boolean, as
desired.
$(2)\Rightarrow (1)$ Since $R/J(R)$ is Boolean, it follows by
~\cite[Corollary 6]{BGDT} that $M_n(R/J(R))$ is nil-clean. That
is, $M_n(R)/J(M_n(R))$ is nil-clean. But $J(M_n(R))=M_n(J(R))$ is
nil. Therefore we complete the proof, by
Lemma~\ref{21}.\end{proof}
We note that the "$(2)\Rightarrow (1)$" in Theorem ~\ref{26} always holds, but "abelian" condition is
necessary in "$(1)\Rightarrow (2)$". Let $R=M_n({\Bbb Z}_2) (n\geq 2)$. Then $R$ is
nil-clean. But $R/J(R)$ is not Boolean. Here, $R$ is not abelian.
\begin{cor}\label{27} Let $R$ be commutative, and let $n\in {\Bbb N}$. Then the following are equivalent:
\begin{enumerate}
\item [(1)]{\it $M_n(R)$ is nil-clean.}
\item [(2)]{\it $R/J(R)$ is Boolean and $J(R)$ is nil.}
\item [(3)]{\it For any $a\in R$, $a-a^2\in R$ is nilpotent.}
\end{enumerate}
\end{cor}
\begin{proof} $(1)\Rightarrow (3)$ Let $a\in R$. In view of Theorem~\ref{26},
$a-a^2\in J(R)$. Since $R$ is commutative, we see that $J(R)$ is
nil if and only if $J(M_n(R))$ is nil. Therefore $a-a^2\in R$ is
nilpotent.
$(3)\Rightarrow (2)$ Clearly, $R/J(R)$ is Boolean. For any $a\in
J(R)$, we have $(a-a^2)^n=0$ for some $n\geq 1$. Hence,
$a^n(1-a)^n=0$, and so $a^n=0$. This implies that $J(R)$ is nil.
$(2)\Rightarrow (1)$ As $R$ is commutative, we see that
$M_n(J(R))$ is nil. This completes the proof, by Theorem~\ref{26}.
\end{proof}
Furthermore, we observe that the converse of ~\cite[Corollary
7]{BGDT} is true as the following shows.
\begin{cor} \label{28} A commutative ring $R$ is nil-clean if and only if $M_n(R)$ is nil-clean.\end{cor} \begin{proof}
One direction is obvious by ~\cite[Corollary 7]{BGDT}. Suppose
that $M_n(R)$ is nil-clean. In view of Corollary~\ref{27} that
$R/J(R)\cong {\Bbb Z}_2$ is nil-clean, and that $J(R)$ is nil.
Therefore $R$ is nil-clean, by Lemma~\ref{21}.\end{proof}
\begin{ex} \label{211} Let $m,n\in {\Bbb N}$. Then $M_n\big({\Bbb Z}_m\big)$ is nil-clean if and only if $m=2^r$ for some $r\in {\Bbb N}$.
Write $m=p_1^{r_1}\cdots p_s^{r_s} (p_1,\cdots ,p_s~\mbox{are distinct primes}, r_1,\cdots ,r_s\in {\Bbb N}$).
Then $Z_m\cong Z_{p_1^{r_1}}\oplus \cdots \oplus Z_{p_m^{r_s}}$. In light of Corollary~\ref{27}, $M_n\big({\Bbb Z}_m\big)$ is nil-clean if and only if $s=1$ and
$Z_{p_1^{r_1}}$ is nil-clean. Therefore we are done
by Lemma~\ref{25}.\end{ex}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.